| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
In the current glock reference counting model, a bias of one is added to
a glock's refcount when it is locked (gl->gl_state != LM_ST_UNLOCKED).
A glock is removed from the lru_list when it is enqueued, and added back
when it is dequeued. This isn't a very appropriate model because most
glocks are held for long periods of time (for example, the inode "owns"
references to its inode and iopen glocks as long as the inode is cached
even when the glock state changes to LM_ST_UNLOCKED), and they can only
be freed when they are no longer referenced, anyway.
Fix this by getting rid of the refcount bias for locked glocks. That
way, we can use lockref_put_or_lock() to efficiently drop all but the
last glock reference, and put the glock onto the lru_list when the last
reference is dropped. When find_insert_glock() returns a reference to a
cached glock, it removes the glock from the lru_list.
Dumping the "glocks" and "glstats" debugfs files also takes glock
references, but instead of removing the glocks from the lru_list in that
case as well, we leave them on the list. This ensures that dumping
those files won't perturb the order of the glocks on the lru_list.
In addition, when the last reference to an *unlocked* glock is dropped,
we immediately free it; this preserves the preexisting behavior. If it
later turns out that caching unlocked glocks is useful in some
situations, we can change the caching strategy.
It is currently unclear if a glock that has no active references can
have the GLF_LFLUSH flag set. To make sure that such a glock won't
accidentally be evicted due to memory pressure, we add a GLF_LFLUSH
check to gfs2_dispose_glock_lru().
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
'timeout' is a vague name for the return value of wait_event_*_timeout
because it actually returns the time left. Because the variable is never
used later, just drop the return value. Since variable 'timeout' is then
only used to carry a fixed timeout value, drop this in favor of a fixed
function argument as in the other call to wait_event_timeout() above.
Signed-off-by: Wolfram Sang <wsa+renesas@sang-engineering.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Function gfs2_freeze_unlock() is always called with &sdp->sd_freeze_gh
as its argument, so clean up the code by passing in sdp instead.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
When a DLM lockspace is released and there ares still locks in that
lockspace, DLM will unlock those locks automatically. Commit
fb6791d100d1b started exploiting this behavior to speed up filesystem
unmount: gfs2 would simply free glocks it didn't want to unlock and then
release the lockspace. This didn't take the bast callbacks for
asynchronous lock contention notifications into account, which remain
active until until a lock is unlocked or its lockspace is released.
To prevent those callbacks from accessing deallocated objects, put the
glocks that should not be unlocked on the sd_dead_glocks list, release
the lockspace, and only then free those glocks.
As an additional measure, ignore unexpected ast and bast callbacks if
the receiving glock is dead.
Fixes: fb6791d100d1b ("GFS2: skip dlm_unlock calls in unmount")
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Cc: David Teigland <teigland@redhat.com>
|
|
Function gfs2_glock_queue_put() puts a glock reference by enqueuing
glock work instead of putting the reference directly. This ensures that
the operation won't sleep, but it is costly and really only necessary
when putting the final glock reference. Replace it with a new
gfs2_glock_put_async() function that only queues glock work when putting
the last glock reference.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
In gfs2_jindex_free(), set sdp->sd_jdesc to NULL under the log flush
lock to provide exclusion against gfs2_log_flush().
In gfs2_log_flush(), check if sdp->sd_jdesc is non-NULL before
dereferencing it. Otherwise, we could run into a NULL pointer
dereference when outstanding glock work races with an unmount
(glock_work_func -> run_queue -> do_xmote -> inode_go_sync ->
gfs2_log_flush).
