Whamcloud - gitweb
LU-13227 LDLM: update LVB if it is server lock
[fs/lustre-release.git] / lustre / ldlm / ldlm_lock.c
index 44680dd..aa2eeea 100644 (file)
@@ -136,8 +136,6 @@ const char *ldlm_it2str(enum ldlm_intent_flags it)
 }
 EXPORT_SYMBOL(ldlm_it2str);
 
-extern struct kmem_cache *ldlm_lock_slab;
-
 #ifdef HAVE_SERVER_SUPPORT
 static ldlm_processing_policy ldlm_processing_policy_table[] = {
        [LDLM_PLAIN]    = ldlm_process_plain_lock,
@@ -151,6 +149,19 @@ ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
         return ldlm_processing_policy_table[res->lr_type];
 }
 EXPORT_SYMBOL(ldlm_get_processing_policy);
+
+static ldlm_reprocessing_policy ldlm_reprocessing_policy_table[] = {
+       [LDLM_PLAIN]    = ldlm_reprocess_queue,
+       [LDLM_EXTENT]   = ldlm_reprocess_queue,
+       [LDLM_FLOCK]    = ldlm_reprocess_queue,
+       [LDLM_IBITS]    = ldlm_reprocess_inodebits_queue,
+};
+
+ldlm_reprocessing_policy ldlm_get_reprocessing_policy(struct ldlm_resource *res)
+{
+       return ldlm_reprocessing_policy_table[res->lr_type];
+}
+
 #endif /* HAVE_SERVER_SUPPORT */
 
 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
@@ -174,11 +185,20 @@ EXPORT_SYMBOL(ldlm_register_intent);
  */
 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
 {
-       atomic_inc(&lock->l_refc);
+       refcount_inc(&lock->l_handle.h_ref);
         return lock;
 }
 EXPORT_SYMBOL(ldlm_lock_get);
 
+static void lock_handle_free(struct rcu_head *rcu)
+{
+       struct ldlm_lock *lock = container_of(rcu, struct ldlm_lock,
+                                             l_handle.h_rcu);
+
+       OBD_FREE_PRE(lock, sizeof(*lock), "slab-freed");
+       kmem_cache_free(ldlm_lock_slab, lock);
+}
+
 /**
  * Release lock reference.
  *
@@ -189,8 +209,8 @@ void ldlm_lock_put(struct ldlm_lock *lock)
         ENTRY;
 
         LASSERT(lock->l_resource != LP_POISON);
-       LASSERT(atomic_read(&lock->l_refc) > 0);
-       if (atomic_dec_and_test(&lock->l_refc)) {
+       LASSERT(refcount_read(&lock->l_handle.h_ref) > 0);
+       if (refcount_dec_and_test(&lock->l_handle.h_ref)) {
                 struct ldlm_resource *res;
 
                 LDLM_DEBUG(lock,
@@ -205,8 +225,6 @@ void ldlm_lock_put(struct ldlm_lock *lock)
                 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
                                      LDLM_NSS_LOCKS);
                 lu_ref_del(&res->lr_reference, "lock", lock);
-                ldlm_resource_putref(res);
-                lock->l_resource = NULL;
                 if (lock->l_export) {
                         class_export_lock_put(lock->l_export, lock);
                         lock->l_export = NULL;
@@ -215,9 +233,17 @@ void ldlm_lock_put(struct ldlm_lock *lock)
                 if (lock->l_lvb_data != NULL)
                         OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
 
-                ldlm_interval_free(ldlm_interval_detach(lock));
+               if (res->lr_type == LDLM_EXTENT) {
+                       ldlm_interval_free(ldlm_interval_detach(lock));
+               } else if (res->lr_type == LDLM_IBITS) {
+                       if (lock->l_ibits_node != NULL)
+                               OBD_SLAB_FREE_PTR(lock->l_ibits_node,
+                                                 ldlm_inodebits_slab);
+               }
+               ldlm_resource_putref(res);
+               lock->l_resource = NULL;
                 lu_ref_fini(&lock->l_reference);
-               OBD_FREE_RCU(lock, sizeof(*lock), &lock->l_handle);
+               call_rcu(&lock->l_handle.h_rcu, lock_handle_free);
         }
 
         EXIT;
@@ -421,22 +447,7 @@ void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
         EXIT;
 }
 
-/* this is called by portals_handle2object with the handle lock taken */
-static void lock_handle_addref(void *lock)
-{
-        LDLM_LOCK_GET((struct ldlm_lock *)lock);
-}
-
-static void lock_handle_free(void *lock, int size)
-{
-       LASSERT(size == sizeof(struct ldlm_lock));
-       OBD_SLAB_FREE(lock, ldlm_lock_slab, size);
-}
-
-static struct portals_handle_ops lock_handle_ops = {
-       .hop_addref = lock_handle_addref,
-       .hop_free   = lock_handle_free,
-};
+static const char lock_handle_owner[] = "ldlm";
 
 /**
  *
@@ -462,7 +473,7 @@ static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
        lock->l_resource = resource;
        lu_ref_add(&resource->lr_reference, "lock", lock);
 
-       atomic_set(&lock->l_refc, 2);
+       refcount_set(&lock->l_handle.h_ref, 2);
        INIT_LIST_HEAD(&lock->l_res_link);
        INIT_LIST_HEAD(&lock->l_lru);
        INIT_LIST_HEAD(&lock->l_pending_chain);
@@ -478,8 +489,8 @@ static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
 
         lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
                              LDLM_NSS_LOCKS);
-       INIT_LIST_HEAD(&lock->l_handle.h_link);
-       class_handle_hash(&lock->l_handle, &lock_handle_ops);
+       INIT_HLIST_NODE(&lock->l_handle.h_link);
+       class_handle_hash(&lock->l_handle, lock_handle_owner);
 
         lu_ref_init(&lock->l_reference);
         lu_ref_add(&lock->l_reference, "hash", lock);
@@ -589,7 +600,10 @@ struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
 
        LASSERT(handle);
 
-       lock = class_handle2object(handle->cookie, NULL);
+       if (!lustre_handle_is_used(handle))
+               RETURN(NULL);
+
+       lock = class_handle2object(handle->cookie, lock_handle_owner);
        if (lock == NULL)
                RETURN(NULL);
 
@@ -603,7 +617,7 @@ struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
        /* It's unlikely but possible that someone marked the lock as
         * destroyed after we did handle2object on it */
        if ((flags == 0) && !ldlm_is_destroyed(lock)) {
-               lu_ref_add(&lock->l_reference, "handle", current);
+               lu_ref_add_atomic(&lock->l_reference, "handle", current);
                RETURN(lock);
        }
 
