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;
489 lock->l_activity = 0;
491 #if LUSTRE_TRACKS_LOCK_EXP_REFS
492 INIT_LIST_HEAD(&lock->l_exp_refs_link);
493 lock->l_exp_refs_nr = 0;
494 lock->l_exp_refs_target = NULL;
496 INIT_LIST_HEAD(&lock->l_exp_list);
502 * Moves LDLM lock \a lock to another resource.
503 * This is used on client when server returns some other lock than requested
504 * (typically as a result of intent operation)
506 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
507 const struct ldlm_res_id *new_resid)
509 struct ldlm_resource *oldres = lock->l_resource;
510 struct ldlm_resource *newres;
514 LASSERT(ns_is_client(ns));
516 lock_res_and_lock(lock);
517 if (memcmp(new_resid, &lock->l_resource->lr_name,
518 sizeof(lock->l_resource->lr_name)) == 0) {
520 unlock_res_and_lock(lock);
524 LASSERT(new_resid->name[0] != 0);
526 /* This function assumes that the lock isn't on any lists */
527 LASSERT(list_empty(&lock->l_res_link));
529 type = oldres->lr_type;
530 unlock_res_and_lock(lock);
532 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
534 RETURN(PTR_ERR(newres));
536 lu_ref_add(&newres->lr_reference, "lock", lock);
538 * To flip the lock from the old to the new resource, lock, oldres and
539 * newres have to be locked. Resource spin-locks are nested within
540 * lock->l_lock, and are taken in the memory address order to avoid
543 spin_lock(&lock->l_lock);
544 oldres = lock->l_resource;
545 if (oldres < newres) {
547 lock_res_nested(newres, LRT_NEW);
550 lock_res_nested(oldres, LRT_NEW);
552 LASSERT(memcmp(new_resid, &oldres->lr_name,
553 sizeof oldres->lr_name) != 0);
554 lock->l_resource = newres;
556 unlock_res_and_lock(lock);
558 /* ...and the flowers are still standing! */
559 lu_ref_del(&oldres->lr_reference, "lock", lock);
560 ldlm_resource_putref(oldres);
565 /** \defgroup ldlm_handles LDLM HANDLES
566 * Ways to get hold of locks without any addresses.
571 * Fills in handle for LDLM lock \a lock into supplied \a lockh
572 * Does not take any references.
574 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
576 lockh->cookie = lock->l_handle.h_cookie;
578 EXPORT_SYMBOL(ldlm_lock2handle);
581 * Obtain a lock reference by handle.
583 * if \a flags: atomically get the lock and set the flags.
584 * Return NULL if flag already set
586 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
589 struct ldlm_lock *lock;
594 lock = class_handle2object(handle->cookie, NULL);
598 if (lock->l_export != NULL && lock->l_export->exp_failed) {
599 CDEBUG(D_INFO, "lock export failed: lock %p, exp %p\n",
600 lock, lock->l_export);
605 /* It's unlikely but possible that someone marked the lock as
606 * destroyed after we did handle2object on it */
607 if ((flags == 0) && !ldlm_is_destroyed(lock)) {
608 lu_ref_add(&lock->l_reference, "handle", current);
612 lock_res_and_lock(lock);
614 LASSERT(lock->l_resource != NULL);
616 lu_ref_add_atomic(&lock->l_reference, "handle", current);
617 if (unlikely(ldlm_is_destroyed(lock))) {
618 unlock_res_and_lock(lock);
619 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
624 /* If we're setting flags, make sure none of them are already set. */
626 if ((lock->l_flags & flags) != 0) {
627 unlock_res_and_lock(lock);
632 lock->l_flags |= flags;
635 unlock_res_and_lock(lock);
638 EXPORT_SYMBOL(__ldlm_handle2lock);
639 /** @} ldlm_handles */
642 * Fill in "on the wire" representation for given LDLM lock into supplied
643 * lock descriptor \a desc structure.
645 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
647 ldlm_res2desc(lock->l_resource, &desc->l_resource);
648 desc->l_req_mode = lock->l_req_mode;
649 desc->l_granted_mode = lock->l_granted_mode;
650 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
651 &lock->l_policy_data,
652 &desc->l_policy_data);
656 * Add a lock to list of conflicting locks to send AST to.
658 * Only add if we have not sent a blocking AST to the lock yet.
660 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
661 struct list_head *work_list)
663 if (!ldlm_is_ast_sent(lock)) {
664 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
665 ldlm_set_ast_sent(lock);
666 /* If the enqueuing client said so, tell the AST recipient to
667 * discard dirty data, rather than writing back. */
668 if (ldlm_is_ast_discard_data(new))
669 ldlm_set_discard_data(lock);
670 LASSERT(list_empty(&lock->l_bl_ast));
671 list_add(&lock->l_bl_ast, work_list);
673 LASSERT(lock->l_blocking_lock == NULL);
674 lock->l_blocking_lock = LDLM_LOCK_GET(new);
679 * Add a lock to list of just granted locks to send completion AST to.
681 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
682 struct list_head *work_list)
684 if (!ldlm_is_cp_reqd(lock)) {
685 ldlm_set_cp_reqd(lock);
686 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
687 LASSERT(list_empty(&lock->l_cp_ast));
688 list_add(&lock->l_cp_ast, work_list);
694 * Aggregator function to add AST work items into a list. Determines
695 * what sort of an AST work needs to be done and calls the proper
697 * Must be called with lr_lock held.
699 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
700 struct list_head *work_list)
703 check_res_locked(lock->l_resource);
705 ldlm_add_bl_work_item(lock, new, work_list);
707 ldlm_add_cp_work_item(lock, work_list);
712 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
713 * r/w reference type is determined by \a mode
714 * Calls ldlm_lock_addref_internal.
716 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
718 struct ldlm_lock *lock;
720 lock = ldlm_handle2lock(lockh);
721 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
722 ldlm_lock_addref_internal(lock, mode);
725 EXPORT_SYMBOL(ldlm_lock_addref);
729 * Add specified reader/writer reference to LDLM lock \a lock.
730 * r/w reference type is determined by \a mode
731 * Removes lock from LRU if it is there.
732 * Assumes the LDLM lock is already locked.
734 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
737 ldlm_lock_remove_from_lru(lock);
738 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
740 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
742 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
744 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
747 lu_ref_add_atomic(&lock->l_reference, "user", lock);
748 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
752 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
753 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
755 * \retval 0 success, lock was addref-ed
757 * \retval -EAGAIN lock is being canceled.