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Functions gfs2_freeze_super() and gfs2_thaw_super() are using the
SDF_FROZEN flag to indicate when the filesystem is frozen, synchronized
by sd_freeze_mutex. However, this doesn't prevent writes from happening
between the point of calling thaw_super() and the point where the
SDF_FROZEN flag is cleared, so the following assert can trigger in
log_write_header():
gfs2_assert_withdraw(sdp, !test_bit(SDF_FROZEN, &sdp->sd_flags));
Fix that by checking for sb->s_writers.frozen != SB_FREEZE_COMPLETE in
log_write_header() instead. To make sure that the filesystem-specific
part of freezing happens before sb->s_writers.frozen is set to
SB_FREEZE_COMPLETE, move that code from gfs2_freeze_locally() into
gfs2_freeze_fs() and hook that up to the .freeze_fs operation.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
It turns out that the .freeze_super and .thaw_super operations require
the filesystem to manage the superblock refcount itself. We are using
the freeze_super() and thaw_super() helpers to mostly take care of that
for us, but this means that the superblock may no longer be around by
when thaw_super() returns, and gfs2_thaw_super() will then access freed
memory. Take an extra superblock reference in gfs2_thaw_super() to fix
that.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
This minor cleanup to gfs2_freeze_super() and gfs2_thaw_super() prepares
for the following refcounting fix.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
This function checks whether the filesystem has been been marked to be
withdrawn eventually or has been withdrawn already. Rename this
function to avoid confusing code like checking for gfs2_withdrawing()
when gfs2_withdrawn() has already returned true.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Mark the gfs2_withdrawn(), gfs2_withdrawing(), and
gfs2_withdraw_in_prog() inline functions as likely to return %false.
This allows to get rid of likely() and unlikely() annotations at the
call sites of those functions.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
... instead of reimplementing it with misguiding name (is_ancestor(x, y)
would normally imply "x is an ancestor of y", not the other way round).
With races, while we are at it...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/gfs2/linux-gfs2
Pull gfs2 updates from Andreas Gruenbacher:
- Don't update inode timestamps for direct writes (performance
regression fix)
- Skip no-op quota records instead of panicing
- Fix a RCU race in gfs2_permission()
- Various other smaller fixes and cleanups all over the place
* tag 'gfs2-v6.6-rc2-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/gfs2/linux-gfs2: (24 commits)
gfs2: don't withdraw if init_threads() got interrupted
gfs2: remove dead code in add_to_queue
gfs2: Fix slab-use-after-free in gfs2_qd_dealloc
gfs2: Silence "suspicious RCU usage in gfs2_permission" warning
gfs2: fs: derive f_fsid from s_uuid
gfs2: No longer use 'extern' in function declarations
gfs2: Rename gfs2_lookup_{ simple => meta }
gfs2: Convert gfs2_internal_read to folios
gfs2: Convert stuffed_readpage to folios
gfs2: Minor gfs2_write_jdata_batch PAGE_SIZE cleanup
gfs2: Get rid of gfs2_alloc_blocks generation parameter
gfs2: Add metapath_dibh helper
gfs2: Clean up quota.c:print_message
gfs2: Clean up gfs2_alloc_parms initializers
gfs2: Two quota=account mode fixes
gfs2: Stop using GFS2_BASIC_BLOCK and GFS2_BASIC_BLOCK_SHIFT
gfs2: setattr_chown: Add missing initialization
gfs2: fix an oops in gfs2_permission
gfs2: ignore negated quota changes
gfs2: Don't update inode timestamps for direct writes
...
|
|
In gfs2_put_super(), whether withdrawn or not, the quota should
be cleaned up by gfs2_quota_cleanup().
Otherwise, struct gfs2_sbd will be freed before gfs2_qd_dealloc (rcu
callback) has run for all gfs2_quota_data objects, resulting in
use-after-free.
Also, gfs2_destroy_threads() and gfs2_quota_cleanup() is already called
by gfs2_make_fs_ro(), so in gfs2_put_super(), after calling
gfs2_make_fs_ro(), there is no need to call them again.
Reported-by: syzbot+29c47e9e51895928698c@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=29c47e9e51895928698c
Signed-off-by: Juntong Deng <juntong.deng@outlook.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
gfs2 already has optional persistent uuid.
Use that uuid to report f_fsid in statfs(2), same as ext2/ext4/zonefs.
This allows gfs2 to be monitored by fanotify filesystem watch.
for example, with inotify-tools 4.23.8.0, the following command can be
used to watch changes over entire filesystem:
fsnotifywatch --filesystem /mnt/gfs2
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
For non-static function declarations, external linkage is implied and
the 'extern' keyword isn't needed. Some static checkers complain about
the overuse of 'extern', so clean up all the function declarations.
