4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_lock.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <libcfs/libcfs.h>
42 #include <lustre_swab.h>
43 #include <obd_class.h>
45 #include "ldlm_internal.h"
47 struct kmem_cache *ldlm_glimpse_work_kmem;
48 EXPORT_SYMBOL(ldlm_glimpse_work_kmem);
51 char *ldlm_lockname[] = {
59 [LCK_GROUP] = "GROUP",
62 EXPORT_SYMBOL(ldlm_lockname);
64 char *ldlm_typename[] = {
66 [LDLM_EXTENT] = "EXT",
71 static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = {
72 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_wire_to_local,
73 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_wire_to_local,
74 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_wire_to_local,
75 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_wire_to_local,
78 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
79 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_local_to_wire,
80 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_local_to_wire,
81 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_local_to_wire,
82 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_local_to_wire,
86 * Converts lock policy from local format to on the wire lock_desc format
88 void ldlm_convert_policy_to_wire(enum ldlm_type type,
89 const union ldlm_policy_data *lpolicy,
90 union ldlm_wire_policy_data *wpolicy)
92 ldlm_policy_local_to_wire_t convert;
94 convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
96 convert(lpolicy, wpolicy);
100 * Converts lock policy from on the wire lock_desc format to local format
102 void ldlm_convert_policy_to_local(struct obd_export *exp, enum ldlm_type type,
103 const union ldlm_wire_policy_data *wpolicy,
104 union ldlm_policy_data *lpolicy)
106 ldlm_policy_wire_to_local_t convert;
108 convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE];
110 convert(wpolicy, lpolicy);
113 const char *ldlm_it2str(enum ldlm_intent_flags it)
120 case (IT_OPEN | IT_CREAT):
133 CERROR("Unknown intent 0x%08x\n", it);
137 EXPORT_SYMBOL(ldlm_it2str);
139 extern struct kmem_cache *ldlm_lock_slab;
141 #ifdef HAVE_SERVER_SUPPORT
142 static ldlm_processing_policy ldlm_processing_policy_table[] = {
143 [LDLM_PLAIN] = ldlm_process_plain_lock,
144 [LDLM_EXTENT] = ldlm_process_extent_lock,
145 [LDLM_FLOCK] = ldlm_process_flock_lock,
146 [LDLM_IBITS] = ldlm_process_inodebits_lock,
149 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
151 return ldlm_processing_policy_table[res->lr_type];
153 EXPORT_SYMBOL(ldlm_get_processing_policy);
154 #endif /* HAVE_SERVER_SUPPORT */
156 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
160 EXPORT_SYMBOL(ldlm_register_intent);
163 * REFCOUNTED LOCK OBJECTS
168 * Get a reference on a lock.
170 * Lock refcounts, during creation:
171 * - one special one for allocation, dec'd only once in destroy
172 * - one for being a lock that's in-use
173 * - one for the addref associated with a new lock
175 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
177 atomic_inc(&lock->l_refc);
180 EXPORT_SYMBOL(ldlm_lock_get);
183 * Release lock reference.
185 * Also frees the lock if it was last reference.
187 void ldlm_lock_put(struct ldlm_lock *lock)
191 LASSERT(lock->l_resource != LP_POISON);
192 LASSERT(atomic_read(&lock->l_refc) > 0);
193 if (atomic_dec_and_test(&lock->l_refc)) {
194 struct ldlm_resource *res;
197 "final lock_put on destroyed lock, freeing it.");
199 res = lock->l_resource;
200 LASSERT(ldlm_is_destroyed(lock));
201 LASSERT(list_empty(&lock->l_exp_list));
202 LASSERT(list_empty(&lock->l_res_link));
203 LASSERT(list_empty(&lock->l_pending_chain));
205 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
207 lu_ref_del(&res->lr_reference, "lock", lock);
208 ldlm_resource_putref(res);
209 lock->l_resource = NULL;
210 if (lock->l_export) {
211 class_export_lock_put(lock->l_export, lock);
212 lock->l_export = NULL;
215 if (lock->l_lvb_data != NULL)
216 OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
218 ldlm_interval_free(ldlm_interval_detach(lock));
219 lu_ref_fini(&lock->l_reference);
220 OBD_FREE_RCU(lock, sizeof(*lock), &lock->l_handle);
225 EXPORT_SYMBOL(ldlm_lock_put);
228 * Removes LDLM lock \a lock from LRU. Assumes LRU is already locked.
230 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
233 if (!list_empty(&lock->l_lru)) {
234 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
236 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
237 if (ns->ns_last_pos == &lock->l_lru)
238 ns->ns_last_pos = lock->l_lru.prev;
239 list_del_init(&lock->l_lru);
240 LASSERT(ns->ns_nr_unused > 0);
248 * Removes LDLM lock \a lock from LRU. Obtains the LRU lock first.
250 * If \a last_use is non-zero, it will remove the lock from LRU only if
251 * it matches lock's l_last_used.
253 * \retval 0 if \a last_use is set, the lock is not in LRU list or \a last_use
254 * doesn't match lock's l_last_used;
255 * otherwise, the lock hasn't been in the LRU list.
256 * \retval 1 the lock was in LRU list and removed.
258 int ldlm_lock_remove_from_lru_check(struct ldlm_lock *lock, ktime_t last_use)
260 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
264 if (ldlm_is_ns_srv(lock)) {
265 LASSERT(list_empty(&lock->l_lru));
269 spin_lock(&ns->ns_lock);
270 if (!ktime_compare(last_use, ktime_set(0, 0)) ||
271 !ktime_compare(last_use, lock->l_last_used))
272 rc = ldlm_lock_remove_from_lru_nolock(lock);
273 spin_unlock(&ns->ns_lock);
279 * Adds LDLM lock \a lock to namespace LRU. Assumes LRU is already locked.
281 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
283 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
285 lock->l_last_used = ktime_get();
286 LASSERT(list_empty(&lock->l_lru));
287 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
288 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
289 LASSERT(ns->ns_nr_unused >= 0);
294 * Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks
297 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
299 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
302 spin_lock(&ns->ns_lock);
303 ldlm_lock_add_to_lru_nolock(lock);
304 spin_unlock(&ns->ns_lock);
309 * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
310 * the LRU. Performs necessary LRU locking
312 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
314 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
317 if (ldlm_is_ns_srv(lock)) {
318 LASSERT(list_empty(&lock->l_lru));
323 spin_lock(&ns->ns_lock);
324 if (!list_empty(&lock->l_lru)) {
325 ldlm_lock_remove_from_lru_nolock(lock);
326 ldlm_lock_add_to_lru_nolock(lock);
328 spin_unlock(&ns->ns_lock);
333 * Helper to destroy a locked lock.
335 * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
336 * Must be called with l_lock and lr_lock held.
338 * Does not actually free the lock data, but rather marks the lock as
339 * destroyed by setting l_destroyed field in the lock to 1. Destroys a
340 * handle->lock association too, so that the lock can no longer be found
341 * and removes the lock from LRU list. Actual lock freeing occurs when
342 * last lock reference goes away.
344 * Original comment (of some historical value):
345 * This used to have a 'strict' flag, which recovery would use to mark an
346 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
347 * shall explain why it's gone: with the new hash table scheme, once you call
348 * ldlm_lock_destroy, you can never drop your final references on this lock.
349 * Because it's not in the hash table anymore. -phil
351 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
355 if (lock->l_readers || lock->l_writers) {
356 LDLM_ERROR(lock, "lock still has references");
360 if (!list_empty(&lock->l_res_link)) {
361 LDLM_ERROR(lock, "lock still on resource");
365 if (ldlm_is_destroyed(lock)) {
366 LASSERT(list_empty(&lock->l_lru));
370 ldlm_set_destroyed(lock);
372 if (lock->l_export && lock->l_export->exp_lock_hash) {
373 /* NB: it's safe to call cfs_hash_del() even lock isn't
374 * in exp_lock_hash. */
375 /* In the function below, .hs_keycmp resolves to
376 * ldlm_export_lock_keycmp() */
377 /* coverity[overrun-buffer-val] */
378 cfs_hash_del(lock->l_export->exp_lock_hash,
379 &lock->l_remote_handle, &lock->l_exp_hash);
382 ldlm_lock_remove_from_lru(lock);
383 class_handle_unhash(&lock->l_handle);
390 * Destroys a LDLM lock \a lock. Performs necessary locking first.
392 void ldlm_lock_destroy(struct ldlm_lock *lock)
396 lock_res_and_lock(lock);
397 first = ldlm_lock_destroy_internal(lock);
398 unlock_res_and_lock(lock);
400 /* drop reference from hashtable only for first destroy */
402 lu_ref_del(&lock->l_reference, "hash", lock);
403 LDLM_LOCK_RELEASE(lock);
409 * Destroys a LDLM lock \a lock that is already locked.
