Whamcloud - gitweb
b=18618
[fs/lustre-release.git] / lustre / ldlm / ldlm_request.c
index 889dac7..1eaa0e1 100644 (file)
@@ -1,25 +1,37 @@
 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
  * vim:expandtab:shiftwidth=8:tabstop=8:
  *
- *  Copyright (C) 2002, 2003 Cluster File Systems, Inc.
+ * 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.
  */
 
 #define DEBUG_SUBSYSTEM S_LDLM
 
 #include "ldlm_internal.h"
 
+int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
+CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644,
+                "lock enqueue timeout minimum");
+
 static void interrupted_completion_wait(void *data)
 {
 }
@@ -61,10 +77,12 @@ int ldlm_expired_completion_wait(void *data)
                 if (ptlrpc_check_suspend())
                         RETURN(0);
 
-                LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago); "
-                           "not entering recovery in server code, just going "
-                           "back to sleep", lock->l_enqueued_time.tv_sec,
-                           CURRENT_SECONDS - lock->l_enqueued_time.tv_sec);
+                LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
+                           CFS_DURATION_T"s ago); not entering recovery in "
+                           "server code, just going back to sleep",
+                           lock->l_last_activity,
+                           cfs_time_sub(cfs_time_current_sec(),
+                           lock->l_last_activity));
                 if (cfs_time_after(cfs_time_current(), next_dump)) {
                         last_dump = next_dump;
                         next_dump = cfs_time_shift(300);
@@ -79,28 +97,103 @@ int ldlm_expired_completion_wait(void *data)
         obd = lock->l_conn_export->exp_obd;
         imp = obd->u.cli.cl_import;
         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
-        LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago), entering "
-                   "recovery for %s@%s", lock->l_enqueued_time.tv_sec,
-                   CURRENT_SECONDS - lock->l_enqueued_time.tv_sec,
-                   obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
+        LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
+                  CFS_DURATION_T"s ago), entering recovery for %s@%s",
+                  lock->l_last_activity,
+                  cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
+                  obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
 
         RETURN(0);
 }
 
-static int is_granted_or_cancelled(struct ldlm_lock *lock)
+/* We use the same basis for both server side and client side functions
+   from a single node. */
+int ldlm_get_enq_timeout(struct ldlm_lock *lock)
 {
-        int ret = 0;
+        int timeout = at_get(&lock->l_resource->lr_namespace->ns_at_estimate);
+        if (AT_OFF)
+                return obd_timeout / 2;
+        /* Since these are non-updating timeouts, we should be conservative.
+           It would be nice to have some kind of "early reply" mechanism for
+           lock callbacks too... */
+        timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
+        return max(timeout, ldlm_enqueue_min);
+}
+EXPORT_SYMBOL(ldlm_get_enq_timeout);
 
-        lock_res_and_lock(lock);
-        if (((lock->l_req_mode == lock->l_granted_mode) &&
-             !(lock->l_flags & LDLM_FL_CP_REQD)) ||
-            (lock->l_flags & LDLM_FL_FAILED))
-                ret = 1;
-        unlock_res_and_lock(lock);
+/**
+ * Helper function for ldlm_completion_ast(), updating timings when lock is
+ * actually granted.
+ */
+static int ldlm_completion_tail(struct ldlm_lock *lock)
+{
+        long delay;
+        int  result;
+
+        if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
+                LDLM_DEBUG(lock, "client-side enqueue: destroyed");
+                result = -EIO;
+        } else {
+                delay = cfs_time_sub(cfs_time_current_sec(),
+                                     lock->l_last_activity);
+                LDLM_DEBUG(lock, "client-side enqueue: granted after "
+                           CFS_DURATION_T"s", delay);
+
+                /* Update our time estimate */
+                at_add(&lock->l_resource->lr_namespace->ns_at_estimate, delay);
+                result = 0;
+        }
+        return result;
+}
+
+/**
+ * Implementation of ->l_completion_ast() for a client, that doesn't wait
+ * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
+ * other threads that cannot block for long.
+ */
+int ldlm_completion_ast_async(struct ldlm_lock *lock, int flags, void *data)
+{
+        ENTRY;
+
+        if (flags == LDLM_FL_WAIT_NOREPROC) {
+                LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
+                RETURN(0);
+        }
+
+        if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
+                       LDLM_FL_BLOCK_CONV))) {
+                cfs_waitq_signal(&lock->l_waitq);
+                RETURN(ldlm_completion_tail(lock));
+        }
 
-        return ret;
+        LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
+                   "going forward");
+        ldlm_lock_dump(D_OTHER, lock, 0);
+        ldlm_reprocess_all(lock->l_resource);
+        RETURN(0);
 }
 
+/**
+ * Client side LDLM "completion" AST. This is called in several cases:
+ *
+ *     - when a reply to an ENQUEUE rpc is received from the server
+ *       (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
+ *       this point (determined by flags);
+ *
+ *     - when LDLM_CP_CALLBACK rpc comes to client to notify it that lock has
+ *       been granted;
+ *
+ *     - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
+ *       gets correct lvb;
+ *
+ *     - to force all locks when resource is destroyed (cleanup_resource());
+ *
+ *     - during lock conversion (not used currently).
+ *
+ * If lock is not granted in the first case, this function waits until second
+ * or penultimate cases happen in some other thread.
+ *
+ */
 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 {
         /* XXX ALLOCATE - 160 bytes */
@@ -108,6 +201,7 @@ int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
         struct obd_device *obd;
         struct obd_import *imp = NULL;
         struct l_wait_info lwi;
+        __u32 timeout;
         int rc = 0;
         ENTRY;
 
@@ -125,15 +219,20 @@ int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
                    "sleeping");
         ldlm_lock_dump(D_OTHER, lock, 0);
-        ldlm_reprocess_all(lock->l_resource);
 
 noreproc:
 
         obd = class_exp2obd(lock->l_conn_export);
 
         /* if this is a local lock, then there is no import */
-        if (obd != NULL)
+        if (obd != NULL) {
                 imp = obd->u.cli.cl_import;
+        }
+
+        /* Wait a long time for enqueue - server may have to callback a
+           lock from another client.  Server will evict the other client if it
+           doesn't respond reasonably, and then give us the lock. */
+        timeout = ldlm_get_enq_timeout(lock) * 2;
 
         lwd.lwd_lock = lock;
 
@@ -141,7 +240,7 @@ noreproc:
                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
         } else {
-                lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
+                lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
                                        ldlm_expired_completion_wait,
                                        interrupted_completion_wait, &lwd);
         }
@@ -152,12 +251,15 @@ noreproc:
                 spin_unlock(&imp->imp_lock);
         }
 
-        /* Go to sleep until the lock is granted or cancelled. */
-        rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
-
-        if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
-                LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
-                RETURN(-EIO);
+        if (ns_is_client(lock->l_resource->lr_namespace) &&
+            OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
+                                 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
+                lock->l_flags |= LDLM_FL_FAIL_LOC;
+                rc = -EINTR;
+        } else {
+                /* Go to sleep until the lock is granted or cancelled. */
+                rc = l_wait_event(lock->l_waitq,
+                                  is_granted_or_cancelled(lock), &lwi);
         }
 
         if (rc) {
@@ -166,35 +268,25 @@ noreproc:
                 RETURN(rc);
         }
 
-        LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
-        RETURN(0);
+        RETURN(ldlm_completion_tail(lock));
 }
 
-/*
- * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
+/**
+ * A helper to build a blocking ast function
+ *
+ * Perform a common operation for blocking asts:
+ * defferred lock cancellation.
+ *
+ * \param lock the lock blocking or canceling ast was called on
+ * \retval 0
+ * \see mdt_blocking_ast
+ * \see ldlm_blocking_ast
  */
-int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
-                      void *data, int flag)
+int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
 {
         int do_ast;
         ENTRY;
 
-        if (flag == LDLM_CB_CANCELING) {
-                /* Don't need to do anything here. */
-                RETURN(0);
-        }
-
-        lock_res_and_lock(lock);
-        /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
-         * that ldlm_blocking_ast is called just before intent_policy method
-         * takes the ns_lock, then by the time we get the lock, we might not
-         * be the correct blocking function anymore.  So check, and return
-         * early, if so. */
-        if (lock->l_blocking_ast != ldlm_blocking_ast) {
-                unlock_res_and_lock(lock);
-                RETURN(0);
-        }
-
         lock->l_flags |= LDLM_FL_CBPENDING;
         do_ast = (!lock->l_readers && !lock->l_writers);
         unlock_res_and_lock(lock);
@@ -215,6 +307,42 @@ int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
         RETURN(0);
 }
 
