Given how rbd_lock_add_request() and rbd_img_exclusive_lock() are
written, lock_dwork may be (re)queued more than it's actually needed:
for example in case a new I/O request comes in while we are in the
middle of rbd_acquire_lock() on behalf of another I/O request. This is
expected and with rbd_release_lock() preemptively canceling lock_dwork
is benign under normal operation.
A more problematic example is maybe_kick_acquire():
if (have_requests || delayed_work_pending(&rbd_dev->lock_dwork)) {
dout("%s rbd_dev %p kicking lock_dwork\n", __func__, rbd_dev);
mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork, 0);
}
It's not unrealistic for lock_dwork to get canceled right after
delayed_work_pending() returns true and for mod_delayed_work() to
requeue it right there anyway. This is a classic TOCTOU race.
When it comes to unmapping the image, there is an implicit assumption
of no self-initiated exclusive lock activity past the point of return
from rbd_dev_image_unlock() which unlocks the lock if it happens to be
held. This unlock is assumed to be final and lock_dwork (as well as
all other exclusive lock tasks, really) isn't expected to get queued
again. However, lock_dwork is canceled only in cancel_tasks_sync()
(i.e. later in the unmap sequence) and on top of that the cancellation
can get in effect nullified by maybe_kick_acquire(). This may result
in rbd_acquire_lock() executing after rbd_dev_device_release() and
rbd_dev_image_release() run and free and/or reset a bunch of things.
One of the possible failure modes then is a violated
rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
in rbd_dev_header_info() which is called via rbd_dev_refresh() from
rbd_post_acquire_action().
Redo exclusive lock task draining to provide saner semantics and try
to meet the assumptions around rbd_dev_image_unlock().
Cc: stable@vger.kernel.org
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>
This was done entirely with mindless brute force, using
git grep -l '\<k[vmz]*alloc_objs*(.*, GFP_KERNEL)' |
xargs sed -i 's/\(alloc_objs*(.*\), GFP_KERNEL)/\1)/'
to convert the new alloc_obj() users that had a simple GFP_KERNEL
argument to just drop that argument.
Note that due to the extreme simplicity of the scripting, any slightly
more complex cases spread over multiple lines would not be triggered:
they definitely exist, but this covers the vast bulk of the cases, and
the resulting diff is also then easier to check automatically.
For the same reason the 'flex' versions will be done as a separate
conversion.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is the result of running the Coccinelle script from
scripts/coccinelle/api/kmalloc_objs.cocci. The script is designed to
avoid scalar types (which need careful case-by-case checking), and
instead replace kmalloc-family calls that allocate struct or union
object instances:
Single allocations: kmalloc(sizeof(TYPE), ...)
are replaced with: kmalloc_obj(TYPE, ...)
Array allocations: kmalloc_array(COUNT, sizeof(TYPE), ...)
are replaced with: kmalloc_objs(TYPE, COUNT, ...)
Flex array allocations: kmalloc(struct_size(PTR, FAM, COUNT), ...)
are replaced with: kmalloc_flex(*PTR, FAM, COUNT, ...)
(where TYPE may also be *VAR)
The resulting allocations no longer return "void *", instead returning
"TYPE *".
Signed-off-by: Kees Cook <kees@kernel.org>
Similar to commit 870611e487 ("rbd: get snapshot context after
exclusive lock is ensured to be held"), move the "beyond EOD" check
into the image request state machine so that it's performed after
exclusive lock is ensured to be held. This avoids various race
conditions which can arise when the image is shrunk under I/O (in
practice, mostly readahead). In one such scenario
rbd_assert(objno < rbd_dev->object_map_size);
can be triggered if a close-to-EOD read gets queued right before the
shrink is initiated and the EOD check is performed against an outdated
mapping_size. After the resize is done on the server side and exclusive
lock is (re)acquired bringing along the new (now shrunk) object map, the
read starts going through the state machine and rbd_obj_may_exist() gets
invoked on an object that is out of bounds of rbd_dev->object_map array.
