linux/fs/btrfs/extent-tree.c

7682 lines
201 KiB
C
Raw Normal View History

/*
* Copyright (C) 2007 Oracle. All rights reserved.
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public
* License v2 as published by the Free Software Foundation.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* General Public License for more details.
*
* You should have received a copy of the GNU General Public
* License along with this program; if not, write to the
* Free Software Foundation, Inc., 59 Temple Place - Suite 330,
* Boston, MA 021110-1307, USA.
*/
#include <linux/sched.h>
#include <linux/pagemap.h>
#include <linux/writeback.h>
#include <linux/blkdev.h>
#include <linux/sort.h>
#include <linux/rcupdate.h>
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
#include <linux/kthread.h>
#include "compat.h"
#include "hash.h"
#include "ctree.h"
#include "disk-io.h"
#include "print-tree.h"
#include "transaction.h"
#include "volumes.h"
#include "locking.h"
#include "free-space-cache.h"
Btrfs: batch extent inserts/updates/deletions on the extent root While profiling the allocator I noticed a good amount of time was being spent in finish_current_insert and del_pending_extents, and as the filesystem filled up more and more time was being spent in those functions. This patch aims to try and reduce that problem. This happens two ways 1) track if we tried to delete an extent that we are going to update or insert. Once we get into finish_current_insert we discard any of the extents that were marked for deletion. This saves us from doing unnecessary work almost every time finish_current_insert runs. 2) Batch insertion/updates/deletions. Instead of doing a btrfs_search_slot for each individual extent and doing the needed operation, we instead keep the leaf around and see if there is anything else we can do on that leaf. On the insert case I introduced a btrfs_insert_some_items, which will take an array of keys with an array of data_sizes and try and squeeze in as many of those keys as possible, and then return how many keys it was able to insert. In the update case we search for an extent ref, update the ref and then loop through the leaf to see if any of the other refs we are looking to update are on that leaf, and then once we are done we release the path and search for the next ref we need to update. And finally for the deletion we try and delete the extent+ref in pairs, so we will try to find extent+ref pairs next to the extent we are trying to free and free them in bulk if possible. This along with the other cluster fix that Chris pushed out a bit ago helps make the allocator preform more uniformly as it fills up the disk. There is still a slight drop as we fill up the disk since we start having to stick new blocks in odd places which results in more COW's than on a empty fs, but the drop is not nearly as severe as it was before. Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-11-13 03:19:50 +08:00
static int update_block_group(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 bytenr, u64 num_bytes, int alloc,
int mark_free);
static int update_reserved_extents(struct btrfs_block_group_cache *cache,
u64 num_bytes, int reserve);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 bytenr, u64 num_bytes, u64 parent,
u64 root_objectid, u64 owner_objectid,
u64 owner_offset, int refs_to_drop,
struct btrfs_delayed_extent_op *extra_op);
static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
struct extent_buffer *leaf,
struct btrfs_extent_item *ei);
static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 parent, u64 root_objectid,
u64 flags, u64 owner, u64 offset,
struct btrfs_key *ins, int ref_mod);
static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 parent, u64 root_objectid,
u64 flags, struct btrfs_disk_key *key,
int level, struct btrfs_key *ins);
static int do_chunk_alloc(struct btrfs_trans_handle *trans,
struct btrfs_root *extent_root, u64 alloc_bytes,
u64 flags, int force);
static int pin_down_bytes(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
u64 bytenr, u64 num_bytes,
int is_data, int reserved,
struct extent_buffer **must_clean);
static int find_next_key(struct btrfs_path *path, int level,
struct btrfs_key *key);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
int dump_block_groups);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
static noinline int
block_group_cache_done(struct btrfs_block_group_cache *cache)
{
smp_mb();
return cache->cached == BTRFS_CACHE_FINISHED;
}
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
{
return (cache->flags & bits) == bits;
}
/*
* this adds the block group to the fs_info rb tree for the block group
* cache
*/
static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
struct btrfs_block_group_cache *block_group)
{
struct rb_node **p;
struct rb_node *parent = NULL;
struct btrfs_block_group_cache *cache;
spin_lock(&info->block_group_cache_lock);
p = &info->block_group_cache_tree.rb_node;
while (*p) {
parent = *p;
cache = rb_entry(parent, struct btrfs_block_group_cache,
cache_node);
if (block_group->key.objectid < cache->key.objectid) {
p = &(*p)->rb_left;
} else if (block_group->key.objectid > cache->key.objectid) {
p = &(*p)->rb_right;
} else {
spin_unlock(&info->block_group_cache_lock);
return -EEXIST;
}
}
rb_link_node(&block_group->cache_node, parent, p);
rb_insert_color(&block_group->cache_node,
&info->block_group_cache_tree);
spin_unlock(&info->block_group_cache_lock);
return 0;
}
/*
* This will return the block group at or after bytenr if contains is 0, else
* it will return the block group that contains the bytenr
*/
static struct btrfs_block_group_cache *
block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
int contains)
{
struct btrfs_block_group_cache *cache, *ret = NULL;
struct rb_node *n;
u64 end, start;
spin_lock(&info->block_group_cache_lock);
n = info->block_group_cache_tree.rb_node;
while (n) {
cache = rb_entry(n, struct btrfs_block_group_cache,
cache_node);
end = cache->key.objectid + cache->key.offset - 1;
start = cache->key.objectid;
if (bytenr < start) {
if (!contains && (!ret || start < ret->key.objectid))
ret = cache;
n = n->rb_left;
} else if (bytenr > start) {
if (contains && bytenr <= end) {
ret = cache;
break;
}
n = n->rb_right;
} else {
ret = cache;
break;
}
}
if (ret)
atomic_inc(&ret->count);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
spin_unlock(&info->block_group_cache_lock);
return ret;
}
static int add_excluded_extent(struct btrfs_root *root,
u64 start, u64 num_bytes)
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
{
u64 end = start + num_bytes - 1;
set_extent_bits(&root->fs_info->freed_extents[0],
start, end, EXTENT_UPTODATE, GFP_NOFS);
set_extent_bits(&root->fs_info->freed_extents[1],
start, end, EXTENT_UPTODATE, GFP_NOFS);
return 0;
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
static void free_excluded_extents(struct btrfs_root *root,
struct btrfs_block_group_cache *cache)
{
u64 start, end;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
start = cache->key.objectid;
end = start + cache->key.offset - 1;
clear_extent_bits(&root->fs_info->freed_extents[0],
start, end, EXTENT_UPTODATE, GFP_NOFS);
clear_extent_bits(&root->fs_info->freed_extents[1],
start, end, EXTENT_UPTODATE, GFP_NOFS);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
}
static int exclude_super_stripes(struct btrfs_root *root,
struct btrfs_block_group_cache *cache)
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
{
u64 bytenr;
u64 *logical;
int stripe_len;
int i, nr, ret;
if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
cache->bytes_super += stripe_len;
ret = add_excluded_extent(root, cache->key.objectid,
stripe_len);
BUG_ON(ret);
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
bytenr = btrfs_sb_offset(i);
ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
cache->key.objectid, bytenr,
0, &logical, &nr, &stripe_len);
BUG_ON(ret);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
while (nr--) {
cache->bytes_super += stripe_len;
ret = add_excluded_extent(root, logical[nr],
stripe_len);
BUG_ON(ret);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
kfree(logical);
}
return 0;
}
static struct btrfs_caching_control *
get_caching_control(struct btrfs_block_group_cache *cache)
{
struct btrfs_caching_control *ctl;
spin_lock(&cache->lock);
if (cache->cached != BTRFS_CACHE_STARTED) {
spin_unlock(&cache->lock);
return NULL;
}
ctl = cache->caching_ctl;
atomic_inc(&ctl->count);
spin_unlock(&cache->lock);
return ctl;
}
static void put_caching_control(struct btrfs_caching_control *ctl)
{
if (atomic_dec_and_test(&ctl->count))
kfree(ctl);
}
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
/*
* this is only called by cache_block_group, since we could have freed extents
* we need to check the pinned_extents for any extents that can't be used yet
* since their free space will be released as soon as the transaction commits.
*/
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
struct btrfs_fs_info *info, u64 start, u64 end)
{
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
u64 extent_start, extent_end, size, total_added = 0;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
int ret;
while (start < end) {
ret = find_first_extent_bit(info->pinned_extents, start,
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
&extent_start, &extent_end,
EXTENT_DIRTY | EXTENT_UPTODATE);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
if (ret)
break;
if (extent_start <= start) {
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
start = extent_end + 1;
} else if (extent_start > start && extent_start < end) {
size = extent_start - start;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
total_added += size;
ret = btrfs_add_free_space(block_group, start,
size);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
BUG_ON(ret);
start = extent_end + 1;
} else {
break;
}
}
if (start < end) {
size = end - start;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
total_added += size;
ret = btrfs_add_free_space(block_group, start, size);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
BUG_ON(ret);
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
return total_added;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
static int caching_kthread(void *data)
{
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
struct btrfs_block_group_cache *block_group = data;
struct btrfs_fs_info *fs_info = block_group->fs_info;
struct btrfs_caching_control *caching_ctl = block_group->caching_ctl;
struct btrfs_root *extent_root = fs_info->extent_root;
struct btrfs_path *path;
struct extent_buffer *leaf;
struct btrfs_key key;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
u64 total_found = 0;
u64 last = 0;
u32 nritems;
int ret = 0;
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
exclude_super_stripes(extent_root, block_group);
spin_lock(&block_group->space_info->lock);
block_group->space_info->bytes_super += block_group->bytes_super;
spin_unlock(&block_group->space_info->lock);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
/*
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
* We don't want to deadlock with somebody trying to allocate a new
* extent for the extent root while also trying to search the extent
* root to add free space. So we skip locking and search the commit
* root, since its read-only
*/
path->skip_locking = 1;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
path->search_commit_root = 1;
path->reada = 2;
key.objectid = last;
key.offset = 0;
key.type = BTRFS_EXTENT_ITEM_KEY;
again:
mutex_lock(&caching_ctl->mutex);
/* need to make sure the commit_root doesn't disappear */
down_read(&fs_info->extent_commit_sem);
ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
if (ret < 0)
goto err;
leaf = path->nodes[0];
nritems = btrfs_header_nritems(leaf);
while (1) {
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
smp_mb();
if (fs_info->closing > 1) {
last = (u64)-1;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
break;
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
if (path->slots[0] < nritems) {
btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
} else {
ret = find_next_key(path, 0, &key);
if (ret)
break;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
caching_ctl->progress = last;
btrfs_release_path(extent_root, path);
up_read(&fs_info->extent_commit_sem);
mutex_unlock(&caching_ctl->mutex);
if (btrfs_transaction_in_commit(fs_info))
schedule_timeout(1);
else
cond_resched();
goto again;
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
if (key.objectid < block_group->key.objectid) {
path->slots[0]++;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
continue;
}
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
if (key.objectid >= block_group->key.objectid +
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
block_group->key.offset)
break;
if (key.type == BTRFS_EXTENT_ITEM_KEY) {
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
total_found += add_new_free_space(block_group,
fs_info, last,
key.objectid);
last = key.objectid + key.offset;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
if (total_found > (1024 * 1024 * 2)) {
total_found = 0;
wake_up(&caching_ctl->wait);
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
}
path->slots[0]++;
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
ret = 0;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
total_found += add_new_free_space(block_group, fs_info, last,
block_group->key.objectid +
block_group->key.offset);
caching_ctl->progress = (u64)-1;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
spin_lock(&block_group->lock);
block_group->caching_ctl = NULL;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
block_group->cached = BTRFS_CACHE_FINISHED;
spin_unlock(&block_group->lock);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
err:
btrfs_free_path(path);
up_read(&fs_info->extent_commit_sem);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
free_excluded_extents(extent_root, block_group);
mutex_unlock(&caching_ctl->mutex);
wake_up(&caching_ctl->wait);
put_caching_control(caching_ctl);
atomic_dec(&block_group->space_info->caching_threads);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
return 0;
}
static int cache_block_group(struct btrfs_block_group_cache *cache)
{
struct btrfs_fs_info *fs_info = cache->fs_info;
struct btrfs_caching_control *caching_ctl;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
struct task_struct *tsk;
int ret = 0;
smp_mb();
if (cache->cached != BTRFS_CACHE_NO)
return 0;
caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_KERNEL);
BUG_ON(!caching_ctl);
INIT_LIST_HEAD(&caching_ctl->list);
mutex_init(&caching_ctl->mutex);
init_waitqueue_head(&caching_ctl->wait);
caching_ctl->block_group = cache;
caching_ctl->progress = cache->key.objectid;
/* one for caching kthread, one for caching block group list */
atomic_set(&caching_ctl->count, 2);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
spin_lock(&cache->lock);
if (cache->cached != BTRFS_CACHE_NO) {
spin_unlock(&cache->lock);
kfree(caching_ctl);
return 0;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
}
cache->caching_ctl = caching_ctl;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
cache->cached = BTRFS_CACHE_STARTED;
spin_unlock(&cache->lock);
down_write(&fs_info->extent_commit_sem);
list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
up_write(&fs_info->extent_commit_sem);
atomic_inc(&cache->space_info->caching_threads);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
tsk = kthread_run(caching_kthread, cache, "btrfs-cache-%llu\n",
cache->key.objectid);
if (IS_ERR(tsk)) {
ret = PTR_ERR(tsk);
printk(KERN_ERR "error running thread %d\n", ret);
BUG();
}
return ret;
}
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
/*
* return the block group that starts at or after bytenr
*/
static struct btrfs_block_group_cache *
btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
{
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
struct btrfs_block_group_cache *cache;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
cache = block_group_cache_tree_search(info, bytenr, 0);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
return cache;
}
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
/*
* return the block group that contains the given bytenr
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
*/
struct btrfs_block_group_cache *btrfs_lookup_block_group(
struct btrfs_fs_info *info,
u64 bytenr)
{
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
struct btrfs_block_group_cache *cache;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
cache = block_group_cache_tree_search(info, bytenr, 1);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
return cache;
}
void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
{
if (atomic_dec_and_test(&cache->count))
kfree(cache);
}
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
u64 flags)
{
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
struct list_head *head = &info->space_info;
struct btrfs_space_info *found;
rcu_read_lock();
list_for_each_entry_rcu(found, head, list) {
if (found->flags == flags) {
rcu_read_unlock();
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
return found;
}
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
}
rcu_read_unlock();
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
return NULL;
}
/*
* after adding space to the filesystem, we need to clear the full flags
* on all the space infos.
*/
void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
{
struct list_head *head = &info->space_info;
struct btrfs_space_info *found;
rcu_read_lock();
list_for_each_entry_rcu(found, head, list)
found->full = 0;
rcu_read_unlock();
}
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
static u64 div_factor(u64 num, int factor)
{
if (factor == 10)
return num;
num *= factor;
do_div(num, 10);
return num;
}
u64 btrfs_find_block_group(struct btrfs_root *root,
u64 search_start, u64 search_hint, int owner)
{
struct btrfs_block_group_cache *cache;
u64 used;
u64 last = max(search_hint, search_start);
u64 group_start = 0;
int full_search = 0;
int factor = 9;
int wrapped = 0;
again:
while (1) {
cache = btrfs_lookup_first_block_group(root->fs_info, last);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
if (!cache)
break;
spin_lock(&cache->lock);
last = cache->key.objectid + cache->key.offset;
used = btrfs_block_group_used(&cache->item);
if ((full_search || !cache->ro) &&
block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
if (used + cache->pinned + cache->reserved <
div_factor(cache->key.offset, factor)) {
group_start = cache->key.objectid;
spin_unlock(&cache->lock);
btrfs_put_block_group(cache);
goto found;
}
}
spin_unlock(&cache->lock);
btrfs_put_block_group(cache);
cond_resched();
}
if (!wrapped) {
last = search_start;
wrapped = 1;
goto again;
}
if (!full_search && factor < 10) {
last = search_start;
full_search = 1;
factor = 10;
goto again;
}
found:
return group_start;
}
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
/* simple helper to search for an existing extent at a given offset */
int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
{
int ret;
struct btrfs_key key;
struct btrfs_path *path;
path = btrfs_alloc_path();
BUG_ON(!path);
key.objectid = start;
key.offset = len;
btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
0, 0);
btrfs_free_path(path);
return ret;
}
/*
* Back reference rules. Back refs have three main goals:
*
* 1) differentiate between all holders of references to an extent so that
* when a reference is dropped we can make sure it was a valid reference
* before freeing the extent.
*
* 2) Provide enough information to quickly find the holders of an extent
* if we notice a given block is corrupted or bad.
*
* 3) Make it easy to migrate blocks for FS shrinking or storage pool
* maintenance. This is actually the same as #2, but with a slightly
* different use case.
*
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* There are two kinds of back refs. The implicit back refs is optimized
* for pointers in non-shared tree blocks. For a given pointer in a block,
* back refs of this kind provide information about the block's owner tree
* and the pointer's key. These information allow us to find the block by
* b-tree searching. The full back refs is for pointers in tree blocks not
* referenced by their owner trees. The location of tree block is recorded
* in the back refs. Actually the full back refs is generic, and can be
* used in all cases the implicit back refs is used. The major shortcoming
* of the full back refs is its overhead. Every time a tree block gets
* COWed, we have to update back refs entry for all pointers in it.
*
* For a newly allocated tree block, we use implicit back refs for
* pointers in it. This means most tree related operations only involve
* implicit back refs. For a tree block created in old transaction, the
* only way to drop a reference to it is COW it. So we can detect the
* event that tree block loses its owner tree's reference and do the
* back refs conversion.
*
* When a tree block is COW'd through a tree, there are four cases:
*
* The reference count of the block is one and the tree is the block's
* owner tree. Nothing to do in this case.
*
* The reference count of the block is one and the tree is not the
* block's owner tree. In this case, full back refs is used for pointers
* in the block. Remove these full back refs, add implicit back refs for
* every pointers in the new block.
*
* The reference count of the block is greater than one and the tree is
* the block's owner tree. In this case, implicit back refs is used for
* pointers in the block. Add full back refs for every pointers in the
* block, increase lower level extents' reference counts. The original
* implicit back refs are entailed to the new block.
*
* The reference count of the block is greater than one and the tree is
* not the block's owner tree. Add implicit back refs for every pointer in
* the new block, increase lower level extents' reference count.
*
* Back Reference Key composing:
*
* The key objectid corresponds to the first byte in the extent,
* The key type is used to differentiate between types of back refs.
* There are different meanings of the key offset for different types
* of back refs.
*
* File extents can be referenced by:
*
* - multiple snapshots, subvolumes, or different generations in one subvol
* - different files inside a single subvolume
* - different offsets inside a file (bookend extents in file.c)
*
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* The extent ref structure for the implicit back refs has fields for:
*
* - Objectid of the subvolume root
* - objectid of the file holding the reference
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* - original offset in the file
* - how many bookend extents
*
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* The key offset for the implicit back refs is hash of the first
* three fields.
*
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* The extent ref structure for the full back refs has field for:
*
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* - number of pointers in the tree leaf
*
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* The key offset for the implicit back refs is the first byte of
* the tree leaf
*
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* When a file extent is allocated, The implicit back refs is used.
* the fields are filled in:
*
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* (root_key.objectid, inode objectid, offset in file, 1)
*
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* When a file extent is removed file truncation, we find the
* corresponding implicit back refs and check the following fields:
*
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* (btrfs_header_owner(leaf), inode objectid, offset in file)
*
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* Btree extents can be referenced by:
*
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* - Different subvolumes
*
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* Both the implicit back refs and the full back refs for tree blocks
* only consist of key. The key offset for the implicit back refs is
* objectid of block's owner tree. The key offset for the full back refs
* is the first byte of parent block.
*
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
* When implicit back refs is used, information about the lowest key and
* level of the tree block are required. These information are stored in
* tree block info structure.
*/
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
u64 owner, u32 extra_size)
{
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_extent_item *item;
struct btrfs_extent_item_v0 *ei0;
struct btrfs_extent_ref_v0 *ref0;
struct btrfs_tree_block_info *bi;
struct extent_buffer *leaf;
struct btrfs_key key;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_key found_key;
u32 new_size = sizeof(*item);
u64 refs;
int ret;
leaf = path->nodes[0];
BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
ei0 = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_extent_item_v0);
refs = btrfs_extent_refs_v0(leaf, ei0);
if (owner == (u64)-1) {
while (1) {
if (path->slots[0] >= btrfs_header_nritems(leaf)) {
ret = btrfs_next_leaf(root, path);
if (ret < 0)
return ret;
BUG_ON(ret > 0);
leaf = path->nodes[0];
}
btrfs_item_key_to_cpu(leaf, &found_key,
path->slots[0]);
BUG_ON(key.objectid != found_key.objectid);
if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
path->slots[0]++;
continue;
}
ref0 = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_extent_ref_v0);
owner = btrfs_ref_objectid_v0(leaf, ref0);
break;
}
}
btrfs_release_path(root, path);
if (owner < BTRFS_FIRST_FREE_OBJECTID)
new_size += sizeof(*bi);
new_size -= sizeof(*ei0);
ret = btrfs_search_slot(trans, root, &key, path,
new_size + extra_size, 1);
if (ret < 0)
return ret;
BUG_ON(ret);
ret = btrfs_extend_item(trans, root, path, new_size);
BUG_ON(ret);
leaf = path->nodes[0];
item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
btrfs_set_extent_refs(leaf, item, refs);
/* FIXME: get real generation */
btrfs_set_extent_generation(leaf, item, 0);
if (owner < BTRFS_FIRST_FREE_OBJECTID) {
btrfs_set_extent_flags(leaf, item,
BTRFS_EXTENT_FLAG_TREE_BLOCK |
BTRFS_BLOCK_FLAG_FULL_BACKREF);
bi = (struct btrfs_tree_block_info *)(item + 1);
/* FIXME: get first key of the block */
memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
btrfs_set_tree_block_level(leaf, bi, (int)owner);
} else {
btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
}
btrfs_mark_buffer_dirty(leaf);
return 0;
}
#endif
static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
{
u32 high_crc = ~(u32)0;
u32 low_crc = ~(u32)0;
__le64 lenum;
lenum = cpu_to_le64(root_objectid);
high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
lenum = cpu_to_le64(owner);
low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
lenum = cpu_to_le64(offset);
low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
return ((u64)high_crc << 31) ^ (u64)low_crc;
}
static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
struct btrfs_extent_data_ref *ref)
{
return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
btrfs_extent_data_ref_objectid(leaf, ref),
btrfs_extent_data_ref_offset(leaf, ref));
}
static int match_extent_data_ref(struct extent_buffer *leaf,
struct btrfs_extent_data_ref *ref,
u64 root_objectid, u64 owner, u64 offset)
{
if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
btrfs_extent_data_ref_offset(leaf, ref) != offset)
return 0;
return 1;
}
static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
u64 bytenr, u64 parent,
u64 root_objectid,
u64 owner, u64 offset)
{
struct btrfs_key key;
struct btrfs_extent_data_ref *ref;
struct extent_buffer *leaf;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
u32 nritems;
int ret;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
int recow;
int err = -ENOENT;
key.objectid = bytenr;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (parent) {
key.type = BTRFS_SHARED_DATA_REF_KEY;
key.offset = parent;
} else {
key.type = BTRFS_EXTENT_DATA_REF_KEY;
key.offset = hash_extent_data_ref(root_objectid,
owner, offset);
}
again:
recow = 0;
ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
if (ret < 0) {
err = ret;
goto fail;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (parent) {
if (!ret)
return 0;
#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
key.type = BTRFS_EXTENT_REF_V0_KEY;
btrfs_release_path(root, path);
ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
if (ret < 0) {
err = ret;
goto fail;
}
if (!ret)
return 0;
#endif
goto fail;
}
leaf = path->nodes[0];
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
nritems = btrfs_header_nritems(leaf);
while (1) {
if (path->slots[0] >= nritems) {
ret = btrfs_next_leaf(root, path);
if (ret < 0)
err = ret;
if (ret)
goto fail;
leaf = path->nodes[0];
nritems = btrfs_header_nritems(leaf);
recow = 1;
}
btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
if (key.objectid != bytenr ||
key.type != BTRFS_EXTENT_DATA_REF_KEY)
goto fail;
ref = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_extent_data_ref);
if (match_extent_data_ref(leaf, ref, root_objectid,
owner, offset)) {
if (recow) {
btrfs_release_path(root, path);
goto again;
}
err = 0;
break;
}
path->slots[0]++;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
fail:
return err;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
u64 bytenr, u64 parent,
u64 root_objectid, u64 owner,
u64 offset, int refs_to_add)
{
struct btrfs_key key;
struct extent_buffer *leaf;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
u32 size;
u32 num_refs;
int ret;
key.objectid = bytenr;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (parent) {
key.type = BTRFS_SHARED_DATA_REF_KEY;
key.offset = parent;
size = sizeof(struct btrfs_shared_data_ref);
} else {
key.type = BTRFS_EXTENT_DATA_REF_KEY;
key.offset = hash_extent_data_ref(root_objectid,
owner, offset);
size = sizeof(struct btrfs_extent_data_ref);
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
ret = btrfs_insert_empty_item(trans, root, path, &key, size);
if (ret && ret != -EEXIST)
goto fail;
leaf = path->nodes[0];
if (parent) {
struct btrfs_shared_data_ref *ref;
ref = btrfs_item_ptr(leaf, path->slots[0],
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_shared_data_ref);
if (ret == 0) {
btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
} else {
num_refs = btrfs_shared_data_ref_count(leaf, ref);
num_refs += refs_to_add;
btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
} else {
struct btrfs_extent_data_ref *ref;
while (ret == -EEXIST) {
ref = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_extent_data_ref);
if (match_extent_data_ref(leaf, ref, root_objectid,
owner, offset))
break;
btrfs_release_path(root, path);
key.offset++;
ret = btrfs_insert_empty_item(trans, root, path, &key,
size);
if (ret && ret != -EEXIST)
goto fail;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
leaf = path->nodes[0];
}
ref = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_extent_data_ref);
if (ret == 0) {
btrfs_set_extent_data_ref_root(leaf, ref,
root_objectid);
btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
btrfs_set_extent_data_ref_offset(leaf, ref, offset);
btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
} else {
num_refs = btrfs_extent_data_ref_count(leaf, ref);
num_refs += refs_to_add;
btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
}
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
btrfs_mark_buffer_dirty(leaf);
ret = 0;
fail:
btrfs_release_path(root, path);
return ret;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
int refs_to_drop)
{
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_key key;
struct btrfs_extent_data_ref *ref1 = NULL;
struct btrfs_shared_data_ref *ref2 = NULL;
struct extent_buffer *leaf;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
u32 num_refs = 0;
int ret = 0;
leaf = path->nodes[0];
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
ref1 = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_extent_data_ref);
num_refs = btrfs_extent_data_ref_count(leaf, ref1);
} else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
ref2 = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_shared_data_ref);
num_refs = btrfs_shared_data_ref_count(leaf, ref2);
#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
} else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
struct btrfs_extent_ref_v0 *ref0;
ref0 = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_extent_ref_v0);
num_refs = btrfs_ref_count_v0(leaf, ref0);
#endif
} else {
BUG();
}
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
BUG_ON(num_refs < refs_to_drop);
num_refs -= refs_to_drop;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (num_refs == 0) {
ret = btrfs_del_item(trans, root, path);
} else {
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
else {
struct btrfs_extent_ref_v0 *ref0;
ref0 = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_extent_ref_v0);
btrfs_set_ref_count_v0(leaf, ref0, num_refs);
}
#endif
btrfs_mark_buffer_dirty(leaf);
}
return ret;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static noinline u32 extent_data_ref_count(struct btrfs_root *root,
struct btrfs_path *path,
struct btrfs_extent_inline_ref *iref)
{
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_key key;
struct extent_buffer *leaf;
struct btrfs_extent_data_ref *ref1;
struct btrfs_shared_data_ref *ref2;
u32 num_refs = 0;
leaf = path->nodes[0];
btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
if (iref) {
if (btrfs_extent_inline_ref_type(leaf, iref) ==
BTRFS_EXTENT_DATA_REF_KEY) {
ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
num_refs = btrfs_extent_data_ref_count(leaf, ref1);
} else {
ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
num_refs = btrfs_shared_data_ref_count(leaf, ref2);
}
} else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
ref1 = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_extent_data_ref);
num_refs = btrfs_extent_data_ref_count(leaf, ref1);
} else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
ref2 = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_shared_data_ref);
num_refs = btrfs_shared_data_ref_count(leaf, ref2);
#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
} else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
struct btrfs_extent_ref_v0 *ref0;
ref0 = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_extent_ref_v0);
num_refs = btrfs_ref_count_v0(leaf, ref0);
#endif
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
} else {
WARN_ON(1);
}
return num_refs;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
u64 bytenr, u64 parent,
u64 root_objectid)
{
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_key key;
int ret;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
key.objectid = bytenr;
if (parent) {
key.type = BTRFS_SHARED_BLOCK_REF_KEY;
key.offset = parent;
} else {
key.type = BTRFS_TREE_BLOCK_REF_KEY;
key.offset = root_objectid;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
if (ret > 0)
ret = -ENOENT;
#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
if (ret == -ENOENT && parent) {
btrfs_release_path(root, path);
key.type = BTRFS_EXTENT_REF_V0_KEY;
ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
if (ret > 0)
ret = -ENOENT;
}
#endif
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
return ret;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
u64 bytenr, u64 parent,
u64 root_objectid)
{
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_key key;
int ret;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
key.objectid = bytenr;
if (parent) {
key.type = BTRFS_SHARED_BLOCK_REF_KEY;
key.offset = parent;
} else {
key.type = BTRFS_TREE_BLOCK_REF_KEY;
key.offset = root_objectid;
}
ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
btrfs_release_path(root, path);
return ret;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static inline int extent_ref_type(u64 parent, u64 owner)
{
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
int type;
if (owner < BTRFS_FIRST_FREE_OBJECTID) {
if (parent > 0)
type = BTRFS_SHARED_BLOCK_REF_KEY;
else
type = BTRFS_TREE_BLOCK_REF_KEY;
} else {
if (parent > 0)
type = BTRFS_SHARED_DATA_REF_KEY;
else
type = BTRFS_EXTENT_DATA_REF_KEY;
}
return type;
}
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
static int find_next_key(struct btrfs_path *path, int level,
struct btrfs_key *key)
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
{
for (; level < BTRFS_MAX_LEVEL; level++) {
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (!path->nodes[level])
break;
if (path->slots[level] + 1 >=
btrfs_header_nritems(path->nodes[level]))
continue;
if (level == 0)
btrfs_item_key_to_cpu(path->nodes[level], key,
path->slots[level] + 1);
else
btrfs_node_key_to_cpu(path->nodes[level], key,
path->slots[level] + 1);
return 0;
}
return 1;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
/*
* look for inline back ref. if back ref is found, *ref_ret is set
* to the address of inline back ref, and 0 is returned.
