4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_lock.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <libcfs/libcfs.h>
42 #include <lustre_swab.h>
43 #include <obd_class.h>
45 #include "ldlm_internal.h"
47 struct kmem_cache *ldlm_glimpse_work_kmem;
48 EXPORT_SYMBOL(ldlm_glimpse_work_kmem);
51 char *ldlm_lockname[] = {
59 [LCK_GROUP] = "GROUP",
62 EXPORT_SYMBOL(ldlm_lockname);
64 char *ldlm_typename[] = {
66 [LDLM_EXTENT] = "EXT",
71 static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = {
72 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_wire_to_local,
73 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_wire_to_local,
74 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_wire_to_local,
75 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_wire_to_local,
78 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
79 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_local_to_wire,
80 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_local_to_wire,
81 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_local_to_wire,
82 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_local_to_wire,
86 * Converts lock policy from local format to on the wire lock_desc format
88 void ldlm_convert_policy_to_wire(enum ldlm_type type,
89 const union ldlm_policy_data *lpolicy,
90 union ldlm_wire_policy_data *wpolicy)
92 ldlm_policy_local_to_wire_t convert;
94 convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
96 convert(lpolicy, wpolicy);
100 * Converts lock policy from on the wire lock_desc format to local format
102 void ldlm_convert_policy_to_local(struct obd_export *exp, enum ldlm_type type,
103 const union ldlm_wire_policy_data *wpolicy,
104 union ldlm_policy_data *lpolicy)
106 ldlm_policy_wire_to_local_t convert;
108 convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE];
110 convert(wpolicy, lpolicy);
113 const char *ldlm_it2str(enum ldlm_intent_flags it)
120 case (IT_OPEN | IT_CREAT):
133 CERROR("Unknown intent 0x%08x\n", it);
137 EXPORT_SYMBOL(ldlm_it2str);
139 extern struct kmem_cache *ldlm_lock_slab;
141 #ifdef HAVE_SERVER_SUPPORT
142 static ldlm_processing_policy ldlm_processing_policy_table[] = {
143 [LDLM_PLAIN] = ldlm_process_plain_lock,
144 [LDLM_EXTENT] = ldlm_process_extent_lock,
145 [LDLM_FLOCK] = ldlm_process_flock_lock,
146 [LDLM_IBITS] = ldlm_process_inodebits_lock,
149 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
151 return ldlm_processing_policy_table[res->lr_type];
153 EXPORT_SYMBOL(ldlm_get_processing_policy);
154 #endif /* HAVE_SERVER_SUPPORT */
156 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
160 EXPORT_SYMBOL(ldlm_register_intent);
163 * REFCOUNTED LOCK OBJECTS
168 * Get a reference on a lock.
170 * Lock refcounts, during creation:
171 * - one special one for allocation, dec'd only once in destroy
172 * - one for being a lock that's in-use
173 * - one for the addref associated with a new lock
175 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
177 atomic_inc(&lock->l_refc);
180 EXPORT_SYMBOL(ldlm_lock_get);
183 * Release lock reference.
185 * Also frees the lock if it was last reference.
187 void ldlm_lock_put(struct ldlm_lock *lock)
191 LASSERT(lock->l_resource != LP_POISON);
192 LASSERT(atomic_read(&lock->l_refc) > 0);
193 if (atomic_dec_and_test(&lock->l_refc)) {
194 struct ldlm_resource *res;
197 "final lock_put on destroyed lock, freeing it.");
199 res = lock->l_resource;
200 LASSERT(ldlm_is_destroyed(lock));
201 LASSERT(list_empty(&lock->l_exp_list));
202 LASSERT(list_empty(&lock->l_res_link));
203 LASSERT(list_empty(&lock->l_pending_chain));
205 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
207 lu_ref_del(&res->lr_reference, "lock", lock);
208 ldlm_resource_putref(res);
209 lock->l_resource = NULL;
210 if (lock->l_export) {
211 class_export_lock_put(lock->l_export, lock);
212 lock->l_export = NULL;
215 if (lock->l_lvb_data != NULL)
216 OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
218 ldlm_interval_free(ldlm_interval_detach(lock));
219 lu_ref_fini(&lock->l_reference);
220 OBD_FREE_RCU(lock, sizeof(*lock), &lock->l_handle);
225 EXPORT_SYMBOL(ldlm_lock_put);
228 * Removes LDLM lock \a lock from LRU. Assumes LRU is already locked.
230 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
233 if (!list_empty(&lock->l_lru)) {
234 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
236 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
237 if (ns->ns_last_pos == &lock->l_lru)
238 ns->ns_last_pos = lock->l_lru.prev;
239 list_del_init(&lock->l_lru);
240 LASSERT(ns->ns_nr_unused > 0);
248 * Removes LDLM lock \a lock from LRU. Obtains the LRU lock first.
250 * If \a last_use is non-zero, it will remove the lock from LRU only if
251 * it matches lock's l_last_used.
253 * \retval 0 if \a last_use is set, the lock is not in LRU list or \a last_use
254 * doesn't match lock's l_last_used;
255 * otherwise, the lock hasn't been in the LRU list.
256 * \retval 1 the lock was in LRU list and removed.
258 int ldlm_lock_remove_from_lru_check(struct ldlm_lock *lock, ktime_t last_use)
260 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
264 if (ldlm_is_ns_srv(lock)) {
265 LASSERT(list_empty(&lock->l_lru));
269 spin_lock(&ns->ns_lock);
270 if (!ktime_compare(last_use, ktime_set(0, 0)) ||
271 !ktime_compare(last_use, lock->l_last_used))
272 rc = ldlm_lock_remove_from_lru_nolock(lock);
273 spin_unlock(&ns->ns_lock);
279 * Adds LDLM lock \a lock to namespace LRU. Assumes LRU is already locked.
281 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
283 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
285 lock->l_last_used = ktime_get();
286 LASSERT(list_empty(&lock->l_lru));
287 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
288 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
289 LASSERT(ns->ns_nr_unused >= 0);
294 * Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks
297 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
299 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
302 spin_lock(&ns->ns_lock);
303 ldlm_lock_add_to_lru_nolock(lock);
304 spin_unlock(&ns->ns_lock);
309 * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
310 * the LRU. Performs necessary LRU locking
312 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
314 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
317 if (ldlm_is_ns_srv(lock)) {
318 LASSERT(list_empty(&lock->l_lru));
323 spin_lock(&ns->ns_lock);
324 if (!list_empty(&lock->l_lru)) {
325 ldlm_lock_remove_from_lru_nolock(lock);
326 ldlm_lock_add_to_lru_nolock(lock);
328 spin_unlock(&ns->ns_lock);
333 * Helper to destroy a locked lock.
335 * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
336 * Must be called with l_lock and lr_lock held.
338 * Does not actually free the lock data, but rather marks the lock as
339 * destroyed by setting l_destroyed field in the lock to 1. Destroys a
340 * handle->lock association too, so that the lock can no longer be found
341 * and removes the lock from LRU list. Actual lock freeing occurs when
342 * last lock reference goes away.
344 * Original comment (of some historical value):
345 * This used to have a 'strict' flag, which recovery would use to mark an
346 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
347 * shall explain why it's gone: with the new hash table scheme, once you call
348 * ldlm_lock_destroy, you can never drop your final references on this lock.
349 * Because it's not in the hash table anymore. -phil
351 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
355 if (lock->l_readers || lock->l_writers) {
356 LDLM_ERROR(lock, "lock still has references");
360 if (!list_empty(&lock->l_res_link)) {
361 LDLM_ERROR(lock, "lock still on resource");
365 if (ldlm_is_destroyed(lock)) {
366 LASSERT(list_empty(&lock->l_lru));
370 ldlm_set_destroyed(lock);
372 if (lock->l_export && lock->l_export->exp_lock_hash) {
373 /* NB: it's safe to call cfs_hash_del() even lock isn't
374 * in exp_lock_hash. */
375 /* In the function below, .hs_keycmp resolves to
376 * ldlm_export_lock_keycmp() */
377 /* coverity[overrun-buffer-val] */
378 cfs_hash_del(lock->l_export->exp_lock_hash,
379 &lock->l_remote_handle, &lock->l_exp_hash);
382 ldlm_lock_remove_from_lru(lock);
383 class_handle_unhash(&lock->l_handle);
390 * Destroys a LDLM lock \a lock. Performs necessary locking first.
392 void ldlm_lock_destroy(struct ldlm_lock *lock)
396 lock_res_and_lock(lock);
397 first = ldlm_lock_destroy_internal(lock);
398 unlock_res_and_lock(lock);
400 /* drop reference from hashtable only for first destroy */
402 lu_ref_del(&lock->l_reference, "hash", lock);
403 LDLM_LOCK_RELEASE(lock);
409 * Destroys a LDLM lock \a lock that is already locked.
