Dave Sterba pointed out a sleeping while atomic bug while doing fsync. This
is because I'm an idiot and didn't realize that rwlock's were spin locks, so
we've been holding this thing while doing allocations and such which is not
good. This patch fixes this by dropping the write lock before we do
anything heavy and re-acquire it when it is done. We also need to take a
ref on the em's in case their corresponding pages are evicted and mark them
as being logged so that releasepage does not remove them and doesn't remove
them from our local list. Thanks,
Reported-by: Dave Sterba <dave@jikos.cz>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
We forget to protect ->log_batch when syncing a file, this patch fix
this problem by atomic operation. And ->log_batch is used to check
if there are parallel sync operations or not, so it is unnecessary to
reset it to 0 after the sync operation of the current log tree complete.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
I audited all users of btrfs_drop_extents and found that nobody actually uses
the hint_byte argument. I'm sure it was used for something at some point but
it's not used now, and the way the pinning works the disk bytenr would never be
immediately useful anyway so lets just remove it. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
This is based on Josef's "Btrfs: turbo charge fsync".
If an inode is a BTRFS_INODE_NODATASUM one, we don't need to look for csum
items any more.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
This is based on Josef's "Btrfs: turbo charge fsync".
The current btrfs checks if an inode is in log by comparing
root's last_log_commit to inode's last_sub_trans[2].
But the problem is that this root->last_log_commit is shared among
inodes.
Say we have N inodes to be logged, after the first inode,
root's last_log_commit is updated and the N-1 remained files will
be skipped.
This fixes the bug by keeping a local copy of root's last_log_commit
inside each inode and this local copy will be maintained itself.
[1]: we regard each log transaction as a subset of btrfs's transaction,
i.e. sub_trans
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
This is based on Josef's "Btrfs: turbo charge fsync".
The above Josef's patch performs very good in random sync write test,
because we won't have too much extents to merge.
However, it does not performs good on the test:
dd if=/dev/zero of=foobar bs=4k count=12500 oflag=sync
The reason is when we do sequencial sync write, we need to merge the
current extent just with the previous one, so that we can get accumulated
extents to log:
A(4k) --> AA(8k) --> AAA(12k) --> AAAA(16k) ...
So we'll have to flush more and more checksum into log tree, which is the
bottleneck according to my tests.
But we can avoid this by telling fsync the real extents that are needed
to be logged.
With this, I did the above dd sync write test (size=50m),
w/o (orig) w/ (josef's) w/ (this)
SATA 104KB/s 109KB/s 121KB/s
ramdisk 1.5MB/s 1.5MB/s 10.7MB/s (613%)
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
This is based on Josef's "Btrfs: turbo charge fsync".
We should cleanup those extents after we've finished logging inode,
otherwise we may do redundant work on them.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
I hit this a couple times while working on my fsync patch (all my bugs, not
normal operation), but with my new stuff we could have new errors from cases
I have not encountered, so instead of BUG()'ing we should be WARN()'ing so
that we are notified there is a problem but the user doesn't lose their
data. We can easily commit the transaction in the case that the tree
logging fails and still be fine, so let's try and be as nice to the user as
possible. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
At least for the vm workload. Currently on fsync we will
1) Truncate all items in the log tree for the given inode if they exist
and
2) Copy all items for a given inode into the log
The problem with this is that for things like VMs you can have lots of
extents from the fragmented writing behavior, and worst yet you may have
only modified a few extents, not the entire thing. This patch fixes this
problem by tracking which transid modified our extent, and then when we do
the tree logging we find all of the extents we've modified in our current
transaction, sort them and commit them. We also only truncate up to the
xattrs of the inode and copy that stuff in normally, and then just drop any
extents in the range we have that exist in the log already. Here are some
numbers of a 50 meg fio job that does random writes and fsync()s after every
write
Original Patched
SATA drive 82KB/s 140KB/s
Fusion drive 431KB/s 2532KB/s
So around 2-6 times faster depending on your hardware. There are a few
corner cases, for example if you truncate at all we have to do it the old
way since there is no way to be sure what is in the log is ok. This
probably could be done smarter, but if you write-fsync-truncate-write-fsync
you deserve what you get. All this work is in RAM of course so if your
inode gets evicted from cache and you read it in and fsync it we'll do it
the slow way if we are still in the same transaction that we last modified
the inode in.
