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);
174 * REFCOUNTED LOCK OBJECTS
179 * Get a reference on a lock.
181 * Lock refcounts, during creation:
182 * - one special one for allocation, dec'd only once in destroy
183 * - one for being a lock that's in-use
184 * - one for the addref associated with a new lock
186 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
188 refcount_inc(&lock->l_handle.h_ref);
191 EXPORT_SYMBOL(ldlm_lock_get);
193 static void lock_handle_free(struct rcu_head *rcu)
195 struct ldlm_lock *lock = container_of(rcu, struct ldlm_lock,
198 OBD_FREE_PRE(lock, sizeof(*lock), "slab-freed");
199 kmem_cache_free(ldlm_lock_slab, lock);
203 * Release lock reference.
205 * Also frees the lock if it was last reference.
207 void ldlm_lock_put(struct ldlm_lock *lock)
211 LASSERT(lock->l_resource != LP_POISON);
212 LASSERT(refcount_read(&lock->l_handle.h_ref) > 0);
213 if (refcount_dec_and_test(&lock->l_handle.h_ref)) {
214 struct ldlm_resource *res;
217 "final lock_put on destroyed lock, freeing it.");
219 res = lock->l_resource;
220 LASSERT(ldlm_is_destroyed(lock));
221 LASSERT(list_empty(&lock->l_exp_list));
222 LASSERT(list_empty(&lock->l_res_link));
223 LASSERT(list_empty(&lock->l_pending_chain));
225 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
227 lu_ref_del(&res->lr_reference, "lock", lock);
228 if (lock->l_export) {
229 class_export_lock_put(lock->l_export, lock);
230 lock->l_export = NULL;
233 if (lock->l_lvb_data != NULL)
234 OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
236 if (res->lr_type == LDLM_EXTENT || res->lr_type == LDLM_FLOCK) {
237 ldlm_interval_free(ldlm_interval_detach(lock));
238 } else if (res->lr_type == LDLM_IBITS) {
239 if (lock->l_ibits_node != NULL)
240 OBD_SLAB_FREE_PTR(lock->l_ibits_node,
241 ldlm_inodebits_slab);
243 ldlm_resource_putref(res);
244 lock->l_resource = NULL;
245 lu_ref_fini(&lock->l_reference);
246 call_rcu(&lock->l_handle.h_rcu, lock_handle_free);
251 EXPORT_SYMBOL(ldlm_lock_put);
254 * Removes LDLM lock \a lock from LRU. Assumes LRU is already locked.
256 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);
305 * Adds LDLM lock \a lock to namespace LRU. Assumes LRU is already locked.
307 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
309 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
311 lock->l_last_used = ktime_get();
312 LASSERT(list_empty(&lock->l_lru));
313 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
314 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
315 LASSERT(ns->ns_nr_unused >= 0);
320 * Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks
323 static void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
325 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
328 spin_lock(&ns->ns_lock);
329 ldlm_lock_add_to_lru_nolock(lock);
330 spin_unlock(&ns->ns_lock);
335 * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
336 * the LRU. Performs necessary LRU locking
338 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
340 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
343 if (ldlm_is_ns_srv(lock)) {
344 LASSERT(list_empty(&lock->l_lru));
349 spin_lock(&ns->ns_lock);
350 if (!list_empty(&lock->l_lru)) {
351 ldlm_lock_remove_from_lru_nolock(lock);
352 ldlm_lock_add_to_lru_nolock(lock);
354 spin_unlock(&ns->ns_lock);
359 * Helper to destroy a locked lock.
361 * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
362 * Must be called with l_lock and lr_lock held.
364 * Does not actually free the lock data, but rather marks the lock as
365 * destroyed by setting l_destroyed field in the lock to 1. Destroys a
366 * handle->lock association too, so that the lock can no longer be found
367 * and removes the lock from LRU list. Actual lock freeing occurs when
368 * last lock reference goes away.
370 * Original comment (of some historical value):
371 * This used to have a 'strict' flag, which recovery would use to mark an
372 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
373 * shall explain why it's gone: with the new hash table scheme, once you call
374 * ldlm_lock_destroy, you can never drop your final references on this lock.
375 * Because it's not in the hash table anymore. -phil
377 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
381 if (lock->l_readers || lock->l_writers) {
382 LDLM_ERROR(lock, "lock still has references");
386 if (!list_empty(&lock->l_res_link)) {
387 LDLM_ERROR(lock, "lock still on resource");
391 if (ldlm_is_destroyed(lock)) {
392 LASSERT(list_empty(&lock->l_lru));
396 ldlm_set_destroyed(lock);
397 wake_up(&lock->l_waitq);
399 if (lock->l_export && lock->l_export->exp_lock_hash) {
400 /* NB: it's safe to call cfs_hash_del() even lock isn't
401 * in exp_lock_hash. */
402 /* In the function below, .hs_keycmp resolves to
403 * ldlm_export_lock_keycmp() */
404 cfs_hash_del(lock->l_export->exp_lock_hash,
405 &lock->l_remote_handle, &lock->l_exp_hash);
408 ldlm_lock_remove_from_lru(lock);
409 class_handle_unhash(&lock->l_handle);
416 * Destroys a LDLM lock \a lock. Performs necessary locking first.
418 void ldlm_lock_destroy(struct ldlm_lock *lock)
422 lock_res_and_lock(lock);
423 first = ldlm_lock_destroy_internal(lock);
424 unlock_res_and_lock(lock);
426 /* drop reference from hashtable only for first destroy */
428 lu_ref_del(&lock->l_reference, "hash", lock);
429 LDLM_LOCK_RELEASE(lock);
435 * Destroys a LDLM lock \a lock that is already locked.
437 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
441 first = ldlm_lock_destroy_internal(lock);
442 /* drop reference from hashtable only for first destroy */
444 lu_ref_del(&lock->l_reference, "hash", lock);
445 LDLM_LOCK_RELEASE(lock);
450 static const char lock_handle_owner[] = "ldlm";
454 * Allocate and initialize new lock structure.
456 * usage: pass in a resource on which you have done ldlm_resource_get
457 * new lock will take over the refcount.
458 * returns: lock with refcount 2 - one for current caller and one for remote
460 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
462 struct ldlm_lock *lock;
465 if (resource == NULL)
468 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, GFP_NOFS);
472 RCU_INIT_POINTER(lock->l_resource, resource);
473 lu_ref_add(&resource->lr_reference, "lock", lock);
475 refcount_set(&lock->l_handle.h_ref, 2);
476 INIT_LIST_HEAD(&lock->l_res_link);
477 INIT_LIST_HEAD(&lock->l_lru);
478 INIT_LIST_HEAD(&lock->l_pending_chain);
479 INIT_LIST_HEAD(&lock->l_bl_ast);
480 INIT_LIST_HEAD(&lock->l_cp_ast);
481 INIT_LIST_HEAD(&lock->l_rk_ast);
482 init_waitqueue_head(&lock->l_waitq);
483 lock->l_blocking_lock = NULL;
484 INIT_LIST_HEAD(&lock->l_sl_mode);
485 INIT_LIST_HEAD(&lock->l_sl_policy);
486 INIT_HLIST_NODE(&lock->l_exp_hash);
487 INIT_HLIST_NODE(&lock->l_exp_flock_hash);
489 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
491 INIT_HLIST_NODE(&lock->l_handle.h_link);
492 class_handle_hash(&lock->l_handle, lock_handle_owner);
494 lu_ref_init(&lock->l_reference);
495 lu_ref_add(&lock->l_reference, "hash", lock);
496 lock->l_callback_timestamp = 0;
497 lock->l_activity = 0;
499 #if LUSTRE_TRACKS_LOCK_EXP_REFS
500 INIT_LIST_HEAD(&lock->l_exp_refs_link);
501 lock->l_exp_refs_nr = 0;
502 lock->l_exp_refs_target = NULL;
504 INIT_LIST_HEAD(&lock->l_exp_list);
510 * Moves LDLM lock \a lock to another resource.
511 * This is used on client when server returns some other lock than requested
512 * (typically as a result of intent operation)
514 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
515 const struct ldlm_res_id *new_resid)
517 struct ldlm_resource *oldres;
518 struct ldlm_resource *newres;
522 LASSERT(ns_is_client(ns));
524 oldres = lock_res_and_lock(lock);
525 if (memcmp(new_resid, &oldres->lr_name,
526 sizeof(oldres->lr_name)) == 0) {
528 unlock_res_and_lock(lock);
532 LASSERT(new_resid->name[0] != 0);
534 /* This function assumes that the lock isn't on any lists */
535 LASSERT(list_empty(&lock->l_res_link));
537 type = oldres->lr_type;
538 unlock_res_and_lock(lock);
540 newres = ldlm_resource_get(ns, new_resid, type, 1);
542 RETURN(PTR_ERR(newres));
544 lu_ref_add(&newres->lr_reference, "lock", lock);
546 * To flip the lock from the old to the new resource, oldres
547 * and newres have to be locked. Resource spin-locks are taken
548 * in the memory address order to avoid dead-locks.
