4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_lock.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <libcfs/libcfs.h>
42 #include <lustre_swab.h>
43 #include <obd_class.h>
45 #include "ldlm_internal.h"
47 struct kmem_cache *ldlm_glimpse_work_kmem;
48 EXPORT_SYMBOL(ldlm_glimpse_work_kmem);
51 char *ldlm_lockname[] = {
59 [LCK_GROUP] = "GROUP",
62 EXPORT_SYMBOL(ldlm_lockname);
64 char *ldlm_typename[] = {
66 [LDLM_EXTENT] = "EXT",
71 static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = {
72 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_wire_to_local,
73 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_wire_to_local,
74 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_wire_to_local,
75 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_wire_to_local,
78 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
79 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_local_to_wire,
80 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_local_to_wire,
81 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_local_to_wire,
82 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_local_to_wire,
86 * Converts lock policy from local format to on the wire lock_desc format
88 void ldlm_convert_policy_to_wire(enum ldlm_type type,
89 const union ldlm_policy_data *lpolicy,
90 union ldlm_wire_policy_data *wpolicy)
92 ldlm_policy_local_to_wire_t convert;
94 convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
96 convert(lpolicy, wpolicy);
100 * Converts lock policy from on the wire lock_desc format to local format
102 void ldlm_convert_policy_to_local(struct obd_export *exp, enum ldlm_type type,
103 const union ldlm_wire_policy_data *wpolicy,
104 union ldlm_policy_data *lpolicy)
106 ldlm_policy_wire_to_local_t convert;
108 convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE];
110 convert(wpolicy, lpolicy);
113 const char *ldlm_it2str(enum ldlm_intent_flags it)
120 case (IT_OPEN | IT_CREAT):
133 CERROR("Unknown intent 0x%08x\n", it);
137 EXPORT_SYMBOL(ldlm_it2str);
139 #ifdef HAVE_SERVER_SUPPORT
140 static ldlm_processing_policy ldlm_processing_policy_table[] = {
141 [LDLM_PLAIN] = ldlm_process_plain_lock,
142 [LDLM_EXTENT] = ldlm_process_extent_lock,
143 [LDLM_FLOCK] = ldlm_process_flock_lock,
144 [LDLM_IBITS] = ldlm_process_inodebits_lock,
147 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
149 return ldlm_processing_policy_table[res->lr_type];
151 EXPORT_SYMBOL(ldlm_get_processing_policy);
152 #endif /* HAVE_SERVER_SUPPORT */
154 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
158 EXPORT_SYMBOL(ldlm_register_intent);
161 * REFCOUNTED LOCK OBJECTS
166 * Get a reference on a lock.
168 * Lock refcounts, during creation:
169 * - one special one for allocation, dec'd only once in destroy
170 * - one for being a lock that's in-use
171 * - one for the addref associated with a new lock
173 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
175 atomic_inc(&lock->l_refc);
178 EXPORT_SYMBOL(ldlm_lock_get);
181 * Release lock reference.
183 * Also frees the lock if it was last reference.
185 void ldlm_lock_put(struct ldlm_lock *lock)
189 LASSERT(lock->l_resource != LP_POISON);
190 LASSERT(atomic_read(&lock->l_refc) > 0);
191 if (atomic_dec_and_test(&lock->l_refc)) {
192 struct ldlm_resource *res;
195 "final lock_put on destroyed lock, freeing it.");
197 res = lock->l_resource;
198 LASSERT(ldlm_is_destroyed(lock));
199 LASSERT(list_empty(&lock->l_exp_list));
200 LASSERT(list_empty(&lock->l_res_link));
201 LASSERT(list_empty(&lock->l_pending_chain));
203 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
205 lu_ref_del(&res->lr_reference, "lock", lock);
206 ldlm_resource_putref(res);
207 lock->l_resource = NULL;
208 if (lock->l_export) {
209 class_export_lock_put(lock->l_export, lock);
210 lock->l_export = NULL;
213 if (lock->l_lvb_data != NULL)
214 OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
216 ldlm_interval_free(ldlm_interval_detach(lock));
217 lu_ref_fini(&lock->l_reference);
218 OBD_FREE_RCU(lock, sizeof(*lock), &lock->l_handle);
223 EXPORT_SYMBOL(ldlm_lock_put);
226 * Removes LDLM lock \a lock from LRU. Assumes LRU is already locked.
228 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
231 if (!list_empty(&lock->l_lru)) {
232 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
234 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
235 if (ns->ns_last_pos == &lock->l_lru)
236 ns->ns_last_pos = lock->l_lru.prev;
237 list_del_init(&lock->l_lru);
238 LASSERT(ns->ns_nr_unused > 0);
246 * Removes LDLM lock \a lock from LRU. Obtains the LRU lock first.
248 * If \a last_use is non-zero, it will remove the lock from LRU only if
249 * it matches lock's l_last_used.
251 * \retval 0 if \a last_use is set, the lock is not in LRU list or \a last_use
252 * doesn't match lock's l_last_used;
253 * otherwise, the lock hasn't been in the LRU list.
254 * \retval 1 the lock was in LRU list and removed.
256 int ldlm_lock_remove_from_lru_check(struct ldlm_lock *lock, ktime_t last_use)
258 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
262 if (ldlm_is_ns_srv(lock)) {
263 LASSERT(list_empty(&lock->l_lru));
267 spin_lock(&ns->ns_lock);
268 if (!ktime_compare(last_use, ktime_set(0, 0)) ||
269 !ktime_compare(last_use, lock->l_last_used))
270 rc = ldlm_lock_remove_from_lru_nolock(lock);
271 spin_unlock(&ns->ns_lock);
277 * Adds LDLM lock \a lock to namespace LRU. Assumes LRU is already locked.
279 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
281 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
283 lock->l_last_used = ktime_get();
284 LASSERT(list_empty(&lock->l_lru));
285 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
286 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
287 LASSERT(ns->ns_nr_unused >= 0);
292 * Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks
295 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
297 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
300 spin_lock(&ns->ns_lock);
301 ldlm_lock_add_to_lru_nolock(lock);
302 spin_unlock(&ns->ns_lock);
307 * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
308 * the LRU. Performs necessary LRU locking
310 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
312 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
315 if (ldlm_is_ns_srv(lock)) {
316 LASSERT(list_empty(&lock->l_lru));
321 spin_lock(&ns->ns_lock);
322 if (!list_empty(&lock->l_lru)) {
323 ldlm_lock_remove_from_lru_nolock(lock);
324 ldlm_lock_add_to_lru_nolock(lock);
326 spin_unlock(&ns->ns_lock);
331 * Helper to destroy a locked lock.
333 * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
334 * Must be called with l_lock and lr_lock held.
336 * Does not actually free the lock data, but rather marks the lock as
337 * destroyed by setting l_destroyed field in the lock to 1. Destroys a
338 * handle->lock association too, so that the lock can no longer be found
339 * and removes the lock from LRU list. Actual lock freeing occurs when
340 * last lock reference goes away.
342 * Original comment (of some historical value):
343 * This used to have a 'strict' flag, which recovery would use to mark an
344 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
345 * shall explain why it's gone: with the new hash table scheme, once you call
346 * ldlm_lock_destroy, you can never drop your final references on this lock.
347 * Because it's not in the hash table anymore. -phil
349 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
353 if (lock->l_readers || lock->l_writers) {
354 LDLM_ERROR(lock, "lock still has references");
358 if (!list_empty(&lock->l_res_link)) {
359 LDLM_ERROR(lock, "lock still on resource");
363 if (ldlm_is_destroyed(lock)) {
364 LASSERT(list_empty(&lock->l_lru));
368 ldlm_set_destroyed(lock);
370 if (lock->l_export && lock->l_export->exp_lock_hash) {
371 /* NB: it's safe to call cfs_hash_del() even lock isn't
372 * in exp_lock_hash. */
373 /* In the function below, .hs_keycmp resolves to
374 * ldlm_export_lock_keycmp() */
375 /* coverity[overrun-buffer-val] */
376 cfs_hash_del(lock->l_export->exp_lock_hash,
377 &lock->l_remote_handle, &lock->l_exp_hash);
380 ldlm_lock_remove_from_lru(lock);
381 class_handle_unhash(&lock->l_handle);
388 * Destroys a LDLM lock \a lock. Performs necessary locking first.
