Currently ext4 direct IO handling is split between ext4_ext_direct_IO()
and ext4_ind_direct_IO(). However the extent based function calls into
the indirect based one for some cases and for example it is not able to
handle file extending. Previously it was not also properly handling
retries in case of ENOSPC errors. With DAX things would get even more
contrieved so just refactor the direct IO code and instead of indirect /
extent split do the split to read vs writes.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Currently, ext4_map_blocks() just returns 0 when it finds a hole and
allocation is not requested. However we have all the information
available to tell how large the hole actually is and there are callers
of ext4_map_blocks() which would save some block-by-block hole iteration
if they knew this information. So fill in struct ext4_map_blocks even
for holes with the information we have. We keep returning 0 for holes to
maintain backward compatibility of the function.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Rename ext4_get_blocks_write() to ext4_get_blocks_unwritten() to better
describe what it does. Also split out get blocks functions for direct
IO. Later we move functionality from _ext4_get_blocks() there. There's no
functional change in this patch.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Create separate predicate functions to test/set/clear feature flags,
thereby replacing the wordy old macros. Furthermore, clean out the
places where we open-coded feature tests.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Instead of overloading EIO for CRC errors and corrupt structures,
return the same error codes that XFS returns for the same issues.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
In order to handle the !CONFIG_TRANSPARENT_HUGEPAGES case, we need to
return VM_FAULT_FALLBACK from the inlined dax_pmd_fault(), which is
defined in linux/mm.h. Given that we don't want to include <linux/mm.h>
in <linux/fs.h>, the easiest solution is to move the DAX-related
functions to a new header, <linux/dax.h>. We could also have moved
VM_FAULT_* definitions to a new header, or a different header that isn't
quite such a boil-the-ocean header as <linux/mm.h>, but this felt like
the best option.
Signed-off-by: Matthew Wilcox <willy@linux.intel.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Jan Kara <jack@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
ext4 isn't willing to map clusters to a non-extent file. Don't signal
this with an out of space error, since the FS will retry the
allocation (which didn't fail) forever. Instead, return EUCLEAN so
that the operation will fail immediately all the way back to userspace.
(The fix is either to run e2fsck -E bmap2extent, or to chattr +e the file.)
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: stable@vger.kernel.org
In order to prevent quota block tracking to be inaccurate when
ext4_quota_write() fails with ENOSPC, we make two changes. The quota
file can now use the reserved block (since the quota file is arguably
file system metadata), and ext4_quota_write() now uses
ext4_should_retry_alloc() to retry the block allocation after a commit
has completed and released some blocks for allocation.
This fixes failures of xfstests generic/270:
Quota error (device vdc): write_blk: dquota write failed
Quota error (device vdc): qtree_write_dquot: Error -28 occurred while creating quota
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
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>
The rw parameter to direct_IO is redundant with iov_iter->type, and
treated slightly differently just about everywhere it's used: some users
do rw & WRITE, and others do rw == WRITE where they should be doing a
bitwise check. Simplify this with the new iov_iter_rw() helper, which
always returns either READ or WRITE.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Most filesystems call through to these at some point, so we'll start
here.
Signed-off-by: Omar Sandoval <osandov@osandov.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
struct kiocb now is a generic I/O container, so move it to fs.h.
Also do a #include diet for aio.h while we're at it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
and read-only images (for which the implementation is mostly just the
reserved code point for a read-only feature :-)
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2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=PF+t
-----END PGP SIGNATURE-----
Merge tag 'ext4_for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4
Pull ext4 fixes from Ted Ts'o:
"Ext4 bug fixes.
We also reserved code points for encryption and read-only images (for
which the implementation is mostly just the reserved code point for a
read-only feature :-)"
* tag 'ext4_for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4:
ext4: fix indirect punch hole corruption
ext4: ignore journal checksum on remount; don't fail
ext4: remove duplicate remount check for JOURNAL_CHECKSUM change
ext4: fix mmap data corruption in nodelalloc mode when blocksize < pagesize
ext4: support read-only images
ext4: change to use setup_timer() instead of init_timer()
ext4: reserve codepoints used by the ext4 encryption feature
jbd2: complain about descriptor block checksum errors
This is a port of the DAX functionality found in the current version of
ext2.
