Whamcloud - gitweb
b=22244 delegate lock cancel to blocking thread
[fs/lustre-release.git] / lustre / ldlm / ldlm_request.c
index 3352317..f808843 100644 (file)
@@ -140,7 +140,8 @@ static int ldlm_completion_tail(struct ldlm_lock *lock)
                            CFS_DURATION_T"s", delay);
 
                 /* Update our time estimate */
-                at_add(&lock->l_resource->lr_namespace->ns_at_estimate, delay);
+                at_measured(&lock->l_resource->lr_namespace->ns_at_estimate,
+                            delay);
                 result = 0;
         }
         return result;
@@ -246,9 +247,9 @@ noreproc:
         }
 
         if (imp != NULL) {
-                spin_lock(&imp->imp_lock);
+                cfs_spin_lock(&imp->imp_lock);
                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
-                spin_unlock(&imp->imp_lock);
+                cfs_spin_unlock(&imp->imp_lock);
         }
 
         if (ns_is_client(lock->l_resource->lr_namespace) &&
@@ -398,7 +399,6 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
         lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len);
         if (unlikely(!lock))
                 GOTO(out_nolock, err = -ENOMEM);
-        LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
 
         ldlm_lock_addref_internal(lock, mode);
         ldlm_lock2handle(lock, lockh);
@@ -421,13 +421,10 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
         if (policy != NULL)
                 *policy = lock->l_policy_data;
 
-        LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
-                          lock);
-
         if (lock->l_completion_ast)
                 lock->l_completion_ast(lock, *flags, NULL);
 
-        LDLM_DEBUG(lock, "client-side local enqueue END");
+        LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
         EXIT;
  out:
         LDLM_LOCK_RELEASE(lock);
@@ -436,8 +433,7 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 }
 
 static void failed_lock_cleanup(struct ldlm_namespace *ns,
-                                struct ldlm_lock *lock,
-                                struct lustre_handle *lockh, int mode)
+                                struct ldlm_lock *lock, int mode)
 {
         int need_cancel = 0;
 
@@ -450,25 +446,31 @@ static void failed_lock_cleanup(struct ldlm_namespace *ns,
                  * bl_ast and -EINVAL reply is sent to server anyways.
                  * bug 17645 */
                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
-                                 LDLM_FL_ATOMIC_CB;
+                                 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
                 need_cancel = 1;
         }
         unlock_res_and_lock(lock);
 
-        if (need_cancel) {
+        if (need_cancel)
                 LDLM_DEBUG(lock,
                            "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
-                           "LDLM_FL_ATOMIC_CB");
-                ldlm_lock_decref_and_cancel(lockh, mode);
-        } else {
+                           "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
+        else
                 LDLM_DEBUG(lock, "lock was granted or failed in race");
-                ldlm_lock_decref(lockh, mode);
-        }
+
+        ldlm_lock_decref_internal(lock, mode);
 
         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
          *       from llite/file.c/ll_file_flock(). */
+        /* This code makes for the fact that we do not have blocking handler on
+         * a client for flock locks. As such this is the place where we must
+         * completely kill failed locks. (interrupted and those that
+         * were waiting to be granted when server evicted us. */
         if (lock->l_resource->lr_type == LDLM_FLOCK) {
-                ldlm_lock_destroy(lock);
+                lock_res_and_lock(lock);
+                ldlm_resource_unlink_lock(lock);
+                ldlm_lock_destroy_nolock(lock);
+                unlock_res_and_lock(lock);
         }
 }
 
@@ -617,7 +619,7 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
                         int err = lock->l_completion_ast(lock, *flags, NULL);
                         if (!rc)
                                 rc = err;
-                        if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
+                        if (rc)
                                 cleanup_phase = 1;
                 }
         }
@@ -632,7 +634,7 @@ int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
         EXIT;
 cleanup:
         if (cleanup_phase == 1 && rc)
-                failed_lock_cleanup(ns, lock, lockh, mode);
+                failed_lock_cleanup(ns, lock, mode);
         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
         LDLM_LOCK_PUT(lock);
         LDLM_LOCK_RELEASE(lock);
@@ -676,7 +678,7 @@ static inline int ldlm_format_handles_avail(struct obd_import *imp,
  * @count locks in @cancels. */
 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
                       int version, int opc, int canceloff,
-                      struct list_head *cancels, int count)
+                      cfs_list_t *cancels, int count)
 {
         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
         struct req_capsule      *pill = &req->rq_pill;
@@ -739,7 +741,7 @@ int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
 }
 
