ocfs2: wait for recovering done after direct unlock request

There is a scenario causing ocfs2 umount hang when multiple hosts are
rebooting at the same time.

NODE1                           NODE2               NODE3
send unlock requset to NODE2
                                dies
                                                    become recovery master
                                                    recover NODE2
find NODE2 dead
mark resource RECOVERING
directly remove lock from grant list
calculate usage but RECOVERING marked
**miss the window of purging
clear RECOVERING

To reproduce this issue, crash a host and then umount ocfs2
from another node.

To solve this, just let unlock progress wait for recovery done.

Link: http://lkml.kernel.org/r/1550124866-20367-1-git-send-email-gechangwei@live.cn
Signed-off-by: Changwei Ge <gechangwei@live.cn>
Reviewed-by: Joseph Qi <joseph.qi@linux.alibaba.com>
Cc: Mark Fasheh <mark@fasheh.com>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Changwei Ge <gechangwei@live.cn>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This commit is contained in:
Changwei Ge 2019-09-23 15:33:37 -07:00 committed by Linus Torvalds
parent a89bd89fae
commit 0a3775e4f8

View File

@ -90,7 +90,8 @@ static enum dlm_status dlmunlock_common(struct dlm_ctxt *dlm,
enum dlm_status status;
int actions = 0;
int in_use;
u8 owner;
u8 owner;
int recovery_wait = 0;
mlog(0, "master_node = %d, valblk = %d\n", master_node,
flags & LKM_VALBLK);
@ -193,9 +194,12 @@ static enum dlm_status dlmunlock_common(struct dlm_ctxt *dlm,
}
if (flags & LKM_CANCEL)
lock->cancel_pending = 0;
else
lock->unlock_pending = 0;
else {
if (!lock->unlock_pending)
recovery_wait = 1;
else
lock->unlock_pending = 0;
}
}
/* get an extra ref on lock. if we are just switching
@ -229,6 +233,17 @@ leave:
spin_unlock(&res->spinlock);
wake_up(&res->wq);
if (recovery_wait) {
spin_lock(&res->spinlock);
/* Unlock request will directly succeed after owner dies,
* and the lock is already removed from grant list. We have to
* wait for RECOVERING done or we miss the chance to purge it
* since the removement is much faster than RECOVERING proc.
*/
__dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_RECOVERING);
spin_unlock(&res->spinlock);
}
/* let the caller's final dlm_lock_put handle the actual kfree */
if (actions & DLM_UNLOCK_FREE_LOCK) {
/* this should always be coupled with list removal */