Whamcloud - gitweb
b=17323
[fs/lustre-release.git] / lustre / ldlm / ldlm_lockd.c
index 36a1c80..83b08d0 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;
@@ -263,11 +278,11 @@ 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));
+                LDLM_ERROR(lock, "lock callback timer expired after %lds: "
+                           "evicting client at %s ",
+                           cfs_time_current_sec()- lock->l_enqueued_time.tv_sec,
+                           libcfs_nid2str(
+                                   lock->l_export->exp_connection->c_peer.nid));
 
                 last = lock;
 
@@ -307,21 +322,25 @@ repeat:
  */
 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock)
 {
+        int 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);
+        timeout = ldlm_get_enq_timeout(lock);
+
+        lock->l_callback_timeout = cfs_time_shift(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;
 }
@@ -467,7 +486,14 @@ static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
 
         if (obd_dump_on_timeout)
                 libcfs_debug_dumplog();
+#ifdef __KERNEL__
+        spin_lock_bh(&waiting_locks_spinlock);
+        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 +543,8 @@ 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 +556,18 @@ 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 
+                /* If client canceled the lock but the cancel has not
                  * been recieved 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, 
+                        ldlm_res_lvbo_update(lock->l_resource, NULL,
                                              0, 1);
-                rc = ldlm_handle_ast_error(lock, req, rc, 
+                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);
@@ -566,7 +593,7 @@ 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);
 }
@@ -649,10 +676,13 @@ 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_prep_req 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);
@@ -678,14 +708,15 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
         total_enqueue_wait = cfs_timeval_sub(&granted_time,
                                              &lock->l_enqueued_time, NULL);
 
-        if (total_enqueue_wait / 1000000 > obd_timeout)
-                LDLM_ERROR(lock, "enqueue wait took %luus from %lu",
+        if (total_enqueue_wait / ONE_MILLION > obd_timeout)
+                /* non-fatal with AT - change to LDLM_DEBUG? */
+                LDLM_ERROR(lock, "enqueue wait took %luus from "CFS_TIME_T,
                            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)
@@ -720,9 +751,17 @@ int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
                    total_enqueue_wait);
 
+        /* 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 / ONE_MILLION);
+
         ptlrpc_request_set_replen(req);
+
         req->rq_send_state = LUSTRE_IMP_FULL;
-        req->rq_timeout = ldlm_get_rq_timeout(ldlm_timeout, obd_timeout);
+        /* ptlrpc_prep_req 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);
@@ -747,8 +786,9 @@ 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 +826,13 @@ 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_prep_req 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);
@@ -801,27 +844,10 @@ int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
                 rc = ldlm_res_lvbo_update(res, req->rq_repmsg,
                                           REPLY_REC_OFF, 1);
         ptlrpc_req_finished(req);
-        RETURN(rc);
-}
+        if (rc == -ERESTART)
+                ldlm_reprocess_all(res);
 
-static struct ldlm_lock *
-find_existing_lock(struct obd_export *exp,
-                   const struct lustre_handle *remote_hdl)
-{
-        struct list_head *iter;
-
-        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__
@@ -892,8 +918,9 @@ 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 +971,9 @@ 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 = lustre_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);
@@ -957,8 +985,7 @@ 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);
@@ -975,10 +1002,11 @@ int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
                 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);
+
+        if (lock->l_export->exp_lock_hash)
+                lustre_hash_add(lock->l_export->exp_lock_hash,
+                                &lock->l_remote_handle,
+                                &lock->l_exp_hash);
 
 existing_lock:
 
@@ -1080,12 +1108,11 @@ 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 +1143,7 @@ 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 +1183,9 @@ 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);
@@ -1213,7 +1241,7 @@ 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)
 {
@@ -1249,10 +1277,12 @@ 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_RESOURCE_ADDREF(res);
                                 ldlm_res_lvbo_update(res, NULL, 0, 1);
                         }
                         pres = res;
@@ -1262,6 +1292,7 @@ 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 +1311,9 @@ 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);
@@ -1326,7 +1358,7 @@ void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
         }
 
         LDLM_DEBUG(lock, "client blocking callback handler END");
-        LDLM_LOCK_PUT(lock);
+        LDLM_LOCK_RELEASE(lock);
         EXIT;
 }
 
@@ -1343,7 +1375,7 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
                 int to = cfs_time_seconds(1);
                 while (to > 0) {
-                        to = schedule_timeout(to);
+                        cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE, to);
                         if (lock->l_granted_mode == lock->l_req_mode ||
                             lock->l_destroyed)
                                 break;
@@ -1356,7 +1388,7 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
                 /* bug 11300: the lock has already been granted */
                 unlock_res_and_lock(lock);
                 LDLM_DEBUG(lock, "Double grant race happened");
-                LDLM_LOCK_PUT(lock);
+                LDLM_LOCK_RELEASE(lock);
                 EXIT;
                 return;
         }
@@ -1378,10 +1410,10 @@ 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);
-                if (ldlm_lock_change_resource(ns, lock, 
+                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_PUT(lock);
+                        LDLM_LOCK_RELEASE(lock);
                         EXIT;
                         return;
                 }
@@ -1420,7 +1452,7 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 
         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
                           lock);
-        LDLM_LOCK_PUT(lock);
+        LDLM_LOCK_RELEASE(lock);
         EXIT;
 }
 
@@ -1458,7 +1490,7 @@ 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;
 }
 
@@ -1468,7 +1500,7 @@ 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;
@@ -1598,6 +1630,8 @@ static int ldlm_callback_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);
         case OBD_QC_CALLBACK:
@@ -1669,7 +1703,7 @@ static int ldlm_callback_handler(struct ptlrpc_request *req)
                         CERROR("ldlm_cli_cancel: %d\n", rc);
         }
 