The biggest cool part of this is that it requires no changes to the recovery
code, so if you fsync with this patch and crash and load an old kernel, it
will run the recovery and be a-ok. I have tested this pretty thoroughly
with an fsync tester and everything comes back fine, as well as xfstests.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
We didn't check error of btrfs_update_inode(), but that error looks
easy to bubble back up.
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
While we are resolving directory modifications in the
tree log, we are triggering delayed metadata updates to
the filesystem btrees.
This commit forces the delayed updates to run so the
replay code can find any modifications done. It stops
us from crashing because the directory deleltion replay
expects items to be removed immediately from the tree.
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
cc: stable@kernel.org
So dpkg fsync()'s the file and the directory containing the file whenever it
writes to a file which is really slow in btrfs. This is partly because
fsync()'ing a directory _always_ committed the transaction instead of just
going to the tree log. This is because drop_objectid_items() would return 1
since it does a btrfs_search_slot() which returns 1. In tree-log jargon
this means that we have to commit the transaction to be safe. So just check
if ret is greater than 0 and set it to 0 if it does. With this patch we now
use the tree-log instead of committing the entire transaction, which is
twice as fast on my box. Thanks,
Signed-off-by: Josef Bacik <josef@redhat.com>
We have this check down in the actual logging code, but this is after we
start a transaction and all that good stuff. So move the helper
inode_in_log() out so we can call it in fsync() and avoid starting a
transaction altogether and just exit if we've already fsync()'ed this file
recently. You would notice this issue if you fsync()'ed a file over and
over again until the transaction committed. Thanks,
Signed-off-by: Josef Bacik <josef@redhat.com>
btrfs_read_buffer() has the possibility of returning the error.
Therefore, I add the code in which the return value of btrfs_read_buffer()
is checked.
Signed-off-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
verify_parent_transid needs to lock the extent range to make
sure no IO is underway, and so it can safely clear the
uptodate bits if our checks fail.
But, a few callers are using it with spinlocks held. Most
of the time, the generation numbers are going to match, and
we don't want to switch to a blocking lock just for the error
case. This adds an atomic flag to verify_parent_transid,
and changes it to return EAGAIN if it needs to block to
properly verifiy things.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
btrfs currently handles most errors with BUG_ON. This patch is a work-in-
progress but aims to handle most errors other than internal logic
errors and ENOMEM more gracefully.
This iteration prevents most crashes but can run into lockups with
the page lock on occasion when the timing "works out."
Signed-off-by: Jeff Mahoney <jeffm@suse.com>
wait_log_commit() and wait_for_writer() were using slightly different
conditions for deciding whether they should call schedule() and whether they
should continue in the wait loop. Thus it could happen that we busylooped when
the first condition was not true while the second one was. That is burning CPU
cycles needlessly and is deadly on UP machines...
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Add a for_cow parameter to add_delayed_*_ref and pass the appropriate value
from every call site. The for_cow parameter will later on be used to
determine if a ref will change anything with respect to qgroups.
Delayed refs coming from relocation are always counted as for_cow, as they
don't change subvol quota.
Also pass in the fs_info for later use.
btrfs_find_all_roots() will use this as an optimization, as changes that are
for_cow will not change anything with respect to which root points to a
certain leaf. Thus, we don't need to add the current sequence number to
those delayed refs.
Signed-off-by: Arne Jansen <sensille@gmx.net>
Signed-off-by: Jan Schmidt <list.btrfs@jan-o-sch.net>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs: (114 commits)
Btrfs: check for a null fs root when writing to the backup root log
Btrfs: fix race during transaction joins
Btrfs: fix a potential btrfs_bio leak on scrub fixups
Btrfs: rename btrfs_bio multi -> bbio for consistency
Btrfs: stop leaking btrfs_bios on readahead
Btrfs: stop the readahead threads on failed mount
Btrfs: fix extent_buffer leak in the metadata IO error handling
Btrfs: fix the new inspection ioctls for 32 bit compat
Btrfs: fix delayed insertion reservation
Btrfs: ClearPageError during writepage and clean_tree_block
Btrfs: be smarter about committing the transaction in reserve_metadata_bytes
Btrfs: make a delayed_block_rsv for the delayed item insertion
Btrfs: add a log of past tree roots
btrfs: separate superblock items out of fs_info
Btrfs: use the global reserve when truncating the free space cache inode
Btrfs: release metadata from global reserve if we have to fallback for unlink
Btrfs: make sure to flush queued bios if write_cache_pages waits
Btrfs: fix extent pinning bugs in the tree log
Btrfs: make sure btrfs_remove_free_space doesn't leak EAGAIN
Btrfs: don't wait as long for more batches during SSD log commit
...
fs_info has now ~9kb, more than fits into one page. This will cause
mount failure when memory is too fragmented. Top space consumers are
super block structures super_copy and super_for_commit, ~2.8kb each.
