4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_lock.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <libcfs/libcfs.h>
42 #include <lustre_swab.h>
43 #include <obd_class.h>
45 #include "ldlm_internal.h"
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 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1077 struct sl_insert_point prev;
1080 LASSERT(lock->l_req_mode == lock->l_granted_mode);
1082 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1083 ldlm_granted_list_add_lock(lock, &prev);
1088 * Perform lock granting bookkeeping.
1090 * Includes putting the lock into granted list and updating lock mode.
1092 * - ldlm_lock_enqueue
1093 * - ldlm_reprocess_queue
1094 * - ldlm_lock_convert
1096 * must be called with lr_lock held
1098 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1100 struct ldlm_resource *res = lock->l_resource;
1103 check_res_locked(res);
1105 lock->l_granted_mode = lock->l_req_mode;
1107 if (work_list && lock->l_completion_ast != NULL)
1108 ldlm_add_ast_work_item(lock, NULL, work_list);
1110 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1111 ldlm_grant_lock_with_skiplist(lock);
1112 else if (res->lr_type == LDLM_EXTENT)
1113 ldlm_extent_add_lock(res, lock);
1114 else if (res->lr_type == LDLM_FLOCK) {
1115 /* We should not add locks to granted list in the following
1117 * - this is an UNLOCK but not a real lock;
1118 * - this is a TEST lock;
1119 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1120 * - this is a deadlock (flock cannot be granted) */
1121 if (lock->l_req_mode == 0 ||
1122 lock->l_req_mode == LCK_NL ||
1123 ldlm_is_test_lock(lock) ||
1124 ldlm_is_flock_deadlock(lock))
1126 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1131 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1136 * Describe the overlap between two locks. itree_overlap_cb data.
1138 struct lock_match_data {
1139 struct ldlm_lock *lmd_old;
1140 struct ldlm_lock *lmd_lock;
1141 enum ldlm_mode *lmd_mode;
1142 union ldlm_policy_data *lmd_policy;
1148 * Check if the given @lock meets the criteria for a match.
1149 * A reference on the lock is taken if matched.
1151 * \param lock test-against this lock
1152 * \param data parameters
1154 static int lock_matches(struct ldlm_lock *lock, struct lock_match_data *data)
1156 union ldlm_policy_data *lpol = &lock->l_policy_data;
1157 enum ldlm_mode match;
1159 if (lock == data->lmd_old)
1160 return INTERVAL_ITER_STOP;
1162 /* Check if this lock can be matched.
1163 * Used by LU-2919(exclusive open) for open lease lock */
1164 if (ldlm_is_excl(lock))
1165 return INTERVAL_ITER_CONT;
1167 /* llite sometimes wants to match locks that will be
1168 * canceled when their users drop, but we allow it to match
1169 * if it passes in CBPENDING and the lock still has users.
1170 * this is generally only going to be used by children
1171 * whose parents already hold a lock so forward progress
1172 * can still happen. */
1173 if (ldlm_is_cbpending(lock) &&
1174 !(data->lmd_flags & LDLM_FL_CBPENDING))
1175 return INTERVAL_ITER_CONT;
1176 if (!data->lmd_unref && ldlm_is_cbpending(lock) &&
1177 lock->l_readers == 0 && lock->l_writers == 0)
1178 return INTERVAL_ITER_CONT;
1180 if (!(lock->l_req_mode & *data->lmd_mode))
1181 return INTERVAL_ITER_CONT;
1182 match = lock->l_req_mode;
1184 switch (lock->l_resource->lr_type) {
1186 if (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1187 lpol->l_extent.end < data->lmd_policy->l_extent.end)
1188 return INTERVAL_ITER_CONT;
1190 if (unlikely(match == LCK_GROUP) &&
1191 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1192 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1193 return INTERVAL_ITER_CONT;
1196 /* We match if we have existing lock with same or wider set
1198 if ((lpol->l_inodebits.bits &
1199 data->lmd_policy->l_inodebits.bits) !=
1200 data->lmd_policy->l_inodebits.bits)
1201 return INTERVAL_ITER_CONT;
1207 /* We match if we have existing lock with same or wider set
1209 if (!data->lmd_unref && LDLM_HAVE_MASK(lock, GONE))
1210 return INTERVAL_ITER_CONT;
1212 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1213 return INTERVAL_ITER_CONT;
1215 if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1216 LDLM_LOCK_GET(lock);
1217 ldlm_lock_touch_in_lru(lock);
1219 ldlm_lock_addref_internal_nolock(lock, match);
1222 *data->lmd_mode = match;
1223 data->lmd_lock = lock;
1225 return INTERVAL_ITER_STOP;
1228 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1230 struct ldlm_interval *node = to_ldlm_interval(in);
1231 struct lock_match_data *data = args;
1232 struct ldlm_lock *lock;
1235 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1236 rc = lock_matches(lock, data);
1237 if (rc == INTERVAL_ITER_STOP)
1238 return INTERVAL_ITER_STOP;
1240 return INTERVAL_ITER_CONT;
1244 * Search for a lock with given parameters in interval trees.
1246 * \param res search for a lock in this resource
1247 * \param data parameters
1249 * \retval a referenced lock or NULL.
1251 static struct ldlm_lock *search_itree(struct ldlm_resource *res,
1252 struct lock_match_data *data)
1254 struct interval_node_extent ext = {
1255 .start = data->lmd_policy->l_extent.start,
1256 .end = data->lmd_policy->l_extent.end
1260 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1261 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1263 if (tree->lit_root == NULL)
1266 if (!(tree->lit_mode & *data->lmd_mode))
1269 interval_search(tree->lit_root, &ext,
1270 itree_overlap_cb, data);
1272 return data->lmd_lock;
1277 * Search for a lock with given properties in a queue.
1279 * \param queue search for a lock in this queue
1280 * \param data parameters
1282 * \retval a referenced lock or NULL.
1284 static struct ldlm_lock *search_queue(struct list_head *queue,
1285 struct lock_match_data *data)
1287 struct ldlm_lock *lock;
1290 list_for_each_entry(lock, queue, l_res_link) {
1291 rc = lock_matches(lock, data);
1292 if (rc == INTERVAL_ITER_STOP)
1293 return data->lmd_lock;
1298 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1300 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1301 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1302 wake_up_all(&lock->l_waitq);
1305 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1307 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1309 lock_res_and_lock(lock);
1310 ldlm_lock_fail_match_locked(lock);
1311 unlock_res_and_lock(lock);
1315 * Mark lock as "matchable" by OST.