759 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
761 struct ldlm_lock *lock;
765 lock = ldlm_handle2lock(lockh);
767 lock_res_and_lock(lock);
768 if (lock->l_readers != 0 || lock->l_writers != 0 ||
769 !ldlm_is_cbpending(lock)) {
770 ldlm_lock_addref_internal_nolock(lock, mode);
773 unlock_res_and_lock(lock);
778 EXPORT_SYMBOL(ldlm_lock_addref_try);
781 * Add specified reader/writer reference to LDLM lock \a lock.
782 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
783 * Only called for local locks.
785 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
787 lock_res_and_lock(lock);
788 ldlm_lock_addref_internal_nolock(lock, mode);
789 unlock_res_and_lock(lock);
793 * Removes reader/writer reference for LDLM lock \a lock.
794 * Assumes LDLM lock is already locked.
795 * only called in ldlm_flock_destroy and for local locks.
796 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
797 * that cannot be placed in LRU.
799 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
802 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
803 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
804 LASSERT(lock->l_readers > 0);
805 lu_ref_del(&lock->l_reference, "reader", lock);
808 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
809 LASSERT(lock->l_writers > 0);
810 lu_ref_del(&lock->l_reference, "writer", lock);
814 lu_ref_del(&lock->l_reference, "user", lock);
815 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
819 * Removes reader/writer reference for LDLM lock \a lock.
820 * Locks LDLM lock first.
821 * If the lock is determined to be client lock on a client and r/w refcount
822 * drops to zero and the lock is not blocked, the lock is added to LRU lock
824 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
826 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
828 struct ldlm_namespace *ns;
831 lock_res_and_lock(lock);
833 ns = ldlm_lock_to_ns(lock);
835 ldlm_lock_decref_internal_nolock(lock, mode);
837 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
838 !lock->l_readers && !lock->l_writers) {
839 /* If this is a local lock on a server namespace and this was
840 * the last reference, cancel the lock.
842 * Group locks are special:
843 * They must not go in LRU, but they are not called back
844 * like non-group locks, instead they are manually released.
845 * They have an l_writers reference which they keep until
846 * they are manually released, so we remove them when they have
847 * no more reader or writer references. - LU-6368 */
848 ldlm_set_cbpending(lock);
851 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
852 /* If we received a blocked AST and this was the last reference,
853 * run the callback. */
854 if (ldlm_is_ns_srv(lock) && lock->l_export)
855 CERROR("FL_CBPENDING set on non-local lock--just a "
858 LDLM_DEBUG(lock, "final decref done on cbpending lock");
860 LDLM_LOCK_GET(lock); /* dropped by bl thread */
861 ldlm_lock_remove_from_lru(lock);
862 unlock_res_and_lock(lock);
864 if (ldlm_is_fail_loc(lock))
865 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
867 if (ldlm_is_atomic_cb(lock) ||
868 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
869 ldlm_handle_bl_callback(ns, NULL, lock);
870 } else if (ns_is_client(ns) &&
871 !lock->l_readers && !lock->l_writers &&
872 !ldlm_is_no_lru(lock) &&
873 !ldlm_is_bl_ast(lock) &&
874 !ldlm_is_converting(lock)) {
876 LDLM_DEBUG(lock, "add lock into lru list");
878 /* If this is a client-side namespace and this was the last
879 * reference, put it on the LRU. */
880 ldlm_lock_add_to_lru(lock);
881 unlock_res_and_lock(lock);
883 if (ldlm_is_fail_loc(lock))
884 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
886 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
887 * are not supported by the server, otherwise, it is done on
889 if (!exp_connect_cancelset(lock->l_conn_export) &&
890 !ns_connect_lru_resize(ns))
891 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
893 LDLM_DEBUG(lock, "do not add lock into lru list");
894 unlock_res_and_lock(lock);
901 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
903 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
905 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
906 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
907 ldlm_lock_decref_internal(lock, mode);
910 EXPORT_SYMBOL(ldlm_lock_decref);
913 * Decrease reader/writer refcount for LDLM lock with handle
914 * \a lockh and mark it for subsequent cancellation once r/w refcount
915 * drops to zero instead of putting into LRU.
918 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
921 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
924 LASSERT(lock != NULL);
926 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
927 lock_res_and_lock(lock);
928 ldlm_set_cbpending(lock);
929 unlock_res_and_lock(lock);
930 ldlm_lock_decref_internal(lock, mode);
933 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
935 struct sl_insert_point {
936 struct list_head *res_link;
937 struct list_head *mode_link;
938 struct list_head *policy_link;
942 * Finds a position to insert the new lock into granted lock list.
944 * Used for locks eligible for skiplist optimization.
947 * queue [input]: the granted list where search acts on;
948 * req [input]: the lock whose position to be located;
949 * prev [output]: positions within 3 lists to insert @req to
953 * - ldlm_grant_lock_with_skiplist
955 static void search_granted_lock(struct list_head *queue,
956 struct ldlm_lock *req,
957 struct sl_insert_point *prev)
959 struct list_head *tmp;
960 struct ldlm_lock *lock, *mode_end, *policy_end;
963 list_for_each(tmp, queue) {
964 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
966 mode_end = list_entry(lock->l_sl_mode.prev,
967 struct ldlm_lock, l_sl_mode);
969 if (lock->l_req_mode != req->l_req_mode) {
970 /* jump to last lock of mode group */
971 tmp = &mode_end->l_res_link;
975 /* suitable mode group is found */
976 if (lock->l_resource->lr_type == LDLM_PLAIN) {
977 /* insert point is last lock of the mode group */
978 prev->res_link = &mode_end->l_res_link;
979 prev->mode_link = &mode_end->l_sl_mode;
980 prev->policy_link = &req->l_sl_policy;
983 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
986 list_entry(lock->l_sl_policy.prev,
990 if (lock->l_policy_data.l_inodebits.bits ==
991 req->l_policy_data.l_inodebits.bits) {
992 /* insert point is last lock of
993 * the policy group */
995 &policy_end->l_res_link;
997 &policy_end->l_sl_mode;
999 &policy_end->l_sl_policy;
1004 if (policy_end == mode_end)
1005 /* done with mode group */
1008 /* go to next policy group within mode group */
1009 tmp = policy_end->l_res_link.next;
1010 lock = list_entry(tmp, struct ldlm_lock,
1012 } /* loop over policy groups within the mode group */
1014 /* insert point is last lock of the mode group,
1015 * new policy group is started */
1016 prev->res_link = &mode_end->l_res_link;
1017 prev->mode_link = &mode_end->l_sl_mode;
1018 prev->policy_link = &req->l_sl_policy;
1022 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1027 /* insert point is last lock on the queue,
1028 * new mode group and new policy group are started */
1029 prev->res_link = queue->prev;
1030 prev->mode_link = &req->l_sl_mode;
1031 prev->policy_link = &req->l_sl_policy;
1037 * Add a lock into resource granted list after a position described by
1040 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1041 struct sl_insert_point *prev)
1043 struct ldlm_resource *res = lock->l_resource;
1046 check_res_locked(res);
1048 ldlm_resource_dump(D_INFO, res);
1049 LDLM_DEBUG(lock, "About to add lock:");
1051 if (ldlm_is_destroyed(lock)) {
1052 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1056 LASSERT(list_empty(&lock->l_res_link));
1057 LASSERT(list_empty(&lock->l_sl_mode));
1058 LASSERT(list_empty(&lock->l_sl_policy));
1061 * lock->link == prev->link means lock is first starting the group.
