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, 2016, 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"
48 char *ldlm_lockname[] = {
56 [LCK_GROUP] = "GROUP",
59 EXPORT_SYMBOL(ldlm_lockname);
61 char *ldlm_typename[] = {
63 [LDLM_EXTENT] = "EXT",
68 static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = {
69 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_wire_to_local,
70 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_wire_to_local,
71 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_wire_to_local,
72 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_wire_to_local,
75 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
76 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_local_to_wire,
77 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_local_to_wire,
78 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_local_to_wire,
79 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_local_to_wire,
83 * Converts lock policy from local format to on the wire lock_desc format
85 void ldlm_convert_policy_to_wire(enum ldlm_type type,
86 const union ldlm_policy_data *lpolicy,
87 union ldlm_wire_policy_data *wpolicy)
89 ldlm_policy_local_to_wire_t convert;
91 convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
93 convert(lpolicy, wpolicy);
97 * Converts lock policy from on the wire lock_desc format to local format
99 void ldlm_convert_policy_to_local(struct obd_export *exp, enum ldlm_type type,
100 const union ldlm_wire_policy_data *wpolicy,
101 union ldlm_policy_data *lpolicy)
103 ldlm_policy_wire_to_local_t convert;
105 convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE];
107 convert(wpolicy, lpolicy);
110 const char *ldlm_it2str(enum ldlm_intent_flags it)
117 case (IT_OPEN | IT_CREAT):
132 CERROR("Unknown intent 0x%08x\n", it);
136 EXPORT_SYMBOL(ldlm_it2str);
138 extern struct kmem_cache *ldlm_lock_slab;
140 #ifdef HAVE_SERVER_SUPPORT
141 static ldlm_processing_policy ldlm_processing_policy_table[] = {
142 [LDLM_PLAIN] = ldlm_process_plain_lock,
143 [LDLM_EXTENT] = ldlm_process_extent_lock,
144 [LDLM_FLOCK] = ldlm_process_flock_lock,
145 [LDLM_IBITS] = ldlm_process_inodebits_lock,
148 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
150 return ldlm_processing_policy_table[res->lr_type];
152 EXPORT_SYMBOL(ldlm_get_processing_policy);
153 #endif /* HAVE_SERVER_SUPPORT */
155 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
159 EXPORT_SYMBOL(ldlm_register_intent);
162 * REFCOUNTED LOCK OBJECTS
167 * Get a reference on a lock.
169 * Lock refcounts, during creation:
170 * - one special one for allocation, dec'd only once in destroy
171 * - one for being a lock that's in-use
172 * - one for the addref associated with a new lock
174 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
176 atomic_inc(&lock->l_refc);
179 EXPORT_SYMBOL(ldlm_lock_get);
182 * Release lock reference.
184 * Also frees the lock if it was last reference.
186 void ldlm_lock_put(struct ldlm_lock *lock)
190 LASSERT(lock->l_resource != LP_POISON);
191 LASSERT(atomic_read(&lock->l_refc) > 0);
192 if (atomic_dec_and_test(&lock->l_refc)) {
193 struct ldlm_resource *res;
196 "final lock_put on destroyed lock, freeing it.");
198 res = lock->l_resource;
199 LASSERT(ldlm_is_destroyed(lock));
200 LASSERT(list_empty(&lock->l_exp_list));
201 LASSERT(list_empty(&lock->l_res_link));
202 LASSERT(list_empty(&lock->l_pending_chain));
204 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
206 lu_ref_del(&res->lr_reference, "lock", lock);
207 ldlm_resource_putref(res);
208 lock->l_resource = NULL;
209 if (lock->l_export) {
210 class_export_lock_put(lock->l_export, lock);
211 lock->l_export = NULL;
214 if (lock->l_lvb_data != NULL)
215 OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
217 ldlm_interval_free(ldlm_interval_detach(lock));
218 lu_ref_fini(&lock->l_reference);
219 OBD_FREE_RCU(lock, sizeof(*lock), &lock->l_handle);
224 EXPORT_SYMBOL(ldlm_lock_put);
227 * Removes LDLM lock \a lock from LRU. Assumes LRU is already locked.
229 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
232 if (!list_empty(&lock->l_lru)) {
233 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
235 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
236 if (ns->ns_last_pos == &lock->l_lru)
237 ns->ns_last_pos = lock->l_lru.prev;
238 list_del_init(&lock->l_lru);
239 LASSERT(ns->ns_nr_unused > 0);
247 * Removes LDLM lock \a lock from LRU. Obtains the LRU lock first.
249 * If \a last_use is non-zero, it will remove the lock from LRU only if
250 * it matches lock's l_last_used.
252 * \retval 0 if \a last_use is set, the lock is not in LRU list or \a last_use
253 * doesn't match lock's l_last_used;
254 * otherwise, the lock hasn't been in the LRU list.
255 * \retval 1 the lock was in LRU list and removed.
257 int ldlm_lock_remove_from_lru_check(struct ldlm_lock *lock, ktime_t last_use)
259 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
263 if (ldlm_is_ns_srv(lock)) {
264 LASSERT(list_empty(&lock->l_lru));
268 spin_lock(&ns->ns_lock);
269 if (!ktime_compare(last_use, ktime_set(0, 0)) ||
270 !ktime_compare(last_use, lock->l_last_used))
271 rc = ldlm_lock_remove_from_lru_nolock(lock);
272 spin_unlock(&ns->ns_lock);
278 * Adds LDLM lock \a lock to namespace LRU. Assumes LRU is already locked.
280 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
282 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
284 lock->l_last_used = ktime_get();
285 LASSERT(list_empty(&lock->l_lru));
286 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
287 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
288 LASSERT(ns->ns_nr_unused >= 0);
293 * Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks
296 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
298 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
301 spin_lock(&ns->ns_lock);
302 ldlm_lock_add_to_lru_nolock(lock);
303 spin_unlock(&ns->ns_lock);
308 * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
309 * the LRU. Performs necessary LRU locking
311 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
313 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
316 if (ldlm_is_ns_srv(lock)) {
317 LASSERT(list_empty(&lock->l_lru));
322 spin_lock(&ns->ns_lock);
323 if (!list_empty(&lock->l_lru)) {
324 ldlm_lock_remove_from_lru_nolock(lock);
325 ldlm_lock_add_to_lru_nolock(lock);
327 spin_unlock(&ns->ns_lock);
332 * Helper to destroy a locked lock.
334 * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
335 * Must be called with l_lock and lr_lock held.
337 * Does not actually free the lock data, but rather marks the lock as
338 * destroyed by setting l_destroyed field in the lock to 1. Destroys a
339 * handle->lock association too, so that the lock can no longer be found
340 * and removes the lock from LRU list. Actual lock freeing occurs when
341 * last lock reference goes away.
343 * Original comment (of some historical value):
344 * This used to have a 'strict' flag, which recovery would use to mark an
345 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
346 * shall explain why it's gone: with the new hash table scheme, once you call
347 * ldlm_lock_destroy, you can never drop your final references on this lock.
348 * Because it's not in the hash table anymore. -phil
350 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
354 if (lock->l_readers || lock->l_writers) {
355 LDLM_ERROR(lock, "lock still has references");
359 if (!list_empty(&lock->l_res_link)) {
360 LDLM_ERROR(lock, "lock still on resource");
364 if (ldlm_is_destroyed(lock)) {
365 LASSERT(list_empty(&lock->l_lru));
369 ldlm_set_destroyed(lock);
371 if (lock->l_export && lock->l_export->exp_lock_hash) {
372 /* NB: it's safe to call cfs_hash_del() even lock isn't
373 * in exp_lock_hash. */
374 /* In the function below, .hs_keycmp resolves to
375 * ldlm_export_lock_keycmp() */
376 /* coverity[overrun-buffer-val] */
377 cfs_hash_del(lock->l_export->exp_lock_hash,
378 &lock->l_remote_handle, &lock->l_exp_hash);
381 ldlm_lock_remove_from_lru(lock);
382 class_handle_unhash(&lock->l_handle);
389 * Destroys a LDLM lock \a lock. Performs necessary locking first.
391 void ldlm_lock_destroy(struct ldlm_lock *lock)
395 lock_res_and_lock(lock);
396 first = ldlm_lock_destroy_internal(lock);
397 unlock_res_and_lock(lock);
399 /* drop reference from hashtable only for first destroy */
401 lu_ref_del(&lock->l_reference, "hash", lock);
402 LDLM_LOCK_RELEASE(lock);
408 * Destroys a LDLM lock \a lock that is already locked.
410 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
414 first = ldlm_lock_destroy_internal(lock);
415 /* drop reference from hashtable only for first destroy */
417 lu_ref_del(&lock->l_reference, "hash", lock);
418 LDLM_LOCK_RELEASE(lock);
423 /* this is called by portals_handle2object with the handle lock taken */
424 static void lock_handle_addref(void *lock)
426 LDLM_LOCK_GET((struct ldlm_lock *)lock);
429 static void lock_handle_free(void *lock, int size)
431 LASSERT(size == sizeof(struct ldlm_lock));
432 OBD_SLAB_FREE(lock, ldlm_lock_slab, size);
435 static struct portals_handle_ops lock_handle_ops = {
436 .hop_addref = lock_handle_addref,
437 .hop_free = lock_handle_free,
442 * Allocate and initialize new lock structure.
