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 if (ns->ns_last_pos == &lock->l_lru)
240 ns->ns_last_pos = lock->l_lru.prev;
241 list_del_init(&lock->l_lru);
242 LASSERT(ns->ns_nr_unused > 0);
250 * Removes LDLM lock \a lock from LRU. Obtains the LRU lock first.
252 * If \a last_use is non-zero, it will remove the lock from LRU only if
253 * it matches lock's l_last_used.
255 * \retval 0 if \a last_use is set, the lock is not in LRU list or \a last_use
256 * doesn't match lock's l_last_used;
257 * otherwise, the lock hasn't been in the LRU list.
258 * \retval 1 the lock was in LRU list and removed.
260 int ldlm_lock_remove_from_lru_check(struct ldlm_lock *lock, ktime_t last_use)
262 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
266 if (ldlm_is_ns_srv(lock)) {
267 LASSERT(list_empty(&lock->l_lru));
271 spin_lock(&ns->ns_lock);
272 if (!ktime_compare(last_use, ktime_set(0, 0)) ||
273 !ktime_compare(last_use, lock->l_last_used))
274 rc = ldlm_lock_remove_from_lru_nolock(lock);
275 spin_unlock(&ns->ns_lock);
281 * Adds LDLM lock \a lock to namespace LRU. Assumes LRU is already locked.
283 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
285 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
287 lock->l_last_used = ktime_get();
288 LASSERT(list_empty(&lock->l_lru));
289 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
290 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
291 LASSERT(ns->ns_nr_unused >= 0);
296 * Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks
299 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
301 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
304 spin_lock(&ns->ns_lock);
305 ldlm_lock_add_to_lru_nolock(lock);
306 spin_unlock(&ns->ns_lock);
311 * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
312 * the LRU. Performs necessary LRU locking
314 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
316 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
319 if (ldlm_is_ns_srv(lock)) {
320 LASSERT(list_empty(&lock->l_lru));
325 spin_lock(&ns->ns_lock);
326 if (!list_empty(&lock->l_lru)) {
327 ldlm_lock_remove_from_lru_nolock(lock);
328 ldlm_lock_add_to_lru_nolock(lock);
330 spin_unlock(&ns->ns_lock);
335 * Helper to destroy a locked lock.
337 * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
338 * Must be called with l_lock and lr_lock held.
340 * Does not actually free the lock data, but rather marks the lock as
341 * destroyed by setting l_destroyed field in the lock to 1. Destroys a
342 * handle->lock association too, so that the lock can no longer be found
343 * and removes the lock from LRU list. Actual lock freeing occurs when
344 * last lock reference goes away.
346 * Original comment (of some historical value):
347 * This used to have a 'strict' flag, which recovery would use to mark an
348 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
349 * shall explain why it's gone: with the new hash table scheme, once you call
350 * ldlm_lock_destroy, you can never drop your final references on this lock.
351 * Because it's not in the hash table anymore. -phil
353 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
357 if (lock->l_readers || lock->l_writers) {
358 LDLM_ERROR(lock, "lock still has references");
362 if (!list_empty(&lock->l_res_link)) {
363 LDLM_ERROR(lock, "lock still on resource");
367 if (ldlm_is_destroyed(lock)) {
368 LASSERT(list_empty(&lock->l_lru));
372 ldlm_set_destroyed(lock);
374 if (lock->l_export && lock->l_export->exp_lock_hash) {
375 /* NB: it's safe to call cfs_hash_del() even lock isn't
376 * in exp_lock_hash. */
377 /* In the function below, .hs_keycmp resolves to
378 * ldlm_export_lock_keycmp() */
379 /* coverity[overrun-buffer-val] */
380 cfs_hash_del(lock->l_export->exp_lock_hash,
381 &lock->l_remote_handle, &lock->l_exp_hash);
384 ldlm_lock_remove_from_lru(lock);
385 class_handle_unhash(&lock->l_handle);
392 * Destroys a LDLM lock \a lock. Performs necessary locking first.
394 void ldlm_lock_destroy(struct ldlm_lock *lock)
398 lock_res_and_lock(lock);
399 first = ldlm_lock_destroy_internal(lock);
400 unlock_res_and_lock(lock);
402 /* drop reference from hashtable only for first destroy */
404 lu_ref_del(&lock->l_reference, "hash", lock);
405 LDLM_LOCK_RELEASE(lock);
411 * Destroys a LDLM lock \a lock that is already locked.
413 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
417 first = ldlm_lock_destroy_internal(lock);
418 /* drop reference from hashtable only for first destroy */
420 lu_ref_del(&lock->l_reference, "hash", lock);
421 LDLM_LOCK_RELEASE(lock);
426 /* this is called by portals_handle2object with the handle lock taken */
427 static void lock_handle_addref(void *lock)
429 LDLM_LOCK_GET((struct ldlm_lock *)lock);
432 static void lock_handle_free(void *lock, int size)
434 LASSERT(size == sizeof(struct ldlm_lock));
435 OBD_SLAB_FREE(lock, ldlm_lock_slab, size);
438 static struct portals_handle_ops lock_handle_ops = {
439 .hop_addref = lock_handle_addref,
440 .hop_free = lock_handle_free,
445 * Allocate and initialize new lock structure.
447 * usage: pass in a resource on which you have done ldlm_resource_get
448 * new lock will take over the refcount.
449 * returns: lock with refcount 2 - one for current caller and one for remote
451 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
453 struct ldlm_lock *lock;
456 if (resource == NULL)
459 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, GFP_NOFS);
463 spin_lock_init(&lock->l_lock);
464 lock->l_resource = resource;
465 lu_ref_add(&resource->lr_reference, "lock", lock);
467 atomic_set(&lock->l_refc, 2);
468 INIT_LIST_HEAD(&lock->l_res_link);
469 INIT_LIST_HEAD(&lock->l_lru);
470 INIT_LIST_HEAD(&lock->l_pending_chain);
471 INIT_LIST_HEAD(&lock->l_bl_ast);
472 INIT_LIST_HEAD(&lock->l_cp_ast);
473 INIT_LIST_HEAD(&lock->l_rk_ast);
474 init_waitqueue_head(&lock->l_waitq);
475 lock->l_blocking_lock = NULL;
476 INIT_LIST_HEAD(&lock->l_sl_mode);
477 INIT_LIST_HEAD(&lock->l_sl_policy);
478 INIT_HLIST_NODE(&lock->l_exp_hash);
479 INIT_HLIST_NODE(&lock->l_exp_flock_hash);
481 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
483 INIT_LIST_HEAD(&lock->l_handle.h_link);
484 class_handle_hash(&lock->l_handle, &lock_handle_ops);
486 lu_ref_init(&lock->l_reference);
487 lu_ref_add(&lock->l_reference, "hash", lock);
488 lock->l_callback_timeout = 0;
490 #if LUSTRE_TRACKS_LOCK_EXP_REFS
491 INIT_LIST_HEAD(&lock->l_exp_refs_link);
492 lock->l_exp_refs_nr = 0;
493 lock->l_exp_refs_target = NULL;
495 INIT_LIST_HEAD(&lock->l_exp_list);
501 * Moves LDLM lock \a lock to another resource.