411 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
415 first = ldlm_lock_destroy_internal(lock);
416 /* drop reference from hashtable only for first destroy */
418 lu_ref_del(&lock->l_reference, "hash", lock);
419 LDLM_LOCK_RELEASE(lock);
424 /* this is called by portals_handle2object with the handle lock taken */
425 static void lock_handle_addref(void *lock)
427 LDLM_LOCK_GET((struct ldlm_lock *)lock);
430 static void lock_handle_free(void *lock, int size)
432 LASSERT(size == sizeof(struct ldlm_lock));
433 OBD_SLAB_FREE(lock, ldlm_lock_slab, size);
436 static struct portals_handle_ops lock_handle_ops = {
437 .hop_addref = lock_handle_addref,
438 .hop_free = lock_handle_free,
443 * Allocate and initialize new lock structure.
445 * usage: pass in a resource on which you have done ldlm_resource_get
446 * new lock will take over the refcount.
447 * returns: lock with refcount 2 - one for current caller and one for remote
449 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
451 struct ldlm_lock *lock;
454 if (resource == NULL)
457 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, GFP_NOFS);
461 spin_lock_init(&lock->l_lock);
462 lock->l_resource = resource;
463 lu_ref_add(&resource->lr_reference, "lock", lock);
465 atomic_set(&lock->l_refc, 2);
466 INIT_LIST_HEAD(&lock->l_res_link);
467 INIT_LIST_HEAD(&lock->l_lru);
468 INIT_LIST_HEAD(&lock->l_pending_chain);
469 INIT_LIST_HEAD(&lock->l_bl_ast);
470 INIT_LIST_HEAD(&lock->l_cp_ast);
471 INIT_LIST_HEAD(&lock->l_rk_ast);
472 init_waitqueue_head(&lock->l_waitq);
473 lock->l_blocking_lock = NULL;
474 INIT_LIST_HEAD(&lock->l_sl_mode);
475 INIT_LIST_HEAD(&lock->l_sl_policy);
476 INIT_HLIST_NODE(&lock->l_exp_hash);
477 INIT_HLIST_NODE(&lock->l_exp_flock_hash);
479 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
481 INIT_LIST_HEAD_RCU(&lock->l_handle.h_link);
482 class_handle_hash(&lock->l_handle, &lock_handle_ops);
484 lu_ref_init(&lock->l_reference);
485 lu_ref_add(&lock->l_reference, "hash", lock);
486 lock->l_callback_timeout = 0;
487 lock->l_activity = 0;
489 #if LUSTRE_TRACKS_LOCK_EXP_REFS
490 INIT_LIST_HEAD(&lock->l_exp_refs_link);
491 lock->l_exp_refs_nr = 0;
492 lock->l_exp_refs_target = NULL;
494 INIT_LIST_HEAD(&lock->l_exp_list);
500 * Moves LDLM lock \a lock to another resource.
501 * This is used on client when server returns some other lock than requested
502 * (typically as a result of intent operation)
504 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
505 const struct ldlm_res_id *new_resid)
507 struct ldlm_resource *oldres = lock->l_resource;
508 struct ldlm_resource *newres;
512 LASSERT(ns_is_client(ns));
514 lock_res_and_lock(lock);
515 if (memcmp(new_resid, &lock->l_resource->lr_name,
516 sizeof(lock->l_resource->lr_name)) == 0) {
518 unlock_res_and_lock(lock);
522 LASSERT(new_resid->name[0] != 0);
524 /* This function assumes that the lock isn't on any lists */
525 LASSERT(list_empty(&lock->l_res_link));
527 type = oldres->lr_type;
528 unlock_res_and_lock(lock);
530 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
532 RETURN(PTR_ERR(newres));
534 lu_ref_add(&newres->lr_reference, "lock", lock);
536 * To flip the lock from the old to the new resource, lock, oldres and
537 * newres have to be locked. Resource spin-locks are nested within
538 * lock->l_lock, and are taken in the memory address order to avoid
541 spin_lock(&lock->l_lock);
542 oldres = lock->l_resource;
543 if (oldres < newres) {
545 lock_res_nested(newres, LRT_NEW);
548 lock_res_nested(oldres, LRT_NEW);
550 LASSERT(memcmp(new_resid, &oldres->lr_name,
551 sizeof oldres->lr_name) != 0);
552 lock->l_resource = newres;
554 unlock_res_and_lock(lock);
556 /* ...and the flowers are still standing! */
557 lu_ref_del(&oldres->lr_reference, "lock", lock);
558 ldlm_resource_putref(oldres);
563 /** \defgroup ldlm_handles LDLM HANDLES
564 * Ways to get hold of locks without any addresses.
569 * Fills in handle for LDLM lock \a lock into supplied \a lockh
570 * Does not take any references.
572 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
574 lockh->cookie = lock->l_handle.h_cookie;
576 EXPORT_SYMBOL(ldlm_lock2handle);
579 * Obtain a lock reference by handle.
581 * if \a flags: atomically get the lock and set the flags.
582 * Return NULL if flag already set
584 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
587 struct ldlm_lock *lock;
592 lock = class_handle2object(handle->cookie, NULL);
596 if (lock->l_export != NULL && lock->l_export->exp_failed) {
597 CDEBUG(D_INFO, "lock export failed: lock %p, exp %p\n",
598 lock, lock->l_export);
603 /* It's unlikely but possible that someone marked the lock as
604 * destroyed after we did handle2object on it */
605 if ((flags == 0) && !ldlm_is_destroyed(lock)) {
606 lu_ref_add(&lock->l_reference, "handle", current);
610 lock_res_and_lock(lock);
612 LASSERT(lock->l_resource != NULL);
614 lu_ref_add_atomic(&lock->l_reference, "handle", current);
615 if (unlikely(ldlm_is_destroyed(lock))) {
616 unlock_res_and_lock(lock);
617 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
622 /* If we're setting flags, make sure none of them are already set. */
624 if ((lock->l_flags & flags) != 0) {
625 unlock_res_and_lock(lock);
630 lock->l_flags |= flags;
633 unlock_res_and_lock(lock);
636 EXPORT_SYMBOL(__ldlm_handle2lock);
637 /** @} ldlm_handles */
640 * Fill in "on the wire" representation for given LDLM lock into supplied
641 * lock descriptor \a desc structure.
643 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
645 ldlm_res2desc(lock->l_resource, &desc->l_resource);
646 desc->l_req_mode = lock->l_req_mode;
647 desc->l_granted_mode = lock->l_granted_mode;
648 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
649 &lock->l_policy_data,
650 &desc->l_policy_data);
654 * Add a lock to list of conflicting locks to send AST to.
656 * Only add if we have not sent a blocking AST to the lock yet.
658 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
659 struct list_head *work_list)
661 if (!ldlm_is_ast_sent(lock)) {
662 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
663 ldlm_set_ast_sent(lock);
664 /* If the enqueuing client said so, tell the AST recipient to
665 * discard dirty data, rather than writing back. */
666 if (ldlm_is_ast_discard_data(new))
667 ldlm_set_discard_data(lock);
669 /* Lock can be converted from a blocking state back to granted
670 * after lock convert or COS downgrade but still be in an
671 * older bl_list because it is controlled only by
672 * ldlm_work_bl_ast_lock(), let it be processed there.
674 if (list_empty(&lock->l_bl_ast)) {
675 list_add(&lock->l_bl_ast, work_list);
678 LASSERT(lock->l_blocking_lock == NULL);
679 lock->l_blocking_lock = LDLM_LOCK_GET(new);
684 * Add a lock to list of just granted locks to send completion AST to.
686 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
687 struct list_head *work_list)
689 if (!ldlm_is_cp_reqd(lock)) {
690 ldlm_set_cp_reqd(lock);
691 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
692 LASSERT(list_empty(&lock->l_cp_ast));
693 list_add(&lock->l_cp_ast, work_list);
699 * Aggregator function to add AST work items into a list. Determines
700 * what sort of an AST work needs to be done and calls the proper
702 * Must be called with lr_lock held.
704 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
705 struct list_head *work_list)
708 check_res_locked(lock->l_resource);
710 ldlm_add_bl_work_item(lock, new, work_list);
712 ldlm_add_cp_work_item(lock, work_list);
717 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
718 * r/w reference type is determined by \a mode
719 * Calls ldlm_lock_addref_internal.
721 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
723 struct ldlm_lock *lock;
725 lock = ldlm_handle2lock(lockh);
726 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
727 ldlm_lock_addref_internal(lock, mode);
730 EXPORT_SYMBOL(ldlm_lock_addref);
734 * Add specified reader/writer reference to LDLM lock \a lock.
735 * r/w reference type is determined by \a mode
736 * Removes lock from LRU if it is there.
737 * Assumes the LDLM lock is already locked.
739 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
742 ldlm_lock_remove_from_lru(lock);
743 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
745 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
747 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
749 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
752 lu_ref_add_atomic(&lock->l_reference, "user", lock);
753 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
757 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
758 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
760 * \retval 0 success, lock was addref-ed
762 * \retval -EAGAIN lock is being canceled.
764 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
766 struct ldlm_lock *lock;
770 lock = ldlm_handle2lock(lockh);
772 lock_res_and_lock(lock);
773 if (lock->l_readers != 0 || lock->l_writers != 0 ||
774 !ldlm_is_cbpending(lock)) {
775 ldlm_lock_addref_internal_nolock(lock, mode);
778 unlock_res_and_lock(lock);
783 EXPORT_SYMBOL(ldlm_lock_addref_try);
786 * Add specified reader/writer reference to LDLM lock \a lock.
