When doing concurrent lseek(2) system calls against the same file
descriptor, using multiple threads belonging to the same process, we have
a short time window where a race happens and can result in a memory leak.
The race happens like this:
1) A program opens a file descriptor for a file and then spawns two
threads (with the pthreads library for example), lets call them
task A and task B;
2) Task A calls lseek with SEEK_DATA or SEEK_HOLE and ends up at
file.c:find_desired_extent() while holding a read lock on the inode;
3) At the start of find_desired_extent(), it extracts the file's
private_data pointer into a local variable named 'private', which has
a value of NULL;
4) Task B also calls lseek with SEEK_DATA or SEEK_HOLE, locks the inode
in shared mode and enters file.c:find_desired_extent(), where it also
extracts file->private_data into its local variable 'private', which
has a NULL value;
5) Because it saw a NULL file private, task A allocates a private
structure and assigns to the file structure;
6) Task B also saw a NULL file private so it also allocates its own file
private and then assigns it to the same file structure, since both
tasks are using the same file descriptor.
At this point we leak the private structure allocated by task A.
Besides the memory leak, there's also the detail that both tasks end up
using the same cached state record in the private structure (struct
btrfs_file_private::llseek_cached_state), which can result in a
use-after-free problem since one task can free it while the other is
still using it (only one task took a reference count on it). Also, sharing
the cached state is not a good idea since it could result in incorrect
results in the future - right now it should not be a problem because it
end ups being used only in extent-io-tree.c:count_range_bits() where we do
range validation before using the cached state.
Fix this by protecting the private assignment and check of a file while
holding the inode's spinlock and keep track of the task that allocated
the private, so that it's used only by that task in order to prevent
user-after-free issues with the cached state record as well as potentially
using it incorrectly in the future.
Fixes: 3c32c7212f ("btrfs: use cached state when looking for delalloc ranges with lseek")
CC: stable@vger.kernel.org # 6.6+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Currently BTRFS_I is a static inline function that takes a const inode
and returns btrfs inode, dropping the 'const' qualifier. This can break
assumptions of compiler though it seems there's no real case.
To make the parameter and return type consistent regardint const we can
use the container_of_const() that preserves it. However this would not
check the parameter type. To fix that use the same _Generic construct
but implement only the two expected types.
Signed-off-by: David Sterba <dsterba@suse.com>
We have a few places that check if we have the inode locked by doing:
ASSERT(inode_is_locked(vfs_inode));
This actually proved to be useful several times as if assertions are
enabled (and by default they are in many distros) it immediately triggers
a crash which is impossible for users to miss.
However that doesn't check if the lock is held by the calling task, so
the check passes if some other task locked the inode.
Using one of the lockdep functions to check the lock is held, like
lockdep_assert_held() for example, does check that the calling task
holds the lock, and if that's not the case it produces a warning and
stack trace in dmesg. However, despite the misleading "assert" in the
name of the lockdep helpers, it does not trigger a crash/BUG_ON(), just
a warning and splat in dmesg, which is easy to get unnoticed by users
who may have lockdep enabled.
So add a helper that does the ASSERT() and calls lockdep_assert_held()
immediately after and use it every where we check the inode is locked.
Like this if the lock is held by some other task we get the warning
in dmesg which is caught by fstests, very helpful during development,
and may also be occassionaly noticed by users with lockdep enabled.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We only pass this into read_inline_extent, change it to take a folio and
update the callers.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Instead of a page, use a folio for btrfs_writepage_cow_fixup. We
already have a folio at the only caller, and the fixup worker uses
folios.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Now that every function that btrfs_run_delalloc_range calls takes a
folio, update it to take a folio and update the callers.
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The direct IO code is over a thousand lines and it's currently spread
between file.c and inode.c, which makes it not easy to locate some parts
of it sometimes. Also inode.c is about 11 thousand lines and file.c about
4 thousand lines, both too big. So move all the direct IO code into a
dedicated file, so that it's easy to locate all its code and reduce the
sizes of inode.c and file.c.