502 * This is used on client when server returns some other lock than requested
503 * (typically as a result of intent operation)
505 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
506 const struct ldlm_res_id *new_resid)
508 struct ldlm_resource *oldres = lock->l_resource;
509 struct ldlm_resource *newres;
513 LASSERT(ns_is_client(ns));
515 lock_res_and_lock(lock);
516 if (memcmp(new_resid, &lock->l_resource->lr_name,
517 sizeof(lock->l_resource->lr_name)) == 0) {
519 unlock_res_and_lock(lock);
523 LASSERT(new_resid->name[0] != 0);
525 /* This function assumes that the lock isn't on any lists */
526 LASSERT(list_empty(&lock->l_res_link));
528 type = oldres->lr_type;
529 unlock_res_and_lock(lock);
531 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
533 RETURN(PTR_ERR(newres));
535 lu_ref_add(&newres->lr_reference, "lock", lock);
537 * To flip the lock from the old to the new resource, lock, oldres and
538 * newres have to be locked. Resource spin-locks are nested within
539 * lock->l_lock, and are taken in the memory address order to avoid
542 spin_lock(&lock->l_lock);
543 oldres = lock->l_resource;
544 if (oldres < newres) {
546 lock_res_nested(newres, LRT_NEW);
549 lock_res_nested(oldres, LRT_NEW);
551 LASSERT(memcmp(new_resid, &oldres->lr_name,
552 sizeof oldres->lr_name) != 0);
553 lock->l_resource = newres;
555 unlock_res_and_lock(lock);
557 /* ...and the flowers are still standing! */
558 lu_ref_del(&oldres->lr_reference, "lock", lock);
559 ldlm_resource_putref(oldres);
564 /** \defgroup ldlm_handles LDLM HANDLES
565 * Ways to get hold of locks without any addresses.
570 * Fills in handle for LDLM lock \a lock into supplied \a lockh
571 * Does not take any references.
573 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
575 lockh->cookie = lock->l_handle.h_cookie;
577 EXPORT_SYMBOL(ldlm_lock2handle);
580 * Obtain a lock reference by handle.
582 * if \a flags: atomically get the lock and set the flags.
583 * Return NULL if flag already set
585 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
588 struct ldlm_lock *lock;
593 lock = class_handle2object(handle->cookie, NULL);
597 if (lock->l_export != NULL && lock->l_export->exp_failed) {
598 CDEBUG(D_INFO, "lock export failed: lock %p, exp %p\n",
599 lock, lock->l_export);
604 /* It's unlikely but possible that someone marked the lock as
605 * destroyed after we did handle2object on it */
606 if ((flags == 0) && !ldlm_is_destroyed(lock)) {
607 lu_ref_add(&lock->l_reference, "handle", current);
611 lock_res_and_lock(lock);
613 LASSERT(lock->l_resource != NULL);
615 lu_ref_add_atomic(&lock->l_reference, "handle", current);
616 if (unlikely(ldlm_is_destroyed(lock))) {
617 unlock_res_and_lock(lock);
618 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
623 /* If we're setting flags, make sure none of them are already set. */
625 if ((lock->l_flags & flags) != 0) {
626 unlock_res_and_lock(lock);
631 lock->l_flags |= flags;
634 unlock_res_and_lock(lock);
637 EXPORT_SYMBOL(__ldlm_handle2lock);
638 /** @} ldlm_handles */
641 * Fill in "on the wire" representation for given LDLM lock into supplied
642 * lock descriptor \a desc structure.
644 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
646 ldlm_res2desc(lock->l_resource, &desc->l_resource);
647 desc->l_req_mode = lock->l_req_mode;
648 desc->l_granted_mode = lock->l_granted_mode;
649 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
650 &lock->l_policy_data,
651 &desc->l_policy_data);
655 * Add a lock to list of conflicting locks to send AST to.
657 * Only add if we have not sent a blocking AST to the lock yet.
659 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
660 struct list_head *work_list)
662 if (!ldlm_is_ast_sent(lock)) {
663 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
664 ldlm_set_ast_sent(lock);
665 /* If the enqueuing client said so, tell the AST recipient to
666 * discard dirty data, rather than writing back. */
667 if (ldlm_is_ast_discard_data(new))
668 ldlm_set_discard_data(lock);
669 LASSERT(list_empty(&lock->l_bl_ast));
670 list_add(&lock->l_bl_ast, work_list);
672 LASSERT(lock->l_blocking_lock == NULL);
673 lock->l_blocking_lock = LDLM_LOCK_GET(new);
678 * Add a lock to list of just granted locks to send completion AST to.
680 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
681 struct list_head *work_list)
683 if (!ldlm_is_cp_reqd(lock)) {
684 ldlm_set_cp_reqd(lock);
685 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
686 LASSERT(list_empty(&lock->l_cp_ast));
687 list_add(&lock->l_cp_ast, work_list);
693 * Aggregator function to add AST work items into a list. Determines
694 * what sort of an AST work needs to be done and calls the proper
696 * Must be called with lr_lock held.
698 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
699 struct list_head *work_list)
702 check_res_locked(lock->l_resource);
704 ldlm_add_bl_work_item(lock, new, work_list);
706 ldlm_add_cp_work_item(lock, work_list);
711 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
712 * r/w reference type is determined by \a mode
713 * Calls ldlm_lock_addref_internal.
715 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
717 struct ldlm_lock *lock;
719 lock = ldlm_handle2lock(lockh);
720 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
721 ldlm_lock_addref_internal(lock, mode);
724 EXPORT_SYMBOL(ldlm_lock_addref);
728 * Add specified reader/writer reference to LDLM lock \a lock.
729 * r/w reference type is determined by \a mode
730 * Removes lock from LRU if it is there.
731 * Assumes the LDLM lock is already locked.
733 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
736 ldlm_lock_remove_from_lru(lock);
737 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
739 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
741 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
743 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
746 lu_ref_add_atomic(&lock->l_reference, "user", lock);
747 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
751 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
752 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
754 * \retval 0 success, lock was addref-ed
756 * \retval -EAGAIN lock is being canceled.
758 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
760 struct ldlm_lock *lock;
764 lock = ldlm_handle2lock(lockh);
766 lock_res_and_lock(lock);
767 if (lock->l_readers != 0 || lock->l_writers != 0 ||
768 !ldlm_is_cbpending(lock)) {
769 ldlm_lock_addref_internal_nolock(lock, mode);
772 unlock_res_and_lock(lock);
777 EXPORT_SYMBOL(ldlm_lock_addref_try);
780 * Add specified reader/writer reference to LDLM lock \a lock.
781 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
782 * Only called for local locks.
784 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
786 lock_res_and_lock(lock);
787 ldlm_lock_addref_internal_nolock(lock, mode);
788 unlock_res_and_lock(lock);
792 * Removes reader/writer reference for LDLM lock \a lock.
793 * Assumes LDLM lock is already locked.
794 * only called in ldlm_flock_destroy and for local locks.
795 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
796 * that cannot be placed in LRU.
798 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
801 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
802 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
803 LASSERT(lock->l_readers > 0);
804 lu_ref_del(&lock->l_reference, "reader", lock);
807 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
808 LASSERT(lock->l_writers > 0);
809 lu_ref_del(&lock->l_reference, "writer", lock);
813 lu_ref_del(&lock->l_reference, "user", lock);
814 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
818 * Removes reader/writer reference for LDLM lock \a lock.
819 * Locks LDLM lock first.
820 * If the lock is determined to be client lock on a client and r/w refcount
821 * drops to zero and the lock is not blocked, the lock is added to LRU lock
823 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
825 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
827 struct ldlm_namespace *ns;
830 lock_res_and_lock(lock);
832 ns = ldlm_lock_to_ns(lock);
834 ldlm_lock_decref_internal_nolock(lock, mode);
836 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
837 !lock->l_readers && !lock->l_writers) {
838 /* If this is a local lock on a server namespace and this was
839 * the last reference, cancel the lock.
841 * Group locks are special:
842 * They must not go in LRU, but they are not called back
843 * like non-group locks, instead they are manually released.
