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):
135 CERROR("Unknown intent 0x%08x\n", it);
139 EXPORT_SYMBOL(ldlm_it2str);
141 extern struct kmem_cache *ldlm_lock_slab;
143 #ifdef HAVE_SERVER_SUPPORT
144 static ldlm_processing_policy ldlm_processing_policy_table[] = {
145 [LDLM_PLAIN] = ldlm_process_plain_lock,
146 [LDLM_EXTENT] = ldlm_process_extent_lock,
147 [LDLM_FLOCK] = ldlm_process_flock_lock,
148 [LDLM_IBITS] = ldlm_process_inodebits_lock,
151 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
153 return ldlm_processing_policy_table[res->lr_type];
155 EXPORT_SYMBOL(ldlm_get_processing_policy);
156 #endif /* HAVE_SERVER_SUPPORT */
158 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
162 EXPORT_SYMBOL(ldlm_register_intent);
165 * REFCOUNTED LOCK OBJECTS
170 * Get a reference on a lock.
172 * Lock refcounts, during creation:
173 * - one special one for allocation, dec'd only once in destroy
174 * - one for being a lock that's in-use
175 * - one for the addref associated with a new lock
177 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
179 atomic_inc(&lock->l_refc);
182 EXPORT_SYMBOL(ldlm_lock_get);
185 * Release lock reference.
187 * Also frees the lock if it was last reference.
189 void ldlm_lock_put(struct ldlm_lock *lock)
193 LASSERT(lock->l_resource != LP_POISON);
194 LASSERT(atomic_read(&lock->l_refc) > 0);
195 if (atomic_dec_and_test(&lock->l_refc)) {
196 struct ldlm_resource *res;
199 "final lock_put on destroyed lock, freeing it.");
201 res = lock->l_resource;
202 LASSERT(ldlm_is_destroyed(lock));
203 LASSERT(list_empty(&lock->l_exp_list));
204 LASSERT(list_empty(&lock->l_res_link));
205 LASSERT(list_empty(&lock->l_pending_chain));
207 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
209 lu_ref_del(&res->lr_reference, "lock", lock);
210 ldlm_resource_putref(res);
211 lock->l_resource = NULL;
212 if (lock->l_export) {
213 class_export_lock_put(lock->l_export, lock);
214 lock->l_export = NULL;
217 if (lock->l_lvb_data != NULL)
218 OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
220 ldlm_interval_free(ldlm_interval_detach(lock));
221 lu_ref_fini(&lock->l_reference);
222 OBD_FREE_RCU(lock, sizeof(*lock), &lock->l_handle);
227 EXPORT_SYMBOL(ldlm_lock_put);
230 * Removes LDLM lock \a lock from LRU. Assumes LRU is already locked.
232 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
235 if (!list_empty(&lock->l_lru)) {
236 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
238 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
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 ldlm_clear_skipped(lock);
290 LASSERT(ns->ns_nr_unused >= 0);
295 * Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks
298 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
300 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
303 spin_lock(&ns->ns_lock);
304 ldlm_lock_add_to_lru_nolock(lock);
305 spin_unlock(&ns->ns_lock);
310 * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
311 * the LRU. Performs necessary LRU locking
313 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
315 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
318 if (ldlm_is_ns_srv(lock)) {
319 LASSERT(list_empty(&lock->l_lru));
324 spin_lock(&ns->ns_lock);
325 if (!list_empty(&lock->l_lru)) {
326 ldlm_lock_remove_from_lru_nolock(lock);
327 ldlm_lock_add_to_lru_nolock(lock);
329 spin_unlock(&ns->ns_lock);
334 * Helper to destroy a locked lock.
336 * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
337 * Must be called with l_lock and lr_lock held.
339 * Does not actually free the lock data, but rather marks the lock as
340 * destroyed by setting l_destroyed field in the lock to 1. Destroys a
341 * handle->lock association too, so that the lock can no longer be found
342 * and removes the lock from LRU list. Actual lock freeing occurs when
343 * last lock reference goes away.
345 * Original comment (of some historical value):
346 * This used to have a 'strict' flag, which recovery would use to mark an
347 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
348 * shall explain why it's gone: with the new hash table scheme, once you call
349 * ldlm_lock_destroy, you can never drop your final references on this lock.
350 * Because it's not in the hash table anymore. -phil
352 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
356 if (lock->l_readers || lock->l_writers) {
357 LDLM_ERROR(lock, "lock still has references");
361 if (!list_empty(&lock->l_res_link)) {
362 LDLM_ERROR(lock, "lock still on resource");
366 if (ldlm_is_destroyed(lock)) {
367 LASSERT(list_empty(&lock->l_lru));
371 ldlm_set_destroyed(lock);
373 if (lock->l_export && lock->l_export->exp_lock_hash) {
374 /* NB: it's safe to call cfs_hash_del() even lock isn't
375 * in exp_lock_hash. */
376 /* In the function below, .hs_keycmp resolves to
377 * ldlm_export_lock_keycmp() */
378 /* coverity[overrun-buffer-val] */
379 cfs_hash_del(lock->l_export->exp_lock_hash,
380 &lock->l_remote_handle, &lock->l_exp_hash);
383 ldlm_lock_remove_from_lru(lock);
384 class_handle_unhash(&lock->l_handle);
391 * Destroys a LDLM lock \a lock. Performs necessary locking first.
393 void ldlm_lock_destroy(struct ldlm_lock *lock)
397 lock_res_and_lock(lock);
398 first = ldlm_lock_destroy_internal(lock);
399 unlock_res_and_lock(lock);
401 /* drop reference from hashtable only for first destroy */
403 lu_ref_del(&lock->l_reference, "hash", lock);
404 LDLM_LOCK_RELEASE(lock);
410 * Destroys a LDLM lock \a lock that is already locked.
412 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
416 first = ldlm_lock_destroy_internal(lock);
417 /* drop reference from hashtable only for first destroy */
419 lu_ref_del(&lock->l_reference, "hash", lock);
420 LDLM_LOCK_RELEASE(lock);
425 /* this is called by portals_handle2object with the handle lock taken */
426 static void lock_handle_addref(void *lock)
428 LDLM_LOCK_GET((struct ldlm_lock *)lock);
431 static void lock_handle_free(void *lock, int size)
433 LASSERT(size == sizeof(struct ldlm_lock));
434 OBD_SLAB_FREE(lock, ldlm_lock_slab, size);
437 static struct portals_handle_ops lock_handle_ops = {
438 .hop_addref = lock_handle_addref,
439 .hop_free = lock_handle_free,
444 * Allocate and initialize new lock structure.
446 * usage: pass in a resource on which you have done ldlm_resource_get
447 * new lock will take over the refcount.
448 * returns: lock with refcount 2 - one for current caller and one for remote
450 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
452 struct ldlm_lock *lock;
455 if (resource == NULL)
458 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, GFP_NOFS);
462 spin_lock_init(&lock->l_lock);
463 lock->l_resource = resource;
464 lu_ref_add(&resource->lr_reference, "lock", lock);
466 atomic_set(&lock->l_refc, 2);
467 INIT_LIST_HEAD(&lock->l_res_link);
468 INIT_LIST_HEAD(&lock->l_lru);
469 INIT_LIST_HEAD(&lock->l_pending_chain);
470 INIT_LIST_HEAD(&lock->l_bl_ast);
471 INIT_LIST_HEAD(&lock->l_cp_ast);
472 INIT_LIST_HEAD(&lock->l_rk_ast);
473 init_waitqueue_head(&lock->l_waitq);
474 lock->l_blocking_lock = NULL;
475 INIT_LIST_HEAD(&lock->l_sl_mode);
476 INIT_LIST_HEAD(&lock->l_sl_policy);
477 INIT_HLIST_NODE(&lock->l_exp_hash);
478 INIT_HLIST_NODE(&lock->l_exp_flock_hash);
480 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
482 INIT_LIST_HEAD(&lock->l_handle.h_link);
483 class_handle_hash(&lock->l_handle, &lock_handle_ops);
485 lu_ref_init(&lock->l_reference);
486 lu_ref_add(&lock->l_reference, "hash", lock);
487 lock->l_callback_timeout = 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);
668 LASSERT(list_empty(&lock->l_bl_ast));
669 list_add(&lock->l_bl_ast, work_list);
671 LASSERT(lock->l_blocking_lock == NULL);
672 lock->l_blocking_lock = LDLM_LOCK_GET(new);
677 * Add a lock to list of just granted locks to send completion AST to.
