Whamcloud - gitweb
b=17682 limit performance impact of rpctrace, dlmtrace & quota
[fs/lustre-release.git] / lustre / ldlm / ldlm_lockd.c
index 8ef4562..7ff6bd8 100644 (file)
@@ -1,27 +1,42 @@
 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
  * vim:expandtab:shiftwidth=8:tabstop=8:
  *
- * Copyright (C) 2002-2004 Cluster File Systems, Inc.
- *   Author: Peter Braam <braam@clusterfs.com>
- *   Author: Phil Schwan <phil@clusterfs.com>
+ * GPL HEADER START
  *
- *   This file is part of the Lustre file system, http://www.lustre.org
- *   Lustre is a trademark of Cluster File Systems, Inc.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
- *   You may have signed or agreed to another license before downloading
- *   this software.  If so, you are bound by the terms and conditions
- *   of that agreement, and the following does not apply to you.  See the
- *   LICENSE file included with this distribution for more information.
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 only,
+ * as published by the Free Software Foundation.
  *
- *   If you did not agree to a different license, then this copy of Lustre
- *   is open source software; you can redistribute it and/or modify it
- *   under the terms of version 2 of the GNU General Public License as
- *   published by the Free Software Foundation.
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License version 2 for more details (a copy is included
+ * in the LICENSE file that accompanied this code).
  *
- *   In either case, Lustre is distributed in the hope that it will be
- *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
- *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- *   license text for more details.
+ * You should have received a copy of the GNU General Public License
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
+ *
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
+ *
+ * GPL HEADER END
+ */
+/*
+ * Copyright  2008 Sun Microsystems, Inc. All rights reserved
+ * Use is subject to license terms.
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ * Lustre is a trademark of Sun Microsystems, Inc.
+ *
+ * lustre/ldlm/ldlm_lockd.c
+ *
+ * Author: Peter Braam <braam@clusterfs.com>
+ * Author: Phil Schwan <phil@clusterfs.com>
  */
 
 #ifndef EXPORT_SYMTAB
@@ -60,10 +75,10 @@ inline cfs_time_t round_timeout(cfs_time_t timeout)
         return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
 }
 
-/* timeout for initial callback (AST) reply */
-static inline unsigned int ldlm_get_rq_timeout(unsigned int ldlm_timeout,
-                                               unsigned int obd_timeout)
+/* timeout for initial callback (AST) reply (bz10399) */
+static inline unsigned int ldlm_get_rq_timeout(void)
 {
+        /* Non-AT value */
         unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
 
         return timeout < 1 ? 1 : timeout;
@@ -187,14 +202,23 @@ static int expired_lock_main(void *arg)
                                        lock->l_export);
                                 lock->l_export = NULL;
                                 LDLM_ERROR(lock, "free export");
+                                /* release extra ref grabbed by
+                                 * ldlm_add_waiting_lock() or
+                                 * ldlm_failed_ast() */
+                                LDLM_LOCK_RELEASE(lock);
                                 continue;
                         }
-                        export = class_export_get(lock->l_export);
+                        export = class_export_lock_get(lock->l_export, lock);
                         spin_unlock_bh(&waiting_locks_spinlock);
 
                         do_dump++;
                         class_fail_export(export);
-                        class_export_put(export);
+                        class_export_lock_put(export, lock);
+
+                        /* release extra ref grabbed by ldlm_add_waiting_lock()
+                         * or ldlm_failed_ast() */
+                        LDLM_LOCK_RELEASE(lock);
                         spin_lock_bh(&waiting_locks_spinlock);
                 }
                 spin_unlock_bh(&waiting_locks_spinlock);
@@ -215,6 +239,31 @@ static int expired_lock_main(void *arg)
 
 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
 
+/**
+ * Check if there is a request in the export request list
+ * which prevents the lock canceling.
+ */
+static int ldlm_lock_busy(struct ldlm_lock *lock)
+{
+        struct ptlrpc_request *req;
+        int match = 0;
+        ENTRY;
+
+        if (lock->l_export == NULL)
+                return 0;
+
+        spin_lock(&lock->l_export->exp_lock);
+        list_for_each_entry(req, &lock->l_export->exp_queued_rpc, rq_exp_list) {
+                if (req->rq_ops->hpreq_lock_match) {
+                        match = req->rq_ops->hpreq_lock_match(req, lock);
+                        if (match)
+                                break;
+                }
+        }
+        spin_unlock(&lock->l_export->exp_lock);
+        RETURN(match);
+}
+
 /* This is called from within a timer interrupt and cannot schedule */
 static void waiting_locks_callback(unsigned long unused)
 {
@@ -225,7 +274,6 @@ repeat:
         while (!list_empty(&waiting_locks_list)) {
                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
                                   l_pending_chain);
-
                 if (cfs_time_after(lock->l_callback_timeout, cfs_time_current()) ||
                     (lock->l_req_mode == LCK_GROUP))
                         break;
@@ -263,14 +311,42 @@ repeat:
                         goto repeat;
                 }
 
-                LDLM_ERROR(lock, "lock callback timer expired: evicting client "
-                           "%s@%s nid %s\n",
-                           lock->l_export->exp_client_uuid.uuid,
-                           lock->l_export->exp_connection->c_remote_uuid.uuid,
-                           libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
+                /* Check if we need to prolong timeout */
+                if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
+                    ldlm_lock_busy(lock)) {
+                        int cont = 1;
+
+                        if (lock->l_pending_chain.next == &waiting_locks_list)
+                                cont = 0;
+
+                        LDLM_LOCK_GET(lock);
+
+                        spin_unlock_bh(&waiting_locks_spinlock);
+                        LDLM_DEBUG(lock, "prolong the busy lock");
+                        ldlm_refresh_waiting_lock(lock,
+                                                  ldlm_get_enq_timeout(lock));
+                        spin_lock_bh(&waiting_locks_spinlock);
+
+                        if (!cont) {
+                                LDLM_LOCK_RELEASE(lock);
+                                break;
+                        }
+
+                        LDLM_LOCK_RELEASE(lock);
+                        continue;
+                }
+                lock->l_resource->lr_namespace->ns_timeouts++;
+                LDLM_ERROR(lock, "lock callback timer expired after %lds: "
+                           "evicting client at %s ",
+                           cfs_time_current_sec()- lock->l_last_activity,
+                           libcfs_nid2str(
+                                   lock->l_export->exp_connection->c_peer.nid));
 
                 last = lock;
 