@@ -665,12 +679,19 @@ static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
                 * discard dirty data, rather than writing back. */
                if (ldlm_is_ast_discard_data(new))
                        ldlm_set_discard_data(lock);
-               LASSERT(list_empty(&lock->l_bl_ast));
-               list_add(&lock->l_bl_ast, work_list);
-                LDLM_LOCK_GET(lock);
-                LASSERT(lock->l_blocking_lock == NULL);
-                lock->l_blocking_lock = LDLM_LOCK_GET(new);
-        }
+
+               /* Lock can be converted from a blocking state back to granted
+                * after lock convert or COS downgrade but still be in an
+                * older bl_list because it is controlled only by
+                * ldlm_work_bl_ast_lock(), let it be processed there.
+                */
+               if (list_empty(&lock->l_bl_ast)) {
+                       list_add(&lock->l_bl_ast, work_list);
+                       LDLM_LOCK_GET(lock);
+               }
+               LASSERT(lock->l_blocking_lock == NULL);
+               lock->l_blocking_lock = LDLM_LOCK_GET(new);
+       }
 }
 
 /**
@@ -1028,7 +1049,6 @@ static void search_granted_lock(struct list_head *queue,
         prev->mode_link = &req->l_sl_mode;
         prev->policy_link = &req->l_sl_policy;
         EXIT;
-        return;
 }
 
 /**
@@ -1077,7 +1097,7 @@ void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
 {
        struct sl_insert_point prev;
 
-       LASSERT(lock->l_req_mode == lock->l_granted_mode);
+       LASSERT(ldlm_is_granted(lock));
 
        search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
        ldlm_granted_list_add_lock(lock, &prev);
@@ -1131,28 +1151,16 @@ void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
 }
 
 /**
- * Describe the overlap between two locks.  itree_overlap_cb data.
- */
-struct lock_match_data {
-       struct ldlm_lock        *lmd_old;
-       struct ldlm_lock        *lmd_lock;
-       enum ldlm_mode          *lmd_mode;
-       union ldlm_policy_data  *lmd_policy;
-       __u64                    lmd_flags;
-       int                      lmd_unref;
-};
-
-/**
  * Check if the given @lock meets the criteria for a match.
  * A reference on the lock is taken if matched.
  *
  * \param lock     test-against this lock
  * \param data    parameters
  */
-static int lock_matches(struct ldlm_lock *lock, struct lock_match_data *data)
+static int lock_matches(struct ldlm_lock *lock, struct ldlm_match_data *data)
 {
        union ldlm_policy_data *lpol = &lock->l_policy_data;
-       enum ldlm_mode match;
+       enum ldlm_mode match = LCK_MINMODE;
 
        if (lock == data->lmd_old)
                return INTERVAL_ITER_STOP;
@@ -1177,6 +1185,17 @@ static int lock_matches(struct ldlm_lock *lock, struct lock_match_data *data)
 
        if (!(lock->l_req_mode & *data->lmd_mode))
                return INTERVAL_ITER_CONT;
+
+       /* When we search for ast_data, we are not doing a traditional match,
+        * so we don't worry about IBITS or extent matching.
+        */
+       if (data->lmd_has_ast_data) {
+               if (!lock->l_ast_data)
+                       return INTERVAL_ITER_CONT;
+
+               goto matched;
+       }
+
        match = lock->l_req_mode;
 
        switch (lock->l_resource->lr_type) {
@@ -1210,6 +1229,11 @@ static int lock_matches(struct ldlm_lock *lock, struct lock_match_data *data)
        if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
                return INTERVAL_ITER_CONT;
 
+       /* Filter locks by skipping flags */
+       if (data->lmd_skip_flags & lock->l_flags)
+               return INTERVAL_ITER_CONT;
+
+matched:
        if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
                LDLM_LOCK_GET(lock);
                ldlm_lock_touch_in_lru(lock);
@@ -1226,7 +1250,7 @@ static int lock_matches(struct ldlm_lock *lock, struct lock_match_data *data)
 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
 {
        struct ldlm_interval *node = to_ldlm_interval(in);
-       struct lock_match_data *data = args;
+       struct ldlm_match_data *data = args;
        struct ldlm_lock *lock;
        int rc;
 
@@ -1246,8 +1270,8 @@ static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
  *
  * \retval a referenced lock or NULL.
  */
-static struct ldlm_lock *search_itree(struct ldlm_resource *res,
-                                     struct lock_match_data *data)
+struct ldlm_lock *search_itree(struct ldlm_resource *res,
+                              struct ldlm_match_data *data)
 {
        struct interval_node_extent ext = {
                .start     = data->lmd_policy->l_extent.start,
@@ -1255,6 +1279,8 @@ static struct ldlm_lock *search_itree(struct ldlm_resource *res,
        };
        int idx;
 
+       data->lmd_lock = NULL;
+
        for (idx = 0; idx < LCK_MODE_NUM; idx++) {
                struct ldlm_interval_tree *tree = &res->lr_itree[idx];
 
@@ -1266,9 +1292,13 @@ static struct ldlm_lock *search_itree(struct ldlm_resource *res,
 
                interval_search(tree->lit_root, &ext,
                                itree_overlap_cb, data);
+               if (data->lmd_lock)
+                       return data->lmd_lock;
        }
-       return data->lmd_lock;
+
+       return NULL;
 }
+EXPORT_SYMBOL(search_itree);
 
 
 /**
@@ -1280,16 +1310,19 @@ static struct ldlm_lock *search_itree(struct ldlm_resource *res,
  * \retval a referenced lock or NULL.
  */
 static struct ldlm_lock *search_queue(struct list_head *queue,
-                                     struct lock_match_data *data)
+                                     struct ldlm_match_data *data)
 {
        struct ldlm_lock *lock;
        int rc;
 
+       data->lmd_lock = NULL;
+
        list_for_each_entry(lock, queue, l_res_link) {
                rc = lock_matches(lock, data);
                if (rc == INTERVAL_ITER_STOP)
                        return data->lmd_lock;
        }
+
        return NULL;
 }
 
@@ -1365,24 +1398,28 @@ EXPORT_SYMBOL(ldlm_lock_allow_match);
  * keep caller code unchanged), the context failure will be discovered by
  * caller sometime later.
  */
-enum ldlm_mode ldlm_lock_match(struct ldlm_namespace *ns, __u64 flags,
-                              const struct ldlm_res_id *res_id,
-                              enum ldlm_type type,
-                              union ldlm_policy_data *policy,
-                              enum ldlm_mode mode,
-                              struct lustre_handle *lockh, int unref)
-{
-       struct lock_match_data data = {
-               .lmd_old        = NULL,
-               .lmd_lock       = NULL,
-               .lmd_mode       = &mode,
-               .lmd_policy     = policy,
-               .lmd_flags      = flags,
-               .lmd_unref      = unref,
+enum ldlm_mode ldlm_lock_match_with_skip(struct ldlm_namespace *ns,
+                                        __u64 flags, __u64 skip_flags,
+                                        const struct ldlm_res_id *res_id,
+                                        enum ldlm_type type,
+                                        union ldlm_policy_data *policy,
+                                        enum ldlm_mode mode,
+                                        struct lustre_handle *lockh, int unref)
+{
+       struct ldlm_match_data data = {
+               .lmd_old = NULL,
+               .lmd_lock = NULL,
+               .lmd_mode = &mode,
+               .lmd_policy = policy,
+               .lmd_flags = flags,
+               .lmd_skip_flags = skip_flags,
+               .lmd_unref = unref,
+               .lmd_has_ast_data = false,
        };
        struct ldlm_resource *res;
        struct ldlm_lock *lock;
-       int rc = 0;
+       int matched;
+
        ENTRY;
 
        if (ns == NULL) {
@@ -1403,98 +1440,76 @@ enum ldlm_mode ldlm_lock_match(struct ldlm_namespace *ns, __u64 flags,
 
