Age | Commit message (Collapse) | Author |
|
These functions return the maximum number of blocks that could be logged
in a particular transaction. "log count" is confusing since there's a
separate concept of a log (operation) count in the reservation code, so
let's change it to "block count" to be less confusing.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Currently, the code that performs CoW remapping after a write has this
odd behavior where it walks /backwards/ through the data fork to remap
extents in reverse order. Earlier, we rewrote the reflink remap
function to use deferred bmap log items instead of trying to cram as
much into the first transaction that we could. Now do the same for the
CoW remap code. There doesn't seem to be any performance impact; we're
just making better use of code that we added for the benefit of reflink.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Before to the introduction of deferred refcount operations, reflink
would try to cram refcount btree updates into the same transaction as an
allocation or a free event. Mainline XFS has never actually done that,
but we never refactored the transaction reservations to reflect that we
now do all refcount updates in separate transactions. Fix this to
reduce the transaction reservation size even farther, so that between
this patch and the previous one, we reduce the tr_write and tr_itruncate
sizes by 66%.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Back in the early days of reflink and rmap development I set the
transaction reservation sizes to be overly generous for rmap+reflink
filesystems, and a little under-generous for rmap-only filesystems.
Since we don't need *eight* transaction rolls to handle three new log
intent items, decrease the logcounts to what we actually need, and amend
the shadow reservation computation function to reflect what we used to
do so that the minimum log size doesn't change.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Move the tracepoint that computes the size of the transaction used to
compute the minimum log size into xfs_log_get_max_trans_res so that we
only have to compute this stuff once.
Leave xfs_log_get_max_trans_res as a non-static function so that xfs_db
can call it to report the results of the userspace computation of the
same value to diagnose mkfs/kernel misinteractions.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Every time someone changes the transaction reservation sizes, they
introduce potential compatibility problems if the changes affect the
minimum log size that we validate at mount time. If the minimum log
size gets larger (which should be avoided because doing so presents a
serious risk of log livelock), filesystems created with old mkfs will
not mount on a newer kernel; if the minimum size shrinks, filesystems
created with newer mkfs will not mount on older kernels.
Therefore, enable the creation of a shadow log reservation structure
where we can "undo" the effects of tweaks when computing minimum log
sizes. These shadow reservations should never be used in practice, but
they insulate us from perturbations in minimum log size.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
This raw call isn't necessary since we can always remove a full delalloc
extent.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
|
|
In commit e1a4e37cc7b6, we clamped the length of bunmapi calls on the
data forks of shared files to avoid two failure scenarios: one where the
extent being unmapped is so sparsely shared that we exceed the
transaction reservation with the sheer number of refcount btree updates
and EFI intent items; and the other where we attach so many deferred
updates to the transaction that we pin the log tail and later the log
head meets the tail, causing the log to livelock.
We avoid triggering the first problem by tracking the number of ops in
the refcount btree cursor and forcing a requeue of the refcount intent
item any time we think that we might be close to overflowing. This has
been baked into XFS since before the original e1a4 patch.
A recent patchset fixed the second problem by changing the deferred ops
code to finish all the work items created by each round of trying to
complete a refcount intent item, which eliminates the long chains of
deferred items (27dad); and causing long-running transactions to relog
their intent log items when space in the log gets low (74f4d).
Because this clamp affects /any/ unmapping request regardless of the
sharing factors of the component blocks, it degrades the performance of
all large unmapping requests -- whereas with an unshared file we can
unmap millions of blocks in one go, shared files are limited to
unmapping a few thousand blocks at a time, which causes the upper level
code to spin in a bunmapi loop even if it wasn't needed.
This also eliminates one more place where log recovery behavior can
differ from online behavior, because bunmapi operations no longer need
to requeue. The fstest generic/447 was created to test the old fix, and
it still passes with this applied.
Partial-revert-of: e1a4e37cc7b6 ("xfs: try to avoid blowing out the transaction reservation when bunmaping a shared extent")
Depends: 27dada070d59 ("xfs: change the order in which child and parent defer ops ar finished")
Depends: 74f4d6a1e065 ("xfs: only relog deferred intent items if free space in the log gets low")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
A long time ago, I added to XFS the ability to use deferred reference
count operations as part of a transaction chain. This enabled us to
avoid blowing out the transaction reservation when the blocks in a
physical extent all had different reference counts because we could ask
the deferred operation manager for a continuation, which would get us a
clean transaction.
The refcount code asks for a continuation when the number of refcount
record updates reaches the point where we think that the transaction has
logged enough full btree blocks due to refcount (and free space) btree
shape changes and refcount record updates that we're in danger of
overflowing the transaction.
We did not previously count the EFIs logged to the refcount update
transaction because the clamps on the length of a bunmap operation were
sufficient to avoid overflowing the transaction reservation even in the
worst case situation where every other block of the unmapped extent is
shared.
Unfortunately, the restrictions on bunmap length avoid failure in the
worst case by imposing a maximum unmap length of ~3000 blocks, even for
non-pathological cases. This seriously limits performance when freeing
large extents.
Therefore, track EFIs with the same counter as refcount record updates,
and use that information as input into when we should ask for a
continuation. This enables the next patch to drop the clumsy bunmap
limitation.
Depends: 27dada070d59 ("xfs: change the order in which child and parent defer ops ar finished")
Depends: 74f4d6a1e065 ("xfs: only relog deferred intent items if free space in the log gets low")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Reverse mapping on a reflink-capable filesystem has some pretty high
overhead when performing file operations. This is because the rmap
records for logically and physically adjacent extents might not be
adjacent in the rmap index due to data block sharing. As a result, we
use expensive overlapped-interval btree search, which walks every record
that overlaps with the supplied key in the hopes of finding the record.
However, profiling data shows that when the index contains a record that
is an exact match for a query key, the non-overlapped btree search
function can find the record much faster than the overlapped version.
Try the non-overlapped lookup first when we're trying to find the left
neighbor rmap record for a given file mapping, which makes unwritten
extent conversion and remap operations run faster if data block sharing
is minimal in this part of the filesystem.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Reverse mapping on a reflink-capable filesystem has some pretty high
overhead when performing file operations. This is because the rmap
records for logically and physically adjacent extents might not be
adjacent in the rmap index due to data block sharing. As a result, we
use expensive overlapped-interval btree search, which walks every record
that overlaps with the supplied key in the hopes of finding the record.
However, profiling data shows that when the index contains a record that
is an exact match for a query key, the non-overlapped btree search
function can find the record much faster than the overlapped version.
Try the non-overlapped lookup first, which will make scrub run much
faster.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Most callers of xfs_rmap_lookup_le will retrieve the btree record
immediately if the lookup succeeds. The overlapped version of this
function (xfs_rmap_lookup_le_range) will return the record if the lookup
succeeds, so make the regular version do it too. Get rid of the useless
len argument, since it's not part of the lookup key.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
|
|
Record the buffer ops in the xfs_buf tracepoints so that we can monitor
the alleged type of the buffer.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
As mentioned in the previous commit, the kernel misuses sb_frextents in
the incore mount to reflect both incore reservations made by running
transactions as well as the actual count of free rt extents on disk.
This results in the superblock being written to the log with an
underestimate of the number of rt extents that are marked free in the
rtbitmap.