390 void ldlm_lock_destroy(struct ldlm_lock *lock)
394 lock_res_and_lock(lock);
395 first = ldlm_lock_destroy_internal(lock);
396 unlock_res_and_lock(lock);
398 /* drop reference from hashtable only for first destroy */
400 lu_ref_del(&lock->l_reference, "hash", lock);
401 LDLM_LOCK_RELEASE(lock);
407 * Destroys a LDLM lock \a lock that is already locked.
409 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
413 first = ldlm_lock_destroy_internal(lock);
414 /* drop reference from hashtable only for first destroy */
416 lu_ref_del(&lock->l_reference, "hash", lock);
417 LDLM_LOCK_RELEASE(lock);
422 /* this is called by portals_handle2object with the handle lock taken */
423 static void lock_handle_addref(void *lock)
425 LDLM_LOCK_GET((struct ldlm_lock *)lock);
428 static void lock_handle_free(void *lock, int size)
430 LASSERT(size == sizeof(struct ldlm_lock));
431 OBD_SLAB_FREE(lock, ldlm_lock_slab, size);
434 static struct portals_handle_ops lock_handle_ops = {
435 .hop_addref = lock_handle_addref,
436 .hop_free = lock_handle_free,
441 * Allocate and initialize new lock structure.
443 * usage: pass in a resource on which you have done ldlm_resource_get
444 * new lock will take over the refcount.
445 * returns: lock with refcount 2 - one for current caller and one for remote
447 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
449 struct ldlm_lock *lock;
452 if (resource == NULL)
455 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, GFP_NOFS);
459 spin_lock_init(&lock->l_lock);
460 lock->l_resource = resource;
461 lu_ref_add(&resource->lr_reference, "lock", lock);
463 atomic_set(&lock->l_refc, 2);
464 INIT_LIST_HEAD(&lock->l_res_link);
465 INIT_LIST_HEAD(&lock->l_lru);
466 INIT_LIST_HEAD(&lock->l_pending_chain);
467 INIT_LIST_HEAD(&lock->l_bl_ast);
468 INIT_LIST_HEAD(&lock->l_cp_ast);
469 INIT_LIST_HEAD(&lock->l_rk_ast);
470 init_waitqueue_head(&lock->l_waitq);
471 lock->l_blocking_lock = NULL;
472 INIT_LIST_HEAD(&lock->l_sl_mode);
473 INIT_LIST_HEAD(&lock->l_sl_policy);
474 INIT_HLIST_NODE(&lock->l_exp_hash);
475 INIT_HLIST_NODE(&lock->l_exp_flock_hash);
477 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
479 INIT_LIST_HEAD_RCU(&lock->l_handle.h_link);
480 class_handle_hash(&lock->l_handle, &lock_handle_ops);
482 lu_ref_init(&lock->l_reference);
483 lu_ref_add(&lock->l_reference, "hash", lock);
484 lock->l_callback_timeout = 0;
485 lock->l_activity = 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, &lock_handle_ops);
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);
667 /* Lock can be converted from a blocking state back to granted
668 * after lock convert or COS downgrade but still be in an
669 * older bl_list because it is controlled only by
670 * ldlm_work_bl_ast_lock(), let it be processed there.
672 if (list_empty(&lock->l_bl_ast)) {
673 list_add(&lock->l_bl_ast, work_list);
676 LASSERT(lock->l_blocking_lock == NULL);
677 lock->l_blocking_lock = LDLM_LOCK_GET(new);
682 * Add a lock to list of just granted locks to send completion AST to.
684 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
685 struct list_head *work_list)
687 if (!ldlm_is_cp_reqd(lock)) {
688 ldlm_set_cp_reqd(lock);
689 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
690 LASSERT(list_empty(&lock->l_cp_ast));
691 list_add(&lock->l_cp_ast, work_list);
697 * Aggregator function to add AST work items into a list. Determines
698 * what sort of an AST work needs to be done and calls the proper
700 * Must be called with lr_lock held.
702 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
703 struct list_head *work_list)
706 check_res_locked(lock->l_resource);
708 ldlm_add_bl_work_item(lock, new, work_list);
710 ldlm_add_cp_work_item(lock, work_list);
715 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
716 * r/w reference type is determined by \a mode
717 * Calls ldlm_lock_addref_internal.
719 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
721 struct ldlm_lock *lock;
723 lock = ldlm_handle2lock(lockh);
724 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
725 ldlm_lock_addref_internal(lock, mode);
728 EXPORT_SYMBOL(ldlm_lock_addref);
732 * Add specified reader/writer reference to LDLM lock \a lock.
733 * r/w reference type is determined by \a mode
734 * Removes lock from LRU if it is there.
735 * Assumes the LDLM lock is already locked.
737 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
740 ldlm_lock_remove_from_lru(lock);
741 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
743 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
745 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
747 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
750 lu_ref_add_atomic(&lock->l_reference, "user", lock);
751 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
755 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
756 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
758 * \retval 0 success, lock was addref-ed
760 * \retval -EAGAIN lock is being canceled.
762 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
764 struct ldlm_lock *lock;
768 lock = ldlm_handle2lock(lockh);
770 lock_res_and_lock(lock);
771 if (lock->l_readers != 0 || lock->l_writers != 0 ||
772 !ldlm_is_cbpending(lock)) {
773 ldlm_lock_addref_internal_nolock(lock, mode);
776 unlock_res_and_lock(lock);
781 EXPORT_SYMBOL(ldlm_lock_addref_try);
784 * Add specified reader/writer reference to LDLM lock \a lock.
785 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
786 * Only called for local locks.
788 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
790 lock_res_and_lock(lock);
791 ldlm_lock_addref_internal_nolock(lock, mode);
792 unlock_res_and_lock(lock);
796 * Removes reader/writer reference for LDLM lock \a lock.
797 * Assumes LDLM lock is already locked.
798 * only called in ldlm_flock_destroy and for local locks.
799 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
800 * that cannot be placed in LRU.
802 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
805 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
806 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
807 LASSERT(lock->l_readers > 0);
808 lu_ref_del(&lock->l_reference, "reader", lock);
811 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
812 LASSERT(lock->l_writers > 0);
813 lu_ref_del(&lock->l_reference, "writer", lock);
817 lu_ref_del(&lock->l_reference, "user", lock);
818 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
822 * Removes reader/writer reference for LDLM lock \a lock.
823 * Locks LDLM lock first.
824 * If the lock is determined to be client lock on a client and r/w refcount
825 * drops to zero and the lock is not blocked, the lock is added to LRU lock
827 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
829 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
831 struct ldlm_namespace *ns;
834 lock_res_and_lock(lock);
836 ns = ldlm_lock_to_ns(lock);
838 ldlm_lock_decref_internal_nolock(lock, mode);
840 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
841 !lock->l_readers && !lock->l_writers) {
842 /* If this is a local lock on a server namespace and this was
843 * the last reference, cancel the lock.
845 * Group locks are special:
846 * They must not go in LRU, but they are not called back
847 * like non-group locks, instead they are manually released.
848 * They have an l_writers reference which they keep until
849 * they are manually released, so we remove them when they have
850 * no more reader or writer references. - LU-6368 */
851 ldlm_set_cbpending(lock);
854 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
855 /* If we received a blocked AST and this was the last reference,
856 * run the callback. */
857 if (ldlm_is_ns_srv(lock) && lock->l_export)
858 CERROR("FL_CBPENDING set on non-local lock--just a "
861 LDLM_DEBUG(lock, "final decref done on cbpending lock");
863 LDLM_LOCK_GET(lock); /* dropped by bl thread */
864 ldlm_lock_remove_from_lru(lock);
865 unlock_res_and_lock(lock);
867 if (ldlm_is_fail_loc(lock))
868 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
870 if (ldlm_is_atomic_cb(lock) ||
871 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
872 ldlm_handle_bl_callback(ns, NULL, lock);
873 } else if (ns_is_client(ns) &&
874 !lock->l_readers && !lock->l_writers &&
875 !ldlm_is_no_lru(lock) &&
876 !ldlm_is_bl_ast(lock) &&
877 !ldlm_is_converting(lock)) {
879 LDLM_DEBUG(lock, "add lock into lru list");
881 /* If this is a client-side namespace and this was the last
882 * reference, put it on the LRU. */
883 ldlm_lock_add_to_lru(lock);
884 unlock_res_and_lock(lock);
886 if (ldlm_is_fail_loc(lock))
887 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
889 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
890 * are not supported by the server, otherwise, it is done on
892 if (!exp_connect_cancelset(lock->l_conn_export) &&
893 !ns_connect_lru_resize(ns))
894 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
896 LDLM_DEBUG(lock, "do not add lock into lru list");
897 unlock_res_and_lock(lock);
904 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
906 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
908 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
909 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
910 ldlm_lock_decref_internal(lock, mode);
913 EXPORT_SYMBOL(ldlm_lock_decref);
916 * Decrease reader/writer refcount for LDLM lock with handle
917 * \a lockh and mark it for subsequent cancellation once r/w refcount
918 * drops to zero instead of putting into LRU.