This is a pure move of code without any other changes except export a
a couple functions from inode.c (get_extent_allocation_hint() and
create_io_em()) because they are used in inode.c and the new direct-io.c
file, and a couple functions from file.c (btrfs_buffered_write() and
btrfs_write_check()) because they are used both in file.c and in the new
direct-io.c file.
Reviewed-by: Boris Burkov <boris@bur.io>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Pass a struct btrfs_inode to is_data_inode() as it's an
internal interface, allowing to remove some use of BTRFS_I.
Reviewed-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>
It's pointless to pass a super block argument to btrfs_iget_path() because
we always pass a root and from it we can get the super block through:
root->fs_info->sb
So remove the super block argument.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
It's pointless to pass a super block argument to btrfs_iget() because we
always pass a root and from it we can get the super block through:
root->fs_info->sb
So remove the super block argument.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We can add const to many parameters, this is for clarity and minor
addition to safety. There are some minor effects, in the assembly
code and .ko measured on release config. This patch does not cover all
possible conversions.
Signed-off-by: David Sterba <dsterba@suse.com>
We have several headers that are including themselves, triggering clangd
warnings.
Such includes are caused by commit 602035d7fe ("btrfs: add forward
declarations and headers, part 2").
Just remove such unnecessary include.
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
All parameters after @filepos of btrfs_alloc_ordered_extent() can be
replaced with btrfs_file_extent structure.
This patch does the cleanup, meanwhile some points to note:
- Move btrfs_file_extent structure to ordered-data.h
The structure is needed by both btrfs_alloc_ordered_extent() and
can_nocow_extent(), but since btrfs_inode.h includes
ordered-data.h, so we need to move the structure to ordered-data.h.
- Move the special handling of NOCOW/PREALLOC into
btrfs_alloc_ordered_extent()
This is to allow btrfs_split_ordered_extent() to properly split them
for DIO.
For now just move the handling into btrfs_alloc_ordered_extent() to
simplify the callers.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The following functions and structures can be simplified using the
btrfs_file_extent structure:
- can_nocow_extent()
No need to return ram_bytes/orig_block_len through the parameter list,
the @file_extent parameter contains all the needed info.
- can_nocow_file_extent_args
The following members are no longer needed:
* disk_bytenr
This one is confusing as it's not really the
btrfs_file_extent_item::disk_bytenr, but where the IO would be,
thus it's file_extent::disk_bytenr + file_extent::offset now.
* num_bytes
Now file_extent::num_bytes.
* extent_offset
Now file_extent::offset.
* disk_num_bytes
Now file_extent::disk_num_bytes.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Since we have extent_map::offset, the old extent_map::orig_start is just
extent_map::start - extent_map::offset for non-hole/inline extents.
And since the new extent_map::offset is already verified by
validate_extent_map() while the old orig_start is not, let's just remove
the old member from all call sites.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Currently function can_nocow_extent() only returns members needed for
extent_map.
However since we will soon change the extent_map structure to be more
like btrfs_file_extent_item, we want to expose the expected file extent
caused by the NOCOW write for future usage.
This introduces a new structure, btrfs_file_extent, to be a more
memory access friendly representation of btrfs_file_extent_item.
And use that structure to expose the expected file extent caused by the
NOCOW write.
For now there is no user of the new structure yet.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
On 64 bits platforms we don't really need to have a dedicated member (the
objectid field) for the inode's number since we store in the VFS inode's
i_ino member, which is an unsigned long and this type is 64 bits wide on
64 bits platforms. We only need that field in case we are on a 32 bits
platform because the unsigned long type is 32 bits wide on such platforms
See commit 33345d0152 ("Btrfs: Always use 64bit inode number") regarding
this 64/32 bits detail.
The objectid field of struct btrfs_inode is also used to store the ID of
a root for directories that are stubs for unreferenced roots. In such
cases the inode is a directory and has the BTRFS_INODE_ROOT_STUB runtime
flag set.
So in order to reduce the size of btrfs_inode structure on 64 bits
platforms we can remove the objectid member and use the VFS inode's i_ino
member instead whenever we need to get the inode number. In case the inode
is a root stub (BTRFS_INODE_ROOT_STUB set) we can use the member
last_reflink_trans to store the ID of the unreferenced root, since such
inode is a directory and reflinks can't be done against directories.
