Whamcloud - gitweb
LU-2901 ldlm: fix resource/fid check, use DLDLMRES
[fs/lustre-release.git] / lustre / ldlm / ldlm_request.c
index 64dc539..95b2f63 100644 (file)
  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2010, 2011, Whamcloud, Inc.
+ * Copyright (c) 2010, 2013, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
  * Lustre is a trademark of Sun Microsystems, Inc.
  */
+/**
+ * This file contains Asynchronous System Trap (AST) handlers and related
+ * LDLM request-processing routines.
+ *
+ * An AST is a callback issued on a lock when its state is changed. There are
+ * several different types of ASTs (callbacks) registered for each lock:
+ *
+ * - completion AST: when a lock is enqueued by some process, but cannot be
+ *   granted immediately due to other conflicting locks on the same resource,
+ *   the completion AST is sent to notify the caller when the lock is
+ *   eventually granted
+ *
+ * - blocking AST: when a lock is granted to some process, if another process
+ *   enqueues a conflicting (blocking) lock on a resource, a blocking AST is
+ *   sent to notify the holder(s) of the lock(s) of the conflicting lock
+ *   request. The lock holder(s) must release their lock(s) on that resource in
+ *   a timely manner or be evicted by the server.
+ *
+ * - glimpse AST: this is used when a process wants information about a lock
+ *   (i.e. the lock value block (LVB)) but does not necessarily require holding
+ *   the lock. If the resource is locked, the lock holder(s) are sent glimpse
+ *   ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL
+ *   their lock(s) if they are idle. If the resource is not locked, the server
+ *   may grant the lock.
+ */
 
 #define DEBUG_SUBSYSTEM S_LDLM
 #ifndef __KERNEL__
@@ -80,12 +105,17 @@ int ldlm_expired_completion_wait(void *data)
                 if (ptlrpc_check_suspend())
                         RETURN(0);
 
-                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));
+               LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", "
+                             CFS_DURATION_T"s ago)\n",
+                             lock->l_last_activity,
+                             cfs_time_sub(cfs_time_current_sec(),
+                                          lock->l_last_activity));
+               LDLM_DEBUG(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);
@@ -131,24 +161,24 @@ EXPORT_SYMBOL(ldlm_get_enq_timeout);
  */
 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_measured(ldlm_lock_to_ns_at(lock),
-                            delay);
-                result = 0;
-        }
-        return result;
+       long delay;
+       int  result;
+
+       if (lock->l_flags & (LDLM_FL_DESTROYED | 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_measured(ldlm_lock_to_ns_at(lock),
+                           delay);
+               result = 0;
+       }
+       return result;
 }
 
 /**
@@ -179,13 +209,13 @@ int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
 EXPORT_SYMBOL(ldlm_completion_ast_async);
 
 /**
- * Client side LDLM "completion" AST. This is called in several cases:
+ * Generic LDLM "completion" AST. This is called in several cases:
  *
- *     - when a reply to an ENQUEUE rpc is received from the server
+ *     - 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
+ *     - 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
@@ -249,11 +279,11 @@ noreproc:
                                        interrupted_completion_wait, &lwd);
         }
 
-        if (imp != NULL) {
-                cfs_spin_lock(&imp->imp_lock);
-                lwd.lwd_conn_cnt = imp->imp_conn_cnt;
-                cfs_spin_unlock(&imp->imp_lock);
-        }
+       if (imp != NULL) {
+               spin_lock(&imp->imp_lock);
+               lwd.lwd_conn_cnt = imp->imp_conn_cnt;
+               spin_unlock(&imp->imp_lock);
+       }
 
         if (ns_is_client(ldlm_lock_to_ns(lock)) &&
             OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
@@ -277,12 +307,12 @@ noreproc:
 EXPORT_SYMBOL(ldlm_completion_ast);
 
 /**
- * A helper to build a blocking ast function
+ * A helper to build a blocking AST function
  *
- * Perform a common operation for blocking asts:
+ * Perform a common operation for blocking ASTs:
  * defferred lock cancellation.
  *
- * \param lock the lock blocking or canceling ast was called on
+ * \param lock the lock blocking or canceling AST was called on
  * \retval 0
  * \see mdt_blocking_ast
  * \see ldlm_blocking_ast
@@ -302,7 +332,7 @@ int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
 
                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
                 ldlm_lock2handle(lock, &lockh);
-                rc = ldlm_cli_cancel(&lockh);
+               rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
                 if (rc < 0)
                         CERROR("ldlm_cli_cancel: %d\n", rc);
         } else {
@@ -350,7 +380,7 @@ int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 }
 EXPORT_SYMBOL(ldlm_blocking_ast);
 
-/*
+/**
  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
  * comment in filter_intent_policy() on why you may need this.
  */
@@ -378,6 +408,9 @@ int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
 }
 EXPORT_SYMBOL(ldlm_glimpse_ast);
 