In addition, remove 'extern' from the definition of
free_local_statfs_inodes(); it isn't needed there, either.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Convert to using the new inode timestamp accessor functions.
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Link: https://lore.kernel.org/r/20231004185347.80880-38-jlayton@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
|
|
In RCU mode, we might race with gfs2_evict_inode(), which zeroes
->i_gl. Freeing of the object it points to is RCU-delayed, so
if we manage to fetch the pointer before it's been replaced with
NULL, we are fine. Check if we'd fetched NULL and treat that
as "bail out and tell the caller to get out of RCU mode".
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
When trying to upgrade the iopen glock, gfs2_upgrade_iopen_glock() tries
to take the iopen glock with the LM_FLAG_TRY_1CB flag set before trying
to take it without the LM_FLAG_TRY or LM_FLAG_TRY_1CB flags set. Both
calls will cause the lock contention bast callbacks to be invoked
throughout the cluster, and we really don't need them to be invoked
twice. Remove the first LM_FLAG_TRY_1CB call to eliminate unnecessary
dlm traffic.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/gfs2/linux-gfs2
Pull gfs2 updates from Andreas Gruenbacher:
- Fix a glock state (non-)transition bug when a dlm request times out
and is canceled, and we have locking requests that can now be granted
immediately
- Various fixes and cleanups in how the logd and quotad daemons are
woken up and terminated
- Fix several bugs in the quota data reference counting and shrinking.
Free quota data objects synchronously in put_super() instead of
letting call_rcu() run wild
- Make sure not to deallocate quota data during a withdraw; rather,
defer quota data deallocation to put_super(). Withdraws can happen in
contexts in which callers on the stack are holding quota data
references
- Many minor quota fixes and cleanups by Bob
- Update the the mailing list address for gfs2 and dlm. (It's the same
list for both and we are moving it to gfs2@lists.linux.dev)
- Various other minor cleanups
* tag 'gfs2-v6.5-rc5-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/gfs2/linux-gfs2: (51 commits)
MAINTAINERS: Update dlm mailing list
MAINTAINERS: Update gfs2 mailing list
gfs2: change qd_slot_count to qd_slot_ref
gfs2: check for no eligible quota changes
gfs2: Remove useless assignment
gfs2: simplify slot_get
gfs2: Simplify qd2offset
gfs2: introduce qd_bh_get_or_undo
gfs2: Remove quota allocation info from quota file
gfs2: use constant for array size
gfs2: Set qd_sync_gen in do_sync
gfs2: Remove useless err set
gfs2: Small gfs2_quota_lock cleanup
gfs2: move qdsb_put and reduce redundancy
gfs2: improvements to sysfs status
gfs2: Don't try to sync non-changes
gfs2: Simplify function need_sync
gfs2: remove unneeded pg_oflow variable
gfs2: remove unneeded variable done
gfs2: pass sdp to gfs2_write_buf_to_page
...
|
|
This patch adds a new mount option quota=quiet which is the same as
quota=on but it suppresses gfs2 quota error messages.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
The kernel threads are currently stopped and destroyed synchronously by
gfs2_make_fs_ro() and gfs2_put_super(), and asynchronously by
signal_our_withdraw(), with no synchronization, so the synchronous and
asynchronous contexts can race with each other.
First, when creating the kernel threads, take an extra task struct
reference so that the task struct won't go away immediately when they
terminate. This allows those kthreads to terminate immediately when
they're done rather than hanging around as zombies until they are reaped
by kthread_stop(). When kthread_stop() is called on a terminated
kthread, it will return immediately.
Second, in signal_our_withdraw(), once the SDF_JOURNAL_LIVE flag has
been cleared, wake up the logd and quotad wait queues instead of
stopping the logd and quotad kthreads. The kthreads are then expected
to terminate automatically within short time, but if they cannot, they
will not block the withdraw.
