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",
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 #ifdef HAVE_SERVER_SUPPORT
140 static ldlm_processing_policy ldlm_processing_policy_table[] = {
141 [LDLM_PLAIN] = ldlm_process_plain_lock,
142 [LDLM_EXTENT] = ldlm_process_extent_lock,
143 [LDLM_FLOCK] = ldlm_process_flock_lock,
144 [LDLM_IBITS] = ldlm_process_inodebits_lock,
147 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
149 return ldlm_processing_policy_table[res->lr_type];
151 EXPORT_SYMBOL(ldlm_get_processing_policy);
153 static ldlm_reprocessing_policy ldlm_reprocessing_policy_table[] = {
154 [LDLM_PLAIN] = ldlm_reprocess_queue,
155 [LDLM_EXTENT] = ldlm_reprocess_queue,
156 [LDLM_FLOCK] = ldlm_reprocess_queue,
157 [LDLM_IBITS] = ldlm_reprocess_inodebits_queue,
160 ldlm_reprocessing_policy ldlm_get_reprocessing_policy(struct ldlm_resource *res)
162 return ldlm_reprocessing_policy_table[res->lr_type];
165 #endif /* HAVE_SERVER_SUPPORT */
167 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
171 EXPORT_SYMBOL(ldlm_register_intent);
173 /* REFCOUNTED LOCK OBJECTS */
177 * Get a reference on a lock.
179 * Lock refcounts, during creation:
180 * - one special one for allocation, dec'd only once in destroy
181 * - one for being a lock that's in-use
182 * - one for the addref associated with a new lock
184 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
186 refcount_inc(&lock->l_handle.h_ref);
189 EXPORT_SYMBOL(ldlm_lock_get);
191 static void lock_handle_free(struct rcu_head *rcu)
193 struct ldlm_lock *lock = container_of(rcu, struct ldlm_lock,
196 OBD_FREE_PRE(lock, sizeof(*lock), "slab-freed");
197 kmem_cache_free(ldlm_lock_slab, lock);
201 * Release lock reference.
203 * Also frees the lock if it was last reference.
205 void ldlm_lock_put(struct ldlm_lock *lock)
209 LASSERT(lock->l_resource != LP_POISON);
210 LASSERT(refcount_read(&lock->l_handle.h_ref) > 0);
211 if (refcount_dec_and_test(&lock->l_handle.h_ref)) {
212 struct ldlm_resource *res;
215 "final lock_put on destroyed lock, freeing it.");
217 res = lock->l_resource;
218 LASSERT(ldlm_is_destroyed(lock));
219 LASSERT(list_empty(&lock->l_exp_list));
220 LASSERT(list_empty(&lock->l_res_link));
221 LASSERT(list_empty(&lock->l_pending_chain));
223 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
225 lu_ref_del(&res->lr_reference, "lock", lock);
226 if (lock->l_export) {
227 class_export_lock_put(lock->l_export, lock);
228 lock->l_export = NULL;
231 if (lock->l_lvb_data != NULL)
232 OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
234 if (res->lr_type == LDLM_IBITS && lock->l_ibits_node) {
235 OBD_SLAB_FREE_PTR(lock->l_ibits_node,
236 ldlm_inodebits_slab);
238 ldlm_resource_putref(res);
239 lock->l_resource = NULL;
240 lu_ref_fini(&lock->l_reference);
241 if (lock->l_flags & BIT(63))
242 /* Performance testing - bypassing RCU removes overhead */
243 lock_handle_free(&lock->l_handle.h_rcu);
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)
259 if (!list_empty(&lock->l_lru)) {
260 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
262 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
263 if (ns->ns_last_pos == &lock->l_lru)
264 ns->ns_last_pos = lock->l_lru.prev;
265 list_del_init(&lock->l_lru);
266 LASSERT(ns->ns_nr_unused > 0);
274 * Removes LDLM lock \a lock from LRU. Obtains the LRU lock first.
276 * If \a last_use is non-zero, it will remove the lock from LRU only if
277 * it matches lock's l_last_used.
279 * \retval 0 if \a last_use is set, the lock is not in LRU list or \a last_use
280 * doesn't match lock's l_last_used;
281 * otherwise, the lock hasn't been in the LRU list.
282 * \retval 1 the lock was in LRU list and removed.
284 int ldlm_lock_remove_from_lru_check(struct ldlm_lock *lock, ktime_t last_use)
286 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
290 if (ldlm_is_ns_srv(lock)) {
291 LASSERT(list_empty(&lock->l_lru));
295 spin_lock(&ns->ns_lock);
296 if (!ktime_compare(last_use, ktime_set(0, 0)) ||
297 !ktime_compare(last_use, lock->l_last_used))
298 rc = ldlm_lock_remove_from_lru_nolock(lock);
299 spin_unlock(&ns->ns_lock);
304 /* Adds LDLM lock \a lock to namespace LRU. Assumes LRU is already locked. */
305 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
307 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
309 lock->l_last_used = ktime_get();
310 LASSERT(list_empty(&lock->l_lru));
311 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
312 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
313 LASSERT(ns->ns_nr_unused >= 0);
317 /* Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks first */
318 static void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
320 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
323 spin_lock(&ns->ns_lock);
324 ldlm_lock_add_to_lru_nolock(lock);
325 spin_unlock(&ns->ns_lock);
330 * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
331 * the LRU. Performs necessary LRU locking
333 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
335 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
338 if (ldlm_is_ns_srv(lock)) {
339 LASSERT(list_empty(&lock->l_lru));
344 spin_lock(&ns->ns_lock);
345 if (!list_empty(&lock->l_lru)) {
346 ldlm_lock_remove_from_lru_nolock(lock);
347 ldlm_lock_add_to_lru_nolock(lock);
349 spin_unlock(&ns->ns_lock);
354 * Helper to destroy a locked lock.
356 * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
357 * Must be called with l_lock and lr_lock held.
359 * Does not actually free the lock data, but rather marks the lock as
360 * destroyed by setting l_destroyed field in the lock to 1. Destroys a
361 * handle->lock association too, so that the lock can no longer be found
362 * and removes the lock from LRU list. Actual lock freeing occurs when
363 * last lock reference goes away.
365 * Original comment (of some historical value):
366 * This used to have a 'strict' flag, which recovery would use to mark an
367 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
368 * shall explain why it's gone: with the new hash table scheme, once you call
369 * ldlm_lock_destroy, you can never drop your final references on this lock.
370 * Because it's not in the hash table anymore. -phil
372 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
376 if (lock->l_readers || lock->l_writers) {
377 LDLM_ERROR(lock, "lock still has references");
381 if (!list_empty(&lock->l_res_link)) {
382 LDLM_ERROR(lock, "lock still on resource");
386 if (ldlm_is_destroyed(lock)) {
387 LASSERT(list_empty(&lock->l_lru));
391 ldlm_set_destroyed(lock);
392 wake_up(&lock->l_waitq);
394 if (lock->l_export && lock->l_export->exp_lock_hash) {
395 /* Safe to call cfs_hash_del as lock isn't in exp_lock_hash. */
396 /* below, .hs_keycmp resolves to ldlm_export_lock_keycmp() */
397 cfs_hash_del(lock->l_export->exp_lock_hash,
398 &lock->l_remote_handle, &lock->l_exp_hash);
401 ldlm_lock_remove_from_lru(lock);
402 class_handle_unhash(&lock->l_handle);
408 /* Destroys a LDLM lock \a lock. Performs necessary locking first. */
409 void ldlm_lock_destroy(struct ldlm_lock *lock)
414 lock_res_and_lock(lock);
415 first = ldlm_lock_destroy_internal(lock);
416 unlock_res_and_lock(lock);
418 /* drop reference from hashtable only for first destroy */
420 lu_ref_del(&lock->l_reference, "hash", lock);
421 LDLM_LOCK_RELEASE(lock);
426 /* Destroys a LDLM lock \a lock that is already locked. */
427 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
432 first = ldlm_lock_destroy_internal(lock);
433 /* drop reference from hashtable only for first destroy */
435 lu_ref_del(&lock->l_reference, "hash", lock);
436 LDLM_LOCK_RELEASE(lock);
441 static const char lock_handle_owner[] = "ldlm";
445 * Allocate and initialize new lock structure.
447 * usage: pass in a resource on which you have done ldlm_resource_get
448 * new lock will take over the refcount.