411 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
415 first = ldlm_lock_destroy_internal(lock);
416 /* drop reference from hashtable only for first destroy */
418 lu_ref_del(&lock->l_reference, "hash", lock);
419 LDLM_LOCK_RELEASE(lock);
424 /* this is called by portals_handle2object with the handle lock taken */
425 static void lock_handle_addref(void *lock)
427 LDLM_LOCK_GET((struct ldlm_lock *)lock);
430 static void lock_handle_free(void *lock, int size)
432 LASSERT(size == sizeof(struct ldlm_lock));
433 OBD_SLAB_FREE(lock, ldlm_lock_slab, size);
436 static struct portals_handle_ops lock_handle_ops = {
437 .hop_addref = lock_handle_addref,
438 .hop_free = lock_handle_free,
443 * Allocate and initialize new lock structure.
445 * usage: pass in a resource on which you have done ldlm_resource_get
446 * new lock will take over the refcount.
447 * returns: lock with refcount 2 - one for current caller and one for remote
449 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
451 struct ldlm_lock *lock;
454 if (resource == NULL)
457 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, GFP_NOFS);
461 spin_lock_init(&lock->l_lock);
462 lock->l_resource = resource;
463 lu_ref_add(&resource->lr_reference, "lock", lock);
465 atomic_set(&lock->l_refc, 2);
466 INIT_LIST_HEAD(&lock->l_res_link);
467 INIT_LIST_HEAD(&lock->l_lru);
468 INIT_LIST_HEAD(&lock->l_pending_chain);
469 INIT_LIST_HEAD(&lock->l_bl_ast);
470 INIT_LIST_HEAD(&lock->l_cp_ast);
471 INIT_LIST_HEAD(&lock->l_rk_ast);
472 init_waitqueue_head(&lock->l_waitq);
473 lock->l_blocking_lock = NULL;
474 INIT_LIST_HEAD(&lock->l_sl_mode);
475 INIT_LIST_HEAD(&lock->l_sl_policy);
476 INIT_HLIST_NODE(&lock->l_exp_hash);
477 INIT_HLIST_NODE(&lock->l_exp_flock_hash);
479 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
481 INIT_LIST_HEAD_RCU(&lock->l_handle.h_link);
482 class_handle_hash(&lock->l_handle, &lock_handle_ops);
484 lu_ref_init(&lock->l_reference);
485 lu_ref_add(&lock->l_reference, "hash", lock);
486 lock->l_callback_timeout = 0;
487 lock->l_activity = 0;
489 #if LUSTRE_TRACKS_LOCK_EXP_REFS
490 INIT_LIST_HEAD(&lock->l_exp_refs_link);
491 lock->l_exp_refs_nr = 0;
492 lock->l_exp_refs_target = NULL;
494 INIT_LIST_HEAD(&lock->l_exp_list);
500 * Moves LDLM lock \a lock to another resource.
501 * This is used on client when server returns some other lock than requested
502 * (typically as a result of intent operation)
504 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
505 const struct ldlm_res_id *new_resid)
507 struct ldlm_resource *oldres = lock->l_resource;
508 struct ldlm_resource *newres;
512 LASSERT(ns_is_client(ns));
514 lock_res_and_lock(lock);
515 if (memcmp(new_resid, &lock->l_resource->lr_name,
516 sizeof(lock->l_resource->lr_name)) == 0) {
518 unlock_res_and_lock(lock);
522 LASSERT(new_resid->name[0] != 0);
524 /* This function assumes that the lock isn't on any lists */
525 LASSERT(list_empty(&lock->l_res_link));
527 type = oldres->lr_type;
528 unlock_res_and_lock(lock);
530 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
532 RETURN(PTR_ERR(newres));
534 lu_ref_add(&newres->lr_reference, "lock", lock);
536 * To flip the lock from the old to the new resource, lock, oldres and
537 * newres have to be locked. Resource spin-locks are nested within
538 * lock->l_lock, and are taken in the memory address order to avoid
541 spin_lock(&lock->l_lock);
542 oldres = lock->l_resource;
543 if (oldres < newres) {
545 lock_res_nested(newres, LRT_NEW);
548 lock_res_nested(oldres, LRT_NEW);
550 LASSERT(memcmp(new_resid, &oldres->lr_name,
551 sizeof oldres->lr_name) != 0);
552 lock->l_resource = newres;
554 unlock_res_and_lock(lock);
556 /* ...and the flowers are still standing! */
557 lu_ref_del(&oldres->lr_reference, "lock", lock);
558 ldlm_resource_putref(oldres);
563 /** \defgroup ldlm_handles LDLM HANDLES
564 * Ways to get hold of locks without any addresses.
569 * Fills in handle for LDLM lock \a lock into supplied \a lockh
570 * Does not take any references.
572 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
574 lockh->cookie = lock->l_handle.h_cookie;
576 EXPORT_SYMBOL(ldlm_lock2handle);
579 * Obtain a lock reference by handle.
581 * if \a flags: atomically get the lock and set the flags.
582 * Return NULL if flag already set
584 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
587 struct ldlm_lock *lock;
592 lock = class_handle2object(handle->cookie, NULL);
596 if (lock->l_export != NULL && lock->l_export->exp_failed) {
597 CDEBUG(D_INFO, "lock export failed: lock %p, exp %p\n",
598 lock, lock->l_export);
603 /* It's unlikely but possible that someone marked the lock as
604 * destroyed after we did handle2object on it */
605 if ((flags == 0) && !ldlm_is_destroyed(lock)) {
606 lu_ref_add(&lock->l_reference, "handle", current);
610 lock_res_and_lock(lock);
612 LASSERT(lock->l_resource != NULL);
614 lu_ref_add_atomic(&lock->l_reference, "handle", current);
615 if (unlikely(ldlm_is_destroyed(lock))) {
616 unlock_res_and_lock(lock);
617 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
622 /* If we're setting flags, make sure none of them are already set. */
624 if ((lock->l_flags & flags) != 0) {
625 unlock_res_and_lock(lock);
630 lock->l_flags |= flags;
633 unlock_res_and_lock(lock);
636 EXPORT_SYMBOL(__ldlm_handle2lock);
637 /** @} ldlm_handles */
640 * Fill in "on the wire" representation for given LDLM lock into supplied
641 * lock descriptor \a desc structure.
643 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
645 ldlm_res2desc(lock->l_resource, &desc->l_resource);
646 desc->l_req_mode = lock->l_req_mode;
647 desc->l_granted_mode = lock->l_granted_mode;
648 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
649 &lock->l_policy_data,
650 &desc->l_policy_data);
654 * Add a lock to list of conflicting locks to send AST to.
656 * Only add if we have not sent a blocking AST to the lock yet.
658 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
659 struct list_head *work_list)
661 if (!ldlm_is_ast_sent(lock)) {
662 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
663 ldlm_set_ast_sent(lock);
664 /* If the enqueuing client said so, tell the AST recipient to
665 * discard dirty data, rather than writing back. */
666 if (ldlm_is_ast_discard_data(new))
667 ldlm_set_discard_data(lock);
669 /* Lock can be converted from a blocking state back to granted
670 * after lock convert or COS downgrade but still be in an
671 * older bl_list because it is controlled only by
672 * ldlm_work_bl_ast_lock(), let it be processed there.
674 if (list_empty(&lock->l_bl_ast)) {
675 list_add(&lock->l_bl_ast, work_list);
678 LASSERT(lock->l_blocking_lock == NULL);
679 lock->l_blocking_lock = LDLM_LOCK_GET(new);
684 * Add a lock to list of just granted locks to send completion AST to.
686 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
687 struct list_head *work_list)
689 if (!ldlm_is_cp_reqd(lock)) {
690 ldlm_set_cp_reqd(lock);
691 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
692 LASSERT(list_empty(&lock->l_cp_ast));
693 list_add(&lock->l_cp_ast, work_list);
699 * Aggregator function to add AST work items into a list. Determines
700 * what sort of an AST work needs to be done and calls the proper
702 * Must be called with lr_lock held.
704 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
705 struct list_head *work_list)
708 check_res_locked(lock->l_resource);
710 ldlm_add_bl_work_item(lock, new, work_list);
712 ldlm_add_cp_work_item(lock, work_list);
717 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
718 * r/w reference type is determined by \a mode
719 * Calls ldlm_lock_addref_internal.
721 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
723 struct ldlm_lock *lock;
725 lock = ldlm_handle2lock(lockh);
726 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
727 ldlm_lock_addref_internal(lock, mode);
730 EXPORT_SYMBOL(ldlm_lock_addref);
734 * Add specified reader/writer reference to LDLM lock \a lock.
735 * r/w reference type is determined by \a mode
736 * Removes lock from LRU if it is there.
737 * Assumes the LDLM lock is already locked.
739 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
742 ldlm_lock_remove_from_lru(lock);
743 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
745 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
747 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
749 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
752 lu_ref_add_atomic(&lock->l_reference, "user", lock);
753 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
757 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
758 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
760 * \retval 0 success, lock was addref-ed
762 * \retval -EAGAIN lock is being canceled.
764 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
766 struct ldlm_lock *lock;
770 lock = ldlm_handle2lock(lockh);
772 lock_res_and_lock(lock);
773 if (lock->l_readers != 0 || lock->l_writers != 0 ||
774 !ldlm_is_cbpending(lock)) {
775 ldlm_lock_addref_internal_nolock(lock, mode);
778 unlock_res_and_lock(lock);
783 EXPORT_SYMBOL(ldlm_lock_addref_try);
786 * Add specified reader/writer reference to LDLM lock \a lock.