        LDLM_RESOURCE_ADDREF(res);
        lock_res(res);
-
        if (res->lr_type == LDLM_EXTENT)
                lock = search_itree(res, &data);
        else
                lock = search_queue(&res->lr_granted, &data);
-       if (lock != NULL)
-               GOTO(out, rc = 1);
-       if (flags & LDLM_FL_BLOCK_GRANTED)
-               GOTO(out, rc = 0);
-       lock = search_queue(&res->lr_waiting, &data);
-       if (lock != NULL)
-               GOTO(out, rc = 1);
-
-        EXIT;
- out:
-        unlock_res(res);
-        LDLM_RESOURCE_DELREF(res);
-        ldlm_resource_putref(res);
+       if (!lock && !(flags & LDLM_FL_BLOCK_GRANTED))
+               lock = search_queue(&res->lr_waiting, &data);
+       matched = lock ? mode : 0;
+       unlock_res(res);
+       LDLM_RESOURCE_DELREF(res);
+       ldlm_resource_putref(res);
 
-        if (lock) {
-                ldlm_lock2handle(lock, lockh);
-                if ((flags & LDLM_FL_LVB_READY) &&
+       if (lock) {
+               ldlm_lock2handle(lock, lockh);
+               if ((flags & LDLM_FL_LVB_READY) &&
                    (!ldlm_is_lvb_ready(lock))) {
                        __u64 wait_flags = LDLM_FL_LVB_READY |
                                LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
-                        struct l_wait_info lwi;
-                        if (lock->l_completion_ast) {
-                                int err = lock->l_completion_ast(lock,
-                                                          LDLM_FL_WAIT_NOREPROC,
-                                                                 NULL);
-                                if (err) {
-                                        if (flags & LDLM_FL_TEST_LOCK)
-                                                LDLM_LOCK_RELEASE(lock);
-                                        else
-                                                ldlm_lock_decref_internal(lock,
-                                                                          mode);
-                                        rc = 0;
-                                        goto out2;
-                                }
-                        }
 
-                        lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
-                                               NULL, LWI_ON_SIGNAL_NOOP, NULL);
-
-                       /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
-                       l_wait_event(lock->l_waitq,
-                                    lock->l_flags & wait_flags,
-                                    &lwi);
-                       if (!ldlm_is_lvb_ready(lock)) {
-                                if (flags & LDLM_FL_TEST_LOCK)
-                                        LDLM_LOCK_RELEASE(lock);
-                                else
-                                        ldlm_lock_decref_internal(lock, mode);
-                                rc = 0;
-                        }
-                }
-        }
- out2:
-        if (rc) {
-               LDLM_DEBUG(lock, "matched (%llu %llu)",
-                           (type == LDLM_PLAIN || type == LDLM_IBITS) ?
-                                res_id->name[2] : policy->l_extent.start,
-                           (type == LDLM_PLAIN || type == LDLM_IBITS) ?
-                                res_id->name[3] : policy->l_extent.end);
-
-                /* check user's security context */
-                if (lock->l_conn_export &&
-                    sptlrpc_import_check_ctx(
-                                class_exp2cliimp(lock->l_conn_export))) {
-                        if (!(flags & LDLM_FL_TEST_LOCK))
-                                ldlm_lock_decref_internal(lock, mode);
-                        rc = 0;
-                }
+                       if (lock->l_completion_ast) {
+                               int err = lock->l_completion_ast(lock,
+                                                       LDLM_FL_WAIT_NOREPROC,
+                                                       NULL);
+                               if (err)
+                                       GOTO(out_fail_match, matched = 0);
+                       }
+
+                       wait_event_idle_timeout(
+                               lock->l_waitq,
+                               lock->l_flags & wait_flags,
+                               cfs_time_seconds(obd_timeout));
+
+                       if (!ldlm_is_lvb_ready(lock))
+                               GOTO(out_fail_match, matched = 0);
+               }
+
+               /* check user's security context */
+               if (lock->l_conn_export &&
+                   sptlrpc_import_check_ctx(
+                               class_exp2cliimp(lock->l_conn_export)))
+                       GOTO(out_fail_match, matched = 0);
 
-                if (flags & LDLM_FL_TEST_LOCK)
-                        LDLM_LOCK_RELEASE(lock);
+               LDLM_DEBUG(lock, "matched (%llu %llu)",
+                          (type == LDLM_PLAIN || type == LDLM_IBITS) ?
+                          res_id->name[2] : policy->l_extent.start,
+                          (type == LDLM_PLAIN || type == LDLM_IBITS) ?
+                          res_id->name[3] : policy->l_extent.end);
+
+out_fail_match:
+               if (flags & LDLM_FL_TEST_LOCK)
+                       LDLM_LOCK_RELEASE(lock);
+               else if (!matched)
+                       ldlm_lock_decref_internal(lock, mode);
+       }
 
-        } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
-                LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
+       /* less verbose for test-only */
+       if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
+               LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
                                  "%llu/%llu (%llu %llu)", ns,
-                                  type, mode, res_id->name[0], res_id->name[1],
-                                  (type == LDLM_PLAIN || type == LDLM_IBITS) ?
-                                        res_id->name[2] :policy->l_extent.start,
-                                  (type == LDLM_PLAIN || type == LDLM_IBITS) ?
-                                        res_id->name[3] : policy->l_extent.end);
-        }
+                                 type, mode, res_id->name[0], res_id->name[1],
+                                 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
+                                 res_id->name[2] : policy->l_extent.start,
+                                 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
+                                 res_id->name[3] : policy->l_extent.end);
+       }
        if (data.lmd_old != NULL)
                LDLM_LOCK_PUT(data.lmd_old);
 
-       return rc ? mode : 0;
+       return matched;
 }
-EXPORT_SYMBOL(ldlm_lock_match);
+EXPORT_SYMBOL(ldlm_lock_match_with_skip);
 
 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
                                           __u64 *bits)
@@ -1665,11 +1680,18 @@ struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
                lock->l_glimpse_ast = cbs->lcs_glimpse;
        }
 
-       lock->l_tree_node = NULL;
-       /* if this is the extent lock, allocate the interval tree node */
-       if (type == LDLM_EXTENT)
-               if (ldlm_interval_alloc(lock) == NULL)
-                       GOTO(out, rc = -ENOMEM);
+       switch (type) {
+       case LDLM_EXTENT:
+               rc = ldlm_extent_alloc_lock(lock);
+               break;
+       case LDLM_IBITS:
+               rc = ldlm_inodebits_alloc_lock(lock);
+               break;
+       default:
+               rc = 0;
+       }
+       if (rc)
+               GOTO(out, rc);
 
        if (lvb_len) {
                lock->l_lvb_len = lvb_len;
@@ -1696,11 +1718,12 @@ static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
 {
        struct ldlm_resource *res = lock->l_resource;
        enum ldlm_error rc = ELDLM_OK;
-       struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
+       LIST_HEAD(rpc_list);
        ldlm_processing_policy policy;
+
        ENTRY;
 