So remove the objectid fields for 64 bits platforms and alias the
last_reflink_trans field with a name of ref_root_id in a union.
On a release kernel config, this reduces the size of struct btrfs_inode
from 1040 bytes down to 1032 bytes.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Currently struct btrfs_inode has a key member, named "location", that is
either:
1) The key of the inode's item. In this case the objectid is the number
of the inode;
2) A key stored in a dir entry with a type of BTRFS_ROOT_ITEM_KEY, for
the case where we have a root that is a snapshot of a subvolume that
points to other subvolumes. In this case the objectid is the ID of
a subvolume inside the snapshotted parent subvolume.
The key is only used to lookup the inode item for the first case, while
for the second it's never used since it corresponds to directory stubs
created with new_simple_dir() and which are marked as dummy, so there's
no actual inode item to ever update. In the second case we only check
the key type at btrfs_ino() for 32 bits platforms and its objectid is
only needed for unlink.
Instead of using a key we can do fine with just the objectid, since we
can generate the key whenever we need it having only the objectid, as
in all use cases the type is always BTRFS_INODE_ITEM_KEY and the offset
is always 0.
So use only an objectid instead of a full key. This reduces the size of
struct btrfs_inode from 1048 bytes down to 1040 bytes on a release kernel.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The index_cnt field of struct btrfs_inode is used only for two purposes:
1) To store the index for the next entry added to a directory;
2) For the data relocation inode to track the logical start address of the
block group currently being relocated.
For the relocation case we use index_cnt because it's not used for
anything else in the relocation use case - we could have used other fields
that are not used by relocation such as defrag_bytes, last_unlink_trans
or last_reflink_trans for example (among others).
Since the csum_bytes field is not used for directories, do the following
changes:
1) Put index_cnt and csum_bytes in a union, and index_cnt is only
initialized when the inode is a directory. The csum_bytes is only
accessed in IO paths for regular files, so we're fine here;
2) Use the defrag_bytes field for relocation, since the data relocation
inode is never used for defrag purposes. And to make the naming better,
alias it to reloc_block_group_start by using a union.
This reduces the size of struct btrfs_inode by 8 bytes in a release
kernel, from 1056 bytes down to 1048 bytes.
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Currently we use a red black tree (rb-tree) to track the currently open
inodes of a root (in struct btrfs_root::inode_tree). This however is not
very efficient when the number of inodes is large since rb-trees are
binary trees. For example for 100K open inodes, the tree has a depth of
17. Besides that, inserting into the tree requires navigating through it
and pulling useless cache lines in the process since the red black tree
nodes are embedded within the btrfs inode - on the other hand, by being
embedded, it requires no extra memory allocations.
We can improve this by using an xarray instead, which is efficient when
indices are densely clustered (such as inode numbers), is more cache
friendly and behaves like a resizable array, with a much better search
and insertion complexity than a red black tree. This only has one small
disadvantage which is that insertion will sometimes require allocating
memory for the xarray - which may fail (not that often since it uses a
kmem_cache) - but on the other hand we can reduce the btrfs inode
structure size by 24 bytes (from 1080 down to 1056 bytes) after removing
the embedded red black tree node, which after the next patches will allow
to reduce the size of the structure to 1024 bytes, meaning we will be able
to store 4 inodes per 4K page instead of 3 inodes.
This change does a straightforward change to use an xarray, and results
in a transaction abort if we can't allocate memory for the xarray when
creating an inode - but the next patch changes things so that we don't
need to abort.