1062 * Don't re-add to itself to suppress kernel warnings.
1064 if (&lock->l_res_link != prev->res_link)
1065 list_add(&lock->l_res_link, prev->res_link);
1066 if (&lock->l_sl_mode != prev->mode_link)
1067 list_add(&lock->l_sl_mode, prev->mode_link);
1068 if (&lock->l_sl_policy != prev->policy_link)
1069 list_add(&lock->l_sl_policy, prev->policy_link);
1075 * Add a lock to granted list on a resource maintaining skiplist
1078 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1080 struct sl_insert_point prev;
1082 LASSERT(lock->l_req_mode == lock->l_granted_mode);
1084 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1085 ldlm_granted_list_add_lock(lock, &prev);
1089 * Perform lock granting bookkeeping.
1091 * Includes putting the lock into granted list and updating lock mode.
1093 * - ldlm_lock_enqueue
1094 * - ldlm_reprocess_queue
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_waiting, &data);
1424 LDLM_RESOURCE_DELREF(res);
1425 ldlm_resource_putref(res);
1428 ldlm_lock2handle(lock, lockh);
1429 if ((flags & LDLM_FL_LVB_READY) &&
1430 (!ldlm_is_lvb_ready(lock))) {
1431 __u64 wait_flags = LDLM_FL_LVB_READY |
1432 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1433 struct l_wait_info lwi;
1434 if (lock->l_completion_ast) {
1435 int err = lock->l_completion_ast(lock,
1436 LDLM_FL_WAIT_NOREPROC,
1439 if (flags & LDLM_FL_TEST_LOCK)
1440 LDLM_LOCK_RELEASE(lock);
1442 ldlm_lock_decref_internal(lock,
1449 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1450 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1452 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1453 l_wait_event(lock->l_waitq,
1454 lock->l_flags & wait_flags,
1456 if (!ldlm_is_lvb_ready(lock)) {
1457 if (flags & LDLM_FL_TEST_LOCK)
1458 LDLM_LOCK_RELEASE(lock);
1460 ldlm_lock_decref_internal(lock, mode);
1467 LDLM_DEBUG(lock, "matched (%llu %llu)",
1468 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1469 res_id->name[2] : policy->l_extent.start,
1470 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1471 res_id->name[3] : policy->l_extent.end);
1473 /* check user's security context */
1474 if (lock->l_conn_export &&
1475 sptlrpc_import_check_ctx(
1476 class_exp2cliimp(lock->l_conn_export))) {
1477 if (!(flags & LDLM_FL_TEST_LOCK))
1478 ldlm_lock_decref_internal(lock, mode);
1482 if (flags & LDLM_FL_TEST_LOCK)
1483 LDLM_LOCK_RELEASE(lock);
1485 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1486 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1487 "%llu/%llu (%llu %llu)", ns,
1488 type, mode, res_id->name[0], res_id->name[1],
1489 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1490 res_id->name[2] :policy->l_extent.start,
1491 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1492 res_id->name[3] : policy->l_extent.end);
1494 if (data.lmd_old != NULL)
1495 LDLM_LOCK_PUT(data.lmd_old);
1497 return rc ? mode : 0;
1499 EXPORT_SYMBOL(ldlm_lock_match);
1501 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1504 struct ldlm_lock *lock;
1505 enum ldlm_mode mode = 0;
1508 lock = ldlm_handle2lock(lockh);
1510 lock_res_and_lock(lock);
1511 if (LDLM_HAVE_MASK(lock, GONE))
1514 if (ldlm_is_cbpending(lock) &&
1515 lock->l_readers == 0 && lock->l_writers == 0)
1519 *bits = lock->l_policy_data.l_inodebits.bits;
1520 mode = lock->l_granted_mode;
1521 ldlm_lock_addref_internal_nolock(lock, mode);
1528 unlock_res_and_lock(lock);
1529 LDLM_LOCK_PUT(lock);
1533 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1535 /** The caller must guarantee that the buffer is large enough. */
1536 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1537 enum req_location loc, void *data, int size)
1542 LASSERT(data != NULL);
1545 switch (lock->l_lvb_type) {
1547 if (size == sizeof(struct ost_lvb)) {
1548 if (loc == RCL_CLIENT)
1549 lvb = req_capsule_client_swab_get(pill,
1551 lustre_swab_ost_lvb);
1553 lvb = req_capsule_server_swab_get(pill,
1555 lustre_swab_ost_lvb);
1556 if (unlikely(lvb == NULL)) {
1557 LDLM_ERROR(lock, "no LVB");
1561 memcpy(data, lvb, size);
1562 } else if (size == sizeof(struct ost_lvb_v1)) {
1563 struct ost_lvb *olvb = data;
1565 if (loc == RCL_CLIENT)
1566 lvb = req_capsule_client_swab_get(pill,
1568 lustre_swab_ost_lvb_v1);
1570 lvb = req_capsule_server_sized_swab_get(pill,
1572 lustre_swab_ost_lvb_v1);
1573 if (unlikely(lvb == NULL)) {
1574 LDLM_ERROR(lock, "no LVB");
1578 memcpy(data, lvb, size);
1579 olvb->lvb_mtime_ns = 0;
1580 olvb->lvb_atime_ns = 0;
1581 olvb->lvb_ctime_ns = 0;
1583 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1589 if (size == sizeof(struct lquota_lvb)) {
1590 if (loc == RCL_CLIENT)
1591 lvb = req_capsule_client_swab_get(pill,
1593 lustre_swab_lquota_lvb);
1595 lvb = req_capsule_server_swab_get(pill,
1597 lustre_swab_lquota_lvb);
1598 if (unlikely(lvb == NULL)) {
1599 LDLM_ERROR(lock, "no LVB");
1603 memcpy(data, lvb, size);
1605 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1614 if (loc == RCL_CLIENT)
1615 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1617 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1618 if (unlikely(lvb == NULL)) {
1619 LDLM_ERROR(lock, "no LVB");
1623 memcpy(data, lvb, size);
1626 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1627 libcfs_debug_dumpstack(NULL);
1635 * Create and fill in new LDLM lock with specified properties.
