From 0a3775e4f883912944481cf2ef36eb6383a9cc74 Mon Sep 17 00:00:00 2001 From: Changwei Ge Date: Mon, 23 Sep 2019 15:33:37 -0700 Subject: [PATCH] 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 Reviewed-by: Joseph Qi Cc: Mark Fasheh Cc: Joel Becker Cc: Junxiao Bi Cc: Changwei Ge Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds --- fs/ocfs2/dlm/dlmunlock.c | 23 +++++++++++++++++++---- 1 file changed, 19 insertions(+), 4 deletions(-) diff --git a/fs/ocfs2/dlm/dlmunlock.c b/fs/ocfs2/dlm/dlmunlock.c index e78657742bd8..3883633e82eb 100644 --- a/fs/ocfs2/dlm/dlmunlock.c +++ b/fs/ocfs2/dlm/dlmunlock.c @@ -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 */