549 * As this is the only circumstance where ->l_resource
550 * can change, and this cannot race with itself, it is safe
551 * to access lock->l_resource without being careful about locking.
553 oldres = lock->l_resource;
554 if (oldres < newres) {
556 lock_res_nested(newres, LRT_NEW);
559 lock_res_nested(oldres, LRT_NEW);
561 LASSERT(memcmp(new_resid, &oldres->lr_name,
562 sizeof oldres->lr_name) != 0);
563 rcu_assign_pointer(lock->l_resource, newres);
567 /* ...and the flowers are still standing! */
568 lu_ref_del(&oldres->lr_reference, "lock", lock);
569 ldlm_resource_putref(oldres);
574 /** \defgroup ldlm_handles LDLM HANDLES
575 * Ways to get hold of locks without any addresses.
580 * Fills in handle for LDLM lock \a lock into supplied \a lockh
581 * Does not take any references.
583 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
585 lockh->cookie = lock->l_handle.h_cookie;
587 EXPORT_SYMBOL(ldlm_lock2handle);
590 * Obtain a lock reference by handle.
592 * if \a flags: atomically get the lock and set the flags.
593 * Return NULL if flag already set
595 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
598 struct ldlm_lock *lock;
603 if (!lustre_handle_is_used(handle))
606 lock = class_handle2object(handle->cookie, lock_handle_owner);
610 if (lock->l_export != NULL && lock->l_export->exp_failed) {
611 CDEBUG(D_INFO, "lock export failed: lock %p, exp %p\n",
612 lock, lock->l_export);
617 /* It's unlikely but possible that someone marked the lock as
618 * destroyed after we did handle2object on it */
619 if ((flags == 0) && !ldlm_is_destroyed(lock)) {
620 lu_ref_add_atomic(&lock->l_reference, "handle", lock);
624 lock_res_and_lock(lock);
626 LASSERT(lock->l_resource != NULL);
628 lu_ref_add_atomic(&lock->l_reference, "handle", lock);
629 if (unlikely(ldlm_is_destroyed(lock))) {
630 unlock_res_and_lock(lock);
631 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
636 /* If we're setting flags, make sure none of them are already set. */
638 if ((lock->l_flags & flags) != 0) {
639 unlock_res_and_lock(lock);
644 lock->l_flags |= flags;
647 unlock_res_and_lock(lock);
650 EXPORT_SYMBOL(__ldlm_handle2lock);
651 /** @} ldlm_handles */
654 * Fill in "on the wire" representation for given LDLM lock into supplied
655 * lock descriptor \a desc structure.
657 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
659 ldlm_res2desc(lock->l_resource, &desc->l_resource);
660 desc->l_req_mode = lock->l_req_mode;
661 desc->l_granted_mode = lock->l_granted_mode;
662 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
663 &lock->l_policy_data,
664 &desc->l_policy_data);
668 * Add a lock to list of conflicting locks to send AST to.
670 * Only add if we have not sent a blocking AST to the lock yet.
672 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
673 struct list_head *work_list)
675 if (!ldlm_is_ast_sent(lock)) {
676 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
677 ldlm_set_ast_sent(lock);
678 /* If the enqueuing client said so, tell the AST recipient to
679 * discard dirty data, rather than writing back. */
680 if (ldlm_is_ast_discard_data(new))
681 ldlm_set_discard_data(lock);
683 /* Lock can be converted from a blocking state back to granted
684 * after lock convert or COS downgrade but still be in an
685 * older bl_list because it is controlled only by
686 * ldlm_work_bl_ast_lock(), let it be processed there.
688 if (list_empty(&lock->l_bl_ast)) {
689 list_add(&lock->l_bl_ast, work_list);
692 LASSERT(lock->l_blocking_lock == NULL);
693 lock->l_blocking_lock = LDLM_LOCK_GET(new);
698 * Add a lock to list of just granted locks to send completion AST to.
700 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
701 struct list_head *work_list)
703 if (!ldlm_is_cp_reqd(lock)) {
704 ldlm_set_cp_reqd(lock);
705 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
706 LASSERT(list_empty(&lock->l_cp_ast));
707 list_add(&lock->l_cp_ast, work_list);
713 * Aggregator function to add AST work items into a list. Determines
714 * what sort of an AST work needs to be done and calls the proper
716 * Must be called with lr_lock held.
718 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
719 struct list_head *work_list)
722 check_res_locked(lock->l_resource);
724 ldlm_add_bl_work_item(lock, new, work_list);
726 ldlm_add_cp_work_item(lock, work_list);
731 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
732 * r/w reference type is determined by \a mode
733 * Calls ldlm_lock_addref_internal.
735 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
737 struct ldlm_lock *lock;
739 lock = ldlm_handle2lock(lockh);
740 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
741 ldlm_lock_addref_internal(lock, mode);
744 EXPORT_SYMBOL(ldlm_lock_addref);
748 * Add specified reader/writer reference to LDLM lock \a lock.
749 * r/w reference type is determined by \a mode
750 * Removes lock from LRU if it is there.
751 * Assumes the LDLM lock is already locked.
753 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
756 ldlm_lock_remove_from_lru(lock);
757 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
759 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
761 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS | LCK_TXN)) {
763 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
766 lu_ref_add_atomic(&lock->l_reference, "user", lock);
767 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
771 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
772 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
774 * \retval 0 success, lock was addref-ed
776 * \retval -EAGAIN lock is being canceled.
778 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
780 struct ldlm_lock *lock;
784 lock = ldlm_handle2lock(lockh);
786 lock_res_and_lock(lock);
787 if (lock->l_readers != 0 || lock->l_writers != 0 ||
788 !ldlm_is_cbpending(lock)) {
789 ldlm_lock_addref_internal_nolock(lock, mode);
792 unlock_res_and_lock(lock);
797 EXPORT_SYMBOL(ldlm_lock_addref_try);
800 * Add specified reader/writer reference to LDLM lock \a lock.
801 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
802 * Only called for local locks.
804 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
806 lock_res_and_lock(lock);
807 ldlm_lock_addref_internal_nolock(lock, mode);
808 unlock_res_and_lock(lock);
812 * Removes reader/writer reference for LDLM lock \a lock.
813 * Assumes LDLM lock is already locked.
814 * only called in ldlm_flock_destroy and for local locks.
815 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
816 * that cannot be placed in LRU.
818 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
821 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
822 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
823 LASSERT(lock->l_readers > 0);
824 lu_ref_del(&lock->l_reference, "reader", lock);
827 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS | LCK_TXN)) {
828 LASSERT(lock->l_writers > 0);
829 lu_ref_del(&lock->l_reference, "writer", lock);
833 lu_ref_del(&lock->l_reference, "user", lock);
834 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
838 * Removes reader/writer reference for LDLM lock \a lock.
839 * Locks LDLM lock first.
840 * If the lock is determined to be client lock on a client and r/w refcount
841 * drops to zero and the lock is not blocked, the lock is added to LRU lock
843 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
845 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
847 struct ldlm_namespace *ns;
851 lock_res_and_lock(lock);
853 ns = ldlm_lock_to_ns(lock);
855 ldlm_lock_decref_internal_nolock(lock, mode);
857 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
858 !lock->l_readers && !lock->l_writers) {
859 /* If this is a local lock on a server namespace and this was
860 * the last reference, cancel the lock.
862 * Group locks are special:
863 * They must not go in LRU, but they are not called back
864 * like non-group locks, instead they are manually released.
865 * They have an l_writers reference which they keep until
866 * they are manually released, so we remove them when they have
867 * no more reader or writer references. - LU-6368 */
868 ldlm_set_cbpending(lock);
871 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
872 unsigned int mask = D_DLMTRACE;
874 /* If we received a blocked AST and this was the last reference,
875 * run the callback. */
876 if (ldlm_is_ns_srv(lock) && lock->l_export)
878 LDLM_DEBUG_LIMIT(mask, lock,
879 "final decref done on %sCBPENDING lock",
880 mask & D_WARNING ? "non-local " : "");
882 LDLM_LOCK_GET(lock); /* dropped by bl thread */
883 ldlm_lock_remove_from_lru(lock);
884 unlock_res_and_lock(lock);
886 if (ldlm_is_fail_loc(lock))
887 CFS_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
889 if (ldlm_is_atomic_cb(lock) || ldlm_is_local(lock) ||
890 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
891 ldlm_handle_bl_callback(ns, NULL, lock);
892 } else if (ns_is_client(ns) &&
893 !lock->l_readers && !lock->l_writers &&
894 !ldlm_is_no_lru(lock) &&
895 !ldlm_is_bl_ast(lock) &&
896 !ldlm_is_converting(lock)) {
898 /* If this is a client-side namespace and this was the last
899 * reference, put it on the LRU.
