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/
31 * lustre/ldlm/ldlm_lock.c
33 * Author: Peter Braam <braam@clusterfs.com>
34 * Author: Phil Schwan <phil@clusterfs.com>
37 #define DEBUG_SUBSYSTEM S_LDLM
39 #include <libcfs/libcfs.h>
41 #include <lustre_swab.h>
42 #include <obd_class.h>
44 #include "ldlm_internal.h"
46 struct kmem_cache *ldlm_glimpse_work_kmem;
47 EXPORT_SYMBOL(ldlm_glimpse_work_kmem);
50 char *ldlm_lockname[] = {
58 [LCK_GROUP] = "GROUP",
61 EXPORT_SYMBOL(ldlm_lockname);
63 char *ldlm_typename[] = {
65 [LDLM_EXTENT] = "EXT",
70 static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = {
71 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_wire_to_local,
72 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_wire_to_local,
73 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_wire_to_local,
74 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_wire_to_local,
77 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
78 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_local_to_wire,
79 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_local_to_wire,
80 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_local_to_wire,
81 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_local_to_wire,
85 * Converts lock policy from local format to on the wire lock_desc format
87 void ldlm_convert_policy_to_wire(enum ldlm_type type,
88 const union ldlm_policy_data *lpolicy,
89 union ldlm_wire_policy_data *wpolicy)
91 ldlm_policy_local_to_wire_t convert;
93 convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
95 convert(lpolicy, wpolicy);
99 * Converts lock policy from on the wire lock_desc format to local format
101 void ldlm_convert_policy_to_local(struct obd_export *exp, enum ldlm_type type,
102 const union ldlm_wire_policy_data *wpolicy,
103 union ldlm_policy_data *lpolicy)
105 ldlm_policy_wire_to_local_t convert;
107 convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE];
109 convert(wpolicy, lpolicy);
112 const char *ldlm_it2str(enum ldlm_intent_flags it)
119 case (IT_OPEN | IT_CREAT):
132 CERROR("Unknown intent 0x%08x\n", it);
136 EXPORT_SYMBOL(ldlm_it2str);
138 #ifdef HAVE_SERVER_SUPPORT
139 static ldlm_processing_policy ldlm_processing_policy_table[] = {
140 [LDLM_PLAIN] = ldlm_process_plain_lock,
141 [LDLM_EXTENT] = ldlm_process_extent_lock,
142 [LDLM_FLOCK] = ldlm_process_flock_lock,
143 [LDLM_IBITS] = ldlm_process_inodebits_lock,
146 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
148 return ldlm_processing_policy_table[res->lr_type];
150 EXPORT_SYMBOL(ldlm_get_processing_policy);
152 static ldlm_reprocessing_policy ldlm_reprocessing_policy_table[] = {
153 [LDLM_PLAIN] = ldlm_reprocess_queue,
154 [LDLM_EXTENT] = ldlm_reprocess_queue,
155 [LDLM_FLOCK] = ldlm_reprocess_queue,
156 [LDLM_IBITS] = ldlm_reprocess_inodebits_queue,
159 ldlm_reprocessing_policy ldlm_get_reprocessing_policy(struct ldlm_resource *res)
161 return ldlm_reprocessing_policy_table[res->lr_type];
164 #endif /* HAVE_SERVER_SUPPORT */
166 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
170 EXPORT_SYMBOL(ldlm_register_intent);
173 * REFCOUNTED LOCK OBJECTS
178 * Get a reference on a lock.
180 * Lock refcounts, during creation:
181 * - one special one for allocation, dec'd only once in destroy
182 * - one for being a lock that's in-use
183 * - one for the addref associated with a new lock
185 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
187 refcount_inc(&lock->l_handle.h_ref);
190 EXPORT_SYMBOL(ldlm_lock_get);
192 static void lock_handle_free(struct rcu_head *rcu)
194 struct ldlm_lock *lock = container_of(rcu, struct ldlm_lock,
197 OBD_FREE_PRE(lock, sizeof(*lock), "slab-freed");
198 kmem_cache_free(ldlm_lock_slab, lock);
202 * Release lock reference.
204 * Also frees the lock if it was last reference.
206 void ldlm_lock_put(struct ldlm_lock *lock)
210 LASSERT(lock->l_resource != LP_POISON);
211 LASSERT(refcount_read(&lock->l_handle.h_ref) > 0);
212 if (refcount_dec_and_test(&lock->l_handle.h_ref)) {
213 struct ldlm_resource *res;
216 "final lock_put on destroyed lock, freeing it.");
218 res = lock->l_resource;
219 LASSERT(ldlm_is_destroyed(lock));
220 LASSERT(list_empty(&lock->l_exp_list));
221 LASSERT(list_empty(&lock->l_res_link));
222 LASSERT(list_empty(&lock->l_pending_chain));
224 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
226 lu_ref_del(&res->lr_reference, "lock", lock);
227 if (lock->l_export) {
228 class_export_lock_put(lock->l_export, lock);
229 lock->l_export = NULL;
232 if (lock->l_lvb_data != NULL)
233 OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
235 if (res->lr_type == LDLM_EXTENT) {
236 ldlm_interval_free(ldlm_interval_detach(lock));
237 } else if (res->lr_type == LDLM_IBITS) {
238 if (lock->l_ibits_node != NULL)
239 OBD_SLAB_FREE_PTR(lock->l_ibits_node,
240 ldlm_inodebits_slab);
242 ldlm_resource_putref(res);
243 lock->l_resource = NULL;
244 lu_ref_fini(&lock->l_reference);
245 call_rcu(&lock->l_handle.h_rcu, lock_handle_free);
250 EXPORT_SYMBOL(ldlm_lock_put);
253 * Removes LDLM lock \a lock from LRU. Assumes LRU is already locked.
255 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
258 if (!list_empty(&lock->l_lru)) {
259 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
261 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
262 if (ns->ns_last_pos == &lock->l_lru)
263 ns->ns_last_pos = lock->l_lru.prev;
264 list_del_init(&lock->l_lru);
265 LASSERT(ns->ns_nr_unused > 0);
273 * Removes LDLM lock \a lock from LRU. Obtains the LRU lock first.
275 * If \a last_use is non-zero, it will remove the lock from LRU only if
276 * it matches lock's l_last_used.
278 * \retval 0 if \a last_use is set, the lock is not in LRU list or \a last_use
279 * doesn't match lock's l_last_used;
280 * otherwise, the lock hasn't been in the LRU list.
281 * \retval 1 the lock was in LRU list and removed.
283 int ldlm_lock_remove_from_lru_check(struct ldlm_lock *lock, ktime_t last_use)
285 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
289 if (ldlm_is_ns_srv(lock)) {
290 LASSERT(list_empty(&lock->l_lru));
294 spin_lock(&ns->ns_lock);
295 if (!ktime_compare(last_use, ktime_set(0, 0)) ||
296 !ktime_compare(last_use, lock->l_last_used))
297 rc = ldlm_lock_remove_from_lru_nolock(lock);
298 spin_unlock(&ns->ns_lock);
304 * Adds LDLM lock \a lock to namespace LRU. Assumes LRU is already locked.
306 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
308 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
310 lock->l_last_used = ktime_get();
311 LASSERT(list_empty(&lock->l_lru));
312 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
313 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
314 LASSERT(ns->ns_nr_unused >= 0);
319 * Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks
322 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
324 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
327 spin_lock(&ns->ns_lock);
328 ldlm_lock_add_to_lru_nolock(lock);
329 spin_unlock(&ns->ns_lock);
334 * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
335 * the LRU. Performs necessary LRU locking
337 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
339 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
342 if (ldlm_is_ns_srv(lock)) {
343 LASSERT(list_empty(&lock->l_lru));
348 spin_lock(&ns->ns_lock);
349 if (!list_empty(&lock->l_lru)) {
350 ldlm_lock_remove_from_lru_nolock(lock);
351 ldlm_lock_add_to_lru_nolock(lock);
353 spin_unlock(&ns->ns_lock);
358 * Helper to destroy a locked lock.
360 * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
361 * Must be called with l_lock and lr_lock held.
363 * Does not actually free the lock data, but rather marks the lock as
364 * destroyed by setting l_destroyed field in the lock to 1. Destroys a
365 * handle->lock association too, so that the lock can no longer be found
366 * and removes the lock from LRU list. Actual lock freeing occurs when
367 * last lock reference goes away.
369 * Original comment (of some historical value):
370 * This used to have a 'strict' flag, which recovery would use to mark an
371 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
372 * shall explain why it's gone: with the new hash table scheme, once you call
373 * ldlm_lock_destroy, you can never drop your final references on this lock.
374 * Because it's not in the hash table anymore. -phil
376 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
380 if (lock->l_readers || lock->l_writers) {
381 LDLM_ERROR(lock, "lock still has references");
385 if (!list_empty(&lock->l_res_link)) {
386 LDLM_ERROR(lock, "lock still on resource");
390 if (ldlm_is_destroyed(lock)) {
391 LASSERT(list_empty(&lock->l_lru));
395 ldlm_set_destroyed(lock);
397 if (lock->l_export && lock->l_export->exp_lock_hash) {
398 /* NB: it's safe to call cfs_hash_del() even lock isn't
399 * in exp_lock_hash. */
400 /* In the function below, .hs_keycmp resolves to
401 * ldlm_export_lock_keycmp() */
402 /* coverity[overrun-buffer-val] */
403 cfs_hash_del(lock->l_export->exp_lock_hash,
404 &lock->l_remote_handle, &lock->l_exp_hash);
407 ldlm_lock_remove_from_lru(lock);
408 class_handle_unhash(&lock->l_handle);
415 * Destroys a LDLM lock \a lock. Performs necessary locking first.