+                /* no needs to take an extra ref on the lock since it was in
+                 * the waiting_locks_list and ldlm_add_waiting_lock()
+                 * already grabbed a ref */
                 list_del(&lock->l_pending_chain);
                 list_add(&lock->l_pending_chain,
                          &expired_lock_thread.elt_expired_locks);
@@ -302,26 +378,36 @@ repeat:
  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
  * timer to fire appropriately.  (We round up to the next second, to avoid
  * floods of timer firings during periods of high lock contention and traffic).
+ * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
+ * if it has been added to the waiting list (1 is returned).
  *
  * Called with the namespace lock held.
  */
-static int __ldlm_add_waiting_lock(struct ldlm_lock *lock)
+static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds)
 {
+        cfs_time_t timeout;
         cfs_time_t timeout_rounded;
 
         if (!list_empty(&lock->l_pending_chain))
                 return 0;
 
-        lock->l_callback_timeout =cfs_time_add(cfs_time_current(),
-                                               cfs_time_seconds(obd_timeout)/2);
+        if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
+            OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
+                seconds = 1;
+
+        timeout = cfs_time_shift(seconds);
+        if (likely(cfs_time_after(timeout, lock->l_callback_timeout)))
+                lock->l_callback_timeout = timeout;
 
         timeout_rounded = round_timeout(lock->l_callback_timeout);
 
-        if (cfs_time_before(timeout_rounded, cfs_timer_deadline(&waiting_locks_timer)) ||
+        if (cfs_time_before(timeout_rounded,
+                            cfs_timer_deadline(&waiting_locks_timer)) ||
             !cfs_timer_is_armed(&waiting_locks_timer)) {
                 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
-
         }
+        /* if the new lock has a shorter timeout than something earlier on
+           the list, we'll wait the longer amount of time; no big deal. */
         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
         return 1;
 }
@@ -329,6 +415,7 @@ static int __ldlm_add_waiting_lock(struct ldlm_lock *lock)
 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 {
         int ret;
+        int timeout = ldlm_get_enq_timeout(lock);
 
         LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
 
@@ -344,11 +431,16 @@ static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
                 return 0;
         }
 
-        ret = __ldlm_add_waiting_lock(lock);
+        ret = __ldlm_add_waiting_lock(lock, timeout);
+        if (ret)
+                /* grab ref on the lock if it has been added to the
+                 * waiting list */
+                LDLM_LOCK_GET(lock);
         spin_unlock_bh(&waiting_locks_spinlock);
 
-        LDLM_DEBUG(lock, "%sadding to wait list",
-                   ret == 0 ? "not re-" : "");
+        LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
+                   ret == 0 ? "not re-" : "", timeout,
+                   AT_OFF ? "off" : "on");
         return ret;
 }
 
@@ -356,10 +448,12 @@ static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
  * Remove a lock from the pending list, likely because it had its cancellation
  * callback arrive without incident.  This adjusts the lock-timeout timer if
  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
+ * As done by ldlm_del_waiting_lock(), the caller must release the lock
+ * reference when the lock is removed from any list (1 is returned).
  *
  * Called with namespace lock held.
  */
-int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
+static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
 {
         struct list_head *list_next;
 
@@ -391,13 +485,17 @@ int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 
         if (lock->l_export == NULL) {
                 /* We don't have a "waiting locks list" on clients. */
-                LDLM_DEBUG(lock, "client lock: no-op");
+                CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
                 return 0;
         }
 
         spin_lock_bh(&waiting_locks_spinlock);
         ret = __ldlm_del_waiting_lock(lock);
         spin_unlock_bh(&waiting_locks_spinlock);
+        if (ret)
+                /* release lock ref if it has indeed been removed
+                 * from a list */
+                LDLM_LOCK_RELEASE(lock);
 
         LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
         return ret;
@@ -408,7 +506,7 @@ int ldlm_del_waiting_lock(struct ldlm_lock *lock)
  *
  * Called with namespace lock held.
  */
-int ldlm_refresh_waiting_lock(struct ldlm_lock *lock)
+int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
 {
         if (lock->l_export == NULL) {
                 /* We don't have a "waiting locks list" on clients. */
@@ -424,14 +522,15 @@ int ldlm_refresh_waiting_lock(struct ldlm_lock *lock)
                 return 0;
         }
 
+        /* we remove/add the lock to the waiting list, so no needs to
+         * release/take a lock reference */
         __ldlm_del_waiting_lock(lock);
-        __ldlm_add_waiting_lock(lock);
+        __ldlm_add_waiting_lock(lock, timeout);
         spin_unlock_bh(&waiting_locks_spinlock);
 
         LDLM_DEBUG(lock, "refreshed");
         return 1;
 }
-
 #else /* !__KERNEL__ */
 
 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
@@ -445,7 +544,7 @@ int ldlm_del_waiting_lock(struct ldlm_lock *lock)
         RETURN(0);
 }
 
-int ldlm_refresh_waiting_lock(struct ldlm_lock *lock)
+int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
 {
         RETURN(0);
 }
@@ -454,20 +553,25 @@ int ldlm_refresh_waiting_lock(struct ldlm_lock *lock)
 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
                             const char *ast_type)
 {
-        struct ptlrpc_connection *conn = lock->l_export->exp_connection;
-        char                     *str = libcfs_nid2str(conn->c_peer.nid);
-
-        LCONSOLE_ERROR_MSG(0x138, "A client on nid %s was evicted from "
-                           "service %s.\n", str,
-                           lock->l_export->exp_obd->obd_name);
-
-        LCONSOLE_ERROR_MSG(0x012, "Lock %s callback to %s timed out for "
-                           "resource %d\n", ast_type,
-                           obd_export_nid2str(lock->l_export), rc);
+        LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
+                           "to a lock %s callback time out: rc %d\n",
+                           lock->l_export->exp_obd->obd_name,
+                           obd_export_nid2str(lock->l_export), ast_type, rc);
 
         if (obd_dump_on_timeout)
                 libcfs_debug_dumplog();
+#ifdef __KERNEL__
+        spin_lock_bh(&waiting_locks_spinlock);
+        if (__ldlm_del_waiting_lock(lock) == 0)
+                /* the lock was not in any list, grab an extra ref before adding
+                 * the lock to the expired list */
+                LDLM_LOCK_GET(lock);
+        list_add(&lock->l_pending_chain, &expired_lock_thread.elt_expired_locks);
+        cfs_waitq_signal(&expired_lock_thread.elt_waitq);
+        spin_unlock_bh(&waiting_locks_spinlock);
+#else
         class_fail_export(lock->l_export);
+#endif
 }
 
 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
@@ -517,7 +621,8 @@ static int ldlm_handle_ast_error(struct ldlm_lock *lock,
         return rc;
 }
 