*
* if back ref isn't found, *ref_ret is set to the address where it
* should be inserted, and -ENOENT is returned.
*
* if insert is true and there are too many inline back refs, the path
* points to the extent item, and -EAGAIN is returned.
*
* NOTE: inline back refs are ordered in the same way that back ref
* items in the tree are ordered.
*/
static noinline_for_stack
int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct btrfs_extent_inline_ref **ref_ret,
u64 bytenr, u64 num_bytes,
u64 parent, u64 root_objectid,
u64 owner, u64 offset, int insert)
{
struct btrfs_key key;
struct extent_buffer *leaf;
struct btrfs_extent_item *ei;
struct btrfs_extent_inline_ref *iref;
u64 flags;
u64 item_size;
unsigned long ptr;
unsigned long end;
int extra_size;
int type;
int want;
int ret;
int err = 0;
key.objectid = bytenr;
key.type = BTRFS_EXTENT_ITEM_KEY;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
key.offset = num_bytes;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
want = extent_ref_type(parent, owner);
if (insert) {
extra_size = btrfs_extent_inline_ref_size(want);
path->keep_locks = 1;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
} else
extra_size = -1;
ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
if (ret < 0) {
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
err = ret;
goto out;
}
BUG_ON(ret);
leaf = path->nodes[0];
item_size = btrfs_item_size_nr(leaf, path->slots[0]);
#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
if (item_size < sizeof(*ei)) {
if (!insert) {
err = -ENOENT;
goto out;
}
ret = convert_extent_item_v0(trans, root, path, owner,
extra_size);
if (ret < 0) {
err = ret;
goto out;
}
leaf = path->nodes[0];
item_size = btrfs_item_size_nr(leaf, path->slots[0]);
}
#endif
BUG_ON(item_size < sizeof(*ei));
ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
flags = btrfs_extent_flags(leaf, ei);
ptr = (unsigned long)(ei + 1);
end = (unsigned long)ei + item_size;
if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
ptr += sizeof(struct btrfs_tree_block_info);
BUG_ON(ptr > end);
} else {
BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
}
err = -ENOENT;
while (1) {
if (ptr >= end) {
WARN_ON(ptr > end);
break;
}
iref = (struct btrfs_extent_inline_ref *)ptr;
type = btrfs_extent_inline_ref_type(leaf, iref);
if (want < type)
break;
if (want > type) {
ptr += btrfs_extent_inline_ref_size(type);
continue;
}
if (type == BTRFS_EXTENT_DATA_REF_KEY) {
struct btrfs_extent_data_ref *dref;
dref = (struct btrfs_extent_data_ref *)(&iref->offset);
if (match_extent_data_ref(leaf, dref, root_objectid,
owner, offset)) {
err = 0;
break;
}
if (hash_extent_data_ref_item(leaf, dref) <
hash_extent_data_ref(root_objectid, owner, offset))
break;
} else {
u64 ref_offset;
ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
if (parent > 0) {
if (parent == ref_offset) {
err = 0;
break;
}
if (ref_offset < parent)
break;
} else {
if (root_objectid == ref_offset) {
err = 0;
break;
}
if (ref_offset < root_objectid)
break;
}
}
ptr += btrfs_extent_inline_ref_size(type);
}
if (err == -ENOENT && insert) {
if (item_size + extra_size >=
BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
err = -EAGAIN;
goto out;
}
/*
* To add new inline back ref, we have to make sure
* there is no corresponding back ref item.
* For simplicity, we just do not add new inline back
* ref if there is any kind of item for this block
*/
if (find_next_key(path, 0, &key) == 0 &&
key.objectid == bytenr &&
key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
err = -EAGAIN;
goto out;
}
}
*ref_ret = (struct btrfs_extent_inline_ref *)ptr;
out:
if (insert) {
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
path->keep_locks = 0;
btrfs_unlock_up_safe(path, 1);
}
return err;
}
/*
* helper to add new inline back ref
*/
static noinline_for_stack
int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct btrfs_extent_inline_ref *iref,
u64 parent, u64 root_objectid,
u64 owner, u64 offset, int refs_to_add,
struct btrfs_delayed_extent_op *extent_op)
{
struct extent_buffer *leaf;
struct btrfs_extent_item *ei;
unsigned long ptr;
unsigned long end;
unsigned long item_offset;
u64 refs;
int size;
int type;
int ret;
leaf = path->nodes[0];
ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
item_offset = (unsigned long)iref - (unsigned long)ei;
type = extent_ref_type(parent, owner);
size = btrfs_extent_inline_ref_size(type);
ret = btrfs_extend_item(trans, root, path, size);
BUG_ON(ret);
ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
refs = btrfs_extent_refs(leaf, ei);
refs += refs_to_add;
btrfs_set_extent_refs(leaf, ei, refs);
if (extent_op)
__run_delayed_extent_op(extent_op, leaf, ei);
ptr = (unsigned long)ei + item_offset;
end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
if (ptr < end - size)
memmove_extent_buffer(leaf, ptr + size, ptr,
end - size - ptr);
iref = (struct btrfs_extent_inline_ref *)ptr;
btrfs_set_extent_inline_ref_type(leaf, iref, type);
if (type == BTRFS_EXTENT_DATA_REF_KEY) {
struct btrfs_extent_data_ref *dref;
dref = (struct btrfs_extent_data_ref *)(&iref->offset);
btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
btrfs_set_extent_data_ref_offset(leaf, dref, offset);
btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
} else if (type == BTRFS_SHARED_DATA_REF_KEY) {
struct btrfs_shared_data_ref *sref;
sref = (struct btrfs_shared_data_ref *)(iref + 1);
btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
} else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
} else {
btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
}
btrfs_mark_buffer_dirty(leaf);
return 0;
}
static int lookup_extent_backref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct btrfs_extent_inline_ref **ref_ret,
u64 bytenr, u64 num_bytes, u64 parent,
u64 root_objectid, u64 owner, u64 offset)
{
int ret;
ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
bytenr, num_bytes, parent,
root_objectid, owner, offset, 0);
if (ret != -ENOENT)
return ret;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
btrfs_release_path(root, path);
*ref_ret = NULL;
if (owner < BTRFS_FIRST_FREE_OBJECTID) {
ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
root_objectid);
} else {
ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
root_objectid, owner, offset);
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
return ret;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
/*
* helper to update/remove inline back ref
*/
static noinline_for_stack
int update_inline_extent_backref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct btrfs_extent_inline_ref *iref,
int refs_to_mod,
struct btrfs_delayed_extent_op *extent_op)
{
struct extent_buffer *leaf;
struct btrfs_extent_item *ei;
struct btrfs_extent_data_ref *dref = NULL;
struct btrfs_shared_data_ref *sref = NULL;
unsigned long ptr;
unsigned long end;
u32 item_size;
int size;
int type;
int ret;
u64 refs;
leaf = path->nodes[0];
ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
refs = btrfs_extent_refs(leaf, ei);
WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
refs += refs_to_mod;
btrfs_set_extent_refs(leaf, ei, refs);
if (extent_op)
__run_delayed_extent_op(extent_op, leaf, ei);
type = btrfs_extent_inline_ref_type(leaf, iref);
if (type == BTRFS_EXTENT_DATA_REF_KEY) {
dref = (struct btrfs_extent_data_ref *)(&iref->offset);
refs = btrfs_extent_data_ref_count(leaf, dref);
} else if (type == BTRFS_SHARED_DATA_REF_KEY) {
sref = (struct btrfs_shared_data_ref *)(iref + 1);
refs = btrfs_shared_data_ref_count(leaf, sref);
} else {
refs = 1;
BUG_ON(refs_to_mod != -1);
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
refs += refs_to_mod;
if (refs > 0) {
if (type == BTRFS_EXTENT_DATA_REF_KEY)
btrfs_set_extent_data_ref_count(leaf, dref, refs);
else
btrfs_set_shared_data_ref_count(leaf, sref, refs);
} else {
size = btrfs_extent_inline_ref_size(type);
item_size = btrfs_item_size_nr(leaf, path->slots[0]);
ptr = (unsigned long)iref;
end = (unsigned long)ei + item_size;
if (ptr + size < end)
memmove_extent_buffer(leaf, ptr, ptr + size,
end - ptr - size);
item_size -= size;
ret = btrfs_truncate_item(trans, root, path, item_size, 1);
BUG_ON(ret);
}
btrfs_mark_buffer_dirty(leaf);
return 0;
}
static noinline_for_stack
int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
u64 bytenr, u64 num_bytes, u64 parent,
u64 root_objectid, u64 owner,
u64 offset, int refs_to_add,
struct btrfs_delayed_extent_op *extent_op)
{
struct btrfs_extent_inline_ref *iref;
int ret;
ret = lookup_inline_extent_backref(trans, root, path, &iref,
bytenr, num_bytes, parent,
root_objectid, owner, offset, 1);
if (ret == 0) {
BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
ret = update_inline_extent_backref(trans, root, path, iref,
refs_to_add, extent_op);
} else if (ret == -ENOENT) {
ret = setup_inline_extent_backref(trans, root, path, iref,
parent, root_objectid,
owner, offset, refs_to_add,
extent_op);
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
return ret;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static int insert_extent_backref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
u64 bytenr, u64 parent, u64 root_objectid,
u64 owner, u64 offset, int refs_to_add)
{
int ret;
if (owner < BTRFS_FIRST_FREE_OBJECTID) {
BUG_ON(refs_to_add != 1);
ret = insert_tree_block_ref(trans, root, path, bytenr,
parent, root_objectid);
} else {
ret = insert_extent_data_ref(trans, root, path, bytenr,
parent, root_objectid,
owner, offset, refs_to_add);
}
return ret;
}
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static int remove_extent_backref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct btrfs_extent_inline_ref *iref,
int refs_to_drop, int is_data)
{
int ret;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
BUG_ON(!is_data && refs_to_drop != 1);
if (iref) {
ret = update_inline_extent_backref(trans, root, path, iref,
-refs_to_drop, NULL);
} else if (is_data) {
ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
} else {
ret = btrfs_del_item(trans, root, path);
}
return ret;
}
static void btrfs_issue_discard(struct block_device *bdev,
u64 start, u64 len)
{
blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_KERNEL,
DISCARD_FL_BARRIER);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
}
static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
u64 num_bytes)
{
int ret;
u64 map_length = num_bytes;
struct btrfs_multi_bio *multi = NULL;
if (!btrfs_test_opt(root, DISCARD))
return 0;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
/* Tell the block device(s) that the sectors can be discarded */
ret = btrfs_map_block(&root->fs_info->mapping_tree, READ,
bytenr, &map_length, &multi, 0);
if (!ret) {
struct btrfs_bio_stripe *stripe = multi->stripes;
int i;
if (map_length > num_bytes)
map_length = num_bytes;
for (i = 0; i < multi->num_stripes; i++, stripe++) {
btrfs_issue_discard(stripe->dev->bdev,
stripe->physical,
map_length);
}
kfree(multi);
}
return ret;
}
int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 bytenr, u64 num_bytes, u64 parent,
u64 root_objectid, u64 owner, u64 offset)
{
int ret;
BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
root_objectid == BTRFS_TREE_LOG_OBJECTID);
if (owner < BTRFS_FIRST_FREE_OBJECTID) {
ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
parent, root_objectid, (int)owner,
BTRFS_ADD_DELAYED_REF, NULL);
} else {
ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
parent, root_objectid, owner, offset,
BTRFS_ADD_DELAYED_REF, NULL);
}
return ret;
}
static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 bytenr, u64 num_bytes,
u64 parent, u64 root_objectid,
u64 owner, u64 offset, int refs_to_add,
struct btrfs_delayed_extent_op *extent_op)
{
struct btrfs_path *path;
struct extent_buffer *leaf;
struct btrfs_extent_item *item;
u64 refs;
int ret;
int err = 0;
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
path->reada = 1;
path->leave_spinning = 1;
/* this will setup the path even if it fails to insert the back ref */
ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
path, bytenr, num_bytes, parent,
root_objectid, owner, offset,
refs_to_add, extent_op);
if (ret == 0)
goto out;
if (ret != -EAGAIN) {
err = ret;
goto out;
}
leaf = path->nodes[0];
item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
refs = btrfs_extent_refs(leaf, item);
btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
if (extent_op)
__run_delayed_extent_op(extent_op, leaf, item);
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
btrfs_mark_buffer_dirty(leaf);
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
btrfs_release_path(root->fs_info->extent_root, path);
path->reada = 1;
path->leave_spinning = 1;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
/* now insert the actual backref */
ret = insert_extent_backref(trans, root->fs_info->extent_root,
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
path, bytenr, parent, root_objectid,
owner, offset, refs_to_add);
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
BUG_ON(ret);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
out:
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
btrfs_free_path(path);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
return err;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_delayed_ref_node *node,
struct btrfs_delayed_extent_op *extent_op,
int insert_reserved)
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
{
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
int ret = 0;
struct btrfs_delayed_data_ref *ref;
struct btrfs_key ins;
u64 parent = 0;
u64 ref_root = 0;
u64 flags = 0;
ins.objectid = node->bytenr;
ins.offset = node->num_bytes;
ins.type = BTRFS_EXTENT_ITEM_KEY;
ref = btrfs_delayed_node_to_data_ref(node);
if (node->type == BTRFS_SHARED_DATA_REF_KEY)
parent = ref->parent;
else
ref_root = ref->root;
if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
if (extent_op) {
BUG_ON(extent_op->update_key);
flags |= extent_op->flags_to_set;
}
ret = alloc_reserved_file_extent(trans, root,
parent, ref_root, flags,
ref->objectid, ref->offset,
&ins, node->ref_mod);
} else if (node->action == BTRFS_ADD_DELAYED_REF) {
ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
node->num_bytes, parent,
ref_root, ref->objectid,
ref->offset, node->ref_mod,
extent_op);
} else if (node->action == BTRFS_DROP_DELAYED_REF) {
ret = __btrfs_free_extent(trans, root, node->bytenr,
node->num_bytes, parent,
ref_root, ref->objectid,
ref->offset, node->ref_mod,
extent_op);
} else {
BUG();
}
return ret;
}
static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
struct extent_buffer *leaf,
struct btrfs_extent_item *ei)
{
u64 flags = btrfs_extent_flags(leaf, ei);
if (extent_op->update_flags) {
flags |= extent_op->flags_to_set;
btrfs_set_extent_flags(leaf, ei, flags);
}
if (extent_op->update_key) {
struct btrfs_tree_block_info *bi;
BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
bi = (struct btrfs_tree_block_info *)(ei + 1);
btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
}
}
static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_delayed_ref_node *node,
struct btrfs_delayed_extent_op *extent_op)
{
struct btrfs_key key;
struct btrfs_path *path;
struct btrfs_extent_item *ei;
struct extent_buffer *leaf;
u32 item_size;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
int ret;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
int err = 0;
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
key.objectid = node->bytenr;
key.type = BTRFS_EXTENT_ITEM_KEY;
key.offset = node->num_bytes;
path->reada = 1;
path->leave_spinning = 1;
ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
path, 0, 1);
if (ret < 0) {
err = ret;
goto out;
}
if (ret > 0) {
err = -EIO;
goto out;
}
leaf = path->nodes[0];
item_size = btrfs_item_size_nr(leaf, path->slots[0]);
#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
if (item_size < sizeof(*ei)) {
ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
path, (u64)-1, 0);
if (ret < 0) {
err = ret;
goto out;
}
leaf = path->nodes[0];
item_size = btrfs_item_size_nr(leaf, path->slots[0]);
}
#endif
BUG_ON(item_size < sizeof(*ei));
ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
__run_delayed_extent_op(extent_op, leaf, ei);
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
btrfs_mark_buffer_dirty(leaf);
out:
btrfs_free_path(path);
return err;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_delayed_ref_node *node,
struct btrfs_delayed_extent_op *extent_op,
int insert_reserved)
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
{
int ret = 0;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_delayed_tree_ref *ref;
struct btrfs_key ins;
u64 parent = 0;
u64 ref_root = 0;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
ins.objectid = node->bytenr;
ins.offset = node->num_bytes;
ins.type = BTRFS_EXTENT_ITEM_KEY;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
ref = btrfs_delayed_node_to_tree_ref(node);
if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
parent = ref->parent;
else
ref_root = ref->root;
BUG_ON(node->ref_mod != 1);
if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
BUG_ON(!extent_op || !extent_op->update_flags ||
!extent_op->update_key);
ret = alloc_reserved_tree_block(trans, root,
parent, ref_root,
extent_op->flags_to_set,
&extent_op->key,
ref->level, &ins);
} else if (node->action == BTRFS_ADD_DELAYED_REF) {
ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
node->num_bytes, parent, ref_root,
ref->level, 0, 1, extent_op);
} else if (node->action == BTRFS_DROP_DELAYED_REF) {
ret = __btrfs_free_extent(trans, root, node->bytenr,
node->num_bytes, parent, ref_root,
ref->level, 0, 1, extent_op);
} else {
BUG();
}
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
return ret;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
/* helper function to actually process a single delayed ref entry */
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_delayed_ref_node *node,
struct btrfs_delayed_extent_op *extent_op,
int insert_reserved)
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
{
int ret;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (btrfs_delayed_ref_is_head(node)) {
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
struct btrfs_delayed_ref_head *head;
/*
* we've hit the end of the chain and we were supposed
* to insert this extent into the tree. But, it got
* deleted before we ever needed to insert it, so all
* we have to do is clean up the accounting
*/
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
BUG_ON(extent_op);
head = btrfs_delayed_node_to_head(node);
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
if (insert_reserved) {
int mark_free = 0;
struct extent_buffer *must_clean = NULL;
ret = pin_down_bytes(trans, root, NULL,
node->bytenr, node->num_bytes,
head->is_data, 1, &must_clean);
if (ret > 0)
mark_free = 1;
if (must_clean) {
clean_tree_block(NULL, root, must_clean);
btrfs_tree_unlock(must_clean);
free_extent_buffer(must_clean);
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (head->is_data) {
ret = btrfs_del_csums(trans, root,
node->bytenr,
node->num_bytes);
BUG_ON(ret);
}
if (mark_free) {
ret = btrfs_free_reserved_extent(root,
node->bytenr,
node->num_bytes);
BUG_ON(ret);
}
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
}
mutex_unlock(&head->mutex);
return 0;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
node->type == BTRFS_SHARED_BLOCK_REF_KEY)
ret = run_delayed_tree_ref(trans, root, node, extent_op,
insert_reserved);
else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
node->type == BTRFS_SHARED_DATA_REF_KEY)
ret = run_delayed_data_ref(trans, root, node, extent_op,
insert_reserved);
else
BUG();
return ret;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
}
static noinline struct btrfs_delayed_ref_node *
select_delayed_ref(struct btrfs_delayed_ref_head *head)
{
struct rb_node *node;
struct btrfs_delayed_ref_node *ref;
int action = BTRFS_ADD_DELAYED_REF;
again:
/*
* select delayed ref of type BTRFS_ADD_DELAYED_REF first.
* this prevents ref count from going down to zero when
* there still are pending delayed ref.
*/
node = rb_prev(&head->node.rb_node);
while (1) {
if (!node)
break;
ref = rb_entry(node, struct btrfs_delayed_ref_node,
rb_node);
if (ref->bytenr != head->node.bytenr)
break;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (ref->action == action)
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
return ref;
node = rb_prev(node);
}
if (action == BTRFS_ADD_DELAYED_REF) {
action = BTRFS_DROP_DELAYED_REF;
goto again;
}
return NULL;
}
static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct list_head *cluster)
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
{
struct btrfs_delayed_ref_root *delayed_refs;
struct btrfs_delayed_ref_node *ref;
struct btrfs_delayed_ref_head *locked_ref = NULL;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_delayed_extent_op *extent_op;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
int ret;
int count = 0;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
int must_insert_reserved = 0;
delayed_refs = &trans->transaction->delayed_refs;
while (1) {
if (!locked_ref) {
/* pick a new head ref from the cluster list */
if (list_empty(cluster))
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
break;
locked_ref = list_entry(cluster->next,
struct btrfs_delayed_ref_head, cluster);
/* grab the lock that says we are going to process
* all the refs for this head */
ret = btrfs_delayed_ref_lock(trans, locked_ref);
/*
* we may have dropped the spin lock to get the head
* mutex lock, and that might have given someone else
* time to free the head. If that's true, it has been
* removed from our list and we can move on.
*/
if (ret == -EAGAIN) {
locked_ref = NULL;
count++;
continue;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
}
}
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
/*
* record the must insert reserved flag before we
* drop the spin lock.
*/
must_insert_reserved = locked_ref->must_insert_reserved;
locked_ref->must_insert_reserved = 0;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
extent_op = locked_ref->extent_op;
locked_ref->extent_op = NULL;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
/*
* locked_ref is the head node, so we have to go one
* node back for any delayed ref updates
*/
ref = select_delayed_ref(locked_ref);
if (!ref) {
/* All delayed refs have been processed, Go ahead
* and send the head node to run_one_delayed_ref,
* so that any accounting fixes can happen
*/
ref = &locked_ref->node;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (extent_op && must_insert_reserved) {
kfree(extent_op);
extent_op = NULL;
}
if (extent_op) {
spin_unlock(&delayed_refs->lock);
ret = run_delayed_extent_op(trans, root,
ref, extent_op);
BUG_ON(ret);
kfree(extent_op);
cond_resched();
spin_lock(&delayed_refs->lock);
continue;
}
list_del_init(&locked_ref->cluster);
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
locked_ref = NULL;
}
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
ref->in_tree = 0;
rb_erase(&ref->rb_node, &delayed_refs->root);
delayed_refs->num_entries--;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
spin_unlock(&delayed_refs->lock);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
ret = run_one_delayed_ref(trans, root, ref, extent_op,
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
must_insert_reserved);
BUG_ON(ret);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
btrfs_put_delayed_ref(ref);
kfree(extent_op);
count++;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
cond_resched();
spin_lock(&delayed_refs->lock);
}
return count;
}
/*
* this starts processing the delayed reference count updates and
* extent insertions we have queued up so far. count can be
* 0, which means to process everything in the tree at the start
* of the run (but not newly added entries), or it can be some target
* number you'd like to process.
*/
int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
struct btrfs_root *root, unsigned long count)
{
struct rb_node *node;
struct btrfs_delayed_ref_root *delayed_refs;
struct btrfs_delayed_ref_node *ref;
struct list_head cluster;
int ret;
int run_all = count == (unsigned long)-1;
int run_most = 0;
if (root == root->fs_info->extent_root)
root = root->fs_info->tree_root;
delayed_refs = &trans->transaction->delayed_refs;
INIT_LIST_HEAD(&cluster);
again:
spin_lock(&delayed_refs->lock);
if (count == 0) {
count = delayed_refs->num_entries * 2;
run_most = 1;
}
while (1) {
if (!(run_all || run_most) &&
delayed_refs->num_heads_ready < 64)
break;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
/*
* go find something we can process in the rbtree. We start at
* the beginning of the tree, and then build a cluster
* of refs to process starting at the first one we are able to
* lock
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
*/
ret = btrfs_find_ref_cluster(trans, &cluster,
delayed_refs->run_delayed_start);
if (ret)
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
break;
ret = run_clustered_refs(trans, root, &cluster);
BUG_ON(ret < 0);
count -= min_t(unsigned long, ret, count);
if (count == 0)
break;
}
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
if (run_all) {
node = rb_first(&delayed_refs->root);
if (!node)
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
goto out;
count = (unsigned long)-1;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
while (node) {
ref = rb_entry(node, struct btrfs_delayed_ref_node,
rb_node);
if (btrfs_delayed_ref_is_head(ref)) {
struct btrfs_delayed_ref_head *head;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
head = btrfs_delayed_node_to_head(ref);
atomic_inc(&ref->refs);
spin_unlock(&delayed_refs->lock);
mutex_lock(&head->mutex);
mutex_unlock(&head->mutex);
btrfs_put_delayed_ref(ref);
cond_resched();
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
goto again;
}
node = rb_next(node);
}
spin_unlock(&delayed_refs->lock);
schedule_timeout(1);
goto again;
}
out:
spin_unlock(&delayed_refs->lock);
return 0;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 bytenr, u64 num_bytes, u64 flags,
int is_data)
{
struct btrfs_delayed_extent_op *extent_op;
int ret;
extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
if (!extent_op)
return -ENOMEM;
extent_op->flags_to_set = flags;
extent_op->update_flags = 1;
extent_op->update_key = 0;
extent_op->is_data = is_data ? 1 : 0;
ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
if (ret)
kfree(extent_op);
return ret;
}
static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
u64 objectid, u64 offset, u64 bytenr)
{
struct btrfs_delayed_ref_head *head;
struct btrfs_delayed_ref_node *ref;
struct btrfs_delayed_data_ref *data_ref;
struct btrfs_delayed_ref_root *delayed_refs;
struct rb_node *node;
int ret = 0;
ret = -ENOENT;
delayed_refs = &trans->transaction->delayed_refs;
spin_lock(&delayed_refs->lock);
head = btrfs_find_delayed_ref_head(trans, bytenr);
if (!head)
goto out;
if (!mutex_trylock(&head->mutex)) {
atomic_inc(&head->node.refs);
spin_unlock(&delayed_refs->lock);
btrfs_release_path(root->fs_info->extent_root, path);
mutex_lock(&head->mutex);
mutex_unlock(&head->mutex);
btrfs_put_delayed_ref(&head->node);
return -EAGAIN;
}
node = rb_prev(&head->node.rb_node);
if (!node)
goto out_unlock;
ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
if (ref->bytenr != bytenr)
goto out_unlock;
ret = 1;
if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
goto out_unlock;
data_ref = btrfs_delayed_node_to_data_ref(ref);
node = rb_prev(node);
if (node) {
ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
if (ref->bytenr == bytenr)
goto out_unlock;
}
if (data_ref->root != root->root_key.objectid ||
data_ref->objectid != objectid || data_ref->offset != offset)
goto out_unlock;
ret = 0;
out_unlock:
mutex_unlock(&head->mutex);
out:
spin_unlock(&delayed_refs->lock);
return ret;
}
static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
u64 objectid, u64 offset, u64 bytenr)
{
struct btrfs_root *extent_root = root->fs_info->extent_root;
struct extent_buffer *leaf;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_extent_data_ref *ref;
struct btrfs_extent_inline_ref *iref;
struct btrfs_extent_item *ei;
struct btrfs_key key;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
u32 item_size;
int ret;
key.objectid = bytenr;
key.offset = (u64)-1;
key.type = BTRFS_EXTENT_ITEM_KEY;
ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
if (ret < 0)
goto out;
BUG_ON(ret == 0);
ret = -ENOENT;
if (path->slots[0] == 0)
goto out;
path->slots[0]--;
leaf = path->nodes[0];
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
goto out;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
ret = 1;
item_size = btrfs_item_size_nr(leaf, path->slots[0]);
#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
if (item_size < sizeof(*ei)) {
WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
goto out;
}
#endif
ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (item_size != sizeof(*ei) +
btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
goto out;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (btrfs_extent_generation(leaf, ei) <=
btrfs_root_last_snapshot(&root->root_item))
goto out;
iref = (struct btrfs_extent_inline_ref *)(ei + 1);
if (btrfs_extent_inline_ref_type(leaf, iref) !=
BTRFS_EXTENT_DATA_REF_KEY)
goto out;
ref = (struct btrfs_extent_data_ref *)(&iref->offset);
if (btrfs_extent_refs(leaf, ei) !=
btrfs_extent_data_ref_count(leaf, ref) ||
btrfs_extent_data_ref_root(leaf, ref) !=
root->root_key.objectid ||
btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
btrfs_extent_data_ref_offset(leaf, ref) != offset)
goto out;
ret = 0;
out:
return ret;
}
int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 objectid, u64 offset, u64 bytenr)
{
struct btrfs_path *path;
int ret;
int ret2;
path = btrfs_alloc_path();
if (!path)
return -ENOENT;
do {
ret = check_committed_ref(trans, root, path, objectid,
offset, bytenr);
if (ret && ret != -ENOENT)
goto out;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
ret2 = check_delayed_ref(trans, root, path, objectid,
offset, bytenr);
} while (ret2 == -EAGAIN);
if (ret2 && ret2 != -ENOENT) {
ret = ret2;
goto out;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (ret != -ENOENT || ret2 != -ENOENT)
ret = 0;
out:
btrfs_free_path(path);
return ret;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
#if 0
int btrfs_cache_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
struct extent_buffer *buf, u32 nr_extents)
{
struct btrfs_key key;
struct btrfs_file_extent_item *fi;
u64 root_gen;
u32 nritems;
int i;
int level;
int ret = 0;
int shared = 0;
if (!root->ref_cows)
return 0;
if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
shared = 0;
root_gen = root->root_key.offset;
} else {
shared = 1;
root_gen = trans->transid - 1;
}
level = btrfs_header_level(buf);
nritems = btrfs_header_nritems(buf);
if (level == 0) {
struct btrfs_leaf_ref *ref;
struct btrfs_extent_info *info;
ref = btrfs_alloc_leaf_ref(root, nr_extents);
if (!ref) {
ret = -ENOMEM;
goto out;
}
ref->root_gen = root_gen;
ref->bytenr = buf->start;
ref->owner = btrfs_header_owner(buf);
ref->generation = btrfs_header_generation(buf);
ref->nritems = nr_extents;
info = ref->extents;
for (i = 0; nr_extents > 0 && i < nritems; i++) {
u64 disk_bytenr;
btrfs_item_key_to_cpu(buf, &key, i);
if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
continue;
fi = btrfs_item_ptr(buf, i,
struct btrfs_file_extent_item);
if (btrfs_file_extent_type(buf, fi) ==
BTRFS_FILE_EXTENT_INLINE)
continue;
disk_bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
if (disk_bytenr == 0)
continue;
info->bytenr = disk_bytenr;
info->num_bytes =
btrfs_file_extent_disk_num_bytes(buf, fi);
info->objectid = key.objectid;
info->offset = key.offset;
info++;
}
ret = btrfs_add_leaf_ref(root, ref, shared);
if (ret == -EEXIST && shared) {
struct btrfs_leaf_ref *old;
old = btrfs_lookup_leaf_ref(root, ref->bytenr);
BUG_ON(!old);
btrfs_remove_leaf_ref(root, old);
btrfs_free_leaf_ref(root, old);
ret = btrfs_add_leaf_ref(root, ref, shared);
}
WARN_ON(ret);
btrfs_free_leaf_ref(root, ref);
}
out:
return ret;
}
/* when a block goes through cow, we update the reference counts of
* everything that block points to. The internal pointers of the block
* can be in just about any order, and it is likely to have clusters of
* things that are close together and clusters of things that are not.
*
* To help reduce the seeks that come with updating all of these reference
* counts, sort them by byte number before actual updates are done.
