Pull ceph fix from Ilya Dryomov:
"A ceph filesystem fix, marked for stable.
There appears to be a deeper issue on the MDS side, but for now we are
going with this one-liner to avoid busy looping and potential soft
lockups"
* tag 'ceph-for-5.19-rc5' of https://github.com/ceph/ceph-client:
ceph: wait on async create before checking caps for syncfs
Pull io_uring fixes from Jens Axboe:
"Two minor tweaks:
- While we still can, adjust the send/recv based flags to be in
->ioprio rather than in ->addr2. This is consistent with eg accept,
and also doesn't waste a full 64-bit field for flags (Pavel)
- 5.18-stable fix for re-importing provided buffers. Not much real
world relevance here as it'll only impact non-pollable files gone
async, which is more of a practical test case rather than something
that is used in the wild (Dylan)"
* tag 'io_uring-5.19-2022-07-01' of git://git.kernel.dk/linux-block:
io_uring: fix provided buffer import
io_uring: keep sendrecv flags in ioprio
The CIL push lock is highly contended on larger machines, becoming a
hard bottleneck that about 700,000 transaction commits/s on >16p
machines. To address this, start moving the CIL tracking
infrastructure to utilise per-CPU structures.
We need to track the space used, the amount of log reservation space
reserved to write the CIL, the log items in the CIL and the busy
extents that need to be completed by the CIL commit. This requires
a couple of per-cpu counters, an unordered per-cpu list and a
globally ordered per-cpu list.
Create a per-cpu structure to hold these and all the management
interfaces needed, as well as the hooks to handle hotplug CPUs.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
For every iclog that a CIL push will use up, we need to ensure we
have space reserved for the iclog header in each iclog. It is
extremely difficult to do this accurately with a per-cpu counter
without expensive summing of the counter in every commit. However,
we know what the maximum CIL size is going to be because of the
hard space limit we have, and hence we know exactly how many iclogs
we are going to need to write out the CIL.
We are constrained by the requirement that small transactions only
have reservation space for a single iclog header built into them.
At commit time we don't know how much of the current transaction
reservation is made up of iclog header reservations as calculated by
xfs_log_calc_unit_res() when the ticket was reserved. As larger
reservations have multiple header spaces reserved, we can steal
more than one iclog header reservation at a time, but we only steal
the exact number needed for the given log vector size delta.
As a result, we don't know exactly when we are going to steal iclog
header reservations, nor do we know exactly how many we are going to
need for a given CIL.
To make things simple, start by calculating the worst case number of
iclog headers a full CIL push will require. Record this into an
atomic variable in the CIL. Then add a byte counter to the log
ticket that records exactly how much iclog header space has been
reserved in this ticket by xfs_log_calc_unit_res(). This tells us
exactly how much space we can steal from the ticket at transaction
commit time.
Now, at transaction commit time, we can check if the CIL has a full
iclog header reservation and, if not, steal the entire reservation
the current ticket holds for iclog headers. This minimises the
number of times we need to do atomic operations in the fast path,
but still guarantees we get all the reservations we need.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
The xc_cil_lock is the most highly contended lock in XFS now. To
start the process of getting rid of it, lift the initial reservation
of the CIL log space out from under the xc_cil_lock.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
In the next patches we are going to make the CIL list itself
per-cpu, and so we cannot use list_empty() to check is the list is
empty. Replace the list_empty() checks with a flag in the CIL to
indicate we have committed at least one transaction to the CIL and
hence the CIL is not empty.
We need this flag to be an atomic so that we can clear it without
holding any locks in the commit fast path, but we also need to be
careful to avoid atomic operations in the fast path. Hence we use
the fact that test_bit() is not an atomic op to first check if the
flag is set and then run the atomic test_and_clear_bit() operation
to clear it and steal the initial unit reservation for the CIL
context checkpoint.
When we are switching to a new context in a push, we place the
setting of the XLOG_CIL_EMPTY flag under the xc_push_lock. THis
allows all the other places that need to check whether the CIL is
empty to use test_bit() and still be serialised correctly with the
CIL context swaps that set the bit.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
KASAN reported the following use after free bug when running
generic/475:
XFS (dm-0): Mounting V5 Filesystem
XFS (dm-0): Starting recovery (logdev: internal)
XFS (dm-0): Ending recovery (logdev: internal)
Buffer I/O error on dev dm-0, logical block 20639616, async page read
Buffer I/O error on dev dm-0, logical block 20639617, async page read
XFS (dm-0): log I/O error -5
XFS (dm-0): Filesystem has been shut down due to log error (0x2).
