Whamcloud - gitweb
b=13872
[fs/lustre-release.git] / lustre / ldlm / ldlm_request.c
index bf7d244..8aff2f6 100644 (file)
@@ -43,6 +43,10 @@ struct lock_wait_data {
         __u32             lwd_conn_cnt;
 };
 
+struct ldlm_async_args {
+        struct lustre_handle lock_handle;
+};
+
 int ldlm_expired_completion_wait(void *data)
 {
         struct lock_wait_data *lwd = data;
@@ -241,7 +245,7 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
         ENTRY;
 
         LASSERT(!(*flags & LDLM_FL_REPLAY));
-        if (unlikely(ns->ns_client)) {
+        if (unlikely(ns_is_client(ns))) {
                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
                 LBUG();
         }
@@ -319,7 +323,11 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 
         lock = ldlm_handle2lock(lockh);
         /* ldlm_cli_enqueue is holding a reference on this lock. */
-        LASSERT(lock != NULL);
+        if (!lock) {
+                LASSERT(type == LDLM_FLOCK);
+                RETURN(-ENOLCK);
+        }
+
         if (rc != ELDLM_OK) {
                 LASSERT(!is_replay);
                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
@@ -465,21 +473,31 @@ cleanup:
 /* 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. */
-#define ldlm_req_handles_avail(exp, size, bufcount, off)                \
-({                                                                      \
-        int _avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512);      \
-        int _s = size[DLM_LOCKREQ_OFF];                                 \
-        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 += LDLM_LOCKREQ_HANDLES - off;                           \
-        size[DLM_LOCKREQ_OFF] = _s;                                     \
-        _avail;                                                         \
-})
+static inline int ldlm_req_handles_avail(struct obd_export *exp,
+                                         int *size, int bufcount, int off)
+{
+        int avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512);
+        int old_size = size[DLM_LOCKREQ_OFF];
+
+        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 += LDLM_LOCKREQ_HANDLES - off;
+        size[DLM_LOCKREQ_OFF] = old_size;
+
+        return avail;
+}
+
+static inline int ldlm_cancel_handles_avail(struct obd_export *exp)
+{
+        int size[2] = { sizeof(struct ptlrpc_body),
+                        sizeof(struct ldlm_request) };
+        return ldlm_req_handles_avail(exp, size, 2, 0);
+}
 
 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
- * @count locks in @cancel. */
+ * @count locks in @cancels. */
 struct ptlrpc_request *ldlm_prep_enqueue_req(struct obd_export *exp,
                                              int bufcount, int *size,
                                              struct list_head *cancels,
@@ -503,8 +521,9 @@ struct ptlrpc_request *ldlm_prep_enqueue_req(struct obd_export *exp,
                  * 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, 1, avail - count,
-                                               LDLM_CANCEL_AGED);
+                count += ldlm_cancel_lru_local(ns, cancels,
+                                               exp_connect_lru_resize(exp) ? 0 : 1,
+                                               avail - count, LDLM_CANCEL_AGED);
                 size[DLM_LOCKREQ_OFF] =
                         ldlm_request_bufsize(count, LDLM_ENQUEUE);
         }
@@ -531,13 +550,10 @@ struct ptlrpc_request *ldlm_prep_enqueue_req(struct obd_export *exp,
  * request was created in ldlm_cli_enqueue and it is the async request,
  * pass it to the caller in @reqp. */
 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
+                     struct ldlm_enqueue_info *einfo,
                      const struct ldlm_res_id *res_id,
-                     ldlm_type_t type, ldlm_policy_data_t *policy,
-                     ldlm_mode_t mode, int *flags,
-                     ldlm_blocking_callback blocking,
-                     ldlm_completion_callback completion,
-                     ldlm_glimpse_callback glimpse,
-                     void *data, void *lvb, __u32 lvb_len, void *lvb_swabber,
+                     ldlm_policy_data_t *policy, int *flags,
+                     void *lvb, __u32 lvb_len, void *lvb_swabber,
                      struct lustre_handle *lockh, int async)
 {
         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
@@ -548,7 +564,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
                         [DLM_LOCKREQ_OFF]     = sizeof(*body),
                         [DLM_REPLY_REC_OFF]   = lvb_len };
         int is_replay = *flags & LDLM_FL_REPLAY;
-        int req_passed_in = 1, rc;
+        int req_passed_in = 1, rc, err;
         struct ptlrpc_request *req;
         ENTRY;
 
