4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_lock.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <libcfs/libcfs.h>
42 #include <lustre_swab.h>
43 #include <obd_class.h>
45 #include "ldlm_internal.h"
47 struct kmem_cache *ldlm_glimpse_work_kmem;
48 EXPORT_SYMBOL(ldlm_glimpse_work_kmem);
51 char *ldlm_lockname[] = {
59 [LCK_GROUP] = "GROUP",
62 EXPORT_SYMBOL(ldlm_lockname);
64 char *ldlm_typename[] = {
66 [LDLM_EXTENT] = "EXT",
71 static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = {
72 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_wire_to_local,
73 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_wire_to_local,
74 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_wire_to_local,
75 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_wire_to_local,
78 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
79 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_local_to_wire,
80 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_local_to_wire,
81 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_local_to_wire,
82 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_local_to_wire,
86 * Converts lock policy from local format to on the wire lock_desc format
88 void ldlm_convert_policy_to_wire(enum ldlm_type type,
89 const union ldlm_policy_data *lpolicy,
90 union ldlm_wire_policy_data *wpolicy)
92 ldlm_policy_local_to_wire_t convert;
94 convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
96 convert(lpolicy, wpolicy);
100 * Converts lock policy from on the wire lock_desc format to local format
102 void ldlm_convert_policy_to_local(struct obd_export *exp, enum ldlm_type type,
103 const union ldlm_wire_policy_data *wpolicy,
104 union ldlm_policy_data *lpolicy)
106 ldlm_policy_wire_to_local_t convert;
108 convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE];
110 convert(wpolicy, lpolicy);
113 const char *ldlm_it2str(enum ldlm_intent_flags it)
120 case (IT_OPEN | IT_CREAT):
133 CERROR("Unknown intent 0x%08x\n", it);
137 EXPORT_SYMBOL(ldlm_it2str);
139 #ifdef HAVE_SERVER_SUPPORT
140 static ldlm_processing_policy ldlm_processing_policy_table[] = {
141 [LDLM_PLAIN] = ldlm_process_plain_lock,
142 [LDLM_EXTENT] = ldlm_process_extent_lock,
143 [LDLM_FLOCK] = ldlm_process_flock_lock,
144 [LDLM_IBITS] = ldlm_process_inodebits_lock,
147 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
149 return ldlm_processing_policy_table[res->lr_type];
151 EXPORT_SYMBOL(ldlm_get_processing_policy);
153 static ldlm_reprocessing_policy ldlm_reprocessing_policy_table[] = {
154 [LDLM_PLAIN] = ldlm_reprocess_queue,
155 [LDLM_EXTENT] = ldlm_reprocess_queue,
156 [LDLM_FLOCK] = ldlm_reprocess_queue,
157 [LDLM_IBITS] = ldlm_reprocess_inodebits_queue,
160 ldlm_reprocessing_policy ldlm_get_reprocessing_policy(struct ldlm_resource *res)
162 return ldlm_reprocessing_policy_table[res->lr_type];
165 #endif /* HAVE_SERVER_SUPPORT */
167 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
171 EXPORT_SYMBOL(ldlm_register_intent);
174 * REFCOUNTED LOCK OBJECTS
179 * Get a reference on a lock.
181 * Lock refcounts, during creation:
182 * - one special one for allocation, dec'd only once in destroy
183 * - one for being a lock that's in-use
184 * - one for the addref associated with a new lock
186 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
188 refcount_inc(&lock->l_handle.h_ref);
191 EXPORT_SYMBOL(ldlm_lock_get);
193 static void lock_handle_free(struct rcu_head *rcu)
195 struct ldlm_lock *lock = container_of(rcu, struct ldlm_lock,
198 OBD_FREE_PRE(lock, sizeof(*lock), "slab-freed");
199 kmem_cache_free(ldlm_lock_slab, lock);
203 * Release lock reference.
205 * Also frees the lock if it was last reference.
207 void ldlm_lock_put(struct ldlm_lock *lock)
211 LASSERT(lock->l_resource != LP_POISON);
212 LASSERT(refcount_read(&lock->l_handle.h_ref) > 0);
213 if (refcount_dec_and_test(&lock->l_handle.h_ref)) {
214 struct ldlm_resource *res;
217 "final lock_put on destroyed lock, freeing it.");
219 res = lock->l_resource;
220 LASSERT(ldlm_is_destroyed(lock));
221 LASSERT(list_empty(&lock->l_exp_list));
222 LASSERT(list_empty(&lock->l_res_link));
223 LASSERT(list_empty(&lock->l_pending_chain));
225 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
227 lu_ref_del(&res->lr_reference, "lock", lock);
228 if (lock->l_export) {
229 class_export_lock_put(lock->l_export, lock);
230 lock->l_export = NULL;
233 if (lock->l_lvb_data != NULL)
234 OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
236 if (res->lr_type == LDLM_EXTENT) {
237 ldlm_interval_free(ldlm_interval_detach(lock));
238 } else if (res->lr_type == LDLM_IBITS) {
239 if (lock->l_ibits_node != NULL)
240 OBD_SLAB_FREE_PTR(lock->l_ibits_node,
241 ldlm_inodebits_slab);
243 ldlm_resource_putref(res);
244 lock->l_resource = NULL;
245 lu_ref_fini(&lock->l_reference);
246 call_rcu(&lock->l_handle.h_rcu, lock_handle_free);
251 EXPORT_SYMBOL(ldlm_lock_put);
254 * Removes LDLM lock \a lock from LRU. Assumes LRU is already locked.
256 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
259 if (!list_empty(&lock->l_lru)) {
260 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
262 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
263 if (ns->ns_last_pos == &lock->l_lru)
264 ns->ns_last_pos = lock->l_lru.prev;
265 list_del_init(&lock->l_lru);
266 LASSERT(ns->ns_nr_unused > 0);
274 * Removes LDLM lock \a lock from LRU. Obtains the LRU lock first.
276 * If \a last_use is non-zero, it will remove the lock from LRU only if
277 * it matches lock's l_last_used.
279 * \retval 0 if \a last_use is set, the lock is not in LRU list or \a last_use
280 * doesn't match lock's l_last_used;
281 * otherwise, the lock hasn't been in the LRU list.
282 * \retval 1 the lock was in LRU list and removed.
284 int ldlm_lock_remove_from_lru_check(struct ldlm_lock *lock, ktime_t last_use)
286 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
290 if (ldlm_is_ns_srv(lock)) {
291 LASSERT(list_empty(&lock->l_lru));
295 spin_lock(&ns->ns_lock);
296 if (!ktime_compare(last_use, ktime_set(0, 0)) ||
297 !ktime_compare(last_use, lock->l_last_used))
298 rc = ldlm_lock_remove_from_lru_nolock(lock);
299 spin_unlock(&ns->ns_lock);
305 * Adds LDLM lock \a lock to namespace LRU. Assumes LRU is already locked.
307 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
309 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
311 lock->l_last_used = ktime_get();
312 LASSERT(list_empty(&lock->l_lru));
313 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
314 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
315 LASSERT(ns->ns_nr_unused >= 0);
320 * Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks
323 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
325 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
328 spin_lock(&ns->ns_lock);
329 ldlm_lock_add_to_lru_nolock(lock);
330 spin_unlock(&ns->ns_lock);
335 * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
336 * the LRU. Performs necessary LRU locking
338 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
340 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
343 if (ldlm_is_ns_srv(lock)) {
344 LASSERT(list_empty(&lock->l_lru));
349 spin_lock(&ns->ns_lock);
350 if (!list_empty(&lock->l_lru)) {
351 ldlm_lock_remove_from_lru_nolock(lock);
352 ldlm_lock_add_to_lru_nolock(lock);
354 spin_unlock(&ns->ns_lock);
359 * Helper to destroy a locked lock.
361 * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
362 * Must be called with l_lock and lr_lock held.
364 * Does not actually free the lock data, but rather marks the lock as
365 * destroyed by setting l_destroyed field in the lock to 1. Destroys a
366 * handle->lock association too, so that the lock can no longer be found
367 * and removes the lock from LRU list. Actual lock freeing occurs when
368 * last lock reference goes away.
370 * Original comment (of some historical value):
371 * This used to have a 'strict' flag, which recovery would use to mark an
372 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
373 * shall explain why it's gone: with the new hash table scheme, once you call
374 * ldlm_lock_destroy, you can never drop your final references on this lock.
375 * Because it's not in the hash table anymore. -phil
377 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
381 if (lock->l_readers || lock->l_writers) {
382 LDLM_ERROR(lock, "lock still has references");
386 if (!list_empty(&lock->l_res_link)) {
387 LDLM_ERROR(lock, "lock still on resource");
391 if (ldlm_is_destroyed(lock)) {
392 LASSERT(list_empty(&lock->l_lru));
396 ldlm_set_destroyed(lock);
398 if (lock->l_export && lock->l_export->exp_lock_hash) {
399 /* NB: it's safe to call cfs_hash_del() even lock isn't
400 * in exp_lock_hash. */
401 /* In the function below, .hs_keycmp resolves to
402 * ldlm_export_lock_keycmp() */
403 /* coverity[overrun-buffer-val] */
404 cfs_hash_del(lock->l_export->exp_lock_hash,
405 &lock->l_remote_handle, &lock->l_exp_hash);
408 ldlm_lock_remove_from_lru(lock);
409 class_handle_unhash(&lock->l_handle);
416 * Destroys a LDLM lock \a lock. Performs necessary locking first.