For example, if a user process and one of the kthread decide to withdraw
at the same time, only one of them will perform the actual withdraw and
the other will wait for it to be done. If the kthread ends up being the
one to wait, the withdrawing user process won't be able to stop it.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
[ 81.372851][ T5532] CPU: 1 PID: 5532 Comm: syz-executor.0 Not tainted 6.2.0-rc1-syzkaller-dirty #0
[ 81.382080][ T5532] Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/12/2023
[ 81.392343][ T5532] Call Trace:
[ 81.395654][ T5532] <TASK>
[ 81.398603][ T5532] dump_stack_lvl+0x1b1/0x290
[ 81.418421][ T5532] gfs2_assert_warn_i+0x19a/0x2e0
[ 81.423480][ T5532] gfs2_quota_cleanup+0x4c6/0x6b0
[ 81.428611][ T5532] gfs2_make_fs_ro+0x517/0x610
[ 81.457802][ T5532] gfs2_withdraw+0x609/0x1540
[ 81.481452][ T5532] gfs2_inode_refresh+0xb2d/0xf60
[ 81.506658][ T5532] gfs2_instantiate+0x15e/0x220
[ 81.511504][ T5532] gfs2_glock_wait+0x1d9/0x2a0
[ 81.516352][ T5532] do_sync+0x485/0xc80
[ 81.554943][ T5532] gfs2_quota_sync+0x3da/0x8b0
[ 81.559738][ T5532] gfs2_sync_fs+0x49/0xb0
[ 81.564063][ T5532] sync_filesystem+0xe8/0x220
[ 81.568740][ T5532] generic_shutdown_super+0x6b/0x310
[ 81.574112][ T5532] kill_block_super+0x79/0xd0
[ 81.578779][ T5532] deactivate_locked_super+0xa7/0xf0
[ 81.584064][ T5532] cleanup_mnt+0x494/0x520
[ 81.593753][ T5532] task_work_run+0x243/0x300
[ 81.608837][ T5532] exit_to_user_mode_loop+0x124/0x150
[ 81.614232][ T5532] exit_to_user_mode_prepare+0xb2/0x140
[ 81.619820][ T5532] syscall_exit_to_user_mode+0x26/0x60
[ 81.625287][ T5532] do_syscall_64+0x49/0xb0
[ 81.629710][ T5532] entry_SYSCALL_64_after_hwframe+0x63/0xcd
In this backtrace, gfs2_quota_sync() takes quota data references and
then calls do_sync(). Function do_sync() encounters filesystem
corruption and withdraws the filesystem, which (among other things) calls
gfs2_quota_cleanup(). Function gfs2_quota_cleanup() wrongly assumes
that nobody is holding any quota data references anymore, and destroys
all quota data objects. When gfs2_quota_sync() then resumes and
dereferences the quota data objects it is holding, those objects are no
longer there.
Function gfs2_quota_cleanup() deals with resource deallocation and can
easily be delayed until gfs2_put_super() in the case of a filesystem
withdraw. In fact, most of the other work gfs2_make_fs_ro() does is
unnecessary during a withdraw as well, so change signal_our_withdraw()
to skip gfs2_make_fs_ro() and perform the necessary steps directly
instead.
Thanks to Edward Adam Davis <eadavis@sina.com> for the initial patches.
Link: https://lore.kernel.org/all/0000000000002b5e2405f14e860f@google.com
Reported-by: syzbot+3f6a670108ce43356017@syzkaller.appspotmail.com
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Rename the SDF_DEACTIVATING flag to SDF_KILL to make it more obvious
that this relates to the kill_sb filesystem operation.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs
Pull superblock updates from Christian Brauner:
"This contains the super rework that was ready for this cycle. The
first part changes the order of how we open block devices and allocate
superblocks, contains various cleanups, simplifications, and a new
mechanism to wait on superblock state changes.
This unblocks work to ultimately limit the number of writers to a
block device. Jan has already scheduled follow-up work that will be
ready for v6.7 and allows us to restrict the number of writers to a
given block device. That series builds on this work right here.
The second part contains filesystem freezing updates.
Overview:
The generic superblock changes are rougly organized as follows
(ignoring additional minor cleanups):
(1) Removal of the bd_super member from struct block_device.
