Currently in ext4_alloc_file_blocks() the number of credits is
calculated only once before we enter the allocation loop. However within
the allocation loop the extent tree depth can change, hence the number
of credits needed can increase potentially exceeding the number of credits
reserved in the handle which can cause journal failures.
Fix this by recalculating number of credits when the inode depth
changes. Note that even though ext4_alloc_file_blocks() is only
currently used by extent base inodes we will avoid recalculating number
of credits unnecessarily in the case of indirect based inodes.
Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Use kernel.h macro definition.
Thanks to Julia Lawall for Coccinelle scripting support.
Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Use kernel.h macro definition.
Thanks to Julia Lawall for Coccinelle scripting support.
Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
The commit cf108bca46: "ext4: Invert the locking order of page_lock
and transaction start" caused __ext4_journalled_writepage() to drop
the page lock before the page was written back, as part of changing
the locking order to jbd2_journal_start -> page_lock. However, this
introduced a potential race if there was a truncate racing with the
data=journalled writeback mode.
Fix this by grabbing the page lock after starting the journal handle,
and then checking to see if page had gotten truncated out from under
us.
This fixes a number of different warnings or BUG_ON's when running
xfstests generic/086 in data=journalled mode, including:
jbd2_journal_dirty_metadata: vdc-8: bad jh for block 115643: transaction (ee3fe7
c0, 164), jh->b_transaction ( (null), 0), jh->b_next_transaction ( (null), 0), jlist 0
- and -
kernel BUG at /usr/projects/linux/ext4/fs/jbd2/transaction.c:2200!
...
Call Trace:
[<c02b2ded>] ? __ext4_journalled_invalidatepage+0x117/0x117
[<c02b2de5>] __ext4_journalled_invalidatepage+0x10f/0x117
[<c02b2ded>] ? __ext4_journalled_invalidatepage+0x117/0x117
[<c027d883>] ? lock_buffer+0x36/0x36
[<c02b2dfa>] ext4_journalled_invalidatepage+0xd/0x22
[<c0229139>] do_invalidatepage+0x22/0x26
[<c0229198>] truncate_inode_page+0x5b/0x85
[<c022934b>] truncate_inode_pages_range+0x156/0x38c
[<c0229592>] truncate_inode_pages+0x11/0x15
[<c022962d>] truncate_pagecache+0x55/0x71
[<c02b913b>] ext4_setattr+0x4a9/0x560
[<c01ca542>] ? current_kernel_time+0x10/0x44
[<c026c4d8>] notify_change+0x1c7/0x2be
[<c0256a00>] do_truncate+0x65/0x85
[<c0226f31>] ? file_ra_state_init+0x12/0x29
- and -
WARNING: CPU: 1 PID: 1331 at /usr/projects/linux/ext4/fs/jbd2/transaction.c:1396
irty_metadata+0x14a/0x1ae()
...
Call Trace:
[<c01b879f>] ? console_unlock+0x3a1/0x3ce
[<c082cbb4>] dump_stack+0x48/0x60
[<c0178b65>] warn_slowpath_common+0x89/0xa0
[<c02ef2cf>] ? jbd2_journal_dirty_metadata+0x14a/0x1ae
[<c0178bef>] warn_slowpath_null+0x14/0x18
[<c02ef2cf>] jbd2_journal_dirty_metadata+0x14a/0x1ae
[<c02d8615>] __ext4_handle_dirty_metadata+0xd4/0x19d
[<c02b2f44>] write_end_fn+0x40/0x53
[<c02b4a16>] ext4_walk_page_buffers+0x4e/0x6a
[<c02b59e7>] ext4_writepage+0x354/0x3b8
[<c02b2f04>] ? mpage_release_unused_pages+0xd4/0xd4
[<c02b1b21>] ? wait_on_buffer+0x2c/0x2c
[<c02b5a4b>] ? ext4_writepage+0x3b8/0x3b8
[<c02b5a5b>] __writepage+0x10/0x2e
[<c0225956>] write_cache_pages+0x22d/0x32c
[<c02b5a4b>] ? ext4_writepage+0x3b8/0x3b8
[<c02b6ee8>] ext4_writepages+0x102/0x607
[<c019adfe>] ? sched_clock_local+0x10/0x10e
[<c01a8a7c>] ? __lock_is_held+0x2e/0x44
[<c01a8ad5>] ? lock_is_held+0x43/0x51
[<c0226dff>] do_writepages+0x1c/0x29
[<c0276bed>] __writeback_single_inode+0xc3/0x545
[<c0277c07>] writeback_sb_inodes+0x21f/0x36d
...