444 * usage: pass in a resource on which you have done ldlm_resource_get
445 * new lock will take over the refcount.
446 * returns: lock with refcount 2 - one for current caller and one for remote
448 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
450 struct ldlm_lock *lock;
453 if (resource == NULL)
456 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, GFP_NOFS);
460 spin_lock_init(&lock->l_lock);
461 lock->l_resource = resource;
462 lu_ref_add(&resource->lr_reference, "lock", lock);
464 atomic_set(&lock->l_refc, 2);
465 INIT_LIST_HEAD(&lock->l_res_link);
466 INIT_LIST_HEAD(&lock->l_lru);
467 INIT_LIST_HEAD(&lock->l_pending_chain);
468 INIT_LIST_HEAD(&lock->l_bl_ast);
469 INIT_LIST_HEAD(&lock->l_cp_ast);
470 INIT_LIST_HEAD(&lock->l_rk_ast);
471 init_waitqueue_head(&lock->l_waitq);
472 lock->l_blocking_lock = NULL;
473 INIT_LIST_HEAD(&lock->l_sl_mode);
474 INIT_LIST_HEAD(&lock->l_sl_policy);
475 INIT_HLIST_NODE(&lock->l_exp_hash);
476 INIT_HLIST_NODE(&lock->l_exp_flock_hash);
478 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
480 INIT_LIST_HEAD(&lock->l_handle.h_link);
481 class_handle_hash(&lock->l_handle, &lock_handle_ops);
483 lu_ref_init(&lock->l_reference);
484 lu_ref_add(&lock->l_reference, "hash", lock);
485 lock->l_callback_timeout = 0;
487 #if LUSTRE_TRACKS_LOCK_EXP_REFS
488 INIT_LIST_HEAD(&lock->l_exp_refs_link);
489 lock->l_exp_refs_nr = 0;
490 lock->l_exp_refs_target = NULL;
492 INIT_LIST_HEAD(&lock->l_exp_list);
498 * Moves LDLM lock \a lock to another resource.
499 * This is used on client when server returns some other lock than requested
500 * (typically as a result of intent operation)
502 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
503 const struct ldlm_res_id *new_resid)
505 struct ldlm_resource *oldres = lock->l_resource;
506 struct ldlm_resource *newres;
510 LASSERT(ns_is_client(ns));
512 lock_res_and_lock(lock);
513 if (memcmp(new_resid, &lock->l_resource->lr_name,
514 sizeof(lock->l_resource->lr_name)) == 0) {
516 unlock_res_and_lock(lock);
520 LASSERT(new_resid->name[0] != 0);
522 /* This function assumes that the lock isn't on any lists */
523 LASSERT(list_empty(&lock->l_res_link));
525 type = oldres->lr_type;
526 unlock_res_and_lock(lock);
528 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
530 RETURN(PTR_ERR(newres));
532 lu_ref_add(&newres->lr_reference, "lock", lock);
534 * To flip the lock from the old to the new resource, lock, oldres and
535 * newres have to be locked. Resource spin-locks are nested within
536 * lock->l_lock, and are taken in the memory address order to avoid
539 spin_lock(&lock->l_lock);
540 oldres = lock->l_resource;
541 if (oldres < newres) {
543 lock_res_nested(newres, LRT_NEW);
546 lock_res_nested(oldres, LRT_NEW);
548 LASSERT(memcmp(new_resid, &oldres->lr_name,
549 sizeof oldres->lr_name) != 0);
550 lock->l_resource = newres;
552 unlock_res_and_lock(lock);
554 /* ...and the flowers are still standing! */
555 lu_ref_del(&oldres->lr_reference, "lock", lock);
556 ldlm_resource_putref(oldres);
561 /** \defgroup ldlm_handles LDLM HANDLES
562 * Ways to get hold of locks without any addresses.
567 * Fills in handle for LDLM lock \a lock into supplied \a lockh
568 * Does not take any references.
570 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
572 lockh->cookie = lock->l_handle.h_cookie;
574 EXPORT_SYMBOL(ldlm_lock2handle);
577 * Obtain a lock reference by handle.
579 * if \a flags: atomically get the lock and set the flags.
580 * Return NULL if flag already set
582 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
585 struct ldlm_lock *lock;
590 lock = class_handle2object(handle->cookie, NULL);
594 if (lock->l_export != NULL && lock->l_export->exp_failed) {
595 CDEBUG(D_INFO, "lock export failed: lock %p, exp %p\n",
596 lock, lock->l_export);
601 /* It's unlikely but possible that someone marked the lock as
602 * destroyed after we did handle2object on it */
603 if ((flags == 0) && !ldlm_is_destroyed(lock)) {
604 lu_ref_add(&lock->l_reference, "handle", current);
608 lock_res_and_lock(lock);
610 LASSERT(lock->l_resource != NULL);
612 lu_ref_add_atomic(&lock->l_reference, "handle", current);
613 if (unlikely(ldlm_is_destroyed(lock))) {
614 unlock_res_and_lock(lock);
615 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
620 /* If we're setting flags, make sure none of them are already set. */
622 if ((lock->l_flags & flags) != 0) {
623 unlock_res_and_lock(lock);
628 lock->l_flags |= flags;
631 unlock_res_and_lock(lock);
634 EXPORT_SYMBOL(__ldlm_handle2lock);
635 /** @} ldlm_handles */
638 * Fill in "on the wire" representation for given LDLM lock into supplied
639 * lock descriptor \a desc structure.
641 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
643 ldlm_res2desc(lock->l_resource, &desc->l_resource);
644 desc->l_req_mode = lock->l_req_mode;
645 desc->l_granted_mode = lock->l_granted_mode;
646 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
647 &lock->l_policy_data,
648 &desc->l_policy_data);
652 * Add a lock to list of conflicting locks to send AST to.
654 * Only add if we have not sent a blocking AST to the lock yet.
656 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
657 struct list_head *work_list)
659 if (!ldlm_is_ast_sent(lock)) {
660 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
661 ldlm_set_ast_sent(lock);
662 /* If the enqueuing client said so, tell the AST recipient to
663 * discard dirty data, rather than writing back. */
664 if (ldlm_is_ast_discard_data(new))
665 ldlm_set_discard_data(lock);
666 LASSERT(list_empty(&lock->l_bl_ast));
667 list_add(&lock->l_bl_ast, work_list);
669 LASSERT(lock->l_blocking_lock == NULL);
670 lock->l_blocking_lock = LDLM_LOCK_GET(new);
675 * Add a lock to list of just granted locks to send completion AST to.
677 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
678 struct list_head *work_list)
680 if (!ldlm_is_cp_reqd(lock)) {
681 ldlm_set_cp_reqd(lock);
682 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
683 LASSERT(list_empty(&lock->l_cp_ast));
684 list_add(&lock->l_cp_ast, work_list);
690 * Aggregator function to add AST work items into a list. Determines
691 * what sort of an AST work needs to be done and calls the proper
693 * Must be called with lr_lock held.
695 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
696 struct list_head *work_list)
699 check_res_locked(lock->l_resource);
701 ldlm_add_bl_work_item(lock, new, work_list);
703 ldlm_add_cp_work_item(lock, work_list);
708 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
709 * r/w reference type is determined by \a mode
710 * Calls ldlm_lock_addref_internal.
712 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
714 struct ldlm_lock *lock;
716 lock = ldlm_handle2lock(lockh);
717 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
718 ldlm_lock_addref_internal(lock, mode);
721 EXPORT_SYMBOL(ldlm_lock_addref);
725 * Add specified reader/writer reference to LDLM lock \a lock.
726 * r/w reference type is determined by \a mode
727 * Removes lock from LRU if it is there.
728 * Assumes the LDLM lock is already locked.
730 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
733 ldlm_lock_remove_from_lru(lock);
734 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
736 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
738 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
740 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
743 lu_ref_add_atomic(&lock->l_reference, "user", lock);
744 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
748 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
749 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
751 * \retval 0 success, lock was addref-ed
753 * \retval -EAGAIN lock is being canceled.
755 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
757 struct ldlm_lock *lock;
761 lock = ldlm_handle2lock(lockh);
763 lock_res_and_lock(lock);
764 if (lock->l_readers != 0 || lock->l_writers != 0 ||
765 !ldlm_is_cbpending(lock)) {
766 ldlm_lock_addref_internal_nolock(lock, mode);
769 unlock_res_and_lock(lock);
774 EXPORT_SYMBOL(ldlm_lock_addref_try);
777 * Add specified reader/writer reference to LDLM lock \a lock.
778 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
779 * Only called for local locks.
781 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
783 lock_res_and_lock(lock);
784 ldlm_lock_addref_internal_nolock(lock, mode);
785 unlock_res_and_lock(lock);
789 * Removes reader/writer reference for LDLM lock \a lock.
790 * Assumes LDLM lock is already locked.
791 * only called in ldlm_flock_destroy and for local locks.
792 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
793 * that cannot be placed in LRU.
795 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
798 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
799 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
800 LASSERT(lock->l_readers > 0);
801 lu_ref_del(&lock->l_reference, "reader", lock);
804 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
805 LASSERT(lock->l_writers > 0);
806 lu_ref_del(&lock->l_reference, "writer", lock);
810 lu_ref_del(&lock->l_reference, "user", lock);
811 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
815 * Removes reader/writer reference for LDLM lock \a lock.