+/**
+ * Enqueue a local lock (typically on a server).
+ */
 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
                            const struct ldlm_res_id *res_id,
                            ldlm_type_t type, ldlm_policy_data_t *policy,
@@ -385,7 +418,7 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
                            ldlm_blocking_callback blocking,
                            ldlm_completion_callback completion,
                            ldlm_glimpse_callback glimpse,
-                           void *data, __u32 lvb_len,
+                          void *data, __u32 lvb_len, enum lvb_type lvb_type,
                            const __u64 *client_cookie,
                            struct lustre_handle *lockh)
 {
@@ -403,7 +436,8 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
                 LBUG();
         }
 
-        lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len);
+       lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
+                               lvb_type);
         if (unlikely(!lock))
                 GOTO(out_nolock, err = -ENOMEM);
 
@@ -484,6 +518,11 @@ static void failed_lock_cleanup(struct ldlm_namespace *ns,
         }
 }
 
+/**
+ * Finishing portion of client lock enqueue code.
+ *
+ * Called after receiving reply from server.
+ */
 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
                          __u64 *flags, void *lvb, __u32 lvb_len,
@@ -493,8 +532,8 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
         int is_replay = *flags & LDLM_FL_REPLAY;
         struct ldlm_lock *lock;
         struct ldlm_reply *reply;
-        struct ost_lvb *tmplvb;
         int cleanup_phase = 1;
+       int size = 0;
         ENTRY;
 
         lock = ldlm_handle2lock(lockh);
@@ -504,35 +543,45 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
                 RETURN(-ENOLCK);
         }
 
+       LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
+                "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
+
         if (rc != ELDLM_OK) {
                 LASSERT(!is_replay);
                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
-                if (rc == ELDLM_LOCK_ABORTED) {
-                        /* Before we return, swab the reply */
-                        reply = req_capsule_server_get(&req->rq_pill,
-                                                       &RMF_DLM_REP);
-                        if (reply == NULL)
-                                rc = -EPROTO;
-                        if (lvb_len) {
-
-                                req_capsule_set_size(&req->rq_pill,
-                                                     &RMF_DLM_LVB, RCL_SERVER,
-                                                     lvb_len);
-                                tmplvb = req_capsule_server_get(&req->rq_pill,
-                                                                 &RMF_DLM_LVB);
-                                if (tmplvb == NULL)
-                                        GOTO(cleanup, rc = -EPROTO);
-                                if (lvb != NULL)
-                                        memcpy(lvb, tmplvb, lvb_len);
-                        }
-                }
-                GOTO(cleanup, rc);
-        }
 
-        reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
-        if (reply == NULL)
-                GOTO(cleanup, rc = -EPROTO);
+               if (rc != ELDLM_LOCK_ABORTED)
+                       GOTO(cleanup, rc);
+       }
+
+       /* Before we return, swab the reply */
+       reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
+       if (reply == NULL)
+               GOTO(cleanup, rc = -EPROTO);
+
+       if (lvb_len != 0) {
+               LASSERT(lvb != NULL);
+
+               size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
+                                           RCL_SERVER);
+               if (size < 0) {
+                       LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
+                       GOTO(cleanup, rc = size);
+               } else if (unlikely(size > lvb_len)) {
+                       LDLM_ERROR(lock, "Replied LVB is larger than "
+                                  "expectation, expected = %d, replied = %d",
+                                  lvb_len, size);
+                       GOTO(cleanup, rc = -EINVAL);
+               }
+       }
+
+       if (rc == ELDLM_LOCK_ABORTED) {
+               if (lvb_len != 0)
+                       rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
+                                          lvb, size);
+               GOTO(cleanup, rc = (rc != 0 ? rc : ELDLM_LOCK_ABORTED));
+       }
 
         /* lock enqueued on the server */
         cleanup_phase = 0;
@@ -575,33 +624,27 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
                         lock->l_req_mode = newmode;
                 }
 