Teaching XFS to recompute frextents after log recovery avoids
operational problems in the current mount, but it doesn't solve the
problem of us writing undercounted frextents which are then recovered by
an older kernel that doesn't have that fix.
Create an incore percpu counter to mirror the ondisk frextents. This
new counter will track transaction reservations and the only time we
will touch the incore super counter (i.e the one that gets logged) is
when those transactions commit updates to the rt bitmap. This is in
contrast to the lazysbcount counters (e.g. fdblocks), where we know that
log recovery will always fix any incorrect counter that we log.
As a bonus, we only take m_sb_lock at transaction commit time.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
I've been observing periodic corruption reports from xfs_scrub involving
the free rt extent counter (frextents) while running xfs/141. That test
uses an error injection knob to induce a torn write to the log, and an
arbitrary number of recovery mounts, frextents will count fewer free rt
extents than can be found the rtbitmap.
The root cause of the problem is a combination of the misuse of
sb_frextents in the incore mount to reflect both incore reservations
made by running transactions as well as the actual count of free rt
extents on disk. The following sequence can reproduce the undercount:
Thread 1 Thread 2
xfs_trans_alloc(rtextents=3)
xfs_mod_frextents(-3)
<blocks>
xfs_attr_set()
xfs_bmap_attr_addfork()
xfs_add_attr2()
xfs_log_sb()
xfs_sb_to_disk()
xfs_trans_commit()
<log flushed to disk>
<log goes down>
Note that thread 1 subtracts 3 from sb_frextents even though it never
commits to using that space. Thread 2 writes the undercounted value to
the ondisk superblock and logs it to the xattr transaction, which is
then flushed to disk. At next mount, log recovery will find the logged
superblock and write that back into the filesystem. At the end of log
recovery, we reread the superblock and install the recovered
undercounted frextents value into the incore superblock. From that
point on, we've effectively leaked thread 1's transaction reservation.
The correct fix for this is to separate the incore reservation from the
ondisk usage, but that's a matter for the next patch. Because the
kernel has been logging superblocks with undercounted frextents for a
very long time and we don't demand that sysadmins run xfs_repair after a
crash, fix the undercount by recomputing frextents after log recovery.
Gating this on log recovery is a reasonable balance (I think) between
correcting the problem and slowing down every mount attempt. Note that
xfs_repair will fix undercounted frextents.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Pass an explicit xfs_mount pointer to the rtalloc query functions so
that they can support transactionless queries.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Since the initial introduction of (posix) fallocate back at the turn of
the century, it has been possible to use this syscall to change the
user-visible contents of files. This can happen by extending the file
size during a preallocation, or through any of the newer modes (punch,
zero range). Because the call can be used to change file contents, we
should treat it like we do any other modification to a file -- update
the mtime, and drop set[ug]id privileges/capabilities.
The VFS function file_modified() does all this for us if pass it a
locked inode, so let's make fallocate drop permissions correctly.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Since the initial introduction of (posix) fallocate back at the turn of
the century, it has been possible to use this syscall to change the
user-visible contents of files. This can happen by extending the file
size during a preallocation, or through any of the newer modes (punch,
zero, collapse, insert range). Because the call can be used to change
file contents, we should treat it like we do any other modification to a
file -- update the mtime, and drop set[ug]id privileges/capabilities.
The VFS function file_modified() does all this for us if pass it a
locked inode, so let's make fallocate drop permissions correctly.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
when get fiemap starting from MAX_LFS_FILESIZE, (maxbytes - *len) < start
will always true , then *len set zero. because of start offset is beyond
file size, for erofs filesystem it will always return iomap.length with
zero,iomap iterate will enter infinite loop. it is necessary cover this
corner case to avoid this situation.
------------[ cut here ]------------
WARNING: CPU: 7 PID: 905 at fs/iomap/iter.c:35 iomap_iter+0x97f/0xc70
Modules linked in: xfs erofs
CPU: 7 PID: 905 Comm: iomap Tainted: G W 5.17.0-rc8 #27
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.13.0-1ubuntu1.1 04/01/2014
RIP: 0010:iomap_iter+0x97f/0xc70
Code: 85 a1 fc ff ff e8 71 be 9c ff 0f 1f 44 00 00 e9 92 fc ff ff e8 62 be 9c ff 0f 0b b8 fb ff ff ff e9 fc f8 ff ff e8 51 be 9c ff <0f> 0b e9 2b fc ff ff e8 45 be 9c ff 0f 0b e9 e1 fb ff ff e8 39 be
RSP: 0018:ffff888060a37ab0 EFLAGS: 00010293
RAX: 0000000000000000 RBX: ffff888060a37bb0 RCX: 0000000000000000
RDX: ffff88807e19a900 RSI: ffffffff81a7da7f RDI: ffff888060a37be0
RBP: 7fffffffffffffff R08: 0000000000000000 R09: ffff888060a37c20
R10: ffff888060a37c67 R11: ffffed100c146f8c R12: 7fffffffffffffff
R13: 0000000000000000 R14: ffff888060a37bd8 R15: ffff888060a37c20
FS: 00007fd3cca01540(0000) GS:ffff888108780000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000000020010820 CR3: 0000000054b92000 CR4: 00000000000006e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Call Trace:
<TASK>
iomap_fiemap+0x1c9/0x2f0
erofs_fiemap+0x64/0x90 [erofs]
do_vfs_ioctl+0x40d/0x12e0
__x64_sys_ioctl+0xaa/0x1c0
do_syscall_64+0x35/0x80
entry_SYSCALL_64_after_hwframe+0x44/0xae
</TASK>
---[ end trace 0000000000000000 ]---
watchdog: BUG: soft lockup - CPU#7 stuck for 26s! [iomap:905]
Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: Guo Xuenan <guoxuenan@huawei.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
[djwong: fix some typos]
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
The O_TMPFILE creation implementation creates a specific order of
operations for inode allocation/freeing and unlinked list
modification. Currently both are serialised by the AGI, so the order
doesn't strictly matter as long as the are both in the same
transaction.
However, if we want to move the unlinked list insertions largely out
from under the AGI lock, then we have to be concerned about the
order in which we do unlinked list modification operations.
O_TMPFILE creation tells us this order is inode allocation/free,
then unlinked list modification.
Change xfs_ifree() to use this same ordering on unlinked list
removal. This way we always guarantee that when we enter the
iunlinked list removal code from this path, we already have the AGI
locked and we don't have to worry about lock nesting AGI reads
inside unlink list locks because it's already locked and attached to
the transaction.
We can do this safely as the inode freeing and unlinked list removal
are done in the same transaction and hence are atomic operations
with respect to log recovery.
Reported-by: Frank Hofmann <fhofmann@cloudflare.com>
Fixes: 298f7bec503f ("xfs: pin inode backing buffer to the inode log item")
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
|
|
Jan Kara reported a performance regression in dbench that he
bisected down to commit bad77c375e8d ("xfs: CIL checkpoint
flushes caches unconditionally").
Whilst developing the journal flush/fua optimisations this cache was
part of, it appeared to made a significant difference to
performance. However, now that this patchset has settled and all the
correctness issues fixed, there does not appear to be any
significant performance benefit to asynchronous cache flushes.
In fact, the opposite is true on some storage types and workloads,
where additional cache flushes that can occur from fsync heavy
workloads have measurable and significant impact on overall
throughput.