1636 * Returns a referenced lock
1638 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1639 const struct ldlm_res_id *res_id,
1640 enum ldlm_type type,
1641 enum ldlm_mode mode,
1642 const struct ldlm_callback_suite *cbs,
1643 void *data, __u32 lvb_len,
1644 enum lvb_type lvb_type)
1646 struct ldlm_lock *lock;
1647 struct ldlm_resource *res;
1651 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1653 RETURN(ERR_CAST(res));
1655 lock = ldlm_lock_new(res);
1657 RETURN(ERR_PTR(-ENOMEM));
1659 lock->l_req_mode = mode;
1660 lock->l_ast_data = data;
1661 lock->l_pid = current_pid();
1662 if (ns_is_server(ns))
1663 ldlm_set_ns_srv(lock);
1665 lock->l_blocking_ast = cbs->lcs_blocking;
1666 lock->l_completion_ast = cbs->lcs_completion;
1667 lock->l_glimpse_ast = cbs->lcs_glimpse;
1670 lock->l_tree_node = NULL;
1671 /* if this is the extent lock, allocate the interval tree node */
1672 if (type == LDLM_EXTENT)
1673 if (ldlm_interval_alloc(lock) == NULL)
1674 GOTO(out, rc = -ENOMEM);
1677 lock->l_lvb_len = lvb_len;
1678 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1679 if (lock->l_lvb_data == NULL)
1680 GOTO(out, rc = -ENOMEM);
1683 lock->l_lvb_type = lvb_type;
1684 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1685 GOTO(out, rc = -ENOENT);
1690 ldlm_lock_destroy(lock);
1691 LDLM_LOCK_RELEASE(lock);
1692 RETURN(ERR_PTR(rc));
1695 #ifdef HAVE_SERVER_SUPPORT
1696 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1699 struct ldlm_resource *res = lock->l_resource;
1700 enum ldlm_error rc = ELDLM_OK;
1701 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1702 ldlm_processing_policy policy;
1705 policy = ldlm_processing_policy_table[res->lr_type];
1707 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1708 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1709 res->lr_type != LDLM_FLOCK) {
1710 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1711 if (rc == -ERESTART)
1715 if (!list_empty(&rpc_list))
1716 ldlm_discard_bl_list(&rpc_list);
1723 * Enqueue (request) a lock.
1725 * Does not block. As a result of enqueue the lock would be put
1726 * into granted or waiting list.
1728 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1729 * set, skip all the enqueueing and delegate lock processing to intent policy
1732 enum ldlm_error ldlm_lock_enqueue(struct ldlm_namespace *ns,
1733 struct ldlm_lock **lockp,
1734 void *cookie, __u64 *flags)
1736 struct ldlm_lock *lock = *lockp;
1737 struct ldlm_resource *res = lock->l_resource;
1738 int local = ns_is_client(ldlm_res_to_ns(res));
1739 enum ldlm_error rc = ELDLM_OK;
1740 struct ldlm_interval *node = NULL;
1743 /* policies are not executed on the client or during replay */
1744 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1745 && !local && ns->ns_policy) {
1746 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1748 if (rc == ELDLM_LOCK_REPLACED) {
1749 /* The lock that was returned has already been granted,
1750 * and placed into lockp. If it's not the same as the
1751 * one we passed in, then destroy the old one and our
1752 * work here is done. */
1753 if (lock != *lockp) {
1754 ldlm_lock_destroy(lock);
1755 LDLM_LOCK_RELEASE(lock);
1757 *flags |= LDLM_FL_LOCK_CHANGED;
1759 } else if (rc != ELDLM_OK &&
1760 lock->l_req_mode == lock->l_granted_mode) {
1761 LASSERT(*flags & LDLM_FL_RESENT);
1762 /* It may happen that ns_policy returns an error in
1763 * resend case, object may be unlinked or just some
1764 * error occurs. It is unclear if lock reached the
1765 * client in the original reply, just leave the lock on
1766 * server, not returning it again to client. Due to
1767 * LU-6529, the server will not OOM. */
1769 } else if (rc != ELDLM_OK ||
1770 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1771 ldlm_lock_destroy(lock);
1776 if (*flags & LDLM_FL_RESENT) {
1777 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1778 * Set LOCK_CHANGED always.
1779 * Check if the lock is granted for BLOCK_GRANTED.
1780 * Take NO_TIMEOUT from the lock as it is inherited through
1781 * LDLM_FL_INHERIT_MASK */
1782 *flags |= LDLM_FL_LOCK_CHANGED;
1783 if (lock->l_req_mode != lock->l_granted_mode)
1784 *flags |= LDLM_FL_BLOCK_GRANTED;
1785 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1789 /* For a replaying lock, it might be already in granted list. So
1790 * unlinking the lock will cause the interval node to be freed, we
1791 * have to allocate the interval node early otherwise we can't regrant
1792 * this lock in the future. - jay */
1793 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1794 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1796 lock_res_and_lock(lock);
1797 if (local && lock->l_req_mode == lock->l_granted_mode) {
1798 /* The server returned a blocked lock, but it was granted
1799 * before we got a chance to actually enqueue it. We don't
1800 * need to do anything else. */
1801 *flags &= ~LDLM_FL_BLOCKED_MASK;
1802 GOTO(out, rc = ELDLM_OK);
1805 ldlm_resource_unlink_lock(lock);
1806 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1808 ldlm_lock_destroy_nolock(lock);
1809 GOTO(out, rc = -ENOMEM);
1812 INIT_LIST_HEAD(&node->li_group);
1813 ldlm_interval_attach(node, lock);
1817 /* Some flags from the enqueue want to make it into the AST, via the
1818 * lock's l_flags. */
1819 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1820 ldlm_set_ast_discard_data(lock);
1821 if (*flags & LDLM_FL_TEST_LOCK)
1822 ldlm_set_test_lock(lock);
1823 if (*flags & LDLM_FL_COS_INCOMPAT)
1824 ldlm_set_cos_incompat(lock);
1825 if (*flags & LDLM_FL_COS_ENABLED)
1826 ldlm_set_cos_enabled(lock);
1828 /* This distinction between local lock trees is very important; a client
1829 * namespace only has information about locks taken by that client, and
1830 * thus doesn't have enough information to decide for itself if it can
1831 * be granted (below). In this case, we do exactly what the server
1832 * tells us to do, as dictated by the 'flags'.