-       policy = ldlm_processing_policy_table[res->lr_type];
+       policy = ldlm_get_processing_policy(res);
 restart:
        policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
        if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
@@ -1727,7 +1750,8 @@ restart:
  * set, skip all the enqueueing and delegate lock processing to intent policy
  * function.
  */
-enum ldlm_error ldlm_lock_enqueue(struct ldlm_namespace *ns,
+enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
+                                 struct ldlm_namespace *ns,
                                  struct ldlm_lock **lockp,
                                  void *cookie, __u64 *flags)
 {
@@ -1736,13 +1760,16 @@ enum ldlm_error ldlm_lock_enqueue(struct ldlm_namespace *ns,
        int local = ns_is_client(ldlm_res_to_ns(res));
        enum ldlm_error rc = ELDLM_OK;
        struct ldlm_interval *node = NULL;
+#ifdef HAVE_SERVER_SUPPORT
+       bool reconstruct = false;
+#endif
        ENTRY;
 
         /* policies are not executed on the client or during replay */
         if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
             && !local && ns->ns_policy) {
-                rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
-                                   NULL);
+               rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
+                                  *flags, NULL);
                 if (rc == ELDLM_LOCK_REPLACED) {
                         /* The lock that was returned has already been granted,
                          * and placed into lockp.  If it's not the same as the
@@ -1755,7 +1782,7 @@ enum ldlm_error ldlm_lock_enqueue(struct ldlm_namespace *ns,
                         *flags |= LDLM_FL_LOCK_CHANGED;
                         RETURN(0);
                } else if (rc != ELDLM_OK &&
-                          lock->l_req_mode == lock->l_granted_mode) {
+                          ldlm_is_granted(lock)) {
                        LASSERT(*flags & LDLM_FL_RESENT);
                        /* It may happen that ns_policy returns an error in
                         * resend case, object may be unlinked or just some
@@ -1778,7 +1805,7 @@ enum ldlm_error ldlm_lock_enqueue(struct ldlm_namespace *ns,
                 * Take NO_TIMEOUT from the lock as it is inherited through
                 * LDLM_FL_INHERIT_MASK */
                *flags |= LDLM_FL_LOCK_CHANGED;
-               if (lock->l_req_mode != lock->l_granted_mode)
+               if (!ldlm_is_granted(lock))
                        *flags |= LDLM_FL_BLOCK_GRANTED;
                *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
                RETURN(ELDLM_OK);
@@ -1791,8 +1818,21 @@ enum ldlm_error ldlm_lock_enqueue(struct ldlm_namespace *ns,
        if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
                OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
 
-        lock_res_and_lock(lock);
-        if (local && lock->l_req_mode == lock->l_granted_mode) {
+#ifdef HAVE_SERVER_SUPPORT
+       reconstruct = !local && res->lr_type == LDLM_FLOCK &&
+                     !(*flags & LDLM_FL_TEST_LOCK);
+       if (reconstruct) {
+               rc = req_can_reconstruct(cookie, NULL);
+               if (rc != 0) {
+                       if (rc == 1)
+                               rc = 0;
+                       RETURN(rc);
+               }
+       }
+#endif
+
+       lock_res_and_lock(lock);
+       if (local && ldlm_is_granted(lock)) {
                 /* The server returned a blocked lock, but it was granted
                  * before we got a chance to actually enqueue it.  We don't
                  * need to do anything else. */
@@ -1864,6 +1904,16 @@ enum ldlm_error ldlm_lock_enqueue(struct ldlm_namespace *ns,
 
 out:
         unlock_res_and_lock(lock);
+
+#ifdef HAVE_SERVER_SUPPORT
+       if (reconstruct) {
+               struct ptlrpc_request *req = cookie;
+
+               tgt_mk_reply_data(NULL, NULL,
+                                 &req->rq_export->exp_target_data,
+                                 req, 0, NULL, false, 0);
+       }
+#endif
         if (node)
                 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
         return rc;
@@ -1878,19 +1928,21 @@ out:
  */
 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
                         struct list_head *work_list,
-                        enum ldlm_process_intention intention)
+                        enum ldlm_process_intention intention,
+                        struct ldlm_lock *hint)
 {
        struct list_head *tmp, *pos;
        ldlm_processing_policy policy;
        __u64 flags;
        int rc = LDLM_ITER_CONTINUE;
        enum ldlm_error err;
-       struct list_head bl_ast_list = LIST_HEAD_INIT(bl_ast_list);
+       LIST_HEAD(bl_ast_list);
+
        ENTRY;
 
        check_res_locked(res);
 
-       policy = ldlm_processing_policy_table[res->lr_type];
+       policy = ldlm_get_processing_policy(res);
        LASSERT(policy);
        LASSERT(intention == LDLM_PROCESS_RESCAN ||
                intention == LDLM_PROCESS_RECOVERY);
@@ -1898,7 +1950,7 @@ int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
 restart:
        list_for_each_safe(tmp, pos, queue) {
                struct ldlm_lock *pending;
-               struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
+               LIST_HEAD(rpc_list);
 
                pending = list_entry(tmp, struct ldlm_lock, l_res_link);
 
@@ -1989,7 +2041,7 @@ int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
                        RETURN(-EAGAIN);
 
                /* lock was granted while resource was unlocked. */
-               if (lock->l_granted_mode == lock->l_req_mode) {
+               if (ldlm_is_granted(lock)) {
                        /* bug 11300: if the lock has been granted,
                         * break earlier because otherwise, we will go
                         * to restart and ldlm_resource_unlink will be
@@ -2015,27 +2067,21 @@ int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
  */
 void ldlm_discard_bl_list(struct list_head *bl_list)
 {
-       struct list_head *tmp, *pos;
-        ENTRY;
+       struct ldlm_lock *lock, *tmp;
 
-       list_for_each_safe(pos, tmp, bl_list) {
-                struct ldlm_lock *lock =
-                       list_entry(pos, struct ldlm_lock, l_bl_ast);
+       ENTRY;
 
+       list_for_each_entry_safe(lock, tmp, bl_list, l_bl_ast) {
+               LASSERT(!list_empty(&lock->l_bl_ast));
                list_del_init(&lock->l_bl_ast);
-               LASSERT(ldlm_is_ast_sent(lock));
                ldlm_clear_ast_sent(lock);
                LASSERT(lock->l_bl_ast_run == 0);
-               LASSERT(lock->l_blocking_lock);
-               LDLM_LOCK_RELEASE(lock->l_blocking_lock);
-               lock->l_blocking_lock = NULL;
+               ldlm_clear_blocking_lock(lock);
                LDLM_LOCK_RELEASE(lock);
        }
        EXIT;
 }
 