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: stable@vger.kernel.org
We currently don't correctly handle the case where blocksize !=
pagesize, so disallow the mount in those cases.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
This patch implements fallocate's FALLOC_FL_INSERT_RANGE for Ext4.
1) Make sure that both offset and len are block size aligned.
2) Update the i_size of inode by len bytes.
3) Compute the file's logical block number against offset. If the computed
block number is not the starting block of the extent, split the extent
such that the block number is the starting block of the extent.
4) Shift all the extents which are lying between [offset, last allocated extent]
towards right by len bytes. This step will make a hole of len bytes
at offset.
Signed-off-by: Namjae Jeon <namjae.jeon@samsung.com>
Signed-off-by: Ashish Sangwan <a.sangwan@samsung.com>
[ Added another sparse fix for EXT4_IOC_GET_ENCRYPTION_POLICY while
we're at it. --tytso ]
Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
During a source code review of fs/ext4/extents.c I noted identical
consecutive lines. An assertion is repeated for inode1 and never done
for inode2. This is not in keeping with the rest of the code in the
ext4_swap_extents function and appears to be a bug.
Assert that the inode2 mutex is not locked.
Signed-off-by: David Moore <dmoorefo@gmail.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Currently ext4_mb_good_group() only returns 0 or 1 depending on whether
the allocation group is suitable for use or not. However we might get
various errors and fail while initializing new group including -EIO
which would never get propagated up the call chain. This might lead to
an endless loop at writeback when we're trying to find a good group to
allocate from and we fail to initialize new group (read error for
example).
Fix this by returning proper error code from ext4_mb_good_group() and
using it in ext4_mb_regular_allocator(). In ext4_mb_regular_allocator()
we will always return only the first occurred error from
ext4_mb_good_group() and we only propagate it back to the caller if we
do not get any other errors and we fail to allocate any blocks.
Note that with other modes than errors=continue, we will fail
immediately in ext4_mb_good_group() in case of error, however with
errors=continue we should try to continue using the file system, that's
why we're not going to fail immediately when we see an error from
ext4_mb_good_group(), but rather when we fail to find a suitable block
group to allocate from due to an problem in group initialization.
Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Currently on the machines with page size > block size when initializing
block group buddy cache we initialize it for all the block group bitmaps
in the page. However in the case of read error, checksum error, or if
a single bitmap is in any way corrupted we would fail to initialize all
of the bitmaps. This is problematic because we will not have access to
the other allocation groups even though those might be perfectly fine
and usable.
Fix this by reading all the bitmaps instead of error out on the first
problem and simply skip the bitmaps which were either not read properly,
or are not valid.
Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
If we want to rely on the buffer_verified() flag of the block bitmap
buffer, we have to set it consistently. However currently if we're
initializing uninitialized block bitmap in
ext4_read_block_bitmap_nowait() we're not going to set buffer verified
at all.
We can do this by simply setting the flag on the buffer, but I think
it's actually better to run ext4_validate_block_bitmap() to make sure
that what we did in the ext4_init_block_bitmap() is right.
So run ext4_validate_block_bitmap() even after the block bitmap
initialization. Also bail out early from ext4_validate_block_bitmap() if
we see corrupt bitmap, since we already know it's corrupt and we do not
need to verify that.
Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Previously we allocated bounce pages using a combination of
alloc_page() and mempool_alloc() with the __GFP_WAIT bit set.