449 * returns: lock with refcount 2 - one for current caller and one for remote
451 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
453 struct ldlm_lock *lock;
457 if (resource == NULL)
460 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, GFP_NOFS);
464 RCU_INIT_POINTER(lock->l_resource, resource);
465 lu_ref_add(&resource->lr_reference, "lock", lock);
467 refcount_set(&lock->l_handle.h_ref, 2);
468 INIT_LIST_HEAD(&lock->l_res_link);
469 INIT_LIST_HEAD(&lock->l_lru);
470 INIT_LIST_HEAD(&lock->l_pending_chain);
471 INIT_LIST_HEAD(&lock->l_bl_ast);
472 INIT_LIST_HEAD(&lock->l_cp_ast);
473 INIT_LIST_HEAD(&lock->l_rk_ast);
474 init_waitqueue_head(&lock->l_waitq);
475 lock->l_blocking_lock = NULL;
476 switch (resource->lr_type) {
479 INIT_LIST_HEAD(&lock->l_sl_mode);
480 INIT_LIST_HEAD(&lock->l_sl_policy);
483 INIT_HLIST_NODE(&lock->l_exp_flock_hash);
489 INIT_HLIST_NODE(&lock->l_exp_hash);
491 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
493 INIT_HLIST_NODE(&lock->l_handle.h_link);
494 class_handle_hash(&lock->l_handle, lock_handle_owner);
496 lu_ref_init(&lock->l_reference);
497 lu_ref_add(&lock->l_reference, "hash", lock);
498 lock->l_callback_timestamp = 0;
499 lock->l_activity = 0;
501 #if LUSTRE_TRACKS_LOCK_EXP_REFS
502 INIT_LIST_HEAD(&lock->l_exp_refs_link);
503 lock->l_exp_refs_nr = 0;
504 lock->l_exp_refs_target = NULL;
506 INIT_LIST_HEAD(&lock->l_exp_list);
511 struct ldlm_lock *ldlm_lock_new_testing(struct ldlm_resource *resource)
513 struct ldlm_lock *lock = ldlm_lock_new(resource);
518 lock->l_flags |= BIT(63);
519 switch (resource->lr_type) {
521 rc = ldlm_inodebits_alloc_lock(lock);
529 ldlm_lock_destroy(lock);
530 LDLM_LOCK_RELEASE(lock);
533 EXPORT_SYMBOL(ldlm_lock_new_testing);
536 * Moves LDLM lock \a lock to another resource.
537 * This is used on client when server returns some other lock than requested
538 * (typically as a result of intent operation)
540 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
541 const struct ldlm_res_id *new_resid)
543 struct ldlm_resource *oldres;
544 struct ldlm_resource *newres;
549 LASSERT(ns_is_client(ns));
551 oldres = lock_res_and_lock(lock);
552 if (memcmp(new_resid, &oldres->lr_name,
553 sizeof(oldres->lr_name)) == 0) {
555 unlock_res_and_lock(lock);
559 LASSERT(new_resid->name[0] != 0);
561 /* This function assumes that the lock isn't on any lists */
562 LASSERT(list_empty(&lock->l_res_link));
564 type = oldres->lr_type;
565 unlock_res_and_lock(lock);
567 newres = ldlm_resource_get(ns, new_resid, type, 1);
569 RETURN(PTR_ERR(newres));
571 lu_ref_add(&newres->lr_reference, "lock", lock);
573 * To flip the lock from the old to the new resource, oldres
574 * and newres have to be locked. Resource spin-locks are taken
575 * in the memory address order to avoid dead-locks.
576 * As this is the only circumstance where ->l_resource
577 * can change, and this cannot race with itself, it is safe
578 * to access lock->l_resource without being careful about locking.
580 oldres = lock->l_resource;
581 if (oldres < newres) {
583 lock_res_nested(newres, LRT_NEW);
586 lock_res_nested(oldres, LRT_NEW);
588 LASSERT(memcmp(new_resid, &oldres->lr_name,
589 sizeof(oldres->lr_name)) != 0);
590 rcu_assign_pointer(lock->l_resource, newres);
594 /* ...and the flowers are still standing! */
595 lu_ref_del(&oldres->lr_reference, "lock", lock);
596 ldlm_resource_putref(oldres);
601 /** \defgroup ldlm_handles LDLM HANDLES
602 * Ways to get hold of locks without any addresses.
607 * Fills in handle for LDLM lock \a lock into supplied \a lockh
608 * Does not take any references.
610 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
612 lockh->cookie = lock->l_handle.h_cookie;
614 EXPORT_SYMBOL(ldlm_lock2handle);
617 * Obtain a lock reference by handle.
619 * if \a flags: atomically get the lock and set the flags.
620 * Return NULL if flag already set
622 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
625 struct ldlm_lock *lock;
631 if (!lustre_handle_is_used(handle))
634 lock = class_handle2object(handle->cookie, lock_handle_owner);
638 if (lock->l_export != NULL && lock->l_export->exp_failed) {
639 CDEBUG(D_INFO, "lock export failed: lock %p, exp %p\n",
640 lock, lock->l_export);
645 /* It's unlikely but possible that someone marked the lock as
646 * destroyed after we did handle2object on it
648 if ((flags == 0) && !ldlm_is_destroyed(lock)) {
649 lu_ref_add_atomic(&lock->l_reference, "handle", lock);
653 lock_res_and_lock(lock);
655 LASSERT(lock->l_resource != NULL);
657 lu_ref_add_atomic(&lock->l_reference, "handle", lock);
658 if (unlikely(ldlm_is_destroyed(lock))) {
659 unlock_res_and_lock(lock);
660 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
665 /* If we're setting flags, make sure none of them are already set. */
667 if ((lock->l_flags & flags) != 0) {
668 unlock_res_and_lock(lock);
673 lock->l_flags |= flags;
676 unlock_res_and_lock(lock);
679 EXPORT_SYMBOL(__ldlm_handle2lock);
680 /** @} ldlm_handles */
683 * Fill in "on the wire" representation for given LDLM lock into supplied
684 * lock descriptor \a desc structure.
686 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
688 ldlm_res2desc(lock->l_resource, &desc->l_resource);
689 desc->l_req_mode = lock->l_req_mode;
690 desc->l_granted_mode = lock->l_granted_mode;
691 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
692 &lock->l_policy_data,
693 &desc->l_policy_data);
697 * Add a lock to list of conflicting locks to send AST to.
699 * Only add if we have not sent a blocking AST to the lock yet.
701 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
702 struct list_head *work_list)
704 if (!ldlm_is_ast_sent(lock)) {
705 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
706 ldlm_set_ast_sent(lock);
707 /* If the enqueuing client said so, tell the AST recipient to
708 * discard dirty data, rather than writing back.
710 if (ldlm_is_ast_discard_data(new))
711 ldlm_set_discard_data(lock);
713 /* Lock can be converted from a blocking state back to granted
714 * after lock convert or COS downgrade but still be in an
715 * older bl_list because it is controlled only by
716 * ldlm_work_bl_ast_lock(), let it be processed there.
718 if (list_empty(&lock->l_bl_ast)) {
719 list_add(&lock->l_bl_ast, work_list);
722 LASSERT(lock->l_blocking_lock == NULL);
723 lock->l_blocking_lock = LDLM_LOCK_GET(new);
727 /* Add a lock to list of just granted locks to send completion AST to. */
728 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
729 struct list_head *work_list)
731 if (!ldlm_is_cp_reqd(lock)) {
732 ldlm_set_cp_reqd(lock);
733 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
734 LASSERT(list_empty(&lock->l_cp_ast));
735 list_add(&lock->l_cp_ast, work_list);
741 * Aggregator function to add AST work items into a list. Determines
742 * what sort of an AST work needs to be done and calls the proper
744 * Must be called with lr_lock held.
746 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
747 struct list_head *work_list)
750 check_res_locked(lock->l_resource);
752 ldlm_add_bl_work_item(lock, new, work_list);
754 ldlm_add_cp_work_item(lock, work_list);
759 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
760 * r/w reference type is determined by \a mode
761 * Calls ldlm_lock_addref_internal.
763 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
765 struct ldlm_lock *lock;
767 lock = ldlm_handle2lock(lockh);
768 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
769 ldlm_lock_addref_internal(lock, mode);
772 EXPORT_SYMBOL(ldlm_lock_addref);
776 * Add specified reader/writer reference to LDLM lock \a lock.
777 * r/w reference type is determined by \a mode
778 * Removes lock from LRU if it is there.
779 * Assumes the LDLM lock is already locked.
781 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
784 ldlm_lock_remove_from_lru(lock);
785 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
787 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
789 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS | LCK_TXN)) {
791 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
794 lu_ref_add_atomic(&lock->l_reference, "user", lock);
795 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
799 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
800 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
802 * \retval 0 success, lock was addref-ed
804 * \retval -EAGAIN lock is being canceled.
806 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
808 struct ldlm_lock *lock;
812 lock = ldlm_handle2lock(lockh);
814 lock_res_and_lock(lock);
815 if (lock->l_readers != 0 || lock->l_writers != 0 ||
816 !ldlm_is_cbpending(lock)) {
817 ldlm_lock_addref_internal_nolock(lock, mode);
820 unlock_res_and_lock(lock);
825 EXPORT_SYMBOL(ldlm_lock_addref_try);
828 * Add specified reader/writer reference to LDLM lock \a lock.
829 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
830 * Only called for local locks.
832 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
834 lock_res_and_lock(lock);
835 ldlm_lock_addref_internal_nolock(lock, mode);
836 unlock_res_and_lock(lock);
840 * Removes reader/writer reference for LDLM lock \a lock.
841 * Assumes LDLM lock is already locked.
842 * only called in ldlm_flock_destroy and for local locks.
843 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
844 * that cannot be placed in LRU.
846 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
849 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
850 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
851 LASSERT(lock->l_readers > 0);
852 lu_ref_del(&lock->l_reference, "reader", lock);
855 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS | LCK_TXN)) {
856 LASSERT(lock->l_writers > 0);
857 lu_ref_del(&lock->l_reference, "writer", lock);
861 lu_ref_del(&lock->l_reference, "user", lock);
862 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
866 * Removes reader/writer reference for LDLM lock \a lock.
867 * Locks LDLM lock first.