418 void ldlm_lock_destroy(struct ldlm_lock *lock)
422 lock_res_and_lock(lock);
423 first = ldlm_lock_destroy_internal(lock);
424 unlock_res_and_lock(lock);
426 /* drop reference from hashtable only for first destroy */
428 lu_ref_del(&lock->l_reference, "hash", lock);
429 LDLM_LOCK_RELEASE(lock);
435 * Destroys a LDLM lock \a lock that is already locked.
437 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
441 first = ldlm_lock_destroy_internal(lock);
442 /* drop reference from hashtable only for first destroy */
444 lu_ref_del(&lock->l_reference, "hash", lock);
445 LDLM_LOCK_RELEASE(lock);
450 static struct portals_handle_ops lock_handle_ops = {
456 * Allocate and initialize new lock structure.
458 * usage: pass in a resource on which you have done ldlm_resource_get
459 * new lock will take over the refcount.
460 * returns: lock with refcount 2 - one for current caller and one for remote
462 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
464 struct ldlm_lock *lock;
467 if (resource == NULL)
470 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, GFP_NOFS);
474 spin_lock_init(&lock->l_lock);
475 lock->l_resource = resource;
476 lu_ref_add(&resource->lr_reference, "lock", lock);
478 refcount_set(&lock->l_handle.h_ref, 2);
479 INIT_LIST_HEAD(&lock->l_res_link);
480 INIT_LIST_HEAD(&lock->l_lru);
481 INIT_LIST_HEAD(&lock->l_pending_chain);
482 INIT_LIST_HEAD(&lock->l_bl_ast);
483 INIT_LIST_HEAD(&lock->l_cp_ast);
484 INIT_LIST_HEAD(&lock->l_rk_ast);
485 init_waitqueue_head(&lock->l_waitq);
486 lock->l_blocking_lock = NULL;
487 INIT_LIST_HEAD(&lock->l_sl_mode);
488 INIT_LIST_HEAD(&lock->l_sl_policy);
489 INIT_HLIST_NODE(&lock->l_exp_hash);
490 INIT_HLIST_NODE(&lock->l_exp_flock_hash);
492 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
494 INIT_LIST_HEAD_RCU(&lock->l_handle.h_link);
495 class_handle_hash(&lock->l_handle, &lock_handle_ops);
497 lu_ref_init(&lock->l_reference);
498 lu_ref_add(&lock->l_reference, "hash", lock);
499 lock->l_callback_timeout = 0;
500 lock->l_activity = 0;
502 #if LUSTRE_TRACKS_LOCK_EXP_REFS
503 INIT_LIST_HEAD(&lock->l_exp_refs_link);
504 lock->l_exp_refs_nr = 0;
505 lock->l_exp_refs_target = NULL;
507 INIT_LIST_HEAD(&lock->l_exp_list);
513 * Moves LDLM lock \a lock to another resource.
514 * This is used on client when server returns some other lock than requested
515 * (typically as a result of intent operation)
517 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
518 const struct ldlm_res_id *new_resid)
520 struct ldlm_resource *oldres = lock->l_resource;
521 struct ldlm_resource *newres;
525 LASSERT(ns_is_client(ns));
527 lock_res_and_lock(lock);
528 if (memcmp(new_resid, &lock->l_resource->lr_name,
529 sizeof(lock->l_resource->lr_name)) == 0) {
531 unlock_res_and_lock(lock);
535 LASSERT(new_resid->name[0] != 0);
537 /* This function assumes that the lock isn't on any lists */
538 LASSERT(list_empty(&lock->l_res_link));
540 type = oldres->lr_type;
541 unlock_res_and_lock(lock);
543 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
545 RETURN(PTR_ERR(newres));
547 lu_ref_add(&newres->lr_reference, "lock", lock);
549 * To flip the lock from the old to the new resource, lock, oldres and
550 * newres have to be locked. Resource spin-locks are nested within
551 * lock->l_lock, and are taken in the memory address order to avoid
554 spin_lock(&lock->l_lock);
555 oldres = lock->l_resource;
556 if (oldres < newres) {
558 lock_res_nested(newres, LRT_NEW);
561 lock_res_nested(oldres, LRT_NEW);
563 LASSERT(memcmp(new_resid, &oldres->lr_name,
564 sizeof oldres->lr_name) != 0);
565 lock->l_resource = newres;
567 unlock_res_and_lock(lock);
569 /* ...and the flowers are still standing! */
570 lu_ref_del(&oldres->lr_reference, "lock", lock);
571 ldlm_resource_putref(oldres);
576 /** \defgroup ldlm_handles LDLM HANDLES
577 * Ways to get hold of locks without any addresses.
582 * Fills in handle for LDLM lock \a lock into supplied \a lockh
583 * Does not take any references.
585 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
587 lockh->cookie = lock->l_handle.h_cookie;
589 EXPORT_SYMBOL(ldlm_lock2handle);
592 * Obtain a lock reference by handle.
594 * if \a flags: atomically get the lock and set the flags.
595 * Return NULL if flag already set
597 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
600 struct ldlm_lock *lock;
605 if (!lustre_handle_is_used(handle))
608 lock = class_handle2object(handle->cookie, &lock_handle_ops);
613 if (lock->l_export != NULL && lock->l_export->exp_failed) {
614 CDEBUG(D_INFO, "lock export failed: lock %p, exp %p\n",
615 lock, lock->l_export);
620 /* It's unlikely but possible that someone marked the lock as
621 * destroyed after we did handle2object on it */
622 if ((flags == 0) && !ldlm_is_destroyed(lock)) {
623 lu_ref_add(&lock->l_reference, "handle", current);
627 lock_res_and_lock(lock);
629 LASSERT(lock->l_resource != NULL);
631 lu_ref_add_atomic(&lock->l_reference, "handle", current);
632 if (unlikely(ldlm_is_destroyed(lock))) {
633 unlock_res_and_lock(lock);
634 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
639 /* If we're setting flags, make sure none of them are already set. */
641 if ((lock->l_flags & flags) != 0) {
642 unlock_res_and_lock(lock);
647 lock->l_flags |= flags;
650 unlock_res_and_lock(lock);
653 EXPORT_SYMBOL(__ldlm_handle2lock);
654 /** @} ldlm_handles */
657 * Fill in "on the wire" representation for given LDLM lock into supplied
658 * lock descriptor \a desc structure.
660 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
662 ldlm_res2desc(lock->l_resource, &desc->l_resource);
663 desc->l_req_mode = lock->l_req_mode;
664 desc->l_granted_mode = lock->l_granted_mode;
665 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
666 &lock->l_policy_data,
667 &desc->l_policy_data);
671 * Add a lock to list of conflicting locks to send AST to.
673 * Only add if we have not sent a blocking AST to the lock yet.
675 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
676 struct list_head *work_list)
678 if (!ldlm_is_ast_sent(lock)) {
679 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
680 ldlm_set_ast_sent(lock);
681 /* If the enqueuing client said so, tell the AST recipient to
682 * discard dirty data, rather than writing back. */
683 if (ldlm_is_ast_discard_data(new))
684 ldlm_set_discard_data(lock);
686 /* Lock can be converted from a blocking state back to granted
687 * after lock convert or COS downgrade but still be in an
688 * older bl_list because it is controlled only by
689 * ldlm_work_bl_ast_lock(), let it be processed there.
691 if (list_empty(&lock->l_bl_ast)) {
692 list_add(&lock->l_bl_ast, work_list);
695 LASSERT(lock->l_blocking_lock == NULL);
696 lock->l_blocking_lock = LDLM_LOCK_GET(new);
701 * Add a lock to list of just granted locks to send completion AST to.
703 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
704 struct list_head *work_list)
706 if (!ldlm_is_cp_reqd(lock)) {
707 ldlm_set_cp_reqd(lock);
708 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
709 LASSERT(list_empty(&lock->l_cp_ast));
710 list_add(&lock->l_cp_ast, work_list);
716 * Aggregator function to add AST work items into a list. Determines
717 * what sort of an AST work needs to be done and calls the proper
719 * Must be called with lr_lock held.
721 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
722 struct list_head *work_list)
725 check_res_locked(lock->l_resource);
727 ldlm_add_bl_work_item(lock, new, work_list);
729 ldlm_add_cp_work_item(lock, work_list);
734 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
735 * r/w reference type is determined by \a mode
736 * Calls ldlm_lock_addref_internal.
738 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
740 struct ldlm_lock *lock;
742 lock = ldlm_handle2lock(lockh);
743 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
744 ldlm_lock_addref_internal(lock, mode);
747 EXPORT_SYMBOL(ldlm_lock_addref);
751 * Add specified reader/writer reference to LDLM lock \a lock.
752 * r/w reference type is determined by \a mode
753 * Removes lock from LRU if it is there.
754 * Assumes the LDLM lock is already locked.
