Skip to content

K
Kernel

openeuler / Kernel
大约 2 年 前同步成功

通知 8
Star 0
Fork 0
K
Kernel
切换分支/标签
  1. 29 3月, 2023 1 次提交
    • D
      xfs, iomap: limit individual ioend chain lengths in writeback · c5883137
      由 Dave Chinner 提交于 
      mainline inclusion
from mainline-v5.17-rc3
commit ebb7fb15
category: bugfix
bugzilla: https://gitee.com/openeuler/kernel/issues/I4KIAO
CVE: NA

Reference: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=ebb7fb1557b1d03b906b668aa2164b51e6b7d19a

--------------------------------

Trond Myklebust reported soft lockups in XFS IO completion such as
this:

 watchdog: BUG: soft lockup - CPU#12 stuck for 23s! [kworker/12:1:3106]
 CPU: 12 PID: 3106 Comm: kworker/12:1 Not tainted 4.18.0-305.10.2.el8_4.x86_64 #1
 Workqueue: xfs-conv/md127 xfs_end_io [xfs]
 RIP: 0010:_raw_spin_unlock_irqrestore+0x11/0x20
 Call Trace:
  wake_up_page_bit+0x8a/0x110
  iomap_finish_ioend+0xd7/0x1c0
  iomap_finish_ioends+0x7f/0xb0
  xfs_end_ioend+0x6b/0x100 [xfs]
  xfs_end_io+0xb9/0xe0 [xfs]
  process_one_work+0x1a7/0x360
  worker_thread+0x1fa/0x390
  kthread+0x116/0x130
  ret_from_fork+0x35/0x40

Ioends are processed as an atomic completion unit when all the
chained bios in the ioend have completed their IO. Logically
contiguous ioends can also be merged and completed as a single,
larger unit.  Both of these things can be problematic as both the
bio chains per ioend and the size of the merged ioends processed as
a single completion are both unbound.

If we have a large sequential dirty region in the page cache,
write_cache_pages() will keep feeding us sequential pages and we
will keep mapping them into ioends and bios until we get a dirty
page at a non-sequential file offset. These large sequential runs
can will result in bio and ioend chaining to optimise the io
patterns. The pages iunder writeback are pinned within these chains
until the submission chaining is broken, allowing the entire chain
to be completed. This can result in huge chains being processed
in IO completion context.

We get deep bio chaining if we have large contiguous physical
extents. We will keep adding pages to the current bio until it is
full, then we'll chain a new bio to keep adding pages for writeback.
Hence we can build bio chains that map millions of pages and tens of
gigabytes of RAM if the page cache contains big enough contiguous
dirty file regions. This long bio chain pins those pages until the
final bio in the chain completes and the ioend can iterate all the
chained bios and complete them.

OTOH, if we have a physically fragmented file, we end up submitting
one ioend per physical fragment that each have a small bio or bio
chain attached to them. We do not chain these at IO submission time,
but instead we chain them at completion time based on file
offset via iomap_ioend_try_merge(). Hence we can end up with unbound
ioend chains being built via completion merging.

XFS can then do COW remapping or unwritten extent conversion on that
merged chain, which involves walking an extent fragment at a time
and running a transaction to modify the physical extent information.
IOWs, we merge all the discontiguous ioends together into a
contiguous file range, only to then process them individually as
discontiguous extents.

This extent manipulation is computationally expensive and can run in
a tight loop, so merging logically contiguous but physically
discontigous ioends gains us nothing except for hiding the fact the
fact we broke the ioends up into individual physical extents at
submission and then need to loop over those individual physical
extents at completion.

Hence we need to have mechanisms to limit ioend sizes and
to break up completion processing of large merged ioend chains:

1. bio chains per ioend need to be bound in length. Pure overwrites
go straight to iomap_finish_ioend() in softirq context with the
exact bio chain attached to the ioend by submission. Hence the only
way to prevent long holdoffs here is to bound ioend submission
sizes because we can't reschedule in softirq context.

2. iomap_finish_ioends() has to handle unbound merged ioend chains
correctly. This relies on any one call to iomap_finish_ioend() being
bound in runtime so that cond_resched() can be issued regularly as
the long ioend chain is processed. i.e. this relies on mechanism #1
to limit individual ioend sizes to work correctly.