868 * If the lock is determined to be client lock on a client and r/w refcount
869 * drops to zero and the lock is not blocked, the lock is added to LRU lock
871 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
873 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
875 struct ldlm_namespace *ns;
879 lock_res_and_lock(lock);
881 ns = ldlm_lock_to_ns(lock);
883 ldlm_lock_decref_internal_nolock(lock, mode);
885 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
886 !lock->l_readers && !lock->l_writers) {
887 /* If this is a local lock on a server namespace and this was
888 * the last reference, cancel the lock.
890 * Group locks are special:
891 * They must not go in LRU, but they are not called back
892 * like non-group locks, instead they are manually released.
893 * They have an l_writers reference which they keep until
894 * they are manually released, so we remove them when they have
895 * no more reader or writer references. - LU-6368
897 ldlm_set_cbpending(lock);
900 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
901 unsigned int mask = D_DLMTRACE;
903 /* If we received a blocked AST and this was the last reference,
906 if (ldlm_is_ns_srv(lock) && lock->l_export)
908 LDLM_DEBUG_LIMIT(mask, lock,
909 "final decref done on %sCBPENDING lock",
910 mask & D_WARNING ? "non-local " : "");
912 LDLM_LOCK_GET(lock); /* dropped by bl thread */
913 ldlm_lock_remove_from_lru(lock);
914 unlock_res_and_lock(lock);
916 if (ldlm_is_fail_loc(lock))
917 CFS_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
919 if (ldlm_is_atomic_cb(lock) || ldlm_is_local(lock) ||
920 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
921 ldlm_handle_bl_callback(ns, NULL, lock);
922 } else if (ns_is_client(ns) &&
923 !lock->l_readers && !lock->l_writers &&
924 !ldlm_is_no_lru(lock) &&
925 !ldlm_is_bl_ast(lock) &&
926 !ldlm_is_converting(lock)) {
928 /* If this is a client-side namespace and this was the last
929 * reference, put it on the LRU.
931 ldlm_lock_add_to_lru(lock);
932 unlock_res_and_lock(lock);
933 LDLM_DEBUG(lock, "add lock into lru list");
935 if (ldlm_is_fail_loc(lock))
936 CFS_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
938 ldlm_pool_recalc(&ns->ns_pool, true);
940 LDLM_DEBUG(lock, "do not add lock into lru list");
941 unlock_res_and_lock(lock);
948 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
950 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
952 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
954 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
955 ldlm_lock_decref_internal(lock, mode);
958 EXPORT_SYMBOL(ldlm_lock_decref);
961 * Decrease reader/writer refcount for LDLM lock with handle
962 * \a lockh and mark it for subsequent cancellation once r/w refcount
963 * drops to zero instead of putting into LRU.
966 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
969 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
973 LASSERT(lock != NULL);
975 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
976 lock_res_and_lock(lock);
977 ldlm_set_cbpending(lock);
978 unlock_res_and_lock(lock);
979 ldlm_lock_decref_internal(lock, mode);
982 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
984 struct sl_insert_point {
985 struct list_head *res_link;
986 struct list_head *mode_link;
987 struct list_head *policy_link;
991 * Finds a position to insert the new lock into granted lock list.
993 * Used for locks eligible for skiplist optimization.
996 * queue [input]: the granted list where search acts on;
997 * req [input]: the lock whose position to be located;
998 * prev [output]: positions within 3 lists to insert @req to
1002 * - ldlm_grant_lock_with_skiplist
1004 static void search_granted_lock(struct list_head *queue,
1005 struct ldlm_lock *req,
1006 struct sl_insert_point *prev)
1008 struct ldlm_lock *lock, *mode_end, *policy_end;
1011 list_for_each_entry(lock, queue, l_res_link) {
1012 mode_end = list_entry(lock->l_sl_mode.prev,
1013 struct ldlm_lock, l_sl_mode);
1015 if (lock->l_req_mode != req->l_req_mode) {
1016 /* jump to last lock of mode group */
1021 /* suitable mode group is found */
1022 if (lock->l_resource->lr_type == LDLM_PLAIN) {
1023 /* insert point is last lock of the mode group */
1024 prev->res_link = &mode_end->l_res_link;
1025 prev->mode_link = &mode_end->l_sl_mode;
1026 prev->policy_link = &req->l_sl_policy;
1029 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
1032 list_entry(lock->l_sl_policy.prev,
1036 if (lock->l_policy_data.l_inodebits.bits ==
1037 req->l_policy_data.l_inodebits.bits) {
1038 /* inserting last lock of policy grp */
1040 &policy_end->l_res_link;
1042 &policy_end->l_sl_mode;
1044 &policy_end->l_sl_policy;
1049 if (policy_end == mode_end)
1050 /* done with mode group */
1053 /* go to next policy group within mode group */
1054 lock = list_next_entry(policy_end, l_res_link);
1055 } /* loop over policy groups within the mode group */
1057 /* insert point is last lock of the mode group,
1058 * new policy group is started
1060 prev->res_link = &mode_end->l_res_link;
1061 prev->mode_link = &mode_end->l_sl_mode;
1062 prev->policy_link = &req->l_sl_policy;
1066 LDLM_ERROR(lock, "is not LDLM_PLAIN or LDLM_IBITS lock");
1070 /* insert point is last lock on the queue,
1071 * new mode group and new policy group are started
1073 prev->res_link = queue->prev;
1074 prev->mode_link = &req->l_sl_mode;
1075 prev->policy_link = &req->l_sl_policy;
1080 * Add a lock into resource granted list after a position described by
1083 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1084 struct sl_insert_point *prev)
1086 struct ldlm_resource *res = lock->l_resource;
1090 check_res_locked(res);
1092 ldlm_resource_dump(D_INFO, res);
1093 LDLM_DEBUG(lock, "About to add lock:");
1095 if (ldlm_is_destroyed(lock)) {
1096 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1100 LASSERT(list_empty(&lock->l_res_link));
1101 LASSERT(list_empty(&lock->l_sl_mode));
1102 LASSERT(list_empty(&lock->l_sl_policy));
1105 * lock->link == prev->link means lock is first starting the group.
1106 * Don't re-add to itself to suppress kernel warnings.
1108 if (&lock->l_res_link != prev->res_link)
1109 list_add(&lock->l_res_link, prev->res_link);
1110 if (&lock->l_sl_mode != prev->mode_link)
1111 list_add(&lock->l_sl_mode, prev->mode_link);
1112 if (&lock->l_sl_policy != prev->policy_link)
1113 list_add(&lock->l_sl_policy, prev->policy_link);
1119 * Add a lock to granted list on a resource maintaining skiplist
1122 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1124 struct sl_insert_point prev;
1126 LASSERT(ldlm_is_granted(lock));
1128 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1129 ldlm_granted_list_add_lock(lock, &prev);
1133 * Perform lock granting bookkeeping.
1135 * Includes putting the lock into granted list and updating lock mode.
1137 * - ldlm_lock_enqueue
1138 * - ldlm_reprocess_queue
1140 * must be called with lr_lock held
1142 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1144 struct ldlm_resource *res = lock->l_resource;
1148 check_res_locked(res);
1150 lock->l_granted_mode = lock->l_req_mode;
1152 if (work_list && lock->l_completion_ast != NULL)
1153 ldlm_add_ast_work_item(lock, NULL, work_list);
1155 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1156 ldlm_grant_lock_with_skiplist(lock);
1157 else if (res->lr_type == LDLM_EXTENT)
1158 ldlm_extent_add_lock(res, lock);
1159 else if (res->lr_type == LDLM_FLOCK) {
1160 /* We should not add locks to granted list in the following
1162 * - this is an UNLOCK but not a real lock;
1163 * - this is a TEST lock;
1164 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1165 * - this is a deadlock (flock cannot be granted)
1167 if (lock->l_req_mode == 0 ||
1168 lock->l_req_mode == LCK_NL ||
1169 ldlm_is_test_lock(lock) ||
1170 ldlm_is_flock_deadlock(lock))
1172 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1177 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1182 * Check if the given @lock meets the criteria for a match.
1183 * A reference on the lock is taken if matched.
1185 * @lock test-against this lock
1188 * RETURN returns true if @lock matches @data, false otherwise
1190 static bool lock_matches(struct ldlm_lock *lock, void *vdata)
1192 struct ldlm_match_data *data = vdata;
1193 union ldlm_policy_data *lpol = &lock->l_policy_data;
1194 enum ldlm_mode match = LCK_MINMODE;
1196 if (lock == data->lmd_old)
1199 /* Check if this lock can be matched.
1200 * Used by LU-2919(exclusive open) for open lease lock
1202 if (ldlm_is_excl(lock))
1205 /* llite sometimes wants to match locks that will be
1206 * canceled when their users drop, but we allow it to match
1207 * if it passes in CBPENDING and the lock still has users.
1208 * this is generally only going to be used by children
1209 * whose parents already hold a lock so forward progress
1212 if (ldlm_is_cbpending(lock) &&
1213 !(data->lmd_flags & LDLM_FL_CBPENDING) &&
1214 !(data->lmd_match & LDLM_MATCH_GROUP))
1217 if (!(data->lmd_match & (LDLM_MATCH_UNREF | LDLM_MATCH_GROUP)) &&
1218 ldlm_is_cbpending(lock) &&
1219 lock->l_readers == 0 && lock->l_writers == 0)
1222 if (!(lock->l_req_mode & *data->lmd_mode))
1225 /* When we search for ast_data, we are not doing a traditional match,
1226 * so we don't worry about IBITS or extent matching.