816 * Locks LDLM lock first.
817 * If the lock is determined to be client lock on a client and r/w refcount
818 * drops to zero and the lock is not blocked, the lock is added to LRU lock
820 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
822 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
824 struct ldlm_namespace *ns;
827 lock_res_and_lock(lock);
829 ns = ldlm_lock_to_ns(lock);
831 ldlm_lock_decref_internal_nolock(lock, mode);
833 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
834 !lock->l_readers && !lock->l_writers) {
835 /* If this is a local lock on a server namespace and this was
836 * the last reference, cancel the lock.
838 * Group locks are special:
839 * They must not go in LRU, but they are not called back
840 * like non-group locks, instead they are manually released.
841 * They have an l_writers reference which they keep until
842 * they are manually released, so we remove them when they have
843 * no more reader or writer references. - LU-6368 */
844 ldlm_set_cbpending(lock);
847 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
848 /* If we received a blocked AST and this was the last reference,
849 * run the callback. */
850 if (ldlm_is_ns_srv(lock) && lock->l_export)
851 CERROR("FL_CBPENDING set on non-local lock--just a "
854 LDLM_DEBUG(lock, "final decref done on cbpending lock");
856 LDLM_LOCK_GET(lock); /* dropped by bl thread */
857 ldlm_lock_remove_from_lru(lock);
858 unlock_res_and_lock(lock);
860 if (ldlm_is_fail_loc(lock))
861 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
863 if (ldlm_is_atomic_cb(lock) ||
864 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
865 ldlm_handle_bl_callback(ns, NULL, lock);
866 } else if (ns_is_client(ns) &&
867 !lock->l_readers && !lock->l_writers &&
868 !ldlm_is_no_lru(lock) &&
869 !ldlm_is_bl_ast(lock)) {
871 LDLM_DEBUG(lock, "add lock into lru list");
873 /* If this is a client-side namespace and this was the last
874 * reference, put it on the LRU. */
875 ldlm_lock_add_to_lru(lock);
876 unlock_res_and_lock(lock);
878 if (ldlm_is_fail_loc(lock))
879 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
881 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
882 * are not supported by the server, otherwise, it is done on
884 if (!exp_connect_cancelset(lock->l_conn_export) &&
885 !ns_connect_lru_resize(ns))
886 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
888 LDLM_DEBUG(lock, "do not add lock into lru list");
889 unlock_res_and_lock(lock);
896 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
898 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
900 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
901 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
902 ldlm_lock_decref_internal(lock, mode);
905 EXPORT_SYMBOL(ldlm_lock_decref);
908 * Decrease reader/writer refcount for LDLM lock with handle
909 * \a lockh and mark it for subsequent cancellation once r/w refcount
910 * drops to zero instead of putting into LRU.
913 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
916 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
919 LASSERT(lock != NULL);
921 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
922 lock_res_and_lock(lock);
923 ldlm_set_cbpending(lock);
924 unlock_res_and_lock(lock);
925 ldlm_lock_decref_internal(lock, mode);
928 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
930 struct sl_insert_point {
931 struct list_head *res_link;
932 struct list_head *mode_link;
933 struct list_head *policy_link;
937 * Finds a position to insert the new lock into granted lock list.
939 * Used for locks eligible for skiplist optimization.
942 * queue [input]: the granted list where search acts on;
943 * req [input]: the lock whose position to be located;
944 * prev [output]: positions within 3 lists to insert @req to
948 * - ldlm_grant_lock_with_skiplist
950 static void search_granted_lock(struct list_head *queue,
951 struct ldlm_lock *req,
952 struct sl_insert_point *prev)
954 struct list_head *tmp;
955 struct ldlm_lock *lock, *mode_end, *policy_end;
958 list_for_each(tmp, queue) {
959 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
961 mode_end = list_entry(lock->l_sl_mode.prev,
962 struct ldlm_lock, l_sl_mode);
964 if (lock->l_req_mode != req->l_req_mode) {
965 /* jump to last lock of mode group */
966 tmp = &mode_end->l_res_link;
970 /* suitable mode group is found */
971 if (lock->l_resource->lr_type == LDLM_PLAIN) {
972 /* insert point is last lock of the mode group */
973 prev->res_link = &mode_end->l_res_link;
974 prev->mode_link = &mode_end->l_sl_mode;
975 prev->policy_link = &req->l_sl_policy;
978 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
981 list_entry(lock->l_sl_policy.prev,
985 if (lock->l_policy_data.l_inodebits.bits ==
986 req->l_policy_data.l_inodebits.bits) {
987 /* insert point is last lock of
988 * the policy group */
990 &policy_end->l_res_link;
992 &policy_end->l_sl_mode;
994 &policy_end->l_sl_policy;
999 if (policy_end == mode_end)
1000 /* done with mode group */
1003 /* go to next policy group within mode group */
1004 tmp = policy_end->l_res_link.next;
1005 lock = list_entry(tmp, struct ldlm_lock,
1007 } /* loop over policy groups within the mode group */
1009 /* insert point is last lock of the mode group,
1010 * new policy group is started */
1011 prev->res_link = &mode_end->l_res_link;
1012 prev->mode_link = &mode_end->l_sl_mode;
1013 prev->policy_link = &req->l_sl_policy;
1017 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1022 /* insert point is last lock on the queue,
1023 * new mode group and new policy group are started */
1024 prev->res_link = queue->prev;
1025 prev->mode_link = &req->l_sl_mode;
1026 prev->policy_link = &req->l_sl_policy;
1032 * Add a lock into resource granted list after a position described by
1035 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1036 struct sl_insert_point *prev)
1038 struct ldlm_resource *res = lock->l_resource;
1041 check_res_locked(res);
1043 ldlm_resource_dump(D_INFO, res);
1044 LDLM_DEBUG(lock, "About to add lock:");
1046 if (ldlm_is_destroyed(lock)) {
1047 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1051 LASSERT(list_empty(&lock->l_res_link));
1052 LASSERT(list_empty(&lock->l_sl_mode));
1053 LASSERT(list_empty(&lock->l_sl_policy));
1056 * lock->link == prev->link means lock is first starting the group.
1057 * Don't re-add to itself to suppress kernel warnings.
1059 if (&lock->l_res_link != prev->res_link)
1060 list_add(&lock->l_res_link, prev->res_link);
1061 if (&lock->l_sl_mode != prev->mode_link)
1062 list_add(&lock->l_sl_mode, prev->mode_link);
1063 if (&lock->l_sl_policy != prev->policy_link)
1064 list_add(&lock->l_sl_policy, prev->policy_link);
1070 * Add a lock to granted list on a resource maintaining skiplist
1073 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1075 struct sl_insert_point prev;
1078 LASSERT(lock->l_req_mode == lock->l_granted_mode);
1080 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1081 ldlm_granted_list_add_lock(lock, &prev);
1086 * Perform lock granting bookkeeping.
1088 * Includes putting the lock into granted list and updating lock mode.
1090 * - ldlm_lock_enqueue
1091 * - ldlm_reprocess_queue
1092 * - ldlm_lock_convert
1094 * must be called with lr_lock held
1096 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1098 struct ldlm_resource *res = lock->l_resource;
1101 check_res_locked(res);
1103 lock->l_granted_mode = lock->l_req_mode;
1105 if (work_list && lock->l_completion_ast != NULL)
1106 ldlm_add_ast_work_item(lock, NULL, work_list);
1108 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1109 ldlm_grant_lock_with_skiplist(lock);
1110 else if (res->lr_type == LDLM_EXTENT)
1111 ldlm_extent_add_lock(res, lock);
1112 else if (res->lr_type == LDLM_FLOCK) {
1113 /* We should not add locks to granted list in the following
1115 * - this is an UNLOCK but not a real lock;
1116 * - this is a TEST lock;
1117 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1118 * - this is a deadlock (flock cannot be granted) */
1119 if (lock->l_req_mode == 0 ||
1120 lock->l_req_mode == LCK_NL ||
1121 ldlm_is_test_lock(lock) ||
1122 ldlm_is_flock_deadlock(lock))
1124 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1129 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1134 * Describe the overlap between two locks. itree_overlap_cb data.
1136 struct lock_match_data {
1137 struct ldlm_lock *lmd_old;
1138 struct ldlm_lock *lmd_lock;
1139 enum ldlm_mode *lmd_mode;
1140 union ldlm_policy_data *lmd_policy;
1146 * Check if the given @lock meets the criteria for a match.
1147 * A reference on the lock is taken if matched.
1149 * \param lock test-against this lock
1150 * \param data parameters
1152 static int lock_matches(struct ldlm_lock *lock, struct lock_match_data *data)
1154 union ldlm_policy_data *lpol = &lock->l_policy_data;
1155 enum ldlm_mode match;
1157 if (lock == data->lmd_old)
1158 return INTERVAL_ITER_STOP;
1160 /* Check if this lock can be matched.
1161 * Used by LU-2919(exclusive open) for open lease lock */
1162 if (ldlm_is_excl(lock))
1163 return INTERVAL_ITER_CONT;
1165 /* llite sometimes wants to match locks that will be
1166 * canceled when their users drop, but we allow it to match
1167 * if it passes in CBPENDING and the lock still has users.