787 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
788 * Only called for local locks.
790 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
792 lock_res_and_lock(lock);
793 ldlm_lock_addref_internal_nolock(lock, mode);
794 unlock_res_and_lock(lock);
798 * Removes reader/writer reference for LDLM lock \a lock.
799 * Assumes LDLM lock is already locked.
800 * only called in ldlm_flock_destroy and for local locks.
801 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
802 * that cannot be placed in LRU.
804 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
807 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
808 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
809 LASSERT(lock->l_readers > 0);
810 lu_ref_del(&lock->l_reference, "reader", lock);
813 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
814 LASSERT(lock->l_writers > 0);
815 lu_ref_del(&lock->l_reference, "writer", lock);
819 lu_ref_del(&lock->l_reference, "user", lock);
820 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
824 * Removes reader/writer reference for LDLM lock \a lock.
825 * Locks LDLM lock first.
826 * If the lock is determined to be client lock on a client and r/w refcount
827 * drops to zero and the lock is not blocked, the lock is added to LRU lock
829 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
831 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
833 struct ldlm_namespace *ns;
836 lock_res_and_lock(lock);
838 ns = ldlm_lock_to_ns(lock);
840 ldlm_lock_decref_internal_nolock(lock, mode);
842 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
843 !lock->l_readers && !lock->l_writers) {
844 /* If this is a local lock on a server namespace and this was
845 * the last reference, cancel the lock.
847 * Group locks are special:
848 * They must not go in LRU, but they are not called back
849 * like non-group locks, instead they are manually released.
850 * They have an l_writers reference which they keep until
851 * they are manually released, so we remove them when they have
852 * no more reader or writer references. - LU-6368 */
853 ldlm_set_cbpending(lock);
856 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
857 /* If we received a blocked AST and this was the last reference,
858 * run the callback. */
859 if (ldlm_is_ns_srv(lock) && lock->l_export)
860 CERROR("FL_CBPENDING set on non-local lock--just a "
863 LDLM_DEBUG(lock, "final decref done on cbpending lock");
865 LDLM_LOCK_GET(lock); /* dropped by bl thread */
866 ldlm_lock_remove_from_lru(lock);
867 unlock_res_and_lock(lock);
869 if (ldlm_is_fail_loc(lock))
870 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
872 if (ldlm_is_atomic_cb(lock) ||
873 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
874 ldlm_handle_bl_callback(ns, NULL, lock);
875 } else if (ns_is_client(ns) &&
876 !lock->l_readers && !lock->l_writers &&
877 !ldlm_is_no_lru(lock) &&
878 !ldlm_is_bl_ast(lock) &&
879 !ldlm_is_converting(lock)) {
881 LDLM_DEBUG(lock, "add lock into lru list");
883 /* If this is a client-side namespace and this was the last
884 * reference, put it on the LRU. */
885 ldlm_lock_add_to_lru(lock);
886 unlock_res_and_lock(lock);
888 if (ldlm_is_fail_loc(lock))
889 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
891 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
892 * are not supported by the server, otherwise, it is done on
894 if (!exp_connect_cancelset(lock->l_conn_export) &&
895 !ns_connect_lru_resize(ns))
896 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
898 LDLM_DEBUG(lock, "do not add lock into lru list");
899 unlock_res_and_lock(lock);
906 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
908 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
910 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
911 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
912 ldlm_lock_decref_internal(lock, mode);
915 EXPORT_SYMBOL(ldlm_lock_decref);
918 * Decrease reader/writer refcount for LDLM lock with handle
919 * \a lockh and mark it for subsequent cancellation once r/w refcount
920 * drops to zero instead of putting into LRU.
923 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
926 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
929 LASSERT(lock != NULL);
931 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
932 lock_res_and_lock(lock);
933 ldlm_set_cbpending(lock);
934 unlock_res_and_lock(lock);
935 ldlm_lock_decref_internal(lock, mode);
938 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
940 struct sl_insert_point {
941 struct list_head *res_link;
942 struct list_head *mode_link;
943 struct list_head *policy_link;
947 * Finds a position to insert the new lock into granted lock list.
949 * Used for locks eligible for skiplist optimization.
952 * queue [input]: the granted list where search acts on;
953 * req [input]: the lock whose position to be located;
954 * prev [output]: positions within 3 lists to insert @req to
958 * - ldlm_grant_lock_with_skiplist
960 static void search_granted_lock(struct list_head *queue,
961 struct ldlm_lock *req,
962 struct sl_insert_point *prev)
964 struct list_head *tmp;
965 struct ldlm_lock *lock, *mode_end, *policy_end;
968 list_for_each(tmp, queue) {
969 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
971 mode_end = list_entry(lock->l_sl_mode.prev,
972 struct ldlm_lock, l_sl_mode);
974 if (lock->l_req_mode != req->l_req_mode) {
975 /* jump to last lock of mode group */
976 tmp = &mode_end->l_res_link;
980 /* suitable mode group is found */
981 if (lock->l_resource->lr_type == LDLM_PLAIN) {
982 /* insert point is last lock of the mode group */
983 prev->res_link = &mode_end->l_res_link;
984 prev->mode_link = &mode_end->l_sl_mode;
985 prev->policy_link = &req->l_sl_policy;
988 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
991 list_entry(lock->l_sl_policy.prev,
995 if (lock->l_policy_data.l_inodebits.bits ==
996 req->l_policy_data.l_inodebits.bits) {
997 /* insert point is last lock of
998 * the policy group */
1000 &policy_end->l_res_link;
1002 &policy_end->l_sl_mode;
1004 &policy_end->l_sl_policy;
1009 if (policy_end == mode_end)
1010 /* done with mode group */
1013 /* go to next policy group within mode group */
1014 tmp = policy_end->l_res_link.next;
1015 lock = list_entry(tmp, struct ldlm_lock,
1017 } /* loop over policy groups within the mode group */
1019 /* insert point is last lock of the mode group,
1020 * new policy group is started */
1021 prev->res_link = &mode_end->l_res_link;
1022 prev->mode_link = &mode_end->l_sl_mode;
1023 prev->policy_link = &req->l_sl_policy;
1027 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1032 /* insert point is last lock on the queue,
1033 * new mode group and new policy group are started */
1034 prev->res_link = queue->prev;
1035 prev->mode_link = &req->l_sl_mode;
1036 prev->policy_link = &req->l_sl_policy;
1042 * Add a lock into resource granted list after a position described by
1045 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1046 struct sl_insert_point *prev)
1048 struct ldlm_resource *res = lock->l_resource;
1051 check_res_locked(res);
1053 ldlm_resource_dump(D_INFO, res);
1054 LDLM_DEBUG(lock, "About to add lock:");
1056 if (ldlm_is_destroyed(lock)) {
1057 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1061 LASSERT(list_empty(&lock->l_res_link));
1062 LASSERT(list_empty(&lock->l_sl_mode));
1063 LASSERT(list_empty(&lock->l_sl_policy));
1066 * lock->link == prev->link means lock is first starting the group.
1067 * Don't re-add to itself to suppress kernel warnings.
1069 if (&lock->l_res_link != prev->res_link)
1070 list_add(&lock->l_res_link, prev->res_link);
1071 if (&lock->l_sl_mode != prev->mode_link)
1072 list_add(&lock->l_sl_mode, prev->mode_link);
1073 if (&lock->l_sl_policy != prev->policy_link)
1074 list_add(&lock->l_sl_policy, prev->policy_link);
1080 * Add a lock to granted list on a resource maintaining skiplist
1083 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1085 struct sl_insert_point prev;
1087 LASSERT(lock->l_req_mode == lock->l_granted_mode);
1089 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1090 ldlm_granted_list_add_lock(lock, &prev);
1094 * Perform lock granting bookkeeping.
1096 * Includes putting the lock into granted list and updating lock mode.
1098 * - ldlm_lock_enqueue
1099 * - ldlm_reprocess_queue
1101 * must be called with lr_lock held
1103 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1105 struct ldlm_resource *res = lock->l_resource;
1108 check_res_locked(res);
1110 lock->l_granted_mode = lock->l_req_mode;
1112 if (work_list && lock->l_completion_ast != NULL)
1113 ldlm_add_ast_work_item(lock, NULL, work_list);
1115 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1116 ldlm_grant_lock_with_skiplist(lock);
1117 else if (res->lr_type == LDLM_EXTENT)
1118 ldlm_extent_add_lock(res, lock);
1119 else if (res->lr_type == LDLM_FLOCK) {
1120 /* We should not add locks to granted list in the following
1122 * - this is an UNLOCK but not a real lock;
1123 * - this is a TEST lock;
1124 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1125 * - this is a deadlock (flock cannot be granted) */
1126 if (lock->l_req_mode == 0 ||
1127 lock->l_req_mode == LCK_NL ||
1128 ldlm_is_test_lock(lock) ||
1129 ldlm_is_flock_deadlock(lock))
1131 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1136 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1141 * Describe the overlap between two locks. itree_overlap_cb data.