756 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
759 ldlm_lock_remove_from_lru(lock);
760 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
762 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
764 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
766 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
769 lu_ref_add_atomic(&lock->l_reference, "user", lock);
770 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
774 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
775 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
777 * \retval 0 success, lock was addref-ed
779 * \retval -EAGAIN lock is being canceled.
781 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
783 struct ldlm_lock *lock;
787 lock = ldlm_handle2lock(lockh);
789 lock_res_and_lock(lock);
790 if (lock->l_readers != 0 || lock->l_writers != 0 ||
791 !ldlm_is_cbpending(lock)) {
792 ldlm_lock_addref_internal_nolock(lock, mode);
795 unlock_res_and_lock(lock);
800 EXPORT_SYMBOL(ldlm_lock_addref_try);
803 * Add specified reader/writer reference to LDLM lock \a lock.
804 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
805 * Only called for local locks.
807 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
809 lock_res_and_lock(lock);
810 ldlm_lock_addref_internal_nolock(lock, mode);
811 unlock_res_and_lock(lock);
815 * Removes reader/writer reference for LDLM lock \a lock.
816 * Assumes LDLM lock is already locked.
817 * only called in ldlm_flock_destroy and for local locks.
818 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
819 * that cannot be placed in LRU.
821 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
824 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
825 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
826 LASSERT(lock->l_readers > 0);
827 lu_ref_del(&lock->l_reference, "reader", lock);
830 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
831 LASSERT(lock->l_writers > 0);
832 lu_ref_del(&lock->l_reference, "writer", lock);
836 lu_ref_del(&lock->l_reference, "user", lock);
837 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
841 * Removes reader/writer reference for LDLM lock \a lock.
842 * Locks LDLM lock first.
843 * If the lock is determined to be client lock on a client and r/w refcount
844 * drops to zero and the lock is not blocked, the lock is added to LRU lock
846 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
848 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
850 struct ldlm_namespace *ns;
853 lock_res_and_lock(lock);
855 ns = ldlm_lock_to_ns(lock);
857 ldlm_lock_decref_internal_nolock(lock, mode);
859 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
860 !lock->l_readers && !lock->l_writers) {
861 /* If this is a local lock on a server namespace and this was
862 * the last reference, cancel the lock.
864 * Group locks are special:
865 * They must not go in LRU, but they are not called back
866 * like non-group locks, instead they are manually released.
867 * They have an l_writers reference which they keep until
868 * they are manually released, so we remove them when they have
869 * no more reader or writer references. - LU-6368 */
870 ldlm_set_cbpending(lock);
873 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
874 /* If we received a blocked AST and this was the last reference,
875 * run the callback. */
876 if (ldlm_is_ns_srv(lock) && lock->l_export)
877 CERROR("FL_CBPENDING set on non-local lock--just a "
880 LDLM_DEBUG(lock, "final decref done on cbpending lock");
882 LDLM_LOCK_GET(lock); /* dropped by bl thread */
883 ldlm_lock_remove_from_lru(lock);
884 unlock_res_and_lock(lock);
886 if (ldlm_is_fail_loc(lock))
887 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
889 if (ldlm_is_atomic_cb(lock) ||
890 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
891 ldlm_handle_bl_callback(ns, NULL, lock);
892 } else if (ns_is_client(ns) &&
893 !lock->l_readers && !lock->l_writers &&
894 !ldlm_is_no_lru(lock) &&
895 !ldlm_is_bl_ast(lock) &&
896 !ldlm_is_converting(lock)) {
898 LDLM_DEBUG(lock, "add lock into lru list");
900 /* If this is a client-side namespace and this was the last
901 * reference, put it on the LRU. */
902 ldlm_lock_add_to_lru(lock);
903 unlock_res_and_lock(lock);
905 if (ldlm_is_fail_loc(lock))
906 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
908 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
909 * are not supported by the server, otherwise, it is done on
911 if (!exp_connect_cancelset(lock->l_conn_export) &&
912 !ns_connect_lru_resize(ns))
913 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
915 LDLM_DEBUG(lock, "do not add lock into lru list");
916 unlock_res_and_lock(lock);
923 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
925 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
927 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
928 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
929 ldlm_lock_decref_internal(lock, mode);
932 EXPORT_SYMBOL(ldlm_lock_decref);
935 * Decrease reader/writer refcount for LDLM lock with handle
936 * \a lockh and mark it for subsequent cancellation once r/w refcount
937 * drops to zero instead of putting into LRU.
940 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
943 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
946 LASSERT(lock != NULL);
948 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
949 lock_res_and_lock(lock);
950 ldlm_set_cbpending(lock);
951 unlock_res_and_lock(lock);
952 ldlm_lock_decref_internal(lock, mode);
955 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
957 struct sl_insert_point {
958 struct list_head *res_link;
959 struct list_head *mode_link;
960 struct list_head *policy_link;
964 * Finds a position to insert the new lock into granted lock list.
966 * Used for locks eligible for skiplist optimization.
969 * queue [input]: the granted list where search acts on;
970 * req [input]: the lock whose position to be located;
971 * prev [output]: positions within 3 lists to insert @req to
975 * - ldlm_grant_lock_with_skiplist
977 static void search_granted_lock(struct list_head *queue,
978 struct ldlm_lock *req,
979 struct sl_insert_point *prev)
981 struct list_head *tmp;
982 struct ldlm_lock *lock, *mode_end, *policy_end;
985 list_for_each(tmp, queue) {
986 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
988 mode_end = list_entry(lock->l_sl_mode.prev,
989 struct ldlm_lock, l_sl_mode);
991 if (lock->l_req_mode != req->l_req_mode) {
992 /* jump to last lock of mode group */
993 tmp = &mode_end->l_res_link;
997 /* suitable mode group is found */
998 if (lock->l_resource->lr_type == LDLM_PLAIN) {
999 /* insert point is last lock of the mode group */
1000 prev->res_link = &mode_end->l_res_link;
1001 prev->mode_link = &mode_end->l_sl_mode;
1002 prev->policy_link = &req->l_sl_policy;
1005 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
1008 list_entry(lock->l_sl_policy.prev,
1012 if (lock->l_policy_data.l_inodebits.bits ==
1013 req->l_policy_data.l_inodebits.bits) {
1014 /* insert point is last lock of
1015 * the policy group */
1017 &policy_end->l_res_link;
1019 &policy_end->l_sl_mode;
1021 &policy_end->l_sl_policy;
1026 if (policy_end == mode_end)
1027 /* done with mode group */
1030 /* go to next policy group within mode group */
1031 tmp = policy_end->l_res_link.next;
1032 lock = list_entry(tmp, struct ldlm_lock,
1034 } /* loop over policy groups within the mode group */
1036 /* insert point is last lock of the mode group,
1037 * new policy group is started */
1038 prev->res_link = &mode_end->l_res_link;
1039 prev->mode_link = &mode_end->l_sl_mode;
1040 prev->policy_link = &req->l_sl_policy;
1044 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1049 /* insert point is last lock on the queue,
1050 * new mode group and new policy group are started */
1051 prev->res_link = queue->prev;
1052 prev->mode_link = &req->l_sl_mode;
1053 prev->policy_link = &req->l_sl_policy;
1058 * Add a lock into resource granted list after a position described by
1061 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1062 struct sl_insert_point *prev)
1064 struct ldlm_resource *res = lock->l_resource;
1067 check_res_locked(res);
1069 ldlm_resource_dump(D_INFO, res);
1070 LDLM_DEBUG(lock, "About to add lock:");
1072 if (ldlm_is_destroyed(lock)) {
1073 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1077 LASSERT(list_empty(&lock->l_res_link));
1078 LASSERT(list_empty(&lock->l_sl_mode));
1079 LASSERT(list_empty(&lock->l_sl_policy));
1082 * lock->link == prev->link means lock is first starting the group.
1083 * Don't re-add to itself to suppress kernel warnings.
1085 if (&lock->l_res_link != prev->res_link)
1086 list_add(&lock->l_res_link, prev->res_link);
1087 if (&lock->l_sl_mode != prev->mode_link)
1088 list_add(&lock->l_sl_mode, prev->mode_link);
1089 if (&lock->l_sl_policy != prev->policy_link)
1090 list_add(&lock->l_sl_policy, prev->policy_link);
1096 * Add a lock to granted list on a resource maintaining skiplist
1099 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1101 struct sl_insert_point prev;
1103 LASSERT(ldlm_is_granted(lock));
1105 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1106 ldlm_granted_list_add_lock(lock, &prev);
1110 * Perform lock granting bookkeeping.
1112 * Includes putting the lock into granted list and updating lock mode.
1114 * - ldlm_lock_enqueue
1115 * - ldlm_reprocess_queue
1117 * must be called with lr_lock held
1119 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1121 struct ldlm_resource *res = lock->l_resource;
1124 check_res_locked(res);
1126 lock->l_granted_mode = lock->l_req_mode;
1128 if (work_list && lock->l_completion_ast != NULL)
1129 ldlm_add_ast_work_item(lock, NULL, work_list);
1131 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1132 ldlm_grant_lock_with_skiplist(lock);
1133 else if (res->lr_type == LDLM_EXTENT)
1134 ldlm_extent_add_lock(res, lock);
1135 else if (res->lr_type == LDLM_FLOCK) {
1136 /* We should not add locks to granted list in the following
1138 * - this is an UNLOCK but not a real lock;
1139 * - this is a TEST lock;
1140 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1141 * - this is a deadlock (flock cannot be granted) */
1142 if (lock->l_req_mode == 0 ||
1143 lock->l_req_mode == LCK_NL ||
1144 ldlm_is_test_lock(lock) ||
1145 ldlm_is_flock_deadlock(lock))
1147 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1152 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1157 * Check if the given @lock meets the criteria for a match.