1168 * this is generally only going to be used by children
1169 * whose parents already hold a lock so forward progress
1170 * can still happen. */
1171 if (ldlm_is_cbpending(lock) &&
1172 !(data->lmd_flags & LDLM_FL_CBPENDING))
1173 return INTERVAL_ITER_CONT;
1174 if (!data->lmd_unref && ldlm_is_cbpending(lock) &&
1175 lock->l_readers == 0 && lock->l_writers == 0)
1176 return INTERVAL_ITER_CONT;
1178 if (!(lock->l_req_mode & *data->lmd_mode))
1179 return INTERVAL_ITER_CONT;
1180 match = lock->l_req_mode;
1182 switch (lock->l_resource->lr_type) {
1184 if (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1185 lpol->l_extent.end < data->lmd_policy->l_extent.end)
1186 return INTERVAL_ITER_CONT;
1188 if (unlikely(match == LCK_GROUP) &&
1189 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1190 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1191 return INTERVAL_ITER_CONT;
1194 /* We match if we have existing lock with same or wider set
1196 if ((lpol->l_inodebits.bits &
1197 data->lmd_policy->l_inodebits.bits) !=
1198 data->lmd_policy->l_inodebits.bits)
1199 return INTERVAL_ITER_CONT;
1205 /* We match if we have existing lock with same or wider set
1207 if (!data->lmd_unref && LDLM_HAVE_MASK(lock, GONE))
1208 return INTERVAL_ITER_CONT;
1210 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1211 return INTERVAL_ITER_CONT;
1213 if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1214 LDLM_LOCK_GET(lock);
1215 ldlm_lock_touch_in_lru(lock);
1217 ldlm_lock_addref_internal_nolock(lock, match);
1220 *data->lmd_mode = match;
1221 data->lmd_lock = lock;
1223 return INTERVAL_ITER_STOP;
1226 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1228 struct ldlm_interval *node = to_ldlm_interval(in);
1229 struct lock_match_data *data = args;
1230 struct ldlm_lock *lock;
1233 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1234 rc = lock_matches(lock, data);
1235 if (rc == INTERVAL_ITER_STOP)
1236 return INTERVAL_ITER_STOP;
1238 return INTERVAL_ITER_CONT;
1242 * Search for a lock with given parameters in interval trees.
1244 * \param res search for a lock in this resource
1245 * \param data parameters
1247 * \retval a referenced lock or NULL.
1249 static struct ldlm_lock *search_itree(struct ldlm_resource *res,
1250 struct lock_match_data *data)
1252 struct interval_node_extent ext = {
1253 .start = data->lmd_policy->l_extent.start,
1254 .end = data->lmd_policy->l_extent.end
1258 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1259 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1261 if (tree->lit_root == NULL)
1264 if (!(tree->lit_mode & *data->lmd_mode))
1267 interval_search(tree->lit_root, &ext,
1268 itree_overlap_cb, data);
1270 return data->lmd_lock;
1275 * Search for a lock with given properties in a queue.
1277 * \param queue search for a lock in this queue
1278 * \param data parameters
1280 * \retval a referenced lock or NULL.
1282 static struct ldlm_lock *search_queue(struct list_head *queue,
1283 struct lock_match_data *data)
1285 struct ldlm_lock *lock;
1288 list_for_each_entry(lock, queue, l_res_link) {
1289 rc = lock_matches(lock, data);
1290 if (rc == INTERVAL_ITER_STOP)
1291 return data->lmd_lock;
1296 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1298 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1299 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1300 wake_up_all(&lock->l_waitq);
1303 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1305 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1307 lock_res_and_lock(lock);
1308 ldlm_lock_fail_match_locked(lock);
1309 unlock_res_and_lock(lock);
1313 * Mark lock as "matchable" by OST.
1315 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1317 * Assumes LDLM lock is already locked.
1319 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1321 ldlm_set_lvb_ready(lock);
1322 wake_up_all(&lock->l_waitq);
1324 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1327 * Mark lock as "matchable" by OST.
1328 * Locks the lock and then \see ldlm_lock_allow_match_locked
1330 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1332 lock_res_and_lock(lock);
1333 ldlm_lock_allow_match_locked(lock);
1334 unlock_res_and_lock(lock);
1336 EXPORT_SYMBOL(ldlm_lock_allow_match);
1339 * Attempt to find a lock with specified properties.
1341 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1344 * Can be called in two ways:
1346 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1347 * for a duplicate of.
1349 * Otherwise, all of the fields must be filled in, to match against.
1351 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1352 * server (ie, connh is NULL)
1353 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1354 * list will be considered
1355 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1356 * to be canceled can still be matched as long as they still have reader
1357 * or writer refernces
1358 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1359 * just tell us if we would have matched.
1361 * \retval 1 if it finds an already-existing lock that is compatible; in this
1362 * case, lockh is filled in with a addref()ed lock
1364 * We also check security context, and if that fails we simply return 0 (to
1365 * keep caller code unchanged), the context failure will be discovered by
1366 * caller sometime later.
1368 enum ldlm_mode ldlm_lock_match(struct ldlm_namespace *ns, __u64 flags,
1369 const struct ldlm_res_id *res_id,
1370 enum ldlm_type type,
1371 union ldlm_policy_data *policy,
1372 enum ldlm_mode mode,
1373 struct lustre_handle *lockh, int unref)
1375 struct lock_match_data data = {
1379 .lmd_policy = policy,
1383 struct ldlm_resource *res;
1384 struct ldlm_lock *lock;
1389 data.lmd_old = ldlm_handle2lock(lockh);
1390 LASSERT(data.lmd_old != NULL);
1392 ns = ldlm_lock_to_ns(data.lmd_old);
1393 res_id = &data.lmd_old->l_resource->lr_name;
1394 type = data.lmd_old->l_resource->lr_type;
1395 *data.lmd_mode = data.lmd_old->l_req_mode;
1398 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1400 LASSERT(data.lmd_old == NULL);
1404 LDLM_RESOURCE_ADDREF(res);
1407 if (res->lr_type == LDLM_EXTENT)
1408 lock = search_itree(res, &data);
1410 lock = search_queue(&res->lr_granted, &data);
1413 if (flags & LDLM_FL_BLOCK_GRANTED)
1415 lock = search_queue(&res->lr_converting, &data);
1418 lock = search_queue(&res->lr_waiting, &data);
1425 LDLM_RESOURCE_DELREF(res);
1426 ldlm_resource_putref(res);
1429 ldlm_lock2handle(lock, lockh);
1430 if ((flags & LDLM_FL_LVB_READY) &&
1431 (!ldlm_is_lvb_ready(lock))) {
1432 __u64 wait_flags = LDLM_FL_LVB_READY |
1433 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1434 struct l_wait_info lwi;
1435 if (lock->l_completion_ast) {
1436 int err = lock->l_completion_ast(lock,
1437 LDLM_FL_WAIT_NOREPROC,
1440 if (flags & LDLM_FL_TEST_LOCK)
1441 LDLM_LOCK_RELEASE(lock);
1443 ldlm_lock_decref_internal(lock,
1450 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1451 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1453 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1454 l_wait_event(lock->l_waitq,
1455 lock->l_flags & wait_flags,
1457 if (!ldlm_is_lvb_ready(lock)) {
1458 if (flags & LDLM_FL_TEST_LOCK)
1459 LDLM_LOCK_RELEASE(lock);
1461 ldlm_lock_decref_internal(lock, mode);
1468 LDLM_DEBUG(lock, "matched (%llu %llu)",
1469 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1470 res_id->name[2] : policy->l_extent.start,
1471 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1472 res_id->name[3] : policy->l_extent.end);
1474 /* check user's security context */
1475 if (lock->l_conn_export &&
1476 sptlrpc_import_check_ctx(
1477 class_exp2cliimp(lock->l_conn_export))) {
1478 if (!(flags & LDLM_FL_TEST_LOCK))
1479 ldlm_lock_decref_internal(lock, mode);
1483 if (flags & LDLM_FL_TEST_LOCK)
1484 LDLM_LOCK_RELEASE(lock);
1486 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1487 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1488 "%llu/%llu (%llu %llu)", ns,
1489 type, mode, res_id->name[0], res_id->name[1],
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);
1495 if (data.lmd_old != NULL)
1496 LDLM_LOCK_PUT(data.lmd_old);
1498 return rc ? mode : 0;
1500 EXPORT_SYMBOL(ldlm_lock_match);
1502 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1505 struct ldlm_lock *lock;
1506 enum ldlm_mode mode = 0;
1509 lock = ldlm_handle2lock(lockh);
1511 lock_res_and_lock(lock);
1512 if (LDLM_HAVE_MASK(lock, GONE))
1515 if (ldlm_is_cbpending(lock) &&
1516 lock->l_readers == 0 && lock->l_writers == 0)
1520 *bits = lock->l_policy_data.l_inodebits.bits;
1521 mode = lock->l_granted_mode;
1522 ldlm_lock_addref_internal_nolock(lock, mode);
1529 unlock_res_and_lock(lock);
1530 LDLM_LOCK_PUT(lock);
1534 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1536 /** The caller must guarantee that the buffer is large enough. */
1537 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1538 enum req_location loc, void *data, int size)
1543 LASSERT(data != NULL);
1546 switch (lock->l_lvb_type) {
1548 if (size == sizeof(struct ost_lvb)) {
1549 if (loc == RCL_CLIENT)
1550 lvb = req_capsule_client_swab_get(pill,
1552 lustre_swab_ost_lvb);
1554 lvb = req_capsule_server_swab_get(pill,
1556 lustre_swab_ost_lvb);
1557 if (unlikely(lvb == NULL)) {
1558 LDLM_ERROR(lock, "no LVB");
1562 memcpy(data, lvb, size);
1563 } else if (size == sizeof(struct ost_lvb_v1)) {
1564 struct ost_lvb *olvb = data;
1566 if (loc == RCL_CLIENT)
1567 lvb = req_capsule_client_swab_get(pill,
1569 lustre_swab_ost_lvb_v1);
1571 lvb = req_capsule_server_sized_swab_get(pill,
1573 lustre_swab_ost_lvb_v1);
1574 if (unlikely(lvb == NULL)) {
1575 LDLM_ERROR(lock, "no LVB");
1579 memcpy(data, lvb, size);
1580 olvb->lvb_mtime_ns = 0;
1581 olvb->lvb_atime_ns = 0;
1582 olvb->lvb_ctime_ns = 0;
1584 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1590 if (size == sizeof(struct lquota_lvb)) {
1591 if (loc == RCL_CLIENT)
1592 lvb = req_capsule_client_swab_get(pill,
1594 lustre_swab_lquota_lvb);
1596 lvb = req_capsule_server_swab_get(pill,
1598 lustre_swab_lquota_lvb);
1599 if (unlikely(lvb == NULL)) {
1600 LDLM_ERROR(lock, "no LVB");
1604 memcpy(data, lvb, size);
1606 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1615 if (loc == RCL_CLIENT)
1616 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1618 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1619 if (unlikely(lvb == NULL)) {
1620 LDLM_ERROR(lock, "no LVB");
1624 memcpy(data, lvb, size);
1627 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1628 libcfs_debug_dumpstack(NULL);
1636 * Create and fill in new LDLM lock with specified properties.