*
* struct refsort is used to match byte number to slot in the btree block.
* we sort based on the byte number and then use the slot to actually
* find the item.
Btrfs: Make btrfs_drop_snapshot work in larger and more efficient chunks Every transaction in btrfs creates a new snapshot, and then schedules the snapshot from the last transaction for deletion. Snapshot deletion works by walking down the btree and dropping the reference counts on each btree block during the walk. If if a given leaf or node has a reference count greater than one, the reference count is decremented and the subtree pointed to by that node is ignored. If the reference count is one, walking continues down into that node or leaf, and the references of everything it points to are decremented. The old code would try to work in small pieces, walking down the tree until it found the lowest leaf or node to free and then returning. This was very friendly to the rest of the FS because it didn't have a huge impact on other operations. But it wouldn't always keep up with the rate that new commits added new snapshots for deletion, and it wasn't very optimal for the extent allocation tree because it wasn't finding leaves that were close together on disk and processing them at the same time. This changes things to walk down to a level 1 node and then process it in bulk. All the leaf pointers are sorted and the leaves are dropped in order based on their extent number. The extent allocation tree and commit code are now fast enough for this kind of bulk processing to work without slowing the rest of the FS down. Overall it does less IO and is better able to keep up with snapshot deletions under high load. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:27:02 +08:00
*
* struct refsort is smaller than strcut btrfs_item and smaller than
* struct btrfs_key_ptr. Since we're currently limited to the page size
* for a btree block, there's no way for a kmalloc of refsorts for a
* single node to be bigger than a page.
*/
struct refsort {
u64 bytenr;
u32 slot;
};
/*
* for passing into sort()
*/
static int refsort_cmp(const void *a_void, const void *b_void)
{
const struct refsort *a = a_void;
const struct refsort *b = b_void;
if (a->bytenr < b->bytenr)
return -1;
if (a->bytenr > b->bytenr)
return 1;
return 0;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
#endif
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct extent_buffer *buf,
int full_backref, int inc)
{
u64 bytenr;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
u64 num_bytes;
u64 parent;
u64 ref_root;
u32 nritems;
struct btrfs_key key;
struct btrfs_file_extent_item *fi;
int i;
int level;
int ret = 0;
int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
u64, u64, u64, u64, u64, u64);
ref_root = btrfs_header_owner(buf);
nritems = btrfs_header_nritems(buf);
level = btrfs_header_level(buf);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (!root->ref_cows && level == 0)
return 0;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (inc)
process_func = btrfs_inc_extent_ref;
else
process_func = btrfs_free_extent;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (full_backref)
parent = buf->start;
else
parent = 0;
for (i = 0; i < nritems; i++) {
if (level == 0) {
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
btrfs_item_key_to_cpu(buf, &key, i);
if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
continue;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
fi = btrfs_item_ptr(buf, i,
struct btrfs_file_extent_item);
if (btrfs_file_extent_type(buf, fi) ==
BTRFS_FILE_EXTENT_INLINE)
continue;
bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
if (bytenr == 0)
continue;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
key.offset -= btrfs_file_extent_offset(buf, fi);
ret = process_func(trans, root, bytenr, num_bytes,
parent, ref_root, key.objectid,
key.offset);
if (ret)
goto fail;
} else {
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
bytenr = btrfs_node_blockptr(buf, i);
num_bytes = btrfs_level_size(root, level - 1);
ret = process_func(trans, root, bytenr, num_bytes,
parent, ref_root, level - 1, 0);
if (ret)
goto fail;
}
}
return 0;
fail:
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
BUG();
return ret;
}
int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
struct extent_buffer *buf, int full_backref)
{
return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
}
int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
struct extent_buffer *buf, int full_backref)
{
return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
}
static int write_one_cache_group(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct btrfs_block_group_cache *cache)
{
int ret;
struct btrfs_root *extent_root = root->fs_info->extent_root;
unsigned long bi;
struct extent_buffer *leaf;
ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
if (ret < 0)
goto fail;
BUG_ON(ret);
leaf = path->nodes[0];
bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
btrfs_mark_buffer_dirty(leaf);
btrfs_release_path(extent_root, path);
fail:
if (ret)
return ret;
return 0;
}
static struct btrfs_block_group_cache *
next_block_group(struct btrfs_root *root,
struct btrfs_block_group_cache *cache)
{
struct rb_node *node;
spin_lock(&root->fs_info->block_group_cache_lock);
node = rb_next(&cache->cache_node);
btrfs_put_block_group(cache);
if (node) {
cache = rb_entry(node, struct btrfs_block_group_cache,
cache_node);
atomic_inc(&cache->count);
} else
cache = NULL;
spin_unlock(&root->fs_info->block_group_cache_lock);
return cache;
}
int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
struct btrfs_root *root)
{
struct btrfs_block_group_cache *cache;
int err = 0;
struct btrfs_path *path;
u64 last = 0;
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
while (1) {
if (last == 0) {
err = btrfs_run_delayed_refs(trans, root,
(unsigned long)-1);
BUG_ON(err);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
}
cache = btrfs_lookup_first_block_group(root->fs_info, last);
while (cache) {
if (cache->dirty)
break;
cache = next_block_group(root, cache);
}
if (!cache) {
if (last == 0)
break;
last = 0;
continue;
}
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
cache->dirty = 0;
last = cache->key.objectid + cache->key.offset;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
err = write_one_cache_group(trans, root, path, cache);
BUG_ON(err);
btrfs_put_block_group(cache);
}
btrfs_free_path(path);
return 0;
}
int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
{
struct btrfs_block_group_cache *block_group;
int readonly = 0;
block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
if (!block_group || block_group->ro)
readonly = 1;
if (block_group)
btrfs_put_block_group(block_group);
return readonly;
}
static int update_space_info(struct btrfs_fs_info *info, u64 flags,
u64 total_bytes, u64 bytes_used,
struct btrfs_space_info **space_info)
{
struct btrfs_space_info *found;
found = __find_space_info(info, flags);
if (found) {
spin_lock(&found->lock);
found->total_bytes += total_bytes;
found->bytes_used += bytes_used;
found->full = 0;
spin_unlock(&found->lock);
*space_info = found;
return 0;
}
found = kzalloc(sizeof(*found), GFP_NOFS);
if (!found)
return -ENOMEM;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
INIT_LIST_HEAD(&found->block_groups);
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
init_rwsem(&found->groups_sem);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
spin_lock_init(&found->lock);
found->flags = flags;
found->total_bytes = total_bytes;
found->bytes_used = bytes_used;
found->bytes_pinned = 0;
found->bytes_reserved = 0;
found->bytes_readonly = 0;
found->bytes_delalloc = 0;
found->full = 0;
found->force_alloc = 0;
*space_info = found;
list_add_rcu(&found->list, &info->space_info);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
atomic_set(&found->caching_threads, 0);
return 0;
}
static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
{
u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
BTRFS_BLOCK_GROUP_RAID1 |
BTRFS_BLOCK_GROUP_RAID10 |
BTRFS_BLOCK_GROUP_DUP);
if (extra_flags) {
if (flags & BTRFS_BLOCK_GROUP_DATA)
fs_info->avail_data_alloc_bits |= extra_flags;
if (flags & BTRFS_BLOCK_GROUP_METADATA)
fs_info->avail_metadata_alloc_bits |= extra_flags;
if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
fs_info->avail_system_alloc_bits |= extra_flags;
}
}
static void set_block_group_readonly(struct btrfs_block_group_cache *cache)
{
spin_lock(&cache->space_info->lock);
spin_lock(&cache->lock);
if (!cache->ro) {
cache->space_info->bytes_readonly += cache->key.offset -
btrfs_block_group_used(&cache->item);
cache->ro = 1;
}
spin_unlock(&cache->lock);
spin_unlock(&cache->space_info->lock);
}
u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
{
u64 num_devices = root->fs_info->fs_devices->rw_devices;
if (num_devices == 1)
flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
if (num_devices < 4)
flags &= ~BTRFS_BLOCK_GROUP_RAID10;
if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
(flags & (BTRFS_BLOCK_GROUP_RAID1 |
BTRFS_BLOCK_GROUP_RAID10))) {
flags &= ~BTRFS_BLOCK_GROUP_DUP;
}
if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
(flags & BTRFS_BLOCK_GROUP_RAID10)) {
flags &= ~BTRFS_BLOCK_GROUP_RAID1;
}
if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
((flags & BTRFS_BLOCK_GROUP_RAID1) |
(flags & BTRFS_BLOCK_GROUP_RAID10) |
(flags & BTRFS_BLOCK_GROUP_DUP)))
flags &= ~BTRFS_BLOCK_GROUP_RAID0;
return flags;
}
static u64 btrfs_get_alloc_profile(struct btrfs_root *root, u64 data)
{
struct btrfs_fs_info *info = root->fs_info;
u64 alloc_profile;
if (data) {
alloc_profile = info->avail_data_alloc_bits &
info->data_alloc_profile;
data = BTRFS_BLOCK_GROUP_DATA | alloc_profile;
} else if (root == root->fs_info->chunk_root) {
alloc_profile = info->avail_system_alloc_bits &
info->system_alloc_profile;
data = BTRFS_BLOCK_GROUP_SYSTEM | alloc_profile;
} else {
alloc_profile = info->avail_metadata_alloc_bits &
info->metadata_alloc_profile;
data = BTRFS_BLOCK_GROUP_METADATA | alloc_profile;
}
return btrfs_reduce_alloc_profile(root, data);
}
void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
{
u64 alloc_target;
alloc_target = btrfs_get_alloc_profile(root, 1);
BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
alloc_target);
}
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
static u64 calculate_bytes_needed(struct btrfs_root *root, int num_items)
{
u64 num_bytes;
int level;
level = BTRFS_MAX_LEVEL - 2;
/*
* NOTE: these calculations are absolutely the worst possible case.
* This assumes that _every_ item we insert will require a new leaf, and
* that the tree has grown to its maximum level size.
*/
/*
* for every item we insert we could insert both an extent item and a
* extent ref item. Then for ever item we insert, we will need to cow
* both the original leaf, plus the leaf to the left and right of it.
*
* Unless we are talking about the extent root, then we just want the
* number of items * 2, since we just need the extent item plus its ref.
*/
if (root == root->fs_info->extent_root)
num_bytes = num_items * 2;
else
num_bytes = (num_items + (2 * num_items)) * 3;
/*
* num_bytes is total number of leaves we could need times the leaf
* size, and then for every leaf we could end up cow'ing 2 nodes per
* level, down to the leaf level.
*/
num_bytes = (num_bytes * root->leafsize) +
(num_bytes * (level * 2)) * root->nodesize;
return num_bytes;
}
/*
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
* Unreserve metadata space for delalloc. If we have less reserved credits than
* we have extents, this function does nothing.
*/
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
int btrfs_unreserve_metadata_for_delalloc(struct btrfs_root *root,
struct inode *inode, int num_items)
{
struct btrfs_fs_info *info = root->fs_info;
struct btrfs_space_info *meta_sinfo;
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
u64 num_bytes;
u64 alloc_target;
bool bug = false;
/* get the space info for where the metadata will live */
alloc_target = btrfs_get_alloc_profile(root, 0);
meta_sinfo = __find_space_info(info, alloc_target);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
num_bytes = calculate_bytes_needed(root->fs_info->extent_root,
num_items);
spin_lock(&meta_sinfo->lock);
spin_lock(&BTRFS_I(inode)->accounting_lock);
if (BTRFS_I(inode)->reserved_extents <=
BTRFS_I(inode)->outstanding_extents) {
spin_unlock(&BTRFS_I(inode)->accounting_lock);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
spin_unlock(&meta_sinfo->lock);
return 0;
}
spin_unlock(&BTRFS_I(inode)->accounting_lock);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
BTRFS_I(inode)->reserved_extents--;
BUG_ON(BTRFS_I(inode)->reserved_extents < 0);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
if (meta_sinfo->bytes_delalloc < num_bytes) {
bug = true;
meta_sinfo->bytes_delalloc = 0;
} else {
meta_sinfo->bytes_delalloc -= num_bytes;
}
spin_unlock(&meta_sinfo->lock);
BUG_ON(bug);
return 0;
}
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
static void check_force_delalloc(struct btrfs_space_info *meta_sinfo)
{
u64 thresh;
thresh = meta_sinfo->bytes_used + meta_sinfo->bytes_reserved +
meta_sinfo->bytes_pinned + meta_sinfo->bytes_readonly +
meta_sinfo->bytes_super + meta_sinfo->bytes_root +
meta_sinfo->bytes_may_use;
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
thresh = meta_sinfo->total_bytes - thresh;
thresh *= 80;
do_div(thresh, 100);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
if (thresh <= meta_sinfo->bytes_delalloc)
meta_sinfo->force_delalloc = 1;
else
meta_sinfo->force_delalloc = 0;
}
struct async_flush {
struct btrfs_root *root;
struct btrfs_space_info *info;
struct btrfs_work work;
};
static noinline void flush_delalloc_async(struct btrfs_work *work)
{
struct async_flush *async;
struct btrfs_root *root;
struct btrfs_space_info *info;
async = container_of(work, struct async_flush, work);
root = async->root;
info = async->info;
btrfs_start_delalloc_inodes(root, 0);
wake_up(&info->flush_wait);
btrfs_wait_ordered_extents(root, 0, 0);
spin_lock(&info->lock);
info->flushing = 0;
spin_unlock(&info->lock);
wake_up(&info->flush_wait);
kfree(async);
}
static void wait_on_flush(struct btrfs_space_info *info)
{
DEFINE_WAIT(wait);
u64 used;
while (1) {
prepare_to_wait(&info->flush_wait, &wait,
TASK_UNINTERRUPTIBLE);
spin_lock(&info->lock);
if (!info->flushing) {
spin_unlock(&info->lock);
break;
}
used = info->bytes_used + info->bytes_reserved +
info->bytes_pinned + info->bytes_readonly +
info->bytes_super + info->bytes_root +
info->bytes_may_use + info->bytes_delalloc;
if (used < info->total_bytes) {
spin_unlock(&info->lock);
break;
}
spin_unlock(&info->lock);
schedule();
}
finish_wait(&info->flush_wait, &wait);
}
static void flush_delalloc(struct btrfs_root *root,
struct btrfs_space_info *info)
{
struct async_flush *async;
bool wait = false;
spin_lock(&info->lock);
if (!info->flushing) {
info->flushing = 1;
init_waitqueue_head(&info->flush_wait);
} else {
wait = true;
}
spin_unlock(&info->lock);
if (wait) {
wait_on_flush(info);
return;
}
async = kzalloc(sizeof(*async), GFP_NOFS);
if (!async)
goto flush;
async->root = root;
async->info = info;
async->work.func = flush_delalloc_async;
btrfs_queue_worker(&root->fs_info->enospc_workers,
&async->work);
wait_on_flush(info);
return;
flush:
btrfs_start_delalloc_inodes(root, 0);
btrfs_wait_ordered_extents(root, 0, 0);
spin_lock(&info->lock);
info->flushing = 0;
spin_unlock(&info->lock);
wake_up(&info->flush_wait);
}
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
static int maybe_allocate_chunk(struct btrfs_root *root,
struct btrfs_space_info *info)
{
struct btrfs_super_block *disk_super = &root->fs_info->super_copy;
struct btrfs_trans_handle *trans;
bool wait = false;
int ret = 0;
u64 min_metadata;
u64 free_space;
free_space = btrfs_super_total_bytes(disk_super);
/*
* we allow the metadata to grow to a max of either 10gb or 5% of the
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
* space in the volume.
*/
min_metadata = min((u64)10 * 1024 * 1024 * 1024,
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
div64_u64(free_space * 5, 100));
if (info->total_bytes >= min_metadata) {
spin_unlock(&info->lock);
return 0;
}
if (info->full) {
spin_unlock(&info->lock);
return 0;
}
if (!info->allocating_chunk) {
info->force_alloc = 1;
info->allocating_chunk = 1;
init_waitqueue_head(&info->allocate_wait);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
} else {
wait = true;
}
spin_unlock(&info->lock);
if (wait) {
wait_event(info->allocate_wait,
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
!info->allocating_chunk);
return 1;
}
trans = btrfs_start_transaction(root, 1);
if (!trans) {
ret = -ENOMEM;
goto out;
}
ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4096 + 2 * 1024 * 1024,
info->flags, 0);
btrfs_end_transaction(trans, root);
if (ret)
goto out;
out:
spin_lock(&info->lock);
info->allocating_chunk = 0;
spin_unlock(&info->lock);
wake_up(&info->allocate_wait);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
if (ret)
return 0;
return 1;
}
/*
* Reserve metadata space for delalloc.
*/
int btrfs_reserve_metadata_for_delalloc(struct btrfs_root *root,
struct inode *inode, int num_items)
{
struct btrfs_fs_info *info = root->fs_info;
struct btrfs_space_info *meta_sinfo;
u64 num_bytes;
u64 used;
u64 alloc_target;
int flushed = 0;
int force_delalloc;
/* get the space info for where the metadata will live */
alloc_target = btrfs_get_alloc_profile(root, 0);
meta_sinfo = __find_space_info(info, alloc_target);
num_bytes = calculate_bytes_needed(root->fs_info->extent_root,
num_items);
again:
spin_lock(&meta_sinfo->lock);
force_delalloc = meta_sinfo->force_delalloc;
if (unlikely(!meta_sinfo->bytes_root))
meta_sinfo->bytes_root = calculate_bytes_needed(root, 6);
if (!flushed)
meta_sinfo->bytes_delalloc += num_bytes;
used = meta_sinfo->bytes_used + meta_sinfo->bytes_reserved +
meta_sinfo->bytes_pinned + meta_sinfo->bytes_readonly +
meta_sinfo->bytes_super + meta_sinfo->bytes_root +
meta_sinfo->bytes_may_use + meta_sinfo->bytes_delalloc;
if (used > meta_sinfo->total_bytes) {
flushed++;
if (flushed == 1) {
if (maybe_allocate_chunk(root, meta_sinfo))
goto again;
flushed++;
} else {
spin_unlock(&meta_sinfo->lock);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
}
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
if (flushed == 2) {
filemap_flush(inode->i_mapping);
goto again;
} else if (flushed == 3) {
flush_delalloc(root, meta_sinfo);
goto again;
}
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
spin_lock(&meta_sinfo->lock);
meta_sinfo->bytes_delalloc -= num_bytes;
spin_unlock(&meta_sinfo->lock);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
printk(KERN_ERR "enospc, has %d, reserved %d\n",
BTRFS_I(inode)->outstanding_extents,
BTRFS_I(inode)->reserved_extents);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
dump_space_info(meta_sinfo, 0, 0);
return -ENOSPC;
}
BTRFS_I(inode)->reserved_extents++;
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
check_force_delalloc(meta_sinfo);
spin_unlock(&meta_sinfo->lock);
if (!flushed && force_delalloc)
filemap_flush(inode->i_mapping);
return 0;
}
/*
* unreserve num_items number of items worth of metadata space. This needs to
* be paired with btrfs_reserve_metadata_space.
*
* NOTE: if you have the option, run this _AFTER_ you do a
* btrfs_end_transaction, since btrfs_end_transaction will run delayed ref
* oprations which will result in more used metadata, so we want to make sure we
* can do that without issue.
*/
int btrfs_unreserve_metadata_space(struct btrfs_root *root, int num_items)
{
struct btrfs_fs_info *info = root->fs_info;
struct btrfs_space_info *meta_sinfo;
u64 num_bytes;
u64 alloc_target;
bool bug = false;
/* get the space info for where the metadata will live */
alloc_target = btrfs_get_alloc_profile(root, 0);
meta_sinfo = __find_space_info(info, alloc_target);
num_bytes = calculate_bytes_needed(root, num_items);
spin_lock(&meta_sinfo->lock);
if (meta_sinfo->bytes_may_use < num_bytes) {
bug = true;
meta_sinfo->bytes_may_use = 0;
} else {
meta_sinfo->bytes_may_use -= num_bytes;
}
spin_unlock(&meta_sinfo->lock);
BUG_ON(bug);
return 0;
}
/*
* Reserve some metadata space for use. We'll calculate the worste case number
* of bytes that would be needed to modify num_items number of items. If we
* have space, fantastic, if not, you get -ENOSPC. Please call
* btrfs_unreserve_metadata_space when you are done for the _SAME_ number of
* items you reserved, since whatever metadata you needed should have already
* been allocated.
*
* This will commit the transaction to make more space if we don't have enough
* metadata space. THe only time we don't do this is if we're reserving space
* inside of a transaction, then we will just return -ENOSPC and it is the
* callers responsibility to handle it properly.
*/
int btrfs_reserve_metadata_space(struct btrfs_root *root, int num_items)
{
struct btrfs_fs_info *info = root->fs_info;
struct btrfs_space_info *meta_sinfo;
u64 num_bytes;
u64 used;
u64 alloc_target;
int retries = 0;
/* get the space info for where the metadata will live */
alloc_target = btrfs_get_alloc_profile(root, 0);
meta_sinfo = __find_space_info(info, alloc_target);
num_bytes = calculate_bytes_needed(root, num_items);
again:
spin_lock(&meta_sinfo->lock);
if (unlikely(!meta_sinfo->bytes_root))
meta_sinfo->bytes_root = calculate_bytes_needed(root, 6);
if (!retries)
meta_sinfo->bytes_may_use += num_bytes;
used = meta_sinfo->bytes_used + meta_sinfo->bytes_reserved +
meta_sinfo->bytes_pinned + meta_sinfo->bytes_readonly +
meta_sinfo->bytes_super + meta_sinfo->bytes_root +
meta_sinfo->bytes_may_use + meta_sinfo->bytes_delalloc;
if (used > meta_sinfo->total_bytes) {
retries++;
if (retries == 1) {
if (maybe_allocate_chunk(root, meta_sinfo))
goto again;
retries++;
} else {
spin_unlock(&meta_sinfo->lock);
}
if (retries == 2) {
flush_delalloc(root, meta_sinfo);
goto again;
}
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
spin_lock(&meta_sinfo->lock);
meta_sinfo->bytes_may_use -= num_bytes;
spin_unlock(&meta_sinfo->lock);
dump_space_info(meta_sinfo, 0, 0);
return -ENOSPC;
}
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
check_force_delalloc(meta_sinfo);
spin_unlock(&meta_sinfo->lock);
return 0;
}
/*
* This will check the space that the inode allocates from to make sure we have
* enough space for bytes.
*/
int btrfs_check_data_free_space(struct btrfs_root *root, struct inode *inode,
u64 bytes)
{
struct btrfs_space_info *data_sinfo;
int ret = 0, committed = 0;
/* make sure bytes are sectorsize aligned */
bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
data_sinfo = BTRFS_I(inode)->space_info;
if (!data_sinfo)
goto alloc;
again:
/* make sure we have enough space to handle the data first */
spin_lock(&data_sinfo->lock);
if (data_sinfo->total_bytes - data_sinfo->bytes_used -
data_sinfo->bytes_delalloc - data_sinfo->bytes_reserved -
data_sinfo->bytes_pinned - data_sinfo->bytes_readonly -
data_sinfo->bytes_may_use - data_sinfo->bytes_super < bytes) {
struct btrfs_trans_handle *trans;
/*
* if we don't have enough free bytes in this space then we need
* to alloc a new chunk.
*/
if (!data_sinfo->full) {
u64 alloc_target;
data_sinfo->force_alloc = 1;
spin_unlock(&data_sinfo->lock);
alloc:
alloc_target = btrfs_get_alloc_profile(root, 1);
trans = btrfs_start_transaction(root, 1);
if (!trans)
return -ENOMEM;
ret = do_chunk_alloc(trans, root->fs_info->extent_root,
bytes + 2 * 1024 * 1024,
alloc_target, 0);
btrfs_end_transaction(trans, root);
if (ret)
return ret;
if (!data_sinfo) {
btrfs_set_inode_space_info(root, inode);
data_sinfo = BTRFS_I(inode)->space_info;
}
goto again;
}
spin_unlock(&data_sinfo->lock);
/* commit the current transaction and try again */
if (!committed && !root->fs_info->open_ioctl_trans) {
committed = 1;
trans = btrfs_join_transaction(root, 1);
if (!trans)
return -ENOMEM;
ret = btrfs_commit_transaction(trans, root);
if (ret)
return ret;
goto again;
}
printk(KERN_ERR "no space left, need %llu, %llu delalloc bytes"
", %llu bytes_used, %llu bytes_reserved, "
"%llu bytes_pinned, %llu bytes_readonly, %llu may use "
"%llu total\n", (unsigned long long)bytes,
(unsigned long long)data_sinfo->bytes_delalloc,
(unsigned long long)data_sinfo->bytes_used,
(unsigned long long)data_sinfo->bytes_reserved,
(unsigned long long)data_sinfo->bytes_pinned,
(unsigned long long)data_sinfo->bytes_readonly,
(unsigned long long)data_sinfo->bytes_may_use,
(unsigned long long)data_sinfo->total_bytes);
return -ENOSPC;
}
data_sinfo->bytes_may_use += bytes;
BTRFS_I(inode)->reserved_bytes += bytes;
spin_unlock(&data_sinfo->lock);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
return 0;
}
/*
* if there was an error for whatever reason after calling
* btrfs_check_data_free_space, call this so we can cleanup the counters.
*/
void btrfs_free_reserved_data_space(struct btrfs_root *root,
struct inode *inode, u64 bytes)
{
struct btrfs_space_info *data_sinfo;
/* make sure bytes are sectorsize aligned */
bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
data_sinfo = BTRFS_I(inode)->space_info;
spin_lock(&data_sinfo->lock);
data_sinfo->bytes_may_use -= bytes;
BTRFS_I(inode)->reserved_bytes -= bytes;
spin_unlock(&data_sinfo->lock);
}
/* called when we are adding a delalloc extent to the inode's io_tree */
void btrfs_delalloc_reserve_space(struct btrfs_root *root, struct inode *inode,
u64 bytes)
{
struct btrfs_space_info *data_sinfo;
/* get the space info for where this inode will be storing its data */
data_sinfo = BTRFS_I(inode)->space_info;
/* make sure we have enough space to handle the data first */
spin_lock(&data_sinfo->lock);
data_sinfo->bytes_delalloc += bytes;
/*
* we are adding a delalloc extent without calling
* btrfs_check_data_free_space first. This happens on a weird
* writepage condition, but shouldn't hurt our accounting
*/
if (unlikely(bytes > BTRFS_I(inode)->reserved_bytes)) {
data_sinfo->bytes_may_use -= BTRFS_I(inode)->reserved_bytes;
BTRFS_I(inode)->reserved_bytes = 0;
} else {
data_sinfo->bytes_may_use -= bytes;
BTRFS_I(inode)->reserved_bytes -= bytes;
}
spin_unlock(&data_sinfo->lock);
}
/* called when we are clearing an delalloc extent from the inode's io_tree */
void btrfs_delalloc_free_space(struct btrfs_root *root, struct inode *inode,
u64 bytes)
{
struct btrfs_space_info *info;
info = BTRFS_I(inode)->space_info;
spin_lock(&info->lock);
info->bytes_delalloc -= bytes;
spin_unlock(&info->lock);
}
static void force_metadata_allocation(struct btrfs_fs_info *info)
{
struct list_head *head = &info->space_info;
struct btrfs_space_info *found;
rcu_read_lock();
list_for_each_entry_rcu(found, head, list) {
if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
found->force_alloc = 1;
}
rcu_read_unlock();
}
static int do_chunk_alloc(struct btrfs_trans_handle *trans,
struct btrfs_root *extent_root, u64 alloc_bytes,
u64 flags, int force)
{
struct btrfs_space_info *space_info;
struct btrfs_fs_info *fs_info = extent_root->fs_info;
u64 thresh;
int ret = 0;
mutex_lock(&fs_info->chunk_mutex);
flags = btrfs_reduce_alloc_profile(extent_root, flags);
space_info = __find_space_info(extent_root->fs_info, flags);
if (!space_info) {
ret = update_space_info(extent_root->fs_info, flags,
0, 0, &space_info);
BUG_ON(ret);
}
BUG_ON(!space_info);
spin_lock(&space_info->lock);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
if (space_info->force_alloc)
force = 1;
if (space_info->full) {
spin_unlock(&space_info->lock);
goto out;
}
thresh = space_info->total_bytes - space_info->bytes_readonly;
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
thresh = div_factor(thresh, 8);
if (!force &&
(space_info->bytes_used + space_info->bytes_pinned +
space_info->bytes_reserved + alloc_bytes) < thresh) {
spin_unlock(&space_info->lock);
goto out;
}
spin_unlock(&space_info->lock);
/*
* if we're doing a data chunk, go ahead and make sure that
* we keep a reasonable number of metadata chunks allocated in the
* FS as well.
