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 extern struct kmem_cache *ldlm_lock_slab;
141 #ifdef HAVE_SERVER_SUPPORT
142 static ldlm_processing_policy ldlm_processing_policy_table[] = {
143 [LDLM_PLAIN] = ldlm_process_plain_lock,
144 [LDLM_EXTENT] = ldlm_process_extent_lock,
145 [LDLM_FLOCK] = ldlm_process_flock_lock,
146 [LDLM_IBITS] = ldlm_process_inodebits_lock,
149 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
151 return ldlm_processing_policy_table[res->lr_type];
153 EXPORT_SYMBOL(ldlm_get_processing_policy);
155 static ldlm_reprocessing_policy ldlm_reprocessing_policy_table[] = {
156 [LDLM_PLAIN] = ldlm_reprocess_queue,
157 [LDLM_EXTENT] = ldlm_reprocess_queue,
158 [LDLM_FLOCK] = ldlm_reprocess_queue,
159 [LDLM_IBITS] = ldlm_reprocess_inodebits_queue,
162 ldlm_reprocessing_policy ldlm_get_reprocessing_policy(struct ldlm_resource *res)
164 return ldlm_reprocessing_policy_table[res->lr_type];
167 #endif /* HAVE_SERVER_SUPPORT */
169 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
173 EXPORT_SYMBOL(ldlm_register_intent);
176 * REFCOUNTED LOCK OBJECTS
181 * Get a reference on a lock.
183 * Lock refcounts, during creation:
184 * - one special one for allocation, dec'd only once in destroy
185 * - one for being a lock that's in-use
186 * - one for the addref associated with a new lock
188 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
190 atomic_inc(&lock->l_refc);
193 EXPORT_SYMBOL(ldlm_lock_get);
196 * Release lock reference.
198 * Also frees the lock if it was last reference.
200 void ldlm_lock_put(struct ldlm_lock *lock)
204 LASSERT(lock->l_resource != LP_POISON);
205 LASSERT(atomic_read(&lock->l_refc) > 0);
206 if (atomic_dec_and_test(&lock->l_refc)) {
207 struct ldlm_resource *res;
210 "final lock_put on destroyed lock, freeing it.");
212 res = lock->l_resource;
213 LASSERT(ldlm_is_destroyed(lock));
214 LASSERT(list_empty(&lock->l_exp_list));
215 LASSERT(list_empty(&lock->l_res_link));
216 LASSERT(list_empty(&lock->l_pending_chain));
218 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
220 lu_ref_del(&res->lr_reference, "lock", lock);
221 if (lock->l_export) {
222 class_export_lock_put(lock->l_export, lock);
223 lock->l_export = NULL;
226 if (lock->l_lvb_data != NULL)
227 OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
229 if (res->lr_type == LDLM_EXTENT) {
230 ldlm_interval_free(ldlm_interval_detach(lock));
231 } else if (res->lr_type == LDLM_IBITS) {
232 if (lock->l_ibits_node != NULL)
233 OBD_SLAB_FREE_PTR(lock->l_ibits_node,
234 ldlm_inodebits_slab);
236 ldlm_resource_putref(res);
237 lock->l_resource = NULL;
238 lu_ref_fini(&lock->l_reference);
239 OBD_FREE_RCU(lock, sizeof(*lock), &lock->l_handle);
244 EXPORT_SYMBOL(ldlm_lock_put);
247 * Removes LDLM lock \a lock from LRU. Assumes LRU is already locked.
249 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
252 if (!list_empty(&lock->l_lru)) {
253 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
255 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
256 if (ns->ns_last_pos == &lock->l_lru)
257 ns->ns_last_pos = lock->l_lru.prev;
258 list_del_init(&lock->l_lru);
259 LASSERT(ns->ns_nr_unused > 0);
267 * Removes LDLM lock \a lock from LRU. Obtains the LRU lock first.
269 * If \a last_use is non-zero, it will remove the lock from LRU only if
270 * it matches lock's l_last_used.
272 * \retval 0 if \a last_use is set, the lock is not in LRU list or \a last_use
273 * doesn't match lock's l_last_used;
274 * otherwise, the lock hasn't been in the LRU list.
275 * \retval 1 the lock was in LRU list and removed.
277 int ldlm_lock_remove_from_lru_check(struct ldlm_lock *lock, ktime_t last_use)
279 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
283 if (ldlm_is_ns_srv(lock)) {
284 LASSERT(list_empty(&lock->l_lru));
288 spin_lock(&ns->ns_lock);
289 if (!ktime_compare(last_use, ktime_set(0, 0)) ||
290 !ktime_compare(last_use, lock->l_last_used))
291 rc = ldlm_lock_remove_from_lru_nolock(lock);
292 spin_unlock(&ns->ns_lock);
298 * Adds LDLM lock \a lock to namespace LRU. Assumes LRU is already locked.
300 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
302 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
304 lock->l_last_used = ktime_get();
305 LASSERT(list_empty(&lock->l_lru));
306 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
307 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
308 LASSERT(ns->ns_nr_unused >= 0);
313 * Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks
316 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
318 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
321 spin_lock(&ns->ns_lock);
322 ldlm_lock_add_to_lru_nolock(lock);
323 spin_unlock(&ns->ns_lock);
328 * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
329 * the LRU. Performs necessary LRU locking
331 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
333 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
336 if (ldlm_is_ns_srv(lock)) {
337 LASSERT(list_empty(&lock->l_lru));
342 spin_lock(&ns->ns_lock);
343 if (!list_empty(&lock->l_lru)) {
344 ldlm_lock_remove_from_lru_nolock(lock);
345 ldlm_lock_add_to_lru_nolock(lock);
347 spin_unlock(&ns->ns_lock);
352 * Helper to destroy a locked lock.
354 * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
355 * Must be called with l_lock and lr_lock held.
357 * Does not actually free the lock data, but rather marks the lock as
358 * destroyed by setting l_destroyed field in the lock to 1. Destroys a
359 * handle->lock association too, so that the lock can no longer be found
360 * and removes the lock from LRU list. Actual lock freeing occurs when
361 * last lock reference goes away.
363 * Original comment (of some historical value):
364 * This used to have a 'strict' flag, which recovery would use to mark an
365 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
366 * shall explain why it's gone: with the new hash table scheme, once you call
367 * ldlm_lock_destroy, you can never drop your final references on this lock.
368 * Because it's not in the hash table anymore. -phil
370 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
374 if (lock->l_readers || lock->l_writers) {
375 LDLM_ERROR(lock, "lock still has references");
379 if (!list_empty(&lock->l_res_link)) {
380 LDLM_ERROR(lock, "lock still on resource");
384 if (ldlm_is_destroyed(lock)) {
385 LASSERT(list_empty(&lock->l_lru));
389 ldlm_set_destroyed(lock);
391 if (lock->l_export && lock->l_export->exp_lock_hash) {
392 /* NB: it's safe to call cfs_hash_del() even lock isn't
393 * in exp_lock_hash. */
394 /* In the function below, .hs_keycmp resolves to
395 * ldlm_export_lock_keycmp() */
396 /* coverity[overrun-buffer-val] */
397 cfs_hash_del(lock->l_export->exp_lock_hash,
398 &lock->l_remote_handle, &lock->l_exp_hash);
401 ldlm_lock_remove_from_lru(lock);
402 class_handle_unhash(&lock->l_handle);
409 * Destroys a LDLM lock \a lock. Performs necessary locking first.
411 void ldlm_lock_destroy(struct ldlm_lock *lock)
415 lock_res_and_lock(lock);
416 first = ldlm_lock_destroy_internal(lock);
417 unlock_res_and_lock(lock);
419 /* drop reference from hashtable only for first destroy */
421 lu_ref_del(&lock->l_reference, "hash", lock);
422 LDLM_LOCK_RELEASE(lock);
428 * Destroys a LDLM lock \a lock that is already locked.
430 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
434 first = ldlm_lock_destroy_internal(lock);
435 /* drop reference from hashtable only for first destroy */
437 lu_ref_del(&lock->l_reference, "hash", lock);
438 LDLM_LOCK_RELEASE(lock);
443 /* this is called by portals_handle2object with the handle lock taken */
444 static void lock_handle_addref(void *lock)
446 LDLM_LOCK_GET((struct ldlm_lock *)lock);
449 static void lock_handle_free(void *lock, int size)
451 LASSERT(size == sizeof(struct ldlm_lock));
452 OBD_SLAB_FREE(lock, ldlm_lock_slab, size);
455 static struct portals_handle_ops lock_handle_ops = {
456 .hop_addref = lock_handle_addref,
457 .hop_free = lock_handle_free,
462 * Allocate and initialize new lock structure.
464 * usage: pass in a resource on which you have done ldlm_resource_get
465 * new lock will take over the refcount.