@@ -562,12 +578,14 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
                 LDLM_DEBUG(lock, "client-side enqueue START");
                 LASSERT(exp == lock->l_conn_export);
         } else {
-                lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
-                                        completion, glimpse, data, lvb_len);
+                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);
                 if (lock == NULL)
                         RETURN(-ENOMEM);
                 /* for the local lock, add the reference */
-                ldlm_lock_addref_internal(lock, mode);
+                ldlm_lock_addref_internal(lock, einfo->ei_mode);
                 ldlm_lock2handle(lock, lockh);
                 lock->l_lvb_swabber = lvb_swabber;
                 if (policy != NULL) {
@@ -576,8 +594,8 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
                          * descriptor (ldlm_lock2desc() below) but use an
                          * inodebits lock internally with both bits set.
                          */
-                        if (type == LDLM_IBITS && !(exp->exp_connect_flags &
-                                                    OBD_CONNECT_IBITS))
+                        if (einfo->ei_type == LDLM_IBITS &&
+                            !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
                                 lock->l_policy_data.l_inodebits.bits =
                                         MDS_INODELOCK_LOOKUP |
                                         MDS_INODELOCK_UPDATE;
@@ -585,7 +603,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
                                 lock->l_policy_data = *policy;
                 }
 
-                if (type == LDLM_EXTENT)
+                if (einfo->ei_type == LDLM_EXTENT)
                         lock->l_req_extent = policy->l_extent;
                 LDLM_DEBUG(lock, "client-side enqueue START");
         }
@@ -595,7 +613,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
         if (reqp == NULL || *reqp == NULL) {
                 req = ldlm_prep_enqueue_req(exp, 2, size, NULL, 0);
                 if (req == NULL) {
-                        failed_lock_cleanup(ns, lock, lockh, mode);
+                        failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
                         LDLM_LOCK_PUT(lock);
                         RETURN(-ENOMEM);
                 }
@@ -613,7 +631,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 
         lock->l_conn_export = exp;
         lock->l_export = NULL;
-        lock->l_blocking_ast = blocking;
+        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));
@@ -632,7 +650,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
          * [i_size, OBD_OBJECT_EOF] lock is taken.
          */
-        LASSERT(ergo(LIBLUSTRE_CLIENT, type != LDLM_EXTENT ||
+        LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
                      policy->l_extent.end == OBD_OBJECT_EOF));
 
         if (async) {
@@ -642,9 +660,16 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 
         LDLM_DEBUG(lock, "sending request");
         rc = ptlrpc_queue_wait(req);
-        rc = ldlm_cli_enqueue_fini(exp, req, type, policy ? 1 : 0,
-                                   mode, flags, lvb, lvb_len, lvb_swabber,
-                                   lockh, rc);
+        err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
+                                    einfo->ei_mode, flags, lvb, lvb_len,
+                                    lvb_swabber, lockh, rc);
+
+        /* 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);
+        else
+                rc = err;
 
         if (!req_passed_in && req != NULL) {
                 ptlrpc_req_finished(req);
@@ -661,7 +686,7 @@ static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
         struct ldlm_resource *res;
         int rc;
         ENTRY;
-        if (lock->l_resource->lr_namespace->ns_client) {
+        if (ns_is_client(lock->l_resource->lr_namespace)) {
                 CERROR("Trying to cancel local lock\n");
                 LBUG();
         }
@@ -757,10 +782,13 @@ int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
 }
 