XFS (dm-0): Unmounting Filesystem
XFS (dm-0): Please unmount the filesystem and rectify the problem(s).
==================================================================
BUG: KASAN: use-after-free in do_raw_spin_lock+0x246/0x270
Read of size 4 at addr ffff888109dd84c4 by task 3:1H/136
CPU: 3 PID: 136 Comm: 3:1H Not tainted 5.19.0-rc4-xfsx #rc4 8e53ab5ad0fddeb31cee5e7063ff9c361915a9c4
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.15.0-1 04/01/2014
Workqueue: xfs-log/dm-0 xlog_ioend_work [xfs]
Call Trace:
<TASK>
dump_stack_lvl+0x34/0x44
print_report.cold+0x2b8/0x661
? do_raw_spin_lock+0x246/0x270
kasan_report+0xab/0x120
? do_raw_spin_lock+0x246/0x270
do_raw_spin_lock+0x246/0x270
? rwlock_bug.part.0+0x90/0x90
xlog_force_shutdown+0xf6/0x370 [xfs 4ad76ae0d6add7e8183a553e624c31e9ed567318]
xlog_ioend_work+0x100/0x190 [xfs 4ad76ae0d6add7e8183a553e624c31e9ed567318]
process_one_work+0x672/0x1040
worker_thread+0x59b/0xec0
? __kthread_parkme+0xc6/0x1f0
? process_one_work+0x1040/0x1040
? process_one_work+0x1040/0x1040
kthread+0x29e/0x340
? kthread_complete_and_exit+0x20/0x20
ret_from_fork+0x1f/0x30
</TASK>
Allocated by task 154099:
kasan_save_stack+0x1e/0x40
__kasan_kmalloc+0x81/0xa0
kmem_alloc+0x8d/0x2e0 [xfs]
xlog_cil_init+0x1f/0x540 [xfs]
xlog_alloc_log+0xd1e/0x1260 [xfs]
xfs_log_mount+0xba/0x640 [xfs]
xfs_mountfs+0xf2b/0x1d00 [xfs]
xfs_fs_fill_super+0x10af/0x1910 [xfs]
get_tree_bdev+0x383/0x670
vfs_get_tree+0x7d/0x240
path_mount+0xdb7/0x1890
__x64_sys_mount+0x1fa/0x270
do_syscall_64+0x2b/0x80
entry_SYSCALL_64_after_hwframe+0x46/0xb0
Freed by task 154151:
kasan_save_stack+0x1e/0x40
kasan_set_track+0x21/0x30
kasan_set_free_info+0x20/0x30
____kasan_slab_free+0x110/0x190
slab_free_freelist_hook+0xab/0x180
kfree+0xbc/0x310
xlog_dealloc_log+0x1b/0x2b0 [xfs]
xfs_unmountfs+0x119/0x200 [xfs]
xfs_fs_put_super+0x6e/0x2e0 [xfs]
generic_shutdown_super+0x12b/0x3a0
kill_block_super+0x95/0xd0
deactivate_locked_super+0x80/0x130
cleanup_mnt+0x329/0x4d0
task_work_run+0xc5/0x160
exit_to_user_mode_prepare+0xd4/0xe0
syscall_exit_to_user_mode+0x1d/0x40
entry_SYSCALL_64_after_hwframe+0x46/0xb0
This appears to be a race between the unmount process, which frees the
CIL and waits for in-flight iclog IO; and the iclog IO completion. When
generic/475 runs, it starts fsstress in the background, waits a few
seconds, and substitutes a dm-error device to simulate a disk falling
out of a machine. If the fsstress encounters EIO on a pure data write,
it will exit but the filesystem will still be online.
The next thing the test does is unmount the filesystem, which tries to
clean the log, free the CIL, and wait for iclog IO completion. If an
iclog was being written when the dm-error switch occurred, it can race
with log unmounting as follows:
Thread 1 Thread 2
xfs_log_unmount
xfs_log_clean
xfs_log_quiesce
xlog_ioend_work
<observe error>
xlog_force_shutdown
test_and_set_bit(XLOG_IOERROR)
xfs_log_force
<log is shut down, nop>
xfs_log_umount_write
<log is shut down, nop>
xlog_dealloc_log
xlog_cil_destroy
<wait for iclogs>
spin_lock(&log->l_cilp->xc_push_lock)
<KABOOM>
Therefore, free the CIL after waiting for the iclogs to complete. I
/think/ this race has existed for quite a few years now, though I don't
remember the ~2014 era logging code well enough to know if it was a real
threat then or if the actual race was exposed only more recently.