921 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
924 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
927 LASSERT(lock != NULL);
929 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
930 lock_res_and_lock(lock);
931 ldlm_set_cbpending(lock);
932 unlock_res_and_lock(lock);
933 ldlm_lock_decref_internal(lock, mode);
936 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
938 struct sl_insert_point {
939 struct list_head *res_link;
940 struct list_head *mode_link;
941 struct list_head *policy_link;
945 * Finds a position to insert the new lock into granted lock list.
947 * Used for locks eligible for skiplist optimization.
950 * queue [input]: the granted list where search acts on;
951 * req [input]: the lock whose position to be located;
952 * prev [output]: positions within 3 lists to insert @req to
956 * - ldlm_grant_lock_with_skiplist
958 static void search_granted_lock(struct list_head *queue,
959 struct ldlm_lock *req,
960 struct sl_insert_point *prev)
962 struct list_head *tmp;
963 struct ldlm_lock *lock, *mode_end, *policy_end;
966 list_for_each(tmp, queue) {
967 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
969 mode_end = list_entry(lock->l_sl_mode.prev,
970 struct ldlm_lock, l_sl_mode);
972 if (lock->l_req_mode != req->l_req_mode) {
973 /* jump to last lock of mode group */
974 tmp = &mode_end->l_res_link;
978 /* suitable mode group is found */
979 if (lock->l_resource->lr_type == LDLM_PLAIN) {
980 /* insert point is last lock of the mode group */
981 prev->res_link = &mode_end->l_res_link;
982 prev->mode_link = &mode_end->l_sl_mode;
983 prev->policy_link = &req->l_sl_policy;
986 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
989 list_entry(lock->l_sl_policy.prev,
993 if (lock->l_policy_data.l_inodebits.bits ==
994 req->l_policy_data.l_inodebits.bits) {
995 /* insert point is last lock of
996 * the policy group */
998 &policy_end->l_res_link;
1000 &policy_end->l_sl_mode;
1002 &policy_end->l_sl_policy;
1007 if (policy_end == mode_end)
1008 /* done with mode group */
1011 /* go to next policy group within mode group */
1012 tmp = policy_end->l_res_link.next;
1013 lock = list_entry(tmp, struct ldlm_lock,
1015 } /* loop over policy groups within the mode group */
1017 /* insert point is last lock of the mode group,
1018 * new policy group is started */
1019 prev->res_link = &mode_end->l_res_link;
1020 prev->mode_link = &mode_end->l_sl_mode;
1021 prev->policy_link = &req->l_sl_policy;
1025 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1030 /* insert point is last lock on the queue,
1031 * new mode group and new policy group are started */
1032 prev->res_link = queue->prev;
1033 prev->mode_link = &req->l_sl_mode;
1034 prev->policy_link = &req->l_sl_policy;
1040 * Add a lock into resource granted list after a position described by
1043 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1044 struct sl_insert_point *prev)
1046 struct ldlm_resource *res = lock->l_resource;
1049 check_res_locked(res);
1051 ldlm_resource_dump(D_INFO, res);
1052 LDLM_DEBUG(lock, "About to add lock:");
1054 if (ldlm_is_destroyed(lock)) {
1055 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1059 LASSERT(list_empty(&lock->l_res_link));
1060 LASSERT(list_empty(&lock->l_sl_mode));
1061 LASSERT(list_empty(&lock->l_sl_policy));
1064 * lock->link == prev->link means lock is first starting the group.
1065 * Don't re-add to itself to suppress kernel warnings.
1067 if (&lock->l_res_link != prev->res_link)
1068 list_add(&lock->l_res_link, prev->res_link);
1069 if (&lock->l_sl_mode != prev->mode_link)
1070 list_add(&lock->l_sl_mode, prev->mode_link);
1071 if (&lock->l_sl_policy != prev->policy_link)
1072 list_add(&lock->l_sl_policy, prev->policy_link);
1078 * Add a lock to granted list on a resource maintaining skiplist
1081 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1083 struct sl_insert_point prev;
1085 LASSERT(ldlm_is_granted(lock));
1087 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1088 ldlm_granted_list_add_lock(lock, &prev);
1092 * Perform lock granting bookkeeping.
1094 * Includes putting the lock into granted list and updating lock mode.
1096 * - ldlm_lock_enqueue
1097 * - ldlm_reprocess_queue
1099 * must be called with lr_lock held
1101 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1103 struct ldlm_resource *res = lock->l_resource;
1106 check_res_locked(res);
1108 lock->l_granted_mode = lock->l_req_mode;
1110 if (work_list && lock->l_completion_ast != NULL)
1111 ldlm_add_ast_work_item(lock, NULL, work_list);
1113 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1114 ldlm_grant_lock_with_skiplist(lock);
1115 else if (res->lr_type == LDLM_EXTENT)
1116 ldlm_extent_add_lock(res, lock);
1117 else if (res->lr_type == LDLM_FLOCK) {
1118 /* We should not add locks to granted list in the following
1120 * - this is an UNLOCK but not a real lock;
1121 * - this is a TEST lock;
1122 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1123 * - this is a deadlock (flock cannot be granted) */
1124 if (lock->l_req_mode == 0 ||
1125 lock->l_req_mode == LCK_NL ||
1126 ldlm_is_test_lock(lock) ||
1127 ldlm_is_flock_deadlock(lock))
1129 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1134 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1139 * Describe the overlap between two locks. itree_overlap_cb data.
1141 struct lock_match_data {
1142 struct ldlm_lock *lmd_old;
1143 struct ldlm_lock *lmd_lock;
1144 enum ldlm_mode *lmd_mode;
1145 union ldlm_policy_data *lmd_policy;
1147 __u64 lmd_skip_flags;
1152 * Check if the given @lock meets the criteria for a match.
1153 * A reference on the lock is taken if matched.
1155 * \param lock test-against this lock
1156 * \param data parameters
1158 static int lock_matches(struct ldlm_lock *lock, struct lock_match_data *data)
1160 union ldlm_policy_data *lpol = &lock->l_policy_data;
1161 enum ldlm_mode match;
1163 if (lock == data->lmd_old)
1164 return INTERVAL_ITER_STOP;
1166 /* Check if this lock can be matched.
1167 * Used by LU-2919(exclusive open) for open lease lock */
1168 if (ldlm_is_excl(lock))
1169 return INTERVAL_ITER_CONT;
1171 /* llite sometimes wants to match locks that will be
1172 * canceled when their users drop, but we allow it to match
1173 * if it passes in CBPENDING and the lock still has users.
1174 * this is generally only going to be used by children
1175 * whose parents already hold a lock so forward progress
1176 * can still happen. */
1177 if (ldlm_is_cbpending(lock) &&
1178 !(data->lmd_flags & LDLM_FL_CBPENDING))
1179 return INTERVAL_ITER_CONT;
1180 if (!data->lmd_unref && ldlm_is_cbpending(lock) &&
1181 lock->l_readers == 0 && lock->l_writers == 0)
1182 return INTERVAL_ITER_CONT;
1184 if (!(lock->l_req_mode & *data->lmd_mode))
1185 return INTERVAL_ITER_CONT;
1186 match = lock->l_req_mode;
1188 switch (lock->l_resource->lr_type) {
1190 if (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1191 lpol->l_extent.end < data->lmd_policy->l_extent.end)
1192 return INTERVAL_ITER_CONT;
1194 if (unlikely(match == LCK_GROUP) &&
1195 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1196 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1197 return INTERVAL_ITER_CONT;
1200 /* We match if we have existing lock with same or wider set
1202 if ((lpol->l_inodebits.bits &
1203 data->lmd_policy->l_inodebits.bits) !=
1204 data->lmd_policy->l_inodebits.bits)
1205 return INTERVAL_ITER_CONT;
1211 /* We match if we have existing lock with same or wider set
1213 if (!data->lmd_unref && LDLM_HAVE_MASK(lock, GONE))
1214 return INTERVAL_ITER_CONT;
1216 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1217 return INTERVAL_ITER_CONT;
1219 /* Filter locks by skipping flags */
1220 if (data->lmd_skip_flags & lock->l_flags)
1221 return INTERVAL_ITER_CONT;
1223 if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1224 LDLM_LOCK_GET(lock);
1225 ldlm_lock_touch_in_lru(lock);
1227 ldlm_lock_addref_internal_nolock(lock, match);
1230 *data->lmd_mode = match;
1231 data->lmd_lock = lock;
1233 return INTERVAL_ITER_STOP;
1236 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1238 struct ldlm_interval *node = to_ldlm_interval(in);
1239 struct lock_match_data *data = args;
1240 struct ldlm_lock *lock;
1243 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1244 rc = lock_matches(lock, data);
1245 if (rc == INTERVAL_ITER_STOP)
1246 return INTERVAL_ITER_STOP;
1248 return INTERVAL_ITER_CONT;
1252 * Search for a lock with given parameters in interval trees.