Allocate them dynamically. fs_info will be ~3.5kb. (measured on x86_64)
Add a wrapper for freeing fs_info and all of it's dynamically allocated
members.
Signed-off-by: David Sterba <dsterba@suse.cz>
The tree log had two important bugs that could cause corruptions after a
crash. Sometimes we were allowing tree log blocks to be reused after
the tree log was committed but before the transaction commit was done.
This allowed a future metadata write to overwrite the tree log data. It
is fixed by adding a new variant of freeing reserved extents that always
pins them. Credit goes to Stefan Behrens and Arne Jansen for many many
hours spent tracking this bug down.
During tree log replay, we do a pass through the tree log and pin all
the extents we find. This makes sure the replay code won't go in and
use any of those blocks for new allocations during replay. The problem
is the free space cache isn't honoring these pinned extents. So the
allocator can end up handing them out, leading to all kinds of problems
during replay.
The fix here is to force any free space cache to load while we pin the
extents, and then to make sure we remove the pinned extents from the
free space rbtree.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Reported-by: Stefan Behrens <sbehrens@giantdisaster.de>
When we're doing log commits, we try to wait for more writers to come in
and make the commit bigger. This helps improve performance on rotating
disks, but on SSDs it adds latencies.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Replace remaining direct i_nlink updates with a new set_nlink()
updater function.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Tested-by: Toshiyuki Okajima <toshi.okajima@jp.fujitsu.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
When btrfs recovers from a crash, it may hit the oops below:
------------[ cut here ]------------
kernel BUG at fs/btrfs/inode.c:4580!
[...]
RIP: 0010:[<ffffffffa03df251>] [<ffffffffa03df251>] btrfs_add_link+0x161/0x1c0 [btrfs]
[...]
Call Trace:
[<ffffffffa03e7b31>] ? btrfs_inode_ref_index+0x31/0x80 [btrfs]
[<ffffffffa04054e9>] add_inode_ref+0x319/0x3f0 [btrfs]
[<ffffffffa0407087>] replay_one_buffer+0x2c7/0x390 [btrfs]
[<ffffffffa040444a>] walk_down_log_tree+0x32a/0x480 [btrfs]
[<ffffffffa0404695>] walk_log_tree+0xf5/0x240 [btrfs]
[<ffffffffa0406cc0>] btrfs_recover_log_trees+0x250/0x350 [btrfs]
[<ffffffffa0406dc0>] ? btrfs_recover_log_trees+0x350/0x350 [btrfs]
[<ffffffffa03d18b2>] open_ctree+0x1442/0x17d0 [btrfs]
[...]
This comes from that while replaying an inode ref item, we forget to
check those old conflicting DIR_ITEM and DIR_INDEX items in fs/file tree,
then we will come to conflict corners which lead to BUG_ON().
Signed-off-by: Liu Bo <liubo2009@cn.fujitsu.com>
Tested-by: Andy Lutomirski <luto@mit.edu>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
The btrfs metadata btree is the source of significant
lock contention, especially in the root node. This
commit changes our locking to use a reader/writer
lock.
The lock is built on top of rw spinlocks, and it
extends the lock tracking to remember if we have a
read lock or a write lock when we go to blocking. Atomics
count the number of blocking readers or writers at any
given time.
It removes all of the adaptive spinning from the old code
and uses only the spinning/blocking hints inside of btrfs
to decide when it should continue spinning.
In read heavy workloads this is dramatically faster. In write
heavy workloads we're still faster because of less contention
on the root node lock.
We suffer slightly in dbench because we schedule more often
during write locks, but all other benchmarks so far are improved.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
The two ->process_func call sites in tree-log.c which were ignoring a return
code have also been updated to gracefully exit as well.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
If return value of btrfs_inc_extent_ref() is not 0, BUG() is called.
Signed-off-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
When read_one_inode() fails, error code is returned to caller instead
of BUG_ON().
Signed-off-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Currently, btrfs_truncate_item and btrfs_extend_item returns only 0.
So, the check by BUG_ON in the caller is unnecessary.
Signed-off-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
The error code is returned instead of calling BUG_ON when
btrfs_del_item returns the error.