1637 * Returns a referenced lock
1639 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1640 const struct ldlm_res_id *res_id,
1641 enum ldlm_type type,
1642 enum ldlm_mode mode,
1643 const struct ldlm_callback_suite *cbs,
1644 void *data, __u32 lvb_len,
1645 enum lvb_type lvb_type)
1647 struct ldlm_lock *lock;
1648 struct ldlm_resource *res;
1652 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1654 RETURN(ERR_CAST(res));
1656 lock = ldlm_lock_new(res);
1658 RETURN(ERR_PTR(-ENOMEM));
1660 lock->l_req_mode = mode;
1661 lock->l_ast_data = data;
1662 lock->l_pid = current_pid();
1663 if (ns_is_server(ns))
1664 ldlm_set_ns_srv(lock);
1666 lock->l_blocking_ast = cbs->lcs_blocking;
1667 lock->l_completion_ast = cbs->lcs_completion;
1668 lock->l_glimpse_ast = cbs->lcs_glimpse;
1671 lock->l_tree_node = NULL;
1672 /* if this is the extent lock, allocate the interval tree node */
1673 if (type == LDLM_EXTENT)
1674 if (ldlm_interval_alloc(lock) == NULL)
1675 GOTO(out, rc = -ENOMEM);
1678 lock->l_lvb_len = lvb_len;
1679 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1680 if (lock->l_lvb_data == NULL)
1681 GOTO(out, rc = -ENOMEM);
1684 lock->l_lvb_type = lvb_type;
1685 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1686 GOTO(out, rc = -ENOENT);
1691 ldlm_lock_destroy(lock);
1692 LDLM_LOCK_RELEASE(lock);
1693 RETURN(ERR_PTR(rc));
1697 * Enqueue (request) a lock.
1699 * Does not block. As a result of enqueue the lock would be put
1700 * into granted or waiting list.
1702 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1703 * set, skip all the enqueueing and delegate lock processing to intent policy
1706 enum ldlm_error ldlm_lock_enqueue(struct ldlm_namespace *ns,
1707 struct ldlm_lock **lockp,
1708 void *cookie, __u64 *flags)
1710 struct ldlm_lock *lock = *lockp;
1711 struct ldlm_resource *res = lock->l_resource;
1712 int local = ns_is_client(ldlm_res_to_ns(res));
1713 #ifdef HAVE_SERVER_SUPPORT
1714 ldlm_processing_policy policy;
1716 enum ldlm_error rc = ELDLM_OK;
1717 struct ldlm_interval *node = NULL;
1720 /* policies are not executed on the client or during replay */
1721 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1722 && !local && ns->ns_policy) {
1723 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1725 if (rc == ELDLM_LOCK_REPLACED) {
1726 /* The lock that was returned has already been granted,
1727 * and placed into lockp. If it's not the same as the
1728 * one we passed in, then destroy the old one and our
1729 * work here is done. */
1730 if (lock != *lockp) {
1731 ldlm_lock_destroy(lock);
1732 LDLM_LOCK_RELEASE(lock);
1734 *flags |= LDLM_FL_LOCK_CHANGED;
1736 } else if (rc != ELDLM_OK &&
1737 lock->l_req_mode == lock->l_granted_mode) {
1738 LASSERT(*flags & LDLM_FL_RESENT);
1739 /* It may happen that ns_policy returns an error in
1740 * resend case, object may be unlinked or just some
1741 * error occurs. It is unclear if lock reached the
1742 * client in the original reply, just leave the lock on
1743 * server, not returning it again to client. Due to
1744 * LU-6529, the server will not OOM. */
1746 } else if (rc != ELDLM_OK ||
1747 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1748 ldlm_lock_destroy(lock);
1753 if (*flags & LDLM_FL_RESENT) {
1754 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1755 * Set LOCK_CHANGED always.
1756 * Check if the lock is granted for BLOCK_GRANTED.
1757 * Take NO_TIMEOUT from the lock as it is inherited through
1758 * LDLM_FL_INHERIT_MASK */
1759 *flags |= LDLM_FL_LOCK_CHANGED;
1760 if (lock->l_req_mode != lock->l_granted_mode)
1761 *flags |= LDLM_FL_BLOCK_GRANTED;
1762 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1766 /* For a replaying lock, it might be already in granted list. So
1767 * unlinking the lock will cause the interval node to be freed, we
1768 * have to allocate the interval node early otherwise we can't regrant
1769 * this lock in the future. - jay */
1770 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1771 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1773 lock_res_and_lock(lock);
1774 if (local && lock->l_req_mode == lock->l_granted_mode) {
1775 /* The server returned a blocked lock, but it was granted
1776 * before we got a chance to actually enqueue it. We don't
1777 * need to do anything else. */
1778 *flags &= ~LDLM_FL_BLOCKED_MASK;
1779 GOTO(out, rc = ELDLM_OK);
1782 ldlm_resource_unlink_lock(lock);
1783 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1785 ldlm_lock_destroy_nolock(lock);
1786 GOTO(out, rc = -ENOMEM);
1789 INIT_LIST_HEAD(&node->li_group);
1790 ldlm_interval_attach(node, lock);
1794 /* Some flags from the enqueue want to make it into the AST, via the
1795 * lock's l_flags. */
1796 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1797 ldlm_set_ast_discard_data(lock);
1798 if (*flags & LDLM_FL_TEST_LOCK)
1799 ldlm_set_test_lock(lock);
1800 if (*flags & LDLM_FL_COS_INCOMPAT)
1801 ldlm_set_cos_incompat(lock);
1802 if (*flags & LDLM_FL_COS_ENABLED)
1803 ldlm_set_cos_enabled(lock);
1805 /* This distinction between local lock trees is very important; a client
1806 * namespace only has information about locks taken by that client, and
1807 * thus doesn't have enough information to decide for itself if it can
1808 * be granted (below). In this case, we do exactly what the server
1809 * tells us to do, as dictated by the 'flags'.
1811 * We do exactly the same thing during recovery, when the server is
1812 * more or less trusting the clients not to lie.
1814 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1815 * granted/converting queues. */
1817 if (*flags & LDLM_FL_BLOCK_CONV)
1818 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1819 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1820 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1822 ldlm_grant_lock(lock, NULL);
1823 GOTO(out, rc = ELDLM_OK);
1824 #ifdef HAVE_SERVER_SUPPORT
1825 } else if (*flags & LDLM_FL_REPLAY) {
1826 if (*flags & LDLM_FL_BLOCK_CONV) {
1827 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1828 GOTO(out, rc = ELDLM_OK);
1829 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1830 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1831 GOTO(out, rc = ELDLM_OK);
1832 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1833 ldlm_grant_lock(lock, NULL);
1834 GOTO(out, rc = ELDLM_OK);
1836 /* If no flags, fall through to normal enqueue path. */
1839 policy = ldlm_processing_policy_table[res->lr_type];
1840 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, NULL);
1844 CERROR("This is client-side-only module, cannot handle "
1845 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1851 unlock_res_and_lock(lock);
1853 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1857 #ifdef HAVE_SERVER_SUPPORT
1859 * Iterate through all waiting locks on a given resource queue and attempt to
1862 * Must be called with resource lock held.