901 ldlm_lock_add_to_lru(lock);
902 unlock_res_and_lock(lock);
903 LDLM_DEBUG(lock, "add lock into lru list");
905 if (ldlm_is_fail_loc(lock))
906 CFS_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
908 ldlm_pool_recalc(&ns->ns_pool, true);
910 LDLM_DEBUG(lock, "do not add lock into lru list");
911 unlock_res_and_lock(lock);
918 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
920 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
922 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
923 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
924 ldlm_lock_decref_internal(lock, mode);
927 EXPORT_SYMBOL(ldlm_lock_decref);
930 * Decrease reader/writer refcount for LDLM lock with handle
931 * \a lockh and mark it for subsequent cancellation once r/w refcount
932 * drops to zero instead of putting into LRU.
935 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
938 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
941 LASSERT(lock != NULL);
943 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
944 lock_res_and_lock(lock);
945 ldlm_set_cbpending(lock);
946 unlock_res_and_lock(lock);
947 ldlm_lock_decref_internal(lock, mode);
950 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
952 struct sl_insert_point {
953 struct list_head *res_link;
954 struct list_head *mode_link;
955 struct list_head *policy_link;
959 * Finds a position to insert the new lock into granted lock list.
961 * Used for locks eligible for skiplist optimization.
964 * queue [input]: the granted list where search acts on;
965 * req [input]: the lock whose position to be located;
966 * prev [output]: positions within 3 lists to insert @req to
970 * - ldlm_grant_lock_with_skiplist
972 static void search_granted_lock(struct list_head *queue,
973 struct ldlm_lock *req,
974 struct sl_insert_point *prev)
976 struct ldlm_lock *lock, *mode_end, *policy_end;
979 list_for_each_entry(lock, queue, l_res_link) {
980 mode_end = list_entry(lock->l_sl_mode.prev,
981 struct ldlm_lock, l_sl_mode);
983 if (lock->l_req_mode != req->l_req_mode) {
984 /* jump to last lock of mode group */
989 /* suitable mode group is found */
990 if (lock->l_resource->lr_type == LDLM_PLAIN) {
991 /* insert point is last lock of the mode group */
992 prev->res_link = &mode_end->l_res_link;
993 prev->mode_link = &mode_end->l_sl_mode;
994 prev->policy_link = &req->l_sl_policy;
997 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
1000 list_entry(lock->l_sl_policy.prev,
1004 if (lock->l_policy_data.l_inodebits.bits ==
1005 req->l_policy_data.l_inodebits.bits) {
1006 /* insert point is last lock of
1007 * the policy group */
1009 &policy_end->l_res_link;
1011 &policy_end->l_sl_mode;
1013 &policy_end->l_sl_policy;
1018 if (policy_end == mode_end)
1019 /* done with mode group */
1022 /* go to next policy group within mode group */
1023 lock = list_next_entry(policy_end, l_res_link);
1024 } /* loop over policy groups within the mode group */
1026 /* insert point is last lock of the mode group,
1027 * new policy group is started */
1028 prev->res_link = &mode_end->l_res_link;
1029 prev->mode_link = &mode_end->l_sl_mode;
1030 prev->policy_link = &req->l_sl_policy;
1034 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1039 /* insert point is last lock on the queue,
1040 * new mode group and new policy group are started */
1041 prev->res_link = queue->prev;
1042 prev->mode_link = &req->l_sl_mode;
1043 prev->policy_link = &req->l_sl_policy;
1048 * Add a lock into resource granted list after a position described by
1051 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1052 struct sl_insert_point *prev)
1054 struct ldlm_resource *res = lock->l_resource;
1057 check_res_locked(res);
1059 ldlm_resource_dump(D_INFO, res);
1060 LDLM_DEBUG(lock, "About to add lock:");
1062 if (ldlm_is_destroyed(lock)) {
1063 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1067 LASSERT(list_empty(&lock->l_res_link));
1068 LASSERT(list_empty(&lock->l_sl_mode));
1069 LASSERT(list_empty(&lock->l_sl_policy));
1072 * lock->link == prev->link means lock is first starting the group.
1073 * Don't re-add to itself to suppress kernel warnings.
1075 if (&lock->l_res_link != prev->res_link)
1076 list_add(&lock->l_res_link, prev->res_link);
1077 if (&lock->l_sl_mode != prev->mode_link)
1078 list_add(&lock->l_sl_mode, prev->mode_link);
1079 if (&lock->l_sl_policy != prev->policy_link)
1080 list_add(&lock->l_sl_policy, prev->policy_link);
1086 * Add a lock to granted list on a resource maintaining skiplist
1089 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1091 struct sl_insert_point prev;
1093 LASSERT(ldlm_is_granted(lock));
1095 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1096 ldlm_granted_list_add_lock(lock, &prev);
1100 * Perform lock granting bookkeeping.
1102 * Includes putting the lock into granted list and updating lock mode.
1104 * - ldlm_lock_enqueue
1105 * - ldlm_reprocess_queue
1107 * must be called with lr_lock held
1109 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1111 struct ldlm_resource *res = lock->l_resource;
1114 check_res_locked(res);
1116 lock->l_granted_mode = lock->l_req_mode;
1118 if (work_list && lock->l_completion_ast != NULL)
1119 ldlm_add_ast_work_item(lock, NULL, work_list);
1121 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1122 ldlm_grant_lock_with_skiplist(lock);
1123 else if (res->lr_type == LDLM_EXTENT)
1124 ldlm_extent_add_lock(res, lock);
1125 else if (res->lr_type == LDLM_FLOCK) {
1126 /* We should not add locks to granted list in the following
1128 * - this is an UNLOCK but not a real lock;
1129 * - this is a TEST lock;
1130 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1131 * - this is a deadlock (flock cannot be granted) */
1132 if (lock->l_req_mode == 0 ||
1133 lock->l_req_mode == LCK_NL ||
1134 ldlm_is_test_lock(lock) ||
1135 ldlm_is_flock_deadlock(lock))
1137 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1142 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1147 * Check if the given @lock meets the criteria for a match.
1148 * A reference on the lock is taken if matched.
1150 * @lock test-against this lock
1153 * RETURN returns true if @lock matches @data, false otherwise
1155 static bool lock_matches(struct ldlm_lock *lock, struct ldlm_match_data *data)
1157 union ldlm_policy_data *lpol = &lock->l_policy_data;
1158 enum ldlm_mode match = LCK_MINMODE;
1160 if (lock == data->lmd_old)
1163 /* Check if this lock can be matched.
1164 * Used by LU-2919(exclusive open) for open lease lock */
1165 if (ldlm_is_excl(lock))
1168 /* llite sometimes wants to match locks that will be
1169 * canceled when their users drop, but we allow it to match
1170 * if it passes in CBPENDING and the lock still has users.
1171 * this is generally only going to be used by children
1172 * whose parents already hold a lock so forward progress
1173 * can still happen. */
1174 if (ldlm_is_cbpending(lock) &&
1175 !(data->lmd_flags & LDLM_FL_CBPENDING) &&
1176 !(data->lmd_match & LDLM_MATCH_GROUP))
1179 if (!(data->lmd_match & (LDLM_MATCH_UNREF | LDLM_MATCH_GROUP)) &&
1180 ldlm_is_cbpending(lock) &&
1181 lock->l_readers == 0 && lock->l_writers == 0)
1184 if (!(lock->l_req_mode & *data->lmd_mode))
1187 /* When we search for ast_data, we are not doing a traditional match,
1188 * so we don't worry about IBITS or extent matching.
1190 if (data->lmd_match & (LDLM_MATCH_AST | LDLM_MATCH_AST_ANY)) {
1191 if (!lock->l_ast_data)
1194 if (data->lmd_match & LDLM_MATCH_AST_ANY)
1198 match = lock->l_req_mode;
1200 switch (lock->l_resource->lr_type) {
1202 if (!(data->lmd_match & LDLM_MATCH_RIGHT) &&
1203 (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1204 lpol->l_extent.end < data->lmd_policy->l_extent.end))
1207 if (unlikely(match == LCK_GROUP) &&
1208 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1209 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1213 /* We match if we have existing lock with same or wider set
1215 if ((lpol->l_inodebits.bits &
1216 data->lmd_policy->l_inodebits.bits) !=
1217 data->lmd_policy->l_inodebits.bits)
1220 if (unlikely(match == LCK_GROUP) &&
1221 data->lmd_policy->l_inodebits.li_gid != LDLM_GID_ANY &&
1222 lpol->l_inodebits.li_gid !=
1223 data->lmd_policy->l_inodebits.li_gid)
1230 /* We match if we have existing lock with same or wider set
1232 if (!(data->lmd_match & LDLM_MATCH_UNREF) && LDLM_HAVE_MASK(lock, GONE))
1235 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1238 /* Filter locks by skipping flags */
1239 if (data->lmd_skip_flags & lock->l_flags)
1244 * In case the lock is a CBPENDING grouplock, just pin it and return,
1245 * we need to wait until it gets to DESTROYED.