 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
-                          struct list_head *cancels, int count)
+                          cfs_list_t *cancels, int count)
 {
         return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
                                  LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
@@ -754,7 +756,7 @@ int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
 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 *policy, int *flags,
+                     ldlm_policy_data_t const *policy, int *flags,
                      void *lvb, __u32 lvb_len, struct lustre_handle *lockh,
                      int async)
 {
@@ -819,7 +821,7 @@ int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
                                                 LUSTRE_DLM_VERSION,
                                                 LDLM_ENQUEUE);
                 if (req == NULL) {
-                        failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
+                        failed_lock_cleanup(ns, lock, einfo->ei_mode);
                         LDLM_LOCK_RELEASE(lock);
                         RETURN(-ENOMEM);
                 }
@@ -1030,7 +1032,6 @@ static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
                 LDLM_DEBUG(lock, "server-side local cancel");
                 ldlm_lock_cancel(lock);
                 ldlm_reprocess_all(lock->l_resource);
-                LDLM_DEBUG(lock, "server-side local cancel handler END");
         }
 
         RETURN(rc);
@@ -1039,7 +1040,7 @@ static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
    of the request @req. */
 static void ldlm_cancel_pack(struct ptlrpc_request *req,
-                             struct list_head *head, int count)
+                             cfs_list_t *head, int count)
 {
         struct ldlm_request *dlm;
         struct ldlm_lock *lock;
@@ -1059,7 +1060,7 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req,
         /* XXX: it would be better to pack lock handles grouped by resource.
          * so that the server cancel would call filter_lvbo_update() less
          * frequently. */
-        list_for_each_entry(lock, head, l_bl_ast) {
+        cfs_list_for_each_entry(lock, head, l_bl_ast) {
                 if (!count--)
                         break;
                 LASSERT(lock->l_conn_export);
@@ -1074,8 +1075,8 @@ static void ldlm_cancel_pack(struct ptlrpc_request *req,
 
 /* Prepare and send a batched cancel rpc, it will include count lock handles
  * of locks given in @head. */
-int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
-                        int count, int flags)
+int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
+                        int count, ldlm_cancel_flags_t flags)
 {
         struct ptlrpc_request *req = NULL;
         struct obd_import *imp;
@@ -1129,7 +1130,7 @@ int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
                 ldlm_cancel_pack(req, cancels, count);
 
                 ptlrpc_request_set_replen(req);
-                if (flags & LDLM_FL_ASYNC) {
+                if (flags & LCF_ASYNC) {
                         ptlrpcd_add_req(req, PSCOPE_OTHER);
                         sent = count;
                         GOTO(out, 0);
@@ -1212,11 +1213,11 @@ int ldlm_cli_update_pool(struct ptlrpc_request *req)
          * alive in cleanup time. Evil races are possible which may cause
          * oops in that time.
          */
-        write_lock(&obd->obd_pool_lock);
+        cfs_write_lock(&obd->obd_pool_lock);
         old_slv = obd->obd_pool_slv;
         obd->obd_pool_slv = new_slv;
         obd->obd_pool_limit = new_limit;
-        write_unlock(&obd->obd_pool_lock);
+        cfs_write_unlock(&obd->obd_pool_lock);
 
         RETURN(0);
 }
@@ -1246,8 +1247,8 @@ int ldlm_cli_cancel(struct lustre_handle *lockh)
         /* 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(list_empty(&lock->l_bl_ast));
-        list_add(&lock->l_bl_ast, &cancels);
+        LASSERT(cfs_list_empty(&lock->l_bl_ast));
+        cfs_list_add(&lock->l_bl_ast, &cancels);
 
         exp = lock->l_conn_export;
         if (exp_connect_cancelset(exp)) {
@@ -1260,7 +1261,7 @@ int ldlm_cli_cancel(struct lustre_handle *lockh)
                 flags = ns_connect_lru_resize(ns) ?
                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
                 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
-                                               LDLM_FL_BL_AST, flags);
+                                               LCF_BL_AST, flags);
         }
         ldlm_cli_cancel_list(&cancels, count, NULL, 0);
         RETURN(0);
@@ -1268,33 +1269,34 @@ int ldlm_cli_cancel(struct lustre_handle *lockh)
 
 /* 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 */
-static int ldlm_cancel_list(struct list_head *cancels, int count, int flags)
+int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
+                               ldlm_cancel_flags_t flags)
 {
         CFS_LIST_HEAD(head);
         struct ldlm_lock *lock, *next;
         int left = 0, bl_ast = 0, rc;
 
         left = count;
-        list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
+        cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
                 if (left-- == 0)
                         break;
 