679 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
680 struct list_head *work_list)
682 if (!ldlm_is_cp_reqd(lock)) {
683 ldlm_set_cp_reqd(lock);
684 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
685 LASSERT(list_empty(&lock->l_cp_ast));
686 list_add(&lock->l_cp_ast, work_list);
692 * Aggregator function to add AST work items into a list. Determines
693 * what sort of an AST work needs to be done and calls the proper
695 * Must be called with lr_lock held.
697 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
698 struct list_head *work_list)
701 check_res_locked(lock->l_resource);
703 ldlm_add_bl_work_item(lock, new, work_list);
705 ldlm_add_cp_work_item(lock, work_list);
710 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
711 * r/w reference type is determined by \a mode
712 * Calls ldlm_lock_addref_internal.
714 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
716 struct ldlm_lock *lock;
718 lock = ldlm_handle2lock(lockh);
719 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
720 ldlm_lock_addref_internal(lock, mode);
723 EXPORT_SYMBOL(ldlm_lock_addref);
727 * Add specified reader/writer reference to LDLM lock \a lock.
728 * r/w reference type is determined by \a mode
729 * Removes lock from LRU if it is there.
730 * Assumes the LDLM lock is already locked.
732 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
735 ldlm_lock_remove_from_lru(lock);
736 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
738 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
740 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
742 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
745 lu_ref_add_atomic(&lock->l_reference, "user", lock);
746 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
750 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
751 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
753 * \retval 0 success, lock was addref-ed
755 * \retval -EAGAIN lock is being canceled.
757 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
759 struct ldlm_lock *lock;
763 lock = ldlm_handle2lock(lockh);
765 lock_res_and_lock(lock);
766 if (lock->l_readers != 0 || lock->l_writers != 0 ||
767 !ldlm_is_cbpending(lock)) {
768 ldlm_lock_addref_internal_nolock(lock, mode);
771 unlock_res_and_lock(lock);
776 EXPORT_SYMBOL(ldlm_lock_addref_try);
779 * Add specified reader/writer reference to LDLM lock \a lock.
780 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
781 * Only called for local locks.
783 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
785 lock_res_and_lock(lock);
786 ldlm_lock_addref_internal_nolock(lock, mode);
787 unlock_res_and_lock(lock);
791 * Removes reader/writer reference for LDLM lock \a lock.
792 * Assumes LDLM lock is already locked.
793 * only called in ldlm_flock_destroy and for local locks.
794 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
795 * that cannot be placed in LRU.
797 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
800 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
801 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
802 LASSERT(lock->l_readers > 0);
803 lu_ref_del(&lock->l_reference, "reader", lock);
806 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
807 LASSERT(lock->l_writers > 0);
808 lu_ref_del(&lock->l_reference, "writer", lock);
812 lu_ref_del(&lock->l_reference, "user", lock);
813 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
817 * Removes reader/writer reference for LDLM lock \a lock.
818 * Locks LDLM lock first.
819 * If the lock is determined to be client lock on a client and r/w refcount
820 * drops to zero and the lock is not blocked, the lock is added to LRU lock
822 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
824 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
826 struct ldlm_namespace *ns;
829 lock_res_and_lock(lock);
831 ns = ldlm_lock_to_ns(lock);
833 ldlm_lock_decref_internal_nolock(lock, mode);
835 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
836 !lock->l_readers && !lock->l_writers) {
837 /* If this is a local lock on a server namespace and this was
838 * the last reference, cancel the lock.
840 * Group locks are special:
841 * They must not go in LRU, but they are not called back
842 * like non-group locks, instead they are manually released.
843 * They have an l_writers reference which they keep until
844 * they are manually released, so we remove them when they have
845 * no more reader or writer references. - LU-6368 */
846 ldlm_set_cbpending(lock);
849 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
850 /* If we received a blocked AST and this was the last reference,
851 * run the callback. */
852 if (ldlm_is_ns_srv(lock) && lock->l_export)
853 CERROR("FL_CBPENDING set on non-local lock--just a "
856 LDLM_DEBUG(lock, "final decref done on cbpending lock");
858 LDLM_LOCK_GET(lock); /* dropped by bl thread */
859 ldlm_lock_remove_from_lru(lock);
860 unlock_res_and_lock(lock);
862 if (ldlm_is_fail_loc(lock))
863 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
865 if (ldlm_is_atomic_cb(lock) ||
866 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
867 ldlm_handle_bl_callback(ns, NULL, lock);
868 } else if (ns_is_client(ns) &&
869 !lock->l_readers && !lock->l_writers &&
870 !ldlm_is_no_lru(lock) &&
871 !ldlm_is_bl_ast(lock)) {
873 LDLM_DEBUG(lock, "add lock into lru list");
875 /* If this is a client-side namespace and this was the last
876 * reference, put it on the LRU. */
877 ldlm_lock_add_to_lru(lock);
878 unlock_res_and_lock(lock);
880 if (ldlm_is_fail_loc(lock))
881 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
883 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
884 * are not supported by the server, otherwise, it is done on
886 if (!exp_connect_cancelset(lock->l_conn_export) &&
887 !ns_connect_lru_resize(ns))
888 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
890 LDLM_DEBUG(lock, "do not add lock into lru list");
891 unlock_res_and_lock(lock);
898 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
900 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
902 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
903 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
904 ldlm_lock_decref_internal(lock, mode);
907 EXPORT_SYMBOL(ldlm_lock_decref);
910 * Decrease reader/writer refcount for LDLM lock with handle
911 * \a lockh and mark it for subsequent cancellation once r/w refcount
912 * drops to zero instead of putting into LRU.
915 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
918 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
921 LASSERT(lock != NULL);
923 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
924 lock_res_and_lock(lock);
925 ldlm_set_cbpending(lock);
926 unlock_res_and_lock(lock);
927 ldlm_lock_decref_internal(lock, mode);
930 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
932 struct sl_insert_point {
933 struct list_head *res_link;
934 struct list_head *mode_link;
935 struct list_head *policy_link;
939 * Finds a position to insert the new lock into granted lock list.
941 * Used for locks eligible for skiplist optimization.