Fixes: ac983517ec ("xfs: don't sleep in xlog_cil_force_lsn on shutdown")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
This flag is a new way to configure ignore mask which allows adding and
removing the event flags FAN_ONDIR and FAN_EVENT_ON_CHILD in ignore mask.
The legacy FAN_MARK_IGNORED_MASK flag would always ignore events on
directories and would ignore events on children depending on whether
the FAN_EVENT_ON_CHILD flag was set in the (non ignored) mask.
FAN_MARK_IGNORE can be used to ignore events on children without setting
FAN_EVENT_ON_CHILD in the mark's mask and will not ignore events on
directories unconditionally, only when FAN_ONDIR is set in ignore mask.
The new behavior is non-downgradable. After calling fanotify_mark() with
FAN_MARK_IGNORE once, calling fanotify_mark() with FAN_MARK_IGNORED_MASK
on the same object will return EEXIST error.
Setting the event flags with FAN_MARK_IGNORE on a non-dir inode mark
has no meaning and will return ENOTDIR error.
The meaning of FAN_MARK_IGNORED_SURV_MODIFY is preserved with the new
FAN_MARK_IGNORE flag, but with a few semantic differences:
1. FAN_MARK_IGNORED_SURV_MODIFY is required for filesystem and mount
marks and on an inode mark on a directory. Omitting this flag
will return EINVAL or EISDIR error.
2. An ignore mask on a non-directory inode that survives modify could
never be downgraded to an ignore mask that does not survive modify.
With new FAN_MARK_IGNORE semantics we make that rule explicit -
trying to update a surviving ignore mask without the flag
FAN_MARK_IGNORED_SURV_MODIFY will return EEXIST error.
The conveniene macro FAN_MARK_IGNORE_SURV is added for
(FAN_MARK_IGNORE | FAN_MARK_IGNORED_SURV_MODIFY), because the
common case should use short constant names.
Link: https://lore.kernel.org/r/20220629144210.2983229-4-amir73il@gmail.com
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Setting flags FAN_ONDIR FAN_EVENT_ON_CHILD in ignore mask has no effect.
The FAN_EVENT_ON_CHILD flag in mask implicitly applies to ignore mask and
ignore mask is always implicitly applied to events on directories.
Define a mark flag that replaces this legacy behavior with logic of
applying the ignore mask according to event flags in ignore mask.
Implement the new logic to prepare for supporting an ignore mask that
ignores events on children and ignore mask that does not ignore events
on directories.
To emphasize the change in terminology, also rename ignored_mask mark
member to ignore_mask and use accessors to get only the effective
ignored events or the ignored events and flags.
This change in terminology finally aligns with the "ignore mask"
language in man pages and in most of the comments.
Link: https://lore.kernel.org/r/20220629144210.2983229-2-amir73il@gmail.com
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
A regression has been reported by Nicolas Boichat, found while using the
copy_file_range syscall to copy a tracefs file.
Before commit 5dae222a5f ("vfs: allow copy_file_range to copy across
devices") the kernel would return -EXDEV to userspace when trying to
copy a file across different filesystems. After this commit, the
syscall doesn't fail anymore and instead returns zero (zero bytes
copied), as this file's content is generated on-the-fly and thus reports
a size of zero.
Another regression has been reported by He Zhe - the assertion of
WARN_ON_ONCE(ret == -EOPNOTSUPP) can be triggered from userspace when
copying from a sysfs file whose read operation may return -EOPNOTSUPP.
Since we do not have test coverage for copy_file_range() between any two
types of filesystems, the best way to avoid these sort of issues in the
future is for the kernel to be more picky about filesystems that are
allowed to do copy_file_range().
This patch restores some cross-filesystem copy restrictions that existed
prior to commit 5dae222a5f ("vfs: allow copy_file_range to copy across
devices"), namely, cross-sb copy is not allowed for filesystems that do
not implement ->copy_file_range().
Filesystems that do implement ->copy_file_range() have full control of
the result - if this method returns an error, the error is returned to
the user. Before this change this was only true for fs that did not
implement the ->remap_file_range() operation (i.e. nfsv3).