Instead, use mempool_alloc() with GFP_NOWAIT. The mempool_alloc()
function will try using alloc_pages() initially, and then only use the
mempool reserve of pages if alloc_pages() is unable to fulfill the
request.
This minimizes the the impact on the mm layer when we need to do a
large amount of writeback of encrypted files, as Jaeguk Kim had
reported that under a heavy fio workload on a system with restricted
amounts memory (which unfortunately, includes many mobile handsets),
he had observed the the OOM killer getting triggered several times.
Using GFP_NOWAIT
If the mempool_alloc() function fails, we will retry the page
writeback at a later time; the function of the mempool is to ensure
that we can writeback at least 32 pages at a time, so we can more
efficiently dispatch I/O under high memory pressure situations. In
the future we should make this be a tunable so we can determine the
best tradeoff between permanently sequestering memory and the ability
to quickly launder pages so we can free up memory quickly when
necessary.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Crypto resource should be released when ext4 module exits, otherwise
it will cause memory leak.
Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Previously we were taking the required padding when allocating space
for the on-disk symlink. This caused a buffer overrun which could
trigger a krenel crash when running fsstress.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Fix a potential memory leak where fname->crypto_buf.name wouldn't get
freed in some error paths, and also make the error handling easier to
understand/audit.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Factor out calls to ext4_inherit_context() and move them to
__ext4_new_inode(); this fixes a problem where ext4_tmpfile() wasn't
calling calling ext4_inherit_context(), so the temporary file wasn't
getting protected. Since the blocks for the tmpfile could end up on
disk, they really should be protected if the tmpfile is created within
the context of an encrypted directory.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Set up the encryption information for newly created inodes immediately
after they inherit their encryption context from their parent
directories.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
ext4_encrypted_zeroout() could end up leaking a bio and bounce page.
Fortunately it's not used much. While we're fixing things up,
refactor out common code into the static function alloc_bounce_page()
and fix up error handling if mempool_alloc() fails.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
As suggested by Herbert Xu, we shouldn't allocate a new tfm each time
we read or write a page. Instead we can use a single tfm hanging off
the inode's crypt_info structure for all of our encryption needs for
that inode, since the tfm can be used by multiple crypto requests in
parallel.
Also use cmpxchg() to avoid races that could result in crypt_info
structure getting doubly allocated or doubly freed.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
On arm64 this is apparently needed for CTS mode to function correctly.
Otherwise attempts to use CTS return ENOENT.
Change-Id: I732ea9a5157acc76de5b89edec195d0365f4ca63
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Some fields are only used when the crypto_ctx is being used on the
read path, some are only used on the write path, and some are only
used when the structure is on free list. Optimize memory use by using
a union.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Use slab caches the ext4_crypto_ctx and ext4_crypt_info structures for
slighly better memory efficiency and debuggability.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
The superblock fields s_file_encryption_mode and s_dir_encryption_mode
are vestigal, so remove them as a cleanup. While we're at it, allow
file systems with both encryption and inline_data enabled at the same
time to work correctly. We can't have encrypted inodes with inline
data, but there's no reason to prohibit unencrypted inodes from using
the inline data feature.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
This is a pretty massive patch which does a number of different things:
1) The per-inode encryption information is now stored in an allocated
data structure, ext4_crypt_info, instead of directly in the node.
This reduces the size usage of an in-memory inode when it is not
using encryption.
2) We drop the ext4_fname_crypto_ctx entirely, and use the per-inode
encryption structure instead. This remove an unnecessary memory
allocation and free for the fname_crypto_ctx as well as allowing us
to reuse the ctfm in a directory for multiple lookups and file
creations.
3) We also cache the inode's policy information in the ext4_crypt_info
structure so we don't have to continually read it out of the
extended attributes.
4) We now keep the keyring key in the inode's encryption structure
instead of releasing it after we are done using it to derive the
per-inode key. This allows us to test to see if the key has been
revoked; if it has, we prevent the use of the derived key and free
it.