Local dbench testing shows little difference on dbench runs with
sync vs async cache flushes on either fast or slow SSD storage, and
no difference in streaming concurrent async transaction workloads
like fs-mark.
Fast NVME storage.
From `dbench -t 30`, CIL scale:
clients async sync
BW Latency BW Latency
1 935.18 0.855 915.64 0.903
8 2404.51 6.873 2341.77 6.511
16 3003.42 6.460 2931.57 6.529
32 3697.23 7.939 3596.28 7.894
128 7237.43 15.495 7217.74 11.588
512 5079.24 90.587 5167.08 95.822
fsmark, 32 threads, create w/ 64 byte xattr w/32k logbsize
create chown unlink
async 1m41s 1m16s 2m03s
sync 1m40s 1m19s 1m54s
Slower SATA SSD storage:
From `dbench -t 30`, CIL scale:
clients async sync
BW Latency BW Latency
1 78.59 15.792 83.78 10.729
8 367.88 92.067 404.63 59.943
16 564.51 72.524 602.71 76.089
32 831.66 105.984 870.26 110.482
128 1659.76 102.969 1624.73 91.356
512 2135.91 223.054 2603.07 161.160
fsmark, 16 threads, create w/32k logbsize
create unlink
async 5m06s 4m15s
sync 5m00s 4m22s
And on Jan's test machine:
5.18-rc8-vanilla 5.18-rc8-patched
Amean 1 71.22 ( 0.00%) 64.94 * 8.81%*
Amean 2 93.03 ( 0.00%) 84.80 * 8.85%*
Amean 4 150.54 ( 0.00%) 137.51 * 8.66%*
Amean 8 252.53 ( 0.00%) 242.24 * 4.08%*
Amean 16 454.13 ( 0.00%) 439.08 * 3.31%*
Amean 32 835.24 ( 0.00%) 829.74 * 0.66%*
Amean 64 1740.59 ( 0.00%) 1686.73 * 3.09%*
Performance and cache flush behaviour is restored to pre-regression
levels.
As such, we can now consider the async cache flush mechanism an
unnecessary exercise in premature optimisation and hence we can
now remove it and the infrastructure it requires completely.
Fixes: bad77c375e8d ("xfs: CIL checkpoint flushes caches unconditionally")
Reported-and-tested-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
When a checkpoint writeback is run by log recovery, corruption
propagated from the log can result in writeback verifiers failing
and calling xfs_force_shutdown() from
xfs_buf_delwri_submit_buffers().
This results in the mount being marked as shutdown, but the log does
not get marked as shut down because:
/*
* If this happens during log recovery then we aren't using the runtime
* log mechanisms yet so there's nothing to shut down.
*/
if (!log || xlog_in_recovery(log))
return false;
If there are other buffers that then fail (say due to detecting the
mount shutdown), they will now hang in xfs_do_force_shutdown()
waiting for the log to shut down like this:
__schedule+0x30d/0x9e0
schedule+0x55/0xd0
xfs_do_force_shutdown+0x1cd/0x200
? init_wait_var_entry+0x50/0x50
xfs_buf_ioend+0x47e/0x530
__xfs_buf_submit+0xb0/0x240
xfs_buf_delwri_submit_buffers+0xfe/0x270
xfs_buf_delwri_submit+0x3a/0xc0
xlog_do_recovery_pass+0x474/0x7b0
? do_raw_spin_unlock+0x30/0xb0
xlog_do_log_recovery+0x91/0x140
xlog_do_recover+0x38/0x1e0
xlog_recover+0xdd/0x170
xfs_log_mount+0x17e/0x2e0
xfs_mountfs+0x457/0x930
xfs_fs_fill_super+0x476/0x830
xlog_force_shutdown() always needs to mark the log as shut down,
regardless of whether recovery is in progress or not, so that
multiple calls to xfs_force_shutdown() during recovery don't end
up waiting for the log to be shut down like this.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
If a shut races with xfs_trans_commit() and we have shut down the
filesystem but not the log, we will still cancel the transaction.
This can result in aborting dirty log items instead of committing and
pinning them whilst the log is still running. Hence we can end up
with dirty, unlogged metadata that isn't in the AIL in memory that
can be flushed to disk via writeback clustering.
This was discovered from a g/388 trace where an inode log item was
having IO completed on it and it wasn't in the AIL, hence tripping
asserts xfs_ail_check(). Inode cluster writeback started long after
the filesystem shutdown started, and long after the transaction
containing the dirty inode was aborted and the log item marked
XFS_LI_ABORTED. The inode was seen as dirty and unpinned, so it
was flushed. IO completion tried to remove the inode from the AIL,
at which point stuff went bad:
XFS (pmem1): Log I/O Error (0x6) detected at xfs_fs_goingdown+0xa3/0xf0 (fs/xfs/xfs_fsops.c:500). Shutting down filesystem.
XFS: Assertion failed: in_ail, file: fs/xfs/xfs_trans_ail.c, line: 67
XFS (pmem1): Please unmount the filesystem and rectify the problem(s)
Workqueue: xfs-buf/pmem1 xfs_buf_ioend_work
RIP: 0010:assfail+0x27/0x2d
Call Trace:
<TASK>
xfs_ail_check+0xa8/0x180
xfs_ail_delete_one+0x3b/0xf0
xfs_buf_inode_iodone+0x329/0x3f0
xfs_buf_ioend+0x1f8/0x530
xfs_buf_ioend_work+0x15/0x20
process_one_work+0x1ac/0x390
worker_thread+0x56/0x3c0
kthread+0xf6/0x120
ret_from_fork+0x1f/0x30
</TASK>
xfs_trans_commit() needs to check log state for shutdown, not mount
state. It cannot abort dirty log items while the log is still
running as dirty items must remained pinned in memory until they are
either committed to the journal or the log has shut down and they
can be safely tossed away. Hence if the log has not shut down, the
xfs_trans_commit() path must allow completed transactions to commit
to the CIL and pin the dirty items even if a mount shutdown has
started.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
When we call xfs_forced_shutdown(), the caller often expects the
filesystem to be completely shut down when it returns. However,
if we have racing xfs_forced_shutdown() calls, the first caller sets
the mount shutdown flag then goes to shutdown the log. The second
caller sees the mount shutdown flag and returns immediately - it
does not wait for the log to be shut down.
Unfortunately, xfs_forced_shutdown() is used in some places that
expect it to completely shut down the filesystem before it returns
(e.g. xfs_trans_log_inode()). As such, returning before the log has
been shut down leaves us in a place where the transaction failed to
complete correctly but we still call xfs_trans_commit(). This
situation arises because xfs_trans_log_inode() does not return an
error and instead calls xfs_force_shutdown() to ensure that the
transaction being committed is aborted.
Unfortunately, we have a race condition where xfs_trans_commit()
needs to check xlog_is_shutdown() because it can't abort log items
before the log is shut down, but it needs to use xfs_is_shutdown()
because xfs_forced_shutdown() does not block waiting for the log to
shut down.
To fix this conundrum, first we make all calls to
xfs_forced_shutdown() block until the log is also shut down. This
means we can then safely use xfs_forced_shutdown() as a mechanism
that ensures the currently running transaction will be aborted by
xfs_trans_commit() regardless of the shutdown check it uses.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
We've got a mess on our hands.