417 void ldlm_lock_destroy(struct ldlm_lock *lock)
421 lock_res_and_lock(lock);
422 first = ldlm_lock_destroy_internal(lock);
423 unlock_res_and_lock(lock);
425 /* drop reference from hashtable only for first destroy */
427 lu_ref_del(&lock->l_reference, "hash", lock);
428 LDLM_LOCK_RELEASE(lock);
434 * Destroys a LDLM lock \a lock that is already locked.
436 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
440 first = ldlm_lock_destroy_internal(lock);
441 /* drop reference from hashtable only for first destroy */
443 lu_ref_del(&lock->l_reference, "hash", lock);
444 LDLM_LOCK_RELEASE(lock);
449 static const char lock_handle_owner[] = "ldlm";
453 * Allocate and initialize new lock structure.
455 * usage: pass in a resource on which you have done ldlm_resource_get
456 * new lock will take over the refcount.
457 * returns: lock with refcount 2 - one for current caller and one for remote
459 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
461 struct ldlm_lock *lock;
464 if (resource == NULL)
467 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, GFP_NOFS);
471 RCU_INIT_POINTER(lock->l_resource, resource);
472 lu_ref_add(&resource->lr_reference, "lock", lock);
474 refcount_set(&lock->l_handle.h_ref, 2);
475 INIT_LIST_HEAD(&lock->l_res_link);
476 INIT_LIST_HEAD(&lock->l_lru);
477 INIT_LIST_HEAD(&lock->l_pending_chain);
478 INIT_LIST_HEAD(&lock->l_bl_ast);
479 INIT_LIST_HEAD(&lock->l_cp_ast);
480 INIT_LIST_HEAD(&lock->l_rk_ast);
481 init_waitqueue_head(&lock->l_waitq);
482 lock->l_blocking_lock = NULL;
483 INIT_LIST_HEAD(&lock->l_sl_mode);
484 INIT_LIST_HEAD(&lock->l_sl_policy);
485 INIT_HLIST_NODE(&lock->l_exp_hash);
486 INIT_HLIST_NODE(&lock->l_exp_flock_hash);
488 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
490 INIT_HLIST_NODE(&lock->l_handle.h_link);
491 class_handle_hash(&lock->l_handle, lock_handle_owner);
493 lu_ref_init(&lock->l_reference);
494 lu_ref_add(&lock->l_reference, "hash", lock);
495 lock->l_callback_timestamp = 0;
496 lock->l_activity = 0;
498 #if LUSTRE_TRACKS_LOCK_EXP_REFS
499 INIT_LIST_HEAD(&lock->l_exp_refs_link);
500 lock->l_exp_refs_nr = 0;
501 lock->l_exp_refs_target = NULL;
503 INIT_LIST_HEAD(&lock->l_exp_list);
509 * Moves LDLM lock \a lock to another resource.
510 * This is used on client when server returns some other lock than requested
511 * (typically as a result of intent operation)
513 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
514 const struct ldlm_res_id *new_resid)
516 struct ldlm_resource *oldres;
517 struct ldlm_resource *newres;
521 LASSERT(ns_is_client(ns));
523 oldres = lock_res_and_lock(lock);
524 if (memcmp(new_resid, &oldres->lr_name,
525 sizeof(oldres->lr_name)) == 0) {
527 unlock_res_and_lock(lock);
531 LASSERT(new_resid->name[0] != 0);
533 /* This function assumes that the lock isn't on any lists */
534 LASSERT(list_empty(&lock->l_res_link));
536 type = oldres->lr_type;
537 unlock_res_and_lock(lock);
539 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
541 RETURN(PTR_ERR(newres));
543 lu_ref_add(&newres->lr_reference, "lock", lock);
545 * To flip the lock from the old to the new resource, oldres
546 * and newres have to be locked. Resource spin-locks are taken
547 * in the memory address order to avoid dead-locks.
548 * As this is the only circumstance where ->l_resource
549 * can change, and this cannot race with itself, it is safe
550 * to access lock->l_resource without being careful about locking.
552 oldres = lock->l_resource;
553 if (oldres < newres) {
555 lock_res_nested(newres, LRT_NEW);
558 lock_res_nested(oldres, LRT_NEW);
560 LASSERT(memcmp(new_resid, &oldres->lr_name,
561 sizeof oldres->lr_name) != 0);
562 rcu_assign_pointer(lock->l_resource, newres);
566 /* ...and the flowers are still standing! */
567 lu_ref_del(&oldres->lr_reference, "lock", lock);
568 ldlm_resource_putref(oldres);
573 /** \defgroup ldlm_handles LDLM HANDLES
574 * Ways to get hold of locks without any addresses.
579 * Fills in handle for LDLM lock \a lock into supplied \a lockh
580 * Does not take any references.
582 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
584 lockh->cookie = lock->l_handle.h_cookie;
586 EXPORT_SYMBOL(ldlm_lock2handle);
589 * Obtain a lock reference by handle.
591 * if \a flags: atomically get the lock and set the flags.
592 * Return NULL if flag already set
594 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
597 struct ldlm_lock *lock;
602 if (!lustre_handle_is_used(handle))
605 lock = class_handle2object(handle->cookie, lock_handle_owner);
609 if (lock->l_export != NULL && lock->l_export->exp_failed) {
610 CDEBUG(D_INFO, "lock export failed: lock %p, exp %p\n",
611 lock, lock->l_export);
616 /* It's unlikely but possible that someone marked the lock as
617 * destroyed after we did handle2object on it */
618 if ((flags == 0) && !ldlm_is_destroyed(lock)) {
619 lu_ref_add_atomic(&lock->l_reference, "handle", lock);
623 lock_res_and_lock(lock);
625 LASSERT(lock->l_resource != NULL);
627 lu_ref_add_atomic(&lock->l_reference, "handle", lock);
628 if (unlikely(ldlm_is_destroyed(lock))) {
629 unlock_res_and_lock(lock);
630 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
635 /* If we're setting flags, make sure none of them are already set. */
637 if ((lock->l_flags & flags) != 0) {
638 unlock_res_and_lock(lock);
643 lock->l_flags |= flags;
646 unlock_res_and_lock(lock);
649 EXPORT_SYMBOL(__ldlm_handle2lock);
650 /** @} ldlm_handles */
653 * Fill in "on the wire" representation for given LDLM lock into supplied
654 * lock descriptor \a desc structure.
656 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
658 ldlm_res2desc(lock->l_resource, &desc->l_resource);
659 desc->l_req_mode = lock->l_req_mode;
660 desc->l_granted_mode = lock->l_granted_mode;
661 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
662 &lock->l_policy_data,
663 &desc->l_policy_data);
667 * Add a lock to list of conflicting locks to send AST to.
669 * Only add if we have not sent a blocking AST to the lock yet.
671 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
672 struct list_head *work_list)
674 if (!ldlm_is_ast_sent(lock)) {
675 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
676 ldlm_set_ast_sent(lock);
677 /* If the enqueuing client said so, tell the AST recipient to
678 * discard dirty data, rather than writing back. */
679 if (ldlm_is_ast_discard_data(new))
680 ldlm_set_discard_data(lock);
682 /* Lock can be converted from a blocking state back to granted
683 * after lock convert or COS downgrade but still be in an
684 * older bl_list because it is controlled only by
685 * ldlm_work_bl_ast_lock(), let it be processed there.
687 if (list_empty(&lock->l_bl_ast)) {
688 list_add(&lock->l_bl_ast, work_list);
691 LASSERT(lock->l_blocking_lock == NULL);
692 lock->l_blocking_lock = LDLM_LOCK_GET(new);
697 * Add a lock to list of just granted locks to send completion AST to.
699 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
700 struct list_head *work_list)
702 if (!ldlm_is_cp_reqd(lock)) {
703 ldlm_set_cp_reqd(lock);
704 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
705 LASSERT(list_empty(&lock->l_cp_ast));
706 list_add(&lock->l_cp_ast, work_list);
712 * Aggregator function to add AST work items into a list. Determines
713 * what sort of an AST work needs to be done and calls the proper
715 * Must be called with lr_lock held.
717 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
718 struct list_head *work_list)
721 check_res_locked(lock->l_resource);
723 ldlm_add_bl_work_item(lock, new, work_list);
725 ldlm_add_cp_work_item(lock, work_list);
730 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
731 * r/w reference type is determined by \a mode
732 * Calls ldlm_lock_addref_internal.
734 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
736 struct ldlm_lock *lock;
738 lock = ldlm_handle2lock(lockh);
739 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
740 ldlm_lock_addref_internal(lock, mode);
743 EXPORT_SYMBOL(ldlm_lock_addref);
747 * Add specified reader/writer reference to LDLM lock \a lock.
748 * r/w reference type is determined by \a mode
749 * Removes lock from LRU if it is there.