1228 if (data->lmd_match & (LDLM_MATCH_AST | LDLM_MATCH_AST_ANY)) {
1229 if (!lock->l_ast_data)
1232 if (data->lmd_match & LDLM_MATCH_AST_ANY)
1236 match = lock->l_req_mode;
1238 switch (lock->l_resource->lr_type) {
1240 if (!(data->lmd_match & LDLM_MATCH_RIGHT) &&
1241 (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1242 lpol->l_extent.end < data->lmd_policy->l_extent.end))
1245 if (unlikely(match == LCK_GROUP) &&
1246 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1247 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1251 /* We match with existing lock with same or wider set of bits */
1252 if ((lpol->l_inodebits.bits &
1253 data->lmd_policy->l_inodebits.bits) !=
1254 data->lmd_policy->l_inodebits.bits)
1257 if (unlikely(match == LCK_GROUP) &&
1258 data->lmd_policy->l_inodebits.li_gid != LDLM_GID_ANY &&
1259 lpol->l_inodebits.li_gid !=
1260 data->lmd_policy->l_inodebits.li_gid)
1267 /* We match if we have existing lock with same or wider set of bits. */
1268 if (!(data->lmd_match & LDLM_MATCH_UNREF) && LDLM_HAVE_MASK(lock, GONE))
1271 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1274 /* Filter locks by skipping flags */
1275 if (data->lmd_skip_flags & lock->l_flags)
1280 * In case the lock is a CBPENDING grouplock, just pin it and return,
1281 * we need to wait until it gets to DESTROYED.
1283 if ((data->lmd_flags & LDLM_FL_TEST_LOCK) ||
1284 (ldlm_is_cbpending(lock) && (data->lmd_match & LDLM_MATCH_GROUP))) {
1285 LDLM_LOCK_GET(lock);
1286 ldlm_lock_touch_in_lru(lock);
1288 ldlm_lock_addref_internal_nolock(lock, match);
1291 *data->lmd_mode = match;
1292 data->lmd_lock = lock;
1298 * Search for a lock with given parameters in interval trees.
1300 * \param res search for a lock in this resource
1301 * \param data parameters
1303 * \retval a referenced lock or NULL.
1305 struct ldlm_lock *search_itree(struct ldlm_resource *res,
1306 struct ldlm_match_data *data)
1308 struct interval_node_extent ext = {
1309 .start = data->lmd_policy->l_extent.start,
1310 .end = data->lmd_policy->l_extent.end
1314 data->lmd_lock = NULL;
1316 if (data->lmd_match & LDLM_MATCH_RIGHT)
1317 ext.end = OBD_OBJECT_EOF;
1319 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1320 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1322 if (tree->lit_root == NULL)
1325 if (!(tree->lit_mode & *data->lmd_mode))
1328 ldlm_extent_search(tree->lit_root, &ext,
1329 lock_matches, data);
1331 return data->lmd_lock;
1336 EXPORT_SYMBOL(search_itree);
1340 * Search for a lock with given properties in a queue.
1342 * \param queue search for a lock in this queue
1343 * \param data parameters
1345 * \retval a referenced lock or NULL.
1347 static struct ldlm_lock *search_queue(struct list_head *queue,
1348 struct ldlm_match_data *data)
1350 struct ldlm_lock *lock;
1352 data->lmd_lock = NULL;
1354 list_for_each_entry(lock, queue, l_res_link)
1355 if (lock_matches(lock, data))
1356 return data->lmd_lock;
1361 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1363 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1364 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1365 wake_up(&lock->l_waitq);
1368 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1370 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1372 lock_res_and_lock(lock);
1373 ldlm_lock_fail_match_locked(lock);
1374 unlock_res_and_lock(lock);
1378 * Mark lock as "matchable" by OST.
1380 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1382 * Assumes LDLM lock is already locked.
1384 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1386 ldlm_set_lvb_ready(lock);
1387 wake_up(&lock->l_waitq);
1389 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1392 * Mark lock as "matchable" by OST.
1393 * Locks the lock and then \see ldlm_lock_allow_match_locked
1395 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1397 lock_res_and_lock(lock);
1398 ldlm_lock_allow_match_locked(lock);
1399 unlock_res_and_lock(lock);
1401 EXPORT_SYMBOL(ldlm_lock_allow_match);
1404 * Attempt to find a lock with specified properties.
1406 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1409 * Can be called in two ways:
1411 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1412 * for a duplicate of.
1414 * Otherwise, all of the fields must be filled in, to match against.
1416 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1417 * server (ie, connh is NULL)
1418 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1419 * list will be considered
1420 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1421 * to be canceled can still be matched as long as they still have reader
1422 * or writer refernces
1423 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1424 * just tell us if we would have matched.
1426 * \retval 1 if it finds an already-existing lock that is compatible; in this
1427 * case, lockh is filled in with a addref()ed lock
1429 * We also check security context, and if that fails we simply return 0 (to
1430 * keep caller code unchanged), the context failure will be discovered by
1431 * caller sometime later.
1433 enum ldlm_mode ldlm_lock_match_with_skip(struct ldlm_namespace *ns,
1434 __u64 flags, __u64 skip_flags,
1435 const struct ldlm_res_id *res_id,
1436 enum ldlm_type type,
1437 union ldlm_policy_data *policy,
1438 enum ldlm_mode mode,
1439 struct lustre_handle *lockh,
1440 enum ldlm_match_flags match_flags)
1442 struct ldlm_match_data data = {
1446 .lmd_policy = policy,
1448 .lmd_skip_flags = skip_flags,
1449 .lmd_match = match_flags,
1451 struct ldlm_resource *res;
1452 struct ldlm_lock *lock;
1453 struct ldlm_lock *group_lock;
1459 data.lmd_old = ldlm_handle2lock(lockh);
1460 LASSERT(data.lmd_old != NULL);
1462 ns = ldlm_lock_to_ns(data.lmd_old);
1463 res_id = &data.lmd_old->l_resource->lr_name;
1464 type = data.lmd_old->l_resource->lr_type;
1465 *data.lmd_mode = data.lmd_old->l_req_mode;
1468 res = ldlm_resource_get(ns, res_id, type, 0);
1470 LASSERT(data.lmd_old == NULL);
1476 LDLM_RESOURCE_ADDREF(res);
1478 if (res->lr_type == LDLM_EXTENT)
1479 lock = search_itree(res, &data);
1481 lock = search_queue(&res->lr_granted, &data);
1482 if (!lock && !(flags & LDLM_FL_BLOCK_GRANTED))
1483 lock = search_queue(&res->lr_waiting, &data);
1484 matched = lock ? mode : 0;
1486 if (lock && ldlm_is_cbpending(lock) &&
1487 (data.lmd_match & LDLM_MATCH_GROUP))
1490 LDLM_RESOURCE_DELREF(res);
1493 l_wait_event_abortable(group_lock->l_waitq,
1494 ldlm_is_destroyed(lock));
1495 LDLM_LOCK_RELEASE(lock);
1498 ldlm_resource_putref(res);
1501 ldlm_lock2handle(lock, lockh);
1502 if ((flags & LDLM_FL_LVB_READY) &&
1503 (!ldlm_is_lvb_ready(lock))) {
1504 __u64 wait_flags = LDLM_FL_LVB_READY |
1505 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1507 if (lock->l_completion_ast) {
1508 int err = lock->l_completion_ast(lock,
1509 LDLM_FL_WAIT_NOREPROC,
1512 GOTO(out_fail_match, matched = 0);
1515 wait_event_idle_timeout(
1517 lock->l_flags & wait_flags,
1518 cfs_time_seconds(obd_timeout));
1520 if (!ldlm_is_lvb_ready(lock))
1521 GOTO(out_fail_match, matched = 0);
1524 /* check user's security context */
1525 if (lock->l_conn_export &&
1526 sptlrpc_import_check_ctx(
1527 class_exp2cliimp(lock->l_conn_export)))
1528 GOTO(out_fail_match, matched = 0);
1530 LDLM_DEBUG(lock, "matched (%llu %llu)",
1531 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1532 res_id->name[2] : policy->l_extent.start,
1533 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1534 res_id->name[3] : policy->l_extent.end);
1537 if (flags & LDLM_FL_TEST_LOCK)
1538 LDLM_LOCK_RELEASE(lock);
1540 ldlm_lock_decref_internal(lock, mode);
1543 /* less verbose for test-only */
1544 if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
1545 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1546 "%llu/%llu (%llu %llu)", ns,
1547 type, mode, res_id->name[0], res_id->name[1],
1548 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1549 res_id->name[2] : policy->l_extent.start,
1550 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1551 res_id->name[3] : policy->l_extent.end);
1553 if (data.lmd_old != NULL)
1554 LDLM_LOCK_PUT(data.lmd_old);
1558 EXPORT_SYMBOL(ldlm_lock_match_with_skip);
1560 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1563 struct ldlm_lock *lock;
1564 enum ldlm_mode mode = 0;
1568 lock = ldlm_handle2lock(lockh);
1570 lock_res_and_lock(lock);
1571 if (LDLM_HAVE_MASK(lock, GONE))
1574 if (ldlm_is_cbpending(lock) &&
1575 lock->l_readers == 0 && lock->l_writers == 0)
1579 *bits = lock->l_policy_data.l_inodebits.bits;
1580 mode = lock->l_granted_mode;
1581 ldlm_lock_addref_internal_nolock(lock, mode);
1588 unlock_res_and_lock(lock);
1589 LDLM_LOCK_PUT(lock);
1593 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1595 /** The caller must guarantee that the buffer is large enough. */
1596 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1597 enum req_location loc, void *data, int size)
1603 LASSERT(data != NULL);
1606 switch (lock->l_lvb_type) {
1608 if (size == sizeof(struct ost_lvb)) {
1609 if (loc == RCL_CLIENT)
1610 lvb = req_capsule_client_swab_get(pill,
1612 lustre_swab_ost_lvb);
1614 lvb = req_capsule_server_swab_get(pill,
1616 lustre_swab_ost_lvb);
1617 if (unlikely(lvb == NULL)) {
1618 LDLM_ERROR(lock, "no LVB");
1622 memcpy(data, lvb, size);
1623 } else if (size == sizeof(struct ost_lvb_v1)) {
1624 struct ost_lvb *olvb = data;
1626 if (loc == RCL_CLIENT)
1627 lvb = req_capsule_client_swab_get(pill,
1629 lustre_swab_ost_lvb_v1);
1631 lvb = req_capsule_server_sized_swab_get(pill,
1633 lustre_swab_ost_lvb_v1);
1634 if (unlikely(lvb == NULL)) {
1635 LDLM_ERROR(lock, "no LVB");
1639 memcpy(data, lvb, size);
1640 olvb->lvb_mtime_ns = 0;
1641 olvb->lvb_atime_ns = 0;
1642 olvb->lvb_ctime_ns = 0;
1644 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1650 if (size == sizeof(struct lquota_lvb)) {
1651 if (loc == RCL_CLIENT)
1652 lvb = req_capsule_client_swab_get(pill,
1654 lustre_swab_lquota_lvb);
1656 lvb = req_capsule_server_swab_get(pill,
1658 lustre_swab_lquota_lvb);
1659 if (unlikely(lvb == NULL)) {
1660 LDLM_ERROR(lock, "no LVB");
1664 memcpy(data, lvb, size);
1667 "Replied unexpected lquota LVB size %d",
1676 if (loc == RCL_CLIENT)
1677 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1679 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1680 if (unlikely(lvb == NULL)) {
1681 LDLM_ERROR(lock, "no LVB");
1685 memcpy(data, lvb, size);
1688 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1696 /* Create and fill in new LDLM lock with specified properties.