 /* Cancel locks locally.
- * Returns: 1 if there is a need to send a cancel RPC to server. 0 otherwise. */
+ * Returns:
+ * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
+ * LDLM_FL_CANCELING otherwise;
+ * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
 {
-        int rc = 0;
+        int rc = LDLM_FL_LOCAL_ONLY;
         ENTRY;
         
         if (lock->l_conn_export) {
@@ -773,17 +801,18 @@ static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
                 local_only = (lock->l_flags &
                               (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
                 ldlm_cancel_callback(lock);
+                rc = (lock->l_flags & LDLM_FL_BL_AST) ?
+                        LDLM_FL_BL_AST : LDLM_FL_CANCELING;
                 unlock_res_and_lock(lock);
 
-                if (local_only)
-                        CDEBUG(D_INFO, "not sending request (at caller's "
+                if (local_only) {
+                        CDEBUG(D_DLMTRACE, "not sending request (at caller's "
                                "instruction)\n");
-                else
-                        rc = 1;
-
+                        rc = LDLM_FL_LOCAL_ONLY;
+                }
                 ldlm_lock_cancel(lock);
         } else {
-                if (lock->l_resource->lr_namespace->ns_client) {
+                if (ns_is_client(lock->l_resource->lr_namespace)) {
                         LDLM_ERROR(lock, "Trying to cancel local lock");
                         LBUG();
                 }
@@ -803,7 +832,7 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req, int off,
 {
         struct ldlm_request *dlm;
         struct ldlm_lock *lock;
-        int max;
+        int max, packed = 0;
         ENTRY;
 
         dlm = lustre_msg_buf(req->rq_reqmsg, off, sizeof(*dlm));
@@ -822,22 +851,13 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req, int off,
         list_for_each_entry(lock, head, l_bl_ast) {
                 if (!count--)
                         break;
-                /* Pack the lock handle to the given request buffer. */
                 LASSERT(lock->l_conn_export);
-                /* Cannot be set on a lock in a resource granted list.*/
-                LASSERT(!(lock->l_flags &
-                          (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK)));
-                /* If @lock is marked CANCEL_ON_BLOCK, cancel
-                 * will not be sent in ldlm_cli_cancel(). It 
-                 * is used for liblustre clients, no cancel on 
-                 * block requests. However, even for liblustre 
-                 * clients, when the flag is set, batched cancel
-                 * should be sent (what if no block rpc has
-                 * come). To not send another separated rpc in
-                 * this case, the caller pass CANCEL_ON_BLOCK
-                 * flag to ldlm_cli_cancel_unused_resource(). */
+                /* Pack the lock handle to the given request buffer. */
+                LDLM_DEBUG(lock, "packing");
                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
+                packed++;
         }
+        CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
         EXIT;
 }
 
@@ -858,6 +878,9 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
         LASSERT(exp != NULL);
         LASSERT(count > 0);
 
+        if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
+                RETURN(count);
+
         free = ldlm_req_handles_avail(exp, size, 2, 0);
         if (count > free)
                 count = free;
@@ -866,9 +889,9 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
         while (1) {
                 imp = class_exp2cliimp(exp);
                 if (imp == NULL || imp->imp_invalid) {
-                        CDEBUG(D_HA, "skipping cancel on invalid import %p\n",
-                               imp);
-                        break;
+                        CDEBUG(D_DLMTRACE,
+                               "skipping cancel on invalid import %p\n", imp);
+                        RETURN(count);
                 }
 
                 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL, 2,
@@ -877,6 +900,7 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
                         GOTO(out, rc = -ENOMEM);
 
                 req->rq_no_resend = 1;
+                req->rq_no_delay = 1;
 
                 /* XXX FIXME bug 249 */
                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
@@ -899,7 +923,9 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
                                "out of sync -- not fatal\n",
                                libcfs_nid2str(req->rq_import->
                                               imp_connection->c_peer.nid));
-                } else if (rc == -ETIMEDOUT) {
+                        rc = 0;
+                } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
+                           req->rq_import_generation == imp->imp_generation) {
                         ptlrpc_req_finished(req);
                         continue;
                 } else if (rc != ELDLM_OK) {
@@ -917,28 +943,85 @@ out:
         return sent ? sent : rc;
 }
 