466 * returns: lock with refcount 2 - one for current caller and one for remote
468 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
470 struct ldlm_lock *lock;
473 if (resource == NULL)
476 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, GFP_NOFS);
480 spin_lock_init(&lock->l_lock);
481 lock->l_resource = resource;
482 lu_ref_add(&resource->lr_reference, "lock", lock);
484 atomic_set(&lock->l_refc, 2);
485 INIT_LIST_HEAD(&lock->l_res_link);
486 INIT_LIST_HEAD(&lock->l_lru);
487 INIT_LIST_HEAD(&lock->l_pending_chain);
488 INIT_LIST_HEAD(&lock->l_bl_ast);
489 INIT_LIST_HEAD(&lock->l_cp_ast);
490 INIT_LIST_HEAD(&lock->l_rk_ast);
491 init_waitqueue_head(&lock->l_waitq);
492 lock->l_blocking_lock = NULL;
493 INIT_LIST_HEAD(&lock->l_sl_mode);
494 INIT_LIST_HEAD(&lock->l_sl_policy);
495 INIT_HLIST_NODE(&lock->l_exp_hash);
496 INIT_HLIST_NODE(&lock->l_exp_flock_hash);
498 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
500 INIT_LIST_HEAD_RCU(&lock->l_handle.h_link);
501 class_handle_hash(&lock->l_handle, &lock_handle_ops);
503 lu_ref_init(&lock->l_reference);
504 lu_ref_add(&lock->l_reference, "hash", lock);
505 lock->l_callback_timeout = 0;
506 lock->l_activity = 0;
508 #if LUSTRE_TRACKS_LOCK_EXP_REFS
509 INIT_LIST_HEAD(&lock->l_exp_refs_link);
510 lock->l_exp_refs_nr = 0;
511 lock->l_exp_refs_target = NULL;
513 INIT_LIST_HEAD(&lock->l_exp_list);
519 * Moves LDLM lock \a lock to another resource.
520 * This is used on client when server returns some other lock than requested
521 * (typically as a result of intent operation)
523 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
524 const struct ldlm_res_id *new_resid)
526 struct ldlm_resource *oldres = lock->l_resource;
527 struct ldlm_resource *newres;
531 LASSERT(ns_is_client(ns));
533 lock_res_and_lock(lock);
534 if (memcmp(new_resid, &lock->l_resource->lr_name,
535 sizeof(lock->l_resource->lr_name)) == 0) {
537 unlock_res_and_lock(lock);
541 LASSERT(new_resid->name[0] != 0);
543 /* This function assumes that the lock isn't on any lists */
544 LASSERT(list_empty(&lock->l_res_link));
546 type = oldres->lr_type;
547 unlock_res_and_lock(lock);
549 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
551 RETURN(PTR_ERR(newres));
553 lu_ref_add(&newres->lr_reference, "lock", lock);
555 * To flip the lock from the old to the new resource, lock, oldres and
556 * newres have to be locked. Resource spin-locks are nested within
557 * lock->l_lock, and are taken in the memory address order to avoid
560 spin_lock(&lock->l_lock);
561 oldres = lock->l_resource;
562 if (oldres < newres) {
564 lock_res_nested(newres, LRT_NEW);
567 lock_res_nested(oldres, LRT_NEW);
569 LASSERT(memcmp(new_resid, &oldres->lr_name,
570 sizeof oldres->lr_name) != 0);
571 lock->l_resource = newres;
573 unlock_res_and_lock(lock);
575 /* ...and the flowers are still standing! */
576 lu_ref_del(&oldres->lr_reference, "lock", lock);
577 ldlm_resource_putref(oldres);
582 /** \defgroup ldlm_handles LDLM HANDLES
583 * Ways to get hold of locks without any addresses.
588 * Fills in handle for LDLM lock \a lock into supplied \a lockh
589 * Does not take any references.
591 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
593 lockh->cookie = lock->l_handle.h_cookie;
595 EXPORT_SYMBOL(ldlm_lock2handle);
598 * Obtain a lock reference by handle.
600 * if \a flags: atomically get the lock and set the flags.
601 * Return NULL if flag already set
603 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
606 struct ldlm_lock *lock;
611 lock = class_handle2object(handle->cookie, NULL);
615 if (lock->l_export != NULL && lock->l_export->exp_failed) {
616 CDEBUG(D_INFO, "lock export failed: lock %p, exp %p\n",
617 lock, lock->l_export);
622 /* It's unlikely but possible that someone marked the lock as
623 * destroyed after we did handle2object on it */
624 if ((flags == 0) && !ldlm_is_destroyed(lock)) {
625 lu_ref_add(&lock->l_reference, "handle", current);
629 lock_res_and_lock(lock);
631 LASSERT(lock->l_resource != NULL);
633 lu_ref_add_atomic(&lock->l_reference, "handle", current);
634 if (unlikely(ldlm_is_destroyed(lock))) {
635 unlock_res_and_lock(lock);
636 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
641 /* If we're setting flags, make sure none of them are already set. */
643 if ((lock->l_flags & flags) != 0) {
644 unlock_res_and_lock(lock);
649 lock->l_flags |= flags;
652 unlock_res_and_lock(lock);
655 EXPORT_SYMBOL(__ldlm_handle2lock);
656 /** @} ldlm_handles */
659 * Fill in "on the wire" representation for given LDLM lock into supplied
660 * lock descriptor \a desc structure.
662 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
664 ldlm_res2desc(lock->l_resource, &desc->l_resource);
665 desc->l_req_mode = lock->l_req_mode;
666 desc->l_granted_mode = lock->l_granted_mode;
667 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
668 &lock->l_policy_data,
669 &desc->l_policy_data);
673 * Add a lock to list of conflicting locks to send AST to.
675 * Only add if we have not sent a blocking AST to the lock yet.
677 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
678 struct list_head *work_list)
680 if (!ldlm_is_ast_sent(lock)) {
681 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
682 ldlm_set_ast_sent(lock);
683 /* If the enqueuing client said so, tell the AST recipient to
684 * discard dirty data, rather than writing back. */
685 if (ldlm_is_ast_discard_data(new))
686 ldlm_set_discard_data(lock);
688 /* Lock can be converted from a blocking state back to granted
689 * after lock convert or COS downgrade but still be in an
690 * older bl_list because it is controlled only by
691 * ldlm_work_bl_ast_lock(), let it be processed there.
693 if (list_empty(&lock->l_bl_ast)) {
694 list_add(&lock->l_bl_ast, work_list);
697 LASSERT(lock->l_blocking_lock == NULL);
698 lock->l_blocking_lock = LDLM_LOCK_GET(new);
703 * Add a lock to list of just granted locks to send completion AST to.
705 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
706 struct list_head *work_list)
708 if (!ldlm_is_cp_reqd(lock)) {
709 ldlm_set_cp_reqd(lock);
710 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
711 LASSERT(list_empty(&lock->l_cp_ast));
712 list_add(&lock->l_cp_ast, work_list);
718 * Aggregator function to add AST work items into a list. Determines
719 * what sort of an AST work needs to be done and calls the proper
721 * Must be called with lr_lock held.
723 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
724 struct list_head *work_list)
727 check_res_locked(lock->l_resource);
729 ldlm_add_bl_work_item(lock, new, work_list);
731 ldlm_add_cp_work_item(lock, work_list);
736 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
737 * r/w reference type is determined by \a mode
738 * Calls ldlm_lock_addref_internal.
740 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
742 struct ldlm_lock *lock;
744 lock = ldlm_handle2lock(lockh);
745 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
746 ldlm_lock_addref_internal(lock, mode);
749 EXPORT_SYMBOL(ldlm_lock_addref);
753 * Add specified reader/writer reference to LDLM lock \a lock.
754 * r/w reference type is determined by \a mode
755 * Removes lock from LRU if it is there.
756 * Assumes the LDLM lock is already locked.
758 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
761 ldlm_lock_remove_from_lru(lock);
762 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
764 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
766 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
768 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
771 lu_ref_add_atomic(&lock->l_reference, "user", lock);
772 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
776 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
777 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
779 * \retval 0 success, lock was addref-ed
781 * \retval -EAGAIN lock is being canceled.
783 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
785 struct ldlm_lock *lock;
789 lock = ldlm_handle2lock(lockh);
791 lock_res_and_lock(lock);
792 if (lock->l_readers != 0 || lock->l_writers != 0 ||
793 !ldlm_is_cbpending(lock)) {
794 ldlm_lock_addref_internal_nolock(lock, mode);
797 unlock_res_and_lock(lock);
802 EXPORT_SYMBOL(ldlm_lock_addref_try);
805 * Add specified reader/writer reference to LDLM lock \a lock.
806 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
807 * Only called for local locks.
809 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
811 lock_res_and_lock(lock);
812 ldlm_lock_addref_internal_nolock(lock, mode);
813 unlock_res_and_lock(lock);
817 * Removes reader/writer reference for LDLM lock \a lock.
818 * Assumes LDLM lock is already locked.
819 * only called in ldlm_flock_destroy and for local locks.
820 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
821 * that cannot be placed in LRU.
823 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
826 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
827 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
828 LASSERT(lock->l_readers > 0);
829 lu_ref_del(&lock->l_reference, "reader", lock);
832 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
833 LASSERT(lock->l_writers > 0);
834 lu_ref_del(&lock->l_reference, "writer", lock);
838 lu_ref_del(&lock->l_reference, "user", lock);
839 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
843 * Removes reader/writer reference for LDLM lock \a lock.
