4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_lock.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <libcfs/libcfs.h>
42 #include <lustre_swab.h>
43 #include <obd_class.h>
45 #include "ldlm_internal.h"
47 struct kmem_cache *ldlm_glimpse_work_kmem;
48 EXPORT_SYMBOL(ldlm_glimpse_work_kmem);
51 char *ldlm_lockname[] = {
59 [LCK_GROUP] = "GROUP",
62 EXPORT_SYMBOL(ldlm_lockname);
64 char *ldlm_typename[] = {
66 [LDLM_EXTENT] = "EXT",
71 static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = {
72 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_wire_to_local,
73 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_wire_to_local,
74 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_wire_to_local,
75 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_wire_to_local,
78 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
79 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_local_to_wire,
80 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_local_to_wire,
81 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_local_to_wire,
82 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_local_to_wire,
86 * Converts lock policy from local format to on the wire lock_desc format
88 void ldlm_convert_policy_to_wire(enum ldlm_type type,
89 const union ldlm_policy_data *lpolicy,
90 union ldlm_wire_policy_data *wpolicy)
92 ldlm_policy_local_to_wire_t convert;
94 convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
96 convert(lpolicy, wpolicy);
100 * Converts lock policy from on the wire lock_desc format to local format
102 void ldlm_convert_policy_to_local(struct obd_export *exp, enum ldlm_type type,
103 const union ldlm_wire_policy_data *wpolicy,
104 union ldlm_policy_data *lpolicy)
106 ldlm_policy_wire_to_local_t convert;
108 convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE];
110 convert(wpolicy, lpolicy);
113 const char *ldlm_it2str(enum ldlm_intent_flags it)
120 case (IT_OPEN | IT_CREAT):
133 CERROR("Unknown intent 0x%08x\n", it);
137 EXPORT_SYMBOL(ldlm_it2str);
139 #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) {
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 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);
398 if (lock->l_export && lock->l_export->exp_lock_hash) {
399 /* NB: it's safe to call cfs_hash_del() even lock isn't
400 * in exp_lock_hash. */
401 /* In the function below, .hs_keycmp resolves to
402 * ldlm_export_lock_keycmp() */
403 /* coverity[overrun-buffer-val] */
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 spin_lock_init(&lock->l_lock);
473 lock->l_resource = resource;
474 lu_ref_add(&resource->lr_reference, "lock", lock);
476 refcount_set(&lock->l_handle.h_ref, 2);
477 INIT_LIST_HEAD(&lock->l_res_link);
478 INIT_LIST_HEAD(&lock->l_lru);
479 INIT_LIST_HEAD(&lock->l_pending_chain);
480 INIT_LIST_HEAD(&lock->l_bl_ast);
481 INIT_LIST_HEAD(&lock->l_cp_ast);
482 INIT_LIST_HEAD(&lock->l_rk_ast);
483 init_waitqueue_head(&lock->l_waitq);
484 lock->l_blocking_lock = NULL;
485 INIT_LIST_HEAD(&lock->l_sl_mode);
486 INIT_LIST_HEAD(&lock->l_sl_policy);
487 INIT_HLIST_NODE(&lock->l_exp_hash);
488 INIT_HLIST_NODE(&lock->l_exp_flock_hash);
490 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
492 INIT_HLIST_NODE(&lock->l_handle.h_link);
493 class_handle_hash(&lock->l_handle, lock_handle_owner);
495 lu_ref_init(&lock->l_reference);
496 lu_ref_add(&lock->l_reference, "hash", lock);
497 lock->l_callback_timestamp = 0;
498 lock->l_activity = 0;
500 #if LUSTRE_TRACKS_LOCK_EXP_REFS
501 INIT_LIST_HEAD(&lock->l_exp_refs_link);
502 lock->l_exp_refs_nr = 0;
503 lock->l_exp_refs_target = NULL;
505 INIT_LIST_HEAD(&lock->l_exp_list);
511 * Moves LDLM lock \a lock to another resource.
512 * This is used on client when server returns some other lock than requested
513 * (typically as a result of intent operation)
515 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
516 const struct ldlm_res_id *new_resid)
518 struct ldlm_resource *oldres;
519 struct ldlm_resource *newres;
523 LASSERT(ns_is_client(ns));
525 oldres = lock_res_and_lock(lock);
526 if (memcmp(new_resid, &oldres->lr_name,
527 sizeof(oldres->lr_name)) == 0) {
529 unlock_res_and_lock(lock);
533 LASSERT(new_resid->name[0] != 0);
535 /* This function assumes that the lock isn't on any lists */
536 LASSERT(list_empty(&lock->l_res_link));
538 type = oldres->lr_type;
539 unlock_res_and_lock(lock);
541 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
543 RETURN(PTR_ERR(newres));
545 lu_ref_add(&newres->lr_reference, "lock", lock);
547 * To flip the lock from the old to the new resource, lock, oldres and
548 * newres have to be locked. Resource spin-locks are nested within
549 * lock->l_lock, and are taken in the memory address order to avoid
552 spin_lock(&lock->l_lock);
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 lock->l_resource = newres;
565 unlock_res_and_lock(lock);
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)) {
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)) {
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 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
889 if (ldlm_is_atomic_cb(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 OBD_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 list_head *tmp;
977 struct ldlm_lock *lock, *mode_end, *policy_end;
980 list_for_each(tmp, queue) {
981 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
983 mode_end = list_entry(lock->l_sl_mode.prev,
984 struct ldlm_lock, l_sl_mode);
986 if (lock->l_req_mode != req->l_req_mode) {
987 /* jump to last lock of mode group */
988 tmp = &mode_end->l_res_link;
992 /* suitable mode group is found */
993 if (lock->l_resource->lr_type == LDLM_PLAIN) {
994 /* insert point is last lock of the mode group */
995 prev->res_link = &mode_end->l_res_link;
996 prev->mode_link = &mode_end->l_sl_mode;
997 prev->policy_link = &req->l_sl_policy;
1000 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
1003 list_entry(lock->l_sl_policy.prev,
1007 if (lock->l_policy_data.l_inodebits.bits ==
1008 req->l_policy_data.l_inodebits.bits) {
1009 /* insert point is last lock of
1010 * the policy group */
1012 &policy_end->l_res_link;
1014 &policy_end->l_sl_mode;
1016 &policy_end->l_sl_policy;
1021 if (policy_end == mode_end)
1022 /* done with mode group */
1025 /* go to next policy group within mode group */
1026 tmp = policy_end->l_res_link.next;
1027 lock = list_entry(tmp, struct ldlm_lock,
1029 } /* loop over policy groups within the mode group */
1031 /* insert point is last lock of the mode group,
1032 * new policy group is started */
1033 prev->res_link = &mode_end->l_res_link;
1034 prev->mode_link = &mode_end->l_sl_mode;
1035 prev->policy_link = &req->l_sl_policy;
1039 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1044 /* insert point is last lock on the queue,
1045 * new mode group and new policy group are started */
1046 prev->res_link = queue->prev;
1047 prev->mode_link = &req->l_sl_mode;
1048 prev->policy_link = &req->l_sl_policy;
1053 * Add a lock into resource granted list after a position described by
1056 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1057 struct sl_insert_point *prev)
1059 struct ldlm_resource *res = lock->l_resource;
1062 check_res_locked(res);
1064 ldlm_resource_dump(D_INFO, res);
1065 LDLM_DEBUG(lock, "About to add lock:");
1067 if (ldlm_is_destroyed(lock)) {
1068 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1072 LASSERT(list_empty(&lock->l_res_link));
1073 LASSERT(list_empty(&lock->l_sl_mode));
1074 LASSERT(list_empty(&lock->l_sl_policy));
1077 * lock->link == prev->link means lock is first starting the group.