[matthew.r.wilcox@intel.com: heavily tweaked]
[akpm@linux-foundation.org: remap_pages went away]
Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Andreas Dilger <andreas.dilger@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Cc: Boaz Harrosh <boaz@plexistor.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit 4f579ae7de (ext4: fix punch hole on files with indirect
mapping) rewrote FALLOC_FL_PUNCH_HOLE for ext4 files with indirect
mapping. However, there are bugs in several corner cases. This fixes 5
distinct bugs:
1. When there is at least one entire level of indirection between the
start and end of the punch range and the end of the punch range is the
first block of its level, we can't return early; we have to free the
intervening levels.
2. When the end is at a higher level of indirection than the start and
ext4_find_shared returns a top branch for the end, we still need to free
the rest of the shared branch it returns; we can't decrement partial2.
3. When a punch happens within one level of indirection, we need to
converge on an indirect block that contains the start and end. However,
because the branches returned from ext4_find_shared do not necessarily
start at the same level (e.g., the partial2 chain will be shallower if
the last block occurs at the beginning of an indirect group), the walk
of the two chains can end up "missing" each other and freeing a bunch of
extra blocks in the process. This mismatch can be handled by first
making sure that the chains are at the same level, then walking them
together until they converge.
4. When the punch happens within one level of indirection and
ext4_find_shared returns a top branch for the start, we must free it,
but only if the end does not occur within that branch.
5. When the punch happens within one level of indirection and
ext4_find_shared returns a top branch for the end, then we shouldn't
free the block referenced by the end of the returned chain (this mirrors
the different levels case).
Signed-off-by: Omar Sandoval <osandov@osandov.com>
The EXT4_STATE_DELALLOC_RESERVED flag was originally implemented
because it was too hard to make sure the mballoc and get_block flags
could be reliably passed down through all of the codepaths that end up
calling ext4_mb_new_blocks().
Since then, we have mb_flags passed down through most of the code
paths, so getting rid of EXT4_STATE_DELALLOC_RESERVED isn't as tricky
as it used to.
This commit plumbs in the last of what is required, and then adds a
WARN_ON check to make sure we haven't missed anything. If this passes
a full regression test run, we can then drop
EXT4_STATE_DELALLOC_RESERVED.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Jan Kara <jack@suse.cz>
Instead of initializing the allocation_request structure in
ext4_alloc_branch(), set it up in ext4_ind_map_blocks(), and then pass
it to ext4_alloc_branch() and ext4_splice_branch().
This allows ext4_ind_map_blocks to pass flags in the allocation
request structure without having to add Yet Another argument to
ext4_alloc_branch().
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Jan Kara <jack@suse.cz>
Currently punch hole code on files with direct/indirect mapping has some
problems which may lead to a data loss. For example (from Jan Kara):
fallocate -n -p 10240000 4096
will punch the range 10240000 - 12632064 instead of the range 1024000 -
10244096.
Also the code is a bit weird and it's not using infrastructure provided
by indirect.c, but rather creating it's own way.
This patch fixes the issues as well as making the operation to run 4
times faster from my testing (punching out 60GB file). It uses similar
approach used in ext4_ind_truncate() which takes advantage of
ext4_free_branches() function.
Also rename the ext4_free_hole_blocks() to something more sensible, like
the equivalent we have for extent mapped files. Call it
ext4_ind_remove_space().
This has been tested mostly with fsx and some xfstests which are testing
punch hole but does not require unwritten extents which are not
supported with direct/indirect mapping. Not problems showed up even with
1024k block size.
CC: stable@vger.kernel.org
Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Hole punching code for files with indirect blocks wrongly computed
number of blocks which need to be cleared when traversing the indirect
block tree. That could result in punching more blocks than actually
requested and thus effectively cause a data loss. For example:
fallocate -n -p 10240000 4096
will punch the range 10240000 - 12632064 instead of the range 1024000 -
10244096. Fix the calculation.
CC: stable@vger.kernel.org
Fixes: 8bad6fc813
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
free_holes_block() passed local variable as a block pointer
to ext4_clear_blocks(). Thus ext4_clear_blocks() zeroed out this local
variable instead of proper place in inode / indirect block. We later
zero out proper place in inode / indirect block but don't dirty the
inode / buffer again which can lead to subtle issues (some changes e.g.
to inode can be lost).