1317 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1319 * Assumes LDLM lock is already locked.
1321 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1323 ldlm_set_lvb_ready(lock);
1324 wake_up_all(&lock->l_waitq);
1326 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1329 * Mark lock as "matchable" by OST.
1330 * Locks the lock and then \see ldlm_lock_allow_match_locked
1332 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1334 lock_res_and_lock(lock);
1335 ldlm_lock_allow_match_locked(lock);
1336 unlock_res_and_lock(lock);
1338 EXPORT_SYMBOL(ldlm_lock_allow_match);
1341 * Attempt to find a lock with specified properties.
1343 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1346 * Can be called in two ways:
1348 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1349 * for a duplicate of.
1351 * Otherwise, all of the fields must be filled in, to match against.
1353 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1354 * server (ie, connh is NULL)
1355 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1356 * list will be considered
1357 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1358 * to be canceled can still be matched as long as they still have reader
1359 * or writer refernces
1360 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1361 * just tell us if we would have matched.
1363 * \retval 1 if it finds an already-existing lock that is compatible; in this
1364 * case, lockh is filled in with a addref()ed lock
1366 * We also check security context, and if that fails we simply return 0 (to
1367 * keep caller code unchanged), the context failure will be discovered by
1368 * caller sometime later.
1370 enum ldlm_mode ldlm_lock_match(struct ldlm_namespace *ns, __u64 flags,
1371 const struct ldlm_res_id *res_id,
1372 enum ldlm_type type,
1373 union ldlm_policy_data *policy,
1374 enum ldlm_mode mode,
1375 struct lustre_handle *lockh, int unref)
1377 struct lock_match_data data = {
1381 .lmd_policy = policy,
1385 struct ldlm_resource *res;
1386 struct ldlm_lock *lock;
1391 data.lmd_old = ldlm_handle2lock(lockh);
1392 LASSERT(data.lmd_old != NULL);
1394 ns = ldlm_lock_to_ns(data.lmd_old);
1395 res_id = &data.lmd_old->l_resource->lr_name;
1396 type = data.lmd_old->l_resource->lr_type;
1397 *data.lmd_mode = data.lmd_old->l_req_mode;
1400 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1402 LASSERT(data.lmd_old == NULL);
1406 LDLM_RESOURCE_ADDREF(res);
1409 if (res->lr_type == LDLM_EXTENT)
1410 lock = search_itree(res, &data);
1412 lock = search_queue(&res->lr_granted, &data);
1415 if (flags & LDLM_FL_BLOCK_GRANTED)
1417 lock = search_queue(&res->lr_converting, &data);
1420 lock = search_queue(&res->lr_waiting, &data);
1427 LDLM_RESOURCE_DELREF(res);
1428 ldlm_resource_putref(res);
1431 ldlm_lock2handle(lock, lockh);
1432 if ((flags & LDLM_FL_LVB_READY) &&
1433 (!ldlm_is_lvb_ready(lock))) {
1434 __u64 wait_flags = LDLM_FL_LVB_READY |
1435 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1436 struct l_wait_info lwi;
1437 if (lock->l_completion_ast) {
1438 int err = lock->l_completion_ast(lock,
1439 LDLM_FL_WAIT_NOREPROC,
1442 if (flags & LDLM_FL_TEST_LOCK)
1443 LDLM_LOCK_RELEASE(lock);
1445 ldlm_lock_decref_internal(lock,
1452 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1453 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1455 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1456 l_wait_event(lock->l_waitq,
1457 lock->l_flags & wait_flags,
1459 if (!ldlm_is_lvb_ready(lock)) {
1460 if (flags & LDLM_FL_TEST_LOCK)
1461 LDLM_LOCK_RELEASE(lock);
1463 ldlm_lock_decref_internal(lock, mode);
1470 LDLM_DEBUG(lock, "matched (%llu %llu)",
1471 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1472 res_id->name[2] : policy->l_extent.start,
1473 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1474 res_id->name[3] : policy->l_extent.end);
1476 /* check user's security context */
1477 if (lock->l_conn_export &&
1478 sptlrpc_import_check_ctx(
1479 class_exp2cliimp(lock->l_conn_export))) {
1480 if (!(flags & LDLM_FL_TEST_LOCK))
1481 ldlm_lock_decref_internal(lock, mode);
1485 if (flags & LDLM_FL_TEST_LOCK)
1486 LDLM_LOCK_RELEASE(lock);
1488 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1489 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1490 "%llu/%llu (%llu %llu)", ns,
1491 type, mode, res_id->name[0], res_id->name[1],
1492 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1493 res_id->name[2] :policy->l_extent.start,
1494 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1495 res_id->name[3] : policy->l_extent.end);
1497 if (data.lmd_old != NULL)
1498 LDLM_LOCK_PUT(data.lmd_old);
1500 return rc ? mode : 0;
1502 EXPORT_SYMBOL(ldlm_lock_match);
1504 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1507 struct ldlm_lock *lock;
1508 enum ldlm_mode mode = 0;
1511 lock = ldlm_handle2lock(lockh);
1513 lock_res_and_lock(lock);
1514 if (LDLM_HAVE_MASK(lock, GONE))
1517 if (ldlm_is_cbpending(lock) &&
1518 lock->l_readers == 0 && lock->l_writers == 0)
1522 *bits = lock->l_policy_data.l_inodebits.bits;
1523 mode = lock->l_granted_mode;
1524 ldlm_lock_addref_internal_nolock(lock, mode);
1531 unlock_res_and_lock(lock);
1532 LDLM_LOCK_PUT(lock);
1536 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1538 /** The caller must guarantee that the buffer is large enough. */
1539 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1540 enum req_location loc, void *data, int size)
1545 LASSERT(data != NULL);
1548 switch (lock->l_lvb_type) {
1550 if (size == sizeof(struct ost_lvb)) {
1551 if (loc == RCL_CLIENT)
1552 lvb = req_capsule_client_swab_get(pill,
1554 lustre_swab_ost_lvb);
1556 lvb = req_capsule_server_swab_get(pill,
1558 lustre_swab_ost_lvb);
1559 if (unlikely(lvb == NULL)) {
1560 LDLM_ERROR(lock, "no LVB");
1564 memcpy(data, lvb, size);
1565 } else if (size == sizeof(struct ost_lvb_v1)) {
1566 struct ost_lvb *olvb = data;
1568 if (loc == RCL_CLIENT)
1569 lvb = req_capsule_client_swab_get(pill,
1571 lustre_swab_ost_lvb_v1);
1573 lvb = req_capsule_server_sized_swab_get(pill,
1575 lustre_swab_ost_lvb_v1);
1576 if (unlikely(lvb == NULL)) {
1577 LDLM_ERROR(lock, "no LVB");
1581 memcpy(data, lvb, size);
1582 olvb->lvb_mtime_ns = 0;
1583 olvb->lvb_atime_ns = 0;
1584 olvb->lvb_ctime_ns = 0;
1586 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1592 if (size == sizeof(struct lquota_lvb)) {
1593 if (loc == RCL_CLIENT)
1594 lvb = req_capsule_client_swab_get(pill,
1596 lustre_swab_lquota_lvb);
1598 lvb = req_capsule_server_swab_get(pill,
1600 lustre_swab_lquota_lvb);
1601 if (unlikely(lvb == NULL)) {
1602 LDLM_ERROR(lock, "no LVB");
1606 memcpy(data, lvb, size);
1608 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1617 if (loc == RCL_CLIENT)
1618 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1620 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1621 if (unlikely(lvb == NULL)) {
1622 LDLM_ERROR(lock, "no LVB");
1626 memcpy(data, lvb, size);
1629 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1630 libcfs_debug_dumpstack(NULL);
1638 * Create and fill in new LDLM lock with specified properties.