787 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
788 * Only called for local locks.
790 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
792 lock_res_and_lock(lock);
793 ldlm_lock_addref_internal_nolock(lock, mode);
794 unlock_res_and_lock(lock);
798 * Removes reader/writer reference for LDLM lock \a lock.
799 * Assumes LDLM lock is already locked.
800 * only called in ldlm_flock_destroy and for local locks.
801 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
802 * that cannot be placed in LRU.
804 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
807 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
808 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
809 LASSERT(lock->l_readers > 0);
810 lu_ref_del(&lock->l_reference, "reader", lock);
813 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
814 LASSERT(lock->l_writers > 0);
815 lu_ref_del(&lock->l_reference, "writer", lock);
819 lu_ref_del(&lock->l_reference, "user", lock);
820 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
824 * Removes reader/writer reference for LDLM lock \a lock.
825 * Locks LDLM lock first.
826 * If the lock is determined to be client lock on a client and r/w refcount
827 * drops to zero and the lock is not blocked, the lock is added to LRU lock
829 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
831 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
833 struct ldlm_namespace *ns;
836 lock_res_and_lock(lock);
838 ns = ldlm_lock_to_ns(lock);
840 ldlm_lock_decref_internal_nolock(lock, mode);
842 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
843 !lock->l_readers && !lock->l_writers) {
844 /* If this is a local lock on a server namespace and this was
845 * the last reference, cancel the lock.
847 * Group locks are special:
848 * They must not go in LRU, but they are not called back
849 * like non-group locks, instead they are manually released.
850 * They have an l_writers reference which they keep until
851 * they are manually released, so we remove them when they have
852 * no more reader or writer references. - LU-6368 */
853 ldlm_set_cbpending(lock);
856 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
857 /* If we received a blocked AST and this was the last reference,
858 * run the callback. */
859 if (ldlm_is_ns_srv(lock) && lock->l_export)
860 CERROR("FL_CBPENDING set on non-local lock--just a "
863 LDLM_DEBUG(lock, "final decref done on cbpending lock");
865 LDLM_LOCK_GET(lock); /* dropped by bl thread */
866 ldlm_lock_remove_from_lru(lock);
867 unlock_res_and_lock(lock);
869 if (ldlm_is_fail_loc(lock))
870 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
872 if (ldlm_is_atomic_cb(lock) ||
873 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
874 ldlm_handle_bl_callback(ns, NULL, lock);
875 } else if (ns_is_client(ns) &&
876 !lock->l_readers && !lock->l_writers &&
877 !ldlm_is_no_lru(lock) &&
878 !ldlm_is_bl_ast(lock) &&
879 !ldlm_is_converting(lock)) {
881 LDLM_DEBUG(lock, "add lock into lru list");
883 /* If this is a client-side namespace and this was the last
884 * reference, put it on the LRU. */
885 ldlm_lock_add_to_lru(lock);
886 unlock_res_and_lock(lock);
888 if (ldlm_is_fail_loc(lock))
889 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
891 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
892 * are not supported by the server, otherwise, it is done on
894 if (!exp_connect_cancelset(lock->l_conn_export) &&
895 !ns_connect_lru_resize(ns))
896 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
898 LDLM_DEBUG(lock, "do not add lock into lru list");
899 unlock_res_and_lock(lock);
906 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
908 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
910 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
911 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
912 ldlm_lock_decref_internal(lock, mode);
915 EXPORT_SYMBOL(ldlm_lock_decref);
918 * Decrease reader/writer refcount for LDLM lock with handle
919 * \a lockh and mark it for subsequent cancellation once r/w refcount
920 * drops to zero instead of putting into LRU.
923 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
926 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
929 LASSERT(lock != NULL);
931 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
932 lock_res_and_lock(lock);
933 ldlm_set_cbpending(lock);
934 unlock_res_and_lock(lock);
935 ldlm_lock_decref_internal(lock, mode);
938 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
940 struct sl_insert_point {
941 struct list_head *res_link;
942 struct list_head *mode_link;
943 struct list_head *policy_link;
947 * Finds a position to insert the new lock into granted lock list.
949 * Used for locks eligible for skiplist optimization.
952 * queue [input]: the granted list where search acts on;
953 * req [input]: the lock whose position to be located;
954 * prev [output]: positions within 3 lists to insert @req to
958 * - ldlm_grant_lock_with_skiplist
960 static void search_granted_lock(struct list_head *queue,
961 struct ldlm_lock *req,
962 struct sl_insert_point *prev)
964 struct list_head *tmp;
965 struct ldlm_lock *lock, *mode_end, *policy_end;
968 list_for_each(tmp, queue) {
969 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
971 mode_end = list_entry(lock->l_sl_mode.prev,
972 struct ldlm_lock, l_sl_mode);
974 if (lock->l_req_mode != req->l_req_mode) {
975 /* jump to last lock of mode group */
976 tmp = &mode_end->l_res_link;
980 /* suitable mode group is found */
981 if (lock->l_resource->lr_type == LDLM_PLAIN) {
982 /* insert point is last lock of the mode group */
983 prev->res_link = &mode_end->l_res_link;
984 prev->mode_link = &mode_end->l_sl_mode;
985 prev->policy_link = &req->l_sl_policy;
988 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
991 list_entry(lock->l_sl_policy.prev,
995 if (lock->l_policy_data.l_inodebits.bits ==
996 req->l_policy_data.l_inodebits.bits) {
997 /* insert point is last lock of
998 * the policy group */
1000 &policy_end->l_res_link;
1002 &policy_end->l_sl_mode;
1004 &policy_end->l_sl_policy;
1009 if (policy_end == mode_end)
1010 /* done with mode group */
1013 /* go to next policy group within mode group */
1014 tmp = policy_end->l_res_link.next;
1015 lock = list_entry(tmp, struct ldlm_lock,
1017 } /* loop over policy groups within the mode group */
1019 /* insert point is last lock of the mode group,
1020 * new policy group is started */
1021 prev->res_link = &mode_end->l_res_link;
1022 prev->mode_link = &mode_end->l_sl_mode;
1023 prev->policy_link = &req->l_sl_policy;
1027 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1032 /* insert point is last lock on the queue,
1033 * new mode group and new policy group are started */
1034 prev->res_link = queue->prev;
1035 prev->mode_link = &req->l_sl_mode;
1036 prev->policy_link = &req->l_sl_policy;
1042 * Add a lock into resource granted list after a position described by
1045 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1046 struct sl_insert_point *prev)
1048 struct ldlm_resource *res = lock->l_resource;
1051 check_res_locked(res);
1053 ldlm_resource_dump(D_INFO, res);
1054 LDLM_DEBUG(lock, "About to add lock:");
1056 if (ldlm_is_destroyed(lock)) {
1057 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1061 LASSERT(list_empty(&lock->l_res_link));
1062 LASSERT(list_empty(&lock->l_sl_mode));
1063 LASSERT(list_empty(&lock->l_sl_policy));
1066 * lock->link == prev->link means lock is first starting the group.
1067 * Don't re-add to itself to suppress kernel warnings.
1069 if (&lock->l_res_link != prev->res_link)
1070 list_add(&lock->l_res_link, prev->res_link);
1071 if (&lock->l_sl_mode != prev->mode_link)
1072 list_add(&lock->l_sl_mode, prev->mode_link);
1073 if (&lock->l_sl_policy != prev->policy_link)
1074 list_add(&lock->l_sl_policy, prev->policy_link);
1080 * Add a lock to granted list on a resource maintaining skiplist
1083 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1085 struct sl_insert_point prev;
1087 LASSERT(lock->l_req_mode == lock->l_granted_mode);
1089 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1090 ldlm_granted_list_add_lock(lock, &prev);
1094 * Perform lock granting bookkeeping.
1096 * Includes putting the lock into granted list and updating lock mode.
1098 * - ldlm_lock_enqueue
1099 * - ldlm_reprocess_queue
1101 * must be called with lr_lock held
1103 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1105 struct ldlm_resource *res = lock->l_resource;
1108 check_res_locked(res);
1110 lock->l_granted_mode = lock->l_req_mode;
1112 if (work_list && lock->l_completion_ast != NULL)
1113 ldlm_add_ast_work_item(lock, NULL, work_list);
1115 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1116 ldlm_grant_lock_with_skiplist(lock);
1117 else if (res->lr_type == LDLM_EXTENT)
1118 ldlm_extent_add_lock(res, lock);
1119 else if (res->lr_type == LDLM_FLOCK) {
1120 /* We should not add locks to granted list in the following
1122 * - this is an UNLOCK but not a real lock;
1123 * - this is a TEST lock;
1124 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1125 * - this is a deadlock (flock cannot be granted) */
1126 if (lock->l_req_mode == 0 ||
1127 lock->l_req_mode == LCK_NL ||
1128 ldlm_is_test_lock(lock) ||
1129 ldlm_is_flock_deadlock(lock))
1131 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1136 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1141 * Describe the overlap between two locks. itree_overlap_cb data.