1158 * A reference on the lock is taken if matched.
1160 * \param lock test-against this lock
1161 * \param data parameters
1163 static int lock_matches(struct ldlm_lock *lock, struct ldlm_match_data *data)
1165 union ldlm_policy_data *lpol = &lock->l_policy_data;
1166 enum ldlm_mode match = LCK_MINMODE;
1168 if (lock == data->lmd_old)
1169 return INTERVAL_ITER_STOP;
1171 /* Check if this lock can be matched.
1172 * Used by LU-2919(exclusive open) for open lease lock */
1173 if (ldlm_is_excl(lock))
1174 return INTERVAL_ITER_CONT;
1176 /* llite sometimes wants to match locks that will be
1177 * canceled when their users drop, but we allow it to match
1178 * if it passes in CBPENDING and the lock still has users.
1179 * this is generally only going to be used by children
1180 * whose parents already hold a lock so forward progress
1181 * can still happen. */
1182 if (ldlm_is_cbpending(lock) &&
1183 !(data->lmd_flags & LDLM_FL_CBPENDING))
1184 return INTERVAL_ITER_CONT;
1185 if (!data->lmd_unref && ldlm_is_cbpending(lock) &&
1186 lock->l_readers == 0 && lock->l_writers == 0)
1187 return INTERVAL_ITER_CONT;
1189 if (!(lock->l_req_mode & *data->lmd_mode))
1190 return INTERVAL_ITER_CONT;
1192 /* When we search for ast_data, we are not doing a traditional match,
1193 * so we don't worry about IBITS or extent matching.
1195 if (data->lmd_has_ast_data) {
1196 if (!lock->l_ast_data)
1197 return INTERVAL_ITER_CONT;
1202 match = lock->l_req_mode;
1204 switch (lock->l_resource->lr_type) {
1206 if (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1207 lpol->l_extent.end < data->lmd_policy->l_extent.end)
1208 return INTERVAL_ITER_CONT;
1210 if (unlikely(match == LCK_GROUP) &&
1211 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1212 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1213 return INTERVAL_ITER_CONT;
1216 /* We match if we have existing lock with same or wider set
1218 if ((lpol->l_inodebits.bits &
1219 data->lmd_policy->l_inodebits.bits) !=
1220 data->lmd_policy->l_inodebits.bits)
1221 return INTERVAL_ITER_CONT;
1227 /* We match if we have existing lock with same or wider set
1229 if (!data->lmd_unref && LDLM_HAVE_MASK(lock, GONE))
1230 return INTERVAL_ITER_CONT;
1232 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1233 return INTERVAL_ITER_CONT;
1235 /* Filter locks by skipping flags */
1236 if (data->lmd_skip_flags & lock->l_flags)
1237 return INTERVAL_ITER_CONT;
1240 if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1241 LDLM_LOCK_GET(lock);
1242 ldlm_lock_touch_in_lru(lock);
1244 ldlm_lock_addref_internal_nolock(lock, match);
1247 *data->lmd_mode = match;
1248 data->lmd_lock = lock;
1250 return INTERVAL_ITER_STOP;
1253 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1255 struct ldlm_interval *node = to_ldlm_interval(in);
1256 struct ldlm_match_data *data = args;
1257 struct ldlm_lock *lock;
1260 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1261 rc = lock_matches(lock, data);
1262 if (rc == INTERVAL_ITER_STOP)
1263 return INTERVAL_ITER_STOP;
1265 return INTERVAL_ITER_CONT;
1269 * Search for a lock with given parameters in interval trees.
1271 * \param res search for a lock in this resource
1272 * \param data parameters
1274 * \retval a referenced lock or NULL.
1276 struct ldlm_lock *search_itree(struct ldlm_resource *res,
1277 struct ldlm_match_data *data)
1279 struct interval_node_extent ext = {
1280 .start = data->lmd_policy->l_extent.start,
1281 .end = data->lmd_policy->l_extent.end
1285 data->lmd_lock = NULL;
1287 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1288 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1290 if (tree->lit_root == NULL)
1293 if (!(tree->lit_mode & *data->lmd_mode))
1296 interval_search(tree->lit_root, &ext,
1297 itree_overlap_cb, data);
1299 return data->lmd_lock;
1304 EXPORT_SYMBOL(search_itree);
1308 * Search for a lock with given properties in a queue.
1310 * \param queue search for a lock in this queue
1311 * \param data parameters
1313 * \retval a referenced lock or NULL.
1315 static struct ldlm_lock *search_queue(struct list_head *queue,
1316 struct ldlm_match_data *data)
1318 struct ldlm_lock *lock;
1321 data->lmd_lock = NULL;
1323 list_for_each_entry(lock, queue, l_res_link) {
1324 rc = lock_matches(lock, data);
1325 if (rc == INTERVAL_ITER_STOP)
1326 return data->lmd_lock;
1332 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1334 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1335 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1336 wake_up_all(&lock->l_waitq);
1339 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1341 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1343 lock_res_and_lock(lock);
1344 ldlm_lock_fail_match_locked(lock);
1345 unlock_res_and_lock(lock);
1349 * Mark lock as "matchable" by OST.
1351 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1353 * Assumes LDLM lock is already locked.
1355 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1357 ldlm_set_lvb_ready(lock);
1358 wake_up_all(&lock->l_waitq);
1360 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1363 * Mark lock as "matchable" by OST.
1364 * Locks the lock and then \see ldlm_lock_allow_match_locked
1366 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1368 lock_res_and_lock(lock);
1369 ldlm_lock_allow_match_locked(lock);
1370 unlock_res_and_lock(lock);
1372 EXPORT_SYMBOL(ldlm_lock_allow_match);
1375 * Attempt to find a lock with specified properties.
1377 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1380 * Can be called in two ways:
1382 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1383 * for a duplicate of.
1385 * Otherwise, all of the fields must be filled in, to match against.
1387 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1388 * server (ie, connh is NULL)
1389 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1390 * list will be considered
1391 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1392 * to be canceled can still be matched as long as they still have reader
1393 * or writer refernces
1394 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1395 * just tell us if we would have matched.
1397 * \retval 1 if it finds an already-existing lock that is compatible; in this
1398 * case, lockh is filled in with a addref()ed lock
1400 * We also check security context, and if that fails we simply return 0 (to
1401 * keep caller code unchanged), the context failure will be discovered by
1402 * caller sometime later.