Filesystems that do not implement ->copy_file_range() still fall-back to
the generic_copy_file_range() implementation when the copy is within the
same sb. This helps the kernel can maintain a more consistent story
about which filesystems support copy_file_range().
nfsd and ksmbd servers are modified to fall-back to the
generic_copy_file_range() implementation in case vfs_copy_file_range()
fails with -EOPNOTSUPP or -EXDEV, which preserves behavior of
server-side-copy.
fall-back to generic_copy_file_range() is not implemented for the smb
operation FSCTL_DUPLICATE_EXTENTS_TO_FILE, which is arguably a correct
change of behavior.
Fixes: 5dae222a5f ("vfs: allow copy_file_range to copy across devices")
Link: https://lore.kernel.org/linux-fsdevel/20210212044405.4120619-1-drinkcat@chromium.org/
Link: https://lore.kernel.org/linux-fsdevel/CANMq1KDZuxir2LM5jOTm0xx+BnvW=ZmpsG47CyHFJwnw7zSX6Q@mail.gmail.com/
Link: https://lore.kernel.org/linux-fsdevel/20210126135012.1.If45b7cdc3ff707bc1efa17f5366057d60603c45f@changeid/
Link: https://lore.kernel.org/linux-fsdevel/20210630161320.29006-1-lhenriques@suse.de/
Reported-by: Nicolas Boichat <drinkcat@chromium.org>
Reported-by: kernel test robot <oliver.sang@intel.com>
Signed-off-by: Luis Henriques <lhenriques@suse.de>
Fixes: 64bf5ff58d ("vfs: no fallback for ->copy_file_range")
Link: https://lore.kernel.org/linux-fsdevel/20f17f64-88cb-4e80-07c1-85cb96c83619@windriver.com/
Reported-by: He Zhe <zhe.he@windriver.com>
Tested-by: Namjae Jeon <linkinjeon@kernel.org>
Tested-by: Luis Henriques <lhenriques@suse.de>
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit f49169c97f ("NFSD: Remove svc_serv_ops::svo_module") removed
calls to module_put_and_kthread_exit() from threads that acted as SUNRPC
servers and had a related svc_serv_ops structure. This was correct.
It ALSO removed the module_put_and_kthread_exit() call from
nfs4_run_state_manager() which is NOT a SUNRPC service.
Consequently every time the NFSv4 state manager runs the module count
increments and won't be decremented. So the nfsv4 module cannot be
unloaded.
So restore the module_put_and_kthread_exit() call.
Fixes: f49169c97f ("NFSD: Remove svc_serv_ops::svo_module")
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
io_import_iovec uses the s pointer, but this was changed immediately
after the iovec was re-imported and so it was imported into the wrong
place.
Change the ordering.
Fixes: 2be2eb02e2 ("io_uring: ensure reads re-import for selected buffers")
Signed-off-by: Dylan Yudaken <dylany@fb.com>
Link: https://lore.kernel.org/r/20220630132006.2825668-1-dylany@fb.com
[axboe: ensure we don't half-import as well]
Signed-off-by: Jens Axboe <axboe@kernel.dk>
It is unnecessary to check and set did_zero value in while() loop
in dax_zero_iter(), we can set did_zero to true only when zeroing
successfully at last.
Signed-off-by: Kaixu Xia <kaixuxia@tencent.com>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
It is unnecessary to check and set did_zero value in while() loop
in iomap_zero_iter(), we can set did_zero to true only when zeroing
successfully at last.
Signed-off-by: Kaixu Xia <kaixuxia@tencent.com>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Pull fanotify fix from Jan Kara:
"A fix for recently added fanotify API to have stricter checks and
refuse some invalid flag combinations to make our life easier in the
future"
* tag 'fsnotify_for_v5.19-rc5' of git://git.kernel.org/pub/scm/linux/kernel/git/jack/linux-fs:
fanotify: refine the validation checks on non-dir inode mask
iomap_do_writepage() sends pages past i_size through
folio_redirty_for_writepage(), which normally isn't a problem because
truncate and friends clean them very quickly.
When the system has cgroups configured, we can end up in situations
where one cgroup has almost no dirty pages at all, and other cgroups
consume the entire background dirty limit. This is especially common in
our XFS workloads in production because they have cgroups using O_DIRECT
for almost all of the IO mixed in with cgroups that do more traditional
buffered IO work.
We've hit storms where the redirty path hits millions of times in a few
seconds, on all a single file that's only ~40 pages long. This leads to
long tail latencies for file writes because the pdflush workers are
hogging the CPU from some kworkers bound to the same CPU.