1697 * Returns: a referenced lock
1699 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1700 const struct ldlm_res_id *res_id,
1701 enum ldlm_type type,
1702 enum ldlm_mode mode,
1703 const struct ldlm_callback_suite *cbs,
1704 void *data, __u32 lvb_len,
1705 enum lvb_type lvb_type)
1707 struct ldlm_lock *lock;
1708 struct ldlm_resource *res;
1713 res = ldlm_resource_get(ns, res_id, type, 1);
1715 RETURN(ERR_CAST(res));
1717 lock = ldlm_lock_new(res);
1719 ldlm_resource_putref(res);
1720 RETURN(ERR_PTR(-ENOMEM));
1723 lock->l_req_mode = mode;
1724 lock->l_ast_data = data;
1725 lock->l_pid = current->pid;
1726 if (ns_is_server(ns))
1727 ldlm_set_ns_srv(lock);
1729 lock->l_blocking_ast = cbs->lcs_blocking;
1730 lock->l_completion_ast = cbs->lcs_completion;
1731 lock->l_glimpse_ast = cbs->lcs_glimpse;
1734 if (type == LDLM_IBITS) {
1735 rc = ldlm_inodebits_alloc_lock(lock);
1741 lock->l_lvb_len = lvb_len;
1742 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1743 if (lock->l_lvb_data == NULL)
1744 GOTO(out, rc = -ENOMEM);
1747 lock->l_lvb_type = lvb_type;
1748 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1749 GOTO(out, rc = -ENOENT);
1754 ldlm_lock_destroy(lock);
1755 LDLM_LOCK_RELEASE(lock);
1756 RETURN(ERR_PTR(rc));
1759 #ifdef HAVE_SERVER_SUPPORT
1760 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1763 struct ldlm_resource *res = lock->l_resource;
1764 enum ldlm_error rc = ELDLM_OK;
1765 LIST_HEAD(rpc_list);
1766 ldlm_processing_policy policy;
1770 policy = ldlm_get_processing_policy(res);
1771 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1772 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1773 res->lr_type != LDLM_FLOCK)
1774 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1776 if (!list_empty(&rpc_list))
1777 ldlm_discard_bl_list(&rpc_list);
1784 * Enqueue (request) a lock.
1786 * Does not block. As a result of enqueue the lock would be put
1787 * into granted or waiting list.
1789 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1790 * set, skip all the enqueueing and delegate lock processing to intent policy
1793 enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
1794 struct ldlm_namespace *ns,
1795 struct ldlm_lock **lockp,
1796 void *cookie, __u64 *flags)
1798 struct ldlm_lock *lock = *lockp;
1799 struct ldlm_resource *res;
1800 int local = ns_is_client(ns);
1801 enum ldlm_error rc = ELDLM_OK;
1802 #ifdef HAVE_SERVER_SUPPORT
1803 bool reconstruct = false;
1807 /* policies are not executed on the client or during replay */
1808 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1809 && !local && ns->ns_policy) {
1810 rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
1812 if (rc == ELDLM_LOCK_REPLACED) {
1813 /* The lock that was returned has already been granted,
1814 * and placed into lockp. If it's not the same as the
1815 * one we passed in, then destroy the old one and our
1816 * work here is done.
1818 if (lock != *lockp) {
1819 ldlm_lock_destroy(lock);
1820 LDLM_LOCK_RELEASE(lock);
1822 *flags |= LDLM_FL_LOCK_CHANGED;
1824 } else if (rc != ELDLM_OK &&
1825 ldlm_is_granted(lock)) {
1826 LASSERT(*flags & LDLM_FL_RESENT);
1827 /* It may happen that ns_policy returns an error in
1828 * resend case, object may be unlinked or just some
1829 * error occurs. It is unclear if lock reached the
1830 * client in the original reply, just leave the lock on
1831 * server, not returning it again to client. Due to
1832 * LU-6529, the server will not OOM.
1835 } else if (rc != ELDLM_OK ||
1836 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1837 ldlm_lock_destroy(lock);
1842 if (*flags & LDLM_FL_RESENT) {
1843 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1844 * Set LOCK_CHANGED always.
1845 * Check if the lock is granted for BLOCK_GRANTED.
1846 * Take NO_TIMEOUT from the lock as it is inherited through
1847 * LDLM_FL_INHERIT_MASK
1849 *flags |= LDLM_FL_LOCK_CHANGED;
1850 if (!ldlm_is_granted(lock))
1851 *flags |= LDLM_FL_BLOCK_GRANTED;
1852 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1856 #ifdef HAVE_SERVER_SUPPORT
1857 reconstruct = !local && lock->l_resource->lr_type == LDLM_FLOCK &&
1858 !(*flags & LDLM_FL_TEST_LOCK);
1860 rc = req_can_reconstruct(cookie, NULL);
1868 if (!local && lock->l_resource->lr_type == LDLM_FLOCK) {
1869 struct ldlm_flock_node *fn = &lock->l_resource->lr_flock_node;
1871 if (lock->l_req_mode == LCK_NL) {
1872 atomic_inc(&fn->lfn_unlock_pending);
1873 res = lock_res_and_lock(lock);
1874 atomic_dec(&fn->lfn_unlock_pending);
1876 res = lock_res_and_lock(lock);
1878 while (atomic_read(&fn->lfn_unlock_pending)) {
1879 unlock_res_and_lock(lock);
1881 lock_res_and_lock(lock);
1887 res = lock_res_and_lock(lock);
1889 if (local && ldlm_is_granted(lock)) {
1890 /* The server returned a blocked lock, but it was granted
1891 * before we got a chance to actually enqueue it. We don't
1892 * need to do anything else.
1894 *flags &= ~LDLM_FL_BLOCKED_MASK;
1895 GOTO(out, rc = ELDLM_OK);
1898 ldlm_resource_unlink_lock(lock);
1900 /* Some flags from the enqueue want to make it into the AST, via the
1903 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1904 ldlm_set_ast_discard_data(lock);
1905 if (*flags & LDLM_FL_TEST_LOCK)
1906 ldlm_set_test_lock(lock);
1908 /* This distinction between local lock trees is very important; a client
1909 * namespace only has information about locks taken by that client, and
1910 * thus doesn't have enough information to decide for itself if it can
1911 * be granted (below). In this case, we do exactly what the server
1912 * tells us to do, as dictated by the 'flags'.
1914 * We do exactly the same thing during recovery, when the server is
1915 * more or less trusting the clients not to lie.