1078 * Don't re-add to itself to suppress kernel warnings.
1080 if (&lock->l_res_link != prev->res_link)
1081 list_add(&lock->l_res_link, prev->res_link);
1082 if (&lock->l_sl_mode != prev->mode_link)
1083 list_add(&lock->l_sl_mode, prev->mode_link);
1084 if (&lock->l_sl_policy != prev->policy_link)
1085 list_add(&lock->l_sl_policy, prev->policy_link);
1091 * Add a lock to granted list on a resource maintaining skiplist
1094 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1096 struct sl_insert_point prev;
1098 LASSERT(ldlm_is_granted(lock));
1100 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1101 ldlm_granted_list_add_lock(lock, &prev);
1105 * Perform lock granting bookkeeping.
1107 * Includes putting the lock into granted list and updating lock mode.
1109 * - ldlm_lock_enqueue
1110 * - ldlm_reprocess_queue
1112 * must be called with lr_lock held
1114 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1116 struct ldlm_resource *res = lock->l_resource;
1119 check_res_locked(res);
1121 lock->l_granted_mode = lock->l_req_mode;
1123 if (work_list && lock->l_completion_ast != NULL)
1124 ldlm_add_ast_work_item(lock, NULL, work_list);
1126 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1127 ldlm_grant_lock_with_skiplist(lock);
1128 else if (res->lr_type == LDLM_EXTENT)
1129 ldlm_extent_add_lock(res, lock);
1130 else if (res->lr_type == LDLM_FLOCK) {
1131 /* We should not add locks to granted list in the following
1133 * - this is an UNLOCK but not a real lock;
1134 * - this is a TEST lock;
1135 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1136 * - this is a deadlock (flock cannot be granted) */
1137 if (lock->l_req_mode == 0 ||
1138 lock->l_req_mode == LCK_NL ||
1139 ldlm_is_test_lock(lock) ||
1140 ldlm_is_flock_deadlock(lock))
1142 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1147 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1152 * Check if the given @lock meets the criteria for a match.
1153 * A reference on the lock is taken if matched.
1155 * \param lock test-against this lock
1156 * \param data parameters
1158 static int lock_matches(struct ldlm_lock *lock, struct ldlm_match_data *data)
1160 union ldlm_policy_data *lpol = &lock->l_policy_data;
1161 enum ldlm_mode match = LCK_MINMODE;
1163 if (lock == data->lmd_old)
1164 return INTERVAL_ITER_STOP;
1166 /* Check if this lock can be matched.
1167 * Used by LU-2919(exclusive open) for open lease lock */
1168 if (ldlm_is_excl(lock))
1169 return INTERVAL_ITER_CONT;
1171 /* llite sometimes wants to match locks that will be
1172 * canceled when their users drop, but we allow it to match
1173 * if it passes in CBPENDING and the lock still has users.
1174 * this is generally only going to be used by children
1175 * whose parents already hold a lock so forward progress
1176 * can still happen. */
1177 if (ldlm_is_cbpending(lock) &&
1178 !(data->lmd_flags & LDLM_FL_CBPENDING))
1179 return INTERVAL_ITER_CONT;
1180 if (!(data->lmd_match & LDLM_MATCH_UNREF) && ldlm_is_cbpending(lock) &&
1181 lock->l_readers == 0 && lock->l_writers == 0)
1182 return INTERVAL_ITER_CONT;
1184 if (!(lock->l_req_mode & *data->lmd_mode))
1185 return INTERVAL_ITER_CONT;
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)
1192 return INTERVAL_ITER_CONT;
1194 if (data->lmd_match & LDLM_MATCH_AST_ANY)
1198 match = lock->l_req_mode;
1200 switch (lock->l_resource->lr_type) {
1202 if (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1203 lpol->l_extent.end < data->lmd_policy->l_extent.end)
1204 return INTERVAL_ITER_CONT;
1206 if (unlikely(match == LCK_GROUP) &&
1207 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1208 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1209 return INTERVAL_ITER_CONT;
1212 /* We match if we have existing lock with same or wider set
1214 if ((lpol->l_inodebits.bits &
1215 data->lmd_policy->l_inodebits.bits) !=
1216 data->lmd_policy->l_inodebits.bits)
1217 return INTERVAL_ITER_CONT;
1219 if (unlikely(match == LCK_GROUP) &&
1220 data->lmd_policy->l_inodebits.li_gid != LDLM_GID_ANY &&
1221 lpol->l_inodebits.li_gid !=
1222 data->lmd_policy->l_inodebits.li_gid)
1223 return INTERVAL_ITER_CONT;
1229 /* We match if we have existing lock with same or wider set
1231 if (!(data->lmd_match & LDLM_MATCH_UNREF) && LDLM_HAVE_MASK(lock, GONE))
1232 return INTERVAL_ITER_CONT;
1234 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1235 return INTERVAL_ITER_CONT;
1237 /* Filter locks by skipping flags */
1238 if (data->lmd_skip_flags & lock->l_flags)
1239 return INTERVAL_ITER_CONT;
1242 if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1243 LDLM_LOCK_GET(lock);
1244 ldlm_lock_touch_in_lru(lock);
1246 ldlm_lock_addref_internal_nolock(lock, match);
1249 *data->lmd_mode = match;
1250 data->lmd_lock = lock;
1252 return INTERVAL_ITER_STOP;
1255 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1257 struct ldlm_interval *node = to_ldlm_interval(in);
1258 struct ldlm_match_data *data = args;
1259 struct ldlm_lock *lock;
1262 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1263 rc = lock_matches(lock, data);
1264 if (rc == INTERVAL_ITER_STOP)
1265 return INTERVAL_ITER_STOP;
1267 return INTERVAL_ITER_CONT;
1271 * Search for a lock with given parameters in interval trees.
1273 * \param res search for a lock in this resource
1274 * \param data parameters
1276 * \retval a referenced lock or NULL.
1278 struct ldlm_lock *search_itree(struct ldlm_resource *res,
1279 struct ldlm_match_data *data)
1281 struct interval_node_extent ext = {
1282 .start = data->lmd_policy->l_extent.start,
1283 .end = data->lmd_policy->l_extent.end
1287 data->lmd_lock = NULL;
1289 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1290 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1292 if (tree->lit_root == NULL)
1295 if (!(tree->lit_mode & *data->lmd_mode))
1298 interval_search(tree->lit_root, &ext,
1299 itree_overlap_cb, data);
1301 return data->lmd_lock;
1306 EXPORT_SYMBOL(search_itree);
1310 * Search for a lock with given properties in a queue.