-static int ldlm_cb_interpret(struct ptlrpc_request *req, void *data, int rc)
+static int ldlm_cb_interpret(const struct lu_env *env,
+                             struct ptlrpc_request *req, void *data, int rc)
 {
         struct ldlm_cb_set_arg *arg;
         struct ldlm_lock *lock;
@@ -529,18 +634,17 @@ static int ldlm_cb_interpret(struct ptlrpc_request *req, void *data, int rc)
         lock = req->rq_async_args.pointer_arg[1];
         LASSERT(lock != NULL);
         if (rc != 0) {
-                /* If client canceled the lock but the cancel has not 
-                 * been recieved yet, we need to update lvbo to have the
+                /* If client canceled the lock but the cancel has not
+                 * been received yet, we need to update lvbo to have the
                  * proper attributes cached. */
                 if (rc == -EINVAL && arg->type == LDLM_BL_CALLBACK)
-                        ldlm_res_lvbo_update(lock->l_resource, NULL, 
-                                             0, 1);
-                rc = ldlm_handle_ast_error(lock, req, rc, 
+                        ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
+                rc = ldlm_handle_ast_error(lock, req, rc,
                                            arg->type == LDLM_BL_CALLBACK
                                            ? "blocking" : "completion");
-        }                
+        }
 
-        LDLM_LOCK_PUT(lock);
+        LDLM_LOCK_RELEASE(lock);
 
         if (rc == -ERESTART)
                 atomic_set(&arg->restart, 1);
@@ -566,11 +670,35 @@ static inline int ldlm_bl_and_cp_ast_fini(struct ptlrpc_request *req,
         } else {
                 LDLM_LOCK_GET(lock);
                 ptlrpc_set_add_req(arg->set, req);
-        }       
+        }
 
         RETURN(rc);
 }
 
+/**
+ * Check if there are requests in the export request list which prevent
+ * the lock canceling and make these requests high priority ones.
+ */
+static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
+{
+        struct ptlrpc_request *req;
+        ENTRY;
+
+        if (lock->l_export == NULL) {
+                LDLM_DEBUG(lock, "client lock: no-op");
+                RETURN_EXIT;
+        }
+
+        spin_lock(&lock->l_export->exp_lock);
+        list_for_each_entry(req, &lock->l_export->exp_queued_rpc, rq_exp_list) {
+                if (!req->rq_hp && req->rq_ops->hpreq_lock_match &&
+                    req->rq_ops->hpreq_lock_match(req, lock))
+                        ptlrpc_hpreq_reorder(req);
+        }
+        spin_unlock(&lock->l_export->exp_lock);
+        EXIT;
+}
+
 /*
  * ->l_blocking_ast() method for server-side locks. This is invoked when newly
  * enqueued server lock conflicts with given one.
@@ -600,6 +728,8 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock,
                 ldlm_lock_dump(D_ERROR, lock, 0);
         }
 
+        ldlm_lock_reorder_req(lock);
+
         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
                                         &RQF_LDLM_BL_CALLBACK,
                                         LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
@@ -649,10 +779,13 @@ int ldlm_server_blocking_ast(struct ldlm_lock *lock,
         }
 
         req->rq_send_state = LUSTRE_IMP_FULL;
-        req->rq_timeout = ldlm_get_rq_timeout(ldlm_timeout, obd_timeout);
+        /* ptlrpc_request_alloc_pack already set timeout */
+        if (AT_OFF)
+                req->rq_timeout = ldlm_get_rq_timeout();
 
-        if (lock->l_export && lock->l_export->exp_ldlm_stats)
-                lprocfs_counter_incr(lock->l_export->exp_ldlm_stats,
+        if (lock->l_export && lock->l_export->exp_nid_stats &&
+            lock->l_export->exp_nid_stats->nid_ldlm_stats)
+                lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
                                      LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
 
         rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel);
@@ -665,7 +798,6 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
         struct ldlm_cb_set_arg *arg = data;
         struct ldlm_request    *body;
         struct ptlrpc_request  *req;
-        struct timeval          granted_time;
         long                    total_enqueue_wait;
         int                     instant_cancel = 0;
         int                     rc = 0;
@@ -674,18 +806,13 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
         LASSERT(lock != NULL);
         LASSERT(data != NULL);
 
-        do_gettimeofday(&granted_time);
-        total_enqueue_wait = cfs_timeval_sub(&granted_time,
-                                             &lock->l_enqueued_time, NULL);
+        total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(),
+                                          lock->l_last_activity);
 
-        if (total_enqueue_wait / 1000000 > obd_timeout)
-                LDLM_ERROR(lock, "enqueue wait took %luus from %lu",
-                           total_enqueue_wait, lock->l_enqueued_time.tv_sec);
-
-         req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
+        req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
                                     &RQF_LDLM_CP_CALLBACK);
-         if (req == NULL)
-                 RETURN(-ENOMEM);
+        if (req == NULL)
+                RETURN(-ENOMEM);
 
         lock_res_and_lock(lock);
         if (lock->l_resource->lr_lvb_len)
@@ -717,17 +844,37 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
                 unlock_res_and_lock(lock);
         }
 
-        LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
+        LDLM_DEBUG(lock, "server preparing completion AST (after %lds wait)",
                    total_enqueue_wait);
 
+        /* Server-side enqueue wait time estimate, used in
+            __ldlm_add_waiting_lock to set future enqueue timers */
+        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);
+
         req->rq_send_state = LUSTRE_IMP_FULL;
-        req->rq_timeout = ldlm_get_rq_timeout(ldlm_timeout, obd_timeout);
+        /* ptlrpc_request_pack already set timeout */
+        if (AT_OFF)
+                req->rq_timeout = ldlm_get_rq_timeout();
 
         /* We only send real blocking ASTs after the lock is granted */
         lock_res_and_lock(lock);
         if (lock->l_flags & LDLM_FL_AST_SENT) {
                 body->lock_flags |= LDLM_FL_AST_SENT;
+                /* copy ast flags like LDLM_FL_DISCARD_DATA */
+                body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 
                 /* We might get here prior to ldlm_handle_enqueue setting
                  * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
@@ -747,8 +894,9 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
         }
         unlock_res_and_lock(lock);
 
-        if (lock->l_export && lock->l_export->exp_ldlm_stats)
-                lprocfs_counter_incr(lock->l_export->exp_ldlm_stats,
+        if (lock->l_export && lock->l_export->exp_nid_stats &&
+            lock->l_export->exp_nid_stats->nid_ldlm_stats)
+                lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
                                      LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
 
         rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel);
@@ -786,10 +934,13 @@ int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 
 
         req->rq_send_state = LUSTRE_IMP_FULL;
-        req->rq_timeout = ldlm_get_rq_timeout(ldlm_timeout, obd_timeout);
+        /* ptlrpc_request_alloc_pack already set timeout */
+        if (AT_OFF)
+                req->rq_timeout = ldlm_get_rq_timeout();
 