-                if (flags & LDLM_FL_LOCAL_ONLY) {
+                if (flags & LCF_LOCAL) {
                         rc = LDLM_FL_LOCAL_ONLY;
                         ldlm_lock_cancel(lock);
                 } else {
                         rc = ldlm_cli_cancel_local(lock);
                 }
-                if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) {
+                if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
                         LDLM_DEBUG(lock, "Cancel lock separately");
-                        list_del_init(&lock->l_bl_ast);
-                        list_add(&lock->l_bl_ast, &head);
+                        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) {
                         /* CANCEL RPC should not be sent to server. */
-                        list_del_init(&lock->l_bl_ast);
+                        cfs_list_del_init(&lock->l_bl_ast);
                         LDLM_LOCK_RELEASE(lock);
                         count--;
                 }
@@ -1349,7 +1351,7 @@ static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
          * Inform pool about current CLV to see it via proc.
          */
         ldlm_pool_set_clv(pl, lv);
-        return (slv == 1 || lv < slv) ?
+        return (slv == 0 || lv < slv) ?
                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
 }
 
@@ -1471,15 +1473,15 @@ ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
  *
  * flags & LDLM_CANCEL_AGED -   cancel alocks according to "aged policy".
  */
-int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
-                          int count, int max, int cancel_flags, int flags)
+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;
         ENTRY;
 
-        spin_lock(&ns->ns_unused_lock);
+        cfs_spin_lock(&ns->ns_unused_lock);
         unused = ns->ns_nr_unused;
 
         if (!ns_connect_lru_resize(ns))
@@ -1488,12 +1490,13 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
         pf = ldlm_cancel_lru_policy(ns, flags);
         LASSERT(pf != NULL);
 
-        while (!list_empty(&ns->ns_unused_list)) {
+        while (!cfs_list_empty(&ns->ns_unused_list)) {
                 /* For any flags, stop scanning if @max is reached. */
                 if (max && added >= max)
                         break;
 
-                list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru){
+                cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
+                                             l_lru){
                         /* No locks which got blocking requests. */
                         LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
 
@@ -1508,7 +1511,7 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
                         break;
 
                 LDLM_LOCK_GET(lock);
-                spin_unlock(&ns->ns_unused_lock);
+                cfs_spin_unlock(&ns->ns_unused_lock);
                 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
 
                 /* Pass the lock through the policy filter and see if it
@@ -1529,7 +1532,7 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
                         lu_ref_del(&lock->l_reference,
                                    __FUNCTION__, cfs_current());
                         LDLM_LOCK_RELEASE(lock);
-                        spin_lock(&ns->ns_unused_lock);
+                        cfs_spin_lock(&ns->ns_unused_lock);
                         break;
                 }
 
@@ -1546,7 +1549,7 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
                         lu_ref_del(&lock->l_reference,
                                    __FUNCTION__, cfs_current());
                         LDLM_LOCK_RELEASE(lock);
-                        spin_lock(&ns->ns_unused_lock);
+                        cfs_spin_lock(&ns->ns_unused_lock);
                         continue;
                 }
                 LASSERT(!lock->l_readers && !lock->l_writers);
@@ -1572,23 +1575,34 @@ int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
                  * and can't use l_pending_chain as it is used both on
                  * server and client nevertheless bug 5666 says it is
                  * used only on server */
-                LASSERT(list_empty(&lock->l_bl_ast));
-                list_add(&lock->l_bl_ast, cancels);
+                LASSERT(cfs_list_empty(&lock->l_bl_ast));
+                cfs_list_add(&lock->l_bl_ast, cancels);
                 unlock_res_and_lock(lock);
                 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
-                spin_lock(&ns->ns_unused_lock);
+                cfs_spin_lock(&ns->ns_unused_lock);
                 added++;
                 unused--;
         }
-        spin_unlock(&ns->ns_unused_lock);
-        RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
+        cfs_spin_unlock(&ns->ns_unused_lock);
+        RETURN(added);
+}
+
+int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
+                          int count, int max, ldlm_cancel_flags_t cancel_flags,
+                          int flags)
+{
+        int added;
+        added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
+        if (added <= 0)
+                return added;
+        return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
 }
 