-                if (memcmp(reply->lock_desc.l_resource.lr_name.name,
-                          lock->l_resource->lr_name.name,
-                          sizeof(struct ldlm_res_id))) {
-                        CDEBUG(D_INFO, "remote intent success, locking "
-                                        "(%ld,%ld,%ld) instead of "
-                                        "(%ld,%ld,%ld)\n",
-                              (long)reply->lock_desc.l_resource.lr_name.name[0],
-                              (long)reply->lock_desc.l_resource.lr_name.name[1],
-                              (long)reply->lock_desc.l_resource.lr_name.name[2],
-                              (long)lock->l_resource->lr_name.name[0],
-                              (long)lock->l_resource->lr_name.name[1],
-                              (long)lock->l_resource->lr_name.name[2]);
-
-                        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)
-                        if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
-                                                    OBD_CONNECT_IBITS)))
-                                /* We assume lock type cannot change on server*/
-                                ldlm_convert_policy_to_local(exp,
-                                                lock->l_resource->lr_type,
-                                                &reply->lock_desc.l_policy_data,
-                                                &lock->l_policy_data);
+               if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
+                                &lock->l_resource->lr_name)) {
+                       CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
+                                      " instead of "DLDLMRES"\n",
+                              PLDLMRES(&reply->lock_desc.l_resource),
+                              PLDLMRES(lock->l_resource));
+
+                       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)
+                       if (!(type == LDLM_IBITS &&
+                             !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
+                               /* We assume lock type cannot change on server*/
+                               ldlm_convert_policy_to_local(exp,
+                                               lock->l_resource->lr_type,
+                                               &reply->lock_desc.l_policy_data,
+                                               &lock->l_policy_data);
                 if (type != LDLM_PLAIN)
                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
         }
@@ -619,25 +662,20 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 
         /* If the lock has already been granted by a completion AST, don't
          * clobber the LVB with an older one. */
-        if (lvb_len) {
-                /* We must lock or a racing completion might update lvb
-                   without letting us know and we'll clobber the correct value.
-                   Cannot unlock after the check either, a that still leaves
-                   a tiny window for completion to get in */
-                lock_res_and_lock(lock);
-                if (lock->l_req_mode != lock->l_granted_mode) {
-
-                        req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
-                                             RCL_SERVER, lvb_len);
-                        tmplvb = req_capsule_server_get(&req->rq_pill,
-                                                             &RMF_DLM_LVB);
-                        if (tmplvb == NULL) {
-                                unlock_res_and_lock(lock);
-                                GOTO(cleanup, rc = -EPROTO);
-                        }
-                        memcpy(lock->l_lvb_data, tmplvb, lvb_len);
-                }
-                unlock_res_and_lock(lock);
+       if (lvb_len != 0) {
+               /* We must lock or a racing completion might update lvb without
+                * letting us know and we'll clobber the correct value.
+                * Cannot unlock after the check either, a that still leaves
+                * a tiny window for completion to get in */
+               lock_res_and_lock(lock);
+               if (lock->l_req_mode != lock->l_granted_mode)
+                       rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
+                                          lock->l_lvb_data, size);
+               unlock_res_and_lock(lock);
+               if (rc < 0) {
+                       cleanup_phase = 1;
+                       GOTO(cleanup, rc);
+               }
         }
 
         if (!is_replay) {
@@ -669,14 +707,17 @@ cleanup:
 }
 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
 
-/* 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. */
+/**
+ * Estimate number of lock handles that would fit into request of given
+ * size.  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(int req_size, int off)
 {
         int avail;
 
-        avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
+       avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
         if (likely(avail >= 0))
                 avail /= (int)sizeof(struct lustre_handle);
         else
@@ -702,8 +743,14 @@ static inline int ldlm_format_handles_avail(struct obd_import *imp,
         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. */
+/**
+ * Cancel LRU locks and pack them into the enqueue request. Pack there the given
+ * \a count locks in \a cancels.
+ *
+ * This is to be called by functions preparing their own requests that
+ * might contain lists of locks to cancel in addition to actual operation
+ * that needs to be performed.
+ */
 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
                       int version, int opc, int canceloff,
                       cfs_list_t *cancels, int count)
@@ -718,7 +765,7 @@ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
 
         if (cancels == NULL)
                 cancels = &head;
-        if (exp_connect_cancelset(exp)) {
+       if (ns_connect_cancelset(ns)) {
                 /* Estimate the amount of available space in the request. */
                 req_capsule_filled_sizes(pill, RCL_CLIENT);
                 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
@@ -728,9 +775,9 @@ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
                 to_free = !ns_connect_lru_resize(ns) &&
                           opc == LDLM_ENQUEUE ? 1 : 0;
 
-                /* Cancel lru locks here _only_ if the server supports
-                 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
-                 * rpc, what will make us slower. */
+               /* Cancel LRU locks here _only_ if the server supports
+                * EARLY_CANCEL. Otherwise we have to send extra CANCEL
+                * RPC, which will make us slower. */
                 if (avail > count)
                         count += ldlm_cancel_lru_local(ns, cancels, to_free,
                                                        avail - count, 0, flags);
@@ -748,7 +795,7 @@ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
                 RETURN(rc);
         }
 
-        if (exp_connect_cancelset(exp)) {
+       if (ns_connect_cancelset(ns)) {
                 if (canceloff) {
                         dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
                         LASSERT(dlm);
@@ -760,7 +807,7 @@ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
                 }
                 /* Pack into the request @pack lock handles. */
                 ldlm_cli_cancel_list(cancels, pack, req, 0);
-                /* Prepare and send separate cancel rpc for others. */
+               /* 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);
@@ -777,20 +824,46 @@ int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
 }
 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
 
