librbd: Truncate of non-existent object results in object map flagged… #7772
Conversation
|
I think this fixes the object map, but it would still send a truncate op to the OSD for a non-existent object. It would be nice to avoid that extra op if the object map says it isn't required. |
… as exists Fixes: ceph#14789 Signed-off-by: xinxin shu <xinxin.shu@intel.com>
|
@dillaman , updated |
| @@ -434,12 +434,12 @@ namespace librbd { | |||
| } | |||
|
|
|||
| bool AbstractAioObjectWrite::send_post() { | |||
| RWLock::RLocker owner_locker(m_ictx->owner_lock); | |||
| RWLock::RLocker snap_locker(m_ictx->snap_lock); | |||
There was a problem hiding this comment.
I believe snap_lock is used to protect m_ictx->object_map during image refreshes, so it can't just be moved. You might need to do something like this to avoid the lock recursion: https://github.com/ceph/ceph/blob/master/src/librbd/AioObjectRequest.cc#L219
|
@dillaman , ok, thanks for your hint, one question, if snap_lock is used to protect m_ictx->object_map, what does m_ictx->object_map_lock protect |
|
@xinxinsh That's used to protect the contents of the object map from being updated concurrently. |
Signed-off-by: xinxin shu <xinxin.shu@intel.com>
|
@dillaman , updated |
|
lgtm now, running through testing |
librbd: truncate does not need to mark the object as existing in the object map Reviewed-by: Josh Durgin <jdurgin@redhat.com>
… as exists
Fixes: #14789
Signed-off-by: xinxin shu xinxin.shu@intel.com