*/
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
fs_info->data_chunk_allocations++;
if (!(fs_info->data_chunk_allocations %
fs_info->metadata_ratio))
force_metadata_allocation(fs_info);
}
ret = btrfs_alloc_chunk(trans, extent_root, flags);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
spin_lock(&space_info->lock);
if (ret)
space_info->full = 1;
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
space_info->force_alloc = 0;
spin_unlock(&space_info->lock);
out:
mutex_unlock(&extent_root->fs_info->chunk_mutex);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
return ret;
}
static int update_block_group(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 bytenr, u64 num_bytes, int alloc,
int mark_free)
{
struct btrfs_block_group_cache *cache;
struct btrfs_fs_info *info = root->fs_info;
u64 total = num_bytes;
u64 old_val;
u64 byte_in_group;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
/* block accounting for super block */
spin_lock(&info->delalloc_lock);
old_val = btrfs_super_bytes_used(&info->super_copy);
if (alloc)
old_val += num_bytes;
else
old_val -= num_bytes;
btrfs_set_super_bytes_used(&info->super_copy, old_val);
spin_unlock(&info->delalloc_lock);
while (total) {
cache = btrfs_lookup_block_group(info, bytenr);
Btrfs: batch extent inserts/updates/deletions on the extent root While profiling the allocator I noticed a good amount of time was being spent in finish_current_insert and del_pending_extents, and as the filesystem filled up more and more time was being spent in those functions. This patch aims to try and reduce that problem. This happens two ways 1) track if we tried to delete an extent that we are going to update or insert. Once we get into finish_current_insert we discard any of the extents that were marked for deletion. This saves us from doing unnecessary work almost every time finish_current_insert runs. 2) Batch insertion/updates/deletions. Instead of doing a btrfs_search_slot for each individual extent and doing the needed operation, we instead keep the leaf around and see if there is anything else we can do on that leaf. On the insert case I introduced a btrfs_insert_some_items, which will take an array of keys with an array of data_sizes and try and squeeze in as many of those keys as possible, and then return how many keys it was able to insert. In the update case we search for an extent ref, update the ref and then loop through the leaf to see if any of the other refs we are looking to update are on that leaf, and then once we are done we release the path and search for the next ref we need to update. And finally for the deletion we try and delete the extent+ref in pairs, so we will try to find extent+ref pairs next to the extent we are trying to free and free them in bulk if possible. This along with the other cluster fix that Chris pushed out a bit ago helps make the allocator preform more uniformly as it fills up the disk. There is still a slight drop as we fill up the disk since we start having to stick new blocks in odd places which results in more COW's than on a empty fs, but the drop is not nearly as severe as it was before. Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-11-13 03:19:50 +08:00
if (!cache)
return -1;
byte_in_group = bytenr - cache->key.objectid;
WARN_ON(byte_in_group > cache->key.offset);
spin_lock(&cache->space_info->lock);
spin_lock(&cache->lock);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
cache->dirty = 1;
old_val = btrfs_block_group_used(&cache->item);
num_bytes = min(total, cache->key.offset - byte_in_group);
if (alloc) {
old_val += num_bytes;
btrfs_set_block_group_used(&cache->item, old_val);
cache->reserved -= num_bytes;
cache->space_info->bytes_used += num_bytes;
cache->space_info->bytes_reserved -= num_bytes;
if (cache->ro)
cache->space_info->bytes_readonly -= num_bytes;
spin_unlock(&cache->lock);
spin_unlock(&cache->space_info->lock);
} else {
old_val -= num_bytes;
cache->space_info->bytes_used -= num_bytes;
if (cache->ro)
cache->space_info->bytes_readonly += num_bytes;
btrfs_set_block_group_used(&cache->item, old_val);
spin_unlock(&cache->lock);
spin_unlock(&cache->space_info->lock);
if (mark_free) {
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
int ret;
ret = btrfs_discard_extent(root, bytenr,
num_bytes);
WARN_ON(ret);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
ret = btrfs_add_free_space(cache, bytenr,
num_bytes);
WARN_ON(ret);
}
}
btrfs_put_block_group(cache);
total -= num_bytes;
bytenr += num_bytes;
}
return 0;
}
static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
{
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
struct btrfs_block_group_cache *cache;
u64 bytenr;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
if (!cache)
return 0;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
bytenr = cache->key.objectid;
btrfs_put_block_group(cache);
return bytenr;
}
/*
* this function must be called within transaction
*/
int btrfs_pin_extent(struct btrfs_root *root,
u64 bytenr, u64 num_bytes, int reserved)
{
struct btrfs_fs_info *fs_info = root->fs_info;
struct btrfs_block_group_cache *cache;
cache = btrfs_lookup_block_group(fs_info, bytenr);
BUG_ON(!cache);
spin_lock(&cache->space_info->lock);
spin_lock(&cache->lock);
cache->pinned += num_bytes;
cache->space_info->bytes_pinned += num_bytes;
if (reserved) {
cache->reserved -= num_bytes;
cache->space_info->bytes_reserved -= num_bytes;
}
spin_unlock(&cache->lock);
spin_unlock(&cache->space_info->lock);
btrfs_put_block_group(cache);
set_extent_dirty(fs_info->pinned_extents,
bytenr, bytenr + num_bytes - 1, GFP_NOFS);
return 0;
}
static int update_reserved_extents(struct btrfs_block_group_cache *cache,
u64 num_bytes, int reserve)
{
spin_lock(&cache->space_info->lock);
spin_lock(&cache->lock);
if (reserve) {
cache->reserved += num_bytes;
cache->space_info->bytes_reserved += num_bytes;
} else {
cache->reserved -= num_bytes;
cache->space_info->bytes_reserved -= num_bytes;
}
spin_unlock(&cache->lock);
spin_unlock(&cache->space_info->lock);
return 0;
}
int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
struct btrfs_root *root)
{
struct btrfs_fs_info *fs_info = root->fs_info;
struct btrfs_caching_control *next;
struct btrfs_caching_control *caching_ctl;
struct btrfs_block_group_cache *cache;
down_write(&fs_info->extent_commit_sem);
list_for_each_entry_safe(caching_ctl, next,
&fs_info->caching_block_groups, list) {
cache = caching_ctl->block_group;
if (block_group_cache_done(cache)) {
cache->last_byte_to_unpin = (u64)-1;
list_del_init(&caching_ctl->list);
put_caching_control(caching_ctl);
} else {
cache->last_byte_to_unpin = caching_ctl->progress;
}
}
if (fs_info->pinned_extents == &fs_info->freed_extents[0])
fs_info->pinned_extents = &fs_info->freed_extents[1];
else
fs_info->pinned_extents = &fs_info->freed_extents[0];
up_write(&fs_info->extent_commit_sem);
return 0;
}
static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
{
struct btrfs_fs_info *fs_info = root->fs_info;
struct btrfs_block_group_cache *cache = NULL;
u64 len;
while (start <= end) {
if (!cache ||
start >= cache->key.objectid + cache->key.offset) {
if (cache)
btrfs_put_block_group(cache);
cache = btrfs_lookup_block_group(fs_info, start);
BUG_ON(!cache);
}
len = cache->key.objectid + cache->key.offset - start;
len = min(len, end + 1 - start);
if (start < cache->last_byte_to_unpin) {
len = min(len, cache->last_byte_to_unpin - start);
btrfs_add_free_space(cache, start, len);
}
spin_lock(&cache->space_info->lock);
spin_lock(&cache->lock);
cache->pinned -= len;
cache->space_info->bytes_pinned -= len;
spin_unlock(&cache->lock);
spin_unlock(&cache->space_info->lock);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
start += len;
}
if (cache)
btrfs_put_block_group(cache);
return 0;
}
int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
struct btrfs_root *root)
{
struct btrfs_fs_info *fs_info = root->fs_info;
struct extent_io_tree *unpin;
u64 start;
u64 end;
int ret;
if (fs_info->pinned_extents == &fs_info->freed_extents[0])
unpin = &fs_info->freed_extents[1];
else
unpin = &fs_info->freed_extents[0];
while (1) {
ret = find_first_extent_bit(unpin, 0, &start, &end,
EXTENT_DIRTY);
if (ret)
break;
ret = btrfs_discard_extent(root, start, end + 1 - start);
clear_extent_dirty(unpin, start, end, GFP_NOFS);
unpin_extent_range(root, start, end);
cond_resched();
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
return ret;
}
static int pin_down_bytes(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
u64 bytenr, u64 num_bytes,
int is_data, int reserved,
struct extent_buffer **must_clean)
{
int err = 0;
struct extent_buffer *buf;
if (is_data)
goto pinit;
/*
* discard is sloooow, and so triggering discards on
* individual btree blocks isn't a good plan. Just
* pin everything in discard mode.
*/
if (btrfs_test_opt(root, DISCARD))
goto pinit;
buf = btrfs_find_tree_block(root, bytenr, num_bytes);
if (!buf)
goto pinit;
/* we can reuse a block if it hasn't been written
* and it is from this transaction. We can't
* reuse anything from the tree log root because
* it has tiny sub-transactions.
*/
if (btrfs_buffer_uptodate(buf, 0) &&
btrfs_try_tree_lock(buf)) {
u64 header_owner = btrfs_header_owner(buf);
u64 header_transid = btrfs_header_generation(buf);
if (header_owner != BTRFS_TREE_LOG_OBJECTID &&
header_transid == trans->transid &&
!btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
*must_clean = buf;
return 1;
}
btrfs_tree_unlock(buf);
}
free_extent_buffer(buf);
pinit:
if (path)
btrfs_set_path_blocking(path);
/* unlocks the pinned mutex */
btrfs_pin_extent(root, bytenr, num_bytes, reserved);
BUG_ON(err < 0);
return 0;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 bytenr, u64 num_bytes, u64 parent,
u64 root_objectid, u64 owner_objectid,
u64 owner_offset, int refs_to_drop,
struct btrfs_delayed_extent_op *extent_op)
{
struct btrfs_key key;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_path *path;
struct btrfs_fs_info *info = root->fs_info;
struct btrfs_root *extent_root = info->extent_root;
struct extent_buffer *leaf;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_extent_item *ei;
struct btrfs_extent_inline_ref *iref;
int ret;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
int is_data;
int extent_slot = 0;
int found_extent = 0;
int num_to_del = 1;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
u32 item_size;
u64 refs;
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
path->reada = 1;
path->leave_spinning = 1;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
BUG_ON(!is_data && refs_to_drop != 1);
ret = lookup_extent_backref(trans, extent_root, path, &iref,
bytenr, num_bytes, parent,
root_objectid, owner_objectid,
owner_offset);
if (ret == 0) {
extent_slot = path->slots[0];
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
while (extent_slot >= 0) {
btrfs_item_key_to_cpu(path->nodes[0], &key,
extent_slot);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (key.objectid != bytenr)
break;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (key.type == BTRFS_EXTENT_ITEM_KEY &&
key.offset == num_bytes) {
found_extent = 1;
break;
}
if (path->slots[0] - extent_slot > 5)
break;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
extent_slot--;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
if (found_extent && item_size < sizeof(*ei))
found_extent = 0;
#endif
if (!found_extent) {
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
BUG_ON(iref);
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
ret = remove_extent_backref(trans, extent_root, path,
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
NULL, refs_to_drop,
is_data);
BUG_ON(ret);
btrfs_release_path(extent_root, path);
path->leave_spinning = 1;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
key.objectid = bytenr;
key.type = BTRFS_EXTENT_ITEM_KEY;
key.offset = num_bytes;
ret = btrfs_search_slot(trans, extent_root,
&key, path, -1, 1);
Btrfs: batch extent inserts/updates/deletions on the extent root While profiling the allocator I noticed a good amount of time was being spent in finish_current_insert and del_pending_extents, and as the filesystem filled up more and more time was being spent in those functions. This patch aims to try and reduce that problem. This happens two ways 1) track if we tried to delete an extent that we are going to update or insert. Once we get into finish_current_insert we discard any of the extents that were marked for deletion. This saves us from doing unnecessary work almost every time finish_current_insert runs. 2) Batch insertion/updates/deletions. Instead of doing a btrfs_search_slot for each individual extent and doing the needed operation, we instead keep the leaf around and see if there is anything else we can do on that leaf. On the insert case I introduced a btrfs_insert_some_items, which will take an array of keys with an array of data_sizes and try and squeeze in as many of those keys as possible, and then return how many keys it was able to insert. In the update case we search for an extent ref, update the ref and then loop through the leaf to see if any of the other refs we are looking to update are on that leaf, and then once we are done we release the path and search for the next ref we need to update. And finally for the deletion we try and delete the extent+ref in pairs, so we will try to find extent+ref pairs next to the extent we are trying to free and free them in bulk if possible. This along with the other cluster fix that Chris pushed out a bit ago helps make the allocator preform more uniformly as it fills up the disk. There is still a slight drop as we fill up the disk since we start having to stick new blocks in odd places which results in more COW's than on a empty fs, but the drop is not nearly as severe as it was before. Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-11-13 03:19:50 +08:00
if (ret) {
printk(KERN_ERR "umm, got %d back from search"
", was looking for %llu\n", ret,
(unsigned long long)bytenr);
Btrfs: batch extent inserts/updates/deletions on the extent root While profiling the allocator I noticed a good amount of time was being spent in finish_current_insert and del_pending_extents, and as the filesystem filled up more and more time was being spent in those functions. This patch aims to try and reduce that problem. This happens two ways 1) track if we tried to delete an extent that we are going to update or insert. Once we get into finish_current_insert we discard any of the extents that were marked for deletion. This saves us from doing unnecessary work almost every time finish_current_insert runs. 2) Batch insertion/updates/deletions. Instead of doing a btrfs_search_slot for each individual extent and doing the needed operation, we instead keep the leaf around and see if there is anything else we can do on that leaf. On the insert case I introduced a btrfs_insert_some_items, which will take an array of keys with an array of data_sizes and try and squeeze in as many of those keys as possible, and then return how many keys it was able to insert. In the update case we search for an extent ref, update the ref and then loop through the leaf to see if any of the other refs we are looking to update are on that leaf, and then once we are done we release the path and search for the next ref we need to update. And finally for the deletion we try and delete the extent+ref in pairs, so we will try to find extent+ref pairs next to the extent we are trying to free and free them in bulk if possible. This along with the other cluster fix that Chris pushed out a bit ago helps make the allocator preform more uniformly as it fills up the disk. There is still a slight drop as we fill up the disk since we start having to stick new blocks in odd places which results in more COW's than on a empty fs, but the drop is not nearly as severe as it was before. Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-11-13 03:19:50 +08:00
btrfs_print_leaf(extent_root, path->nodes[0]);
}
BUG_ON(ret);
extent_slot = path->slots[0];
}
} else {
btrfs_print_leaf(extent_root, path->nodes[0]);
WARN_ON(1);
printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
"parent %llu root %llu owner %llu offset %llu\n",
(unsigned long long)bytenr,
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
(unsigned long long)parent,
(unsigned long long)root_objectid,
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
(unsigned long long)owner_objectid,
(unsigned long long)owner_offset);
}
leaf = path->nodes[0];
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
item_size = btrfs_item_size_nr(leaf, extent_slot);
#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
if (item_size < sizeof(*ei)) {
BUG_ON(found_extent || extent_slot != path->slots[0]);
ret = convert_extent_item_v0(trans, extent_root, path,
owner_objectid, 0);
BUG_ON(ret < 0);
btrfs_release_path(extent_root, path);
path->leave_spinning = 1;
key.objectid = bytenr;
key.type = BTRFS_EXTENT_ITEM_KEY;
key.offset = num_bytes;
ret = btrfs_search_slot(trans, extent_root, &key, path,
-1, 1);
if (ret) {
printk(KERN_ERR "umm, got %d back from search"
", was looking for %llu\n", ret,
(unsigned long long)bytenr);
btrfs_print_leaf(extent_root, path->nodes[0]);
}
BUG_ON(ret);
extent_slot = path->slots[0];
leaf = path->nodes[0];
item_size = btrfs_item_size_nr(leaf, extent_slot);
}
#endif
BUG_ON(item_size < sizeof(*ei));
ei = btrfs_item_ptr(leaf, extent_slot,
struct btrfs_extent_item);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
struct btrfs_tree_block_info *bi;
BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
bi = (struct btrfs_tree_block_info *)(ei + 1);
WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
}
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
refs = btrfs_extent_refs(leaf, ei);
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
BUG_ON(refs < refs_to_drop);
refs -= refs_to_drop;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (refs > 0) {
if (extent_op)
__run_delayed_extent_op(extent_op, leaf, ei);
/*
* In the case of inline back ref, reference count will
* be updated by remove_extent_backref
*/
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (iref) {
BUG_ON(!found_extent);
} else {
btrfs_set_extent_refs(leaf, ei, refs);
btrfs_mark_buffer_dirty(leaf);
}
if (found_extent) {
ret = remove_extent_backref(trans, extent_root, path,
iref, refs_to_drop,
is_data);
BUG_ON(ret);
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
} else {
int mark_free = 0;
struct extent_buffer *must_clean = NULL;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (found_extent) {
BUG_ON(is_data && refs_to_drop !=
extent_data_ref_count(root, path, iref));
if (iref) {
BUG_ON(path->slots[0] != extent_slot);
} else {
BUG_ON(path->slots[0] != extent_slot + 1);
path->slots[0] = extent_slot;
num_to_del = 2;
}
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
ret = pin_down_bytes(trans, root, path, bytenr,
num_bytes, is_data, 0, &must_clean);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (ret > 0)
mark_free = 1;
BUG_ON(ret < 0);
/*
* it is going to be very rare for someone to be waiting
* on the block we're freeing. del_items might need to
* schedule, so rather than get fancy, just force it
* to blocking here
*/
if (must_clean)
btrfs_set_lock_blocking(must_clean);
ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
num_to_del);
BUG_ON(ret);
btrfs_release_path(extent_root, path);
if (must_clean) {
clean_tree_block(NULL, root, must_clean);
btrfs_tree_unlock(must_clean);
free_extent_buffer(must_clean);
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (is_data) {
ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
BUG_ON(ret);
} else {
invalidate_mapping_pages(info->btree_inode->i_mapping,
bytenr >> PAGE_CACHE_SHIFT,
(bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
}
ret = update_block_group(trans, root, bytenr, num_bytes, 0,
mark_free);
BUG_ON(ret);
}
btrfs_free_path(path);
return ret;
}
/*
* when we free an extent, it is possible (and likely) that we free the last
* delayed ref for that extent as well. This searches the delayed ref tree for
* a given extent, and if there are no other delayed refs to be processed, it
* removes it from the tree.
*/
static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
struct btrfs_root *root, u64 bytenr)
{
struct btrfs_delayed_ref_head *head;
struct btrfs_delayed_ref_root *delayed_refs;
struct btrfs_delayed_ref_node *ref;
struct rb_node *node;
int ret;
delayed_refs = &trans->transaction->delayed_refs;
spin_lock(&delayed_refs->lock);
head = btrfs_find_delayed_ref_head(trans, bytenr);
if (!head)
goto out;
node = rb_prev(&head->node.rb_node);
if (!node)
goto out;
ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
/* there are still entries for this ref, we can't drop it */
if (ref->bytenr == bytenr)
goto out;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (head->extent_op) {
if (!head->must_insert_reserved)
goto out;
kfree(head->extent_op);
head->extent_op = NULL;
}
/*
* waiting for the lock here would deadlock. If someone else has it
* locked they are already in the process of dropping it anyway
*/
if (!mutex_trylock(&head->mutex))
goto out;
/*
* at this point we have a head with no other entries. Go
* ahead and process it.
*/
head->node.in_tree = 0;
rb_erase(&head->node.rb_node, &delayed_refs->root);
delayed_refs->num_entries--;
/*
* we don't take a ref on the node because we're removing it from the
* tree, so we just steal the ref the tree was holding.
*/
delayed_refs->num_heads--;
if (list_empty(&head->cluster))
delayed_refs->num_heads_ready--;
list_del_init(&head->cluster);
spin_unlock(&delayed_refs->lock);
ret = run_one_delayed_ref(trans, root->fs_info->tree_root,
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
&head->node, head->extent_op,
head->must_insert_reserved);
BUG_ON(ret);
btrfs_put_delayed_ref(&head->node);
return 0;
out:
spin_unlock(&delayed_refs->lock);
return 0;
}
int btrfs_free_extent(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 bytenr, u64 num_bytes, u64 parent,
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
u64 root_objectid, u64 owner, u64 offset)
{
int ret;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
/*
* tree log blocks never actually go into the extent allocation
* tree, just update pinning info and exit early.
*/
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
/* unlocks the pinned mutex */
btrfs_pin_extent(root, bytenr, num_bytes, 1);
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
ret = 0;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
} else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
parent, root_objectid, (int)owner,
BTRFS_DROP_DELAYED_REF, NULL);
BUG_ON(ret);
ret = check_ref_cleanup(trans, root, bytenr);
BUG_ON(ret);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
} else {
ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
parent, root_objectid, owner,
offset, BTRFS_DROP_DELAYED_REF, NULL);
BUG_ON(ret);
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
}
return ret;
}
int btrfs_free_tree_block(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 bytenr, u32 blocksize,
u64 parent, u64 root_objectid, int level)
{
u64 used;
spin_lock(&root->node_lock);
used = btrfs_root_used(&root->root_item) - blocksize;
btrfs_set_root_used(&root->root_item, used);
spin_unlock(&root->node_lock);
return btrfs_free_extent(trans, root, bytenr, blocksize,
parent, root_objectid, level, 0);
}
static u64 stripe_align(struct btrfs_root *root, u64 val)
{
u64 mask = ((u64)root->stripesize - 1);
u64 ret = (val + mask) & ~mask;
return ret;
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
/*
* when we wait for progress in the block group caching, its because
* our allocation attempt failed at least once. So, we must sleep
* and let some progress happen before we try again.
*
* This function will sleep at least once waiting for new free space to
* show up, and then it will check the block group free space numbers
* for our min num_bytes. Another option is to have it go ahead
* and look in the rbtree for a free extent of a given size, but this
* is a good start.
*/
static noinline int
wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
u64 num_bytes)
{
struct btrfs_caching_control *caching_ctl;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
DEFINE_WAIT(wait);
caching_ctl = get_caching_control(cache);
if (!caching_ctl)
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
return 0;
wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
(cache->free_space >= num_bytes));
put_caching_control(caching_ctl);
return 0;
}
static noinline int
wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
{
struct btrfs_caching_control *caching_ctl;
DEFINE_WAIT(wait);
caching_ctl = get_caching_control(cache);
if (!caching_ctl)
return 0;
wait_event(caching_ctl->wait, block_group_cache_done(cache));
put_caching_control(caching_ctl);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
return 0;
}
enum btrfs_loop_type {
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
LOOP_FIND_IDEAL = 0,
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
LOOP_CACHING_NOWAIT = 1,
LOOP_CACHING_WAIT = 2,
LOOP_ALLOC_CHUNK = 3,
LOOP_NO_EMPTY_SIZE = 4,
};
/*
* walks the btree of allocated extents and find a hole of a given size.
* The key ins is changed to record the hole:
* ins->objectid == block start
* ins->flags = BTRFS_EXTENT_ITEM_KEY
* ins->offset == number of blocks
* Any available blocks before search_start are skipped.
*/
static noinline int find_free_extent(struct btrfs_trans_handle *trans,
struct btrfs_root *orig_root,
u64 num_bytes, u64 empty_size,
u64 search_start, u64 search_end,
u64 hint_byte, struct btrfs_key *ins,
u64 exclude_start, u64 exclude_nr,
int data)
{
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
int ret = 0;
struct btrfs_root *root = orig_root->fs_info->extent_root;
struct btrfs_free_cluster *last_ptr = NULL;
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
struct btrfs_block_group_cache *block_group = NULL;
int empty_cluster = 2 * 1024 * 1024;
int allowed_chunk_alloc = 0;
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
int done_chunk_alloc = 0;
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
struct btrfs_space_info *space_info;
int last_ptr_loop = 0;
int loop = 0;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
bool found_uncached_bg = false;
bool failed_cluster_refill = false;
bool failed_alloc = false;
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
u64 ideal_cache_percent = 0;
u64 ideal_cache_offset = 0;
WARN_ON(num_bytes < root->sectorsize);
btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
ins->objectid = 0;
ins->offset = 0;
space_info = __find_space_info(root->fs_info, data);
if (orig_root->ref_cows || empty_size)
allowed_chunk_alloc = 1;
if (data & BTRFS_BLOCK_GROUP_METADATA) {
last_ptr = &root->fs_info->meta_alloc_cluster;
if (!btrfs_test_opt(root, SSD))
empty_cluster = 64 * 1024;
}
if ((data & BTRFS_BLOCK_GROUP_DATA) && btrfs_test_opt(root, SSD)) {
last_ptr = &root->fs_info->data_alloc_cluster;
}
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
if (last_ptr) {
spin_lock(&last_ptr->lock);
if (last_ptr->block_group)
hint_byte = last_ptr->window_start;
spin_unlock(&last_ptr->lock);
}
search_start = max(search_start, first_logical_byte(root, 0));
search_start = max(search_start, hint_byte);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
if (!last_ptr)
empty_cluster = 0;
if (search_start == hint_byte) {
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
ideal_cache:
block_group = btrfs_lookup_block_group(root->fs_info,
search_start);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
/*
* we don't want to use the block group if it doesn't match our
* allocation bits, or if its not cached.
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
*
* However if we are re-searching with an ideal block group
* picked out then we don't care that the block group is cached.
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
*/
if (block_group && block_group_bits(block_group, data) &&
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
(block_group->cached != BTRFS_CACHE_NO ||
search_start == ideal_cache_offset)) {
down_read(&space_info->groups_sem);
if (list_empty(&block_group->list) ||
block_group->ro) {
/*
* someone is removing this block group,
* we can't jump into the have_block_group
* target because our list pointers are not
* valid
*/
btrfs_put_block_group(block_group);
up_read(&space_info->groups_sem);
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
} else {
goto have_block_group;
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
}
} else if (block_group) {
btrfs_put_block_group(block_group);
}
}
search:
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
down_read(&space_info->groups_sem);
list_for_each_entry(block_group, &space_info->block_groups, list) {
u64 offset;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
int cached;
atomic_inc(&block_group->count);
search_start = block_group->key.objectid;
have_block_group:
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
if (unlikely(block_group->cached == BTRFS_CACHE_NO)) {
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
u64 free_percent;
free_percent = btrfs_block_group_used(&block_group->item);
free_percent *= 100;
free_percent = div64_u64(free_percent,
block_group->key.offset);
free_percent = 100 - free_percent;
if (free_percent > ideal_cache_percent &&
likely(!block_group->ro)) {
ideal_cache_offset = block_group->key.objectid;
ideal_cache_percent = free_percent;
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
/*
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
* We only want to start kthread caching if we are at
* the point where we will wait for caching to make
* progress, or if our ideal search is over and we've
* found somebody to start caching.
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
*/
if (loop > LOOP_CACHING_NOWAIT ||
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
(loop > LOOP_FIND_IDEAL &&
atomic_read(&space_info->caching_threads) < 2)) {
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
ret = cache_block_group(block_group);
BUG_ON(ret);
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
found_uncached_bg = true;
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
/*
* If loop is set for cached only, try the next block
* group.
*/
if (loop == LOOP_FIND_IDEAL)
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
goto loop;
}
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
cached = block_group_cache_done(block_group);
if (unlikely(!cached))
found_uncached_bg = true;
if (unlikely(block_group->ro))
goto loop;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
/*
* Ok we want to try and use the cluster allocator, so lets look
* there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
* have tried the cluster allocator plenty of times at this
* point and not have found anything, so we are likely way too
* fragmented for the clustering stuff to find anything, so lets
* just skip it and let the allocator find whatever block it can
* find
*/
if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
/*
* the refill lock keeps out other
* people trying to start a new cluster
*/
spin_lock(&last_ptr->refill_lock);
if (last_ptr->block_group &&
(last_ptr->block_group->ro ||
!block_group_bits(last_ptr->block_group, data))) {
offset = 0;
goto refill_cluster;
}
offset = btrfs_alloc_from_cluster(block_group, last_ptr,
num_bytes, search_start);
if (offset) {
/* we have a block, we're done */
spin_unlock(&last_ptr->refill_lock);
goto checks;
}
spin_lock(&last_ptr->lock);
/*
* whoops, this cluster doesn't actually point to
* this block group. Get a ref on the block
* group is does point to and try again
*/
if (!last_ptr_loop && last_ptr->block_group &&
last_ptr->block_group != block_group) {
btrfs_put_block_group(block_group);
block_group = last_ptr->block_group;
atomic_inc(&block_group->count);
spin_unlock(&last_ptr->lock);
spin_unlock(&last_ptr->refill_lock);
last_ptr_loop = 1;
search_start = block_group->key.objectid;
/*
* we know this block group is properly
* in the list because
* btrfs_remove_block_group, drops the
* cluster before it removes the block
* group from the list
*/
goto have_block_group;
}
spin_unlock(&last_ptr->lock);
refill_cluster:
/*
* this cluster didn't work out, free it and
* start over
*/
btrfs_return_cluster_to_free_space(NULL, last_ptr);
last_ptr_loop = 0;
/* allocate a cluster in this block group */
ret = btrfs_find_space_cluster(trans, root,
block_group, last_ptr,
offset, num_bytes,
empty_cluster + empty_size);
if (ret == 0) {
/*
* now pull our allocation out of this
* cluster
*/
offset = btrfs_alloc_from_cluster(block_group,
last_ptr, num_bytes,
search_start);
if (offset) {
/* we found one, proceed */
spin_unlock(&last_ptr->refill_lock);
goto checks;
}
} else if (!cached && loop > LOOP_CACHING_NOWAIT
&& !failed_cluster_refill) {
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
spin_unlock(&last_ptr->refill_lock);
failed_cluster_refill = true;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
wait_block_group_cache_progress(block_group,
num_bytes + empty_cluster + empty_size);
goto have_block_group;
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
/*
* at this point we either didn't find a cluster
* or we weren't able to allocate a block from our
* cluster. Free the cluster we've been trying
* to use, and go to the next block group
*/
btrfs_return_cluster_to_free_space(NULL, last_ptr);
spin_unlock(&last_ptr->refill_lock);
goto loop;
}
offset = btrfs_find_space_for_alloc(block_group, search_start,
num_bytes, empty_size);
/*
* If we didn't find a chunk, and we haven't failed on this
* block group before, and this block group is in the middle of
* caching and we are ok with waiting, then go ahead and wait
* for progress to be made, and set failed_alloc to true.
*
* If failed_alloc is true then we've already waited on this
* block group once and should move on to the next block group.
*/
if (!offset && !failed_alloc && !cached &&
loop > LOOP_CACHING_NOWAIT) {
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
wait_block_group_cache_progress(block_group,
num_bytes + empty_size);
failed_alloc = true;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
goto have_block_group;
} else if (!offset) {
goto loop;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
}
checks:
search_start = stripe_align(root, offset);
/* move on to the next group */
if (search_start + num_bytes >= search_end) {
btrfs_add_free_space(block_group, offset, num_bytes);
goto loop;
}
/* move on to the next group */
if (search_start + num_bytes >
block_group->key.objectid + block_group->key.offset) {
btrfs_add_free_space(block_group, offset, num_bytes);
goto loop;
}
if (exclude_nr > 0 &&
(search_start + num_bytes > exclude_start &&
search_start < exclude_start + exclude_nr)) {
search_start = exclude_start + exclude_nr;
btrfs_add_free_space(block_group, offset, num_bytes);
/*
* if search_start is still in this block group
* then we just re-search this block group
*/
if (search_start >= block_group->key.objectid &&
search_start < (block_group->key.objectid +
block_group->key.offset))
goto have_block_group;
goto loop;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
}
ins->objectid = search_start;
ins->offset = num_bytes;
if (offset < search_start)
btrfs_add_free_space(block_group, offset,
search_start - offset);
BUG_ON(offset > search_start);
update_reserved_extents(block_group, num_bytes, 1);
/* we are all good, lets return */
break;
loop:
failed_cluster_refill = false;
failed_alloc = false;
btrfs_put_block_group(block_group);
}
up_read(&space_info->groups_sem);
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
/* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
* for them to make caching progress. Also
* determine the best possible bg to cache
* LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
* caching kthreads as we move along
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
* LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
* LOOP_ALLOC_CHUNK, force a chunk allocation and try again
* LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
* again
*/
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE &&
(found_uncached_bg || empty_size || empty_cluster ||
allowed_chunk_alloc)) {
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
found_uncached_bg = false;
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
loop++;
if (!ideal_cache_percent &&
atomic_read(&space_info->caching_threads))
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
goto search;
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
/*
* 1 of the following 2 things have happened so far
*
* 1) We found an ideal block group for caching that
* is mostly full and will cache quickly, so we might
* as well wait for it.