+static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
+{
+        LASSERT(imp != NULL);
+        return &imp->imp_obd->obd_namespace->ns_pool;
+}
+
+int ldlm_cli_update_pool(struct ptlrpc_request *req)
+{
+        struct ldlm_pool *pl;
+        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);
+#ifdef __KERNEL__
+        {
+                __u64 old_slv, fast_slv_change;
+
+                old_slv = ldlm_pool_get_slv(pl);
+                fast_slv_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE;
+                do_div(fast_slv_change, 100);
+#endif
+                pl->pl_update_time = cfs_time_current();
+                ldlm_pool_set_slv(pl, lustre_msg_get_slv(req->rq_repmsg));
+                ldlm_pool_set_limit(pl, lustre_msg_get_limit(req->rq_repmsg));
+#ifdef __KERNEL__
+                /* Wake up pools thread only if SLV has changed more than 
+                 * 5% 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 1s anyways. */
+                if (old_slv > ldlm_pool_get_slv(pl) && 
+                    old_slv - ldlm_pool_get_slv(pl) > fast_slv_change)
+                        ldlm_pools_wakeup();
+        }
+#endif
+        spin_unlock(&pl->pl_lock);
+
+        RETURN(0);
+}
+EXPORT_SYMBOL(ldlm_cli_update_pool);
+
 int ldlm_cli_cancel(struct lustre_handle *lockh)
 {
         struct ldlm_lock *lock;
-        CFS_LIST_HEAD(head);
+        CFS_LIST_HEAD(cancels);
         int rc = 0;
         ENTRY;
 
         /* concurrent cancels on the same handle can happen */
         lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
-        if (lock == NULL)
+        if (lock == NULL) {
+                LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
                 RETURN(0);
-
+        }
+        
         rc = ldlm_cli_cancel_local(lock);
-        if (rc <= 0)
-                GOTO(out, rc);
-
-        list_add(&lock->l_bl_ast, &head);
-        rc = ldlm_cli_cancel_req(lock->l_conn_export, &head, 1, 0);
-        EXIT;
-out:
-        LDLM_LOCK_PUT(lock);
-        return rc < 0 ? rc : 0;
+        list_add(&lock->l_bl_ast, &cancels);
+
+        if (rc == LDLM_FL_BL_AST) {
+                rc = ldlm_cli_cancel_req(lock->l_conn_export, &cancels, 1, 0);
+        } else if (rc == LDLM_FL_CANCELING) {
+                int avail = ldlm_cancel_handles_avail(lock->l_conn_export);
+                int count = 1;
+                LASSERT(avail > 0);
+                count += ldlm_cancel_lru_local(lock->l_resource->lr_namespace,
+                                               &cancels, 0, avail - 1,
+                                               LDLM_CANCEL_AGED);
+                ldlm_cli_cancel_list(&cancels, count, NULL, 0, 0);
+        }
+        if (rc != LDLM_FL_CANCELING)
+                LDLM_LOCK_PUT(lock);
+        RETURN(rc < 0 ? rc : 0);
 }
 
 /* - Free space in lru for @count new locks,
@@ -957,33 +1040,80 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
                           int count, int max, int flags)
 {
         cfs_time_t cur = cfs_time_current();
+        int rc, added = 0, left, unused;
         struct ldlm_lock *lock, *next;
-        int rc, added = 0, left;
+        __u64 slv, lvf, lv;
         ENTRY;
 
         spin_lock(&ns->ns_unused_lock);
-        count += ns->ns_nr_unused - ns->ns_max_unused;
+        unused = ns->ns_nr_unused;
+        
+        if (!ns_connect_lru_resize(ns))
+                count += unused - ns->ns_max_unused;
+
         while (!list_empty(&ns->ns_unused_list)) {
-                struct list_head *tmp = ns->ns_unused_list.next;
-                lock = list_entry(tmp, struct ldlm_lock, l_lru);
+                struct ldlm_pool *pl = &ns->ns_pool;
+
+                LASSERT(unused >= 0);
 
                 if (max && added >= max)
                         break;
 