1404 enum ldlm_mode ldlm_lock_match_with_skip(struct ldlm_namespace *ns,
1405 __u64 flags, __u64 skip_flags,
1406 const struct ldlm_res_id *res_id,
1407 enum ldlm_type type,
1408 union ldlm_policy_data *policy,
1409 enum ldlm_mode mode,
1410 struct lustre_handle *lockh, int unref)
1412 struct ldlm_match_data data = {
1416 .lmd_policy = policy,
1418 .lmd_skip_flags = skip_flags,
1420 .lmd_has_ast_data = false,
1422 struct ldlm_resource *res;
1423 struct ldlm_lock *lock;
1429 data.lmd_old = ldlm_handle2lock(lockh);
1430 LASSERT(data.lmd_old != NULL);
1432 ns = ldlm_lock_to_ns(data.lmd_old);
1433 res_id = &data.lmd_old->l_resource->lr_name;
1434 type = data.lmd_old->l_resource->lr_type;
1435 *data.lmd_mode = data.lmd_old->l_req_mode;
1438 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1440 LASSERT(data.lmd_old == NULL);
1444 LDLM_RESOURCE_ADDREF(res);
1446 if (res->lr_type == LDLM_EXTENT)
1447 lock = search_itree(res, &data);
1449 lock = search_queue(&res->lr_granted, &data);
1450 if (!lock && !(flags & LDLM_FL_BLOCK_GRANTED))
1451 lock = search_queue(&res->lr_waiting, &data);
1452 matched = lock ? mode : 0;
1454 LDLM_RESOURCE_DELREF(res);
1455 ldlm_resource_putref(res);
1458 ldlm_lock2handle(lock, lockh);
1459 if ((flags & LDLM_FL_LVB_READY) &&
1460 (!ldlm_is_lvb_ready(lock))) {
1461 __u64 wait_flags = LDLM_FL_LVB_READY |
1462 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1463 struct l_wait_info lwi;
1465 if (lock->l_completion_ast) {
1466 int err = lock->l_completion_ast(lock,
1467 LDLM_FL_WAIT_NOREPROC,
1470 GOTO(out_fail_match, matched = 0);
1473 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1474 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1476 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1477 l_wait_event(lock->l_waitq, lock->l_flags & wait_flags,
1479 if (!ldlm_is_lvb_ready(lock))
1480 GOTO(out_fail_match, matched = 0);
1483 /* check user's security context */
1484 if (lock->l_conn_export &&
1485 sptlrpc_import_check_ctx(
1486 class_exp2cliimp(lock->l_conn_export)))
1487 GOTO(out_fail_match, matched = 0);
1489 LDLM_DEBUG(lock, "matched (%llu %llu)",
1490 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1491 res_id->name[2] : policy->l_extent.start,
1492 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1493 res_id->name[3] : policy->l_extent.end);
1496 if (flags & LDLM_FL_TEST_LOCK)
1497 LDLM_LOCK_RELEASE(lock);
1499 ldlm_lock_decref_internal(lock, mode);
1502 /* less verbose for test-only */
1503 if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
1504 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1505 "%llu/%llu (%llu %llu)", ns,
1506 type, mode, res_id->name[0], res_id->name[1],
1507 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1508 res_id->name[2] : policy->l_extent.start,
1509 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1510 res_id->name[3] : policy->l_extent.end);
1512 if (data.lmd_old != NULL)
1513 LDLM_LOCK_PUT(data.lmd_old);
1517 EXPORT_SYMBOL(ldlm_lock_match_with_skip);
1519 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1522 struct ldlm_lock *lock;
1523 enum ldlm_mode mode = 0;
1526 lock = ldlm_handle2lock(lockh);
1528 lock_res_and_lock(lock);
1529 if (LDLM_HAVE_MASK(lock, GONE))
1532 if (ldlm_is_cbpending(lock) &&
1533 lock->l_readers == 0 && lock->l_writers == 0)
1537 *bits = lock->l_policy_data.l_inodebits.bits;
1538 mode = lock->l_granted_mode;
1539 ldlm_lock_addref_internal_nolock(lock, mode);
1546 unlock_res_and_lock(lock);
1547 LDLM_LOCK_PUT(lock);
1551 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1553 /** The caller must guarantee that the buffer is large enough. */
1554 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1555 enum req_location loc, void *data, int size)
1560 LASSERT(data != NULL);
1563 switch (lock->l_lvb_type) {
1565 if (size == sizeof(struct ost_lvb)) {
1566 if (loc == RCL_CLIENT)
1567 lvb = req_capsule_client_swab_get(pill,
1569 lustre_swab_ost_lvb);
1571 lvb = req_capsule_server_swab_get(pill,
1573 lustre_swab_ost_lvb);
1574 if (unlikely(lvb == NULL)) {
1575 LDLM_ERROR(lock, "no LVB");
1579 memcpy(data, lvb, size);
1580 } else if (size == sizeof(struct ost_lvb_v1)) {
1581 struct ost_lvb *olvb = data;
1583 if (loc == RCL_CLIENT)
1584 lvb = req_capsule_client_swab_get(pill,
1586 lustre_swab_ost_lvb_v1);
1588 lvb = req_capsule_server_sized_swab_get(pill,
1590 lustre_swab_ost_lvb_v1);
1591 if (unlikely(lvb == NULL)) {
1592 LDLM_ERROR(lock, "no LVB");
1596 memcpy(data, lvb, size);
1597 olvb->lvb_mtime_ns = 0;
1598 olvb->lvb_atime_ns = 0;
1599 olvb->lvb_ctime_ns = 0;
1601 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1607 if (size == sizeof(struct lquota_lvb)) {
1608 if (loc == RCL_CLIENT)
1609 lvb = req_capsule_client_swab_get(pill,
1611 lustre_swab_lquota_lvb);
1613 lvb = req_capsule_server_swab_get(pill,
1615 lustre_swab_lquota_lvb);
1616 if (unlikely(lvb == NULL)) {
1617 LDLM_ERROR(lock, "no LVB");
1621 memcpy(data, lvb, size);
1623 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1632 if (loc == RCL_CLIENT)
1633 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1635 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1636 if (unlikely(lvb == NULL)) {
1637 LDLM_ERROR(lock, "no LVB");
1641 memcpy(data, lvb, size);
1644 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1645 libcfs_debug_dumpstack(NULL);
1653 * Create and fill in new LDLM lock with specified properties.
1654 * Returns a referenced lock
1656 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1657 const struct ldlm_res_id *res_id,
1658 enum ldlm_type type,
1659 enum ldlm_mode mode,
1660 const struct ldlm_callback_suite *cbs,
1661 void *data, __u32 lvb_len,
1662 enum lvb_type lvb_type)
1664 struct ldlm_lock *lock;
1665 struct ldlm_resource *res;
1669 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1671 RETURN(ERR_CAST(res));
1673 lock = ldlm_lock_new(res);
1675 RETURN(ERR_PTR(-ENOMEM));
1677 lock->l_req_mode = mode;
1678 lock->l_ast_data = data;
1679 lock->l_pid = current_pid();
1680 if (ns_is_server(ns))
1681 ldlm_set_ns_srv(lock);
1683 lock->l_blocking_ast = cbs->lcs_blocking;
1684 lock->l_completion_ast = cbs->lcs_completion;
1685 lock->l_glimpse_ast = cbs->lcs_glimpse;
1690 rc = ldlm_extent_alloc_lock(lock);
1693 rc = ldlm_inodebits_alloc_lock(lock);
1702 lock->l_lvb_len = lvb_len;
1703 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1704 if (lock->l_lvb_data == NULL)
1705 GOTO(out, rc = -ENOMEM);
1708 lock->l_lvb_type = lvb_type;
1709 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1710 GOTO(out, rc = -ENOENT);
1715 ldlm_lock_destroy(lock);
1716 LDLM_LOCK_RELEASE(lock);
1717 RETURN(ERR_PTR(rc));
1720 #ifdef HAVE_SERVER_SUPPORT
1721 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1724 struct ldlm_resource *res = lock->l_resource;
1725 enum ldlm_error rc = ELDLM_OK;
1726 LIST_HEAD(rpc_list);
1727 ldlm_processing_policy policy;
1731 policy = ldlm_get_processing_policy(res);
1733 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1734 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1735 res->lr_type != LDLM_FLOCK) {
1736 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1737 if (rc == -ERESTART)
1741 if (!list_empty(&rpc_list))
1742 ldlm_discard_bl_list(&rpc_list);
1749 * Enqueue (request) a lock.
1751 * Does not block. As a result of enqueue the lock would be put
1752 * into granted or waiting list.
1754 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1755 * set, skip all the enqueueing and delegate lock processing to intent policy
1758 enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
1759 struct ldlm_namespace *ns,
1760 struct ldlm_lock **lockp,
1761 void *cookie, __u64 *flags)
1763 struct ldlm_lock *lock = *lockp;
1764 struct ldlm_resource *res = lock->l_resource;
1765 int local = ns_is_client(ldlm_res_to_ns(res));
1766 enum ldlm_error rc = ELDLM_OK;
1767 struct ldlm_interval *node = NULL;
1768 #ifdef HAVE_SERVER_SUPPORT
1769 bool reconstruct = false;
1773 /* policies are not executed on the client or during replay */
1774 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1775 && !local && ns->ns_policy) {
1776 rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
1778 if (rc == ELDLM_LOCK_REPLACED) {
1779 /* The lock that was returned has already been granted,
1780 * and placed into lockp. If it's not the same as the
1781 * one we passed in, then destroy the old one and our
1782 * work here is done. */
1783 if (lock != *lockp) {
1784 ldlm_lock_destroy(lock);
1785 LDLM_LOCK_RELEASE(lock);
1787 *flags |= LDLM_FL_LOCK_CHANGED;
1789 } else if (rc != ELDLM_OK &&
1790 ldlm_is_granted(lock)) {
1791 LASSERT(*flags & LDLM_FL_RESENT);
1792 /* It may happen that ns_policy returns an error in
1793 * resend case, object may be unlinked or just some
1794 * error occurs. It is unclear if lock reached the
1795 * client in the original reply, just leave the lock on
1796 * server, not returning it again to client. Due to
1797 * LU-6529, the server will not OOM. */
1799 } else if (rc != ELDLM_OK ||
1800 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1801 ldlm_lock_destroy(lock);
1806 if (*flags & LDLM_FL_RESENT) {
1807 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1808 * Set LOCK_CHANGED always.
1809 * Check if the lock is granted for BLOCK_GRANTED.