This was a very odd back pointer to struct super_block with
unclear rules. For all relevant places we have other means to get
the same information so just get rid of this.
(2) Simplify rules for superblock cleanup.
Roughly, everything that is allocated during fs_context
initialization and that's stored in fs_context->s_fs_info needs
to be cleaned up by the fs_context->free() implementation before
the superblock allocation function has been called successfully.
After sget_fc() returned fs_context->s_fs_info has been
transferred to sb->s_fs_info at which point sb->kill_sb() if
fully responsible for cleanup. Adhering to these rules means that
cleanup of sb->s_fs_info in fill_super() is to be avoided as it's
brittle and inconsistent.
Cleanup shouldn't be duplicated between sb->put_super() as
sb->put_super() is only called if sb->s_root has been set aka
when the filesystem has been successfully born (SB_BORN). That
complexity should be avoided.
This also means that block devices are to be closed in
sb->kill_sb() instead of sb->put_super(). More details in the
lower section.
(3) Make it possible to lookup or create a superblock before opening
block devices
There's a subtle dependency on (2) as some filesystems did rely
on fill_super() to be called in order to correctly clean up
sb->s_fs_info. All these filesystems have been fixed.
(4) Switch most filesystem to follow the same logic as the generic
mount code now does as outlined in (3).
(5) Use the superblock as the holder of the block device. We can now
easily go back from block device to owning superblock.
(6) Export and extend the generic fs_holder_ops and use them as
holder ops everywhere and remove the filesystem specific holder
ops.
(7) Call from the block layer up into the filesystem layer when the
block device is removed, allowing to shut down the filesystem
without risk of deadlocks.
(8) Get rid of get_super().
We can now easily go back from the block device to owning
superblock and can call up from the block layer into the
filesystem layer when the device is removed. So no need to wade
through all registered superblock to find the owning superblock
anymore"
Link: https://lore.kernel.org/lkml/20230824-prall-intakt-95dbffdee4a0@brauner/
* tag 'v6.6-vfs.super' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs: (47 commits)
super: use higher-level helper for {freeze,thaw}
super: wait until we passed kill super
super: wait for nascent superblocks
super: make locking naming consistent
super: use locking helpers
fs: simplify invalidate_inodes
fs: remove get_super
block: call into the file system for ioctl BLKFLSBUF
block: call into the file system for bdev_mark_dead
block: consolidate __invalidate_device and fsync_bdev
block: drop the "busy inodes on changed media" log message
dasd: also call __invalidate_device when setting the device offline
amiflop: don't call fsync_bdev in FDFMTBEG
floppy: call disk_force_media_change when changing the format
block: simplify the disk_force_media_change interface
nbd: call blk_mark_disk_dead in nbd_clear_sock_ioctl
xfs use fs_holder_ops for the log and RT devices
xfs: drop s_umount over opening the log and RT devices
ext4: use fs_holder_ops for the log device
ext4: drop s_umount over opening the log device
...
|
|
In later patches, we're going to change how the inode's ctime field is
used. Switch to using accessor functions instead of raw accesses of
inode->i_ctime.
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Message-Id: <20230705190309.579783-45-jlayton@kernel.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
|
|
Userspace can freeze a filesystem using the FIFREEZE ioctl or by
suspending the block device; this state persists until userspace thaws
the filesystem with the FITHAW ioctl or resuming the block device.
Since commit 18e9e5104fcd ("Introduce freeze_super and thaw_super for
the fsfreeze ioctl") we only allow the first freeze command to succeed.
The kernel may decide that it is necessary to freeze a filesystem for
its own internal purposes, such as suspends in progress, filesystem fsck
activities, or quiescing a device prior to removal. Userspace thaw
commands must never break a kernel freeze, and kernel thaw commands
shouldn't undo userspace's freeze command.
Introduce a couple of freeze holder flags and wire it into the
sb_writers state. One kernel and one userspace freeze are allowed to
coexist at the same time; the filesystem will not thaw until both are
lifted.
I wonder if the f2fs/gfs2 code should be using a kernel freeze here, but
for now we'll use FREEZE_HOLDER_USERSPACE to preserve existing
behaviors.