1312 * \param queue search for a lock in this queue
1313 * \param data parameters
1315 * \retval a referenced lock or NULL.
1317 static struct ldlm_lock *search_queue(struct list_head *queue,
1318 struct ldlm_match_data *data)
1320 struct ldlm_lock *lock;
1323 data->lmd_lock = NULL;
1325 list_for_each_entry(lock, queue, l_res_link) {
1326 rc = lock_matches(lock, data);
1327 if (rc == INTERVAL_ITER_STOP)
1328 return data->lmd_lock;
1334 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1336 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1337 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1338 wake_up_all(&lock->l_waitq);
1341 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1343 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1345 lock_res_and_lock(lock);
1346 ldlm_lock_fail_match_locked(lock);
1347 unlock_res_and_lock(lock);
1351 * Mark lock as "matchable" by OST.
1353 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1355 * Assumes LDLM lock is already locked.
1357 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1359 ldlm_set_lvb_ready(lock);
1360 wake_up_all(&lock->l_waitq);
1362 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1365 * Mark lock as "matchable" by OST.
1366 * Locks the lock and then \see ldlm_lock_allow_match_locked
1368 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1370 lock_res_and_lock(lock);
1371 ldlm_lock_allow_match_locked(lock);
1372 unlock_res_and_lock(lock);
1374 EXPORT_SYMBOL(ldlm_lock_allow_match);
1377 * Attempt to find a lock with specified properties.
1379 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1382 * Can be called in two ways:
1384 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1385 * for a duplicate of.
1387 * Otherwise, all of the fields must be filled in, to match against.
1389 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1390 * server (ie, connh is NULL)
1391 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1392 * list will be considered
1393 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1394 * to be canceled can still be matched as long as they still have reader
1395 * or writer refernces
1396 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1397 * just tell us if we would have matched.
1399 * \retval 1 if it finds an already-existing lock that is compatible; in this
1400 * case, lockh is filled in with a addref()ed lock
1402 * We also check security context, and if that fails we simply return 0 (to
1403 * keep caller code unchanged), the context failure will be discovered by
1404 * caller sometime later.
1406 enum ldlm_mode ldlm_lock_match_with_skip(struct ldlm_namespace *ns,
1407 __u64 flags, __u64 skip_flags,
1408 const struct ldlm_res_id *res_id,
1409 enum ldlm_type type,
1410 union ldlm_policy_data *policy,
1411 enum ldlm_mode mode,
1412 struct lustre_handle *lockh,
1413 enum ldlm_match_flags match_flags)
1415 struct ldlm_match_data data = {
1419 .lmd_policy = policy,
1421 .lmd_skip_flags = skip_flags,
1422 .lmd_match = match_flags,
1424 struct ldlm_resource *res;
1425 struct ldlm_lock *lock;
1431 data.lmd_old = ldlm_handle2lock(lockh);
1432 LASSERT(data.lmd_old != NULL);
1434 ns = ldlm_lock_to_ns(data.lmd_old);
1435 res_id = &data.lmd_old->l_resource->lr_name;
1436 type = data.lmd_old->l_resource->lr_type;
1437 *data.lmd_mode = data.lmd_old->l_req_mode;
1440 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1442 LASSERT(data.lmd_old == NULL);
1446 LDLM_RESOURCE_ADDREF(res);
1448 if (res->lr_type == LDLM_EXTENT)
1449 lock = search_itree(res, &data);
1451 lock = search_queue(&res->lr_granted, &data);
1452 if (!lock && !(flags & LDLM_FL_BLOCK_GRANTED))
1453 lock = search_queue(&res->lr_waiting, &data);
1454 matched = lock ? mode : 0;
1456 LDLM_RESOURCE_DELREF(res);
1457 ldlm_resource_putref(res);
1460 ldlm_lock2handle(lock, lockh);
1461 if ((flags & LDLM_FL_LVB_READY) &&
1462 (!ldlm_is_lvb_ready(lock))) {
1463 __u64 wait_flags = LDLM_FL_LVB_READY |
1464 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1466 if (lock->l_completion_ast) {
1467 int err = lock->l_completion_ast(lock,
1468 LDLM_FL_WAIT_NOREPROC,
1471 GOTO(out_fail_match, matched = 0);
1474 wait_event_idle_timeout(
1476 lock->l_flags & wait_flags,
1477 cfs_time_seconds(obd_timeout));
1479 if (!ldlm_is_lvb_ready(lock))
1480 GOTO(out_fail_match, matched = 0);
1483 /* check user's security context */
1484 if (lock->l_conn_export &&
1485 sptlrpc_import_check_ctx(
1486 class_exp2cliimp(lock->l_conn_export)))
1487 GOTO(out_fail_match, matched = 0);
1489 LDLM_DEBUG(lock, "matched (%llu %llu)",
1490 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1491 res_id->name[2] : policy->l_extent.start,
1492 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1493 res_id->name[3] : policy->l_extent.end);
1496 if (flags & LDLM_FL_TEST_LOCK)
1497 LDLM_LOCK_RELEASE(lock);
1499 ldlm_lock_decref_internal(lock, mode);
1502 /* less verbose for test-only */
1503 if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
1504 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1505 "%llu/%llu (%llu %llu)", ns,
1506 type, mode, res_id->name[0], res_id->name[1],
1507 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1508 res_id->name[2] : policy->l_extent.start,
1509 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1510 res_id->name[3] : policy->l_extent.end);
1512 if (data.lmd_old != NULL)
1513 LDLM_LOCK_PUT(data.lmd_old);
1517 EXPORT_SYMBOL(ldlm_lock_match_with_skip);
1519 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1522 struct ldlm_lock *lock;
1523 enum ldlm_mode mode = 0;
1526 lock = ldlm_handle2lock(lockh);
1528 lock_res_and_lock(lock);
1529 if (LDLM_HAVE_MASK(lock, GONE))
1532 if (ldlm_is_cbpending(lock) &&
1533 lock->l_readers == 0 && lock->l_writers == 0)
1537 *bits = lock->l_policy_data.l_inodebits.bits;
1538 mode = lock->l_granted_mode;
1539 ldlm_lock_addref_internal_nolock(lock, mode);
1546 unlock_res_and_lock(lock);
1547 LDLM_LOCK_PUT(lock);
1551 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1553 /** The caller must guarantee that the buffer is large enough. */
1554 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1555 enum req_location loc, void *data, int size)
1560 LASSERT(data != NULL);
1563 switch (lock->l_lvb_type) {
1565 if (size == sizeof(struct ost_lvb)) {
1566 if (loc == RCL_CLIENT)
1567 lvb = req_capsule_client_swab_get(pill,
1569 lustre_swab_ost_lvb);
1571 lvb = req_capsule_server_swab_get(pill,
1573 lustre_swab_ost_lvb);
1574 if (unlikely(lvb == NULL)) {
1575 LDLM_ERROR(lock, "no LVB");
1579 memcpy(data, lvb, size);
1580 } else if (size == sizeof(struct ost_lvb_v1)) {
1581 struct ost_lvb *olvb = data;
1583 if (loc == RCL_CLIENT)
1584 lvb = req_capsule_client_swab_get(pill,
1586 lustre_swab_ost_lvb_v1);
1588 lvb = req_capsule_server_sized_swab_get(pill,
1590 lustre_swab_ost_lvb_v1);
1591 if (unlikely(lvb == NULL)) {
1592 LDLM_ERROR(lock, "no LVB");
1596 memcpy(data, lvb, size);
1597 olvb->lvb_mtime_ns = 0;
1598 olvb->lvb_atime_ns = 0;
1599 olvb->lvb_ctime_ns = 0;
1601 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1607 if (size == sizeof(struct lquota_lvb)) {
1608 if (loc == RCL_CLIENT)
1609 lvb = req_capsule_client_swab_get(pill,
1611 lustre_swab_lquota_lvb);
1613 lvb = req_capsule_server_swab_get(pill,
1615 lustre_swab_lquota_lvb);
1616 if (unlikely(lvb == NULL)) {
1617 LDLM_ERROR(lock, "no LVB");
1621 memcpy(data, lvb, size);
1623 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1632 if (loc == RCL_CLIENT)
1633 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1635 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1636 if (unlikely(lvb == NULL)) {
1637 LDLM_ERROR(lock, "no LVB");
1641 memcpy(data, lvb, size);
1644 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1645 libcfs_debug_dumpstack(NULL);
1653 * Create and fill in new LDLM lock with specified properties.