 /* when called with LDLM_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 sync,
+int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t mode,
                     int flags)
 {
         CFS_LIST_HEAD(cancels);
@@ -1596,36 +1610,33 @@ int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
         ENTRY;
 
 #ifndef __KERNEL__
-        sync = LDLM_SYNC; /* force to be sync in user space */
+        mode = LDLM_SYNC; /* force to be sync in user space */
 #endif
-        count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
-        if (sync == LDLM_ASYNC) {
-                rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
-                if (rc == 0)
-                        RETURN(count);
-        }
+        /* 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);
 
-        /* If an error occured in ASYNC mode, or this is SYNC mode,
-         * cancel the list. */
-        ldlm_cli_cancel_list(&cancels, count, NULL, 0);
-        RETURN(count);
+        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. */
 int ldlm_cancel_resource_local(struct ldlm_resource *res,
-                               struct list_head *cancels,
+                               cfs_list_t *cancels,
                                ldlm_policy_data_t *policy,
                                ldlm_mode_t mode, int lock_flags,
-                               int cancel_flags, void *opaque)
+                               ldlm_cancel_flags_t cancel_flags, void *opaque)
 {
         struct ldlm_lock *lock;
         int count = 0;
         ENTRY;
 
         lock_res(res);
-        list_for_each_entry(lock, &res->lr_granted, l_res_link) {
+        cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
                 if (opaque != NULL && lock->l_ast_data != opaque) {
                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
                                    lock->l_ast_data, opaque);
@@ -1633,13 +1644,8 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res,
                         continue;
                 }
 
-                if (lock->l_readers || lock->l_writers) {
-                        if (cancel_flags & LDLM_FL_WARN) {
-                                LDLM_ERROR(lock, "lock in use");
-                                //LBUG();
-                        }
+                if (lock->l_readers || lock->l_writers)
                         continue;
-                }
 
                 /* If somebody is already doing CANCEL, or blocking ast came,
                  * skip this lock. */
@@ -1661,14 +1667,14 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res,
                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
                                  lock_flags;
 
-                LASSERT(list_empty(&lock->l_bl_ast));
-                list_add(&lock->l_bl_ast, cancels);
+                LASSERT(cfs_list_empty(&lock->l_bl_ast));
+                cfs_list_add(&lock->l_bl_ast, cancels);
                 LDLM_LOCK_GET(lock);
                 count++;
         }
         unlock_res(res);
 
-        RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
+        RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
 }
 
 /* If @req is NULL, send CANCEL request to server with handles of locks
@@ -1677,14 +1683,14 @@ int ldlm_cancel_resource_local(struct ldlm_resource *res,
  * If @req is not NULL, put handles of locks in @cancels into the request
  * buffer at the offset @off.
  * Destroy @cancels at the end. */
-int ldlm_cli_cancel_list(struct list_head *cancels, int count,
-                         struct ptlrpc_request *req, int flags)
+int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
+                         struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
 {
         struct ldlm_lock *lock;
         int res = 0;
         ENTRY;
 
-        if (list_empty(cancels) || count == 0)
+        if (cfs_list_empty(cancels) || count == 0)
                 RETURN(0);
 
         /* XXX: requests (both batched and not) could be sent in parallel.
@@ -1693,8 +1699,9 @@ int ldlm_cli_cancel_list(struct list_head *cancels, int count,
          * It would also speed up the case when the server does not support
          * the feature. */
         while (count > 0) {
-                LASSERT(!list_empty(cancels));
-                lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
+                LASSERT(!cfs_list_empty(cancels));
+                lock = cfs_list_entry(cancels->next, struct ldlm_lock,
+                                      l_bl_ast);
                 LASSERT(lock->l_conn_export);
 
                 if (exp_connect_cancelset(lock->l_conn_export)) {
@@ -1725,7 +1732,9 @@ 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,
-                                    ldlm_mode_t mode, int flags, void *opaque)
+                                    ldlm_mode_t mode,
+                                    ldlm_cancel_flags_t flags,
+                                    void *opaque)
 {
         struct ldlm_resource *res;
         CFS_LIST_HEAD(cancels);
@@ -1756,10 +1765,10 @@ static inline int have_no_nsresource(struct ldlm_namespace *ns)
 {
         int no_resource = 0;
 
-        spin_lock(&ns->ns_hash_lock);
+        cfs_spin_lock(&ns->ns_hash_lock);
         if (ns->ns_resources == 0)
                 no_resource = 1;
-        spin_unlock(&ns->ns_hash_lock);
+        cfs_spin_unlock(&ns->ns_hash_lock);
 