1254 * \param res search for a lock in this resource
1255 * \param data parameters
1257 * \retval a referenced lock or NULL.
1259 static struct ldlm_lock *search_itree(struct ldlm_resource *res,
1260 struct lock_match_data *data)
1262 struct interval_node_extent ext = {
1263 .start = data->lmd_policy->l_extent.start,
1264 .end = data->lmd_policy->l_extent.end
1268 data->lmd_lock = NULL;
1270 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1271 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1273 if (tree->lit_root == NULL)
1276 if (!(tree->lit_mode & *data->lmd_mode))
1279 interval_search(tree->lit_root, &ext,
1280 itree_overlap_cb, data);
1282 return data->lmd_lock;
1290 * Search for a lock with given properties in a queue.
1292 * \param queue search for a lock in this queue
1293 * \param data parameters
1295 * \retval a referenced lock or NULL.
1297 static struct ldlm_lock *search_queue(struct list_head *queue,
1298 struct lock_match_data *data)
1300 struct ldlm_lock *lock;
1303 data->lmd_lock = NULL;
1305 list_for_each_entry(lock, queue, l_res_link) {
1306 rc = lock_matches(lock, data);
1307 if (rc == INTERVAL_ITER_STOP)
1308 return data->lmd_lock;
1314 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1316 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1317 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1318 wake_up_all(&lock->l_waitq);
1321 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1323 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1325 lock_res_and_lock(lock);
1326 ldlm_lock_fail_match_locked(lock);
1327 unlock_res_and_lock(lock);
1331 * Mark lock as "matchable" by OST.
1333 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1335 * Assumes LDLM lock is already locked.
1337 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1339 ldlm_set_lvb_ready(lock);
1340 wake_up_all(&lock->l_waitq);
1342 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1345 * Mark lock as "matchable" by OST.
1346 * Locks the lock and then \see ldlm_lock_allow_match_locked
1348 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1350 lock_res_and_lock(lock);
1351 ldlm_lock_allow_match_locked(lock);
1352 unlock_res_and_lock(lock);
1354 EXPORT_SYMBOL(ldlm_lock_allow_match);
1357 * Attempt to find a lock with specified properties.
1359 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1362 * Can be called in two ways:
1364 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1365 * for a duplicate of.
1367 * Otherwise, all of the fields must be filled in, to match against.
1369 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1370 * server (ie, connh is NULL)
1371 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1372 * list will be considered
1373 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1374 * to be canceled can still be matched as long as they still have reader
1375 * or writer refernces
1376 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1377 * just tell us if we would have matched.
1379 * \retval 1 if it finds an already-existing lock that is compatible; in this
1380 * case, lockh is filled in with a addref()ed lock
1382 * We also check security context, and if that fails we simply return 0 (to
1383 * keep caller code unchanged), the context failure will be discovered by
1384 * caller sometime later.
1386 enum ldlm_mode ldlm_lock_match_with_skip(struct ldlm_namespace *ns,
1387 __u64 flags, __u64 skip_flags,
1388 const struct ldlm_res_id *res_id,
1389 enum ldlm_type type,
1390 union ldlm_policy_data *policy,
1391 enum ldlm_mode mode,
1392 struct lustre_handle *lockh, int unref)
1394 struct lock_match_data data = {
1398 .lmd_policy = policy,
1400 .lmd_skip_flags = skip_flags,
1403 struct ldlm_resource *res;
1404 struct ldlm_lock *lock;
1410 data.lmd_old = ldlm_handle2lock(lockh);
1411 LASSERT(data.lmd_old != NULL);
1413 ns = ldlm_lock_to_ns(data.lmd_old);
1414 res_id = &data.lmd_old->l_resource->lr_name;
1415 type = data.lmd_old->l_resource->lr_type;
1416 *data.lmd_mode = data.lmd_old->l_req_mode;
1419 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1421 LASSERT(data.lmd_old == NULL);
1425 LDLM_RESOURCE_ADDREF(res);
1427 if (res->lr_type == LDLM_EXTENT)
1428 lock = search_itree(res, &data);
1430 lock = search_queue(&res->lr_granted, &data);
1431 if (!lock && !(flags & LDLM_FL_BLOCK_GRANTED))
1432 lock = search_queue(&res->lr_waiting, &data);
1433 matched = lock ? mode : 0;
1435 LDLM_RESOURCE_DELREF(res);
1436 ldlm_resource_putref(res);
1439 ldlm_lock2handle(lock, lockh);
1440 if ((flags & LDLM_FL_LVB_READY) &&
1441 (!ldlm_is_lvb_ready(lock))) {
1442 __u64 wait_flags = LDLM_FL_LVB_READY |
1443 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1444 struct l_wait_info lwi;
1446 if (lock->l_completion_ast) {
1447 int err = lock->l_completion_ast(lock,
1448 LDLM_FL_WAIT_NOREPROC,
1451 GOTO(out_fail_match, matched = 0);
1454 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1455 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1457 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1458 l_wait_event(lock->l_waitq, lock->l_flags & wait_flags,
1460 if (!ldlm_is_lvb_ready(lock))
1461 GOTO(out_fail_match, matched = 0);
1464 /* check user's security context */
1465 if (lock->l_conn_export &&
1466 sptlrpc_import_check_ctx(
1467 class_exp2cliimp(lock->l_conn_export)))
1468 GOTO(out_fail_match, matched = 0);
1470 LDLM_DEBUG(lock, "matched (%llu %llu)",
1471 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1472 res_id->name[2] : policy->l_extent.start,
1473 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1474 res_id->name[3] : policy->l_extent.end);
1477 if (flags & LDLM_FL_TEST_LOCK)
1478 LDLM_LOCK_RELEASE(lock);
1480 ldlm_lock_decref_internal(lock, mode);
1483 /* less verbose for test-only */
1484 if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
1485 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1486 "%llu/%llu (%llu %llu)", ns,
1487 type, mode, res_id->name[0], res_id->name[1],
1488 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1489 res_id->name[2] : policy->l_extent.start,
1490 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1491 res_id->name[3] : policy->l_extent.end);
1493 if (data.lmd_old != NULL)
1494 LDLM_LOCK_PUT(data.lmd_old);
1498 EXPORT_SYMBOL(ldlm_lock_match_with_skip);
1500 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1503 struct ldlm_lock *lock;
1504 enum ldlm_mode mode = 0;
1507 lock = ldlm_handle2lock(lockh);
1509 lock_res_and_lock(lock);
1510 if (LDLM_HAVE_MASK(lock, GONE))
1513 if (ldlm_is_cbpending(lock) &&
1514 lock->l_readers == 0 && lock->l_writers == 0)
1518 *bits = lock->l_policy_data.l_inodebits.bits;
1519 mode = lock->l_granted_mode;
1520 ldlm_lock_addref_internal_nolock(lock, mode);
1527 unlock_res_and_lock(lock);
1528 LDLM_LOCK_PUT(lock);
1532 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1534 /** The caller must guarantee that the buffer is large enough. */
1535 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1536 enum req_location loc, void *data, int size)
1541 LASSERT(data != NULL);
1544 switch (lock->l_lvb_type) {
1546 if (size == sizeof(struct ost_lvb)) {
1547 if (loc == RCL_CLIENT)
1548 lvb = req_capsule_client_swab_get(pill,
1550 lustre_swab_ost_lvb);
1552 lvb = req_capsule_server_swab_get(pill,
1554 lustre_swab_ost_lvb);
1555 if (unlikely(lvb == NULL)) {
1556 LDLM_ERROR(lock, "no LVB");
1560 memcpy(data, lvb, size);
1561 } else if (size == sizeof(struct ost_lvb_v1)) {
1562 struct ost_lvb *olvb = data;
1564 if (loc == RCL_CLIENT)
1565 lvb = req_capsule_client_swab_get(pill,
1567 lustre_swab_ost_lvb_v1);
1569 lvb = req_capsule_server_sized_swab_get(pill,
1571 lustre_swab_ost_lvb_v1);
1572 if (unlikely(lvb == NULL)) {
1573 LDLM_ERROR(lock, "no LVB");
1577 memcpy(data, lvb, size);
1578 olvb->lvb_mtime_ns = 0;
1579 olvb->lvb_atime_ns = 0;
1580 olvb->lvb_ctime_ns = 0;
1582 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1588 if (size == sizeof(struct lquota_lvb)) {
1589 if (loc == RCL_CLIENT)
1590 lvb = req_capsule_client_swab_get(pill,
1592 lustre_swab_lquota_lvb);
1594 lvb = req_capsule_server_swab_get(pill,
1596 lustre_swab_lquota_lvb);
1597 if (unlikely(lvb == NULL)) {
1598 LDLM_ERROR(lock, "no LVB");
1602 memcpy(data, lvb, size);
1604 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1613 if (loc == RCL_CLIENT)
1614 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1616 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1617 if (unlikely(lvb == NULL)) {
1618 LDLM_ERROR(lock, "no LVB");
1622 memcpy(data, lvb, size);
1625 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1626 libcfs_debug_dumpstack(NULL);
1634 * Create and fill in new LDLM lock with specified properties.