1654 * Returns a referenced lock
1656 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1657 const struct ldlm_res_id *res_id,
1658 enum ldlm_type type,
1659 enum ldlm_mode mode,
1660 const struct ldlm_callback_suite *cbs,
1661 void *data, __u32 lvb_len,
1662 enum lvb_type lvb_type)
1664 struct ldlm_lock *lock;
1665 struct ldlm_resource *res;
1669 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1671 RETURN(ERR_CAST(res));
1673 lock = ldlm_lock_new(res);
1675 RETURN(ERR_PTR(-ENOMEM));
1677 lock->l_req_mode = mode;
1678 lock->l_ast_data = data;
1679 lock->l_pid = current->pid;
1680 if (ns_is_server(ns))
1681 ldlm_set_ns_srv(lock);
1683 lock->l_blocking_ast = cbs->lcs_blocking;
1684 lock->l_completion_ast = cbs->lcs_completion;
1685 lock->l_glimpse_ast = cbs->lcs_glimpse;
1690 rc = ldlm_extent_alloc_lock(lock);
1693 rc = ldlm_inodebits_alloc_lock(lock);
1702 lock->l_lvb_len = lvb_len;
1703 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1704 if (lock->l_lvb_data == NULL)
1705 GOTO(out, rc = -ENOMEM);
1708 lock->l_lvb_type = lvb_type;
1709 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1710 GOTO(out, rc = -ENOENT);
1715 ldlm_lock_destroy(lock);
1716 LDLM_LOCK_RELEASE(lock);
1717 RETURN(ERR_PTR(rc));
1720 #ifdef HAVE_SERVER_SUPPORT
1721 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1724 struct ldlm_resource *res = lock->l_resource;
1725 enum ldlm_error rc = ELDLM_OK;
1726 LIST_HEAD(rpc_list);
1727 ldlm_processing_policy policy;
1731 policy = ldlm_get_processing_policy(res);
1732 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1733 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1734 res->lr_type != LDLM_FLOCK)
1735 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1737 if (!list_empty(&rpc_list))
1738 ldlm_discard_bl_list(&rpc_list);
1745 * Enqueue (request) a lock.
1747 * Does not block. As a result of enqueue the lock would be put
1748 * into granted or waiting list.
1750 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1751 * set, skip all the enqueueing and delegate lock processing to intent policy
1754 enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
1755 struct ldlm_namespace *ns,
1756 struct ldlm_lock **lockp,
1757 void *cookie, __u64 *flags)
1759 struct ldlm_lock *lock = *lockp;
1760 struct ldlm_resource *res;
1761 int local = ns_is_client(ns);
1762 enum ldlm_error rc = ELDLM_OK;
1763 struct ldlm_interval *node = NULL;
1764 #ifdef HAVE_SERVER_SUPPORT
1765 bool reconstruct = false;
1769 /* policies are not executed on the client or during replay */
1770 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1771 && !local && ns->ns_policy) {
1772 rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
1774 if (rc == ELDLM_LOCK_REPLACED) {
1775 /* The lock that was returned has already been granted,
1776 * and placed into lockp. If it's not the same as the
1777 * one we passed in, then destroy the old one and our
1778 * work here is done. */
1779 if (lock != *lockp) {
1780 ldlm_lock_destroy(lock);
1781 LDLM_LOCK_RELEASE(lock);
1783 *flags |= LDLM_FL_LOCK_CHANGED;
1785 } else if (rc != ELDLM_OK &&
1786 ldlm_is_granted(lock)) {
1787 LASSERT(*flags & LDLM_FL_RESENT);
1788 /* It may happen that ns_policy returns an error in
1789 * resend case, object may be unlinked or just some
1790 * error occurs. It is unclear if lock reached the
1791 * client in the original reply, just leave the lock on
1792 * server, not returning it again to client. Due to
1793 * LU-6529, the server will not OOM. */
1795 } else if (rc != ELDLM_OK ||
1796 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1797 ldlm_lock_destroy(lock);
1802 if (*flags & LDLM_FL_RESENT) {
1803 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1804 * Set LOCK_CHANGED always.
1805 * Check if the lock is granted for BLOCK_GRANTED.
1806 * Take NO_TIMEOUT from the lock as it is inherited through
1807 * LDLM_FL_INHERIT_MASK */
1808 *flags |= LDLM_FL_LOCK_CHANGED;
1809 if (!ldlm_is_granted(lock))
1810 *flags |= LDLM_FL_BLOCK_GRANTED;
1811 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1815 #ifdef HAVE_SERVER_SUPPORT
1816 /* For a replaying lock, it might be already in granted list. So
1817 * unlinking the lock will cause the interval node to be freed, we
1818 * have to allocate the interval node early otherwise we can't regrant
1819 * this lock in the future. - jay
1821 * The only time the ldlm_resource changes for the ldlm_lock is when
1822 * ldlm_lock_change_resource() is called and that only happens for
1823 * the Lustre client case.