+/**
+ * Server blocking AST
+ *
+ * ->l_blocking_ast() callback for LDLM locks acquired by server-side
+ * OBDs.
+ *
+ * \param lock the lock which blocks a request or cancelling lock
+ * \param desc unused
+ * \param data unused
+ * \param flag indicates whether this cancelling or blocking callback
+ * \retval 0
+ * \see ldlm_blocking_ast_nocheck
+ */
+int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
+                      void *data, int flag)
+{
+        ENTRY;
+
+        if (flag == LDLM_CB_CANCELING) {
+                /* Don't need to do anything here. */
+                RETURN(0);
+        }
+
+        lock_res_and_lock(lock);
+        /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
+         * that ldlm_blocking_ast is called just before intent_policy method
+         * takes the ns_lock, then by the time we get the lock, we might not
+         * be the correct blocking function anymore.  So check, and return
+         * early, if so. */
+        if (lock->l_blocking_ast != ldlm_blocking_ast) {
+                unlock_res_and_lock(lock);
+                RETURN(0);
+        }
+        RETURN(ldlm_blocking_ast_nocheck(lock));
+}
+
 /*
  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
  * comment in filter_intent_policy() on why you may need this.
@@ -250,10 +378,15 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
                            ldlm_completion_callback completion,
                            ldlm_glimpse_callback glimpse,
                            void *data, __u32 lvb_len, void *lvb_swabber,
+                           const __u64 *client_cookie,
                            struct lustre_handle *lockh)
 {
         struct ldlm_lock *lock;
         int err;
+        const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
+                                                 .lcs_blocking   = blocking,
+                                                 .lcs_glimpse    = glimpse,
+        };
         ENTRY;
 
         LASSERT(!(*flags & LDLM_FL_REPLAY));
@@ -262,8 +395,7 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
                 LBUG();
         }
 
-        lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
-                                completion, glimpse, data, lvb_len);
+        lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len);
         if (unlikely(!lock))
                 GOTO(out_nolock, err = -ENOMEM);
         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
@@ -278,6 +410,8 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
         unlock_res_and_lock(lock);
         if (policy != NULL)
                 lock->l_policy_data = *policy;
+        if (client_cookie != NULL)
+                lock->l_client_cookie = *client_cookie;
         if (type == LDLM_EXTENT)
                 lock->l_req_extent = policy->l_extent;
 
@@ -297,7 +431,7 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
         LDLM_DEBUG(lock, "client-side local enqueue END");
         EXIT;
  out:
-        LDLM_LOCK_PUT(lock);
+        LDLM_LOCK_RELEASE(lock);
  out_nolock:
         return err;
 }
@@ -306,13 +440,31 @@ static void failed_lock_cleanup(struct ldlm_namespace *ns,
                                 struct ldlm_lock *lock,
                                 struct lustre_handle *lockh, int mode)
 {
+        int need_cancel = 0;
+
         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
         lock_res_and_lock(lock);
-        lock->l_flags |= LDLM_FL_LOCAL_ONLY;
+        /* Check that lock is not granted or failed, we might race. */
+        if ((lock->l_req_mode != lock->l_granted_mode) &&
+            !(lock->l_flags & LDLM_FL_FAILED)) {
+                /* Make sure that this lock will not be found by raced
+                 * bl_ast and -EINVAL reply is sent to server anyways.
+                 * bug 17645 */
+                lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
+                                 LDLM_FL_ATOMIC_CB;
+                need_cancel = 1;
+        }
         unlock_res_and_lock(lock);
-        LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 
-        ldlm_lock_decref_and_cancel(lockh, mode);
+        if (need_cancel) {
+                LDLM_DEBUG(lock,
+                           "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
+                           "LDLM_FL_ATOMIC_CB");
+                ldlm_lock_decref_and_cancel(lockh, mode);
+        } else {
+                LDLM_DEBUG(lock, "lock was granted or failed in race");
+                ldlm_lock_decref(lockh, mode);
+        }
 
         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
          *       from llite/file.c/ll_file_flock(). */
@@ -328,6 +480,7 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 {
         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
         int is_replay = *flags & LDLM_FL_REPLAY;
+        struct lustre_handle old_hash_key;
         struct ldlm_lock *lock;
         struct ldlm_reply *reply;
         int cleanup_phase = 1;
@@ -346,19 +499,19 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
                 if (rc == ELDLM_LOCK_ABORTED) {
                         /* Before we return, swab the reply */
-                        reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF,
-                                                   sizeof(*reply),
-                                                   lustre_swab_ldlm_reply);
-                        if (reply == NULL) {
-                                CERROR("Can't unpack ldlm_reply\n");
+                        reply = req_capsule_server_get(&req->rq_pill,
+                                                       &RMF_DLM_REP);
+                        if (reply == NULL)
                                 rc = -EPROTO;
-                        }
                         if (lvb_len) {
-                                void *tmplvb;
-                                tmplvb = lustre_swab_repbuf(req,
-                                                            DLM_REPLY_REC_OFF,
-                                                            lvb_len,
-                                                            lvb_swabber);
+                                struct ost_lvb *tmplvb;
+
+                                req_capsule_set_size(&req->rq_pill,
+                                                     &RMF_DLM_LVB, RCL_SERVER,
+                                                     lvb_len);
+                            tmplvb = req_capsule_server_swab_get(&req->rq_pill,
+                                                                 &RMF_DLM_LVB,
+                                                                 lvb_swabber);
                                 if (tmplvb == NULL)
                                         GOTO(cleanup, rc = -EPROTO);
                                 if (lvb != NULL)
@@ -368,18 +521,23 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
                 GOTO(cleanup, rc);
         }
 
-        reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
-                                   lustre_swab_ldlm_reply);
-        if (reply == NULL) {
-                CERROR("Can't unpack ldlm_reply\n");
+        reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
+        if (reply == NULL)
                 GOTO(cleanup, rc = -EPROTO);
-        }
 
         /* lock enqueued on the server */
         cleanup_phase = 0;
 
         lock_res_and_lock(lock);
+        old_hash_key = lock->l_remote_handle;
         lock->l_remote_handle = reply->lock_handle;
+
+        /* Key change rehash lock in per-export hash with new key */
+        if (exp->exp_lock_hash)
+                lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
+                                       &lock->l_remote_handle,
+                                       &lock->l_exp_hash);
+
         *flags = reply->lock_flags;
         lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
@@ -415,12 +573,10 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
                               (long)lock->l_resource->lr_name.name[1],
                               (long)lock->l_resource->lr_name.name[2]);
 
-                        ldlm_lock_change_resource(ns, lock,
-                                          &reply->lock_desc.l_resource.lr_name);
-                        if (lock->l_resource == NULL) {
-                                LBUG();
+                        rc = ldlm_lock_change_resource(ns, lock,
+                                        &reply->lock_desc.l_resource.lr_name);
+                        if (rc || lock->l_resource == NULL)
                                 GOTO(cleanup, rc = -ENOMEM);
-                        }
                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
                 }
                 if (with_policy)
@@ -447,8 +603,12 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
          * clobber the LVB with an older one. */
         if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
                 void *tmplvb;
-                tmplvb = lustre_swab_repbuf(req, DLM_REPLY_REC_OFF, lvb_len,
-                                            lvb_swabber);
+
+                req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
+                                     lvb_len);
+                tmplvb = req_capsule_server_swab_get(&req->rq_pill,
+                                                     &RMF_DLM_LVB,
+                                                     lvb_swabber);
                 if (tmplvb == NULL)
                         GOTO(cleanup, rc = -EPROTO);
                 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
@@ -478,86 +638,114 @@ cleanup:
                 failed_lock_cleanup(ns, lock, lockh, mode);
         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
         LDLM_LOCK_PUT(lock);
-        LDLM_LOCK_PUT(lock);
+        LDLM_LOCK_RELEASE(lock);
         return rc;
 }
 