750 * Assumes the LDLM lock is already locked.
752 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
755 ldlm_lock_remove_from_lru(lock);
756 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
758 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
760 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
762 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
765 lu_ref_add_atomic(&lock->l_reference, "user", lock);
766 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
770 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
771 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
773 * \retval 0 success, lock was addref-ed
775 * \retval -EAGAIN lock is being canceled.
777 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
779 struct ldlm_lock *lock;
783 lock = ldlm_handle2lock(lockh);
785 lock_res_and_lock(lock);
786 if (lock->l_readers != 0 || lock->l_writers != 0 ||
787 !ldlm_is_cbpending(lock)) {
788 ldlm_lock_addref_internal_nolock(lock, mode);
791 unlock_res_and_lock(lock);
796 EXPORT_SYMBOL(ldlm_lock_addref_try);
799 * Add specified reader/writer reference to LDLM lock \a lock.
800 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
801 * Only called for local locks.
803 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
805 lock_res_and_lock(lock);
806 ldlm_lock_addref_internal_nolock(lock, mode);
807 unlock_res_and_lock(lock);
811 * Removes reader/writer reference for LDLM lock \a lock.
812 * Assumes LDLM lock is already locked.
813 * only called in ldlm_flock_destroy and for local locks.
814 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
815 * that cannot be placed in LRU.
817 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
820 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
821 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
822 LASSERT(lock->l_readers > 0);
823 lu_ref_del(&lock->l_reference, "reader", lock);
826 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
827 LASSERT(lock->l_writers > 0);
828 lu_ref_del(&lock->l_reference, "writer", lock);
832 lu_ref_del(&lock->l_reference, "user", lock);
833 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
837 * Removes reader/writer reference for LDLM lock \a lock.
838 * Locks LDLM lock first.
839 * If the lock is determined to be client lock on a client and r/w refcount
840 * drops to zero and the lock is not blocked, the lock is added to LRU lock
842 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
844 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
846 struct ldlm_namespace *ns;
850 lock_res_and_lock(lock);
852 ns = ldlm_lock_to_ns(lock);
854 ldlm_lock_decref_internal_nolock(lock, mode);
856 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
857 !lock->l_readers && !lock->l_writers) {
858 /* If this is a local lock on a server namespace and this was
859 * the last reference, cancel the lock.
861 * Group locks are special:
862 * They must not go in LRU, but they are not called back
863 * like non-group locks, instead they are manually released.
864 * They have an l_writers reference which they keep until
865 * they are manually released, so we remove them when they have
866 * no more reader or writer references. - LU-6368 */
867 ldlm_set_cbpending(lock);
870 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
871 unsigned int mask = D_DLMTRACE;
873 /* If we received a blocked AST and this was the last reference,
874 * run the callback. */
875 if (ldlm_is_ns_srv(lock) && lock->l_export)
877 LDLM_DEBUG_LIMIT(mask, lock,
878 "final decref done on %sCBPENDING lock",
879 mask & D_WARNING ? "non-local " : "");
881 LDLM_LOCK_GET(lock); /* dropped by bl thread */
882 ldlm_lock_remove_from_lru(lock);
883 unlock_res_and_lock(lock);
885 if (ldlm_is_fail_loc(lock))
886 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
888 if (ldlm_is_atomic_cb(lock) ||
889 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
890 ldlm_handle_bl_callback(ns, NULL, lock);
891 } else if (ns_is_client(ns) &&
892 !lock->l_readers && !lock->l_writers &&
893 !ldlm_is_no_lru(lock) &&
894 !ldlm_is_bl_ast(lock) &&
895 !ldlm_is_converting(lock)) {
897 /* If this is a client-side namespace and this was the last
898 * reference, put it on the LRU.
900 ldlm_lock_add_to_lru(lock);
901 unlock_res_and_lock(lock);
902 LDLM_DEBUG(lock, "add lock into lru list");
904 if (ldlm_is_fail_loc(lock))
905 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
907 ldlm_pool_recalc(&ns->ns_pool, true);
909 LDLM_DEBUG(lock, "do not add lock into lru list");
910 unlock_res_and_lock(lock);
917 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
919 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
921 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
922 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
923 ldlm_lock_decref_internal(lock, mode);
926 EXPORT_SYMBOL(ldlm_lock_decref);
929 * Decrease reader/writer refcount for LDLM lock with handle
930 * \a lockh and mark it for subsequent cancellation once r/w refcount
931 * drops to zero instead of putting into LRU.
934 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
937 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
940 LASSERT(lock != NULL);
942 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
943 lock_res_and_lock(lock);
944 ldlm_set_cbpending(lock);
945 unlock_res_and_lock(lock);
946 ldlm_lock_decref_internal(lock, mode);
949 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
951 struct sl_insert_point {
952 struct list_head *res_link;
953 struct list_head *mode_link;
954 struct list_head *policy_link;
958 * Finds a position to insert the new lock into granted lock list.
960 * Used for locks eligible for skiplist optimization.
963 * queue [input]: the granted list where search acts on;
964 * req [input]: the lock whose position to be located;
965 * prev [output]: positions within 3 lists to insert @req to
969 * - ldlm_grant_lock_with_skiplist
971 static void search_granted_lock(struct list_head *queue,
972 struct ldlm_lock *req,
973 struct sl_insert_point *prev)
975 struct list_head *tmp;
976 struct ldlm_lock *lock, *mode_end, *policy_end;
979 list_for_each(tmp, queue) {
980 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
982 mode_end = list_entry(lock->l_sl_mode.prev,
983 struct ldlm_lock, l_sl_mode);
985 if (lock->l_req_mode != req->l_req_mode) {
986 /* jump to last lock of mode group */
987 tmp = &mode_end->l_res_link;
991 /* suitable mode group is found */
992 if (lock->l_resource->lr_type == LDLM_PLAIN) {
993 /* insert point is last lock of the mode group */
994 prev->res_link = &mode_end->l_res_link;
995 prev->mode_link = &mode_end->l_sl_mode;
996 prev->policy_link = &req->l_sl_policy;
999 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
1002 list_entry(lock->l_sl_policy.prev,
1006 if (lock->l_policy_data.l_inodebits.bits ==
1007 req->l_policy_data.l_inodebits.bits) {
1008 /* insert point is last lock of
1009 * the policy group */
1011 &policy_end->l_res_link;
1013 &policy_end->l_sl_mode;
1015 &policy_end->l_sl_policy;
1020 if (policy_end == mode_end)
1021 /* done with mode group */
1024 /* go to next policy group within mode group */
1025 tmp = policy_end->l_res_link.next;
1026 lock = list_entry(tmp, struct ldlm_lock,
1028 } /* loop over policy groups within the mode group */
1030 /* insert point is last lock of the mode group,
1031 * new policy group is started */
1032 prev->res_link = &mode_end->l_res_link;
1033 prev->mode_link = &mode_end->l_sl_mode;
1034 prev->policy_link = &req->l_sl_policy;
1038 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1043 /* insert point is last lock on the queue,
1044 * new mode group and new policy group are started */
1045 prev->res_link = queue->prev;
1046 prev->mode_link = &req->l_sl_mode;
1047 prev->policy_link = &req->l_sl_policy;
1052 * Add a lock into resource granted list after a position described by
1055 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1056 struct sl_insert_point *prev)
1058 struct ldlm_resource *res = lock->l_resource;
1061 check_res_locked(res);
1063 ldlm_resource_dump(D_INFO, res);
1064 LDLM_DEBUG(lock, "About to add lock:");
1066 if (ldlm_is_destroyed(lock)) {
1067 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1071 LASSERT(list_empty(&lock->l_res_link));
1072 LASSERT(list_empty(&lock->l_sl_mode));
1073 LASSERT(list_empty(&lock->l_sl_policy));
1076 * lock->link == prev->link means lock is first starting the group.
1077 * Don't re-add to itself to suppress kernel warnings.
1079 if (&lock->l_res_link != prev->res_link)
1080 list_add(&lock->l_res_link, prev->res_link);
1081 if (&lock->l_sl_mode != prev->mode_link)
1082 list_add(&lock->l_sl_mode, prev->mode_link);
1083 if (&lock->l_sl_policy != prev->policy_link)
1084 list_add(&lock->l_sl_policy, prev->policy_link);
1090 * Add a lock to granted list on a resource maintaining skiplist
1093 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1095 struct sl_insert_point prev;
1097 LASSERT(ldlm_is_granted(lock));
1099 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1100 ldlm_granted_list_add_lock(lock, &prev);
1104 * Perform lock granting bookkeeping.
1106 * Includes putting the lock into granted list and updating lock mode.
1108 * - ldlm_lock_enqueue
1109 * - ldlm_reprocess_queue
1111 * must be called with lr_lock held
1113 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1115 struct ldlm_resource *res = lock->l_resource;
1118 check_res_locked(res);
1120 lock->l_granted_mode = lock->l_req_mode;
1122 if (work_list && lock->l_completion_ast != NULL)
1123 ldlm_add_ast_work_item(lock, NULL, work_list);
1125 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1126 ldlm_grant_lock_with_skiplist(lock);
1127 else if (res->lr_type == LDLM_EXTENT)
1128 ldlm_extent_add_lock(res, lock);
1129 else if (res->lr_type == LDLM_FLOCK) {
1130 /* We should not add locks to granted list in the following
1132 * - this is an UNLOCK but not a real lock;
1133 * - this is a TEST lock;
1134 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1135 * - this is a deadlock (flock cannot be granted) */
1136 if (lock->l_req_mode == 0 ||
1137 lock->l_req_mode == LCK_NL ||
1138 ldlm_is_test_lock(lock) ||
1139 ldlm_is_flock_deadlock(lock))
1141 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1146 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1151 * Check if the given @lock meets the criteria for a match.