1864 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1865 struct list_head *work_list,
1866 enum ldlm_process_intention intention)
1868 struct list_head *tmp, *pos;
1869 ldlm_processing_policy policy;
1871 int rc = LDLM_ITER_CONTINUE;
1872 enum ldlm_error err;
1875 check_res_locked(res);
1877 policy = ldlm_processing_policy_table[res->lr_type];
1879 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1880 intention == LDLM_PROCESS_RECOVERY);
1882 list_for_each_safe(tmp, pos, queue) {
1883 struct ldlm_lock *pending;
1885 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1887 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1890 rc = policy(pending, &flags, intention, &err, work_list);
1892 * When this is called from recovery done, we always want
1893 * to scan the whole list no matter what 'rc' is returned.
1895 if (rc != LDLM_ITER_CONTINUE &&
1896 intention == LDLM_PROCESS_RESCAN)
1900 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
1904 * Conflicting locks are detected for a lock to be enqueued, add the lock
1905 * into waiting list and send blocking ASTs to the conflicting locks.
1907 * \param[in] lock The lock to be enqueued.
1908 * \param[out] flags Lock flags for the lock to be enqueued.
1909 * \param[in] rpc_list Conflicting locks list.
1910 * \param[in] grant_flags extra flags when granting a lock.
1912 * \retval -ERESTART: Some lock was instantly canceled while sending
1913 * blocking ASTs, caller needs to re-check conflicting
1915 * \retval -EAGAIN: Lock was destroyed, caller should return error.
1916 * \reval 0: Lock is successfully added in waiting list.
1918 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
1919 struct list_head *rpc_list, __u64 grant_flags)
1921 struct ldlm_resource *res = lock->l_resource;
1925 check_res_locked(res);
1927 /* If either of the compat_queue()s returned failure, then we
1928 * have ASTs to send and must go onto the waiting list.
1930 * bug 2322: we used to unlink and re-add here, which was a
1931 * terrible folly -- if we goto restart, we could get
1932 * re-ordered! Causes deadlock, because ASTs aren't sent! */
1933 if (list_empty(&lock->l_res_link))
1934 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1937 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
1940 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
1941 !ns_is_client(ldlm_res_to_ns(res)))
1942 class_fail_export(lock->l_export);
1945 if (rc == -ERESTART) {
1946 /* 15715: The lock was granted and destroyed after
1947 * resource lock was dropped. Interval node was freed
1948 * in ldlm_lock_destroy. Anyway, this always happens
1949 * when a client is being evicted. So it would be
1950 * ok to return an error. -jay */
1951 if (ldlm_is_destroyed(lock))
1954 /* lock was granted while resource was unlocked. */
1955 if (lock->l_granted_mode == lock->l_req_mode) {
1956 /* bug 11300: if the lock has been granted,
1957 * break earlier because otherwise, we will go
1958 * to restart and ldlm_resource_unlink will be
1959 * called and it causes the interval node to be
1960 * freed. Then we will fail at
1961 * ldlm_extent_add_lock() */
1962 *flags &= ~LDLM_FL_BLOCKED_MASK;
1968 *flags |= (LDLM_FL_BLOCK_GRANTED | grant_flags);
1974 * Discard all AST work items from list.
1976 * If for whatever reason we do not want to send ASTs to conflicting locks
1977 * anymore, disassemble the list with this function.
1979 void ldlm_discard_bl_list(struct list_head *bl_list)
1981 struct list_head *tmp, *pos;
1984 list_for_each_safe(pos, tmp, bl_list) {
1985 struct ldlm_lock *lock =
1986 list_entry(pos, struct ldlm_lock, l_bl_ast);
1988 list_del_init(&lock->l_bl_ast);
1989 LASSERT(ldlm_is_ast_sent(lock));
1990 ldlm_clear_ast_sent(lock);
1991 LASSERT(lock->l_bl_ast_run == 0);
1992 LASSERT(lock->l_blocking_lock);
1993 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1994 lock->l_blocking_lock = NULL;
1995 LDLM_LOCK_RELEASE(lock);
2003 * Process a call to blocking AST callback for a lock in ast_work list
2006 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2008 struct ldlm_cb_set_arg *arg = opaq;
2009 struct ldlm_lock_desc d;
2011 struct ldlm_lock *lock;
2014 if (list_empty(arg->list))
2017 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2019 /* nobody should touch l_bl_ast */
2020 lock_res_and_lock(lock);
2021 list_del_init(&lock->l_bl_ast);
2023 LASSERT(ldlm_is_ast_sent(lock));
2024 LASSERT(lock->l_bl_ast_run == 0);
2025 LASSERT(lock->l_blocking_lock);
2026 lock->l_bl_ast_run++;
2027 unlock_res_and_lock(lock);
2029 ldlm_lock2desc(lock->l_blocking_lock, &d);
2031 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2032 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
2033 lock->l_blocking_lock = NULL;
2034 LDLM_LOCK_RELEASE(lock);
2040 * Process a call to completion AST callback for a lock in ast_work list
2043 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2045 struct ldlm_cb_set_arg *arg = opaq;
2047 struct ldlm_lock *lock;
2048 ldlm_completion_callback completion_callback;
2051 if (list_empty(arg->list))
2054 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2056 /* It's possible to receive a completion AST before we've set
2057 * the l_completion_ast pointer: either because the AST arrived
2058 * before the reply, or simply because there's a small race
2059 * window between receiving the reply and finishing the local
2060 * enqueue. (bug 842)
2062 * This can't happen with the blocking_ast, however, because we
2063 * will never call the local blocking_ast until we drop our
2064 * reader/writer reference, which we won't do until we get the
2065 * reply and finish enqueueing. */
2067 /* nobody should touch l_cp_ast */
2068 lock_res_and_lock(lock);
2069 list_del_init(&lock->l_cp_ast);
2070 LASSERT(ldlm_is_cp_reqd(lock));
2071 /* save l_completion_ast since it can be changed by
2072 * mds_intent_policy(), see bug 14225 */
2073 completion_callback = lock->l_completion_ast;
2074 ldlm_clear_cp_reqd(lock);
2075 unlock_res_and_lock(lock);
2077 if (completion_callback != NULL)
2078 rc = completion_callback(lock, 0, (void *)arg);
2079 LDLM_LOCK_RELEASE(lock);
2085 * Process a call to revocation AST callback for a lock in ast_work list
2088 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2090 struct ldlm_cb_set_arg *arg = opaq;
2091 struct ldlm_lock_desc desc;
2093 struct ldlm_lock *lock;
2096 if (list_empty(arg->list))
2099 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2100 list_del_init(&lock->l_rk_ast);
2102 /* the desc just pretend to exclusive */
2103 ldlm_lock2desc(lock, &desc);
2104 desc.l_req_mode = LCK_EX;
2105 desc.l_granted_mode = 0;
2107 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2108 LDLM_LOCK_RELEASE(lock);
2114 * Process a call to glimpse AST callback for a lock in ast_work list
2116 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2118 struct ldlm_cb_set_arg *arg = opaq;
2119 struct ldlm_glimpse_work *gl_work;
2120 struct ldlm_lock *lock;
2124 if (list_empty(arg->list))
2127 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2129 list_del_init(&gl_work->gl_list);
2131 lock = gl_work->gl_lock;
2133 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2134 arg->gl_desc = gl_work->gl_desc;
2135 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2136 arg->gl_interpret_data = gl_work->gl_interpret_data;
2138 /* invoke the actual glimpse callback */
2139 if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
2142 LDLM_LOCK_RELEASE(lock);
2144 if ((gl_work->gl_flags & LDLM_GL_WORK_NOFREE) == 0)
2145 OBD_FREE_PTR(gl_work);
2151 * Process list of locks in need of ASTs being sent.
2153 * Used on server to send multiple ASTs together instead of sending one by
2156 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2157 ldlm_desc_ast_t ast_type)
2159 struct ldlm_cb_set_arg *arg;
2160 set_producer_func work_ast_lock;
2163 if (list_empty(rpc_list))
2170 atomic_set(&arg->restart, 0);
2171 arg->list = rpc_list;
2174 case LDLM_WORK_BL_AST:
2175 arg->type = LDLM_BL_CALLBACK;
2176 work_ast_lock = ldlm_work_bl_ast_lock;
2178 case LDLM_WORK_CP_AST:
2179 arg->type = LDLM_CP_CALLBACK;
2180 work_ast_lock = ldlm_work_cp_ast_lock;
2182 case LDLM_WORK_REVOKE_AST:
2183 arg->type = LDLM_BL_CALLBACK;
2184 work_ast_lock = ldlm_work_revoke_ast_lock;
2186 case LDLM_WORK_GL_AST:
2187 arg->type = LDLM_GL_CALLBACK;
2188 work_ast_lock = ldlm_work_gl_ast_lock;
2194 /* We create a ptlrpc request set with flow control extension.
2195 * This request set will use the work_ast_lock function to produce new
2196 * requests and will send a new request each time one completes in order
2197 * to keep the number of requests in flight to ns_max_parallel_ast */
2198 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2199 work_ast_lock, arg);
2200 if (arg->set == NULL)
2201 GOTO(out, rc = -ENOMEM);
2203 ptlrpc_set_wait(arg->set);
2204 ptlrpc_set_destroy(arg->set);
2206 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2214 * Try to grant all waiting locks on a resource.