844 * They have an l_writers reference which they keep until
845 * they are manually released, so we remove them when they have
846 * no more reader or writer references. - LU-6368 */
847 ldlm_set_cbpending(lock);
850 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
851 /* If we received a blocked AST and this was the last reference,
852 * run the callback. */
853 if (ldlm_is_ns_srv(lock) && lock->l_export)
854 CERROR("FL_CBPENDING set on non-local lock--just a "
857 LDLM_DEBUG(lock, "final decref done on cbpending lock");
859 LDLM_LOCK_GET(lock); /* dropped by bl thread */
860 ldlm_lock_remove_from_lru(lock);
861 unlock_res_and_lock(lock);
863 if (ldlm_is_fail_loc(lock))
864 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
866 if (ldlm_is_atomic_cb(lock) ||
867 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
868 ldlm_handle_bl_callback(ns, NULL, lock);
869 } else if (ns_is_client(ns) &&
870 !lock->l_readers && !lock->l_writers &&
871 !ldlm_is_no_lru(lock) &&
872 !ldlm_is_bl_ast(lock) &&
873 !ldlm_is_converting(lock)) {
875 LDLM_DEBUG(lock, "add lock into lru list");
877 /* If this is a client-side namespace and this was the last
878 * reference, put it on the LRU. */
879 ldlm_lock_add_to_lru(lock);
880 unlock_res_and_lock(lock);
882 if (ldlm_is_fail_loc(lock))
883 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
885 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
886 * are not supported by the server, otherwise, it is done on
888 if (!exp_connect_cancelset(lock->l_conn_export) &&
889 !ns_connect_lru_resize(ns))
890 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
892 LDLM_DEBUG(lock, "do not add lock into lru list");
893 unlock_res_and_lock(lock);
900 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
902 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
904 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
905 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
906 ldlm_lock_decref_internal(lock, mode);
909 EXPORT_SYMBOL(ldlm_lock_decref);
912 * Decrease reader/writer refcount for LDLM lock with handle
913 * \a lockh and mark it for subsequent cancellation once r/w refcount
914 * drops to zero instead of putting into LRU.
917 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
920 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
923 LASSERT(lock != NULL);
925 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
926 lock_res_and_lock(lock);
927 ldlm_set_cbpending(lock);
928 unlock_res_and_lock(lock);
929 ldlm_lock_decref_internal(lock, mode);
932 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
934 struct sl_insert_point {
935 struct list_head *res_link;
936 struct list_head *mode_link;
937 struct list_head *policy_link;
941 * Finds a position to insert the new lock into granted lock list.
943 * Used for locks eligible for skiplist optimization.
946 * queue [input]: the granted list where search acts on;
947 * req [input]: the lock whose position to be located;
948 * prev [output]: positions within 3 lists to insert @req to
952 * - ldlm_grant_lock_with_skiplist
954 static void search_granted_lock(struct list_head *queue,
955 struct ldlm_lock *req,
956 struct sl_insert_point *prev)
958 struct list_head *tmp;
959 struct ldlm_lock *lock, *mode_end, *policy_end;
962 list_for_each(tmp, queue) {
963 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
965 mode_end = list_entry(lock->l_sl_mode.prev,
966 struct ldlm_lock, l_sl_mode);
968 if (lock->l_req_mode != req->l_req_mode) {
969 /* jump to last lock of mode group */
970 tmp = &mode_end->l_res_link;
974 /* suitable mode group is found */
975 if (lock->l_resource->lr_type == LDLM_PLAIN) {
976 /* insert point is last lock of the mode group */
977 prev->res_link = &mode_end->l_res_link;
978 prev->mode_link = &mode_end->l_sl_mode;
979 prev->policy_link = &req->l_sl_policy;
982 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
985 list_entry(lock->l_sl_policy.prev,
989 if (lock->l_policy_data.l_inodebits.bits ==
990 req->l_policy_data.l_inodebits.bits) {
991 /* insert point is last lock of
992 * the policy group */
994 &policy_end->l_res_link;
996 &policy_end->l_sl_mode;
998 &policy_end->l_sl_policy;
1003 if (policy_end == mode_end)
1004 /* done with mode group */
1007 /* go to next policy group within mode group */
1008 tmp = policy_end->l_res_link.next;
1009 lock = list_entry(tmp, struct ldlm_lock,
1011 } /* loop over policy groups within the mode group */
1013 /* insert point is last lock of the mode group,
1014 * new policy group is started */
1015 prev->res_link = &mode_end->l_res_link;
1016 prev->mode_link = &mode_end->l_sl_mode;
1017 prev->policy_link = &req->l_sl_policy;
1021 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1026 /* insert point is last lock on the queue,
1027 * new mode group and new policy group are started */
1028 prev->res_link = queue->prev;
1029 prev->mode_link = &req->l_sl_mode;
1030 prev->policy_link = &req->l_sl_policy;
1036 * Add a lock into resource granted list after a position described by
1039 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1040 struct sl_insert_point *prev)
1042 struct ldlm_resource *res = lock->l_resource;
1045 check_res_locked(res);
1047 ldlm_resource_dump(D_INFO, res);
1048 LDLM_DEBUG(lock, "About to add lock:");
1050 if (ldlm_is_destroyed(lock)) {
1051 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1055 LASSERT(list_empty(&lock->l_res_link));
1056 LASSERT(list_empty(&lock->l_sl_mode));
1057 LASSERT(list_empty(&lock->l_sl_policy));
1060 * lock->link == prev->link means lock is first starting the group.
1061 * Don't re-add to itself to suppress kernel warnings.
1063 if (&lock->l_res_link != prev->res_link)
1064 list_add(&lock->l_res_link, prev->res_link);
1065 if (&lock->l_sl_mode != prev->mode_link)
1066 list_add(&lock->l_sl_mode, prev->mode_link);
1067 if (&lock->l_sl_policy != prev->policy_link)
1068 list_add(&lock->l_sl_policy, prev->policy_link);
1074 * Add a lock to granted list on a resource maintaining skiplist
1077 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1079 struct sl_insert_point prev;
1081 LASSERT(lock->l_req_mode == lock->l_granted_mode);
1083 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1084 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
1095 * must be called with lr_lock held
1097 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1099 struct ldlm_resource *res = lock->l_resource;
1102 check_res_locked(res);
1104 lock->l_granted_mode = lock->l_req_mode;
1106 if (work_list && lock->l_completion_ast != NULL)
1107 ldlm_add_ast_work_item(lock, NULL, work_list);
1109 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1110 ldlm_grant_lock_with_skiplist(lock);
1111 else if (res->lr_type == LDLM_EXTENT)
1112 ldlm_extent_add_lock(res, lock);
1113 else if (res->lr_type == LDLM_FLOCK) {
1114 /* We should not add locks to granted list in the following
1116 * - this is an UNLOCK but not a real lock;
1117 * - this is a TEST lock;
1118 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1119 * - this is a deadlock (flock cannot be granted) */
1120 if (lock->l_req_mode == 0 ||
1121 lock->l_req_mode == LCK_NL ||
1122 ldlm_is_test_lock(lock) ||
1123 ldlm_is_flock_deadlock(lock))
1125 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1130 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1135 * Describe the overlap between two locks. itree_overlap_cb data.
1137 struct lock_match_data {
1138 struct ldlm_lock *lmd_old;
1139 struct ldlm_lock *lmd_lock;
1140 enum ldlm_mode *lmd_mode;
1141 union ldlm_policy_data *lmd_policy;
1147 * Check if the given @lock meets the criteria for a match.
1148 * A reference on the lock is taken if matched.
1150 * \param lock test-against this lock
1151 * \param data parameters
1153 static int lock_matches(struct ldlm_lock *lock, struct lock_match_data *data)
1155 union ldlm_policy_data *lpol = &lock->l_policy_data;
1156 enum ldlm_mode match;
1158 if (lock == data->lmd_old)
1159 return INTERVAL_ITER_STOP;
1161 /* Check if this lock can be matched.
1162 * Used by LU-2919(exclusive open) for open lease lock */
1163 if (ldlm_is_excl(lock))
1164 return INTERVAL_ITER_CONT;
1166 /* llite sometimes wants to match locks that will be
1167 * canceled when their users drop, but we allow it to match
1168 * if it passes in CBPENDING and the lock still has users.