-/* If a request has some specific initialisation it is passed in @reqp,
+struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
+{
+       struct ptlrpc_request *req;
+       int rc;
+       ENTRY;
+
+       req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
+       if (req == NULL)
+               RETURN(ERR_PTR(-ENOMEM));
+
+       rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
+       if (rc) {
+               ptlrpc_request_free(req);
+               RETURN(ERR_PTR(rc));
+       }
+
+       req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
+       ptlrpc_request_set_replen(req);
+       RETURN(req);
+}
+EXPORT_SYMBOL(ldlm_enqueue_pack);
+
+/**
+ * Client-side lock enqueue.
+ *
+ * If a request has some specific initialisation it is passed in \a reqp,
  * otherwise it is created in ldlm_cli_enqueue.
  *
- * Supports sync and async requests, pass @async flag accordingly. If a
+ * Supports sync and async requests, pass \a async flag accordingly. If a
  * request was created in ldlm_cli_enqueue and it is the async request,
- * pass it to the caller in @reqp. */
+ * pass it to the caller in \a 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_policy_data_t const *policy, __u64 *flags,
-                     void *lvb, __u32 lvb_len, struct lustre_handle *lockh,
-                     int async)
+                    void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
+                    struct lustre_handle *lockh, int async)
 {
-        struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
+       struct ldlm_namespace *ns;
         struct ldlm_lock      *lock;
         struct ldlm_request   *body;
         int                    is_replay = *flags & LDLM_FL_REPLAY;
@@ -801,6 +874,8 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 
         LASSERT(exp != NULL);
 
+       ns = exp->exp_obd->obd_namespace;
+
         /* 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) {
@@ -809,17 +884,16 @@ 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 {
-                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, &cbs, einfo->ei_cbdata,
-                                        lvb_len);
-                if (lock == NULL)
-                        RETURN(-ENOMEM);
+               const struct ldlm_callback_suite cbs = {
+                       .lcs_completion = einfo->ei_cb_cp,
+                       .lcs_blocking   = einfo->ei_cb_bl,
+                       .lcs_glimpse    = einfo->ei_cb_gl
+               };
+               lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
+                                       einfo->ei_mode, &cbs, einfo->ei_cbdata,
+                                       lvb_len, lvb_type);
+               if (lock == NULL)
+                       RETURN(-ENOMEM);
                 /* for the local lock, add the reference */
                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
                 ldlm_lock2handle(lock, lockh);
@@ -830,7 +904,8 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
                          * inodebits lock internally with both bits set.
                          */
                         if (einfo->ei_type == LDLM_IBITS &&
-                            !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
+                           !(exp_connect_flags(exp) &
+                             OBD_CONNECT_IBITS))
                                 lock->l_policy_data.l_inodebits.bits =
                                         MDS_INODELOCK_LOOKUP |
                                         MDS_INODELOCK_UPDATE;
@@ -840,13 +915,14 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 
                 if (einfo->ei_type == LDLM_EXTENT)
                         lock->l_req_extent = policy->l_extent;
-                LDLM_DEBUG(lock, "client-side enqueue START");
+                LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n",
+                          *flags);
         }
 
        lock->l_conn_export = exp;
        lock->l_export = NULL;
        lock->l_blocking_ast = einfo->ei_cb_bl;
-       lock->l_flags |= (*flags & LDLM_FL_NO_LRU);
+       lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
 
         /* lock not sent to server yet */
 
@@ -881,13 +957,12 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 
         /* Continue as normal. */
         if (!req_passed_in) {
-                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);
+               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);
         }
 
         /*
@@ -945,7 +1020,7 @@ static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
                 ldlm_reprocess_all(res);
                 rc = 0;
         } else {
-                rc = EDEADLOCK;
+               rc = LUSTRE_EDEADLK;
         }
         LDLM_DEBUG(lock, "client-side local convert handler END");
         LDLM_LOCK_PUT(lock);
@@ -1017,7 +1092,7 @@ int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
                                 GOTO(out, rc);
                 }
         } else {
-                rc = EDEADLOCK;
+               rc = LUSTRE_EDEADLK;
         }
         EXIT;
  out:
@@ -1027,11 +1102,13 @@ int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
 }
 EXPORT_SYMBOL(ldlm_cli_convert);
 
-/* Cancel locks locally.
+/**
+ * Cancel locks locally.
  * 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. */
+ * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
+ * \retval LDLM_FL_CANCELING otherwise;
+ * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
+ */
 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
 {
        __u64 rc = LDLM_FL_LOCAL_ONLY;
@@ -1070,8 +1147,9 @@ static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
         RETURN(rc);
 }
 
-/* Pack @count locks in @head into ldlm_request buffer at the offset @off,
-   of the request @req. */
+/**
+ * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
+ */
 static void ldlm_cancel_pack(struct ptlrpc_request *req,
                              cfs_list_t *head, int count)
 {
@@ -1106,8 +1184,9 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req,
         EXIT;
 }
 
-/* Prepare and send a batched cancel rpc, it will include count lock handles
- * of locks given in @head. */
+/**
+ * Prepare and send a batched cancel RPC. It will include \a count lock
+ * handles of locks given in \a cancels list. */
 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
                         int count, ldlm_cancel_flags_t flags)
 {
@@ -1151,8 +1230,6 @@ int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
                         ptlrpc_request_free(req);
                         GOTO(out, rc);
                 }
-                req->rq_no_resend = 1;
-                req->rq_no_delay = 1;
 