1834 * We do exactly the same thing during recovery, when the server is
1835 * more or less trusting the clients not to lie.
1837 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1840 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1841 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1843 ldlm_grant_lock(lock, NULL);
1844 GOTO(out, rc = ELDLM_OK);
1845 #ifdef HAVE_SERVER_SUPPORT
1846 } else if (*flags & LDLM_FL_REPLAY) {
1847 if (*flags & LDLM_FL_BLOCK_WAIT) {
1848 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1849 GOTO(out, rc = ELDLM_OK);
1850 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1851 ldlm_grant_lock(lock, NULL);
1852 GOTO(out, rc = ELDLM_OK);
1854 /* If no flags, fall through to normal enqueue path. */
1857 rc = ldlm_lock_enqueue_helper(lock, flags);
1861 CERROR("This is client-side-only module, cannot handle "
1862 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1868 unlock_res_and_lock(lock);
1870 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1874 #ifdef HAVE_SERVER_SUPPORT
1876 * Iterate through all waiting locks on a given resource queue and attempt to
1879 * Must be called with resource lock held.
1881 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1882 struct list_head *work_list,
1883 enum ldlm_process_intention intention)
1885 struct list_head *tmp, *pos;
1886 ldlm_processing_policy policy;
1888 int rc = LDLM_ITER_CONTINUE;
1889 enum ldlm_error err;
1890 struct list_head bl_ast_list = LIST_HEAD_INIT(bl_ast_list);
1893 check_res_locked(res);
1895 policy = ldlm_processing_policy_table[res->lr_type];
1897 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1898 intention == LDLM_PROCESS_RECOVERY);
1901 list_for_each_safe(tmp, pos, queue) {
1902 struct ldlm_lock *pending;
1903 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1905 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1907 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1910 rc = policy(pending, &flags, intention, &err, &rpc_list);
1911 if (pending->l_granted_mode == pending->l_req_mode ||
1912 res->lr_type == LDLM_FLOCK) {
1913 list_splice(&rpc_list, work_list);
1915 list_splice(&rpc_list, &bl_ast_list);
1918 * When this is called from recovery done, we always want
1919 * to scan the whole list no matter what 'rc' is returned.
1921 if (rc != LDLM_ITER_CONTINUE &&
1922 intention == LDLM_PROCESS_RESCAN)
1926 if (!list_empty(&bl_ast_list)) {
1929 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
1933 if (rc == -ERESTART)
1937 if (!list_empty(&bl_ast_list))
1938 ldlm_discard_bl_list(&bl_ast_list);
1940 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
1944 * Conflicting locks are detected for a lock to be enqueued, add the lock
1945 * into waiting list and send blocking ASTs to the conflicting locks.
1947 * \param[in] lock The lock to be enqueued.
1948 * \param[out] flags Lock flags for the lock to be enqueued.
1949 * \param[in] rpc_list Conflicting locks list.
1951 * \retval -ERESTART: Some lock was instantly canceled while sending
1952 * blocking ASTs, caller needs to re-check conflicting
1954 * \retval -EAGAIN: Lock was destroyed, caller should return error.
1955 * \reval 0: Lock is successfully added in waiting list.
1957 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
1958 struct list_head *rpc_list)
1960 struct ldlm_resource *res = lock->l_resource;
1964 check_res_locked(res);
1966 /* If either of the compat_queue()s returned failure, then we
1967 * have ASTs to send and must go onto the waiting list.
1969 * bug 2322: we used to unlink and re-add here, which was a
1970 * terrible folly -- if we goto restart, we could get
1971 * re-ordered! Causes deadlock, because ASTs aren't sent! */
1972 if (list_empty(&lock->l_res_link))
1973 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1976 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
1979 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
1980 !ns_is_client(ldlm_res_to_ns(res)))
1981 class_fail_export(lock->l_export);
1984 if (rc == -ERESTART) {
1985 /* 15715: The lock was granted and destroyed after
1986 * resource lock was dropped. Interval node was freed
1987 * in ldlm_lock_destroy. Anyway, this always happens
1988 * when a client is being evicted. So it would be
1989 * ok to return an error. -jay */
1990 if (ldlm_is_destroyed(lock))
1993 /* lock was granted while resource was unlocked. */
1994 if (lock->l_granted_mode == lock->l_req_mode) {
1995 /* bug 11300: if the lock has been granted,
1996 * break earlier because otherwise, we will go
1997 * to restart and ldlm_resource_unlink will be
1998 * called and it causes the interval node to be
1999 * freed. Then we will fail at
2000 * ldlm_extent_add_lock() */
2001 *flags &= ~LDLM_FL_BLOCKED_MASK;
2007 *flags |= LDLM_FL_BLOCK_GRANTED;
2013 * Discard all AST work items from list.
2015 * If for whatever reason we do not want to send ASTs to conflicting locks
2016 * anymore, disassemble the list with this function.