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Error recovery in ext4_alloc_branch() calls ext4_forget() even for
buffer corresponding to indirect block it did not allocate. This leads
to brelse() being called twice for that buffer (once from ext4_forget()
and once from cleanup in ext4_ind_map_blocks()) leading to buffer use
count misaccounting. Eventually (but often much later because there
are other users of the buffer) we will see messages like:
VFS: brelse: Trying to free free buffer
Another manifestation of this problem is an error:
JBD2 unexpected failure: jbd2_journal_revoke: !buffer_revoked(bh);
inconsistent data on disk
The fix is easy - don't forget buffer we did not allocate. Also add an
explanatory comment because the indexing at ext4_alloc_branch() is
somewhat subtle.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: stable@vger.kernel.org
After applied the commit (4a092d73), we have reduced the number of
source files that need to #include ext4_extents.h. But we can do
better.
This commit defines ext4_zeroout_es() in extents.c and move
EXT_MAX_BLOCKS into ext4.h in order not to include ext4_extents.h in
indirect.c and ioctl.c. Meanwhile we just need to include this file in
extent_status.c when ES_AGGRESSIVE_TEST is defined. Otherwise, this
commit removes a duplicated declaration in trace/events/ext4.h.
After applied this patch, we just need to include ext4_extents.h file
in {super,migrate,move_extents,extents}.c, and it is easy for us to
define a new extent disk layout.
Signed-off-by: Zheng Liu <wenqing.lz@taobao.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Translate the bitfields used in various flags argument to strings to
make the tracepoint output more human-readable.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Commit 18888cf088: "ext4: speed up truncate/unlink by not using
bforget() unless needed" removed the use of EXT4_FREE_BLOCKS_FORGET in
the most important codepath for file systems using extents, but a
similar optimization also can be done for file systems using indirect
blocks, and for the two special cases in the ext4 extents code.
Cc: Andrey Sidorov <qrxd43@motorola.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
We don't have to wait for unwritten extent conversion in
ext4_ind_direct_IO() as all writes that happened before DIO are
flushed by the generic code and extent conversion has happened before
we cleared PageWriteback bit.
Reviewed-by: Zheng Liu <wenqing.lz@taobao.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
ext4_ind_trans_blocks() wrongly used 'chunk' argument to decide whether
blocks mapped are logically contiguous. That is wrong since the argument
informs whether the blocks are physically contiguous. As the blocks
mapped are always logically contiguous and that's all
ext4_ind_trans_blocks() cares about, just remove the 'chunk' argument.
Reviewed-by: Zheng Liu <wenqing.lz@taobao.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Move common code in ext4_ind_truncate() and ext4_ext_truncate() into
ext4_truncate(). This saves over 60 lines of code.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Move common code in ext4_ind_punch_hole() and ext4_ext_punch_hole()
into ext4_punch_hole(). This saves over 150 lines of code.
This also fixes a potential bug when the punch_hole() code is racing
against indirect-to-extents or extents-to-indirect migation. We are
currently using i_mutex to protect against changes to the inode flag;
specifically, the append-only, immutable, and extents inode flags. So
we need to take i_mutex before deciding whether to use the
extents-specific or indirect-specific punch_hole code.
Also, there was a missing call to ext4_inode_block_unlocked_dio() in
the indirect punch codepath. This was added in commit 02d262dffc
to block DIO readers racing against the punch operation in the
codepath for extent-mapped inodes, but it was missing for
indirect-block mapped inodes. One of the advantages of refactoring
the code is that it makes such oversights much less likely.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
The older code was far more complicated than it needed to be because
of how we spliced in the ext4's new multiblock allocator into ext3's
indirect block code. By folding ext4_alloc_blocks() into
ext4_alloc_branch(), we make the code far more understable, shave off
over 130 lines of code and half a kilobyte of compiled object code.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
When an extent was zeroed out, we forgot to do convert from cpu to le16.
It could make us hit a BUG_ON when we try to write dirty pages out. So
fix it.
[ Also fix a bug found by Dmitry Monakhov where we were missing
le32_to_cpu() calls in the new indirect punch hole code.
There are a number of other big endian warnings found by static code
analyzers, but we'll wait for the next merge window to fix them all
up. These fixes are designed to be Obviously Correct by code
inspection, and easy to demonstrate that it won't make any
difference (and hence, won't introduce any bugs) on little endian
architectures such as x86. --tytso ]
Signed-off-by: Zheng Liu <wenqing.lz@taobao.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Reported-by: CAI Qian <caiqian@redhat.com>
Reported-by: Christian Kujau <lists@nerdbynature.de>
Cc: Dmitry Monakhov <dmonakhov@openvz.org>
So we can better understand what bits of ext4 are responsible for
long-running jbd2 handles, use jbd2__journal_start() so we can pass
context information for logging purposes.