1143 struct lock_match_data {
1144 struct ldlm_lock *lmd_old;
1145 struct ldlm_lock *lmd_lock;
1146 enum ldlm_mode *lmd_mode;
1147 union ldlm_policy_data *lmd_policy;
1149 __u64 lmd_skip_flags;
1154 * Check if the given @lock meets the criteria for a match.
1155 * A reference on the lock is taken if matched.
1157 * \param lock test-against this lock
1158 * \param data parameters
1160 static int lock_matches(struct ldlm_lock *lock, struct lock_match_data *data)
1162 union ldlm_policy_data *lpol = &lock->l_policy_data;
1163 enum ldlm_mode match;
1165 if (lock == data->lmd_old)
1166 return INTERVAL_ITER_STOP;
1168 /* Check if this lock can be matched.
1169 * Used by LU-2919(exclusive open) for open lease lock */
1170 if (ldlm_is_excl(lock))
1171 return INTERVAL_ITER_CONT;
1173 /* llite sometimes wants to match locks that will be
1174 * canceled when their users drop, but we allow it to match
1175 * if it passes in CBPENDING and the lock still has users.
1176 * this is generally only going to be used by children
1177 * whose parents already hold a lock so forward progress
1178 * can still happen. */
1179 if (ldlm_is_cbpending(lock) &&
1180 !(data->lmd_flags & LDLM_FL_CBPENDING))
1181 return INTERVAL_ITER_CONT;
1182 if (!data->lmd_unref && ldlm_is_cbpending(lock) &&
1183 lock->l_readers == 0 && lock->l_writers == 0)
1184 return INTERVAL_ITER_CONT;
1186 if (!(lock->l_req_mode & *data->lmd_mode))
1187 return INTERVAL_ITER_CONT;
1188 match = lock->l_req_mode;
1190 switch (lock->l_resource->lr_type) {
1192 if (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1193 lpol->l_extent.end < data->lmd_policy->l_extent.end)
1194 return INTERVAL_ITER_CONT;
1196 if (unlikely(match == LCK_GROUP) &&
1197 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1198 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1199 return INTERVAL_ITER_CONT;
1202 /* We match if we have existing lock with same or wider set
1204 if ((lpol->l_inodebits.bits &
1205 data->lmd_policy->l_inodebits.bits) !=
1206 data->lmd_policy->l_inodebits.bits)
1207 return INTERVAL_ITER_CONT;
1213 /* We match if we have existing lock with same or wider set
1215 if (!data->lmd_unref && LDLM_HAVE_MASK(lock, GONE))
1216 return INTERVAL_ITER_CONT;
1218 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1219 return INTERVAL_ITER_CONT;
1221 /* Filter locks by skipping flags */
1222 if (data->lmd_skip_flags & lock->l_flags)
1223 return INTERVAL_ITER_CONT;
1225 if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1226 LDLM_LOCK_GET(lock);
1227 ldlm_lock_touch_in_lru(lock);
1229 ldlm_lock_addref_internal_nolock(lock, match);
1232 *data->lmd_mode = match;
1233 data->lmd_lock = lock;
1235 return INTERVAL_ITER_STOP;
1238 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1240 struct ldlm_interval *node = to_ldlm_interval(in);
1241 struct lock_match_data *data = args;
1242 struct ldlm_lock *lock;
1245 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1246 rc = lock_matches(lock, data);
1247 if (rc == INTERVAL_ITER_STOP)
1248 return INTERVAL_ITER_STOP;
1250 return INTERVAL_ITER_CONT;
1254 * Search for a lock with given parameters in interval trees.
1256 * \param res search for a lock in this resource
1257 * \param data parameters
1259 * \retval a referenced lock or NULL.
1261 static struct ldlm_lock *search_itree(struct ldlm_resource *res,
1262 struct lock_match_data *data)
1264 struct interval_node_extent ext = {
1265 .start = data->lmd_policy->l_extent.start,
1266 .end = data->lmd_policy->l_extent.end
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;
1287 * Search for a lock with given properties in a queue.
1289 * \param queue search for a lock in this queue
1290 * \param data parameters
1292 * \retval a referenced lock or NULL.
1294 static struct ldlm_lock *search_queue(struct list_head *queue,
1295 struct lock_match_data *data)
1297 struct ldlm_lock *lock;
1300 list_for_each_entry(lock, queue, l_res_link) {
1301 rc = lock_matches(lock, data);
1302 if (rc == INTERVAL_ITER_STOP)
1303 return data->lmd_lock;
1308 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1310 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1311 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1312 wake_up_all(&lock->l_waitq);
1315 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1317 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1319 lock_res_and_lock(lock);
1320 ldlm_lock_fail_match_locked(lock);
1321 unlock_res_and_lock(lock);
1325 * Mark lock as "matchable" by OST.
1327 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1329 * Assumes LDLM lock is already locked.
1331 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1333 ldlm_set_lvb_ready(lock);
1334 wake_up_all(&lock->l_waitq);
1336 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1339 * Mark lock as "matchable" by OST.
1340 * Locks the lock and then \see ldlm_lock_allow_match_locked
1342 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1344 lock_res_and_lock(lock);
1345 ldlm_lock_allow_match_locked(lock);
1346 unlock_res_and_lock(lock);
1348 EXPORT_SYMBOL(ldlm_lock_allow_match);
1351 * Attempt to find a lock with specified properties.
1353 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1356 * Can be called in two ways:
1358 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1359 * for a duplicate of.
1361 * Otherwise, all of the fields must be filled in, to match against.
1363 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1364 * server (ie, connh is NULL)
1365 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1366 * list will be considered
1367 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1368 * to be canceled can still be matched as long as they still have reader
1369 * or writer refernces
1370 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1371 * just tell us if we would have matched.
1373 * \retval 1 if it finds an already-existing lock that is compatible; in this
1374 * case, lockh is filled in with a addref()ed lock
1376 * We also check security context, and if that fails we simply return 0 (to
1377 * keep caller code unchanged), the context failure will be discovered by
1378 * caller sometime later.
1380 enum ldlm_mode ldlm_lock_match_with_skip(struct ldlm_namespace *ns,
1381 __u64 flags, __u64 skip_flags,
1382 const struct ldlm_res_id *res_id,
1383 enum ldlm_type type,
1384 union ldlm_policy_data *policy,
1385 enum ldlm_mode mode,
1386 struct lustre_handle *lockh, int unref)
1388 struct lock_match_data data = {
1392 .lmd_policy = policy,
1394 .lmd_skip_flags = skip_flags,
1397 struct ldlm_resource *res;
1398 struct ldlm_lock *lock;
1404 data.lmd_old = ldlm_handle2lock(lockh);
1405 LASSERT(data.lmd_old != NULL);
1407 ns = ldlm_lock_to_ns(data.lmd_old);
1408 res_id = &data.lmd_old->l_resource->lr_name;
1409 type = data.lmd_old->l_resource->lr_type;
1410 *data.lmd_mode = data.lmd_old->l_req_mode;
1413 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1415 LASSERT(data.lmd_old == NULL);
1419 LDLM_RESOURCE_ADDREF(res);
1421 if (res->lr_type == LDLM_EXTENT)
1422 lock = search_itree(res, &data);
1424 lock = search_queue(&res->lr_granted, &data);
1425 if (!lock && !(flags & LDLM_FL_BLOCK_GRANTED))
1426 lock = search_queue(&res->lr_waiting, &data);
1427 matched = lock ? mode : 0;
1429 LDLM_RESOURCE_DELREF(res);
1430 ldlm_resource_putref(res);
1433 ldlm_lock2handle(lock, lockh);
1434 if ((flags & LDLM_FL_LVB_READY) &&
1435 (!ldlm_is_lvb_ready(lock))) {
1436 __u64 wait_flags = LDLM_FL_LVB_READY |
1437 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1438 struct l_wait_info lwi;
1440 if (lock->l_completion_ast) {
1441 int err = lock->l_completion_ast(lock,
1442 LDLM_FL_WAIT_NOREPROC,
1445 GOTO(out_fail_match, matched = 0);
1448 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1449 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1451 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1452 l_wait_event(lock->l_waitq, lock->l_flags & wait_flags,
1454 if (!ldlm_is_lvb_ready(lock))
1455 GOTO(out_fail_match, matched = 0);
1458 /* check user's security context */
1459 if (lock->l_conn_export &&
1460 sptlrpc_import_check_ctx(
1461 class_exp2cliimp(lock->l_conn_export)))
1462 GOTO(out_fail_match, matched = 0);
1464 LDLM_DEBUG(lock, "matched (%llu %llu)",
1465 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1466 res_id->name[2] : policy->l_extent.start,
1467 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1468 res_id->name[3] : policy->l_extent.end);
1471 if (flags & LDLM_FL_TEST_LOCK)
1472 LDLM_LOCK_RELEASE(lock);
1474 ldlm_lock_decref_internal(lock, mode);
1477 /* less verbose for test-only */
1478 if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
1479 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1480 "%llu/%llu (%llu %llu)", ns,