2216 * Calls ldlm_reprocess_queue on converting and waiting queues.
2218 * Typically called after some resource locks are cancelled to see
2219 * if anything could be granted as a result of the cancellation.
2221 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2222 enum ldlm_process_intention intention)
2224 struct list_head rpc_list;
2225 #ifdef HAVE_SERVER_SUPPORT
2226 struct obd_device *obd;
2230 INIT_LIST_HEAD(&rpc_list);
2231 /* Local lock trees don't get reprocessed. */
2232 if (ns_is_client(ldlm_res_to_ns(res))) {
2237 /* Disable reprocess during lock replay stage but allow during
2238 * request replay stage.
2240 obd = ldlm_res_to_ns(res)->ns_obd;
2241 if (obd->obd_recovering &&
2242 atomic_read(&obd->obd_req_replay_clients) == 0)
2246 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list,
2248 if (rc == LDLM_ITER_CONTINUE)
2249 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list,
2253 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2255 if (rc == -ERESTART) {
2256 LASSERT(list_empty(&rpc_list));
2262 INIT_LIST_HEAD(&rpc_list);
2263 if (!ns_is_client(ldlm_res_to_ns(res))) {
2264 CERROR("This is client-side-only module, cannot handle "
2265 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2272 void ldlm_reprocess_all(struct ldlm_resource *res)
2274 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN);
2276 EXPORT_SYMBOL(ldlm_reprocess_all);
2278 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2279 struct hlist_node *hnode, void *arg)
2281 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2283 /* This is only called once after recovery done. LU-8306. */
2284 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY);
2289 * Iterate through all resources on a namespace attempting to grant waiting
2292 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2297 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2298 ldlm_reprocess_res, NULL, 0);
2304 * Helper function to call blocking AST for LDLM lock \a lock in a
2305 * "cancelling" mode.
2307 void ldlm_cancel_callback(struct ldlm_lock *lock)
2309 check_res_locked(lock->l_resource);
2310 if (!ldlm_is_cancel(lock)) {
2311 ldlm_set_cancel(lock);
2312 if (lock->l_blocking_ast) {
2313 unlock_res_and_lock(lock);
2314 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2316 lock_res_and_lock(lock);
2318 LDLM_DEBUG(lock, "no blocking ast");
2321 /* only canceller can set bl_done bit */
2322 ldlm_set_bl_done(lock);
2323 wake_up_all(&lock->l_waitq);
2324 } else if (!ldlm_is_bl_done(lock)) {
2325 struct l_wait_info lwi = { 0 };
2327 /* The lock is guaranteed to have been canceled once
2328 * returning from this function. */
2329 unlock_res_and_lock(lock);
2330 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
2331 lock_res_and_lock(lock);
2336 * Remove skiplist-enabled LDLM lock \a req from granted list
2338 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2340 if (req->l_resource->lr_type != LDLM_PLAIN &&
2341 req->l_resource->lr_type != LDLM_IBITS)
2344 list_del_init(&req->l_sl_policy);
2345 list_del_init(&req->l_sl_mode);
2349 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2351 void ldlm_lock_cancel(struct ldlm_lock *lock)
2353 struct ldlm_resource *res;
2354 struct ldlm_namespace *ns;
2357 lock_res_and_lock(lock);
2359 res = lock->l_resource;
2360 ns = ldlm_res_to_ns(res);
2362 /* Please do not, no matter how tempting, remove this LBUG without
2363 * talking to me first. -phik */
2364 if (lock->l_readers || lock->l_writers) {
2365 LDLM_ERROR(lock, "lock still has references");
2369 if (ldlm_is_waited(lock))
2370 ldlm_del_waiting_lock(lock);
2372 /* Releases cancel callback. */
2373 ldlm_cancel_callback(lock);
2375 /* Yes, second time, just in case it was added again while we were
2376 * running with no res lock in ldlm_cancel_callback */
2377 if (ldlm_is_waited(lock))
2378 ldlm_del_waiting_lock(lock);
2380 ldlm_resource_unlink_lock(lock);
2381 ldlm_lock_destroy_nolock(lock);
2383 if (lock->l_granted_mode == lock->l_req_mode)
2384 ldlm_pool_del(&ns->ns_pool, lock);
2386 /* Make sure we will not be called again for same lock what is possible
2387 * if not to zero out lock->l_granted_mode */
2388 lock->l_granted_mode = LCK_MINMODE;
2389 unlock_res_and_lock(lock);
2393 EXPORT_SYMBOL(ldlm_lock_cancel);
2396 * Set opaque data into the lock that only makes sense to upper layer.
2398 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2400 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2405 if (lock->l_ast_data == NULL)
2406 lock->l_ast_data = data;
2407 if (lock->l_ast_data == data)
2409 LDLM_LOCK_PUT(lock);
2413 EXPORT_SYMBOL(ldlm_lock_set_data);
2415 struct export_cl_data {
2416 struct obd_export *ecl_exp;
2420 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2421 struct ldlm_lock *lock,
2422 struct export_cl_data *ecl)
2424 struct ldlm_resource *res;
2426 res = ldlm_resource_getref(lock->l_resource);
2428 ldlm_res_lvbo_update(res, NULL, 1);
2429 ldlm_lock_cancel(lock);
2430 if (!exp->exp_obd->obd_stopping)
2431 ldlm_reprocess_all(res);
2432 ldlm_resource_putref(res);
2435 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2436 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2437 exp, ecl->ecl_loop);
2442 * Iterator function for ldlm_export_cancel_locks.
2443 * Cancels passed locks.
2446 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2447 struct hlist_node *hnode, void *data)
2450 struct export_cl_data *ecl = (struct export_cl_data *)data;
2451 struct obd_export *exp = ecl->ecl_exp;
2452 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2454 LDLM_LOCK_GET(lock);
2455 ldlm_cancel_lock_for_export(exp, lock, ecl);
2456 LDLM_LOCK_RELEASE(lock);
2462 * Cancel all blocked locks for given export.
2464 * Typically called on client disconnection/eviction
2466 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2468 struct export_cl_data ecl = {
2473 while (!list_empty(&exp->exp_bl_list)) {
2474 struct ldlm_lock *lock;
2476 spin_lock_bh(&exp->exp_bl_list_lock);
2477 if (!list_empty(&exp->exp_bl_list)) {
2478 lock = list_entry(exp->exp_bl_list.next,
2479 struct ldlm_lock, l_exp_list);
2480 LDLM_LOCK_GET(lock);
2481 list_del_init(&lock->l_exp_list);
2485 spin_unlock_bh(&exp->exp_bl_list_lock);
2490 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2491 LDLM_LOCK_RELEASE(lock);
2494 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2495 "left on hash table %d.\n", exp, ecl.ecl_loop,
2496 atomic_read(&exp->exp_lock_hash->hs_count));
2498 return ecl.ecl_loop;
2502 * Cancel all locks for given export.
2504 * Typically called after client disconnection/eviction
2506 int ldlm_export_cancel_locks(struct obd_export *exp)
2508 struct export_cl_data ecl = {
2513 cfs_hash_for_each_empty(exp->exp_lock_hash,
2514 ldlm_cancel_locks_for_export_cb, &ecl);
2516 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2517 "left on hash table %d.\n", exp, ecl.ecl_loop,
2518 atomic_read(&exp->exp_lock_hash->hs_count));
2520 if (ecl.ecl_loop > 0 &&
2521 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2522 exp->exp_obd->obd_stopping)
2523 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2525 return ecl.ecl_loop;
2529 * Downgrade an exclusive lock.
2531 * A fast variant of ldlm_lock_convert for convertion of exclusive locks. The
2532 * convertion may fail if lock was canceled before downgrade, but it doesn't
2533 * indicate any problem, because such lock has no reader or writer, and will
2535 * Used by Commit on Sharing (COS) code.
2537 * \param lock A lock to convert
2538 * \param new_mode new lock mode
2540 void ldlm_lock_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2544 LASSERT(new_mode == LCK_COS);
2546 lock_res_and_lock(lock);
2548 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2549 unlock_res_and_lock(lock);
2551 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2552 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2556 ldlm_resource_unlink_lock(lock);
2558 * Remove the lock from pool as it will be added again in
2559 * ldlm_grant_lock() called below.
2561 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2562 lock->l_req_mode = new_mode;
2563 ldlm_grant_lock(lock, NULL);
2565 unlock_res_and_lock(lock);
2567 ldlm_reprocess_all(lock->l_resource);
2571 EXPORT_SYMBOL(ldlm_lock_downgrade);
2574 * Attempt to convert already granted lock to a different mode.