1152 * A reference on the lock is taken if matched.
1154 * @lock test-against this lock
1157 * RETURN returns true if @lock matches @data, false otherwise
1159 static bool lock_matches(struct ldlm_lock *lock, struct ldlm_match_data *data)
1161 union ldlm_policy_data *lpol = &lock->l_policy_data;
1162 enum ldlm_mode match = LCK_MINMODE;
1164 if (lock == data->lmd_old)
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))
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))
1182 if (!(data->lmd_match & LDLM_MATCH_UNREF) && ldlm_is_cbpending(lock) &&
1183 lock->l_readers == 0 && lock->l_writers == 0)
1186 if (!(lock->l_req_mode & *data->lmd_mode))
1189 /* When we search for ast_data, we are not doing a traditional match,
1190 * so we don't worry about IBITS or extent matching.
1192 if (data->lmd_match & (LDLM_MATCH_AST | LDLM_MATCH_AST_ANY)) {
1193 if (!lock->l_ast_data)
1196 if (data->lmd_match & LDLM_MATCH_AST_ANY)
1200 match = lock->l_req_mode;
1202 switch (lock->l_resource->lr_type) {
1204 if (!(data->lmd_match & LDLM_MATCH_RIGHT) &&
1205 (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1206 lpol->l_extent.end < data->lmd_policy->l_extent.end))
1209 if (unlikely(match == LCK_GROUP) &&
1210 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1211 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1215 /* We match if we have existing lock with same or wider set
1217 if ((lpol->l_inodebits.bits &
1218 data->lmd_policy->l_inodebits.bits) !=
1219 data->lmd_policy->l_inodebits.bits)
1222 if (unlikely(match == LCK_GROUP) &&
1223 data->lmd_policy->l_inodebits.li_gid != LDLM_GID_ANY &&
1224 lpol->l_inodebits.li_gid !=
1225 data->lmd_policy->l_inodebits.li_gid)
1232 /* We match if we have existing lock with same or wider set
1234 if (!(data->lmd_match & LDLM_MATCH_UNREF) && LDLM_HAVE_MASK(lock, GONE))
1237 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1240 /* Filter locks by skipping flags */
1241 if (data->lmd_skip_flags & lock->l_flags)
1245 if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1246 LDLM_LOCK_GET(lock);
1247 ldlm_lock_touch_in_lru(lock);
1249 ldlm_lock_addref_internal_nolock(lock, match);
1252 *data->lmd_mode = match;
1253 data->lmd_lock = lock;
1258 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1260 struct ldlm_interval *node = to_ldlm_interval(in);
1261 struct ldlm_match_data *data = args;
1262 struct ldlm_lock *lock;
1264 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1265 if (lock_matches(lock, data))
1266 return INTERVAL_ITER_STOP;
1268 return INTERVAL_ITER_CONT;
1272 * Search for a lock with given parameters in interval trees.
1274 * \param res search for a lock in this resource
1275 * \param data parameters
1277 * \retval a referenced lock or NULL.
1279 struct ldlm_lock *search_itree(struct ldlm_resource *res,
1280 struct ldlm_match_data *data)
1282 struct interval_node_extent ext = {
1283 .start = data->lmd_policy->l_extent.start,
1284 .end = data->lmd_policy->l_extent.end
1288 data->lmd_lock = NULL;
1290 if (data->lmd_match & LDLM_MATCH_RIGHT)
1291 ext.end = OBD_OBJECT_EOF;
1293 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1294 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1296 if (tree->lit_root == NULL)
1299 if (!(tree->lit_mode & *data->lmd_mode))
1302 interval_search(tree->lit_root, &ext,
1303 itree_overlap_cb, data);
1305 return data->lmd_lock;
1310 EXPORT_SYMBOL(search_itree);
1314 * Search for a lock with given properties in a queue.
1316 * \param queue search for a lock in this queue
1317 * \param data parameters
1319 * \retval a referenced lock or NULL.
1321 static struct ldlm_lock *search_queue(struct list_head *queue,
1322 struct ldlm_match_data *data)
1324 struct ldlm_lock *lock;
1326 data->lmd_lock = NULL;
1328 list_for_each_entry(lock, queue, l_res_link)
1329 if (lock_matches(lock, data))
1330 return data->lmd_lock;
1335 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1337 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1338 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1339 wake_up(&lock->l_waitq);
1342 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1344 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1346 lock_res_and_lock(lock);
1347 ldlm_lock_fail_match_locked(lock);
1348 unlock_res_and_lock(lock);
1352 * Mark lock as "matchable" by OST.
1354 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1356 * Assumes LDLM lock is already locked.
1358 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1360 ldlm_set_lvb_ready(lock);
1361 wake_up(&lock->l_waitq);
1363 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1366 * Mark lock as "matchable" by OST.
1367 * Locks the lock and then \see ldlm_lock_allow_match_locked
1369 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1371 lock_res_and_lock(lock);
1372 ldlm_lock_allow_match_locked(lock);
1373 unlock_res_and_lock(lock);
1375 EXPORT_SYMBOL(ldlm_lock_allow_match);
1378 * Attempt to find a lock with specified properties.
1380 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1383 * Can be called in two ways:
1385 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1386 * for a duplicate of.
1388 * Otherwise, all of the fields must be filled in, to match against.
1390 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1391 * server (ie, connh is NULL)
1392 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1393 * list will be considered
1394 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1395 * to be canceled can still be matched as long as they still have reader
1396 * or writer refernces
1397 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1398 * just tell us if we would have matched.
1400 * \retval 1 if it finds an already-existing lock that is compatible; in this
1401 * case, lockh is filled in with a addref()ed lock
1403 * We also check security context, and if that fails we simply return 0 (to
1404 * keep caller code unchanged), the context failure will be discovered by
1405 * caller sometime later.
1407 enum ldlm_mode ldlm_lock_match_with_skip(struct ldlm_namespace *ns,
1408 __u64 flags, __u64 skip_flags,
1409 const struct ldlm_res_id *res_id,
1410 enum ldlm_type type,
1411 union ldlm_policy_data *policy,
1412 enum ldlm_mode mode,
1413 struct lustre_handle *lockh,
1414 enum ldlm_match_flags match_flags)
1416 struct ldlm_match_data data = {
1420 .lmd_policy = policy,
1422 .lmd_skip_flags = skip_flags,
1423 .lmd_match = match_flags,
1425 struct ldlm_resource *res;
1426 struct ldlm_lock *lock;
1432 data.lmd_old = ldlm_handle2lock(lockh);
1433 LASSERT(data.lmd_old != NULL);
1435 ns = ldlm_lock_to_ns(data.lmd_old);
1436 res_id = &data.lmd_old->l_resource->lr_name;
1437 type = data.lmd_old->l_resource->lr_type;
1438 *data.lmd_mode = data.lmd_old->l_req_mode;
1441 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1443 LASSERT(data.lmd_old == NULL);
1447 LDLM_RESOURCE_ADDREF(res);
1449 if (res->lr_type == LDLM_EXTENT)
1450 lock = search_itree(res, &data);
1452 lock = search_queue(&res->lr_granted, &data);
1453 if (!lock && !(flags & LDLM_FL_BLOCK_GRANTED))
1454 lock = search_queue(&res->lr_waiting, &data);
1455 matched = lock ? mode : 0;
1457 LDLM_RESOURCE_DELREF(res);
1458 ldlm_resource_putref(res);
1461 ldlm_lock2handle(lock, lockh);
1462 if ((flags & LDLM_FL_LVB_READY) &&
1463 (!ldlm_is_lvb_ready(lock))) {
1464 __u64 wait_flags = LDLM_FL_LVB_READY |
1465 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1467 if (lock->l_completion_ast) {
1468 int err = lock->l_completion_ast(lock,
1469 LDLM_FL_WAIT_NOREPROC,
1472 GOTO(out_fail_match, matched = 0);
1475 wait_event_idle_timeout(
1477 lock->l_flags & wait_flags,
1478 cfs_time_seconds(obd_timeout));
1480 if (!ldlm_is_lvb_ready(lock))
1481 GOTO(out_fail_match, matched = 0);
1484 /* check user's security context */
1485 if (lock->l_conn_export &&
1486 sptlrpc_import_check_ctx(
1487 class_exp2cliimp(lock->l_conn_export)))