1143 struct lock_match_data {
1144 struct ldlm_lock *lmd_old;
1145 struct ldlm_lock *lmd_lock;
1146 enum ldlm_mode *lmd_mode;
1147 union ldlm_policy_data *lmd_policy;
1153 * Check if the given @lock meets the criteria for a match.
1154 * A reference on the lock is taken if matched.
1156 * \param lock test-against this lock
1157 * \param data parameters
1159 static int lock_matches(struct ldlm_lock *lock, struct lock_match_data *data)
1161 union ldlm_policy_data *lpol = &lock->l_policy_data;
1162 enum ldlm_mode match;
1164 if (lock == data->lmd_old)
1165 return INTERVAL_ITER_STOP;
1167 /* Check if this lock can be matched.
1168 * Used by LU-2919(exclusive open) for open lease lock */
1169 if (ldlm_is_excl(lock))
1170 return INTERVAL_ITER_CONT;
1172 /* llite sometimes wants to match locks that will be
1173 * canceled when their users drop, but we allow it to match
1174 * if it passes in CBPENDING and the lock still has users.
1175 * this is generally only going to be used by children
1176 * whose parents already hold a lock so forward progress
1177 * can still happen. */
1178 if (ldlm_is_cbpending(lock) &&
1179 !(data->lmd_flags & LDLM_FL_CBPENDING))
1180 return INTERVAL_ITER_CONT;
1181 if (!data->lmd_unref && ldlm_is_cbpending(lock) &&
1182 lock->l_readers == 0 && lock->l_writers == 0)
1183 return INTERVAL_ITER_CONT;
1185 if (!(lock->l_req_mode & *data->lmd_mode))
1186 return INTERVAL_ITER_CONT;
1187 match = lock->l_req_mode;
1189 switch (lock->l_resource->lr_type) {
1191 if (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1192 lpol->l_extent.end < data->lmd_policy->l_extent.end)
1193 return INTERVAL_ITER_CONT;
1195 if (unlikely(match == LCK_GROUP) &&
1196 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1197 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1198 return INTERVAL_ITER_CONT;
1201 /* We match if we have existing lock with same or wider set
1203 if ((lpol->l_inodebits.bits &
1204 data->lmd_policy->l_inodebits.bits) !=
1205 data->lmd_policy->l_inodebits.bits)
1206 return INTERVAL_ITER_CONT;
1212 /* We match if we have existing lock with same or wider set
1214 if (!data->lmd_unref && LDLM_HAVE_MASK(lock, GONE))
1215 return INTERVAL_ITER_CONT;
1217 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1218 return INTERVAL_ITER_CONT;
1220 if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1221 LDLM_LOCK_GET(lock);
1222 ldlm_lock_touch_in_lru(lock);
1224 ldlm_lock_addref_internal_nolock(lock, match);
1227 *data->lmd_mode = match;
1228 data->lmd_lock = lock;
1230 return INTERVAL_ITER_STOP;
1233 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1235 struct ldlm_interval *node = to_ldlm_interval(in);
1236 struct lock_match_data *data = args;
1237 struct ldlm_lock *lock;
1240 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1241 rc = lock_matches(lock, data);
1242 if (rc == INTERVAL_ITER_STOP)
1243 return INTERVAL_ITER_STOP;
1245 return INTERVAL_ITER_CONT;
1249 * Search for a lock with given parameters in interval trees.
1251 * \param res search for a lock in this resource
1252 * \param data parameters
1254 * \retval a referenced lock or NULL.
1256 static struct ldlm_lock *search_itree(struct ldlm_resource *res,
1257 struct lock_match_data *data)
1259 struct interval_node_extent ext = {
1260 .start = data->lmd_policy->l_extent.start,
1261 .end = data->lmd_policy->l_extent.end
1265 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1266 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1268 if (tree->lit_root == NULL)
1271 if (!(tree->lit_mode & *data->lmd_mode))
1274 interval_search(tree->lit_root, &ext,
1275 itree_overlap_cb, data);
1277 return data->lmd_lock;
1282 * Search for a lock with given properties in a queue.
1284 * \param queue search for a lock in this queue
1285 * \param data parameters
1287 * \retval a referenced lock or NULL.
1289 static struct ldlm_lock *search_queue(struct list_head *queue,
1290 struct lock_match_data *data)
1292 struct ldlm_lock *lock;
1295 list_for_each_entry(lock, queue, l_res_link) {
1296 rc = lock_matches(lock, data);
1297 if (rc == INTERVAL_ITER_STOP)
1298 return data->lmd_lock;
1303 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1305 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1306 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1307 wake_up_all(&lock->l_waitq);
1310 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1312 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1314 lock_res_and_lock(lock);
1315 ldlm_lock_fail_match_locked(lock);
1316 unlock_res_and_lock(lock);
1320 * Mark lock as "matchable" by OST.
1322 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1324 * Assumes LDLM lock is already locked.
1326 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1328 ldlm_set_lvb_ready(lock);
1329 wake_up_all(&lock->l_waitq);
1331 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1334 * Mark lock as "matchable" by OST.
1335 * Locks the lock and then \see ldlm_lock_allow_match_locked
1337 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1339 lock_res_and_lock(lock);
1340 ldlm_lock_allow_match_locked(lock);
1341 unlock_res_and_lock(lock);
1343 EXPORT_SYMBOL(ldlm_lock_allow_match);
1346 * Attempt to find a lock with specified properties.
1348 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1351 * Can be called in two ways:
1353 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1354 * for a duplicate of.
1356 * Otherwise, all of the fields must be filled in, to match against.
1358 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1359 * server (ie, connh is NULL)
1360 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1361 * list will be considered
1362 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1363 * to be canceled can still be matched as long as they still have reader
1364 * or writer refernces
1365 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1366 * just tell us if we would have matched.
1368 * \retval 1 if it finds an already-existing lock that is compatible; in this
1369 * case, lockh is filled in with a addref()ed lock
1371 * We also check security context, and if that fails we simply return 0 (to
1372 * keep caller code unchanged), the context failure will be discovered by
1373 * caller sometime later.
1375 enum ldlm_mode ldlm_lock_match(struct ldlm_namespace *ns, __u64 flags,
1376 const struct ldlm_res_id *res_id,
1377 enum ldlm_type type,
1378 union ldlm_policy_data *policy,
1379 enum ldlm_mode mode,
1380 struct lustre_handle *lockh, int unref)
1382 struct lock_match_data data = {
1386 .lmd_policy = policy,
1390 struct ldlm_resource *res;
1391 struct ldlm_lock *lock;
1396 data.lmd_old = ldlm_handle2lock(lockh);
1397 LASSERT(data.lmd_old != NULL);
1399 ns = ldlm_lock_to_ns(data.lmd_old);
1400 res_id = &data.lmd_old->l_resource->lr_name;
1401 type = data.lmd_old->l_resource->lr_type;
1402 *data.lmd_mode = data.lmd_old->l_req_mode;
1405 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1407 LASSERT(data.lmd_old == NULL);
1411 LDLM_RESOURCE_ADDREF(res);
1414 if (res->lr_type == LDLM_EXTENT)
1415 lock = search_itree(res, &data);
1417 lock = search_queue(&res->lr_granted, &data);
1420 if (flags & LDLM_FL_BLOCK_GRANTED)
1422 lock = search_queue(&res->lr_waiting, &data);
1429 LDLM_RESOURCE_DELREF(res);
1430 ldlm_resource_putref(res);
1433 ldlm_lock2handle(lock, lockh);
1434 if ((flags & LDLM_FL_LVB_READY) &&
1435 (!ldlm_is_lvb_ready(lock))) {
1436 __u64 wait_flags = LDLM_FL_LVB_READY |
1437 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1438 struct l_wait_info lwi;
1439 if (lock->l_completion_ast) {
1440 int err = lock->l_completion_ast(lock,
1441 LDLM_FL_WAIT_NOREPROC,
1444 if (flags & LDLM_FL_TEST_LOCK)
1445 LDLM_LOCK_RELEASE(lock);
1447 ldlm_lock_decref_internal(lock,
1454 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1455 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1457 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1458 l_wait_event(lock->l_waitq,
1459 lock->l_flags & wait_flags,
1461 if (!ldlm_is_lvb_ready(lock)) {
1462 if (flags & LDLM_FL_TEST_LOCK)
1463 LDLM_LOCK_RELEASE(lock);
1465 ldlm_lock_decref_internal(lock, mode);
1472 LDLM_DEBUG(lock, "matched (%llu %llu)",
1473 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1474 res_id->name[2] : policy->l_extent.start,
1475 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1476 res_id->name[3] : policy->l_extent.end);
1478 /* check user's security context */
1479 if (lock->l_conn_export &&
1480 sptlrpc_import_check_ctx(
1481 class_exp2cliimp(lock->l_conn_export))) {
1482 if (!(flags & LDLM_FL_TEST_LOCK))
1483 ldlm_lock_decref_internal(lock, mode);
1487 if (flags & LDLM_FL_TEST_LOCK)
1488 LDLM_LOCK_RELEASE(lock);
1490 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1491 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1492 "%llu/%llu (%llu %llu)", ns,