1825 if (!local && (*flags & LDLM_FL_REPLAY) &&
1826 lock->l_resource->lr_type == LDLM_EXTENT)
1827 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1829 reconstruct = !local && lock->l_resource->lr_type == LDLM_FLOCK &&
1830 !(*flags & LDLM_FL_TEST_LOCK);
1832 rc = req_can_reconstruct(cookie, NULL);
1840 res = lock_res_and_lock(lock);
1841 if (local && ldlm_is_granted(lock)) {
1842 /* The server returned a blocked lock, but it was granted
1843 * before we got a chance to actually enqueue it. We don't
1844 * need to do anything else. */
1845 *flags &= ~LDLM_FL_BLOCKED_MASK;
1846 GOTO(out, rc = ELDLM_OK);
1849 ldlm_resource_unlink_lock(lock);
1850 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1852 ldlm_lock_destroy_nolock(lock);
1853 GOTO(out, rc = -ENOMEM);
1856 INIT_LIST_HEAD(&node->li_group);
1857 ldlm_interval_attach(node, lock);
1861 /* Some flags from the enqueue want to make it into the AST, via the
1862 * lock's l_flags. */
1863 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1864 ldlm_set_ast_discard_data(lock);
1865 if (*flags & LDLM_FL_TEST_LOCK)
1866 ldlm_set_test_lock(lock);
1867 if (*flags & LDLM_FL_COS_INCOMPAT)
1868 ldlm_set_cos_incompat(lock);
1869 if (*flags & LDLM_FL_COS_ENABLED)
1870 ldlm_set_cos_enabled(lock);
1872 /* This distinction between local lock trees is very important; a client
1873 * namespace only has information about locks taken by that client, and
1874 * thus doesn't have enough information to decide for itself if it can
1875 * be granted (below). In this case, we do exactly what the server
1876 * tells us to do, as dictated by the 'flags'.
1878 * We do exactly the same thing during recovery, when the server is
1879 * more or less trusting the clients not to lie.
1881 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1884 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1885 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1887 ldlm_grant_lock(lock, NULL);
1888 GOTO(out, rc = ELDLM_OK);
1889 #ifdef HAVE_SERVER_SUPPORT
1890 } else if (*flags & LDLM_FL_REPLAY) {
1891 if (*flags & LDLM_FL_BLOCK_WAIT) {
1892 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1893 GOTO(out, rc = ELDLM_OK);
1894 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1895 ldlm_grant_lock(lock, NULL);
1896 GOTO(out, rc = ELDLM_OK);
1898 /* If no flags, fall through to normal enqueue path. */
1901 rc = ldlm_lock_enqueue_helper(lock, flags);
1905 CERROR("This is client-side-only module, cannot handle "
1906 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1912 unlock_res_and_lock(lock);
1914 #ifdef HAVE_SERVER_SUPPORT
1916 struct ptlrpc_request *req = cookie;
1918 tgt_mk_reply_data(NULL, NULL,
1919 &req->rq_export->exp_target_data,
1920 req, 0, NULL, false, 0);
1924 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1928 #ifdef HAVE_SERVER_SUPPORT
1930 * Iterate through all waiting locks on a given resource queue and attempt to
1933 * Must be called with resource lock held.
1935 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1936 struct list_head *work_list,
1937 enum ldlm_process_intention intention,
1938 struct ldlm_lock *hint)
1940 struct list_head *tmp, *pos;
1941 ldlm_processing_policy policy;
1943 int rc = LDLM_ITER_CONTINUE;
1944 enum ldlm_error err;
1945 LIST_HEAD(bl_ast_list);
1949 check_res_locked(res);
1951 policy = ldlm_get_processing_policy(res);
1953 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1954 intention == LDLM_PROCESS_RECOVERY);
1957 list_for_each_safe(tmp, pos, queue) {
1958 struct ldlm_lock *pending;
1959 LIST_HEAD(rpc_list);
1961 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1963 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1966 rc = policy(pending, &flags, intention, &err, &rpc_list);
1967 if (pending->l_granted_mode == pending->l_req_mode ||
1968 res->lr_type == LDLM_FLOCK) {
1969 list_splice(&rpc_list, work_list);
1971 list_splice(&rpc_list, &bl_ast_list);
1974 * When this is called from recovery done, we always want
1975 * to scan the whole list no matter what 'rc' is returned.
1977 if (rc != LDLM_ITER_CONTINUE &&
1978 intention == LDLM_PROCESS_RESCAN)
1982 if (!list_empty(&bl_ast_list)) {
1985 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
1989 if (rc == -ERESTART)
1993 if (!list_empty(&bl_ast_list))
1994 ldlm_discard_bl_list(&bl_ast_list);
1996 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
2000 * Conflicting locks are detected for a lock to be enqueued, add the lock
2001 * into waiting list and send blocking ASTs to the conflicting locks.
2003 * \param[in] lock The lock to be enqueued.
2004 * \param[out] flags Lock flags for the lock to be enqueued.
2005 * \param[in] rpc_list Conflicting locks list.
2007 * \retval -ERESTART: Some lock was instantly canceled while sending
2008 * blocking ASTs, caller needs to re-check conflicting
2010 * \retval -EAGAIN: Lock was destroyed, caller should return error.
2011 * \reval 0: Lock is successfully added in waiting list.
2013 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
2014 struct list_head *rpc_list)
2016 struct ldlm_resource *res = lock->l_resource;
2020 check_res_locked(res);
2022 /* If either of the compat_queue()s returned failure, then we
2023 * have ASTs to send and must go onto the waiting list.
2025 * bug 2322: we used to unlink and re-add here, which was a
2026 * terrible folly -- if we goto restart, we could get
2027 * re-ordered! Causes deadlock, because ASTs aren't sent! */
2028 if (list_empty(&lock->l_res_link))
2029 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
2032 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
2035 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
2036 !ns_is_client(ldlm_res_to_ns(res)))
2037 class_fail_export(lock->l_export);
2039 if (rc == -ERESTART)
2040 ldlm_reprocess_all(res, NULL);
2043 if (rc == -ERESTART) {
2044 /* 15715: The lock was granted and destroyed after
2045 * resource lock was dropped. Interval node was freed
2046 * in ldlm_lock_destroy. Anyway, this always happens
2047 * when a client is being evicted. So it would be
2048 * ok to return an error. -jay */
2049 if (ldlm_is_destroyed(lock))
2052 /* lock was granted while resource was unlocked. */
2053 if (ldlm_is_granted(lock)) {
2054 /* bug 11300: if the lock has been granted,
2055 * break earlier because otherwise, we will go
2056 * to restart and ldlm_resource_unlink will be
2057 * called and it causes the interval node to be
2058 * freed. Then we will fail at
2059 * ldlm_extent_add_lock() */
2060 *flags &= ~LDLM_FL_BLOCKED_MASK;
2064 *flags |= LDLM_FL_BLOCK_GRANTED;
2070 * Discard all AST work items from list.
2072 * If for whatever reason we do not want to send ASTs to conflicting locks
2073 * anymore, disassemble the list with this function.
2075 void ldlm_discard_bl_list(struct list_head *bl_list)
2077 struct ldlm_lock *lock, *tmp;
2081 list_for_each_entry_safe(lock, tmp, bl_list, l_bl_ast) {
2082 LASSERT(!list_empty(&lock->l_bl_ast));
2083 list_del_init(&lock->l_bl_ast);
2084 ldlm_clear_ast_sent(lock);
2085 LASSERT(lock->l_bl_ast_run == 0);
2086 ldlm_clear_blocking_lock(lock);
2087 LDLM_LOCK_RELEASE(lock);
2093 * Process a call to blocking AST callback for a lock in ast_work list
2096 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2098 struct ldlm_cb_set_arg *arg = opaq;
2099 struct ldlm_lock *lock;
2100 struct ldlm_lock_desc d;
2101 struct ldlm_bl_desc bld;
2106 if (list_empty(arg->list))
2109 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2111 /* nobody should touch l_bl_ast but some locks in the list may become
2112 * granted after lock convert or COS downgrade, these locks should be
2113 * just skipped here and removed from the list.