Cc: stable@vger.kernel.org
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Dongsheng Yang <dongsheng.yang@linux.dev>
Currently if a user enqueue a work item using schedule_delayed_work() the
used wq is "system_wq" (per-cpu wq) while queue_delayed_work() use
WORK_CPU_UNBOUND (used when a cpu is not specified). The same applies to
schedule_work() that is using system_wq and queue_work(), that makes use
again of WORK_CPU_UNBOUND.
This lack of consistentcy cannot be addressed without refactoring the API.
alloc_workqueue() treats all queues as per-CPU by default, while unbound
workqueues must opt-in via WQ_UNBOUND.
This default is suboptimal: most workloads benefit from unbound queues,
allowing the scheduler to place worker threads where they’re needed and
reducing noise when CPUs are isolated.
This default is suboptimal: most workloads benefit from unbound queues,
allowing the scheduler to place worker threads where they’re needed and
reducing noise when CPUs are isolated.
This patch adds a new WQ_PERCPU flag to explicitly request the use of
the per-CPU behavior. Both flags coexist for one release cycle to allow
callers to transition their calls.
Once migration is complete, WQ_UNBOUND can be removed and unbound will
become the implicit default.
With the introduction of the WQ_PERCPU flag (equivalent to !WQ_UNBOUND),
any alloc_workqueue() caller that doesn’t explicitly specify WQ_UNBOUND
must now use WQ_PERCPU.
All existing users have been updated accordingly.
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Marco Crivellari <marco.crivellari@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When block drivers or the core block code perform allocations with a
frozen queue, this could try to recurse into the block device to
reclaim memory and deadlock. Thus all allocations done by a process
that froze a queue need to be done without __GFP_IO and __GFP_FS.
Instead of tying to track all of them down, force a noio scope as
part of freezing the queue.
Note that nvme is a bit of a mess here due to the non-owner freezes,
and they will be addressed separately.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20250131120352.1315351-2-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
BLK_MQ_F_SHOULD_MERGE is set for all tag_sets except those that purely
process passthrough commands (bsg-lib, ufs tmf, various nvme admin
queues) and thus don't even check the flag. Remove it to simplify the
driver interface.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20241219060214.1928848-1-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pull ceph updates from Ilya Dryomov:
"A small patchset to address bogus I/O errors and ultimately an
assertion failure in the face of watch errors with -o exclusive
mappings in RBD marked for stable and some assorted CephFS fixes"
* tag 'ceph-for-6.11-rc1' of https://github.com/ceph/ceph-client:
rbd: don't assume rbd_is_lock_owner() for exclusive mappings
rbd: don't assume RBD_LOCK_STATE_LOCKED for exclusive mappings
rbd: rename RBD_LOCK_STATE_RELEASING and releasing_wait
ceph: fix incorrect kmalloc size of pagevec mempool
ceph: periodically flush the cap releases
ceph: convert comma to semicolon in __ceph_dentry_dir_lease_touch()
ceph: use cap_wait_list only if debugfs is enabled
Expanding on the previous commit, assuming that rbd_is_lock_owner()
always returns true (i.e. that we are either in RBD_LOCK_STATE_LOCKED
or RBD_LOCK_STATE_QUIESCING) if the mapping is exclusive is wrong too.
In case ceph_cls_set_cookie() fails, the lock would be temporarily
released even if the mapping is exclusive, meaning that we can end up
even in RBD_LOCK_STATE_UNLOCKED.
IOW, exclusive mappings are really "just" about disabling automatic
lock transitions (as documented in the man page), not about grabbing
the lock and holding on to it whatever it takes.
Cc: stable@vger.kernel.org
Fixes: 637cd06053 ("rbd: new exclusive lock wait/wake code")
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Dongsheng Yang <dongsheng.yang@easystack.cn>
Every time a watch is reestablished after getting lost, we need to
update the cookie which involves quiescing exclusive lock. For this,
we transition from RBD_LOCK_STATE_LOCKED to RBD_LOCK_STATE_QUIESCING
roughly for the duration of rbd_reacquire_lock() call. If the mapping
is exclusive and I/O happens to arrive in this time window, it's failed
with EROFS (later translated to EIO) based on the wrong assumption in
rbd_img_exclusive_lock() -- "lock got released?" check there stopped
making sense with commit a2b1da0979 ("rbd: lock should be quiesced on
reacquire").