944 * queue [input]: the granted list where search acts on;
945 * req [input]: the lock whose position to be located;
946 * prev [output]: positions within 3 lists to insert @req to
950 * - ldlm_grant_lock_with_skiplist
952 static void search_granted_lock(struct list_head *queue,
953 struct ldlm_lock *req,
954 struct sl_insert_point *prev)
956 struct list_head *tmp;
957 struct ldlm_lock *lock, *mode_end, *policy_end;
960 list_for_each(tmp, queue) {
961 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
963 mode_end = list_entry(lock->l_sl_mode.prev,
964 struct ldlm_lock, l_sl_mode);
966 if (lock->l_req_mode != req->l_req_mode) {
967 /* jump to last lock of mode group */
968 tmp = &mode_end->l_res_link;
972 /* suitable mode group is found */
973 if (lock->l_resource->lr_type == LDLM_PLAIN) {
974 /* insert point is last lock of the mode group */
975 prev->res_link = &mode_end->l_res_link;
976 prev->mode_link = &mode_end->l_sl_mode;
977 prev->policy_link = &req->l_sl_policy;
980 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
983 list_entry(lock->l_sl_policy.prev,
987 if (lock->l_policy_data.l_inodebits.bits ==
988 req->l_policy_data.l_inodebits.bits) {
989 /* insert point is last lock of
990 * the policy group */
992 &policy_end->l_res_link;
994 &policy_end->l_sl_mode;
996 &policy_end->l_sl_policy;
1001 if (policy_end == mode_end)
1002 /* done with mode group */
1005 /* go to next policy group within mode group */
1006 tmp = policy_end->l_res_link.next;
1007 lock = list_entry(tmp, struct ldlm_lock,
1009 } /* loop over policy groups within the mode group */
1011 /* insert point is last lock of the mode group,
1012 * new policy group is started */
1013 prev->res_link = &mode_end->l_res_link;
1014 prev->mode_link = &mode_end->l_sl_mode;
1015 prev->policy_link = &req->l_sl_policy;
1019 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1024 /* insert point is last lock on the queue,
1025 * new mode group and new policy group are started */
1026 prev->res_link = queue->prev;
1027 prev->mode_link = &req->l_sl_mode;
1028 prev->policy_link = &req->l_sl_policy;
1034 * Add a lock into resource granted list after a position described by
1037 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1038 struct sl_insert_point *prev)
1040 struct ldlm_resource *res = lock->l_resource;
1043 check_res_locked(res);
1045 ldlm_resource_dump(D_INFO, res);
1046 LDLM_DEBUG(lock, "About to add lock:");
1048 if (ldlm_is_destroyed(lock)) {
1049 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1053 LASSERT(list_empty(&lock->l_res_link));
1054 LASSERT(list_empty(&lock->l_sl_mode));
1055 LASSERT(list_empty(&lock->l_sl_policy));
1058 * lock->link == prev->link means lock is first starting the group.
1059 * Don't re-add to itself to suppress kernel warnings.
1061 if (&lock->l_res_link != prev->res_link)
1062 list_add(&lock->l_res_link, prev->res_link);
1063 if (&lock->l_sl_mode != prev->mode_link)
1064 list_add(&lock->l_sl_mode, prev->mode_link);
1065 if (&lock->l_sl_policy != prev->policy_link)
1066 list_add(&lock->l_sl_policy, prev->policy_link);
1072 * Add a lock to granted list on a resource maintaining skiplist
1075 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1077 struct sl_insert_point prev;
1079 LASSERT(lock->l_req_mode == lock->l_granted_mode);
1081 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1082 ldlm_granted_list_add_lock(lock, &prev);
1086 * Perform lock granting bookkeeping.
1088 * Includes putting the lock into granted list and updating lock mode.
1090 * - ldlm_lock_enqueue
1091 * - ldlm_reprocess_queue
1093 * must be called with lr_lock held
1095 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1097 struct ldlm_resource *res = lock->l_resource;
1100 check_res_locked(res);
1102 lock->l_granted_mode = lock->l_req_mode;
1104 if (work_list && lock->l_completion_ast != NULL)
1105 ldlm_add_ast_work_item(lock, NULL, work_list);
1107 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1108 ldlm_grant_lock_with_skiplist(lock);
1109 else if (res->lr_type == LDLM_EXTENT)
1110 ldlm_extent_add_lock(res, lock);
1111 else if (res->lr_type == LDLM_FLOCK) {
1112 /* We should not add locks to granted list in the following
1114 * - this is an UNLOCK but not a real lock;
1115 * - this is a TEST lock;
1116 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1117 * - this is a deadlock (flock cannot be granted) */
1118 if (lock->l_req_mode == 0 ||
1119 lock->l_req_mode == LCK_NL ||
1120 ldlm_is_test_lock(lock) ||
1121 ldlm_is_flock_deadlock(lock))
1123 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1128 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1133 * Describe the overlap between two locks. itree_overlap_cb data.
1135 struct lock_match_data {
1136 struct ldlm_lock *lmd_old;
1137 struct ldlm_lock *lmd_lock;
1138 enum ldlm_mode *lmd_mode;
1139 union ldlm_policy_data *lmd_policy;
1145 * Check if the given @lock meets the criteria for a match.
1146 * A reference on the lock is taken if matched.
1148 * \param lock test-against this lock
1149 * \param data parameters
1151 static int lock_matches(struct ldlm_lock *lock, struct lock_match_data *data)
1153 union ldlm_policy_data *lpol = &lock->l_policy_data;
1154 enum ldlm_mode match;
1156 if (lock == data->lmd_old)
1157 return INTERVAL_ITER_STOP;
1159 /* Check if this lock can be matched.
1160 * Used by LU-2919(exclusive open) for open lease lock */
1161 if (ldlm_is_excl(lock))
1162 return INTERVAL_ITER_CONT;
1164 /* llite sometimes wants to match locks that will be
1165 * canceled when their users drop, but we allow it to match
1166 * if it passes in CBPENDING and the lock still has users.
1167 * this is generally only going to be used by children
1168 * whose parents already hold a lock so forward progress
1169 * can still happen. */
1170 if (ldlm_is_cbpending(lock) &&
1171 !(data->lmd_flags & LDLM_FL_CBPENDING))
1172 return INTERVAL_ITER_CONT;
1173 if (!data->lmd_unref && ldlm_is_cbpending(lock) &&
1174 lock->l_readers == 0 && lock->l_writers == 0)
1175 return INTERVAL_ITER_CONT;
1177 if (!(lock->l_req_mode & *data->lmd_mode))
1178 return INTERVAL_ITER_CONT;
1179 match = lock->l_req_mode;
1181 switch (lock->l_resource->lr_type) {
1183 if (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1184 lpol->l_extent.end < data->lmd_policy->l_extent.end)
1185 return INTERVAL_ITER_CONT;
1187 if (unlikely(match == LCK_GROUP) &&
1188 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1189 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1190 return INTERVAL_ITER_CONT;
1193 /* We match if we have existing lock with same or wider set
1195 if ((lpol->l_inodebits.bits &
1196 data->lmd_policy->l_inodebits.bits) !=
1197 data->lmd_policy->l_inodebits.bits)
1198 return INTERVAL_ITER_CONT;
1204 /* We match if we have existing lock with same or wider set
1206 if (!data->lmd_unref && LDLM_HAVE_MASK(lock, GONE))
1207 return INTERVAL_ITER_CONT;
1209 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1210 return INTERVAL_ITER_CONT;
1212 if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1213 LDLM_LOCK_GET(lock);
1214 ldlm_lock_touch_in_lru(lock);
1216 ldlm_lock_addref_internal_nolock(lock, match);
1219 *data->lmd_mode = match;
1220 data->lmd_lock = lock;
1222 return INTERVAL_ITER_STOP;
1225 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1227 struct ldlm_interval *node = to_ldlm_interval(in);
1228 struct lock_match_data *data = args;
1229 struct ldlm_lock *lock;
1232 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1233 rc = lock_matches(lock, data);
1234 if (rc == INTERVAL_ITER_STOP)
1235 return INTERVAL_ITER_STOP;
1237 return INTERVAL_ITER_CONT;
1241 * Search for a lock with given parameters in interval trees.
1243 * \param res search for a lock in this resource
1244 * \param data parameters
1246 * \retval a referenced lock or NULL.
1248 static struct ldlm_lock *search_itree(struct ldlm_resource *res,
1249 struct lock_match_data *data)
1251 struct interval_node_extent ext = {
1252 .start = data->lmd_policy->l_extent.start,
1253 .end = data->lmd_policy->l_extent.end
1257 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1258 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1260 if (tree->lit_root == NULL)
1263 if (!(tree->lit_mode & *data->lmd_mode))
1266 interval_search(tree->lit_root, &ext,
1267 itree_overlap_cb, data);
1269 return data->lmd_lock;
1274 * Search for a lock with given properties in a queue.
1276 * \param queue search for a lock in this queue
1277 * \param data parameters
1279 * \retval a referenced lock or NULL.