Reproducing this on 5.18 was tricky because 869ae85dae ("xfs: flush new
eof page on truncate...") ends up writing/waiting most of these dirty pages
before truncate gets a chance to wait on them.
The actual repro looks like this:
/*
* run me in a cgroup all alone. Start a second cgroup with dd
* streaming IO into the block device.
*/
int main(int ac, char **av) {
int fd;
int ret;
char buf[BUFFER_SIZE];
char *filename = av[1];
memset(buf, 0, BUFFER_SIZE);
if (ac != 2) {
fprintf(stderr, "usage: looper filename\n");
exit(1);
}
fd = open(filename, O_WRONLY | O_CREAT, 0600);
if (fd < 0) {
err(errno, "failed to open");
}
fprintf(stderr, "looping on %s\n", filename);
while(1) {
/*
* skip past page 0 so truncate doesn't write and wait
* on our extent before changing i_size
*/
ret = lseek(fd, 8192, SEEK_SET);
if (ret < 0)
err(errno, "lseek");
ret = write(fd, buf, BUFFER_SIZE);
if (ret != BUFFER_SIZE)
err(errno, "write failed");
/* start IO so truncate has to wait after i_size is 0 */
ret = sync_file_range(fd, 16384, 4095, SYNC_FILE_RANGE_WRITE);
if (ret < 0)
err(errno, "sync_file_range");
ret = ftruncate(fd, 0);
if (ret < 0)
err(errno, "truncate");
usleep(1000);
}
}
And this bpftrace script will show when you've hit a redirty storm:
kretprobe:xfs_vm_writepages {
delete(@dirty[pid]);
}
kprobe:xfs_vm_writepages {
@dirty[pid] = 1;
}
kprobe:folio_redirty_for_writepage /@dirty[pid] > 0/ {
$inode = ((struct folio *)arg1)->mapping->host->i_ino;
@inodes[$inode] = count();
@redirty++;
if (@redirty > 90000) {
printf("inode %d redirty was %d", $inode, @redirty);
exit();
}
}
This patch has the same number of failures on xfstests as unpatched 5.18:
Failures: generic/648 xfs/019 xfs/050 xfs/168 xfs/299 xfs/348 xfs/506
xfs/543
I also ran it through a long stress of multiple fsx processes hammering.
(Johannes Weiner did significant tracing and debugging on this as well)
Signed-off-by: Chris Mason <clm@fb.com>
Co-authored-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: Domas Mituzas <domas@fb.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
We waste a u64 SQE field for flags even though we don't need as many
bits and it can be used for something more useful later. Store io_uring
specific send/recv flags in sqe->ioprio instead of ->addr2.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Fixes: 0455d4ccec ("io_uring: add POLL_FIRST support for send/sendmsg and recv/recvmsg")
[axboe: change comment in io_uring.h as well]
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pull ksmbd server fixes from Steve French:
- seek null check (don't use f_seek op directly and blindly)
- offset validation in FSCTL_SET_ZERO_DATA
- fallocate fix (relates e.g. to xfstests generic/091 and 263)
- two cleanup fixes
- fix socket settings on some arch
* tag '5.19-rc4-ksmbd-server-fixes' of git://git.samba.org/ksmbd:
ksmbd: use vfs_llseek instead of dereferencing NULL
ksmbd: check invalid FileOffset and BeyondFinalZero in FSCTL_ZERO_DATA
ksmbd: set the range of bytes to zero without extending file size in FSCTL_ZERO_DATA
ksmbd: remove duplicate flag set in smb2_write
ksmbd: smbd: Remove useless license text when SPDX-License-Identifier is already used
ksmbd: use SOCK_NONBLOCK type for kernel_accept()
Currently, we'll call ceph_check_caps, but if we're still waiting
on the reply, we'll end up spinning around on the same inode in
flush_dirty_session_caps. Wait for the async create reply before
flushing caps.
Cc: stable@vger.kernel.org
URL: https://tracker.ceph.com/issues/55823
Fixes: fbed7045f5 ("ceph: wait for async create reply before sending any cap messages")
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
On a system with a realtime volume and a 28k realtime extent,
generic/491 fails because the test opens a file on a frozen filesystem
and closing it causes xfs_release -> xfs_can_free_eofblocks to
mistakenly think that the the blocks of the realtime extent beyond EOF
are posteof blocks to be freed. Realtime extents cannot be partially
unmapped, so this is pointless. Worse yet, this triggers posteof
cleanup, which stalls on a transaction allocation, which is why the test
fails.
Teach the predicate to account for realtime extents properly.