-#endif
-
 /**
  * Process a call to blocking AST callback for a lock in ast_work list
  */
@@ -2043,9 +2089,11 @@ static int
 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
 {
        struct ldlm_cb_set_arg *arg = opaq;
-       struct ldlm_lock_desc   d;
-       int                     rc;
-       struct ldlm_lock       *lock;
+       struct ldlm_lock *lock;
+       struct ldlm_lock_desc d;
+       struct ldlm_bl_desc bld;
+       int rc;
+
        ENTRY;
 
        if (list_empty(arg->list))
@@ -2053,16 +2101,23 @@ ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
 
        lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
 
-       /* nobody should touch l_bl_ast */
+       /* nobody should touch l_bl_ast but some locks in the list may become
+        * granted after lock convert or COS downgrade, these locks should be
+        * just skipped here and removed from the list.
+        */
        lock_res_and_lock(lock);
        list_del_init(&lock->l_bl_ast);
 
-       LASSERT(ldlm_is_ast_sent(lock));
-       LASSERT(lock->l_bl_ast_run == 0);
-       LASSERT(lock->l_blocking_lock);
-       lock->l_bl_ast_run++;
-       unlock_res_and_lock(lock);
+       /* lock is not blocking lock anymore, but was kept in the list because
+        * it can managed only here.
+        */
+       if (!ldlm_is_ast_sent(lock)) {
+               unlock_res_and_lock(lock);
+               LDLM_LOCK_RELEASE(lock);
+               RETURN(0);
+       }
 
+       LASSERT(lock->l_blocking_lock);
        ldlm_lock2desc(lock->l_blocking_lock, &d);
        /* copy blocking lock ibits in cancel_bits as well,
         * new client may use them for lock convert and it is
@@ -2072,54 +2127,23 @@ ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
        d.l_policy_data.l_inodebits.cancel_bits =
                lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
 
-       rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
-       LDLM_LOCK_RELEASE(lock->l_blocking_lock);
-       lock->l_blocking_lock = NULL;
-       LDLM_LOCK_RELEASE(lock);
-
-       RETURN(rc);
-}
-
-/**
- * Process a call to completion AST callback for a lock in ast_work list
- */
-static int
-ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
-{
-       struct ldlm_cb_set_arg  *arg = opaq;
-       int                      rc = 0;
-       struct ldlm_lock        *lock;
-       ldlm_completion_callback completion_callback;
-       ENTRY;
-
-       if (list_empty(arg->list))
-               RETURN(-ENOENT);
-
-       lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
-
-       /* It's possible to receive a completion AST before we've set
-        * the l_completion_ast pointer: either because the AST arrived
-        * before the reply, or simply because there's a small race
-        * window between receiving the reply and finishing the local
-        * enqueue. (bug 842)
-        *
-        * This can't happen with the blocking_ast, however, because we
-        * will never call the local blocking_ast until we drop our
-        * reader/writer reference, which we won't do until we get the
-        * reply and finish enqueueing. */
+       /* Blocking lock is being destroyed here but some information about it
+        * may be needed inside l_blocking_ast() function below,
+        * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
+        */
+       bld.bl_same_client = lock->l_client_cookie ==
+                            lock->l_blocking_lock->l_client_cookie;
+       bld.bl_cos_incompat = ldlm_is_cos_incompat(lock->l_blocking_lock);
+       arg->bl_desc = &bld;
 
-       /* nobody should touch l_cp_ast */
-       lock_res_and_lock(lock);
-       list_del_init(&lock->l_cp_ast);
-       LASSERT(ldlm_is_cp_reqd(lock));
-       /* save l_completion_ast since it can be changed by
-        * mds_intent_policy(), see bug 14225 */
-       completion_callback = lock->l_completion_ast;
-       ldlm_clear_cp_reqd(lock);
+       LASSERT(ldlm_is_ast_sent(lock));
+       LASSERT(lock->l_bl_ast_run == 0);
+       lock->l_bl_ast_run++;
+       ldlm_clear_blocking_lock(lock);
        unlock_res_and_lock(lock);
 
-       if (completion_callback != NULL)
-               rc = completion_callback(lock, 0, (void *)arg);
+       rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
+
        LDLM_LOCK_RELEASE(lock);
 
        RETURN(rc);
@@ -2180,8 +2204,11 @@ int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
        arg->gl_interpret_data = gl_work->gl_interpret_data;
 
        /* invoke the actual glimpse callback */
-       if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
-               rc = 1;
+       rc = lock->l_glimpse_ast(lock, (void *)arg);
+       if (rc == 0)
+               rc = 1; /* update LVB if this is server lock */
+       else if (rc == -ELDLM_NO_LOCK_DATA)
+               ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
 
        LDLM_LOCK_RELEASE(lock);
        if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
@@ -2191,6 +2218,53 @@ int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
 
        RETURN(rc);
 }
+#endif
+
+/**
+ * Process a call to completion AST callback for a lock in ast_work list
+ */
+static int
+ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
+{
+       struct ldlm_cb_set_arg *arg = opaq;
+       struct ldlm_lock *lock;
+       ldlm_completion_callback completion_callback;
+       int rc = 0;
+
+       ENTRY;
+
+       if (list_empty(arg->list))
+               RETURN(-ENOENT);
+
+       lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
+
+       /* It's possible to receive a completion AST before we've set
+        * the l_completion_ast pointer: either because the AST arrived
+        * before the reply, or simply because there's a small race
+        * window between receiving the reply and finishing the local
+        * enqueue. (bug 842)
+        *
+        * This can't happen with the blocking_ast, however, because we
+        * will never call the local blocking_ast until we drop our
+        * reader/writer reference, which we won't do until we get the
+        * reply and finish enqueueing. */
+
+       /* nobody should touch l_cp_ast */
+       lock_res_and_lock(lock);
+       list_del_init(&lock->l_cp_ast);
+       LASSERT(ldlm_is_cp_reqd(lock));
+       /* save l_completion_ast since it can be changed by
+        * mds_intent_policy(), see bug 14225 */
+       completion_callback = lock->l_completion_ast;
+       ldlm_clear_cp_reqd(lock);
+       unlock_res_and_lock(lock);
+
+       if (completion_callback != NULL)
+               rc = completion_callback(lock, 0, (void *)arg);
+       LDLM_LOCK_RELEASE(lock);
+
+       RETURN(rc);
+}
 