1481 type, mode, res_id->name[0], res_id->name[1],
1482 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1483 res_id->name[2] : policy->l_extent.start,
1484 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1485 res_id->name[3] : policy->l_extent.end);
1487 if (data.lmd_old != NULL)
1488 LDLM_LOCK_PUT(data.lmd_old);
1492 EXPORT_SYMBOL(ldlm_lock_match_with_skip);
1494 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1497 struct ldlm_lock *lock;
1498 enum ldlm_mode mode = 0;
1501 lock = ldlm_handle2lock(lockh);
1503 lock_res_and_lock(lock);
1504 if (LDLM_HAVE_MASK(lock, GONE))
1507 if (ldlm_is_cbpending(lock) &&
1508 lock->l_readers == 0 && lock->l_writers == 0)
1512 *bits = lock->l_policy_data.l_inodebits.bits;
1513 mode = lock->l_granted_mode;
1514 ldlm_lock_addref_internal_nolock(lock, mode);
1521 unlock_res_and_lock(lock);
1522 LDLM_LOCK_PUT(lock);
1526 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1528 /** The caller must guarantee that the buffer is large enough. */
1529 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1530 enum req_location loc, void *data, int size)
1535 LASSERT(data != NULL);
1538 switch (lock->l_lvb_type) {
1540 if (size == sizeof(struct ost_lvb)) {
1541 if (loc == RCL_CLIENT)
1542 lvb = req_capsule_client_swab_get(pill,
1544 lustre_swab_ost_lvb);
1546 lvb = req_capsule_server_swab_get(pill,
1548 lustre_swab_ost_lvb);
1549 if (unlikely(lvb == NULL)) {
1550 LDLM_ERROR(lock, "no LVB");
1554 memcpy(data, lvb, size);
1555 } else if (size == sizeof(struct ost_lvb_v1)) {
1556 struct ost_lvb *olvb = data;
1558 if (loc == RCL_CLIENT)
1559 lvb = req_capsule_client_swab_get(pill,
1561 lustre_swab_ost_lvb_v1);
1563 lvb = req_capsule_server_sized_swab_get(pill,
1565 lustre_swab_ost_lvb_v1);
1566 if (unlikely(lvb == NULL)) {
1567 LDLM_ERROR(lock, "no LVB");
1571 memcpy(data, lvb, size);
1572 olvb->lvb_mtime_ns = 0;
1573 olvb->lvb_atime_ns = 0;
1574 olvb->lvb_ctime_ns = 0;
1576 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1582 if (size == sizeof(struct lquota_lvb)) {
1583 if (loc == RCL_CLIENT)
1584 lvb = req_capsule_client_swab_get(pill,
1586 lustre_swab_lquota_lvb);
1588 lvb = req_capsule_server_swab_get(pill,
1590 lustre_swab_lquota_lvb);
1591 if (unlikely(lvb == NULL)) {
1592 LDLM_ERROR(lock, "no LVB");
1596 memcpy(data, lvb, size);
1598 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1607 if (loc == RCL_CLIENT)
1608 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1610 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1611 if (unlikely(lvb == NULL)) {
1612 LDLM_ERROR(lock, "no LVB");
1616 memcpy(data, lvb, size);
1619 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1620 libcfs_debug_dumpstack(NULL);
1628 * Create and fill in new LDLM lock with specified properties.
1629 * Returns a referenced lock
1631 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1632 const struct ldlm_res_id *res_id,
1633 enum ldlm_type type,
1634 enum ldlm_mode mode,
1635 const struct ldlm_callback_suite *cbs,
1636 void *data, __u32 lvb_len,
1637 enum lvb_type lvb_type)
1639 struct ldlm_lock *lock;
1640 struct ldlm_resource *res;
1644 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1646 RETURN(ERR_CAST(res));
1648 lock = ldlm_lock_new(res);
1650 RETURN(ERR_PTR(-ENOMEM));
1652 lock->l_req_mode = mode;
1653 lock->l_ast_data = data;
1654 lock->l_pid = current_pid();
1655 if (ns_is_server(ns))
1656 ldlm_set_ns_srv(lock);
1658 lock->l_blocking_ast = cbs->lcs_blocking;
1659 lock->l_completion_ast = cbs->lcs_completion;
1660 lock->l_glimpse_ast = cbs->lcs_glimpse;
1663 lock->l_tree_node = NULL;
1664 /* if this is the extent lock, allocate the interval tree node */
1665 if (type == LDLM_EXTENT)
1666 if (ldlm_interval_alloc(lock) == NULL)
1667 GOTO(out, rc = -ENOMEM);
1670 lock->l_lvb_len = lvb_len;
1671 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1672 if (lock->l_lvb_data == NULL)
1673 GOTO(out, rc = -ENOMEM);
1676 lock->l_lvb_type = lvb_type;
1677 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1678 GOTO(out, rc = -ENOENT);
1683 ldlm_lock_destroy(lock);
1684 LDLM_LOCK_RELEASE(lock);
1685 RETURN(ERR_PTR(rc));
1688 #ifdef HAVE_SERVER_SUPPORT
1689 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1692 struct ldlm_resource *res = lock->l_resource;
1693 enum ldlm_error rc = ELDLM_OK;
1694 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1695 ldlm_processing_policy policy;
1698 policy = ldlm_processing_policy_table[res->lr_type];
1700 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1701 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1702 res->lr_type != LDLM_FLOCK) {
1703 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1704 if (rc == -ERESTART)
1708 if (!list_empty(&rpc_list))
1709 ldlm_discard_bl_list(&rpc_list);
1716 * Enqueue (request) a lock.
1718 * Does not block. As a result of enqueue the lock would be put
1719 * into granted or waiting list.
1721 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1722 * set, skip all the enqueueing and delegate lock processing to intent policy
1725 enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
1726 struct ldlm_namespace *ns,
1727 struct ldlm_lock **lockp,
1728 void *cookie, __u64 *flags)
1730 struct ldlm_lock *lock = *lockp;
1731 struct ldlm_resource *res = lock->l_resource;
1732 int local = ns_is_client(ldlm_res_to_ns(res));
1733 enum ldlm_error rc = ELDLM_OK;
1734 struct ldlm_interval *node = NULL;
1737 /* policies are not executed on the client or during replay */
1738 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1739 && !local && ns->ns_policy) {
1740 rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
1742 if (rc == ELDLM_LOCK_REPLACED) {
1743 /* The lock that was returned has already been granted,
1744 * and placed into lockp. If it's not the same as the
1745 * one we passed in, then destroy the old one and our
1746 * work here is done. */
1747 if (lock != *lockp) {
1748 ldlm_lock_destroy(lock);
1749 LDLM_LOCK_RELEASE(lock);
1751 *flags |= LDLM_FL_LOCK_CHANGED;
1753 } else if (rc != ELDLM_OK &&
1754 lock->l_req_mode == lock->l_granted_mode) {
1755 LASSERT(*flags & LDLM_FL_RESENT);
1756 /* It may happen that ns_policy returns an error in
1757 * resend case, object may be unlinked or just some
1758 * error occurs. It is unclear if lock reached the
1759 * client in the original reply, just leave the lock on
1760 * server, not returning it again to client. Due to
1761 * LU-6529, the server will not OOM. */
1763 } else if (rc != ELDLM_OK ||
1764 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1765 ldlm_lock_destroy(lock);
1770 if (*flags & LDLM_FL_RESENT) {
1771 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1772 * Set LOCK_CHANGED always.
1773 * Check if the lock is granted for BLOCK_GRANTED.
1774 * Take NO_TIMEOUT from the lock as it is inherited through
1775 * LDLM_FL_INHERIT_MASK */
1776 *flags |= LDLM_FL_LOCK_CHANGED;
1777 if (lock->l_req_mode != lock->l_granted_mode)
1778 *flags |= LDLM_FL_BLOCK_GRANTED;
1779 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1783 /* For a replaying lock, it might be already in granted list. So
1784 * unlinking the lock will cause the interval node to be freed, we
1785 * have to allocate the interval node early otherwise we can't regrant
1786 * this lock in the future. - jay */
1787 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1788 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1790 lock_res_and_lock(lock);
1791 if (local && lock->l_req_mode == lock->l_granted_mode) {
1792 /* The server returned a blocked lock, but it was granted
1793 * before we got a chance to actually enqueue it. We don't
1794 * need to do anything else. */
1795 *flags &= ~LDLM_FL_BLOCKED_MASK;
1796 GOTO(out, rc = ELDLM_OK);
1799 ldlm_resource_unlink_lock(lock);
1800 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1802 ldlm_lock_destroy_nolock(lock);
1803 GOTO(out, rc = -ENOMEM);
1806 INIT_LIST_HEAD(&node->li_group);
1807 ldlm_interval_attach(node, lock);
1811 /* Some flags from the enqueue want to make it into the AST, via the
1812 * lock's l_flags. */
1813 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1814 ldlm_set_ast_discard_data(lock);
1815 if (*flags & LDLM_FL_TEST_LOCK)
1816 ldlm_set_test_lock(lock);
1817 if (*flags & LDLM_FL_COS_INCOMPAT)
1818 ldlm_set_cos_incompat(lock);
1819 if (*flags & LDLM_FL_COS_ENABLED)
1820 ldlm_set_cos_enabled(lock);
1822 /* This distinction between local lock trees is very important; a client
1823 * namespace only has information about locks taken by that client, and
1824 * thus doesn't have enough information to decide for itself if it can
1825 * be granted (below). In this case, we do exactly what the server
1826 * tells us to do, as dictated by the 'flags'.