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Now that we've established (again!) that empty xattr leaf buffers are
ok, we no longer need to bhold them to transactions when we're creating
new leaf blocks. Get rid of the entire mechanism, which should simplify
the xattr code quite a bit.
The original justification for using bhold here was to prevent the AIL
from trying to write the empty leaf block into the fs during the brief
time that we release the buffer lock. The reason for /that/ was to
prevent recovery from tripping over the empty ondisk block.
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
TLDR: Revert commit 51e6104fdb ("xfs: detect empty attr leaf blocks in
xfs_attr3_leaf_verify") because it was wrong.
Every now and then we get a corruption report from the kernel or
xfs_repair about empty leaf blocks in the extended attribute structure.
We've long thought that these shouldn't be possible, but prior to 5.18
one would shake loose in the recoveryloop fstests about once a month.
A new addition to the xattr leaf block verifier in 5.19-rc1 makes this
happen every 7 minutes on my testing cloud. I added a ton of logging to
detect any time we set the header count on an xattr leaf block to zero.
This produced the following dmesg output on generic/388:
XFS (sda4): ino 0x21fcbaf leaf 0x129bf78 hdcount==0!
Call Trace:
<TASK>
dump_stack_lvl+0x34/0x44
xfs_attr3_leaf_create+0x187/0x230
xfs_attr_shortform_to_leaf+0xd1/0x2f0
xfs_attr_set_iter+0x73e/0xa90
xfs_xattri_finish_update+0x45/0x80
xfs_attr_finish_item+0x1b/0xd0
xfs_defer_finish_noroll+0x19c/0x770
__xfs_trans_commit+0x153/0x3e0
xfs_attr_set+0x36b/0x740
xfs_xattr_set+0x89/0xd0
__vfs_setxattr+0x67/0x80
__vfs_setxattr_noperm+0x6e/0x120
vfs_setxattr+0x97/0x180
setxattr+0x88/0xa0
path_setxattr+0xc3/0xe0
__x64_sys_setxattr+0x27/0x30
do_syscall_64+0x35/0x80
entry_SYSCALL_64_after_hwframe+0x46/0xb0
So now we know that someone is creating empty xattr leaf blocks as part
of converting a sf xattr structure into a leaf xattr structure. The
conversion routine logs any existing sf attributes in the same
transaction that creates the leaf block, so we know this is a setxattr
to a file that has no attributes at all.
Next, g/388 calls the shutdown ioctl and cycles the mount to trigger log
recovery. I also augmented buffer item recovery to call ->verify_struct
on any attr leaf blocks and complain if it finds a failure:
XFS (sda4): Unmounting Filesystem
XFS (sda4): Mounting V5 Filesystem
XFS (sda4): Starting recovery (logdev: internal)
XFS (sda4): xattr leaf daddr 0x129bf78 hdrcount == 0!
Call Trace:
<TASK>
dump_stack_lvl+0x34/0x44
xfs_attr3_leaf_verify+0x3b8/0x420
xlog_recover_buf_commit_pass2+0x60a/0x6c0
xlog_recover_items_pass2+0x4e/0xc0
xlog_recover_commit_trans+0x33c/0x350
xlog_recovery_process_trans+0xa5/0xe0
xlog_recover_process_data+0x8d/0x140
xlog_do_recovery_pass+0x19b/0x720
xlog_do_log_recovery+0x62/0xc0
xlog_do_recover+0x33/0x1d0
xlog_recover+0xda/0x190
xfs_log_mount+0x14c/0x360
xfs_mountfs+0x517/0xa60
xfs_fs_fill_super+0x6bc/0x950
get_tree_bdev+0x175/0x280
vfs_get_tree+0x1a/0x80
path_mount+0x6f5/0xaa0
__x64_sys_mount+0x103/0x140
do_syscall_64+0x35/0x80
entry_SYSCALL_64_after_hwframe+0x46/0xb0
RIP: 0033:0x7fc61e241eae
And a moment later, the _delwri_submit of the recovered buffers trips
the same verifier and recovery fails:
XFS (sda4): Metadata corruption detected at xfs_attr3_leaf_verify+0x393/0x420 [xfs], xfs_attr3_leaf block 0x129bf78
XFS (sda4): Unmount and run xfs_repair
XFS (sda4): First 128 bytes of corrupted metadata buffer:
00000000: 00 00 00 00 00 00 00 00 3b ee 00 00 00 00 00 00 ........;.......