1. xfs_trans_commit() cannot cancel transactions because the mount is
shut down - that causes dirty, aborted, unlogged log items to sit
unpinned in memory and potentially get written to disk before the
log is shut down. Hence xfs_trans_commit() can only abort
transactions when xlog_is_shutdown() is true.
2. xfs_force_shutdown() is used in places to cause the current
modification to be aborted via xfs_trans_commit() because it may be
impractical or impossible to cancel the transaction directly, and
hence xfs_trans_commit() must cancel transactions when
xfs_is_shutdown() is true in this situation. But we can't do that
because of #1.
3. Log IO errors cause log shutdowns by calling xfs_force_shutdown()
to shut down the mount and then the log from log IO completion.
4. xfs_force_shutdown() can result in a log force being issued,
which has to wait for log IO completion before it will mark the log
as shut down. If #3 races with some other shutdown trigger that runs
a log force, we rely on xfs_force_shutdown() silently ignoring #3
and avoiding shutting down the log until the failed log force
completes.
5. To ensure #2 always works, we have to ensure that
xfs_force_shutdown() does not return until the the log is shut down.
But in the case of #4, this will result in a deadlock because the
log Io completion will block waiting for a log force to complete
which is blocked waiting for log IO to complete....
So the very first thing we have to do here to untangle this mess is
dissociate log shutdown triggers from mount shutdowns. We already
have xlog_forced_shutdown, which will atomically transistion to the
log a shutdown state. Due to internal asserts it cannot be called
multiple times, but was done simply because the only place that
could call it was xfs_do_force_shutdown() (i.e. the mount shutdown!)
and that could only call it once and once only. So the first thing
we do is remove the asserts.
We then convert all the internal log shutdown triggers to call
xlog_force_shutdown() directly instead of xfs_force_shutdown(). This
allows the log shutdown triggers to shut down the log without
needing to care about mount based shutdown constraints. This means
we shut down the log independently of the mount and the mount may
not notice this until it's next attempt to read or modify metadata.
At that point (e.g. xfs_trans_commit()) it will see that the log is
shutdown, error out and shutdown the mount.
To ensure that all the unmount behaviours and asserts track
correctly as a result of a log shutdown, propagate the shutdown up
to the mount if it is not already set. This keeps the mount and log
state in sync, and saves a huge amount of hassle where code fails
because of a log shutdown but only checks for mount shutdowns and
hence ends up doing the wrong thing. Cleaning up that mess is
an exercise for another day.
This enables us to address the other problems noted above in
followup patches.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Brian reported a null pointer dereference failure during unmount in
xfs/006. He tracked the problem down to the AIL being torn down
before a log shutdown had completed and removed all the items from
the AIL. The failure occurred in this path while unmount was
proceeding in another task:
xfs_trans_ail_delete+0x102/0x130 [xfs]
xfs_buf_item_done+0x22/0x30 [xfs]
xfs_buf_ioend+0x73/0x4d0 [xfs]
xfs_trans_committed_bulk+0x17e/0x2f0 [xfs]
xlog_cil_committed+0x2a9/0x300 [xfs]
xlog_cil_process_committed+0x69/0x80 [xfs]
xlog_state_shutdown_callbacks+0xce/0xf0 [xfs]
xlog_force_shutdown+0xdf/0x150 [xfs]
xfs_do_force_shutdown+0x5f/0x150 [xfs]
xlog_ioend_work+0x71/0x80 [xfs]
process_one_work+0x1c5/0x390
worker_thread+0x30/0x350
kthread+0xd7/0x100
ret_from_fork+0x1f/0x30
This is processing an EIO error to a log write, and it's
triggering a force shutdown. This causes the log to be shut down,
and then it is running attached iclog callbacks from the shutdown
context. That means the fs and log has already been marked as
xfs_is_shutdown/xlog_is_shutdown and so high level code will abort
(e.g. xfs_trans_commit(), xfs_log_force(), etc) with an error
because of shutdown.
The umount would have been blocked waiting for a log force
completion inside xfs_log_cover() -> xfs_sync_sb(). The first thing
for this situation to occur is for xfs_sync_sb() to exit without
waiting for the iclog buffer to be comitted to disk. The
above trace is the completion routine for the iclog buffer, and
it is shutting down the filesystem.
xlog_state_shutdown_callbacks() does this:
{
struct xlog_in_core *iclog;
LIST_HEAD(cb_list);
spin_lock(&log->l_icloglock);
iclog = log->l_iclog;
do {
if (atomic_read(&iclog->ic_refcnt)) {
/* Reference holder will re-run iclog callbacks. */
continue;
}
list_splice_init(&iclog->ic_callbacks, &cb_list);
>>>>>> wake_up_all(&iclog->ic_write_wait);
>>>>>> wake_up_all(&iclog->ic_force_wait);
} while ((iclog = iclog->ic_next) != log->l_iclog);
wake_up_all(&log->l_flush_wait);
spin_unlock(&log->l_icloglock);
>>>>>> xlog_cil_process_committed(&cb_list);
}
This wakes any thread waiting on IO completion of the iclog (in this
case the umount log force) before shutdown processes all the pending
callbacks. That means the xfs_sync_sb() waiting on a sync
transaction in xfs_log_force() on iclog->ic_force_wait will get
woken before the callbacks attached to that iclog are run. This
results in xfs_sync_sb() returning an error, and so unmount unblocks
and continues to run whilst the log shutdown is still in progress.
Normally this is just fine because the force waiter has nothing to
do with AIL operations. But in the case of this unmount path, the
log force waiter goes on to tear down the AIL because the log is now
shut down and so nothing ever blocks it again from the wait point in
xfs_log_cover().
Hence it's a race to see who gets to the AIL first - the unmount
code or xlog_cil_process_committed() killing the superblock buffer.
To fix this, we just have to change the order of processing in
xlog_state_shutdown_callbacks() to run the callbacks before it wakes
any task waiting on completion of the iclog.
Reported-by: Brian Foster <bfoster@redhat.com>
Fixes: aad7272a9208 ("xfs: separate out log shutdown callback processing")
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
generic/388 triggered a failure in RUI recovery due to a corrupted
btree record and the system then locked up hard due to a subsequent
assert failure while holding a spinlock cancelling intents:
XFS (pmem1): Corruption of in-memory data (0x8) detected at xfs_do_force_shutdown+0x1a/0x20 (fs/xfs/xfs_trans.c:964). Shutting down filesystem.
XFS (pmem1): Please unmount the filesystem and rectify the problem(s)
XFS: Assertion failed: !xlog_item_is_intent(lip), file: fs/xfs/xfs_log_recover.c, line: 2632
Call Trace:
<TASK>
xlog_recover_cancel_intents.isra.0+0xd1/0x120
xlog_recover_finish+0xb9/0x110
xfs_log_mount_finish+0x15a/0x1e0
xfs_mountfs+0x540/0x910
xfs_fs_fill_super+0x476/0x830
get_tree_bdev+0x171/0x270
? xfs_init_fs_context+0x1e0/0x1e0
xfs_fs_get_tree+0x15/0x20
vfs_get_tree+0x24/0xc0
path_mount+0x304/0xba0
? putname+0x55/0x60
__x64_sys_mount+0x108/0x140
do_syscall_64+0x35/0x80
entry_SYSCALL_64_after_hwframe+0x44/0xae
Essentially, there's dirty metadata in the AIL from intent recovery
transactions, so when we go to cancel the remaining intents we assume
that all objects after the first non-intent log item in the AIL are
not intents.