Cc: mcgrof@kernel.org
Cc: jack@suse.cz
Cc: hch@infradead.org
Cc: ruansy.fnst@fujitsu.com
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Jan Kara <jack@suse.cz>
|
|
All the remaining users of gfs2_freeze_lock_shared() set freeze_gh to
&sdp->sd_freeze_gh and flags to 0, so remove those two parameters.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Replace sd_freeze_state with a new SDF_FROZEN flag.
There no longer is a need for indicating that a freeze is in progress
(SDF_STARTING_FREEZE); we are now protecting the critical sections with
the sd_freeze_mutex.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
So far, at mount time, gfs2 would take the freeze glock in shared mode
and then immediately drop it again, turning it into a cached glock that
can be reclaimed at any time. To freeze the filesystem cluster-wide,
the node initiating the freeze would take the freeze glock in exclusive
mode, which would cause the freeze glock's freeze_go_sync() callback to
run on each node. There, gfs2 would freeze the filesystem and schedule
gfs2_freeze_func() to run. gfs2_freeze_func() would re-acquire the
freeze glock in shared mode, thaw the filesystem, and drop the freeze
glock again. The initiating node would keep the freeze glock held in
exclusive mode. To thaw the filesystem, the initiating node would drop
the freeze glock again, which would allow gfs2_freeze_func() to resume
on all nodes, leaving the filesystem in the thawed state.
It turns out that in freeze_go_sync(), we cannot reliably and safely
freeze the filesystem. This is primarily because the final unmount of a
filesystem takes a write lock on the s_umount rw semaphore before
calling into gfs2_put_super(), and freeze_go_sync() needs to call
freeze_super() which also takes a write lock on the same semaphore,
causing a deadlock. We could work around this by trying to take an
active reference on the super block first, which would prevent unmount
from running at the same time. But that can fail, and freeze_go_sync()
isn't actually allowed to fail.
To get around this, this patch changes the freeze glock locking scheme
as follows:
At mount time, each node takes the freeze glock in shared mode. To
freeze a filesystem, the initiating node first freezes the filesystem
locally and then drops and re-acquires the freeze glock in exclusive
mode. All other nodes notice that there is contention on the freeze
glock in their go_callback callbacks, and they schedule
gfs2_freeze_func() to run. There, they freeze the filesystem locally
and drop and re-acquire the freeze glock before re-thawing the
filesystem. This is happening outside of the glock state engine, so
there, we are allowed to fail.
From a cluster point of view, taking and immediately dropping a glock is
indistinguishable from taking the glock and only dropping it upon
contention, so this new scheme is compatible with the old one.
Thanks to Li Dong <lidong@vivo.com> for reporting a locking bug in
gfs2_freeze_func() in a previous version of this commit.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Rename the SDF_FS_FROZEN flag to SDF_FREEZE_INITIATOR to indicate more
clearly that the node that has this flag set is the initiator of the
freeze.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com
|
|
Rename gfs2_freeze_lock to gfs2_freeze_lock_shared to make it a bit more
obvious that this function establishes the "thawed" state of the freeze
glock.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Rename gfs2_freeze to gfs2_freeze_super and gfs2_unfreeze to
gfs2_thaw_super to match the names of the corresponding super
operations.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
The transaction glock was repurposed to serve as the new freeze glock
years ago. Don't refer to it as the transaction glock anymore.
Also, to be more precise, call it the "freeze glock" instead of the
"freeze lock". Ditto for the journal glock.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Some fields such as gt_logd_secs of the struct gfs2_tune are accessed
without holding the lock gt_spin in gfs2_show_options():
val = sdp->sd_tune.gt_logd_secs;
if (val != 30)
seq_printf(s, ",commit=%d", val);
And thus can cause data races when gfs2_show_options() and other functions
such as gfs2_reconfigure() are concurrently executed:
spin_lock(>->gt_spin);
gt->gt_logd_secs = newargs->ar_commit;
To fix these possible data races, the lock sdp->sd_tune.gt_spin is
acquired before accessing the fields of gfs2_tune and released after these
accesses.
Further changes by Andreas:
- Don't hold the spin lock over the seq_printf operations.