-                if ((added >= count) && 
-                    (!(flags & LDLM_CANCEL_AGED) ||
-                     cfs_time_before_64(cur, ns->ns_max_age +
-                                        lock->l_last_used)))
+                list_for_each_entry(lock, &ns->ns_unused_list, l_lru) {
+                        /* somebody is already doing CANCEL or there is a
+                         * blocking request will send cancel. */
+                        if (!(lock->l_flags & LDLM_FL_CANCELING) &&
+                            !(lock->l_flags & LDLM_FL_BL_AST))
+                                break;
+                }
+                if (&lock->l_lru == &ns->ns_unused_list)
                         break;
 
+                if (ns_connect_lru_resize(ns)) {
+                        cfs_time_t la;
+                        
+                        /* Take into account SLV only if cpount == 0. */
+                        if (count == 0) {
+                                /* Calculate lv for every lock. */
+                                spin_lock(&pl->pl_lock);
+                                slv = ldlm_pool_get_slv(pl);
+                                lvf = atomic_read(&pl->pl_lock_volume_factor);
+                                spin_unlock(&pl->pl_lock);
+
+                                la = cfs_duration_sec(cfs_time_sub(cur, 
+                                                      lock->l_last_used));
+                                if (la == 0)
+                                        la = 1;
+                                
+                                /* Stop when slv is not yet come from server 
+                                 * or lv is smaller than it is. */
+                                lv = lvf * la * unused;
+                                if (slv == 1 || lv < slv)
+                                        break;
+                        } else {
+                                if (added >= count)
+                                        break;
+                        }
+                } else {
+                        if ((added >= count) && 
+                            (!(flags & LDLM_CANCEL_AGED) ||
+                             cfs_time_before_64(cur, ns->ns_max_age +
+                                                lock->l_last_used)))
+                                break;
+                }
+
                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
                 spin_unlock(&ns->ns_unused_lock);
 
                 lock_res_and_lock(lock);
-                if ((ldlm_lock_remove_from_lru(lock) == 0) ||
-                    (lock->l_flags & LDLM_FL_CANCELING)) {
+                /* 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. */
+                         * 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);
@@ -1015,6 +1145,7 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
                 unlock_res_and_lock(lock);
                 spin_lock(&ns->ns_unused_lock);
                 added++;
+                unused--;
         }
         spin_unlock(&ns->ns_unused_lock);
 
@@ -1023,13 +1154,24 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
                 if (left-- == 0)
                         break;
+
                 rc = ldlm_cli_cancel_local(lock);
-                if (rc == 0) {
+                if (rc == LDLM_FL_BL_AST) {
+                        CFS_LIST_HEAD(head);
+
+                        LDLM_DEBUG(lock, "Cancel lock separately");
+                        list_del_init(&lock->l_bl_ast);
+                        list_add(&lock->l_bl_ast, &head);
+                        ldlm_cli_cancel_req(lock->l_conn_export, &head, 1, 0);
+                        rc = LDLM_FL_LOCAL_ONLY;
+                }
+                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);
                         added--;
                 }
+
         } 
         RETURN(added);
 }
@@ -1038,7 +1180,7 @@ 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, ldlm_sync_t sync)
+int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync)
 {
         CFS_LIST_HEAD(cancels);
         int count, rc;
@@ -1047,24 +1189,17 @@ int ldlm_cancel_lru(struct ldlm_namespace *ns, ldlm_sync_t sync)
 #ifndef __KERNEL__
         sync = LDLM_SYNC; /* force to be sync in user space */
 #endif
-        count = ldlm_cancel_lru_local(ns, &cancels, 0, 0, 0);
+        count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0);
         if (sync == LDLM_ASYNC) {
-                struct ldlm_lock *lock, *next;
-                list_for_each_entry_safe(lock, next, &cancels, l_bl_ast) {
-                        /* Remove from the list to allow blocking thread to
-                         * re-use l_bl_ast. */
-                        list_del_init(&lock->l_bl_ast);
-                        rc = ldlm_bl_to_thread(ns, NULL, lock,
-                                               LDLM_FL_CANCELING);
-                        if (rc)
-                                list_add_tail(&lock->l_bl_ast, &next->l_bl_ast);
-                }
+                rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
+                if (rc == 0)
+                        RETURN(count);
         }
 