1493 type, mode, res_id->name[0], res_id->name[1],
1494 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1495 res_id->name[2] :policy->l_extent.start,
1496 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1497 res_id->name[3] : policy->l_extent.end);
1499 if (data.lmd_old != NULL)
1500 LDLM_LOCK_PUT(data.lmd_old);
1502 return rc ? mode : 0;
1504 EXPORT_SYMBOL(ldlm_lock_match);
1506 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1509 struct ldlm_lock *lock;
1510 enum ldlm_mode mode = 0;
1513 lock = ldlm_handle2lock(lockh);
1515 lock_res_and_lock(lock);
1516 if (LDLM_HAVE_MASK(lock, GONE))
1519 if (ldlm_is_cbpending(lock) &&
1520 lock->l_readers == 0 && lock->l_writers == 0)
1524 *bits = lock->l_policy_data.l_inodebits.bits;
1525 mode = lock->l_granted_mode;
1526 ldlm_lock_addref_internal_nolock(lock, mode);
1533 unlock_res_and_lock(lock);
1534 LDLM_LOCK_PUT(lock);
1538 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1540 /** The caller must guarantee that the buffer is large enough. */
1541 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1542 enum req_location loc, void *data, int size)
1547 LASSERT(data != NULL);
1550 switch (lock->l_lvb_type) {
1552 if (size == sizeof(struct ost_lvb)) {
1553 if (loc == RCL_CLIENT)
1554 lvb = req_capsule_client_swab_get(pill,
1556 lustre_swab_ost_lvb);
1558 lvb = req_capsule_server_swab_get(pill,
1560 lustre_swab_ost_lvb);
1561 if (unlikely(lvb == NULL)) {
1562 LDLM_ERROR(lock, "no LVB");
1566 memcpy(data, lvb, size);
1567 } else if (size == sizeof(struct ost_lvb_v1)) {
1568 struct ost_lvb *olvb = data;
1570 if (loc == RCL_CLIENT)
1571 lvb = req_capsule_client_swab_get(pill,
1573 lustre_swab_ost_lvb_v1);
1575 lvb = req_capsule_server_sized_swab_get(pill,
1577 lustre_swab_ost_lvb_v1);
1578 if (unlikely(lvb == NULL)) {
1579 LDLM_ERROR(lock, "no LVB");
1583 memcpy(data, lvb, size);
1584 olvb->lvb_mtime_ns = 0;
1585 olvb->lvb_atime_ns = 0;
1586 olvb->lvb_ctime_ns = 0;
1588 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1594 if (size == sizeof(struct lquota_lvb)) {
1595 if (loc == RCL_CLIENT)
1596 lvb = req_capsule_client_swab_get(pill,
1598 lustre_swab_lquota_lvb);
1600 lvb = req_capsule_server_swab_get(pill,
1602 lustre_swab_lquota_lvb);
1603 if (unlikely(lvb == NULL)) {
1604 LDLM_ERROR(lock, "no LVB");
1608 memcpy(data, lvb, size);
1610 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1619 if (loc == RCL_CLIENT)
1620 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1622 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1623 if (unlikely(lvb == NULL)) {
1624 LDLM_ERROR(lock, "no LVB");
1628 memcpy(data, lvb, size);
1631 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1632 libcfs_debug_dumpstack(NULL);
1640 * Create and fill in new LDLM lock with specified properties.
1641 * Returns a referenced lock
1643 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1644 const struct ldlm_res_id *res_id,
1645 enum ldlm_type type,
1646 enum ldlm_mode mode,
1647 const struct ldlm_callback_suite *cbs,
1648 void *data, __u32 lvb_len,
1649 enum lvb_type lvb_type)
1651 struct ldlm_lock *lock;
1652 struct ldlm_resource *res;
1656 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1658 RETURN(ERR_CAST(res));
1660 lock = ldlm_lock_new(res);
1662 RETURN(ERR_PTR(-ENOMEM));
1664 lock->l_req_mode = mode;
1665 lock->l_ast_data = data;
1666 lock->l_pid = current_pid();
1667 if (ns_is_server(ns))
1668 ldlm_set_ns_srv(lock);
1670 lock->l_blocking_ast = cbs->lcs_blocking;
1671 lock->l_completion_ast = cbs->lcs_completion;
1672 lock->l_glimpse_ast = cbs->lcs_glimpse;
1675 lock->l_tree_node = NULL;
1676 /* if this is the extent lock, allocate the interval tree node */
1677 if (type == LDLM_EXTENT)
1678 if (ldlm_interval_alloc(lock) == NULL)
1679 GOTO(out, rc = -ENOMEM);
1682 lock->l_lvb_len = lvb_len;
1683 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1684 if (lock->l_lvb_data == NULL)
1685 GOTO(out, rc = -ENOMEM);
1688 lock->l_lvb_type = lvb_type;
1689 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1690 GOTO(out, rc = -ENOENT);
1695 ldlm_lock_destroy(lock);
1696 LDLM_LOCK_RELEASE(lock);
1697 RETURN(ERR_PTR(rc));
1700 #ifdef HAVE_SERVER_SUPPORT
1701 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1704 struct ldlm_resource *res = lock->l_resource;
1705 enum ldlm_error rc = ELDLM_OK;
1706 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1707 ldlm_processing_policy policy;
1710 policy = ldlm_processing_policy_table[res->lr_type];
1712 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1713 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1714 res->lr_type != LDLM_FLOCK) {
1715 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1716 if (rc == -ERESTART)
1720 if (!list_empty(&rpc_list))
1721 ldlm_discard_bl_list(&rpc_list);
1728 * Enqueue (request) a lock.
1730 * Does not block. As a result of enqueue the lock would be put
1731 * into granted or waiting list.
1733 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1734 * set, skip all the enqueueing and delegate lock processing to intent policy
1737 enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
1738 struct ldlm_namespace *ns,
1739 struct ldlm_lock **lockp,
1740 void *cookie, __u64 *flags)
1742 struct ldlm_lock *lock = *lockp;
1743 struct ldlm_resource *res = lock->l_resource;
1744 int local = ns_is_client(ldlm_res_to_ns(res));
1745 enum ldlm_error rc = ELDLM_OK;
1746 struct ldlm_interval *node = NULL;
1749 /* policies are not executed on the client or during replay */
1750 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1751 && !local && ns->ns_policy) {
1752 rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
1754 if (rc == ELDLM_LOCK_REPLACED) {
1755 /* The lock that was returned has already been granted,
1756 * and placed into lockp. If it's not the same as the
1757 * one we passed in, then destroy the old one and our
1758 * work here is done. */
1759 if (lock != *lockp) {
1760 ldlm_lock_destroy(lock);
1761 LDLM_LOCK_RELEASE(lock);
1763 *flags |= LDLM_FL_LOCK_CHANGED;
1765 } else if (rc != ELDLM_OK &&
1766 lock->l_req_mode == lock->l_granted_mode) {
1767 LASSERT(*flags & LDLM_FL_RESENT);
1768 /* It may happen that ns_policy returns an error in
1769 * resend case, object may be unlinked or just some
1770 * error occurs. It is unclear if lock reached the
1771 * client in the original reply, just leave the lock on
1772 * server, not returning it again to client. Due to
1773 * LU-6529, the server will not OOM. */
1775 } else if (rc != ELDLM_OK ||
1776 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1777 ldlm_lock_destroy(lock);
1782 if (*flags & LDLM_FL_RESENT) {
1783 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1784 * Set LOCK_CHANGED always.
1785 * Check if the lock is granted for BLOCK_GRANTED.
1786 * Take NO_TIMEOUT from the lock as it is inherited through
1787 * LDLM_FL_INHERIT_MASK */
1788 *flags |= LDLM_FL_LOCK_CHANGED;
1789 if (lock->l_req_mode != lock->l_granted_mode)
1790 *flags |= LDLM_FL_BLOCK_GRANTED;
1791 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1795 /* For a replaying lock, it might be already in granted list. So
1796 * unlinking the lock will cause the interval node to be freed, we
1797 * have to allocate the interval node early otherwise we can't regrant
1798 * this lock in the future. - jay */
1799 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1800 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1802 lock_res_and_lock(lock);
1803 if (local && lock->l_req_mode == lock->l_granted_mode) {
1804 /* The server returned a blocked lock, but it was granted
1805 * before we got a chance to actually enqueue it. We don't
1806 * need to do anything else. */
1807 *flags &= ~LDLM_FL_BLOCKED_MASK;
1808 GOTO(out, rc = ELDLM_OK);
1811 ldlm_resource_unlink_lock(lock);
1812 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1814 ldlm_lock_destroy_nolock(lock);
1815 GOTO(out, rc = -ENOMEM);
1818 INIT_LIST_HEAD(&node->li_group);
1819 ldlm_interval_attach(node, lock);
1823 /* Some flags from the enqueue want to make it into the AST, via the
1824 * lock's l_flags. */
1825 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1826 ldlm_set_ast_discard_data(lock);
1827 if (*flags & LDLM_FL_TEST_LOCK)
1828 ldlm_set_test_lock(lock);
1829 if (*flags & LDLM_FL_COS_INCOMPAT)
1830 ldlm_set_cos_incompat(lock);
1831 if (*flags & LDLM_FL_COS_ENABLED)
1832 ldlm_set_cos_enabled(lock);
1834 /* This distinction between local lock trees is very important; a client
1835 * namespace only has information about locks taken by that client, and
1836 * thus doesn't have enough information to decide for itself if it can
1837 * be granted (below). In this case, we do exactly what the server
1838 * tells us to do, as dictated by the 'flags'.