This is not true. Intent recovery can log new intents to continue
the operations the original intent could not complete in a single
transaction. The new intents are committed before they are deferred,
which means if the CIL commits in the background they will get
inserted into the AIL at the head.
Hence if we shut down the filesystem while processing intent
recovery, the AIL may have new intents active at the current head.
Hence this check:
/*
* We're done when we see something other than an intent.
* There should be no intents left in the AIL now.
*/
if (!xlog_item_is_intent(lip)) {
#ifdef DEBUG
for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
ASSERT(!xlog_item_is_intent(lip));
#endif
break;
}
in both xlog_recover_process_intents() and
log_recover_cancel_intents() is simply not valid. It was valid back
when we only had EFI/EFD intents and didn't chain intents, but it
hasn't been valid ever since intent recovery could create and commit
new intents.
Given that crashing the mount task like this pretty much prevents
diagnosing what went wrong that lead to the initial failure that
triggered intent cancellation, just remove the checks altogether.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Most buffer io list operations are run with the bp->b_lock held, but
xfs_iflush_abort() can be called without the buffer lock being held
resulting in inodes being removed from the buffer list while other
list operations are occurring. This causes problems with corrupted
bp->b_io_list inode lists during filesystem shutdown, leading to
traversals that never end, double removals from the AIL, etc.
Fix this by passing the buffer to xfs_iflush_abort() if we have
it locked. If the inode is attached to the buffer, we're going to
have to remove it from the buffer list and we'd have to get the
buffer off the inode log item to do that anyway.
If we don't have a buffer passed in (e.g. from xfs_reclaim_inode())
then we can determine if the inode has a log item and if it is
attached to a buffer before we do anything else. If it does have an
attached buffer, we can lock it safely (because the inode has a
reference to it) and then perform the inode abort.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
On a modern filesystem, we don't allow userspace to allocate blocks for
data storage from the per-AG space reservations, the user-controlled
reservation pool that prevents ENOSPC in the middle of internal
operations, or the internal per-AG set-aside that prevents unwanted
filesystem shutdowns due to ENOSPC during a bmap btree split.
Since we now consider freespace btree blocks as unavailable for
allocation for data storage, we shouldn't report those blocks via statfs
either. This makes the numbers that we return via the statfs f_bavail
and f_bfree fields a more conservative estimate of actual free space.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Due to cycling of m_sb_lock, it's possible for multiple callers of
xfs_reserve_blocks to race at changing the pool size, subtracting blocks
from fdblocks, and actually putting it in the pool. The result of all
this is that we can overfill the reserve pool to hilarious levels.
xfs_mod_fdblocks, when called with a positive value, already knows how
to take freed blocks and either fill the reserve until it's full, or put
them in fdblocks. Use that instead of setting m_resblks_avail directly.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Nowadays, xfs_mod_fdblocks will always choose to fill the reserve pool
with freed blocks before adding to fdblocks. Therefore, we can change
the behavior of xfs_reserve_blocks slightly -- setting the target size
of the pool should always succeed, since a deficiency will eventually
be made up as blocks get freed.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Infinite loops in kernel code are scary. Calls to xfs_reserve_blocks
should be rare (people should just use the defaults!) so we really don't
need to try so hard. Simplify the logic here by removing the infinite
loop.
Cc: Brian Foster <bfoster@redhat.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
xfs_reserve_blocks controls the size of the user-visible free space
reserve pool. Given the difference between the current and requested
pool sizes, it will try to reserve free space from fdblocks. However,
the amount requested from fdblocks is also constrained by the amount of
space that we think xfs_mod_fdblocks will give us. If we forget to
subtract m_allocbt_blks before calling xfs_mod_fdblocks, it will will
return ENOSPC and we'll hang the kernel at mount due to the infinite
loop.
In commit fd43cf600cf6, we decided that xfs_mod_fdblocks should not hand
out the "free space" used by the free space btrees, because some portion
of the free space btrees hold in reserve space for future btree
expansion. Unfortunately, xfs_reserve_blocks' estimation of the number
of blocks that it could request from xfs_mod_fdblocks was not updated to
include m_allocbt_blks, so if space is extremely low, the caller hangs.
Fix this by creating a function to estimate the number of blocks that
can be reserved from fdblocks, which needs to exclude the set-aside and
m_allocbt_blks.
Found by running xfs/306 (which formats a single-AG 20MB filesystem)
with an fstests configuration that specifies a 1k blocksize and a
specially crafted log size that will consume 7/8 of the space (17920
blocks, specifically) in that AG.
Cc: Brian Foster <bfoster@redhat.com>
Fixes: fd43cf600cf6 ("xfs: set aside allocation btree blocks from block reservation")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
|
|
Currently, we use this undocumented macro to encode the minimum number
of blocks needed to replenish a completely empty AGFL when an AG is
nearly full. This has lead to confusion on the part of the maintainers,
so let's document what the value actually means, and move it to
xfs_alloc.c since it's not used outside of that module.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
I've been chasing a recent resurgence in generic/388 recovery
failure and/or corruption events. The events have largely been
uninitialised inode chunks being tripped over in log recovery
such as:
XFS (pmem1): User initiated shutdown received.
pmem1: writeback error on inode 12621949, offset 1019904, sector 12968096
XFS (pmem1): Log I/O Error (0x6) detected at xfs_fs_goingdown+0xa3/0xf0 (fs/xfs/xfs_fsops.c:500). Shutting down filesystem.
XFS (pmem1): Please unmount the filesystem and rectify the problem(s)
XFS (pmem1): Unmounting Filesystem
XFS (pmem1): Mounting V5 Filesystem
XFS (pmem1): Starting recovery (logdev: internal)
XFS (pmem1): bad inode magic/vsn daddr 8723584 #0 (magic=1818)
XFS (pmem1): Metadata corruption detected at xfs_inode_buf_verify+0x180/0x190, xfs_inode block 0x851c80 xfs_inode_buf_verify
XFS (pmem1): Unmount and run xfs_repair
XFS (pmem1): First 128 bytes of corrupted metadata buffer:
00000000: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 ................
00000010: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 ................
00000020: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 ................
00000030: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 ................
00000040: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 ................
00000050: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 ................
00000060: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 ................
00000070: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 ................
XFS (pmem1): metadata I/O error in "xlog_recover_items_pass2+0x52/0xc0" at daddr 0x851c80 len 32 error 117
XFS (pmem1): log mount/recovery failed: error -117
XFS (pmem1): log mount failed
There have been isolated random other issues, too - xfs_repair fails
because it finds some corruption in symlink blocks, rmap
inconsistencies, etc - but they are nowhere near as common as the
uninitialised inode chunk failure.
The problem has clearly happened at runtime before recovery has run;
I can see the ICREATE log item in the log shortly before the
actively recovered range of the log. This means the ICREATE was
definitely created and written to the log, but for some reason the
tail of the log has been moved past the ordered buffer log item that
tracks INODE_ALLOC buffers and, supposedly, prevents the tail of the
log moving past the ICREATE log item before the inode chunk buffer
is written to disk.