1247 if ((data->lmd_flags & LDLM_FL_TEST_LOCK) ||
1248 (ldlm_is_cbpending(lock) && (data->lmd_match & LDLM_MATCH_GROUP))) {
1249 LDLM_LOCK_GET(lock);
1250 ldlm_lock_touch_in_lru(lock);
1252 ldlm_lock_addref_internal_nolock(lock, match);
1255 *data->lmd_mode = match;
1256 data->lmd_lock = lock;
1261 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1263 struct ldlm_interval *node = to_ldlm_interval(in);
1264 struct ldlm_match_data *data = args;
1265 struct ldlm_lock *lock;
1267 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1268 if (lock_matches(lock, data))
1269 return INTERVAL_ITER_STOP;
1271 return INTERVAL_ITER_CONT;
1275 * Search for a lock with given parameters in interval trees.
1277 * \param res search for a lock in this resource
1278 * \param data parameters
1280 * \retval a referenced lock or NULL.
1282 struct ldlm_lock *search_itree(struct ldlm_resource *res,
1283 struct ldlm_match_data *data)
1285 struct interval_node_extent ext = {
1286 .start = data->lmd_policy->l_extent.start,
1287 .end = data->lmd_policy->l_extent.end
1291 data->lmd_lock = NULL;
1293 if (data->lmd_match & LDLM_MATCH_RIGHT)
1294 ext.end = OBD_OBJECT_EOF;
1296 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1297 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1299 if (tree->lit_root == NULL)
1302 if (!(tree->lit_mode & *data->lmd_mode))
1305 interval_search(tree->lit_root, &ext,
1306 itree_overlap_cb, data);
1308 return data->lmd_lock;
1313 EXPORT_SYMBOL(search_itree);
1317 * Search for a lock with given properties in a queue.
1319 * \param queue search for a lock in this queue
1320 * \param data parameters
1322 * \retval a referenced lock or NULL.
1324 static struct ldlm_lock *search_queue(struct list_head *queue,
1325 struct ldlm_match_data *data)
1327 struct ldlm_lock *lock;
1329 data->lmd_lock = NULL;
1331 list_for_each_entry(lock, queue, l_res_link)
1332 if (lock_matches(lock, data))
1333 return data->lmd_lock;
1338 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1340 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1341 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1342 wake_up(&lock->l_waitq);
1345 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1347 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1349 lock_res_and_lock(lock);
1350 ldlm_lock_fail_match_locked(lock);
1351 unlock_res_and_lock(lock);
1355 * Mark lock as "matchable" by OST.
1357 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1359 * Assumes LDLM lock is already locked.
1361 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1363 ldlm_set_lvb_ready(lock);
1364 wake_up(&lock->l_waitq);
1366 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1369 * Mark lock as "matchable" by OST.
1370 * Locks the lock and then \see ldlm_lock_allow_match_locked
1372 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1374 lock_res_and_lock(lock);
1375 ldlm_lock_allow_match_locked(lock);
1376 unlock_res_and_lock(lock);
1378 EXPORT_SYMBOL(ldlm_lock_allow_match);
1381 * Attempt to find a lock with specified properties.
1383 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1386 * Can be called in two ways:
1388 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1389 * for a duplicate of.
1391 * Otherwise, all of the fields must be filled in, to match against.
1393 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1394 * server (ie, connh is NULL)
1395 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1396 * list will be considered
1397 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1398 * to be canceled can still be matched as long as they still have reader
1399 * or writer refernces
1400 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1401 * just tell us if we would have matched.
1403 * \retval 1 if it finds an already-existing lock that is compatible; in this
1404 * case, lockh is filled in with a addref()ed lock
1406 * We also check security context, and if that fails we simply return 0 (to
1407 * keep caller code unchanged), the context failure will be discovered by
1408 * caller sometime later.
1410 enum ldlm_mode ldlm_lock_match_with_skip(struct ldlm_namespace *ns,
1411 __u64 flags, __u64 skip_flags,
1412 const struct ldlm_res_id *res_id,
1413 enum ldlm_type type,
1414 union ldlm_policy_data *policy,
1415 enum ldlm_mode mode,
1416 struct lustre_handle *lockh,
1417 enum ldlm_match_flags match_flags)
1419 struct ldlm_match_data data = {
1423 .lmd_policy = policy,
1425 .lmd_skip_flags = skip_flags,
1426 .lmd_match = match_flags,
1428 struct ldlm_resource *res;
1429 struct ldlm_lock *lock;
1430 struct ldlm_lock *group_lock;
1436 data.lmd_old = ldlm_handle2lock(lockh);
1437 LASSERT(data.lmd_old != NULL);
1439 ns = ldlm_lock_to_ns(data.lmd_old);
1440 res_id = &data.lmd_old->l_resource->lr_name;
1441 type = data.lmd_old->l_resource->lr_type;
1442 *data.lmd_mode = data.lmd_old->l_req_mode;
1445 res = ldlm_resource_get(ns, res_id, type, 0);
1447 LASSERT(data.lmd_old == NULL);
1453 LDLM_RESOURCE_ADDREF(res);
1455 if (res->lr_type == LDLM_EXTENT)
1456 lock = search_itree(res, &data);
1458 lock = search_queue(&res->lr_granted, &data);
1459 if (!lock && !(flags & LDLM_FL_BLOCK_GRANTED))
1460 lock = search_queue(&res->lr_waiting, &data);
1461 matched = lock ? mode : 0;
1463 if (lock && ldlm_is_cbpending(lock) &&
1464 (data.lmd_match & LDLM_MATCH_GROUP))
1467 LDLM_RESOURCE_DELREF(res);
1470 l_wait_event_abortable(group_lock->l_waitq,
1471 ldlm_is_destroyed(lock));
1472 LDLM_LOCK_RELEASE(lock);
1475 ldlm_resource_putref(res);
1478 ldlm_lock2handle(lock, lockh);
1479 if ((flags & LDLM_FL_LVB_READY) &&
1480 (!ldlm_is_lvb_ready(lock))) {
1481 __u64 wait_flags = LDLM_FL_LVB_READY |
1482 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1484 if (lock->l_completion_ast) {
1485 int err = lock->l_completion_ast(lock,
1486 LDLM_FL_WAIT_NOREPROC,
1489 GOTO(out_fail_match, matched = 0);
1492 wait_event_idle_timeout(
1494 lock->l_flags & wait_flags,
1495 cfs_time_seconds(obd_timeout));
1497 if (!ldlm_is_lvb_ready(lock))
1498 GOTO(out_fail_match, matched = 0);
1501 /* check user's security context */
1502 if (lock->l_conn_export &&
1503 sptlrpc_import_check_ctx(
1504 class_exp2cliimp(lock->l_conn_export)))
1505 GOTO(out_fail_match, matched = 0);
1507 LDLM_DEBUG(lock, "matched (%llu %llu)",
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);
1514 if (flags & LDLM_FL_TEST_LOCK)
1515 LDLM_LOCK_RELEASE(lock);
1517 ldlm_lock_decref_internal(lock, mode);
1520 /* less verbose for test-only */
1521 if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
1522 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1523 "%llu/%llu (%llu %llu)", ns,
1524 type, mode, res_id->name[0], res_id->name[1],
1525 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1526 res_id->name[2] : policy->l_extent.start,
1527 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1528 res_id->name[3] : policy->l_extent.end);
1530 if (data.lmd_old != NULL)
1531 LDLM_LOCK_PUT(data.lmd_old);
1535 EXPORT_SYMBOL(ldlm_lock_match_with_skip);
1537 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1540 struct ldlm_lock *lock;
1541 enum ldlm_mode mode = 0;
1544 lock = ldlm_handle2lock(lockh);
1546 lock_res_and_lock(lock);
1547 if (LDLM_HAVE_MASK(lock, GONE))
1550 if (ldlm_is_cbpending(lock) &&
1551 lock->l_readers == 0 && lock->l_writers == 0)
1555 *bits = lock->l_policy_data.l_inodebits.bits;
1556 mode = lock->l_granted_mode;
1557 ldlm_lock_addref_internal_nolock(lock, mode);
1564 unlock_res_and_lock(lock);
1565 LDLM_LOCK_PUT(lock);
1569 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1571 /** The caller must guarantee that the buffer is large enough. */
1572 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1573 enum req_location loc, void *data, int size)
1578 LASSERT(data != NULL);
1581 switch (lock->l_lvb_type) {
1583 if (size == sizeof(struct ost_lvb)) {
1584 if (loc == RCL_CLIENT)
1585 lvb = req_capsule_client_swab_get(pill,
1587 lustre_swab_ost_lvb);
1589 lvb = req_capsule_server_swab_get(pill,
1591 lustre_swab_ost_lvb);
1592 if (unlikely(lvb == NULL)) {
1593 LDLM_ERROR(lock, "no LVB");
1597 memcpy(data, lvb, size);
1598 } else if (size == sizeof(struct ost_lvb_v1)) {
1599 struct ost_lvb *olvb = data;
1601 if (loc == RCL_CLIENT)
1602 lvb = req_capsule_client_swab_get(pill,
1604 lustre_swab_ost_lvb_v1);
1606 lvb = req_capsule_server_sized_swab_get(pill,
1608 lustre_swab_ost_lvb_v1);
1609 if (unlikely(lvb == NULL)) {
1610 LDLM_ERROR(lock, "no LVB");
1614 memcpy(data, lvb, size);
1615 olvb->lvb_mtime_ns = 0;
1616 olvb->lvb_atime_ns = 0;
1617 olvb->lvb_ctime_ns = 0;
1619 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1625 if (size == sizeof(struct lquota_lvb)) {
1626 if (loc == RCL_CLIENT)
1627 lvb = req_capsule_client_swab_get(pill,
1629 lustre_swab_lquota_lvb);
1631 lvb = req_capsule_server_swab_get(pill,
1633 lustre_swab_lquota_lvb);
1634 if (unlikely(lvb == NULL)) {
1635 LDLM_ERROR(lock, "no LVB");
1639 memcpy(data, lvb, size);
1641 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1650 if (loc == RCL_CLIENT)
1651 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1653 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1654 if (unlikely(lvb == NULL)) {
1655 LDLM_ERROR(lock, "no LVB");
1659 memcpy(data, lvb, size);
1662 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1671 * Create and fill in new LDLM lock with specified properties.