1281 static struct ldlm_lock *search_queue(struct list_head *queue,
1282 struct lock_match_data *data)
1284 struct ldlm_lock *lock;
1287 list_for_each_entry(lock, queue, l_res_link) {
1288 rc = lock_matches(lock, data);
1289 if (rc == INTERVAL_ITER_STOP)
1290 return data->lmd_lock;
1295 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1297 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1298 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1299 wake_up_all(&lock->l_waitq);
1302 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1304 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1306 lock_res_and_lock(lock);
1307 ldlm_lock_fail_match_locked(lock);
1308 unlock_res_and_lock(lock);
1312 * Mark lock as "matchable" by OST.
1314 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1316 * Assumes LDLM lock is already locked.
1318 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1320 ldlm_set_lvb_ready(lock);
1321 wake_up_all(&lock->l_waitq);
1323 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1326 * Mark lock as "matchable" by OST.
1327 * Locks the lock and then \see ldlm_lock_allow_match_locked
1329 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1331 lock_res_and_lock(lock);
1332 ldlm_lock_allow_match_locked(lock);
1333 unlock_res_and_lock(lock);
1335 EXPORT_SYMBOL(ldlm_lock_allow_match);
1338 * Attempt to find a lock with specified properties.
1340 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1343 * Can be called in two ways:
1345 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1346 * for a duplicate of.
1348 * Otherwise, all of the fields must be filled in, to match against.
1350 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1351 * server (ie, connh is NULL)
1352 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1353 * list will be considered
1354 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1355 * to be canceled can still be matched as long as they still have reader
1356 * or writer refernces
1357 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1358 * just tell us if we would have matched.
1360 * \retval 1 if it finds an already-existing lock that is compatible; in this
1361 * case, lockh is filled in with a addref()ed lock
1363 * We also check security context, and if that fails we simply return 0 (to
1364 * keep caller code unchanged), the context failure will be discovered by
1365 * caller sometime later.
1367 enum ldlm_mode ldlm_lock_match(struct ldlm_namespace *ns, __u64 flags,
1368 const struct ldlm_res_id *res_id,
1369 enum ldlm_type type,
1370 union ldlm_policy_data *policy,
1371 enum ldlm_mode mode,
1372 struct lustre_handle *lockh, int unref)
1374 struct lock_match_data data = {
1378 .lmd_policy = policy,
1382 struct ldlm_resource *res;
1383 struct ldlm_lock *lock;
1388 data.lmd_old = ldlm_handle2lock(lockh);
1389 LASSERT(data.lmd_old != NULL);
1391 ns = ldlm_lock_to_ns(data.lmd_old);
1392 res_id = &data.lmd_old->l_resource->lr_name;
1393 type = data.lmd_old->l_resource->lr_type;
1394 *data.lmd_mode = data.lmd_old->l_req_mode;
1397 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1399 LASSERT(data.lmd_old == NULL);
1403 LDLM_RESOURCE_ADDREF(res);
1406 if (res->lr_type == LDLM_EXTENT)
1407 lock = search_itree(res, &data);
1409 lock = search_queue(&res->lr_granted, &data);
1412 if (flags & LDLM_FL_BLOCK_GRANTED)
1414 lock = search_queue(&res->lr_waiting, &data);
1421 LDLM_RESOURCE_DELREF(res);
1422 ldlm_resource_putref(res);
1425 ldlm_lock2handle(lock, lockh);
1426 if ((flags & LDLM_FL_LVB_READY) &&
1427 (!ldlm_is_lvb_ready(lock))) {
1428 __u64 wait_flags = LDLM_FL_LVB_READY |
1429 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1430 struct l_wait_info lwi;
1431 if (lock->l_completion_ast) {
1432 int err = lock->l_completion_ast(lock,
1433 LDLM_FL_WAIT_NOREPROC,
1436 if (flags & LDLM_FL_TEST_LOCK)
1437 LDLM_LOCK_RELEASE(lock);
1439 ldlm_lock_decref_internal(lock,
1446 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1447 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1449 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1450 l_wait_event(lock->l_waitq,
1451 lock->l_flags & wait_flags,
1453 if (!ldlm_is_lvb_ready(lock)) {
1454 if (flags & LDLM_FL_TEST_LOCK)
1455 LDLM_LOCK_RELEASE(lock);
1457 ldlm_lock_decref_internal(lock, mode);
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);
1470 /* check user's security context */
1471 if (lock->l_conn_export &&
1472 sptlrpc_import_check_ctx(
1473 class_exp2cliimp(lock->l_conn_export))) {
1474 if (!(flags & LDLM_FL_TEST_LOCK))
1475 ldlm_lock_decref_internal(lock, mode);
1479 if (flags & LDLM_FL_TEST_LOCK)
1480 LDLM_LOCK_RELEASE(lock);
1482 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1483 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1484 "%llu/%llu (%llu %llu)", ns,
1485 type, mode, res_id->name[0], res_id->name[1],
1486 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1487 res_id->name[2] :policy->l_extent.start,
1488 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1489 res_id->name[3] : policy->l_extent.end);
1491 if (data.lmd_old != NULL)
1492 LDLM_LOCK_PUT(data.lmd_old);
1494 return rc ? mode : 0;
1496 EXPORT_SYMBOL(ldlm_lock_match);
1498 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1501 struct ldlm_lock *lock;
1502 enum ldlm_mode mode = 0;
1505 lock = ldlm_handle2lock(lockh);
1507 lock_res_and_lock(lock);
1508 if (LDLM_HAVE_MASK(lock, GONE))
1511 if (ldlm_is_cbpending(lock) &&
1512 lock->l_readers == 0 && lock->l_writers == 0)
1516 *bits = lock->l_policy_data.l_inodebits.bits;
1517 mode = lock->l_granted_mode;
1518 ldlm_lock_addref_internal_nolock(lock, mode);
1525 unlock_res_and_lock(lock);
1526 LDLM_LOCK_PUT(lock);
1530 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1532 /** The caller must guarantee that the buffer is large enough. */
1533 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1534 enum req_location loc, void *data, int size)
1539 LASSERT(data != NULL);
1542 switch (lock->l_lvb_type) {
1544 if (size == sizeof(struct ost_lvb)) {
1545 if (loc == RCL_CLIENT)
1546 lvb = req_capsule_client_swab_get(pill,
1548 lustre_swab_ost_lvb);
1550 lvb = req_capsule_server_swab_get(pill,
1552 lustre_swab_ost_lvb);
1553 if (unlikely(lvb == NULL)) {
1554 LDLM_ERROR(lock, "no LVB");
1558 memcpy(data, lvb, size);
1559 } else if (size == sizeof(struct ost_lvb_v1)) {
1560 struct ost_lvb *olvb = data;
1562 if (loc == RCL_CLIENT)
1563 lvb = req_capsule_client_swab_get(pill,
1565 lustre_swab_ost_lvb_v1);
1567 lvb = req_capsule_server_sized_swab_get(pill,
1569 lustre_swab_ost_lvb_v1);
1570 if (unlikely(lvb == NULL)) {
1571 LDLM_ERROR(lock, "no LVB");
1575 memcpy(data, lvb, size);
1576 olvb->lvb_mtime_ns = 0;
1577 olvb->lvb_atime_ns = 0;
1578 olvb->lvb_ctime_ns = 0;
1580 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1586 if (size == sizeof(struct lquota_lvb)) {
1587 if (loc == RCL_CLIENT)
1588 lvb = req_capsule_client_swab_get(pill,
1590 lustre_swab_lquota_lvb);
1592 lvb = req_capsule_server_swab_get(pill,
1594 lustre_swab_lquota_lvb);
1595 if (unlikely(lvb == NULL)) {
1596 LDLM_ERROR(lock, "no LVB");
1600 memcpy(data, lvb, size);
1602 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1611 if (loc == RCL_CLIENT)
1612 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1614 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1615 if (unlikely(lvb == NULL)) {
1616 LDLM_ERROR(lock, "no LVB");
1620 memcpy(data, lvb, size);
1623 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1624 libcfs_debug_dumpstack(NULL);
1632 * Create and fill in new LDLM lock with specified properties.