00000010: 00 00 00 00 01 29 bf 78 00 00 00 00 00 00 00 00 .....).x........
00000020: a5 1b d0 02 b2 9a 49 df 8e 9c fb 8d f8 31 3e 9d ......I......1>.
00000030: 00 00 00 00 02 1f cb af 00 00 00 00 10 00 00 00 ................
00000040: 00 50 0f b0 00 00 00 00 00 00 00 00 00 00 00 00 .P..............
00000050: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000060: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000070: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
XFS (sda4): Corruption of in-memory data (0x8) detected at _xfs_buf_ioapply+0x37f/0x3b0 [xfs] (fs/xfs/xfs_buf.c:1518). Shutting down filesystem.
XFS (sda4): Please unmount the filesystem and rectify the problem(s)
XFS (sda4): log mount/recovery failed: error -117
XFS (sda4): log mount failed
I think I see what's going on here -- setxattr is racing with something
that shuts down the filesystem:
Thread 1 Thread 2
-------- --------
xfs_attr_sf_addname
xfs_attr_shortform_to_leaf
<create empty leaf>
xfs_trans_bhold(leaf)
xattri_dela_state = XFS_DAS_LEAF_ADD
<roll transaction>
<flush log>
<shut down filesystem>
xfs_trans_bhold_release(leaf)
<discover fs is dead, bail>
Thread 3
--------
<cycle mount, start recovery>
xlog_recover_buf_commit_pass2
xlog_recover_do_reg_buffer
<replay empty leaf buffer from recovered buf item>
xfs_buf_delwri_queue(leaf)
xfs_buf_delwri_submit
_xfs_buf_ioapply(leaf)
xfs_attr3_leaf_write_verify
<trip over empty leaf buffer>
<fail recovery>
As you can see, the bhold keeps the leaf buffer locked and thus prevents
the *AIL* from tripping over the ichdr.count==0 check in the write
verifier. Unfortunately, it doesn't prevent the log from getting
flushed to disk, which sets up log recovery to fail.
So. It's clear that the kernel has always had the ability to persist
attr leaf blocks with ichdr.count==0, which means that it's part of the
ondisk format now.
Unfortunately, this check has been added and removed multiple times
throughout history. It first appeared in[1] kernel 3.10 as part of the
early V5 format patches. The check was later discovered to break log
recovery and hence disabled[2] during log recovery in kernel 4.10.
Simultaneously, the check was added[3] to xfs_repair 4.9.0 to try to
weed out the empty leaf blocks. This was still not correct because log
recovery would recover an empty attr leaf block successfully only for
regular xattr operations to trip over the empty block during of the
block during regular operation. Therefore, the check was removed
entirely[4] in kernel 5.7 but removal of the xfs_repair check was
forgotten. The continued complaints from xfs_repair lead to us
mistakenly re-adding[5] the verifier check for kernel 5.19. Remove it
once again.
[1] 517c22207b ("xfs: add CRCs to attr leaf blocks")
[2] 2e1d23370e ("xfs: ignore leaf attr ichdr.count in verifier
during log replay")
[3] f7140161 ("xfs_repair: junk leaf attribute if count == 0")
[4] f28cef9e4d ("xfs: don't fail verifier on empty attr3 leaf
block")
[5] 51e6104fdb ("xfs: detect empty attr leaf blocks in
xfs_attr3_leaf_verify")
Looking at the rest of the xattr code, it seems that files with empty
leaf blocks behave as expected -- listxattr reports no attributes;
getxattr on any xattr returns nothing as expected; removexattr does
nothing; and setxattr can add attributes just fine.
Original-bug: 517c22207b ("xfs: add CRCs to attr leaf blocks")
Still-not-fixed-by: 2e1d23370e ("xfs: ignore leaf attr ichdr.count in verifier during log replay")
Removed-in: f28cef9e4d ("xfs: don't fail verifier on empty attr3 leaf block")
Fixes: 51e6104fdb ("xfs: detect empty attr leaf blocks in xfs_attr3_leaf_verify")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
In do_promote(), we're never removing the current entry from the list
and so the list traversal is actually safe. Switch back to
list_for_each_entry().
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
In do_promote(), when the glock had no strong holders, we were
accidentally calling demote_incompat_holders() with new_gh == NULL, so
no weak holders were considered incompatible. Instead, the new holder
should have been passed in.
For doing that, the HIF_HOLDER flag needs to be set in new_gh to prevent
may_grant() from complaining. This means that the new holder will now
be recognized as a current holder, so skip over it explicitly in
demote_incompat_holders() to prevent it from being dequeued.