1840 * We do exactly the same thing during recovery, when the server is
1841 * more or less trusting the clients not to lie.
1843 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1846 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1847 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1849 ldlm_grant_lock(lock, NULL);
1850 GOTO(out, rc = ELDLM_OK);
1851 #ifdef HAVE_SERVER_SUPPORT
1852 } else if (*flags & LDLM_FL_REPLAY) {
1853 if (*flags & LDLM_FL_BLOCK_WAIT) {
1854 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1855 GOTO(out, rc = ELDLM_OK);
1856 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1857 ldlm_grant_lock(lock, NULL);
1858 GOTO(out, rc = ELDLM_OK);
1860 /* If no flags, fall through to normal enqueue path. */
1863 rc = ldlm_lock_enqueue_helper(lock, flags);
1867 CERROR("This is client-side-only module, cannot handle "
1868 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1874 unlock_res_and_lock(lock);
1876 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1880 #ifdef HAVE_SERVER_SUPPORT
1882 * Iterate through all waiting locks on a given resource queue and attempt to
1885 * Must be called with resource lock held.
1887 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1888 struct list_head *work_list,
1889 enum ldlm_process_intention intention)
1891 struct list_head *tmp, *pos;
1892 ldlm_processing_policy policy;
1894 int rc = LDLM_ITER_CONTINUE;
1895 enum ldlm_error err;
1896 struct list_head bl_ast_list = LIST_HEAD_INIT(bl_ast_list);
1899 check_res_locked(res);
1901 policy = ldlm_processing_policy_table[res->lr_type];
1903 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1904 intention == LDLM_PROCESS_RECOVERY);
1907 list_for_each_safe(tmp, pos, queue) {
1908 struct ldlm_lock *pending;
1909 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1911 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1913 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1916 rc = policy(pending, &flags, intention, &err, &rpc_list);
1917 if (pending->l_granted_mode == pending->l_req_mode ||
1918 res->lr_type == LDLM_FLOCK) {
1919 list_splice(&rpc_list, work_list);
1921 list_splice(&rpc_list, &bl_ast_list);
1924 * When this is called from recovery done, we always want
1925 * to scan the whole list no matter what 'rc' is returned.
1927 if (rc != LDLM_ITER_CONTINUE &&
1928 intention == LDLM_PROCESS_RESCAN)
1932 if (!list_empty(&bl_ast_list)) {
1935 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
1939 if (rc == -ERESTART)
1943 if (!list_empty(&bl_ast_list))
1944 ldlm_discard_bl_list(&bl_ast_list);
1946 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
1950 * Conflicting locks are detected for a lock to be enqueued, add the lock
1951 * into waiting list and send blocking ASTs to the conflicting locks.
1953 * \param[in] lock The lock to be enqueued.
1954 * \param[out] flags Lock flags for the lock to be enqueued.
1955 * \param[in] rpc_list Conflicting locks list.
1957 * \retval -ERESTART: Some lock was instantly canceled while sending
1958 * blocking ASTs, caller needs to re-check conflicting
1960 * \retval -EAGAIN: Lock was destroyed, caller should return error.
1961 * \reval 0: Lock is successfully added in waiting list.
1963 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
1964 struct list_head *rpc_list)
1966 struct ldlm_resource *res = lock->l_resource;
1970 check_res_locked(res);
1972 /* If either of the compat_queue()s returned failure, then we
1973 * have ASTs to send and must go onto the waiting list.
1975 * bug 2322: we used to unlink and re-add here, which was a
1976 * terrible folly -- if we goto restart, we could get
1977 * re-ordered! Causes deadlock, because ASTs aren't sent! */
1978 if (list_empty(&lock->l_res_link))
1979 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1982 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
1985 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
1986 !ns_is_client(ldlm_res_to_ns(res)))
1987 class_fail_export(lock->l_export);
1990 if (rc == -ERESTART) {
1991 /* 15715: The lock was granted and destroyed after
1992 * resource lock was dropped. Interval node was freed
1993 * in ldlm_lock_destroy. Anyway, this always happens
1994 * when a client is being evicted. So it would be
1995 * ok to return an error. -jay */
1996 if (ldlm_is_destroyed(lock))
1999 /* lock was granted while resource was unlocked. */
2000 if (lock->l_granted_mode == lock->l_req_mode) {
2001 /* bug 11300: if the lock has been granted,
2002 * break earlier because otherwise, we will go
2003 * to restart and ldlm_resource_unlink will be
2004 * called and it causes the interval node to be
2005 * freed. Then we will fail at
2006 * ldlm_extent_add_lock() */
2007 *flags &= ~LDLM_FL_BLOCKED_MASK;
2013 *flags |= LDLM_FL_BLOCK_GRANTED;
2019 * Discard all AST work items from list.
2021 * If for whatever reason we do not want to send ASTs to conflicting locks
2022 * anymore, disassemble the list with this function.
2024 void ldlm_discard_bl_list(struct list_head *bl_list)
2026 struct ldlm_lock *lock, *tmp;
2030 list_for_each_entry_safe(lock, tmp, bl_list, l_bl_ast) {
2031 LASSERT(!list_empty(&lock->l_bl_ast));
2032 list_del_init(&lock->l_bl_ast);
2033 ldlm_clear_ast_sent(lock);
2034 LASSERT(lock->l_bl_ast_run == 0);
2035 ldlm_clear_blocking_lock(lock);
2036 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 *lock;
2049 struct ldlm_lock_desc d;
2050 struct ldlm_bl_desc bld;
2055 if (list_empty(arg->list))
2058 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2060 /* nobody should touch l_bl_ast but some locks in the list may become
2061 * granted after lock convert or COS downgrade, these locks should be
2062 * just skipped here and removed from the list.
2064 lock_res_and_lock(lock);
2065 list_del_init(&lock->l_bl_ast);
2067 /* lock is not blocking lock anymore, but was kept in the list because
2068 * it can managed only here.
2070 if (!ldlm_is_ast_sent(lock)) {
2071 unlock_res_and_lock(lock);
2072 LDLM_LOCK_RELEASE(lock);
2076 LASSERT(lock->l_blocking_lock);
2077 ldlm_lock2desc(lock->l_blocking_lock, &d);
2078 /* copy blocking lock ibits in cancel_bits as well,
2079 * new client may use them for lock convert and it is
2080 * important to use new field to convert locks from
2083 d.l_policy_data.l_inodebits.cancel_bits =
2084 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2086 /* Blocking lock is being destroyed here but some information about it
2087 * may be needed inside l_blocking_ast() function below,
2088 * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
2090 bld.bl_same_client = lock->l_client_cookie ==
2091 lock->l_blocking_lock->l_client_cookie;
2092 bld.bl_cos_incompat = ldlm_is_cos_incompat(lock->l_blocking_lock);
2093 arg->bl_desc = &bld;
2095 LASSERT(ldlm_is_ast_sent(lock));
2096 LASSERT(lock->l_bl_ast_run == 0);
2097 lock->l_bl_ast_run++;
2098 ldlm_clear_blocking_lock(lock);
2099 unlock_res_and_lock(lock);
2101 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2103 LDLM_LOCK_RELEASE(lock);
2109 * Process a call to revocation AST callback for a lock in ast_work list
2112 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2114 struct ldlm_cb_set_arg *arg = opaq;
2115 struct ldlm_lock_desc desc;
2117 struct ldlm_lock *lock;
2120 if (list_empty(arg->list))
2123 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2124 list_del_init(&lock->l_rk_ast);
2126 /* the desc just pretend to exclusive */
2127 ldlm_lock2desc(lock, &desc);
2128 desc.l_req_mode = LCK_EX;
2129 desc.l_granted_mode = 0;
2131 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2132 LDLM_LOCK_RELEASE(lock);
2138 * Process a call to glimpse AST callback for a lock in ast_work list
2140 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2142 struct ldlm_cb_set_arg *arg = opaq;
2143 struct ldlm_glimpse_work *gl_work;
2144 struct ldlm_lock *lock;
2148 if (list_empty(arg->list))
2151 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2153 list_del_init(&gl_work->gl_list);
2155 lock = gl_work->gl_lock;
2157 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2158 arg->gl_desc = gl_work->gl_desc;
2159 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2160 arg->gl_interpret_data = gl_work->gl_interpret_data;
2162 /* invoke the actual glimpse callback */
2163 if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
2166 LDLM_LOCK_RELEASE(lock);
2167 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2168 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2170 OBD_FREE_PTR(gl_work);
2177 * Process a call to completion AST callback for a lock in ast_work list
2180 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2182 struct ldlm_cb_set_arg *arg = opaq;
2183 struct ldlm_lock *lock;
2184 ldlm_completion_callback completion_callback;
2189 if (list_empty(arg->list))
2192 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2194 /* It's possible to receive a completion AST before we've set
2195 * the l_completion_ast pointer: either because the AST arrived
2196 * before the reply, or simply because there's a small race
2197 * window between receiving the reply and finishing the local
2198 * enqueue. (bug 842)
2200 * This can't happen with the blocking_ast, however, because we
2201 * will never call the local blocking_ast until we drop our
2202 * reader/writer reference, which we won't do until we get the
2203 * reply and finish enqueueing. */
2205 /* nobody should touch l_cp_ast */
2206 lock_res_and_lock(lock);
2207 list_del_init(&lock->l_cp_ast);
2208 LASSERT(ldlm_is_cp_reqd(lock));
2209 /* save l_completion_ast since it can be changed by
2210 * mds_intent_policy(), see bug 14225 */
2211 completion_callback = lock->l_completion_ast;
2212 ldlm_clear_cp_reqd(lock);
2213 unlock_res_and_lock(lock);
2215 if (completion_callback != NULL)
2216 rc = completion_callback(lock, 0, (void *)arg);
2217 LDLM_LOCK_RELEASE(lock);
2223 * Process list of locks in need of ASTs being sent.