1635 * Returns a referenced lock
1637 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1638 const struct ldlm_res_id *res_id,
1639 enum ldlm_type type,
1640 enum ldlm_mode mode,
1641 const struct ldlm_callback_suite *cbs,
1642 void *data, __u32 lvb_len,
1643 enum lvb_type lvb_type)
1645 struct ldlm_lock *lock;
1646 struct ldlm_resource *res;
1650 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1652 RETURN(ERR_CAST(res));
1654 lock = ldlm_lock_new(res);
1656 RETURN(ERR_PTR(-ENOMEM));
1658 lock->l_req_mode = mode;
1659 lock->l_ast_data = data;
1660 lock->l_pid = current_pid();
1661 if (ns_is_server(ns))
1662 ldlm_set_ns_srv(lock);
1664 lock->l_blocking_ast = cbs->lcs_blocking;
1665 lock->l_completion_ast = cbs->lcs_completion;
1666 lock->l_glimpse_ast = cbs->lcs_glimpse;
1669 lock->l_tree_node = NULL;
1670 /* if this is the extent lock, allocate the interval tree node */
1671 if (type == LDLM_EXTENT)
1672 if (ldlm_interval_alloc(lock) == NULL)
1673 GOTO(out, rc = -ENOMEM);
1676 lock->l_lvb_len = lvb_len;
1677 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1678 if (lock->l_lvb_data == NULL)
1679 GOTO(out, rc = -ENOMEM);
1682 lock->l_lvb_type = lvb_type;
1683 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1684 GOTO(out, rc = -ENOENT);
1689 ldlm_lock_destroy(lock);
1690 LDLM_LOCK_RELEASE(lock);
1691 RETURN(ERR_PTR(rc));
1694 #ifdef HAVE_SERVER_SUPPORT
1695 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1698 struct ldlm_resource *res = lock->l_resource;
1699 enum ldlm_error rc = ELDLM_OK;
1700 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1701 ldlm_processing_policy policy;
1704 policy = ldlm_processing_policy_table[res->lr_type];
1706 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1707 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1708 res->lr_type != LDLM_FLOCK) {
1709 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1710 if (rc == -ERESTART)
1714 if (!list_empty(&rpc_list))
1715 ldlm_discard_bl_list(&rpc_list);
1722 * Enqueue (request) a lock.
1724 * Does not block. As a result of enqueue the lock would be put
1725 * into granted or waiting list.
1727 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1728 * set, skip all the enqueueing and delegate lock processing to intent policy
1731 enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
1732 struct ldlm_namespace *ns,
1733 struct ldlm_lock **lockp,
1734 void *cookie, __u64 *flags)
1736 struct ldlm_lock *lock = *lockp;
1737 struct ldlm_resource *res = lock->l_resource;
1738 int local = ns_is_client(ldlm_res_to_ns(res));
1739 enum ldlm_error rc = ELDLM_OK;
1740 struct ldlm_interval *node = NULL;
1743 /* policies are not executed on the client or during replay */
1744 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1745 && !local && ns->ns_policy) {
1746 rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
1748 if (rc == ELDLM_LOCK_REPLACED) {
1749 /* The lock that was returned has already been granted,
1750 * and placed into lockp. If it's not the same as the
1751 * one we passed in, then destroy the old one and our
1752 * work here is done. */
1753 if (lock != *lockp) {
1754 ldlm_lock_destroy(lock);
1755 LDLM_LOCK_RELEASE(lock);
1757 *flags |= LDLM_FL_LOCK_CHANGED;
1759 } else if (rc != ELDLM_OK &&
1760 ldlm_is_granted(lock)) {
1761 LASSERT(*flags & LDLM_FL_RESENT);
1762 /* It may happen that ns_policy returns an error in
1763 * resend case, object may be unlinked or just some
1764 * error occurs. It is unclear if lock reached the
1765 * client in the original reply, just leave the lock on
1766 * server, not returning it again to client. Due to
1767 * LU-6529, the server will not OOM. */
1769 } else if (rc != ELDLM_OK ||
1770 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1771 ldlm_lock_destroy(lock);
1776 if (*flags & LDLM_FL_RESENT) {
1777 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1778 * Set LOCK_CHANGED always.
1779 * Check if the lock is granted for BLOCK_GRANTED.
1780 * Take NO_TIMEOUT from the lock as it is inherited through
1781 * LDLM_FL_INHERIT_MASK */
1782 *flags |= LDLM_FL_LOCK_CHANGED;
1783 if (!ldlm_is_granted(lock))
1784 *flags |= LDLM_FL_BLOCK_GRANTED;
1785 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1789 /* For a replaying lock, it might be already in granted list. So
1790 * unlinking the lock will cause the interval node to be freed, we
1791 * have to allocate the interval node early otherwise we can't regrant
1792 * this lock in the future. - jay */
1793 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1794 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1796 lock_res_and_lock(lock);
1797 if (local && ldlm_is_granted(lock)) {
1798 /* The server returned a blocked lock, but it was granted
1799 * before we got a chance to actually enqueue it. We don't
1800 * need to do anything else. */
1801 *flags &= ~LDLM_FL_BLOCKED_MASK;
1802 GOTO(out, rc = ELDLM_OK);
1805 ldlm_resource_unlink_lock(lock);
1806 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1808 ldlm_lock_destroy_nolock(lock);
1809 GOTO(out, rc = -ENOMEM);
1812 INIT_LIST_HEAD(&node->li_group);
1813 ldlm_interval_attach(node, lock);
1817 /* Some flags from the enqueue want to make it into the AST, via the
1818 * lock's l_flags. */
1819 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1820 ldlm_set_ast_discard_data(lock);
1821 if (*flags & LDLM_FL_TEST_LOCK)
1822 ldlm_set_test_lock(lock);
1823 if (*flags & LDLM_FL_COS_INCOMPAT)
1824 ldlm_set_cos_incompat(lock);
1825 if (*flags & LDLM_FL_COS_ENABLED)
1826 ldlm_set_cos_enabled(lock);
1828 /* This distinction between local lock trees is very important; a client
1829 * namespace only has information about locks taken by that client, and
1830 * thus doesn't have enough information to decide for itself if it can
1831 * be granted (below). In this case, we do exactly what the server
1832 * tells us to do, as dictated by the 'flags'.
1834 * We do exactly the same thing during recovery, when the server is
1835 * more or less trusting the clients not to lie.
1837 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1840 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1841 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1843 ldlm_grant_lock(lock, NULL);
1844 GOTO(out, rc = ELDLM_OK);
1845 #ifdef HAVE_SERVER_SUPPORT
1846 } else if (*flags & LDLM_FL_REPLAY) {
1847 if (*flags & LDLM_FL_BLOCK_WAIT) {
1848 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1849 GOTO(out, rc = ELDLM_OK);
1850 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1851 ldlm_grant_lock(lock, NULL);
1852 GOTO(out, rc = ELDLM_OK);
1854 /* If no flags, fall through to normal enqueue path. */
1857 rc = ldlm_lock_enqueue_helper(lock, flags);
1861 CERROR("This is client-side-only module, cannot handle "
1862 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1868 unlock_res_and_lock(lock);
1870 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1874 #ifdef HAVE_SERVER_SUPPORT
1876 * Iterate through all waiting locks on a given resource queue and attempt to
1879 * Must be called with resource lock held.