 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
  * a single page on the send/receive side. XXX: 512 should be changed
  * to more adequate value. */
-static inline int ldlm_req_handles_avail(struct obd_export *exp,
-                                         int *size, int bufcount, int off)
+static inline int ldlm_req_handles_avail(int req_size, int off)
 {
-        int avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512);
-        int old_size = size[DLM_LOCKREQ_OFF];
+        int avail;
 
-        size[DLM_LOCKREQ_OFF] = sizeof(struct ldlm_request);
-        avail -= lustre_msg_size(class_exp2cliimp(exp)->imp_msg_magic,
-                                 bufcount, size);
-        avail /= sizeof(struct lustre_handle);
+        avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
+        if (likely(avail >= 0))
+                avail /= (int)sizeof(struct lustre_handle);
+        else
+                avail = 0;
         avail += LDLM_LOCKREQ_HANDLES - off;
-        size[DLM_LOCKREQ_OFF] = old_size;
 
         return avail;
 }
 
-static inline int ldlm_cancel_handles_avail(struct obd_export *exp)
+static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
+                                             enum req_location loc,
+                                             int off)
 {
-        int size[2] = { sizeof(struct ptlrpc_body),
-                        sizeof(struct ldlm_request) };
-        return ldlm_req_handles_avail(exp, size, 2, 0);
+        int size = req_capsule_msg_size(pill, loc);
+        return ldlm_req_handles_avail(size, off);
+}
+
+static inline int ldlm_format_handles_avail(struct obd_import *imp,
+                                            const struct req_format *fmt,
+                                            enum req_location loc, int off)
+{
+        int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
+        return ldlm_req_handles_avail(size, off);
 }
 
 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
  * @count locks in @cancels. */
-struct ptlrpc_request *ldlm_prep_enqueue_req(struct obd_export *exp,
-                                             int bufcount, int *size,
-                                             struct list_head *cancels,
-                                             int count)
+int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
+                      int version, int opc, int canceloff,
+                      struct list_head *cancels, int count)
 {
-        struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
-        struct ldlm_request *dlm = NULL;
-        struct ptlrpc_request *req;
+        struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
+        struct req_capsule      *pill = &req->rq_pill;
+        struct ldlm_request     *dlm = NULL;
+        int flags, avail, to_free, bufcount, pack = 0;
         CFS_LIST_HEAD(head);
+        int rc;
         ENTRY;
 
         if (cancels == NULL)
                 cancels = &head;
         if (exp_connect_cancelset(exp)) {
                 /* Estimate the amount of available space in the request. */
-                int avail = ldlm_req_handles_avail(exp, size, bufcount,
-                                                   LDLM_ENQUEUE_CANCEL_OFF);
-                int flags, cancel;
+                bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
+                avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
 
-                LASSERT(avail >= count);
-
-                flags = ns_connect_lru_resize(ns) ? 
+                flags = ns_connect_lru_resize(ns) ?
                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
-                cancel = ns_connect_lru_resize(ns) ? 0 : 1;
+                to_free = !ns_connect_lru_resize(ns) &&
+                          opc == LDLM_ENQUEUE ? 1 : 0;
 
-                /* Cancel lru locks here _only_ if the server supports 
+                /* Cancel lru locks here _only_ if the server supports
                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
-                 * rpc right on enqueue, what will make it slower, vs. 
-                 * asynchronous rpc in blocking thread. */
-                count += ldlm_cancel_lru_local(ns, cancels, cancel,
-                                               avail - count, 0, flags);
-                size[DLM_LOCKREQ_OFF] =
-                        ldlm_request_bufsize(count, LDLM_ENQUEUE);
-        }
-        req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
-                              LDLM_ENQUEUE, bufcount, size, NULL);
-        if (exp_connect_cancelset(exp) && req) {
-                dlm = lustre_msg_buf(req->rq_reqmsg,
-                                     DLM_LOCKREQ_OFF, sizeof(*dlm));
-                /* Skip first lock handler in ldlm_request_pack(), this method
-                 * will incrment @lock_count according to the lock handle amount
-                 * actually written to the buffer. */
-                dlm->lock_count = LDLM_ENQUEUE_CANCEL_OFF;
-                ldlm_cli_cancel_list(cancels, count, req, DLM_LOCKREQ_OFF, 0);
+                 * rpc, what will make us slower. */
+                if (avail > count)
+                        count += ldlm_cancel_lru_local(ns, cancels, to_free,
+                                                       avail - count, 0, flags);
+                if (avail > count)
+                        pack = count;
+                else
+                        pack = avail;
+                req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
+                                     ldlm_request_bufsize(pack, opc));
+        }
+
+        rc = ptlrpc_request_pack(req, version, opc);
+        if (rc) {
+                ldlm_lock_list_put(cancels, l_bl_ast, count);
+                RETURN(rc);
+        }
+
+        if (exp_connect_cancelset(exp)) {
+                if (canceloff) {
+                        dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
+                        LASSERT(dlm);
+                        /* Skip first lock handler in ldlm_request_pack(),
+                         * this method will incrment @lock_count according
+                         * to the lock handle amount actually written to
+                         * the buffer. */
+                        dlm->lock_count = canceloff;
+                }
+                /* Pack into the request @pack lock handles. */
+                ldlm_cli_cancel_list(cancels, pack, req, 0);
+                /* Prepare and send separate cancel rpc for others. */
+                ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
         } else {
                 ldlm_lock_list_put(cancels, l_bl_ast, count);
         }
-        RETURN(req);
+        RETURN(0);
+}
+
+int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
+                          struct list_head *cancels, int count)
+{
+        return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
+                                 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
 }
 
 /* If a request has some specific initialisation it is passed in @reqp,
@@ -574,14 +762,11 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
                      struct lustre_handle *lockh, int async)
 {
         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
-        struct ldlm_lock *lock;
-        struct ldlm_request *body;
-        struct ldlm_reply *reply;
-        int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
-                        [DLM_LOCKREQ_OFF]     = sizeof(*body),
-                        [DLM_REPLY_REC_OFF]   = lvb_len };
-        int is_replay = *flags & LDLM_FL_REPLAY;
-        int req_passed_in = 1, rc, err;
+        struct ldlm_lock      *lock;
+        struct ldlm_request   *body;
+        int                    is_replay = *flags & LDLM_FL_REPLAY;
+        int                    req_passed_in = 1;
+        int                    rc, err;
         struct ptlrpc_request *req;
         ENTRY;
 
@@ -590,15 +775,20 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
         /* If we're replaying this lock, just check some invariants.
          * If we're creating a new lock, get everything all setup nice. */
         if (is_replay) {
-                lock = ldlm_handle2lock(lockh);
+                lock = ldlm_handle2lock_long(lockh, 0);
                 LASSERT(lock != NULL);
                 LDLM_DEBUG(lock, "client-side enqueue START");
                 LASSERT(exp == lock->l_conn_export);
         } else {
+                const struct ldlm_callback_suite cbs = {
+                        .lcs_completion = einfo->ei_cb_cp,
+                        .lcs_blocking   = einfo->ei_cb_bl,
+                        .lcs_glimpse    = einfo->ei_cb_gl,
+                        .lcs_weigh      = einfo->ei_cb_wg
+                };
                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
-                                        einfo->ei_mode, einfo->ei_cb_bl,
-                                        einfo->ei_cb_cp, einfo->ei_cb_gl,
-                                        einfo->ei_cbdata, lvb_len);
+                                        einfo->ei_mode, &cbs, einfo->ei_cbdata,
+                                        lvb_len);
                 if (lock == NULL)
                         RETURN(-ENOMEM);
                 /* for the local lock, add the reference */
@@ -628,22 +818,26 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
         /* lock not sent to server yet */
 
         if (reqp == NULL || *reqp == NULL) {
-                req = ldlm_prep_enqueue_req(exp, 2, size, NULL, 0);
+                req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
+                                                &RQF_LDLM_ENQUEUE,
+                                                LUSTRE_DLM_VERSION,
+                                                LDLM_ENQUEUE);
                 if (req == NULL) {
                         failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
-                        LDLM_LOCK_PUT(lock);
+                        LDLM_LOCK_RELEASE(lock);
                         RETURN(-ENOMEM);
                 }
                 req_passed_in = 0;
                 if (reqp)
                         *reqp = req;
         } else {
+                int len;
+
                 req = *reqp;
-                LASSERTF(lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF) >=
-                         sizeof(*body), "buflen[%d] = %d, not "LPSZ"\n",
-                         DLM_LOCKREQ_OFF,
-                         lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF),
-                         sizeof(*body));
+                len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
+                                           RCL_CLIENT);
+                LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
+                         DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
         }
 
         lock->l_conn_export = exp;
@@ -651,15 +845,20 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
         lock->l_blocking_ast = einfo->ei_cb_bl;
 