                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
@@ -1168,7 +1245,7 @@ int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
                 } else {
                         rc = ptlrpc_queue_wait(req);
                 }
-                if (rc == ESTALE) {
+               if (rc == LUSTRE_ESTALE) {
                         CDEBUG(D_DLMTRACE, "client/server (nid %s) "
                                "out of sync -- not fatal\n",
                                libcfs_nid2str(req->rq_import->
@@ -1179,10 +1256,11 @@ int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
                         ptlrpc_req_finished(req);
                         continue;
                 } else if (rc != ELDLM_OK) {
-                       if (rc != -ESHUTDOWN)
-                               CERROR("Got rc %d from cancel RPC: canceling "
-                                      "anyway\n", rc);
-                        break;
+                       /* -ESHUTDOWN is common on umount */
+                       CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
+                                    "Got rc %d from cancel RPC: "
+                                    "canceling anyway\n", rc);
+                       break;
                 }
                 sent = count;
                 break;
@@ -1202,7 +1280,7 @@ static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
 }
 
 /**
- * Update client's obd pool related fields with new SLV and Limit from \a req.
+ * Update client's OBD pool related fields with new SLV and Limit from \a req.
  */
 int ldlm_cli_update_pool(struct ptlrpc_request *req)
 {
@@ -1219,13 +1297,11 @@ int ldlm_cli_update_pool(struct ptlrpc_request *req)
                 RETURN(0);
         }
 
-        /*
-         * 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.
-         */
+       /* 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 reference to the OBD export and thus access to
+        * server-side namespace is not 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 "
@@ -1239,23 +1315,27 @@ int ldlm_cli_update_pool(struct ptlrpc_request *req)
         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.
-         */
-        cfs_write_lock(&obd->obd_pool_lock);
+       /* Set new SLV and limit in OBD fields to make them accessible
+        * to the 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 at cleanup time. Evil races are possible
+        * which may cause Oops at that time. */
+       write_lock(&obd->obd_pool_lock);
         obd->obd_pool_slv = new_slv;
         obd->obd_pool_limit = new_limit;
-        cfs_write_unlock(&obd->obd_pool_lock);
+       write_unlock(&obd->obd_pool_lock);
 
         RETURN(0);
 }
 EXPORT_SYMBOL(ldlm_cli_update_pool);
 
-int ldlm_cli_cancel(struct lustre_handle *lockh)
+/**
+ * Client side lock cancel.
+ *
+ * Lock must not have any readers or writers by this time.
+ */
+int ldlm_cli_cancel(struct lustre_handle *lockh,
+                   ldlm_cancel_flags_t cancel_flags)
 {
         struct obd_export *exp;
        int avail, flags, count = 1;
@@ -1273,13 +1353,13 @@ int ldlm_cli_cancel(struct lustre_handle *lockh)
         }
 
         rc = ldlm_cli_cancel_local(lock);
-       if (rc == LDLM_FL_LOCAL_ONLY) {
-                LDLM_LOCK_RELEASE(lock);
+       if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
+               LDLM_LOCK_RELEASE(lock);
                RETURN(0);
-        }
-        /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
-         * rpc which goes to canceld portal, so we can cancel other lru locks
-         * here and send them all as one LDLM_CANCEL rpc. */
+       }
+       /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
+        * RPC which goes to canceld portal, so we can cancel other LRU locks
+        * here and send them all as one LDLM_CANCEL RPC. */
         LASSERT(cfs_list_empty(&lock->l_bl_ast));
         cfs_list_add(&lock->l_bl_ast, &cancels);
 
@@ -1296,13 +1376,15 @@ int ldlm_cli_cancel(struct lustre_handle *lockh)
                 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
                                                LCF_BL_AST, flags);
         }
-        ldlm_cli_cancel_list(&cancels, count, NULL, 0);
+       ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
         RETURN(0);
 }
 EXPORT_SYMBOL(ldlm_cli_cancel);
 
-/* XXX until we will have compound requests and can cut cancels from generic rpc
- * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
+/**
+ * Locally cancel up to \a count locks in list \a cancels.
+ * Return the number of cancelled locks.
+ */
 int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
                                ldlm_cancel_flags_t flags)
 {
@@ -1322,11 +1404,15 @@ int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
                 } else {
                         rc = ldlm_cli_cancel_local(lock);
                 }
-                if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
-                        LDLM_DEBUG(lock, "Cancel lock separately");
-                        cfs_list_del_init(&lock->l_bl_ast);
-                        cfs_list_add(&lock->l_bl_ast, &head);
-                        bl_ast ++;
+               /* Until we have compound requests and can send LDLM_CANCEL
+                * requests batched with generic RPCs, we need to send cancels
+                * with the LDLM_FL_BL_AST flag in a separate RPC from
+                * the one being generated now. */
+               if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
+                       LDLM_DEBUG(lock, "Cancel lock separately");
+                       cfs_list_del_init(&lock->l_bl_ast);
+                       cfs_list_add(&lock->l_bl_ast, &head);
+                       bl_ast++;
                         continue;
                 }
                 if (rc == LDLM_FL_LOCAL_ONLY) {
@@ -1335,7 +1421,6 @@ int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
                         LDLM_LOCK_RELEASE(lock);
                         count--;
                 }
-
         }
         if (bl_ast > 0) {
                 count -= bl_ast;
@@ -1347,8 +1432,8 @@ int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
 