2225 * Used on server to send multiple ASTs together instead of sending one by
2228 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2229 ldlm_desc_ast_t ast_type)
2231 struct ldlm_cb_set_arg *arg;
2232 set_producer_func work_ast_lock;
2235 if (list_empty(rpc_list))
2242 atomic_set(&arg->restart, 0);
2243 arg->list = rpc_list;
2246 case LDLM_WORK_CP_AST:
2247 arg->type = LDLM_CP_CALLBACK;
2248 work_ast_lock = ldlm_work_cp_ast_lock;
2250 #ifdef HAVE_SERVER_SUPPORT
2251 case LDLM_WORK_BL_AST:
2252 arg->type = LDLM_BL_CALLBACK;
2253 work_ast_lock = ldlm_work_bl_ast_lock;
2255 case LDLM_WORK_REVOKE_AST:
2256 arg->type = LDLM_BL_CALLBACK;
2257 work_ast_lock = ldlm_work_revoke_ast_lock;
2259 case LDLM_WORK_GL_AST:
2260 arg->type = LDLM_GL_CALLBACK;
2261 work_ast_lock = ldlm_work_gl_ast_lock;
2268 /* We create a ptlrpc request set with flow control extension.
2269 * This request set will use the work_ast_lock function to produce new
2270 * requests and will send a new request each time one completes in order
2271 * to keep the number of requests in flight to ns_max_parallel_ast */
2272 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2273 work_ast_lock, arg);
2274 if (arg->set == NULL)
2275 GOTO(out, rc = -ENOMEM);
2277 ptlrpc_set_wait(NULL, arg->set);
2278 ptlrpc_set_destroy(arg->set);
2280 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2288 * Try to grant all waiting locks on a resource.
2290 * Calls ldlm_reprocess_queue on waiting queue.
2292 * Typically called after some resource locks are cancelled to see
2293 * if anything could be granted as a result of the cancellation.
2295 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2296 enum ldlm_process_intention intention)
2298 struct list_head rpc_list;
2299 #ifdef HAVE_SERVER_SUPPORT
2300 struct obd_device *obd;
2304 INIT_LIST_HEAD(&rpc_list);
2305 /* Local lock trees don't get reprocessed. */
2306 if (ns_is_client(ldlm_res_to_ns(res))) {
2311 /* Disable reprocess during lock replay stage but allow during
2312 * request replay stage.
2314 obd = ldlm_res_to_ns(res)->ns_obd;
2315 if (obd->obd_recovering &&
2316 atomic_read(&obd->obd_req_replay_clients) == 0)
2320 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list, intention);
2323 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2325 if (rc == -ERESTART) {
2326 LASSERT(list_empty(&rpc_list));
2332 INIT_LIST_HEAD(&rpc_list);
2333 if (!ns_is_client(ldlm_res_to_ns(res))) {
2334 CERROR("This is client-side-only module, cannot handle "
2335 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2342 void ldlm_reprocess_all(struct ldlm_resource *res)
2344 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN);
2346 EXPORT_SYMBOL(ldlm_reprocess_all);
2348 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2349 struct hlist_node *hnode, void *arg)
2351 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2353 /* This is only called once after recovery done. LU-8306. */
2354 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY);
2359 * Iterate through all resources on a namespace attempting to grant waiting
2362 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2367 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2368 ldlm_reprocess_res, NULL, 0);
2374 * Helper function to call blocking AST for LDLM lock \a lock in a
2375 * "cancelling" mode.
2377 void ldlm_cancel_callback(struct ldlm_lock *lock)
2379 check_res_locked(lock->l_resource);
2380 if (!ldlm_is_cancel(lock)) {
2381 ldlm_set_cancel(lock);
2382 if (lock->l_blocking_ast) {
2383 unlock_res_and_lock(lock);
2384 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2386 lock_res_and_lock(lock);
2388 LDLM_DEBUG(lock, "no blocking ast");
2391 /* only canceller can set bl_done bit */
2392 ldlm_set_bl_done(lock);
2393 wake_up_all(&lock->l_waitq);
2394 } else if (!ldlm_is_bl_done(lock)) {
2395 struct l_wait_info lwi = { 0 };
2397 /* The lock is guaranteed to have been canceled once
2398 * returning from this function. */
2399 unlock_res_and_lock(lock);
2400 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
2401 lock_res_and_lock(lock);
2406 * Remove skiplist-enabled LDLM lock \a req from granted list
2408 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2410 if (req->l_resource->lr_type != LDLM_PLAIN &&
2411 req->l_resource->lr_type != LDLM_IBITS)
2414 list_del_init(&req->l_sl_policy);
2415 list_del_init(&req->l_sl_mode);
2419 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2421 void ldlm_lock_cancel(struct ldlm_lock *lock)
2423 struct ldlm_resource *res;
2424 struct ldlm_namespace *ns;
2427 lock_res_and_lock(lock);
2429 res = lock->l_resource;
2430 ns = ldlm_res_to_ns(res);
2432 /* Please do not, no matter how tempting, remove this LBUG without
2433 * talking to me first. -phik */
2434 if (lock->l_readers || lock->l_writers) {
2435 LDLM_ERROR(lock, "lock still has references");
2436 unlock_res_and_lock(lock);
2440 if (ldlm_is_waited(lock))
2441 ldlm_del_waiting_lock(lock);
2443 /* Releases cancel callback. */
2444 ldlm_cancel_callback(lock);
2446 /* Yes, second time, just in case it was added again while we were
2447 * running with no res lock in ldlm_cancel_callback */
2448 if (ldlm_is_waited(lock))
2449 ldlm_del_waiting_lock(lock);
2451 ldlm_resource_unlink_lock(lock);
2452 ldlm_lock_destroy_nolock(lock);
2454 if (lock->l_granted_mode == lock->l_req_mode)
2455 ldlm_pool_del(&ns->ns_pool, lock);
2457 /* Make sure we will not be called again for same lock what is possible
2458 * if not to zero out lock->l_granted_mode */
2459 lock->l_granted_mode = LCK_MINMODE;
2460 unlock_res_and_lock(lock);
2464 EXPORT_SYMBOL(ldlm_lock_cancel);
2467 * Set opaque data into the lock that only makes sense to upper layer.
2469 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2471 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2476 if (lock->l_ast_data == NULL)
2477 lock->l_ast_data = data;
2478 if (lock->l_ast_data == data)
2480 LDLM_LOCK_PUT(lock);
2484 EXPORT_SYMBOL(ldlm_lock_set_data);
2486 struct export_cl_data {
2487 const struct lu_env *ecl_env;
2488 struct obd_export *ecl_exp;
2492 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2493 struct ldlm_lock *lock,
2494 struct export_cl_data *ecl)
2496 struct ldlm_resource *res;
2498 res = ldlm_resource_getref(lock->l_resource);
2500 ldlm_lvbo_update(ecl->ecl_env, res, lock, NULL, 1);
2501 ldlm_lock_cancel(lock);
2502 if (!exp->exp_obd->obd_stopping)
2503 ldlm_reprocess_all(res);
2504 ldlm_resource_putref(res);
2507 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2508 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2509 exp, ecl->ecl_loop);
2514 * Iterator function for ldlm_export_cancel_locks.
2515 * Cancels passed locks.
2518 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2519 struct hlist_node *hnode, void *data)
2522 struct export_cl_data *ecl = (struct export_cl_data *)data;
2523 struct obd_export *exp = ecl->ecl_exp;
2524 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2526 LDLM_LOCK_GET(lock);
2527 ldlm_cancel_lock_for_export(exp, lock, ecl);
2528 LDLM_LOCK_RELEASE(lock);
2534 * Cancel all blocked locks for given export.