2115 lock_res_and_lock(lock);
2116 list_del_init(&lock->l_bl_ast);
2118 /* lock is not blocking lock anymore, but was kept in the list because
2119 * it can managed only here.
2121 if (!ldlm_is_ast_sent(lock)) {
2122 unlock_res_and_lock(lock);
2123 LDLM_LOCK_RELEASE(lock);
2127 LASSERT(lock->l_blocking_lock);
2128 ldlm_lock2desc(lock->l_blocking_lock, &d);
2129 /* copy blocking lock ibits in cancel_bits as well,
2130 * new client may use them for lock convert and it is
2131 * important to use new field to convert locks from
2134 d.l_policy_data.l_inodebits.cancel_bits =
2135 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2137 /* Blocking lock is being destroyed here but some information about it
2138 * may be needed inside l_blocking_ast() function below,
2139 * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
2141 bld.bl_same_client = lock->l_client_cookie ==
2142 lock->l_blocking_lock->l_client_cookie;
2143 bld.bl_cos_incompat = ldlm_is_cos_incompat(lock->l_blocking_lock);
2144 arg->bl_desc = &bld;
2146 LASSERT(ldlm_is_ast_sent(lock));
2147 LASSERT(lock->l_bl_ast_run == 0);
2148 lock->l_bl_ast_run++;
2149 ldlm_clear_blocking_lock(lock);
2150 unlock_res_and_lock(lock);
2152 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2154 LDLM_LOCK_RELEASE(lock);
2160 * Process a call to revocation AST callback for a lock in ast_work list
2163 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2165 struct ldlm_cb_set_arg *arg = opaq;
2166 struct ldlm_lock_desc desc;
2168 struct ldlm_lock *lock;
2171 if (list_empty(arg->list))
2174 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2175 list_del_init(&lock->l_rk_ast);
2177 /* the desc just pretend to exclusive */
2178 ldlm_lock2desc(lock, &desc);
2179 desc.l_req_mode = LCK_EX;
2180 desc.l_granted_mode = 0;
2182 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2183 LDLM_LOCK_RELEASE(lock);
2189 * Process a call to glimpse AST callback for a lock in ast_work list
2191 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2193 struct ldlm_cb_set_arg *arg = opaq;
2194 struct ldlm_glimpse_work *gl_work;
2195 struct ldlm_lock *lock;
2199 if (list_empty(arg->list))
2202 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2204 list_del_init(&gl_work->gl_list);
2206 lock = gl_work->gl_lock;
2208 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2209 arg->gl_desc = gl_work->gl_desc;
2210 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2211 arg->gl_interpret_data = gl_work->gl_interpret_data;
2213 /* invoke the actual glimpse callback */
2214 rc = lock->l_glimpse_ast(lock, (void *)arg);
2216 rc = 1; /* update LVB if this is server lock */
2217 else if (rc == -ELDLM_NO_LOCK_DATA)
2218 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
2220 LDLM_LOCK_RELEASE(lock);
2221 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2222 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2224 OBD_FREE_PTR(gl_work);
2231 * Process a call to completion AST callback for a lock in ast_work list
2234 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2236 struct ldlm_cb_set_arg *arg = opaq;
2237 struct ldlm_lock *lock;
2238 ldlm_completion_callback completion_callback;
2243 if (list_empty(arg->list))
2246 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2248 /* It's possible to receive a completion AST before we've set
2249 * the l_completion_ast pointer: either because the AST arrived
2250 * before the reply, or simply because there's a small race
2251 * window between receiving the reply and finishing the local
2252 * enqueue. (bug 842)
2254 * This can't happen with the blocking_ast, however, because we
2255 * will never call the local blocking_ast until we drop our
2256 * reader/writer reference, which we won't do until we get the
2257 * reply and finish enqueueing. */
2259 /* nobody should touch l_cp_ast */
2260 lock_res_and_lock(lock);
2261 list_del_init(&lock->l_cp_ast);
2262 LASSERT(ldlm_is_cp_reqd(lock));
2263 /* save l_completion_ast since it can be changed by
2264 * mds_intent_policy(), see bug 14225 */
2265 completion_callback = lock->l_completion_ast;
2266 ldlm_clear_cp_reqd(lock);
2267 unlock_res_and_lock(lock);
2269 if (completion_callback != NULL)
2270 rc = completion_callback(lock, 0, (void *)arg);
2271 LDLM_LOCK_RELEASE(lock);
2277 * Process list of locks in need of ASTs being sent.
2279 * Used on server to send multiple ASTs together instead of sending one by
2282 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2283 ldlm_desc_ast_t ast_type)
2285 struct ldlm_cb_set_arg *arg;
2286 set_producer_func work_ast_lock;
2289 if (list_empty(rpc_list))
2296 atomic_set(&arg->restart, 0);
2297 arg->list = rpc_list;
2300 case LDLM_WORK_CP_AST:
2301 arg->type = LDLM_CP_CALLBACK;
2302 work_ast_lock = ldlm_work_cp_ast_lock;
2304 #ifdef HAVE_SERVER_SUPPORT
2305 case LDLM_WORK_BL_AST:
2306 arg->type = LDLM_BL_CALLBACK;
2307 work_ast_lock = ldlm_work_bl_ast_lock;
2309 case LDLM_WORK_REVOKE_AST:
2310 arg->type = LDLM_BL_CALLBACK;
2311 work_ast_lock = ldlm_work_revoke_ast_lock;
2313 case LDLM_WORK_GL_AST:
2314 arg->type = LDLM_GL_CALLBACK;
2315 work_ast_lock = ldlm_work_gl_ast_lock;
2322 /* We create a ptlrpc request set with flow control extension.
2323 * This request set will use the work_ast_lock function to produce new
2324 * requests and will send a new request each time one completes in order
2325 * to keep the number of requests in flight to ns_max_parallel_ast */
2326 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2327 work_ast_lock, arg);
2328 if (arg->set == NULL)
2329 GOTO(out, rc = -ENOMEM);
2331 ptlrpc_set_wait(NULL, arg->set);
2332 ptlrpc_set_destroy(arg->set);
2334 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2342 * Try to grant all waiting locks on a resource.
2344 * Calls ldlm_reprocess_queue on waiting queue.
2346 * Typically called after some resource locks are cancelled to see
2347 * if anything could be granted as a result of the cancellation.