1881 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1882 struct list_head *work_list,
1883 enum ldlm_process_intention intention)
1885 struct list_head *tmp, *pos;
1886 ldlm_processing_policy policy;
1888 int rc = LDLM_ITER_CONTINUE;
1889 enum ldlm_error err;
1890 struct list_head bl_ast_list = LIST_HEAD_INIT(bl_ast_list);
1893 check_res_locked(res);
1895 policy = ldlm_processing_policy_table[res->lr_type];
1897 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1898 intention == LDLM_PROCESS_RECOVERY);
1901 list_for_each_safe(tmp, pos, queue) {
1902 struct ldlm_lock *pending;
1903 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1905 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1907 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1910 rc = policy(pending, &flags, intention, &err, &rpc_list);
1911 if (pending->l_granted_mode == pending->l_req_mode ||
1912 res->lr_type == LDLM_FLOCK) {
1913 list_splice(&rpc_list, work_list);
1915 list_splice(&rpc_list, &bl_ast_list);
1918 * When this is called from recovery done, we always want
1919 * to scan the whole list no matter what 'rc' is returned.
1921 if (rc != LDLM_ITER_CONTINUE &&
1922 intention == LDLM_PROCESS_RESCAN)
1926 if (!list_empty(&bl_ast_list)) {
1929 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
1933 if (rc == -ERESTART)
1937 if (!list_empty(&bl_ast_list))
1938 ldlm_discard_bl_list(&bl_ast_list);
1940 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
1944 * Conflicting locks are detected for a lock to be enqueued, add the lock
1945 * into waiting list and send blocking ASTs to the conflicting locks.
1947 * \param[in] lock The lock to be enqueued.
1948 * \param[out] flags Lock flags for the lock to be enqueued.
1949 * \param[in] rpc_list Conflicting locks list.
1951 * \retval -ERESTART: Some lock was instantly canceled while sending
1952 * blocking ASTs, caller needs to re-check conflicting
1954 * \retval -EAGAIN: Lock was destroyed, caller should return error.
1955 * \reval 0: Lock is successfully added in waiting list.
1957 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
1958 struct list_head *rpc_list)
1960 struct ldlm_resource *res = lock->l_resource;
1964 check_res_locked(res);
1966 /* If either of the compat_queue()s returned failure, then we
1967 * have ASTs to send and must go onto the waiting list.
1969 * bug 2322: we used to unlink and re-add here, which was a
1970 * terrible folly -- if we goto restart, we could get
1971 * re-ordered! Causes deadlock, because ASTs aren't sent! */
1972 if (list_empty(&lock->l_res_link))
1973 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1976 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
1979 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
1980 !ns_is_client(ldlm_res_to_ns(res)))
1981 class_fail_export(lock->l_export);
1984 if (rc == -ERESTART) {
1985 /* 15715: The lock was granted and destroyed after
1986 * resource lock was dropped. Interval node was freed
1987 * in ldlm_lock_destroy. Anyway, this always happens
1988 * when a client is being evicted. So it would be
1989 * ok to return an error. -jay */
1990 if (ldlm_is_destroyed(lock))
1993 /* lock was granted while resource was unlocked. */
1994 if (ldlm_is_granted(lock)) {
1995 /* bug 11300: if the lock has been granted,
1996 * break earlier because otherwise, we will go
1997 * to restart and ldlm_resource_unlink will be
1998 * called and it causes the interval node to be
1999 * freed. Then we will fail at
2000 * ldlm_extent_add_lock() */
2001 *flags &= ~LDLM_FL_BLOCKED_MASK;
2007 *flags |= LDLM_FL_BLOCK_GRANTED;
2013 * Discard all AST work items from list.
2015 * If for whatever reason we do not want to send ASTs to conflicting locks
2016 * anymore, disassemble the list with this function.
2018 void ldlm_discard_bl_list(struct list_head *bl_list)
2020 struct ldlm_lock *lock, *tmp;
2024 list_for_each_entry_safe(lock, tmp, bl_list, l_bl_ast) {
2025 LASSERT(!list_empty(&lock->l_bl_ast));
2026 list_del_init(&lock->l_bl_ast);
2027 ldlm_clear_ast_sent(lock);
2028 LASSERT(lock->l_bl_ast_run == 0);
2029 ldlm_clear_blocking_lock(lock);
2030 LDLM_LOCK_RELEASE(lock);
2036 * Process a call to blocking AST callback for a lock in ast_work list
2039 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2041 struct ldlm_cb_set_arg *arg = opaq;
2042 struct ldlm_lock *lock;
2043 struct ldlm_lock_desc d;
2044 struct ldlm_bl_desc bld;
2049 if (list_empty(arg->list))
2052 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2054 /* nobody should touch l_bl_ast but some locks in the list may become
2055 * granted after lock convert or COS downgrade, these locks should be
2056 * just skipped here and removed from the list.
2058 lock_res_and_lock(lock);
2059 list_del_init(&lock->l_bl_ast);
2061 /* lock is not blocking lock anymore, but was kept in the list because
2062 * it can managed only here.
2064 if (!ldlm_is_ast_sent(lock)) {
2065 unlock_res_and_lock(lock);
2066 LDLM_LOCK_RELEASE(lock);
2070 LASSERT(lock->l_blocking_lock);
2071 ldlm_lock2desc(lock->l_blocking_lock, &d);
2072 /* copy blocking lock ibits in cancel_bits as well,
2073 * new client may use them for lock convert and it is
2074 * important to use new field to convert locks from
2077 d.l_policy_data.l_inodebits.cancel_bits =
2078 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2080 /* Blocking lock is being destroyed here but some information about it
2081 * may be needed inside l_blocking_ast() function below,
2082 * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
2084 bld.bl_same_client = lock->l_client_cookie ==
2085 lock->l_blocking_lock->l_client_cookie;
2086 bld.bl_cos_incompat = ldlm_is_cos_incompat(lock->l_blocking_lock);
2087 arg->bl_desc = &bld;
2089 LASSERT(ldlm_is_ast_sent(lock));
2090 LASSERT(lock->l_bl_ast_run == 0);
2091 lock->l_bl_ast_run++;
2092 ldlm_clear_blocking_lock(lock);
2093 unlock_res_and_lock(lock);
2095 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2097 LDLM_LOCK_RELEASE(lock);
2103 * Process a call to revocation AST callback for a lock in ast_work list
2106 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2108 struct ldlm_cb_set_arg *arg = opaq;
2109 struct ldlm_lock_desc desc;
2111 struct ldlm_lock *lock;
2114 if (list_empty(arg->list))
2117 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2118 list_del_init(&lock->l_rk_ast);
2120 /* the desc just pretend to exclusive */
2121 ldlm_lock2desc(lock, &desc);
2122 desc.l_req_mode = LCK_EX;
2123 desc.l_granted_mode = 0;
2125 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2126 LDLM_LOCK_RELEASE(lock);
2132 * Process a call to glimpse AST callback for a lock in ast_work list
2134 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2136 struct ldlm_cb_set_arg *arg = opaq;
2137 struct ldlm_glimpse_work *gl_work;
2138 struct ldlm_lock *lock;
2142 if (list_empty(arg->list))
2145 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2147 list_del_init(&gl_work->gl_list);
2149 lock = gl_work->gl_lock;
2151 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2152 arg->gl_desc = gl_work->gl_desc;
2153 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2154 arg->gl_interpret_data = gl_work->gl_interpret_data;
2156 /* invoke the actual glimpse callback */
2157 if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
2160 LDLM_LOCK_RELEASE(lock);
2161 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2162 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2164 OBD_FREE_PTR(gl_work);
2171 * Process a call to completion AST callback for a lock in ast_work list
2174 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2176 struct ldlm_cb_set_arg *arg = opaq;
2177 struct ldlm_lock *lock;
2178 ldlm_completion_callback completion_callback;
2183 if (list_empty(arg->list))
2186 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2188 /* It's possible to receive a completion AST before we've set
2189 * the l_completion_ast pointer: either because the AST arrived
2190 * before the reply, or simply because there's a small race
2191 * window between receiving the reply and finishing the local
2192 * enqueue. (bug 842)
2194 * This can't happen with the blocking_ast, however, because we
2195 * will never call the local blocking_ast until we drop our
2196 * reader/writer reference, which we won't do until we get the
2197 * reply and finish enqueueing. */
2199 /* nobody should touch l_cp_ast */
2200 lock_res_and_lock(lock);
2201 list_del_init(&lock->l_cp_ast);
2202 LASSERT(ldlm_is_cp_reqd(lock));
2203 /* save l_completion_ast since it can be changed by
2204 * mds_intent_policy(), see bug 14225 */
2205 completion_callback = lock->l_completion_ast;
2206 ldlm_clear_cp_reqd(lock);
2207 unlock_res_and_lock(lock);
2209 if (completion_callback != NULL)
2210 rc = completion_callback(lock, 0, (void *)arg);
2211 LDLM_LOCK_RELEASE(lock);
2217 * Process list of locks in need of ASTs being sent.