1639 * Returns a referenced lock
1641 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1642 const struct ldlm_res_id *res_id,
1643 enum ldlm_type type,
1644 enum ldlm_mode mode,
1645 const struct ldlm_callback_suite *cbs,
1646 void *data, __u32 lvb_len,
1647 enum lvb_type lvb_type)
1649 struct ldlm_lock *lock;
1650 struct ldlm_resource *res;
1654 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1656 RETURN(ERR_CAST(res));
1658 lock = ldlm_lock_new(res);
1660 RETURN(ERR_PTR(-ENOMEM));
1662 lock->l_req_mode = mode;
1663 lock->l_ast_data = data;
1664 lock->l_pid = current_pid();
1665 if (ns_is_server(ns))
1666 ldlm_set_ns_srv(lock);
1668 lock->l_blocking_ast = cbs->lcs_blocking;
1669 lock->l_completion_ast = cbs->lcs_completion;
1670 lock->l_glimpse_ast = cbs->lcs_glimpse;
1673 lock->l_tree_node = NULL;
1674 /* if this is the extent lock, allocate the interval tree node */
1675 if (type == LDLM_EXTENT)
1676 if (ldlm_interval_alloc(lock) == NULL)
1677 GOTO(out, rc = -ENOMEM);
1680 lock->l_lvb_len = lvb_len;
1681 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1682 if (lock->l_lvb_data == NULL)
1683 GOTO(out, rc = -ENOMEM);
1686 lock->l_lvb_type = lvb_type;
1687 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1688 GOTO(out, rc = -ENOENT);
1693 ldlm_lock_destroy(lock);
1694 LDLM_LOCK_RELEASE(lock);
1695 RETURN(ERR_PTR(rc));
1699 * Enqueue (request) a lock.
1701 * Does not block. As a result of enqueue the lock would be put
1702 * into granted or waiting list.
1704 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1705 * set, skip all the enqueueing and delegate lock processing to intent policy
1708 enum ldlm_error ldlm_lock_enqueue(struct ldlm_namespace *ns,
1709 struct ldlm_lock **lockp,
1710 void *cookie, __u64 *flags)
1712 struct ldlm_lock *lock = *lockp;
1713 struct ldlm_resource *res = lock->l_resource;
1714 int local = ns_is_client(ldlm_res_to_ns(res));
1715 #ifdef HAVE_SERVER_SUPPORT
1716 ldlm_processing_policy policy;
1718 enum ldlm_error rc = ELDLM_OK;
1719 struct ldlm_interval *node = NULL;
1722 /* policies are not executed on the client or during replay */
1723 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1724 && !local && ns->ns_policy) {
1725 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1727 if (rc == ELDLM_LOCK_REPLACED) {
1728 /* The lock that was returned has already been granted,
1729 * and placed into lockp. If it's not the same as the
1730 * one we passed in, then destroy the old one and our
1731 * work here is done. */
1732 if (lock != *lockp) {
1733 ldlm_lock_destroy(lock);
1734 LDLM_LOCK_RELEASE(lock);
1736 *flags |= LDLM_FL_LOCK_CHANGED;
1738 } else if (rc != ELDLM_OK &&
1739 lock->l_req_mode == lock->l_granted_mode) {
1740 LASSERT(*flags & LDLM_FL_RESENT);
1741 /* It may happen that ns_policy returns an error in
1742 * resend case, object may be unlinked or just some
1743 * error occurs. It is unclear if lock reached the
1744 * client in the original reply, just leave the lock on
1745 * server, not returning it again to client. Due to
1746 * LU-6529, the server will not OOM. */
1748 } else if (rc != ELDLM_OK ||
1749 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1750 ldlm_lock_destroy(lock);
1755 if (*flags & LDLM_FL_RESENT) {
1756 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1757 * Set LOCK_CHANGED always.
1758 * Check if the lock is granted for BLOCK_GRANTED.