-        if (lock->l_export && lock->l_export->exp_ldlm_stats)
-                lprocfs_counter_incr(lock->l_export->exp_ldlm_stats,
+        if (lock->l_export && lock->l_export->exp_nid_stats &&
+            lock->l_export->exp_nid_stats->nid_ldlm_stats)
+                lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
                                      LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
 
         rc = ptlrpc_queue_wait(req);
@@ -798,30 +949,13 @@ int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
         else if (rc != 0)
                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
         else
-                rc = ldlm_res_lvbo_update(res, req->rq_repmsg,
-                                          REPLY_REC_OFF, 1);
-        ptlrpc_req_finished(req);
-        RETURN(rc);
-}
+                rc = ldlm_res_lvbo_update(res, req, 1);
 
-static struct ldlm_lock *
-find_existing_lock(struct obd_export *exp,
-                   const struct lustre_handle *remote_hdl)
-{
-        struct list_head *iter;
+        ptlrpc_req_finished(req);
+        if (rc == -ERESTART)
+                ldlm_reprocess_all(res);
 
-        spin_lock(&exp->exp_ldlm_data.led_lock);
-        list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
-                struct ldlm_lock *lock;
-                lock = list_entry(iter, struct ldlm_lock, l_export_chain);
-                if (lock->l_remote_handle.cookie == remote_hdl->cookie) {
-                        LDLM_LOCK_GET(lock);
-                        spin_unlock(&exp->exp_ldlm_data.led_lock);
-                        return lock;
-                }
-        }
-        spin_unlock(&exp->exp_ldlm_data.led_lock);
-        return NULL;
+        RETURN(rc);
 }
 
 #ifdef __KERNEL__
@@ -861,7 +995,7 @@ static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
         if (op)
                 lprocfs_counter_incr(srv_stats, op);
 
-        return ;
+        return;
 }
 
 /*
@@ -892,8 +1026,9 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
                 ldlm_svc_get_eopc(dlm_req,
                                   req->rq_rqbd->rqbd_service->srv_stats);
 
-        if (req->rq_export->exp_ldlm_stats)
-                lprocfs_counter_incr(req->rq_export->exp_ldlm_stats,
+        if (req->rq_export && req->rq_export->exp_nid_stats &&
+            req->rq_export->exp_nid_stats->nid_ldlm_stats)
+                lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
                                      LDLM_ENQUEUE - LDLM_FIRST_OPC);
 
         if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
@@ -944,8 +1079,9 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
 #endif
 
         if (unlikely(flags & LDLM_FL_REPLAY)) {
-                lock = find_existing_lock(req->rq_export,
-                                          &dlm_req->lock_handle[0]);
+                /* Find an existing lock in the per-export lock hash */
+                lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
+                                       (void *)&dlm_req->lock_handle[0]);
                 if (lock != NULL) {
                         DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
                                   LPX64, lock->l_handle.h_cookie);
@@ -957,13 +1093,12 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
         lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
                                 dlm_req->lock_desc.l_resource.lr_type,
                                 dlm_req->lock_desc.l_req_mode,
-                                cbs->lcs_blocking, cbs->lcs_completion,
-                                cbs->lcs_glimpse, NULL, 0);
+                                cbs, NULL, 0);
 
         if (!lock)
                 GOTO(out, rc = -ENOMEM);
 
-        do_gettimeofday(&lock->l_enqueued_time);
+        lock->l_last_activity = cfs_time_current_sec();
         lock->l_remote_handle = dlm_req->lock_handle[0];
         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 
@@ -974,11 +1109,12 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
                 GOTO(out, rc = -ENOTCONN);
         }
-        lock->l_export = class_export_get(req->rq_export);
-        spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
-        list_add(&lock->l_export_chain,
-                 &lock->l_export->exp_ldlm_data.led_held_locks);
-        spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
+
+        lock->l_export = class_export_lock_get(req->rq_export, lock);
+        if (lock->l_export->exp_lock_hash)
+                cfs_hash_add(lock->l_export->exp_lock_hash,
+                             &lock->l_remote_handle,
+                             &lock->l_exp_hash);
 
 existing_lock:
 
@@ -1038,9 +1174,9 @@ existing_lock:
                 if (lock->l_granted_mode == lock->l_req_mode) {
                         /*
                          * Only cancel lock if it was granted, because it would
-                         * be destroyed immediatelly and would never be granted
+                         * be destroyed immediately and would never be granted
                          * in the future, causing timeouts on client.  Not
-                         * granted lock will be cancelled immediatelly after
+                         * granted lock will be cancelled immediately after
                          * sending completion AST.
                          */
                         if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
@@ -1059,7 +1195,7 @@ existing_lock:
                 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
                              !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
                         CERROR("Granting sync lock to libclient. "
-                               "req fl %d, rep fl %d, lock fl %d\n",
+                               "req fl %d, rep fl %d, lock fl "LPX64"\n",
                                dlm_req->lock_flags, dlm_rep->lock_flags,
                                lock->l_flags);
                         LDLM_ERROR(lock, "sync lock");
@@ -1080,12 +1216,11 @@ existing_lock:
 
         EXIT;
  out:
-        req->rq_status = err;
-        if (req->rq_reply_state == NULL) {
+        req->rq_status = rc ?: err; /* return either error - bug 11190 */
+        if (!req->rq_packed_final) {
                 err = lustre_pack_reply(req, 1, NULL, NULL);
                 if (rc == 0)
                         rc = err;
-                req->rq_status = rc;
         }
 
         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
@@ -1116,7 +1251,7 @@ existing_lock:
                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
                         ldlm_reprocess_all(lock->l_resource);
 
-                LDLM_LOCK_PUT(lock);
+                LDLM_LOCK_RELEASE(lock);
         }
 
         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
@@ -1156,8 +1291,9 @@ int ldlm_handle_convert0(struct ptlrpc_request *req,
         int rc;
         ENTRY;
 
-        if (req->rq_export && req->rq_export->exp_ldlm_stats)
-                lprocfs_counter_incr(req->rq_export->exp_ldlm_stats,
+        if (req->rq_export && req->rq_export->exp_nid_stats &&
+            req->rq_export->exp_nid_stats->nid_ldlm_stats)
+                lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
                                      LDLM_CONVERT - LDLM_FIRST_OPC);
 
         rc = req_capsule_server_pack(&req->rq_pill);
@@ -1175,7 +1311,7 @@ int ldlm_handle_convert0(struct ptlrpc_request *req,
 