2219 * Used on server to send multiple ASTs together instead of sending one by
2222 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2223 ldlm_desc_ast_t ast_type)
2225 struct ldlm_cb_set_arg *arg;
2226 set_producer_func work_ast_lock;
2229 if (list_empty(rpc_list))
2236 atomic_set(&arg->restart, 0);
2237 arg->list = rpc_list;
2240 case LDLM_WORK_CP_AST:
2241 arg->type = LDLM_CP_CALLBACK;
2242 work_ast_lock = ldlm_work_cp_ast_lock;
2244 #ifdef HAVE_SERVER_SUPPORT
2245 case LDLM_WORK_BL_AST:
2246 arg->type = LDLM_BL_CALLBACK;
2247 work_ast_lock = ldlm_work_bl_ast_lock;
2249 case LDLM_WORK_REVOKE_AST:
2250 arg->type = LDLM_BL_CALLBACK;
2251 work_ast_lock = ldlm_work_revoke_ast_lock;
2253 case LDLM_WORK_GL_AST:
2254 arg->type = LDLM_GL_CALLBACK;
2255 work_ast_lock = ldlm_work_gl_ast_lock;
2262 /* We create a ptlrpc request set with flow control extension.
2263 * This request set will use the work_ast_lock function to produce new
2264 * requests and will send a new request each time one completes in order
2265 * to keep the number of requests in flight to ns_max_parallel_ast */
2266 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2267 work_ast_lock, arg);
2268 if (arg->set == NULL)
2269 GOTO(out, rc = -ENOMEM);
2271 ptlrpc_set_wait(NULL, arg->set);
2272 ptlrpc_set_destroy(arg->set);
2274 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2282 * Try to grant all waiting locks on a resource.
2284 * Calls ldlm_reprocess_queue on waiting queue.
2286 * Typically called after some resource locks are cancelled to see
2287 * if anything could be granted as a result of the cancellation.
2289 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2290 enum ldlm_process_intention intention)
2292 struct list_head rpc_list;
2293 #ifdef HAVE_SERVER_SUPPORT
2294 struct obd_device *obd;
2298 INIT_LIST_HEAD(&rpc_list);
2299 /* Local lock trees don't get reprocessed. */
2300 if (ns_is_client(ldlm_res_to_ns(res))) {
2305 /* Disable reprocess during lock replay stage but allow during
2306 * request replay stage.
2308 obd = ldlm_res_to_ns(res)->ns_obd;
2309 if (obd->obd_recovering &&
2310 atomic_read(&obd->obd_req_replay_clients) == 0)
2314 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list, intention);
2317 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2319 if (rc == -ERESTART) {
2320 LASSERT(list_empty(&rpc_list));
2326 INIT_LIST_HEAD(&rpc_list);
2327 if (!ns_is_client(ldlm_res_to_ns(res))) {
2328 CERROR("This is client-side-only module, cannot handle "
2329 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2336 void ldlm_reprocess_all(struct ldlm_resource *res)
2338 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN);
2340 EXPORT_SYMBOL(ldlm_reprocess_all);
2342 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2343 struct hlist_node *hnode, void *arg)
2345 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2347 /* This is only called once after recovery done. LU-8306. */
2348 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY);
2353 * Iterate through all resources on a namespace attempting to grant waiting
2356 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2361 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2362 ldlm_reprocess_res, NULL, 0);
2368 * Helper function to call blocking AST for LDLM lock \a lock in a
2369 * "cancelling" mode.
2371 void ldlm_cancel_callback(struct ldlm_lock *lock)
2373 check_res_locked(lock->l_resource);
2374 if (!ldlm_is_cancel(lock)) {
2375 ldlm_set_cancel(lock);
2376 if (lock->l_blocking_ast) {
2377 unlock_res_and_lock(lock);
2378 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2380 lock_res_and_lock(lock);
2382 LDLM_DEBUG(lock, "no blocking ast");
2385 /* only canceller can set bl_done bit */
2386 ldlm_set_bl_done(lock);
2387 wake_up_all(&lock->l_waitq);
2388 } else if (!ldlm_is_bl_done(lock)) {
2389 struct l_wait_info lwi = { 0 };
2391 /* The lock is guaranteed to have been canceled once
2392 * returning from this function. */
2393 unlock_res_and_lock(lock);
2394 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
2395 lock_res_and_lock(lock);
2400 * Remove skiplist-enabled LDLM lock \a req from granted list
2402 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2404 if (req->l_resource->lr_type != LDLM_PLAIN &&
2405 req->l_resource->lr_type != LDLM_IBITS)
2408 list_del_init(&req->l_sl_policy);
2409 list_del_init(&req->l_sl_mode);
2413 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2415 void ldlm_lock_cancel(struct ldlm_lock *lock)
2417 struct ldlm_resource *res;
2418 struct ldlm_namespace *ns;
2421 lock_res_and_lock(lock);
2423 res = lock->l_resource;
2424 ns = ldlm_res_to_ns(res);
2426 /* Please do not, no matter how tempting, remove this LBUG without
2427 * talking to me first. -phik */
2428 if (lock->l_readers || lock->l_writers) {
2429 LDLM_ERROR(lock, "lock still has references");
2430 unlock_res_and_lock(lock);
2434 if (ldlm_is_waited(lock))
2435 ldlm_del_waiting_lock(lock);
2437 /* Releases cancel callback. */
2438 ldlm_cancel_callback(lock);
2440 /* Yes, second time, just in case it was added again while we were
2441 * running with no res lock in ldlm_cancel_callback */
2442 if (ldlm_is_waited(lock))
2443 ldlm_del_waiting_lock(lock);
2445 ldlm_resource_unlink_lock(lock);
2446 ldlm_lock_destroy_nolock(lock);
2448 if (ldlm_is_granted(lock))
2449 ldlm_pool_del(&ns->ns_pool, lock);
2451 /* Make sure we will not be called again for same lock what is possible
2452 * if not to zero out lock->l_granted_mode */
2453 lock->l_granted_mode = LCK_MINMODE;
2454 unlock_res_and_lock(lock);
2458 EXPORT_SYMBOL(ldlm_lock_cancel);
2461 * Set opaque data into the lock that only makes sense to upper layer.
2463 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2465 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2470 if (lock->l_ast_data == NULL)
2471 lock->l_ast_data = data;
2472 if (lock->l_ast_data == data)
2474 LDLM_LOCK_PUT(lock);
2478 EXPORT_SYMBOL(ldlm_lock_set_data);
2480 struct export_cl_data {
2481 const struct lu_env *ecl_env;
2482 struct obd_export *ecl_exp;
2486 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2487 struct ldlm_lock *lock,
2488 struct export_cl_data *ecl)
2490 struct ldlm_resource *res;
2492 res = ldlm_resource_getref(lock->l_resource);
2494 ldlm_lvbo_update(res, lock, NULL, 1);
2495 ldlm_lock_cancel(lock);
2496 if (!exp->exp_obd->obd_stopping)
2497 ldlm_reprocess_all(res);
2498 ldlm_resource_putref(res);
2501 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2502 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2503 exp, ecl->ecl_loop);
2508 * Iterator function for ldlm_export_cancel_locks.
2509 * Cancels passed locks.
2512 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2513 struct hlist_node *hnode, void *data)
2516 struct export_cl_data *ecl = (struct export_cl_data *)data;
2517 struct obd_export *exp = ecl->ecl_exp;
2518 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2520 LDLM_LOCK_GET(lock);
2521 ldlm_cancel_lock_for_export(exp, lock, ecl);
2522 LDLM_LOCK_RELEASE(lock);
2528 * Cancel all blocked locks for given export.