         /* Dump lock data into the request buffer */
-        body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
+        body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
         ldlm_lock2desc(lock, &body->lock_desc);
         body->lock_flags = *flags;
         body->lock_handle[0] = *lockh;
 
         /* Continue as normal. */
         if (!req_passed_in) {
-                size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
-                ptlrpc_req_set_repsize(req, 2 + (lvb_len > 0), size);
+                if (lvb_len > 0) {
+                        req_capsule_extend(&req->rq_pill,
+                                           &RQF_LDLM_ENQUEUE_LVB);
+                        req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
+                                             RCL_SERVER, lvb_len);
+                }
+                ptlrpc_request_set_replen(req);
         }
 
         /*
@@ -676,7 +875,9 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
         }
 
         LDLM_DEBUG(lock, "sending request");
+
         rc = ptlrpc_queue_wait(req);
+
         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
                                     einfo->ei_mode, flags, lvb, lvb_len,
                                     lvb_swabber, lockh, rc);
@@ -684,7 +885,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
          * one reference that we took */
         if (err == -ENOLCK)
-                LDLM_LOCK_PUT(lock);
+                LDLM_LOCK_RELEASE(lock);
         else
                 rc = err;
 
@@ -698,7 +899,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 }
 
 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
-                                  int *flags)
+                                  __u32 *flags)
 {
         struct ldlm_resource *res;
         int rc;
@@ -725,16 +926,14 @@ static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
  * conversion of locks which are on the waiting or converting queue */
 /* Caller of this code is supposed to take care of lock readers/writers
    accounting */
-int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
+int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
 {
-        struct ldlm_request *body;
-        struct ldlm_reply *reply;
-        struct ldlm_lock *lock;
-        struct ldlm_resource *res;
+        struct ldlm_request   *body;
+        struct ldlm_reply     *reply;
+        struct ldlm_lock      *lock;
+        struct ldlm_resource  *res;
         struct ptlrpc_request *req;
-        int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
-                        [DLM_LOCKREQ_OFF]     = sizeof(*body) };
-        int rc;
+        int                    rc;
         ENTRY;
 
         lock = ldlm_handle2lock(lockh);
@@ -749,30 +948,29 @@ int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 
         LDLM_DEBUG(lock, "client-side convert");
 
-        req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
-                              LUSTRE_DLM_VERSION, LDLM_CONVERT, 2, size, NULL);
-        if (!req)
-                GOTO(out, rc = -ENOMEM);
+        req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
+                                        &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
+                                        LDLM_CONVERT);
+        if (req == NULL) {
+                LDLM_LOCK_PUT(lock);
+                RETURN(-ENOMEM);
+        }
 
-        body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
+        body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
         body->lock_handle[0] = lock->l_remote_handle;
 
         body->lock_desc.l_req_mode = new_mode;
         body->lock_flags = *flags;
 
-        size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
-        ptlrpc_req_set_repsize(req, 2, size);
 
+        ptlrpc_request_set_replen(req);
         rc = ptlrpc_queue_wait(req);
         if (rc != ELDLM_OK)
                 GOTO(out, rc);
 
-        reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
-                                   lustre_swab_ldlm_reply);
-        if (reply == NULL) {
-                CERROR ("Can't unpack ldlm_reply\n");
-                GOTO (out, rc = -EPROTO);
-        }
+        reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
+        if (reply == NULL)
+                GOTO(out, rc = -EPROTO);
 
         if (req->rq_status)
                 GOTO(out, rc = req->rq_status);
@@ -807,7 +1005,7 @@ static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
 {
         int rc = LDLM_FL_LOCAL_ONLY;
         ENTRY;
-        
+
         if (lock->l_conn_export) {
                 int local_only;
 
@@ -844,7 +1042,7 @@ static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
 
 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
    of the request @req. */
-static void ldlm_cancel_pack(struct ptlrpc_request *req, int off,
+static void ldlm_cancel_pack(struct ptlrpc_request *req,
                              struct list_head *head, int count)
 {
         struct ldlm_request *dlm;
@@ -852,11 +1050,11 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req, int off,
         int max, packed = 0;
         ENTRY;
 
-        dlm = lustre_msg_buf(req->rq_reqmsg, off, sizeof(*dlm));
+        dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
         LASSERT(dlm != NULL);
 
         /* Check the room in the request buffer. */
-        max = lustre_msg_buflen(req->rq_reqmsg, off) - 
+        max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
                 sizeof(struct ldlm_request);
         max /= sizeof(struct lustre_handle);
         max += LDLM_LOCKREQ_HANDLES;
@@ -884,9 +1082,6 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
                         int count, int flags)
 {
         struct ptlrpc_request *req = NULL;
-        struct ldlm_request *body;
-        int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
-                [DLM_LOCKREQ_OFF]     = sizeof(*body) };
         struct obd_import *imp;
         int free, sent = 0;
         int rc = 0;
@@ -895,15 +1090,19 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
         LASSERT(exp != NULL);
         LASSERT(count > 0);
 
+        OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val);
+
         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
                 RETURN(count);
 
-        free = ldlm_req_handles_avail(exp, size, 2, 0);
+        free = ldlm_format_handles_avail(class_exp2cliimp(exp),
+                                         &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
         if (count > free)
                 count = free;
 
-        size[DLM_LOCKREQ_OFF] = ldlm_request_bufsize(count, LDLM_CANCEL);
         while (1) {
+                int bufcount;
+
                 imp = class_exp2cliimp(exp);
                 if (imp == NULL || imp->imp_invalid) {
                         CDEBUG(D_DLMTRACE,
@@ -911,25 +1110,31 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
                         RETURN(count);
                 }
 
-                req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL, 2,
-                                      size, NULL);
-                if (!req)
+                req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
+                if (req == NULL)
                         GOTO(out, rc = -ENOMEM);
 
+                bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
+                req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
+                                     ldlm_request_bufsize(count, LDLM_CANCEL));
+
+                rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
+                if (rc) {
+                        ptlrpc_request_free(req);
+                        GOTO(out, rc);
+                }
                 req->rq_no_resend = 1;
                 req->rq_no_delay = 1;
 
-                /* XXX FIXME bug 249 */
                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
+                ptlrpc_at_set_req_timeout(req);
 
-                body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF,
-                                      sizeof(*body));
-                ldlm_cancel_pack(req, DLM_LOCKREQ_OFF, cancels, count);
+                ldlm_cancel_pack(req, cancels, count);
 
-                ptlrpc_req_set_repsize(req, 1, NULL);
+                ptlrpc_request_set_replen(req);
                 if (flags & LDLM_FL_ASYNC) {
-                        ptlrpcd_add_req(req);
+                        ptlrpcd_add_req(req, PSCOPE_OTHER);
                         sent = count;
                         GOTO(out, 0);
                 } else {
@@ -966,58 +1171,56 @@ static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
         return &imp->imp_obd->obd_namespace->ns_pool;
 }
 