1759 * Take NO_TIMEOUT from the lock as it is inherited through
1760 * LDLM_FL_INHERIT_MASK */
1761 *flags |= LDLM_FL_LOCK_CHANGED;
1762 if (lock->l_req_mode != lock->l_granted_mode)
1763 *flags |= LDLM_FL_BLOCK_GRANTED;
1764 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1768 /* For a replaying lock, it might be already in granted list. So
1769 * unlinking the lock will cause the interval node to be freed, we
1770 * have to allocate the interval node early otherwise we can't regrant
1771 * this lock in the future. - jay */
1772 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1773 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1775 lock_res_and_lock(lock);
1776 if (local && lock->l_req_mode == lock->l_granted_mode) {
1777 /* The server returned a blocked lock, but it was granted
1778 * before we got a chance to actually enqueue it. We don't
1779 * need to do anything else. */
1780 *flags &= ~LDLM_FL_BLOCKED_MASK;
1781 GOTO(out, rc = ELDLM_OK);
1784 ldlm_resource_unlink_lock(lock);
1785 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1787 ldlm_lock_destroy_nolock(lock);
1788 GOTO(out, rc = -ENOMEM);
1791 INIT_LIST_HEAD(&node->li_group);
1792 ldlm_interval_attach(node, lock);
1796 /* Some flags from the enqueue want to make it into the AST, via the
1797 * lock's l_flags. */
1798 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1799 ldlm_set_ast_discard_data(lock);
1800 if (*flags & LDLM_FL_TEST_LOCK)
1801 ldlm_set_test_lock(lock);
1802 if (*flags & LDLM_FL_COS_INCOMPAT)
1803 ldlm_set_cos_incompat(lock);
1804 if (*flags & LDLM_FL_COS_ENABLED)
1805 ldlm_set_cos_enabled(lock);
1807 /* This distinction between local lock trees is very important; a client
1808 * namespace only has information about locks taken by that client, and
1809 * thus doesn't have enough information to decide for itself if it can
1810 * be granted (below). In this case, we do exactly what the server
1811 * tells us to do, as dictated by the 'flags'.
1813 * We do exactly the same thing during recovery, when the server is
1814 * more or less trusting the clients not to lie.
1816 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1817 * granted/converting queues. */
1819 if (*flags & LDLM_FL_BLOCK_CONV)
1820 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1821 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1822 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1824 ldlm_grant_lock(lock, NULL);
1825 GOTO(out, rc = ELDLM_OK);
1826 #ifdef HAVE_SERVER_SUPPORT
1827 } else if (*flags & LDLM_FL_REPLAY) {
1828 if (*flags & LDLM_FL_BLOCK_CONV) {
1829 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1830 GOTO(out, rc = ELDLM_OK);
1831 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1832 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1833 GOTO(out, rc = ELDLM_OK);
1834 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1835 ldlm_grant_lock(lock, NULL);
1836 GOTO(out, rc = ELDLM_OK);
1838 /* If no flags, fall through to normal enqueue path. */
1841 policy = ldlm_processing_policy_table[res->lr_type];
1842 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, NULL);
1846 CERROR("This is client-side-only module, cannot handle "
1847 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1853 unlock_res_and_lock(lock);
1855 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1859 #ifdef HAVE_SERVER_SUPPORT
1861 * Iterate through all waiting locks on a given resource queue and attempt to
1864 * Must be called with resource lock held.
1866 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1867 struct list_head *work_list,
1868 enum ldlm_process_intention intention)
1870 struct list_head *tmp, *pos;
1871 ldlm_processing_policy policy;
1873 int rc = LDLM_ITER_CONTINUE;
1874 enum ldlm_error err;
1877 check_res_locked(res);
1879 policy = ldlm_processing_policy_table[res->lr_type];
1881 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1882 intention == LDLM_PROCESS_RECOVERY);
1884 list_for_each_safe(tmp, pos, queue) {
1885 struct ldlm_lock *pending;
1887 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1889 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1892 rc = policy(pending, &flags, intention, &err, work_list);
1894 * When this is called from recovery done, we always want
1895 * to scan the whole list no matter what 'rc' is returned.
1897 if (rc != LDLM_ITER_CONTINUE &&
1898 intention == LDLM_PROCESS_RESCAN)
1902 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
1906 * Conflicting locks are detected for a lock to be enqueued, add the lock
1907 * into waiting list and send blocking ASTs to the conflicting locks.
1909 * \param[in] lock The lock to be enqueued.
1910 * \param[out] flags Lock flags for the lock to be enqueued.
1911 * \param[in] rpc_list Conflicting locks list.
1912 * \param[in] grant_flags extra flags when granting a lock.
1914 * \retval -ERESTART: Some lock was instantly canceled while sending
1915 * blocking ASTs, caller needs to re-check conflicting
1917 * \retval -EAGAIN: Lock was destroyed, caller should return error.
1918 * \reval 0: Lock is successfully added in waiting list.
1920 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
1921 struct list_head *rpc_list, __u64 grant_flags)
1923 struct ldlm_resource *res = lock->l_resource;
1927 check_res_locked(res);
1929 /* If either of the compat_queue()s returned failure, then we
1930 * have ASTs to send and must go onto the waiting list.
1932 * bug 2322: we used to unlink and re-add here, which was a
1933 * terrible folly -- if we goto restart, we could get
1934 * re-ordered! Causes deadlock, because ASTs aren't sent! */
1935 if (list_empty(&lock->l_res_link))
1936 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1939 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
1942 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
1943 !ns_is_client(ldlm_res_to_ns(res)))
1944 class_fail_export(lock->l_export);
1947 if (rc == -ERESTART) {
1948 /* 15715: The lock was granted and destroyed after
1949 * resource lock was dropped. Interval node was freed
1950 * in ldlm_lock_destroy. Anyway, this always happens
1951 * when a client is being evicted. So it would be
1952 * ok to return an error. -jay */
1953 if (ldlm_is_destroyed(lock))
1956 /* lock was granted while resource was unlocked. */
1957 if (lock->l_granted_mode == lock->l_req_mode) {
1958 /* bug 11300: if the lock has been granted,
1959 * break earlier because otherwise, we will go
1960 * to restart and ldlm_resource_unlink will be
1961 * called and it causes the interval node to be
1962 * freed. Then we will fail at
1963 * ldlm_extent_add_lock() */
1964 *flags &= ~LDLM_FL_BLOCKED_MASK;
1970 *flags |= (LDLM_FL_BLOCK_GRANTED | grant_flags);
1976 * Discard all AST work items from list.
1978 * If for whatever reason we do not want to send ASTs to conflicting locks
1979 * anymore, disassemble the list with this function.