2530 * Typically called on client disconnection/eviction
2532 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2535 struct export_cl_data ecl = {
2541 rc = lu_env_init(&env, LCT_DT_THREAD);
2546 while (!list_empty(&exp->exp_bl_list)) {
2547 struct ldlm_lock *lock;
2549 spin_lock_bh(&exp->exp_bl_list_lock);
2550 if (!list_empty(&exp->exp_bl_list)) {
2551 lock = list_entry(exp->exp_bl_list.next,
2552 struct ldlm_lock, l_exp_list);
2553 LDLM_LOCK_GET(lock);
2554 list_del_init(&lock->l_exp_list);
2558 spin_unlock_bh(&exp->exp_bl_list_lock);
2563 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2564 LDLM_LOCK_RELEASE(lock);
2569 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2570 "left on hash table %d.\n", exp, ecl.ecl_loop,
2571 atomic_read(&exp->exp_lock_hash->hs_count));
2573 return ecl.ecl_loop;
2577 * Cancel all locks for given export.
2579 * Typically called after client disconnection/eviction
2581 int ldlm_export_cancel_locks(struct obd_export *exp)
2583 struct export_cl_data ecl;
2587 rc = lu_env_init(&env, LCT_DT_THREAD);
2594 cfs_hash_for_each_empty(exp->exp_lock_hash,
2595 ldlm_cancel_locks_for_export_cb, &ecl);
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 if (ecl.ecl_loop > 0 &&
2602 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2603 exp->exp_obd->obd_stopping)
2604 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2608 return ecl.ecl_loop;
2612 * Downgrade an PW/EX lock to COS mode.
2614 * A lock mode convertion from PW/EX mode to less conflict mode. The
2615 * convertion may fail if lock was canceled before downgrade, but it doesn't
2616 * indicate any problem, because such lock has no reader or writer, and will
2619 * Used by Commit on Sharing (COS) code to force object changes commit in case
2620 * of conflict. Converted lock is considered as new lock and all blocking AST
2621 * things are cleared, so any pending or new blocked lock on that lock will
2622 * cause new call to blocking_ast and force resource object commit.
2624 * \param lock A lock to convert
2625 * \param new_mode new lock mode
2627 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2629 #ifdef HAVE_SERVER_SUPPORT
2632 LASSERT(new_mode == LCK_COS);
2634 lock_res_and_lock(lock);
2636 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2637 unlock_res_and_lock(lock);
2639 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2640 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2644 ldlm_resource_unlink_lock(lock);
2646 * Remove the lock from pool as it will be added again in
2647 * ldlm_grant_lock() called below.
2649 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2651 /* Consider downgraded lock as a new lock and clear all states
2652 * related to a previous blocking AST processing.
2654 ldlm_clear_blocking_data(lock);
2656 lock->l_req_mode = new_mode;
2657 ldlm_grant_lock(lock, NULL);
2658 unlock_res_and_lock(lock);
2660 ldlm_reprocess_all(lock->l_resource);
2665 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2668 * Print lock with lock handle \a lockh description into debug log.
2670 * Used when printing all locks on a resource for debug purposes.
2672 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2674 struct ldlm_lock *lock;
2676 if (!((libcfs_debug | D_ERROR) & level))
2679 lock = ldlm_handle2lock(lockh);
2683 LDLM_DEBUG_LIMIT(level, lock, "###");
2685 LDLM_LOCK_PUT(lock);
2687 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2690 * Print lock information with custom message into debug log.
2693 void _ldlm_lock_debug(struct ldlm_lock *lock,
2694 struct libcfs_debug_msg_data *msgdata,
2695 const char *fmt, ...)
2698 struct obd_export *exp = lock->l_export;
2699 struct ldlm_resource *resource = NULL;
2700 struct va_format vaf;
2701 char *nid = "local";
2703 /* on server-side resource of lock doesn't change */
2704 if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2705 if (lock->l_resource != NULL)
2706 resource = ldlm_resource_getref(lock->l_resource);
2707 } else if (spin_trylock(&lock->l_lock)) {
2708 if (lock->l_resource != NULL)
2709 resource = ldlm_resource_getref(lock->l_resource);
2710 spin_unlock(&lock->l_lock);
2713 va_start(args, fmt);
2717 if (exp && exp->exp_connection) {
2718 nid = obd_export_nid2str(exp);
2719 } else if (exp && exp->exp_obd != NULL) {
2720 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2721 nid = obd_import_nid2str(imp);
2724 if (resource == NULL) {
2725 libcfs_debug_msg(msgdata,
2726 "%pV ns: \?\? lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: \?\? rrc=\?\? type: \?\?\? flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2729 lock->l_handle.h_cookie,
2730 atomic_read(&lock->l_refc),
2731 lock->l_readers, lock->l_writers,
2732 ldlm_lockname[lock->l_granted_mode],
2733 ldlm_lockname[lock->l_req_mode],
2735 lock->l_remote_handle.cookie,
2736 exp ? atomic_read(&exp->exp_refcount) : -99,
2737 lock->l_pid, lock->l_callback_timeout,
2743 switch (resource->lr_type) {
2745 libcfs_debug_msg(msgdata,
2746 "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s [%llu->%llu] (req %llu->%llu) flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2748 ldlm_lock_to_ns_name(lock), lock,
2749 lock->l_handle.h_cookie,
2750 atomic_read(&lock->l_refc),
2751 lock->l_readers, lock->l_writers,
2752 ldlm_lockname[lock->l_granted_mode],
2753 ldlm_lockname[lock->l_req_mode],
2755 atomic_read(&resource->lr_refcount),
2756 ldlm_typename[resource->lr_type],
2757 lock->l_policy_data.l_extent.start,
2758 lock->l_policy_data.l_extent.end,
2759 lock->l_req_extent.start, lock->l_req_extent.end,
2761 lock->l_remote_handle.cookie,
2762 exp ? atomic_read(&exp->exp_refcount) : -99,
2763 lock->l_pid, lock->l_callback_timeout,
2768 libcfs_debug_msg(msgdata,
2769 "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s pid: %d [%llu->%llu] flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld\n",
2771 ldlm_lock_to_ns_name(lock), lock,
2772 lock->l_handle.h_cookie,
2773 atomic_read(&lock->l_refc),
2774 lock->l_readers, lock->l_writers,
2775 ldlm_lockname[lock->l_granted_mode],
2776 ldlm_lockname[lock->l_req_mode],
2778 atomic_read(&resource->lr_refcount),
2779 ldlm_typename[resource->lr_type],
2780 lock->l_policy_data.l_flock.pid,
2781 lock->l_policy_data.l_flock.start,
2782 lock->l_policy_data.l_flock.end,
2784 lock->l_remote_handle.cookie,
2785 exp ? atomic_read(&exp->exp_refcount) : -99,
2786 lock->l_pid, lock->l_callback_timeout);
2790 libcfs_debug_msg(msgdata,
2791 "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " bits %#llx/%#llx rrc: %d type: %s flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2793 ldlm_lock_to_ns_name(lock),
2794 lock, lock->l_handle.h_cookie,
2795 atomic_read(&lock->l_refc),
2796 lock->l_readers, lock->l_writers,
2797 ldlm_lockname[lock->l_granted_mode],
2798 ldlm_lockname[lock->l_req_mode],
2800 lock->l_policy_data.l_inodebits.bits,
2801 lock->l_policy_data.l_inodebits.try_bits,
2802 atomic_read(&resource->lr_refcount),
2803 ldlm_typename[resource->lr_type],
2805 lock->l_remote_handle.cookie,
2806 exp ? atomic_read(&exp->exp_refcount) : -99,
2807 lock->l_pid, lock->l_callback_timeout,
2812 libcfs_debug_msg(msgdata,
2813 "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2815 ldlm_lock_to_ns_name(lock),
2816 lock, lock->l_handle.h_cookie,
2817 atomic_read(&lock->l_refc),
2818 lock->l_readers, lock->l_writers,
2819 ldlm_lockname[lock->l_granted_mode],
2820 ldlm_lockname[lock->l_req_mode],
2822 atomic_read(&resource->lr_refcount),
2823 ldlm_typename[resource->lr_type],
2825 lock->l_remote_handle.cookie,
2826 exp ? atomic_read(&exp->exp_refcount) : -99,
2827 lock->l_pid, lock->l_callback_timeout,
2832 ldlm_resource_putref(resource);
2834 EXPORT_SYMBOL(_ldlm_lock_debug);