1828 * We do exactly the same thing during recovery, when the server is
1829 * more or less trusting the clients not to lie.
1831 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1834 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1835 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1837 ldlm_grant_lock(lock, NULL);
1838 GOTO(out, rc = ELDLM_OK);
1839 #ifdef HAVE_SERVER_SUPPORT
1840 } else if (*flags & LDLM_FL_REPLAY) {
1841 if (*flags & LDLM_FL_BLOCK_WAIT) {
1842 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1843 GOTO(out, rc = ELDLM_OK);
1844 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1845 ldlm_grant_lock(lock, NULL);
1846 GOTO(out, rc = ELDLM_OK);
1848 /* If no flags, fall through to normal enqueue path. */
1851 rc = ldlm_lock_enqueue_helper(lock, flags);
1855 CERROR("This is client-side-only module, cannot handle "
1856 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1862 unlock_res_and_lock(lock);
1864 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1868 #ifdef HAVE_SERVER_SUPPORT
1870 * Iterate through all waiting locks on a given resource queue and attempt to
1873 * Must be called with resource lock held.
1875 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1876 struct list_head *work_list,
1877 enum ldlm_process_intention intention)
1879 struct list_head *tmp, *pos;
1880 ldlm_processing_policy policy;
1882 int rc = LDLM_ITER_CONTINUE;
1883 enum ldlm_error err;
1884 struct list_head bl_ast_list = LIST_HEAD_INIT(bl_ast_list);
1887 check_res_locked(res);
1889 policy = ldlm_processing_policy_table[res->lr_type];
1891 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1892 intention == LDLM_PROCESS_RECOVERY);
1895 list_for_each_safe(tmp, pos, queue) {
1896 struct ldlm_lock *pending;
1897 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1899 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1901 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1904 rc = policy(pending, &flags, intention, &err, &rpc_list);
1905 if (pending->l_granted_mode == pending->l_req_mode ||
1906 res->lr_type == LDLM_FLOCK) {
1907 list_splice(&rpc_list, work_list);
1909 list_splice(&rpc_list, &bl_ast_list);
1912 * When this is called from recovery done, we always want
1913 * to scan the whole list no matter what 'rc' is returned.
1915 if (rc != LDLM_ITER_CONTINUE &&
1916 intention == LDLM_PROCESS_RESCAN)
1920 if (!list_empty(&bl_ast_list)) {
1923 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
1927 if (rc == -ERESTART)
1931 if (!list_empty(&bl_ast_list))
1932 ldlm_discard_bl_list(&bl_ast_list);
1934 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
1938 * Conflicting locks are detected for a lock to be enqueued, add the lock
1939 * into waiting list and send blocking ASTs to the conflicting locks.
1941 * \param[in] lock The lock to be enqueued.
1942 * \param[out] flags Lock flags for the lock to be enqueued.
1943 * \param[in] rpc_list Conflicting locks list.
1945 * \retval -ERESTART: Some lock was instantly canceled while sending
1946 * blocking ASTs, caller needs to re-check conflicting
1948 * \retval -EAGAIN: Lock was destroyed, caller should return error.
1949 * \reval 0: Lock is successfully added in waiting list.
1951 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
1952 struct list_head *rpc_list)
1954 struct ldlm_resource *res = lock->l_resource;
1958 check_res_locked(res);
1960 /* If either of the compat_queue()s returned failure, then we
1961 * have ASTs to send and must go onto the waiting list.
1963 * bug 2322: we used to unlink and re-add here, which was a
1964 * terrible folly -- if we goto restart, we could get
1965 * re-ordered! Causes deadlock, because ASTs aren't sent! */
1966 if (list_empty(&lock->l_res_link))
1967 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1970 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
1973 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
1974 !ns_is_client(ldlm_res_to_ns(res)))
1975 class_fail_export(lock->l_export);
1978 if (rc == -ERESTART) {
1979 /* 15715: The lock was granted and destroyed after
1980 * resource lock was dropped. Interval node was freed
1981 * in ldlm_lock_destroy. Anyway, this always happens
1982 * when a client is being evicted. So it would be
1983 * ok to return an error. -jay */
1984 if (ldlm_is_destroyed(lock))
1987 /* lock was granted while resource was unlocked. */
1988 if (lock->l_granted_mode == lock->l_req_mode) {
1989 /* bug 11300: if the lock has been granted,
1990 * break earlier because otherwise, we will go
1991 * to restart and ldlm_resource_unlink will be
1992 * called and it causes the interval node to be
1993 * freed. Then we will fail at
1994 * ldlm_extent_add_lock() */
1995 *flags &= ~LDLM_FL_BLOCKED_MASK;
2001 *flags |= LDLM_FL_BLOCK_GRANTED;
2007 * Discard all AST work items from list.
2009 * If for whatever reason we do not want to send ASTs to conflicting locks
2010 * anymore, disassemble the list with this function.
2012 void ldlm_discard_bl_list(struct list_head *bl_list)
2014 struct ldlm_lock *lock, *tmp;
2018 list_for_each_entry_safe(lock, tmp, bl_list, l_bl_ast) {
2019 LASSERT(!list_empty(&lock->l_bl_ast));
2020 list_del_init(&lock->l_bl_ast);
2021 ldlm_clear_ast_sent(lock);
2022 LASSERT(lock->l_bl_ast_run == 0);
2023 ldlm_clear_blocking_lock(lock);
2024 LDLM_LOCK_RELEASE(lock);
2030 * Process a call to blocking AST callback for a lock in ast_work list
2033 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2035 struct ldlm_cb_set_arg *arg = opaq;
2036 struct ldlm_lock *lock;
2037 struct ldlm_lock_desc d;
2038 struct ldlm_bl_desc bld;
2043 if (list_empty(arg->list))
2046 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2048 /* nobody should touch l_bl_ast but some locks in the list may become
2049 * granted after lock convert or COS downgrade, these locks should be
2050 * just skipped here and removed from the list.
2052 lock_res_and_lock(lock);
2053 list_del_init(&lock->l_bl_ast);
2055 /* lock is not blocking lock anymore, but was kept in the list because
2056 * it can managed only here.
2058 if (!ldlm_is_ast_sent(lock)) {
2059 unlock_res_and_lock(lock);
2060 LDLM_LOCK_RELEASE(lock);
2064 LASSERT(lock->l_blocking_lock);
2065 ldlm_lock2desc(lock->l_blocking_lock, &d);
2066 /* copy blocking lock ibits in cancel_bits as well,
2067 * new client may use them for lock convert and it is
2068 * important to use new field to convert locks from
2071 d.l_policy_data.l_inodebits.cancel_bits =
2072 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2074 /* Blocking lock is being destroyed here but some information about it
2075 * may be needed inside l_blocking_ast() function below,
2076 * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
2078 bld.bl_same_client = lock->l_client_cookie ==
2079 lock->l_blocking_lock->l_client_cookie;
2080 bld.bl_cos_incompat = ldlm_is_cos_incompat(lock->l_blocking_lock);
2081 arg->bl_desc = &bld;
2083 LASSERT(ldlm_is_ast_sent(lock));
2084 LASSERT(lock->l_bl_ast_run == 0);
2085 lock->l_bl_ast_run++;
2086 ldlm_clear_blocking_lock(lock);
2087 unlock_res_and_lock(lock);
2089 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2091 LDLM_LOCK_RELEASE(lock);
2097 * Process a call to revocation AST callback for a lock in ast_work list
2100 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2102 struct ldlm_cb_set_arg *arg = opaq;
2103 struct ldlm_lock_desc desc;
2105 struct ldlm_lock *lock;
2108 if (list_empty(arg->list))
2111 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2112 list_del_init(&lock->l_rk_ast);
2114 /* the desc just pretend to exclusive */
2115 ldlm_lock2desc(lock, &desc);
2116 desc.l_req_mode = LCK_EX;
2117 desc.l_granted_mode = 0;
2119 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2120 LDLM_LOCK_RELEASE(lock);
2126 * Process a call to glimpse AST callback for a lock in ast_work list
2128 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2130 struct ldlm_cb_set_arg *arg = opaq;
2131 struct ldlm_glimpse_work *gl_work;
2132 struct ldlm_lock *lock;
2136 if (list_empty(arg->list))
2139 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2141 list_del_init(&gl_work->gl_list);
2143 lock = gl_work->gl_lock;
2145 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2146 arg->gl_desc = gl_work->gl_desc;
2147 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2148 arg->gl_interpret_data = gl_work->gl_interpret_data;
2150 /* invoke the actual glimpse callback */
2151 if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
2154 LDLM_LOCK_RELEASE(lock);
2155 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2156 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2158 OBD_FREE_PTR(gl_work);
2165 * Process a call to completion AST callback for a lock in ast_work list
2168 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2170 struct ldlm_cb_set_arg *arg = opaq;
2171 struct ldlm_lock *lock;
2172 ldlm_completion_callback completion_callback;
2177 if (list_empty(arg->list))
2180 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2182 /* It's possible to receive a completion AST before we've set
2183 * the l_completion_ast pointer: either because the AST arrived
2184 * before the reply, or simply because there's a small race
2185 * window between receiving the reply and finishing the local
2186 * enqueue. (bug 842)
2188 * This can't happen with the blocking_ast, however, because we
2189 * will never call the local blocking_ast until we drop our
2190 * reader/writer reference, which we won't do until we get the
2191 * reply and finish enqueueing. */
2193 /* nobody should touch l_cp_ast */
2194 lock_res_and_lock(lock);
2195 list_del_init(&lock->l_cp_ast);
2196 LASSERT(ldlm_is_cp_reqd(lock));
2197 /* save l_completion_ast since it can be changed by
2198 * mds_intent_policy(), see bug 14225 */
2199 completion_callback = lock->l_completion_ast;
2200 ldlm_clear_cp_reqd(lock);
2201 unlock_res_and_lock(lock);
2203 if (completion_callback != NULL)
2204 rc = completion_callback(lock, 0, (void *)arg);
2205 LDLM_LOCK_RELEASE(lock);
2211 * Process list of locks in need of ASTs being sent.