 /**
  * Process list of locks in need of ASTs being sent.
@@ -2199,11 +2273,11 @@ int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
  * one.
  */
 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
-                      ldlm_desc_ast_t ast_type)
+                     ldlm_desc_ast_t ast_type)
 {
        struct ldlm_cb_set_arg *arg;
-       set_producer_func       work_ast_lock;
-       int                     rc;
+       set_producer_func work_ast_lock;
+       int rc;
 
        if (list_empty(rpc_list))
                RETURN(0);
@@ -2216,24 +2290,26 @@ int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
        arg->list = rpc_list;
 
        switch (ast_type) {
-               case LDLM_WORK_BL_AST:
-                       arg->type = LDLM_BL_CALLBACK;
-                       work_ast_lock = ldlm_work_bl_ast_lock;
-                       break;
-               case LDLM_WORK_CP_AST:
-                       arg->type = LDLM_CP_CALLBACK;
-                       work_ast_lock = ldlm_work_cp_ast_lock;
-                       break;
-               case LDLM_WORK_REVOKE_AST:
-                       arg->type = LDLM_BL_CALLBACK;
-                       work_ast_lock = ldlm_work_revoke_ast_lock;
-                       break;
-               case LDLM_WORK_GL_AST:
-                       arg->type = LDLM_GL_CALLBACK;
-                       work_ast_lock = ldlm_work_gl_ast_lock;
-                       break;
-               default:
-                       LBUG();
+       case LDLM_WORK_CP_AST:
+               arg->type = LDLM_CP_CALLBACK;
+               work_ast_lock = ldlm_work_cp_ast_lock;
+               break;
+#ifdef HAVE_SERVER_SUPPORT
+       case LDLM_WORK_BL_AST:
+               arg->type = LDLM_BL_CALLBACK;
+               work_ast_lock = ldlm_work_bl_ast_lock;
+               break;
+       case LDLM_WORK_REVOKE_AST:
+               arg->type = LDLM_BL_CALLBACK;
+               work_ast_lock = ldlm_work_revoke_ast_lock;
+               break;
+       case LDLM_WORK_GL_AST:
+               arg->type = LDLM_GL_CALLBACK;
+               work_ast_lock = ldlm_work_gl_ast_lock;
+               break;
+#endif
+       default:
+               LBUG();
        }
 
        /* We create a ptlrpc request set with flow control extension.
@@ -2245,7 +2321,7 @@ int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
        if (arg->set == NULL)
                GOTO(out, rc = -ENOMEM);
 
-       ptlrpc_set_wait(arg->set);
+       ptlrpc_set_wait(NULL, arg->set);
        ptlrpc_set_destroy(arg->set);
 
        rc = atomic_read(&arg->restart) ? -ERESTART : 0;
@@ -2264,20 +2340,22 @@ out:
  * if anything could be granted as a result of the cancellation.
  */
 static void __ldlm_reprocess_all(struct ldlm_resource *res,
-                                enum ldlm_process_intention intention)
+                                enum ldlm_process_intention intention,
+                                struct ldlm_lock *hint)
 {
-       struct list_head rpc_list;
+       LIST_HEAD(rpc_list);
 #ifdef HAVE_SERVER_SUPPORT
+       ldlm_reprocessing_policy reprocess;
        struct obd_device *obd;
-        int rc;
-        ENTRY;
+       int rc;
 
-       INIT_LIST_HEAD(&rpc_list);
-        /* Local lock trees don't get reprocessed. */
-        if (ns_is_client(ldlm_res_to_ns(res))) {
-                EXIT;
-                return;
-        }
+       ENTRY;
+
+       /* Local lock trees don't get reprocessed. */
+       if (ns_is_client(ldlm_res_to_ns(res))) {
+               EXIT;
+               return;
+       }
 
        /* Disable reprocess during lock replay stage but allow during
         * request replay stage.
@@ -2288,7 +2366,8 @@ static void __ldlm_reprocess_all(struct ldlm_resource *res,
                RETURN_EXIT;
 restart:
        lock_res(res);
-       ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list, intention);
+       reprocess = ldlm_get_reprocessing_policy(res);
+       reprocess(res, &res->lr_waiting, &rpc_list, intention, hint);
        unlock_res(res);
 
        rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
@@ -2298,21 +2377,20 @@ restart:
                goto restart;
        }
 #else
-        ENTRY;
+       ENTRY;
 
-       INIT_LIST_HEAD(&rpc_list);
-        if (!ns_is_client(ldlm_res_to_ns(res))) {
-                CERROR("This is client-side-only module, cannot handle "
-                       "LDLM_NAMESPACE_SERVER resource type lock.\n");
-                LBUG();
-        }
+       if (!ns_is_client(ldlm_res_to_ns(res))) {
+               CERROR("This is client-side-only module, cannot handle "
+                      "LDLM_NAMESPACE_SERVER resource type lock.\n");
+               LBUG();
+       }
 #endif
-        EXIT;
+       EXIT;
 }
 
-void ldlm_reprocess_all(struct ldlm_resource *res)
+void ldlm_reprocess_all(struct ldlm_resource *res, struct ldlm_lock *hint)
 {
-       __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN);
+       __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN, hint);
 }
 EXPORT_SYMBOL(ldlm_reprocess_all);
 
@@ -2322,7 +2400,7 @@ static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
        struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 
        /* This is only called once after recovery done. LU-8306. */
-       __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY);
+       __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY, NULL);
        return 0;
 }
 
@@ -2351,24 +2429,22 @@ void ldlm_cancel_callback(struct ldlm_lock *lock)
        if (!ldlm_is_cancel(lock)) {
                ldlm_set_cancel(lock);
                if (lock->l_blocking_ast) {
-                        unlock_res_and_lock(lock);
-                        lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
-                                             LDLM_CB_CANCELING);
-                        lock_res_and_lock(lock);
-                } else {
-                        LDLM_DEBUG(lock, "no blocking ast");
-                }
+                       unlock_res_and_lock(lock);
+                       lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
+                                            LDLM_CB_CANCELING);
+                       lock_res_and_lock(lock);
+               } else {
+                       LDLM_DEBUG(lock, "no blocking ast");
+               }
 
                /* only canceller can set bl_done bit */
                ldlm_set_bl_done(lock);
                wake_up_all(&lock->l_waitq);
        } else if (!ldlm_is_bl_done(lock)) {
-               struct l_wait_info lwi = { 0 };
-
                /* The lock is guaranteed to have been canceled once
                 * returning from this function. */
                unlock_res_and_lock(lock);
-               l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
+               wait_event_idle(lock->l_waitq, is_bl_done(lock));
                lock_res_and_lock(lock);
        }
 }
@@ -2422,8 +2498,8 @@ void ldlm_lock_cancel(struct ldlm_lock *lock)
         ldlm_resource_unlink_lock(lock);
         ldlm_lock_destroy_nolock(lock);
 
-        if (lock->l_granted_mode == lock->l_req_mode)
-                ldlm_pool_del(&ns->ns_pool, lock);
+       if (ldlm_is_granted(lock))
+               ldlm_pool_del(&ns->ns_pool, lock);
 