2576 * While lock conversion is not currently used, future client-side
2577 * optimizations could take advantage of it to avoid discarding cached
2580 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock,
2581 enum ldlm_mode new_mode, __u32 *flags)
2583 struct list_head rpc_list;
2584 struct ldlm_resource *res;
2585 struct ldlm_namespace *ns;
2587 #ifdef HAVE_SERVER_SUPPORT
2589 struct sl_insert_point prev;
2591 struct ldlm_interval *node;
2594 INIT_LIST_HEAD(&rpc_list);
2595 /* Just return if mode is unchanged. */
2596 if (new_mode == lock->l_granted_mode) {
2597 *flags |= LDLM_FL_BLOCK_GRANTED;
2598 RETURN(lock->l_resource);
2601 /* I can't check the type of lock here because the bitlock of lock
2602 * is not held here, so do the allocation blindly. -jay */
2603 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
2604 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
2607 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
2608 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
2610 lock_res_and_lock(lock);
2612 res = lock->l_resource;
2613 ns = ldlm_res_to_ns(res);
2615 #ifdef HAVE_SERVER_SUPPORT
2616 old_mode = lock->l_req_mode;
2618 lock->l_req_mode = new_mode;
2619 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
2620 #ifdef HAVE_SERVER_SUPPORT
2621 /* remember the lock position where the lock might be
2622 * added back to the granted list later and also
2623 * remember the join mode for skiplist fixing. */
2624 prev.res_link = lock->l_res_link.prev;
2625 prev.mode_link = lock->l_sl_mode.prev;
2626 prev.policy_link = lock->l_sl_policy.prev;
2628 ldlm_resource_unlink_lock(lock);
2630 ldlm_resource_unlink_lock(lock);
2631 if (res->lr_type == LDLM_EXTENT) {
2632 /* FIXME: ugly code, I have to attach the lock to a
2633 * interval node again since perhaps it will be granted
2635 INIT_LIST_HEAD(&node->li_group);
2636 ldlm_interval_attach(node, lock);
2642 * Remove old lock from the pool before adding the lock with new
2643 * mode below in ->policy()
2645 ldlm_pool_del(&ns->ns_pool, lock);
2647 /* If this is a local resource, put it on the appropriate list. */
2648 if (ns_is_client(ldlm_res_to_ns(res))) {
2649 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
2650 ldlm_resource_add_lock(res, &res->lr_converting, lock);
2652 /* This should never happen, because of the way the
2653 * server handles conversions. */
2654 LDLM_ERROR(lock, "Erroneous flags %x on local lock\n",
2658 ldlm_grant_lock(lock, &rpc_list);
2660 /* FIXME: completion handling not with lr_lock held ! */
2661 if (lock->l_completion_ast)
2662 lock->l_completion_ast(lock, 0, NULL);
2664 #ifdef HAVE_SERVER_SUPPORT
2667 enum ldlm_error err;
2669 ldlm_processing_policy policy;
2671 policy = ldlm_processing_policy_table[res->lr_type];
2672 rc = policy(lock, &pflags, LDLM_PROCESS_RESCAN, &err,
2674 if (rc == LDLM_ITER_STOP) {
2675 lock->l_req_mode = old_mode;
2676 if (res->lr_type == LDLM_EXTENT)
2677 ldlm_extent_add_lock(res, lock);
2679 ldlm_granted_list_add_lock(lock, &prev);
2683 *flags |= LDLM_FL_BLOCK_GRANTED;
2689 CERROR("This is client-side-only module, cannot handle "
2690 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2694 unlock_res_and_lock(lock);
2697 ldlm_run_ast_work(ns, &rpc_list, LDLM_WORK_CP_AST);
2699 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
2704 * Print lock with lock handle \a lockh description into debug log.
2706 * Used when printing all locks on a resource for debug purposes.
2708 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2710 struct ldlm_lock *lock;
2712 if (!((libcfs_debug | D_ERROR) & level))
2715 lock = ldlm_handle2lock(lockh);
2719 LDLM_DEBUG_LIMIT(level, lock, "###");
2721 LDLM_LOCK_PUT(lock);
2723 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2726 * Print lock information with custom message into debug log.
2729 void _ldlm_lock_debug(struct ldlm_lock *lock,
2730 struct libcfs_debug_msg_data *msgdata,
2731 const char *fmt, ...)
2734 struct obd_export *exp = lock->l_export;
2735 struct ldlm_resource *resource = NULL;
2736 char *nid = "local";
2738 /* on server-side resource of lock doesn't change */
2739 if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2740 if (lock->l_resource != NULL)
2741 resource = ldlm_resource_getref(lock->l_resource);
2742 } else if (spin_trylock(&lock->l_lock)) {
2743 if (lock->l_resource != NULL)
2744 resource = ldlm_resource_getref(lock->l_resource);
2745 spin_unlock(&lock->l_lock);
2748 va_start(args, fmt);
2750 if (exp && exp->exp_connection) {
2751 nid = libcfs_nid2str(exp->exp_connection->c_peer.nid);
2752 } else if (exp && exp->exp_obd != NULL) {
2753 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2754 nid = libcfs_nid2str(imp->imp_connection->c_peer.nid);
2757 if (resource == NULL) {
2758 libcfs_debug_vmsg2(msgdata, fmt, args,
2759 " ns: \?\? lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2760 "res: \?\? rrc=\?\? type: \?\?\? flags: %#llx nid: %s "
2761 "remote: %#llx expref: %d pid: %u timeout: %lu "
2764 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2765 lock->l_readers, lock->l_writers,
2766 ldlm_lockname[lock->l_granted_mode],
2767 ldlm_lockname[lock->l_req_mode],
2768 lock->l_flags, nid, lock->l_remote_handle.cookie,
2769 exp ? atomic_read(&exp->exp_refcount) : -99,
2770 lock->l_pid, lock->l_callback_timeout, lock->l_lvb_type);
2775 switch (resource->lr_type) {
2777 libcfs_debug_vmsg2(msgdata, fmt, args,
2778 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2779 "res: "DLDLMRES" rrc: %d type: %s [%llu->%llu] "
2780 "(req %llu->%llu) flags: %#llx nid: %s remote: "
2781 "%#llx expref: %d pid: %u timeout: %lu lvb_type: %d\n",
2782 ldlm_lock_to_ns_name(lock), lock,
2783 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2784 lock->l_readers, lock->l_writers,
2785 ldlm_lockname[lock->l_granted_mode],
2786 ldlm_lockname[lock->l_req_mode],
2788 atomic_read(&resource->lr_refcount),
2789 ldlm_typename[resource->lr_type],
2790 lock->l_policy_data.l_extent.start,
2791 lock->l_policy_data.l_extent.end,
2792 lock->l_req_extent.start, lock->l_req_extent.end,
2793 lock->l_flags, nid, lock->l_remote_handle.cookie,
2794 exp ? atomic_read(&exp->exp_refcount) : -99,
2795 lock->l_pid, lock->l_callback_timeout,
2800 libcfs_debug_vmsg2(msgdata, fmt, args,
2801 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2802 "res: "DLDLMRES" rrc: %d type: %s pid: %d "
2803 "[%llu->%llu] flags: %#llx nid: %s "
2804 "remote: %#llx expref: %d pid: %u timeout: %lu\n",
2805 ldlm_lock_to_ns_name(lock), lock,
2806 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2807 lock->l_readers, lock->l_writers,
2808 ldlm_lockname[lock->l_granted_mode],
2809 ldlm_lockname[lock->l_req_mode],
2811 atomic_read(&resource->lr_refcount),
2812 ldlm_typename[resource->lr_type],
2813 lock->l_policy_data.l_flock.pid,
2814 lock->l_policy_data.l_flock.start,
2815 lock->l_policy_data.l_flock.end,
2816 lock->l_flags, nid, lock->l_remote_handle.cookie,
2817 exp ? atomic_read(&exp->exp_refcount) : -99,
2818 lock->l_pid, lock->l_callback_timeout);
2822 libcfs_debug_vmsg2(msgdata, fmt, args,
2823 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2824 "res: "DLDLMRES" bits %#llx rrc: %d type: %s "
2825 "flags: %#llx nid: %s remote: %#llx expref: %d "
2826 "pid: %u timeout: %lu lvb_type: %d\n",
2827 ldlm_lock_to_ns_name(lock),
2828 lock, lock->l_handle.h_cookie,
2829 atomic_read(&lock->l_refc),
2830 lock->l_readers, lock->l_writers,
2831 ldlm_lockname[lock->l_granted_mode],
2832 ldlm_lockname[lock->l_req_mode],
2834 lock->l_policy_data.l_inodebits.bits,
2835 atomic_read(&resource->lr_refcount),
2836 ldlm_typename[resource->lr_type],
2837 lock->l_flags, nid, lock->l_remote_handle.cookie,
2838 exp ? atomic_read(&exp->exp_refcount) : -99,
2839 lock->l_pid, lock->l_callback_timeout,
2844 libcfs_debug_vmsg2(msgdata, fmt, args,
2845 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2846 "res: "DLDLMRES" rrc: %d type: %s flags: %#llx "
2847 "nid: %s remote: %#llx expref: %d pid: %u "
2848 "timeout: %lu lvb_type: %d\n",
2849 ldlm_lock_to_ns_name(lock),
2850 lock, lock->l_handle.h_cookie,
2851 atomic_read(&lock->l_refc),
2852 lock->l_readers, lock->l_writers,
2853 ldlm_lockname[lock->l_granted_mode],
2854 ldlm_lockname[lock->l_req_mode],
2856 atomic_read(&resource->lr_refcount),
2857 ldlm_typename[resource->lr_type],
2858 lock->l_flags, nid, lock->l_remote_handle.cookie,
2859 exp ? atomic_read(&exp->exp_refcount) : -99,
2860 lock->l_pid, lock->l_callback_timeout,
2865 ldlm_resource_putref(resource);
2867 EXPORT_SYMBOL(_ldlm_lock_debug);