1810 * Take NO_TIMEOUT from the lock as it is inherited through
1811 * LDLM_FL_INHERIT_MASK */
1812 *flags |= LDLM_FL_LOCK_CHANGED;
1813 if (!ldlm_is_granted(lock))
1814 *flags |= LDLM_FL_BLOCK_GRANTED;
1815 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1819 /* For a replaying lock, it might be already in granted list. So
1820 * unlinking the lock will cause the interval node to be freed, we
1821 * have to allocate the interval node early otherwise we can't regrant
1822 * this lock in the future. - jay */
1823 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1824 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1826 #ifdef HAVE_SERVER_SUPPORT
1827 reconstruct = !local && res->lr_type == LDLM_FLOCK &&
1828 !(*flags & LDLM_FL_TEST_LOCK);
1830 rc = req_can_reconstruct(cookie, NULL);
1839 lock_res_and_lock(lock);
1840 if (local && ldlm_is_granted(lock)) {
1841 /* The server returned a blocked lock, but it was granted
1842 * before we got a chance to actually enqueue it. We don't
1843 * need to do anything else. */
1844 *flags &= ~LDLM_FL_BLOCKED_MASK;
1845 GOTO(out, rc = ELDLM_OK);
1848 ldlm_resource_unlink_lock(lock);
1849 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1851 ldlm_lock_destroy_nolock(lock);
1852 GOTO(out, rc = -ENOMEM);
1855 INIT_LIST_HEAD(&node->li_group);
1856 ldlm_interval_attach(node, lock);
1860 /* Some flags from the enqueue want to make it into the AST, via the
1861 * lock's l_flags. */
1862 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1863 ldlm_set_ast_discard_data(lock);
1864 if (*flags & LDLM_FL_TEST_LOCK)
1865 ldlm_set_test_lock(lock);
1866 if (*flags & LDLM_FL_COS_INCOMPAT)
1867 ldlm_set_cos_incompat(lock);
1868 if (*flags & LDLM_FL_COS_ENABLED)
1869 ldlm_set_cos_enabled(lock);
1871 /* This distinction between local lock trees is very important; a client
1872 * namespace only has information about locks taken by that client, and
1873 * thus doesn't have enough information to decide for itself if it can
1874 * be granted (below). In this case, we do exactly what the server
1875 * tells us to do, as dictated by the 'flags'.
1877 * We do exactly the same thing during recovery, when the server is
1878 * more or less trusting the clients not to lie.
1880 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1883 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1884 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1886 ldlm_grant_lock(lock, NULL);
1887 GOTO(out, rc = ELDLM_OK);
1888 #ifdef HAVE_SERVER_SUPPORT
1889 } else if (*flags & LDLM_FL_REPLAY) {
1890 if (*flags & LDLM_FL_BLOCK_WAIT) {
1891 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1892 GOTO(out, rc = ELDLM_OK);
1893 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1894 ldlm_grant_lock(lock, NULL);
1895 GOTO(out, rc = ELDLM_OK);
1897 /* If no flags, fall through to normal enqueue path. */
1900 rc = ldlm_lock_enqueue_helper(lock, flags);
1904 CERROR("This is client-side-only module, cannot handle "
1905 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1911 unlock_res_and_lock(lock);
1913 #ifdef HAVE_SERVER_SUPPORT
1915 struct ptlrpc_request *req = cookie;
1917 tgt_mk_reply_data(NULL, NULL,
1918 &req->rq_export->exp_target_data,
1919 req, 0, NULL, false, 0);
1923 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1927 #ifdef HAVE_SERVER_SUPPORT
1929 * Iterate through all waiting locks on a given resource queue and attempt to
1932 * Must be called with resource lock held.
1934 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1935 struct list_head *work_list,
1936 enum ldlm_process_intention intention,
1937 struct ldlm_lock *hint)
1939 struct list_head *tmp, *pos;
1940 ldlm_processing_policy policy;
1942 int rc = LDLM_ITER_CONTINUE;
1943 enum ldlm_error err;
1944 LIST_HEAD(bl_ast_list);
1948 check_res_locked(res);
1950 policy = ldlm_get_processing_policy(res);
1952 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1953 intention == LDLM_PROCESS_RECOVERY);
1956 list_for_each_safe(tmp, pos, queue) {
1957 struct ldlm_lock *pending;
1958 LIST_HEAD(rpc_list);
1960 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1962 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1965 rc = policy(pending, &flags, intention, &err, &rpc_list);
1966 if (pending->l_granted_mode == pending->l_req_mode ||
1967 res->lr_type == LDLM_FLOCK) {
1968 list_splice(&rpc_list, work_list);
1970 list_splice(&rpc_list, &bl_ast_list);
1973 * When this is called from recovery done, we always want
1974 * to scan the whole list no matter what 'rc' is returned.
1976 if (rc != LDLM_ITER_CONTINUE &&
1977 intention == LDLM_PROCESS_RESCAN)
1981 if (!list_empty(&bl_ast_list)) {
1984 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
1988 if (rc == -ERESTART)
1992 if (!list_empty(&bl_ast_list))
1993 ldlm_discard_bl_list(&bl_ast_list);
1995 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
1999 * Conflicting locks are detected for a lock to be enqueued, add the lock
2000 * into waiting list and send blocking ASTs to the conflicting locks.
2002 * \param[in] lock The lock to be enqueued.
2003 * \param[out] flags Lock flags for the lock to be enqueued.
2004 * \param[in] rpc_list Conflicting locks list.
2006 * \retval -ERESTART: Some lock was instantly canceled while sending
2007 * blocking ASTs, caller needs to re-check conflicting
2009 * \retval -EAGAIN: Lock was destroyed, caller should return error.
2010 * \reval 0: Lock is successfully added in waiting list.
2012 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
2013 struct list_head *rpc_list)
2015 struct ldlm_resource *res = lock->l_resource;
2019 check_res_locked(res);
2021 /* If either of the compat_queue()s returned failure, then we
2022 * have ASTs to send and must go onto the waiting list.
2024 * bug 2322: we used to unlink and re-add here, which was a
2025 * terrible folly -- if we goto restart, we could get
2026 * re-ordered! Causes deadlock, because ASTs aren't sent! */
2027 if (list_empty(&lock->l_res_link))
2028 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
2031 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
2034 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
2035 !ns_is_client(ldlm_res_to_ns(res)))
2036 class_fail_export(lock->l_export);
2039 if (rc == -ERESTART) {
2040 /* 15715: The lock was granted and destroyed after
2041 * resource lock was dropped. Interval node was freed
2042 * in ldlm_lock_destroy. Anyway, this always happens
2043 * when a client is being evicted. So it would be
2044 * ok to return an error. -jay */
2045 if (ldlm_is_destroyed(lock))
2048 /* lock was granted while resource was unlocked. */
2049 if (ldlm_is_granted(lock)) {
2050 /* bug 11300: if the lock has been granted,
2051 * break earlier because otherwise, we will go
2052 * to restart and ldlm_resource_unlink will be
2053 * called and it causes the interval node to be
2054 * freed. Then we will fail at
2055 * ldlm_extent_add_lock() */
2056 *flags &= ~LDLM_FL_BLOCKED_MASK;
2062 *flags |= LDLM_FL_BLOCK_GRANTED;
2068 * Discard all AST work items from list.
2070 * If for whatever reason we do not want to send ASTs to conflicting locks
2071 * anymore, disassemble the list with this function.
2073 void ldlm_discard_bl_list(struct list_head *bl_list)
2075 struct ldlm_lock *lock, *tmp;
2079 list_for_each_entry_safe(lock, tmp, bl_list, l_bl_ast) {
2080 LASSERT(!list_empty(&lock->l_bl_ast));
2081 list_del_init(&lock->l_bl_ast);
2082 ldlm_clear_ast_sent(lock);
2083 LASSERT(lock->l_bl_ast_run == 0);
2084 ldlm_clear_blocking_lock(lock);
2085 LDLM_LOCK_RELEASE(lock);
2091 * Process a call to blocking AST callback for a lock in ast_work list
2094 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2096 struct ldlm_cb_set_arg *arg = opaq;
2097 struct ldlm_lock *lock;
2098 struct ldlm_lock_desc d;
2099 struct ldlm_bl_desc bld;
2104 if (list_empty(arg->list))
2107 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2109 /* nobody should touch l_bl_ast but some locks in the list may become
2110 * granted after lock convert or COS downgrade, these locks should be
2111 * just skipped here and removed from the list.
2113 lock_res_and_lock(lock);
2114 list_del_init(&lock->l_bl_ast);
2116 /* lock is not blocking lock anymore, but was kept in the list because
2117 * it can managed only here.