Tracing the fsstress processes that are running when the filesystem
shut down immediately pin-pointed the problem:
user shutdown marks xfs_mount as shutdown
godown-213341 [008] 6398.022871: console: [ 6397.915392] XFS (pmem1): User initiated shutdown received.
.....
aild tries to push ordered inode cluster buffer
xfsaild/pmem1-213314 [001] 6398.022974: xfs_buf_trylock: dev 259:1 daddr 0x851c80 bbcount 0x20 hold 16 pincount 0 lock 0 flags DONE|INODES|PAGES caller xfs_inode_item_push+0x8e
xfsaild/pmem1-213314 [001] 6398.022976: xfs_ilock_nowait: dev 259:1 ino 0x851c80 flags ILOCK_SHARED caller xfs_iflush_cluster+0xae
xfs_iflush_cluster() checks xfs_is_shutdown(), returns true,
calls xfs_iflush_abort() to kill writeback of the inode.
Inode is removed from AIL, drops cluster buffer reference.
xfsaild/pmem1-213314 [001] 6398.022977: xfs_ail_delete: dev 259:1 lip 0xffff88880247ed80 old lsn 7/20344 new lsn 7/21000 type XFS_LI_INODE flags IN_AIL
xfsaild/pmem1-213314 [001] 6398.022978: xfs_buf_rele: dev 259:1 daddr 0x851c80 bbcount 0x20 hold 17 pincount 0 lock 0 flags DONE|INODES|PAGES caller xfs_iflush_abort+0xd7
.....
All inodes on cluster buffer are aborted, then the cluster buffer
itself is aborted and removed from the AIL *without writeback*:
xfsaild/pmem1-213314 [001] 6398.023011: xfs_buf_error_relse: dev 259:1 daddr 0x851c80 bbcount 0x20 hold 2 pincount 0 lock 0 flags ASYNC|DONE|STALE|INODES|PAGES caller xfs_buf_ioend_fail+0x33
xfsaild/pmem1-213314 [001] 6398.023012: xfs_ail_delete: dev 259:1 lip 0xffff8888053efde8 old lsn 7/20344 new lsn 7/20344 type XFS_LI_BUF flags IN_AIL
The inode buffer was at 7/20344 when it was removed from the AIL.
xfsaild/pmem1-213314 [001] 6398.023012: xfs_buf_item_relse: dev 259:1 daddr 0x851c80 bbcount 0x20 hold 2 pincount 0 lock 0 flags ASYNC|DONE|STALE|INODES|PAGES caller xfs_buf_item_done+0x31
xfsaild/pmem1-213314 [001] 6398.023012: xfs_buf_rele: dev 259:1 daddr 0x851c80 bbcount 0x20 hold 2 pincount 0 lock 0 flags ASYNC|DONE|STALE|INODES|PAGES caller xfs_buf_item_relse+0x39
.....
Userspace is still running, doing stuff. an fsstress process runs
syncfs() or sync() and we end up in sync_fs_one_sb() which issues
a log force. This pushes on the CIL:
fsstress-213322 [001] 6398.024430: xfs_fs_sync_fs: dev 259:1 m_features 0x20000000019ff6e9 opstate (clean|shutdown|inodegc|blockgc) s_flags 0x70810000 caller sync_fs_one_sb+0x26
fsstress-213322 [001] 6398.024430: xfs_log_force: dev 259:1 lsn 0x0 caller xfs_fs_sync_fs+0x82
fsstress-213322 [001] 6398.024430: xfs_log_force: dev 259:1 lsn 0x5f caller xfs_log_force+0x7c
<...>-194402 [001] 6398.024467: kmem_alloc: size 176 flags 0x14 caller xlog_cil_push_work+0x9f
And the CIL fills up iclogs with pending changes. This picks up
the current tail from the AIL:
<...>-194402 [001] 6398.024497: xlog_iclog_get_space: dev 259:1 state XLOG_STATE_ACTIVE refcnt 1 offset 0 lsn 0x0 flags caller xlog_write+0x149
<...>-194402 [001] 6398.024498: xlog_iclog_switch: dev 259:1 state XLOG_STATE_ACTIVE refcnt 1 offset 0 lsn 0x700005408 flags caller xlog_state_get_iclog_space+0x37e
<...>-194402 [001] 6398.024521: xlog_iclog_release: dev 259:1 state XLOG_STATE_WANT_SYNC refcnt 1 offset 32256 lsn 0x700005408 flags caller xlog_write+0x5f9
<...>-194402 [001] 6398.024522: xfs_log_assign_tail_lsn: dev 259:1 new tail lsn 7/21000, old lsn 7/20344, last sync 7/21448
And it moves the tail of the log to 7/21000 from 7/20344. This
*moves the tail of the log beyond the ICREATE transaction* that was
at 7/20344 and pinned by the inode cluster buffer that was cancelled
above.
....
godown-213341 [008] 6398.027005: xfs_force_shutdown: dev 259:1 tag logerror flags log_io|force_umount file fs/xfs/xfs_fsops.c line_num 500
godown-213341 [008] 6398.027022: console: [ 6397.915406] pmem1: writeback error on inode 12621949, offset 1019904, sector 12968096
godown-213341 [008] 6398.030551: console: [ 6397.919546] XFS (pmem1): Log I/O Error (0x6) detected at xfs_fs_goingdown+0xa3/0xf0 (fs/
And finally the log itself is now shutdown, stopping all further
writes to the log. But this is too late to prevent the corruption
that moving the tail of the log forwards after we start cancelling
writeback causes.
The fundamental problem here is that we are using the wrong shutdown
checks for log items. We've long conflated mount shutdown with log
shutdown state, and I started separating that recently with the
atomic shutdown state changes in commit b36d4651e165 ("xfs: make
forced shutdown processing atomic"). The changes in that commit
series are directly responsible for being able to diagnose this
issue because it clearly separated mount shutdown from log shutdown.
Essentially, once we start cancelling writeback of log items and
removing them from the AIL because the filesystem is shut down, we
*cannot* update the journal because we may have cancelled the items
that pin the tail of the log. That moves the tail of the log
forwards without having written the metadata back, hence we have
corrupt in memory state and writing to the journal propagates that
to the on-disk state.
What commit b36d4651e165 makes clear is that log item state needs to
change relative to log shutdown, not mount shutdown. IOWs, anything
that aborts metadata writeback needs to check log shutdown state
because log items directly affect log consistency. Having them check
mount shutdown state introduces the above race condition where we
cancel metadata writeback before the log shuts down.
To fix this, this patch works through all log items and converts
shutdown checks to use xlog_is_shutdown() rather than
xfs_is_shutdown(), so that we don't start aborting metadata
writeback before we shut off journal writes.