1169 * this is generally only going to be used by children
1170 * whose parents already hold a lock so forward progress
1171 * can still happen. */
1172 if (ldlm_is_cbpending(lock) &&
1173 !(data->lmd_flags & LDLM_FL_CBPENDING))
1174 return INTERVAL_ITER_CONT;
1175 if (!data->lmd_unref && ldlm_is_cbpending(lock) &&
1176 lock->l_readers == 0 && lock->l_writers == 0)
1177 return INTERVAL_ITER_CONT;
1179 if (!(lock->l_req_mode & *data->lmd_mode))
1180 return INTERVAL_ITER_CONT;
1181 match = lock->l_req_mode;
1183 switch (lock->l_resource->lr_type) {
1185 if (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1186 lpol->l_extent.end < data->lmd_policy->l_extent.end)
1187 return INTERVAL_ITER_CONT;
1189 if (unlikely(match == LCK_GROUP) &&
1190 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1191 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1192 return INTERVAL_ITER_CONT;
1195 /* We match if we have existing lock with same or wider set
1197 if ((lpol->l_inodebits.bits &
1198 data->lmd_policy->l_inodebits.bits) !=
1199 data->lmd_policy->l_inodebits.bits)
1200 return INTERVAL_ITER_CONT;
1206 /* We match if we have existing lock with same or wider set
1208 if (!data->lmd_unref && LDLM_HAVE_MASK(lock, GONE))
1209 return INTERVAL_ITER_CONT;
1211 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1212 return INTERVAL_ITER_CONT;
1214 if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1215 LDLM_LOCK_GET(lock);
1216 ldlm_lock_touch_in_lru(lock);
1218 ldlm_lock_addref_internal_nolock(lock, match);
1221 *data->lmd_mode = match;
1222 data->lmd_lock = lock;
1224 return INTERVAL_ITER_STOP;
1227 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1229 struct ldlm_interval *node = to_ldlm_interval(in);
1230 struct lock_match_data *data = args;
1231 struct ldlm_lock *lock;
1234 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1235 rc = lock_matches(lock, data);
1236 if (rc == INTERVAL_ITER_STOP)
1237 return INTERVAL_ITER_STOP;
1239 return INTERVAL_ITER_CONT;
1243 * Search for a lock with given parameters in interval trees.
1245 * \param res search for a lock in this resource
1246 * \param data parameters
1248 * \retval a referenced lock or NULL.
1250 static struct ldlm_lock *search_itree(struct ldlm_resource *res,
1251 struct lock_match_data *data)
1253 struct interval_node_extent ext = {
1254 .start = data->lmd_policy->l_extent.start,
1255 .end = data->lmd_policy->l_extent.end
1259 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1260 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1262 if (tree->lit_root == NULL)
1265 if (!(tree->lit_mode & *data->lmd_mode))
1268 interval_search(tree->lit_root, &ext,
1269 itree_overlap_cb, data);
1271 return data->lmd_lock;
1276 * Search for a lock with given properties in a queue.
1278 * \param queue search for a lock in this queue
1279 * \param data parameters
1281 * \retval a referenced lock or NULL.
1283 static struct ldlm_lock *search_queue(struct list_head *queue,
1284 struct lock_match_data *data)
1286 struct ldlm_lock *lock;
1289 list_for_each_entry(lock, queue, l_res_link) {
1290 rc = lock_matches(lock, data);
1291 if (rc == INTERVAL_ITER_STOP)
1292 return data->lmd_lock;
1297 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1299 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1300 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1301 wake_up_all(&lock->l_waitq);
1304 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1306 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1308 lock_res_and_lock(lock);
1309 ldlm_lock_fail_match_locked(lock);
1310 unlock_res_and_lock(lock);
1314 * Mark lock as "matchable" by OST.
1316 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1318 * Assumes LDLM lock is already locked.
1320 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1322 ldlm_set_lvb_ready(lock);
1323 wake_up_all(&lock->l_waitq);
1325 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1328 * Mark lock as "matchable" by OST.
1329 * Locks the lock and then \see ldlm_lock_allow_match_locked
1331 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1333 lock_res_and_lock(lock);
1334 ldlm_lock_allow_match_locked(lock);
1335 unlock_res_and_lock(lock);
1337 EXPORT_SYMBOL(ldlm_lock_allow_match);
1340 * Attempt to find a lock with specified properties.
1342 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1345 * Can be called in two ways:
1347 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1348 * for a duplicate of.
1350 * Otherwise, all of the fields must be filled in, to match against.
1352 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1353 * server (ie, connh is NULL)
1354 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1355 * list will be considered
1356 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1357 * to be canceled can still be matched as long as they still have reader
1358 * or writer refernces
1359 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1360 * just tell us if we would have matched.
1362 * \retval 1 if it finds an already-existing lock that is compatible; in this
1363 * case, lockh is filled in with a addref()ed lock
1365 * We also check security context, and if that fails we simply return 0 (to
1366 * keep caller code unchanged), the context failure will be discovered by
1367 * caller sometime later.
1369 enum ldlm_mode ldlm_lock_match(struct ldlm_namespace *ns, __u64 flags,
1370 const struct ldlm_res_id *res_id,
1371 enum ldlm_type type,
1372 union ldlm_policy_data *policy,
1373 enum ldlm_mode mode,
1374 struct lustre_handle *lockh, int unref)
1376 struct lock_match_data data = {
1380 .lmd_policy = policy,
1384 struct ldlm_resource *res;
1385 struct ldlm_lock *lock;
1390 data.lmd_old = ldlm_handle2lock(lockh);
1391 LASSERT(data.lmd_old != NULL);
1393 ns = ldlm_lock_to_ns(data.lmd_old);
1394 res_id = &data.lmd_old->l_resource->lr_name;
1395 type = data.lmd_old->l_resource->lr_type;
1396 *data.lmd_mode = data.lmd_old->l_req_mode;
1399 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1401 LASSERT(data.lmd_old == NULL);
1405 LDLM_RESOURCE_ADDREF(res);
1408 if (res->lr_type == LDLM_EXTENT)
1409 lock = search_itree(res, &data);
1411 lock = search_queue(&res->lr_granted, &data);
1414 if (flags & LDLM_FL_BLOCK_GRANTED)
1416 lock = search_queue(&res->lr_waiting, &data);
1423 LDLM_RESOURCE_DELREF(res);
1424 ldlm_resource_putref(res);
1427 ldlm_lock2handle(lock, lockh);
1428 if ((flags & LDLM_FL_LVB_READY) &&
1429 (!ldlm_is_lvb_ready(lock))) {
1430 __u64 wait_flags = LDLM_FL_LVB_READY |
1431 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1432 struct l_wait_info lwi;
1433 if (lock->l_completion_ast) {
1434 int err = lock->l_completion_ast(lock,
1435 LDLM_FL_WAIT_NOREPROC,
1438 if (flags & LDLM_FL_TEST_LOCK)
1439 LDLM_LOCK_RELEASE(lock);
1441 ldlm_lock_decref_internal(lock,
1448 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1449 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1451 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1452 l_wait_event(lock->l_waitq,
1453 lock->l_flags & wait_flags,
1455 if (!ldlm_is_lvb_ready(lock)) {
1456 if (flags & LDLM_FL_TEST_LOCK)
1457 LDLM_LOCK_RELEASE(lock);
1459 ldlm_lock_decref_internal(lock, mode);
1466 LDLM_DEBUG(lock, "matched (%llu %llu)",
1467 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1468 res_id->name[2] : policy->l_extent.start,