The recommended way for finding the longer-running handles is:
T=/sys/kernel/debug/tracing
EVENT=$T/events/jbd2/jbd2_handle_stats
echo "interval > 5" > $EVENT/filter
echo 1 > $EVENT/enable
./run-my-fs-benchmark
cat $T/trace > /tmp/problem-handles
This will list handles that were active for longer than 20ms. Having
longer-running handles is bad, because a commit started at the wrong
time could stall for those 20+ milliseconds, which could delay an
fsync() or an O_SYNC operation. Here is an example line from the
trace file describing a handle which lived on for 311 jiffies, or over
1.2 seconds:
postmark-2917 [000] .... 196.435786: jbd2_handle_stats: dev 254,32
tid 570 type 2 line_no 2541 interval 311 sync 0 requested_blocks 1
dirtied_blocks 0
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Use WARN rather than printk followed by WARN_ON(1), for conciseness.
A simplified version of the semantic patch that makes this transformation
is as follows: (http://coccinelle.lip6.fr/)
// <smpl>
@@
expression list es;
@@
-printk(
+WARN(1,
es);
-WARN_ON(1);
// </smpl>
Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
This patch add supports for indirect file support punching hole. It
is almost the same as ext4_ext_punch_hole. First, we invalidate all
pages between this hole, and then we try to deallocate all blocks of
this hole.
A recursive function is used to handle deallocation of blocks. In
this function, it iterates over the entries in inode's i_blocks or
indirect blocks, and try to free the block for each one of them.
After applying this patch, xfstest #255 will not pass w/o extent because
indirect-based file doesn't support unwritten extents.
Signed-off-by: Zheng Liu <wenqing.lz@taobao.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
The only reason for sb_getblk() failing is if it can't allocate the
buffer_head. So ENOMEM is more appropriate than EIO. In addition,
make sure that the file system is marked as being inconsistent if
sb_getblk() fails.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@vger.kernel.org
Previously, ext4_extents.h was being included at the end of ext4.h,
which was bad for a number of reasons: (a) it was not being included
in the expected place, and (b) it caused the header to be included
multiple times. There were #ifdef's to prevent this from causing any
problems, but it still was unnecessary.
By moving the function declarations that were in ext4_extents.h to
ext4.h, which is standard practice for where the function declarations
for the rest of ext4.h can be found, we can remove ext4_extents.h from
being included in ext4.h at all, and then we can only include
ext4_extents.h where it is needed in ext4's source files.
It should be possible to move a few more things into ext4.h, and
further reduce the number of source files that need to #include
ext4_extents.h, but that's a cleanup for another day.
Reported-by: Sachin Kamat <sachin.kamat@linaro.org>
Reported-by: Wei Yongjun <weiyj.lk@gmail.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
This patch lets ext4 maintain extent status tree.
Currently it only tracks delay extent status in extent status tree. When a
delay allocation is issued, the related delay extent will be inserted into
extent status tree. When a delay extent is written out or invalidated, it will
be removed from this tree.
Signed-off-by: Yongqiang Yang <xiaoqiangnk@gmail.com>
Signed-off-by: Allison Henderson <achender@linux.vnet.ibm.com>
Signed-off-by: Zheng Liu <wenqing.lz@taobao.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
When we use trace_ext4_ext/ind_map_blocks_exit, print the value of
map->m_flags in order that we can understand the extent's current
status.
Reviewed-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: Zheng Liu <wenqing.lz@taobao.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
BUG #1) All places where we call ext4_flush_completed_IO are broken
because buffered io and DIO/AIO goes through three stages
1) submitted io,
2) completed io (in i_completed_io_list) conversion pended
3) finished io (conversion done)
And by calling ext4_flush_completed_IO we will flush only
requests which were in (2) stage, which is wrong because:
1) punch_hole and truncate _must_ wait for all outstanding unwritten io
regardless to it's state.
2) fsync and nolock_dio_read should also wait because there is
a time window between end_page_writeback() and ext4_add_complete_io()
As result integrity fsync is broken in case of buffered write
to fallocated region:
fsync blkdev_completion
->filemap_write_and_wait_range
->ext4_end_bio
->end_page_writeback
<-- filemap_write_and_wait_range return
->ext4_flush_completed_IO
sees empty i_completed_io_list but pended
conversion still exist
->ext4_add_complete_io
BUG #2) Race window becomes wider due to the 'ext4: completed_io
locking cleanup V4' patch series
This patch make following changes:
1) ext4_flush_completed_io() now first try to flush completed io and when
wait for any outstanding unwritten io via ext4_unwritten_wait()
2) Rename function to more appropriate name.