+/**
+ * Update client's obd pool related fields with new SLV and Limit from \a req.
+ */
 int ldlm_cli_update_pool(struct ptlrpc_request *req)
 {
-        struct ldlm_pool *pl;
+        struct obd_device *obd;
+        __u64 old_slv, new_slv;
+        __u32 new_limit;
         ENTRY;
-    
-        if (!imp_connect_lru_resize(req->rq_import))
-                RETURN(0);
-
-        if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
-            lustre_msg_get_limit(req->rq_repmsg) == 0)
-                RETURN(0);
-
-        pl = ldlm_imp2pl(req->rq_import);
-        
-        spin_lock(&pl->pl_lock);
-
-        /* Check if we need to wakeup pools thread for fast SLV change. 
-         * This is only done when threads period is noticably long like 
-         * 10s or more. */
-#if defined(__KERNEL__) && (LDLM_POOLS_THREAD_PERIOD >= 10)
+        if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
+                     !imp_connect_lru_resize(req->rq_import)))
         {
-                __u64 old_slv, new_slv, fast_change;
-
-                old_slv = ldlm_pool_get_slv(pl);
-                new_slv = lustre_msg_get_slv(req->rq_repmsg);
-                fast_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE;
-                do_div(fast_change, 100);
-
-                /* Wake up pools thread only if SLV has changed more than 
-                 * 50% since last update. In this case we want to react asap. 
-                 * Otherwise it is no sense to wake up pools as they are 
-                 * re-calculated every LDLM_POOLS_THREAD_PERIOD anyways. */
-                if (old_slv > new_slv && old_slv - new_slv > fast_change)
-                        ldlm_pools_wakeup();
+                /*
+                 * Do nothing for corner cases.
+                 */
+                RETURN(0);
         }
-#endif
-        /* In some cases RPC may contain slv and limit zeroed out. This is 
+
+        /*
+         * In some cases RPC may contain slv and limit zeroed out. This is
          * the case when server does not support lru resize feature. This is
          * also possible in some recovery cases when server side reqs have no
-         * ref to obd export and thus access to server side namespace is no 
-         * possible. */
-        if (lustre_msg_get_slv(req->rq_repmsg) != 0 && 
-            lustre_msg_get_limit(req->rq_repmsg) != 0) {
-                ldlm_pool_set_slv(pl, lustre_msg_get_slv(req->rq_repmsg));
-                ldlm_pool_set_limit(pl, lustre_msg_get_limit(req->rq_repmsg));
-        } else {
-                DEBUG_REQ(D_HA, req, "zero SLV or Limit found "
-                          "(SLV: "LPU64", Limit: %u)", 
-                          lustre_msg_get_slv(req->rq_repmsg), 
+         * ref to obd export and thus access to server side namespace is no
+         * possible.
+         */
+        if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
+            lustre_msg_get_limit(req->rq_repmsg) == 0) {
+                DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
+                          "(SLV: "LPU64", Limit: %u)",
+                          lustre_msg_get_slv(req->rq_repmsg),
                           lustre_msg_get_limit(req->rq_repmsg));
+                RETURN(0);
         }
-        spin_unlock(&pl->pl_lock);
+
+        new_limit = lustre_msg_get_limit(req->rq_repmsg);
+        new_slv = lustre_msg_get_slv(req->rq_repmsg);
+        obd = req->rq_import->imp_obd;
+
+        /*
+         * Set new SLV and Limit to obd fields to make accessible for pool
+         * thread. We do not access obd_namespace and pool directly here
+         * as there is no reliable way to make sure that they are still
+         * alive in cleanup time. Evil races are possible which may cause
+         * oops in that time.
+         */
+        write_lock(&obd->obd_pool_lock);
+        old_slv = obd->obd_pool_slv;
+        obd->obd_pool_slv = new_slv;
+        obd->obd_pool_limit = new_limit;
+        write_unlock(&obd->obd_pool_lock);
 
         RETURN(0);
 }
@@ -1025,6 +1228,7 @@ EXPORT_SYMBOL(ldlm_cli_update_pool);
 
 int ldlm_cli_cancel(struct lustre_handle *lockh)
 {
+        struct obd_export *exp;
         int avail, flags, count = 1, rc = 0;
         struct ldlm_namespace *ns;
         struct ldlm_lock *lock;
@@ -1032,7 +1236,7 @@ int ldlm_cli_cancel(struct lustre_handle *lockh)
         ENTRY;
 
         /* concurrent cancels on the same handle can happen */
-        lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
+        lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
         if (lock == NULL) {
                 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
                 RETURN(0);
@@ -1040,7 +1244,7 @@ int ldlm_cli_cancel(struct lustre_handle *lockh)
 
         rc = ldlm_cli_cancel_local(lock);
         if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
-                LDLM_LOCK_PUT(lock);
+                LDLM_LOCK_RELEASE(lock);
                 RETURN(rc < 0 ? rc : 0);
         }
         /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
@@ -1048,14 +1252,21 @@ int ldlm_cli_cancel(struct lustre_handle *lockh)
          * here and send them all as one LDLM_CANCEL rpc. */
         LASSERT(list_empty(&lock->l_bl_ast));
         list_add(&lock->l_bl_ast, &cancels);
-        avail = ldlm_cancel_handles_avail(lock->l_conn_export);
-        LASSERT(avail > 0);
-
-        ns = lock->l_resource->lr_namespace;
-        flags = ns_connect_lru_resize(ns) ? LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
-        count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - count,
-                                       LDLM_FL_BL_AST, flags);
-        ldlm_cli_cancel_list(&cancels, count, NULL, 0, 0);
+
+        exp = lock->l_conn_export;
+        if (exp_connect_cancelset(exp)) {
+                avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
+                                                  &RQF_LDLM_CANCEL,
+                                                  RCL_CLIENT, 0);
+                LASSERT(avail > 0);
+
+                ns = lock->l_resource->lr_namespace;
+                flags = ns_connect_lru_resize(ns) ?
+                        LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
+                count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
+                                               LDLM_FL_BL_AST, flags);
+        }
+        ldlm_cli_cancel_list(&cancels, count, NULL, 0);
         RETURN(0);
 }
 
@@ -1088,125 +1299,144 @@ static int ldlm_cancel_list(struct list_head *cancels, int count, int flags)
                 if (rc == LDLM_FL_LOCAL_ONLY) {
                         /* CANCEL RPC should not be sent to server. */
                         list_del_init(&lock->l_bl_ast);
-                        LDLM_LOCK_PUT(lock);
+                        LDLM_LOCK_RELEASE(lock);
                         count--;
                 }
 
         }
         if (bl_ast > 0) {
                 count -= bl_ast;
-                ldlm_cli_cancel_list(&head, bl_ast, NULL, 0, 0);
+                ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
         }
 
         RETURN(count);
 }
 
-/* Return 1 if @lock should be canceled according to shrinker policy. 
- * Return zero otherwise. */
-static int ldlm_cancel_shrink_policy(struct ldlm_namespace *ns,
-                                     struct ldlm_lock *lock,
-                                     int unused, int added, 
-                                     int asked)
-{
-        int lock_cost;
-        __u64 page_nr;
-
-        if (lock->l_resource->lr_type == LDLM_EXTENT) {
-                struct ldlm_extent *l_extent;
-
-                /* For all extent locks cost is 1 + number of pages in
-                 * their extent. */
-                l_extent = &lock->l_policy_data.l_extent;
-                page_nr = (l_extent->end - l_extent->start);
-                do_div(page_nr, CFS_PAGE_SIZE);
-
-#ifdef __KERNEL__
-                /* XXX: In fact this is evil hack, we can't access inode
-                 * here. For doing it right we need somehow to have number
-                 * of covered by lock. This should be fixed later when 10718 
-                 * is landed. */
-                if (lock->l_ast_data != NULL) {
-                        struct inode *inode = lock->l_ast_data;
-                        if (page_nr > inode->i_mapping->nrpages)
-                                page_nr = inode->i_mapping->nrpages;
-                }
-#endif
-                lock_cost = 1 + page_nr;
-        } else {
-                /* For all locks which are not extent ones cost is 1 */
-                lock_cost = 1;
-        }
-
-        /* Keep all expensive locks in lru for the memory pressure time
-         * cancel policy. They anyways may be canceled by lru resize
-         * pplicy if they have not small enough CLV. */
-        return (lock_cost <= ns->ns_shrink_thumb);
-}
-
-/* Return 1 if @lock should be canceled according to lru resize policy. 
- * Return zero otherwise. */
-static int ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
-                                   struct ldlm_lock *lock, 
-                                   int unused, int added, 
-                                   int asked)
+/**
+ * Callback function for lru-resize policy. Makes decision whether to keep
+ * \a lock in LRU for current \a LRU size \a unused, added in current scan
+ * \a added and number of locks to be preferably canceled \a count.
+ *
+ * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
+ *
+ * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
+ */
+static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
+                                                 struct ldlm_lock *lock,
+                                                 int unused, int added,
+                                                 int count)
 {
         cfs_time_t cur = cfs_time_current();
         struct ldlm_pool *pl = &ns->ns_pool;
         __u64 slv, lvf, lv;
         cfs_time_t la;
 