1672 * Returns a referenced lock
1674 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1675 const struct ldlm_res_id *res_id,
1676 enum ldlm_type type,
1677 enum ldlm_mode mode,
1678 const struct ldlm_callback_suite *cbs,
1679 void *data, __u32 lvb_len,
1680 enum lvb_type lvb_type)
1682 struct ldlm_lock *lock;
1683 struct ldlm_resource *res;
1687 res = ldlm_resource_get(ns, res_id, type, 1);
1689 RETURN(ERR_CAST(res));
1691 lock = ldlm_lock_new(res);
1693 ldlm_resource_putref(res);
1694 RETURN(ERR_PTR(-ENOMEM));
1697 lock->l_req_mode = mode;
1698 lock->l_ast_data = data;
1699 lock->l_pid = current->pid;
1700 if (ns_is_server(ns))
1701 ldlm_set_ns_srv(lock);
1703 lock->l_blocking_ast = cbs->lcs_blocking;
1704 lock->l_completion_ast = cbs->lcs_completion;
1705 lock->l_glimpse_ast = cbs->lcs_glimpse;
1711 rc = ldlm_extent_alloc_lock(lock);
1714 rc = ldlm_inodebits_alloc_lock(lock);
1723 lock->l_lvb_len = lvb_len;
1724 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1725 if (lock->l_lvb_data == NULL)
1726 GOTO(out, rc = -ENOMEM);
1729 lock->l_lvb_type = lvb_type;
1730 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1731 GOTO(out, rc = -ENOENT);
1736 ldlm_lock_destroy(lock);
1737 LDLM_LOCK_RELEASE(lock);
1738 RETURN(ERR_PTR(rc));
1741 #ifdef HAVE_SERVER_SUPPORT
1742 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1745 struct ldlm_resource *res = lock->l_resource;
1746 enum ldlm_error rc = ELDLM_OK;
1747 LIST_HEAD(rpc_list);
1748 ldlm_processing_policy policy;
1752 policy = ldlm_get_processing_policy(res);
1753 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1754 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1755 res->lr_type != LDLM_FLOCK)
1756 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1758 if (!list_empty(&rpc_list))
1759 ldlm_discard_bl_list(&rpc_list);
1766 * Enqueue (request) a lock.
1768 * Does not block. As a result of enqueue the lock would be put
1769 * into granted or waiting list.
1771 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1772 * set, skip all the enqueueing and delegate lock processing to intent policy
1775 enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
1776 struct ldlm_namespace *ns,
1777 struct ldlm_lock **lockp,
1778 void *cookie, __u64 *flags)
1780 struct ldlm_lock *lock = *lockp;
1781 struct ldlm_resource *res;
1782 int local = ns_is_client(ns);
1783 enum ldlm_error rc = ELDLM_OK;
1784 struct ldlm_interval *node = NULL;
1785 #ifdef HAVE_SERVER_SUPPORT
1786 bool reconstruct = false;
1790 /* policies are not executed on the client or during replay */
1791 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1792 && !local && ns->ns_policy) {
1793 rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
1795 if (rc == ELDLM_LOCK_REPLACED) {
1796 /* The lock that was returned has already been granted,
1797 * and placed into lockp. If it's not the same as the
1798 * one we passed in, then destroy the old one and our
1799 * work here is done.
1801 if (lock != *lockp) {
1802 ldlm_lock_destroy(lock);
1803 LDLM_LOCK_RELEASE(lock);
1805 *flags |= LDLM_FL_LOCK_CHANGED;
1807 } else if (rc != ELDLM_OK &&
1808 ldlm_is_granted(lock)) {
1809 LASSERT(*flags & LDLM_FL_RESENT);
1810 /* It may happen that ns_policy returns an error in
1811 * resend case, object may be unlinked or just some
1812 * error occurs. It is unclear if lock reached the
1813 * client in the original reply, just leave the lock on
1814 * server, not returning it again to client. Due to
1815 * LU-6529, the server will not OOM.
1818 } else if (rc != ELDLM_OK ||
1819 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1820 ldlm_lock_destroy(lock);
1825 if (*flags & LDLM_FL_RESENT) {
1826 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1827 * Set LOCK_CHANGED always.
1828 * Check if the lock is granted for BLOCK_GRANTED.
1829 * Take NO_TIMEOUT from the lock as it is inherited through
1830 * LDLM_FL_INHERIT_MASK */
1831 *flags |= LDLM_FL_LOCK_CHANGED;
1832 if (!ldlm_is_granted(lock))
1833 *flags |= LDLM_FL_BLOCK_GRANTED;
1834 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1838 #ifdef HAVE_SERVER_SUPPORT
1839 /* For a replaying lock, it might be already in granted list. So
1840 * unlinking the lock will cause the interval node to be freed, we
1841 * have to allocate the interval node early otherwise we can't regrant
1842 * this lock in the future. - jay
1844 * The only time the ldlm_resource changes for the ldlm_lock is when
1845 * ldlm_lock_change_resource() is called and that only happens for
1846 * the Lustre client case.
1848 if (!local && (*flags & LDLM_FL_REPLAY) &&
1849 lock->l_resource->lr_type == LDLM_EXTENT)
1850 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1852 reconstruct = !local && lock->l_resource->lr_type == LDLM_FLOCK &&
1853 !(*flags & LDLM_FL_TEST_LOCK);
1855 rc = req_can_reconstruct(cookie, NULL);
1863 if (!local && lock->l_resource->lr_type == LDLM_FLOCK) {
1864 struct ldlm_flock_node *fn = &lock->l_resource->lr_flock_node;
1865 if (lock->l_req_mode == LCK_NL) {
1866 atomic_inc(&fn->lfn_unlock_pending);
1867 res = lock_res_and_lock(lock);
1868 atomic_dec(&fn->lfn_unlock_pending);
1870 res = lock_res_and_lock(lock);
1872 while (atomic_read(&fn->lfn_unlock_pending)) {
1873 unlock_res_and_lock(lock);
1875 lock_res_and_lock(lock);
1881 res = lock_res_and_lock(lock);
1883 if (local && ldlm_is_granted(lock)) {
1884 /* The server returned a blocked lock, but it was granted
1885 * before we got a chance to actually enqueue it. We don't
1886 * need to do anything else. */
1887 *flags &= ~LDLM_FL_BLOCKED_MASK;
1888 GOTO(out, rc = ELDLM_OK);
1891 ldlm_resource_unlink_lock(lock);
1892 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1894 ldlm_lock_destroy_nolock(lock);
1895 GOTO(out, rc = -ENOMEM);
1898 INIT_LIST_HEAD(&node->li_group);
1899 ldlm_interval_attach(node, lock);
1903 /* Some flags from the enqueue want to make it into the AST, via the
1904 * lock's l_flags. */
1905 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1906 ldlm_set_ast_discard_data(lock);
1907 if (*flags & LDLM_FL_TEST_LOCK)
1908 ldlm_set_test_lock(lock);
1910 /* This distinction between local lock trees is very important; a client
1911 * namespace only has information about locks taken by that client, and
1912 * thus doesn't have enough information to decide for itself if it can
1913 * be granted (below). In this case, we do exactly what the server
1914 * tells us to do, as dictated by the 'flags'.
1916 * We do exactly the same thing during recovery, when the server is
1917 * more or less trusting the clients not to lie.