Running the following fs_mark test showed some improvements:
$ cat test.sh
#!/bin/bash
DEV=/dev/nullb0
MNT=/mnt/nullb0
MOUNT_OPTIONS="-o ssd"
FILES=100000
THREADS=$(nproc --all)
echo "performance" | \
tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
mkfs.btrfs -f $DEV
mount $MOUNT_OPTIONS $DEV $MNT
OPTS="-S 0 -L 5 -n $FILES -s 0 -t $THREADS -k"
for ((i = 1; i <= $THREADS; i++)); do
OPTS="$OPTS -d $MNT/d$i"
done
fs_mark $OPTS
umount $MNT
Before this patch:
FSUse% Count Size Files/sec App Overhead
10 1200000 0 92081.6 12505547
16 2400000 0 138222.6 13067072
23 3600000 0 148833.1 13290336
43 4800000 0 97864.7 13931248
53 6000000 0 85597.3 14384313
After this patch:
FSUse% Count Size Files/sec App Overhead
10 1200000 0 93225.1 12571078
16 2400000 0 146720.3 12805007
23 3600000 0 160626.4 13073835
46 4800000 0 116286.2 13802927
53 6000000 0 90087.9 14754892
The test was run with a release kernel config (Debian's default config).
Also capturing the insertion times into the rb tree and into the xarray,
that is measuring the duration of the old function inode_tree_add() and
the duration of the new btrfs_add_inode_to_root() function, gave the
following results (in nanoseconds):
Before this patch, inode_tree_add() execution times:
Count: 5000000
Range: 0.000 - 5536887.000; Mean: 775.674; Median: 729.000; Stddev: 4820.961
Percentiles: 90th: 1015.000; 95th: 1139.000; 99th: 1397.000
0.000 - 7.816: 40 |
7.816 - 37.858: 209 |
37.858 - 170.278: 6059 |
170.278 - 753.961: 2754890 #####################################################
753.961 - 3326.728: 2232312 ###########################################
3326.728 - 14667.018: 4366 |
14667.018 - 64652.943: 852 |
64652.943 - 284981.761: 550 |
284981.761 - 1256150.914: 221 |
1256150.914 - 5536887.000: 7 |
After this patch, btrfs_add_inode_to_root() execution times:
Count: 5000000
Range: 0.000 - 2900652.000; Mean: 272.148; Median: 241.000; Stddev: 2873.369
Percentiles: 90th: 342.000; 95th: 432.000; 99th: 572.000
0.000 - 7.264: 104 |
7.264 - 33.145: 352 |
33.145 - 140.081: 109606 #
140.081 - 581.930: 4840090 #####################################################
581.930 - 2407.590: 43532 |
2407.590 - 9950.979: 2245 |
9950.979 - 41119.278: 514 |
41119.278 - 169902.616: 155 |
169902.616 - 702018.539: 47 |
702018.539 - 2900652.000: 9 |
Average, percentiles, standard deviation, etc, are all much better.
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
If a write path in COW mode fails, either before submitting a bio for the
new extents or an actual IO error happens, we can end up allowing a fast
fsync to log file extent items that point to unwritten extents.
This is because dropping the extent maps happens when completing ordered
extents, at btrfs_finish_one_ordered(), and the completion of an ordered
extent is executed in a work queue.
This can result in a fast fsync to start logging file extent items based
on existing extent maps before the ordered extents complete, therefore
resulting in a log that has file extent items that point to unwritten
extents, resulting in a corrupt file if a crash happens after and the log
tree is replayed the next time the fs is mounted.
This can happen for both direct IO writes and buffered writes.
For example consider a direct IO write, in COW mode, that fails at
btrfs_dio_submit_io() because btrfs_extract_ordered_extent() returned an
error:
1) We call btrfs_finish_ordered_extent() with the 'uptodate' parameter
set to false, meaning an error happened;
2) That results in marking the ordered extent with the BTRFS_ORDERED_IOERR
flag;
3) btrfs_finish_ordered_extent() queues the completion of the ordered
extent - so that btrfs_finish_one_ordered() will be executed later in
a work queue. That function will drop extent maps in the range when
it's executed, since the extent maps point to unwritten locations
(signaled by the BTRFS_ORDERED_IOERR flag);
4) After calling btrfs_finish_ordered_extent() we keep going down the
write path and unlock the inode;
5) After that a fast fsync starts and locks the inode;
6) Before the work queue executes btrfs_finish_one_ordered(), the fsync
task sees the extent maps that point to the unwritten locations and
logs file extent items based on them - it does not know they are
unwritten, and the fast fsync path does not wait for ordered extents
to complete, which is an intentional behaviour in order to reduce
latency.