-        /* If some locks are left in the list in ASYNC mode, or
+        /* If an error occured in ASYNC mode, or
          * this is SYNC mode, cancel the list. */
-        ldlm_cli_cancel_list(&cancels, count, NULL, DLM_LOCKREQ_OFF, 0);
-        RETURN(0);
+        ldlm_cli_cancel_list(&cancels, count, NULL, 0, 0);
+        RETURN(count);
 }
 
 /* Find and cancel locally unused locks found on resource, matched to the
@@ -1097,6 +1232,12 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res,
                         continue;
                 }
 
+                /* If somebody is already doing CANCEL, or blocking ast came,
+                 * skip this lock. */
+                if (lock->l_flags & LDLM_FL_BL_AST || 
+                    lock->l_flags & LDLM_FL_CANCELING)
+                        continue;
+
                 if (lockmode_compat(lock->l_granted_mode, mode))
                         continue;
 
@@ -1107,13 +1248,9 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res,
                       policy->l_inodebits.bits))
                         continue;
 
-                /* If somebody is already doing CANCEL, skip it. */
-                if (lock->l_flags & LDLM_FL_CANCELING)
-                        continue;
-                
                 /* See CBPENDING comment in ldlm_cancel_lru */
                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
-                        lock_flags;
+                                 lock_flags;
 
                 LASSERT(list_empty(&lock->l_bl_ast));
                 list_add(&lock->l_bl_ast, cancels);
@@ -1125,7 +1262,7 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res,
         /* Handle only @count inserted locks. */
         left = count;
         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
-                int rc = 0;
+                int rc = LDLM_FL_LOCAL_ONLY;
 
                 if (left-- == 0)
                         break;
@@ -1134,7 +1271,16 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res,
                 else
                         rc = ldlm_cli_cancel_local(lock);
 
-                if (rc == 0) {
+                if (rc == LDLM_FL_BL_AST) {
+                        CFS_LIST_HEAD(head);
+
+                        LDLM_DEBUG(lock, "Cancel lock separately");
+                        list_del_init(&lock->l_bl_ast);
+                        list_add(&lock->l_bl_ast, &head);
+                        ldlm_cli_cancel_req(lock->l_conn_export, &head, 1, 0);
+                        rc = LDLM_FL_LOCAL_ONLY;
+                }
+                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);
@@ -1176,7 +1322,8 @@ int ldlm_cli_cancel_list(struct list_head *cancels, int count,
                                 ldlm_cancel_pack(req, off, cancels, count);
                         else
                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
-                                                          cancels, count, flags);
+                                                          cancels, count,
+                                                          flags);
                 } else {
                         res = ldlm_cli_cancel_req(lock->l_conn_export,
                                                   cancels, 1, flags);
@@ -1198,7 +1345,7 @@ int ldlm_cli_cancel_list(struct list_head *cancels, int count,
 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
                                     const struct ldlm_res_id *res_id,
                                     ldlm_policy_data_t *policy,
-                                    int mode, int flags, void *opaque)
+                                    ldlm_mode_t mode, int flags, void *opaque)
 {
         struct ldlm_resource *res;
         CFS_LIST_HEAD(cancels);
@@ -1215,8 +1362,7 @@ int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
 
         count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
                                            0, flags, opaque);
-        rc = ldlm_cli_cancel_list(&cancels, count, NULL,
-                                  DLM_LOCKREQ_OFF, flags);
+        rc = ldlm_cli_cancel_list(&cancels, count, NULL, 0, flags);
         if (rc != ELDLM_OK)
                 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
 