3. filesystems have to loop over the merged ioends to process
physical extent manipulations. This means they can loop internally,
and so we break merging at physical extent boundaries so the
filesystem can easily insert reschedule points between individual
extent manipulations.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reported-and-tested-by: NTrond Myklebust <trondmy@hammerspace.com>
Reviewed-by: NDarrick J. Wong <djwong@kernel.org>
Signed-off-by: NDarrick J. Wong <djwong@kernel.org>
Conflicts:
	include/linux/iomap.h
	fs/iomap/buffered-io.c
	fs/xfs/xfs_aops.c

	[ 6e552494 ("iomap: remove unused private field from ioend")
	  is not applied.
	  95c4cd05 ("iomap: Convert to_iomap_page to take a folio") is
	  not applied.
	  8ffd74e9 ("iomap: Convert bio completions to use folios") is
	  not applied.
	  044c6449 ("xfs: drop unused ioend private merge and
	  setfilesize code") is not applied. ]
Signed-off-by: NZhihao Cheng <chengzhihao1@huawei.com>
Reviewed-by: NZhang Yi <yi.zhang@huawei.com>
Signed-off-by: NJialin Zhang <zhangjialin11@huawei.com>
      c5883137
  3. 18 1月, 2023 1 次提交
  5. 27 12月, 2021 2 次提交
    • B
      xfs: punch out data fork delalloc blocks on COW writeback failure · 22987db8
      由 Brian Foster 提交于 
      mainline-inclusion
from mainline-v5.15-rc4
commit 5ca5916b
category: bugfix
bugzilla: https://gitee.com/openeuler/kernel/issues/I4KIAO
CVE: NA

Reference: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=5ca5916b6bc93577c360c06cb7cdf71adb9b5faf

-------------------------------------------------

If writeback I/O to a COW extent fails, the COW fork blocks are
punched out and the data fork blocks left alone. It is possible for
COW fork blocks to overlap non-shared data fork blocks (due to
cowextsz hint prealloc), however, and writeback unconditionally maps
to the COW fork whenever blocks exist at the corresponding offset of
the page undergoing writeback. This means it's quite possible for a
COW fork extent to overlap delalloc data fork blocks, writeback to
convert and map to the COW fork blocks, writeback to fail, and
finally for ioend completion to cancel the COW fork blocks and leave
stale data fork delalloc blocks around in the inode. The blocks are
effectively stale because writeback failure also discards dirty page
state.

If this occurs, it is likely to trigger assert failures, free space
accounting corruption and failures in unrelated file operations. For
example, a subsequent reflink attempt of the affected file to a new
target file will trip over the stale delalloc in the source file and
fail. Several of these issues are occasionally reproduced by
generic/648, but are reproducible on demand with the right sequence
of operations and timely I/O error injection.

To fix this problem, update the ioend failure path to also punch out
underlying data fork delalloc blocks on I/O error. This is analogous
to the writeback submission failure path in xfs_discard_page() where
we might fail to map data fork delalloc blocks and consistent with
the successful COW writeback completion path, which is responsible
for unmapping from the data fork and remapping in COW fork blocks.

Fixes: 787eb485 ("xfs: fix and streamline error handling in xfs_end_io")
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDarrick J. Wong <djwong@kernel.org>
Signed-off-by: NDarrick J. Wong <djwong@kernel.org>
Signed-off-by: NGuo Xuenan <guoxuenan@huawei.com>
Reviewed-by: NLihong Kou <koulihong@huawei.com>
Reviewed-by: NZhang Yi <yi.zhang@huawei.com>
Signed-off-by: NZheng Zengkai <zhengzengkai@huawei.com>
      22987db8
    • B
      xfs: drop submit side trans alloc for append ioends · b906e741
      由 Brian Foster 提交于 
      mainline-inclusion
from mainline-v5.12-rc4
commit 7cd3099f
category: bugfix
bugzilla: https://gitee.com/openeuler/kernel/issues/I4KIAO
CVE: NA

Reference: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=7cd3099f4925d7c15887d1940ebd65acd66100f5

-------------------------------------------------

Per-inode ioend completion batching has a log reservation deadlock
vector between preallocated append transactions and transactions
that are acquired at completion time for other purposes (i.e.,
unwritten extent conversion or COW fork remaps). For example, if the
ioend completion workqueue task executes on a batch of ioends that
are sorted such that an append ioend sits at the tail, it's possible
for the outstanding append transaction reservation to block
allocation of transactions required to process preceding ioends in
the list.

Append ioend completion is historically the common path for on-disk
inode size updates. While file extending writes may have completed
sometime earlier, the on-disk inode size is only updated after
successful writeback completion. These transactions are preallocated
serially from writeback context to mitigate concurrency and
associated log reservation pressure across completions processed by
multi-threaded workqueue tasks.