1633 * Returns a referenced lock
1635 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1636 const struct ldlm_res_id *res_id,
1637 enum ldlm_type type,
1638 enum ldlm_mode mode,
1639 const struct ldlm_callback_suite *cbs,
1640 void *data, __u32 lvb_len,
1641 enum lvb_type lvb_type)
1643 struct ldlm_lock *lock;
1644 struct ldlm_resource *res;
1648 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1650 RETURN(ERR_CAST(res));
1652 lock = ldlm_lock_new(res);
1654 RETURN(ERR_PTR(-ENOMEM));
1656 lock->l_req_mode = mode;
1657 lock->l_ast_data = data;
1658 lock->l_pid = current_pid();
1659 if (ns_is_server(ns))
1660 ldlm_set_ns_srv(lock);
1662 lock->l_blocking_ast = cbs->lcs_blocking;
1663 lock->l_completion_ast = cbs->lcs_completion;
1664 lock->l_glimpse_ast = cbs->lcs_glimpse;
1667 lock->l_tree_node = NULL;
1668 /* if this is the extent lock, allocate the interval tree node */
1669 if (type == LDLM_EXTENT)
1670 if (ldlm_interval_alloc(lock) == NULL)
1671 GOTO(out, rc = -ENOMEM);
1674 lock->l_lvb_len = lvb_len;
1675 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1676 if (lock->l_lvb_data == NULL)
1677 GOTO(out, rc = -ENOMEM);
1680 lock->l_lvb_type = lvb_type;
1681 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1682 GOTO(out, rc = -ENOENT);
1687 ldlm_lock_destroy(lock);
1688 LDLM_LOCK_RELEASE(lock);
1689 RETURN(ERR_PTR(rc));
1692 #ifdef HAVE_SERVER_SUPPORT
1693 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1696 struct ldlm_resource *res = lock->l_resource;
1697 enum ldlm_error rc = ELDLM_OK;
1698 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1699 ldlm_processing_policy policy;
1702 policy = ldlm_processing_policy_table[res->lr_type];
1704 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1705 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1706 res->lr_type != LDLM_FLOCK) {
1707 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1708 if (rc == -ERESTART)
1712 if (!list_empty(&rpc_list))
1713 ldlm_discard_bl_list(&rpc_list);
1720 * Enqueue (request) a lock.
1722 * Does not block. As a result of enqueue the lock would be put
1723 * into granted or waiting list.
1725 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1726 * set, skip all the enqueueing and delegate lock processing to intent policy
1729 enum ldlm_error ldlm_lock_enqueue(struct ldlm_namespace *ns,
1730 struct ldlm_lock **lockp,
1731 void *cookie, __u64 *flags)
1733 struct ldlm_lock *lock = *lockp;
1734 struct ldlm_resource *res = lock->l_resource;
1735 int local = ns_is_client(ldlm_res_to_ns(res));
1736 enum ldlm_error rc = ELDLM_OK;
1737 struct ldlm_interval *node = NULL;
1740 /* policies are not executed on the client or during replay */
1741 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1742 && !local && ns->ns_policy) {
1743 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1745 if (rc == ELDLM_LOCK_REPLACED) {
1746 /* The lock that was returned has already been granted,
1747 * and placed into lockp. If it's not the same as the
1748 * one we passed in, then destroy the old one and our
1749 * work here is done. */
1750 if (lock != *lockp) {
1751 ldlm_lock_destroy(lock);
1752 LDLM_LOCK_RELEASE(lock);
1754 *flags |= LDLM_FL_LOCK_CHANGED;
1756 } else if (rc != ELDLM_OK &&
1757 lock->l_req_mode == lock->l_granted_mode) {
1758 LASSERT(*flags & LDLM_FL_RESENT);
1759 /* It may happen that ns_policy returns an error in
1760 * resend case, object may be unlinked or just some
1761 * error occurs. It is unclear if lock reached the
1762 * client in the original reply, just leave the lock on
1763 * server, not returning it again to client. Due to
1764 * LU-6529, the server will not OOM. */
1766 } else if (rc != ELDLM_OK ||
1767 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1768 ldlm_lock_destroy(lock);
1773 if (*flags & LDLM_FL_RESENT) {
1774 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1775 * Set LOCK_CHANGED always.
1776 * Check if the lock is granted for BLOCK_GRANTED.
1777 * Take NO_TIMEOUT from the lock as it is inherited through
1778 * LDLM_FL_INHERIT_MASK */
1779 *flags |= LDLM_FL_LOCK_CHANGED;
1780 if (lock->l_req_mode != lock->l_granted_mode)
1781 *flags |= LDLM_FL_BLOCK_GRANTED;
1782 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1786 /* For a replaying lock, it might be already in granted list. So
1787 * unlinking the lock will cause the interval node to be freed, we
1788 * have to allocate the interval node early otherwise we can't regrant
1789 * this lock in the future. - jay */
1790 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1791 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1793 lock_res_and_lock(lock);
1794 if (local && lock->l_req_mode == lock->l_granted_mode) {
1795 /* The server returned a blocked lock, but it was granted
1796 * before we got a chance to actually enqueue it. We don't
1797 * need to do anything else. */
1798 *flags &= ~LDLM_FL_BLOCKED_MASK;
1799 GOTO(out, rc = ELDLM_OK);
1802 ldlm_resource_unlink_lock(lock);
1803 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1805 ldlm_lock_destroy_nolock(lock);
1806 GOTO(out, rc = -ENOMEM);
1809 INIT_LIST_HEAD(&node->li_group);
1810 ldlm_interval_attach(node, lock);
1814 /* Some flags from the enqueue want to make it into the AST, via the
1815 * lock's l_flags. */
1816 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1817 ldlm_set_ast_discard_data(lock);
1818 if (*flags & LDLM_FL_TEST_LOCK)
1819 ldlm_set_test_lock(lock);
1820 if (*flags & LDLM_FL_COS_INCOMPAT)
1821 ldlm_set_cos_incompat(lock);
1822 if (*flags & LDLM_FL_COS_ENABLED)
1823 ldlm_set_cos_enabled(lock);
1825 /* This distinction between local lock trees is very important; a client
1826 * namespace only has information about locks taken by that client, and
1827 * thus doesn't have enough information to decide for itself if it can
1828 * be granted (below). In this case, we do exactly what the server
1829 * tells us to do, as dictated by the 'flags'.
1831 * We do exactly the same thing during recovery, when the server is
1832 * more or less trusting the clients not to lie.
1834 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1837 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1838 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1840 ldlm_grant_lock(lock, NULL);
1841 GOTO(out, rc = ELDLM_OK);
1842 #ifdef HAVE_SERVER_SUPPORT
1843 } else if (*flags & LDLM_FL_REPLAY) {
1844 if (*flags & LDLM_FL_BLOCK_WAIT) {
1845 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1846 GOTO(out, rc = ELDLM_OK);
1847 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1848 ldlm_grant_lock(lock, NULL);
1849 GOTO(out, rc = ELDLM_OK);
1851 /* If no flags, fall through to normal enqueue path. */
1854 rc = ldlm_lock_enqueue_helper(lock, flags);
1858 CERROR("This is client-side-only module, cannot handle "
1859 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1865 unlock_res_and_lock(lock);
1867 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1871 #ifdef HAVE_SERVER_SUPPORT
1873 * Iterate through all waiting locks on a given resource queue and attempt to
1876 * Must be called with resource lock held.
1878 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1879 struct list_head *work_list,
1880 enum ldlm_process_intention intention)
1882 struct list_head *tmp, *pos;
1883 ldlm_processing_policy policy;
1885 int rc = LDLM_ITER_CONTINUE;
1886 enum ldlm_error err;
1887 struct list_head bl_ast_list = LIST_HEAD_INIT(bl_ast_list);
1890 check_res_locked(res);
1892 policy = ldlm_processing_policy_table[res->lr_type];
1894 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1895 intention == LDLM_PROCESS_RECOVERY);
1898 list_for_each_safe(tmp, pos, queue) {
1899 struct ldlm_lock *pending;
1900 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1902 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1904 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1907 rc = policy(pending, &flags, intention, &err, &rpc_list);
1908 if (pending->l_granted_mode == pending->l_req_mode ||
1909 res->lr_type == LDLM_FLOCK) {
1910 list_splice(&rpc_list, work_list);
1912 list_splice(&rpc_list, &bl_ast_list);
1915 * When this is called from recovery done, we always want
1916 * to scan the whole list no matter what 'rc' is returned.