 /**
- * Cancel as many locks as possible w/o sending any rpcs (e.g. to write back
- * dirty data, to close a file, ...) or waiting for any rpcs in-flight (e.g.
+ * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
+ * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
  * readahead requests, ...)
  */
 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
@@ -1378,9 +1463,9 @@ static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
 }
 
 /**
- * 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.
+ * Callback function for LRU-resize policy. Decides 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
  *
@@ -1396,10 +1481,8 @@ static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
         __u64 slv, lvf, lv;
         cfs_time_t la;
 
-        /*
-         * Stop lru processing when we reached passed @count or checked all
-         * locks in lru.
-         */
+       /* Stop LRU processing when we reach past @count or have checked all
+        * locks in LRU. */
         if (count && added >= count)
                 return LDLM_POLICY_KEEP_LOCK;
 
@@ -1407,25 +1490,21 @@ static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
         lvf = ldlm_pool_get_lvf(pl);
         la = cfs_duration_sec(cfs_time_sub(cur,
                               lock->l_last_used));
+       lv = lvf * la * unused;
 
-        /*
-         * Stop when slv is not yet come from server or lv is smaller than
-         * it is.
-         */
-        lv = lvf * la * unused;
+       /* Inform pool about current CLV to see it via proc. */
+       ldlm_pool_set_clv(pl, lv);
 
-        /*
-         * Inform pool about current CLV to see it via proc.
-         */
-        ldlm_pool_set_clv(pl, lv);
-        return (slv == 0 || lv < slv) ?
-                LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
+       /* Stop when SLV is not yet come from server or lv is smaller than
+        * it is. */
+       return (slv == 0 || lv < slv) ?
+               LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
 }
 
 /**
  * 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.
+ * \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
  *
@@ -1436,32 +1515,27 @@ static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
                                                    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;
+       /* Stop LRU processing when we reach past @count or have 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.
+ * Callback function for aged policy. Makes decision whether to keep \a lock in
+ * LRU for current 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)
+                                                struct ldlm_lock *lock,
+                                                int unused, int added,
+                                                int count)
 {
-        /*
-         * Stop lru processing if young lock is found and we reached passed
-         * @count.
-         */
+       /* Stop LRU processing if young lock is found and we reach past count */
         return ((added >= count) &&
                 cfs_time_before(cfs_time_current(),
                                 cfs_time_add(lock->l_last_used,
@@ -1470,23 +1544,21 @@ static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
 }
 
 /**
- * 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.
+ * Callback function for default policy. Makes decision whether to keep \a lock
+ * in LRU for current 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)
+                                                   struct ldlm_lock *lock,
+                                                   int unused, int added,
+                                                   int count)
 {
-        /*
-         * Stop lru processing when we reached passed @count or checked all
-         * locks in lru.
-         */
+       /* Stop LRU processing when we reach past count or have checked all
+        * locks in LRU. */
         return (added >= count) ?
                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
 }
@@ -1517,46 +1589,48 @@ ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
         return ldlm_cancel_default_policy;
 }
 
-/* - Free space in lru for @count new locks,
+/**
+ * - Free space in LRU for \a count new locks,
  *   redundant unused locks are canceled locally;
  * - also cancel locally unused aged locks;
- * - do not cancel more than @max locks;
- * - GET the found locks and add them into the @cancels list.
+ * - do not cancel more than \a max locks;
+ * - GET the found locks and add them into the \a cancels list.
  *
  * A client lock can be added to the l_bl_ast list only when it is
- * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
- * There are the following use cases: ldlm_cancel_resource_local(),
- * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
- * flag properly. As any attempt to cancel a lock rely on this flag,
- * l_bl_ast list is accessed later without any special locking.
+ * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
+ * CANCEL.  There are the following use cases:
+ * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
+ * ldlm_cli_cancel(), which check and set this flag properly. As any
+ * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
+ * later without any special locking.
  *
- * Calling policies for enabled lru resize:
+ * Calling policies for enabled LRU resize:
  * ----------------------------------------
- * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
- *                            cancel not more than @count locks;
+ * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
+ *                            cancel not more than \a count locks;
  *
- * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
- *                              the beginning of lru list);
+ * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
+ *                              the beginning of LRU list);
  *
- * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
+ * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
  *                              memory pressre policy function;
  *
- * flags & LDLM_CANCEL_AGED -   cancel alocks according to "aged policy".
+ * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
  *
  * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
  *                               (typically before replaying locks) w/o
- *                               sending any rpcs or waiting for any
- *                               outstanding rpc to complete.
+ *                               sending any RPCs or waiting for any
+ *                               outstanding RPC to complete.
  */
 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
                                  int count, int max, int flags)
 {
-        ldlm_cancel_lru_policy_t pf;
-        struct ldlm_lock *lock, *next;
-        int added = 0, unused, remained;
-        ENTRY;
+       ldlm_cancel_lru_policy_t pf;
+       struct ldlm_lock *lock, *next;
+       int added = 0, unused, remained;
+       ENTRY;
 
