Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

librbd: Possible crash while concurrently writing and shrinking an image #5318

Merged
32 commits merged into from Aug 30, 2015

Conversation

Projects
None yet
3 participants
@dillaman
Copy link
Contributor

dillaman commented Jul 22, 2015

http://tracker.ceph.com/issues/12236

Note: contains commits included in PR #5296 that can be rebased away once they merge.

@ghost ghost changed the title Possible crash while concurrently writing and shrinking an image DNM: Possible crash while concurrently writing and shrinking an image Jul 28, 2015

@ghost

This comment has been minimized.

Copy link

ghost commented Jul 28, 2015

Marked DNM to ack that it is on top of PR #5280 and PR #5296

@ghost ghost self-assigned this Jul 28, 2015

@ghost

This comment has been minimized.

Copy link

ghost commented Jul 28, 2015

@dillaman could you please rebase now that #5280 is merged ?

@dillaman

This comment has been minimized.

Copy link
Contributor Author

dillaman commented Jul 28, 2015

@dachary should I just wait until #5296 merges as well?

dillaman and others added some commits Apr 8, 2015

WorkQueue: ContextWQ can now accept a return code
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit e5ffae5)
librbd: add work queue for op completions
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 21f990e)
common: add valgrind.h convenience wrapper
Conditionally support helgrind annotations if valgrind support is
enabled during the build.

Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 650ad32)
common: lockdep now support unregistering once destructed
librbd use of an image hierarchy resulted in lock names being
re-used and incorrectly analyzed.  librbd now uses unique lock
names per instance, but to prevent an unbounded growth of
tracked locks, we now remove lock tracking once a lock is
destructed.

Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 7c7df2c)
librados_test_stub: add support for flushing watches
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 6e400b9)
librados_test_stub: fix helgrind warnings
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit b65ae4b)
log: fix helgrind warnings regarding possible data race
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit c1e1445)
librbd: require callers to ObjectMap::aio_update to acquire lock
This is needed to allow an atomic compare and update operation
from the rebuild object map utility.

Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 2db758c)
librbd/AioRequest.h: fix UNINIT_CTOR
Fix for:

CID 1274319: Uninitialized scalar field (UNINIT_CTOR)
 uninit_member: Non-static class member m_object_state is not
 initialized in this constructor nor in any functions that it calls.

Signed-off-by: Danny Al-Gaaf <danny.al-gaaf@bisect.de>
(cherry picked from commit 48f18ea)
librbd: simplify AioRequest constructor parameters
Moved all parent overlap computation to within AioRequest so that
callers don't need to independently compute the overlap.  Also
removed the need to pass the snap_id for write operations since
it can only be CEPH_NOSNAP.

Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 4651597)
librbd: move copyup class method call to CopyupRequest
Move AbstractWrite's invocation of copyup to the CopyupRequest
class.  The AioRequest write path will now always create a
CopyupRequest, which will now append the actual write ops to the
copyup.

Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 7be3df6)
librbd: complete cache read in a new thread context
The ObjectCacher complete the read callback while still holding
the cache lock.  This introduces lock ordering issues which are
resolved by queuing the completion to execute in a clean (unlocked)
context.

Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 0024677)
librbd: AioCompletion shouldn't hold its lock during callback
The callback routine most likely will attempt to retrieve the result
code, which will result in a recursive lock attempt.

Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 3ad19ae)
librbd: give locks unique names to prevent false lockdep failures
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit c474ee4)
librbd: add AsyncRequest task enqueue helper method
In order to support the invariant that all state machine
callbacks occur without holding locks, transitions that
don't always involve a librados call should queue their
callback.

Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 218bc2d)
librbd: execute flush completion outside of cache_lock
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 5f157f2)
librbd: AsyncObjectThrottle should always hold owner_lock
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit c352bcd)
librbd: disable lockdep on AioCompletion
It is only used by clients and it causes a large slowdown
in performance due to the rate at which the lock is constructed/
destructed for each IO request.

Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 66e7464)
librbd: add object state accessor to ObjectMap
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
librbd: ObjectMap::aio_update can acquire snap_lock out-of-order
Detected during an fsx run where a refresh and CoR were occurring
concurrently.  The refresh held the snap_lock and was waiting on
the object_map_lock, while the CoR held object_map_lock and was
waiting for snap_lock.

Fixes: #11577
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 8cbd92b)

dillaman added some commits Apr 30, 2015

librbd: simplify state machine handling of exclusive lock
It is expected that all IO is flushed and all async ops are cancelled
prior to releasing the exclusive lock.  Therefore, replace handling of
lost exclusive locks in state machines with an assertion.

Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit d6b733d)
librbd: fix recursive locking issues
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 1b57cc1)
osdc: ObjectCacher flusher might needs additional locks
librbd requires the ObjectCacher flusher thread to acquire
an additional lock in order to maintain lock ordering
constraints.

Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit a38f9e5)
librbd: owner_lock should be held during flush request
Flush might result in the cache writing out dirty objects, which
would require that the owner_lock be held.

Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit c9142fe)
tests: enable lockdep for librbd unit tests
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 45cb9cb)
tests: fix possible deadlock in librbd ImageWatcher tests
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 742a85d)
librbd: flush pending ops while not holding lock
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 3d5cef3)
librbd: avoid infinite loop if copyup fails
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 43e0e3c)
librbd: flush operations need to acquire owner lock
Cache writeback operations will expect the owner lock to be held.

Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit adfa2e0)
lockdep: do not automatically collect all backtraces
It is expensive to collect backtraces every time a lock is
checked in order to provide cycle backtraces.  The backtraces
can be forced on for specific locks or globally via the new
config option "lockdep_force_backtrace".

Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 7354d25)

@ghost ghost changed the title DNM: Possible crash while concurrently writing and shrinking an image Possible crash while concurrently writing and shrinking an image Jul 29, 2015

@ghost ghost changed the title Possible crash while concurrently writing and shrinking an image DNM: Possible crash while concurrently writing and shrinking an image Jul 29, 2015

@dillaman dillaman force-pushed the wip-12236-hammer branch from 45c513a to e1e8b20 Jul 29, 2015

@dillaman dillaman changed the title DNM: Possible crash while concurrently writing and shrinking an image Possible crash while concurrently writing and shrinking an image Jul 29, 2015

dillaman added some commits Jun 23, 2015

librbd: invalidate cache outside cache callback context
When shrinking an image, it's possible that the op flush callback
will be from within the cache callback context.  This would result
in a deadlock when attempting to re-lock the cache lock in order to
invalidate the cache.

Fixes: #11743
Backport: hammer
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 726d699)
tests: verify that image shrink properly handles flush op
Signed-off-by: Jason Dillaman <dillaman@redhat.com>
(cherry picked from commit 8dfcc72)

@dillaman dillaman force-pushed the wip-12236-hammer branch from e1e8b20 to b575ecc Jul 29, 2015

ghost pushed a commit that referenced this pull request Jul 29, 2015

@ghost

This comment has been minimized.

Copy link

ghost commented Aug 30, 2015

ghost pushed a commit that referenced this pull request Aug 30, 2015

Loic Dachary
Merge pull request #5318 from ceph/wip-12236-hammer
Possible crash while concurrently writing and shrinking an image

Reviewed-by: Loic Dachary <ldachary@redhat.com>

@ghost ghost merged commit 260e24b into hammer Aug 30, 2015

@dillaman dillaman deleted the wip-12236-hammer branch Sep 7, 2015

@ghost ghost changed the title Possible crash while concurrently writing and shrinking an image librbd: Possible crash while concurrently writing and shrinking an image Sep 13, 2015

This issue was closed.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
You can’t perform that action at this time.