For the buffered write case, here's one example:
1) A fast fsync begins, and it starts by flushing delalloc and waiting for
the writeback to complete by calling filemap_fdatawait_range();
2) Flushing the dellaloc created a new extent map X;
3) During the writeback some IO error happened, and at the end io callback
(end_bbio_data_write()) we call btrfs_finish_ordered_extent(), which
sets the BTRFS_ORDERED_IOERR flag in the ordered extent and queues its
completion;
4) After queuing the ordered extent completion, the end io callback clears
the writeback flag from all pages (or folios), and from that moment the
fast fsync can proceed;
5) The fast fsync proceeds sees extent map X and logs a file extent item
based on extent map X, resulting in a log that points to an unwritten
data extent - because the ordered extent completion hasn't run yet, it
happens only after the logging.
To fix this make btrfs_finish_ordered_extent() set the inode flag
BTRFS_INODE_NEEDS_FULL_SYNC in case an error happened for a COW write,
so that a fast fsync will wait for ordered extent completion.
Note that this issues of using extent maps that point to unwritten
locations can not happen for reads, because in read paths we start by
locking the extent range and wait for any ordered extents in the range
to complete before looking for extent maps.
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Nowadays we have a lock used to synchronize mmap writes with reflink and
fsync operations (struct btrfs_inode::i_mmap_lock), so update the comment
for btrfs_set_inode_full_sync() to mention that it can also be called
while holding that mmap lock. Besides being a valid alternative to the
inode's VFS lock, we already have the extent map shrinker using that mmap
lock instead.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Export the relocation private helper find_next_inode() to inode.c, as this
same logic is also used at btrfs_prune_dentries() and will be used by an
upcoming change that adds an extent map shrinker. The next patch will
change btrfs_prune_dentries() to use this helper.
Reviewed-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
btrfs_page_mkwrite() is a struct vm_operations_struct callback and we
define that structure in file.c. Currently the function is in inode.c and
has to be exported to be used in file.c, which makes no sense because it's
not used anywhere else. So move btrfs_page_mkwrite() from inode.c and into
file.c.
While at it do a few minor style changes:
1) Capitalize the first word of every comment and end each sentence with
punctuation;
2) Avoid splitting some statements into two lines when everything fits in
85 characters or less.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The helpers btrfs_del_delalloc_inode() and __btrfs_del_delalloc_inode()
don't follow the pattern when the "__" helper does a special case and
are in fact reversed regarding the naming. We can merge them into one as
there's only one place that needs to be open coded.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When setting and clearing a delalloc range, at btrfs_set_delalloc_extent()
and btrfs_clear_delalloc_extent(), we are adding/removing the inode
to/from the root's list of delalloc inodes while under the protection of
the inode's lock. This however is not needed, we can add and remove the
inode to the root's list without holding the inode's lock because here
we are under the protection of the io tree's lock, reducing the size of
the critical section delimited by the inode's lock. The inode's lock is
used in many other places such as when finishing an ordered extent (when
calling btrfs_update_inode_bytes() or btrfs_delalloc_release_metadata(),
or decreasing the number of outstanding extents) or when reserving space
when doing a buffered or direct IO write (calls to functions from
delalloc-space.c).
So move the inode add/remove operations to the root's list of delalloc
inodes to outside the critical section delimited by the inode's lock.
This also allows us to get rid of the BTRFS_INODE_IN_DELALLOC_LIST flag
since we can rely on the inode's delalloc bytes counter to determine if
the inode is or is not in the list.
The following fio based test, that exercises IO to multiple files in the
same subvolume, was used to test:
$ cat test.sh
#!/bin/bash
DEV=/dev/nullb0
MNT=/mnt/nullb0
MOUNT_OPTIONS="-o ssd"
mkfs.btrfs -f $DEV &> /dev/null
mount $MOUNT_OPTIONS $DEV $MNT
fio --direct=0 --ioengine=sync --thread --directory=$MNT \
--invalidate=1 --group_reporting=1 \
--new_group --rw=randwrite --size=50m --numjobs=200 \
--bs=4k --fsync_on_close=0 --fallocate=none --end_fsync=0 \
--name=foo --filename_format=FioWorkloads.\$jobnum
umount $MNT
The test was run on a non-debug kernel (Debian's default kernel config)
against a 16G null block device.