                 LDLM_DEBUG(lock, "server-side convert handler START");
 
-                do_gettimeofday(&lock->l_enqueued_time);
+                lock->l_last_activity = cfs_time_current_sec();
                 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
                                         &dlm_rep->lock_flags);
                 if (res) {
@@ -1213,7 +1349,7 @@ int ldlm_handle_convert(struct ptlrpc_request *req)
         return rc;
 }
 
-/* Cancel all the locks, which handles are packed into ldlm_request */
+/* Cancel all the locks whos handles are packed into ldlm_request */
 int ldlm_request_cancel(struct ptlrpc_request *req,
                         const struct ldlm_request *dlm_req, int first)
 {
@@ -1249,11 +1385,13 @@ int ldlm_request_cancel(struct ptlrpc_request *req,
                 if (res != pres) {
                         if (pres != NULL) {
                                 ldlm_reprocess_all(pres);
+                                LDLM_RESOURCE_DELREF(pres);
                                 ldlm_resource_putref(pres);
                         }
                         if (res != NULL) {
                                 ldlm_resource_getref(res);
-                                ldlm_res_lvbo_update(res, NULL, 0, 1);
+                                LDLM_RESOURCE_ADDREF(res);
+                                ldlm_res_lvbo_update(res, NULL, 1);
                         }
                         pres = res;
                 }
@@ -1262,6 +1400,7 @@ int ldlm_request_cancel(struct ptlrpc_request *req,
         }
         if (pres != NULL) {
                 ldlm_reprocess_all(pres);
+                LDLM_RESOURCE_DELREF(pres);
                 ldlm_resource_putref(pres);
         }
         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
@@ -1280,8 +1419,9 @@ int ldlm_handle_cancel(struct ptlrpc_request *req)
                 RETURN(-EFAULT);
         }
 
-        if (req->rq_export && req->rq_export->exp_ldlm_stats)
-                lprocfs_counter_incr(req->rq_export->exp_ldlm_stats,
+        if (req->rq_export && req->rq_export->exp_nid_stats &&
+            req->rq_export->exp_nid_stats->nid_ldlm_stats)
+                lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
                                      LDLM_CANCEL - LDLM_FIRST_OPC);
 
         rc = req_capsule_server_pack(&req->rq_pill);
@@ -1303,7 +1443,7 @@ void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
         int do_ast;
         ENTRY;
 
-        LDLM_DEBUG(lock, "client blocking AST callback handler START");
+        LDLM_DEBUG(lock, "client blocking AST callback handler");
 
         lock_res_and_lock(lock);
         lock->l_flags |= LDLM_FL_CBPENDING;
@@ -1315,18 +1455,18 @@ void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
         unlock_res_and_lock(lock);
 
         if (do_ast) {
-                LDLM_DEBUG(lock, "already unused, calling "
-                           "callback (%p)", lock->l_blocking_ast);
+                CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
+                       lock, lock->l_blocking_ast);
                 if (lock->l_blocking_ast != NULL)
                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
                                              LDLM_CB_BLOCKING);
         } else {
-                LDLM_DEBUG(lock, "Lock still has references, will be"
-                           " cancelled later");
+                CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
+                       lock);
         }
 
         LDLM_DEBUG(lock, "client blocking callback handler END");
-        LDLM_LOCK_PUT(lock);
+        LDLM_LOCK_RELEASE(lock);
         EXIT;
 }
 
@@ -1340,7 +1480,26 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 
         LDLM_DEBUG(lock, "client completion callback handler START");
 
+        if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
+                int to = cfs_time_seconds(1);
+                while (to > 0) {
+                        cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE, to);
+                        if (lock->l_granted_mode == lock->l_req_mode ||
+                            lock->l_destroyed)
+                                break;
+                }
+        }
+
         lock_res_and_lock(lock);
+        if (lock->l_destroyed ||
+            lock->l_granted_mode == lock->l_req_mode) {
+                /* bug 11300: the lock has already been granted */
+                unlock_res_and_lock(lock);
+                LDLM_DEBUG(lock, "Double grant race happened");
+                LDLM_LOCK_RELEASE(lock);
+                EXIT;
+                return;
+        }
 
         /* If we receive the completion AST before the actual enqueue returned,
          * then we might need to switch lock modes, resources, or extents. */
@@ -1359,14 +1518,22 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
                    &lock->l_resource->lr_name,
                    sizeof(lock->l_resource->lr_name)) != 0) {
                 unlock_res_and_lock(lock);
-                ldlm_lock_change_resource(ns, lock,
-                                        &dlm_req->lock_desc.l_resource.lr_name);
+                if (ldlm_lock_change_resource(ns, lock,
+                                &dlm_req->lock_desc.l_resource.lr_name) != 0) {
+                        LDLM_ERROR(lock, "Failed to allocate resource");
+                        LDLM_LOCK_RELEASE(lock);
+                        EXIT;
+                        return;
+                }
                 LDLM_DEBUG(lock, "completion AST, new resource");
                 CERROR("change resource!\n");
                 lock_res_and_lock(lock);
         }
 
         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
+                /* BL_AST locks are not needed in lru.
+                 * let ldlm_cancel_lru() be fast. */
+                ldlm_lock_remove_from_lru(lock);
                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
         }
@@ -1377,9 +1544,8 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
                         LDLM_ERROR(lock, "completion AST did not contain "
                                    "expected LVB!");
                 } else {
-                        void *lvb = req_capsule_client_swab_get(&req->rq_pill,
-                                                                &RMF_DLM_LVB,
-                                                  (void *)lock->l_lvb_swabber);
+                        void *lvb = req_capsule_client_get(&req->rq_pill,
+                                                           &RMF_DLM_LVB);
                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
                 }
         }
@@ -1389,11 +1555,11 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 
         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 
-        ldlm_run_cp_ast_work(&ast_list);
+        ldlm_run_ast_work(&ast_list, LDLM_WORK_CP_AST);
 
         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
                           lock);
-        LDLM_LOCK_PUT(lock);
+        LDLM_LOCK_RELEASE(lock);
         EXIT;
 }
 
@@ -1431,7 +1597,7 @@ static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
                 return;
         }
         unlock_res_and_lock(lock);
-        LDLM_LOCK_PUT(lock);
+        LDLM_LOCK_RELEASE(lock);
         EXIT;
 }
 