844 * Locks LDLM lock first.
845 * If the lock is determined to be client lock on a client and r/w refcount
846 * drops to zero and the lock is not blocked, the lock is added to LRU lock
848 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
850 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
852 struct ldlm_namespace *ns;
855 lock_res_and_lock(lock);
857 ns = ldlm_lock_to_ns(lock);
859 ldlm_lock_decref_internal_nolock(lock, mode);
861 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
862 !lock->l_readers && !lock->l_writers) {
863 /* If this is a local lock on a server namespace and this was
864 * the last reference, cancel the lock.
866 * Group locks are special:
867 * They must not go in LRU, but they are not called back
868 * like non-group locks, instead they are manually released.
869 * They have an l_writers reference which they keep until
870 * they are manually released, so we remove them when they have
871 * no more reader or writer references. - LU-6368 */
872 ldlm_set_cbpending(lock);
875 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
876 /* If we received a blocked AST and this was the last reference,
877 * run the callback. */
878 if (ldlm_is_ns_srv(lock) && lock->l_export)
879 CERROR("FL_CBPENDING set on non-local lock--just a "
882 LDLM_DEBUG(lock, "final decref done on cbpending lock");
884 LDLM_LOCK_GET(lock); /* dropped by bl thread */
885 ldlm_lock_remove_from_lru(lock);
886 unlock_res_and_lock(lock);
888 if (ldlm_is_fail_loc(lock))
889 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
891 if (ldlm_is_atomic_cb(lock) ||
892 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
893 ldlm_handle_bl_callback(ns, NULL, lock);
894 } else if (ns_is_client(ns) &&
895 !lock->l_readers && !lock->l_writers &&
896 !ldlm_is_no_lru(lock) &&
897 !ldlm_is_bl_ast(lock) &&
898 !ldlm_is_converting(lock)) {
900 LDLM_DEBUG(lock, "add lock into lru list");
902 /* If this is a client-side namespace and this was the last
903 * reference, put it on the LRU. */
904 ldlm_lock_add_to_lru(lock);
905 unlock_res_and_lock(lock);
907 if (ldlm_is_fail_loc(lock))
908 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
910 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
911 * are not supported by the server, otherwise, it is done on
913 if (!exp_connect_cancelset(lock->l_conn_export) &&
914 !ns_connect_lru_resize(ns))
915 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
917 LDLM_DEBUG(lock, "do not add lock into lru list");
918 unlock_res_and_lock(lock);
925 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
927 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
929 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
930 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
931 ldlm_lock_decref_internal(lock, mode);
934 EXPORT_SYMBOL(ldlm_lock_decref);
937 * Decrease reader/writer refcount for LDLM lock with handle
938 * \a lockh and mark it for subsequent cancellation once r/w refcount
939 * drops to zero instead of putting into LRU.
942 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
945 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
948 LASSERT(lock != NULL);
950 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
951 lock_res_and_lock(lock);
952 ldlm_set_cbpending(lock);
953 unlock_res_and_lock(lock);
954 ldlm_lock_decref_internal(lock, mode);
957 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
959 struct sl_insert_point {
960 struct list_head *res_link;
961 struct list_head *mode_link;
962 struct list_head *policy_link;
966 * Finds a position to insert the new lock into granted lock list.
968 * Used for locks eligible for skiplist optimization.
971 * queue [input]: the granted list where search acts on;
972 * req [input]: the lock whose position to be located;
973 * prev [output]: positions within 3 lists to insert @req to
977 * - ldlm_grant_lock_with_skiplist
979 static void search_granted_lock(struct list_head *queue,
980 struct ldlm_lock *req,
981 struct sl_insert_point *prev)
983 struct list_head *tmp;
984 struct ldlm_lock *lock, *mode_end, *policy_end;
987 list_for_each(tmp, queue) {
988 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
990 mode_end = list_entry(lock->l_sl_mode.prev,
991 struct ldlm_lock, l_sl_mode);
993 if (lock->l_req_mode != req->l_req_mode) {
994 /* jump to last lock of mode group */
995 tmp = &mode_end->l_res_link;
999 /* suitable mode group is found */
1000 if (lock->l_resource->lr_type == LDLM_PLAIN) {
1001 /* insert point is last lock of the mode group */
1002 prev->res_link = &mode_end->l_res_link;
1003 prev->mode_link = &mode_end->l_sl_mode;
1004 prev->policy_link = &req->l_sl_policy;
1007 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
1010 list_entry(lock->l_sl_policy.prev,
1014 if (lock->l_policy_data.l_inodebits.bits ==
1015 req->l_policy_data.l_inodebits.bits) {
1016 /* insert point is last lock of
1017 * the policy group */
1019 &policy_end->l_res_link;
1021 &policy_end->l_sl_mode;
1023 &policy_end->l_sl_policy;
1028 if (policy_end == mode_end)
1029 /* done with mode group */
1032 /* go to next policy group within mode group */
1033 tmp = policy_end->l_res_link.next;
1034 lock = list_entry(tmp, struct ldlm_lock,
1036 } /* loop over policy groups within the mode group */
1038 /* insert point is last lock of the mode group,
1039 * new policy group is started */
1040 prev->res_link = &mode_end->l_res_link;
1041 prev->mode_link = &mode_end->l_sl_mode;
1042 prev->policy_link = &req->l_sl_policy;
1046 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1051 /* insert point is last lock on the queue,
1052 * new mode group and new policy group are started */
1053 prev->res_link = queue->prev;
1054 prev->mode_link = &req->l_sl_mode;
1055 prev->policy_link = &req->l_sl_policy;
1061 * Add a lock into resource granted list after a position described by
1064 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1065 struct sl_insert_point *prev)
1067 struct ldlm_resource *res = lock->l_resource;
1070 check_res_locked(res);
1072 ldlm_resource_dump(D_INFO, res);
1073 LDLM_DEBUG(lock, "About to add lock:");
1075 if (ldlm_is_destroyed(lock)) {
1076 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1080 LASSERT(list_empty(&lock->l_res_link));
1081 LASSERT(list_empty(&lock->l_sl_mode));
1082 LASSERT(list_empty(&lock->l_sl_policy));
1085 * lock->link == prev->link means lock is first starting the group.
1086 * Don't re-add to itself to suppress kernel warnings.
1088 if (&lock->l_res_link != prev->res_link)
1089 list_add(&lock->l_res_link, prev->res_link);
1090 if (&lock->l_sl_mode != prev->mode_link)
1091 list_add(&lock->l_sl_mode, prev->mode_link);
1092 if (&lock->l_sl_policy != prev->policy_link)
1093 list_add(&lock->l_sl_policy, prev->policy_link);
1099 * Add a lock to granted list on a resource maintaining skiplist
1102 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1104 struct sl_insert_point prev;
1106 LASSERT(ldlm_is_granted(lock));
1108 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1109 ldlm_granted_list_add_lock(lock, &prev);
1113 * Perform lock granting bookkeeping.
1115 * Includes putting the lock into granted list and updating lock mode.
1117 * - ldlm_lock_enqueue
1118 * - ldlm_reprocess_queue
1120 * must be called with lr_lock held
1122 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1124 struct ldlm_resource *res = lock->l_resource;
1127 check_res_locked(res);
1129 lock->l_granted_mode = lock->l_req_mode;
1131 if (work_list && lock->l_completion_ast != NULL)
1132 ldlm_add_ast_work_item(lock, NULL, work_list);
1134 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1135 ldlm_grant_lock_with_skiplist(lock);
1136 else if (res->lr_type == LDLM_EXTENT)
1137 ldlm_extent_add_lock(res, lock);
1138 else if (res->lr_type == LDLM_FLOCK) {
1139 /* We should not add locks to granted list in the following
1141 * - this is an UNLOCK but not a real lock;
1142 * - this is a TEST lock;
1143 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1144 * - this is a deadlock (flock cannot be granted) */
1145 if (lock->l_req_mode == 0 ||
1146 lock->l_req_mode == LCK_NL ||
1147 ldlm_is_test_lock(lock) ||
1148 ldlm_is_flock_deadlock(lock))
1150 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1155 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1160 * Check if the given @lock meets the criteria for a match.
1161 * A reference on the lock is taken if matched.
1163 * \param lock test-against this lock
1164 * \param data parameters
1166 static int lock_matches(struct ldlm_lock *lock, struct ldlm_match_data *data)
1168 union ldlm_policy_data *lpol = &lock->l_policy_data;
1169 enum ldlm_mode match = LCK_MINMODE;
1171 if (lock == data->lmd_old)
1172 return INTERVAL_ITER_STOP;
1174 /* Check if this lock can be matched.
1175 * Used by LU-2919(exclusive open) for open lease lock */
1176 if (ldlm_is_excl(lock))
1177 return INTERVAL_ITER_CONT;
1179 /* llite sometimes wants to match locks that will be
1180 * canceled when their users drop, but we allow it to match
1181 * if it passes in CBPENDING and the lock still has users.