-        lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle[0]);
+        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);
@@ -1690,7 +1724,7 @@ static int ldlm_callback_handler(struct ptlrpc_request *req)
                                    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);
                 }
@@ -1783,6 +1817,8 @@ 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:
@@ -1795,47 +1831,51 @@ 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 list_head   *rpc_list = data;
+        struct ldlm_lock   *lock = obj;
 
-        ENTRY;
-        CFS_INIT_LIST_HEAD(&rpc_list);
+        lock_res_and_lock(lock);
 
-        spin_lock(&exp->exp_ldlm_data.led_lock);
-        list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
-                lock_res_and_lock(lock);
+        if (lock->l_req_mode != lock->l_granted_mode) {
+                unlock_res_and_lock(lock);
+                return;
+        }
 
-                if (lock->l_req_mode != lock->l_granted_mode) {
-                        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;
+        }
 
-                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;
-                }
+        if (lock->l_flags & LDLM_FL_AST_SENT) {
+                unlock_res_and_lock(lock);
+                return;
+        }
 
-                if (lock->l_flags & LDLM_FL_AST_SENT) {
-                        unlock_res_and_lock(lock);
-                        continue;
-                }
+        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;
+        if (lock->l_export && lock->l_export->exp_lock_hash)
+                lustre_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);
 
-                lock->l_flags |= LDLM_FL_AST_SENT;
-                list_move(&lock->l_export_chain, &rpc_list);
-                LDLM_LOCK_GET(lock);
+        unlock_res_and_lock(lock);
+}
 
-                unlock_res_and_lock(lock);
-        }
-        spin_unlock(&exp->exp_ldlm_data.led_lock);
+void ldlm_revoke_export_locks(struct obd_export *exp)
+{
+        struct list_head  rpc_list;
+        ENTRY;
 
+        CFS_INIT_LIST_HEAD(&rpc_list);
+        lustre_hash_for_each_empty(exp->exp_lock_hash,
+                                   ldlm_revoke_lock_cb, &rpc_list);
         ldlm_run_ast_work(&rpc_list, LDLM_WORK_REVOKE_AST);
 
         EXIT;
@@ -1969,7 +2009,7 @@ 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)
 {
@@ -1986,12 +2026,12 @@ 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
@@ -2004,6 +2044,88 @@ void ldlm_put_ref(int force)
         EXIT;
 }
 
+/*
+ * Export handle<->lock hash operations.
+ */
+static unsigned
+ldlm_export_lock_hash(lustre_hash_t *lh, void *key, unsigned mask)
+{
+        return lh_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 lustre_hash_ops_t ldlm_export_lock_ops = {
+        .lh_hash    = ldlm_export_lock_hash,
+        .lh_key     = ldlm_export_lock_key,
+        .lh_compare = ldlm_export_lock_compare,
+        .lh_get     = ldlm_export_lock_get,
+        .lh_put     = ldlm_export_lock_put
+};
+
+int ldlm_init_export(struct obd_export *exp)
+{
+        ENTRY;
+
+        exp->exp_lock_hash =
+                lustre_hash_init(obd_uuid2str(&exp->exp_client_uuid),
+                                 128, 65536, &ldlm_export_lock_ops, LH_REHASH);
+
+        if (!exp->exp_lock_hash)
+                RETURN(-ENOMEM);
+
+        RETURN(0);
+}
+EXPORT_SYMBOL(ldlm_init_export);
+
+void ldlm_destroy_export(struct obd_export *exp)
+{
+        ENTRY;
+        lustre_hash_exit(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;
@@ -2042,7 +2164,7 @@ 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, ldlm_timeout * 900,
+                                LDLM_CB_REPLY_PORTAL, 1800,
                                 ldlm_callback_handler, "ldlm_cbd",
                                 ldlm_svc_proc_dir, NULL,
                                 ldlm_min_threads, ldlm_max_threads,
@@ -2057,7 +2179,7 @@ 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, ldlm_timeout * 6000,
+                                LDLM_CANCEL_REPLY_PORTAL, 6000,
                                 ldlm_cancel_handler, "ldlm_canceld",
                                 ldlm_svc_proc_dir, NULL,
                                 ldlm_min_threads, ldlm_max_threads,
@@ -2139,14 +2261,14 @@ 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);
@@ -2249,9 +2371,12 @@ EXPORT_SYMBOL(ldlm_lock2handle);
 EXPORT_SYMBOL(__ldlm_handle2lock);
 EXPORT_SYMBOL(ldlm_lock_get);
 EXPORT_SYMBOL(ldlm_lock_put);
+EXPORT_SYMBOL(ldlm_lock_fast_match);
+EXPORT_SYMBOL(ldlm_lock_fast_release);
 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);
@@ -2264,6 +2389,7 @@ EXPORT_SYMBOL(ldlm_reprocess_all_ns);
 EXPORT_SYMBOL(ldlm_lock_allow_match);
 
 /* ldlm_request.c */
+EXPORT_SYMBOL(ldlm_completion_ast_async);
 EXPORT_SYMBOL(ldlm_completion_ast);
 EXPORT_SYMBOL(ldlm_blocking_ast);
 EXPORT_SYMBOL(ldlm_glimpse_ast);
@@ -2278,7 +2404,6 @@ 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);
@@ -2335,4 +2460,3 @@ EXPORT_SYMBOL(target_handle_disconnect);
 /* l_lock.c */
 EXPORT_SYMBOL(lock_res_and_lock);
 EXPORT_SYMBOL(unlock_res_and_lock);
-