*
* 2) We searched for cached only and we didn't find
* anything, and we didn't start any caching kthreads
* either, so chances are we will loop through and
* start a couple caching kthreads, and then come back
* around and just wait for them. This will be slower
* because we will have 2 caching kthreads reading at
* the same time when we could have just started one
* and waited for it to get far enough to give us an
* allocation, so go ahead and go to the wait caching
* loop.
*/
loop = LOOP_CACHING_WAIT;
search_start = ideal_cache_offset;
ideal_cache_percent = 0;
goto ideal_cache;
} else if (loop == LOOP_FIND_IDEAL) {
/*
* Didn't find a uncached bg, wait on anything we find
* next.
*/
loop = LOOP_CACHING_WAIT;
goto search;
}
if (loop < LOOP_CACHING_WAIT) {
loop++;
goto search;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
}
if (loop == LOOP_ALLOC_CHUNK) {
empty_size = 0;
empty_cluster = 0;
}
if (allowed_chunk_alloc) {
ret = do_chunk_alloc(trans, root, num_bytes +
2 * 1024 * 1024, data, 1);
allowed_chunk_alloc = 0;
Btrfs: find ideal block group for caching This patch changes a few things. Hopefully the comments are helpfull, but I'll try and be as verbose here. Problem: My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root. Part of this problem was we pick the first block group we can find and start caching it, even if it may not have enough free space. The other problem is we only search for cached block groups the first time around, which we won't find any cached block groups because this is a newly mounted fs, so we end up caching several block groups during bootup, which with alot of fragmentation takes around 30-45 seconds to complete, which bogs down the system. So Solution: 1) Don't cache block groups willy-nilly at first. Instead try and figure out which block group has the most free, and therefore will take the least amount of time to cache. 2) Don't be so picky about cached block groups. The other problem is once we've filled up a cluster, if the block group isn't finished caching the next time we try and do the allocation we'll completely ignore the cluster and start searching from the beginning of the space, which makes us cache more block groups, which slows us down even more. So instead of skipping block groups that are not finished caching when we have a hint, only skip the block group if it hasn't started caching yet. There is one other tweak in here. Before if we allocated a chunk and still couldn't find new space, we'd end up switching the space info to force another chunk allocation. This could make us end up with way too many chunks, so keep track of this particular case. With this patch and my previous cluster fixes my fedora box now boots in 43 seconds, and according to the bootchart is not held up by our block group caching at all. Signed-off-by: Josef Bacik <josef@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-11-11 10:23:48 +08:00
done_chunk_alloc = 1;
} else if (!done_chunk_alloc) {
space_info->force_alloc = 1;
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
if (loop < LOOP_NO_EMPTY_SIZE) {
loop++;
goto search;
}
ret = -ENOSPC;
} else if (!ins->objectid) {
ret = -ENOSPC;
}
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
/* we found what we needed */
if (ins->objectid) {
if (!(data & BTRFS_BLOCK_GROUP_DATA))
trans->block_group = block_group->key.objectid;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
btrfs_put_block_group(block_group);
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
ret = 0;
}
return ret;
}
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
int dump_block_groups)
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
{
struct btrfs_block_group_cache *cache;
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
spin_lock(&info->lock);
printk(KERN_INFO "space_info has %llu free, is %sfull\n",
(unsigned long long)(info->total_bytes - info->bytes_used -
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
info->bytes_pinned - info->bytes_reserved -
info->bytes_super),
(info->full) ? "" : "not ");
printk(KERN_INFO "space_info total=%llu, pinned=%llu, delalloc=%llu,"
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
" may_use=%llu, used=%llu, root=%llu, super=%llu, reserved=%llu"
"\n",
(unsigned long long)info->total_bytes,
(unsigned long long)info->bytes_pinned,
(unsigned long long)info->bytes_delalloc,
(unsigned long long)info->bytes_may_use,
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
(unsigned long long)info->bytes_used,
(unsigned long long)info->bytes_root,
(unsigned long long)info->bytes_super,
(unsigned long long)info->bytes_reserved);
spin_unlock(&info->lock);
if (!dump_block_groups)
return;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
down_read(&info->groups_sem);
list_for_each_entry(cache, &info->block_groups, list) {
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
spin_lock(&cache->lock);
printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
"%llu pinned %llu reserved\n",
(unsigned long long)cache->key.objectid,
(unsigned long long)cache->key.offset,
(unsigned long long)btrfs_block_group_used(&cache->item),
(unsigned long long)cache->pinned,
(unsigned long long)cache->reserved);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
btrfs_dump_free_space(cache, bytes);
spin_unlock(&cache->lock);
}
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
up_read(&info->groups_sem);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
}
int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 num_bytes, u64 min_alloc_size,
u64 empty_size, u64 hint_byte,
u64 search_end, struct btrfs_key *ins,
u64 data)
{
int ret;
u64 search_start = 0;
data = btrfs_get_alloc_profile(root, data);
again:
/*
* the only place that sets empty_size is btrfs_realloc_node, which
* is not called recursively on allocations
*/
if (empty_size || root->ref_cows)
ret = do_chunk_alloc(trans, root->fs_info->extent_root,
num_bytes + 2 * 1024 * 1024, data, 0);
WARN_ON(num_bytes < root->sectorsize);
ret = find_free_extent(trans, root, num_bytes, empty_size,
search_start, search_end, hint_byte, ins,
trans->alloc_exclude_start,
trans->alloc_exclude_nr, data);
if (ret == -ENOSPC && num_bytes > min_alloc_size) {
num_bytes = num_bytes >> 1;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
num_bytes = num_bytes & ~(root->sectorsize - 1);
num_bytes = max(num_bytes, min_alloc_size);
do_chunk_alloc(trans, root->fs_info->extent_root,
num_bytes, data, 1);
goto again;
}
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
if (ret == -ENOSPC) {
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
struct btrfs_space_info *sinfo;
sinfo = __find_space_info(root->fs_info, data);
printk(KERN_ERR "btrfs allocation failed flags %llu, "
"wanted %llu\n", (unsigned long long)data,
(unsigned long long)num_bytes);
Btrfs: proper -ENOSPC handling At the start of a transaction we do a btrfs_reserve_metadata_space() and specify how many items we plan on modifying. Then once we've done our modifications and such, just call btrfs_unreserve_metadata_space() for the same number of items we reserved. For keeping track of metadata needed for data I've had to add an extent_io op for when we merge extents. This lets us track space properly when we are doing sequential writes, so we don't end up reserving way more metadata space than what we need. The only place where the metadata space accounting is not done is in the relocation code. This is because Yan is going to be reworking that code in the near future, so running btrfs-vol -b could still possibly result in a ENOSPC related panic. This patch also turns off the metadata_ratio stuff in order to allow users to more efficiently use their disk space. This patch makes it so we track how much metadata we need for an inode's delayed allocation extents by tracking how many extents are currently waiting for allocation. It introduces two new callbacks for the extent_io tree's, merge_extent_hook and split_extent_hook. These help us keep track of when we merge delalloc extents together and split them up. Reservations are handled prior to any actually dirty'ing occurs, and then we unreserve after we dirty. btrfs_unreserve_metadata_for_delalloc() will make the appropriate unreservations as needed based on the number of reservations we currently have and the number of extents we currently have. Doing the reservation outside of doing any of the actual dirty'ing lets us do things like filemap_flush() the inode to try and force delalloc to happen, or as a last resort actually start allocation on all delalloc inodes in the fs. This has survived dbench, fs_mark and an fsx torture test. Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-09-12 04:12:44 +08:00
dump_space_info(sinfo, num_bytes, 1);
}
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
return ret;
}
int btrfs_free_reserved_extent(struct btrfs_root *root, u64 start, u64 len)
{
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
struct btrfs_block_group_cache *cache;
int ret = 0;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
cache = btrfs_lookup_block_group(root->fs_info, start);
if (!cache) {
printk(KERN_ERR "Unable to find block group for %llu\n",
(unsigned long long)start);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
return -ENOSPC;
}
ret = btrfs_discard_extent(root, start, len);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
btrfs_add_free_space(cache, start, len);
update_reserved_extents(cache, len, 0);
btrfs_put_block_group(cache);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
return ret;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 parent, u64 root_objectid,
u64 flags, u64 owner, u64 offset,
struct btrfs_key *ins, int ref_mod)
{
int ret;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_fs_info *fs_info = root->fs_info;
struct btrfs_extent_item *extent_item;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_extent_inline_ref *iref;
struct btrfs_path *path;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct extent_buffer *leaf;
int type;
u32 size;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (parent > 0)
type = BTRFS_SHARED_DATA_REF_KEY;
else
type = BTRFS_EXTENT_DATA_REF_KEY;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
path = btrfs_alloc_path();
BUG_ON(!path);
path->leave_spinning = 1;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
ins, size);
BUG_ON(ret);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
leaf = path->nodes[0];
extent_item = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_extent_item);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
btrfs_set_extent_refs(leaf, extent_item, ref_mod);
btrfs_set_extent_generation(leaf, extent_item, trans->transid);
btrfs_set_extent_flags(leaf, extent_item,
flags | BTRFS_EXTENT_FLAG_DATA);
iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
btrfs_set_extent_inline_ref_type(leaf, iref, type);
if (parent > 0) {
struct btrfs_shared_data_ref *ref;
ref = (struct btrfs_shared_data_ref *)(iref + 1);
btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
} else {
struct btrfs_extent_data_ref *ref;
ref = (struct btrfs_extent_data_ref *)(&iref->offset);
btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
btrfs_set_extent_data_ref_offset(leaf, ref, offset);
btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
}
btrfs_mark_buffer_dirty(path->nodes[0]);
btrfs_free_path(path);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
ret = update_block_group(trans, root, ins->objectid, ins->offset,
1, 0);
if (ret) {
printk(KERN_ERR "btrfs update block group failed for %llu "
"%llu\n", (unsigned long long)ins->objectid,
(unsigned long long)ins->offset);
BUG();
}
return ret;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 parent, u64 root_objectid,
u64 flags, struct btrfs_disk_key *key,
int level, struct btrfs_key *ins)
{
int ret;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_fs_info *fs_info = root->fs_info;
struct btrfs_extent_item *extent_item;
struct btrfs_tree_block_info *block_info;
struct btrfs_extent_inline_ref *iref;
struct btrfs_path *path;
struct extent_buffer *leaf;
u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
path = btrfs_alloc_path();
BUG_ON(!path);
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
path->leave_spinning = 1;
ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
ins, size);
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
BUG_ON(ret);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
leaf = path->nodes[0];
extent_item = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_extent_item);
btrfs_set_extent_refs(leaf, extent_item, 1);
btrfs_set_extent_generation(leaf, extent_item, trans->transid);
btrfs_set_extent_flags(leaf, extent_item,
flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
btrfs_set_tree_block_key(leaf, block_info, key);
btrfs_set_tree_block_level(leaf, block_info, level);
iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
if (parent > 0) {
BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
btrfs_set_extent_inline_ref_type(leaf, iref,
BTRFS_SHARED_BLOCK_REF_KEY);
btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
} else {
btrfs_set_extent_inline_ref_type(leaf, iref,
BTRFS_TREE_BLOCK_REF_KEY);
btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
}
btrfs_mark_buffer_dirty(leaf);
btrfs_free_path(path);
ret = update_block_group(trans, root, ins->objectid, ins->offset,
1, 0);
if (ret) {
printk(KERN_ERR "btrfs update block group failed for %llu "
"%llu\n", (unsigned long long)ins->objectid,
(unsigned long long)ins->offset);
BUG();
}
return ret;
}
int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 root_objectid, u64 owner,
u64 offset, struct btrfs_key *ins)
{
int ret;
BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
ret = btrfs_add_delayed_data_ref(trans, ins->objectid, ins->offset,
0, root_objectid, owner, offset,
BTRFS_ADD_DELAYED_EXTENT, NULL);
return ret;
}
/*
* this is used by the tree logging recovery code. It records that
* an extent has been allocated and makes sure to clear the free
* space cache bits as well
*/
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 root_objectid, u64 owner, u64 offset,
struct btrfs_key *ins)
{
int ret;
struct btrfs_block_group_cache *block_group;
struct btrfs_caching_control *caching_ctl;
u64 start = ins->objectid;
u64 num_bytes = ins->offset;
block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
cache_block_group(block_group);
caching_ctl = get_caching_control(block_group);
if (!caching_ctl) {
BUG_ON(!block_group_cache_done(block_group));
ret = btrfs_remove_free_space(block_group, start, num_bytes);
BUG_ON(ret);
} else {
mutex_lock(&caching_ctl->mutex);
if (start >= caching_ctl->progress) {
ret = add_excluded_extent(root, start, num_bytes);
BUG_ON(ret);
} else if (start + num_bytes <= caching_ctl->progress) {
ret = btrfs_remove_free_space(block_group,
start, num_bytes);
BUG_ON(ret);
} else {
num_bytes = caching_ctl->progress - start;
ret = btrfs_remove_free_space(block_group,
start, num_bytes);
BUG_ON(ret);
start = caching_ctl->progress;
num_bytes = ins->objectid + ins->offset -
caching_ctl->progress;
ret = add_excluded_extent(root, start, num_bytes);
BUG_ON(ret);
}
mutex_unlock(&caching_ctl->mutex);
put_caching_control(caching_ctl);
}
update_reserved_extents(block_group, ins->offset, 1);
btrfs_put_block_group(block_group);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
0, owner, offset, ins, 1);
return ret;
}
/*
* finds a free extent and does all the dirty work required for allocation
* returns the key for the extent through ins, and a tree buffer for
* the first block of the extent through buf.
*
* returns 0 if everything worked, non-zero otherwise.
*/
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
static int alloc_tree_block(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 num_bytes, u64 parent, u64 root_objectid,
struct btrfs_disk_key *key, int level,
u64 empty_size, u64 hint_byte, u64 search_end,
struct btrfs_key *ins)
{
int ret;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
u64 flags = 0;
ret = btrfs_reserve_extent(trans, root, num_bytes, num_bytes,
empty_size, hint_byte, search_end,
ins, 0);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
if (ret)
return ret;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
if (parent == 0)
parent = ins->objectid;
flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
} else
BUG_ON(parent > 0);
if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_delayed_extent_op *extent_op;
extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
BUG_ON(!extent_op);
if (key)
memcpy(&extent_op->key, key, sizeof(extent_op->key));
else
memset(&extent_op->key, 0, sizeof(extent_op->key));
extent_op->flags_to_set = flags;
extent_op->update_key = 1;
extent_op->update_flags = 1;
extent_op->is_data = 0;
ret = btrfs_add_delayed_tree_ref(trans, ins->objectid,
ins->offset, parent, root_objectid,
level, BTRFS_ADD_DELAYED_EXTENT,
extent_op);
BUG_ON(ret);
}
if (root_objectid == root->root_key.objectid) {
u64 used;
spin_lock(&root->node_lock);
used = btrfs_root_used(&root->root_item) + num_bytes;
btrfs_set_root_used(&root->root_item, used);
spin_unlock(&root->node_lock);
}
return ret;
}
struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 bytenr, u32 blocksize,
int level)
{
struct extent_buffer *buf;
buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
if (!buf)
return ERR_PTR(-ENOMEM);
btrfs_set_header_generation(buf, trans->transid);
btrfs_set_buffer_lockdep_class(buf, level);
btrfs_tree_lock(buf);
clean_tree_block(trans, root, buf);
Btrfs: Change btree locking to use explicit blocking points Most of the btrfs metadata operations can be protected by a spinlock, but some operations still need to schedule. So far, btrfs has been using a mutex along with a trylock loop, most of the time it is able to avoid going for the full mutex, so the trylock loop is a big performance gain. This commit is step one for getting rid of the blocking locks entirely. btrfs_tree_lock takes a spinlock, and the code explicitly switches to a blocking lock when it starts an operation that can schedule. We'll be able get rid of the blocking locks in smaller pieces over time. Tracing allows us to find the most common cause of blocking, so we can start with the hot spots first. The basic idea is: btrfs_tree_lock() returns with the spin lock held btrfs_set_lock_blocking() sets the EXTENT_BUFFER_BLOCKING bit in the extent buffer flags, and then drops the spin lock. The buffer is still considered locked by all of the btrfs code. If btrfs_tree_lock gets the spinlock but finds the blocking bit set, it drops the spin lock and waits on a wait queue for the blocking bit to go away. Much of the code that needs to set the blocking bit finishes without actually blocking a good percentage of the time. So, an adaptive spin is still used against the blocking bit to avoid very high context switch rates. btrfs_clear_lock_blocking() clears the blocking bit and returns with the spinlock held again. btrfs_tree_unlock() can be called on either blocking or spinning locks, it does the right thing based on the blocking bit. ctree.c has a helper function to set/clear all the locked buffers in a path as blocking. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:25:08 +08:00
btrfs_set_lock_blocking(buf);
btrfs_set_buffer_uptodate(buf);
Btrfs: Change btree locking to use explicit blocking points Most of the btrfs metadata operations can be protected by a spinlock, but some operations still need to schedule. So far, btrfs has been using a mutex along with a trylock loop, most of the time it is able to avoid going for the full mutex, so the trylock loop is a big performance gain. This commit is step one for getting rid of the blocking locks entirely. btrfs_tree_lock takes a spinlock, and the code explicitly switches to a blocking lock when it starts an operation that can schedule. We'll be able get rid of the blocking locks in smaller pieces over time. Tracing allows us to find the most common cause of blocking, so we can start with the hot spots first. The basic idea is: btrfs_tree_lock() returns with the spin lock held btrfs_set_lock_blocking() sets the EXTENT_BUFFER_BLOCKING bit in the extent buffer flags, and then drops the spin lock. The buffer is still considered locked by all of the btrfs code. If btrfs_tree_lock gets the spinlock but finds the blocking bit set, it drops the spin lock and waits on a wait queue for the blocking bit to go away. Much of the code that needs to set the blocking bit finishes without actually blocking a good percentage of the time. So, an adaptive spin is still used against the blocking bit to avoid very high context switch rates. btrfs_clear_lock_blocking() clears the blocking bit and returns with the spinlock held again. btrfs_tree_unlock() can be called on either blocking or spinning locks, it does the right thing based on the blocking bit. ctree.c has a helper function to set/clear all the locked buffers in a path as blocking. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:25:08 +08:00
if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
/*
* we allow two log transactions at a time, use different
* EXENT bit to differentiate dirty pages.
*/
if (root->log_transid % 2 == 0)
set_extent_dirty(&root->dirty_log_pages, buf->start,
buf->start + buf->len - 1, GFP_NOFS);
else
set_extent_new(&root->dirty_log_pages, buf->start,
buf->start + buf->len - 1, GFP_NOFS);
} else {
set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
buf->start + buf->len - 1, GFP_NOFS);
}
trans->blocks_used++;
Btrfs: Change btree locking to use explicit blocking points Most of the btrfs metadata operations can be protected by a spinlock, but some operations still need to schedule. So far, btrfs has been using a mutex along with a trylock loop, most of the time it is able to avoid going for the full mutex, so the trylock loop is a big performance gain. This commit is step one for getting rid of the blocking locks entirely. btrfs_tree_lock takes a spinlock, and the code explicitly switches to a blocking lock when it starts an operation that can schedule. We'll be able get rid of the blocking locks in smaller pieces over time. Tracing allows us to find the most common cause of blocking, so we can start with the hot spots first. The basic idea is: btrfs_tree_lock() returns with the spin lock held btrfs_set_lock_blocking() sets the EXTENT_BUFFER_BLOCKING bit in the extent buffer flags, and then drops the spin lock. The buffer is still considered locked by all of the btrfs code. If btrfs_tree_lock gets the spinlock but finds the blocking bit set, it drops the spin lock and waits on a wait queue for the blocking bit to go away. Much of the code that needs to set the blocking bit finishes without actually blocking a good percentage of the time. So, an adaptive spin is still used against the blocking bit to avoid very high context switch rates. btrfs_clear_lock_blocking() clears the blocking bit and returns with the spinlock held again. btrfs_tree_unlock() can be called on either blocking or spinning locks, it does the right thing based on the blocking bit. ctree.c has a helper function to set/clear all the locked buffers in a path as blocking. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:25:08 +08:00
/* this returns a buffer locked for blocking */
return buf;
}
/*
* helper function to allocate a block for a given tree
* returns the tree buffer or NULL.
*/
struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_root *root, u32 blocksize,
u64 parent, u64 root_objectid,
struct btrfs_disk_key *key, int level,
u64 hint, u64 empty_size)
{
struct btrfs_key ins;
int ret;
struct extent_buffer *buf;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
ret = alloc_tree_block(trans, root, blocksize, parent, root_objectid,
key, level, empty_size, hint, (u64)-1, &ins);
if (ret) {
BUG_ON(ret > 0);
return ERR_PTR(ret);
}
buf = btrfs_init_new_buffer(trans, root, ins.objectid,
blocksize, level);
return buf;
}
struct walk_control {
u64 refs[BTRFS_MAX_LEVEL];
u64 flags[BTRFS_MAX_LEVEL];
struct btrfs_key update_progress;
int stage;
int level;
int shared_level;
int update_ref;
int keep_locks;
int reada_slot;
int reada_count;
};
#define DROP_REFERENCE 1
#define UPDATE_BACKREF 2
static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct walk_control *wc,
struct btrfs_path *path)
{
u64 bytenr;
u64 generation;
u64 refs;
u64 flags;
u64 last = 0;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
u32 nritems;
u32 blocksize;
struct btrfs_key key;
struct extent_buffer *eb;
int ret;
int slot;
int nread = 0;
if (path->slots[wc->level] < wc->reada_slot) {
wc->reada_count = wc->reada_count * 2 / 3;
wc->reada_count = max(wc->reada_count, 2);
} else {
wc->reada_count = wc->reada_count * 3 / 2;
wc->reada_count = min_t(int, wc->reada_count,
BTRFS_NODEPTRS_PER_BLOCK(root));
}
eb = path->nodes[wc->level];
nritems = btrfs_header_nritems(eb);
blocksize = btrfs_level_size(root, wc->level - 1);
Btrfs: Make btrfs_drop_snapshot work in larger and more efficient chunks Every transaction in btrfs creates a new snapshot, and then schedules the snapshot from the last transaction for deletion. Snapshot deletion works by walking down the btree and dropping the reference counts on each btree block during the walk. If if a given leaf or node has a reference count greater than one, the reference count is decremented and the subtree pointed to by that node is ignored. If the reference count is one, walking continues down into that node or leaf, and the references of everything it points to are decremented. The old code would try to work in small pieces, walking down the tree until it found the lowest leaf or node to free and then returning. This was very friendly to the rest of the FS because it didn't have a huge impact on other operations. But it wouldn't always keep up with the rate that new commits added new snapshots for deletion, and it wasn't very optimal for the extent allocation tree because it wasn't finding leaves that were close together on disk and processing them at the same time. This changes things to walk down to a level 1 node and then process it in bulk. All the leaf pointers are sorted and the leaves are dropped in order based on their extent number. The extent allocation tree and commit code are now fast enough for this kind of bulk processing to work without slowing the rest of the FS down. Overall it does less IO and is better able to keep up with snapshot deletions under high load. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:27:02 +08:00
for (slot = path->slots[wc->level]; slot < nritems; slot++) {
if (nread >= wc->reada_count)
break;
Btrfs: Make btrfs_drop_snapshot work in larger and more efficient chunks Every transaction in btrfs creates a new snapshot, and then schedules the snapshot from the last transaction for deletion. Snapshot deletion works by walking down the btree and dropping the reference counts on each btree block during the walk. If if a given leaf or node has a reference count greater than one, the reference count is decremented and the subtree pointed to by that node is ignored. If the reference count is one, walking continues down into that node or leaf, and the references of everything it points to are decremented. The old code would try to work in small pieces, walking down the tree until it found the lowest leaf or node to free and then returning. This was very friendly to the rest of the FS because it didn't have a huge impact on other operations. But it wouldn't always keep up with the rate that new commits added new snapshots for deletion, and it wasn't very optimal for the extent allocation tree because it wasn't finding leaves that were close together on disk and processing them at the same time. This changes things to walk down to a level 1 node and then process it in bulk. All the leaf pointers are sorted and the leaves are dropped in order based on their extent number. The extent allocation tree and commit code are now fast enough for this kind of bulk processing to work without slowing the rest of the FS down. Overall it does less IO and is better able to keep up with snapshot deletions under high load. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:27:02 +08:00
cond_resched();
bytenr = btrfs_node_blockptr(eb, slot);
generation = btrfs_node_ptr_generation(eb, slot);
if (slot == path->slots[wc->level])
goto reada;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (wc->stage == UPDATE_BACKREF &&
generation <= root->root_key.offset)
Btrfs: Make btrfs_drop_snapshot work in larger and more efficient chunks Every transaction in btrfs creates a new snapshot, and then schedules the snapshot from the last transaction for deletion. Snapshot deletion works by walking down the btree and dropping the reference counts on each btree block during the walk. If if a given leaf or node has a reference count greater than one, the reference count is decremented and the subtree pointed to by that node is ignored. If the reference count is one, walking continues down into that node or leaf, and the references of everything it points to are decremented. The old code would try to work in small pieces, walking down the tree until it found the lowest leaf or node to free and then returning. This was very friendly to the rest of the FS because it didn't have a huge impact on other operations. But it wouldn't always keep up with the rate that new commits added new snapshots for deletion, and it wasn't very optimal for the extent allocation tree because it wasn't finding leaves that were close together on disk and processing them at the same time. This changes things to walk down to a level 1 node and then process it in bulk. All the leaf pointers are sorted and the leaves are dropped in order based on their extent number. The extent allocation tree and commit code are now fast enough for this kind of bulk processing to work without slowing the rest of the FS down. Overall it does less IO and is better able to keep up with snapshot deletions under high load. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:27:02 +08:00
continue;
/* We don't lock the tree block, it's OK to be racy here */
ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
&refs, &flags);
BUG_ON(ret);
BUG_ON(refs == 0);
if (wc->stage == DROP_REFERENCE) {
if (refs == 1)
goto reada;
Btrfs: Make btrfs_drop_snapshot work in larger and more efficient chunks Every transaction in btrfs creates a new snapshot, and then schedules the snapshot from the last transaction for deletion. Snapshot deletion works by walking down the btree and dropping the reference counts on each btree block during the walk. If if a given leaf or node has a reference count greater than one, the reference count is decremented and the subtree pointed to by that node is ignored. If the reference count is one, walking continues down into that node or leaf, and the references of everything it points to are decremented. The old code would try to work in small pieces, walking down the tree until it found the lowest leaf or node to free and then returning. This was very friendly to the rest of the FS because it didn't have a huge impact on other operations. But it wouldn't always keep up with the rate that new commits added new snapshots for deletion, and it wasn't very optimal for the extent allocation tree because it wasn't finding leaves that were close together on disk and processing them at the same time. This changes things to walk down to a level 1 node and then process it in bulk. All the leaf pointers are sorted and the leaves are dropped in order based on their extent number. The extent allocation tree and commit code are now fast enough for this kind of bulk processing to work without slowing the rest of the FS down. Overall it does less IO and is better able to keep up with snapshot deletions under high load. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:27:02 +08:00
if (wc->level == 1 &&
(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
continue;
if (!wc->update_ref ||
generation <= root->root_key.offset)
continue;
btrfs_node_key_to_cpu(eb, &key, slot);
ret = btrfs_comp_cpu_keys(&key,
&wc->update_progress);
if (ret < 0)
continue;
} else {
if (wc->level == 1 &&
(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
continue;
}
reada:
ret = readahead_tree_block(root, bytenr, blocksize,
generation);
if (ret)
Btrfs: Make btrfs_drop_snapshot work in larger and more efficient chunks Every transaction in btrfs creates a new snapshot, and then schedules the snapshot from the last transaction for deletion. Snapshot deletion works by walking down the btree and dropping the reference counts on each btree block during the walk. If if a given leaf or node has a reference count greater than one, the reference count is decremented and the subtree pointed to by that node is ignored. If the reference count is one, walking continues down into that node or leaf, and the references of everything it points to are decremented. The old code would try to work in small pieces, walking down the tree until it found the lowest leaf or node to free and then returning. This was very friendly to the rest of the FS because it didn't have a huge impact on other operations. But it wouldn't always keep up with the rate that new commits added new snapshots for deletion, and it wasn't very optimal for the extent allocation tree because it wasn't finding leaves that were close together on disk and processing them at the same time. This changes things to walk down to a level 1 node and then process it in bulk. All the leaf pointers are sorted and the leaves are dropped in order based on their extent number. The extent allocation tree and commit code are now fast enough for this kind of bulk processing to work without slowing the rest of the FS down. Overall it does less IO and is better able to keep up with snapshot deletions under high load. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:27:02 +08:00
break;
last = bytenr + blocksize;
nread++;
}
wc->reada_slot = slot;
}
/*
* hepler to process tree block while walking down the tree.
*
* when wc->stage == UPDATE_BACKREF, this function updates
* back refs for pointers in the block.
*
* NOTE: return value 1 means we should stop walking down.
*/
static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
struct btrfs_root *root,
struct btrfs_path *path,
struct walk_control *wc, int lookup_info)
{
int level = wc->level;
struct extent_buffer *eb = path->nodes[level];
u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
int ret;
if (wc->stage == UPDATE_BACKREF &&
btrfs_header_owner(eb) != root->root_key.objectid)
return 1;
/*
* when reference count of tree block is 1, it won't increase
* again. once full backref flag is set, we never clear it.
*/
if (lookup_info &&
((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
(wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
BUG_ON(!path->locks[level]);
ret = btrfs_lookup_extent_info(trans, root,
eb->start, eb->len,
&wc->refs[level],
&wc->flags[level]);
BUG_ON(ret);
BUG_ON(wc->refs[level] == 0);
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (wc->stage == DROP_REFERENCE) {
if (wc->refs[level] > 1)
return 1;
if (path->locks[level] && !wc->keep_locks) {
btrfs_tree_unlock(eb);
path->locks[level] = 0;
}
return 0;
}
/* wc->stage == UPDATE_BACKREF */
if (!(wc->flags[level] & flag)) {
BUG_ON(!path->locks[level]);
ret = btrfs_inc_ref(trans, root, eb, 1);
BUG_ON(ret);
ret = btrfs_dec_ref(trans, root, eb, 0);
BUG_ON(ret);
ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
eb->len, flag, 0);
BUG_ON(ret);
wc->flags[level] |= flag;
}
/*
* the block is shared by multiple trees, so it's not good to
* keep the tree lock
*/
if (path->locks[level] && level > 0) {
btrfs_tree_unlock(eb);
path->locks[level] = 0;
}
return 0;
}
/*
* hepler to process tree block pointer.
*
* when wc->stage == DROP_REFERENCE, this function checks
* reference count of the block pointed to. if the block
* is shared and we need update back refs for the subtree
* rooted at the block, this function changes wc->stage to
* UPDATE_BACKREF. if the block is shared and there is no
* need to update back, this function drops the reference
* to the block.