1981 void ldlm_discard_bl_list(struct list_head *bl_list)
1983 struct list_head *tmp, *pos;
1986 list_for_each_safe(pos, tmp, bl_list) {
1987 struct ldlm_lock *lock =
1988 list_entry(pos, struct ldlm_lock, l_bl_ast);
1990 list_del_init(&lock->l_bl_ast);
1991 LASSERT(ldlm_is_ast_sent(lock));
1992 ldlm_clear_ast_sent(lock);
1993 LASSERT(lock->l_bl_ast_run == 0);
1994 LASSERT(lock->l_blocking_lock);
1995 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1996 lock->l_blocking_lock = NULL;
1997 LDLM_LOCK_RELEASE(lock);
2005 * Process a call to blocking AST callback for a lock in ast_work list
2008 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2010 struct ldlm_cb_set_arg *arg = opaq;
2011 struct ldlm_lock_desc d;
2013 struct ldlm_lock *lock;
2016 if (list_empty(arg->list))
2019 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2021 /* nobody should touch l_bl_ast */
2022 lock_res_and_lock(lock);
2023 list_del_init(&lock->l_bl_ast);
2025 LASSERT(ldlm_is_ast_sent(lock));
2026 LASSERT(lock->l_bl_ast_run == 0);
2027 LASSERT(lock->l_blocking_lock);
2028 lock->l_bl_ast_run++;
2029 unlock_res_and_lock(lock);
2031 ldlm_lock2desc(lock->l_blocking_lock, &d);
2033 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2034 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
2035 lock->l_blocking_lock = NULL;
2036 LDLM_LOCK_RELEASE(lock);
2042 * Process a call to completion AST callback for a lock in ast_work list
2045 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2047 struct ldlm_cb_set_arg *arg = opaq;
2049 struct ldlm_lock *lock;
2050 ldlm_completion_callback completion_callback;
2053 if (list_empty(arg->list))
2056 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2058 /* It's possible to receive a completion AST before we've set
2059 * the l_completion_ast pointer: either because the AST arrived
2060 * before the reply, or simply because there's a small race
2061 * window between receiving the reply and finishing the local
2062 * enqueue. (bug 842)
2064 * This can't happen with the blocking_ast, however, because we
2065 * will never call the local blocking_ast until we drop our
2066 * reader/writer reference, which we won't do until we get the
2067 * reply and finish enqueueing. */
2069 /* nobody should touch l_cp_ast */
2070 lock_res_and_lock(lock);
2071 list_del_init(&lock->l_cp_ast);
2072 LASSERT(ldlm_is_cp_reqd(lock));
2073 /* save l_completion_ast since it can be changed by
2074 * mds_intent_policy(), see bug 14225 */
2075 completion_callback = lock->l_completion_ast;
2076 ldlm_clear_cp_reqd(lock);
2077 unlock_res_and_lock(lock);
2079 if (completion_callback != NULL)
2080 rc = completion_callback(lock, 0, (void *)arg);
2081 LDLM_LOCK_RELEASE(lock);
2087 * Process a call to revocation AST callback for a lock in ast_work list
2090 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2092 struct ldlm_cb_set_arg *arg = opaq;
2093 struct ldlm_lock_desc desc;
2095 struct ldlm_lock *lock;
2098 if (list_empty(arg->list))
2101 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2102 list_del_init(&lock->l_rk_ast);
2104 /* the desc just pretend to exclusive */
2105 ldlm_lock2desc(lock, &desc);
2106 desc.l_req_mode = LCK_EX;
2107 desc.l_granted_mode = 0;
2109 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2110 LDLM_LOCK_RELEASE(lock);
2116 * Process a call to glimpse AST callback for a lock in ast_work list
2118 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2120 struct ldlm_cb_set_arg *arg = opaq;
2121 struct ldlm_glimpse_work *gl_work;
2122 struct ldlm_lock *lock;
2126 if (list_empty(arg->list))
2129 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2131 list_del_init(&gl_work->gl_list);
2133 lock = gl_work->gl_lock;
2135 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2136 arg->gl_desc = gl_work->gl_desc;
2137 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2138 arg->gl_interpret_data = gl_work->gl_interpret_data;
2140 /* invoke the actual glimpse callback */
2141 if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
2144 LDLM_LOCK_RELEASE(lock);
2145 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2146 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2148 OBD_FREE_PTR(gl_work);
2154 * Process list of locks in need of ASTs being sent.
2156 * Used on server to send multiple ASTs together instead of sending one by
2159 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2160 ldlm_desc_ast_t ast_type)
2162 struct ldlm_cb_set_arg *arg;
2163 set_producer_func work_ast_lock;
2166 if (list_empty(rpc_list))
2173 atomic_set(&arg->restart, 0);
2174 arg->list = rpc_list;
2177 case LDLM_WORK_BL_AST:
2178 arg->type = LDLM_BL_CALLBACK;
2179 work_ast_lock = ldlm_work_bl_ast_lock;
2181 case LDLM_WORK_CP_AST:
2182 arg->type = LDLM_CP_CALLBACK;
2183 work_ast_lock = ldlm_work_cp_ast_lock;
2185 case LDLM_WORK_REVOKE_AST:
2186 arg->type = LDLM_BL_CALLBACK;
2187 work_ast_lock = ldlm_work_revoke_ast_lock;
2189 case LDLM_WORK_GL_AST:
2190 arg->type = LDLM_GL_CALLBACK;
2191 work_ast_lock = ldlm_work_gl_ast_lock;
2197 /* We create a ptlrpc request set with flow control extension.
2198 * This request set will use the work_ast_lock function to produce new
2199 * requests and will send a new request each time one completes in order
2200 * to keep the number of requests in flight to ns_max_parallel_ast */
2201 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2202 work_ast_lock, arg);
2203 if (arg->set == NULL)
2204 GOTO(out, rc = -ENOMEM);
2206 ptlrpc_set_wait(arg->set);
2207 ptlrpc_set_destroy(arg->set);
2209 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2217 * Try to grant all waiting locks on a resource.
2219 * Calls ldlm_reprocess_queue on converting and waiting queues.
2221 * Typically called after some resource locks are cancelled to see
2222 * if anything could be granted as a result of the cancellation.