1917 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1921 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1922 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1924 ldlm_grant_lock(lock, NULL);
1925 GOTO(out, rc = ELDLM_OK);
1926 #ifdef HAVE_SERVER_SUPPORT
1927 } else if (*flags & LDLM_FL_REPLAY) {
1928 if (*flags & LDLM_FL_BLOCK_WAIT) {
1929 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1930 GOTO(out, rc = ELDLM_OK);
1931 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1932 ldlm_grant_lock(lock, NULL);
1933 GOTO(out, rc = ELDLM_OK);
1935 /* If no flags, fall through to normal enqueue path. */
1938 rc = ldlm_lock_enqueue_helper(lock, flags);
1942 CERROR("This is client-side-only module, cannot handle LDLM_NAMESPACE_SERVER resource type lock.\n");
1948 unlock_res_and_lock(lock);
1950 #ifdef HAVE_SERVER_SUPPORT
1952 struct ptlrpc_request *req = cookie;
1954 tgt_mk_reply_data(NULL, NULL,
1955 &req->rq_export->exp_target_data,
1956 req, 0, NULL, false, 0);
1962 #ifdef HAVE_SERVER_SUPPORT
1964 * Iterate through all waiting locks on a given resource queue and attempt to
1967 * Must be called with resource lock held.
1969 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1970 struct list_head *work_list,
1971 enum ldlm_process_intention intention, __u64 hint)
1973 struct list_head *tmp, *pos;
1974 ldlm_processing_policy policy;
1976 int rc = LDLM_ITER_CONTINUE;
1977 enum ldlm_error err;
1978 LIST_HEAD(bl_ast_list);
1982 check_res_locked(res);
1984 policy = ldlm_get_processing_policy(res);
1986 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1987 intention == LDLM_PROCESS_RECOVERY);
1990 list_for_each_safe(tmp, pos, queue) {
1991 struct ldlm_lock *pending;
1992 LIST_HEAD(rpc_list);
1994 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1996 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1999 rc = policy(pending, &flags, intention, &err, &rpc_list);
2000 if (pending->l_granted_mode == pending->l_req_mode ||
2001 res->lr_type == LDLM_FLOCK) {
2002 list_splice(&rpc_list, work_list);
2004 list_splice(&rpc_list, &bl_ast_list);
2007 * When this is called from recovery done, we always want
2008 * to scan the whole list no matter what 'rc' is returned.
2010 if (rc != LDLM_ITER_CONTINUE &&
2011 intention == LDLM_PROCESS_RESCAN)
2015 if (!list_empty(&bl_ast_list)) {
2018 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
2022 if (rc == -ERESTART)
2026 if (!list_empty(&bl_ast_list))
2027 ldlm_discard_bl_list(&bl_ast_list);
2029 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
2033 * Conflicting locks are detected for a lock to be enqueued, add the lock
2034 * into waiting list and send blocking ASTs to the conflicting locks.
2036 * \param[in] lock The lock to be enqueued.
2037 * \param[out] flags Lock flags for the lock to be enqueued.
2038 * \param[in] rpc_list Conflicting locks list.
2040 * \retval -ERESTART: Some lock was instantly canceled while sending
2041 * blocking ASTs, caller needs to re-check conflicting
2043 * \retval -EAGAIN: Lock was destroyed, caller should return error.
2044 * \reval 0: Lock is successfully added in waiting list.
2046 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
2047 struct list_head *rpc_list)
2049 struct ldlm_resource *res = lock->l_resource;
2054 check_res_locked(res);
2056 /* If either of the compat_queue()s returned failure, then we
2057 * have ASTs to send and must go onto the waiting list.
2059 * bug 2322: we used to unlink and re-add here, which was a
2060 * terrible folly -- if we goto restart, we could get
2061 * re-ordered! Causes deadlock, because ASTs aren't sent!
2063 if (list_empty(&lock->l_res_link))
2064 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
2067 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
2070 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
2071 !ns_is_client(ldlm_res_to_ns(res)))
2072 class_fail_export(lock->l_export);
2074 if (rc == -ERESTART)
2075 ldlm_reprocess_all(res, 0);
2078 if (rc == -ERESTART) {
2079 /* 15715: The lock was granted and destroyed after
2080 * resource lock was dropped. Interval node was freed
2081 * in ldlm_lock_destroy. Anyway, this always happens
2082 * when a client is being evicted. So it would be
2083 * ok to return an error. -jay
2085 if (ldlm_is_destroyed(lock))
2088 /* lock was granted while resource was unlocked. */
2089 if (ldlm_is_granted(lock)) {
2090 /* bug 11300: if the lock has been granted,
2091 * break earlier because otherwise, we will go
2092 * to restart and ldlm_resource_unlink will be
2093 * called and it causes the interval node to be
2094 * freed. Then we will fail at
2095 * ldlm_extent_add_lock()
2097 *flags &= ~LDLM_FL_BLOCKED_MASK;
2101 *flags |= LDLM_FL_BLOCK_GRANTED;
2107 * Discard all AST work items from list.
2109 * If for whatever reason we do not want to send ASTs to conflicting locks
2110 * anymore, disassemble the list with this function.
2112 void ldlm_discard_bl_list(struct list_head *bl_list)
2114 struct ldlm_lock *lock, *tmp;
2118 list_for_each_entry_safe(lock, tmp, bl_list, l_bl_ast) {
2119 LASSERT(!list_empty(&lock->l_bl_ast));
2120 list_del_init(&lock->l_bl_ast);
2121 ldlm_clear_ast_sent(lock);
2122 LASSERT(lock->l_bl_ast_run == 0);
2123 ldlm_clear_blocking_lock(lock);
2124 LDLM_LOCK_RELEASE(lock);
2129 /* Process a call to blocking AST callback for a lock in ast_work list */
2131 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2133 struct ldlm_cb_set_arg *arg = opaq;
2134 struct ldlm_lock *lock;
2135 struct ldlm_lock_desc d;
2136 struct ldlm_bl_desc bld;
2141 if (list_empty(arg->list))
2144 lock = list_first_entry(arg->list, struct ldlm_lock, l_bl_ast);
2146 /* nobody should touch l_bl_ast but some locks in the list may become
2147 * granted after lock convert or COS downgrade, these locks should be
2148 * just skipped here and removed from the list.
2150 lock_res_and_lock(lock);
2151 list_del_init(&lock->l_bl_ast);
2153 /* lock is not blocking lock anymore, but was kept in the list because
2154 * it can managed only here.
2156 if (!ldlm_is_ast_sent(lock)) {
2157 unlock_res_and_lock(lock);
2158 LDLM_LOCK_RELEASE(lock);
2162 LASSERT(lock->l_blocking_lock);
2163 ldlm_lock2desc(lock->l_blocking_lock, &d);
2164 /* copy blocking lock ibits in cancel_bits as well,
2165 * new client may use them for lock convert and it is
2166 * important to use new field to convert locks from
2169 d.l_policy_data.l_inodebits.cancel_bits =
2170 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2172 /* Blocking lock is being destroyed here but some information about it
2173 * may be needed inside l_blocking_ast() function below,
2174 * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
2176 bld.bl_same_client = lock->l_client_cookie ==
2177 lock->l_blocking_lock->l_client_cookie;
2178 /* if two locks are initiated from the same MDT, transactions are
2179 * independent, or the request lock mode isn't EX|PW, no need to trigger
2180 * CoS because current lock will be downgraded to TXN mode soon, then
2181 * the blocking lock can be granted.
2183 if (lock->l_blocking_lock->l_policy_data.l_inodebits.li_initiator_id ==
2184 lock->l_policy_data.l_inodebits.li_initiator_id ||
2185 !(lock->l_blocking_lock->l_req_mode & (LCK_EX | LCK_PW)))
2186 bld.bl_txn_dependent = false;
2188 bld.bl_txn_dependent = true;
2189 arg->bl_desc = &bld;
2191 LASSERT(ldlm_is_ast_sent(lock));
2192 LASSERT(lock->l_bl_ast_run == 0);
2193 lock->l_bl_ast_run++;
2194 ldlm_clear_blocking_lock(lock);
2195 unlock_res_and_lock(lock);
2197 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2199 LDLM_LOCK_RELEASE(lock);
2204 /* Process a call to revocation AST callback for a lock in ast_work list */
2206 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2208 struct ldlm_cb_set_arg *arg = opaq;
2209 struct ldlm_lock_desc desc;
2211 struct ldlm_lock *lock;
2215 if (list_empty(arg->list))
2218 lock = list_first_entry(arg->list, struct ldlm_lock, l_rk_ast);
2219 list_del_init(&lock->l_rk_ast);
2221 /* the desc just pretend to exclusive */
2222 ldlm_lock2desc(lock, &desc);
2223 desc.l_req_mode = LCK_EX;
2224 desc.l_granted_mode = 0;
2226 rc = lock->l_blocking_ast(lock, &desc, (void *)arg, LDLM_CB_BLOCKING);
2227 LDLM_LOCK_RELEASE(lock);
2232 /* Process a call to glimpse AST callback for a lock in ast_work list */
2233 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2235 struct ldlm_cb_set_arg *arg = opaq;
2236 struct ldlm_glimpse_work *gl_work;
2237 struct ldlm_lock *lock;
2242 if (list_empty(arg->list))
2245 gl_work = list_first_entry(arg->list, struct ldlm_glimpse_work,
2247 list_del_init(&gl_work->gl_list);
2249 lock = gl_work->gl_lock;
2251 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2252 arg->gl_desc = gl_work->gl_desc;
2253 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2254 arg->gl_interpret_data = gl_work->gl_interpret_data;
2256 /* invoke the actual glimpse callback */
2257 rc = lock->l_glimpse_ast(lock, (void *)arg);
2259 rc = 1; /* update LVB if this is server lock */
2260 else if (rc == -ELDLM_NO_LOCK_DATA)
2261 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
2263 LDLM_LOCK_RELEASE(lock);
2264 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2265 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2267 OBD_FREE_PTR(gl_work);
2273 /* Process a call to completion AST callback for a lock in ast_work list */
2275 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2277 struct ldlm_cb_set_arg *arg = opaq;
2278 struct ldlm_lock *lock;
2279 ldlm_completion_callback completion_callback;
2284 if (list_empty(arg->list))
2287 lock = list_first_entry(arg->list, struct ldlm_lock, l_cp_ast);
2289 /* It's possible to receive a completion AST before we've set
2290 * the l_completion_ast pointer: either because the AST arrived
2291 * before the reply, or simply because there's a small race
2292 * window between receiving the reply and finishing the local
2293 * enqueue. (bug 842)
2295 * This can't happen with the blocking_ast, however, because we
2296 * will never call the local blocking_ast until we drop our
2297 * reader/writer reference, which we won't do until we get the
2298 * reply and finish enqueueing.