Reported-by: BassCheck <bass@buaa.edu.cn>
Signed-off-by: Tuo Li <islituo@gmail.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
This patch changes function evict_unlinked_inode so it does not call
gfs2_inode_remember_delete until it gets a good return code from
gfs2_dinode_dealloc.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Before this patch, function gfs2_dinode_dealloc would abort if it got a
bad return code from gfs2_rindex_update(). The problem is that it left the
dinode in the unlinked (not free) state, which meant subsequent fsck
would clean it up and flag an error. That meant some of our QE tests
would fail.
The sole purpose of gfs2_rindex_update(), in this code path, is to read in
any newer rgrps added by gfs2_grow. But since this is a delete operation
it won't actually use any of those new rgrps. It can really only twiddle
the bits from "Unlinked" to "Free" in an existing rgrp. Therefore the
error should not prevent the transition from unlinked to free.
This patch makes gfs2_dinode_dealloc ignore the bad return code and
proceed with freeing the dinode so the QE tests will not be tripped up.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
On corrupt gfs2 file systems the evict code can try to reference the
journal descriptor structure, jdesc, after it has been freed and set to
NULL. The sequence of events is:
init_journal()
...
fail_jindex:
gfs2_jindex_free(sdp); <------frees journals, sets jdesc = NULL
if (gfs2_holder_initialized(&ji_gh))
gfs2_glock_dq_uninit(&ji_gh);
fail:
iput(sdp->sd_jindex); <--references jdesc in evict_linked_inode
evict()
gfs2_evict_inode()
evict_linked_inode()
ret = gfs2_trans_begin(sdp, 0, sdp->sd_jdesc->jd_blocks);
<------references the now freed/zeroed sd_jdesc pointer.
The call to gfs2_trans_begin is done because the truncate_inode_pages
call can cause gfs2 events that require a transaction, such as removing
journaled data (jdata) blocks from the journal.
This patch fixes the problem by adding a check for sdp->sd_jdesc to
function gfs2_evict_inode. In theory, this should only happen to corrupt
gfs2 file systems, when gfs2 detects the problem, reports it, then tries
to evict all the system inodes it has read in up to that point.
Reported-by: Yang Lan <lanyang0908@gmail.com>
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Before this patch, function gfs2_make_fs_ro called gfs2_log_flush once to
finalize the log. However, if there's dirty metadata, log flushes tend
to sync the metadata and formulate revokes. Before this patch, those
revokes may not be written out to the journal immediately, which meant
unresolved glocks could still have revokes in their ail lists. When the
glock worker runs, it tries to transition the glock, but the unresolved
revokes in the ail still need to be written, so it tries to start a
transaction. It's impossible to start a transaction because at that
point, the SDF_JOURNAL_LIVE flag has been cleared by gfs2_make_fs_ro.
That causes the glock worker to fail, unable to write the revokes. The
calling sequence looked something like this:
gfs2_make_fs_ro
gfs2_log_flush - with GFS2_LOG_HEAD_FLUSH_SHUTDOWN flag set
if (flags & GFS2_LOG_HEAD_FLUSH_SHUTDOWN)
clear_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags);
...meanwhile...
glock_work_func
do_xmote
rgrp_go_sync (or possibly inode_go_sync)
...
gfs2_ail_empty_gl
__gfs2_trans_begin
if (unlikely(!test_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags))) {
...
return -EROFS;
The previous patch in the series ("gfs2: return errors from
gfs2_ail_empty_gl") now causes the transaction error to no longer be
ignored, so it causes a warning from MOST of the xfstests:
WARNING: CPU: 11 PID: X at fs/gfs2/super.c:603 gfs2_put_super [gfs2]
which corresponds to:
WARN_ON(gfs2_withdrawing(sdp));
The withdraw was triggered silently from do_xmote by:
if (unlikely(sdp->sd_log_error && !gfs2_withdrawn(sdp)))
gfs2_withdraw_delayed(sdp);
This patch adds a second log_flush to gfs2_make_fs_ro: one to sync the
data and one to sync any outstanding revokes and finalize the journal.