2018 void ldlm_discard_bl_list(struct list_head *bl_list)
2020 struct list_head *tmp, *pos;
2023 list_for_each_safe(pos, tmp, bl_list) {
2024 struct ldlm_lock *lock =
2025 list_entry(pos, struct ldlm_lock, l_bl_ast);
2027 list_del_init(&lock->l_bl_ast);
2028 LASSERT(ldlm_is_ast_sent(lock));
2029 ldlm_clear_ast_sent(lock);
2030 LASSERT(lock->l_bl_ast_run == 0);
2031 LASSERT(lock->l_blocking_lock);
2032 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
2033 lock->l_blocking_lock = NULL;
2034 LDLM_LOCK_RELEASE(lock);
2042 * Process a call to blocking AST callback for a lock in ast_work list
2045 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2047 struct ldlm_cb_set_arg *arg = opaq;
2048 struct ldlm_lock_desc d;
2050 struct ldlm_lock *lock;
2053 if (list_empty(arg->list))
2056 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2058 /* nobody should touch l_bl_ast */
2059 lock_res_and_lock(lock);
2060 list_del_init(&lock->l_bl_ast);
2062 LASSERT(ldlm_is_ast_sent(lock));
2063 LASSERT(lock->l_bl_ast_run == 0);
2064 LASSERT(lock->l_blocking_lock);
2065 lock->l_bl_ast_run++;
2066 unlock_res_and_lock(lock);
2068 ldlm_lock2desc(lock->l_blocking_lock, &d);
2069 /* copy blocking lock ibits in cancel_bits as well,
2070 * new client may use them for lock convert and it is
2071 * important to use new field to convert locks from
2074 d.l_policy_data.l_inodebits.cancel_bits =
2075 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2077 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2078 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
2079 lock->l_blocking_lock = NULL;
2080 LDLM_LOCK_RELEASE(lock);
2086 * Process a call to completion AST callback for a lock in ast_work list
2089 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2091 struct ldlm_cb_set_arg *arg = opaq;
2093 struct ldlm_lock *lock;
2094 ldlm_completion_callback completion_callback;
2097 if (list_empty(arg->list))
2100 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2102 /* It's possible to receive a completion AST before we've set
2103 * the l_completion_ast pointer: either because the AST arrived
2104 * before the reply, or simply because there's a small race
2105 * window between receiving the reply and finishing the local
2106 * enqueue. (bug 842)
2108 * This can't happen with the blocking_ast, however, because we
2109 * will never call the local blocking_ast until we drop our
2110 * reader/writer reference, which we won't do until we get the
2111 * reply and finish enqueueing. */
2113 /* nobody should touch l_cp_ast */
2114 lock_res_and_lock(lock);
2115 list_del_init(&lock->l_cp_ast);
2116 LASSERT(ldlm_is_cp_reqd(lock));
2117 /* save l_completion_ast since it can be changed by
2118 * mds_intent_policy(), see bug 14225 */
2119 completion_callback = lock->l_completion_ast;
2120 ldlm_clear_cp_reqd(lock);
2121 unlock_res_and_lock(lock);
2123 if (completion_callback != NULL)
2124 rc = completion_callback(lock, 0, (void *)arg);
2125 LDLM_LOCK_RELEASE(lock);
2131 * Process a call to revocation AST callback for a lock in ast_work list
2134 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2136 struct ldlm_cb_set_arg *arg = opaq;
2137 struct ldlm_lock_desc desc;
2139 struct ldlm_lock *lock;
2142 if (list_empty(arg->list))
2145 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2146 list_del_init(&lock->l_rk_ast);
2148 /* the desc just pretend to exclusive */
2149 ldlm_lock2desc(lock, &desc);
2150 desc.l_req_mode = LCK_EX;
2151 desc.l_granted_mode = 0;
2153 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2154 LDLM_LOCK_RELEASE(lock);
2160 * Process a call to glimpse AST callback for a lock in ast_work list
2162 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2164 struct ldlm_cb_set_arg *arg = opaq;
2165 struct ldlm_glimpse_work *gl_work;
2166 struct ldlm_lock *lock;
2170 if (list_empty(arg->list))
2173 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2175 list_del_init(&gl_work->gl_list);
2177 lock = gl_work->gl_lock;
2179 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2180 arg->gl_desc = gl_work->gl_desc;
2181 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2182 arg->gl_interpret_data = gl_work->gl_interpret_data;
2184 /* invoke the actual glimpse callback */
2185 if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
2188 LDLM_LOCK_RELEASE(lock);
2189 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2190 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2192 OBD_FREE_PTR(gl_work);
2198 * Process list of locks in need of ASTs being sent.
2200 * Used on server to send multiple ASTs together instead of sending one by
2203 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2204 ldlm_desc_ast_t ast_type)
2206 struct ldlm_cb_set_arg *arg;
2207 set_producer_func work_ast_lock;
2210 if (list_empty(rpc_list))
2217 atomic_set(&arg->restart, 0);
2218 arg->list = rpc_list;
2221 case LDLM_WORK_BL_AST:
2222 arg->type = LDLM_BL_CALLBACK;
2223 work_ast_lock = ldlm_work_bl_ast_lock;
2225 case LDLM_WORK_CP_AST:
2226 arg->type = LDLM_CP_CALLBACK;
2227 work_ast_lock = ldlm_work_cp_ast_lock;
2229 case LDLM_WORK_REVOKE_AST:
2230 arg->type = LDLM_BL_CALLBACK;
2231 work_ast_lock = ldlm_work_revoke_ast_lock;
2233 case LDLM_WORK_GL_AST:
2234 arg->type = LDLM_GL_CALLBACK;
2235 work_ast_lock = ldlm_work_gl_ast_lock;
2241 /* We create a ptlrpc request set with flow control extension.
2242 * This request set will use the work_ast_lock function to produce new
2243 * requests and will send a new request each time one completes in order
2244 * to keep the number of requests in flight to ns_max_parallel_ast */
2245 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2246 work_ast_lock, arg);
2247 if (arg->set == NULL)
2248 GOTO(out, rc = -ENOMEM);
2250 ptlrpc_set_wait(arg->set);
2251 ptlrpc_set_destroy(arg->set);
2253 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2261 * Try to grant all waiting locks on a resource.
2263 * Calls ldlm_reprocess_queue on waiting queue.
2265 * Typically called after some resource locks are cancelled to see
2266 * if anything could be granted as a result of the cancellation.
2268 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2269 enum ldlm_process_intention intention)
2271 struct list_head rpc_list;
2272 #ifdef HAVE_SERVER_SUPPORT
2273 struct obd_device *obd;
2277 INIT_LIST_HEAD(&rpc_list);
2278 /* Local lock trees don't get reprocessed. */
2279 if (ns_is_client(ldlm_res_to_ns(res))) {
2284 /* Disable reprocess during lock replay stage but allow during
2285 * request replay stage.
2287 obd = ldlm_res_to_ns(res)->ns_obd;
2288 if (obd->obd_recovering &&
2289 atomic_read(&obd->obd_req_replay_clients) == 0)
2293 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list, intention);
2296 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2298 if (rc == -ERESTART) {
2299 LASSERT(list_empty(&rpc_list));
2305 INIT_LIST_HEAD(&rpc_list);
2306 if (!ns_is_client(ldlm_res_to_ns(res))) {
2307 CERROR("This is client-side-only module, cannot handle "
2308 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2315 void ldlm_reprocess_all(struct ldlm_resource *res)
2317 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN);
2319 EXPORT_SYMBOL(ldlm_reprocess_all);
2321 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2322 struct hlist_node *hnode, void *arg)
2324 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2326 /* This is only called once after recovery done. LU-8306. */
2327 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY);
2332 * Iterate through all resources on a namespace attempting to grant waiting
2335 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2340 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2341 ldlm_reprocess_res, NULL, 0);
2347 * Helper function to call blocking AST for LDLM lock \a lock in a
2348 * "cancelling" mode.