2213 * Used on server to send multiple ASTs together instead of sending one by
2216 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2217 ldlm_desc_ast_t ast_type)
2219 struct ldlm_cb_set_arg *arg;
2220 set_producer_func work_ast_lock;
2223 if (list_empty(rpc_list))
2230 atomic_set(&arg->restart, 0);
2231 arg->list = rpc_list;
2234 case LDLM_WORK_CP_AST:
2235 arg->type = LDLM_CP_CALLBACK;
2236 work_ast_lock = ldlm_work_cp_ast_lock;
2238 #ifdef HAVE_SERVER_SUPPORT
2239 case LDLM_WORK_BL_AST:
2240 arg->type = LDLM_BL_CALLBACK;
2241 work_ast_lock = ldlm_work_bl_ast_lock;
2243 case LDLM_WORK_REVOKE_AST:
2244 arg->type = LDLM_BL_CALLBACK;
2245 work_ast_lock = ldlm_work_revoke_ast_lock;
2247 case LDLM_WORK_GL_AST:
2248 arg->type = LDLM_GL_CALLBACK;
2249 work_ast_lock = ldlm_work_gl_ast_lock;
2256 /* We create a ptlrpc request set with flow control extension.
2257 * This request set will use the work_ast_lock function to produce new
2258 * requests and will send a new request each time one completes in order
2259 * to keep the number of requests in flight to ns_max_parallel_ast */
2260 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2261 work_ast_lock, arg);
2262 if (arg->set == NULL)
2263 GOTO(out, rc = -ENOMEM);
2265 ptlrpc_set_wait(NULL, arg->set);
2266 ptlrpc_set_destroy(arg->set);
2268 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2276 * Try to grant all waiting locks on a resource.
2278 * Calls ldlm_reprocess_queue on waiting queue.
2280 * Typically called after some resource locks are cancelled to see
2281 * if anything could be granted as a result of the cancellation.
2283 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2284 enum ldlm_process_intention intention)
2286 struct list_head rpc_list;
2287 #ifdef HAVE_SERVER_SUPPORT
2288 struct obd_device *obd;
2292 INIT_LIST_HEAD(&rpc_list);
2293 /* Local lock trees don't get reprocessed. */
2294 if (ns_is_client(ldlm_res_to_ns(res))) {
2299 /* Disable reprocess during lock replay stage but allow during
2300 * request replay stage.
2302 obd = ldlm_res_to_ns(res)->ns_obd;
2303 if (obd->obd_recovering &&
2304 atomic_read(&obd->obd_req_replay_clients) == 0)
2308 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list, intention);
2311 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2313 if (rc == -ERESTART) {
2314 LASSERT(list_empty(&rpc_list));
2320 INIT_LIST_HEAD(&rpc_list);
2321 if (!ns_is_client(ldlm_res_to_ns(res))) {
2322 CERROR("This is client-side-only module, cannot handle "
2323 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2330 void ldlm_reprocess_all(struct ldlm_resource *res)
2332 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN);
2334 EXPORT_SYMBOL(ldlm_reprocess_all);
2336 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2337 struct hlist_node *hnode, void *arg)
2339 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2341 /* This is only called once after recovery done. LU-8306. */
2342 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY);
2347 * Iterate through all resources on a namespace attempting to grant waiting
2350 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2355 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2356 ldlm_reprocess_res, NULL, 0);
2362 * Helper function to call blocking AST for LDLM lock \a lock in a
2363 * "cancelling" mode.
2365 void ldlm_cancel_callback(struct ldlm_lock *lock)
2367 check_res_locked(lock->l_resource);
2368 if (!ldlm_is_cancel(lock)) {
2369 ldlm_set_cancel(lock);
2370 if (lock->l_blocking_ast) {
2371 unlock_res_and_lock(lock);
2372 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2374 lock_res_and_lock(lock);
2376 LDLM_DEBUG(lock, "no blocking ast");
2379 /* only canceller can set bl_done bit */
2380 ldlm_set_bl_done(lock);
2381 wake_up_all(&lock->l_waitq);
2382 } else if (!ldlm_is_bl_done(lock)) {
2383 struct l_wait_info lwi = { 0 };
2385 /* The lock is guaranteed to have been canceled once
2386 * returning from this function. */
2387 unlock_res_and_lock(lock);
2388 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
2389 lock_res_and_lock(lock);
2394 * Remove skiplist-enabled LDLM lock \a req from granted list
2396 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2398 if (req->l_resource->lr_type != LDLM_PLAIN &&
2399 req->l_resource->lr_type != LDLM_IBITS)
2402 list_del_init(&req->l_sl_policy);
2403 list_del_init(&req->l_sl_mode);
2407 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2409 void ldlm_lock_cancel(struct ldlm_lock *lock)
2411 struct ldlm_resource *res;
2412 struct ldlm_namespace *ns;
2415 lock_res_and_lock(lock);
2417 res = lock->l_resource;
2418 ns = ldlm_res_to_ns(res);
2420 /* Please do not, no matter how tempting, remove this LBUG without
2421 * talking to me first. -phik */
2422 if (lock->l_readers || lock->l_writers) {
2423 LDLM_ERROR(lock, "lock still has references");
2424 unlock_res_and_lock(lock);
2428 if (ldlm_is_waited(lock))
2429 ldlm_del_waiting_lock(lock);
2431 /* Releases cancel callback. */
2432 ldlm_cancel_callback(lock);
2434 /* Yes, second time, just in case it was added again while we were
2435 * running with no res lock in ldlm_cancel_callback */
2436 if (ldlm_is_waited(lock))
2437 ldlm_del_waiting_lock(lock);
2439 ldlm_resource_unlink_lock(lock);
2440 ldlm_lock_destroy_nolock(lock);
2442 if (lock->l_granted_mode == lock->l_req_mode)
2443 ldlm_pool_del(&ns->ns_pool, lock);
2445 /* Make sure we will not be called again for same lock what is possible
2446 * if not to zero out lock->l_granted_mode */
2447 lock->l_granted_mode = LCK_MINMODE;
2448 unlock_res_and_lock(lock);
2452 EXPORT_SYMBOL(ldlm_lock_cancel);
2455 * Set opaque data into the lock that only makes sense to upper layer.
2457 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2459 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2464 if (lock->l_ast_data == NULL)
2465 lock->l_ast_data = data;
2466 if (lock->l_ast_data == data)
2468 LDLM_LOCK_PUT(lock);
2472 EXPORT_SYMBOL(ldlm_lock_set_data);
2474 struct export_cl_data {
2475 const struct lu_env *ecl_env;
2476 struct obd_export *ecl_exp;
2480 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2481 struct ldlm_lock *lock,
2482 struct export_cl_data *ecl)
2484 struct ldlm_resource *res;
2486 res = ldlm_resource_getref(lock->l_resource);
2488 ldlm_lvbo_update(ecl->ecl_env, res, lock, NULL, 1);
2489 ldlm_lock_cancel(lock);
2490 if (!exp->exp_obd->obd_stopping)
2491 ldlm_reprocess_all(res);
2492 ldlm_resource_putref(res);
2495 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2496 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2497 exp, ecl->ecl_loop);
2502 * Iterator function for ldlm_export_cancel_locks.
2503 * Cancels passed locks.
2506 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2507 struct hlist_node *hnode, void *data)
2510 struct export_cl_data *ecl = (struct export_cl_data *)data;
2511 struct obd_export *exp = ecl->ecl_exp;
2512 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2514 LDLM_LOCK_GET(lock);
2515 ldlm_cancel_lock_for_export(exp, lock, ecl);
2516 LDLM_LOCK_RELEASE(lock);
2522 * Cancel all blocked locks for given export.