1919 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1922 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1923 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1925 ldlm_grant_lock(lock, NULL);
1926 GOTO(out, rc = ELDLM_OK);
1927 #ifdef HAVE_SERVER_SUPPORT
1928 } else if (*flags & LDLM_FL_REPLAY) {
1929 if (*flags & LDLM_FL_BLOCK_WAIT) {
1930 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1931 GOTO(out, rc = ELDLM_OK);
1932 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1933 ldlm_grant_lock(lock, NULL);
1934 GOTO(out, rc = ELDLM_OK);
1936 /* If no flags, fall through to normal enqueue path. */
1939 rc = ldlm_lock_enqueue_helper(lock, flags);
1943 CERROR("This is client-side-only module, cannot handle "
1944 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1950 unlock_res_and_lock(lock);
1952 #ifdef HAVE_SERVER_SUPPORT
1954 struct ptlrpc_request *req = cookie;
1956 tgt_mk_reply_data(NULL, NULL,
1957 &req->rq_export->exp_target_data,
1958 req, 0, NULL, false, 0);
1962 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1966 #ifdef HAVE_SERVER_SUPPORT
1968 * Iterate through all waiting locks on a given resource queue and attempt to
1971 * Must be called with resource lock held.
1973 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1974 struct list_head *work_list,
1975 enum ldlm_process_intention intention, __u64 hint)
1977 struct list_head *tmp, *pos;
1978 ldlm_processing_policy policy;
1980 int rc = LDLM_ITER_CONTINUE;
1981 enum ldlm_error err;
1982 LIST_HEAD(bl_ast_list);
1986 check_res_locked(res);
1988 policy = ldlm_get_processing_policy(res);
1990 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1991 intention == LDLM_PROCESS_RECOVERY);
1994 list_for_each_safe(tmp, pos, queue) {
1995 struct ldlm_lock *pending;
1996 LIST_HEAD(rpc_list);
1998 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
2000 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
2003 rc = policy(pending, &flags, intention, &err, &rpc_list);
2004 if (pending->l_granted_mode == pending->l_req_mode ||
2005 res->lr_type == LDLM_FLOCK) {
2006 list_splice(&rpc_list, work_list);
2008 list_splice(&rpc_list, &bl_ast_list);
2011 * When this is called from recovery done, we always want
2012 * to scan the whole list no matter what 'rc' is returned.
2014 if (rc != LDLM_ITER_CONTINUE &&
2015 intention == LDLM_PROCESS_RESCAN)
2019 if (!list_empty(&bl_ast_list)) {
2022 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
2026 if (rc == -ERESTART)
2030 if (!list_empty(&bl_ast_list))
2031 ldlm_discard_bl_list(&bl_ast_list);
2033 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
2037 * Conflicting locks are detected for a lock to be enqueued, add the lock
2038 * into waiting list and send blocking ASTs to the conflicting locks.
2040 * \param[in] lock The lock to be enqueued.
2041 * \param[out] flags Lock flags for the lock to be enqueued.
2042 * \param[in] rpc_list Conflicting locks list.
2044 * \retval -ERESTART: Some lock was instantly canceled while sending
2045 * blocking ASTs, caller needs to re-check conflicting
2047 * \retval -EAGAIN: Lock was destroyed, caller should return error.
2048 * \reval 0: Lock is successfully added in waiting list.
2050 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
2051 struct list_head *rpc_list)
2053 struct ldlm_resource *res = lock->l_resource;
2057 check_res_locked(res);
2059 /* If either of the compat_queue()s returned failure, then we
2060 * have ASTs to send and must go onto the waiting list.
2062 * bug 2322: we used to unlink and re-add here, which was a
2063 * terrible folly -- if we goto restart, we could get
2064 * re-ordered! Causes deadlock, because ASTs aren't sent! */
2065 if (list_empty(&lock->l_res_link))
2066 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
2069 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
2072 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
2073 !ns_is_client(ldlm_res_to_ns(res)))
2074 class_fail_export(lock->l_export);
2076 if (rc == -ERESTART)
2077 ldlm_reprocess_all(res, 0);
2080 if (rc == -ERESTART) {
2081 /* 15715: The lock was granted and destroyed after
2082 * resource lock was dropped. Interval node was freed
2083 * in ldlm_lock_destroy. Anyway, this always happens
2084 * when a client is being evicted. So it would be
2085 * ok to return an error. -jay */
2086 if (ldlm_is_destroyed(lock))
2089 /* lock was granted while resource was unlocked. */
2090 if (ldlm_is_granted(lock)) {
2091 /* bug 11300: if the lock has been granted,
2092 * break earlier because otherwise, we will go
2093 * to restart and ldlm_resource_unlink will be
2094 * called and it causes the interval node to be
2095 * freed. Then we will fail at
2096 * 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);
2130 * Process a call to blocking AST callback for a lock in ast_work list
2133 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2135 struct ldlm_cb_set_arg *arg = opaq;
2136 struct ldlm_lock *lock;
2137 struct ldlm_lock_desc d;
2138 struct ldlm_bl_desc bld;
2143 if (list_empty(arg->list))
2146 lock = list_first_entry(arg->list, struct ldlm_lock, l_bl_ast);
2148 /* nobody should touch l_bl_ast but some locks in the list may become
2149 * granted after lock convert or COS downgrade, these locks should be
2150 * just skipped here and removed from the list.
2152 lock_res_and_lock(lock);
2153 list_del_init(&lock->l_bl_ast);
2155 /* lock is not blocking lock anymore, but was kept in the list because
2156 * it can managed only here.
2158 if (!ldlm_is_ast_sent(lock)) {
2159 unlock_res_and_lock(lock);
2160 LDLM_LOCK_RELEASE(lock);
2164 LASSERT(lock->l_blocking_lock);
2165 ldlm_lock2desc(lock->l_blocking_lock, &d);
2166 /* copy blocking lock ibits in cancel_bits as well,
2167 * new client may use them for lock convert and it is
2168 * important to use new field to convert locks from
2171 d.l_policy_data.l_inodebits.cancel_bits =
2172 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2174 /* Blocking lock is being destroyed here but some information about it
2175 * may be needed inside l_blocking_ast() function below,
2176 * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
2178 bld.bl_same_client = lock->l_client_cookie ==
2179 lock->l_blocking_lock->l_client_cookie;
2180 /* if two locks are initiated from the same MDT, transactions are
2181 * independent, or the request lock mode isn't EX|PW, no need to trigger
2182 * CoS because current lock will be downgraded to TXN mode soon, then
2183 * the blocking lock can be granted.
2185 if (lock->l_blocking_lock->l_policy_data.l_inodebits.li_initiator_id ==
2186 lock->l_policy_data.l_inodebits.li_initiator_id ||
2187 !(lock->l_blocking_lock->l_req_mode & (LCK_EX | LCK_PW)))
2188 bld.bl_txn_dependent = false;
2190 bld.bl_txn_dependent = true;
2191 arg->bl_desc = &bld;
2193 LASSERT(ldlm_is_ast_sent(lock));
2194 LASSERT(lock->l_bl_ast_run == 0);
2195 lock->l_bl_ast_run++;
2196 ldlm_clear_blocking_lock(lock);
2197 unlock_res_and_lock(lock);
2199 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2201 LDLM_LOCK_RELEASE(lock);
2207 * Process a call to revocation AST callback for a lock in ast_work list
2210 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2212 struct ldlm_cb_set_arg *arg = opaq;
2213 struct ldlm_lock_desc desc;
2215 struct ldlm_lock *lock;
2218 if (list_empty(arg->list))
2221 lock = list_first_entry(arg->list, struct ldlm_lock, l_rk_ast);
2222 list_del_init(&lock->l_rk_ast);
2224 /* the desc just pretend to exclusive */
2225 ldlm_lock2desc(lock, &desc);
2226 desc.l_req_mode = LCK_EX;
2227 desc.l_granted_mode = 0;
2229 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2230 LDLM_LOCK_RELEASE(lock);
2236 * Process a call to glimpse AST callback for a lock in ast_work list
2238 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2240 struct ldlm_cb_set_arg *arg = opaq;
2241 struct ldlm_glimpse_work *gl_work;
2242 struct ldlm_lock *lock;
2246 if (list_empty(arg->list))
2249 gl_work = list_first_entry(arg->list, struct ldlm_glimpse_work,
2251 list_del_init(&gl_work->gl_list);
2253 lock = gl_work->gl_lock;
2255 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2256 arg->gl_desc = gl_work->gl_desc;
2257 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2258 arg->gl_interpret_data = gl_work->gl_interpret_data;
2260 /* invoke the actual glimpse callback */
2261 rc = lock->l_glimpse_ast(lock, (void *)arg);
2263 rc = 1; /* update LVB if this is server lock */
2264 else if (rc == -ELDLM_NO_LOCK_DATA)
2265 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
2267 LDLM_LOCK_RELEASE(lock);
2268 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2269 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2271 OBD_FREE_PTR(gl_work);
2278 * Process a call to completion AST callback for a lock in ast_work list
2281 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2283 struct ldlm_cb_set_arg *arg = opaq;
2284 struct ldlm_lock *lock;
2285 ldlm_completion_callback completion_callback;
2290 if (list_empty(arg->list))
2293 lock = list_first_entry(arg->list, struct ldlm_lock, l_cp_ast);
2295 /* It's possible to receive a completion AST before we've set
2296 * the l_completion_ast pointer: either because the AST arrived
2297 * before the reply, or simply because there's a small race
2298 * window between receiving the reply and finishing the local
2299 * enqueue. (bug 842)
2301 * This can't happen with the blocking_ast, however, because we
2302 * will never call the local blocking_ast until we drop our
2303 * reader/writer reference, which we won't do until we get the
2304 * reply and finish enqueueing. */
2306 /* nobody should touch l_cp_ast */
2307 lock_res_and_lock(lock);
2308 list_del_init(&lock->l_cp_ast);
2309 LASSERT(ldlm_is_cp_reqd(lock));
2310 /* save l_completion_ast since it can be changed by
2311 * mds_intent_policy(), see bug 14225 */
2312 completion_callback = lock->l_completion_ast;
2313 ldlm_clear_cp_reqd(lock);
2314 unlock_res_and_lock(lock);
2316 if (completion_callback != NULL)
2317 rc = completion_callback(lock, 0, (void *)arg);
2318 LDLM_LOCK_RELEASE(lock);
2324 * Process list of locks in need of ASTs being sent.