*
* NOTE: return value 1 means we should stop walking down.
*/
static noinline int do_walk_down(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct walk_control *wc, int *lookup_info)
{
u64 bytenr;
u64 generation;
u64 parent;
u32 blocksize;
struct btrfs_key key;
struct extent_buffer *next;
int level = wc->level;
int reada = 0;
int ret = 0;
generation = btrfs_node_ptr_generation(path->nodes[level],
path->slots[level]);
/*
* if the lower level block was created before the snapshot
* was created, we know there is no need to update back refs
* for the subtree
*/
if (wc->stage == UPDATE_BACKREF &&
generation <= root->root_key.offset) {
*lookup_info = 1;
return 1;
}
bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
blocksize = btrfs_level_size(root, level - 1);
next = btrfs_find_tree_block(root, bytenr, blocksize);
if (!next) {
next = btrfs_find_create_tree_block(root, bytenr, blocksize);
reada = 1;
}
btrfs_tree_lock(next);
btrfs_set_lock_blocking(next);
ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
&wc->refs[level - 1],
&wc->flags[level - 1]);
BUG_ON(ret);
BUG_ON(wc->refs[level - 1] == 0);
*lookup_info = 0;
if (wc->stage == DROP_REFERENCE) {
if (wc->refs[level - 1] > 1) {
if (level == 1 &&
(wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
goto skip;
if (!wc->update_ref ||
generation <= root->root_key.offset)
goto skip;
btrfs_node_key_to_cpu(path->nodes[level], &key,
path->slots[level]);
ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
if (ret < 0)
goto skip;
wc->stage = UPDATE_BACKREF;
wc->shared_level = level - 1;
}
} else {
if (level == 1 &&
(wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
goto skip;
}
if (!btrfs_buffer_uptodate(next, generation)) {
btrfs_tree_unlock(next);
free_extent_buffer(next);
next = NULL;
*lookup_info = 1;
}
if (!next) {
if (reada && level == 1)
reada_walk_down(trans, root, wc, path);
next = read_tree_block(root, bytenr, blocksize, generation);
btrfs_tree_lock(next);
btrfs_set_lock_blocking(next);
}
level--;
BUG_ON(level != btrfs_header_level(next));
path->nodes[level] = next;
path->slots[level] = 0;
path->locks[level] = 1;
wc->level = level;
if (wc->level == 1)
wc->reada_slot = 0;
return 0;
skip:
wc->refs[level - 1] = 0;
wc->flags[level - 1] = 0;
if (wc->stage == DROP_REFERENCE) {
if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
parent = path->nodes[level]->start;
} else {
BUG_ON(root->root_key.objectid !=
btrfs_header_owner(path->nodes[level]));
parent = 0;
}
ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
root->root_key.objectid, level - 1, 0);
BUG_ON(ret);
}
btrfs_tree_unlock(next);
free_extent_buffer(next);
*lookup_info = 1;
return 1;
}
/*
* hepler to process tree block while walking up the tree.
*
* when wc->stage == DROP_REFERENCE, this function drops
* reference count on the block.
*
* when wc->stage == UPDATE_BACKREF, this function changes
* wc->stage back to DROP_REFERENCE if we changed wc->stage
* to UPDATE_BACKREF previously while processing the block.
*
* NOTE: return value 1 means we should stop walking up.
*/
static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct walk_control *wc)
{
int ret = 0;
int level = wc->level;
struct extent_buffer *eb = path->nodes[level];
u64 parent = 0;
if (wc->stage == UPDATE_BACKREF) {
BUG_ON(wc->shared_level < level);
if (level < wc->shared_level)
goto out;
ret = find_next_key(path, level + 1, &wc->update_progress);
if (ret > 0)
wc->update_ref = 0;
wc->stage = DROP_REFERENCE;
wc->shared_level = -1;
path->slots[level] = 0;
/*
* check reference count again if the block isn't locked.
* we should start walking down the tree again if reference
* count is one.
*/
if (!path->locks[level]) {
BUG_ON(level == 0);
btrfs_tree_lock(eb);
btrfs_set_lock_blocking(eb);
path->locks[level] = 1;
ret = btrfs_lookup_extent_info(trans, root,
eb->start, eb->len,
&wc->refs[level],
&wc->flags[level]);
BUG_ON(ret);
BUG_ON(wc->refs[level] == 0);
if (wc->refs[level] == 1) {
btrfs_tree_unlock(eb);
path->locks[level] = 0;
return 1;
}
}
}
/* wc->stage == DROP_REFERENCE */
BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (wc->refs[level] == 1) {
if (level == 0) {
if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
ret = btrfs_dec_ref(trans, root, eb, 1);
else
ret = btrfs_dec_ref(trans, root, eb, 0);
BUG_ON(ret);
}
/* make block locked assertion in clean_tree_block happy */
if (!path->locks[level] &&
btrfs_header_generation(eb) == trans->transid) {
btrfs_tree_lock(eb);
btrfs_set_lock_blocking(eb);
path->locks[level] = 1;
}
clean_tree_block(trans, root, eb);
}
if (eb == root->node) {
if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
parent = eb->start;
else
BUG_ON(root->root_key.objectid !=
btrfs_header_owner(eb));
} else {
if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
parent = path->nodes[level + 1]->start;
else
BUG_ON(root->root_key.objectid !=
btrfs_header_owner(path->nodes[level + 1]));
}
ret = btrfs_free_extent(trans, root, eb->start, eb->len, parent,
root->root_key.objectid, level, 0);
BUG_ON(ret);
out:
wc->refs[level] = 0;
wc->flags[level] = 0;
return ret;
}
static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct walk_control *wc)
{
int level = wc->level;
int lookup_info = 1;
int ret;
while (level >= 0) {
if (path->slots[level] >=
btrfs_header_nritems(path->nodes[level]))
break;
ret = walk_down_proc(trans, root, path, wc, lookup_info);
if (ret > 0)
break;
if (level == 0)
break;
ret = do_walk_down(trans, root, path, wc, &lookup_info);
if (ret > 0) {
path->slots[level]++;
continue;
}
level = wc->level;
}
return 0;
}
static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct walk_control *wc, int max_level)
{
int level = wc->level;
int ret;
path->slots[level] = btrfs_header_nritems(path->nodes[level]);
while (level < max_level && path->nodes[level]) {
wc->level = level;
if (path->slots[level] + 1 <
btrfs_header_nritems(path->nodes[level])) {
path->slots[level]++;
return 0;
} else {
ret = walk_up_proc(trans, root, path, wc);
if (ret > 0)
return 0;
Btrfs: Make btrfs_drop_snapshot work in larger and more efficient chunks Every transaction in btrfs creates a new snapshot, and then schedules the snapshot from the last transaction for deletion. Snapshot deletion works by walking down the btree and dropping the reference counts on each btree block during the walk. If if a given leaf or node has a reference count greater than one, the reference count is decremented and the subtree pointed to by that node is ignored. If the reference count is one, walking continues down into that node or leaf, and the references of everything it points to are decremented. The old code would try to work in small pieces, walking down the tree until it found the lowest leaf or node to free and then returning. This was very friendly to the rest of the FS because it didn't have a huge impact on other operations. But it wouldn't always keep up with the rate that new commits added new snapshots for deletion, and it wasn't very optimal for the extent allocation tree because it wasn't finding leaves that were close together on disk and processing them at the same time. This changes things to walk down to a level 1 node and then process it in bulk. All the leaf pointers are sorted and the leaves are dropped in order based on their extent number. The extent allocation tree and commit code are now fast enough for this kind of bulk processing to work without slowing the rest of the FS down. Overall it does less IO and is better able to keep up with snapshot deletions under high load. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:27:02 +08:00
if (path->locks[level]) {
btrfs_tree_unlock(path->nodes[level]);
path->locks[level] = 0;
}
free_extent_buffer(path->nodes[level]);
path->nodes[level] = NULL;
level++;
}
}
return 1;
}
/*
* drop a subvolume tree.
*
* this function traverses the tree freeing any blocks that only
* referenced by the tree.
*
* when a shared tree block is found. this function decreases its
* reference count by one. if update_ref is true, this function
* also make sure backrefs for the shared block and all lower level
* blocks are properly updated.
*/
int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref)
{
struct btrfs_path *path;
struct btrfs_trans_handle *trans;
struct btrfs_root *tree_root = root->fs_info->tree_root;
struct btrfs_root_item *root_item = &root->root_item;
struct walk_control *wc;
struct btrfs_key key;
int err = 0;
int ret;
int level;
path = btrfs_alloc_path();
BUG_ON(!path);
wc = kzalloc(sizeof(*wc), GFP_NOFS);
BUG_ON(!wc);
trans = btrfs_start_transaction(tree_root, 1);
if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
level = btrfs_header_level(root->node);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
path->nodes[level] = btrfs_lock_root_node(root);
btrfs_set_lock_blocking(path->nodes[level]);
path->slots[level] = 0;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
path->locks[level] = 1;
memset(&wc->update_progress, 0,
sizeof(wc->update_progress));
} else {
btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
memcpy(&wc->update_progress, &key,
sizeof(wc->update_progress));
level = root_item->drop_level;
BUG_ON(level == 0);
path->lowest_level = level;
ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
path->lowest_level = 0;
if (ret < 0) {
err = ret;
goto out;
}
WARN_ON(ret > 0);
/*
* unlock our path, this is safe because only this
* function is allowed to delete this snapshot
*/
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
btrfs_unlock_up_safe(path, 0);
level = btrfs_header_level(root->node);
while (1) {
btrfs_tree_lock(path->nodes[level]);
btrfs_set_lock_blocking(path->nodes[level]);
ret = btrfs_lookup_extent_info(trans, root,
path->nodes[level]->start,
path->nodes[level]->len,
&wc->refs[level],
&wc->flags[level]);
BUG_ON(ret);
BUG_ON(wc->refs[level] == 0);
if (level == root_item->drop_level)
break;
btrfs_tree_unlock(path->nodes[level]);
WARN_ON(wc->refs[level] != 1);
level--;
}
}
wc->level = level;
wc->shared_level = -1;
wc->stage = DROP_REFERENCE;
wc->update_ref = update_ref;
wc->keep_locks = 0;
wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
while (1) {
ret = walk_down_tree(trans, root, path, wc);
if (ret < 0) {
err = ret;
break;
}
ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
if (ret < 0) {
err = ret;
break;
}
if (ret > 0) {
BUG_ON(wc->stage != DROP_REFERENCE);
break;
}
if (wc->stage == DROP_REFERENCE) {
level = wc->level;
btrfs_node_key(path->nodes[level],
&root_item->drop_progress,
path->slots[level]);
root_item->drop_level = level;
}
BUG_ON(wc->level == 0);
if (trans->transaction->in_commit ||
trans->transaction->delayed_refs.flushing) {
ret = btrfs_update_root(trans, tree_root,
&root->root_key,
root_item);
BUG_ON(ret);
btrfs_end_transaction(trans, tree_root);
trans = btrfs_start_transaction(tree_root, 1);
} else {
unsigned long update;
update = trans->delayed_ref_updates;
trans->delayed_ref_updates = 0;
if (update)
btrfs_run_delayed_refs(trans, tree_root,
update);
}
}
btrfs_release_path(root, path);
BUG_ON(err);
ret = btrfs_del_root(trans, tree_root, &root->root_key);
BUG_ON(ret);
if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
NULL, NULL);
BUG_ON(ret < 0);
if (ret > 0) {
ret = btrfs_del_orphan_item(trans, tree_root,
root->root_key.objectid);
BUG_ON(ret);
}
}
if (root->in_radix) {
btrfs_free_fs_root(tree_root->fs_info, root);
} else {
free_extent_buffer(root->node);
free_extent_buffer(root->commit_root);
kfree(root);
}
out:
btrfs_end_transaction(trans, tree_root);
kfree(wc);
btrfs_free_path(path);
return err;
}
/*
* drop subtree rooted at tree block 'node'.
*
* NOTE: this function will unlock and release tree block 'node'
*/
int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct extent_buffer *node,
struct extent_buffer *parent)
{
struct btrfs_path *path;
struct walk_control *wc;
int level;
int parent_level;
int ret = 0;
int wret;
BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
path = btrfs_alloc_path();
BUG_ON(!path);
wc = kzalloc(sizeof(*wc), GFP_NOFS);
BUG_ON(!wc);
btrfs_assert_tree_locked(parent);
parent_level = btrfs_header_level(parent);
extent_buffer_get(parent);
path->nodes[parent_level] = parent;
path->slots[parent_level] = btrfs_header_nritems(parent);
btrfs_assert_tree_locked(node);
level = btrfs_header_level(node);
path->nodes[level] = node;
path->slots[level] = 0;
path->locks[level] = 1;
wc->refs[parent_level] = 1;
wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
wc->level = level;
wc->shared_level = -1;
wc->stage = DROP_REFERENCE;
wc->update_ref = 0;
wc->keep_locks = 1;
wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
while (1) {
wret = walk_down_tree(trans, root, path, wc);
if (wret < 0) {
ret = wret;
break;
}
wret = walk_up_tree(trans, root, path, wc, parent_level);
if (wret < 0)
ret = wret;
if (wret != 0)
break;
}
kfree(wc);
btrfs_free_path(path);
return ret;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
#if 0
static unsigned long calc_ra(unsigned long start, unsigned long last,
unsigned long nr)
{
return min(last, start + nr - 1);
}
static noinline int relocate_inode_pages(struct inode *inode, u64 start,
u64 len)
{
u64 page_start;
u64 page_end;
2008-09-26 22:09:34 +08:00
unsigned long first_index;
unsigned long last_index;
unsigned long i;
struct page *page;
struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
struct file_ra_state *ra;
struct btrfs_ordered_extent *ordered;
2008-09-26 22:09:34 +08:00
unsigned int total_read = 0;
unsigned int total_dirty = 0;
int ret = 0;
ra = kzalloc(sizeof(*ra), GFP_NOFS);
mutex_lock(&inode->i_mutex);
2008-09-26 22:09:34 +08:00
first_index = start >> PAGE_CACHE_SHIFT;
last_index = (start + len - 1) >> PAGE_CACHE_SHIFT;
2008-09-26 22:09:34 +08:00
/* make sure the dirty trick played by the caller work */
ret = invalidate_inode_pages2_range(inode->i_mapping,
first_index, last_index);
if (ret)
goto out_unlock;
file_ra_state_init(ra, inode->i_mapping);
2008-09-26 22:09:34 +08:00
for (i = first_index ; i <= last_index; i++) {
if (total_read % ra->ra_pages == 0) {
btrfs_force_ra(inode->i_mapping, ra, NULL, i,
2008-09-26 22:09:34 +08:00
calc_ra(i, last_index, ra->ra_pages));
}
total_read++;
again:
if (((u64)i << PAGE_CACHE_SHIFT) > i_size_read(inode))
2008-09-26 22:09:34 +08:00
BUG_ON(1);
page = grab_cache_page(inode->i_mapping, i);
if (!page) {
2008-09-26 22:09:34 +08:00
ret = -ENOMEM;
goto out_unlock;
}
if (!PageUptodate(page)) {
btrfs_readpage(NULL, page);
lock_page(page);
if (!PageUptodate(page)) {
unlock_page(page);
page_cache_release(page);
2008-09-26 22:09:34 +08:00
ret = -EIO;
goto out_unlock;
}
}
wait_on_page_writeback(page);
page_start = (u64)page->index << PAGE_CACHE_SHIFT;
page_end = page_start + PAGE_CACHE_SIZE - 1;
lock_extent(io_tree, page_start, page_end, GFP_NOFS);
ordered = btrfs_lookup_ordered_extent(inode, page_start);
if (ordered) {
unlock_extent(io_tree, page_start, page_end, GFP_NOFS);
unlock_page(page);
page_cache_release(page);
btrfs_start_ordered_extent(inode, ordered, 1);
btrfs_put_ordered_extent(ordered);
goto again;
}
set_page_extent_mapped(page);
2008-09-26 22:09:34 +08:00
if (i == first_index)
set_extent_bits(io_tree, page_start, page_end,
EXTENT_BOUNDARY, GFP_NOFS);
btrfs_set_extent_delalloc(inode, page_start, page_end);
2008-09-26 22:09:34 +08:00
set_page_dirty(page);
2008-09-26 22:09:34 +08:00
total_dirty++;
unlock_extent(io_tree, page_start, page_end, GFP_NOFS);
unlock_page(page);
page_cache_release(page);
}
out_unlock:
kfree(ra);
mutex_unlock(&inode->i_mutex);
2008-09-26 22:09:34 +08:00
balance_dirty_pages_ratelimited_nr(inode->i_mapping, total_dirty);
return ret;
}
static noinline int relocate_data_extent(struct inode *reloc_inode,
2008-09-26 22:09:34 +08:00
struct btrfs_key *extent_key,
u64 offset)
{
struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
struct extent_map_tree *em_tree = &BTRFS_I(reloc_inode)->extent_tree;
struct extent_map *em;
u64 start = extent_key->objectid - offset;
u64 end = start + extent_key->offset - 1;
2008-09-26 22:09:34 +08:00
em = alloc_extent_map(GFP_NOFS);
BUG_ON(!em || IS_ERR(em));
em->start = start;
2008-09-26 22:09:34 +08:00
em->len = extent_key->offset;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
em->block_len = extent_key->offset;
2008-09-26 22:09:34 +08:00
em->block_start = extent_key->objectid;
em->bdev = root->fs_info->fs_devices->latest_bdev;
set_bit(EXTENT_FLAG_PINNED, &em->flags);
/* setup extent map to cheat btrfs_readpage */
lock_extent(&BTRFS_I(reloc_inode)->io_tree, start, end, GFP_NOFS);
2008-09-26 22:09:34 +08:00
while (1) {
int ret;
write_lock(&em_tree->lock);
2008-09-26 22:09:34 +08:00
ret = add_extent_mapping(em_tree, em);
write_unlock(&em_tree->lock);
2008-09-26 22:09:34 +08:00
if (ret != -EEXIST) {
free_extent_map(em);
break;
}
btrfs_drop_extent_cache(reloc_inode, start, end, 0);
}
unlock_extent(&BTRFS_I(reloc_inode)->io_tree, start, end, GFP_NOFS);
return relocate_inode_pages(reloc_inode, start, extent_key->offset);
2008-09-26 22:09:34 +08:00
}
2008-09-26 22:09:34 +08:00
struct btrfs_ref_path {
u64 extent_start;
u64 nodes[BTRFS_MAX_LEVEL];
u64 root_objectid;
u64 root_generation;
u64 owner_objectid;
u32 num_refs;
int lowest_level;
int current_level;
int shared_level;
struct btrfs_key node_keys[BTRFS_MAX_LEVEL];
u64 new_nodes[BTRFS_MAX_LEVEL];
2008-09-26 22:09:34 +08:00
};
2008-09-26 22:09:34 +08:00
struct disk_extent {
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
u64 ram_bytes;
2008-09-26 22:09:34 +08:00
u64 disk_bytenr;
u64 disk_num_bytes;
u64 offset;
u64 num_bytes;
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
u8 compression;
u8 encryption;
u16 other_encoding;
2008-09-26 22:09:34 +08:00
};
2008-09-26 22:09:34 +08:00
static int is_cowonly_root(u64 root_objectid)
{
if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
root_objectid == BTRFS_DEV_TREE_OBJECTID ||
root_objectid == BTRFS_TREE_LOG_OBJECTID ||
root_objectid == BTRFS_CSUM_TREE_OBJECTID)
2008-09-26 22:09:34 +08:00
return 1;
return 0;
}
static noinline int __next_ref_path(struct btrfs_trans_handle *trans,
2008-09-26 22:09:34 +08:00
struct btrfs_root *extent_root,
struct btrfs_ref_path *ref_path,
int first_time)
{
struct extent_buffer *leaf;
struct btrfs_path *path;
struct btrfs_extent_ref *ref;
struct btrfs_key key;
struct btrfs_key found_key;
u64 bytenr;
u32 nritems;
int level;
int ret = 1;
2008-09-26 22:09:34 +08:00
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
2008-09-26 22:09:34 +08:00
if (first_time) {
ref_path->lowest_level = -1;
ref_path->current_level = -1;
ref_path->shared_level = -1;
2008-09-26 22:09:34 +08:00
goto walk_up;
}
walk_down:
level = ref_path->current_level - 1;
while (level >= -1) {
u64 parent;
if (level < ref_path->lowest_level)
break;
if (level >= 0)
2008-09-26 22:09:34 +08:00
bytenr = ref_path->nodes[level];
else
2008-09-26 22:09:34 +08:00
bytenr = ref_path->extent_start;
BUG_ON(bytenr == 0);
2008-09-26 22:09:34 +08:00
parent = ref_path->nodes[level + 1];
ref_path->nodes[level + 1] = 0;
ref_path->current_level = level;
BUG_ON(parent == 0);
2008-09-26 22:09:34 +08:00
key.objectid = bytenr;
key.offset = parent + 1;
key.type = BTRFS_EXTENT_REF_KEY;
2008-09-26 22:09:34 +08:00
ret = btrfs_search_slot(trans, extent_root, &key, path, 0, 0);
if (ret < 0)
goto out;
2008-09-26 22:09:34 +08:00
BUG_ON(ret == 0);
2008-09-26 22:09:34 +08:00
leaf = path->nodes[0];
nritems = btrfs_header_nritems(leaf);
if (path->slots[0] >= nritems) {
ret = btrfs_next_leaf(extent_root, path);
if (ret < 0)
goto out;
if (ret > 0)
goto next;
leaf = path->nodes[0];
}
2008-09-26 22:09:34 +08:00
btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
if (found_key.objectid == bytenr &&
found_key.type == BTRFS_EXTENT_REF_KEY) {
if (level < ref_path->shared_level)
ref_path->shared_level = level;
2008-09-26 22:09:34 +08:00
goto found;
}
2008-09-26 22:09:34 +08:00
next:
level--;
btrfs_release_path(extent_root, path);
cond_resched();
2008-09-26 22:09:34 +08:00
}
/* reached lowest level */
ret = 1;
goto out;
walk_up:
level = ref_path->current_level;
while (level < BTRFS_MAX_LEVEL - 1) {
u64 ref_objectid;
if (level >= 0)
2008-09-26 22:09:34 +08:00
bytenr = ref_path->nodes[level];
else
2008-09-26 22:09:34 +08:00
bytenr = ref_path->extent_start;
2008-09-26 22:09:34 +08:00
BUG_ON(bytenr == 0);
2008-09-26 22:09:34 +08:00
key.objectid = bytenr;
key.offset = 0;
key.type = BTRFS_EXTENT_REF_KEY;
2008-09-26 22:09:34 +08:00
ret = btrfs_search_slot(trans, extent_root, &key, path, 0, 0);
if (ret < 0)
goto out;
2008-09-26 22:09:34 +08:00
leaf = path->nodes[0];
nritems = btrfs_header_nritems(leaf);
if (path->slots[0] >= nritems) {
ret = btrfs_next_leaf(extent_root, path);
if (ret < 0)
goto out;
if (ret > 0) {
/* the extent was freed by someone */
if (ref_path->lowest_level == level)
goto out;
btrfs_release_path(extent_root, path);
goto walk_down;
}
leaf = path->nodes[0];
}
2008-09-26 22:09:34 +08:00
btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
if (found_key.objectid != bytenr ||
found_key.type != BTRFS_EXTENT_REF_KEY) {
/* the extent was freed by someone */
if (ref_path->lowest_level == level) {
ret = 1;
goto out;
}
btrfs_release_path(extent_root, path);
goto walk_down;
}
found:
ref = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_extent_ref);
ref_objectid = btrfs_ref_objectid(leaf, ref);
if (ref_objectid < BTRFS_FIRST_FREE_OBJECTID) {
if (first_time) {
level = (int)ref_objectid;
BUG_ON(level >= BTRFS_MAX_LEVEL);
ref_path->lowest_level = level;
ref_path->current_level = level;
ref_path->nodes[level] = bytenr;
} else {
WARN_ON(ref_objectid != level);
}
} else {
WARN_ON(level != -1);
}
first_time = 0;
2008-09-26 22:09:34 +08:00
if (ref_path->lowest_level == level) {
ref_path->owner_objectid = ref_objectid;
ref_path->num_refs = btrfs_ref_num_refs(leaf, ref);
}
/*
2008-09-26 22:09:34 +08:00
* the block is tree root or the block isn't in reference
* counted tree.
*/
2008-09-26 22:09:34 +08:00
if (found_key.objectid == found_key.offset ||
is_cowonly_root(btrfs_ref_root(leaf, ref))) {
ref_path->root_objectid = btrfs_ref_root(leaf, ref);
ref_path->root_generation =
btrfs_ref_generation(leaf, ref);
if (level < 0) {
/* special reference from the tree log */
ref_path->nodes[0] = found_key.offset;
ref_path->current_level = 0;
}
ret = 0;
goto out;
}
2008-09-26 22:09:34 +08:00
level++;
BUG_ON(ref_path->nodes[level] != 0);
ref_path->nodes[level] = found_key.offset;
ref_path->current_level = level;
2008-09-26 22:09:34 +08:00
/*
* the reference was created in the running transaction,
* no need to continue walking up.
*/
if (btrfs_ref_generation(leaf, ref) == trans->transid) {
ref_path->root_objectid = btrfs_ref_root(leaf, ref);
ref_path->root_generation =
btrfs_ref_generation(leaf, ref);
ret = 0;
goto out;
}
2008-09-26 22:09:34 +08:00
btrfs_release_path(extent_root, path);
cond_resched();
}
2008-09-26 22:09:34 +08:00
/* reached max tree level, but no tree root found. */
BUG();
out:
2008-09-26 22:09:34 +08:00
btrfs_free_path(path);
return ret;
}
2008-09-26 22:09:34 +08:00
static int btrfs_first_ref_path(struct btrfs_trans_handle *trans,
struct btrfs_root *extent_root,
struct btrfs_ref_path *ref_path,
u64 extent_start)
{
2008-09-26 22:09:34 +08:00
memset(ref_path, 0, sizeof(*ref_path));
ref_path->extent_start = extent_start;
2008-09-26 22:09:34 +08:00
return __next_ref_path(trans, extent_root, ref_path, 1);
}
2008-09-26 22:09:34 +08:00
static int btrfs_next_ref_path(struct btrfs_trans_handle *trans,
struct btrfs_root *extent_root,
struct btrfs_ref_path *ref_path)
{
2008-09-26 22:09:34 +08:00
return __next_ref_path(trans, extent_root, ref_path, 0);
}
static noinline int get_new_locations(struct inode *reloc_inode,
2008-09-26 22:09:34 +08:00
struct btrfs_key *extent_key,
u64 offset, int no_fragment,
struct disk_extent **extents,
int *nr_extents)
{
struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
struct btrfs_path *path;
struct btrfs_file_extent_item *fi;
struct extent_buffer *leaf;
2008-09-26 22:09:34 +08:00
struct disk_extent *exts = *extents;
struct btrfs_key found_key;
u64 cur_pos;
u64 last_byte;
u32 nritems;
2008-09-26 22:09:34 +08:00
int nr = 0;
int max = *nr_extents;
int ret;
2008-09-26 22:09:34 +08:00
WARN_ON(!no_fragment && *extents);
if (!exts) {
max = 1;
exts = kmalloc(sizeof(*exts) * max, GFP_NOFS);
if (!exts)
return -ENOMEM;
}
2008-09-26 22:09:34 +08:00
path = btrfs_alloc_path();
BUG_ON(!path);
2008-09-26 22:09:34 +08:00
cur_pos = extent_key->objectid - offset;
last_byte = extent_key->objectid + extent_key->offset;
ret = btrfs_lookup_file_extent(NULL, root, path, reloc_inode->i_ino,
cur_pos, 0);
if (ret < 0)
goto out;
if (ret > 0) {
ret = -ENOENT;
goto out;
}
2008-09-26 22:09:34 +08:00
while (1) {
leaf = path->nodes[0];
nritems = btrfs_header_nritems(leaf);
2008-09-26 22:09:34 +08:00
if (path->slots[0] >= nritems) {
ret = btrfs_next_leaf(root, path);
if (ret < 0)
goto out;
2008-09-26 22:09:34 +08:00
if (ret > 0)
break;
leaf = path->nodes[0];
}
btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2008-09-26 22:09:34 +08:00
if (found_key.offset != cur_pos ||
found_key.type != BTRFS_EXTENT_DATA_KEY ||
found_key.objectid != reloc_inode->i_ino)
break;
2008-09-26 22:09:34 +08:00
fi = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_file_extent_item);
if (btrfs_file_extent_type(leaf, fi) !=
BTRFS_FILE_EXTENT_REG ||
btrfs_file_extent_disk_bytenr(leaf, fi) == 0)
break;
2008-09-26 22:09:34 +08:00
if (nr == max) {
struct disk_extent *old = exts;
max *= 2;
exts = kzalloc(sizeof(*exts) * max, GFP_NOFS);
memcpy(exts, old, sizeof(*exts) * nr);
if (old != *extents)
kfree(old);
}
2008-09-26 22:09:34 +08:00
exts[nr].disk_bytenr =
btrfs_file_extent_disk_bytenr(leaf, fi);
exts[nr].disk_num_bytes =
btrfs_file_extent_disk_num_bytes(leaf, fi);
exts[nr].offset = btrfs_file_extent_offset(leaf, fi);
exts[nr].num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
exts[nr].ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
exts[nr].compression = btrfs_file_extent_compression(leaf, fi);
exts[nr].encryption = btrfs_file_extent_encryption(leaf, fi);
exts[nr].other_encoding = btrfs_file_extent_other_encoding(leaf,
fi);
BUG_ON(exts[nr].offset > 0);
BUG_ON(exts[nr].compression || exts[nr].encryption);
BUG_ON(exts[nr].num_bytes != exts[nr].disk_num_bytes);
2008-09-26 22:09:34 +08:00
cur_pos += exts[nr].num_bytes;
nr++;
if (cur_pos + offset >= last_byte)
break;
if (no_fragment) {
ret = 1;
goto out;
2008-09-26 22:09:34 +08:00
}
path->slots[0]++;
}
BUG_ON(cur_pos + offset > last_byte);
2008-09-26 22:09:34 +08:00
if (cur_pos + offset < last_byte) {
ret = -ENOENT;
goto out;
}
ret = 0;
out:
2008-09-26 22:09:34 +08:00
btrfs_free_path(path);
if (ret) {
if (exts != *extents)
kfree(exts);
} else {
*extents = exts;
*nr_extents = nr;
}
return ret;
}
static noinline int replace_one_extent(struct btrfs_trans_handle *trans,
2008-09-26 22:09:34 +08:00
struct btrfs_root *root,
struct btrfs_path *path,
struct btrfs_key *extent_key,
struct btrfs_key *leaf_key,
struct btrfs_ref_path *ref_path,
struct disk_extent *new_extents,
int nr_extents)
{
struct extent_buffer *leaf;
struct btrfs_file_extent_item *fi;
struct inode *inode = NULL;
struct btrfs_key key;
u64 lock_start = 0;
u64 lock_end = 0;
u64 num_bytes;
u64 ext_offset;
u64 search_end = (u64)-1;
2008-09-26 22:09:34 +08:00
u32 nritems;
int nr_scaned = 0;
2008-09-26 22:09:34 +08:00
int extent_locked = 0;
int extent_type;
2008-09-26 22:09:34 +08:00
int ret;
memcpy(&key, leaf_key, sizeof(key));
2008-09-26 22:09:34 +08:00
if (ref_path->owner_objectid != BTRFS_MULTIPLE_OBJECTIDS) {
if (key.objectid < ref_path->owner_objectid ||
(key.objectid == ref_path->owner_objectid &&
key.type < BTRFS_EXTENT_DATA_KEY)) {
key.objectid = ref_path->owner_objectid;
key.type = BTRFS_EXTENT_DATA_KEY;
key.offset = 0;
}
2008-09-26 22:09:34 +08:00
}
while (1) {
ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
if (ret < 0)
goto out;
leaf = path->nodes[0];
nritems = btrfs_header_nritems(leaf);
next:
if (extent_locked && ret > 0) {
/*
* the file extent item was modified by someone
* before the extent got locked.