2350 void ldlm_cancel_callback(struct ldlm_lock *lock)
2352 check_res_locked(lock->l_resource);
2353 if (!ldlm_is_cancel(lock)) {
2354 ldlm_set_cancel(lock);
2355 if (lock->l_blocking_ast) {
2356 unlock_res_and_lock(lock);
2357 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2359 lock_res_and_lock(lock);
2361 LDLM_DEBUG(lock, "no blocking ast");
2364 /* only canceller can set bl_done bit */
2365 ldlm_set_bl_done(lock);
2366 wake_up_all(&lock->l_waitq);
2367 } else if (!ldlm_is_bl_done(lock)) {
2368 struct l_wait_info lwi = { 0 };
2370 /* The lock is guaranteed to have been canceled once
2371 * returning from this function. */
2372 unlock_res_and_lock(lock);
2373 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
2374 lock_res_and_lock(lock);
2379 * Remove skiplist-enabled LDLM lock \a req from granted list
2381 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2383 if (req->l_resource->lr_type != LDLM_PLAIN &&
2384 req->l_resource->lr_type != LDLM_IBITS)
2387 list_del_init(&req->l_sl_policy);
2388 list_del_init(&req->l_sl_mode);
2392 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2394 void ldlm_lock_cancel(struct ldlm_lock *lock)
2396 struct ldlm_resource *res;
2397 struct ldlm_namespace *ns;
2400 lock_res_and_lock(lock);
2402 res = lock->l_resource;
2403 ns = ldlm_res_to_ns(res);
2405 /* Please do not, no matter how tempting, remove this LBUG without
2406 * talking to me first. -phik */
2407 if (lock->l_readers || lock->l_writers) {
2408 LDLM_ERROR(lock, "lock still has references");
2409 unlock_res_and_lock(lock);
2413 if (ldlm_is_waited(lock))
2414 ldlm_del_waiting_lock(lock);
2416 /* Releases cancel callback. */
2417 ldlm_cancel_callback(lock);
2419 /* Yes, second time, just in case it was added again while we were
2420 * running with no res lock in ldlm_cancel_callback */
2421 if (ldlm_is_waited(lock))
2422 ldlm_del_waiting_lock(lock);
2424 ldlm_resource_unlink_lock(lock);
2425 ldlm_lock_destroy_nolock(lock);
2427 if (lock->l_granted_mode == lock->l_req_mode)
2428 ldlm_pool_del(&ns->ns_pool, lock);
2430 /* Make sure we will not be called again for same lock what is possible
2431 * if not to zero out lock->l_granted_mode */
2432 lock->l_granted_mode = LCK_MINMODE;
2433 unlock_res_and_lock(lock);
2437 EXPORT_SYMBOL(ldlm_lock_cancel);
2440 * Set opaque data into the lock that only makes sense to upper layer.
2442 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2444 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2449 if (lock->l_ast_data == NULL)
2450 lock->l_ast_data = data;
2451 if (lock->l_ast_data == data)
2453 LDLM_LOCK_PUT(lock);
2457 EXPORT_SYMBOL(ldlm_lock_set_data);
2459 struct export_cl_data {
2460 struct obd_export *ecl_exp;
2464 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2465 struct ldlm_lock *lock,
2466 struct export_cl_data *ecl)
2468 struct ldlm_resource *res;
2470 res = ldlm_resource_getref(lock->l_resource);
2472 ldlm_lvbo_update(res, lock, NULL, 1);
2473 ldlm_lock_cancel(lock);
2474 if (!exp->exp_obd->obd_stopping)
2475 ldlm_reprocess_all(res);
2476 ldlm_resource_putref(res);
2479 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2480 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2481 exp, ecl->ecl_loop);
2486 * Iterator function for ldlm_export_cancel_locks.
2487 * Cancels passed locks.
2490 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2491 struct hlist_node *hnode, void *data)
2494 struct export_cl_data *ecl = (struct export_cl_data *)data;
2495 struct obd_export *exp = ecl->ecl_exp;
2496 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2498 LDLM_LOCK_GET(lock);
2499 ldlm_cancel_lock_for_export(exp, lock, ecl);
2500 LDLM_LOCK_RELEASE(lock);
2506 * Cancel all blocked locks for given export.
2508 * Typically called on client disconnection/eviction
2510 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2512 struct export_cl_data ecl = {
2517 while (!list_empty(&exp->exp_bl_list)) {
2518 struct ldlm_lock *lock;
2520 spin_lock_bh(&exp->exp_bl_list_lock);
2521 if (!list_empty(&exp->exp_bl_list)) {
2522 lock = list_entry(exp->exp_bl_list.next,
2523 struct ldlm_lock, l_exp_list);
2524 LDLM_LOCK_GET(lock);
2525 list_del_init(&lock->l_exp_list);
2529 spin_unlock_bh(&exp->exp_bl_list_lock);
2534 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2535 LDLM_LOCK_RELEASE(lock);
2538 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2539 "left on hash table %d.\n", exp, ecl.ecl_loop,
2540 atomic_read(&exp->exp_lock_hash->hs_count));
2542 return ecl.ecl_loop;
2546 * Cancel all locks for given export.
2548 * Typically called after client disconnection/eviction
2550 int ldlm_export_cancel_locks(struct obd_export *exp)
2552 struct export_cl_data ecl = {
2557 cfs_hash_for_each_empty(exp->exp_lock_hash,
2558 ldlm_cancel_locks_for_export_cb, &ecl);
2560 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2561 "left on hash table %d.\n", exp, ecl.ecl_loop,
2562 atomic_read(&exp->exp_lock_hash->hs_count));
2564 if (ecl.ecl_loop > 0 &&
2565 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2566 exp->exp_obd->obd_stopping)
2567 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2569 return ecl.ecl_loop;
2573 * Downgrade an PW/EX lock to COS mode.
2575 * A lock mode convertion from PW/EX mode to less conflict mode. The
2576 * convertion may fail if lock was canceled before downgrade, but it doesn't
2577 * indicate any problem, because such lock has no reader or writer, and will
2579 * Used by Commit on Sharing (COS) code only for now.
2581 * \param lock A lock to convert
2582 * \param new_mode new lock mode
2584 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2588 LASSERT(new_mode == LCK_COS);
2590 lock_res_and_lock(lock);
2592 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2593 unlock_res_and_lock(lock);
2595 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2596 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2600 ldlm_resource_unlink_lock(lock);
2602 * Remove the lock from pool as it will be added again in
2603 * ldlm_grant_lock() called below.