2224 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2225 enum ldlm_process_intention intention)
2227 struct list_head rpc_list;
2228 #ifdef HAVE_SERVER_SUPPORT
2229 struct obd_device *obd;
2233 INIT_LIST_HEAD(&rpc_list);
2234 /* Local lock trees don't get reprocessed. */
2235 if (ns_is_client(ldlm_res_to_ns(res))) {
2240 /* Disable reprocess during lock replay stage but allow during
2241 * request replay stage.
2243 obd = ldlm_res_to_ns(res)->ns_obd;
2244 if (obd->obd_recovering &&
2245 atomic_read(&obd->obd_req_replay_clients) == 0)
2249 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list,
2251 if (rc == LDLM_ITER_CONTINUE)
2252 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list,
2256 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2258 if (rc == -ERESTART) {
2259 LASSERT(list_empty(&rpc_list));
2265 INIT_LIST_HEAD(&rpc_list);
2266 if (!ns_is_client(ldlm_res_to_ns(res))) {
2267 CERROR("This is client-side-only module, cannot handle "
2268 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2275 void ldlm_reprocess_all(struct ldlm_resource *res)
2277 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN);
2279 EXPORT_SYMBOL(ldlm_reprocess_all);
2281 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2282 struct hlist_node *hnode, void *arg)
2284 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2286 /* This is only called once after recovery done. LU-8306. */
2287 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY);
2292 * Iterate through all resources on a namespace attempting to grant waiting
2295 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2300 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2301 ldlm_reprocess_res, NULL, 0);
2306 static bool is_bl_done(struct ldlm_lock *lock)
2308 bool bl_done = true;
2310 if (!ldlm_is_bl_done(lock)) {
2311 lock_res_and_lock(lock);
2312 bl_done = ldlm_is_bl_done(lock);
2313 unlock_res_and_lock(lock);
2320 * Helper function to call blocking AST for LDLM lock \a lock in a
2321 * "cancelling" mode.
2323 void ldlm_cancel_callback(struct ldlm_lock *lock)
2325 check_res_locked(lock->l_resource);
2326 if (!ldlm_is_cancel(lock)) {
2327 ldlm_set_cancel(lock);
2328 if (lock->l_blocking_ast) {
2329 unlock_res_and_lock(lock);
2330 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2332 lock_res_and_lock(lock);
2334 LDLM_DEBUG(lock, "no blocking ast");
2337 /* only canceller can set bl_done bit */
2338 ldlm_set_bl_done(lock);
2339 wake_up_all(&lock->l_waitq);
2340 } else if (!ldlm_is_bl_done(lock)) {
2341 struct l_wait_info lwi = { 0 };
2343 /* The lock is guaranteed to have been canceled once
2344 * returning from this function. */
2345 unlock_res_and_lock(lock);
2346 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
2347 lock_res_and_lock(lock);
2352 * Remove skiplist-enabled LDLM lock \a req from granted list
2354 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2356 if (req->l_resource->lr_type != LDLM_PLAIN &&
2357 req->l_resource->lr_type != LDLM_IBITS)
2360 list_del_init(&req->l_sl_policy);
2361 list_del_init(&req->l_sl_mode);
2365 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2367 void ldlm_lock_cancel(struct ldlm_lock *lock)
2369 struct ldlm_resource *res;
2370 struct ldlm_namespace *ns;
2373 lock_res_and_lock(lock);
2375 res = lock->l_resource;
2376 ns = ldlm_res_to_ns(res);
2378 /* Please do not, no matter how tempting, remove this LBUG without
2379 * talking to me first. -phik */
2380 if (lock->l_readers || lock->l_writers) {
2381 LDLM_ERROR(lock, "lock still has references");
2385 if (ldlm_is_waited(lock))
2386 ldlm_del_waiting_lock(lock);
2388 /* Releases cancel callback. */
2389 ldlm_cancel_callback(lock);
2391 /* Yes, second time, just in case it was added again while we were
2392 * running with no res lock in ldlm_cancel_callback */
2393 if (ldlm_is_waited(lock))
2394 ldlm_del_waiting_lock(lock);
2396 ldlm_resource_unlink_lock(lock);
2397 ldlm_lock_destroy_nolock(lock);
2399 if (lock->l_granted_mode == lock->l_req_mode)
2400 ldlm_pool_del(&ns->ns_pool, lock);
2402 /* Make sure we will not be called again for same lock what is possible
2403 * if not to zero out lock->l_granted_mode */
2404 lock->l_granted_mode = LCK_MINMODE;
2405 unlock_res_and_lock(lock);
2409 EXPORT_SYMBOL(ldlm_lock_cancel);
2412 * Set opaque data into the lock that only makes sense to upper layer.
2414 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2416 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2421 if (lock->l_ast_data == NULL)
2422 lock->l_ast_data = data;
2423 if (lock->l_ast_data == data)
2425 LDLM_LOCK_PUT(lock);
2429 EXPORT_SYMBOL(ldlm_lock_set_data);
2431 struct export_cl_data {
2432 struct obd_export *ecl_exp;
2436 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2437 struct ldlm_lock *lock,
2438 struct export_cl_data *ecl)
2440 struct ldlm_resource *res;
2442 res = ldlm_resource_getref(lock->l_resource);
2444 ldlm_res_lvbo_update(res, NULL, 1);
2445 ldlm_lock_cancel(lock);
2446 if (!exp->exp_obd->obd_stopping)
2447 ldlm_reprocess_all(res);
2448 ldlm_resource_putref(res);
2451 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2452 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2453 exp, ecl->ecl_loop);
2458 * Iterator function for ldlm_export_cancel_locks.
2459 * Cancels passed locks.
2462 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2463 struct hlist_node *hnode, void *data)
2466 struct export_cl_data *ecl = (struct export_cl_data *)data;
2467 struct obd_export *exp = ecl->ecl_exp;
2468 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2470 LDLM_LOCK_GET(lock);
2471 ldlm_cancel_lock_for_export(exp, lock, ecl);
2472 LDLM_LOCK_RELEASE(lock);
2478 * Cancel all blocked locks for given export.