-        cfs_spin_lock(&ns->ns_lock);
+       spin_lock(&ns->ns_lock);
         unused = ns->ns_nr_unused;
         remained = unused;
 
@@ -1578,7 +1652,7 @@ static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
                         break;
 
                 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
-                                             l_lru){
+                                            l_lru) {
                         /* No locks which got blocking requests. */
                         LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
 
@@ -1587,8 +1661,8 @@ static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
                                 /* already processed */
                                 continue;
 
-                        /* Somebody is already doing CANCEL. No need in this
-                         * lock in lru, do not traverse it again. */
+                       /* Somebody is already doing CANCEL. No need for this
+                        * lock in LRU, do not traverse it again. */
                         if (!(lock->l_flags & LDLM_FL_CANCELING))
                                 break;
 
@@ -1598,35 +1672,35 @@ static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
                         break;
 
                 LDLM_LOCK_GET(lock);
-                cfs_spin_unlock(&ns->ns_lock);
+               spin_unlock(&ns->ns_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. */
+               /* 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 a new lock is a valuable resource even if
+                * it has a low weight.
+                *
+                * That is, for shrinker policy we drop only
+                * old locks, but additionally choose them by
+                * their weight. Big extent locks will stay in
+                * the cache. */
                 result = pf(ns, lock, unused, added, count);
                 if (result == LDLM_POLICY_KEEP_LOCK) {
                         lu_ref_del(&lock->l_reference,
                                    __FUNCTION__, cfs_current());
                         LDLM_LOCK_RELEASE(lock);
-                        cfs_spin_lock(&ns->ns_lock);
-                        break;
-                }
-                if (result == LDLM_POLICY_SKIP_LOCK) {
-                        lu_ref_del(&lock->l_reference,
-                                   __FUNCTION__, cfs_current());
-                        LDLM_LOCK_RELEASE(lock);
-                        cfs_spin_lock(&ns->ns_lock);
+                       spin_lock(&ns->ns_lock);
+                       break;
+               }
+               if (result == LDLM_POLICY_SKIP_LOCK) {
+                       lu_ref_del(&lock->l_reference,
+                                  __func__, cfs_current());
+                       LDLM_LOCK_RELEASE(lock);
+                       spin_lock(&ns->ns_lock);
                         continue;
                 }
 
@@ -1634,16 +1708,15 @@ static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
                 /* 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. */
+                       /* Another thread is removing lock from LRU, or
+                        * somebody is already doing CANCEL, or there
+                        * is a blocking request which will send cancel
+                        * by itself, or the lock is no longer unused. */
                         unlock_res_and_lock(lock);
                         lu_ref_del(&lock->l_reference,
                                    __FUNCTION__, cfs_current());
                         LDLM_LOCK_RELEASE(lock);
-                        cfs_spin_lock(&ns->ns_lock);
+                       spin_lock(&ns->ns_lock);
                         continue;
                 }
                 LASSERT(!lock->l_readers && !lock->l_writers);
@@ -1673,12 +1746,12 @@ static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
                 cfs_list_add(&lock->l_bl_ast, cancels);
                 unlock_res_and_lock(lock);
                 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
-                cfs_spin_lock(&ns->ns_lock);
-                added++;
-                unused--;
-        }
-        cfs_spin_unlock(&ns->ns_lock);
-        RETURN(added);
+               spin_lock(&ns->ns_lock);
+               added++;
+               unused--;
+       }
+       spin_unlock(&ns->ns_lock);
+       RETURN(added);
 }
 
 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
@@ -1692,33 +1765,40 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
         return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
 }
 
-/* when called with LDLM_ASYNC the blocking callback will be handled
+/**
+ * Cancel at least \a nr locks from given namespace LRU.
+ *
+ * When called with LCF_ASYNC the blocking callback will be handled
  * 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 mode,
-                    int flags)
+ * asked to call the callback.  When called with LCF_ASYNC the blocking
+ * callback will be performed in this function.
+ */
+int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
+                   ldlm_cancel_flags_t cancel_flags,
+                   int flags)
 {
-        CFS_LIST_HEAD(cancels);
-        int count, rc;
-        ENTRY;
+       CFS_LIST_HEAD(cancels);
+       int count, rc;
+       ENTRY;
 
 #ifndef __KERNEL__
-        mode = LDLM_SYNC; /* force to be sync in user space */
+       cancel_flags &= ~LCF_ASYNC; /* force to be sync in user space */
 #endif
-        /* Just prepare the list of locks, do not actually cancel them yet.
-         * Locks are cancelled later in a separate thread. */
-        count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
-        rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, mode);
-        if (rc == 0)
-                RETURN(count);
-
-        RETURN(0);
+       /* Just prepare the list of locks, do not actually cancel them yet.
+        * Locks are cancelled later in a separate thread. */
+       count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
+       rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
+       if (rc == 0)
+               RETURN(count);
+
+       RETURN(0);
 }
 