-        spin_lock(&pl->pl_lock);
-        slv = ldlm_pool_get_slv(pl);
-        lvf = atomic_read(&pl->pl_lock_volume_factor);
-        spin_unlock(&pl->pl_lock);
+        /*
+         * Stop lru processing when we reached passed @count or checked all
+         * locks in lru.
+         */
+        if (count && added >= count)
+                return LDLM_POLICY_KEEP_LOCK;
 
-        la = cfs_duration_sec(cfs_time_sub(cur, 
+        slv = ldlm_pool_get_slv(pl);
+        lvf = ldlm_pool_get_lvf(pl);
+        la = cfs_duration_sec(cfs_time_sub(cur,
                               lock->l_last_used));
 
-        /* Stop when slv is not yet come from server or 
-         * lv is smaller than it is. */
+        /*
+         * Stop when slv is not yet come from server or lv is smaller than
+         * it is.
+         */
         lv = lvf * la * unused;
-        return (slv > 1 && lv >= slv);
+
+        /*
+         * Inform pool about current CLV to see it via proc.
+         */
+        ldlm_pool_set_clv(pl, lv);
+        return (slv == 1 || lv < slv) ?
+                LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
 }
 
-/* Return 1 if @lock should be canceled according to passed policy. 
- * Return zero otherwise. */
-static int ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
-                                     struct ldlm_lock *lock, 
-                                     int unused, int added,
-                                     int asked)
+/**
+ * Callback function for proc used policy. Makes decision whether to keep
+ * \a lock in LRU for current \a LRU size \a unused, added in current scan
+ * \a added and number of locks to be preferably canceled \a count.
+ *
+ * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
+ *
+ * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
+ */
+static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
+                                                   struct ldlm_lock *lock,
+                                                   int unused, int added,
+                                                   int count)
+{
+        /*
+         * Stop lru processing when we reached passed @count or checked all
+         * locks in lru.
+         */
+        return (added >= count) ?
+                LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
+}
+
+/**
+ * Callback function for aged policy. Makes decision whether to keep
+ * \a lock in LRU for current \a LRU size \a unused, added in current scan
+ * \a added and number of locks to be preferably canceled \a count.
+ *
+ * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
+ *
+ * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
+ */
+static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
+                                                 struct ldlm_lock *lock,
+                                                 int unused, int added,
+                                                 int count)
 {
-        /* Do nothing here, we allow canceling all locks which
-         * are passed here from upper layer logic. So that locks
-         * number to be canceled will be limited by @count and
-         * @max in ldlm_cancel_lru_local(). */
-        return 1;
+        /*
+         * Stop lru processing if young lock is found and we reached passed
+         * @count.
+         */
+        return ((added >= count) &&
+                cfs_time_before(cfs_time_current(),
+                                cfs_time_add(lock->l_last_used,
+                                             ns->ns_max_age))) ?
+                LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
 }
 
-/* Return 1 if @lock should be canceled according to aged policy. 
- * Return zero otherwise. */
-static int ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
-                                   struct ldlm_lock *lock, 
-                                   int unused, int added,
-                                   int asked)
+/**
+ * Callback function for default policy. Makes decision whether to keep
+ * \a lock in LRU for current \a LRU size \a unused, added in current scan
+ * \a added and number of locks to be preferably canceled \a count.
+ *
+ * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
+ *
+ * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
+ */
+static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
+                                                    struct ldlm_lock *lock,
+                                                    int unused, int added,
+                                                    int count)
 {
-        /* Cancel old locks if reached asked limit. */
-        return !((added >= asked) && 
-                 cfs_time_before_64(cfs_time_current(),
-                                    cfs_time_add(lock->l_last_used,
-                                                 ns->ns_max_age)));
+        /*
+         * Stop lru processing when we reached passed @count or checked all
+         * locks in lru.
+         */
+        return (added >= count) ?
+                LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
 }
 
-typedef int (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *, 
-                                        struct ldlm_lock *, int, 
-                                        int, int);
+typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
+                                                      struct ldlm_lock *, int,
+                                                      int, int);
 
 static ldlm_cancel_lru_policy_t
 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
 {
         if (ns_connect_lru_resize(ns)) {
                 if (flags & LDLM_CANCEL_SHRINK)
-                        return ldlm_cancel_shrink_policy;
+                        /* We kill passed number of old locks. */
+                        return ldlm_cancel_passed_policy;
                 else if (flags & LDLM_CANCEL_LRUR)
                         return ldlm_cancel_lrur_policy;
                 else if (flags & LDLM_CANCEL_PASSED)
@@ -1215,9 +1445,10 @@ ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
                 if (flags & LDLM_CANCEL_AGED)
                         return ldlm_cancel_aged_policy;
         }
-        return NULL;
+
+        return ldlm_cancel_default_policy;
 }
+
 /* - Free space in lru for @count new locks,
  *   redundant unused locks are canceled locally;
  * - also cancel locally unused aged locks;
@@ -1240,14 +1471,16 @@ ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
  *                              the beginning of lru list);
  *
  * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
- *                              memory pressre policy function.
+ *                              memory pressre policy function;
+ *
+ * flags & LDLM_CANCEL_AGED -   cancel alocks according to "aged policy".
  */
 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
                           int count, int max, int cancel_flags, int flags)
 {
-        ldlm_cancel_lru_policy_t cancel_lru_policy_func;
-        int added = 0, unused, cancel;
+        ldlm_cancel_lru_policy_t pf;
         struct ldlm_lock *lock, *next;
+        int added = 0, unused;
         ENTRY;
 
         spin_lock(&ns->ns_unused_lock);
@@ -1256,87 +1489,102 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
         if (!ns_connect_lru_resize(ns))
                 count += unused - ns->ns_max_unused;
 
-        cancel_lru_policy_func = ldlm_cancel_lru_policy(ns, flags);
-     
-        list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru) {
-                /* Make sure that we skip locks being already in cancel. */
-                if ((lock->l_flags & LDLM_FL_CANCELING) ||
-                    (lock->l_flags & LDLM_FL_BL_AST))
-                        continue;
+        pf = ldlm_cancel_lru_policy(ns, flags);
+        LASSERT(pf != NULL);
 
-                /* For any flags, stop scanning if @max or passed @count is
-                 * reached. */
-                if ((max && added >= max) || (count && added >= count))
+        while (!list_empty(&ns->ns_unused_list)) {
+                /* For any flags, stop scanning if @max is reached. */
+                if (max && added >= max)
                         break;
 
-                /* Pass the lock through the policy filter and see if it
-                 * should stay in lru. */
-                if (cancel_lru_policy_func != NULL) {
-                        cancel = cancel_lru_policy_func(ns, lock, unused, 
-                                                        added, count);
-                     
-                        /* Take next lock for shrink policy, we need to check
-                         * whole list. Stop scanning for other policies. */
-                        if ((flags & LDLM_CANCEL_SHRINK) && !cancel)
-                                continue;
-                        else if (!cancel)
+                list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru){
+                        /* No locks which got blocking requests. */
+                        LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
+
+                        /* Somebody is already doing CANCEL. No need in this
+                         * lock in lru, do not traverse it again. */
+                        if (!(lock->l_flags & LDLM_FL_CANCELING))
                                 break;
+
+                        ldlm_lock_remove_from_lru_nolock(lock);
                 }
+                if (&lock->l_lru == &ns->ns_unused_list)
+                        break;
 