@@ -1441,7 +1607,7 @@ static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
                 return 0;
 
         req->rq_status = rc;
-        if (req->rq_reply_state == NULL) {
+        if (!req->rq_packed_final) {
                 rc = lustre_pack_reply(req, 1, NULL, NULL);
                 if (rc)
                         return rc;
@@ -1510,6 +1676,49 @@ int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
 #endif
 }
 
+/* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
+static int ldlm_handle_setinfo(struct ptlrpc_request *req)
+{
+        struct obd_device *obd = req->rq_export->exp_obd;
+        char *key;
+        void *val;
+        int keylen, vallen;
+        int rc = -ENOSYS;
+        ENTRY;
+
+        DEBUG_REQ(D_ERROR, req, "%s: handle setinfo\n", obd->obd_name);
+
+        req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
+
+        key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
+        if (key == NULL) {
+                DEBUG_REQ(D_IOCTL, req, "no set_info key");
+                RETURN(-EFAULT);
+        }
+        keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
+                                      RCL_CLIENT);
+        val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
+        if (val == NULL) {
+                DEBUG_REQ(D_IOCTL, req, "no set_info val");
+                RETURN(-EFAULT);
+        }
+        vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
+                                      RCL_CLIENT);
+
+        /* We are responsible for swabbing contents of val */
+
+        if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
+                /* Pass it on to mdc (the "export" in this case) */
+                rc = obd_set_info_async(req->rq_export,
+                                        sizeof(KEY_HSM_COPYTOOL_SEND),
+                                        KEY_HSM_COPYTOOL_SEND,
+                                        vallen, val, NULL);
+        else
+                DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
+
+        return rc;
+}
+
 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
 static int ldlm_callback_handler(struct ptlrpc_request *req)
 {
@@ -1524,24 +1733,13 @@ static int ldlm_callback_handler(struct ptlrpc_request *req)
          * incoming request message body, but I am responsible for the
          * message buffers. */
 
+        /* do nothing for sec context finalize */
+        if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
+                RETURN(0);
+
         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 
         if (req->rq_export == NULL) {
-                struct ldlm_request *dlm_req;
-
-                CDEBUG(D_RPCTRACE, "operation %d from %s with bad "
-                       "export cookie "LPX64"; this is "
-                       "normal if this node rebooted with a lock held\n",
-                       lustre_msg_get_opc(req->rq_reqmsg),
-                       libcfs_id2str(req->rq_peer),
-                       lustre_msg_get_handle(req->rq_reqmsg)->cookie);
-
-                req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
-                dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
-                if (dlm_req != NULL)
-                        CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
-                               dlm_req->lock_handle[0].cookie);
-
                 ldlm_callback_reply(req, -ENOTCONN);
                 RETURN(0);
         }
@@ -1562,11 +1760,18 @@ static int ldlm_callback_handler(struct ptlrpc_request *req)
                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK))
                         RETURN(0);
                 break;
+        case LDLM_SET_INFO:
+                rc = ldlm_handle_setinfo(req);
+                ldlm_callback_reply(req, rc);
+                RETURN(0);
         case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
+                CERROR("shouldn't be handling OBD_LOG_CANCEL on DLM thread\n");
                 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
                         RETURN(0);
                 rc = llog_origin_handle_cancel(req);
+                if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
+                        RETURN(0);
                 ldlm_callback_reply(req, rc);
                 RETURN(0);
         case OBD_QC_CALLBACK:
@@ -1611,9 +1816,6 @@ static int ldlm_callback_handler(struct ptlrpc_request *req)
                 rc = llog_origin_handle_close(req);
                 ldlm_callback_reply(req, rc);
                 RETURN(0);
-        case SEC_CTX_FINI:
-                /* do nothing */
-                RETURN(0);
         default:
                 CERROR("unknown opcode %u\n",
                        lustre_msg_get_opc(req->rq_reqmsg));
@@ -1632,7 +1834,16 @@ static int ldlm_callback_handler(struct ptlrpc_request *req)
                 RETURN(0);
         }
 
-        lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle[0]);
+        /* Force a known safe race, send a cancel to the server for a lock
+         * which the server has already started a blocking callback on. */
+        if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
+            lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
+                rc = ldlm_cli_cancel(&dlm_req->lock_handle[0]);
+                if (rc < 0)
+                        CERROR("ldlm_cli_cancel: %d\n", rc);
+        }
+
+        lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
         if (!lock) {
                 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
                        "disappeared\n", dlm_req->lock_handle[0].cookie);
@@ -1640,23 +1851,32 @@ static int ldlm_callback_handler(struct ptlrpc_request *req)
                 RETURN(0);
         }
 
+        if ((lock->l_flags & LDLM_FL_FAIL_LOC) &&
+            lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
+                OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
+
         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
         lock_res_and_lock(lock);
         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
-                /* If somebody cancels locks and cache is already droped,
+                /* If somebody cancels lock and cache is already dropped,
+                 * or lock is failed before cp_ast received on client,
                  * we can tell the server we have no lock. Otherwise, we
                  * should send cancel after dropping the cache. */
-                if ((lock->l_flags & LDLM_FL_CANCELING) &&
-                    (lock->l_flags & LDLM_FL_BL_DONE)) {
+                if (((lock->l_flags & LDLM_FL_CANCELING) &&
+                    (lock->l_flags & LDLM_FL_BL_DONE)) ||
+                    (lock->l_flags & LDLM_FL_FAILED)) {
                         LDLM_DEBUG(lock, "callback on lock "
                                    LPX64" - lock disappeared\n",
                                    dlm_req->lock_handle[0].cookie);
                         unlock_res_and_lock(lock);
-                        LDLM_LOCK_PUT(lock);
+                        LDLM_LOCK_RELEASE(lock);
                         ldlm_callback_reply(req, -EINVAL);
                         RETURN(0);
                 }
+                /* BL_AST locks are not needed in lru.
+                 * let ldlm_cancel_lru() be fast. */
+                ldlm_lock_remove_from_lru(lock);
                 lock->l_flags |= LDLM_FL_BL_AST;
         }
         unlock_res_and_lock(lock);
@@ -1717,11 +1937,14 @@ static int ldlm_cancel_handler(struct ptlrpc_request *req)
                        libcfs_id2str(req->rq_peer),
                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
 