-/* Find and cancel locally unused locks found on resource, matched to the
- * given policy, mode. GET the found locks and add them into the @cancels
- * list. */
+/**
+ * Find and cancel locally unused locks found on resource, matched to the
+ * given policy, mode. GET the found locks and add them into the \a cancels
+ * list.
+ */
 int ldlm_cancel_resource_local(struct ldlm_resource *res,
                                cfs_list_t *cancels,
                                ldlm_policy_data_t *policy,
@@ -1741,8 +1821,8 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res,
                 if (lock->l_readers || lock->l_writers)
                         continue;
 
-                /* If somebody is already doing CANCEL, or blocking ast came,
-                 * skip this lock. */
+               /* 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;
@@ -1772,12 +1852,16 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res,
 }
 EXPORT_SYMBOL(ldlm_cancel_resource_local);
 
-/* 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
+/**
+ * Cancel client-side locks from a list and send/prepare cancel RPCs to the
+ * server.
+ * If \a req is NULL, send CANCEL request to server with handles of locks
+ * in the \a 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
- * buffer at the offset @off.
- * Destroy @cancels at the end. */
+ * If \a req is not NULL, put handles of locks in \a cancels into the request
+ * buffer at the offset \a off.
+ * Destroy \a cancels at the end.
+ */
 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
                          struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
 {
@@ -1790,7 +1874,7 @@ int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
 
         /* 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.
+         * there are too many locks to be cancelled in LRU or on a resource.
          * It would also speed up the case when the server does not support
          * the feature. */
         while (count > 0) {
@@ -1812,11 +1896,11 @@ int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
                                                   cancels, 1, flags);
                 }
 
-                if (res < 0) {
-                       if (res != -ESHUTDOWN)
-                               CERROR("ldlm_cli_cancel_list: %d\n", res);
-                        res = count;
-                }
+               if (res < 0) {
+                       CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
+                                    "ldlm_cli_cancel_list: %d\n", res);
+                       res = count;
+               }
 
                 count -= res;
                 ldlm_lock_list_put(cancels, l_bl_ast, res);
@@ -1826,6 +1910,11 @@ int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
 }
 EXPORT_SYMBOL(ldlm_cli_cancel_list);
 
+/**
+ * Cancel all locks on a resource that have 0 readers/writers.
+ *
+ * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
+ * to notify the server. */
 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
                                     const struct ldlm_res_id *res_id,
                                     ldlm_policy_data_t *policy,
@@ -1851,7 +1940,8 @@ int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
                                            0, flags | LCF_BL_AST, opaque);
         rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
         if (rc != ELDLM_OK)
-                CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
+               CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
+                      PLDLMRES(res), rc);
 
         LDLM_RESOURCE_DELREF(res);
         ldlm_resource_putref(res);
@@ -1865,24 +1955,20 @@ struct ldlm_cli_cancel_arg {
 };
 
 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
-                                       cfs_hlist_node_t *hnode, void *arg)
+                                      cfs_hlist_node_t *hnode, void *arg)
 {
-        struct ldlm_resource           *res = cfs_hash_object(hs, hnode);
-        struct ldlm_cli_cancel_arg     *lc = arg;
-        int                             rc;
-
-        rc = ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
-                                             NULL, LCK_MINMODE,
-                                             lc->lc_flags, lc->lc_opaque);
-        if (rc != 0) {
-                CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
-                       res->lr_name.name[0], rc);
-        }
-        /* must return 0 for hash iteration */
-        return 0;
+       struct ldlm_resource           *res = cfs_hash_object(hs, hnode);
+       struct ldlm_cli_cancel_arg     *lc = arg;
+
+       ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
+                                       NULL, LCK_MINMODE, lc->lc_flags,
+                                       lc->lc_opaque);
+       /* must return 0 for hash iteration */
+       return 0;
 }
 
-/* Cancel all locks on a namespace (or a specific resource, if given)
+/**
+ * Cancel all locks on a namespace (or a specific resource, if given)
  * that have 0 readers/writers.
  *
  * If flags & LCF_LOCAL, throw the locks away without trying
@@ -2114,6 +2200,7 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
                 ldlm_lock_cancel(lock);
                 RETURN(0);
         }
+
         /*
          * If granted mode matches the requested mode, this lock is granted.
          *
@@ -2150,11 +2237,10 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
        body->lock_flags = ldlm_flags_to_wire(flags);
 
         ldlm_lock2handle(lock, &body->lock_handle[0]);
-        if (lock->l_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,
-                                     lock->l_lvb_len);
-        }
+       if (lock->l_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,
+                            lock->l_lvb_len);
         ptlrpc_request_set_replen(req);
         /* notify the server we've replayed all requests.
          * also, we mark the request to be put on a dedicated