Result before this patch:
WRITE: bw=81.9MiB/s (85.9MB/s), 81.9MiB/s-81.9MiB/s (85.9MB/s-85.9MB/s), io=9.77GiB (10.5GB), run=122136-122136msec
Result after this patch:
WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=9.77GiB (10.5GB), run=115180-115180msec
Reviewed-by: Boris Burkov <boris@bur.io>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
There's no need to pass a root argument to __btrfs_del_delalloc_inode()
and btrfs_del_delalloc_inode(), we can just pass the inode since the root
is always the root associated to that inode. Some remove the root argument
from these functions.
Reviewed-by: Boris Burkov <boris@bur.io>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Do a cleanup in the rest of the headers:
- add forward declarations for types referenced by pointers
- add includes when types need them
This fixes potential compilation problems if the headers are reordered
or the missing includes are not provided indirectly.
Signed-off-by: David Sterba <dsterba@suse.com>
The parameter @pg_offset of btrfs_get_extent() is only utilized for
inlined extent, and we already have an ASSERT() and tree-checker, to
make sure we can only get inline extent at file offset 0.
Any invalid inline extent with non-zero file offset would be rejected by
tree-checker in the first place.
Thus the @pg_offset parameter is not really necessary, just remove it.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The definition for btrfs_get_extent() is using "u64 end" as the last
parameter, but in implementation we go "u64 len", and all call sites
follows the implementation.
This can be very confusing during development, as most developers
including me, would just use the snippet returned by LSP (clangd in my
case), which would only check the definition.
Unfortunately this mismatch is introduced from the very beginning of
btrfs.
Fix it to prevent further confusion.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The file_extent_tree was added in 41a2ee75aa ("btrfs: introduce
per-inode file extent tree") so we have an explicit mapping of the file
extents to know where it is safe to update i_size. When the feature
NO_HOLES is enabled, and it's been a mkfs default since 5.15, the tree
is not necessary.
To save some space in the inode, allocate the tree only when necessary.
This reduces size by 16 bytes from 1096 to 1080 on a x86_64 release
config.
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When profiling a workload I noticed we were constantly calling getxattr.
These were mostly coming from __remove_privs, which will lookup if
security.capability exists to remove it. However instrumenting getxattr
showed we get called nearly constantly on an idle machine on a lot of
accesses.
These are wasteful and not free. Other security LSMs have a way to
cache their results, but capability doesn't have this, so it's asking us
all the time for the xattr.
Fix this by setting a flag in our inode that it doesn't have a
security.capability xattr. We set this on new inodes and after a failed
lookup of security.capability. If we set this xattr at all we'll clear
the flag.
I haven't found a test in fsperf that this makes a visible difference
on, but I assume fs_mark related tests would show it clearly. This is a
perf report output of the smallfiles100k run where it shows 20% of our
time spent in __remove_privs because we're looking up the non-existent
xattr.
--21.86%--btrfs_write_check.constprop.0
--21.62%--__file_remove_privs
--21.55%--security_inode_need_killpriv
--21.54%--cap_inode_need_killpriv
--21.53%--__vfs_getxattr
--20.89%--btrfs_getxattr
Obviously this is just CPU time in a mostly IO bound test, so the actual
effect of removing this callchain is minimal. However in just normal
testing of an idle system tracing showed around 100 getxattr calls per
minute, and with this patch there are 0.
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The type of timespec64::tv_nsec is 'unsigned long', while we have only
u32 for on-disk and in-memory. This wastes a few bytes in btrfs_inode.
Add separate members for sec and nsec with the corresponding type width.
This creates a 4 byte hole in btrfs_inode which can be utilized in the
future.
Signed-off-by: David Sterba <dsterba@suse.com>
Update the comment for the lock named "lock" in struct btrfs_inode because
it does not mention that the fields "delalloc_bytes", "defrag_bytes",
"csum_bytes", "outstanding_extents" and "disk_i_size" are also protected
by that lock.