1469 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1470 res_id->name[3] : policy->l_extent.end);
1472 /* check user's security context */
1473 if (lock->l_conn_export &&
1474 sptlrpc_import_check_ctx(
1475 class_exp2cliimp(lock->l_conn_export))) {
1476 if (!(flags & LDLM_FL_TEST_LOCK))
1477 ldlm_lock_decref_internal(lock, mode);
1481 if (flags & LDLM_FL_TEST_LOCK)
1482 LDLM_LOCK_RELEASE(lock);
1484 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1485 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1486 "%llu/%llu (%llu %llu)", ns,
1487 type, mode, res_id->name[0], res_id->name[1],
1488 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1489 res_id->name[2] :policy->l_extent.start,
1490 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1491 res_id->name[3] : policy->l_extent.end);
1493 if (data.lmd_old != NULL)
1494 LDLM_LOCK_PUT(data.lmd_old);
1496 return rc ? mode : 0;
1498 EXPORT_SYMBOL(ldlm_lock_match);
1500 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1503 struct ldlm_lock *lock;
1504 enum ldlm_mode mode = 0;
1507 lock = ldlm_handle2lock(lockh);
1509 lock_res_and_lock(lock);
1510 if (LDLM_HAVE_MASK(lock, GONE))
1513 if (ldlm_is_cbpending(lock) &&
1514 lock->l_readers == 0 && lock->l_writers == 0)
1518 *bits = lock->l_policy_data.l_inodebits.bits;
1519 mode = lock->l_granted_mode;
1520 ldlm_lock_addref_internal_nolock(lock, mode);
1527 unlock_res_and_lock(lock);
1528 LDLM_LOCK_PUT(lock);
1532 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1534 /** The caller must guarantee that the buffer is large enough. */
1535 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1536 enum req_location loc, void *data, int size)
1541 LASSERT(data != NULL);
1544 switch (lock->l_lvb_type) {
1546 if (size == sizeof(struct ost_lvb)) {
1547 if (loc == RCL_CLIENT)
1548 lvb = req_capsule_client_swab_get(pill,
1550 lustre_swab_ost_lvb);
1552 lvb = req_capsule_server_swab_get(pill,
1554 lustre_swab_ost_lvb);
1555 if (unlikely(lvb == NULL)) {
1556 LDLM_ERROR(lock, "no LVB");
1560 memcpy(data, lvb, size);
1561 } else if (size == sizeof(struct ost_lvb_v1)) {
1562 struct ost_lvb *olvb = data;
1564 if (loc == RCL_CLIENT)
1565 lvb = req_capsule_client_swab_get(pill,
1567 lustre_swab_ost_lvb_v1);
1569 lvb = req_capsule_server_sized_swab_get(pill,
1571 lustre_swab_ost_lvb_v1);
1572 if (unlikely(lvb == NULL)) {
1573 LDLM_ERROR(lock, "no LVB");
1577 memcpy(data, lvb, size);
1578 olvb->lvb_mtime_ns = 0;
1579 olvb->lvb_atime_ns = 0;
1580 olvb->lvb_ctime_ns = 0;
1582 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1588 if (size == sizeof(struct lquota_lvb)) {
1589 if (loc == RCL_CLIENT)
1590 lvb = req_capsule_client_swab_get(pill,
1592 lustre_swab_lquota_lvb);
1594 lvb = req_capsule_server_swab_get(pill,
1596 lustre_swab_lquota_lvb);
1597 if (unlikely(lvb == NULL)) {
1598 LDLM_ERROR(lock, "no LVB");
1602 memcpy(data, lvb, size);
1604 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1613 if (loc == RCL_CLIENT)
1614 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1616 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1617 if (unlikely(lvb == NULL)) {
1618 LDLM_ERROR(lock, "no LVB");
1622 memcpy(data, lvb, size);
1625 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1626 libcfs_debug_dumpstack(NULL);
1634 * Create and fill in new LDLM lock with specified properties.
1635 * Returns a referenced lock
1637 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1638 const struct ldlm_res_id *res_id,
1639 enum ldlm_type type,
1640 enum ldlm_mode mode,
1641 const struct ldlm_callback_suite *cbs,
1642 void *data, __u32 lvb_len,
1643 enum lvb_type lvb_type)
1645 struct ldlm_lock *lock;
1646 struct ldlm_resource *res;
1650 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1652 RETURN(ERR_CAST(res));
1654 lock = ldlm_lock_new(res);
1656 RETURN(ERR_PTR(-ENOMEM));
1658 lock->l_req_mode = mode;
1659 lock->l_ast_data = data;
1660 lock->l_pid = current_pid();
1661 if (ns_is_server(ns))
1662 ldlm_set_ns_srv(lock);
1664 lock->l_blocking_ast = cbs->lcs_blocking;
1665 lock->l_completion_ast = cbs->lcs_completion;
1666 lock->l_glimpse_ast = cbs->lcs_glimpse;
1669 lock->l_tree_node = NULL;
1670 /* if this is the extent lock, allocate the interval tree node */
1671 if (type == LDLM_EXTENT)
1672 if (ldlm_interval_alloc(lock) == NULL)
1673 GOTO(out, rc = -ENOMEM);
1676 lock->l_lvb_len = lvb_len;
1677 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1678 if (lock->l_lvb_data == NULL)
1679 GOTO(out, rc = -ENOMEM);
1682 lock->l_lvb_type = lvb_type;
1683 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1684 GOTO(out, rc = -ENOENT);
1689 ldlm_lock_destroy(lock);
1690 LDLM_LOCK_RELEASE(lock);
1691 RETURN(ERR_PTR(rc));
1694 #ifdef HAVE_SERVER_SUPPORT
1695 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1698 struct ldlm_resource *res = lock->l_resource;
1699 enum ldlm_error rc = ELDLM_OK;
1700 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1701 ldlm_processing_policy policy;
1704 policy = ldlm_processing_policy_table[res->lr_type];
1706 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1707 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1708 res->lr_type != LDLM_FLOCK) {
1709 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1710 if (rc == -ERESTART)
1714 if (!list_empty(&rpc_list))
1715 ldlm_discard_bl_list(&rpc_list);
1722 * Enqueue (request) a lock.
1724 * Does not block. As a result of enqueue the lock would be put
1725 * into granted or waiting list.
1727 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1728 * set, skip all the enqueueing and delegate lock processing to intent policy
1731 enum ldlm_error ldlm_lock_enqueue(struct ldlm_namespace *ns,
1732 struct ldlm_lock **lockp,
1733 void *cookie, __u64 *flags)
1735 struct ldlm_lock *lock = *lockp;
1736 struct ldlm_resource *res = lock->l_resource;
1737 int local = ns_is_client(ldlm_res_to_ns(res));
1738 enum ldlm_error rc = ELDLM_OK;
1739 struct ldlm_interval *node = NULL;
1742 /* policies are not executed on the client or during replay */
1743 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1744 && !local && ns->ns_policy) {
1745 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1747 if (rc == ELDLM_LOCK_REPLACED) {
1748 /* The lock that was returned has already been granted,
1749 * and placed into lockp. If it's not the same as the
1750 * one we passed in, then destroy the old one and our
1751 * work here is done. */
1752 if (lock != *lockp) {
1753 ldlm_lock_destroy(lock);
1754 LDLM_LOCK_RELEASE(lock);
1756 *flags |= LDLM_FL_LOCK_CHANGED;
1758 } else if (rc != ELDLM_OK &&
1759 lock->l_req_mode == lock->l_granted_mode) {
1760 LASSERT(*flags & LDLM_FL_RESENT);
1761 /* It may happen that ns_policy returns an error in
1762 * resend case, object may be unlinked or just some
1763 * error occurs. It is unclear if lock reached the
1764 * client in the original reply, just leave the lock on
1765 * server, not returning it again to client. Due to
1766 * LU-6529, the server will not OOM. */
1768 } else if (rc != ELDLM_OK ||
1769 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1770 ldlm_lock_destroy(lock);
1775 if (*flags & LDLM_FL_RESENT) {
1776 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1777 * Set LOCK_CHANGED always.
1778 * Check if the lock is granted for BLOCK_GRANTED.