AFAICT, this race condition is a zero day IO error handling bug in
XFS that dates back to the introduction of XLOG_IO_ERROR,
XLOG_STATE_IOERROR and XFS_FORCED_SHUTDOWN back in January 1997.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
The AIL operates purely on log items, so it is a log centric
subsystem. Divorce it from the xfs_mount and instead have it pass
around xlog pointers.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Log items belong to the log, not the xfs_mount. Convert the mount
pointer in the log item to a xlog pointer in preparation for
upcoming log centric changes to the log items.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
When the AIL tries to flush the CIL, it relies on the CIL push
ending up on stable storage without having to wait for and
manipulate iclog state directly. However, if there is already a
pending CIL push when the AIL tries to flush the CIL, it won't set
the cil->xc_push_commit_stable flag and so the CIL push will not
actively flush the commit record iclog.
generic/530 when run on a single CPU test VM can trigger this fairly
reliably. This test exercises unlinked inode recovery, and can
result in inodes being pinned in memory by ongoing modifications to
the inode cluster buffer to record unlinked list modifications. As a
result, the first inode unlinked in a buffer can pin the tail of the
log whilst the inode cluster buffer is pinned by the current
checkpoint that has been pushed but isn't on stable storage because
because the cil->xc_push_commit_stable was not set. This results in
the log/AIL effectively deadlocking until something triggers the
commit record iclog to be pushed to stable storage (i.e. the
periodic log worker calling xfs_log_force()).
The fix is two-fold - first we should always set the
cil->xc_push_commit_stable when xlog_cil_flush() is called,
regardless of whether there is already a pending push or not.
Second, if the CIL is empty, we should trigger an iclog flush to
ensure that the iclogs of the last checkpoint have actually been
submitted to disk as that checkpoint may not have been run under
stable completion constraints.
Reported-and-tested-by: Matthew Wilcox <willy@infradead.org>
Fixes: 0020a190cf3e ("xfs: AIL needs asynchronous CIL forcing")
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
xfs_ail_push_all_sync() has a loop like this:
while max_ail_lsn {
prepare_to_wait(ail_empty)
target = max_ail_lsn
wake_up(ail_task);
schedule()
}
Which is designed to sleep until the AIL is emptied. When
xfs_ail_update_finish() moves the tail of the log, it does:
if (list_empty(&ailp->ail_head))
wake_up_all(&ailp->ail_empty);
So it will only wake up the sync push waiter when the AIL goes
empty. If, by the time the push waiter has woken, the AIL has more
in it, it will reset the target, wake the push task and go back to
sleep.
The problem here is that if the AIL is having items added to it
when xfs_ail_push_all_sync() is called, then they may get inserted
into the AIL at a LSN higher than the target LSN. At this point,
xfsaild_push() will see that the target is X, the item LSNs are
(X+N) and skip over them, hence never pushing the out.
The result of this the AIL will not get emptied by the AIL push
thread, hence xfs_ail_finish_update() will never see the AIL being
empty even if it moves the tail. Hence xfs_ail_push_all_sync() never
gets woken and hence cannot update the push target to capture the
items beyond the current target on the LSN.
This is a TOCTOU type of issue so the way to avoid it is to not
use the push target at all for sync pushes. We know that a sync push
is being requested by the fact the ail_empty wait queue is active,
hence the xfsaild can just set the target to max_ail_lsn on every
push that we see the wait queue active. Hence we no longer will
leave items on the AIL that are beyond the LSN sampled at the start
of a sync push.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
AIL flushing can get stuck here:
[316649.005769] INFO: task xfsaild/pmem1:324525 blocked for more than 123 seconds.
[316649.007807] Not tainted 5.17.0-rc6-dgc+ #975
[316649.009186] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[316649.011720] task:xfsaild/pmem1 state:D stack:14544 pid:324525 ppid: 2 flags:0x00004000
[316649.014112] Call Trace:
[316649.014841] <TASK>
[316649.015492] __schedule+0x30d/0x9e0
[316649.017745] schedule+0x55/0xd0
[316649.018681] io_schedule+0x4b/0x80
[316649.019683] xfs_buf_wait_unpin+0x9e/0xf0
[316649.021850] __xfs_buf_submit+0x14a/0x230
[316649.023033] xfs_buf_delwri_submit_buffers+0x107/0x280
[316649.024511] xfs_buf_delwri_submit_nowait+0x10/0x20
[316649.025931] xfsaild+0x27e/0x9d0
[316649.028283] kthread+0xf6/0x120
[316649.030602] ret_from_fork+0x1f/0x30
in the situation where flushing gets preempted between the unpin
check and the buffer trylock under nowait conditions:
blk_start_plug(&plug);
list_for_each_entry_safe(bp, n, buffer_list, b_list) {
if (!wait_list) {
if (xfs_buf_ispinned(bp)) {
pinned++;
continue;
}
Here >>>>>>
if (!xfs_buf_trylock(bp))
continue;
This means submission is stuck until something else triggers a log
force to unpin the buffer.
To get onto the delwri list to begin with, the buffer pin state has
already been checked, and hence it's relatively rare we get a race
between flushing and encountering a pinned buffer in delwri
submission to begin with. Further, to increase the pin count the
buffer has to be locked, so the only way we can hit this race
without failing the trylock is to be preempted between the pincount
check seeing zero and the trylock being run.
Hence to avoid this problem, just invert the order of trylock vs
pin check. We shouldn't hit that many pinned buffers here, so
optimising away the trylock for pinned buffers should not matter for
performance at all.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
After 963 iterations of generic/530, it deadlocked during recovery
on a pinned inode cluster buffer like so:
XFS (pmem1): Starting recovery (logdev: internal)
INFO: task kworker/8:0:306037 blocked for more than 122 seconds.
Not tainted 5.17.0-rc6-dgc+ #975
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
task:kworker/8:0 state:D stack:13024 pid:306037 ppid: 2 flags:0x00004000
Workqueue: xfs-inodegc/pmem1 xfs_inodegc_worker
Call Trace:
<TASK>
__schedule+0x30d/0x9e0
schedule+0x55/0xd0
schedule_timeout+0x114/0x160
__down+0x99/0xf0
down+0x5e/0x70
xfs_buf_lock+0x36/0xf0
xfs_buf_find+0x418/0x850
xfs_buf_get_map+0x47/0x380
xfs_buf_read_map+0x54/0x240
xfs_trans_read_buf_map+0x1bd/0x490
xfs_imap_to_bp+0x4f/0x70
xfs_iunlink_map_ino+0x66/0xd0
xfs_iunlink_map_prev.constprop.0+0x148/0x2f0
xfs_iunlink_remove_inode+0xf2/0x1d0
xfs_inactive_ifree+0x1a3/0x900
xfs_inode_unlink+0xcc/0x210
xfs_inodegc_worker+0x1ac/0x2f0
process_one_work+0x1ac/0x390
worker_thread+0x56/0x3c0
kthread+0xf6/0x120
ret_from_fork+0x1f/0x30
</TASK>
task:mount state:D stack:13248 pid:324509 ppid:324233 flags:0x00004000
Call Trace:
<TASK>
__schedule+0x30d/0x9e0
schedule+0x55/0xd0
schedule_timeout+0x114/0x160
__down+0x99/0xf0
down+0x5e/0x70
xfs_buf_lock+0x36/0xf0
xfs_buf_find+0x418/0x850
xfs_buf_get_map+0x47/0x380
xfs_buf_read_map+0x54/0x240
xfs_trans_read_buf_map+0x1bd/0x490
xfs_imap_to_bp+0x4f/0x70
xfs_iget+0x300/0xb40
xlog_recover_process_one_iunlink+0x4c/0x170
xlog_recover_process_iunlinks.isra.0+0xee/0x130
xlog_recover_finish+0x57/0x110
xfs_log_mount_finish+0xfc/0x1e0
xfs_mountfs+0x540/0x910
xfs_fs_fill_super+0x495/0x850
get_tree_bdev+0x171/0x270
xfs_fs_get_tree+0x15/0x20
vfs_get_tree+0x24/0xc0
path_mount+0x304/0xba0
__x64_sys_mount+0x108/0x140
do_syscall_64+0x35/0x80
entry_SYSCALL_64_after_hwframe+0x44/0xae
</TASK>
task:xfsaild/pmem1 state:D stack:14544 pid:324525 ppid: 2 flags:0x00004000
Call Trace:
<TASK>
__schedule+0x30d/0x9e0
schedule+0x55/0xd0
io_schedule+0x4b/0x80
xfs_buf_wait_unpin+0x9e/0xf0
__xfs_buf_submit+0x14a/0x230
xfs_buf_delwri_submit_buffers+0x107/0x280
xfs_buf_delwri_submit_nowait+0x10/0x20
xfsaild+0x27e/0x9d0
kthread+0xf6/0x120
ret_from_fork+0x1f/0x30
We have the mount process waiting on an inode cluster buffer read,
inodegc doing unlink waiting on the same inode cluster buffer, and
the AIL push thread blocked in writeback waiting for the inode
cluster buffer to become unpinned.