-                if (cancels != NULL) {
-                        LDLM_LOCK_GET(lock); /* dropped by bl thread */
-                        spin_unlock(&ns->ns_unused_lock);
-
-                        lock_res_and_lock(lock);
-                        /* Check flags again under the lock. */
-                        if ((lock->l_flags & LDLM_FL_CANCELING) ||
-                            (lock->l_flags & LDLM_FL_BL_AST) ||
-                            (ldlm_lock_remove_from_lru(lock) == 0)) {
-                                /* other thread is removing lock from lru or
-                                 * somebody is already doing CANCEL or
-                                 * there is a blocking request which will send
-                                 * cancel by itseft. */
-                                unlock_res_and_lock(lock);
-                                LDLM_LOCK_PUT(lock);
-                                spin_lock(&ns->ns_unused_lock);
-                                continue;
-                        }
-                        LASSERT(!lock->l_readers && !lock->l_writers);
-
-                        /* If we have chosen to cancel this lock voluntarily, we
-                         * better send cancel notification to server, so that it
-                         * frees appropriate state. This might lead to a race 
-                         * where while we are doing cancel here, server is also 
-                         * silently cancelling this lock. */
-                        lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
-
-                        /* Setting the CBPENDING flag is a little misleading,
-                         * but prevents an important race; namely, once
-                         * CBPENDING is set, the lock can accumulate no more
-                         * readers/writers. Since readers and writers are
-                         * already zero here, ldlm_lock_decref() won't see
-                         * this flag and call l_blocking_ast */
-                        lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
-
-                        /* We can't re-add to l_lru as it confuses the
-                         * refcounting in ldlm_lock_remove_from_lru() if an AST
-                         * arrives after we drop ns_lock below. We use l_bl_ast
-                         * and can't use l_pending_chain as it is used both on
-                         * server and client nevertheless bug 5666 says it is
-                         * used only on server */
-                        LASSERT(list_empty(&lock->l_bl_ast));
-                        list_add(&lock->l_bl_ast, cancels);
+                LDLM_LOCK_GET(lock);
+                spin_unlock(&ns->ns_unused_lock);
+                lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
+
+                /* Pass the lock through the policy filter and see if it
+                 * should stay in lru.
+                 *
+                 * Even for shrinker policy we stop scanning if
+                 * we find a lock that should stay in the cache.
+                 * We should take into account lock age anyway
+                 * as new lock even if it is small of weight is
+                 * valuable resource.
+                 *
+                 * That is, for shrinker policy we drop only
+                 * old locks, but additionally chose them by
+                 * their weight. Big extent locks will stay in
+                 * the cache. */
+                if (pf(ns, lock, unused, added, count) ==
+                    LDLM_POLICY_KEEP_LOCK) {
+                        lu_ref_del(&lock->l_reference,
+                                   __FUNCTION__, cfs_current());
+                        LDLM_LOCK_RELEASE(lock);
+                        spin_lock(&ns->ns_unused_lock);
+                        break;
+                }
+
+                lock_res_and_lock(lock);
+                /* Check flags again under the lock. */
+                if ((lock->l_flags & LDLM_FL_CANCELING) ||
+                    (ldlm_lock_remove_from_lru(lock) == 0)) {
+                        /* other thread is removing lock from lru or
+                         * somebody is already doing CANCEL or
+                         * there is a blocking request which will send
+                         * cancel by itseft or the lock is matched
+                         * is already not unused. */
                         unlock_res_and_lock(lock);
+                        lu_ref_del(&lock->l_reference,
+                                   __FUNCTION__, cfs_current());
+                        LDLM_LOCK_RELEASE(lock);
                         spin_lock(&ns->ns_unused_lock);
+                        continue;
                 }
+                LASSERT(!lock->l_readers && !lock->l_writers);
+
+                /* If we have chosen to cancel this lock voluntarily, we
+                 * better send cancel notification to server, so that it
+                 * frees appropriate state. This might lead to a race
+                 * where while we are doing cancel here, server is also
+                 * silently cancelling this lock. */
+                lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
+
+                /* Setting the CBPENDING flag is a little misleading,
+                 * but prevents an important race; namely, once
+                 * CBPENDING is set, the lock can accumulate no more
+                 * readers/writers. Since readers and writers are
+                 * already zero here, ldlm_lock_decref() won't see
+                 * this flag and call l_blocking_ast */
+                lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
+
+                /* We can't re-add to l_lru as it confuses the
+                 * refcounting in ldlm_lock_remove_from_lru() if an AST
+                 * arrives after we drop ns_lock below. We use l_bl_ast
+                 * and can't use l_pending_chain as it is used both on
+                 * server and client nevertheless bug 5666 says it is
+                 * used only on server */
+                LASSERT(list_empty(&lock->l_bl_ast));
+                list_add(&lock->l_bl_ast, cancels);
+                unlock_res_and_lock(lock);
+                lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
+                spin_lock(&ns->ns_unused_lock);
                 added++;
                 unused--;
         }
         spin_unlock(&ns->ns_unused_lock);
-  
-        if (cancels == NULL)
-                RETURN(added);
-
         RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
 }
 
@@ -1344,7 +1592,7 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
  * in a thread and this function will return after the thread has been
  * asked to call the callback.  when called with LDLM_SYNC the blocking
  * callback will be performed in this function. */
-int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync, 
+int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
                     int flags)
 {
         CFS_LIST_HEAD(cancels);
@@ -1361,9 +1609,9 @@ int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
                         RETURN(count);
         }
 
-        /* If an error occured in ASYNC mode, or
-         * this is SYNC mode, cancel the list. */
-        ldlm_cli_cancel_list(&cancels, count, NULL, 0, 0);
+        /* If an error occured in ASYNC mode, or this is SYNC mode,
+         * cancel the list. */
+        ldlm_cli_cancel_list(&cancels, count, NULL, 0);
         RETURN(count);
 }
 
@@ -1399,7 +1647,7 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res,
 
                 /* If somebody is already doing CANCEL, or blocking ast came,
                  * skip this lock. */
-                if (lock->l_flags & LDLM_FL_BL_AST || 
+                if (lock->l_flags & LDLM_FL_BL_AST ||
                     lock->l_flags & LDLM_FL_CANCELING)
                         continue;
 
@@ -1427,14 +1675,14 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res,
         RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
 }
 
-/* If @req is NULL, send CANCEL request to server with handles of locks 
- * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests 
+/* If @req is NULL, send CANCEL request to server with handles of locks
+ * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
  * separately per lock.
- * If @req is not NULL, put handles of locks in @cancels into the request 
+ * If @req is not NULL, put handles of locks in @cancels into the request
  * buffer at the offset @off.
  * Destroy @cancels at the end. */
 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
-                         struct ptlrpc_request *req, int off, int flags)
+                         struct ptlrpc_request *req, int flags)
 {
         struct ldlm_lock *lock;
         int res = 0;
@@ -1442,8 +1690,8 @@ int ldlm_cli_cancel_list(struct list_head *cancels, int count,
 
         if (list_empty(cancels) || count == 0)
                 RETURN(0);
-        
-        /* XXX: requests (both batched and not) could be sent in parallel. 
+
+        /* XXX: requests (both batched and not) could be sent in parallel.
          * Usually it is enough to have just 1 RPC, but it is possible that
          * there are to many locks to be cancelled in LRU or on a resource.
          * It would also speed up the case when the server does not support
@@ -1456,7 +1704,7 @@ int ldlm_cli_cancel_list(struct list_head *cancels, int count,
                 if (exp_connect_cancelset(lock->l_conn_export)) {
                         res = count;
                         if (req)
-                                ldlm_cancel_pack(req, off, cancels, count);
+                                ldlm_cancel_pack(req, cancels, count);
                         else
                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
                                                           cancels, count,
@@ -1474,7 +1722,6 @@ int ldlm_cli_cancel_list(struct list_head *cancels, int count,
                 count -= res;
                 ldlm_lock_list_put(cancels, l_bl_ast, res);
         }
-        LASSERT(list_empty(cancels));
         LASSERT(count == 0);
         RETURN(0);
 }
@@ -1497,12 +1744,14 @@ int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
                 RETURN(0);
         }
 
+        LDLM_RESOURCE_ADDREF(res);
         count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
                                            0, flags, opaque);
-        rc = ldlm_cli_cancel_list(&cancels, count, NULL, 0, flags);
+        rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
         if (rc != ELDLM_OK)
                 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
 
+        LDLM_RESOURCE_DELREF(res);
         ldlm_resource_putref(res);
         RETURN(0);
 }
@@ -1551,6 +1800,7 @@ int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
                         ldlm_resource_getref(res);
                         spin_unlock(&ns->ns_hash_lock);
 
+                        LDLM_RESOURCE_ADDREF(res);
                         rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
                                                              NULL, LCK_MINMODE,
                                                              flags, opaque);
@@ -1559,6 +1809,7 @@ int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
                                 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
                                        res->lr_name.name[0], rc);
 