2536 * Typically called on client disconnection/eviction
2538 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2541 struct export_cl_data ecl = {
2547 rc = lu_env_init(&env, LCT_DT_THREAD);
2552 while (!list_empty(&exp->exp_bl_list)) {
2553 struct ldlm_lock *lock;
2555 spin_lock_bh(&exp->exp_bl_list_lock);
2556 if (!list_empty(&exp->exp_bl_list)) {
2557 lock = list_entry(exp->exp_bl_list.next,
2558 struct ldlm_lock, l_exp_list);
2559 LDLM_LOCK_GET(lock);
2560 list_del_init(&lock->l_exp_list);
2564 spin_unlock_bh(&exp->exp_bl_list_lock);
2569 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2570 LDLM_LOCK_RELEASE(lock);
2575 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2576 "left on hash table %d.\n", exp, ecl.ecl_loop,
2577 atomic_read(&exp->exp_lock_hash->hs_count));
2579 return ecl.ecl_loop;
2583 * Cancel all locks for given export.
2585 * Typically called after client disconnection/eviction
2587 int ldlm_export_cancel_locks(struct obd_export *exp)
2589 struct export_cl_data ecl;
2593 rc = lu_env_init(&env, LCT_DT_THREAD);
2600 cfs_hash_for_each_empty(exp->exp_lock_hash,
2601 ldlm_cancel_locks_for_export_cb, &ecl);
2603 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2604 "left on hash table %d.\n", exp, ecl.ecl_loop,
2605 atomic_read(&exp->exp_lock_hash->hs_count));
2607 if (ecl.ecl_loop > 0 &&
2608 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2609 exp->exp_obd->obd_stopping)
2610 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2614 return ecl.ecl_loop;
2618 * Downgrade an PW/EX lock to COS mode.
2620 * A lock mode convertion from PW/EX mode to less conflict mode. The
2621 * convertion may fail if lock was canceled before downgrade, but it doesn't
2622 * indicate any problem, because such lock has no reader or writer, and will
2625 * Used by Commit on Sharing (COS) code to force object changes commit in case
2626 * of conflict. Converted lock is considered as new lock and all blocking AST
2627 * things are cleared, so any pending or new blocked lock on that lock will
2628 * cause new call to blocking_ast and force resource object commit.
2630 * \param lock A lock to convert
2631 * \param new_mode new lock mode
2633 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2635 #ifdef HAVE_SERVER_SUPPORT
2638 LASSERT(new_mode == LCK_COS);
2640 lock_res_and_lock(lock);
2642 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2643 unlock_res_and_lock(lock);
2645 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2646 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2650 ldlm_resource_unlink_lock(lock);
2652 * Remove the lock from pool as it will be added again in
2653 * ldlm_grant_lock() called below.
2655 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2657 /* Consider downgraded lock as a new lock and clear all states
2658 * related to a previous blocking AST processing.
2660 ldlm_clear_blocking_data(lock);
2662 lock->l_req_mode = new_mode;
2663 ldlm_grant_lock(lock, NULL);
2664 unlock_res_and_lock(lock);
2666 ldlm_reprocess_all(lock->l_resource);
2671 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2674 * Print lock with lock handle \a lockh description into debug log.
2676 * Used when printing all locks on a resource for debug purposes.
2678 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2680 struct ldlm_lock *lock;
2682 if (!((libcfs_debug | D_ERROR) & level))
2685 lock = ldlm_handle2lock(lockh);
2689 LDLM_DEBUG_LIMIT(level, lock, "###");
2691 LDLM_LOCK_PUT(lock);
2693 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2696 * Print lock information with custom message into debug log.
2699 void _ldlm_lock_debug(struct ldlm_lock *lock,
2700 struct libcfs_debug_msg_data *msgdata,
2701 const char *fmt, ...)
2704 struct obd_export *exp = lock->l_export;
2705 struct ldlm_resource *resource = NULL;
2706 char *nid = "local";
2708 /* on server-side resource of lock doesn't change */
2709 if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2710 if (lock->l_resource != NULL)
2711 resource = ldlm_resource_getref(lock->l_resource);
2712 } else if (spin_trylock(&lock->l_lock)) {
2713 if (lock->l_resource != NULL)
2714 resource = ldlm_resource_getref(lock->l_resource);
2715 spin_unlock(&lock->l_lock);
2718 va_start(args, fmt);
2720 if (exp && exp->exp_connection) {
2721 nid = obd_export_nid2str(exp);
2722 } else if (exp && exp->exp_obd != NULL) {
2723 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2724 nid = obd_import_nid2str(imp);
2727 if (resource == NULL) {
2728 libcfs_debug_vmsg2(msgdata, fmt, args,
2729 " ns: \?\? lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2730 "res: \?\? rrc=\?\? type: \?\?\? flags: %#llx nid: %s "
2731 "remote: %#llx expref: %d pid: %u timeout: %lld "
2734 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2735 lock->l_readers, lock->l_writers,
2736 ldlm_lockname[lock->l_granted_mode],
2737 ldlm_lockname[lock->l_req_mode],
2738 lock->l_flags, nid, lock->l_remote_handle.cookie,
2739 exp ? atomic_read(&exp->exp_refcount) : -99,
2740 lock->l_pid, lock->l_callback_timeout, lock->l_lvb_type);
2745 switch (resource->lr_type) {
2747 libcfs_debug_vmsg2(msgdata, fmt, args,
2748 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2749 "res: "DLDLMRES" rrc: %d type: %s [%llu->%llu] "
2750 "(req %llu->%llu) flags: %#llx nid: %s remote: "
2751 "%#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2752 ldlm_lock_to_ns_name(lock), lock,
2753 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2754 lock->l_readers, lock->l_writers,
2755 ldlm_lockname[lock->l_granted_mode],
2756 ldlm_lockname[lock->l_req_mode],
2758 atomic_read(&resource->lr_refcount),
2759 ldlm_typename[resource->lr_type],
2760 lock->l_policy_data.l_extent.start,
2761 lock->l_policy_data.l_extent.end,
2762 lock->l_req_extent.start, lock->l_req_extent.end,
2763 lock->l_flags, nid, lock->l_remote_handle.cookie,
2764 exp ? atomic_read(&exp->exp_refcount) : -99,
2765 lock->l_pid, lock->l_callback_timeout,
2770 libcfs_debug_vmsg2(msgdata, fmt, args,
2771 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2772 "res: "DLDLMRES" rrc: %d type: %s pid: %d "
2773 "[%llu->%llu] flags: %#llx nid: %s "
2774 "remote: %#llx expref: %d pid: %u timeout: %lld\n",
2775 ldlm_lock_to_ns_name(lock), lock,
2776 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2777 lock->l_readers, lock->l_writers,
2778 ldlm_lockname[lock->l_granted_mode],
2779 ldlm_lockname[lock->l_req_mode],
2781 atomic_read(&resource->lr_refcount),
2782 ldlm_typename[resource->lr_type],
2783 lock->l_policy_data.l_flock.pid,
2784 lock->l_policy_data.l_flock.start,
2785 lock->l_policy_data.l_flock.end,
2786 lock->l_flags, nid, lock->l_remote_handle.cookie,
2787 exp ? atomic_read(&exp->exp_refcount) : -99,
2788 lock->l_pid, lock->l_callback_timeout);
2792 libcfs_debug_vmsg2(msgdata, fmt, args,
2793 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2794 "res: "DLDLMRES" bits %#llx/%#llx rrc: %d type: %s "
2795 "flags: %#llx nid: %s remote: %#llx expref: %d "
2796 "pid: %u timeout: %lld lvb_type: %d\n",
2797 ldlm_lock_to_ns_name(lock),
2798 lock, lock->l_handle.h_cookie,
2799 atomic_read(&lock->l_refc),
2800 lock->l_readers, lock->l_writers,
2801 ldlm_lockname[lock->l_granted_mode],
2802 ldlm_lockname[lock->l_req_mode],
2804 lock->l_policy_data.l_inodebits.bits,
2805 lock->l_policy_data.l_inodebits.try_bits,
2806 atomic_read(&resource->lr_refcount),
2807 ldlm_typename[resource->lr_type],
2808 lock->l_flags, nid, lock->l_remote_handle.cookie,
2809 exp ? atomic_read(&exp->exp_refcount) : -99,
2810 lock->l_pid, lock->l_callback_timeout,
2815 libcfs_debug_vmsg2(msgdata, fmt, args,
2816 " ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s "
2817 "res: "DLDLMRES" rrc: %d type: %s flags: %#llx "
2818 "nid: %s remote: %#llx expref: %d pid: %u "
2819 "timeout: %lld lvb_type: %d\n",
2820 ldlm_lock_to_ns_name(lock),
2821 lock, lock->l_handle.h_cookie,
2822 atomic_read(&lock->l_refc),
2823 lock->l_readers, lock->l_writers,
2824 ldlm_lockname[lock->l_granted_mode],
2825 ldlm_lockname[lock->l_req_mode],
2827 atomic_read(&resource->lr_refcount),
2828 ldlm_typename[resource->lr_type],
2829 lock->l_flags, nid, lock->l_remote_handle.cookie,
2830 exp ? atomic_read(&exp->exp_refcount) : -99,
2831 lock->l_pid, lock->l_callback_timeout,
2836 ldlm_resource_putref(resource);
2838 EXPORT_SYMBOL(_ldlm_lock_debug);