2326 * Used on server to send multiple ASTs together instead of sending one by
2329 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2330 ldlm_desc_ast_t ast_type)
2332 struct ldlm_cb_set_arg *arg;
2333 set_producer_func work_ast_lock;
2336 if (list_empty(rpc_list))
2343 atomic_set(&arg->restart, 0);
2344 arg->list = rpc_list;
2347 case LDLM_WORK_CP_AST:
2348 arg->type = LDLM_CP_CALLBACK;
2349 work_ast_lock = ldlm_work_cp_ast_lock;
2351 #ifdef HAVE_SERVER_SUPPORT
2352 case LDLM_WORK_BL_AST:
2353 arg->type = LDLM_BL_CALLBACK;
2354 work_ast_lock = ldlm_work_bl_ast_lock;
2356 case LDLM_WORK_REVOKE_AST:
2357 arg->type = LDLM_BL_CALLBACK;
2358 work_ast_lock = ldlm_work_revoke_ast_lock;
2360 case LDLM_WORK_GL_AST:
2361 arg->type = LDLM_GL_CALLBACK;
2362 work_ast_lock = ldlm_work_gl_ast_lock;
2369 /* We create a ptlrpc request set with flow control extension.
2370 * This request set will use the work_ast_lock function to produce new
2371 * requests and will send a new request each time one completes in order
2372 * to keep the number of requests in flight to ns_max_parallel_ast */
2373 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2374 work_ast_lock, arg);
2375 if (arg->set == NULL)
2376 GOTO(out, rc = -ENOMEM);
2378 ptlrpc_set_wait(NULL, arg->set);
2379 ptlrpc_set_destroy(arg->set);
2381 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2389 * Try to grant all waiting locks on a resource.
2391 * Calls ldlm_reprocess_queue on waiting queue.
2393 * Typically called after some resource locks are cancelled to see
2394 * if anything could be granted as a result of the cancellation.
2396 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2397 enum ldlm_process_intention intention,
2400 LIST_HEAD(rpc_list);
2401 #ifdef HAVE_SERVER_SUPPORT
2402 ldlm_reprocessing_policy reprocess;
2403 struct obd_device *obd;
2408 /* Local lock trees don't get reprocessed. */
2409 if (ns_is_client(ldlm_res_to_ns(res))) {
2414 /* Disable reprocess during lock replay stage but allow during
2415 * request replay stage.
2417 obd = ldlm_res_to_ns(res)->ns_obd;
2418 if (obd->obd_recovering &&
2419 atomic_read(&obd->obd_req_replay_clients) == 0)
2423 reprocess = ldlm_get_reprocessing_policy(res);
2424 reprocess(res, &res->lr_waiting, &rpc_list, intention, hint);
2427 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2429 if (rc == -ERESTART) {
2430 LASSERT(list_empty(&rpc_list));
2437 if (!ns_is_client(ldlm_res_to_ns(res))) {
2438 CERROR("This is client-side-only module, cannot handle "
2439 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2446 void ldlm_reprocess_all(struct ldlm_resource *res, __u64 hint)
2448 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN, hint);
2450 EXPORT_SYMBOL(ldlm_reprocess_all);
2452 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2453 struct hlist_node *hnode, void *arg)
2455 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2457 /* This is only called once after recovery done. LU-8306. */
2458 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY, 0);
2463 * Iterate through all resources on a namespace attempting to grant waiting
2466 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2471 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2472 ldlm_reprocess_res, NULL, 0);
2478 * Helper function to call blocking AST for LDLM lock \a lock in a
2479 * "cancelling" mode.
2481 void ldlm_cancel_callback(struct ldlm_lock *lock)
2483 check_res_locked(lock->l_resource);
2484 if (!ldlm_is_cancel(lock)) {
2485 ldlm_set_cancel(lock);
2486 if (lock->l_blocking_ast) {
2487 unlock_res_and_lock(lock);
2488 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2490 lock_res_and_lock(lock);
2492 LDLM_DEBUG(lock, "no blocking ast");
2495 /* only canceller can set bl_done bit */
2496 ldlm_set_bl_done(lock);
2497 wake_up(&lock->l_waitq);
2498 } else if (!ldlm_is_bl_done(lock)) {
2499 /* The lock is guaranteed to have been canceled once
2500 * returning from this function. */
2501 unlock_res_and_lock(lock);
2502 wait_event_idle(lock->l_waitq, is_bl_done(lock));
2503 lock_res_and_lock(lock);
2508 * Remove skiplist-enabled LDLM lock \a req from granted list
2510 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2512 if (req->l_resource->lr_type != LDLM_PLAIN &&
2513 req->l_resource->lr_type != LDLM_IBITS)
2516 list_del_init(&req->l_sl_policy);
2517 list_del_init(&req->l_sl_mode);
2521 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2523 void ldlm_lock_cancel(struct ldlm_lock *lock)
2525 struct ldlm_resource *res;
2526 struct ldlm_namespace *ns;
2529 lock_res_and_lock(lock);
2531 res = lock->l_resource;
2532 ns = ldlm_res_to_ns(res);
2534 /* Please do not, no matter how tempting, remove this LBUG without
2535 * talking to me first. -phik */
2536 if (lock->l_readers || lock->l_writers) {
2537 LDLM_ERROR(lock, "lock still has references");
2538 unlock_res_and_lock(lock);
2542 if (ldlm_is_waited(lock))
2543 ldlm_del_waiting_lock(lock);
2545 /* Releases cancel callback. */
2546 ldlm_cancel_callback(lock);
2548 /* Yes, second time, just in case it was added again while we were
2549 * running with no res lock in ldlm_cancel_callback */
2550 if (ldlm_is_waited(lock))
2551 ldlm_del_waiting_lock(lock);
2553 ldlm_resource_unlink_lock(lock);
2554 ldlm_lock_destroy_nolock(lock);
2556 if (ldlm_is_granted(lock))
2557 ldlm_pool_del(&ns->ns_pool, lock);
2559 /* Make sure we will not be called again for same lock what is possible
2560 * if not to zero out lock->l_granted_mode */
2561 lock->l_granted_mode = LCK_MINMODE;
2562 unlock_res_and_lock(lock);
2566 EXPORT_SYMBOL(ldlm_lock_cancel);
2569 * Set opaque data into the lock that only makes sense to upper layer.
2571 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2573 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2578 if (lock->l_ast_data == NULL)
2579 lock->l_ast_data = data;
2580 if (lock->l_ast_data == data)
2582 LDLM_LOCK_PUT(lock);
2586 EXPORT_SYMBOL(ldlm_lock_set_data);
2588 struct export_cl_data {
2589 const struct lu_env *ecl_env;
2590 struct obd_export *ecl_exp;
2594 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2595 struct ldlm_lock *lock,
2596 struct export_cl_data *ecl)
2598 struct ldlm_resource *res;
2600 res = ldlm_resource_getref(lock->l_resource);
2602 ldlm_lvbo_update(res, lock, NULL, 1);
2603 ldlm_lock_cancel(lock);
2604 if (!exp->exp_obd->obd_stopping)
2605 ldlm_reprocess_all(res, lock->l_policy_data.l_inodebits.bits);
2606 ldlm_resource_putref(res);
2609 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2610 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2611 exp, ecl->ecl_loop);
2616 * Iterator function for ldlm_export_cancel_locks.
2617 * Cancels passed locks.
2620 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2621 struct hlist_node *hnode, void *data)
2624 struct export_cl_data *ecl = (struct export_cl_data *)data;
2625 struct obd_export *exp = ecl->ecl_exp;
2626 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2628 LDLM_LOCK_GET(lock);
2629 ldlm_cancel_lock_for_export(exp, lock, ecl);
2630 LDLM_LOCK_RELEASE(lock);
2636 * Cancel all blocked locks for given export.