         /* Make sure we will not be called again for same lock what is possible
          * if not to zero out lock->l_granted_mode */
@@ -2455,6 +2531,7 @@ int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
 EXPORT_SYMBOL(ldlm_lock_set_data);
 
 struct export_cl_data {
+       const struct lu_env     *ecl_env;
        struct obd_export       *ecl_exp;
        int                     ecl_loop;
 };
@@ -2470,7 +2547,7 @@ static void ldlm_cancel_lock_for_export(struct obd_export *exp,
        ldlm_lvbo_update(res, lock, NULL, 1);
        ldlm_lock_cancel(lock);
        if (!exp->exp_obd->obd_stopping)
-               ldlm_reprocess_all(res);
+               ldlm_reprocess_all(res, lock);
        ldlm_resource_putref(res);
 
        ecl->ecl_loop++;
@@ -2507,10 +2584,17 @@ ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
  */
 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
 {
+       struct lu_env env;
        struct export_cl_data   ecl = {
                .ecl_exp        = exp,
                .ecl_loop       = 0,
        };
+       int rc;
+
+       rc = lu_env_init(&env, LCT_DT_THREAD);
+       if (rc)
+               RETURN(rc);
+       ecl.ecl_env = &env;
 
        while (!list_empty(&exp->exp_bl_list)) {
                struct ldlm_lock *lock;
@@ -2533,6 +2617,8 @@ int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
                LDLM_LOCK_RELEASE(lock);
        }
 
+       lu_env_fini(&env);
+
        CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
               "left on hash table %d.\n", exp, ecl.ecl_loop,
               atomic_read(&exp->exp_lock_hash->hs_count));
@@ -2547,10 +2633,16 @@ int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
  */
 int ldlm_export_cancel_locks(struct obd_export *exp)
 {
-       struct export_cl_data   ecl = {
-               .ecl_exp        = exp,
-               .ecl_loop       = 0,
-       };
+       struct export_cl_data ecl;
+       struct lu_env env;
+       int rc;
+
+       rc = lu_env_init(&env, LCT_DT_THREAD);
+       if (rc)
+               RETURN(rc);
+       ecl.ecl_env = &env;
+       ecl.ecl_exp = exp;
+       ecl.ecl_loop = 0;
 
        cfs_hash_for_each_empty(exp->exp_lock_hash,
                                ldlm_cancel_locks_for_export_cb, &ecl);
@@ -2564,26 +2656,35 @@ int ldlm_export_cancel_locks(struct obd_export *exp)
            exp->exp_obd->obd_stopping)
                ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
 
+       lu_env_fini(&env);
+
        return ecl.ecl_loop;
 }
 
 /**
- * Downgrade an PW/EX lock to COS mode.
+ * Downgrade an PW/EX lock to COS | CR mode.
  *
  * A lock mode convertion from PW/EX mode to less conflict mode. The
  * convertion may fail if lock was canceled before downgrade, but it doesn't
  * indicate any problem, because such lock has no reader or writer, and will
  * be released soon.
- * Used by Commit on Sharing (COS) code only for now.
+ *
+ * Used by Commit on Sharing (COS) code to force object changes commit in case
+ * of conflict. Converted lock is considered as new lock and all blocking AST
+ * things are cleared, so any pending or new blocked lock on that lock will
+ * cause new call to blocking_ast and force resource object commit.
+ *
+ * Also used by layout_change to replace EX lock to CR lock.
  *
  * \param lock A lock to convert
  * \param new_mode new lock mode
  */
 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
 {
+#ifdef HAVE_SERVER_SUPPORT
        ENTRY;
 
-       LASSERT(new_mode == LCK_COS);
+       LASSERT(new_mode == LCK_COS || new_mode == LCK_CR);
 
        lock_res_and_lock(lock);
 
@@ -2601,14 +2702,20 @@ void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
         * ldlm_grant_lock() called below.
         */
        ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
+
+       /* Consider downgraded lock as a new lock and clear all states
+        * related to a previous blocking AST processing.
+        */
+       ldlm_clear_blocking_data(lock);
+
        lock->l_req_mode = new_mode;
        ldlm_grant_lock(lock, NULL);
-
        unlock_res_and_lock(lock);
 
-       ldlm_reprocess_all(lock->l_resource);
+       ldlm_reprocess_all(lock->l_resource, lock);
 
        EXIT;
+#endif
 }
 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
 
@@ -2645,6 +2752,7 @@ void _ldlm_lock_debug(struct ldlm_lock *lock,
         va_list args;
         struct obd_export *exp = lock->l_export;
        struct ldlm_resource *resource = NULL;
+       struct va_format vaf;
         char *nid = "local";
 
        /* on server-side resource of lock doesn't change */
@@ -2658,6 +2766,8 @@ void _ldlm_lock_debug(struct ldlm_lock *lock,
        }
 
         va_start(args, fmt);
+       vaf.fmt = fmt;
+       vaf.va = &args;
 
         if (exp && exp->exp_connection) {
                nid = obd_export_nid2str(exp);
@@ -2667,111 +2777,110 @@ void _ldlm_lock_debug(struct ldlm_lock *lock,
         }
 
         if (resource == NULL) {
-                libcfs_debug_vmsg2(msgdata, fmt, args,
-                      " ns: \?\? lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
-                      "res: \?\? rrc=\?\? type: \?\?\? flags: %#llx nid: %s "
-                      "remote: %#llx expref: %d pid: %u timeout: %lld "
-                      "lvb_type: %d\n",
-                       lock,
-                      lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
-                       lock->l_readers, lock->l_writers,
-                       ldlm_lockname[lock->l_granted_mode],
-                       ldlm_lockname[lock->l_req_mode],
-                       lock->l_flags, nid, lock->l_remote_handle.cookie,
-                      exp ? atomic_read(&exp->exp_refcount) : -99,
-                       lock->l_pid, lock->l_callback_timeout, lock->l_lvb_type);
+               libcfs_debug_msg(msgdata,
+                                "%pV ns: \?\? lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: \?\? rrc=\?\? type: \?\?\? flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
+                                &vaf,
+                                lock,
+                                lock->l_handle.h_cookie,
+                                refcount_read(&lock->l_handle.h_ref),
+                                lock->l_readers, lock->l_writers,
+                                ldlm_lockname[lock->l_granted_mode],
+                                ldlm_lockname[lock->l_req_mode],
+                                lock->l_flags, nid,
+                                lock->l_remote_handle.cookie,
+                                exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
+                                lock->l_pid, lock->l_callback_timeout,
+                                lock->l_lvb_type);
                 va_end(args);
                 return;
         }
 
        switch (resource->lr_type) {
        case LDLM_EXTENT:
-               libcfs_debug_vmsg2(msgdata, fmt, args,
-                       " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
-                       "res: "DLDLMRES" rrc: %d type: %s [%llu->%llu] "
-                       "(req %llu->%llu) flags: %#llx nid: %s remote: "
-                       "%#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
-                       ldlm_lock_to_ns_name(lock), lock,
-                       lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
-                       lock->l_readers, lock->l_writers,
-                       ldlm_lockname[lock->l_granted_mode],
-                       ldlm_lockname[lock->l_req_mode],
-                       PLDLMRES(resource),
-                       atomic_read(&resource->lr_refcount),
-                       ldlm_typename[resource->lr_type],
-                       lock->l_policy_data.l_extent.start,
-                       lock->l_policy_data.l_extent.end,
-                       lock->l_req_extent.start, lock->l_req_extent.end,
-                       lock->l_flags, nid, lock->l_remote_handle.cookie,
-                       exp ? atomic_read(&exp->exp_refcount) : -99,
-                       lock->l_pid, lock->l_callback_timeout,
-                       lock->l_lvb_type);
+               libcfs_debug_msg(msgdata,
+                                "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s [%llu->%llu] (req %llu->%llu) flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
+                                &vaf,
+                                ldlm_lock_to_ns_name(lock), lock,
+                                lock->l_handle.h_cookie,
+                                refcount_read(&lock->l_handle.h_ref),
+                                lock->l_readers, lock->l_writers,
+                                ldlm_lockname[lock->l_granted_mode],
+                                ldlm_lockname[lock->l_req_mode],
+                                PLDLMRES(resource),
+                                atomic_read(&resource->lr_refcount),
+                                ldlm_typename[resource->lr_type],
+                                lock->l_policy_data.l_extent.start,
+                                lock->l_policy_data.l_extent.end,
+                                lock->l_req_extent.start, lock->l_req_extent.end,
+                                lock->l_flags, nid,
+                                lock->l_remote_handle.cookie,
+                                exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
+                                lock->l_pid, lock->l_callback_timeout,
+                                lock->l_lvb_type);
                break;
 