1182 * this is generally only going to be used by children
1183 * whose parents already hold a lock so forward progress
1184 * can still happen. */
1185 if (ldlm_is_cbpending(lock) &&
1186 !(data->lmd_flags & LDLM_FL_CBPENDING))
1187 return INTERVAL_ITER_CONT;
1188 if (!data->lmd_unref && ldlm_is_cbpending(lock) &&
1189 lock->l_readers == 0 && lock->l_writers == 0)
1190 return INTERVAL_ITER_CONT;
1192 if (!(lock->l_req_mode & *data->lmd_mode))
1193 return INTERVAL_ITER_CONT;
1195 /* When we search for ast_data, we are not doing a traditional match,
1196 * so we don't worry about IBITS or extent matching.
1198 if (data->lmd_has_ast_data) {
1199 if (!lock->l_ast_data)
1200 return INTERVAL_ITER_CONT;
1205 match = lock->l_req_mode;
1207 switch (lock->l_resource->lr_type) {
1209 if (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1210 lpol->l_extent.end < data->lmd_policy->l_extent.end)
1211 return INTERVAL_ITER_CONT;
1213 if (unlikely(match == LCK_GROUP) &&
1214 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1215 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1216 return INTERVAL_ITER_CONT;
1219 /* We match if we have existing lock with same or wider set
1221 if ((lpol->l_inodebits.bits &
1222 data->lmd_policy->l_inodebits.bits) !=
1223 data->lmd_policy->l_inodebits.bits)
1224 return INTERVAL_ITER_CONT;
1230 /* We match if we have existing lock with same or wider set
1232 if (!data->lmd_unref && LDLM_HAVE_MASK(lock, GONE))
1233 return INTERVAL_ITER_CONT;
1235 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1236 return INTERVAL_ITER_CONT;
1238 /* Filter locks by skipping flags */
1239 if (data->lmd_skip_flags & lock->l_flags)
1240 return INTERVAL_ITER_CONT;
1243 if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1244 LDLM_LOCK_GET(lock);
1245 ldlm_lock_touch_in_lru(lock);
1247 ldlm_lock_addref_internal_nolock(lock, match);
1250 *data->lmd_mode = match;
1251 data->lmd_lock = lock;
1253 return INTERVAL_ITER_STOP;
1256 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1258 struct ldlm_interval *node = to_ldlm_interval(in);
1259 struct ldlm_match_data *data = args;
1260 struct ldlm_lock *lock;
1263 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1264 rc = lock_matches(lock, data);
1265 if (rc == INTERVAL_ITER_STOP)
1266 return INTERVAL_ITER_STOP;
1268 return INTERVAL_ITER_CONT;
1272 * Search for a lock with given parameters in interval trees.
1274 * \param res search for a lock in this resource
1275 * \param data parameters
1277 * \retval a referenced lock or NULL.
1279 struct ldlm_lock *search_itree(struct ldlm_resource *res,
1280 struct ldlm_match_data *data)
1282 struct interval_node_extent ext = {
1283 .start = data->lmd_policy->l_extent.start,
1284 .end = data->lmd_policy->l_extent.end
1288 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1289 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1291 if (tree->lit_root == NULL)
1294 if (!(tree->lit_mode & *data->lmd_mode))
1297 interval_search(tree->lit_root, &ext,
1298 itree_overlap_cb, data);
1300 return data->lmd_lock;
1302 EXPORT_SYMBOL(search_itree);
1306 * Search for a lock with given properties in a queue.
1308 * \param queue search for a lock in this queue
1309 * \param data parameters
1311 * \retval a referenced lock or NULL.
1313 static struct ldlm_lock *search_queue(struct list_head *queue,
1314 struct ldlm_match_data *data)
1316 struct ldlm_lock *lock;
1319 list_for_each_entry(lock, queue, l_res_link) {
1320 rc = lock_matches(lock, data);
1321 if (rc == INTERVAL_ITER_STOP)
1322 return data->lmd_lock;
1327 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1329 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1330 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1331 wake_up_all(&lock->l_waitq);
1334 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1336 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1338 lock_res_and_lock(lock);
1339 ldlm_lock_fail_match_locked(lock);
1340 unlock_res_and_lock(lock);
1344 * Mark lock as "matchable" by OST.
1346 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1348 * Assumes LDLM lock is already locked.
1350 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1352 ldlm_set_lvb_ready(lock);
1353 wake_up_all(&lock->l_waitq);
1355 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1358 * Mark lock as "matchable" by OST.
1359 * Locks the lock and then \see ldlm_lock_allow_match_locked
1361 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1363 lock_res_and_lock(lock);
1364 ldlm_lock_allow_match_locked(lock);
1365 unlock_res_and_lock(lock);
1367 EXPORT_SYMBOL(ldlm_lock_allow_match);
1370 * Attempt to find a lock with specified properties.
1372 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1375 * Can be called in two ways:
1377 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1378 * for a duplicate of.
1380 * Otherwise, all of the fields must be filled in, to match against.
1382 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1383 * server (ie, connh is NULL)
1384 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1385 * list will be considered
1386 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1387 * to be canceled can still be matched as long as they still have reader
1388 * or writer refernces
1389 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1390 * just tell us if we would have matched.
1392 * \retval 1 if it finds an already-existing lock that is compatible; in this
1393 * case, lockh is filled in with a addref()ed lock
1395 * We also check security context, and if that fails we simply return 0 (to
1396 * keep caller code unchanged), the context failure will be discovered by
1397 * caller sometime later.
1399 enum ldlm_mode ldlm_lock_match_with_skip(struct ldlm_namespace *ns,
1400 __u64 flags, __u64 skip_flags,
1401 const struct ldlm_res_id *res_id,
1402 enum ldlm_type type,
1403 union ldlm_policy_data *policy,
1404 enum ldlm_mode mode,
1405 struct lustre_handle *lockh, int unref)
1407 struct ldlm_match_data data = {
1411 .lmd_policy = policy,
1413 .lmd_skip_flags = skip_flags,
1415 .lmd_has_ast_data = false,
1417 struct ldlm_resource *res;
1418 struct ldlm_lock *lock;
1424 data.lmd_old = ldlm_handle2lock(lockh);
1425 LASSERT(data.lmd_old != NULL);
1427 ns = ldlm_lock_to_ns(data.lmd_old);
1428 res_id = &data.lmd_old->l_resource->lr_name;
1429 type = data.lmd_old->l_resource->lr_type;
1430 *data.lmd_mode = data.lmd_old->l_req_mode;
1433 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1435 LASSERT(data.lmd_old == NULL);
1439 LDLM_RESOURCE_ADDREF(res);
1441 if (res->lr_type == LDLM_EXTENT)
1442 lock = search_itree(res, &data);
1444 lock = search_queue(&res->lr_granted, &data);
1445 if (!lock && !(flags & LDLM_FL_BLOCK_GRANTED))
1446 lock = search_queue(&res->lr_waiting, &data);
1447 matched = lock ? mode : 0;
1449 LDLM_RESOURCE_DELREF(res);
1450 ldlm_resource_putref(res);
1453 ldlm_lock2handle(lock, lockh);
1454 if ((flags & LDLM_FL_LVB_READY) &&
1455 (!ldlm_is_lvb_ready(lock))) {
1456 __u64 wait_flags = LDLM_FL_LVB_READY |
1457 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1458 struct l_wait_info lwi;
1460 if (lock->l_completion_ast) {
1461 int err = lock->l_completion_ast(lock,
1462 LDLM_FL_WAIT_NOREPROC,
1465 GOTO(out_fail_match, matched = 0);
1468 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1469 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1471 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1472 l_wait_event(lock->l_waitq, lock->l_flags & wait_flags,
1474 if (!ldlm_is_lvb_ready(lock))
1475 GOTO(out_fail_match, matched = 0);
1478 /* check user's security context */
1479 if (lock->l_conn_export &&
1480 sptlrpc_import_check_ctx(
1481 class_exp2cliimp(lock->l_conn_export)))
1482 GOTO(out_fail_match, matched = 0);
1484 LDLM_DEBUG(lock, "matched (%llu %llu)",
1485 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1486 res_id->name[2] : policy->l_extent.start,
1487 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1488 res_id->name[3] : policy->l_extent.end);
1491 if (flags & LDLM_FL_TEST_LOCK)
1492 LDLM_LOCK_RELEASE(lock);
1494 ldlm_lock_decref_internal(lock, mode);
1497 /* less verbose for test-only */
1498 if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
1499 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1500 "%llu/%llu (%llu %llu)", ns,
1501 type, mode, res_id->name[0], res_id->name[1],
1502 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1503 res_id->name[2] : policy->l_extent.start,