To make it worse, any such I/O is added to the acquiring list before
EROFS is returned and this sets up for violating rbd_lock_del_request()
precondition that the request is either on the running list or not on
any list at all -- see commit ded080c86b ("rbd: don't move requests
to the running list on errors"). rbd_lock_del_request() ends up
processing these requests as if they were on the running list which
screws up quiescing_wait completion counter and ultimately leads to
rbd_assert(!completion_done(&rbd_dev->quiescing_wait));
being triggered on the next watch error.
Cc: stable@vger.kernel.org # 06ef84c4e9c4: rbd: rename RBD_LOCK_STATE_RELEASING and releasing_wait
Cc: stable@vger.kernel.org
Fixes: 637cd06053 ("rbd: new exclusive lock wait/wake code")
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Dongsheng Yang <dongsheng.yang@easystack.cn>
... to RBD_LOCK_STATE_QUIESCING and quiescing_wait to recognize that
this state and the associated completion are backing rbd_quiesce_lock(),
which isn't specific to releasing the lock.
While exclusive lock does get quiesced before it's released, it also
gets quiesced before an attempt to update the cookie is made and there
the lock is not released as long as ceph_cls_set_cookie() succeeds.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Dongsheng Yang <dongsheng.yang@easystack.cn>
Move the stable_writes flag into the queue_limits feature field so that
it can be set atomically with the queue frozen.
The flag is now inherited by blk_stack_limits, which greatly simplifies
the code in dm, and fixed md which previously did not pass on the flag
set on lower devices.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Link: https://lore.kernel.org/r/20240617060532.127975-18-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Move the nonrot flag into the queue_limits feature field so that it can
be set atomically with the queue frozen.
Use the chance to switch to defaulting to non-rotational and require
the driver to opt into rotational, which matches the polarity of the
sysfs interface.
For the z2ram, ps3vram, 2x memstick, ubiblock and dcssblk the new
rotational flag is not set as they clearly are not rotational despite
this being a behavior change. There are some other drivers that
unconditionally set the rotational flag to keep the existing behavior
as they arguably can be used on rotational devices even if that is
probably not their main use today (e.g. virtio_blk and drbd).
The flag is automatically inherited in blk_stack_limits matching the
existing behavior in dm and md.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Link: https://lore.kernel.org/r/20240617060532.127975-15-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The soft max_sectors limit is normally capped by the hardware limits and
an arbitrary upper limit enforced by the kernel, but can be modified by
the user. A few drivers want to increase this limit (nbd, rbd) or
adjust it up or down based on hardware capabilities (sd).
Change blk_validate_limits to default max_sectors to the optimal I/O
size, or upgrade it to the preferred minimal I/O size if that is
larger than the kernel default if no optimal I/O size is provided based
on the logic in the SD driver.
This keeps the existing kernel default for drivers that do not provide
an io_opt or very big io_min value, but picks a much more useful
default for those who provide these hints, and allows to remove the
hacks to set the user max_sectors limit in nbd, rbd and sd.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Acked-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Link: https://lore.kernel.org/r/20240531074837.1648501-5-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Commit 16d80c54ad ("rbd: set io_min, io_opt and discard_granularity to
alloc_size") lowered the io_opt size for rbd from objset_bytes which is
4MB for typical setup to alloc_size which is typically 64KB.
The commit mostly talks about discard behavior and does mention io_min
in passing. Reducing io_opt means reducing the readahead size, which
seems counter-intuitive given that rbd currently abuses the user
max_sectors setting to actually increase the I/O size. Switch back
to the old setting to allow larger reads (the readahead size despite it's
name actually limits the size of any buffered read) and to prepare
for using io_opt in the max_sectors calculation and getting drivers out
of the business of overriding the max_user_sectors value.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Link: https://lore.kernel.org/r/20240531074837.1648501-4-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The running list is supposed to contain requests that are pinning the
exclusive lock, i.e. those that must be flushed before exclusive lock
is released. When wake_lock_waiters() is called to handle an error,
requests on the acquiring list are failed with that error and no
flushing takes place. Briefly moving them to the running list is not
only pointless but also harmful: if exclusive lock gets acquired
before all of their state machines are scheduled and go through
rbd_lock_del_request(), we trigger
rbd_assert(list_empty(&rbd_dev->running_list));
in rbd_try_acquire_lock().