Also add a comment on top of each field protected by this lock to mention
that the lock protects them.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Currently, the last_log_commit of a root can be accessed concurrently
without any lock protection. Readers can be calling btrfs_inode_in_log()
early in a fsync call, which reads a root's last_log_commit, while a
writer can change the last_log_commit while a log tree if being synced,
at btrfs_sync_log(). Any races here should be harmless, and in the worst
case they may cause a fsync to log an inode when it's not really needed,
so nothing bad from a functional perspective.
To avoid data race warnings from tools like KCSAN and other issues such
as load and store tearing (amongst others, see [1]), create helpers to
access the last_log_commit field of a root using READ_ONCE() and
WRITE_ONCE(), and use these helpers everywhere.
[1] https://lwn.net/Articles/793253/
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Previous commit created a hole in struct btrfs_inode, we can move
outstanding_extents there. This reduces size by 8 bytes from 1120 to
1112 on a release config.
Signed-off-by: David Sterba <dsterba@suse.com>
The structure btrfs_ordered_inode_tree is used only in one place, in
btrfs_inode. The structure itself has a 4 byte hole which is wasted
space.
Move the btrfs_ordered_inode_tree members to btrfs_inode with a common
prefix 'ordered_tree_' where the hole can be utilized and shrink inode
size.
Signed-off-by: David Sterba <dsterba@suse.com>
The root argument for btrfs_update_inode() always matches the root of the
given inode, so remove the root argument and get it from the inode
argument.
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The root argument for btrfs_update_inode_fallback() always matches the
root of the given inode, so remove the root argument and get it from the
inode argument.
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Currently the compression type values are bounded and fit to an u8, we
can pack the btrfs_inode a bit by reordering them to the space created
by the location key. This reduces size from 1112 to 1104.
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
If we no longer include the tracepoints from ctree.h we fail to compile
because we have the dependency in some of the header files and source
files. Add the include where we have these dependencies to allow us to
remove the include from ctree.h.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
These headers have struct fscrypt_str as function arguments, so add
struct fscrypt_str to the theader, and include linux/fscrypt.h in
btrfs_inode.h as it also needs the definition of struct fscrypt_name for
the new inode args.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Instead of a separate page_started argument that tells the callers that
btrfs_run_delalloc_range already started writeback by itself, overload
the return value with a positive 1 in additio to 0 and a negative error
code to indicate that is has already started writeback, and remove the
nr_written argument as that caller can calculate it directly based on
the range, and in fact already does so for the case where writeback
wasn't started yet.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
btrfs_writepage_endio_finish_ordered is a small wrapper around
btrfs_mark_ordered_io_finished that just changs the argument passing
slightly, and adds a tracepoint.
Move the tracpoint to btrfs_mark_ordered_io_finished, which means
it now also covers the error handling in btrfs_cleanup_ordered_extent
and switch all callers to just call btrfs_mark_ordered_io_finished
directly.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The btrfs_inode_mod_outstanding_extents trace event only shows the modified
number to the number of outstanding extents. It would be helpful if we can
see the resulting extent number as well.
Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The btrfs zoned completion code currently needs an ordered_extent and
extent_map per bio so that it can account for the non-predictable
write location from Zone Append. To archive that it currently splits
the ordered_extent and extent_map at I/O submission time, and then
records the actual physical address in the ->physical field of the
ordered_extent.
This patch instead switches to record the "original" physical address
that the btrfs allocator assigned in spare space in the btrfs_bio,
and then rewrites the logical address in the btrfs_ordered_sum
structure at I/O completion time. This allows the ordered extent
completion handler to simply walk the list of ordered csums and
split the ordered extent as needed. This removes an extra ordered
extent and extent_map lookup and manipulation during the I/O
submission path, and instead batches it in the I/O completion path
where we need to touch these anyway.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: David Sterba <dsterba@suse.com>
These are more related to the inode item flags on disk than the
in-memory btrfs_inode, move the helpers to inode-item.h.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The writeback_control structure already passes down the information about
a writeback being synchronous from the core VM code, and thus information
is propagated into the bio REQ_SYNC flag through the wbc_to_write_flags
helper.