        case LDLM_FLOCK:
-               libcfs_debug_vmsg2(msgdata, fmt, args,
-                       " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
-                       "res: "DLDLMRES" rrc: %d type: %s pid: %d "
-                       "[%llu->%llu] flags: %#llx nid: %s "
-                       "remote: %#llx expref: %d pid: %u timeout: %lld\n",
-                       ldlm_lock_to_ns_name(lock), lock,
-                       lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
-                       lock->l_readers, lock->l_writers,
-                       ldlm_lockname[lock->l_granted_mode],
-                       ldlm_lockname[lock->l_req_mode],
-                       PLDLMRES(resource),
-                       atomic_read(&resource->lr_refcount),
-                       ldlm_typename[resource->lr_type],
-                       lock->l_policy_data.l_flock.pid,
-                       lock->l_policy_data.l_flock.start,
-                       lock->l_policy_data.l_flock.end,
-                       lock->l_flags, nid, lock->l_remote_handle.cookie,
-                       exp ? atomic_read(&exp->exp_refcount) : -99,
-                       lock->l_pid, lock->l_callback_timeout);
+               libcfs_debug_msg(msgdata,
+                                "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s pid: %d [%llu->%llu] flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld\n",
+                                &vaf,
+                                ldlm_lock_to_ns_name(lock), lock,
+                                lock->l_handle.h_cookie,
+                                refcount_read(&lock->l_handle.h_ref),
+                                lock->l_readers, lock->l_writers,
+                                ldlm_lockname[lock->l_granted_mode],
+                                ldlm_lockname[lock->l_req_mode],
+                                PLDLMRES(resource),
+                                atomic_read(&resource->lr_refcount),
+                                ldlm_typename[resource->lr_type],
+                                lock->l_policy_data.l_flock.pid,
+                                lock->l_policy_data.l_flock.start,
+                                lock->l_policy_data.l_flock.end,
+                                lock->l_flags, nid,
+                                lock->l_remote_handle.cookie,
+                                exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
+                                lock->l_pid, lock->l_callback_timeout);
                break;
 
        case LDLM_IBITS:
-               libcfs_debug_vmsg2(msgdata, fmt, args,
-                       " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
-                       "res: "DLDLMRES" bits %#llx/%#llx rrc: %d type: %s "
-                       "flags: %#llx nid: %s remote: %#llx expref: %d "
-                       "pid: %u timeout: %lld lvb_type: %d\n",
-                       ldlm_lock_to_ns_name(lock),
-                       lock, lock->l_handle.h_cookie,
-                       atomic_read(&lock->l_refc),
-                       lock->l_readers, lock->l_writers,
-                       ldlm_lockname[lock->l_granted_mode],
-                       ldlm_lockname[lock->l_req_mode],
-                       PLDLMRES(resource),
-                       lock->l_policy_data.l_inodebits.bits,
-                       lock->l_policy_data.l_inodebits.try_bits,
-                       atomic_read(&resource->lr_refcount),
-                       ldlm_typename[resource->lr_type],
-                       lock->l_flags, nid, lock->l_remote_handle.cookie,
-                       exp ? atomic_read(&exp->exp_refcount) : -99,
-                       lock->l_pid, lock->l_callback_timeout,
-                       lock->l_lvb_type);
+               libcfs_debug_msg(msgdata,
+                                "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " bits %#llx/%#llx rrc: %d type: %s flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
+                                &vaf,
+                                ldlm_lock_to_ns_name(lock),
+                                lock, lock->l_handle.h_cookie,
+                                refcount_read(&lock->l_handle.h_ref),
+                                lock->l_readers, lock->l_writers,
+                                ldlm_lockname[lock->l_granted_mode],
+                                ldlm_lockname[lock->l_req_mode],
+                                PLDLMRES(resource),
+                                lock->l_policy_data.l_inodebits.bits,
+                                lock->l_policy_data.l_inodebits.try_bits,
+                                atomic_read(&resource->lr_refcount),
+                                ldlm_typename[resource->lr_type],
+                                lock->l_flags, nid,
+                                lock->l_remote_handle.cookie,
+                                exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
+                                lock->l_pid, lock->l_callback_timeout,
+                                lock->l_lvb_type);
                break;
 
        default:
-               libcfs_debug_vmsg2(msgdata, fmt, args,
-                       " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
-                       "res: "DLDLMRES" rrc: %d type: %s flags: %#llx "
-                       "nid: %s remote: %#llx expref: %d pid: %u "
-                       "timeout: %lld lvb_type: %d\n",
-                       ldlm_lock_to_ns_name(lock),
-                       lock, lock->l_handle.h_cookie,
-                       atomic_read(&lock->l_refc),
-                       lock->l_readers, lock->l_writers,
-                       ldlm_lockname[lock->l_granted_mode],
-                       ldlm_lockname[lock->l_req_mode],
-                       PLDLMRES(resource),
-                       atomic_read(&resource->lr_refcount),
-                       ldlm_typename[resource->lr_type],
-                       lock->l_flags, nid, lock->l_remote_handle.cookie,
-                       exp ? atomic_read(&exp->exp_refcount) : -99,
-                       lock->l_pid, lock->l_callback_timeout,
-                       lock->l_lvb_type);
+               libcfs_debug_msg(msgdata,
+                                "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
+                                &vaf,
+                                ldlm_lock_to_ns_name(lock),
+                                lock, lock->l_handle.h_cookie,
+                                refcount_read(&lock->l_handle.h_ref),
+                                lock->l_readers, lock->l_writers,
+                                ldlm_lockname[lock->l_granted_mode],
+                                ldlm_lockname[lock->l_req_mode],
+                                PLDLMRES(resource),
+                                atomic_read(&resource->lr_refcount),
+                                ldlm_typename[resource->lr_type],
+                                lock->l_flags, nid,
+                                lock->l_remote_handle.cookie,
+                                exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
+                                lock->l_pid, lock->l_callback_timeout,
+                                lock->l_lvb_type);
                break;
        }
        va_end(args);