*/
unlock_extent(&BTRFS_I(inode)->io_tree, lock_start,
lock_end, GFP_NOFS);
extent_locked = 0;
}
if (path->slots[0] >= nritems) {
if (++nr_scaned > 2)
2008-09-26 22:09:34 +08:00
break;
BUG_ON(extent_locked);
ret = btrfs_next_leaf(root, path);
if (ret < 0)
goto out;
if (ret > 0)
break;
leaf = path->nodes[0];
nritems = btrfs_header_nritems(leaf);
}
btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
if (ref_path->owner_objectid != BTRFS_MULTIPLE_OBJECTIDS) {
if ((key.objectid > ref_path->owner_objectid) ||
(key.objectid == ref_path->owner_objectid &&
key.type > BTRFS_EXTENT_DATA_KEY) ||
key.offset >= search_end)
2008-09-26 22:09:34 +08:00
break;
}
if (inode && key.objectid != inode->i_ino) {
BUG_ON(extent_locked);
btrfs_release_path(root, path);
mutex_unlock(&inode->i_mutex);
iput(inode);
inode = NULL;
continue;
}
if (key.type != BTRFS_EXTENT_DATA_KEY) {
path->slots[0]++;
ret = 1;
goto next;
}
fi = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_file_extent_item);
extent_type = btrfs_file_extent_type(leaf, fi);
if ((extent_type != BTRFS_FILE_EXTENT_REG &&
extent_type != BTRFS_FILE_EXTENT_PREALLOC) ||
2008-09-26 22:09:34 +08:00
(btrfs_file_extent_disk_bytenr(leaf, fi) !=
extent_key->objectid)) {
path->slots[0]++;
ret = 1;
goto next;
}
num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
ext_offset = btrfs_file_extent_offset(leaf, fi);
if (search_end == (u64)-1) {
search_end = key.offset - ext_offset +
btrfs_file_extent_ram_bytes(leaf, fi);
}
2008-09-26 22:09:34 +08:00
if (!extent_locked) {
lock_start = key.offset;
lock_end = lock_start + num_bytes - 1;
} else {
if (lock_start > key.offset ||
lock_end + 1 < key.offset + num_bytes) {
unlock_extent(&BTRFS_I(inode)->io_tree,
lock_start, lock_end, GFP_NOFS);
extent_locked = 0;
}
2008-09-26 22:09:34 +08:00
}
if (!inode) {
btrfs_release_path(root, path);
inode = btrfs_iget_locked(root->fs_info->sb,
key.objectid, root);
if (inode->i_state & I_NEW) {
BTRFS_I(inode)->root = root;
BTRFS_I(inode)->location.objectid =
key.objectid;
BTRFS_I(inode)->location.type =
BTRFS_INODE_ITEM_KEY;
BTRFS_I(inode)->location.offset = 0;
btrfs_read_locked_inode(inode);
unlock_new_inode(inode);
}
/*
* some code call btrfs_commit_transaction while
* holding the i_mutex, so we can't use mutex_lock
* here.
*/
if (is_bad_inode(inode) ||
!mutex_trylock(&inode->i_mutex)) {
iput(inode);
inode = NULL;
key.offset = (u64)-1;
goto skip;
}
}
if (!extent_locked) {
struct btrfs_ordered_extent *ordered;
btrfs_release_path(root, path);
lock_extent(&BTRFS_I(inode)->io_tree, lock_start,
lock_end, GFP_NOFS);
ordered = btrfs_lookup_first_ordered_extent(inode,
lock_end);
if (ordered &&
ordered->file_offset <= lock_end &&
ordered->file_offset + ordered->len > lock_start) {
unlock_extent(&BTRFS_I(inode)->io_tree,
lock_start, lock_end, GFP_NOFS);
btrfs_start_ordered_extent(inode, ordered, 1);
btrfs_put_ordered_extent(ordered);
key.offset += num_bytes;
goto skip;
}
if (ordered)
btrfs_put_ordered_extent(ordered);
extent_locked = 1;
continue;
}
if (nr_extents == 1) {
/* update extent pointer in place */
btrfs_set_file_extent_disk_bytenr(leaf, fi,
new_extents[0].disk_bytenr);
btrfs_set_file_extent_disk_num_bytes(leaf, fi,
new_extents[0].disk_num_bytes);
btrfs_mark_buffer_dirty(leaf);
btrfs_drop_extent_cache(inode, key.offset,
key.offset + num_bytes - 1, 0);
ret = btrfs_inc_extent_ref(trans, root,
new_extents[0].disk_bytenr,
new_extents[0].disk_num_bytes,
leaf->start,
root->root_key.objectid,
trans->transid,
key.objectid);
2008-09-26 22:09:34 +08:00
BUG_ON(ret);
ret = btrfs_free_extent(trans, root,
extent_key->objectid,
extent_key->offset,
leaf->start,
btrfs_header_owner(leaf),
btrfs_header_generation(leaf),
key.objectid, 0);
2008-09-26 22:09:34 +08:00
BUG_ON(ret);
btrfs_release_path(root, path);
key.offset += num_bytes;
} else {
BUG_ON(1);
#if 0
2008-09-26 22:09:34 +08:00
u64 alloc_hint;
u64 extent_len;
int i;
/*
* drop old extent pointer at first, then insert the
* new pointers one bye one
*/
btrfs_release_path(root, path);
ret = btrfs_drop_extents(trans, root, inode, key.offset,
key.offset + num_bytes,
key.offset, &alloc_hint);
BUG_ON(ret);
for (i = 0; i < nr_extents; i++) {
if (ext_offset >= new_extents[i].num_bytes) {
ext_offset -= new_extents[i].num_bytes;
continue;
}
extent_len = min(new_extents[i].num_bytes -
ext_offset, num_bytes);
ret = btrfs_insert_empty_item(trans, root,
path, &key,
sizeof(*fi));
BUG_ON(ret);
leaf = path->nodes[0];
fi = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_file_extent_item);
btrfs_set_file_extent_generation(leaf, fi,
trans->transid);
btrfs_set_file_extent_type(leaf, fi,
BTRFS_FILE_EXTENT_REG);
btrfs_set_file_extent_disk_bytenr(leaf, fi,
new_extents[i].disk_bytenr);
btrfs_set_file_extent_disk_num_bytes(leaf, fi,
new_extents[i].disk_num_bytes);
Btrfs: Add zlib compression support This is a large change for adding compression on reading and writing, both for inline and regular extents. It does some fairly large surgery to the writeback paths. Compression is off by default and enabled by mount -o compress. Even when the -o compress mount option is not used, it is possible to read compressed extents off the disk. If compression for a given set of pages fails to make them smaller, the file is flagged to avoid future compression attempts later. * While finding delalloc extents, the pages are locked before being sent down to the delalloc handler. This allows the delalloc handler to do complex things such as cleaning the pages, marking them writeback and starting IO on their behalf. * Inline extents are inserted at delalloc time now. This allows us to compress the data before inserting the inline extent, and it allows us to insert an inline extent that spans multiple pages. * All of the in-memory extent representations (extent_map.c, ordered-data.c etc) are changed to record both an in-memory size and an on disk size, as well as a flag for compression. From a disk format point of view, the extent pointers in the file are changed to record the on disk size of a given extent and some encoding flags. Space in the disk format is allocated for compression encoding, as well as encryption and a generic 'other' field. Neither the encryption or the 'other' field are currently used. In order to limit the amount of data read for a single random read in the file, the size of a compressed extent is limited to 128k. This is a software only limit, the disk format supports u64 sized compressed extents. In order to limit the ram consumed while processing extents, the uncompressed size of a compressed extent is limited to 256k. This is a software only limit and will be subject to tuning later. Checksumming is still done on compressed extents, and it is done on the uncompressed version of the data. This way additional encodings can be layered on without having to figure out which encoding to checksum. Compression happens at delalloc time, which is basically singled threaded because it is usually done by a single pdflush thread. This makes it tricky to spread the compression load across all the cpus on the box. We'll have to look at parallel pdflush walks of dirty inodes at a later time. Decompression is hooked into readpages and it does spread across CPUs nicely. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-10-30 02:49:59 +08:00
btrfs_set_file_extent_ram_bytes(leaf, fi,
new_extents[i].ram_bytes);
btrfs_set_file_extent_compression(leaf, fi,
new_extents[i].compression);
btrfs_set_file_extent_encryption(leaf, fi,
new_extents[i].encryption);
btrfs_set_file_extent_other_encoding(leaf, fi,
new_extents[i].other_encoding);
2008-09-26 22:09:34 +08:00
btrfs_set_file_extent_num_bytes(leaf, fi,
extent_len);
ext_offset += new_extents[i].offset;
btrfs_set_file_extent_offset(leaf, fi,
ext_offset);
btrfs_mark_buffer_dirty(leaf);
btrfs_drop_extent_cache(inode, key.offset,
key.offset + extent_len - 1, 0);
ret = btrfs_inc_extent_ref(trans, root,
new_extents[i].disk_bytenr,
new_extents[i].disk_num_bytes,
leaf->start,
root->root_key.objectid,
trans->transid, key.objectid);
2008-09-26 22:09:34 +08:00
BUG_ON(ret);
btrfs_release_path(root, path);
inode_add_bytes(inode, extent_len);
2008-09-26 22:09:34 +08:00
ext_offset = 0;
num_bytes -= extent_len;
key.offset += extent_len;
if (num_bytes == 0)
break;
}
BUG_ON(i >= nr_extents);
#endif
2008-09-26 22:09:34 +08:00
}
if (extent_locked) {
unlock_extent(&BTRFS_I(inode)->io_tree, lock_start,
lock_end, GFP_NOFS);
extent_locked = 0;
}
skip:
if (ref_path->owner_objectid != BTRFS_MULTIPLE_OBJECTIDS &&
key.offset >= search_end)
2008-09-26 22:09:34 +08:00
break;
cond_resched();
}
ret = 0;
out:
btrfs_release_path(root, path);
if (inode) {
mutex_unlock(&inode->i_mutex);
if (extent_locked) {
unlock_extent(&BTRFS_I(inode)->io_tree, lock_start,
lock_end, GFP_NOFS);
}
iput(inode);
}
return ret;
}
int btrfs_reloc_tree_cache_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct extent_buffer *buf, u64 orig_start)
{
int level;
int ret;
BUG_ON(btrfs_header_generation(buf) != trans->transid);
BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
level = btrfs_header_level(buf);
if (level == 0) {
struct btrfs_leaf_ref *ref;
struct btrfs_leaf_ref *orig_ref;
orig_ref = btrfs_lookup_leaf_ref(root, orig_start);
if (!orig_ref)
return -ENOENT;
ref = btrfs_alloc_leaf_ref(root, orig_ref->nritems);
if (!ref) {
btrfs_free_leaf_ref(root, orig_ref);
return -ENOMEM;
}
ref->nritems = orig_ref->nritems;
memcpy(ref->extents, orig_ref->extents,
sizeof(ref->extents[0]) * ref->nritems);
btrfs_free_leaf_ref(root, orig_ref);
ref->root_gen = trans->transid;
ref->bytenr = buf->start;
ref->owner = btrfs_header_owner(buf);
ref->generation = btrfs_header_generation(buf);
Btrfs: Make btrfs_drop_snapshot work in larger and more efficient chunks Every transaction in btrfs creates a new snapshot, and then schedules the snapshot from the last transaction for deletion. Snapshot deletion works by walking down the btree and dropping the reference counts on each btree block during the walk. If if a given leaf or node has a reference count greater than one, the reference count is decremented and the subtree pointed to by that node is ignored. If the reference count is one, walking continues down into that node or leaf, and the references of everything it points to are decremented. The old code would try to work in small pieces, walking down the tree until it found the lowest leaf or node to free and then returning. This was very friendly to the rest of the FS because it didn't have a huge impact on other operations. But it wouldn't always keep up with the rate that new commits added new snapshots for deletion, and it wasn't very optimal for the extent allocation tree because it wasn't finding leaves that were close together on disk and processing them at the same time. This changes things to walk down to a level 1 node and then process it in bulk. All the leaf pointers are sorted and the leaves are dropped in order based on their extent number. The extent allocation tree and commit code are now fast enough for this kind of bulk processing to work without slowing the rest of the FS down. Overall it does less IO and is better able to keep up with snapshot deletions under high load. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-02-04 22:27:02 +08:00
2008-09-26 22:09:34 +08:00
ret = btrfs_add_leaf_ref(root, ref, 0);
WARN_ON(ret);
btrfs_free_leaf_ref(root, ref);
}
return 0;
}
static noinline int invalidate_extent_cache(struct btrfs_root *root,
2008-09-26 22:09:34 +08:00
struct extent_buffer *leaf,
struct btrfs_block_group_cache *group,
struct btrfs_root *target_root)
{
struct btrfs_key key;
struct inode *inode = NULL;
struct btrfs_file_extent_item *fi;
u64 num_bytes;
u64 skip_objectid = 0;
u32 nritems;
u32 i;
nritems = btrfs_header_nritems(leaf);
for (i = 0; i < nritems; i++) {
btrfs_item_key_to_cpu(leaf, &key, i);
if (key.objectid == skip_objectid ||
key.type != BTRFS_EXTENT_DATA_KEY)
continue;
fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
if (btrfs_file_extent_type(leaf, fi) ==
BTRFS_FILE_EXTENT_INLINE)
continue;
if (btrfs_file_extent_disk_bytenr(leaf, fi) == 0)
continue;
if (!inode || inode->i_ino != key.objectid) {
iput(inode);
inode = btrfs_ilookup(target_root->fs_info->sb,
key.objectid, target_root, 1);
}
if (!inode) {
skip_objectid = key.objectid;
continue;
}
num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
lock_extent(&BTRFS_I(inode)->io_tree, key.offset,
key.offset + num_bytes - 1, GFP_NOFS);
btrfs_drop_extent_cache(inode, key.offset,
key.offset + num_bytes - 1, 1);
unlock_extent(&BTRFS_I(inode)->io_tree, key.offset,
key.offset + num_bytes - 1, GFP_NOFS);
cond_resched();
}
iput(inode);
return 0;
}
static noinline int replace_extents_in_leaf(struct btrfs_trans_handle *trans,
2008-09-26 22:09:34 +08:00
struct btrfs_root *root,
struct extent_buffer *leaf,
struct btrfs_block_group_cache *group,
struct inode *reloc_inode)
{
struct btrfs_key key;
struct btrfs_key extent_key;
struct btrfs_file_extent_item *fi;
struct btrfs_leaf_ref *ref;
struct disk_extent *new_extent;
u64 bytenr;
u64 num_bytes;
u32 nritems;
u32 i;
int ext_index;
int nr_extent;
int ret;
new_extent = kmalloc(sizeof(*new_extent), GFP_NOFS);
BUG_ON(!new_extent);
ref = btrfs_lookup_leaf_ref(root, leaf->start);
BUG_ON(!ref);
ext_index = -1;
nritems = btrfs_header_nritems(leaf);
for (i = 0; i < nritems; i++) {
btrfs_item_key_to_cpu(leaf, &key, i);
if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
continue;
fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
if (btrfs_file_extent_type(leaf, fi) ==
BTRFS_FILE_EXTENT_INLINE)
continue;
bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
if (bytenr == 0)
continue;
ext_index++;
if (bytenr >= group->key.objectid + group->key.offset ||
bytenr + num_bytes <= group->key.objectid)
continue;
extent_key.objectid = bytenr;
extent_key.offset = num_bytes;
extent_key.type = BTRFS_EXTENT_ITEM_KEY;
nr_extent = 1;
ret = get_new_locations(reloc_inode, &extent_key,
group->key.objectid, 1,
&new_extent, &nr_extent);
if (ret > 0)
continue;
BUG_ON(ret < 0);
BUG_ON(ref->extents[ext_index].bytenr != bytenr);
BUG_ON(ref->extents[ext_index].num_bytes != num_bytes);
ref->extents[ext_index].bytenr = new_extent->disk_bytenr;
ref->extents[ext_index].num_bytes = new_extent->disk_num_bytes;
btrfs_set_file_extent_disk_bytenr(leaf, fi,
new_extent->disk_bytenr);
btrfs_set_file_extent_disk_num_bytes(leaf, fi,
new_extent->disk_num_bytes);
btrfs_mark_buffer_dirty(leaf);
ret = btrfs_inc_extent_ref(trans, root,
new_extent->disk_bytenr,
new_extent->disk_num_bytes,
leaf->start,
root->root_key.objectid,
trans->transid, key.objectid);
2008-09-26 22:09:34 +08:00
BUG_ON(ret);
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
2008-09-26 22:09:34 +08:00
ret = btrfs_free_extent(trans, root,
bytenr, num_bytes, leaf->start,
btrfs_header_owner(leaf),
btrfs_header_generation(leaf),
key.objectid, 0);
2008-09-26 22:09:34 +08:00
BUG_ON(ret);
cond_resched();
}
kfree(new_extent);
BUG_ON(ext_index + 1 != ref->nritems);
btrfs_free_leaf_ref(root, ref);
return 0;
}
int btrfs_free_reloc_root(struct btrfs_trans_handle *trans,
struct btrfs_root *root)
2008-09-26 22:09:34 +08:00
{
struct btrfs_root *reloc_root;
int ret;
2008-09-26 22:09:34 +08:00
if (root->reloc_root) {
reloc_root = root->reloc_root;
root->reloc_root = NULL;
list_add(&reloc_root->dead_list,
&root->fs_info->dead_reloc_roots);
btrfs_set_root_bytenr(&reloc_root->root_item,
reloc_root->node->start);
btrfs_set_root_level(&root->root_item,
btrfs_header_level(reloc_root->node));
memset(&reloc_root->root_item.drop_progress, 0,
sizeof(struct btrfs_disk_key));
reloc_root->root_item.drop_level = 0;
ret = btrfs_update_root(trans, root->fs_info->tree_root,
&reloc_root->root_key,
&reloc_root->root_item);
BUG_ON(ret);
2008-09-26 22:09:34 +08:00
}
return 0;
}
int btrfs_drop_dead_reloc_roots(struct btrfs_root *root)
{
struct btrfs_trans_handle *trans;
struct btrfs_root *reloc_root;
struct btrfs_root *prev_root = NULL;
struct list_head dead_roots;
int ret;
unsigned long nr;
INIT_LIST_HEAD(&dead_roots);
list_splice_init(&root->fs_info->dead_reloc_roots, &dead_roots);
while (!list_empty(&dead_roots)) {
reloc_root = list_entry(dead_roots.prev,
struct btrfs_root, dead_list);
list_del_init(&reloc_root->dead_list);
BUG_ON(reloc_root->commit_root != NULL);
while (1) {
trans = btrfs_join_transaction(root, 1);
BUG_ON(!trans);
mutex_lock(&root->fs_info->drop_mutex);
ret = btrfs_drop_snapshot(trans, reloc_root);
if (ret != -EAGAIN)
break;
mutex_unlock(&root->fs_info->drop_mutex);
nr = trans->blocks_used;
ret = btrfs_end_transaction(trans, root);
BUG_ON(ret);
btrfs_btree_balance_dirty(root, nr);
}
free_extent_buffer(reloc_root->node);
ret = btrfs_del_root(trans, root->fs_info->tree_root,
&reloc_root->root_key);
BUG_ON(ret);
mutex_unlock(&root->fs_info->drop_mutex);
nr = trans->blocks_used;
ret = btrfs_end_transaction(trans, root);
BUG_ON(ret);
btrfs_btree_balance_dirty(root, nr);
kfree(prev_root);
prev_root = reloc_root;
}
if (prev_root) {
btrfs_remove_leaf_refs(prev_root, (u64)-1, 0);
kfree(prev_root);
}
return 0;
}
int btrfs_add_dead_reloc_root(struct btrfs_root *root)
{
list_add(&root->dead_list, &root->fs_info->dead_reloc_roots);
return 0;
}
int btrfs_cleanup_reloc_trees(struct btrfs_root *root)
{
struct btrfs_root *reloc_root;
struct btrfs_trans_handle *trans;
struct btrfs_key location;
int found;
int ret;
mutex_lock(&root->fs_info->tree_reloc_mutex);
ret = btrfs_find_dead_roots(root, BTRFS_TREE_RELOC_OBJECTID, NULL);
BUG_ON(ret);
found = !list_empty(&root->fs_info->dead_reloc_roots);
mutex_unlock(&root->fs_info->tree_reloc_mutex);
if (found) {
trans = btrfs_start_transaction(root, 1);
BUG_ON(!trans);
ret = btrfs_commit_transaction(trans, root);
BUG_ON(ret);
}
location.objectid = BTRFS_DATA_RELOC_TREE_OBJECTID;
location.offset = (u64)-1;
location.type = BTRFS_ROOT_ITEM_KEY;
reloc_root = btrfs_read_fs_root_no_name(root->fs_info, &location);
BUG_ON(!reloc_root);
btrfs_orphan_cleanup(reloc_root);
return 0;
}
static noinline int init_reloc_tree(struct btrfs_trans_handle *trans,
2008-09-26 22:09:34 +08:00
struct btrfs_root *root)
{
struct btrfs_root *reloc_root;
struct extent_buffer *eb;
struct btrfs_root_item *root_item;
struct btrfs_key root_key;
int ret;
BUG_ON(!root->ref_cows);
if (root->reloc_root)
return 0;
root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
BUG_ON(!root_item);
ret = btrfs_copy_root(trans, root, root->commit_root,
&eb, BTRFS_TREE_RELOC_OBJECTID);
BUG_ON(ret);
root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
root_key.offset = root->root_key.objectid;
root_key.type = BTRFS_ROOT_ITEM_KEY;
memcpy(root_item, &root->root_item, sizeof(root_item));
btrfs_set_root_refs(root_item, 0);
btrfs_set_root_bytenr(root_item, eb->start);
btrfs_set_root_level(root_item, btrfs_header_level(eb));
btrfs_set_root_generation(root_item, trans->transid);
2008-09-26 22:09:34 +08:00
btrfs_tree_unlock(eb);
free_extent_buffer(eb);
ret = btrfs_insert_root(trans, root->fs_info->tree_root,
&root_key, root_item);
BUG_ON(ret);
kfree(root_item);
reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root,
&root_key);
BUG_ON(!reloc_root);
reloc_root->last_trans = trans->transid;
reloc_root->commit_root = NULL;
reloc_root->ref_tree = &root->fs_info->reloc_ref_tree;
root->reloc_root = reloc_root;
return 0;
}
/*
* Core function of space balance.
*
* The idea is using reloc trees to relocate tree blocks in reference
* counted roots. There is one reloc tree for each subvol, and all
* reloc trees share same root key objectid. Reloc trees are snapshots
* of the latest committed roots of subvols (root->commit_root).
*
* To relocate a tree block referenced by a subvol, there are two steps.
* COW the block through subvol's reloc tree, then update block pointer
* in the subvol to point to the new block. Since all reloc trees share
* same root key objectid, doing special handing for tree blocks owned
* by them is easy. Once a tree block has been COWed in one reloc tree,
* we can use the resulting new block directly when the same block is
* required to COW again through other reloc trees. By this way, relocated
* tree blocks are shared between reloc trees, so they are also shared
* between subvols.
2008-09-26 22:09:34 +08:00
*/
static noinline int relocate_one_path(struct btrfs_trans_handle *trans,
2008-09-26 22:09:34 +08:00
struct btrfs_root *root,
struct btrfs_path *path,
struct btrfs_key *first_key,
struct btrfs_ref_path *ref_path,
struct btrfs_block_group_cache *group,
struct inode *reloc_inode)
{
struct btrfs_root *reloc_root;
struct extent_buffer *eb = NULL;
struct btrfs_key *keys;
u64 *nodes;
int level;
int shared_level;
2008-09-26 22:09:34 +08:00
int lowest_level = 0;
int ret;
if (ref_path->owner_objectid < BTRFS_FIRST_FREE_OBJECTID)
lowest_level = ref_path->owner_objectid;
if (!root->ref_cows) {
2008-09-26 22:09:34 +08:00
path->lowest_level = lowest_level;
ret = btrfs_search_slot(trans, root, first_key, path, 0, 1);
BUG_ON(ret < 0);
path->lowest_level = 0;
btrfs_release_path(root, path);
return 0;
}
mutex_lock(&root->fs_info->tree_reloc_mutex);
ret = init_reloc_tree(trans, root);
BUG_ON(ret);
reloc_root = root->reloc_root;
shared_level = ref_path->shared_level;
ref_path->shared_level = BTRFS_MAX_LEVEL - 1;
2008-09-26 22:09:34 +08:00
keys = ref_path->node_keys;
nodes = ref_path->new_nodes;
memset(&keys[shared_level + 1], 0,
sizeof(*keys) * (BTRFS_MAX_LEVEL - shared_level - 1));
memset(&nodes[shared_level + 1], 0,
sizeof(*nodes) * (BTRFS_MAX_LEVEL - shared_level - 1));
2008-09-26 22:09:34 +08:00
if (nodes[lowest_level] == 0) {
path->lowest_level = lowest_level;
ret = btrfs_search_slot(trans, reloc_root, first_key, path,
0, 1);
BUG_ON(ret);
for (level = lowest_level; level < BTRFS_MAX_LEVEL; level++) {
eb = path->nodes[level];
if (!eb || eb == reloc_root->node)
break;
nodes[level] = eb->start;
if (level == 0)
btrfs_item_key_to_cpu(eb, &keys[level], 0);
else
btrfs_node_key_to_cpu(eb, &keys[level], 0);
}
if (nodes[0] &&
ref_path->owner_objectid >= BTRFS_FIRST_FREE_OBJECTID) {
eb = path->nodes[0];
ret = replace_extents_in_leaf(trans, reloc_root, eb,
group, reloc_inode);
BUG_ON(ret);
}
btrfs_release_path(reloc_root, path);
} else {
2008-09-26 22:09:34 +08:00
ret = btrfs_merge_path(trans, reloc_root, keys, nodes,
lowest_level);
2008-09-26 22:09:34 +08:00
BUG_ON(ret);
}
/*
* replace tree blocks in the fs tree with tree blocks in
* the reloc tree.
*/
ret = btrfs_merge_path(trans, root, keys, nodes, lowest_level);
BUG_ON(ret < 0);
if (ref_path->owner_objectid >= BTRFS_FIRST_FREE_OBJECTID) {
ret = btrfs_search_slot(trans, reloc_root, first_key, path,
0, 0);
BUG_ON(ret);
extent_buffer_get(path->nodes[0]);
eb = path->nodes[0];
btrfs_release_path(reloc_root, path);
2008-09-26 22:09:34 +08:00
ret = invalidate_extent_cache(reloc_root, eb, group, root);
BUG_ON(ret);
free_extent_buffer(eb);
}
mutex_unlock(&root->fs_info->tree_reloc_mutex);
2008-09-26 22:09:34 +08:00
path->lowest_level = 0;
return 0;
}
static noinline int relocate_tree_block(struct btrfs_trans_handle *trans,
2008-09-26 22:09:34 +08:00
struct btrfs_root *root,
struct btrfs_path *path,
struct btrfs_key *first_key,
struct btrfs_ref_path *ref_path)
{
int ret;
ret = relocate_one_path(trans, root, path, first_key,
ref_path, NULL, NULL);
BUG_ON(ret);
return 0;
}
static noinline int del_extent_zero(struct btrfs_trans_handle *trans,
2008-09-26 22:09:34 +08:00
struct btrfs_root *extent_root,
struct btrfs_path *path,
struct btrfs_key *extent_key)
{
int ret;
ret = btrfs_search_slot(trans, extent_root, extent_key, path, -1, 1);
if (ret)
goto out;
ret = btrfs_del_item(trans, extent_root, path);
out:
btrfs_release_path(extent_root, path);
return ret;
}
static noinline struct btrfs_root *read_ref_root(struct btrfs_fs_info *fs_info,
2008-09-26 22:09:34 +08:00
struct btrfs_ref_path *ref_path)
{
struct btrfs_key root_key;
root_key.objectid = ref_path->root_objectid;
root_key.type = BTRFS_ROOT_ITEM_KEY;
if (is_cowonly_root(ref_path->root_objectid))
root_key.offset = 0;
else
root_key.offset = (u64)-1;
return btrfs_read_fs_root_no_name(fs_info, &root_key);
}
static noinline int relocate_one_extent(struct btrfs_root *extent_root,
2008-09-26 22:09:34 +08:00
struct btrfs_path *path,
struct btrfs_key *extent_key,
struct btrfs_block_group_cache *group,
struct inode *reloc_inode, int pass)
{
struct btrfs_trans_handle *trans;
struct btrfs_root *found_root;
struct btrfs_ref_path *ref_path = NULL;
struct disk_extent *new_extents = NULL;
int nr_extents = 0;
int loops;
int ret;
int level;
struct btrfs_key first_key;
u64 prev_block = 0;
trans = btrfs_start_transaction(extent_root, 1);
BUG_ON(!trans);
if (extent_key->objectid == 0) {
ret = del_extent_zero(trans, extent_root, path, extent_key);
goto out;
}
ref_path = kmalloc(sizeof(*ref_path), GFP_NOFS);
if (!ref_path) {
ret = -ENOMEM;
goto out;
2008-09-26 22:09:34 +08:00
}
for (loops = 0; ; loops++) {
if (loops == 0) {
ret = btrfs_first_ref_path(trans, extent_root, ref_path,
extent_key->objectid);
} else {
ret = btrfs_next_ref_path(trans, extent_root, ref_path);
}
if (ret < 0)
goto out;
if (ret > 0)
break;
if (ref_path->root_objectid == BTRFS_TREE_LOG_OBJECTID ||
ref_path->root_objectid == BTRFS_TREE_RELOC_OBJECTID)
continue;
found_root = read_ref_root(extent_root->fs_info, ref_path);
BUG_ON(!found_root);
/*
* for reference counted tree, only process reference paths
* rooted at the latest committed root.