Note that both of these log flushes need to be "special," in other
words, not GFS2_LOG_HEAD_FLUSH_NORMAL.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
In gfs2_make_fs_rw(), make sure to call gfs2_consist() to report an
inconsistency and mark the filesystem as withdrawn when
gfs2_find_jhead() fails.
At the end of gfs2_make_fs_rw(), when we discover that the filesystem
has been withdrawn, make sure we report an error. This also replaces
the gfs2_withdrawn() check after gfs2_find_jhead().
Reported-by: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Cc: syzbot+f51cb4b9afbd87ec06f2@syzkaller.appspotmail.com
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Add a gfs2_evict_inodes() helper that evicts inodes cooperatively across
the cluster. This avoids running into timeouts during unmount
unnecessarily.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
In gfs2_kill_sb(), flush the delete work queue after setting the
SDF_DEACTIVATING flag. This ensures that no new inodes will be
instantiated anymore, and the inode cache will be empty after the
following kill_block_super() -> generic_shutdown_super() ->
evict_inodes() call.
With that, function gfs2_make_fs_ro() now calls gfs2_flush_delete_work()
after the workqueue has been destroyed. Skip that by checking for the
presence of the SDF_DEACTIVATING flag.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Function delete_work_func() is used for two purposes:
* to immediately try to evict the glock's inode, and
* to verify after a little while that the inode has been deleted as
expected, and didn't just get skipped.
These two operations are not separated very well, so introduce two new
glock flags to improved that. Split gfs2_queue_delete_work() into
gfs2_queue_try_to_evict and gfs2_queue_verify_evict().
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Move the global delete workqueue into struct gfs2_sbd so that we can
flush / drain it without interfering with other filesystems.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Improve the comment describing the inode and iopen glock interactions
and the glock poking related to inode evict.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Function glock_clear_object() checks if the specified glock is still
pointing at the right object and clears the gl_object pointer. To
handle the case of incompletely constructed inodes, glock_clear_object()
also allows gl_object to be NULL.
However, in the teardown case, when iget_failed() is called and the
inode is removed from the inode hash, by the time we get to the
glock_clear_object() calls in gfs2_put_super() and its helpers, we don't
have exclusion against concurrent gfs2_inode_lookup() and
gfs2_create_inode() calls, and the inode and iopen glocks may already be
pointing at another inode, so the checks in glock_clear_object() are
incorrect.
To better handle this case, always completely disassociate an inode from
its glocks before tearing it down. In addition, get rid of a duplicate
glock_clear_object() call in gfs2_evict_inode(). That way,
glock_clear_object() will only ever be called when the glock points at
the current inode, and the NULL check in glock_clear_object() can be
removed.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
With the previous change, to simplify things, we can always just dequeue
and uninitialize the iopen glock in gfs2_evict_inode() even if it isn't
queued anymore.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
Since commit 3d36e57ff768 ("gfs2: gfs2_create_inode rework"),
gfs2_evict_inode() and gfs2_create_inode() / gfs2_inode_lookup() will
synchronize via the inode hash table and we can be certain that once a
new inode is inserted into the inode hash table(), gfs2_evict_inode()
has completely destroyed any previous versions. We no longer need to
worry about overlapping inode object lifespans. Update the code and
comments accordingly.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
In each of the two functions, add an inode variable that points to
&ip->i_inode and use that throughout the rest of the function.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
|
When gfs2_create_inode() fails after creating a new inode, it uses the
GIF_FREE_VFS_INODE and GIF_ALLOC_FAILED inode flags to communicate to
gfs2_evict_inode() which parts of the inode need to be deallocated and
destroyed. In some error cases, the inode ends up being allocated on
disk and then accidentally left behind. In others, the inode is
partially constructed and then not properly destroyed. Clean this up by
completely handling the inode deallocation and destruction in
gfs2_evict_inode().
This means that gfs2_evict_inode() may now be faced with partially
constructed inodes, so add the necessary checks to cope with that. In
particular, make sure that for incompletely constructed inodes, we're
not accessing the buffers backing the on-disk blocks; the contents may
be undefined.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
|