Signed-off-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
The current code relogs the entire inode every time during fsync log,
and it is much better suited to small files rather than large ones.
During my performance test, the fsync performace of large files sucks,
and we can ascribe this to the tremendous amount of csum infos of the
large ones, cause we have to flush all of these csum infos into log trees
even when there are only _one_ change in the whole file data. Apparently,
to optimize fsync, we need to create a filter to skip the unnecessary csum
ones, that is, the corresponding file data remains unchanged before this fsync.
Here I have some test results to show, I use sysbench to do "random write + fsync".
===
sysbench --test=fileio --num-threads=1 --file-num=2 --file-block-size=4K --file-total-size=8G --file-test-mode=rndwr --file-io-mode=sync --file-extra-flags= [prepare, run]
===
Sysbench args:
- Number of threads: 1
- Extra file open flags: 0
- 2 files, 4Gb each
- Block size 4Kb
- Number of random requests for random IO: 10000
- Read/Write ratio for combined random IO test: 1.50
- Periodic FSYNC enabled, calling fsync() each 100 requests.
- Calling fsync() at the end of test, Enabled.
- Using synchronous I/O mode
- Doing random write test
Sysbench results:
===
Operations performed: 0 Read, 10000 Write, 200 Other = 10200 Total
Read 0b Written 39.062Mb Total transferred 39.062Mb
===
a) without patch: (*SPEED* : 451.01Kb/sec)
112.75 Requests/sec executed
b) with patch: (*SPEED* : 4.7533Mb/sec)
1216.84 Requests/sec executed
PS: I've made a _sub transid_ stuff patch, but it does not perform as effectively as this patch,
and I'm wanderring where the problem is and trying to improve it more.
Signed-off-by: Liu Bo <liubo2009@cn.fujitsu.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Changelog V5 -> V6:
- Fix oom when the memory load is high, by storing the delayed nodes into the
root's radix tree, and letting btrfs inodes go.
Changelog V4 -> V5:
- Fix the race on adding the delayed node to the inode, which is spotted by
Chris Mason.
- Merge Chris Mason's incremental patch into this patch.
- Fix deadlock between readdir() and memory fault, which is reported by
Itaru Kitayama.
Changelog V3 -> V4:
- Fix nested lock, which is reported by Itaru Kitayama, by updating space cache
inode in time.
Changelog V2 -> V3:
- Fix the race between the delayed worker and the task which does delayed items
balance, which is reported by Tsutomu Itoh.
- Modify the patch address David Sterba's comment.
- Fix the bug of the cpu recursion spinlock, reported by Chris Mason
Changelog V1 -> V2:
- break up the global rb-tree, use a list to manage the delayed nodes,
which is created for every directory and file, and used to manage the
delayed directory name index items and the delayed inode item.
- introduce a worker to deal with the delayed nodes.
Compare with Ext3/4, the performance of file creation and deletion on btrfs
is very poor. the reason is that btrfs must do a lot of b+ tree insertions,
such as inode item, directory name item, directory name index and so on.
If we can do some delayed b+ tree insertion or deletion, we can improve the
performance, so we made this patch which implemented delayed directory name
index insertion/deletion and delayed inode update.
Implementation:
- introduce a delayed root object into the filesystem, that use two lists to
manage the delayed nodes which are created for every file/directory.
One is used to manage all the delayed nodes that have delayed items. And the
other is used to manage the delayed nodes which is waiting to be dealt with
by the work thread.
- Every delayed node has two rb-tree, one is used to manage the directory name
index which is going to be inserted into b+ tree, and the other is used to
manage the directory name index which is going to be deleted from b+ tree.
- introduce a worker to deal with the delayed operation. This worker is used
to deal with the works of the delayed directory name index items insertion
and deletion and the delayed inode update.
When the delayed items is beyond the lower limit, we create works for some
delayed nodes and insert them into the work queue of the worker, and then
go back.
When the delayed items is beyond the upper bound, we create works for all
the delayed nodes that haven't been dealt with, and insert them into the work
queue of the worker, and then wait for that the untreated items is below some
threshold value.
- When we want to insert a directory name index into b+ tree, we just add the
information into the delayed inserting rb-tree.
And then we check the number of the delayed items and do delayed items
balance. (The balance policy is above.)
- When we want to delete a directory name index from the b+ tree, we search it
in the inserting rb-tree at first. If we look it up, just drop it. If not,
add the key of it into the delayed deleting rb-tree.
Similar to the delayed inserting rb-tree, we also check the number of the
delayed items and do delayed items balance.