1488 GOTO(out_fail_match, matched = 0);
1490 LDLM_DEBUG(lock, "matched (%llu %llu)",
1491 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1492 res_id->name[2] : policy->l_extent.start,
1493 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1494 res_id->name[3] : policy->l_extent.end);
1497 if (flags & LDLM_FL_TEST_LOCK)
1498 LDLM_LOCK_RELEASE(lock);
1500 ldlm_lock_decref_internal(lock, mode);
1503 /* less verbose for test-only */
1504 if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
1505 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1506 "%llu/%llu (%llu %llu)", ns,
1507 type, mode, res_id->name[0], res_id->name[1],
1508 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1509 res_id->name[2] : policy->l_extent.start,
1510 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1511 res_id->name[3] : policy->l_extent.end);
1513 if (data.lmd_old != NULL)
1514 LDLM_LOCK_PUT(data.lmd_old);
1518 EXPORT_SYMBOL(ldlm_lock_match_with_skip);
1520 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1523 struct ldlm_lock *lock;
1524 enum ldlm_mode mode = 0;
1527 lock = ldlm_handle2lock(lockh);
1529 lock_res_and_lock(lock);
1530 if (LDLM_HAVE_MASK(lock, GONE))
1533 if (ldlm_is_cbpending(lock) &&
1534 lock->l_readers == 0 && lock->l_writers == 0)
1538 *bits = lock->l_policy_data.l_inodebits.bits;
1539 mode = lock->l_granted_mode;
1540 ldlm_lock_addref_internal_nolock(lock, mode);
1547 unlock_res_and_lock(lock);
1548 LDLM_LOCK_PUT(lock);
1552 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1554 /** The caller must guarantee that the buffer is large enough. */
1555 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1556 enum req_location loc, void *data, int size)
1561 LASSERT(data != NULL);
1564 switch (lock->l_lvb_type) {
1566 if (size == sizeof(struct ost_lvb)) {
1567 if (loc == RCL_CLIENT)
1568 lvb = req_capsule_client_swab_get(pill,
1570 lustre_swab_ost_lvb);
1572 lvb = req_capsule_server_swab_get(pill,
1574 lustre_swab_ost_lvb);
1575 if (unlikely(lvb == NULL)) {
1576 LDLM_ERROR(lock, "no LVB");
1580 memcpy(data, lvb, size);
1581 } else if (size == sizeof(struct ost_lvb_v1)) {
1582 struct ost_lvb *olvb = data;
1584 if (loc == RCL_CLIENT)
1585 lvb = req_capsule_client_swab_get(pill,
1587 lustre_swab_ost_lvb_v1);
1589 lvb = req_capsule_server_sized_swab_get(pill,
1591 lustre_swab_ost_lvb_v1);
1592 if (unlikely(lvb == NULL)) {
1593 LDLM_ERROR(lock, "no LVB");
1597 memcpy(data, lvb, size);
1598 olvb->lvb_mtime_ns = 0;
1599 olvb->lvb_atime_ns = 0;
1600 olvb->lvb_ctime_ns = 0;
1602 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1608 if (size == sizeof(struct lquota_lvb)) {
1609 if (loc == RCL_CLIENT)
1610 lvb = req_capsule_client_swab_get(pill,
1612 lustre_swab_lquota_lvb);
1614 lvb = req_capsule_server_swab_get(pill,
1616 lustre_swab_lquota_lvb);
1617 if (unlikely(lvb == NULL)) {
1618 LDLM_ERROR(lock, "no LVB");
1622 memcpy(data, lvb, size);
1624 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1633 if (loc == RCL_CLIENT)
1634 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1636 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1637 if (unlikely(lvb == NULL)) {
1638 LDLM_ERROR(lock, "no LVB");
1642 memcpy(data, lvb, size);
1645 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1646 libcfs_debug_dumpstack(NULL);
1654 * Create and fill in new LDLM lock with specified properties.
1655 * Returns a referenced lock
1657 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1658 const struct ldlm_res_id *res_id,
1659 enum ldlm_type type,
1660 enum ldlm_mode mode,
1661 const struct ldlm_callback_suite *cbs,
1662 void *data, __u32 lvb_len,
1663 enum lvb_type lvb_type)
1665 struct ldlm_lock *lock;
1666 struct ldlm_resource *res;
1670 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1672 RETURN(ERR_CAST(res));
1674 lock = ldlm_lock_new(res);
1676 RETURN(ERR_PTR(-ENOMEM));
1678 lock->l_req_mode = mode;
1679 lock->l_ast_data = data;
1680 lock->l_pid = current->pid;
1681 if (ns_is_server(ns))
1682 ldlm_set_ns_srv(lock);
1684 lock->l_blocking_ast = cbs->lcs_blocking;
1685 lock->l_completion_ast = cbs->lcs_completion;
1686 lock->l_glimpse_ast = cbs->lcs_glimpse;
1691 rc = ldlm_extent_alloc_lock(lock);
1694 rc = ldlm_inodebits_alloc_lock(lock);
1703 lock->l_lvb_len = lvb_len;
1704 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1705 if (lock->l_lvb_data == NULL)
1706 GOTO(out, rc = -ENOMEM);
1709 lock->l_lvb_type = lvb_type;
1710 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1711 GOTO(out, rc = -ENOENT);
1716 ldlm_lock_destroy(lock);
1717 LDLM_LOCK_RELEASE(lock);
1718 RETURN(ERR_PTR(rc));
1721 #ifdef HAVE_SERVER_SUPPORT
1722 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1725 struct ldlm_resource *res = lock->l_resource;
1726 enum ldlm_error rc = ELDLM_OK;
1727 LIST_HEAD(rpc_list);
1728 ldlm_processing_policy policy;
1732 policy = ldlm_get_processing_policy(res);
1733 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1734 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1735 res->lr_type != LDLM_FLOCK)
1736 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1738 if (!list_empty(&rpc_list))
1739 ldlm_discard_bl_list(&rpc_list);
1746 * Enqueue (request) a lock.
1748 * Does not block. As a result of enqueue the lock would be put
1749 * into granted or waiting list.
1751 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1752 * set, skip all the enqueueing and delegate lock processing to intent policy
1755 enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
1756 struct ldlm_namespace *ns,
1757 struct ldlm_lock **lockp,
1758 void *cookie, __u64 *flags)
1760 struct ldlm_lock *lock = *lockp;
1761 struct ldlm_resource *res;
1762 int local = ns_is_client(ns);
1763 enum ldlm_error rc = ELDLM_OK;
1764 struct ldlm_interval *node = NULL;
1765 #ifdef HAVE_SERVER_SUPPORT
1766 bool reconstruct = false;
1770 /* policies are not executed on the client or during replay */
1771 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1772 && !local && ns->ns_policy) {
1773 rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
1775 if (rc == ELDLM_LOCK_REPLACED) {
1776 /* The lock that was returned has already been granted,
1777 * and placed into lockp. If it's not the same as the
1778 * one we passed in, then destroy the old one and our
1779 * work here is done. */
1780 if (lock != *lockp) {
1781 ldlm_lock_destroy(lock);
1782 LDLM_LOCK_RELEASE(lock);
1784 *flags |= LDLM_FL_LOCK_CHANGED;
1786 } else if (rc != ELDLM_OK &&
1787 ldlm_is_granted(lock)) {
1788 LASSERT(*flags & LDLM_FL_RESENT);
1789 /* It may happen that ns_policy returns an error in
1790 * resend case, object may be unlinked or just some
1791 * error occurs. It is unclear if lock reached the
1792 * client in the original reply, just leave the lock on
1793 * server, not returning it again to client. Due to
1794 * LU-6529, the server will not OOM. */
1796 } else if (rc != ELDLM_OK ||
1797 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1798 ldlm_lock_destroy(lock);
1803 if (*flags & LDLM_FL_RESENT) {
1804 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1805 * Set LOCK_CHANGED always.
1806 * Check if the lock is granted for BLOCK_GRANTED.
1807 * Take NO_TIMEOUT from the lock as it is inherited through
1808 * LDLM_FL_INHERIT_MASK */
1809 *flags |= LDLM_FL_LOCK_CHANGED;
1810 if (!ldlm_is_granted(lock))
1811 *flags |= LDLM_FL_BLOCK_GRANTED;
1812 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1816 #ifdef HAVE_SERVER_SUPPORT
1817 /* For a replaying lock, it might be already in granted list. So
1818 * unlinking the lock will cause the interval node to be freed, we
1819 * have to allocate the interval node early otherwise we can't regrant
1820 * this lock in the future. - jay
1822 * The only time the ldlm_resource changes for the ldlm_lock is when
1823 * ldlm_lock_change_resource() is called and that only happens for
1824 * the Lustre client case.