5) When an inode is released (or when the derived key is freed), we
will use memset_explicit() to zero out the derived key, so it's not
left hanging around in memory. This implies that when a user logs
out, it is important to first revoke the key, and then unlink it,
and then finally, to use "echo 3 > /proc/sys/vm/drop_caches" to
release any decrypted pages and dcache entries from the system
caches.
6) All this, and we also shrink the number of lines of code by around
100. :-)
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Use struct ext4_encryption_key only for the master key passed via the
kernel keyring.
For internal kernel space users, we now use struct ext4_crypt_info.
This will allow us to put information from the policy structure so we
can cache it and avoid needing to constantly looking up the extended
attribute. We will do this in a spearate patch. This patch is mostly
mechnical to make it easier for patch review.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Encrypt the filename as soon it is passed in by the user. This avoids
our needing to encrypt the filename 2 or 3 times while in the process
of creating a filename.
Similarly, when looking up a directory entry, encrypt the filename
early, or if the encryption key is not available, base-64 decode the
file syystem so that the hash value and the last 16 bytes of the
encrypted filename is available in the new struct ext4_filename data
structure.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
The xfstests test suite assumes that an attempt to collapse range on
the range (0, 1) will return EOPNOTSUPP if the file system does not
support collapse range. Commit 280227a75b: "ext4: move check under
lock scope to close a race" broke this, and this caused xfstests to
fail when run when testing file systems that did not have the extents
feature enabled.
Reported-by: Eric Whitney <enwlinux@gmail.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
The following commit introduced a bug when checking for zero length extent
5946d08 ext4: check for overlapping extents in ext4_valid_extent_entries()
Zero length extent could pass the check if lblock is zero.
Adding the explicit check for zero length back.
Signed-off-by: Eryu Guan <guaneryu@gmail.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: stable@vger.kernel.org
Currently when journal restart fails, we'll have the h_transaction of
the handle set to NULL to indicate that the handle has been effectively
aborted. We handle this situation quietly in the jbd2_journal_stop() and just
free the handle and exit because everything else has been done before we
attempted (and failed) to restart the journal.
Unfortunately there are a number of problems with that approach
introduced with commit
41a5b91319 "jbd2: invalidate handle if jbd2_journal_restart()
fails"
First of all in ext4 jbd2_journal_stop() will be called through
__ext4_journal_stop() where we would try to get a hold of the superblock
by dereferencing h_transaction which in this case would lead to NULL
pointer dereference and crash.
In addition we're going to free the handle regardless of the refcount
which is bad as well, because others up the call chain will still
reference the handle so we might potentially reference already freed
memory.
Moreover it's expected that we'll get aborted handle as well as detached
handle in some of the journalling function as the error propagates up
the stack, so it's unnecessary to call WARN_ON every time we get
detached handle.
And finally we might leak some memory by forgetting to free reserved
handle in jbd2_journal_stop() in the case where handle was detached from
the transaction (h_transaction is NULL).
Fix the NULL pointer dereference in __ext4_journal_stop() by just
calling jbd2_journal_stop() quietly as suggested by Jan Kara. Also fix
the potential memory leak in jbd2_journal_stop() and use proper
handle refcounting before we attempt to free it to avoid use-after-free
issues.
And finally remove all WARN_ON(!transaction) from the code so that we do
not get random traces when something goes wrong because when journal
restart fails we will get to some of those functions.
Cc: stable@vger.kernel.org
Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Jan Kara <jack@suse.cz>
The ext4_extent_tree_init() function hasn't been in the ext4 code for
a long time ago, except in an unused function prototype in ext4.h
Google-Bug-Id: 4530137
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
We had a fencepost error in the lazytime optimization which means that
timestamp would get written to the wrong inode.