         RETURN(no_resource);
 }
@@ -1767,11 +1776,11 @@ static inline int have_no_nsresource(struct ldlm_namespace *ns)
 /* Cancel all locks on a namespace (or a specific resource, if given)
  * that have 0 readers/writers.
  *
- * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
+ * If flags & LCF_LOCAL, throw the locks away without trying
  * to notify the server. */
 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
                            const struct ldlm_res_id *res_id,
-                           int flags, void *opaque)
+                           ldlm_cancel_flags_t flags, void *opaque)
 {
         int i;
         ENTRY;
@@ -1784,17 +1793,18 @@ int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
                                                        LCK_MINMODE, flags,
                                                        opaque));
 
-        spin_lock(&ns->ns_hash_lock);
+        cfs_spin_lock(&ns->ns_hash_lock);
         for (i = 0; i < RES_HASH_SIZE; i++) {
-                struct list_head *tmp;
+                cfs_list_t *tmp;
                 tmp = ns->ns_hash[i].next;
                 while (tmp != &(ns->ns_hash[i])) {
                         struct ldlm_resource *res;
                         int rc;
 
-                        res = list_entry(tmp, struct ldlm_resource, lr_hash);
+                        res = cfs_list_entry(tmp, struct ldlm_resource,
+                                             lr_hash);
                         ldlm_resource_getref(res);
-                        spin_unlock(&ns->ns_hash_lock);
+                        cfs_spin_unlock(&ns->ns_hash_lock);
 
                         LDLM_RESOURCE_ADDREF(res);
                         rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
@@ -1806,12 +1816,12 @@ int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
                                        res->lr_name.name[0], rc);
 
                         LDLM_RESOURCE_DELREF(res);
-                        spin_lock(&ns->ns_hash_lock);
+                        cfs_spin_lock(&ns->ns_hash_lock);
                         tmp = tmp->next;
                         ldlm_resource_putref_locked(res);
                 }
         }
-        spin_unlock(&ns->ns_hash_lock);
+        cfs_spin_unlock(&ns->ns_hash_lock);
 
         RETURN(ELDLM_OK);
 }
@@ -1821,7 +1831,7 @@ int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
                           void *closure)
 {
-        struct list_head *tmp, *next;
+        cfs_list_t *tmp, *next;
         struct ldlm_lock *lock;
         int rc = LDLM_ITER_CONTINUE;
 
@@ -1831,22 +1841,22 @@ int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
                 RETURN(LDLM_ITER_CONTINUE);
 
         lock_res(res);
-        list_for_each_safe(tmp, next, &res->lr_granted) {
-                lock = list_entry(tmp, struct ldlm_lock, l_res_link);
+        cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
+                lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
 
                 if (iter(lock, closure) == LDLM_ITER_STOP)
                         GOTO(out, rc = LDLM_ITER_STOP);
         }
 
-        list_for_each_safe(tmp, next, &res->lr_converting) {
-                lock = list_entry(tmp, struct ldlm_lock, l_res_link);
+        cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
+                lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
 
                 if (iter(lock, closure) == LDLM_ITER_STOP)
                         GOTO(out, rc = LDLM_ITER_STOP);
         }
 
-        list_for_each_safe(tmp, next, &res->lr_waiting) {
-                lock = list_entry(tmp, struct ldlm_lock, l_res_link);
+        cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
+                lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
 
                 if (iter(lock, closure) == LDLM_ITER_STOP)
                         GOTO(out, rc = LDLM_ITER_STOP);
@@ -1884,22 +1894,23 @@ int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 {
         int i, rc = LDLM_ITER_CONTINUE;
         struct ldlm_resource *res;
-        struct list_head *tmp;
+        cfs_list_t *tmp;
 
         ENTRY;
-        spin_lock(&ns->ns_hash_lock);
+        cfs_spin_lock(&ns->ns_hash_lock);
         for (i = 0; i < RES_HASH_SIZE; i++) {
                 tmp = ns->ns_hash[i].next;
                 while (tmp != &(ns->ns_hash[i])) {
-                        res = list_entry(tmp, struct ldlm_resource, lr_hash);
+                        res = cfs_list_entry(tmp, struct ldlm_resource,
+                                             lr_hash);
                         ldlm_resource_getref(res);
-                        spin_unlock(&ns->ns_hash_lock);
+                        cfs_spin_unlock(&ns->ns_hash_lock);
                         LDLM_RESOURCE_ADDREF(res);
 