2480 * Typically called on client disconnection/eviction
2482 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2484 struct export_cl_data ecl = {
2489 while (!list_empty(&exp->exp_bl_list)) {
2490 struct ldlm_lock *lock;
2492 spin_lock_bh(&exp->exp_bl_list_lock);
2493 if (!list_empty(&exp->exp_bl_list)) {
2494 lock = list_entry(exp->exp_bl_list.next,
2495 struct ldlm_lock, l_exp_list);
2496 LDLM_LOCK_GET(lock);
2497 list_del_init(&lock->l_exp_list);
2501 spin_unlock_bh(&exp->exp_bl_list_lock);
2506 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2507 LDLM_LOCK_RELEASE(lock);
2510 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2511 "left on hash table %d.\n", exp, ecl.ecl_loop,
2512 atomic_read(&exp->exp_lock_hash->hs_count));
2514 return ecl.ecl_loop;
2518 * Cancel all locks for given export.
2520 * Typically called after client disconnection/eviction
2522 int ldlm_export_cancel_locks(struct obd_export *exp)
2524 struct export_cl_data ecl = {
2529 cfs_hash_for_each_empty(exp->exp_lock_hash,
2530 ldlm_cancel_locks_for_export_cb, &ecl);
2532 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2533 "left on hash table %d.\n", exp, ecl.ecl_loop,
2534 atomic_read(&exp->exp_lock_hash->hs_count));
2536 if (ecl.ecl_loop > 0 &&
2537 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2538 exp->exp_obd->obd_stopping)
2539 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2541 return ecl.ecl_loop;
2545 * Downgrade an exclusive lock.
2547 * A fast variant of ldlm_lock_convert for convertion of exclusive locks. The
2548 * convertion may fail if lock was canceled before downgrade, but it doesn't
2549 * indicate any problem, because such lock has no reader or writer, and will
2551 * Used by Commit on Sharing (COS) code.
2553 * \param lock A lock to convert
2554 * \param new_mode new lock mode
2556 void ldlm_lock_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2560 LASSERT(new_mode == LCK_COS);
2562 lock_res_and_lock(lock);
2564 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2565 unlock_res_and_lock(lock);
2567 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2568 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2572 ldlm_resource_unlink_lock(lock);
2574 * Remove the lock from pool as it will be added again in
2575 * ldlm_grant_lock() called below.
2577 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2578 lock->l_req_mode = new_mode;
2579 ldlm_grant_lock(lock, NULL);
2581 unlock_res_and_lock(lock);
2583 ldlm_reprocess_all(lock->l_resource);
2587 EXPORT_SYMBOL(ldlm_lock_downgrade);
2590 * Attempt to convert already granted lock to a different mode.
2592 * While lock conversion is not currently used, future client-side
2593 * optimizations could take advantage of it to avoid discarding cached
2596 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock,
2597 enum ldlm_mode new_mode, __u32 *flags)
2599 struct list_head rpc_list;
2600 struct ldlm_resource *res;
2601 struct ldlm_namespace *ns;
2603 #ifdef HAVE_SERVER_SUPPORT
2605 struct sl_insert_point prev;
2607 struct ldlm_interval *node;
2610 INIT_LIST_HEAD(&rpc_list);
2611 /* Just return if mode is unchanged. */
2612 if (new_mode == lock->l_granted_mode) {
2613 *flags |= LDLM_FL_BLOCK_GRANTED;
2614 RETURN(lock->l_resource);
2617 /* I can't check the type of lock here because the bitlock of lock
2618 * is not held here, so do the allocation blindly. -jay */
2619 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
2620 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
2623 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
2624 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
2626 lock_res_and_lock(lock);
2628 res = lock->l_resource;
2629 ns = ldlm_res_to_ns(res);
2631 #ifdef HAVE_SERVER_SUPPORT
2632 old_mode = lock->l_req_mode;
2634 lock->l_req_mode = new_mode;
2635 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
2636 #ifdef HAVE_SERVER_SUPPORT
2637 /* remember the lock position where the lock might be
2638 * added back to the granted list later and also
2639 * remember the join mode for skiplist fixing. */
2640 prev.res_link = lock->l_res_link.prev;
2641 prev.mode_link = lock->l_sl_mode.prev;
2642 prev.policy_link = lock->l_sl_policy.prev;
2644 ldlm_resource_unlink_lock(lock);
2646 ldlm_resource_unlink_lock(lock);
2647 if (res->lr_type == LDLM_EXTENT) {
2648 /* FIXME: ugly code, I have to attach the lock to a
2649 * interval node again since perhaps it will be granted
2651 INIT_LIST_HEAD(&node->li_group);
2652 ldlm_interval_attach(node, lock);
2658 * Remove old lock from the pool before adding the lock with new
2659 * mode below in ->policy()
2661 ldlm_pool_del(&ns->ns_pool, lock);
2663 /* If this is a local resource, put it on the appropriate list. */
2664 if (ns_is_client(ldlm_res_to_ns(res))) {
2665 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
2666 ldlm_resource_add_lock(res, &res->lr_converting, lock);
2668 /* This should never happen, because of the way the
2669 * server handles conversions. */
2670 LDLM_ERROR(lock, "Erroneous flags %x on local lock\n",
2674 ldlm_grant_lock(lock, &rpc_list);
2676 /* FIXME: completion handling not with lr_lock held ! */
2677 if (lock->l_completion_ast)
2678 lock->l_completion_ast(lock, 0, NULL);
2680 #ifdef HAVE_SERVER_SUPPORT
2683 enum ldlm_error err;
2685 ldlm_processing_policy policy;
2687 policy = ldlm_processing_policy_table[res->lr_type];
2688 rc = policy(lock, &pflags, LDLM_PROCESS_RESCAN, &err,
2690 if (rc == LDLM_ITER_STOP) {
2691 lock->l_req_mode = old_mode;
2692 if (res->lr_type == LDLM_EXTENT)
2693 ldlm_extent_add_lock(res, lock);
2695 ldlm_granted_list_add_lock(lock, &prev);
2699 *flags |= LDLM_FL_BLOCK_GRANTED;
2705 CERROR("This is client-side-only module, cannot handle "
2706 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2710 unlock_res_and_lock(lock);
2713 ldlm_run_ast_work(ns, &rpc_list, LDLM_WORK_CP_AST);
2715 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
2720 * Print lock with lock handle \a lockh description into debug log.
2722 * Used when printing all locks on a resource for debug purposes.