1826 if (!local && (*flags & LDLM_FL_REPLAY) &&
1827 lock->l_resource->lr_type == LDLM_EXTENT)
1828 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1830 reconstruct = !local && lock->l_resource->lr_type == LDLM_FLOCK &&
1831 !(*flags & LDLM_FL_TEST_LOCK);
1833 rc = req_can_reconstruct(cookie, NULL);
1841 res = lock_res_and_lock(lock);
1842 if (local && ldlm_is_granted(lock)) {
1843 /* The server returned a blocked lock, but it was granted
1844 * before we got a chance to actually enqueue it. We don't
1845 * need to do anything else. */
1846 *flags &= ~LDLM_FL_BLOCKED_MASK;
1847 GOTO(out, rc = ELDLM_OK);
1850 ldlm_resource_unlink_lock(lock);
1851 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1853 ldlm_lock_destroy_nolock(lock);
1854 GOTO(out, rc = -ENOMEM);
1857 INIT_LIST_HEAD(&node->li_group);
1858 ldlm_interval_attach(node, lock);
1862 /* Some flags from the enqueue want to make it into the AST, via the
1863 * lock's l_flags. */
1864 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1865 ldlm_set_ast_discard_data(lock);
1866 if (*flags & LDLM_FL_TEST_LOCK)
1867 ldlm_set_test_lock(lock);
1868 if (*flags & LDLM_FL_COS_INCOMPAT)
1869 ldlm_set_cos_incompat(lock);
1870 if (*flags & LDLM_FL_COS_ENABLED)
1871 ldlm_set_cos_enabled(lock);
1873 /* This distinction between local lock trees is very important; a client
1874 * namespace only has information about locks taken by that client, and
1875 * thus doesn't have enough information to decide for itself if it can
1876 * be granted (below). In this case, we do exactly what the server
1877 * tells us to do, as dictated by the 'flags'.
1879 * We do exactly the same thing during recovery, when the server is
1880 * more or less trusting the clients not to lie.
1882 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1885 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1886 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1888 ldlm_grant_lock(lock, NULL);
1889 GOTO(out, rc = ELDLM_OK);
1890 #ifdef HAVE_SERVER_SUPPORT
1891 } else if (*flags & LDLM_FL_REPLAY) {
1892 if (*flags & LDLM_FL_BLOCK_WAIT) {
1893 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1894 GOTO(out, rc = ELDLM_OK);
1895 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1896 ldlm_grant_lock(lock, NULL);
1897 GOTO(out, rc = ELDLM_OK);
1899 /* If no flags, fall through to normal enqueue path. */
1902 rc = ldlm_lock_enqueue_helper(lock, flags);
1906 CERROR("This is client-side-only module, cannot handle "
1907 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1913 unlock_res_and_lock(lock);
1915 #ifdef HAVE_SERVER_SUPPORT
1917 struct ptlrpc_request *req = cookie;
1919 tgt_mk_reply_data(NULL, NULL,
1920 &req->rq_export->exp_target_data,
1921 req, 0, NULL, false, 0);
1925 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1929 #ifdef HAVE_SERVER_SUPPORT
1931 * Iterate through all waiting locks on a given resource queue and attempt to
1934 * Must be called with resource lock held.
1936 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1937 struct list_head *work_list,
1938 enum ldlm_process_intention intention,
1939 struct ldlm_lock *hint)
1941 struct list_head *tmp, *pos;
1942 ldlm_processing_policy policy;
1944 int rc = LDLM_ITER_CONTINUE;
1945 enum ldlm_error err;
1946 LIST_HEAD(bl_ast_list);
1950 check_res_locked(res);
1952 policy = ldlm_get_processing_policy(res);
1954 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1955 intention == LDLM_PROCESS_RECOVERY);
1958 list_for_each_safe(tmp, pos, queue) {
1959 struct ldlm_lock *pending;
1960 LIST_HEAD(rpc_list);
1962 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1964 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1967 rc = policy(pending, &flags, intention, &err, &rpc_list);
1968 if (pending->l_granted_mode == pending->l_req_mode ||
1969 res->lr_type == LDLM_FLOCK) {
1970 list_splice(&rpc_list, work_list);
1972 list_splice(&rpc_list, &bl_ast_list);
1975 * When this is called from recovery done, we always want
1976 * to scan the whole list no matter what 'rc' is returned.
1978 if (rc != LDLM_ITER_CONTINUE &&
1979 intention == LDLM_PROCESS_RESCAN)
1983 if (!list_empty(&bl_ast_list)) {
1986 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
1990 if (rc == -ERESTART)
1994 if (!list_empty(&bl_ast_list))
1995 ldlm_discard_bl_list(&bl_ast_list);
1997 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
2001 * Conflicting locks are detected for a lock to be enqueued, add the lock
2002 * into waiting list and send blocking ASTs to the conflicting locks.
2004 * \param[in] lock The lock to be enqueued.
2005 * \param[out] flags Lock flags for the lock to be enqueued.
2006 * \param[in] rpc_list Conflicting locks list.
2008 * \retval -ERESTART: Some lock was instantly canceled while sending
2009 * blocking ASTs, caller needs to re-check conflicting
2011 * \retval -EAGAIN: Lock was destroyed, caller should return error.
2012 * \reval 0: Lock is successfully added in waiting list.
2014 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
2015 struct list_head *rpc_list)
2017 struct ldlm_resource *res = lock->l_resource;
2021 check_res_locked(res);
2023 /* If either of the compat_queue()s returned failure, then we
2024 * have ASTs to send and must go onto the waiting list.
2026 * bug 2322: we used to unlink and re-add here, which was a
2027 * terrible folly -- if we goto restart, we could get
2028 * re-ordered! Causes deadlock, because ASTs aren't sent! */
2029 if (list_empty(&lock->l_res_link))
2030 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
2033 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
2036 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
2037 !ns_is_client(ldlm_res_to_ns(res)))
2038 class_fail_export(lock->l_export);
2040 if (rc == -ERESTART)
2041 ldlm_reprocess_all(res, NULL);
2044 if (rc == -ERESTART) {
2045 /* 15715: The lock was granted and destroyed after
2046 * resource lock was dropped. Interval node was freed
2047 * in ldlm_lock_destroy. Anyway, this always happens
2048 * when a client is being evicted. So it would be
2049 * ok to return an error. -jay */
2050 if (ldlm_is_destroyed(lock))
2053 /* lock was granted while resource was unlocked. */
2054 if (ldlm_is_granted(lock)) {
2055 /* bug 11300: if the lock has been granted,
2056 * break earlier because otherwise, we will go
2057 * to restart and ldlm_resource_unlink will be
2058 * called and it causes the interval node to be
2059 * freed. Then we will fail at
2060 * ldlm_extent_add_lock() */
2061 *flags &= ~LDLM_FL_BLOCKED_MASK;
2065 *flags |= LDLM_FL_BLOCK_GRANTED;
2071 * Discard all AST work items from list.
2073 * If for whatever reason we do not want to send ASTs to conflicting locks
2074 * anymore, disassemble the list with this function.
2076 void ldlm_discard_bl_list(struct list_head *bl_list)
2078 struct ldlm_lock *lock, *tmp;
2082 list_for_each_entry_safe(lock, tmp, bl_list, l_bl_ast) {
2083 LASSERT(!list_empty(&lock->l_bl_ast));
2084 list_del_init(&lock->l_bl_ast);
2085 ldlm_clear_ast_sent(lock);
2086 LASSERT(lock->l_bl_ast_run == 0);
2087 ldlm_clear_blocking_lock(lock);
2088 LDLM_LOCK_RELEASE(lock);
2094 * Process a call to blocking AST callback for a lock in ast_work list
2097 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2099 struct ldlm_cb_set_arg *arg = opaq;
2100 struct ldlm_lock *lock;
2101 struct ldlm_lock_desc d;
2102 struct ldlm_bl_desc bld;
2107 if (list_empty(arg->list))
2110 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2112 /* nobody should touch l_bl_ast but some locks in the list may become
2113 * granted after lock convert or COS downgrade, these locks should be
2114 * just skipped here and removed from the list.
2116 lock_res_and_lock(lock);
2117 list_del_init(&lock->l_bl_ast);
2119 /* lock is not blocking lock anymore, but was kept in the list because
2120 * it can managed only here.