Cc: stable@vger.kernel.org
Fixes: 637cd06053 ("rbd: new exclusive lock wait/wake code")
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Dongsheng Yang <dongsheng.yang@easystack.cn>
ida_alloc() and ida_free() should be preferred to the deprecated
ida_simple_get() and ida_simple_remove().
Note that the upper limit of ida_simple_get() is exclusive, while that
of ida_alloc_max() is inclusive, so 1 has been subtracted.
[ idryomov: tweak changelog ]
Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Reviewed-by: Ilya Dryomov <idryomov@gmail.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
rbd_dev_refresh() has been holding header_rwsem across header and
parent info read-in unnecessarily for ages. With commit 870611e487
("rbd: get snapshot context after exclusive lock is ensured to be
held"), the potential for deadlocks became much more real owning to
a) header_rwsem now nesting inside lock_rwsem and b) rw_semaphores
not allowing new readers after a writer is registered.
For example, assuming that I/O request 1, I/O request 2 and header
read-in request all target the same OSD:
1. I/O request 1 comes in and gets submitted
2. watch error occurs
3. rbd_watch_errcb() takes lock_rwsem for write, clears owner_cid and
releases lock_rwsem
4. after reestablishing the watch, rbd_reregister_watch() calls
rbd_dev_refresh() which takes header_rwsem for write and submits
a header read-in request
5. I/O request 2 comes in: after taking lock_rwsem for read in
__rbd_img_handle_request(), it blocks trying to take header_rwsem
for read in rbd_img_object_requests()
6. another watch error occurs
7. rbd_watch_errcb() blocks trying to take lock_rwsem for write
8. I/O request 1 completion is received by the messenger but can't be
processed because lock_rwsem won't be granted anymore
9. header read-in request completion can't be received, let alone
processed, because the messenger is stranded
Change rbd_dev_refresh() to take header_rwsem only for actually
updating rbd_dev->header. Header and parent info read-in don't need
any locking.
Cc: stable@vger.kernel.org # 0b035401c5: rbd: move rbd_dev_refresh() definition
Cc: stable@vger.kernel.org # 510a7330c8: rbd: decouple header read-in from updating rbd_dev->header
Cc: stable@vger.kernel.org # c10311776f: rbd: decouple parent info read-in from updating rbd_dev
Cc: stable@vger.kernel.org
Fixes: 870611e487 ("rbd: get snapshot context after exclusive lock is ensured to be held")
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Dongsheng Yang <dongsheng.yang@easystack.cn>
Unlike header read-in, parent info read-in is already decoupled in
get_parent_info(), but it's buried in rbd_dev_v2_parent_info() along
with the processing logic.
Separate the initial read-in and update read-in logic into
rbd_dev_setup_parent() and rbd_dev_update_parent() respectively and
have rbd_dev_v2_parent_info() just populate struct parent_image_info
(i.e. what get_parent_info() did). Some existing QoI issues, like
flatten of a standalone clone being disregarded on refresh, remain.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Dongsheng Yang <dongsheng.yang@easystack.cn>
Make rbd_dev_header_info() populate a passed struct rbd_image_header
instead of rbd_dev->header and introduce rbd_dev_update_header() for
updating mutable fields in rbd_dev->header upon refresh. The initial
read-in of both mutable and immutable fields in rbd_dev_image_probe()
passes in rbd_dev->header so no update step is required there.
rbd_init_layout() is now called directly from rbd_dev_image_probe()
instead of individually in format 1 and format 2 implementations.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Dongsheng Yang <dongsheng.yang@easystack.cn>
Ilya Dryomov
Dawei FengandIlya Dryomov
Linus Torvalds
Kees Cook
Marco CrivellariandJens Axboe
Christoph HellwigandJens Axboe
Ming LeiandJens Axboe
Ricardo B. MarliereandJens Axboe
Christophe JAILLETandIlya Dryomov