2349 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2350 enum ldlm_process_intention intention,
2351 struct ldlm_lock *hint)
2353 LIST_HEAD(rpc_list);
2354 #ifdef HAVE_SERVER_SUPPORT
2355 ldlm_reprocessing_policy reprocess;
2356 struct obd_device *obd;
2361 /* Local lock trees don't get reprocessed. */
2362 if (ns_is_client(ldlm_res_to_ns(res))) {
2367 /* Disable reprocess during lock replay stage but allow during
2368 * request replay stage.
2370 obd = ldlm_res_to_ns(res)->ns_obd;
2371 if (obd->obd_recovering &&
2372 atomic_read(&obd->obd_req_replay_clients) == 0)
2376 reprocess = ldlm_get_reprocessing_policy(res);
2377 reprocess(res, &res->lr_waiting, &rpc_list, intention, hint);
2380 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2382 if (rc == -ERESTART) {
2383 LASSERT(list_empty(&rpc_list));
2389 if (!ns_is_client(ldlm_res_to_ns(res))) {
2390 CERROR("This is client-side-only module, cannot handle "
2391 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2398 void ldlm_reprocess_all(struct ldlm_resource *res, struct ldlm_lock *hint)
2400 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN, hint);
2402 EXPORT_SYMBOL(ldlm_reprocess_all);
2404 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2405 struct hlist_node *hnode, void *arg)
2407 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2409 /* This is only called once after recovery done. LU-8306. */
2410 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY, NULL);
2415 * Iterate through all resources on a namespace attempting to grant waiting
2418 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2423 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2424 ldlm_reprocess_res, NULL, 0);
2430 * Helper function to call blocking AST for LDLM lock \a lock in a
2431 * "cancelling" mode.
2433 void ldlm_cancel_callback(struct ldlm_lock *lock)
2435 check_res_locked(lock->l_resource);
2436 if (!ldlm_is_cancel(lock)) {
2437 ldlm_set_cancel(lock);
2438 if (lock->l_blocking_ast) {
2439 unlock_res_and_lock(lock);
2440 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2442 lock_res_and_lock(lock);
2444 LDLM_DEBUG(lock, "no blocking ast");
2447 /* only canceller can set bl_done bit */
2448 ldlm_set_bl_done(lock);
2449 wake_up_all(&lock->l_waitq);
2450 } else if (!ldlm_is_bl_done(lock)) {
2451 /* The lock is guaranteed to have been canceled once
2452 * returning from this function. */
2453 unlock_res_and_lock(lock);
2454 wait_event_idle(lock->l_waitq, is_bl_done(lock));
2455 lock_res_and_lock(lock);
2460 * Remove skiplist-enabled LDLM lock \a req from granted list
2462 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2464 if (req->l_resource->lr_type != LDLM_PLAIN &&
2465 req->l_resource->lr_type != LDLM_IBITS)
2468 list_del_init(&req->l_sl_policy);
2469 list_del_init(&req->l_sl_mode);
2473 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2475 void ldlm_lock_cancel(struct ldlm_lock *lock)
2477 struct ldlm_resource *res;
2478 struct ldlm_namespace *ns;
2481 lock_res_and_lock(lock);
2483 res = lock->l_resource;
2484 ns = ldlm_res_to_ns(res);
2486 /* Please do not, no matter how tempting, remove this LBUG without
2487 * talking to me first. -phik */
2488 if (lock->l_readers || lock->l_writers) {
2489 LDLM_ERROR(lock, "lock still has references");
2490 unlock_res_and_lock(lock);
2494 if (ldlm_is_waited(lock))
2495 ldlm_del_waiting_lock(lock);
2497 /* Releases cancel callback. */
2498 ldlm_cancel_callback(lock);
2500 /* Yes, second time, just in case it was added again while we were
2501 * running with no res lock in ldlm_cancel_callback */
2502 if (ldlm_is_waited(lock))
2503 ldlm_del_waiting_lock(lock);
2505 ldlm_resource_unlink_lock(lock);
2506 ldlm_lock_destroy_nolock(lock);
2508 if (ldlm_is_granted(lock))
2509 ldlm_pool_del(&ns->ns_pool, lock);
2511 /* Make sure we will not be called again for same lock what is possible
2512 * if not to zero out lock->l_granted_mode */
2513 lock->l_granted_mode = LCK_MINMODE;
2514 unlock_res_and_lock(lock);
2518 EXPORT_SYMBOL(ldlm_lock_cancel);
2521 * Set opaque data into the lock that only makes sense to upper layer.
2523 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2525 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2530 if (lock->l_ast_data == NULL)
2531 lock->l_ast_data = data;
2532 if (lock->l_ast_data == data)
2534 LDLM_LOCK_PUT(lock);
2538 EXPORT_SYMBOL(ldlm_lock_set_data);
2540 struct export_cl_data {
2541 const struct lu_env *ecl_env;
2542 struct obd_export *ecl_exp;
2546 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2547 struct ldlm_lock *lock,
2548 struct export_cl_data *ecl)
2550 struct ldlm_resource *res;
2552 res = ldlm_resource_getref(lock->l_resource);
2554 ldlm_lvbo_update(res, lock, NULL, 1);
2555 ldlm_lock_cancel(lock);
2556 if (!exp->exp_obd->obd_stopping)
2557 ldlm_reprocess_all(res, lock);
2558 ldlm_resource_putref(res);
2561 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2562 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2563 exp, ecl->ecl_loop);
2568 * Iterator function for ldlm_export_cancel_locks.
2569 * Cancels passed locks.
2572 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2573 struct hlist_node *hnode, void *data)
2576 struct export_cl_data *ecl = (struct export_cl_data *)data;
2577 struct obd_export *exp = ecl->ecl_exp;
2578 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2580 LDLM_LOCK_GET(lock);
2581 ldlm_cancel_lock_for_export(exp, lock, ecl);
2582 LDLM_LOCK_RELEASE(lock);
2588 * Cancel all blocked locks for given export.
2590 * Typically called on client disconnection/eviction
2592 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2595 struct export_cl_data ecl = {
2601 rc = lu_env_init(&env, LCT_DT_THREAD);
2606 while (!list_empty(&exp->exp_bl_list)) {
2607 struct ldlm_lock *lock;
2609 spin_lock_bh(&exp->exp_bl_list_lock);
2610 if (!list_empty(&exp->exp_bl_list)) {
2611 lock = list_entry(exp->exp_bl_list.next,
2612 struct ldlm_lock, l_exp_list);
2613 LDLM_LOCK_GET(lock);
2614 list_del_init(&lock->l_exp_list);
2618 spin_unlock_bh(&exp->exp_bl_list_lock);
2623 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2624 LDLM_LOCK_RELEASE(lock);
2629 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2630 "left on hash table %d.\n", exp, ecl.ecl_loop,
2631 atomic_read(&exp->exp_lock_hash->hs_count));
2633 return ecl.ecl_loop;
2637 * Cancel all locks for given export.