2524 * Typically called on client disconnection/eviction
2526 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2529 struct export_cl_data ecl = {
2535 rc = lu_env_init(&env, LCT_DT_THREAD);
2540 while (!list_empty(&exp->exp_bl_list)) {
2541 struct ldlm_lock *lock;
2543 spin_lock_bh(&exp->exp_bl_list_lock);
2544 if (!list_empty(&exp->exp_bl_list)) {
2545 lock = list_entry(exp->exp_bl_list.next,
2546 struct ldlm_lock, l_exp_list);
2547 LDLM_LOCK_GET(lock);
2548 list_del_init(&lock->l_exp_list);
2552 spin_unlock_bh(&exp->exp_bl_list_lock);
2557 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2558 LDLM_LOCK_RELEASE(lock);
2563 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2564 "left on hash table %d.\n", exp, ecl.ecl_loop,
2565 atomic_read(&exp->exp_lock_hash->hs_count));
2567 return ecl.ecl_loop;
2571 * Cancel all locks for given export.
2573 * Typically called after client disconnection/eviction
2575 int ldlm_export_cancel_locks(struct obd_export *exp)
2577 struct export_cl_data ecl;
2581 rc = lu_env_init(&env, LCT_DT_THREAD);
2588 cfs_hash_for_each_empty(exp->exp_lock_hash,
2589 ldlm_cancel_locks_for_export_cb, &ecl);
2591 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2592 "left on hash table %d.\n", exp, ecl.ecl_loop,
2593 atomic_read(&exp->exp_lock_hash->hs_count));
2595 if (ecl.ecl_loop > 0 &&
2596 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2597 exp->exp_obd->obd_stopping)
2598 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2602 return ecl.ecl_loop;
2606 * Downgrade an PW/EX lock to COS mode.
2608 * A lock mode convertion from PW/EX mode to less conflict mode. The
2609 * convertion may fail if lock was canceled before downgrade, but it doesn't
2610 * indicate any problem, because such lock has no reader or writer, and will
2613 * Used by Commit on Sharing (COS) code to force object changes commit in case
2614 * of conflict. Converted lock is considered as new lock and all blocking AST
2615 * things are cleared, so any pending or new blocked lock on that lock will
2616 * cause new call to blocking_ast and force resource object commit.
2618 * \param lock A lock to convert
2619 * \param new_mode new lock mode
2621 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2623 #ifdef HAVE_SERVER_SUPPORT
2626 LASSERT(new_mode == LCK_COS);
2628 lock_res_and_lock(lock);
2630 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2631 unlock_res_and_lock(lock);
2633 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2634 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2638 ldlm_resource_unlink_lock(lock);
2640 * Remove the lock from pool as it will be added again in
2641 * ldlm_grant_lock() called below.
2643 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2645 /* Consider downgraded lock as a new lock and clear all states
2646 * related to a previous blocking AST processing.
2648 ldlm_clear_blocking_data(lock);
2650 lock->l_req_mode = new_mode;
2651 ldlm_grant_lock(lock, NULL);
2652 unlock_res_and_lock(lock);
2654 ldlm_reprocess_all(lock->l_resource);
2659 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2662 * Print lock with lock handle \a lockh description into debug log.
2664 * Used when printing all locks on a resource for debug purposes.
2666 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2668 struct ldlm_lock *lock;
2670 if (!((libcfs_debug | D_ERROR) & level))
2673 lock = ldlm_handle2lock(lockh);
2677 LDLM_DEBUG_LIMIT(level, lock, "###");
2679 LDLM_LOCK_PUT(lock);
2681 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2684 * Print lock information with custom message into debug log.
2687 void _ldlm_lock_debug(struct ldlm_lock *lock,
2688 struct libcfs_debug_msg_data *msgdata,
2689 const char *fmt, ...)
2692 struct obd_export *exp = lock->l_export;
2693 struct ldlm_resource *resource = NULL;
2694 char *nid = "local";
2696 /* on server-side resource of lock doesn't change */
2697 if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2698 if (lock->l_resource != NULL)
2699 resource = ldlm_resource_getref(lock->l_resource);
2700 } else if (spin_trylock(&lock->l_lock)) {
2701 if (lock->l_resource != NULL)
2702 resource = ldlm_resource_getref(lock->l_resource);
2703 spin_unlock(&lock->l_lock);
2706 va_start(args, fmt);
2708 if (exp && exp->exp_connection) {
2709 nid = obd_export_nid2str(exp);
2710 } else if (exp && exp->exp_obd != NULL) {
2711 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2712 nid = obd_import_nid2str(imp);
2715 if (resource == NULL) {
2716 libcfs_debug_vmsg2(msgdata, fmt, args,
2717 " ns: \?\? lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2718 "res: \?\? rrc=\?\? type: \?\?\? flags: %#llx nid: %s "
2719 "remote: %#llx expref: %d pid: %u timeout: %lld "
2722 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2723 lock->l_readers, lock->l_writers,
2724 ldlm_lockname[lock->l_granted_mode],
2725 ldlm_lockname[lock->l_req_mode],
2726 lock->l_flags, nid, lock->l_remote_handle.cookie,
2727 exp ? atomic_read(&exp->exp_refcount) : -99,
2728 lock->l_pid, lock->l_callback_timeout, lock->l_lvb_type);
2733 switch (resource->lr_type) {
2735 libcfs_debug_vmsg2(msgdata, fmt, args,
2736 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2737 "res: "DLDLMRES" rrc: %d type: %s [%llu->%llu] "
2738 "(req %llu->%llu) flags: %#llx nid: %s remote: "
2739 "%#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2740 ldlm_lock_to_ns_name(lock), lock,
2741 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2742 lock->l_readers, lock->l_writers,
2743 ldlm_lockname[lock->l_granted_mode],
2744 ldlm_lockname[lock->l_req_mode],
2746 atomic_read(&resource->lr_refcount),
2747 ldlm_typename[resource->lr_type],
2748 lock->l_policy_data.l_extent.start,
2749 lock->l_policy_data.l_extent.end,
2750 lock->l_req_extent.start, lock->l_req_extent.end,
2751 lock->l_flags, nid, lock->l_remote_handle.cookie,
2752 exp ? atomic_read(&exp->exp_refcount) : -99,
2753 lock->l_pid, lock->l_callback_timeout,
2758 libcfs_debug_vmsg2(msgdata, fmt, args,
2759 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2760 "res: "DLDLMRES" rrc: %d type: %s pid: %d "
2761 "[%llu->%llu] flags: %#llx nid: %s "
2762 "remote: %#llx expref: %d pid: %u timeout: %lld\n",
2763 ldlm_lock_to_ns_name(lock), lock,
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],
2769 atomic_read(&resource->lr_refcount),
2770 ldlm_typename[resource->lr_type],
2771 lock->l_policy_data.l_flock.pid,
2772 lock->l_policy_data.l_flock.start,
2773 lock->l_policy_data.l_flock.end,
2774 lock->l_flags, nid, lock->l_remote_handle.cookie,
2775 exp ? atomic_read(&exp->exp_refcount) : -99,
2776 lock->l_pid, lock->l_callback_timeout);
2780 libcfs_debug_vmsg2(msgdata, fmt, args,
2781 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2782 "res: "DLDLMRES" bits %#llx/%#llx rrc: %d type: %s "
2783 "flags: %#llx nid: %s remote: %#llx expref: %d "
2784 "pid: %u timeout: %lld lvb_type: %d\n",
2785 ldlm_lock_to_ns_name(lock),
2786 lock, lock->l_handle.h_cookie,
2787 atomic_read(&lock->l_refc),
2788 lock->l_readers, lock->l_writers,
2789 ldlm_lockname[lock->l_granted_mode],
2790 ldlm_lockname[lock->l_req_mode],
2792 lock->l_policy_data.l_inodebits.bits,
2793 lock->l_policy_data.l_inodebits.try_bits,
2794 atomic_read(&resource->lr_refcount),
2795 ldlm_typename[resource->lr_type],
2796 lock->l_flags, nid, lock->l_remote_handle.cookie,
2797 exp ? atomic_read(&exp->exp_refcount) : -99,
2798 lock->l_pid, lock->l_callback_timeout,
2803 libcfs_debug_vmsg2(msgdata, fmt, args,
2804 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2805 "res: "DLDLMRES" rrc: %d type: %s flags: %#llx "
2806 "nid: %s remote: %#llx expref: %d pid: %u "
2807 "timeout: %lld lvb_type: %d\n",
2808 ldlm_lock_to_ns_name(lock),
2809 lock, lock->l_handle.h_cookie,
2810 atomic_read(&lock->l_refc),
2811 lock->l_readers, lock->l_writers,
2812 ldlm_lockname[lock->l_granted_mode],
2813 ldlm_lockname[lock->l_req_mode],
2815 atomic_read(&resource->lr_refcount),
2816 ldlm_typename[resource->lr_type],
2817 lock->l_flags, nid, lock->l_remote_handle.cookie,
2818 exp ? atomic_read(&exp->exp_refcount) : -99,
2819 lock->l_pid, lock->l_callback_timeout,
2824 ldlm_resource_putref(resource);
2826 EXPORT_SYMBOL(_ldlm_lock_debug);