2119 if (!ldlm_is_ast_sent(lock)) {
2120 unlock_res_and_lock(lock);
2121 LDLM_LOCK_RELEASE(lock);
2125 LASSERT(lock->l_blocking_lock);
2126 ldlm_lock2desc(lock->l_blocking_lock, &d);
2127 /* copy blocking lock ibits in cancel_bits as well,
2128 * new client may use them for lock convert and it is
2129 * important to use new field to convert locks from
2132 d.l_policy_data.l_inodebits.cancel_bits =
2133 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2135 /* Blocking lock is being destroyed here but some information about it
2136 * may be needed inside l_blocking_ast() function below,
2137 * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
2139 bld.bl_same_client = lock->l_client_cookie ==
2140 lock->l_blocking_lock->l_client_cookie;
2141 bld.bl_cos_incompat = ldlm_is_cos_incompat(lock->l_blocking_lock);
2142 arg->bl_desc = &bld;
2144 LASSERT(ldlm_is_ast_sent(lock));
2145 LASSERT(lock->l_bl_ast_run == 0);
2146 lock->l_bl_ast_run++;
2147 ldlm_clear_blocking_lock(lock);
2148 unlock_res_and_lock(lock);
2150 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2152 LDLM_LOCK_RELEASE(lock);
2158 * Process a call to revocation AST callback for a lock in ast_work list
2161 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2163 struct ldlm_cb_set_arg *arg = opaq;
2164 struct ldlm_lock_desc desc;
2166 struct ldlm_lock *lock;
2169 if (list_empty(arg->list))
2172 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2173 list_del_init(&lock->l_rk_ast);
2175 /* the desc just pretend to exclusive */
2176 ldlm_lock2desc(lock, &desc);
2177 desc.l_req_mode = LCK_EX;
2178 desc.l_granted_mode = 0;
2180 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2181 LDLM_LOCK_RELEASE(lock);
2187 * Process a call to glimpse AST callback for a lock in ast_work list
2189 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2191 struct ldlm_cb_set_arg *arg = opaq;
2192 struct ldlm_glimpse_work *gl_work;
2193 struct ldlm_lock *lock;
2197 if (list_empty(arg->list))
2200 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2202 list_del_init(&gl_work->gl_list);
2204 lock = gl_work->gl_lock;
2206 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2207 arg->gl_desc = gl_work->gl_desc;
2208 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2209 arg->gl_interpret_data = gl_work->gl_interpret_data;
2211 /* invoke the actual glimpse callback */
2212 if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
2215 LDLM_LOCK_RELEASE(lock);
2216 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2217 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2219 OBD_FREE_PTR(gl_work);
2226 * Process a call to completion AST callback for a lock in ast_work list
2229 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2231 struct ldlm_cb_set_arg *arg = opaq;
2232 struct ldlm_lock *lock;
2233 ldlm_completion_callback completion_callback;
2238 if (list_empty(arg->list))
2241 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2243 /* It's possible to receive a completion AST before we've set
2244 * the l_completion_ast pointer: either because the AST arrived
2245 * before the reply, or simply because there's a small race
2246 * window between receiving the reply and finishing the local
2247 * enqueue. (bug 842)
2249 * This can't happen with the blocking_ast, however, because we
2250 * will never call the local blocking_ast until we drop our
2251 * reader/writer reference, which we won't do until we get the
2252 * reply and finish enqueueing. */
2254 /* nobody should touch l_cp_ast */
2255 lock_res_and_lock(lock);
2256 list_del_init(&lock->l_cp_ast);
2257 LASSERT(ldlm_is_cp_reqd(lock));
2258 /* save l_completion_ast since it can be changed by
2259 * mds_intent_policy(), see bug 14225 */
2260 completion_callback = lock->l_completion_ast;
2261 ldlm_clear_cp_reqd(lock);
2262 unlock_res_and_lock(lock);
2264 if (completion_callback != NULL)
2265 rc = completion_callback(lock, 0, (void *)arg);
2266 LDLM_LOCK_RELEASE(lock);
2272 * Process list of locks in need of ASTs being sent.
2274 * Used on server to send multiple ASTs together instead of sending one by
2277 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2278 ldlm_desc_ast_t ast_type)
2280 struct ldlm_cb_set_arg *arg;
2281 set_producer_func work_ast_lock;
2284 if (list_empty(rpc_list))
2291 atomic_set(&arg->restart, 0);
2292 arg->list = rpc_list;
2295 case LDLM_WORK_CP_AST:
2296 arg->type = LDLM_CP_CALLBACK;
2297 work_ast_lock = ldlm_work_cp_ast_lock;
2299 #ifdef HAVE_SERVER_SUPPORT
2300 case LDLM_WORK_BL_AST:
2301 arg->type = LDLM_BL_CALLBACK;
2302 work_ast_lock = ldlm_work_bl_ast_lock;
2304 case LDLM_WORK_REVOKE_AST:
2305 arg->type = LDLM_BL_CALLBACK;
2306 work_ast_lock = ldlm_work_revoke_ast_lock;
2308 case LDLM_WORK_GL_AST:
2309 arg->type = LDLM_GL_CALLBACK;
2310 work_ast_lock = ldlm_work_gl_ast_lock;
2317 /* We create a ptlrpc request set with flow control extension.
2318 * This request set will use the work_ast_lock function to produce new
2319 * requests and will send a new request each time one completes in order
2320 * to keep the number of requests in flight to ns_max_parallel_ast */
2321 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2322 work_ast_lock, arg);
2323 if (arg->set == NULL)
2324 GOTO(out, rc = -ENOMEM);
2326 ptlrpc_set_wait(NULL, arg->set);
2327 ptlrpc_set_destroy(arg->set);
2329 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2337 * Try to grant all waiting locks on a resource.
2339 * Calls ldlm_reprocess_queue on waiting queue.
2341 * Typically called after some resource locks are cancelled to see
2342 * if anything could be granted as a result of the cancellation.
2344 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2345 enum ldlm_process_intention intention,
2346 struct ldlm_lock *hint)
2348 struct list_head rpc_list;
2349 #ifdef HAVE_SERVER_SUPPORT
2350 ldlm_reprocessing_policy reprocess;
2351 struct obd_device *obd;
2356 INIT_LIST_HEAD(&rpc_list);
2357 /* Local lock trees don't get reprocessed. */
2358 if (ns_is_client(ldlm_res_to_ns(res))) {
2363 /* Disable reprocess during lock replay stage but allow during
2364 * request replay stage.
2366 obd = ldlm_res_to_ns(res)->ns_obd;
2367 if (obd->obd_recovering &&
2368 atomic_read(&obd->obd_req_replay_clients) == 0)
2372 reprocess = ldlm_get_reprocessing_policy(res);
2373 reprocess(res, &res->lr_waiting, &rpc_list, intention, hint);
2376 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2378 if (rc == -ERESTART) {
2379 LASSERT(list_empty(&rpc_list));
2385 INIT_LIST_HEAD(&rpc_list);
2386 if (!ns_is_client(ldlm_res_to_ns(res))) {
2387 CERROR("This is client-side-only module, cannot handle "
2388 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2395 void ldlm_reprocess_all(struct ldlm_resource *res, struct ldlm_lock *hint)
2397 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN, hint);
2399 EXPORT_SYMBOL(ldlm_reprocess_all);
2401 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2402 struct hlist_node *hnode, void *arg)
2404 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2406 /* This is only called once after recovery done. LU-8306. */
2407 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY, NULL);
2412 * Iterate through all resources on a namespace attempting to grant waiting
2415 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2420 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2421 ldlm_reprocess_res, NULL, 0);
2427 * Helper function to call blocking AST for LDLM lock \a lock in a
2428 * "cancelling" mode.
2430 void ldlm_cancel_callback(struct ldlm_lock *lock)
2432 check_res_locked(lock->l_resource);
2433 if (!ldlm_is_cancel(lock)) {
2434 ldlm_set_cancel(lock);
2435 if (lock->l_blocking_ast) {
2436 unlock_res_and_lock(lock);
2437 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2439 lock_res_and_lock(lock);
2441 LDLM_DEBUG(lock, "no blocking ast");
2444 /* only canceller can set bl_done bit */
2445 ldlm_set_bl_done(lock);
2446 wake_up_all(&lock->l_waitq);
2447 } else if (!ldlm_is_bl_done(lock)) {
2448 /* The lock is guaranteed to have been canceled once
2449 * returning from this function. */
2450 unlock_res_and_lock(lock);
2451 wait_event_idle(lock->l_waitq, is_bl_done(lock));
2452 lock_res_and_lock(lock);
2457 * Remove skiplist-enabled LDLM lock \a req from granted list
2459 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2461 if (req->l_resource->lr_type != LDLM_PLAIN &&
2462 req->l_resource->lr_type != LDLM_IBITS)
2465 list_del_init(&req->l_sl_policy);
2466 list_del_init(&req->l_sl_mode);
2470 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2472 void ldlm_lock_cancel(struct ldlm_lock *lock)
2474 struct ldlm_resource *res;
2475 struct ldlm_namespace *ns;
2478 lock_res_and_lock(lock);
2480 res = lock->l_resource;
2481 ns = ldlm_res_to_ns(res);
2483 /* Please do not, no matter how tempting, remove this LBUG without
2484 * talking to me first. -phik */
2485 if (lock->l_readers || lock->l_writers) {
2486 LDLM_ERROR(lock, "lock still has references");
2487 unlock_res_and_lock(lock);
2491 if (ldlm_is_waited(lock))
2492 ldlm_del_waiting_lock(lock);
2494 /* Releases cancel callback. */
2495 ldlm_cancel_callback(lock);
2497 /* Yes, second time, just in case it was added again while we were
2498 * running with no res lock in ldlm_cancel_callback */
2499 if (ldlm_is_waited(lock))
2500 ldlm_del_waiting_lock(lock);
2502 ldlm_resource_unlink_lock(lock);
2503 ldlm_lock_destroy_nolock(lock);
2505 if (ldlm_is_granted(lock))
2506 ldlm_pool_del(&ns->ns_pool, lock);
2508 /* Make sure we will not be called again for same lock what is possible
2509 * if not to zero out lock->l_granted_mode */
2510 lock->l_granted_mode = LCK_MINMODE;
2511 unlock_res_and_lock(lock);
2515 EXPORT_SYMBOL(ldlm_lock_cancel);
2518 * Set opaque data into the lock that only makes sense to upper layer.