1918 if (rc != LDLM_ITER_CONTINUE &&
1919 intention == LDLM_PROCESS_RESCAN)
1923 if (!list_empty(&bl_ast_list)) {
1926 LASSERT(intention == LDLM_PROCESS_RECOVERY);
1928 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
1932 if (rc == -ERESTART)
1936 if (!list_empty(&bl_ast_list))
1937 ldlm_discard_bl_list(&bl_ast_list);
1939 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
1943 * Conflicting locks are detected for a lock to be enqueued, add the lock
1944 * into waiting list and send blocking ASTs to the conflicting locks.
1946 * \param[in] lock The lock to be enqueued.
1947 * \param[out] flags Lock flags for the lock to be enqueued.
1948 * \param[in] rpc_list Conflicting locks list.
1950 * \retval -ERESTART: Some lock was instantly canceled while sending
1951 * blocking ASTs, caller needs to re-check conflicting
1953 * \retval -EAGAIN: Lock was destroyed, caller should return error.
1954 * \reval 0: Lock is successfully added in waiting list.
1956 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
1957 struct list_head *rpc_list)
1959 struct ldlm_resource *res = lock->l_resource;
1963 check_res_locked(res);
1965 /* If either of the compat_queue()s returned failure, then we
1966 * have ASTs to send and must go onto the waiting list.
1968 * bug 2322: we used to unlink and re-add here, which was a
1969 * terrible folly -- if we goto restart, we could get
1970 * re-ordered! Causes deadlock, because ASTs aren't sent! */
1971 if (list_empty(&lock->l_res_link))
1972 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1975 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
1978 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
1979 !ns_is_client(ldlm_res_to_ns(res)))
1980 class_fail_export(lock->l_export);
1983 if (rc == -ERESTART) {
1984 /* 15715: The lock was granted and destroyed after
1985 * resource lock was dropped. Interval node was freed
1986 * in ldlm_lock_destroy. Anyway, this always happens
1987 * when a client is being evicted. So it would be
1988 * ok to return an error. -jay */
1989 if (ldlm_is_destroyed(lock))
1992 /* lock was granted while resource was unlocked. */
1993 if (lock->l_granted_mode == lock->l_req_mode) {
1994 /* bug 11300: if the lock has been granted,
1995 * break earlier because otherwise, we will go
1996 * to restart and ldlm_resource_unlink will be
1997 * called and it causes the interval node to be
1998 * freed. Then we will fail at
1999 * ldlm_extent_add_lock() */
2000 *flags &= ~LDLM_FL_BLOCKED_MASK;
2006 *flags |= LDLM_FL_BLOCK_GRANTED;
2012 * Discard all AST work items from list.
2014 * If for whatever reason we do not want to send ASTs to conflicting locks
2015 * anymore, disassemble the list with this function.
2017 void ldlm_discard_bl_list(struct list_head *bl_list)
2019 struct list_head *tmp, *pos;
2022 list_for_each_safe(pos, tmp, bl_list) {
2023 struct ldlm_lock *lock =
2024 list_entry(pos, struct ldlm_lock, l_bl_ast);
2026 list_del_init(&lock->l_bl_ast);
2027 LASSERT(ldlm_is_ast_sent(lock));
2028 ldlm_clear_ast_sent(lock);
2029 LASSERT(lock->l_bl_ast_run == 0);
2030 LASSERT(lock->l_blocking_lock);
2031 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
2032 lock->l_blocking_lock = NULL;
2033 LDLM_LOCK_RELEASE(lock);
2041 * Process a call to blocking AST callback for a lock in ast_work list
2044 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2046 struct ldlm_cb_set_arg *arg = opaq;
2047 struct ldlm_lock_desc d;
2049 struct ldlm_lock *lock;
2052 if (list_empty(arg->list))
2055 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2057 /* nobody should touch l_bl_ast */
2058 lock_res_and_lock(lock);
2059 list_del_init(&lock->l_bl_ast);
2061 LASSERT(ldlm_is_ast_sent(lock));
2062 LASSERT(lock->l_bl_ast_run == 0);
2063 LASSERT(lock->l_blocking_lock);
2064 lock->l_bl_ast_run++;
2065 unlock_res_and_lock(lock);
2067 ldlm_lock2desc(lock->l_blocking_lock, &d);
2069 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2070 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
2071 lock->l_blocking_lock = NULL;
2072 LDLM_LOCK_RELEASE(lock);
2078 * Process a call to completion AST callback for a lock in ast_work list
2081 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2083 struct ldlm_cb_set_arg *arg = opaq;
2085 struct ldlm_lock *lock;
2086 ldlm_completion_callback completion_callback;
2089 if (list_empty(arg->list))
2092 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2094 /* It's possible to receive a completion AST before we've set
2095 * the l_completion_ast pointer: either because the AST arrived
2096 * before the reply, or simply because there's a small race
2097 * window between receiving the reply and finishing the local
2098 * enqueue. (bug 842)
2100 * This can't happen with the blocking_ast, however, because we
2101 * will never call the local blocking_ast until we drop our
2102 * reader/writer reference, which we won't do until we get the
2103 * reply and finish enqueueing. */
2105 /* nobody should touch l_cp_ast */
2106 lock_res_and_lock(lock);
2107 list_del_init(&lock->l_cp_ast);
2108 LASSERT(ldlm_is_cp_reqd(lock));
2109 /* save l_completion_ast since it can be changed by
2110 * mds_intent_policy(), see bug 14225 */
2111 completion_callback = lock->l_completion_ast;
2112 ldlm_clear_cp_reqd(lock);
2113 unlock_res_and_lock(lock);
2115 if (completion_callback != NULL)
2116 rc = completion_callback(lock, 0, (void *)arg);
2117 LDLM_LOCK_RELEASE(lock);
2123 * Process a call to revocation AST callback for a lock in ast_work list
2126 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2128 struct ldlm_cb_set_arg *arg = opaq;
2129 struct ldlm_lock_desc desc;
2131 struct ldlm_lock *lock;
2134 if (list_empty(arg->list))
2137 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2138 list_del_init(&lock->l_rk_ast);
2140 /* the desc just pretend to exclusive */
2141 ldlm_lock2desc(lock, &desc);
2142 desc.l_req_mode = LCK_EX;
2143 desc.l_granted_mode = 0;
2145 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2146 LDLM_LOCK_RELEASE(lock);
2152 * Process a call to glimpse AST callback for a lock in ast_work list
2154 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2156 struct ldlm_cb_set_arg *arg = opaq;
2157 struct ldlm_glimpse_work *gl_work;
2158 struct ldlm_lock *lock;
2162 if (list_empty(arg->list))
2165 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2167 list_del_init(&gl_work->gl_list);
2169 lock = gl_work->gl_lock;
2171 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2172 arg->gl_desc = gl_work->gl_desc;
2173 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2174 arg->gl_interpret_data = gl_work->gl_interpret_data;
2176 /* invoke the actual glimpse callback */
2177 if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
2180 LDLM_LOCK_RELEASE(lock);
2181 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2182 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2184 OBD_FREE_PTR(gl_work);
2190 * Process list of locks in need of ASTs being sent.