What has happened here is that the AIL push thread has raced with
the inodegc process modifying, committing and pinning the inode
cluster buffer here in xfs_buf_delwri_submit_buffers() here:
blk_start_plug(&plug);
list_for_each_entry_safe(bp, n, buffer_list, b_list) {
if (!wait_list) {
if (xfs_buf_ispinned(bp)) {
pinned++;
continue;
}
Here >>>>>>
if (!xfs_buf_trylock(bp))
continue;
Basically, the AIL has found the buffer wasn't pinned and got the
lock without blocking, but then the buffer was pinned. This implies
the processing here was pre-empted between the pin check and the
lock, because the pin count can only be increased while holding the
buffer locked. Hence when it has gone to submit the IO, it has
blocked waiting for the buffer to be unpinned.
With all executing threads now waiting on the buffer to be unpinned,
we normally get out of situations like this via the background log
worker issuing a log force which will unpinned stuck buffers like
this. But at this point in recovery, we haven't started the log
worker. In fact, the first thing we do after processing intents and
unlinked inodes is *start the log worker*. IOWs, we start it too
late to have it break deadlocks like this.
Avoid this and any other similar deadlock vectors in intent and
unlinked inode recovery by starting the log worker before we recover
intents and unlinked inodes. This part of recovery runs as though
the filesystem is fully active, so we really should have the same
infrastructure running as we normally do at runtime.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
The symbol xfs_name_dotdot is a global variable that the xfs codebase
uses here and there to look up directory dotdot entries. Currently it's
a non-const variable, which means that it's a mutable global variable.
So far nobody's abused this to cause problems, but let's use the
compiler to enforce that.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Various directory functions do not modify their @name parameter,
so mark it const to make that clear. This will enable us to mark
the global xfs_name_dotdot variable as const to prevent mischief.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
XFS does not reserve quota for directory expansion when renaming
children into a directory. This means that we don't reject the
expansion with EDQUOT when we're at or near a hard limit, which means
that unprivileged userspace can use rename() to exceed quota.
Rename operations don't always expand the target directory, and we allow
a rename to proceed with no space reservation if we don't need to add a
block to the target directory to handle the addition. Moreover, the
unlink operation on the source directory generally does not expand the
directory (you'd have to free a block and then cause a btree split) and
it's probably of little consequence to leave the corner case that
renaming a file out of a directory can increase its size.
As with link and unlink, there is a further bug in that we do not
trigger the blockgc workers to try to clear space when we're out of
quota.
Because rename is its own special tricky animal, we'll patch xfs_rename
directly to reserve quota to the rename transaction. We'll leave
cleaning up the rest of xfs_rename for the metadata directory tree
patchset.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
XFS does not reserve quota for directory expansion when linking or
unlinking children from a directory. This means that we don't reject
the expansion with EDQUOT when we're at or near a hard limit, which
means that unprivileged userspace can use link()/unlink() to exceed
quota.
The fix for this is nuanced -- link operations don't always expand the
directory, and we allow a link to proceed with no space reservation if
we don't need to add a block to the directory to handle the addition.
Unlink operations generally do not expand the directory (you'd have to
free a block and then cause a btree split) and we can defer the
directory block freeing if there is no space reservation.
Moreover, there is a further bug in that we do not trigger the blockgc
workers to try to clear space when we're out of quota.
To fix both cases, create a new xfs_trans_alloc_dir function that
allocates the transaction, locks and joins the inodes, and reserves
quota for the directory. If there isn't sufficient space or quota,
we'll switch the caller to reservationless mode. This should prevent
quota usage overruns with the least restriction in functionality.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Combine if tests to reduce the indentation levels of the quota chown
calls in xfs_setattr_nonsize.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Christian Brauner <brauner@kernel.org>
|
|
Filipe Manana pointed out that XFS' behavior w.r.t. setuid/setgid
revocation isn't consistent with btrfs[1] or ext4. Those two
filesystems use the VFS function setattr_copy to convey certain
attributes from struct iattr into the VFS inode structure.
Andrey Zhadchenko reported[2] that XFS uses the wrong user namespace to
decide if it should clear setgid and setuid on a file attribute update.
This is a second symptom of the problem that Filipe noticed.
XFS, on the other hand, open-codes setattr_copy in xfs_setattr_mode,
xfs_setattr_nonsize, and xfs_setattr_time. Regrettably, setattr_copy is
/not/ a simple copy function; it contains additional logic to clear the
setgid bit when setting the mode, and XFS' version no longer matches.
The VFS implements its own setuid/setgid stripping logic, which
establishes consistent behavior. It's a tad unfortunate that it's
scattered across notify_change, should_remove_suid, and setattr_copy but
XFS should really follow the Linux VFS. Adapt XFS to use the VFS
functions and get rid of the old functions.
[1] https://lore.kernel.org/fstests/CAL3q7H47iNQ=Wmk83WcGB-KBJVOEtR9+qGczzCeXJ9Y2KCV25Q@mail.gmail.com/
[2] https://lore.kernel.org/linux-xfs/20220221182218.748084-1-andrey.zhadchenko@virtuozzo.com/
Fixes: 7fa294c8991c ("userns: Allow chown and setgid preservation")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Christian Brauner <brauner@kernel.org>
|
|
There are a few places where we test the current process' capability set
to decide if we're going to be more or less generous with resource
acquisition for a system call. If the process doesn't have the
capability, we can continue the call, albeit in a degraded mode.
These are /not/ the actual security decisions, so it's not proper to use
capable(), which (in certain selinux setups) causes audit messages to
get logged. Switch them to has_capability_noaudit.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Cc: Ondrej Mosnacek <omosnace@redhat.com>
Cc: Dave Chinner <david@fromorbit.com>
|
|
COW extents are already converted into written real extents after
xfs_reflink_convert_cow_locked(), therefore cmap->br_state should
reflect it.
Otherwise, there is another necessary unwritten convertion
triggered in xfs_dio_write_end_io() for direct I/O cases.
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
|
|
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|