2520 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2522 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2527 if (lock->l_ast_data == NULL)
2528 lock->l_ast_data = data;
2529 if (lock->l_ast_data == data)
2531 LDLM_LOCK_PUT(lock);
2535 EXPORT_SYMBOL(ldlm_lock_set_data);
2537 struct export_cl_data {
2538 const struct lu_env *ecl_env;
2539 struct obd_export *ecl_exp;
2543 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2544 struct ldlm_lock *lock,
2545 struct export_cl_data *ecl)
2547 struct ldlm_resource *res;
2549 res = ldlm_resource_getref(lock->l_resource);
2551 ldlm_lvbo_update(res, lock, NULL, 1);
2552 ldlm_lock_cancel(lock);
2553 if (!exp->exp_obd->obd_stopping)
2554 ldlm_reprocess_all(res, lock);
2555 ldlm_resource_putref(res);
2558 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2559 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2560 exp, ecl->ecl_loop);
2565 * Iterator function for ldlm_export_cancel_locks.
2566 * Cancels passed locks.
2569 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2570 struct hlist_node *hnode, void *data)
2573 struct export_cl_data *ecl = (struct export_cl_data *)data;
2574 struct obd_export *exp = ecl->ecl_exp;
2575 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2577 LDLM_LOCK_GET(lock);
2578 ldlm_cancel_lock_for_export(exp, lock, ecl);
2579 LDLM_LOCK_RELEASE(lock);
2585 * Cancel all blocked locks for given export.
2587 * Typically called on client disconnection/eviction
2589 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2592 struct export_cl_data ecl = {
2598 rc = lu_env_init(&env, LCT_DT_THREAD);
2603 while (!list_empty(&exp->exp_bl_list)) {
2604 struct ldlm_lock *lock;
2606 spin_lock_bh(&exp->exp_bl_list_lock);
2607 if (!list_empty(&exp->exp_bl_list)) {
2608 lock = list_entry(exp->exp_bl_list.next,
2609 struct ldlm_lock, l_exp_list);
2610 LDLM_LOCK_GET(lock);
2611 list_del_init(&lock->l_exp_list);
2615 spin_unlock_bh(&exp->exp_bl_list_lock);
2620 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2621 LDLM_LOCK_RELEASE(lock);
2626 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2627 "left on hash table %d.\n", exp, ecl.ecl_loop,
2628 atomic_read(&exp->exp_lock_hash->hs_count));
2630 return ecl.ecl_loop;
2634 * Cancel all locks for given export.
2636 * Typically called after client disconnection/eviction
2638 int ldlm_export_cancel_locks(struct obd_export *exp)
2640 struct export_cl_data ecl;
2644 rc = lu_env_init(&env, LCT_DT_THREAD);
2651 cfs_hash_for_each_empty(exp->exp_lock_hash,
2652 ldlm_cancel_locks_for_export_cb, &ecl);
2654 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2655 "left on hash table %d.\n", exp, ecl.ecl_loop,
2656 atomic_read(&exp->exp_lock_hash->hs_count));
2658 if (ecl.ecl_loop > 0 &&
2659 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2660 exp->exp_obd->obd_stopping)
2661 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2665 return ecl.ecl_loop;
2669 * Downgrade an PW/EX lock to COS | CR mode.
2671 * A lock mode convertion from PW/EX mode to less conflict mode. The
2672 * convertion may fail if lock was canceled before downgrade, but it doesn't
2673 * indicate any problem, because such lock has no reader or writer, and will
2676 * Used by Commit on Sharing (COS) code to force object changes commit in case
2677 * of conflict. Converted lock is considered as new lock and all blocking AST
2678 * things are cleared, so any pending or new blocked lock on that lock will
2679 * cause new call to blocking_ast and force resource object commit.
2681 * Also used by layout_change to replace EX lock to CR lock.
2683 * \param lock A lock to convert
2684 * \param new_mode new lock mode
2686 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2688 #ifdef HAVE_SERVER_SUPPORT
2691 LASSERT(new_mode == LCK_COS || new_mode == LCK_CR);
2693 lock_res_and_lock(lock);
2695 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2696 unlock_res_and_lock(lock);
2698 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2699 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2703 ldlm_resource_unlink_lock(lock);
2705 * Remove the lock from pool as it will be added again in
2706 * ldlm_grant_lock() called below.
2708 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2710 /* Consider downgraded lock as a new lock and clear all states
2711 * related to a previous blocking AST processing.
2713 ldlm_clear_blocking_data(lock);
2715 lock->l_req_mode = new_mode;
2716 ldlm_grant_lock(lock, NULL);
2717 unlock_res_and_lock(lock);
2719 ldlm_reprocess_all(lock->l_resource, lock);
2724 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2727 * Print lock with lock handle \a lockh description into debug log.
2729 * Used when printing all locks on a resource for debug purposes.
2731 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2733 struct ldlm_lock *lock;
2735 if (!((libcfs_debug | D_ERROR) & level))
2738 lock = ldlm_handle2lock(lockh);
2742 LDLM_DEBUG_LIMIT(level, lock, "###");
2744 LDLM_LOCK_PUT(lock);
2746 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2749 * Print lock information with custom message into debug log.
2752 void _ldlm_lock_debug(struct ldlm_lock *lock,
2753 struct libcfs_debug_msg_data *msgdata,
2754 const char *fmt, ...)
2757 struct obd_export *exp = lock->l_export;
2758 struct ldlm_resource *resource = NULL;
2759 struct va_format vaf;
2760 char *nid = "local";
2762 /* on server-side resource of lock doesn't change */
2763 if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2764 if (lock->l_resource != NULL)
2765 resource = ldlm_resource_getref(lock->l_resource);
2766 } else if (spin_trylock(&lock->l_lock)) {
2767 if (lock->l_resource != NULL)
2768 resource = ldlm_resource_getref(lock->l_resource);
2769 spin_unlock(&lock->l_lock);
2772 va_start(args, fmt);
2776 if (exp && exp->exp_connection) {
2777 nid = obd_export_nid2str(exp);
2778 } else if (exp && exp->exp_obd != NULL) {
2779 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2780 nid = obd_import_nid2str(imp);
2783 if (resource == NULL) {
2784 libcfs_debug_msg(msgdata,
2785 "%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",
2788 lock->l_handle.h_cookie,
2789 refcount_read(&lock->l_handle.h_ref),
2790 lock->l_readers, lock->l_writers,
2791 ldlm_lockname[lock->l_granted_mode],
2792 ldlm_lockname[lock->l_req_mode],
2794 lock->l_remote_handle.cookie,
2795 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2796 lock->l_pid, lock->l_callback_timeout,
2802 switch (resource->lr_type) {
2804 libcfs_debug_msg(msgdata,
2805 "%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",
2807 ldlm_lock_to_ns_name(lock), lock,
2808 lock->l_handle.h_cookie,
2809 refcount_read(&lock->l_handle.h_ref),
2810 lock->l_readers, lock->l_writers,
2811 ldlm_lockname[lock->l_granted_mode],
2812 ldlm_lockname[lock->l_req_mode],
2814 atomic_read(&resource->lr_refcount),
2815 ldlm_typename[resource->lr_type],
2816 lock->l_policy_data.l_extent.start,
2817 lock->l_policy_data.l_extent.end,
2818 lock->l_req_extent.start, lock->l_req_extent.end,
2820 lock->l_remote_handle.cookie,
2821 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2822 lock->l_pid, lock->l_callback_timeout,
2827 libcfs_debug_msg(msgdata,
2828 "%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",
2830 ldlm_lock_to_ns_name(lock), lock,
2831 lock->l_handle.h_cookie,
2832 refcount_read(&lock->l_handle.h_ref),
2833 lock->l_readers, lock->l_writers,
2834 ldlm_lockname[lock->l_granted_mode],
2835 ldlm_lockname[lock->l_req_mode],
2837 atomic_read(&resource->lr_refcount),
2838 ldlm_typename[resource->lr_type],
2839 lock->l_policy_data.l_flock.pid,
2840 lock->l_policy_data.l_flock.start,
2841 lock->l_policy_data.l_flock.end,
2843 lock->l_remote_handle.cookie,
2844 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2845 lock->l_pid, lock->l_callback_timeout);
2849 libcfs_debug_msg(msgdata,
2850 "%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",
2852 ldlm_lock_to_ns_name(lock),
2853 lock, lock->l_handle.h_cookie,
2854 refcount_read(&lock->l_handle.h_ref),
2855 lock->l_readers, lock->l_writers,
2856 ldlm_lockname[lock->l_granted_mode],
2857 ldlm_lockname[lock->l_req_mode],
2859 lock->l_policy_data.l_inodebits.bits,
2860 lock->l_policy_data.l_inodebits.try_bits,
2861 atomic_read(&resource->lr_refcount),
2862 ldlm_typename[resource->lr_type],
2864 lock->l_remote_handle.cookie,
2865 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2866 lock->l_pid, lock->l_callback_timeout,
2871 libcfs_debug_msg(msgdata,
2872 "%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",
2874 ldlm_lock_to_ns_name(lock),
2875 lock, lock->l_handle.h_cookie,
2876 refcount_read(&lock->l_handle.h_ref),
2877 lock->l_readers, lock->l_writers,
2878 ldlm_lockname[lock->l_granted_mode],
2879 ldlm_lockname[lock->l_req_mode],
2881 atomic_read(&resource->lr_refcount),
2882 ldlm_typename[resource->lr_type],
2884 lock->l_remote_handle.cookie,
2885 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2886 lock->l_pid, lock->l_callback_timeout,
2891 ldlm_resource_putref(resource);
2893 EXPORT_SYMBOL(_ldlm_lock_debug);