1504 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1505 res_id->name[3] : policy->l_extent.end);
1507 if (data.lmd_old != NULL)
1508 LDLM_LOCK_PUT(data.lmd_old);
1512 EXPORT_SYMBOL(ldlm_lock_match_with_skip);
1514 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1517 struct ldlm_lock *lock;
1518 enum ldlm_mode mode = 0;
1521 lock = ldlm_handle2lock(lockh);
1523 lock_res_and_lock(lock);
1524 if (LDLM_HAVE_MASK(lock, GONE))
1527 if (ldlm_is_cbpending(lock) &&
1528 lock->l_readers == 0 && lock->l_writers == 0)
1532 *bits = lock->l_policy_data.l_inodebits.bits;
1533 mode = lock->l_granted_mode;
1534 ldlm_lock_addref_internal_nolock(lock, mode);
1541 unlock_res_and_lock(lock);
1542 LDLM_LOCK_PUT(lock);
1546 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1548 /** The caller must guarantee that the buffer is large enough. */
1549 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1550 enum req_location loc, void *data, int size)
1555 LASSERT(data != NULL);
1558 switch (lock->l_lvb_type) {
1560 if (size == sizeof(struct ost_lvb)) {
1561 if (loc == RCL_CLIENT)
1562 lvb = req_capsule_client_swab_get(pill,
1564 lustre_swab_ost_lvb);
1566 lvb = req_capsule_server_swab_get(pill,
1568 lustre_swab_ost_lvb);
1569 if (unlikely(lvb == NULL)) {
1570 LDLM_ERROR(lock, "no LVB");
1574 memcpy(data, lvb, size);
1575 } else if (size == sizeof(struct ost_lvb_v1)) {
1576 struct ost_lvb *olvb = data;
1578 if (loc == RCL_CLIENT)
1579 lvb = req_capsule_client_swab_get(pill,
1581 lustre_swab_ost_lvb_v1);
1583 lvb = req_capsule_server_sized_swab_get(pill,
1585 lustre_swab_ost_lvb_v1);
1586 if (unlikely(lvb == NULL)) {
1587 LDLM_ERROR(lock, "no LVB");
1591 memcpy(data, lvb, size);
1592 olvb->lvb_mtime_ns = 0;
1593 olvb->lvb_atime_ns = 0;
1594 olvb->lvb_ctime_ns = 0;
1596 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1602 if (size == sizeof(struct lquota_lvb)) {
1603 if (loc == RCL_CLIENT)
1604 lvb = req_capsule_client_swab_get(pill,
1606 lustre_swab_lquota_lvb);
1608 lvb = req_capsule_server_swab_get(pill,
1610 lustre_swab_lquota_lvb);
1611 if (unlikely(lvb == NULL)) {
1612 LDLM_ERROR(lock, "no LVB");
1616 memcpy(data, lvb, size);
1618 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1627 if (loc == RCL_CLIENT)
1628 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1630 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1631 if (unlikely(lvb == NULL)) {
1632 LDLM_ERROR(lock, "no LVB");
1636 memcpy(data, lvb, size);
1639 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1640 libcfs_debug_dumpstack(NULL);
1648 * Create and fill in new LDLM lock with specified properties.
1649 * Returns a referenced lock
1651 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1652 const struct ldlm_res_id *res_id,
1653 enum ldlm_type type,
1654 enum ldlm_mode mode,
1655 const struct ldlm_callback_suite *cbs,
1656 void *data, __u32 lvb_len,
1657 enum lvb_type lvb_type)
1659 struct ldlm_lock *lock;
1660 struct ldlm_resource *res;
1664 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1666 RETURN(ERR_CAST(res));
1668 lock = ldlm_lock_new(res);
1670 RETURN(ERR_PTR(-ENOMEM));
1672 lock->l_req_mode = mode;
1673 lock->l_ast_data = data;
1674 lock->l_pid = current_pid();
1675 if (ns_is_server(ns))
1676 ldlm_set_ns_srv(lock);
1678 lock->l_blocking_ast = cbs->lcs_blocking;
1679 lock->l_completion_ast = cbs->lcs_completion;
1680 lock->l_glimpse_ast = cbs->lcs_glimpse;
1685 rc = ldlm_extent_alloc_lock(lock);
1688 rc = ldlm_inodebits_alloc_lock(lock);
1697 lock->l_lvb_len = lvb_len;
1698 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1699 if (lock->l_lvb_data == NULL)
1700 GOTO(out, rc = -ENOMEM);
1703 lock->l_lvb_type = lvb_type;
1704 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1705 GOTO(out, rc = -ENOENT);
1710 ldlm_lock_destroy(lock);
1711 LDLM_LOCK_RELEASE(lock);
1712 RETURN(ERR_PTR(rc));
1715 #ifdef HAVE_SERVER_SUPPORT
1716 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1719 struct ldlm_resource *res = lock->l_resource;
1720 enum ldlm_error rc = ELDLM_OK;
1721 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1722 ldlm_processing_policy policy;
1726 policy = ldlm_get_processing_policy(res);
1728 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1729 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1730 res->lr_type != LDLM_FLOCK) {
1731 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1732 if (rc == -ERESTART)
1736 if (!list_empty(&rpc_list))
1737 ldlm_discard_bl_list(&rpc_list);
1744 * Enqueue (request) a lock.
1746 * Does not block. As a result of enqueue the lock would be put
1747 * into granted or waiting list.
1749 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1750 * set, skip all the enqueueing and delegate lock processing to intent policy
1753 enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
1754 struct ldlm_namespace *ns,
1755 struct ldlm_lock **lockp,
1756 void *cookie, __u64 *flags)
1758 struct ldlm_lock *lock = *lockp;
1759 struct ldlm_resource *res = lock->l_resource;
1760 int local = ns_is_client(ldlm_res_to_ns(res));
1761 enum ldlm_error rc = ELDLM_OK;
1762 struct ldlm_interval *node = NULL;
1765 /* policies are not executed on the client or during replay */
1766 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1767 && !local && ns->ns_policy) {
1768 rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
1770 if (rc == ELDLM_LOCK_REPLACED) {
1771 /* The lock that was returned has already been granted,
1772 * and placed into lockp. If it's not the same as the
1773 * one we passed in, then destroy the old one and our
1774 * work here is done. */
1775 if (lock != *lockp) {
1776 ldlm_lock_destroy(lock);
1777 LDLM_LOCK_RELEASE(lock);
1779 *flags |= LDLM_FL_LOCK_CHANGED;
1781 } else if (rc != ELDLM_OK &&
1782 ldlm_is_granted(lock)) {
1783 LASSERT(*flags & LDLM_FL_RESENT);
1784 /* It may happen that ns_policy returns an error in
1785 * resend case, object may be unlinked or just some
1786 * error occurs. It is unclear if lock reached the
1787 * client in the original reply, just leave the lock on
1788 * server, not returning it again to client. Due to
1789 * LU-6529, the server will not OOM. */
1791 } else if (rc != ELDLM_OK ||
1792 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1793 ldlm_lock_destroy(lock);
1798 if (*flags & LDLM_FL_RESENT) {
1799 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1800 * Set LOCK_CHANGED always.
1801 * Check if the lock is granted for BLOCK_GRANTED.
1802 * Take NO_TIMEOUT from the lock as it is inherited through
1803 * LDLM_FL_INHERIT_MASK */
1804 *flags |= LDLM_FL_LOCK_CHANGED;
1805 if (!ldlm_is_granted(lock))
1806 *flags |= LDLM_FL_BLOCK_GRANTED;
1807 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1811 /* For a replaying lock, it might be already in granted list. So
1812 * unlinking the lock will cause the interval node to be freed, we
1813 * have to allocate the interval node early otherwise we can't regrant
1814 * this lock in the future. - jay */
1815 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1816 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1818 lock_res_and_lock(lock);
1819 if (local && ldlm_is_granted(lock)) {
1820 /* The server returned a blocked lock, but it was granted
1821 * before we got a chance to actually enqueue it. We don't
1822 * need to do anything else. */
1823 *flags &= ~LDLM_FL_BLOCKED_MASK;
1824 GOTO(out, rc = ELDLM_OK);
1827 ldlm_resource_unlink_lock(lock);
1828 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1830 ldlm_lock_destroy_nolock(lock);
1831 GOTO(out, rc = -ENOMEM);
1834 INIT_LIST_HEAD(&node->li_group);
1835 ldlm_interval_attach(node, lock);
1839 /* Some flags from the enqueue want to make it into the AST, via the
1840 * lock's l_flags. */
1841 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1842 ldlm_set_ast_discard_data(lock);
1843 if (*flags & LDLM_FL_TEST_LOCK)
1844 ldlm_set_test_lock(lock);
1845 if (*flags & LDLM_FL_COS_INCOMPAT)
1846 ldlm_set_cos_incompat(lock);
1847 if (*flags & LDLM_FL_COS_ENABLED)
1848 ldlm_set_cos_enabled(lock);
1850 /* This distinction between local lock trees is very important; a client
1851 * namespace only has information about locks taken by that client, and
1852 * thus doesn't have enough information to decide for itself if it can
1853 * be granted (below). In this case, we do exactly what the server
1854 * tells us to do, as dictated by the 'flags'.
1856 * We do exactly the same thing during recovery, when the server is
1857 * more or less trusting the clients not to lie.