3) Assert that all callers of ext4_flush_unwritten_io should hold i_mutex to
prevent endless wait
Signed-off-by: Dmitry Monakhov <dmonakhov@openvz.org>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Reviewed-by: Jan Kara <jack@suse.cz>
Inode's block defrag and ext4_change_inode_journal_flag() may
affect nonlocked DIO reads result, so proper synchronization
required.
- Add missed inode_dio_wait() calls where appropriate
- Check inode state under extra i_dio_count reference.
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dmitry Monakhov <dmonakhov@openvz.org>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Current unwritten extent conversion state-machine is very fuzzy.
- For unknown reason it performs conversion under i_mutex. What for?
My diagnosis:
We already protect extent tree with i_data_sem, truncate and punch_hole
should wait for DIO, so the only data we have to protect is end_io->flags
modification, but only flush_completed_IO and end_io_work modified this
flags and we can serialize them via i_completed_io_lock.
Currently all these games with mutex_trylock result in the following deadlock
truncate: kworker:
ext4_setattr ext4_end_io_work
mutex_lock(i_mutex)
inode_dio_wait(inode) ->BLOCK
DEADLOCK<- mutex_trylock()
inode_dio_done()
#TEST_CASE1_BEGIN
MNT=/mnt_scrach
unlink $MNT/file
fallocate -l $((1024*1024*1024)) $MNT/file
aio-stress -I 100000 -O -s 100m -n -t 1 -c 10 -o 2 -o 3 $MNT/file
sleep 2
truncate -s 0 $MNT/file
#TEST_CASE1_END
Or use 286's xfstests https://github.com/dmonakhov/xfstests/blob/devel/286
This patch makes state machine simple and clean:
(1) xxx_end_io schedule final extent conversion simply by calling
ext4_add_complete_io(), which append it to ei->i_completed_io_list
NOTE1: because of (2A) work should be queued only if
->i_completed_io_list was empty, otherwise the work is scheduled already.
(2) ext4_flush_completed_IO is responsible for handling all pending
end_io from ei->i_completed_io_list
Flushing sequence consists of following stages:
A) LOCKED: Atomically drain completed_io_list to local_list
B) Perform extents conversion
C) LOCKED: move converted io's to to_free list for final deletion
This logic depends on context which we was called from.
D) Final end_io context destruction
NOTE1: i_mutex is no longer required because end_io->flags modification
is protected by ei->ext4_complete_io_lock
Full list of changes:
- Move all completion end_io related routines to page-io.c in order to improve
logic locality
- Move open coded logic from various xx_end_xx routines to ext4_add_complete_io()
- remove EXT4_IO_END_FSYNC
- Improve SMP scalability by removing useless i_mutex which does not
protect io->flags anymore.
- Reduce lock contention on i_completed_io_lock by optimizing list walk.
- Rename ext4_end_io_nolock to end4_end_io and make it static
- Check flush completion status to ext4_ext_punch_hole(). Because it is
not good idea to punch blocks from corrupted inode.
Changes since V3 (in request to Jan's comments):
Fall back to active flush_completed_IO() approach in order to prevent
performance issues with nolocked DIO reads.
Changes since V2:
Fix use-after-free caused by race truncate vs end_io_work
Signed-off-by: Dmitry Monakhov <dmonakhov@openvz.org>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Delete any instances of include module.h that were not strictly
required. In the case of ext2, the declaration of MODULE_LICENSE
etc. were in inode.c but the module_init/exit were in super.c, so
relocate the MODULE_LICENCE/AUTHOR block to super.c which makes it
consistent with ext3 and ext4 at the same time.
Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com>
Signed-off-by: Jan Kara <jack@suse.cz>
At least initially if the bigalloc feature is enabled, we will not
support non-extent mapped inodes, online resizing, online defrag, or
the FITRIM ioctl. This simplifies the initial implementation.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
While running extended fsx tests to verify the first
two patches, a similar bug was also found in the
truncate operation.
This bug happens because the truncate routine only zeros
the unblock aligned portion of the last page. This means
that the block aligned portions of the page appearing after
i_size are left unzeroed, and the buffer heads still mapped.
This bug is corrected by using ext4_discard_partial_page_buffers
in the truncate routine to zero the partial page and unmap
the buffer headers.
Signed-off-by: Allison Henderson <achender@linux.vnet.ibm.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
