LU-1126 ldlm: Fix a race during FLock handling

Protect against race where lock could have been just destroyed
due to overlap, in ldlm_process_flock_lock().
Easy reproducer is BULL's NFS Locktests in pthread mode.
(http://nfsv4.bullopensource.org/tools/tests/locktest.php)

Signed-off-by: Bruno Faccini <bruno.faccini@intel.com>
Change-Id: I4f4183098e8e7aaacb2c6b80787d3c7a7ad5688b
Reviewed-on: http://review.whamcloud.com/7134
Tested-by: Hudson
Tested-by: Maloo <whamcloud.maloo@gmail.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>
Reviewed-by: John L. Hammond <john.hammond@intel.com>

diff --git a/lustre/ldlm/ldlm_flock.c b/lustre/ldlm/ldlm_flock.c
index 6273b72..d9cc416 100644 (file)
--- a/lustre/ldlm/ldlm_flock.c
+++ b/lustre/ldlm/ldlm_flock.c
@@ -697,11 +697,6 @@ ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 granted:
         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
 
-       if (lock->l_flags & LDLM_FL_DESTROYED) {
-               LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
-               RETURN(0);
-       }
-
         if (lock->l_flags & LDLM_FL_FAILED) {
                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed");
                 RETURN(-EIO);
@@ -711,6 +706,16 @@ granted:
 
        lock_res_and_lock(lock);
 
+
+       /* Protect against race where lock could have been just destroyed
+        * due to overlap in ldlm_process_flock_lock().
+        */
+       if (lock->l_flags & LDLM_FL_DESTROYED) {
+               unlock_res_and_lock(lock);
+               LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
+               RETURN(0);
+       }
+
        /* take lock off the deadlock detection hash list. */
         ldlm_flock_blocking_unlink(lock);