1779 * Take NO_TIMEOUT from the lock as it is inherited through
1780 * LDLM_FL_INHERIT_MASK */
1781 *flags |= LDLM_FL_LOCK_CHANGED;
1782 if (lock->l_req_mode != lock->l_granted_mode)
1783 *flags |= LDLM_FL_BLOCK_GRANTED;
1784 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1788 /* For a replaying lock, it might be already in granted list. So
1789 * unlinking the lock will cause the interval node to be freed, we
1790 * have to allocate the interval node early otherwise we can't regrant
1791 * this lock in the future. - jay */
1792 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1793 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1795 lock_res_and_lock(lock);
1796 if (local && lock->l_req_mode == lock->l_granted_mode) {
1797 /* The server returned a blocked lock, but it was granted
1798 * before we got a chance to actually enqueue it. We don't
1799 * need to do anything else. */
1800 *flags &= ~LDLM_FL_BLOCKED_MASK;
1801 GOTO(out, rc = ELDLM_OK);
1804 ldlm_resource_unlink_lock(lock);
1805 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1807 ldlm_lock_destroy_nolock(lock);
1808 GOTO(out, rc = -ENOMEM);
1811 INIT_LIST_HEAD(&node->li_group);
1812 ldlm_interval_attach(node, lock);
1816 /* Some flags from the enqueue want to make it into the AST, via the
1817 * lock's l_flags. */
1818 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1819 ldlm_set_ast_discard_data(lock);
1820 if (*flags & LDLM_FL_TEST_LOCK)
1821 ldlm_set_test_lock(lock);
1822 if (*flags & LDLM_FL_COS_INCOMPAT)
1823 ldlm_set_cos_incompat(lock);
1824 if (*flags & LDLM_FL_COS_ENABLED)
1825 ldlm_set_cos_enabled(lock);
1827 /* This distinction between local lock trees is very important; a client
1828 * namespace only has information about locks taken by that client, and
1829 * thus doesn't have enough information to decide for itself if it can
1830 * be granted (below). In this case, we do exactly what the server
1831 * tells us to do, as dictated by the 'flags'.
1833 * We do exactly the same thing during recovery, when the server is
1834 * more or less trusting the clients not to lie.
1836 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1839 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1840 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1842 ldlm_grant_lock(lock, NULL);
1843 GOTO(out, rc = ELDLM_OK);
1844 #ifdef HAVE_SERVER_SUPPORT
1845 } else if (*flags & LDLM_FL_REPLAY) {
1846 if (*flags & LDLM_FL_BLOCK_WAIT) {
1847 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1848 GOTO(out, rc = ELDLM_OK);
1849 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1850 ldlm_grant_lock(lock, NULL);
1851 GOTO(out, rc = ELDLM_OK);
1853 /* If no flags, fall through to normal enqueue path. */
1856 rc = ldlm_lock_enqueue_helper(lock, flags);
1860 CERROR("This is client-side-only module, cannot handle "
1861 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1867 unlock_res_and_lock(lock);
1869 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1873 #ifdef HAVE_SERVER_SUPPORT
1875 * Iterate through all waiting locks on a given resource queue and attempt to
1878 * Must be called with resource lock held.
1880 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1881 struct list_head *work_list,
1882 enum ldlm_process_intention intention)
1884 struct list_head *tmp, *pos;
1885 ldlm_processing_policy policy;
1887 int rc = LDLM_ITER_CONTINUE;
1888 enum ldlm_error err;
1889 struct list_head bl_ast_list = LIST_HEAD_INIT(bl_ast_list);
1892 check_res_locked(res);
1894 policy = ldlm_processing_policy_table[res->lr_type];
1896 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1897 intention == LDLM_PROCESS_RECOVERY);
1900 list_for_each_safe(tmp, pos, queue) {
1901 struct ldlm_lock *pending;
1902 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1904 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1906 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1909 rc = policy(pending, &flags, intention, &err, &rpc_list);
1910 if (pending->l_granted_mode == pending->l_req_mode ||
1911 res->lr_type == LDLM_FLOCK) {
1912 list_splice(&rpc_list, work_list);
1914 list_splice(&rpc_list, &bl_ast_list);
1917 * When this is called from recovery done, we always want
1918 * to scan the whole list no matter what 'rc' is returned.
1920 if (rc != LDLM_ITER_CONTINUE &&
1921 intention == LDLM_PROCESS_RESCAN)
1925 if (!list_empty(&bl_ast_list)) {
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);
2068 /* copy blocking lock ibits in cancel_bits as well,
2069 * new client may use them for lock convert and it is
2070 * important to use new field to convert locks from
2073 d.l_policy_data.l_inodebits.cancel_bits =
2074 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2076 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2077 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
2078 lock->l_blocking_lock = NULL;
2079 LDLM_LOCK_RELEASE(lock);
2085 * Process a call to completion AST callback for a lock in ast_work list
2088 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2090 struct ldlm_cb_set_arg *arg = opaq;
2092 struct ldlm_lock *lock;
2093 ldlm_completion_callback completion_callback;
2096 if (list_empty(arg->list))
2099 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2101 /* It's possible to receive a completion AST before we've set
2102 * the l_completion_ast pointer: either because the AST arrived
2103 * before the reply, or simply because there's a small race
2104 * window between receiving the reply and finishing the local
2105 * enqueue. (bug 842)
2107 * This can't happen with the blocking_ast, however, because we
2108 * will never call the local blocking_ast until we drop our
2109 * reader/writer reference, which we won't do until we get the
2110 * reply and finish enqueueing. */
2112 /* nobody should touch l_cp_ast */
2113 lock_res_and_lock(lock);
2114 list_del_init(&lock->l_cp_ast);
2115 LASSERT(ldlm_is_cp_reqd(lock));
2116 /* save l_completion_ast since it can be changed by
2117 * mds_intent_policy(), see bug 14225 */
2118 completion_callback = lock->l_completion_ast;
2119 ldlm_clear_cp_reqd(lock);
2120 unlock_res_and_lock(lock);
2122 if (completion_callback != NULL)
2123 rc = completion_callback(lock, 0, (void *)arg);
2124 LDLM_LOCK_RELEASE(lock);
2130 * Process a call to revocation AST callback for a lock in ast_work list
2133 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2135 struct ldlm_cb_set_arg *arg = opaq;
2136 struct ldlm_lock_desc desc;
2138 struct ldlm_lock *lock;
2141 if (list_empty(arg->list))
2144 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2145 list_del_init(&lock->l_rk_ast);
2147 /* the desc just pretend to exclusive */
2148 ldlm_lock2desc(lock, &desc);
2149 desc.l_req_mode = LCK_EX;
2150 desc.l_granted_mode = 0;
2152 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2153 LDLM_LOCK_RELEASE(lock);
2159 * Process a call to glimpse AST callback for a lock in ast_work list
2161 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2163 struct ldlm_cb_set_arg *arg = opaq;
2164 struct ldlm_glimpse_work *gl_work;
2165 struct ldlm_lock *lock;
2169 if (list_empty(arg->list))
2172 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2174 list_del_init(&gl_work->gl_list);
2176 lock = gl_work->gl_lock;
2178 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2179 arg->gl_desc = gl_work->gl_desc;
2180 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2181 arg->gl_interpret_data = gl_work->gl_interpret_data;
2183 /* invoke the actual glimpse callback */
2184 if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
2187 LDLM_LOCK_RELEASE(lock);
2188 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2189 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2191 OBD_FREE_PTR(gl_work);
2197 * Process list of locks in need of ASTs being sent.
2199 * Used on server to send multiple ASTs together instead of sending one by
2202 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2203 ldlm_desc_ast_t ast_type)
2205 struct ldlm_cb_set_arg *arg;
2206 set_producer_func work_ast_lock;
2209 if (list_empty(rpc_list))
2216 atomic_set(&arg->restart, 0);
2217 arg->list = rpc_list;
2220 case LDLM_WORK_BL_AST:
2221 arg->type = LDLM_BL_CALLBACK;
2222 work_ast_lock = ldlm_work_bl_ast_lock;
2224 case LDLM_WORK_CP_AST:
2225 arg->type = LDLM_CP_CALLBACK;
2226 work_ast_lock = ldlm_work_cp_ast_lock;
2228 case LDLM_WORK_REVOKE_AST:
2229 arg->type = LDLM_BL_CALLBACK;
2230 work_ast_lock = ldlm_work_revoke_ast_lock;
2232 case LDLM_WORK_GL_AST:
2233 arg->type = LDLM_GL_CALLBACK;
2234 work_ast_lock = ldlm_work_gl_ast_lock;
2240 /* We create a ptlrpc request set with flow control extension.