However, now that delalloc blocks unconditionally map to unwritten
extents at physical block allocation time, size updates via append
ioends are relatively rare. This means that inode size updates most
commonly occur as part of the preexisting completion time
transaction to convert unwritten extents. As a result, there is no
longer a strong need to preallocate size update transactions.

Remove the preallocation of inode size update transactions to avoid
the ioend completion processing log reservation deadlock. Instead,
continue to send all potential size extending ioends to workqueue
context for completion and allocate the transaction from that
context. This ensures that no outstanding log reservation is owned
by the ioend completion worker task when it begins to process
ioends.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDarrick J. Wong <djwong@kernel.org>
Signed-off-by: NDarrick J. Wong <djwong@kernel.org>
Signed-off-by: NGuo Xuenan <guoxuenan@huawei.com>
Reviewed-by: NLihong Kou <koulihong@huawei.com>
Reviewed-by: NZhang Yi <yi.zhang@huawei.com>
Signed-off-by: NZheng Zengkai <zhengzengkai@huawei.com>
      b906e741
  7. 05 11月, 2020 2 次提交
    • D
      xfs: fix missing CoW blocks writeback conversion retry · c2f09217
      由 Darrick J. Wong 提交于 
      In commit 7588cbee, we tried to fix a race stemming from the lack of
coordination between higher level code that wants to allocate and remap
CoW fork extents into the data fork.  Christoph cites as examples the
always_cow mode, and a directio write completion racing with writeback.

According to the comments before the goto retry, we want to restart the
lookup to catch the extent in the data fork, but we don't actually reset
whichfork or cow_fsb, which means the second try executes using stale
information.  Up until now I think we've gotten lucky that either
there's something left in the CoW fork to cause cow_fsb to be reset, or
either data/cow fork sequence numbers have advanced enough to force a
fresh lookup from the data fork.  However, if we reach the retry with an
empty stable CoW fork and a stable data fork, neither of those things
happens.  The retry foolishly re-calls xfs_convert_blocks on the CoW
fork which fails again.  This time, we toss the write.

I've recently been working on extending reflink to the realtime device.
When the realtime extent size is larger than a single block, we have to
force the page cache to CoW the entire rt extent if a write (or
fallocate) are not aligned with the rt extent size.  The strategy I've
chosen to deal with this is derived from Dave's blocksize > pagesize
series: dirtying around the write range, and ensuring that writeback
always starts mapping on an rt extent boundary.  This has brought this
race front and center, since generic/522 blows up immediately.

However, I'm pretty sure this is a bug outright, independent of that.

Fixes: 7588cbee ("xfs: retry COW fork delalloc conversion when no extent was found")
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
      c2f09217
    • B
      iomap: support partial page discard on writeback block mapping failure · 763e4cdc
      由 Brian Foster 提交于 
      iomap writeback mapping failure only calls into ->discard_page() if
the current page has not been added to the ioend. Accordingly, the
XFS callback assumes a full page discard and invalidation. This is
problematic for sub-page block size filesystems where some portion
of a page might have been mapped successfully before a failure to
map a delalloc block occurs. ->discard_page() is not called in that
error scenario and the bio is explicitly failed by iomap via the
error return from ->prepare_ioend(). As a result, the filesystem
leaks delalloc blocks and corrupts the filesystem block counters.

Since XFS is the only user of ->discard_page(), tweak the semantics
to invoke the callback unconditionally on mapping errors and provide
the file offset that failed to map. Update xfs_discard_page() to
discard the corresponding portion of the file and pass the range
along to iomap_invalidatepage(). The latter already properly handles
both full and sub-page scenarios by not changing any iomap or page
state on sub-page invalidations.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDarrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
      763e4cdc
  9. 21 9月, 2020 1 次提交
  11. 03 6月, 2020 1 次提交
  13. 20 5月, 2020 1 次提交
  15. 03 3月, 2020 1 次提交
  17. 04 1月, 2020 1 次提交
  19. 28 10月, 2019 1 次提交
  21. 22 10月, 2019 1 次提交
  23. 21 10月, 2019 6 次提交
  25. 01 7月, 2019 5 次提交
  27. 29 6月, 2019 3 次提交
  29. 17 6月, 2019 1 次提交
  31. 30 4月, 2019 1 次提交
  33. 17 4月, 2019 2 次提交
    • D
      xfs: merge adjacent io completions of the same type · 3994fc48
      由 Darrick J. Wong 提交于 
      It's possible for pagecache writeback to split up a large amount of work
into smaller pieces for throttling purposes or to reduce the amount of
time a writeback operation is pending.  Whatever the reason, XFS can end
up with a bunch of IO completions that call for the same operation to be
performed on a contiguous extent mapping.  Since mappings are extent
based in XFS, we'd prefer to run fewer transactions when we can.

