b=18618

[fs/lustre-release.git] / lustre / ldlm / ldlm_lockd.c

diff --git a/lustre/ldlm/ldlm_lockd.c b/lustre/ldlm/ldlm_lockd.c
index f6fe4a7..82c37cf 100644 (file)
--- a/lustre/ldlm/ldlm_lockd.c
+++ b/lustre/ldlm/ldlm_lockd.c
@@ -809,11 +809,6 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
         total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(),
                                           lock->l_last_activity);
 
-        if (total_enqueue_wait > obd_timeout)
-                /* non-fatal with AT - change to LDLM_DEBUG? */
-                LDLM_WARN(lock, "enqueue wait took %lus from "CFS_TIME_T,
-                          total_enqueue_wait, lock->l_last_activity);
-
         req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
                                     &RQF_LDLM_CP_CALLBACK);
         if (req == NULL)
@@ -854,8 +849,18 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 
         /* Server-side enqueue wait time estimate, used in
             __ldlm_add_waiting_lock to set future enqueue timers */
-        at_add(&lock->l_resource->lr_namespace->ns_at_estimate,
-               total_enqueue_wait);
+        if (total_enqueue_wait < ldlm_get_enq_timeout(lock))
+                at_add(&lock->l_resource->lr_namespace->ns_at_estimate,
+                       total_enqueue_wait);
+        else
+                /* bz18618. Don't add lock enqueue time we spend waiting for a
+                   previous callback to fail. Locks waiting legitimately will
+                   get extended by ldlm_refresh_waiting_lock regardless of the
+                   estimate, so it's okay to underestimate here. */
+                LDLM_DEBUG(lock, "lock completed after %lus; estimate was %ds. "
+                       "It is likely that a previous callback timed out.",
+                       total_enqueue_wait,
+                       at_get(&lock->l_resource->lr_namespace->ns_at_estimate));
 
         ptlrpc_request_set_replen(req);
 
@@ -2260,7 +2265,7 @@ static int ldlm_setup(void)
         ldlm_state->ldlm_cb_service =
                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
                                 LDLM_MAXREPSIZE, LDLM_CB_REQUEST_PORTAL,
-                                LDLM_CB_REPLY_PORTAL, 1800,
+                                LDLM_CB_REPLY_PORTAL, 2,
                                 ldlm_callback_handler, "ldlm_cbd",
                                 ldlm_svc_proc_dir, NULL,
                                 ldlm_min_threads, ldlm_max_threads,
@@ -2275,7 +2280,7 @@ static int ldlm_setup(void)
         ldlm_state->ldlm_cancel_service =
                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
                                 LDLM_MAXREPSIZE, LDLM_CANCEL_REQUEST_PORTAL,
-                                LDLM_CANCEL_REPLY_PORTAL, 6000,
+                                LDLM_CANCEL_REPLY_PORTAL, 6,
                                 ldlm_cancel_handler, "ldlm_canceld",
                                 ldlm_svc_proc_dir, NULL,
                                 ldlm_min_threads, ldlm_max_threads,
@@ -2449,6 +2454,12 @@ void __exit ldlm_exit(void)
                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
         rc = cfs_mem_cache_destroy(ldlm_resource_slab);
         LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
+#ifdef __KERNEL__
+        /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
+         * synchronize_rcu() to wait a grace period elapsed, so that
+         * ldlm_lock_free() get a chance to be called. */
+        synchronize_rcu();
+#endif
         rc = cfs_mem_cache_destroy(ldlm_lock_slab);
         LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
         rc = cfs_mem_cache_destroy(ldlm_interval_slab);