2301 /* nobody should touch l_cp_ast */
2302 lock_res_and_lock(lock);
2303 list_del_init(&lock->l_cp_ast);
2304 LASSERT(ldlm_is_cp_reqd(lock));
2305 /* save l_completion_ast since it can be changed by
2306 * mds_intent_policy(), see bug 14225
2308 completion_callback = lock->l_completion_ast;
2309 ldlm_clear_cp_reqd(lock);
2310 unlock_res_and_lock(lock);
2312 if (completion_callback != NULL)
2313 rc = completion_callback(lock, 0, (void *)arg);
2314 LDLM_LOCK_RELEASE(lock);
2320 * Process list of locks in need of ASTs being sent.
2322 * Used on server to send multiple ASTs together instead of sending one by
2325 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2326 ldlm_desc_ast_t ast_type)
2328 struct ldlm_cb_set_arg *arg;
2329 set_producer_func work_ast_lock;
2332 if (list_empty(rpc_list))
2339 atomic_set(&arg->restart, 0);
2340 arg->list = rpc_list;
2343 case LDLM_WORK_CP_AST:
2344 arg->type = LDLM_CP_CALLBACK;
2345 work_ast_lock = ldlm_work_cp_ast_lock;
2347 #ifdef HAVE_SERVER_SUPPORT
2348 case LDLM_WORK_BL_AST:
2349 arg->type = LDLM_BL_CALLBACK;
2350 work_ast_lock = ldlm_work_bl_ast_lock;
2352 case LDLM_WORK_REVOKE_AST:
2353 arg->type = LDLM_BL_CALLBACK;
2354 work_ast_lock = ldlm_work_revoke_ast_lock;
2356 case LDLM_WORK_GL_AST:
2357 arg->type = LDLM_GL_CALLBACK;
2358 work_ast_lock = ldlm_work_gl_ast_lock;
2365 /* We create a ptlrpc request set with flow control extension.
2366 * This request set will use the work_ast_lock function to produce new
2367 * requests and will send a new request each time one completes in order
2368 * to keep the number of requests in flight to ns_max_parallel_ast
2370 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2371 work_ast_lock, arg);
2372 if (arg->set == NULL)
2373 GOTO(out, rc = -ENOMEM);
2375 ptlrpc_set_wait(NULL, arg->set);
2376 ptlrpc_set_destroy(arg->set);
2378 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2386 * Try to grant all waiting locks on a resource.
2388 * Calls ldlm_reprocess_queue on waiting queue.
2390 * Typically called after some resource locks are cancelled to see
2391 * if anything could be granted as a result of the cancellation.
2393 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2394 enum ldlm_process_intention intention,
2397 LIST_HEAD(rpc_list);
2398 #ifdef HAVE_SERVER_SUPPORT
2399 ldlm_reprocessing_policy reprocess;
2400 struct obd_device *obd;
2405 /* Local lock trees don't get reprocessed. */
2406 if (ns_is_client(ldlm_res_to_ns(res))) {
2411 /* Disable reprocess during lock replay stage but allow during
2412 * request replay stage.
2414 obd = ldlm_res_to_ns(res)->ns_obd;
2415 if (obd->obd_recovering &&
2416 atomic_read(&obd->obd_req_replay_clients) == 0)
2420 reprocess = ldlm_get_reprocessing_policy(res);
2421 reprocess(res, &res->lr_waiting, &rpc_list, intention, hint);
2424 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2426 if (rc == -ERESTART) {
2427 LASSERT(list_empty(&rpc_list));
2434 if (!ns_is_client(ldlm_res_to_ns(res))) {
2435 CERROR("This is client-side-only module, cannot handle LDLM_NAMESPACE_SERVER resource type lock.\n");
2442 void ldlm_reprocess_all(struct ldlm_resource *res, __u64 hint)
2444 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN, hint);
2446 EXPORT_SYMBOL(ldlm_reprocess_all);
2448 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2449 struct hlist_node *hnode, void *arg)
2451 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2453 /* This is only called once after recovery done. LU-8306. */
2454 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY, 0);
2458 /* Iterate on all resources on namespace attempting to grant waiting locks. */
2459 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2464 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2465 ldlm_reprocess_res, NULL, 0);
2470 /* Helper to call blocking AST for LDLM lock \a lock in a "cancelling" mode. */
2471 void ldlm_cancel_callback(struct ldlm_lock *lock)
2473 check_res_locked(lock->l_resource);
2474 if (!ldlm_is_cancel(lock)) {
2475 ldlm_set_cancel(lock);
2476 if (lock->l_blocking_ast) {
2477 unlock_res_and_lock(lock);
2478 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2480 lock_res_and_lock(lock);
2482 LDLM_DEBUG(lock, "no blocking ast");
2485 /* only canceller can set bl_done bit */
2486 ldlm_set_bl_done(lock);
2487 wake_up(&lock->l_waitq);
2488 } else if (!ldlm_is_bl_done(lock)) {
2489 /* lock is guaranteed to be canceled returning from function. */
2490 unlock_res_and_lock(lock);
2491 wait_event_idle(lock->l_waitq, is_bl_done(lock));
2492 lock_res_and_lock(lock);
2496 /* Remove skiplist-enabled LDLM lock \a req from granted list */
2497 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2499 if (req->l_resource->lr_type != LDLM_PLAIN &&
2500 req->l_resource->lr_type != LDLM_IBITS)
2503 list_del_init(&req->l_sl_policy);
2504 list_del_init(&req->l_sl_mode);
2507 /* Attempts to cancel LDLM lock \a lock that has no reader/writer references. */
2508 void ldlm_lock_cancel(struct ldlm_lock *lock)
2510 struct ldlm_resource *res;
2511 struct ldlm_namespace *ns;
2515 lock_res_and_lock(lock);
2517 res = lock->l_resource;
2518 ns = ldlm_res_to_ns(res);
2520 /* Please do not remove this LBUG without talking to me first. -phik */
2521 if (lock->l_readers || lock->l_writers) {
2522 LDLM_ERROR(lock, "lock still has references");
2523 unlock_res_and_lock(lock);
2527 if (ldlm_is_waited(lock))
2528 ldlm_del_waiting_lock(lock);
2530 /* Releases cancel callback. */
2531 ldlm_cancel_callback(lock);
2533 /* Yes, second time, just in case it was added again while we were
2534 * running with no res lock in ldlm_cancel_callback
2536 if (ldlm_is_waited(lock))
2537 ldlm_del_waiting_lock(lock);
2539 ldlm_resource_unlink_lock(lock);
2540 ldlm_lock_destroy_nolock(lock);
2542 if (ldlm_is_granted(lock))
2543 ldlm_pool_del(&ns->ns_pool, lock);
2545 /* should not be called again for same lock(zero out l_granted_mode) */
2546 lock->l_granted_mode = LCK_MINMODE;
2547 unlock_res_and_lock(lock);
2551 EXPORT_SYMBOL(ldlm_lock_cancel);
2553 /* Set opaque data into the lock that only makes sense to upper layer. */
2554 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2556 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2562 if (lock->l_ast_data == NULL)
2563 lock->l_ast_data = data;
2564 if (lock->l_ast_data == data)
2566 LDLM_LOCK_PUT(lock);
2570 EXPORT_SYMBOL(ldlm_lock_set_data);
2572 struct export_cl_data {
2573 const struct lu_env *ecl_env;
2574 struct obd_export *ecl_exp;
2578 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2579 struct ldlm_lock *lock,
2580 struct export_cl_data *ecl)
2582 struct ldlm_resource *res;
2584 res = ldlm_resource_getref(lock->l_resource);
2586 ldlm_lvbo_update(res, lock, NULL, 1);
2587 ldlm_lock_cancel(lock);
2588 if (!exp->exp_obd->obd_stopping)
2589 ldlm_reprocess_all(res, lock->l_policy_data.l_inodebits.bits);
2590 ldlm_resource_putref(res);
2593 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2594 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2595 exp, ecl->ecl_loop);
2600 * Iterator function for ldlm_export_cancel_locks.
2601 * Cancels passed locks.
2604 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2605 struct hlist_node *hnode, void *data)
2608 struct export_cl_data *ecl = (struct export_cl_data *)data;
2609 struct obd_export *exp = ecl->ecl_exp;
2610 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2612 LDLM_LOCK_GET(lock);
2613 ldlm_cancel_lock_for_export(exp, lock, ecl);
2614 LDLM_LOCK_RELEASE(lock);
2620 * Cancel all blocked locks for given export.