@@ -1295,7 +1441,7 @@ int ldlm_cli_join_lru(struct ldlm_namespace *ns,
         int count = 0;
         ENTRY;
 
-        LASSERT(ns->ns_client == LDLM_NAMESPACE_CLIENT);
+        LASSERT(ns_is_client(ns));
 
         res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
         if (res == NULL)
@@ -1311,12 +1457,7 @@ int ldlm_cli_join_lru(struct ldlm_namespace *ns,
                     !lock->l_readers && !lock->l_writers &&
                     !(lock->l_flags & LDLM_FL_LOCAL) &&
                     !(lock->l_flags & LDLM_FL_CBPENDING)) {
-                        lock->l_last_used = cfs_time_current();
-                        spin_lock(&ns->ns_unused_lock);
-                        LASSERT(ns->ns_nr_unused >= 0);
-                        list_add_tail(&lock->l_lru, &ns->ns_unused_list);
-                        ns->ns_nr_unused++;
-                        spin_unlock(&ns->ns_unused_lock);
+                        ldlm_lock_add_to_lru(lock);
                         lock->l_flags &= ~LDLM_FL_NO_LRU;
                         LDLM_DEBUG(lock, "join lock to lru");
                         count++;
@@ -1473,7 +1614,7 @@ static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
 }
 
 static int replay_lock_interpret(struct ptlrpc_request *req,
-                                    void * data, int rc)
+                                 struct ldlm_async_args *aa, int rc)
 {
         struct ldlm_lock *lock;
         struct ldlm_reply *reply;
@@ -1483,8 +1624,6 @@ static int replay_lock_interpret(struct ptlrpc_request *req,
         if (rc != ELDLM_OK)
                 GOTO(out, rc);
 
-        lock = req->rq_async_args.pointer_arg[0];
-        LASSERT(lock != NULL);
 
         reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
                                    lustre_swab_ldlm_reply);
@@ -1493,10 +1632,21 @@ static int replay_lock_interpret(struct ptlrpc_request *req,
                 GOTO (out, rc = -EPROTO);
         }
 
+        lock = ldlm_handle2lock(&aa->lock_handle);
+        if (!lock) {
+                CERROR("received replay ack for unknown local cookie "LPX64
+                       " remote cookie "LPX64 " from server %s id %s\n",
+                       aa->lock_handle.cookie, reply->lock_handle.cookie,
+                       req->rq_export->exp_client_uuid.uuid,
+                       libcfs_id2str(req->rq_peer));
+                GOTO(out, rc = -ESTALE);
+        }
+
         lock->l_remote_handle = reply->lock_handle;
         LDLM_DEBUG(lock, "replayed lock:");
         ptlrpc_import_recovery_state_machine(req->rq_import);
- out:
+        LDLM_LOCK_PUT(lock);
+out:
         if (rc != ELDLM_OK)
                 ptlrpc_connect_import(req->rq_import, NULL);
 
@@ -1509,11 +1659,19 @@ 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) };
         int flags;
         ENTRY;
 
+
+        /* Bug 11974: Do not replay a lock which is actively being canceled */
+        if (lock->l_flags & LDLM_FL_CANCELING) {
+                LDLM_DEBUG(lock, "Not replaying canceled lock:");
+                RETURN(0);
+        }
+
         /* If this is reply-less callback lock, we cannot replay it, since
          * server might have long dropped it, but notification of that event was
          * lost by network. (and server granted conflicting lock already) */
@@ -1574,7 +1732,9 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
         LDLM_DEBUG(lock, "replaying lock:");
 
         atomic_inc(&req->rq_import->imp_replay_inflight);
-        req->rq_async_args.pointer_arg[0] = lock;
+        CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
+        aa = (struct ldlm_async_args *)&req->rq_async_args;
+        aa->lock_handle = body->lock_handle[0];
         req->rq_interpret_reply = replay_lock_interpret;
         ptlrpcd_add_req(req);