(The same to inserting manipulation)
- When we want to update the metadata of some inode, we cached the data of the
inode into the delayed node. the worker will flush it into the b+ tree after
dealing with the delayed insertion and deletion.
- We will move the delayed node to the tail of the list after we access the
delayed node, By this way, we can cache more delayed items and merge more
inode updates.
- If we want to commit transaction, we will deal with all the delayed node.
- the delayed node will be freed when we free the btrfs inode.
- Before we log the inode items, we commit all the directory name index items
and the delayed inode update.
I did a quick test by the benchmark tool[1] and found we can improve the
performance of file creation by ~15%, and file deletion by ~20%.
Before applying this patch:
Create files:
Total files: 50000
Total time: 1.096108
Average time: 0.000022
Delete files:
Total files: 50000
Total time: 1.510403
Average time: 0.000030
After applying this patch:
Create files:
Total files: 50000
Total time: 0.932899
Average time: 0.000019
Delete files:
Total files: 50000
Total time: 1.215732
Average time: 0.000024
[1] http://marc.info/?l=linux-btrfs&m=128212635122920&q=p3
Many thanks for Kitayama-san's help!
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Reviewed-by: David Sterba <dave@jikos.cz>
Tested-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Tested-by: Itaru Kitayama <kitayama@cl.bb4u.ne.jp>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
This adds an initial implementation for scrub. It works quite
straightforward. The usermode issues an ioctl for each device in the
fs. For each device, it enumerates the allocated device chunks. For
each chunk, the contained extents are enumerated and the data checksums
fetched. The extents are read sequentially and the checksums verified.
If an error occurs (checksum or EIO), a good copy is searched for. If
one is found, the bad copy will be rewritten.
All enumerations happen from the commit roots. During a transaction
commit, the scrubs get paused and afterwards continue from the new
roots.
This commit is based on the series originally posted to linux-btrfs
with some improvements that resulted from comments from David Sterba,
Ilya Dryomov and Jan Schmidt.
Signed-off-by: Arne Jansen <sensille@gmx.net>
parameter tree root it's not used since commit
5f39d397df ("Btrfs: Create extent_buffer
interface for large blocksizes")
Signed-off-by: David Sterba <dsterba@suse.cz>
It is necessary to unlock mutex_lock before it return an error when
btrfs_alloc_path() fails.
Signed-off-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
There's a potential problem in 32bit system when we exhaust 32bit inode
numbers and start to allocate big inode numbers, because btrfs uses
inode->i_ino in many places.
So here we always use BTRFS_I(inode)->location.objectid, which is an
u64 variable.
There are 2 exceptions that BTRFS_I(inode)->location.objectid !=
inode->i_ino: the btree inode (0 vs 1) and empty subvol dirs (256 vs 2),
and inode->i_ino will be used in those cases.
Another reason to make this change is I'm going to use a special inode
to save free ino cache, and the inode number must be > (u64)-256.
Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
When we recover from crash via write-ahead log tree and process
the inode refs, for each btrfs_inode_ref item, we will
1) check if we already have a perfect match in fs/file tree, if
we have, then we're done.
2) search the corresponding back reference in fs/file tree, and
check all the names in this back reference to see if they are
also in the log to avoid conflict corners.
3) recover the logged inode refs to fs/file tree.
In current btrfs, however,
- for 2)'s check, once is enough, since the checked back reference
will remain unchanged after processing all the inode refs belonged
to the key.
- it has no need to do another 1) between 2) and 3).
I've made a small test to show how it improves,
$dd if=/dev/zero of=foobar bs=4K count=1
$sync
$make 100 hard links continuously, like ln foobar link_i
$fsync foobar
$echo b > /proc/sysrq-trigger
after reboot
$time mount DEV PATH
without patch:
real 0m0.285s
user 0m0.001s
sys 0m0.009s
with patch:
real 0m0.123s
user 0m0.000s
sys 0m0.010s
Changelog v1->v2:
- fix double free - pointed by David Sterba
Changelog v2->v3:
- adjust free order
Signed-off-by: Liu Bo <liubo2009@cn.fujitsu.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
This patch changes some BUG_ON() to the error return.
(but, most callers still use BUG_ON())
Signed-off-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
We need to make sure the dir items we get are valid dir items. So any time we
try and read one check it with verify_dir_item, which will do various sanity
checks to make sure it looks sane. Thanks,
Signed-off-by: Josef Bacik <josef@redhat.com>
