staging: lustre: ldlm: handle ldlm lock cancel race when evicting client.

A ldlm lock could be canceled simutaneously by ldlm bl thread and
cleanup_resource(). In this case, only one side will win the race
and the other side should wait for the work to complete. Eviction
on group lock is now well supported.

Signed-off-by: Jinshan Xiong <jinshan.xiong@intel.com>
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6271
Reviewed-on: http://review.whamcloud.com/16456
Reviewed-by: Bobi Jam <bobijam@hotmail.com>
Reviewed-by: John L. Hammond <john.hammond@intel.com>
Reviewed-by: James Simmons <uja.ornl@yahoo.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>
Signed-off-by: James Simmons <jsimmons@infradead.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

1 files changed, 12 insertions, 2 deletions

diff --git a/drivers/staging/lustre/lustre/ldlm/ldlm_request.c b/drivers/staging/lustre/lustre/ldlm/ldlm_request.c
index ebfda36..84eeaa5 100644
--- a/drivers/staging/lustre/lustre/ldlm/ldlm_request.c
+++ b/drivers/staging/lustre/lustre/ldlm/ldlm_request.c
@@ -1029,13 +1029,23 @@ int ldlm_cli_cancel(const struct lustre_handle *lockh,
 	struct ldlm_lock *lock;
 	LIST_HEAD(cancels);
 
-	/* concurrent cancels on the same handle can happen */
-	lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
+	lock = ldlm_handle2lock_long(lockh, 0);
 	if (!lock) {
 		LDLM_DEBUG_NOLOCK("lock is already being destroyed");
 		return 0;
 	}
 
+	lock_res_and_lock(lock);
+	/* Lock is being canceled and the caller doesn't want to wait */
+	if (ldlm_is_canceling(lock) && (cancel_flags & LCF_ASYNC)) {
+		unlock_res_and_lock(lock);
+		LDLM_LOCK_RELEASE(lock);
+		return 0;
+	}
+
+	ldlm_set_canceling(lock);
+	unlock_res_and_lock(lock);
+
 	rc = ldlm_cli_cancel_local(lock);
 	if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
 		LDLM_LOCK_RELEASE(lock);