2192 * Used on server to send multiple ASTs together instead of sending one by
2195 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2196 ldlm_desc_ast_t ast_type)
2198 struct ldlm_cb_set_arg *arg;
2199 set_producer_func work_ast_lock;
2202 if (list_empty(rpc_list))
2209 atomic_set(&arg->restart, 0);
2210 arg->list = rpc_list;
2213 case LDLM_WORK_BL_AST:
2214 arg->type = LDLM_BL_CALLBACK;
2215 work_ast_lock = ldlm_work_bl_ast_lock;
2217 case LDLM_WORK_CP_AST:
2218 arg->type = LDLM_CP_CALLBACK;
2219 work_ast_lock = ldlm_work_cp_ast_lock;
2221 case LDLM_WORK_REVOKE_AST:
2222 arg->type = LDLM_BL_CALLBACK;
2223 work_ast_lock = ldlm_work_revoke_ast_lock;
2225 case LDLM_WORK_GL_AST:
2226 arg->type = LDLM_GL_CALLBACK;
2227 work_ast_lock = ldlm_work_gl_ast_lock;
2233 /* We create a ptlrpc request set with flow control extension.
2234 * This request set will use the work_ast_lock function to produce new
2235 * requests and will send a new request each time one completes in order
2236 * to keep the number of requests in flight to ns_max_parallel_ast */
2237 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2238 work_ast_lock, arg);
2239 if (arg->set == NULL)
2240 GOTO(out, rc = -ENOMEM);
2242 ptlrpc_set_wait(arg->set);
2243 ptlrpc_set_destroy(arg->set);
2245 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2253 * Try to grant all waiting locks on a resource.
2255 * Calls ldlm_reprocess_queue on waiting queue.
2257 * Typically called after some resource locks are cancelled to see
2258 * if anything could be granted as a result of the cancellation.
2260 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2261 enum ldlm_process_intention intention)
2263 struct list_head rpc_list;
2264 #ifdef HAVE_SERVER_SUPPORT
2265 struct obd_device *obd;
2269 INIT_LIST_HEAD(&rpc_list);
2270 /* Local lock trees don't get reprocessed. */
2271 if (ns_is_client(ldlm_res_to_ns(res))) {
2276 /* Disable reprocess during lock replay stage but allow during
2277 * request replay stage.
2279 obd = ldlm_res_to_ns(res)->ns_obd;
2280 if (obd->obd_recovering &&
2281 atomic_read(&obd->obd_req_replay_clients) == 0)
2285 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list, intention);
2288 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2290 if (rc == -ERESTART) {
2291 LASSERT(list_empty(&rpc_list));
2297 INIT_LIST_HEAD(&rpc_list);
2298 if (!ns_is_client(ldlm_res_to_ns(res))) {
2299 CERROR("This is client-side-only module, cannot handle "
2300 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2307 void ldlm_reprocess_all(struct ldlm_resource *res)
2309 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN);
2311 EXPORT_SYMBOL(ldlm_reprocess_all);
2313 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2314 struct hlist_node *hnode, void *arg)
2316 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2318 /* This is only called once after recovery done. LU-8306. */
2319 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY);
2324 * Iterate through all resources on a namespace attempting to grant waiting
2327 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2332 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2333 ldlm_reprocess_res, NULL, 0);
2339 * Helper function to call blocking AST for LDLM lock \a lock in a
2340 * "cancelling" mode.
2342 void ldlm_cancel_callback(struct ldlm_lock *lock)
2344 check_res_locked(lock->l_resource);
2345 if (!ldlm_is_cancel(lock)) {
2346 ldlm_set_cancel(lock);
2347 if (lock->l_blocking_ast) {
2348 unlock_res_and_lock(lock);
2349 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2351 lock_res_and_lock(lock);
2353 LDLM_DEBUG(lock, "no blocking ast");
2356 /* only canceller can set bl_done bit */
2357 ldlm_set_bl_done(lock);
2358 wake_up_all(&lock->l_waitq);
2359 } else if (!ldlm_is_bl_done(lock)) {
2360 struct l_wait_info lwi = { 0 };
2362 /* The lock is guaranteed to have been canceled once
2363 * returning from this function. */
2364 unlock_res_and_lock(lock);
2365 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
2366 lock_res_and_lock(lock);
2371 * Remove skiplist-enabled LDLM lock \a req from granted list
2373 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2375 if (req->l_resource->lr_type != LDLM_PLAIN &&
2376 req->l_resource->lr_type != LDLM_IBITS)
2379 list_del_init(&req->l_sl_policy);
2380 list_del_init(&req->l_sl_mode);
2384 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2386 void ldlm_lock_cancel(struct ldlm_lock *lock)
2388 struct ldlm_resource *res;
2389 struct ldlm_namespace *ns;
2392 lock_res_and_lock(lock);
2394 res = lock->l_resource;
2395 ns = ldlm_res_to_ns(res);
2397 /* Please do not, no matter how tempting, remove this LBUG without
2398 * talking to me first. -phik */
2399 if (lock->l_readers || lock->l_writers) {
2400 LDLM_ERROR(lock, "lock still has references");
2404 if (ldlm_is_waited(lock))
2405 ldlm_del_waiting_lock(lock);
2407 /* Releases cancel callback. */
2408 ldlm_cancel_callback(lock);
2410 /* Yes, second time, just in case it was added again while we were
2411 * running with no res lock in ldlm_cancel_callback */
2412 if (ldlm_is_waited(lock))
2413 ldlm_del_waiting_lock(lock);
2415 ldlm_resource_unlink_lock(lock);
2416 ldlm_lock_destroy_nolock(lock);
2418 if (lock->l_granted_mode == lock->l_req_mode)
2419 ldlm_pool_del(&ns->ns_pool, lock);
2421 /* Make sure we will not be called again for same lock what is possible
2422 * if not to zero out lock->l_granted_mode */
2423 lock->l_granted_mode = LCK_MINMODE;
2424 unlock_res_and_lock(lock);
2428 EXPORT_SYMBOL(ldlm_lock_cancel);
2431 * Set opaque data into the lock that only makes sense to upper layer.
2433 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2435 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2440 if (lock->l_ast_data == NULL)
2441 lock->l_ast_data = data;
2442 if (lock->l_ast_data == data)
2444 LDLM_LOCK_PUT(lock);
2448 EXPORT_SYMBOL(ldlm_lock_set_data);
2450 struct export_cl_data {
2451 struct obd_export *ecl_exp;
2455 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2456 struct ldlm_lock *lock,
2457 struct export_cl_data *ecl)
2459 struct ldlm_resource *res;
2461 res = ldlm_resource_getref(lock->l_resource);
2463 ldlm_lvbo_update(res, lock, NULL, 1);
2464 ldlm_lock_cancel(lock);
2465 if (!exp->exp_obd->obd_stopping)
2466 ldlm_reprocess_all(res);
2467 ldlm_resource_putref(res);
2470 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2471 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2472 exp, ecl->ecl_loop);
2477 * Iterator function for ldlm_export_cancel_locks.
2478 * Cancels passed locks.
2481 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2482 struct hlist_node *hnode, void *data)
2485 struct export_cl_data *ecl = (struct export_cl_data *)data;
2486 struct obd_export *exp = ecl->ecl_exp;
2487 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2489 LDLM_LOCK_GET(lock);
2490 ldlm_cancel_lock_for_export(exp, lock, ecl);
2491 LDLM_LOCK_RELEASE(lock);
2497 * Cancel all blocked locks for given export.
2499 * Typically called on client disconnection/eviction
2501 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2503 struct export_cl_data ecl = {
2508 while (!list_empty(&exp->exp_bl_list)) {
2509 struct ldlm_lock *lock;
2511 spin_lock_bh(&exp->exp_bl_list_lock);
2512 if (!list_empty(&exp->exp_bl_list)) {
2513 lock = list_entry(exp->exp_bl_list.next,
2514 struct ldlm_lock, l_exp_list);
2515 LDLM_LOCK_GET(lock);
2516 list_del_init(&lock->l_exp_list);
2520 spin_unlock_bh(&exp->exp_bl_list_lock);
2525 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2526 LDLM_LOCK_RELEASE(lock);
2529 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2530 "left on hash table %d.\n", exp, ecl.ecl_loop,
2531 atomic_read(&exp->exp_lock_hash->hs_count));
2533 return ecl.ecl_loop;
2537 * Cancel all locks for given export.