Cc: stable@vger.kernel.org
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
for ext4 encryption which provide better security and performance.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2
iQEcBAABCAAGBQJVRsVDAAoJEPL5WVaVDYGj/UUIAI6zLGhq3I8uQLZQC22Ew2Ph
TPj6eABDuTrB/7QpAu21Dk59N70MQpsBTES6yLWWLf/eHp0gsH7gCNY/C9185vOh
tQjzw18hRH2IfPftOBrjDlPGbbBD8Gu9jAmpm5kKKOtBuSVbKQ4GeN6BTECkgwlg
U5EJHJJ5Ahl4MalODFreOE5ZrVC7FWGEpc1y/MquQ0qcGSGlNd35leK5FE2bfHWZ
M1IJfXH5RRVPUBp26uNvzEg0TtpqkigmCJUT6gOVLfSYBw+lYEbGl4lCflrJmbgt
8EZh3Q0plsDbNhMzqSvOE4RvsOZ28oMjRNbzxkAaoz/FlatWX2hrfAoI2nqRrKg=
=Unbp
-----END PGP SIGNATURE-----
Merge tag 'for_linus_stable' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4
Pull ext4 fixes from Ted Ts'o:
"Some miscellaneous bug fixes and some final on-disk and ABI changes
for ext4 encryption which provide better security and performance"
* tag 'for_linus_stable' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4:
ext4: fix growing of tiny filesystems
ext4: move check under lock scope to close a race.
ext4: fix data corruption caused by unwritten and delayed extents
ext4 crypto: remove duplicated encryption mode definitions
ext4 crypto: do not select from EXT4_FS_ENCRYPTION
ext4 crypto: add padding to filenames before encrypting
ext4 crypto: simplify and speed up filename encryption
The estimate of necessary transaction credits in ext4_flex_group_add()
is too pessimistic. It reserves credit for sb, resize inode, and resize
inode dindirect block for each group added in a flex group although they
are always the same block and thus it is enough to account them only
once. Also the number of modified GDT block is overestimated since we
fit EXT4_DESC_PER_BLOCK(sb) descriptors in one block.
Make the estimation more precise. That reduces number of requested
credits enough that we can grow 20 MB filesystem (which has 1 MB
journal, 79 reserved GDT blocks, and flex group size 16 by default).
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
fallocate() checks that the file is extent-based and returns
EOPNOTSUPP in case is not. Other tasks can convert from and to
indirect and extent so it's safe to check only after grabbing
the inode mutex.
Signed-off-by: Davide Italiano <dccitaliano@gmail.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: stable@vger.kernel.org
Currently it is possible to lose whole file system block worth of data
when we hit the specific interaction with unwritten and delayed extents
in status extent tree.
The problem is that when we insert delayed extent into extent status
tree the only way to get rid of it is when we write out delayed buffer.
However there is a limitation in the extent status tree implementation
so that when inserting unwritten extent should there be even a single
delayed block the whole unwritten extent would be marked as delayed.
At this point, there is no way to get rid of the delayed extents,
because there are no delayed buffers to write out. So when a we write
into said unwritten extent we will convert it to written, but it still
remains delayed.
When we try to write into that block later ext4_da_map_blocks() will set
the buffer new and delayed and map it to invalid block which causes
the rest of the block to be zeroed loosing already written data.
For now we can fix this by simply not allowing to set delayed status on
written extent in the extent status tree. Also add WARN_ON() to make
sure that we notice if this happens in the future.
This problem can be easily reproduced by running the following xfs_io.
xfs_io -f -c "pwrite -S 0xaa 4096 2048" \
-c "falloc 0 131072" \
-c "pwrite -S 0xbb 65536 2048" \
-c "fsync" /mnt/test/fff
echo 3 > /proc/sys/vm/drop_caches
xfs_io -c "pwrite -S 0xdd 67584 2048" /mnt/test/fff
This can be theoretically also reproduced by at random by running fsx,
but it's not very reliable, though on machines with bigger page size
(like ppc) this can be seen more often (especially xfstest generic/127)
Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: stable@vger.kernel.org
This patch removes duplicated encryption modes which were already in
ext4.h. They were duplicated from commit 3edc18d and commit f542fb.