To further clarify things, we can now rename new_gh to current_gh in
demote_incompat_holders(); after all, the HIF_HOLDER flag is already set,
which means the new holder is already a current holder.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
In do_promote() and add_to_queue(), use current_gh as the variable name
for the first strong holder we could find: this matches the variable
name is may_grant(), and more clearly indicates that we're interested in
one (any) of the current strong holders.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Make go_instantiate take a glock instead of a glock holder as its argument:
this handler is supposed to instantiate the object associated with the glock.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Right now, inode_go_instantiate() contains functionality that relates to
how a glock is held rather than the glock itself, like waiting for
pending direct I/O to complete and completing interrupted truncates.
This code is meant to be run each time a holder is acquired, but
go_instantiate is actually only called once, when the glock is
instantiated.
To fix that, introduce a new go_held glock operation that is called each
time a glock holder is acquired. Move the holder specific code in
inode_go_instantiate() over to inode_go_held().
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Now that interrupted truncates are completed in the context of the
process taking the glock, there is no need for the glock state engine to
delegate that task to gfs2_quotad or for quotad to perform those
truncates anymore. Get rid of the obsolete associated infrastructure.
Reverts commit 813e0c46c9 ("GFS2: Fix "truncate in progress" hang").
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Instantiate glocks outside of the glock state engine: there is no real
reason for instantiating them inside the glock state engine; it only
complicates the code.
Instead, instantiate them in gfs2_glock_wait() and gfs2_glock_async_wait()
using the new gfs2_glock_holder_ready() helper. On top of that, the only
other place that acquires a glock without using gfs2_glock_wait() or
gfs2_glock_async_wait() is gfs2_upgrade_iopen_glock(), so call
gfs2_glock_holder_ready() there as well.
If a dinode has a pending truncate, the glock-specific instantiate function
for inodes wakes up the truncate function in the quota daemon. Waiting for
the completion of the truncate was previously done by the glock state
engine, but we now need to wait in inode_go_instantiate().
This also means that gfs2_instantiate() will now no longer return any
"special" error codes.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Since commit 1fc05c8d84 ("gfs2: cancel timed-out glock requests"), a
pending locking request can be canceled by calling gfs2_glock_dq() on
the pending holder. In gfs2_glock_async_wait(), when we time out, use
that to cancel the remaining locking requests and dequeue the locking
requests already granted. That's simpler as well as more efficient than
waiting for all locking requests to eventually be granted and dequeuing
them then.
In addition, gfs2_glock_async_wait() promises that by the time the
function completes, all glocks are either granted or dequeued, but the
implementation doesn't keep that promise if individual locking requests
fail. Fix that as well.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
The pagecache handles readpage failing by itself; it doesn't want
filesystems to remove pages from under it.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
If a buffer is completed with an error, its uptodate flag will be clear,
so the page_uptodate variable will have been set to 0. There's no
need to check PageError here.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
The page cache clears the error bit before calling ->read_folio(),
so this condition could never have been true.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Just because there has been a read error doesn't mean we should avoid
marking this part of the folio as uptodate. Indeed, it may overwrite
the error part of the folio and let us mark the entire folio uptodate.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Pages returned from read_mapping_page() are always uptodate, so
this check is unnecessary.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
read_mapping_folio() returns an ERR_PTR if the folio is not
uptodate, so this check is simply dead code.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
If read_mapping_page() encounters an error, it returns an errno, not
a page with PageError set, or a page that is not Uptodate, so this is
dead code.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
If read_mapping_page() encounters an error, it returns an errno, not a
page with PageError set, so this is dead code.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
If read_mapping_page() encounters an error, it returns an errno, not a
page with PageError set, so this is dead code.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
If read_mapping_page() encounters an error, it returns an errno, not a
page with PageError set, so this is dead code.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
If read_mapping_page() encounters an error, it returns an errno, not a
page with PageError set, so this test is not needed.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
If read_mapping_page() encounters an error, it returns an errno, not a
page with PageError set, so this test is not needed.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
If read_mapping_page() encounters an error, it returns an errno, not a
page with PageError set, so this is dead code.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
If read_mapping_page() encounters an error, it returns an errno, not a
page with PageError set, so this is dead code.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
If read_mapping_page() encounters an error, it returns an errno, not a
page with PageError set, so this is dead code.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
If read_mapping_page() encounters an error, it returns an errno, not a
page with PageError set, so this is dead code.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