2605 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2606 lock->l_req_mode = new_mode;
2607 ldlm_grant_lock(lock, NULL);
2609 unlock_res_and_lock(lock);
2611 ldlm_reprocess_all(lock->l_resource);
2615 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2618 * Print lock with lock handle \a lockh description into debug log.
2620 * Used when printing all locks on a resource for debug purposes.
2622 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2624 struct ldlm_lock *lock;
2626 if (!((libcfs_debug | D_ERROR) & level))
2629 lock = ldlm_handle2lock(lockh);
2633 LDLM_DEBUG_LIMIT(level, lock, "###");
2635 LDLM_LOCK_PUT(lock);
2637 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2640 * Print lock information with custom message into debug log.
2643 void _ldlm_lock_debug(struct ldlm_lock *lock,
2644 struct libcfs_debug_msg_data *msgdata,
2645 const char *fmt, ...)
2648 struct obd_export *exp = lock->l_export;
2649 struct ldlm_resource *resource = NULL;
2650 char *nid = "local";
2652 /* on server-side resource of lock doesn't change */
2653 if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2654 if (lock->l_resource != NULL)
2655 resource = ldlm_resource_getref(lock->l_resource);
2656 } else if (spin_trylock(&lock->l_lock)) {
2657 if (lock->l_resource != NULL)
2658 resource = ldlm_resource_getref(lock->l_resource);
2659 spin_unlock(&lock->l_lock);
2662 va_start(args, fmt);
2664 if (exp && exp->exp_connection) {
2665 nid = obd_export_nid2str(exp);
2666 } else if (exp && exp->exp_obd != NULL) {
2667 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2668 nid = obd_import_nid2str(imp);
2671 if (resource == NULL) {
2672 libcfs_debug_vmsg2(msgdata, fmt, args,
2673 " ns: \?\? lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2674 "res: \?\? rrc=\?\? type: \?\?\? flags: %#llx nid: %s "
2675 "remote: %#llx expref: %d pid: %u timeout: %lld "
2678 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2679 lock->l_readers, lock->l_writers,
2680 ldlm_lockname[lock->l_granted_mode],
2681 ldlm_lockname[lock->l_req_mode],
2682 lock->l_flags, nid, lock->l_remote_handle.cookie,
2683 exp ? atomic_read(&exp->exp_refcount) : -99,
2684 lock->l_pid, lock->l_callback_timeout, lock->l_lvb_type);
2689 switch (resource->lr_type) {
2691 libcfs_debug_vmsg2(msgdata, fmt, args,
2692 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2693 "res: "DLDLMRES" rrc: %d type: %s [%llu->%llu] "
2694 "(req %llu->%llu) flags: %#llx nid: %s remote: "
2695 "%#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2696 ldlm_lock_to_ns_name(lock), lock,
2697 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2698 lock->l_readers, lock->l_writers,
2699 ldlm_lockname[lock->l_granted_mode],
2700 ldlm_lockname[lock->l_req_mode],
2702 atomic_read(&resource->lr_refcount),
2703 ldlm_typename[resource->lr_type],
2704 lock->l_policy_data.l_extent.start,
2705 lock->l_policy_data.l_extent.end,
2706 lock->l_req_extent.start, lock->l_req_extent.end,
2707 lock->l_flags, nid, lock->l_remote_handle.cookie,
2708 exp ? atomic_read(&exp->exp_refcount) : -99,
2709 lock->l_pid, lock->l_callback_timeout,
2714 libcfs_debug_vmsg2(msgdata, fmt, args,
2715 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2716 "res: "DLDLMRES" rrc: %d type: %s pid: %d "
2717 "[%llu->%llu] flags: %#llx nid: %s "
2718 "remote: %#llx expref: %d pid: %u timeout: %lld\n",
2719 ldlm_lock_to_ns_name(lock), lock,
2720 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2721 lock->l_readers, lock->l_writers,
2722 ldlm_lockname[lock->l_granted_mode],
2723 ldlm_lockname[lock->l_req_mode],
2725 atomic_read(&resource->lr_refcount),
2726 ldlm_typename[resource->lr_type],
2727 lock->l_policy_data.l_flock.pid,
2728 lock->l_policy_data.l_flock.start,
2729 lock->l_policy_data.l_flock.end,
2730 lock->l_flags, nid, lock->l_remote_handle.cookie,
2731 exp ? atomic_read(&exp->exp_refcount) : -99,
2732 lock->l_pid, lock->l_callback_timeout);
2736 libcfs_debug_vmsg2(msgdata, fmt, args,
2737 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2738 "res: "DLDLMRES" bits %#llx/%#llx rrc: %d type: %s "
2739 "flags: %#llx nid: %s remote: %#llx expref: %d "
2740 "pid: %u timeout: %lld lvb_type: %d\n",
2741 ldlm_lock_to_ns_name(lock),
2742 lock, lock->l_handle.h_cookie,
2743 atomic_read(&lock->l_refc),
2744 lock->l_readers, lock->l_writers,
2745 ldlm_lockname[lock->l_granted_mode],
2746 ldlm_lockname[lock->l_req_mode],
2748 lock->l_policy_data.l_inodebits.bits,
2749 lock->l_policy_data.l_inodebits.try_bits,
2750 atomic_read(&resource->lr_refcount),
2751 ldlm_typename[resource->lr_type],
2752 lock->l_flags, nid, lock->l_remote_handle.cookie,
2753 exp ? atomic_read(&exp->exp_refcount) : -99,
2754 lock->l_pid, lock->l_callback_timeout,
2759 libcfs_debug_vmsg2(msgdata, fmt, args,
2760 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2761 "res: "DLDLMRES" rrc: %d type: %s flags: %#llx "
2762 "nid: %s remote: %#llx expref: %d pid: %u "
2763 "timeout: %lld lvb_type: %d\n",
2764 ldlm_lock_to_ns_name(lock),
2765 lock, lock->l_handle.h_cookie,
2766 atomic_read(&lock->l_refc),
2767 lock->l_readers, lock->l_writers,
2768 ldlm_lockname[lock->l_granted_mode],
2769 ldlm_lockname[lock->l_req_mode],
2771 atomic_read(&resource->lr_refcount),
2772 ldlm_typename[resource->lr_type],
2773 lock->l_flags, nid, lock->l_remote_handle.cookie,
2774 exp ? atomic_read(&exp->exp_refcount) : -99,
2775 lock->l_pid, lock->l_callback_timeout,
2780 ldlm_resource_putref(resource);
2782 EXPORT_SYMBOL(_ldlm_lock_debug);