1859 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1862 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1863 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1865 ldlm_grant_lock(lock, NULL);
1866 GOTO(out, rc = ELDLM_OK);
1867 #ifdef HAVE_SERVER_SUPPORT
1868 } else if (*flags & LDLM_FL_REPLAY) {
1869 if (*flags & LDLM_FL_BLOCK_WAIT) {
1870 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1871 GOTO(out, rc = ELDLM_OK);
1872 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1873 ldlm_grant_lock(lock, NULL);
1874 GOTO(out, rc = ELDLM_OK);
1876 /* If no flags, fall through to normal enqueue path. */
1879 rc = ldlm_lock_enqueue_helper(lock, flags);
1883 CERROR("This is client-side-only module, cannot handle "
1884 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1890 unlock_res_and_lock(lock);
1892 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1896 #ifdef HAVE_SERVER_SUPPORT
1898 * Iterate through all waiting locks on a given resource queue and attempt to
1901 * Must be called with resource lock held.
1903 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1904 struct list_head *work_list,
1905 enum ldlm_process_intention intention,
1906 struct ldlm_lock *hint)
1908 struct list_head *tmp, *pos;
1909 ldlm_processing_policy policy;
1911 int rc = LDLM_ITER_CONTINUE;
1912 enum ldlm_error err;
1913 struct list_head bl_ast_list = LIST_HEAD_INIT(bl_ast_list);
1917 check_res_locked(res);
1919 policy = ldlm_get_processing_policy(res);
1921 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1922 intention == LDLM_PROCESS_RECOVERY);
1925 list_for_each_safe(tmp, pos, queue) {
1926 struct ldlm_lock *pending;
1927 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1929 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1931 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1934 rc = policy(pending, &flags, intention, &err, &rpc_list);
1935 if (pending->l_granted_mode == pending->l_req_mode ||
1936 res->lr_type == LDLM_FLOCK) {
1937 list_splice(&rpc_list, work_list);
1939 list_splice(&rpc_list, &bl_ast_list);
1942 * When this is called from recovery done, we always want
1943 * to scan the whole list no matter what 'rc' is returned.
1945 if (rc != LDLM_ITER_CONTINUE &&
1946 intention == LDLM_PROCESS_RESCAN)
1950 if (!list_empty(&bl_ast_list)) {
1953 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
1957 if (rc == -ERESTART)
1961 if (!list_empty(&bl_ast_list))
1962 ldlm_discard_bl_list(&bl_ast_list);
1964 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
1968 * Conflicting locks are detected for a lock to be enqueued, add the lock
1969 * into waiting list and send blocking ASTs to the conflicting locks.
1971 * \param[in] lock The lock to be enqueued.
1972 * \param[out] flags Lock flags for the lock to be enqueued.
1973 * \param[in] rpc_list Conflicting locks list.
1975 * \retval -ERESTART: Some lock was instantly canceled while sending
1976 * blocking ASTs, caller needs to re-check conflicting
1978 * \retval -EAGAIN: Lock was destroyed, caller should return error.
1979 * \reval 0: Lock is successfully added in waiting list.
1981 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
1982 struct list_head *rpc_list)
1984 struct ldlm_resource *res = lock->l_resource;
1988 check_res_locked(res);
1990 /* If either of the compat_queue()s returned failure, then we
1991 * have ASTs to send and must go onto the waiting list.
1993 * bug 2322: we used to unlink and re-add here, which was a
1994 * terrible folly -- if we goto restart, we could get
1995 * re-ordered! Causes deadlock, because ASTs aren't sent! */
1996 if (list_empty(&lock->l_res_link))
1997 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
2000 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
2003 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
2004 !ns_is_client(ldlm_res_to_ns(res)))
2005 class_fail_export(lock->l_export);
2008 if (rc == -ERESTART) {
2009 /* 15715: The lock was granted and destroyed after
2010 * resource lock was dropped. Interval node was freed
2011 * in ldlm_lock_destroy. Anyway, this always happens
2012 * when a client is being evicted. So it would be
2013 * ok to return an error. -jay */
2014 if (ldlm_is_destroyed(lock))
2017 /* lock was granted while resource was unlocked. */
2018 if (ldlm_is_granted(lock)) {
2019 /* bug 11300: if the lock has been granted,
2020 * break earlier because otherwise, we will go
2021 * to restart and ldlm_resource_unlink will be
2022 * called and it causes the interval node to be
2023 * freed. Then we will fail at
2024 * ldlm_extent_add_lock() */
2025 *flags &= ~LDLM_FL_BLOCKED_MASK;
2031 *flags |= LDLM_FL_BLOCK_GRANTED;
2037 * Discard all AST work items from list.
2039 * If for whatever reason we do not want to send ASTs to conflicting locks
2040 * anymore, disassemble the list with this function.
2042 void ldlm_discard_bl_list(struct list_head *bl_list)
2044 struct ldlm_lock *lock, *tmp;
2048 list_for_each_entry_safe(lock, tmp, bl_list, l_bl_ast) {
2049 LASSERT(!list_empty(&lock->l_bl_ast));
2050 list_del_init(&lock->l_bl_ast);
2051 ldlm_clear_ast_sent(lock);
2052 LASSERT(lock->l_bl_ast_run == 0);
2053 ldlm_clear_blocking_lock(lock);
2054 LDLM_LOCK_RELEASE(lock);
2060 * Process a call to blocking AST callback for a lock in ast_work list
2063 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2065 struct ldlm_cb_set_arg *arg = opaq;
2066 struct ldlm_lock *lock;
2067 struct ldlm_lock_desc d;
2068 struct ldlm_bl_desc bld;
2073 if (list_empty(arg->list))
2076 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2078 /* nobody should touch l_bl_ast but some locks in the list may become
2079 * granted after lock convert or COS downgrade, these locks should be
2080 * just skipped here and removed from the list.
2082 lock_res_and_lock(lock);
2083 list_del_init(&lock->l_bl_ast);
2085 /* lock is not blocking lock anymore, but was kept in the list because
2086 * it can managed only here.
2088 if (!ldlm_is_ast_sent(lock)) {
2089 unlock_res_and_lock(lock);
2090 LDLM_LOCK_RELEASE(lock);
2094 LASSERT(lock->l_blocking_lock);
2095 ldlm_lock2desc(lock->l_blocking_lock, &d);
2096 /* copy blocking lock ibits in cancel_bits as well,
2097 * new client may use them for lock convert and it is
2098 * important to use new field to convert locks from
2101 d.l_policy_data.l_inodebits.cancel_bits =
2102 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2104 /* Blocking lock is being destroyed here but some information about it
2105 * may be needed inside l_blocking_ast() function below,
2106 * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
2108 bld.bl_same_client = lock->l_client_cookie ==
2109 lock->l_blocking_lock->l_client_cookie;
2110 bld.bl_cos_incompat = ldlm_is_cos_incompat(lock->l_blocking_lock);
2111 arg->bl_desc = &bld;
2113 LASSERT(ldlm_is_ast_sent(lock));
2114 LASSERT(lock->l_bl_ast_run == 0);
2115 lock->l_bl_ast_run++;
2116 ldlm_clear_blocking_lock(lock);
2117 unlock_res_and_lock(lock);
2119 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2121 LDLM_LOCK_RELEASE(lock);
2127 * Process a call to revocation AST callback for a lock in ast_work list
2130 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2132 struct ldlm_cb_set_arg *arg = opaq;
2133 struct ldlm_lock_desc desc;
2135 struct ldlm_lock *lock;
2138 if (list_empty(arg->list))
2141 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2142 list_del_init(&lock->l_rk_ast);
2144 /* the desc just pretend to exclusive */
2145 ldlm_lock2desc(lock, &desc);
2146 desc.l_req_mode = LCK_EX;
2147 desc.l_granted_mode = 0;
2149 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2150 LDLM_LOCK_RELEASE(lock);
2156 * Process a call to glimpse AST callback for a lock in ast_work list
2158 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2160 struct ldlm_cb_set_arg *arg = opaq;
2161 struct ldlm_glimpse_work *gl_work;
2162 struct ldlm_lock *lock;
2166 if (list_empty(arg->list))
2169 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2171 list_del_init(&gl_work->gl_list);
2173 lock = gl_work->gl_lock;
2175 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2176 arg->gl_desc = gl_work->gl_desc;
2177 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2178 arg->gl_interpret_data = gl_work->gl_interpret_data;
2180 /* invoke the actual glimpse callback */
2181 if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
2184 LDLM_LOCK_RELEASE(lock);
2185 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2186 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2188 OBD_FREE_PTR(gl_work);
2195 * Process a call to completion AST callback for a lock in ast_work list
2198 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2200 struct ldlm_cb_set_arg *arg = opaq;
2201 struct ldlm_lock *lock;
2202 ldlm_completion_callback completion_callback;
2207 if (list_empty(arg->list))
2210 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2212 /* It's possible to receive a completion AST before we've set
2213 * the l_completion_ast pointer: either because the AST arrived
2214 * before the reply, or simply because there's a small race
2215 * window between receiving the reply and finishing the local
2216 * enqueue. (bug 842)
2218 * This can't happen with the blocking_ast, however, because we
2219 * will never call the local blocking_ast until we drop our
2220 * reader/writer reference, which we won't do until we get the
2221 * reply and finish enqueueing. */
2223 /* nobody should touch l_cp_ast */
2224 lock_res_and_lock(lock);
2225 list_del_init(&lock->l_cp_ast);
2226 LASSERT(ldlm_is_cp_reqd(lock));
2227 /* save l_completion_ast since it can be changed by
2228 * mds_intent_policy(), see bug 14225 */
2229 completion_callback = lock->l_completion_ast;
2230 ldlm_clear_cp_reqd(lock);
2231 unlock_res_and_lock(lock);
2233 if (completion_callback != NULL)
2234 rc = completion_callback(lock, 0, (void *)arg);
2235 LDLM_LOCK_RELEASE(lock);
2241 * Process list of locks in need of ASTs being sent.