2122 if (!ldlm_is_ast_sent(lock)) {
2123 unlock_res_and_lock(lock);
2124 LDLM_LOCK_RELEASE(lock);
2128 LASSERT(lock->l_blocking_lock);
2129 ldlm_lock2desc(lock->l_blocking_lock, &d);
2130 /* copy blocking lock ibits in cancel_bits as well,
2131 * new client may use them for lock convert and it is
2132 * important to use new field to convert locks from
2135 d.l_policy_data.l_inodebits.cancel_bits =
2136 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2138 /* Blocking lock is being destroyed here but some information about it
2139 * may be needed inside l_blocking_ast() function below,
2140 * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
2142 bld.bl_same_client = lock->l_client_cookie ==
2143 lock->l_blocking_lock->l_client_cookie;
2144 bld.bl_cos_incompat = ldlm_is_cos_incompat(lock->l_blocking_lock);
2145 arg->bl_desc = &bld;
2147 LASSERT(ldlm_is_ast_sent(lock));
2148 LASSERT(lock->l_bl_ast_run == 0);
2149 lock->l_bl_ast_run++;
2150 ldlm_clear_blocking_lock(lock);
2151 unlock_res_and_lock(lock);
2153 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2155 LDLM_LOCK_RELEASE(lock);
2161 * Process a call to revocation AST callback for a lock in ast_work list
2164 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2166 struct ldlm_cb_set_arg *arg = opaq;
2167 struct ldlm_lock_desc desc;
2169 struct ldlm_lock *lock;
2172 if (list_empty(arg->list))
2175 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2176 list_del_init(&lock->l_rk_ast);
2178 /* the desc just pretend to exclusive */
2179 ldlm_lock2desc(lock, &desc);
2180 desc.l_req_mode = LCK_EX;
2181 desc.l_granted_mode = 0;
2183 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2184 LDLM_LOCK_RELEASE(lock);
2190 * Process a call to glimpse AST callback for a lock in ast_work list
2192 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2194 struct ldlm_cb_set_arg *arg = opaq;
2195 struct ldlm_glimpse_work *gl_work;
2196 struct ldlm_lock *lock;
2200 if (list_empty(arg->list))
2203 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2205 list_del_init(&gl_work->gl_list);
2207 lock = gl_work->gl_lock;
2209 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2210 arg->gl_desc = gl_work->gl_desc;
2211 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2212 arg->gl_interpret_data = gl_work->gl_interpret_data;
2214 /* invoke the actual glimpse callback */
2215 rc = lock->l_glimpse_ast(lock, (void *)arg);
2217 rc = 1; /* update LVB if this is server lock */
2218 else if (rc == -ELDLM_NO_LOCK_DATA)
2219 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
2221 LDLM_LOCK_RELEASE(lock);
2222 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2223 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2225 OBD_FREE_PTR(gl_work);
2232 * Process a call to completion AST callback for a lock in ast_work list
2235 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2237 struct ldlm_cb_set_arg *arg = opaq;
2238 struct ldlm_lock *lock;
2239 ldlm_completion_callback completion_callback;
2244 if (list_empty(arg->list))
2247 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2249 /* It's possible to receive a completion AST before we've set
2250 * the l_completion_ast pointer: either because the AST arrived
2251 * before the reply, or simply because there's a small race
2252 * window between receiving the reply and finishing the local
2253 * enqueue. (bug 842)
2255 * This can't happen with the blocking_ast, however, because we
2256 * will never call the local blocking_ast until we drop our
2257 * reader/writer reference, which we won't do until we get the
2258 * reply and finish enqueueing. */
2260 /* nobody should touch l_cp_ast */
2261 lock_res_and_lock(lock);
2262 list_del_init(&lock->l_cp_ast);
2263 LASSERT(ldlm_is_cp_reqd(lock));
2264 /* save l_completion_ast since it can be changed by
2265 * mds_intent_policy(), see bug 14225 */
2266 completion_callback = lock->l_completion_ast;
2267 ldlm_clear_cp_reqd(lock);
2268 unlock_res_and_lock(lock);
2270 if (completion_callback != NULL)
2271 rc = completion_callback(lock, 0, (void *)arg);
2272 LDLM_LOCK_RELEASE(lock);
2278 * Process list of locks in need of ASTs being sent.
2280 * Used on server to send multiple ASTs together instead of sending one by
2283 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2284 ldlm_desc_ast_t ast_type)
2286 struct ldlm_cb_set_arg *arg;
2287 set_producer_func work_ast_lock;
2290 if (list_empty(rpc_list))
2297 atomic_set(&arg->restart, 0);
2298 arg->list = rpc_list;
2301 case LDLM_WORK_CP_AST:
2302 arg->type = LDLM_CP_CALLBACK;
2303 work_ast_lock = ldlm_work_cp_ast_lock;
2305 #ifdef HAVE_SERVER_SUPPORT
2306 case LDLM_WORK_BL_AST:
2307 arg->type = LDLM_BL_CALLBACK;
2308 work_ast_lock = ldlm_work_bl_ast_lock;
2310 case LDLM_WORK_REVOKE_AST:
2311 arg->type = LDLM_BL_CALLBACK;
2312 work_ast_lock = ldlm_work_revoke_ast_lock;
2314 case LDLM_WORK_GL_AST:
2315 arg->type = LDLM_GL_CALLBACK;
2316 work_ast_lock = ldlm_work_gl_ast_lock;
2323 /* We create a ptlrpc request set with flow control extension.
2324 * This request set will use the work_ast_lock function to produce new
2325 * requests and will send a new request each time one completes in order
2326 * to keep the number of requests in flight to ns_max_parallel_ast */
2327 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2328 work_ast_lock, arg);
2329 if (arg->set == NULL)
2330 GOTO(out, rc = -ENOMEM);
2332 ptlrpc_set_wait(NULL, arg->set);
2333 ptlrpc_set_destroy(arg->set);
2335 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2343 * Try to grant all waiting locks on a resource.
2345 * Calls ldlm_reprocess_queue on waiting queue.
2347 * Typically called after some resource locks are cancelled to see
2348 * if anything could be granted as a result of the cancellation.
2350 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2351 enum ldlm_process_intention intention,
2352 struct ldlm_lock *hint)
2354 LIST_HEAD(rpc_list);
2355 #ifdef HAVE_SERVER_SUPPORT
2356 ldlm_reprocessing_policy reprocess;
2357 struct obd_device *obd;
2362 /* Local lock trees don't get reprocessed. */
2363 if (ns_is_client(ldlm_res_to_ns(res))) {
2368 /* Disable reprocess during lock replay stage but allow during
2369 * request replay stage.
2371 obd = ldlm_res_to_ns(res)->ns_obd;
2372 if (obd->obd_recovering &&
2373 atomic_read(&obd->obd_req_replay_clients) == 0)
2377 reprocess = ldlm_get_reprocessing_policy(res);
2378 reprocess(res, &res->lr_waiting, &rpc_list, intention, hint);
2381 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2383 if (rc == -ERESTART) {
2384 LASSERT(list_empty(&rpc_list));
2390 if (!ns_is_client(ldlm_res_to_ns(res))) {
2391 CERROR("This is client-side-only module, cannot handle "
2392 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2399 void ldlm_reprocess_all(struct ldlm_resource *res, struct ldlm_lock *hint)
2401 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN, hint);
2403 EXPORT_SYMBOL(ldlm_reprocess_all);
2405 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2406 struct hlist_node *hnode, void *arg)
2408 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2410 /* This is only called once after recovery done. LU-8306. */
2411 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY, NULL);
2416 * Iterate through all resources on a namespace attempting to grant waiting
2419 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2424 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2425 ldlm_reprocess_res, NULL, 0);
2431 * Helper function to call blocking AST for LDLM lock \a lock in a
2432 * "cancelling" mode.
2434 void ldlm_cancel_callback(struct ldlm_lock *lock)
2436 check_res_locked(lock->l_resource);
2437 if (!ldlm_is_cancel(lock)) {
2438 ldlm_set_cancel(lock);
2439 if (lock->l_blocking_ast) {
2440 unlock_res_and_lock(lock);
2441 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2443 lock_res_and_lock(lock);
2445 LDLM_DEBUG(lock, "no blocking ast");
2448 /* only canceller can set bl_done bit */
2449 ldlm_set_bl_done(lock);
2450 wake_up(&lock->l_waitq);
2451 } else if (!ldlm_is_bl_done(lock)) {
2452 /* The lock is guaranteed to have been canceled once
2453 * returning from this function. */
2454 unlock_res_and_lock(lock);
2455 wait_event_idle(lock->l_waitq, is_bl_done(lock));
2456 lock_res_and_lock(lock);
2461 * Remove skiplist-enabled LDLM lock \a req from granted list
2463 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2465 if (req->l_resource->lr_type != LDLM_PLAIN &&
2466 req->l_resource->lr_type != LDLM_IBITS)
2469 list_del_init(&req->l_sl_policy);
2470 list_del_init(&req->l_sl_mode);
2474 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2476 void ldlm_lock_cancel(struct ldlm_lock *lock)
2478 struct ldlm_resource *res;
2479 struct ldlm_namespace *ns;
2482 lock_res_and_lock(lock);
2484 res = lock->l_resource;
2485 ns = ldlm_res_to_ns(res);
2487 /* Please do not, no matter how tempting, remove this LBUG without
2488 * talking to me first. -phik */
2489 if (lock->l_readers || lock->l_writers) {
2490 LDLM_ERROR(lock, "lock still has references");
2491 unlock_res_and_lock(lock);
2495 if (ldlm_is_waited(lock))
2496 ldlm_del_waiting_lock(lock);
2498 /* Releases cancel callback. */
2499 ldlm_cancel_callback(lock);
2501 /* Yes, second time, just in case it was added again while we were
2502 * running with no res lock in ldlm_cancel_callback */
2503 if (ldlm_is_waited(lock))
2504 ldlm_del_waiting_lock(lock);
2506 ldlm_resource_unlink_lock(lock);
2507 ldlm_lock_destroy_nolock(lock);
2509 if (ldlm_is_granted(lock))
2510 ldlm_pool_del(&ns->ns_pool, lock);
2512 /* Make sure we will not be called again for same lock what is possible
2513 * if not to zero out lock->l_granted_mode */
2514 lock->l_granted_mode = LCK_MINMODE;
2515 unlock_res_and_lock(lock);
2519 EXPORT_SYMBOL(ldlm_lock_cancel);
2522 * Set opaque data into the lock that only makes sense to upper layer.
2524 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2526 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2531 if (lock->l_ast_data == NULL)
2532 lock->l_ast_data = data;
2533 if (lock->l_ast_data == data)
2535 LDLM_LOCK_PUT(lock);
2539 EXPORT_SYMBOL(ldlm_lock_set_data);
2541 struct export_cl_data {
2542 const struct lu_env *ecl_env;
2543 struct obd_export *ecl_exp;
2547 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2548 struct ldlm_lock *lock,
2549 struct export_cl_data *ecl)
2551 struct ldlm_resource *res;
2553 res = ldlm_resource_getref(lock->l_resource);
2555 ldlm_lvbo_update(res, lock, NULL, 1);
2556 ldlm_lock_cancel(lock);
2557 if (!exp->exp_obd->obd_stopping)
2558 ldlm_reprocess_all(res, lock);
2559 ldlm_resource_putref(res);
2562 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2563 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2564 exp, ecl->ecl_loop);
2569 * Iterator function for ldlm_export_cancel_locks.