                         rc = iter(res, closure);
 
                         LDLM_RESOURCE_DELREF(res);
-                        spin_lock(&ns->ns_hash_lock);
+                        cfs_spin_lock(&ns->ns_hash_lock);
                         tmp = tmp->next;
                         ldlm_resource_putref_locked(res);
                         if (rc == LDLM_ITER_STOP)
@@ -1907,16 +1918,21 @@ int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
                 }
         }
  out:
-        spin_unlock(&ns->ns_hash_lock);
+        cfs_spin_unlock(&ns->ns_hash_lock);
         RETURN(rc);
 }
 
-/* non-blocking function to manipulate a lock whose cb_data is being put away.*/
-void ldlm_resource_iterate(struct ldlm_namespace *ns,
-                           const struct ldlm_res_id *res_id,
-                           ldlm_iterator_t iter, void *data)
+/* non-blocking function to manipulate a lock whose cb_data is being put away.
+ * return  0:  find no resource
+ *       > 0:  must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
+ *       < 0:  errors
+ */
+int ldlm_resource_iterate(struct ldlm_namespace *ns,
+                          const struct ldlm_res_id *res_id,
+                          ldlm_iterator_t iter, void *data)
 {
         struct ldlm_resource *res;
+        int rc;
         ENTRY;
 
         if (ns == NULL) {
@@ -1925,33 +1941,32 @@ void ldlm_resource_iterate(struct ldlm_namespace *ns,
         }
 
         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
-        if (res == NULL) {
-                EXIT;
-                return;
-        }
+        if (res == NULL)
+                RETURN(0);
 
         LDLM_RESOURCE_ADDREF(res);
-        ldlm_resource_foreach(res, iter, data);
+        rc = ldlm_resource_foreach(res, iter, data);
         LDLM_RESOURCE_DELREF(res);
         ldlm_resource_putref(res);
-        EXIT;
+        RETURN(rc);
 }
 
 /* Lock replay */
 
 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
 {
-        struct list_head *list = closure;
+        cfs_list_t *list = closure;
 
         /* we use l_pending_chain here, because it's unused on clients. */
-        LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
+        LASSERTF(cfs_list_empty(&lock->l_pending_chain),
+                 "lock %p next %p prev %p\n",
                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
         /* bug 9573: don't replay locks left after eviction, or
          * bug 17614: locks being actively cancelled. Get a reference
          * on a lock so that it does not disapear under us (e.g. due to cancel)
          */
         if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
-                list_add(&lock->l_pending_chain, list);
+                cfs_list_add(&lock->l_pending_chain, list);
                 LDLM_LOCK_GET(lock);
         }
 
@@ -1968,7 +1983,7 @@ static int replay_lock_interpret(const struct lu_env *env,
         struct obd_export    *exp;
 
         ENTRY;
-        atomic_dec(&req->rq_import->imp_replay_inflight);
+        cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
         if (rc != ELDLM_OK)
                 GOTO(out, rc);
 
@@ -2048,7 +2063,7 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
         else if (lock->l_granted_mode)
                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
-        else if (!list_empty(&lock->l_res_link))
+        else if (!cfs_list_empty(&lock->l_res_link))
                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
         else
                 flags = LDLM_FL_REPLAY;
@@ -2080,7 +2095,7 @@ 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);
+        cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
         aa = ptlrpc_req_async_args(req);
         aa->lock_handle = body->lock_handle[0];
@@ -2099,15 +2114,19 @@ int ldlm_replay_locks(struct obd_import *imp)
 
         ENTRY;
 
-        LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
+        LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
+
+        /* don't replay locks if import failed recovery */
+        if (imp->imp_vbr_failed)
+                RETURN(0);
 
         /* ensure this doesn't fall to 0 before all have been queued */
-        atomic_inc(&imp->imp_replay_inflight);
+        cfs_atomic_inc(&imp->imp_replay_inflight);
 
         (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
 
-        list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
-                list_del_init(&lock->l_pending_chain);
+        cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
+                cfs_list_del_init(&lock->l_pending_chain);
                 if (rc) {
                         LDLM_LOCK_PUT(lock);
                         continue; /* or try to do the rest? */
@@ -2116,7 +2135,7 @@ int ldlm_replay_locks(struct obd_import *imp)
                 LDLM_LOCK_PUT(lock);
         }
 
-        atomic_dec(&imp->imp_replay_inflight);
+        cfs_atomic_dec(&imp->imp_replay_inflight);
 
         RETURN(rc);
 }