2241 * This request set will use the work_ast_lock function to produce new
2242 * requests and will send a new request each time one completes in order
2243 * to keep the number of requests in flight to ns_max_parallel_ast */
2244 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2245 work_ast_lock, arg);
2246 if (arg->set == NULL)
2247 GOTO(out, rc = -ENOMEM);
2249 ptlrpc_set_wait(arg->set);
2250 ptlrpc_set_destroy(arg->set);
2252 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2260 * Try to grant all waiting locks on a resource.
2262 * Calls ldlm_reprocess_queue on waiting queue.
2264 * Typically called after some resource locks are cancelled to see
2265 * if anything could be granted as a result of the cancellation.
2267 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2268 enum ldlm_process_intention intention)
2270 struct list_head rpc_list;
2271 #ifdef HAVE_SERVER_SUPPORT
2272 struct obd_device *obd;
2276 INIT_LIST_HEAD(&rpc_list);
2277 /* Local lock trees don't get reprocessed. */
2278 if (ns_is_client(ldlm_res_to_ns(res))) {
2283 /* Disable reprocess during lock replay stage but allow during
2284 * request replay stage.
2286 obd = ldlm_res_to_ns(res)->ns_obd;
2287 if (obd->obd_recovering &&
2288 atomic_read(&obd->obd_req_replay_clients) == 0)
2292 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list, intention);
2295 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2297 if (rc == -ERESTART) {
2298 LASSERT(list_empty(&rpc_list));
2304 INIT_LIST_HEAD(&rpc_list);
2305 if (!ns_is_client(ldlm_res_to_ns(res))) {
2306 CERROR("This is client-side-only module, cannot handle "
2307 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2314 void ldlm_reprocess_all(struct ldlm_resource *res)
2316 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN);
2318 EXPORT_SYMBOL(ldlm_reprocess_all);
2320 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2321 struct hlist_node *hnode, void *arg)
2323 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2325 /* This is only called once after recovery done. LU-8306. */
2326 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY);
2331 * Iterate through all resources on a namespace attempting to grant waiting
2334 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2339 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2340 ldlm_reprocess_res, NULL, 0);
2346 * Helper function to call blocking AST for LDLM lock \a lock in a
2347 * "cancelling" mode.
2349 void ldlm_cancel_callback(struct ldlm_lock *lock)
2351 check_res_locked(lock->l_resource);
2352 if (!ldlm_is_cancel(lock)) {
2353 ldlm_set_cancel(lock);
2354 if (lock->l_blocking_ast) {
2355 unlock_res_and_lock(lock);
2356 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2358 lock_res_and_lock(lock);
2360 LDLM_DEBUG(lock, "no blocking ast");
2363 /* only canceller can set bl_done bit */
2364 ldlm_set_bl_done(lock);
2365 wake_up_all(&lock->l_waitq);
2366 } else if (!ldlm_is_bl_done(lock)) {
2367 struct l_wait_info lwi = { 0 };
2369 /* The lock is guaranteed to have been canceled once
2370 * returning from this function. */
2371 unlock_res_and_lock(lock);
2372 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
2373 lock_res_and_lock(lock);
2378 * Remove skiplist-enabled LDLM lock \a req from granted list
2380 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2382 if (req->l_resource->lr_type != LDLM_PLAIN &&
2383 req->l_resource->lr_type != LDLM_IBITS)
2386 list_del_init(&req->l_sl_policy);
2387 list_del_init(&req->l_sl_mode);
2391 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2393 void ldlm_lock_cancel(struct ldlm_lock *lock)
2395 struct ldlm_resource *res;
2396 struct ldlm_namespace *ns;
2399 lock_res_and_lock(lock);
2401 res = lock->l_resource;
2402 ns = ldlm_res_to_ns(res);
2404 /* Please do not, no matter how tempting, remove this LBUG without
2405 * talking to me first. -phik */
2406 if (lock->l_readers || lock->l_writers) {
2407 LDLM_ERROR(lock, "lock still has references");
2408 unlock_res_and_lock(lock);
2412 if (ldlm_is_waited(lock))
2413 ldlm_del_waiting_lock(lock);
2415 /* Releases cancel callback. */
2416 ldlm_cancel_callback(lock);
2418 /* Yes, second time, just in case it was added again while we were
2419 * running with no res lock in ldlm_cancel_callback */
2420 if (ldlm_is_waited(lock))
2421 ldlm_del_waiting_lock(lock);
2423 ldlm_resource_unlink_lock(lock);
2424 ldlm_lock_destroy_nolock(lock);
2426 if (lock->l_granted_mode == lock->l_req_mode)
2427 ldlm_pool_del(&ns->ns_pool, lock);
2429 /* Make sure we will not be called again for same lock what is possible
2430 * if not to zero out lock->l_granted_mode */
2431 lock->l_granted_mode = LCK_MINMODE;
2432 unlock_res_and_lock(lock);
2436 EXPORT_SYMBOL(ldlm_lock_cancel);
2439 * Set opaque data into the lock that only makes sense to upper layer.
2441 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2443 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2448 if (lock->l_ast_data == NULL)
2449 lock->l_ast_data = data;
2450 if (lock->l_ast_data == data)
2452 LDLM_LOCK_PUT(lock);
2456 EXPORT_SYMBOL(ldlm_lock_set_data);
2458 struct export_cl_data {
2459 struct obd_export *ecl_exp;
2463 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2464 struct ldlm_lock *lock,
2465 struct export_cl_data *ecl)
2467 struct ldlm_resource *res;
2469 res = ldlm_resource_getref(lock->l_resource);
2471 ldlm_lvbo_update(res, lock, NULL, 1);
2472 ldlm_lock_cancel(lock);
2473 if (!exp->exp_obd->obd_stopping)
2474 ldlm_reprocess_all(res);
2475 ldlm_resource_putref(res);
2478 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2479 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2480 exp, ecl->ecl_loop);
2485 * Iterator function for ldlm_export_cancel_locks.
2486 * Cancels passed locks.
2489 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2490 struct hlist_node *hnode, void *data)
2493 struct export_cl_data *ecl = (struct export_cl_data *)data;
2494 struct obd_export *exp = ecl->ecl_exp;
2495 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2497 LDLM_LOCK_GET(lock);
2498 ldlm_cancel_lock_for_export(exp, lock, ecl);
2499 LDLM_LOCK_RELEASE(lock);
2505 * Cancel all blocked locks for given export.
2507 * Typically called on client disconnection/eviction
2509 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2511 struct export_cl_data ecl = {
2516 while (!list_empty(&exp->exp_bl_list)) {
2517 struct ldlm_lock *lock;
2519 spin_lock_bh(&exp->exp_bl_list_lock);
2520 if (!list_empty(&exp->exp_bl_list)) {
2521 lock = list_entry(exp->exp_bl_list.next,
2522 struct ldlm_lock, l_exp_list);
2523 LDLM_LOCK_GET(lock);
2524 list_del_init(&lock->l_exp_list);
2528 spin_unlock_bh(&exp->exp_bl_list_lock);
2533 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2534 LDLM_LOCK_RELEASE(lock);
2537 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2538 "left on hash table %d.\n", exp, ecl.ecl_loop,
2539 atomic_read(&exp->exp_lock_hash->hs_count));
2541 return ecl.ecl_loop;
2545 * Cancel all locks for given export.