2570 * Cancels passed locks.
2573 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2574 struct hlist_node *hnode, void *data)
2577 struct export_cl_data *ecl = (struct export_cl_data *)data;
2578 struct obd_export *exp = ecl->ecl_exp;
2579 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2581 LDLM_LOCK_GET(lock);
2582 ldlm_cancel_lock_for_export(exp, lock, ecl);
2583 LDLM_LOCK_RELEASE(lock);
2589 * Cancel all blocked locks for given export.
2591 * Typically called on client disconnection/eviction
2593 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2596 struct export_cl_data ecl = {
2602 rc = lu_env_init(&env, LCT_DT_THREAD);
2607 while (!list_empty(&exp->exp_bl_list)) {
2608 struct ldlm_lock *lock;
2610 spin_lock_bh(&exp->exp_bl_list_lock);
2611 if (!list_empty(&exp->exp_bl_list)) {
2612 lock = list_entry(exp->exp_bl_list.next,
2613 struct ldlm_lock, l_exp_list);
2614 LDLM_LOCK_GET(lock);
2615 list_del_init(&lock->l_exp_list);
2619 spin_unlock_bh(&exp->exp_bl_list_lock);
2624 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2625 LDLM_LOCK_RELEASE(lock);
2630 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2631 "left on hash table %d.\n", exp, ecl.ecl_loop,
2632 atomic_read(&exp->exp_lock_hash->hs_count));
2634 return ecl.ecl_loop;
2638 * Cancel all locks for given export.
2640 * Typically called after client disconnection/eviction
2642 int ldlm_export_cancel_locks(struct obd_export *exp)
2644 struct export_cl_data ecl;
2648 rc = lu_env_init(&env, LCT_DT_THREAD);
2655 cfs_hash_for_each_empty(exp->exp_lock_hash,
2656 ldlm_cancel_locks_for_export_cb, &ecl);
2658 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2659 "left on hash table %d.\n", exp, ecl.ecl_loop,
2660 atomic_read(&exp->exp_lock_hash->hs_count));
2662 if (ecl.ecl_loop > 0 &&
2663 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2664 exp->exp_obd->obd_stopping)
2665 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2669 return ecl.ecl_loop;
2673 * Downgrade an PW/EX lock to COS | CR mode.
2675 * A lock mode convertion from PW/EX mode to less conflict mode. The
2676 * convertion may fail if lock was canceled before downgrade, but it doesn't
2677 * indicate any problem, because such lock has no reader or writer, and will
2680 * Used by Commit on Sharing (COS) code to force object changes commit in case
2681 * of conflict. Converted lock is considered as new lock and all blocking AST
2682 * things are cleared, so any pending or new blocked lock on that lock will
2683 * cause new call to blocking_ast and force resource object commit.
2685 * Also used by layout_change to replace EX lock to CR lock.
2687 * \param lock A lock to convert
2688 * \param new_mode new lock mode
2690 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2692 #ifdef HAVE_SERVER_SUPPORT
2695 LASSERT(new_mode == LCK_COS || new_mode == LCK_CR);
2697 lock_res_and_lock(lock);
2699 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2700 unlock_res_and_lock(lock);
2702 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2703 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2707 ldlm_resource_unlink_lock(lock);
2709 * Remove the lock from pool as it will be added again in
2710 * ldlm_grant_lock() called below.
2712 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2714 /* Consider downgraded lock as a new lock and clear all states
2715 * related to a previous blocking AST processing.
2717 ldlm_clear_blocking_data(lock);
2719 lock->l_req_mode = new_mode;
2720 ldlm_grant_lock(lock, NULL);
2721 unlock_res_and_lock(lock);
2723 ldlm_reprocess_all(lock->l_resource, lock);
2728 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2731 * Print lock with lock handle \a lockh description into debug log.
2733 * Used when printing all locks on a resource for debug purposes.
2735 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2737 struct ldlm_lock *lock;
2739 if (!((libcfs_debug | D_ERROR) & level))
2742 lock = ldlm_handle2lock(lockh);
2746 LDLM_DEBUG_LIMIT(level, lock, "###");
2748 LDLM_LOCK_PUT(lock);
2750 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2753 * Print lock information with custom message into debug log.
2756 void _ldlm_lock_debug(struct ldlm_lock *lock,
2757 struct libcfs_debug_msg_data *msgdata,
2758 const char *fmt, ...)
2761 struct obd_export *exp = lock->l_export;
2762 struct ldlm_resource *resource = NULL;
2763 struct va_format vaf;
2764 char *nid = "local";
2767 resource = rcu_dereference(lock->l_resource);
2768 if (resource && !atomic_inc_not_zero(&resource->lr_refcount))
2772 va_start(args, fmt);
2776 if (exp && exp->exp_connection) {
2777 nid = obd_export_nid2str(exp);
2778 } else if (exp && exp->exp_obd != NULL) {
2779 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2780 nid = obd_import_nid2str(imp);
2783 if (resource == NULL) {
2784 libcfs_debug_msg(msgdata,
2785 "%pV ns: \?\? lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: \?\? rrc=\?\? type: \?\?\? flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2788 lock->l_handle.h_cookie,
2789 refcount_read(&lock->l_handle.h_ref),
2790 lock->l_readers, lock->l_writers,
2791 ldlm_lockname[lock->l_granted_mode],
2792 ldlm_lockname[lock->l_req_mode],
2794 lock->l_remote_handle.cookie,
2795 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2796 lock->l_pid, lock->l_callback_timestamp,
2802 switch (resource->lr_type) {
2804 libcfs_debug_msg(msgdata,
2805 "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s [%llu->%llu] (req %llu->%llu) gid %llu flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2807 ldlm_lock_to_ns_name(lock), lock,
2808 lock->l_handle.h_cookie,
2809 refcount_read(&lock->l_handle.h_ref),
2810 lock->l_readers, lock->l_writers,
2811 ldlm_lockname[lock->l_granted_mode],
2812 ldlm_lockname[lock->l_req_mode],
2814 atomic_read(&resource->lr_refcount),
2815 ldlm_typename[resource->lr_type],
2816 lock->l_policy_data.l_extent.start,
2817 lock->l_policy_data.l_extent.end,
2818 lock->l_req_extent.start, lock->l_req_extent.end,
2819 lock->l_req_extent.gid,
2821 lock->l_remote_handle.cookie,
2822 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2823 lock->l_pid, lock->l_callback_timestamp,
2828 libcfs_debug_msg(msgdata,
2829 "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s pid: %d [%llu->%llu] flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld\n",
2831 ldlm_lock_to_ns_name(lock), lock,
2832 lock->l_handle.h_cookie,
2833 refcount_read(&lock->l_handle.h_ref),
2834 lock->l_readers, lock->l_writers,
2835 ldlm_lockname[lock->l_granted_mode],
2836 ldlm_lockname[lock->l_req_mode],
2838 atomic_read(&resource->lr_refcount),
2839 ldlm_typename[resource->lr_type],
2840 lock->l_policy_data.l_flock.pid,
2841 lock->l_policy_data.l_flock.start,
2842 lock->l_policy_data.l_flock.end,
2844 lock->l_remote_handle.cookie,
2845 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2846 lock->l_pid, lock->l_callback_timestamp);
2850 libcfs_debug_msg(msgdata,
2851 "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " bits %#llx/%#llx rrc: %d type: %s gid %llu flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2853 ldlm_lock_to_ns_name(lock),
2854 lock, lock->l_handle.h_cookie,
2855 refcount_read(&lock->l_handle.h_ref),
2856 lock->l_readers, lock->l_writers,
2857 ldlm_lockname[lock->l_granted_mode],
2858 ldlm_lockname[lock->l_req_mode],
2860 lock->l_policy_data.l_inodebits.bits,
2861 lock->l_policy_data.l_inodebits.try_bits,
2862 atomic_read(&resource->lr_refcount),
2863 ldlm_typename[resource->lr_type],
2864 lock->l_policy_data.l_inodebits.li_gid,
2866 lock->l_remote_handle.cookie,
2867 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2868 lock->l_pid, lock->l_callback_timestamp,
2873 libcfs_debug_msg(msgdata,
2874 "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2876 ldlm_lock_to_ns_name(lock),
2877 lock, lock->l_handle.h_cookie,
2878 refcount_read(&lock->l_handle.h_ref),
2879 lock->l_readers, lock->l_writers,
2880 ldlm_lockname[lock->l_granted_mode],
2881 ldlm_lockname[lock->l_req_mode],
2883 atomic_read(&resource->lr_refcount),
2884 ldlm_typename[resource->lr_type],
2886 lock->l_remote_handle.cookie,
2887 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2888 lock->l_pid, lock->l_callback_timestamp,
2893 ldlm_resource_putref(resource);
2895 EXPORT_SYMBOL(_ldlm_lock_debug);