2622 * Typically called on client disconnection/eviction
2624 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2627 struct export_cl_data ecl = {
2633 rc = lu_env_init(&env, LCT_DT_THREAD);
2638 while (!list_empty(&exp->exp_bl_list)) {
2639 struct ldlm_lock *lock;
2641 spin_lock_bh(&exp->exp_bl_list_lock);
2642 if (!list_empty(&exp->exp_bl_list)) {
2643 lock = list_first_entry(&exp->exp_bl_list,
2644 struct ldlm_lock, l_exp_list);
2645 LDLM_LOCK_GET(lock);
2646 list_del_init(&lock->l_exp_list);
2650 spin_unlock_bh(&exp->exp_bl_list_lock);
2655 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2656 LDLM_LOCK_RELEASE(lock);
2662 "Export %p, canceled %d locks, left on hash table %d.\n", exp,
2663 ecl.ecl_loop, atomic_read(&exp->exp_lock_hash->hs_count));
2665 return ecl.ecl_loop;
2669 * Cancel all locks for given export.
2671 * Typically called after client disconnection/eviction
2673 int ldlm_export_cancel_locks(struct obd_export *exp)
2675 struct export_cl_data ecl;
2679 rc = lu_env_init(&env, LCT_DT_THREAD);
2686 cfs_hash_for_each_empty(exp->exp_lock_hash,
2687 ldlm_cancel_locks_for_export_cb, &ecl);
2690 "Export %p, canceled %d locks, left on hash table %d.\n", exp,
2691 ecl.ecl_loop, atomic_read(&exp->exp_lock_hash->hs_count));
2693 if (ecl.ecl_loop > 0 &&
2694 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2695 exp->exp_obd->obd_stopping)
2696 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2700 return ecl.ecl_loop;
2704 * Downgrade an PW/EX lock to COS, TXN or CR mode.
2706 * A lock mode convertion from PW/EX mode to less conflict mode. The
2707 * convertion may fail if lock was canceled before downgrade, but it doesn't
2708 * indicate any problem, because such lock has no reader or writer, and will
2711 * Used by Commit on Sharing (COS) code to force object changes commit in case
2712 * of conflict. Converted lock is considered as new lock and all blocking AST
2713 * things are cleared, so any pending or new blocked lock on that lock will
2714 * cause new call to blocking_ast and force resource object commit.
2716 * Used by DNE to force commit upon operation dependency.
2718 * Also used by layout_change to replace EX lock to CR lock.
2720 * \param lock A lock to convert
2721 * \param new_mode new lock mode
2723 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2725 #ifdef HAVE_SERVER_SUPPORT
2728 LASSERT(new_mode == LCK_COS || new_mode == LCK_TXN ||
2729 new_mode == LCK_CR);
2731 lock_res_and_lock(lock);
2733 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2734 unlock_res_and_lock(lock);
2736 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2737 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2741 ldlm_resource_unlink_lock(lock);
2743 * Remove the lock from pool as it will be added again in
2744 * ldlm_grant_lock() called below.
2746 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2748 /* Consider downgraded lock as a new lock and clear all states
2749 * related to a previous blocking AST processing.
2751 ldlm_clear_blocking_data(lock);
2753 lock->l_req_mode = new_mode;
2754 ldlm_grant_lock(lock, NULL);
2755 unlock_res_and_lock(lock);
2757 ldlm_reprocess_all(lock->l_resource,
2758 lock->l_policy_data.l_inodebits.bits);
2763 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2766 * Print lock with lock handle \a lockh description into debug log.
2768 * Used when printing all locks on a resource for debug purposes.
2770 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2772 struct ldlm_lock *lock;
2774 if (!((libcfs_debug | D_ERROR) & level))
2777 lock = ldlm_handle2lock(lockh);
2781 LDLM_DEBUG_LIMIT(level, lock, "###");
2783 LDLM_LOCK_PUT(lock);
2785 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2788 * Print lock information with custom message into debug log.
2791 void _ldlm_lock_debug(struct ldlm_lock *lock,
2792 struct libcfs_debug_msg_data *msgdata,
2793 const char *fmt, ...)
2796 struct obd_export *exp = lock->l_export;
2797 struct ldlm_resource *resource = NULL;
2798 struct va_format vaf;
2799 char *nid = "local";
2802 resource = rcu_dereference(lock->l_resource);
2803 if (resource && !refcount_inc_not_zero(&resource->lr_refcount))
2807 va_start(args, fmt);
2811 if (exp && exp->exp_connection) {
2812 nid = obd_export_nid2str(exp);
2813 } else if (exp && exp->exp_obd != NULL) {
2814 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2816 nid = obd_import_nid2str(imp);
2819 if (resource == NULL) {
2820 libcfs_debug_msg(msgdata,
2821 "%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",
2824 lock->l_handle.h_cookie,
2825 refcount_read(&lock->l_handle.h_ref),
2826 lock->l_readers, lock->l_writers,
2827 ldlm_lockname[lock->l_granted_mode],
2828 ldlm_lockname[lock->l_req_mode],
2830 lock->l_remote_handle.cookie,
2831 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2832 lock->l_pid, lock->l_callback_timestamp,
2838 switch (resource->lr_type) {
2840 libcfs_debug_msg(msgdata,
2841 "%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",
2843 ldlm_lock_to_ns_name(lock), lock,
2844 lock->l_handle.h_cookie,
2845 refcount_read(&lock->l_handle.h_ref),
2846 lock->l_readers, lock->l_writers,
2847 ldlm_lockname[lock->l_granted_mode],
2848 ldlm_lockname[lock->l_req_mode],
2850 refcount_read(&resource->lr_refcount),
2851 ldlm_typename[resource->lr_type],
2852 lock->l_policy_data.l_extent.start,
2853 lock->l_policy_data.l_extent.end,
2854 lock->l_req_extent.start, lock->l_req_extent.end,
2855 lock->l_req_extent.gid,
2857 lock->l_remote_handle.cookie,
2858 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2859 lock->l_pid, lock->l_callback_timestamp,
2864 libcfs_debug_msg(msgdata,
2865 "%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",
2867 ldlm_lock_to_ns_name(lock), lock,
2868 lock->l_handle.h_cookie,
2869 refcount_read(&lock->l_handle.h_ref),
2870 lock->l_readers, lock->l_writers,
2871 ldlm_lockname[lock->l_granted_mode],
2872 ldlm_lockname[lock->l_req_mode],
2874 refcount_read(&resource->lr_refcount),
2875 ldlm_typename[resource->lr_type],
2876 lock->l_policy_data.l_flock.pid,
2877 lock->l_policy_data.l_flock.start,
2878 lock->l_policy_data.l_flock.end,
2880 lock->l_remote_handle.cookie,
2881 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2882 lock->l_pid, lock->l_callback_timestamp);
2886 if (!lock->l_remote_handle.cookie)
2887 libcfs_debug_msg(msgdata,
2888 "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " bits %#llx/%#llx rrc: %d type: %s flags: %#llx pid: %u initiator: MDT%d\n",
2890 ldlm_lock_to_ns_name(lock),
2891 lock, lock->l_handle.h_cookie,
2892 refcount_read(&lock->l_handle.h_ref),
2893 lock->l_readers, lock->l_writers,
2894 ldlm_lockname[lock->l_granted_mode],
2895 ldlm_lockname[lock->l_req_mode],
2897 lock->l_policy_data.l_inodebits.bits,
2898 lock->l_policy_data.l_inodebits.try_bits,
2899 refcount_read(&resource->lr_refcount),
2900 ldlm_typename[resource->lr_type],
2901 lock->l_flags, lock->l_pid,
2902 lock->l_policy_data.l_inodebits.li_initiator_id);
2904 libcfs_debug_msg(msgdata,
2905 "%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",
2907 ldlm_lock_to_ns_name(lock),
2908 lock, lock->l_handle.h_cookie,
2909 refcount_read(&lock->l_handle.h_ref),
2910 lock->l_readers, lock->l_writers,
2911 ldlm_lockname[lock->l_granted_mode],
2912 ldlm_lockname[lock->l_req_mode],
2914 lock->l_policy_data.l_inodebits.bits,
2915 lock->l_policy_data.l_inodebits.try_bits,
2916 refcount_read(&resource->lr_refcount),
2917 ldlm_typename[resource->lr_type],
2918 lock->l_policy_data.l_inodebits.li_gid,
2920 lock->l_remote_handle.cookie,
2921 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2922 lock->l_pid, lock->l_callback_timestamp,
2927 libcfs_debug_msg(msgdata,
2928 "%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",
2930 ldlm_lock_to_ns_name(lock),
2931 lock, lock->l_handle.h_cookie,
2932 refcount_read(&lock->l_handle.h_ref),
2933 lock->l_readers, lock->l_writers,
2934 ldlm_lockname[lock->l_granted_mode],
2935 ldlm_lockname[lock->l_req_mode],
2937 refcount_read(&resource->lr_refcount),
2938 ldlm_typename[resource->lr_type],
2940 lock->l_remote_handle.cookie,
2941 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2942 lock->l_pid, lock->l_callback_timestamp,
2947 ldlm_resource_putref(resource);
2949 EXPORT_SYMBOL(_ldlm_lock_debug);