2243 * Used on server to send multiple ASTs together instead of sending one by
2246 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2247 ldlm_desc_ast_t ast_type)
2249 struct ldlm_cb_set_arg *arg;
2250 set_producer_func work_ast_lock;
2253 if (list_empty(rpc_list))
2260 atomic_set(&arg->restart, 0);
2261 arg->list = rpc_list;
2264 case LDLM_WORK_CP_AST:
2265 arg->type = LDLM_CP_CALLBACK;
2266 work_ast_lock = ldlm_work_cp_ast_lock;
2268 #ifdef HAVE_SERVER_SUPPORT
2269 case LDLM_WORK_BL_AST:
2270 arg->type = LDLM_BL_CALLBACK;
2271 work_ast_lock = ldlm_work_bl_ast_lock;
2273 case LDLM_WORK_REVOKE_AST:
2274 arg->type = LDLM_BL_CALLBACK;
2275 work_ast_lock = ldlm_work_revoke_ast_lock;
2277 case LDLM_WORK_GL_AST:
2278 arg->type = LDLM_GL_CALLBACK;
2279 work_ast_lock = ldlm_work_gl_ast_lock;
2286 /* We create a ptlrpc request set with flow control extension.
2287 * This request set will use the work_ast_lock function to produce new
2288 * requests and will send a new request each time one completes in order
2289 * to keep the number of requests in flight to ns_max_parallel_ast */
2290 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2291 work_ast_lock, arg);
2292 if (arg->set == NULL)
2293 GOTO(out, rc = -ENOMEM);
2295 ptlrpc_set_wait(NULL, arg->set);
2296 ptlrpc_set_destroy(arg->set);
2298 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2306 * Try to grant all waiting locks on a resource.
2308 * Calls ldlm_reprocess_queue on waiting queue.
2310 * Typically called after some resource locks are cancelled to see
2311 * if anything could be granted as a result of the cancellation.
2313 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2314 enum ldlm_process_intention intention,
2315 struct ldlm_lock *hint)
2317 struct list_head rpc_list;
2318 #ifdef HAVE_SERVER_SUPPORT
2319 ldlm_reprocessing_policy reprocess;
2320 struct obd_device *obd;
2325 INIT_LIST_HEAD(&rpc_list);
2326 /* Local lock trees don't get reprocessed. */
2327 if (ns_is_client(ldlm_res_to_ns(res))) {
2332 /* Disable reprocess during lock replay stage but allow during
2333 * request replay stage.
2335 obd = ldlm_res_to_ns(res)->ns_obd;
2336 if (obd->obd_recovering &&
2337 atomic_read(&obd->obd_req_replay_clients) == 0)
2341 reprocess = ldlm_get_reprocessing_policy(res);
2342 reprocess(res, &res->lr_waiting, &rpc_list, intention, hint);
2345 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2347 if (rc == -ERESTART) {
2348 LASSERT(list_empty(&rpc_list));
2354 INIT_LIST_HEAD(&rpc_list);
2355 if (!ns_is_client(ldlm_res_to_ns(res))) {
2356 CERROR("This is client-side-only module, cannot handle "
2357 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2364 void ldlm_reprocess_all(struct ldlm_resource *res, struct ldlm_lock *hint)
2366 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN, hint);
2368 EXPORT_SYMBOL(ldlm_reprocess_all);
2370 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2371 struct hlist_node *hnode, void *arg)
2373 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2375 /* This is only called once after recovery done. LU-8306. */
2376 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY, NULL);
2381 * Iterate through all resources on a namespace attempting to grant waiting
2384 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2389 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2390 ldlm_reprocess_res, NULL, 0);
2396 * Helper function to call blocking AST for LDLM lock \a lock in a
2397 * "cancelling" mode.
2399 void ldlm_cancel_callback(struct ldlm_lock *lock)
2401 check_res_locked(lock->l_resource);
2402 if (!ldlm_is_cancel(lock)) {
2403 ldlm_set_cancel(lock);
2404 if (lock->l_blocking_ast) {
2405 unlock_res_and_lock(lock);
2406 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2408 lock_res_and_lock(lock);
2410 LDLM_DEBUG(lock, "no blocking ast");
2413 /* only canceller can set bl_done bit */
2414 ldlm_set_bl_done(lock);
2415 wake_up_all(&lock->l_waitq);
2416 } else if (!ldlm_is_bl_done(lock)) {
2417 struct l_wait_info lwi = { 0 };
2419 /* The lock is guaranteed to have been canceled once
2420 * returning from this function. */
2421 unlock_res_and_lock(lock);
2422 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
2423 lock_res_and_lock(lock);
2428 * Remove skiplist-enabled LDLM lock \a req from granted list
2430 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2432 if (req->l_resource->lr_type != LDLM_PLAIN &&
2433 req->l_resource->lr_type != LDLM_IBITS)
2436 list_del_init(&req->l_sl_policy);
2437 list_del_init(&req->l_sl_mode);
2441 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2443 void ldlm_lock_cancel(struct ldlm_lock *lock)
2445 struct ldlm_resource *res;
2446 struct ldlm_namespace *ns;
2449 lock_res_and_lock(lock);
2451 res = lock->l_resource;
2452 ns = ldlm_res_to_ns(res);
2454 /* Please do not, no matter how tempting, remove this LBUG without
2455 * talking to me first. -phik */
2456 if (lock->l_readers || lock->l_writers) {
2457 LDLM_ERROR(lock, "lock still has references");
2458 unlock_res_and_lock(lock);
2462 if (ldlm_is_waited(lock))
2463 ldlm_del_waiting_lock(lock);
2465 /* Releases cancel callback. */
2466 ldlm_cancel_callback(lock);
2468 /* Yes, second time, just in case it was added again while we were
2469 * running with no res lock in ldlm_cancel_callback */
2470 if (ldlm_is_waited(lock))
2471 ldlm_del_waiting_lock(lock);
2473 ldlm_resource_unlink_lock(lock);
2474 ldlm_lock_destroy_nolock(lock);
2476 if (ldlm_is_granted(lock))
2477 ldlm_pool_del(&ns->ns_pool, lock);
2479 /* Make sure we will not be called again for same lock what is possible
2480 * if not to zero out lock->l_granted_mode */
2481 lock->l_granted_mode = LCK_MINMODE;
2482 unlock_res_and_lock(lock);
2486 EXPORT_SYMBOL(ldlm_lock_cancel);
2489 * Set opaque data into the lock that only makes sense to upper layer.
2491 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2493 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2498 if (lock->l_ast_data == NULL)
2499 lock->l_ast_data = data;
2500 if (lock->l_ast_data == data)
2502 LDLM_LOCK_PUT(lock);
2506 EXPORT_SYMBOL(ldlm_lock_set_data);
2508 struct export_cl_data {
2509 const struct lu_env *ecl_env;
2510 struct obd_export *ecl_exp;
2514 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2515 struct ldlm_lock *lock,
2516 struct export_cl_data *ecl)
2518 struct ldlm_resource *res;
2520 res = ldlm_resource_getref(lock->l_resource);
2522 ldlm_lvbo_update(res, lock, NULL, 1);
2523 ldlm_lock_cancel(lock);
2524 if (!exp->exp_obd->obd_stopping)
2525 ldlm_reprocess_all(res, lock);
2526 ldlm_resource_putref(res);
2529 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2530 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2531 exp, ecl->ecl_loop);
2536 * Iterator function for ldlm_export_cancel_locks.
2537 * Cancels passed locks.
2540 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2541 struct hlist_node *hnode, void *data)
2544 struct export_cl_data *ecl = (struct export_cl_data *)data;
2545 struct obd_export *exp = ecl->ecl_exp;
2546 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2548 LDLM_LOCK_GET(lock);
2549 ldlm_cancel_lock_for_export(exp, lock, ecl);
2550 LDLM_LOCK_RELEASE(lock);
2556 * Cancel all blocked locks for given export.