2547 * Typically called after client disconnection/eviction
2549 int ldlm_export_cancel_locks(struct obd_export *exp)
2551 struct export_cl_data ecl = {
2556 cfs_hash_for_each_empty(exp->exp_lock_hash,
2557 ldlm_cancel_locks_for_export_cb, &ecl);
2559 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2560 "left on hash table %d.\n", exp, ecl.ecl_loop,
2561 atomic_read(&exp->exp_lock_hash->hs_count));
2563 if (ecl.ecl_loop > 0 &&
2564 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2565 exp->exp_obd->obd_stopping)
2566 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2568 return ecl.ecl_loop;
2572 * Downgrade an PW/EX lock to COS mode.
2574 * A lock mode convertion from PW/EX mode to less conflict mode. The
2575 * convertion may fail if lock was canceled before downgrade, but it doesn't
2576 * indicate any problem, because such lock has no reader or writer, and will
2578 * Used by Commit on Sharing (COS) code only for now.
2580 * \param lock A lock to convert
2581 * \param new_mode new lock mode
2583 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2587 LASSERT(new_mode == LCK_COS);
2589 lock_res_and_lock(lock);
2591 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2592 unlock_res_and_lock(lock);
2594 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2595 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2599 ldlm_resource_unlink_lock(lock);
2601 * Remove the lock from pool as it will be added again in
2602 * ldlm_grant_lock() called below.
2604 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2605 lock->l_req_mode = new_mode;
2606 ldlm_grant_lock(lock, NULL);
2608 unlock_res_and_lock(lock);
2610 ldlm_reprocess_all(lock->l_resource);
2614 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2617 * Print lock with lock handle \a lockh description into debug log.
2619 * Used when printing all locks on a resource for debug purposes.
2621 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2623 struct ldlm_lock *lock;
2625 if (!((libcfs_debug | D_ERROR) & level))
2628 lock = ldlm_handle2lock(lockh);
2632 LDLM_DEBUG_LIMIT(level, lock, "###");
2634 LDLM_LOCK_PUT(lock);
2636 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2639 * Print lock information with custom message into debug log.
2642 void _ldlm_lock_debug(struct ldlm_lock *lock,
2643 struct libcfs_debug_msg_data *msgdata,
2644 const char *fmt, ...)
2647 struct obd_export *exp = lock->l_export;
2648 struct ldlm_resource *resource = NULL;
2649 char *nid = "local";
2651 /* on server-side resource of lock doesn't change */
2652 if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2653 if (lock->l_resource != NULL)
2654 resource = ldlm_resource_getref(lock->l_resource);
2655 } else if (spin_trylock(&lock->l_lock)) {
2656 if (lock->l_resource != NULL)
2657 resource = ldlm_resource_getref(lock->l_resource);
2658 spin_unlock(&lock->l_lock);
2661 va_start(args, fmt);
2663 if (exp && exp->exp_connection) {
2664 nid = obd_export_nid2str(exp);
2665 } else if (exp && exp->exp_obd != NULL) {
2666 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2667 nid = obd_import_nid2str(imp);
2670 if (resource == NULL) {
2671 libcfs_debug_vmsg2(msgdata, fmt, args,
2672 " ns: \?\? lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2673 "res: \?\? rrc=\?\? type: \?\?\? flags: %#llx nid: %s "
2674 "remote: %#llx expref: %d pid: %u timeout: %lld "
2677 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2678 lock->l_readers, lock->l_writers,
2679 ldlm_lockname[lock->l_granted_mode],
2680 ldlm_lockname[lock->l_req_mode],
2681 lock->l_flags, nid, lock->l_remote_handle.cookie,
2682 exp ? atomic_read(&exp->exp_refcount) : -99,
2683 lock->l_pid, lock->l_callback_timeout, lock->l_lvb_type);
2688 switch (resource->lr_type) {
2690 libcfs_debug_vmsg2(msgdata, fmt, args,
2691 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2692 "res: "DLDLMRES" rrc: %d type: %s [%llu->%llu] "
2693 "(req %llu->%llu) flags: %#llx nid: %s remote: "
2694 "%#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2695 ldlm_lock_to_ns_name(lock), lock,
2696 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2697 lock->l_readers, lock->l_writers,
2698 ldlm_lockname[lock->l_granted_mode],
2699 ldlm_lockname[lock->l_req_mode],
2701 atomic_read(&resource->lr_refcount),
2702 ldlm_typename[resource->lr_type],
2703 lock->l_policy_data.l_extent.start,
2704 lock->l_policy_data.l_extent.end,
2705 lock->l_req_extent.start, lock->l_req_extent.end,
2706 lock->l_flags, nid, lock->l_remote_handle.cookie,
2707 exp ? atomic_read(&exp->exp_refcount) : -99,
2708 lock->l_pid, lock->l_callback_timeout,
2713 libcfs_debug_vmsg2(msgdata, fmt, args,
2714 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2715 "res: "DLDLMRES" rrc: %d type: %s pid: %d "
2716 "[%llu->%llu] flags: %#llx nid: %s "
2717 "remote: %#llx expref: %d pid: %u timeout: %lld\n",
2718 ldlm_lock_to_ns_name(lock), lock,
2719 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2720 lock->l_readers, lock->l_writers,
2721 ldlm_lockname[lock->l_granted_mode],
2722 ldlm_lockname[lock->l_req_mode],
2724 atomic_read(&resource->lr_refcount),
2725 ldlm_typename[resource->lr_type],
2726 lock->l_policy_data.l_flock.pid,
2727 lock->l_policy_data.l_flock.start,
2728 lock->l_policy_data.l_flock.end,
2729 lock->l_flags, nid, lock->l_remote_handle.cookie,
2730 exp ? atomic_read(&exp->exp_refcount) : -99,
2731 lock->l_pid, lock->l_callback_timeout);
2735 libcfs_debug_vmsg2(msgdata, fmt, args,
2736 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2737 "res: "DLDLMRES" bits %#llx/%#llx rrc: %d type: %s "
2738 "flags: %#llx nid: %s remote: %#llx expref: %d "
2739 "pid: %u timeout: %lld lvb_type: %d\n",
2740 ldlm_lock_to_ns_name(lock),
2741 lock, lock->l_handle.h_cookie,
2742 atomic_read(&lock->l_refc),
2743 lock->l_readers, lock->l_writers,
2744 ldlm_lockname[lock->l_granted_mode],
2745 ldlm_lockname[lock->l_req_mode],
2747 lock->l_policy_data.l_inodebits.bits,
2748 lock->l_policy_data.l_inodebits.try_bits,
2749 atomic_read(&resource->lr_refcount),
2750 ldlm_typename[resource->lr_type],
2751 lock->l_flags, nid, lock->l_remote_handle.cookie,
2752 exp ? atomic_read(&exp->exp_refcount) : -99,
2753 lock->l_pid, lock->l_callback_timeout,
2758 libcfs_debug_vmsg2(msgdata, fmt, args,
2759 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2760 "res: "DLDLMRES" rrc: %d type: %s flags: %#llx "
2761 "nid: %s remote: %#llx expref: %d pid: %u "
2762 "timeout: %lld lvb_type: %d\n",
2763 ldlm_lock_to_ns_name(lock),
2764 lock, lock->l_handle.h_cookie,
2765 atomic_read(&lock->l_refc),
2766 lock->l_readers, lock->l_writers,
2767 ldlm_lockname[lock->l_granted_mode],
2768 ldlm_lockname[lock->l_req_mode],
2770 atomic_read(&resource->lr_refcount),
2771 ldlm_typename[resource->lr_type],
2772 lock->l_flags, nid, lock->l_remote_handle.cookie,
2773 exp ? atomic_read(&exp->exp_refcount) : -99,
2774 lock->l_pid, lock->l_callback_timeout,
2779 ldlm_resource_putref(resource);
2781 EXPORT_SYMBOL(_ldlm_lock_debug);