2639 * Typically called after client disconnection/eviction
2641 int ldlm_export_cancel_locks(struct obd_export *exp)
2643 struct export_cl_data ecl;
2647 rc = lu_env_init(&env, LCT_DT_THREAD);
2654 cfs_hash_for_each_empty(exp->exp_lock_hash,
2655 ldlm_cancel_locks_for_export_cb, &ecl);
2657 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2658 "left on hash table %d.\n", exp, ecl.ecl_loop,
2659 atomic_read(&exp->exp_lock_hash->hs_count));
2661 if (ecl.ecl_loop > 0 &&
2662 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2663 exp->exp_obd->obd_stopping)
2664 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2668 return ecl.ecl_loop;
2672 * Downgrade an PW/EX lock to COS | CR mode.
2674 * A lock mode convertion from PW/EX mode to less conflict mode. The
2675 * convertion may fail if lock was canceled before downgrade, but it doesn't
2676 * indicate any problem, because such lock has no reader or writer, and will
2679 * Used by Commit on Sharing (COS) code to force object changes commit in case
2680 * of conflict. Converted lock is considered as new lock and all blocking AST
2681 * things are cleared, so any pending or new blocked lock on that lock will
2682 * cause new call to blocking_ast and force resource object commit.
2684 * Also used by layout_change to replace EX lock to CR lock.
2686 * \param lock A lock to convert
2687 * \param new_mode new lock mode
2689 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2691 #ifdef HAVE_SERVER_SUPPORT
2694 LASSERT(new_mode == LCK_COS || new_mode == LCK_CR);
2696 lock_res_and_lock(lock);
2698 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2699 unlock_res_and_lock(lock);
2701 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2702 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2706 ldlm_resource_unlink_lock(lock);
2708 * Remove the lock from pool as it will be added again in
2709 * ldlm_grant_lock() called below.
2711 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2713 /* Consider downgraded lock as a new lock and clear all states
2714 * related to a previous blocking AST processing.
2716 ldlm_clear_blocking_data(lock);
2718 lock->l_req_mode = new_mode;
2719 ldlm_grant_lock(lock, NULL);
2720 unlock_res_and_lock(lock);
2722 ldlm_reprocess_all(lock->l_resource, lock);
2727 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2730 * Print lock with lock handle \a lockh description into debug log.
2732 * Used when printing all locks on a resource for debug purposes.
2734 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2736 struct ldlm_lock *lock;
2738 if (!((libcfs_debug | D_ERROR) & level))
2741 lock = ldlm_handle2lock(lockh);
2745 LDLM_DEBUG_LIMIT(level, lock, "###");
2747 LDLM_LOCK_PUT(lock);
2749 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2752 * Print lock information with custom message into debug log.
2755 void _ldlm_lock_debug(struct ldlm_lock *lock,
2756 struct libcfs_debug_msg_data *msgdata,
2757 const char *fmt, ...)
2760 struct obd_export *exp = lock->l_export;
2761 struct ldlm_resource *resource = NULL;
2762 struct va_format vaf;
2763 char *nid = "local";
2765 /* on server-side resource of lock doesn't change */
2766 if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2767 if (lock->l_resource != NULL)
2768 resource = ldlm_resource_getref(lock->l_resource);
2769 } else if (spin_trylock(&lock->l_lock)) {
2770 if (lock->l_resource != NULL)
2771 resource = ldlm_resource_getref(lock->l_resource);
2772 spin_unlock(&lock->l_lock);
2775 va_start(args, fmt);
2779 if (exp && exp->exp_connection) {
2780 nid = obd_export_nid2str(exp);
2781 } else if (exp && exp->exp_obd != NULL) {
2782 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2783 nid = obd_import_nid2str(imp);
2786 if (resource == NULL) {
2787 libcfs_debug_msg(msgdata,
2788 "%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",
2791 lock->l_handle.h_cookie,
2792 refcount_read(&lock->l_handle.h_ref),
2793 lock->l_readers, lock->l_writers,
2794 ldlm_lockname[lock->l_granted_mode],
2795 ldlm_lockname[lock->l_req_mode],
2797 lock->l_remote_handle.cookie,
2798 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2799 lock->l_pid, lock->l_callback_timestamp,
2805 switch (resource->lr_type) {
2807 libcfs_debug_msg(msgdata,
2808 "%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",
2810 ldlm_lock_to_ns_name(lock), lock,
2811 lock->l_handle.h_cookie,
2812 refcount_read(&lock->l_handle.h_ref),
2813 lock->l_readers, lock->l_writers,
2814 ldlm_lockname[lock->l_granted_mode],
2815 ldlm_lockname[lock->l_req_mode],
2817 atomic_read(&resource->lr_refcount),
2818 ldlm_typename[resource->lr_type],
2819 lock->l_policy_data.l_extent.start,
2820 lock->l_policy_data.l_extent.end,
2821 lock->l_req_extent.start, lock->l_req_extent.end,
2822 lock->l_req_extent.gid,
2824 lock->l_remote_handle.cookie,
2825 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2826 lock->l_pid, lock->l_callback_timestamp,
2831 libcfs_debug_msg(msgdata,
2832 "%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",
2834 ldlm_lock_to_ns_name(lock), lock,
2835 lock->l_handle.h_cookie,
2836 refcount_read(&lock->l_handle.h_ref),
2837 lock->l_readers, lock->l_writers,
2838 ldlm_lockname[lock->l_granted_mode],
2839 ldlm_lockname[lock->l_req_mode],
2841 atomic_read(&resource->lr_refcount),
2842 ldlm_typename[resource->lr_type],
2843 lock->l_policy_data.l_flock.pid,
2844 lock->l_policy_data.l_flock.start,
2845 lock->l_policy_data.l_flock.end,
2847 lock->l_remote_handle.cookie,
2848 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2849 lock->l_pid, lock->l_callback_timestamp);
2853 libcfs_debug_msg(msgdata,
2854 "%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",
2856 ldlm_lock_to_ns_name(lock),
2857 lock, lock->l_handle.h_cookie,
2858 refcount_read(&lock->l_handle.h_ref),
2859 lock->l_readers, lock->l_writers,
2860 ldlm_lockname[lock->l_granted_mode],
2861 ldlm_lockname[lock->l_req_mode],
2863 lock->l_policy_data.l_inodebits.bits,
2864 lock->l_policy_data.l_inodebits.try_bits,
2865 atomic_read(&resource->lr_refcount),
2866 ldlm_typename[resource->lr_type],
2867 lock->l_policy_data.l_inodebits.li_gid,
2869 lock->l_remote_handle.cookie,
2870 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2871 lock->l_pid, lock->l_callback_timestamp,
2876 libcfs_debug_msg(msgdata,
2877 "%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",
2879 ldlm_lock_to_ns_name(lock),
2880 lock, lock->l_handle.h_cookie,
2881 refcount_read(&lock->l_handle.h_ref),
2882 lock->l_readers, lock->l_writers,
2883 ldlm_lockname[lock->l_granted_mode],
2884 ldlm_lockname[lock->l_req_mode],
2886 atomic_read(&resource->lr_refcount),
2887 ldlm_typename[resource->lr_type],
2889 lock->l_remote_handle.cookie,
2890 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2891 lock->l_pid, lock->l_callback_timestamp,
2896 ldlm_resource_putref(resource);
2898 EXPORT_SYMBOL(_ldlm_lock_debug);