-                req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
-                dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
-                if (dlm_req != NULL)
-                        ldlm_lock_dump_handle(D_ERROR,
-                                              &dlm_req->lock_handle[0]);
+                if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
+                        req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
+                        dlm_req = req_capsule_client_get(&req->rq_pill,
+                                                         &RMF_DLM_REQ);
+                        if (dlm_req != NULL)
+                                ldlm_lock_dump_handle(D_ERROR,
+                                                      &dlm_req->lock_handle[0]);
+                }
                 ldlm_callback_reply(req, -ENOTCONN);
                 RETURN(0);
         }
@@ -1743,6 +1966,8 @@ static int ldlm_cancel_handler(struct ptlrpc_request *req)
                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
                         RETURN(0);
                 rc = llog_origin_handle_cancel(req);
+                if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
+                        RETURN(0);
                 ldlm_callback_reply(req, rc);
                 RETURN(0);
         default:
@@ -1755,62 +1980,54 @@ static int ldlm_cancel_handler(struct ptlrpc_request *req)
         RETURN(0);
 }
 
-void ldlm_revoke_export_locks(struct obd_export *exp)
+void ldlm_revoke_lock_cb(void *obj, void *data)
 {
-        struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
-        struct list_head  rpc_list;
-        struct ldlm_lock *lock, *next;
-        struct ldlm_lock_desc desc;
-
-        ENTRY;
-        INIT_LIST_HEAD(&rpc_list);
+        struct list_head   *rpc_list = data;
+        struct ldlm_lock   *lock = obj;
 
-        spin_lock(&exp->exp_ldlm_data.led_lock);
-        list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
-                lock_res_and_lock(lock);
+        lock_res_and_lock(lock);
 
-                if (lock->l_req_mode != lock->l_granted_mode) {
-                        unlock_res_and_lock(lock);
-                        continue;
-                }
+        if (lock->l_req_mode != lock->l_granted_mode) {
+                unlock_res_and_lock(lock);
+                return;
+        }
 
-                LASSERT(lock->l_resource);
-                if (lock->l_resource->lr_type != LDLM_IBITS &&
-                    lock->l_resource->lr_type != LDLM_PLAIN) {
-                        unlock_res_and_lock(lock);
-                        continue;
-                }
+        LASSERT(lock->l_resource);
+        if (lock->l_resource->lr_type != LDLM_IBITS &&
+            lock->l_resource->lr_type != LDLM_PLAIN) {
+                unlock_res_and_lock(lock);
+                return;
+        }
 
-                if (lock->l_flags & LDLM_FL_AST_SENT) {
-                        unlock_res_and_lock(lock);
-                        continue;
-                }
+        if (lock->l_flags & LDLM_FL_AST_SENT) {
+                unlock_res_and_lock(lock);
+                return;
+        }
 
-                LASSERT(lock->l_blocking_ast);
-                LASSERT(!lock->l_blocking_lock);
+        LASSERT(lock->l_blocking_ast);
+        LASSERT(!lock->l_blocking_lock);
 
-                lock->l_flags |= LDLM_FL_AST_SENT;
-                list_move(&lock->l_export_chain, &rpc_list);
+        lock->l_flags |= LDLM_FL_AST_SENT;
+        if (lock->l_export && lock->l_export->exp_lock_hash &&
+            !hlist_unhashed(&lock->l_exp_hash))
+                cfs_hash_del(lock->l_export->exp_lock_hash,
+                             &lock->l_remote_handle, &lock->l_exp_hash);
+        list_add_tail(&lock->l_rk_ast, rpc_list);
+        LDLM_LOCK_GET(lock);
 
-                unlock_res_and_lock(lock);
-        }
-        spin_unlock(&exp->exp_ldlm_data.led_lock);
+        unlock_res_and_lock(lock);
+}
 
-        while (!list_empty(&rpc_list)) {
-                lock = list_entry(rpc_list.next, struct ldlm_lock,
-                                  l_export_chain);
-                list_del_init(&lock->l_export_chain);
+void ldlm_revoke_export_locks(struct obd_export *exp)
+{
+        struct list_head  rpc_list;
+        ENTRY;
 
-                /* the desc just pretend to exclusive */
-                ldlm_lock2desc(lock, &desc);
-                desc.l_req_mode = LCK_EX;
-                desc.l_granted_mode = 0;
+        CFS_INIT_LIST_HEAD(&rpc_list);
+        cfs_hash_for_each_empty(exp->exp_lock_hash,
+                                ldlm_revoke_lock_cb, &rpc_list);
+        ldlm_run_ast_work(&rpc_list, LDLM_WORK_REVOKE_AST);
 
-                LDLM_LOCK_GET(lock);
-                lock->l_blocking_ast(lock, &desc, lock->l_ast_data,
-                                     LDLM_CB_BLOCKING);
-                LDLM_LOCK_PUT(lock);
-        }
         EXIT;
 }
 
@@ -1942,7 +2159,7 @@ static int ldlm_bl_thread_main(void *arg)
 #endif
 
 static int ldlm_setup(void);
-static int ldlm_cleanup(int force);
+static int ldlm_cleanup(void);
 
 int ldlm_get_ref(void)
 {
@@ -1959,12 +2176,12 @@ int ldlm_get_ref(void)
         RETURN(rc);
 }
 
-void ldlm_put_ref(int force)
+void ldlm_put_ref(void)
 {
         ENTRY;
         mutex_down(&ldlm_ref_sem);
         if (ldlm_refcount == 1) {
-                int rc = ldlm_cleanup(force);
+                int rc = ldlm_cleanup();
                 if (rc)
                         CERROR("ldlm_cleanup failed: %d\n", rc);
                 else
@@ -1977,6 +2194,89 @@ void ldlm_put_ref(int force)
         EXIT;
 }
 