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Michael Halcrow <mhalcrow@google.com>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Signed-off-by: Chanho Park <chanho61.park@samsung.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
This patch adds a tristate EXT4_ENCRYPTION to do the selections
for EXT4_FS_ENCRYPTION because selecting from a bool causes all
the selected options to be built-in, even if EXT4 itself is a
module.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
This obscures the length of the filenames, to decrease the amount of
information leakage. By default, we pad the filenames to the next 4
byte boundaries. This costs nothing, since the directory entries are
aligned to 4 byte boundaries anyway. Filenames can also be padded to
8, 16, or 32 bytes, which will consume more directory space.
Change-Id: Ibb7a0fb76d2c48e2061240a709358ff40b14f322
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Avoid using SHA-1 when calculating the user-visible filename when the
encryption key is available, and avoid decrypting lots of filenames
when searching for a directory entry in a directory block.
Change-Id: If4655f144784978ba0305b597bfa1c8d7bb69e63
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Pull fourth vfs update from Al Viro:
"d_inode() annotations from David Howells (sat in for-next since before
the beginning of merge window) + four assorted fixes"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
RCU pathwalk breakage when running into a symlink overmounting something
fix I_DIO_WAKEUP definition
direct-io: only inc/dec inode->i_dio_count for file systems
fs/9p: fix readdir()
VFS: assorted d_backing_inode() annotations
VFS: fs/inode.c helpers: d_inode() annotations
VFS: fs/cachefiles: d_backing_inode() annotations
VFS: fs library helpers: d_inode() annotations
VFS: assorted weird filesystems: d_inode() annotations
VFS: normal filesystems (and lustre): d_inode() annotations
VFS: security/: d_inode() annotations
VFS: security/: d_backing_inode() annotations
VFS: net/: d_inode() annotations
VFS: net/unix: d_backing_inode() annotations
VFS: kernel/: d_inode() annotations
VFS: audit: d_backing_inode() annotations
VFS: Fix up some ->d_inode accesses in the chelsio driver
VFS: Cachefiles should perform fs modifications on the top layer only
VFS: AF_UNIX sockets should call mknod on the top layer only
do_blockdev_direct_IO() increments and decrements the inode
->i_dio_count for each IO operation. It does this to protect against
truncate of a file. Block devices don't need this sort of protection.
For a capable multiqueue setup, this atomic int is the only shared
state between applications accessing the device for O_DIRECT, and it
presents a scaling wall for that. In my testing, as much as 30% of
system time is spent incrementing and decrementing this value. A mixed
read/write workload improved from ~2.5M IOPS to ~9.6M IOPS, with
better latencies too. Before:
clat percentiles (usec):
| 1.00th=[ 33], 5.00th=[ 34], 10.00th=[ 34], 20.00th=[ 34],
| 30.00th=[ 34], 40.00th=[ 34], 50.00th=[ 35], 60.00th=[ 35],
| 70.00th=[ 35], 80.00th=[ 35], 90.00th=[ 37], 95.00th=[ 80],
| 99.00th=[ 98], 99.50th=[ 151], 99.90th=[ 155], 99.95th=[ 155],
| 99.99th=[ 165]
After:
clat percentiles (usec):
| 1.00th=[ 95], 5.00th=[ 108], 10.00th=[ 129], 20.00th=[ 149],
| 30.00th=[ 155], 40.00th=[ 161], 50.00th=[ 167], 60.00th=[ 171],
| 70.00th=[ 177], 80.00th=[ 185], 90.00th=[ 201], 95.00th=[ 270],
| 99.00th=[ 390], 99.50th=[ 398], 99.90th=[ 418], 99.95th=[ 422],
| 99.99th=[ 438]
In other setups, Robert Elliott reported seeing good performance
improvements:
https://lkml.org/lkml/2015/4/3/557
The more applications accessing the device, the worse it gets.
Add a new direct-io flags, DIO_SKIP_DIO_COUNT, which tells
do_blockdev_direct_IO() that it need not worry about incrementing
or decrementing the inode i_dio_count for this caller.
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Elliott, Robert (Server Storage) <elliott@hp.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Jens Axboe <axboe@fb.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