2539 * Typically called after client disconnection/eviction
2541 int ldlm_export_cancel_locks(struct obd_export *exp)
2543 struct export_cl_data ecl = {
2548 cfs_hash_for_each_empty(exp->exp_lock_hash,
2549 ldlm_cancel_locks_for_export_cb, &ecl);
2551 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2552 "left on hash table %d.\n", exp, ecl.ecl_loop,
2553 atomic_read(&exp->exp_lock_hash->hs_count));
2555 if (ecl.ecl_loop > 0 &&
2556 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2557 exp->exp_obd->obd_stopping)
2558 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2560 return ecl.ecl_loop;
2564 * Downgrade an PW/EX lock to COS mode.
2566 * A lock mode convertion from PW/EX mode to less conflict mode. The
2567 * convertion may fail if lock was canceled before downgrade, but it doesn't
2568 * indicate any problem, because such lock has no reader or writer, and will
2570 * Used by Commit on Sharing (COS) code only for now.
2572 * \param lock A lock to convert
2573 * \param new_mode new lock mode
2575 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2579 LASSERT(new_mode == LCK_COS);
2581 lock_res_and_lock(lock);
2583 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2584 unlock_res_and_lock(lock);
2586 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2587 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2591 ldlm_resource_unlink_lock(lock);
2593 * Remove the lock from pool as it will be added again in
2594 * ldlm_grant_lock() called below.
2596 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2597 lock->l_req_mode = new_mode;
2598 ldlm_grant_lock(lock, NULL);
2600 unlock_res_and_lock(lock);
2602 ldlm_reprocess_all(lock->l_resource);
2606 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2609 * Print lock with lock handle \a lockh description into debug log.
2611 * Used when printing all locks on a resource for debug purposes.
2613 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2615 struct ldlm_lock *lock;
2617 if (!((libcfs_debug | D_ERROR) & level))
2620 lock = ldlm_handle2lock(lockh);
2624 LDLM_DEBUG_LIMIT(level, lock, "###");
2626 LDLM_LOCK_PUT(lock);
2628 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2631 * Print lock information with custom message into debug log.
2634 void _ldlm_lock_debug(struct ldlm_lock *lock,
2635 struct libcfs_debug_msg_data *msgdata,
2636 const char *fmt, ...)
2639 struct obd_export *exp = lock->l_export;
2640 struct ldlm_resource *resource = NULL;
2641 char *nid = "local";
2643 /* on server-side resource of lock doesn't change */
2644 if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2645 if (lock->l_resource != NULL)
2646 resource = ldlm_resource_getref(lock->l_resource);
2647 } else if (spin_trylock(&lock->l_lock)) {
2648 if (lock->l_resource != NULL)
2649 resource = ldlm_resource_getref(lock->l_resource);
2650 spin_unlock(&lock->l_lock);
2653 va_start(args, fmt);
2655 if (exp && exp->exp_connection) {
2656 nid = obd_export_nid2str(exp);
2657 } else if (exp && exp->exp_obd != NULL) {
2658 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2659 nid = obd_import_nid2str(imp);
2662 if (resource == NULL) {
2663 libcfs_debug_vmsg2(msgdata, fmt, args,
2664 " ns: \?\? lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2665 "res: \?\? rrc=\?\? type: \?\?\? flags: %#llx nid: %s "
2666 "remote: %#llx expref: %d pid: %u timeout: %lld "
2669 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2670 lock->l_readers, lock->l_writers,
2671 ldlm_lockname[lock->l_granted_mode],
2672 ldlm_lockname[lock->l_req_mode],
2673 lock->l_flags, nid, lock->l_remote_handle.cookie,
2674 exp ? atomic_read(&exp->exp_refcount) : -99,
2675 lock->l_pid, lock->l_callback_timeout, lock->l_lvb_type);
2680 switch (resource->lr_type) {
2682 libcfs_debug_vmsg2(msgdata, fmt, args,
2683 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2684 "res: "DLDLMRES" rrc: %d type: %s [%llu->%llu] "
2685 "(req %llu->%llu) flags: %#llx nid: %s remote: "
2686 "%#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2687 ldlm_lock_to_ns_name(lock), lock,
2688 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2689 lock->l_readers, lock->l_writers,
2690 ldlm_lockname[lock->l_granted_mode],
2691 ldlm_lockname[lock->l_req_mode],
2693 atomic_read(&resource->lr_refcount),
2694 ldlm_typename[resource->lr_type],
2695 lock->l_policy_data.l_extent.start,
2696 lock->l_policy_data.l_extent.end,
2697 lock->l_req_extent.start, lock->l_req_extent.end,
2698 lock->l_flags, nid, lock->l_remote_handle.cookie,
2699 exp ? atomic_read(&exp->exp_refcount) : -99,
2700 lock->l_pid, lock->l_callback_timeout,
2705 libcfs_debug_vmsg2(msgdata, fmt, args,
2706 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2707 "res: "DLDLMRES" rrc: %d type: %s pid: %d "
2708 "[%llu->%llu] flags: %#llx nid: %s "
2709 "remote: %#llx expref: %d pid: %u timeout: %lld\n",
2710 ldlm_lock_to_ns_name(lock), lock,
2711 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2712 lock->l_readers, lock->l_writers,
2713 ldlm_lockname[lock->l_granted_mode],
2714 ldlm_lockname[lock->l_req_mode],
2716 atomic_read(&resource->lr_refcount),
2717 ldlm_typename[resource->lr_type],
2718 lock->l_policy_data.l_flock.pid,
2719 lock->l_policy_data.l_flock.start,
2720 lock->l_policy_data.l_flock.end,
2721 lock->l_flags, nid, lock->l_remote_handle.cookie,
2722 exp ? atomic_read(&exp->exp_refcount) : -99,
2723 lock->l_pid, lock->l_callback_timeout);
2727 libcfs_debug_vmsg2(msgdata, fmt, args,
2728 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2729 "res: "DLDLMRES" bits %#llx/%#llx rrc: %d type: %s "
2730 "flags: %#llx nid: %s remote: %#llx expref: %d "
2731 "pid: %u timeout: %lld lvb_type: %d\n",
2732 ldlm_lock_to_ns_name(lock),
2733 lock, lock->l_handle.h_cookie,
2734 atomic_read(&lock->l_refc),
2735 lock->l_readers, lock->l_writers,
2736 ldlm_lockname[lock->l_granted_mode],
2737 ldlm_lockname[lock->l_req_mode],
2739 lock->l_policy_data.l_inodebits.bits,
2740 lock->l_policy_data.l_inodebits.try_bits,
2741 atomic_read(&resource->lr_refcount),
2742 ldlm_typename[resource->lr_type],
2743 lock->l_flags, nid, lock->l_remote_handle.cookie,
2744 exp ? atomic_read(&exp->exp_refcount) : -99,
2745 lock->l_pid, lock->l_callback_timeout,
2750 libcfs_debug_vmsg2(msgdata, fmt, args,
2751 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2752 "res: "DLDLMRES" rrc: %d type: %s flags: %#llx "
2753 "nid: %s remote: %#llx expref: %d pid: %u "
2754 "timeout: %lld lvb_type: %d\n",
2755 ldlm_lock_to_ns_name(lock),
2756 lock, lock->l_handle.h_cookie,
2757 atomic_read(&lock->l_refc),
2758 lock->l_readers, lock->l_writers,
2759 ldlm_lockname[lock->l_granted_mode],
2760 ldlm_lockname[lock->l_req_mode],
2762 atomic_read(&resource->lr_refcount),
2763 ldlm_typename[resource->lr_type],
2764 lock->l_flags, nid, lock->l_remote_handle.cookie,
2765 exp ? atomic_read(&exp->exp_refcount) : -99,
2766 lock->l_pid, lock->l_callback_timeout,
2771 ldlm_resource_putref(resource);
2773 EXPORT_SYMBOL(_ldlm_lock_debug);