2724 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2726 struct ldlm_lock *lock;
2728 if (!((libcfs_debug | D_ERROR) & level))
2731 lock = ldlm_handle2lock(lockh);
2735 LDLM_DEBUG_LIMIT(level, lock, "###");
2737 LDLM_LOCK_PUT(lock);
2739 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2742 * Print lock information with custom message into debug log.
2745 void _ldlm_lock_debug(struct ldlm_lock *lock,
2746 struct libcfs_debug_msg_data *msgdata,
2747 const char *fmt, ...)
2750 struct obd_export *exp = lock->l_export;
2751 struct ldlm_resource *resource = NULL;
2752 char *nid = "local";
2754 /* on server-side resource of lock doesn't change */
2755 if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2756 if (lock->l_resource != NULL)
2757 resource = ldlm_resource_getref(lock->l_resource);
2758 } else if (spin_trylock(&lock->l_lock)) {
2759 if (lock->l_resource != NULL)
2760 resource = ldlm_resource_getref(lock->l_resource);
2761 spin_unlock(&lock->l_lock);
2764 va_start(args, fmt);
2766 if (exp && exp->exp_connection) {
2767 nid = libcfs_nid2str(exp->exp_connection->c_peer.nid);
2768 } else if (exp && exp->exp_obd != NULL) {
2769 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2770 nid = libcfs_nid2str(imp->imp_connection->c_peer.nid);
2773 if (resource == NULL) {
2774 libcfs_debug_vmsg2(msgdata, fmt, args,
2775 " ns: \?\? lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2776 "res: \?\? rrc=\?\? type: \?\?\? flags: %#llx nid: %s "
2777 "remote: %#llx expref: %d pid: %u timeout: %lld "
2780 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2781 lock->l_readers, lock->l_writers,
2782 ldlm_lockname[lock->l_granted_mode],
2783 ldlm_lockname[lock->l_req_mode],
2784 lock->l_flags, nid, lock->l_remote_handle.cookie,
2785 exp ? atomic_read(&exp->exp_refcount) : -99,
2786 lock->l_pid, lock->l_callback_timeout, lock->l_lvb_type);
2791 switch (resource->lr_type) {
2793 libcfs_debug_vmsg2(msgdata, fmt, args,
2794 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2795 "res: "DLDLMRES" rrc: %d type: %s [%llu->%llu] "
2796 "(req %llu->%llu) flags: %#llx nid: %s remote: "
2797 "%#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2798 ldlm_lock_to_ns_name(lock), lock,
2799 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2800 lock->l_readers, lock->l_writers,
2801 ldlm_lockname[lock->l_granted_mode],
2802 ldlm_lockname[lock->l_req_mode],
2804 atomic_read(&resource->lr_refcount),
2805 ldlm_typename[resource->lr_type],
2806 lock->l_policy_data.l_extent.start,
2807 lock->l_policy_data.l_extent.end,
2808 lock->l_req_extent.start, lock->l_req_extent.end,
2809 lock->l_flags, nid, lock->l_remote_handle.cookie,
2810 exp ? atomic_read(&exp->exp_refcount) : -99,
2811 lock->l_pid, lock->l_callback_timeout,
2816 libcfs_debug_vmsg2(msgdata, fmt, args,
2817 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2818 "res: "DLDLMRES" rrc: %d type: %s pid: %d "
2819 "[%llu->%llu] flags: %#llx nid: %s "
2820 "remote: %#llx expref: %d pid: %u timeout: %lld\n",
2821 ldlm_lock_to_ns_name(lock), lock,
2822 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2823 lock->l_readers, lock->l_writers,
2824 ldlm_lockname[lock->l_granted_mode],
2825 ldlm_lockname[lock->l_req_mode],
2827 atomic_read(&resource->lr_refcount),
2828 ldlm_typename[resource->lr_type],
2829 lock->l_policy_data.l_flock.pid,
2830 lock->l_policy_data.l_flock.start,
2831 lock->l_policy_data.l_flock.end,
2832 lock->l_flags, nid, lock->l_remote_handle.cookie,
2833 exp ? atomic_read(&exp->exp_refcount) : -99,
2834 lock->l_pid, lock->l_callback_timeout);
2838 libcfs_debug_vmsg2(msgdata, fmt, args,
2839 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2840 "res: "DLDLMRES" bits %#llx/%#llx rrc: %d type: %s "
2841 "flags: %#llx nid: %s remote: %#llx expref: %d "
2842 "pid: %u timeout: %lld lvb_type: %d\n",
2843 ldlm_lock_to_ns_name(lock),
2844 lock, lock->l_handle.h_cookie,
2845 atomic_read(&lock->l_refc),
2846 lock->l_readers, lock->l_writers,
2847 ldlm_lockname[lock->l_granted_mode],
2848 ldlm_lockname[lock->l_req_mode],
2850 lock->l_policy_data.l_inodebits.bits,
2851 lock->l_policy_data.l_inodebits.try_bits,
2852 atomic_read(&resource->lr_refcount),
2853 ldlm_typename[resource->lr_type],
2854 lock->l_flags, nid, lock->l_remote_handle.cookie,
2855 exp ? atomic_read(&exp->exp_refcount) : -99,
2856 lock->l_pid, lock->l_callback_timeout,
2861 libcfs_debug_vmsg2(msgdata, fmt, args,
2862 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2863 "res: "DLDLMRES" rrc: %d type: %s flags: %#llx "
2864 "nid: %s remote: %#llx expref: %d pid: %u "
2865 "timeout: %lld lvb_type: %d\n",
2866 ldlm_lock_to_ns_name(lock),
2867 lock, lock->l_handle.h_cookie,
2868 atomic_read(&lock->l_refc),
2869 lock->l_readers, lock->l_writers,
2870 ldlm_lockname[lock->l_granted_mode],
2871 ldlm_lockname[lock->l_req_mode],
2873 atomic_read(&resource->lr_refcount),
2874 ldlm_typename[resource->lr_type],
2875 lock->l_flags, nid, lock->l_remote_handle.cookie,
2876 exp ? atomic_read(&exp->exp_refcount) : -99,
2877 lock->l_pid, lock->l_callback_timeout,
2882 ldlm_resource_putref(resource);
2884 EXPORT_SYMBOL(_ldlm_lock_debug);