Use that information to decide if checksums calculation is offloaded to
a workqueue instead of btrfs_inode::sync_writers field that not only
bloats the inode but also has too wide scope, being inode wide instead
of limited to the actual writeback request.
The sync writes were set in:
- btrfs_do_write_iter - regular IO, sync status is set
- start_ordered_ops - ordered write start, writeback with WB_SYNC_ALL
mode
- btrfs_write_marked_extents - write marked extents, writeback with
WB_SYNC_ALL mode
Reviewed-by: Chris Mason <clm@fb.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: David Sterba <dsterba@suse.com>
[ update changelog ]
Signed-off-by: David Sterba <dsterba@suse.com>
When logging a directory, after copying all directory index items from the
subvolume tree to the log tree, we iterate over the subvolume tree to find
all dir index items that are located in leaves COWed (or created) in the
current transaction. If we keep logging a directory several times during
the same transaction, we end up iterating over the same dir index items
everytime we log the directory, wasting time and adding extra lock
contention on the subvolume tree.
So just keep track of the last logged dir index offset in order to start
the search for that index (+1) the next time the directory is logged, as
dir index values (key offsets) come from a monotonically increasing
counter.
The following test measures the difference before and after this change:
$ cat test.sh
#!/bin/bash
DEV=/dev/nullb0
MNT=/mnt/nullb0
umount $DEV &> /dev/null
mkfs.btrfs -f $DEV
mount -o ssd $DEV $MNT
# Time values in milliseconds.
declare -a fsync_times
# Total number of files added to the test directory.
num_files=1000000
# Fsync directory after every N files are added.
fsync_period=100
mkdir $MNT/testdir
fsync_total_time=0
for ((i = 1; i <= $num_files; i++)); do
echo -n > $MNT/testdir/file_$i
if [ $((i % fsync_period)) -eq 0 ]; then
start=$(date +%s%N)
xfs_io -c "fsync" $MNT/testdir
end=$(date +%s%N)
fsync_total_time=$((fsync_total_time + (end - start)))
fsync_times[i]=$(( (end - start) / 1000000 ))
echo -n -e "Progress $i / $num_files\r"
fi
done
echo -e "\nHistogram of directory fsync duration in ms:\n"
printf '%s\n' "${fsync_times[@]}" | \
perl -MStatistics::Histogram -e '@d = <>; print get_histogram(\@d);'
fsync_total_time=$((fsync_total_time / 1000000))
echo -e "\nTotal time spent in fsync: $fsync_total_time ms\n"
echo
umount $MNT
The test was run on a non-debug kernel (Debian's default kernel config)
against a 15G null block device.
Result before this change:
Histogram of directory fsync duration in ms:
Count: 10000
Range: 3.000 - 362.000; Mean: 34.556; Median: 31.000; Stddev: 25.751
Percentiles: 90th: 71.000; 95th: 77.000; 99th: 81.000
3.000 - 5.278: 1423 #################################
5.278 - 8.854: 1173 ###########################
8.854 - 14.467: 591 ##############
14.467 - 23.277: 1025 #######################
23.277 - 37.105: 1422 #################################
37.105 - 58.809: 2036 ###############################################
58.809 - 92.876: 2316 #####################################################
92.876 - 146.346: 6 |
146.346 - 230.271: 6 |
230.271 - 362.000: 2 |
Total time spent in fsync: 350527 ms
Result after this change:
Histogram of directory fsync duration in ms:
Count: 10000
Range: 3.000 - 1088.000; Mean: 8.704; Median: 8.000; Stddev: 12.576
Percentiles: 90th: 12.000; 95th: 14.000; 99th: 17.000
3.000 - 6.007: 3222 #################################
6.007 - 11.276: 5197 #####################################################
11.276 - 20.506: 1551 ################
20.506 - 36.674: 24 |
36.674 - 201.552: 1 |
201.552 - 353.841: 4 |
353.841 - 1088.000: 1 |
Total time spent in fsync: 92114 ms
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
To prepare for a new caller that already has the ordered_extent
available, change btrfs_extract_ordered_extent to take an argument
for it. Add a wrapper for the bio case that still has to do the
lookup (for now).
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Tested-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