+/*
+ * Export handle<->lock hash operations.
+ */
+static unsigned
+ldlm_export_lock_hash(cfs_hash_t *hs, void *key, unsigned mask)
+{
+        return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
+}
+
+static void *
+ldlm_export_lock_key(struct hlist_node *hnode)
+{
+        struct ldlm_lock *lock;
+        ENTRY;
+
+        lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
+        RETURN(&lock->l_remote_handle);
+}
+
+static int
+ldlm_export_lock_compare(void *key, struct hlist_node *hnode)
+{
+        ENTRY;
+        RETURN(lustre_handle_equal(ldlm_export_lock_key(hnode), key));
+}
+
+static void *
+ldlm_export_lock_get(struct hlist_node *hnode)
+{
+        struct ldlm_lock *lock;
+        ENTRY;
+
+        lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
+        LDLM_LOCK_GET(lock);
+
+        RETURN(lock);
+}
+
+static void *
+ldlm_export_lock_put(struct hlist_node *hnode)
+{
+        struct ldlm_lock *lock;
+        ENTRY;
+
+        lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
+        LDLM_LOCK_RELEASE(lock);
+
+        RETURN(lock);
+}
+
+static cfs_hash_ops_t ldlm_export_lock_ops = {
+        .hs_hash    = ldlm_export_lock_hash,
+        .hs_key     = ldlm_export_lock_key,
+        .hs_compare = ldlm_export_lock_compare,
+        .hs_get     = ldlm_export_lock_get,
+        .hs_put     = ldlm_export_lock_put
+};
+
+int ldlm_init_export(struct obd_export *exp)
+{
+        ENTRY;
+
+        exp->exp_lock_hash =
+                cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
+                                HASH_EXP_LOCK_CUR_BITS, HASH_EXP_LOCK_MAX_BITS,
+                                &ldlm_export_lock_ops, CFS_HASH_REHASH);
+
+        if (!exp->exp_lock_hash)
+                RETURN(-ENOMEM);
+
+        RETURN(0);
+}
+EXPORT_SYMBOL(ldlm_init_export);
+
+void ldlm_destroy_export(struct obd_export *exp)
+{
+        ENTRY;
+        cfs_hash_destroy(exp->exp_lock_hash);
+        exp->exp_lock_hash = NULL;
+        EXIT;
+}
+EXPORT_SYMBOL(ldlm_destroy_export);
+
 static int ldlm_setup(void)
 {
         struct ldlm_bl_pool *blp;
@@ -2015,12 +2315,12 @@ 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, ldlm_timeout * 900,
+                                LDLM_CB_REPLY_PORTAL, 2,
                                 ldlm_callback_handler, "ldlm_cbd",
                                 ldlm_svc_proc_dir, NULL,
                                 ldlm_min_threads, ldlm_max_threads,
                                 "ldlm_cb",
-                                LCT_MD_THREAD|LCT_DT_THREAD);
+                                LCT_MD_THREAD|LCT_DT_THREAD, NULL);
 
         if (!ldlm_state->ldlm_cb_service) {
                 CERROR("failed to start service\n");
@@ -2030,12 +2330,13 @@ 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, ldlm_timeout * 6000,
+                                LDLM_CANCEL_REPLY_PORTAL, 6,
                                 ldlm_cancel_handler, "ldlm_canceld",
                                 ldlm_svc_proc_dir, NULL,
                                 ldlm_min_threads, ldlm_max_threads,
                                 "ldlm_cn",
-                                LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD);
+                                LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD,
+                                NULL);
 
         if (!ldlm_state->ldlm_cancel_service) {
                 CERROR("failed to start service\n");
@@ -2112,14 +2413,14 @@ static int ldlm_setup(void)
         return rc;
 }
 
-static int ldlm_cleanup(int force)
+static int ldlm_cleanup(void)
 {
 #ifdef __KERNEL__
         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
 #endif
         ENTRY;
 
-        if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) || 
+        if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
             !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
                 CERROR("ldlm still has namespaces; clean these up first.\n");
                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
@@ -2177,13 +2478,24 @@ int __init ldlm_init(void)
                 return -ENOMEM;
 
         ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
-                                           sizeof(struct ldlm_lock), 0,
-                                           SLAB_HWCACHE_ALIGN);
+                                      sizeof(struct ldlm_lock), 0,
+                                      SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU);
         if (ldlm_lock_slab == NULL) {
                 cfs_mem_cache_destroy(ldlm_resource_slab);
                 return -ENOMEM;
         }
 
+        ldlm_interval_slab = cfs_mem_cache_create("interval_node",
+                                        sizeof(struct ldlm_interval),
+                                        0, SLAB_HWCACHE_ALIGN);
+        if (ldlm_interval_slab == NULL) {
+                cfs_mem_cache_destroy(ldlm_resource_slab);
+                cfs_mem_cache_destroy(ldlm_lock_slab);
+                return -ENOMEM;
+        }
+#if LUSTRE_TRACKS_LOCK_EXP_REFS
+        class_export_dump_hook = ldlm_dump_export_locks;
+#endif
         return 0;
 }
 
@@ -2194,8 +2506,16 @@ 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);
+        LASSERTF(rc == 0, "couldn't free interval node slab\n");
 }
 
 /* ldlm_extent.c */
@@ -2214,18 +2534,22 @@ EXPORT_SYMBOL(ldlm_lock_put);
 EXPORT_SYMBOL(ldlm_lock_match);
 EXPORT_SYMBOL(ldlm_lock_cancel);
 EXPORT_SYMBOL(ldlm_lock_addref);
+EXPORT_SYMBOL(ldlm_lock_addref_try);
 EXPORT_SYMBOL(ldlm_lock_decref);
 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
 EXPORT_SYMBOL(ldlm_lock_change_resource);
-EXPORT_SYMBOL(ldlm_lock_set_data);
 EXPORT_SYMBOL(ldlm_it2str);
 EXPORT_SYMBOL(ldlm_lock_dump);
 EXPORT_SYMBOL(ldlm_lock_dump_handle);
-EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
+EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
 EXPORT_SYMBOL(ldlm_lock_allow_match);
+EXPORT_SYMBOL(ldlm_lock_downgrade);
+EXPORT_SYMBOL(ldlm_lock_convert);
 
 /* ldlm_request.c */
+EXPORT_SYMBOL(ldlm_completion_ast_async);
+EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
 EXPORT_SYMBOL(ldlm_completion_ast);
 EXPORT_SYMBOL(ldlm_blocking_ast);
 EXPORT_SYMBOL(ldlm_glimpse_ast);
@@ -2240,7 +2564,6 @@ EXPORT_SYMBOL(ldlm_cli_cancel);
 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
 EXPORT_SYMBOL(ldlm_cli_cancel_req);
-EXPORT_SYMBOL(ldlm_cli_join_lru);
 EXPORT_SYMBOL(ldlm_replay_locks);
 EXPORT_SYMBOL(ldlm_resource_foreach);
 EXPORT_SYMBOL(ldlm_namespace_foreach);
@@ -2282,7 +2605,7 @@ EXPORT_SYMBOL(client_obd_setup);
 EXPORT_SYMBOL(client_obd_cleanup);
 EXPORT_SYMBOL(client_connect_import);
 EXPORT_SYMBOL(client_disconnect_export);
-EXPORT_SYMBOL(target_start_recovery_thread);
+EXPORT_SYMBOL(server_disconnect_export);
 EXPORT_SYMBOL(target_stop_recovery_thread);
 EXPORT_SYMBOL(target_handle_connect);
 EXPORT_SYMBOL(target_cleanup_recovery);
@@ -2297,4 +2620,3 @@ EXPORT_SYMBOL(target_handle_disconnect);
 /* l_lock.c */
 EXPORT_SYMBOL(lock_res_and_lock);
 EXPORT_SYMBOL(unlock_res_and_lock);
-