2558 * Typically called on client disconnection/eviction
2560 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2563 struct export_cl_data ecl = {
2569 rc = lu_env_init(&env, LCT_DT_THREAD);
2574 while (!list_empty(&exp->exp_bl_list)) {
2575 struct ldlm_lock *lock;
2577 spin_lock_bh(&exp->exp_bl_list_lock);
2578 if (!list_empty(&exp->exp_bl_list)) {
2579 lock = list_entry(exp->exp_bl_list.next,
2580 struct ldlm_lock, l_exp_list);
2581 LDLM_LOCK_GET(lock);
2582 list_del_init(&lock->l_exp_list);
2586 spin_unlock_bh(&exp->exp_bl_list_lock);
2591 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2592 LDLM_LOCK_RELEASE(lock);
2597 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2598 "left on hash table %d.\n", exp, ecl.ecl_loop,
2599 atomic_read(&exp->exp_lock_hash->hs_count));
2601 return ecl.ecl_loop;
2605 * Cancel all locks for given export.
2607 * Typically called after client disconnection/eviction
2609 int ldlm_export_cancel_locks(struct obd_export *exp)
2611 struct export_cl_data ecl;
2615 rc = lu_env_init(&env, LCT_DT_THREAD);
2622 cfs_hash_for_each_empty(exp->exp_lock_hash,
2623 ldlm_cancel_locks_for_export_cb, &ecl);
2625 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2626 "left on hash table %d.\n", exp, ecl.ecl_loop,
2627 atomic_read(&exp->exp_lock_hash->hs_count));
2629 if (ecl.ecl_loop > 0 &&
2630 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2631 exp->exp_obd->obd_stopping)
2632 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2636 return ecl.ecl_loop;
2640 * Downgrade an PW/EX lock to COS | CR mode.
2642 * A lock mode convertion from PW/EX mode to less conflict mode. The
2643 * convertion may fail if lock was canceled before downgrade, but it doesn't
2644 * indicate any problem, because such lock has no reader or writer, and will
2647 * Used by Commit on Sharing (COS) code to force object changes commit in case
2648 * of conflict. Converted lock is considered as new lock and all blocking AST
2649 * things are cleared, so any pending or new blocked lock on that lock will
2650 * cause new call to blocking_ast and force resource object commit.
2652 * Also used by layout_change to replace EX lock to CR lock.
2654 * \param lock A lock to convert
2655 * \param new_mode new lock mode
2657 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2659 #ifdef HAVE_SERVER_SUPPORT
2662 LASSERT(new_mode == LCK_COS || new_mode == LCK_CR);
2664 lock_res_and_lock(lock);
2666 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2667 unlock_res_and_lock(lock);
2669 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2670 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2674 ldlm_resource_unlink_lock(lock);
2676 * Remove the lock from pool as it will be added again in
2677 * ldlm_grant_lock() called below.
2679 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2681 /* Consider downgraded lock as a new lock and clear all states
2682 * related to a previous blocking AST processing.
2684 ldlm_clear_blocking_data(lock);
2686 lock->l_req_mode = new_mode;
2687 ldlm_grant_lock(lock, NULL);
2688 unlock_res_and_lock(lock);
2690 ldlm_reprocess_all(lock->l_resource, lock);
2695 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2698 * Print lock with lock handle \a lockh description into debug log.
2700 * Used when printing all locks on a resource for debug purposes.
2702 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2704 struct ldlm_lock *lock;
2706 if (!((libcfs_debug | D_ERROR) & level))
2709 lock = ldlm_handle2lock(lockh);
2713 LDLM_DEBUG_LIMIT(level, lock, "###");
2715 LDLM_LOCK_PUT(lock);
2717 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2720 * Print lock information with custom message into debug log.
2723 void _ldlm_lock_debug(struct ldlm_lock *lock,
2724 struct libcfs_debug_msg_data *msgdata,
2725 const char *fmt, ...)
2728 struct obd_export *exp = lock->l_export;
2729 struct ldlm_resource *resource = NULL;
2730 char *nid = "local";
2732 /* on server-side resource of lock doesn't change */
2733 if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2734 if (lock->l_resource != NULL)
2735 resource = ldlm_resource_getref(lock->l_resource);
2736 } else if (spin_trylock(&lock->l_lock)) {
2737 if (lock->l_resource != NULL)
2738 resource = ldlm_resource_getref(lock->l_resource);
2739 spin_unlock(&lock->l_lock);
2742 va_start(args, fmt);
2744 if (exp && exp->exp_connection) {
2745 nid = obd_export_nid2str(exp);
2746 } else if (exp && exp->exp_obd != NULL) {
2747 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2748 nid = obd_import_nid2str(imp);
2751 if (resource == NULL) {
2752 libcfs_debug_vmsg2(msgdata, fmt, args,
2753 " ns: \?\? lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2754 "res: \?\? rrc=\?\? type: \?\?\? flags: %#llx nid: %s "
2755 "remote: %#llx expref: %d pid: %u timeout: %lld "
2758 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2759 lock->l_readers, lock->l_writers,
2760 ldlm_lockname[lock->l_granted_mode],
2761 ldlm_lockname[lock->l_req_mode],
2762 lock->l_flags, nid, lock->l_remote_handle.cookie,
2763 exp ? atomic_read(&exp->exp_refcount) : -99,
2764 lock->l_pid, lock->l_callback_timeout, lock->l_lvb_type);
2769 switch (resource->lr_type) {
2771 libcfs_debug_vmsg2(msgdata, fmt, args,
2772 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2773 "res: "DLDLMRES" rrc: %d type: %s [%llu->%llu] "
2774 "(req %llu->%llu) flags: %#llx nid: %s remote: "
2775 "%#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2776 ldlm_lock_to_ns_name(lock), lock,
2777 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2778 lock->l_readers, lock->l_writers,
2779 ldlm_lockname[lock->l_granted_mode],
2780 ldlm_lockname[lock->l_req_mode],
2782 atomic_read(&resource->lr_refcount),
2783 ldlm_typename[resource->lr_type],
2784 lock->l_policy_data.l_extent.start,
2785 lock->l_policy_data.l_extent.end,
2786 lock->l_req_extent.start, lock->l_req_extent.end,
2787 lock->l_flags, nid, lock->l_remote_handle.cookie,
2788 exp ? atomic_read(&exp->exp_refcount) : -99,
2789 lock->l_pid, lock->l_callback_timeout,
2794 libcfs_debug_vmsg2(msgdata, fmt, args,
2795 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2796 "res: "DLDLMRES" rrc: %d type: %s pid: %d "
2797 "[%llu->%llu] flags: %#llx nid: %s "
2798 "remote: %#llx expref: %d pid: %u timeout: %lld\n",
2799 ldlm_lock_to_ns_name(lock), lock,
2800 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2801 lock->l_readers, lock->l_writers,
2802 ldlm_lockname[lock->l_granted_mode],
2803 ldlm_lockname[lock->l_req_mode],
2805 atomic_read(&resource->lr_refcount),
2806 ldlm_typename[resource->lr_type],
2807 lock->l_policy_data.l_flock.pid,
2808 lock->l_policy_data.l_flock.start,
2809 lock->l_policy_data.l_flock.end,
2810 lock->l_flags, nid, lock->l_remote_handle.cookie,
2811 exp ? atomic_read(&exp->exp_refcount) : -99,
2812 lock->l_pid, lock->l_callback_timeout);
2816 libcfs_debug_vmsg2(msgdata, fmt, args,
2817 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2818 "res: "DLDLMRES" bits %#llx/%#llx rrc: %d type: %s "
2819 "flags: %#llx nid: %s remote: %#llx expref: %d "
2820 "pid: %u timeout: %lld lvb_type: %d\n",
2821 ldlm_lock_to_ns_name(lock),
2822 lock, lock->l_handle.h_cookie,
2823 atomic_read(&lock->l_refc),
2824 lock->l_readers, lock->l_writers,
2825 ldlm_lockname[lock->l_granted_mode],
2826 ldlm_lockname[lock->l_req_mode],
2828 lock->l_policy_data.l_inodebits.bits,
2829 lock->l_policy_data.l_inodebits.try_bits,
2830 atomic_read(&resource->lr_refcount),
2831 ldlm_typename[resource->lr_type],
2832 lock->l_flags, nid, lock->l_remote_handle.cookie,
2833 exp ? atomic_read(&exp->exp_refcount) : -99,
2834 lock->l_pid, lock->l_callback_timeout,
2839 libcfs_debug_vmsg2(msgdata, fmt, args,
2840 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2841 "res: "DLDLMRES" rrc: %d type: %s flags: %#llx "
2842 "nid: %s remote: %#llx expref: %d pid: %u "
2843 "timeout: %lld lvb_type: %d\n",
2844 ldlm_lock_to_ns_name(lock),
2845 lock, lock->l_handle.h_cookie,
2846 atomic_read(&lock->l_refc),
2847 lock->l_readers, lock->l_writers,
2848 ldlm_lockname[lock->l_granted_mode],
2849 ldlm_lockname[lock->l_req_mode],
2851 atomic_read(&resource->lr_refcount),
2852 ldlm_typename[resource->lr_type],
2853 lock->l_flags, nid, lock->l_remote_handle.cookie,
2854 exp ? atomic_read(&exp->exp_refcount) : -99,
2855 lock->l_pid, lock->l_callback_timeout,
2860 ldlm_resource_putref(resource);
2862 EXPORT_SYMBOL(_ldlm_lock_debug);