When we're processing an ioend on the list of io completions, check to
see if the next items on the list are both adjacent and of the same
type.  If so, we can merge the completions to reduce transaction
overhead.

On fast storage this doesn't seem to make much of a difference in
performance, though the number of transactions for an overnight xfstests
run seems to drop by ~5%.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
      3994fc48
    • D
      xfs: implement per-inode writeback completion queues · cb357bf3
      由 Darrick J. Wong 提交于 
      When scheduling writeback of dirty file data in the page cache, XFS uses
IO completion workqueue items to ensure that filesystem metadata only
updates after the write completes successfully.  This is essential for
converting unwritten extents to real extents at the right time and
performing COW remappings.

Unfortunately, XFS queues each IO completion work item to an unbounded
workqueue, which means that the kernel can spawn dozens of threads to
try to handle the items quickly.  These threads need to take the ILOCK
to update file metadata, which results in heavy ILOCK contention if a
large number of the work items target a single file, which is
inefficient.

Worse yet, the writeback completion threads get stuck waiting for the
ILOCK while holding transaction reservations, which can use up all
available log reservation space.  When that happens, metadata updates to
other parts of the filesystem grind to a halt, even if the filesystem
could otherwise have handled it.

Even worse, if one of the things grinding to a halt happens to be a
thread in the middle of a defer-ops finish holding the same ILOCK and
trying to obtain more log reservation having exhausted the permanent
reservation, we now have an ABBA deadlock - writeback completion has a
transaction reserved and wants the ILOCK, and someone else has the ILOCK
and wants a transaction reservation.

Therefore, we create a per-inode writeback io completion queue + work
item.  When writeback finishes, it can add the ioend to the per-inode
queue and let the single worker item process that queue.  This
dramatically cuts down on the number of kworkers and ILOCK contention in
the system, and seems to have eliminated an occasional deadlock I was
seeing while running generic/476.

Testing with a program that simulates a heavy random-write workload to a
single file demonstrates that the number of kworkers drops from
approximately 120 threads per file to 1, without dramatically changing
write bandwidth or pagecache access latency.

Note that we leave the xfs-conv workqueue's max_active alone because we
still want to be able to run ioend processing for as many inodes as the
system can handle.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
      cb357bf3
  35. 21 2月, 2019 2 次提交
    • C
      xfs: introduce an always_cow mode · 66ae56a5
      由 Christoph Hellwig 提交于 
      Add a mode where XFS never overwrites existing blocks in place.  This
is to aid debugging our COW code, and also put infatructure in place
for things like possible future support for zoned block devices, which
can't support overwrites.

This mode is enabled globally by doing a:

    echo 1 > /sys/fs/xfs/debug/always_cow

Note that the parameter is global to allow running all tests in xfstests
easily in this mode, which would not easily be possible with a per-fs
sysfs file.

In always_cow mode persistent preallocations are disabled, and fallocate
will fail when called with a 0 mode (with our without
FALLOC_FL_KEEP_SIZE), and not create unwritten extent for zeroed space
when called with FALLOC_FL_ZERO_RANGE or FALLOC_FL_UNSHARE_RANGE.

There are a few interesting xfstests failures when run in always_cow
mode:

 - generic/392 fails because the bytes used in the file used to test
   hole punch recovery are less after the log replay.  This is
   because the blocks written and then punched out are only freed
   with a delay due to the logging mechanism.
 - xfs/170 will fail as the already fragile file streams mechanism
   doesn't seem to interact well with the COW allocator
 - xfs/180 xfs/182 xfs/192 xfs/198 xfs/204 and xfs/208 will claim
   the file system is badly fragmented, but there is not much we
   can do to avoid that when always writing out of place
 - xfs/205 fails because overwriting a file in always_cow mode
   will require new space allocation and the assumption in the
   test thus don't work anymore.
 - xfs/326 fails to modify the file at all in always_cow mode after
   injecting the refcount error, leading to an unexpected md5sum
   after the remount, but that again is expected
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDarrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
      66ae56a5
    • C
      xfs: also truncate holes covered by COW blocks · 12df89f2
      由 Christoph Hellwig 提交于 
      This only matters if we want to write data through the COW fork that is
not actually an overwrite of existing data.  Reasons for that are
speculative COW fork allocations using the cowextsize, or a mode where
we always write through the COW fork.  Currently both can't actually
happen, but I plan to enable them.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDarrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
      12df89f2
  37. 18 2月, 2019 5 次提交
  39. 15 2月, 2019 2 次提交