+                        LDLM_RESOURCE_DELREF(res);
                         spin_lock(&ns->ns_hash_lock);
                         tmp = tmp->next;
                         ldlm_resource_putref_locked(res);
@@ -1569,55 +1820,6 @@ int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
         RETURN(ELDLM_OK);
 }
 
-/* join/split resource locks to/from lru list */
-int ldlm_cli_join_lru(struct ldlm_namespace *ns,
-                      const struct ldlm_res_id *res_id, int join)
-{
-        struct ldlm_resource *res;
-        struct ldlm_lock *lock, *n;
-        int count = 0;
-        ENTRY;
-
-        LASSERT(ns_is_client(ns));
-
-        res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
-        if (res == NULL)
-                RETURN(count);
-        LASSERT(res->lr_type == LDLM_EXTENT);
-
-        lock_res(res);
-        if (!join)
-                goto split;
-
-        list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
-                if (list_empty(&lock->l_lru) &&
-                    !lock->l_readers && !lock->l_writers &&
-                    !(lock->l_flags & LDLM_FL_LOCAL) &&
-                    !(lock->l_flags & LDLM_FL_CBPENDING)) {
-                        ldlm_lock_add_to_lru(lock);
-                        lock->l_flags &= ~LDLM_FL_NO_LRU;
-                        LDLM_DEBUG(lock, "join lock to lru");
-                        count++;
-                }
-        }
-        goto unlock;
-split:
-        spin_lock(&ns->ns_unused_lock);
-        list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
-                if (lock->l_resource == res) {
-                        ldlm_lock_remove_from_lru_nolock(lock);
-                        lock->l_flags |= LDLM_FL_NO_LRU;
-                        LDLM_DEBUG(lock, "split lock from lru");
-                        count++;
-                }
-        }
-        spin_unlock(&ns->ns_unused_lock);
-unlock:
-        unlock_res(res);
-        ldlm_resource_putref(res);
-        RETURN(count);
-}
-
 /* Lock iterators. */
 
 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
@@ -1696,9 +1898,11 @@ int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
                         ldlm_resource_getref(res);
                         spin_unlock(&ns->ns_hash_lock);
+                        LDLM_RESOURCE_ADDREF(res);
 
                         rc = iter(res, closure);
 
+                        LDLM_RESOURCE_DELREF(res);
                         spin_lock(&ns->ns_hash_lock);
                         tmp = tmp->next;
                         ldlm_resource_putref_locked(res);
@@ -1730,7 +1934,9 @@ void ldlm_resource_iterate(struct ldlm_namespace *ns,
                 return;
         }
 
+        LDLM_RESOURCE_ADDREF(res);
         ldlm_resource_foreach(res, iter, data);
+        LDLM_RESOURCE_DELREF(res);
         ldlm_resource_putref(res);
         EXIT;
 }
@@ -1744,17 +1950,26 @@ static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
         /* we use l_pending_chain here, because it's unused on clients. */
         LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
-        /* bug 9573: don't replay locks left after eviction */
-        if (!(lock->l_flags & LDLM_FL_FAILED))
+        /* bug 9573: don't replay locks left after eviction, or
+         * bug 17614: locks being actively cancelled. Get a reference
+         * on a lock so that it does not disapear under us (e.g. due to cancel)
+         */
+        if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
                 list_add(&lock->l_pending_chain, list);
+                LDLM_LOCK_GET(lock);
+        }
+
         return LDLM_ITER_CONTINUE;
 }
 
-static int replay_lock_interpret(struct ptlrpc_request *req,
+static int replay_lock_interpret(const struct lu_env *env,
+                                 struct ptlrpc_request *req,
                                  struct ldlm_async_args *aa, int rc)
 {
-        struct ldlm_lock *lock;
-        struct ldlm_reply *reply;
+        struct lustre_handle  old_hash_key;
+        struct ldlm_lock     *lock;
+        struct ldlm_reply    *reply;
+        struct obd_export    *exp;
 
         ENTRY;
         atomic_dec(&req->rq_import->imp_replay_inflight);
@@ -1762,12 +1977,9 @@ static int replay_lock_interpret(struct ptlrpc_request *req,
                 GOTO(out, rc);
 
 
-        reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
-                                   lustre_swab_ldlm_reply);
-        if (reply == NULL) {
-                CERROR("Can't unpack ldlm_reply\n");
-                GOTO (out, rc = -EPROTO);
-        }
+        reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
+        if (reply == NULL)
+                GOTO(out, rc = -EPROTO);
 
         lock = ldlm_handle2lock(&aa->lock_handle);
         if (!lock) {
@@ -1779,7 +1991,16 @@ static int replay_lock_interpret(struct ptlrpc_request *req,
                 GOTO(out, rc = -ESTALE);
         }
 
+        old_hash_key = lock->l_remote_handle;
         lock->l_remote_handle = reply->lock_handle;
+
+        /* Key change rehash lock in per-export hash with new key */
+        exp = req->rq_export;
+        if (exp && exp->exp_lock_hash)
+                lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
+                                       &lock->l_remote_handle,
+                                       &lock->l_exp_hash);
+
         LDLM_DEBUG(lock, "replayed lock:");
         ptlrpc_import_recovery_state_machine(req->rq_import);
         LDLM_LOCK_PUT(lock);
@@ -1787,18 +2008,14 @@ out:
         if (rc != ELDLM_OK)
                 ptlrpc_connect_import(req->rq_import, NULL);
 
-
         RETURN(rc);
 }
 
 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
 {
         struct ptlrpc_request *req;
-        struct ldlm_request *body;
-        struct ldlm_reply *reply;
         struct ldlm_async_args *aa;
-        int buffers = 2;
-        int size[3] = { sizeof(struct ptlrpc_body) };
+        struct ldlm_request   *body;
         int flags;
         ENTRY;
 
@@ -1840,26 +2057,25 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
         else
                 flags = LDLM_FL_REPLAY;
 
-        size[DLM_LOCKREQ_OFF] = sizeof(*body);
-        req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, 2, size,
-                              NULL);
-        if (!req)
+        req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
+                                        LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
+        if (req == NULL)
                 RETURN(-ENOMEM);
 
         /* We're part of recovery, so don't wait for it. */
         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
 
-        body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
+        body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
         ldlm_lock2desc(lock, &body->lock_desc);
         body->lock_flags = flags;
 
         ldlm_lock2handle(lock, &body->lock_handle[0]);
-        size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
         if (lock->l_lvb_len != 0) {
-                buffers = 3;
-                size[DLM_REPLY_REC_OFF] = lock->l_lvb_len;
+                req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
+                req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
+                                     lock->l_lvb_len);
         }
-        ptlrpc_req_set_repsize(req, buffers, size);
+        ptlrpc_request_set_replen(req);
         /* notify the server we've replayed all requests.
          * also, we mark the request to be put on a dedicated
          * queue to be processed after all request replayes.
@@ -1870,10 +2086,10 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
 
         atomic_inc(&req->rq_import->imp_replay_inflight);
         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
-        aa = (struct ldlm_async_args *)&req->rq_async_args;
+        aa = ptlrpc_req_async_args(req);
         aa->lock_handle = body->lock_handle[0];
-        req->rq_interpret_reply = replay_lock_interpret;
-        ptlrpcd_add_req(req);
+        req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
+        ptlrpcd_add_req(req, PSCOPE_OTHER);
 
         RETURN(0);
 }
@@ -1881,12 +2097,11 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
 int ldlm_replay_locks(struct obd_import *imp)
 {
         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
-        struct list_head list;
+        CFS_LIST_HEAD(list);
         struct ldlm_lock *lock, *next;
         int rc = 0;
 
         ENTRY;
-        CFS_INIT_LIST_HEAD(&list);
 
         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 
@@ -1897,9 +2112,12 @@ int ldlm_replay_locks(struct obd_import *imp)
 
         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
                 list_del_init(&lock->l_pending_chain);
-                if (rc)
+                if (rc) {
+                        LDLM_LOCK_PUT(lock);
                         continue; /* or try to do the rest? */
+                }
                 rc = replay_one_lock(imp, lock);
+                LDLM_LOCK_PUT(lock);
         }
 
         atomic_dec(&imp->imp_replay_inflight);