2638 * Typically called on client disconnection/eviction
2640 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2643 struct export_cl_data ecl = {
2649 rc = lu_env_init(&env, LCT_DT_THREAD);
2654 while (!list_empty(&exp->exp_bl_list)) {
2655 struct ldlm_lock *lock;
2657 spin_lock_bh(&exp->exp_bl_list_lock);
2658 if (!list_empty(&exp->exp_bl_list)) {
2659 lock = list_first_entry(&exp->exp_bl_list,
2660 struct ldlm_lock, l_exp_list);
2661 LDLM_LOCK_GET(lock);
2662 list_del_init(&lock->l_exp_list);
2666 spin_unlock_bh(&exp->exp_bl_list_lock);
2671 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2672 LDLM_LOCK_RELEASE(lock);
2677 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2678 "left on hash table %d.\n", exp, ecl.ecl_loop,
2679 atomic_read(&exp->exp_lock_hash->hs_count));
2681 return ecl.ecl_loop;
2685 * Cancel all locks for given export.
2687 * Typically called after client disconnection/eviction
2689 int ldlm_export_cancel_locks(struct obd_export *exp)
2691 struct export_cl_data ecl;
2695 rc = lu_env_init(&env, LCT_DT_THREAD);
2702 cfs_hash_for_each_empty(exp->exp_lock_hash,
2703 ldlm_cancel_locks_for_export_cb, &ecl);
2705 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2706 "left on hash table %d.\n", exp, ecl.ecl_loop,
2707 atomic_read(&exp->exp_lock_hash->hs_count));
2709 if (ecl.ecl_loop > 0 &&
2710 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2711 exp->exp_obd->obd_stopping)
2712 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2716 return ecl.ecl_loop;
2720 * Downgrade an PW/EX lock to COS, TXN or CR mode.
2722 * A lock mode convertion from PW/EX mode to less conflict mode. The
2723 * convertion may fail if lock was canceled before downgrade, but it doesn't
2724 * indicate any problem, because such lock has no reader or writer, and will
2727 * Used by Commit on Sharing (COS) code to force object changes commit in case
2728 * of conflict. Converted lock is considered as new lock and all blocking AST
2729 * things are cleared, so any pending or new blocked lock on that lock will
2730 * cause new call to blocking_ast and force resource object commit.
2732 * Used by DNE to force commit upon operation dependency.
2734 * Also used by layout_change to replace EX lock to CR lock.
2736 * \param lock A lock to convert
2737 * \param new_mode new lock mode
2739 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2741 #ifdef HAVE_SERVER_SUPPORT
2744 LASSERT(new_mode == LCK_COS || new_mode == LCK_TXN ||
2745 new_mode == LCK_CR);
2747 lock_res_and_lock(lock);
2749 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2750 unlock_res_and_lock(lock);
2752 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2753 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2757 ldlm_resource_unlink_lock(lock);
2759 * Remove the lock from pool as it will be added again in
2760 * ldlm_grant_lock() called below.
2762 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2764 /* Consider downgraded lock as a new lock and clear all states
2765 * related to a previous blocking AST processing.
2767 ldlm_clear_blocking_data(lock);
2769 lock->l_req_mode = new_mode;
2770 ldlm_grant_lock(lock, NULL);
2771 unlock_res_and_lock(lock);
2773 ldlm_reprocess_all(lock->l_resource,
2774 lock->l_policy_data.l_inodebits.bits);
2779 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2782 * Print lock with lock handle \a lockh description into debug log.
2784 * Used when printing all locks on a resource for debug purposes.
2786 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2788 struct ldlm_lock *lock;
2790 if (!((libcfs_debug | D_ERROR) & level))
2793 lock = ldlm_handle2lock(lockh);
2797 LDLM_DEBUG_LIMIT(level, lock, "###");
2799 LDLM_LOCK_PUT(lock);
2801 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2804 * Print lock information with custom message into debug log.
2807 void _ldlm_lock_debug(struct ldlm_lock *lock,
2808 struct libcfs_debug_msg_data *msgdata,
2809 const char *fmt, ...)
2812 struct obd_export *exp = lock->l_export;
2813 struct ldlm_resource *resource = NULL;
2814 struct va_format vaf;
2815 char *nid = "local";
2818 resource = rcu_dereference(lock->l_resource);
2819 if (resource && !refcount_inc_not_zero(&resource->lr_refcount))
2823 va_start(args, fmt);
2827 if (exp && exp->exp_connection) {
2828 nid = obd_export_nid2str(exp);
2829 } else if (exp && exp->exp_obd != NULL) {
2830 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2832 nid = obd_import_nid2str(imp);
2835 if (resource == NULL) {
2836 libcfs_debug_msg(msgdata,
2837 "%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",
2840 lock->l_handle.h_cookie,
2841 refcount_read(&lock->l_handle.h_ref),
2842 lock->l_readers, lock->l_writers,
2843 ldlm_lockname[lock->l_granted_mode],
2844 ldlm_lockname[lock->l_req_mode],
2846 lock->l_remote_handle.cookie,
2847 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2848 lock->l_pid, lock->l_callback_timestamp,
2854 switch (resource->lr_type) {
2856 libcfs_debug_msg(msgdata,
2857 "%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",
2859 ldlm_lock_to_ns_name(lock), lock,
2860 lock->l_handle.h_cookie,
2861 refcount_read(&lock->l_handle.h_ref),
2862 lock->l_readers, lock->l_writers,
2863 ldlm_lockname[lock->l_granted_mode],
2864 ldlm_lockname[lock->l_req_mode],
2866 refcount_read(&resource->lr_refcount),
2867 ldlm_typename[resource->lr_type],
2868 lock->l_policy_data.l_extent.start,
2869 lock->l_policy_data.l_extent.end,
2870 lock->l_req_extent.start, lock->l_req_extent.end,
2871 lock->l_req_extent.gid,
2873 lock->l_remote_handle.cookie,
2874 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2875 lock->l_pid, lock->l_callback_timestamp,
2880 libcfs_debug_msg(msgdata,
2881 "%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",
2883 ldlm_lock_to_ns_name(lock), lock,
2884 lock->l_handle.h_cookie,
2885 refcount_read(&lock->l_handle.h_ref),
2886 lock->l_readers, lock->l_writers,
2887 ldlm_lockname[lock->l_granted_mode],
2888 ldlm_lockname[lock->l_req_mode],
2890 refcount_read(&resource->lr_refcount),
2891 ldlm_typename[resource->lr_type],
2892 lock->l_policy_data.l_flock.pid,
2893 lock->l_policy_data.l_flock.start,
2894 lock->l_policy_data.l_flock.end,
2896 lock->l_remote_handle.cookie,
2897 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2898 lock->l_pid, lock->l_callback_timestamp);
2902 if (!lock->l_remote_handle.cookie)
2903 libcfs_debug_msg(msgdata,
2904 "%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",
2906 ldlm_lock_to_ns_name(lock),
2907 lock, lock->l_handle.h_cookie,
2908 refcount_read(&lock->l_handle.h_ref),
2909 lock->l_readers, lock->l_writers,
2910 ldlm_lockname[lock->l_granted_mode],
2911 ldlm_lockname[lock->l_req_mode],
2913 lock->l_policy_data.l_inodebits.bits,
2914 lock->l_policy_data.l_inodebits.try_bits,
2915 refcount_read(&resource->lr_refcount),
2916 ldlm_typename[resource->lr_type],
2917 lock->l_flags, lock->l_pid,
2918 lock->l_policy_data.l_inodebits.li_initiator_id);
2920 libcfs_debug_msg(msgdata,
2921 "%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",
2923 ldlm_lock_to_ns_name(lock),
2924 lock, lock->l_handle.h_cookie,
2925 refcount_read(&lock->l_handle.h_ref),
2926 lock->l_readers, lock->l_writers,
2927 ldlm_lockname[lock->l_granted_mode],
2928 ldlm_lockname[lock->l_req_mode],
2930 lock->l_policy_data.l_inodebits.bits,
2931 lock->l_policy_data.l_inodebits.try_bits,
2932 refcount_read(&resource->lr_refcount),
2933 ldlm_typename[resource->lr_type],
2934 lock->l_policy_data.l_inodebits.li_gid,
2936 lock->l_remote_handle.cookie,
2937 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2938 lock->l_pid, lock->l_callback_timestamp,
2943 libcfs_debug_msg(msgdata,
2944 "%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",
2946 ldlm_lock_to_ns_name(lock),
2947 lock, lock->l_handle.h_cookie,
2948 refcount_read(&lock->l_handle.h_ref),
2949 lock->l_readers, lock->l_writers,
2950 ldlm_lockname[lock->l_granted_mode],
2951 ldlm_lockname[lock->l_req_mode],
2953 refcount_read(&resource->lr_refcount),
2954 ldlm_typename[resource->lr_type],
2956 lock->l_remote_handle.cookie,
2957 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2958 lock->l_pid, lock->l_callback_timestamp,
2963 ldlm_resource_putref(resource);
2965 EXPORT_SYMBOL(_ldlm_lock_debug);