*/
if (found_root->ref_cows &&
ref_path->root_generation != found_root->root_key.offset)
continue;
if (ref_path->owner_objectid >= BTRFS_FIRST_FREE_OBJECTID) {
if (pass == 0) {
/*
* copy data extents to new locations
*/
u64 group_start = group->key.objectid;
ret = relocate_data_extent(reloc_inode,
extent_key,
group_start);
if (ret < 0)
goto out;
break;
}
level = 0;
} else {
level = ref_path->owner_objectid;
}
if (prev_block != ref_path->nodes[level]) {
struct extent_buffer *eb;
u64 block_start = ref_path->nodes[level];
u64 block_size = btrfs_level_size(found_root, level);
eb = read_tree_block(found_root, block_start,
block_size, 0);
btrfs_tree_lock(eb);
BUG_ON(level != btrfs_header_level(eb));
if (level == 0)
btrfs_item_key_to_cpu(eb, &first_key, 0);
else
btrfs_node_key_to_cpu(eb, &first_key, 0);
btrfs_tree_unlock(eb);
free_extent_buffer(eb);
prev_block = block_start;
}
mutex_lock(&extent_root->fs_info->trans_mutex);
btrfs_record_root_in_trans(found_root);
mutex_unlock(&extent_root->fs_info->trans_mutex);
if (ref_path->owner_objectid >= BTRFS_FIRST_FREE_OBJECTID) {
/*
* try to update data extent references while
* keeping metadata shared between snapshots.
*/
if (pass == 1) {
ret = relocate_one_path(trans, found_root,
path, &first_key, ref_path,
group, reloc_inode);
if (ret < 0)
goto out;
continue;
}
2008-09-26 22:09:34 +08:00
/*
* use fallback method to process the remaining
* references.
*/
if (!new_extents) {
u64 group_start = group->key.objectid;
new_extents = kmalloc(sizeof(*new_extents),
GFP_NOFS);
nr_extents = 1;
2008-09-26 22:09:34 +08:00
ret = get_new_locations(reloc_inode,
extent_key,
group_start, 1,
2008-09-26 22:09:34 +08:00
&new_extents,
&nr_extents);
if (ret)
2008-09-26 22:09:34 +08:00
goto out;
}
ret = replace_one_extent(trans, found_root,
path, extent_key,
&first_key, ref_path,
new_extents, nr_extents);
} else {
2008-09-26 22:09:34 +08:00
ret = relocate_tree_block(trans, found_root, path,
&first_key, ref_path);
}
if (ret < 0)
goto out;
}
ret = 0;
out:
btrfs_end_transaction(trans, extent_root);
kfree(new_extents);
kfree(ref_path);
return ret;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
#endif
2008-09-26 22:09:34 +08:00
static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
{
u64 num_devices;
u64 stripped = BTRFS_BLOCK_GROUP_RAID0 |
BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
num_devices = root->fs_info->fs_devices->rw_devices;
if (num_devices == 1) {
stripped |= BTRFS_BLOCK_GROUP_DUP;
stripped = flags & ~stripped;
/* turn raid0 into single device chunks */
if (flags & BTRFS_BLOCK_GROUP_RAID0)
return stripped;
/* turn mirroring into duplication */
if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
BTRFS_BLOCK_GROUP_RAID10))
return stripped | BTRFS_BLOCK_GROUP_DUP;
return flags;
} else {
/* they already had raid on here, just return */
if (flags & stripped)
return flags;
stripped |= BTRFS_BLOCK_GROUP_DUP;
stripped = flags & ~stripped;
/* switch duplicated blocks with raid1 */
if (flags & BTRFS_BLOCK_GROUP_DUP)
return stripped | BTRFS_BLOCK_GROUP_RAID1;
/* turn single device chunks into raid0 */
return stripped | BTRFS_BLOCK_GROUP_RAID0;
}
return flags;
}
static int __alloc_chunk_for_shrink(struct btrfs_root *root,
struct btrfs_block_group_cache *shrink_block_group,
int force)
{
struct btrfs_trans_handle *trans;
u64 new_alloc_flags;
u64 calc;
spin_lock(&shrink_block_group->lock);
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
if (btrfs_block_group_used(&shrink_block_group->item) +
shrink_block_group->reserved > 0) {
spin_unlock(&shrink_block_group->lock);
trans = btrfs_start_transaction(root, 1);
spin_lock(&shrink_block_group->lock);
new_alloc_flags = update_block_group_flags(root,
shrink_block_group->flags);
if (new_alloc_flags != shrink_block_group->flags) {
calc =
btrfs_block_group_used(&shrink_block_group->item);
} else {
calc = shrink_block_group->key.offset;
}
spin_unlock(&shrink_block_group->lock);
do_chunk_alloc(trans, root->fs_info->extent_root,
calc + 2 * 1024 * 1024, new_alloc_flags, force);
btrfs_end_transaction(trans, root);
} else
spin_unlock(&shrink_block_group->lock);
return 0;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 22:45:14 +08:00
int btrfs_prepare_block_group_relocation(struct btrfs_root *root,
struct btrfs_block_group_cache *group)
{
__alloc_chunk_for_shrink(root, group, 1);
set_block_group_readonly(group);
return 0;
}
/*
* checks to see if its even possible to relocate this block group.
*
* @return - -1 if it's not a good idea to relocate this block group, 0 if its
* ok to go ahead and try.
*/
int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
2008-09-26 22:09:34 +08:00
{
struct btrfs_block_group_cache *block_group;
struct btrfs_space_info *space_info;
struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
struct btrfs_device *device;
int full = 0;
int ret = 0;
2008-09-26 22:09:34 +08:00
block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2008-09-26 22:09:34 +08:00
/* odd, couldn't find the block group, leave it alone */
if (!block_group)
return -1;
2008-09-26 22:09:34 +08:00
/* no bytes used, we're good */
if (!btrfs_block_group_used(&block_group->item))
2008-09-26 22:09:34 +08:00
goto out;
space_info = block_group->space_info;
spin_lock(&space_info->lock);
full = space_info->full;
/*
* if this is the last block group we have in this space, we can't
* relocate it unless we're able to allocate a new chunk below.
*
* Otherwise, we need to make sure we have room in the space to handle
* all of the extents from this block group. If we can, we're good
*/
if ((space_info->total_bytes != block_group->key.offset) &&
(space_info->bytes_used + space_info->bytes_reserved +
space_info->bytes_pinned + space_info->bytes_readonly +
btrfs_block_group_used(&block_group->item) <
space_info->total_bytes)) {
spin_unlock(&space_info->lock);
goto out;
}
spin_unlock(&space_info->lock);
/*
* ok we don't have enough space, but maybe we have free space on our
* devices to allocate new chunks for relocation, so loop through our
* alloc devices and guess if we have enough space. However, if we
* were marked as full, then we know there aren't enough chunks, and we
* can just return.
*/
ret = -1;
if (full)
goto out;
mutex_lock(&root->fs_info->chunk_mutex);
list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
u64 min_free = btrfs_block_group_used(&block_group->item);
u64 dev_offset, max_avail;
Btrfs: do extent allocation and reference count updates in the background The extent allocation tree maintains a reference count and full back reference information for every extent allocated in the filesystem. For subvolume and snapshot trees, every time a block goes through COW, the new copy of the block adds a reference on every block it points to. If a btree node points to 150 leaves, then the COW code needs to go and add backrefs on 150 different extents, which might be spread all over the extent allocation tree. These updates currently happen during btrfs_cow_block, and most COWs happen during btrfs_search_slot. btrfs_search_slot has locks held on both the parent and the node we are COWing, and so we really want to avoid IO during the COW if we can. This commit adds an rbtree of pending reference count updates and extent allocations. The tree is ordered by byte number of the extent and byte number of the parent for the back reference. The tree allows us to: 1) Modify back references in something close to disk order, reducing seeks 2) Significantly reduce the number of modifications made as block pointers are balanced around 3) Do all of the extent insertion and back reference modifications outside of the performance critical btrfs_search_slot code. #3 has the added benefit of greatly reducing the btrfs stack footprint. The extent allocation tree modifications are done without the deep (and somewhat recursive) call chains used in the past. These delayed back reference updates must be done before the transaction commits, and so the rbtree is tied to the transaction. Throttling is implemented to help keep the queue of backrefs at a reasonable size. Since there was a similar mechanism in place for the extent tree extents, that is removed and replaced by the delayed reference tree. Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-13 22:10:06 +08:00
/*
* check to make sure we can actually find a chunk with enough
* space to fit our block group in.
*/
if (device->total_bytes > device->bytes_used + min_free) {
ret = find_free_dev_extent(NULL, device, min_free,
&dev_offset, &max_avail);
if (!ret)
break;
ret = -1;
}
}
mutex_unlock(&root->fs_info->chunk_mutex);
out:
btrfs_put_block_group(block_group);
return ret;
}
static int find_first_block_group(struct btrfs_root *root,
struct btrfs_path *path, struct btrfs_key *key)
{
int ret = 0;
struct btrfs_key found_key;
struct extent_buffer *leaf;
int slot;
ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
if (ret < 0)
goto out;
while (1) {
slot = path->slots[0];
leaf = path->nodes[0];
if (slot >= btrfs_header_nritems(leaf)) {
ret = btrfs_next_leaf(root, path);
if (ret == 0)
continue;
if (ret < 0)
goto out;
break;
}
btrfs_item_key_to_cpu(leaf, &found_key, slot);
if (found_key.objectid >= key->objectid &&
found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
ret = 0;
goto out;
}
path->slots[0]++;
}
ret = -ENOENT;
out:
return ret;
}
2008-09-26 22:09:34 +08:00
int btrfs_free_block_groups(struct btrfs_fs_info *info)
{
struct btrfs_block_group_cache *block_group;
struct btrfs_space_info *space_info;
struct btrfs_caching_control *caching_ctl;
2008-09-26 22:09:34 +08:00
struct rb_node *n;
down_write(&info->extent_commit_sem);
while (!list_empty(&info->caching_block_groups)) {
caching_ctl = list_entry(info->caching_block_groups.next,
struct btrfs_caching_control, list);
list_del(&caching_ctl->list);
put_caching_control(caching_ctl);
}
up_write(&info->extent_commit_sem);
2008-09-26 22:09:34 +08:00
spin_lock(&info->block_group_cache_lock);
while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
block_group = rb_entry(n, struct btrfs_block_group_cache,
cache_node);
rb_erase(&block_group->cache_node,
&info->block_group_cache_tree);
spin_unlock(&info->block_group_cache_lock);
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
down_write(&block_group->space_info->groups_sem);
2008-09-26 22:09:34 +08:00
list_del(&block_group->list);
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
up_write(&block_group->space_info->groups_sem);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
if (block_group->cached == BTRFS_CACHE_STARTED)
wait_block_group_cache_done(block_group);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
btrfs_remove_free_space_cache(block_group);
WARN_ON(atomic_read(&block_group->count) != 1);
2008-09-26 22:09:34 +08:00
kfree(block_group);
spin_lock(&info->block_group_cache_lock);
2008-09-26 22:09:34 +08:00
}
spin_unlock(&info->block_group_cache_lock);
/* now that all the block groups are freed, go through and
* free all the space_info structs. This is only called during
* the final stages of unmount, and so we know nobody is
* using them. We call synchronize_rcu() once before we start,
* just to be on the safe side.
*/
synchronize_rcu();
while(!list_empty(&info->space_info)) {
space_info = list_entry(info->space_info.next,
struct btrfs_space_info,
list);
list_del(&space_info->list);
kfree(space_info);
}
2008-09-26 22:09:34 +08:00
return 0;
}
int btrfs_read_block_groups(struct btrfs_root *root)
{
struct btrfs_path *path;
int ret;
struct btrfs_block_group_cache *cache;
struct btrfs_fs_info *info = root->fs_info;
struct btrfs_space_info *space_info;
struct btrfs_key key;
struct btrfs_key found_key;
struct extent_buffer *leaf;
root = info->extent_root;
key.objectid = 0;
key.offset = 0;
btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
while (1) {
ret = find_first_block_group(root, path, &key);
if (ret > 0) {
ret = 0;
goto error;
}
if (ret != 0)
goto error;
leaf = path->nodes[0];
btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
cache = kzalloc(sizeof(*cache), GFP_NOFS);
if (!cache) {
ret = -ENOMEM;
break;
}
atomic_set(&cache->count, 1);
spin_lock_init(&cache->lock);
spin_lock_init(&cache->tree_lock);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
cache->fs_info = info;
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
INIT_LIST_HEAD(&cache->list);
INIT_LIST_HEAD(&cache->cluster_list);
Btrfs: use hybrid extents+bitmap rb tree for free space Currently btrfs has a problem where it can use a ridiculous amount of RAM simply tracking free space. As free space gets fragmented, we end up with thousands of entries on an rb-tree per block group, which usually spans 1 gig of area. Since we currently don't ever flush free space cache back to disk this gets to be a bit unweildly on large fs's with lots of fragmentation. This patch solves this problem by using PAGE_SIZE bitmaps for parts of the free space cache. Initially we calculate a threshold of extent entries we can handle, which is however many extent entries we can cram into 16k of ram. The maximum amount of RAM that should ever be used to track 1 gigabyte of diskspace will be 32k of RAM, which scales much better than we did before. Once we pass the extent threshold, we start adding bitmaps and using those instead for tracking the free space. This patch also makes it so that any free space thats less than 4 * sectorsize we go ahead and put into a bitmap. This is nice since we try and allocate out of the front of a block group, so if the front of a block group is heavily fragmented and then has a huge chunk of free space at the end, we go ahead and add the fragmented areas to bitmaps and use a normal extent entry to track the big chunk at the back of the block group. I've also taken the opportunity to revamp how we search for free space. Previously we indexed free space via an offset indexed rb tree and a bytes indexed rb tree. I've dropped the bytes indexed rb tree and use only the offset indexed rb tree. This cuts the number of tree operations we were doing previously down by half, and gives us a little bit of a better allocation pattern since we will always start from a specific offset and search forward from there, instead of searching for the size we need and try and get it as close as possible to the offset we want. I've given this a healthy amount of testing pre-new format stuff, as well as post-new format stuff. I've booted up my fedora box which is installed on btrfs with this patch and ran with it for a few days without issues. I've not seen any performance regressions in any of my tests. Since the last patch Yan Zheng fixed a problem where we could have overlapping entries, so updating their offset inline would cause problems. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
/*
* we only want to have 32k of ram per block group for keeping
* track of free space, and if we pass 1/2 of that we want to
* start converting things over to using bitmaps
*/
cache->extents_thresh = ((1024 * 32) / 2) /
sizeof(struct btrfs_free_space);
read_extent_buffer(leaf, &cache->item,
btrfs_item_ptr_offset(leaf, path->slots[0]),
sizeof(cache->item));
memcpy(&cache->key, &found_key, sizeof(found_key));
key.objectid = found_key.objectid + found_key.offset;
btrfs_release_path(root, path);
cache->flags = btrfs_block_group_flags(&cache->item);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
cache->sectorsize = root->sectorsize;
/*
* check for two cases, either we are full, and therefore
* don't need to bother with the caching work since we won't
* find any space, or we are empty, and we can just add all
* the space in and be done with it. This saves us _alot_ of
* time, particularly in the full case.
*/
if (found_key.offset == btrfs_block_group_used(&cache->item)) {
exclude_super_stripes(root, cache);
cache->last_byte_to_unpin = (u64)-1;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
cache->cached = BTRFS_CACHE_FINISHED;
free_excluded_extents(root, cache);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
} else if (btrfs_block_group_used(&cache->item) == 0) {
exclude_super_stripes(root, cache);
cache->last_byte_to_unpin = (u64)-1;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
cache->cached = BTRFS_CACHE_FINISHED;
add_new_free_space(cache, root->fs_info,
found_key.objectid,
found_key.objectid +
found_key.offset);
free_excluded_extents(root, cache);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
}
ret = update_space_info(info, cache->flags, found_key.offset,
btrfs_block_group_used(&cache->item),
&space_info);
BUG_ON(ret);
cache->space_info = space_info;
spin_lock(&cache->space_info->lock);
cache->space_info->bytes_super += cache->bytes_super;
spin_unlock(&cache->space_info->lock);
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
down_write(&space_info->groups_sem);
list_add_tail(&cache->list, &space_info->block_groups);
up_write(&space_info->groups_sem);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
ret = btrfs_add_block_group_cache(root->fs_info, cache);
BUG_ON(ret);
set_avail_alloc_bits(root->fs_info, cache->flags);
if (btrfs_chunk_readonly(root, cache->key.objectid))
set_block_group_readonly(cache);
}
ret = 0;
error:
btrfs_free_path(path);
return ret;
}
int btrfs_make_block_group(struct btrfs_trans_handle *trans,
struct btrfs_root *root, u64 bytes_used,
u64 type, u64 chunk_objectid, u64 chunk_offset,
u64 size)
{
int ret;
struct btrfs_root *extent_root;
struct btrfs_block_group_cache *cache;
extent_root = root->fs_info->extent_root;
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 22:24:20 +08:00
root->fs_info->last_trans_log_full_commit = trans->transid;
cache = kzalloc(sizeof(*cache), GFP_NOFS);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
if (!cache)
return -ENOMEM;
cache->key.objectid = chunk_offset;
cache->key.offset = size;
cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
Btrfs: use hybrid extents+bitmap rb tree for free space Currently btrfs has a problem where it can use a ridiculous amount of RAM simply tracking free space. As free space gets fragmented, we end up with thousands of entries on an rb-tree per block group, which usually spans 1 gig of area. Since we currently don't ever flush free space cache back to disk this gets to be a bit unweildly on large fs's with lots of fragmentation. This patch solves this problem by using PAGE_SIZE bitmaps for parts of the free space cache. Initially we calculate a threshold of extent entries we can handle, which is however many extent entries we can cram into 16k of ram. The maximum amount of RAM that should ever be used to track 1 gigabyte of diskspace will be 32k of RAM, which scales much better than we did before. Once we pass the extent threshold, we start adding bitmaps and using those instead for tracking the free space. This patch also makes it so that any free space thats less than 4 * sectorsize we go ahead and put into a bitmap. This is nice since we try and allocate out of the front of a block group, so if the front of a block group is heavily fragmented and then has a huge chunk of free space at the end, we go ahead and add the fragmented areas to bitmaps and use a normal extent entry to track the big chunk at the back of the block group. I've also taken the opportunity to revamp how we search for free space. Previously we indexed free space via an offset indexed rb tree and a bytes indexed rb tree. I've dropped the bytes indexed rb tree and use only the offset indexed rb tree. This cuts the number of tree operations we were doing previously down by half, and gives us a little bit of a better allocation pattern since we will always start from a specific offset and search forward from there, instead of searching for the size we need and try and get it as close as possible to the offset we want. I've given this a healthy amount of testing pre-new format stuff, as well as post-new format stuff. I've booted up my fedora box which is installed on btrfs with this patch and ran with it for a few days without issues. I've not seen any performance regressions in any of my tests. Since the last patch Yan Zheng fixed a problem where we could have overlapping entries, so updating their offset inline would cause problems. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
cache->sectorsize = root->sectorsize;
/*
* we only want to have 32k of ram per block group for keeping track
* of free space, and if we pass 1/2 of that we want to start
* converting things over to using bitmaps
*/
cache->extents_thresh = ((1024 * 32) / 2) /
sizeof(struct btrfs_free_space);
atomic_set(&cache->count, 1);
spin_lock_init(&cache->lock);
spin_lock_init(&cache->tree_lock);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
INIT_LIST_HEAD(&cache->list);
INIT_LIST_HEAD(&cache->cluster_list);
btrfs_set_block_group_used(&cache->item, bytes_used);
btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
cache->flags = type;
btrfs_set_block_group_flags(&cache->item, type);
cache->last_byte_to_unpin = (u64)-1;
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
cache->cached = BTRFS_CACHE_FINISHED;
exclude_super_stripes(root, cache);
Btrfs: use hybrid extents+bitmap rb tree for free space Currently btrfs has a problem where it can use a ridiculous amount of RAM simply tracking free space. As free space gets fragmented, we end up with thousands of entries on an rb-tree per block group, which usually spans 1 gig of area. Since we currently don't ever flush free space cache back to disk this gets to be a bit unweildly on large fs's with lots of fragmentation. This patch solves this problem by using PAGE_SIZE bitmaps for parts of the free space cache. Initially we calculate a threshold of extent entries we can handle, which is however many extent entries we can cram into 16k of ram. The maximum amount of RAM that should ever be used to track 1 gigabyte of diskspace will be 32k of RAM, which scales much better than we did before. Once we pass the extent threshold, we start adding bitmaps and using those instead for tracking the free space. This patch also makes it so that any free space thats less than 4 * sectorsize we go ahead and put into a bitmap. This is nice since we try and allocate out of the front of a block group, so if the front of a block group is heavily fragmented and then has a huge chunk of free space at the end, we go ahead and add the fragmented areas to bitmaps and use a normal extent entry to track the big chunk at the back of the block group. I've also taken the opportunity to revamp how we search for free space. Previously we indexed free space via an offset indexed rb tree and a bytes indexed rb tree. I've dropped the bytes indexed rb tree and use only the offset indexed rb tree. This cuts the number of tree operations we were doing previously down by half, and gives us a little bit of a better allocation pattern since we will always start from a specific offset and search forward from there, instead of searching for the size we need and try and get it as close as possible to the offset we want. I've given this a healthy amount of testing pre-new format stuff, as well as post-new format stuff. I've booted up my fedora box which is installed on btrfs with this patch and ran with it for a few days without issues. I've not seen any performance regressions in any of my tests. Since the last patch Yan Zheng fixed a problem where we could have overlapping entries, so updating their offset inline would cause problems. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
add_new_free_space(cache, root->fs_info, chunk_offset,
chunk_offset + size);
free_excluded_extents(root, cache);
ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
&cache->space_info);
BUG_ON(ret);
spin_lock(&cache->space_info->lock);
cache->space_info->bytes_super += cache->bytes_super;
spin_unlock(&cache->space_info->lock);
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
down_write(&cache->space_info->groups_sem);
list_add_tail(&cache->list, &cache->space_info->block_groups);
up_write(&cache->space_info->groups_sem);
Btrfs: free space accounting redo 1) replace the per fs_info extent_io_tree that tracked free space with two rb-trees per block group to track free space areas via offset and size. The reason to do this is because most allocations come with a hint byte where to start, so we can usually find a chunk of free space at that hint byte to satisfy the allocation and get good space packing. If we cannot find free space at or after the given offset we fall back on looking for a chunk of the given size as close to that given offset as possible. When we fall back on the size search we also try to find a slot as close to the size we want as possible, to avoid breaking small chunks off of huge areas if possible. 2) remove the extent_io_tree that tracked the block group cache from fs_info and replaced it with an rb-tree thats tracks block group cache via offset. also added a per space_info list that tracks the block group cache for the particular space so we can lookup related block groups easily. 3) cleaned up the allocation code to make it a little easier to read and a little less complicated. Basically there are 3 steps, first look from our provided hint. If we couldn't find from that given hint, start back at our original search start and look for space from there. If that fails try to allocate space if we can and start looking again. If not we're screwed and need to start over again. 4) small fixes. there were some issues in volumes.c where we wouldn't allocate the rest of the disk. fixed cow_file_range to actually pass the alloc_hint, which has helped a good bit in making the fs_mark test I run have semi-normal results as we run out of space. Generally with data allocations we don't track where we last allocated from, so everytime we did a data allocation we'd search through every block group that we have looking for free space. Now searching a block group with no free space isn't terribly time consuming, it was causing a slight degradation as we got more data block groups. The alloc_hint has fixed this slight degredation and made things semi-normal. There is still one nagging problem I'm working on where we will get ENOSPC when there is definitely plenty of space. This only happens with metadata allocations, and only when we are almost full. So you generally hit the 85% mark first, but sometimes you'll hit the BUG before you hit the 85% wall. I'm still tracking it down, but until then this seems to be pretty stable and make a significant performance gain. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-09-24 01:14:11 +08:00
ret = btrfs_add_block_group_cache(root->fs_info, cache);
BUG_ON(ret);
ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
sizeof(cache->item));
BUG_ON(ret);
set_avail_alloc_bits(extent_root->fs_info, type);
return 0;
}
2008-09-26 22:09:34 +08:00
int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
struct btrfs_root *root, u64 group_start)
{
struct btrfs_path *path;
struct btrfs_block_group_cache *block_group;
struct btrfs_free_cluster *cluster;
2008-09-26 22:09:34 +08:00
struct btrfs_key key;
int ret;
root = root->fs_info->extent_root;
block_group = btrfs_lookup_block_group(root->fs_info, group_start);
BUG_ON(!block_group);
BUG_ON(!block_group->ro);
2008-09-26 22:09:34 +08:00
memcpy(&key, &block_group->key, sizeof(key));
/* make sure this block group isn't part of an allocation cluster */
cluster = &root->fs_info->data_alloc_cluster;
spin_lock(&cluster->refill_lock);
btrfs_return_cluster_to_free_space(block_group, cluster);
spin_unlock(&cluster->refill_lock);
/*
* make sure this block group isn't part of a metadata
* allocation cluster
*/
cluster = &root->fs_info->meta_alloc_cluster;
spin_lock(&cluster->refill_lock);
btrfs_return_cluster_to_free_space(block_group, cluster);
spin_unlock(&cluster->refill_lock);
2008-09-26 22:09:34 +08:00
path = btrfs_alloc_path();
BUG_ON(!path);
spin_lock(&root->fs_info->block_group_cache_lock);
2008-09-26 22:09:34 +08:00
rb_erase(&block_group->cache_node,
&root->fs_info->block_group_cache_tree);
spin_unlock(&root->fs_info->block_group_cache_lock);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
down_write(&block_group->space_info->groups_sem);
/*
* we must use list_del_init so people can check to see if they
* are still on the list after taking the semaphore
*/
list_del_init(&block_group->list);
Btrfs: fix enospc when there is plenty of space So there is an odd case where we can possibly return -ENOSPC when there is in fact space to be had. It only happens with Metadata writes, and happens _very_ infrequently. What has to happen is we have to allocate have allocated out of the first logical byte on the disk, which would set last_alloc to first_logical_byte(root, 0), so search_start == orig_search_start. We then need to allocate for normal metadata, so BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DUP. We will do a block lookup for the given search_start, block_group_bits() won't match and we'll go to choose another block group. However because search_start matches orig_search_start we go to see if we can allocate a chunk. If we are in the situation that we cannot allocate a chunk, we fail and ENOSPC. This is kind of a big flaw of the way find_free_extent works, as it along with find_free_space loop through _all_ of the block groups, not just the ones that we want to allocate out of. This patch completely kills find_free_space and rolls it into find_free_extent. I've introduced a sort of state machine into this, which will make it easier to get cache miss information out of the allocator, and will work well with my locking changes. The basic flow is this: We have the variable loop which is 0, meaning we are in the hint phase. We lookup the block group for the hint, and lookup the space_info for what we want to allocate out of. If the block group we were pointed at by the hint either isn't of the correct type, or just doesn't have the space we need, we set head to space_info->block_groups, so we start at the beginning of the block groups for this particular space info, and loop through. This is also where we add the empty_cluster to total_needed. At this point loop is set to 1 and we just loop through all of the block groups for this particular space_info looking for the space we need, just as find_free_space would have done, except we only hit the block groups we want and not _all_ of the block groups. If we come full circle we see if we can allocate a chunk. If we cannot of course we exit with -ENOSPC and we are good. If not we start over at space_info->block_groups and loop through again, with loop == 2. If we come full circle and haven't found what we need then we exit with -ENOSPC. I've been running this for a couple of days now and it seems stable, and I haven't yet hit a -ENOSPC when there was plenty of space left. Also I've made a groups_sem to handle the group list for the space_info. This is part of my locking changes, but is relatively safe and seems better than holding the space_info spinlock over that entire search time. Thanks, Signed-off-by: Josef Bacik <jbacik@redhat.com>
2008-10-30 02:49:05 +08:00
up_write(&block_group->space_info->groups_sem);
2008-09-26 22:09:34 +08:00
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
if (block_group->cached == BTRFS_CACHE_STARTED)
wait_block_group_cache_done(block_group);
Btrfs: async block group caching This patch moves the caching of the block group off to a kthread in order to allow people to allocate sooner. Instead of blocking up behind the caching mutex, we instead kick of the caching kthread, and then attempt to make an allocation. If we cannot, we wait on the block groups caching waitqueue, which the caching kthread will wake the waiting threads up everytime it finds 2 meg worth of space, and then again when its finished caching. This is how I tested the speedup from this mkfs the disk mount the disk fill the disk up with fs_mark unmount the disk mount the disk time touch /mnt/foo Without my changes this took 11 seconds on my box, with these changes it now takes 1 second. Another change thats been put in place is we lock the super mirror's in the pinned extent map in order to keep us from adding that stuff as free space when caching the block group. This doesn't really change anything else as far as the pinned extent map is concerned, since for actual pinned extents we use EXTENT_DIRTY, but it does mean that when we unmount we have to go in and unlock those extents to keep from leaking memory. I've also added a check where when we are reading block groups from disk, if the amount of space used == the size of the block group, we go ahead and mark the block group as cached. This drastically reduces the amount of time it takes to cache the block groups. Using the same test as above, except doing a dd to a file and then unmounting, it used to take 33 seconds to umount, now it takes 3 seconds. This version uses the commit_root in the caching kthread, and then keeps track of how many async caching threads are running at any given time so if one of the async threads is still running as we cross transactions we can wait until its finished before handling the pinned extents. Thank you, Signed-off-by: Josef Bacik <jbacik@redhat.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-07-14 09:29:25 +08:00
btrfs_remove_free_space_cache(block_group);
spin_lock(&block_group->space_info->lock);
block_group->space_info->total_bytes -= block_group->key.offset;
block_group->space_info->bytes_readonly -= block_group->key.offset;
spin_unlock(&block_group->space_info->lock);
btrfs_clear_space_info_full(root->fs_info);
btrfs_put_block_group(block_group);
btrfs_put_block_group(block_group);
2008-09-26 22:09:34 +08:00
ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
if (ret > 0)
ret = -EIO;
if (ret < 0)
goto out;
ret = btrfs_del_item(trans, root, path);
out:
btrfs_free_path(path);
return ret;
}