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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_lock.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_LDLM
44 #include <libcfs/libcfs.h>
45 #include <obd_class.h>
46 #include "ldlm_internal.h"
49 char *ldlm_lockname[] = {
57 [LCK_GROUP] = "GROUP",
60 EXPORT_SYMBOL(ldlm_lockname);
62 char *ldlm_typename[] = {
64 [LDLM_EXTENT] = "EXT",
69 static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = {
70 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_wire_to_local,
71 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_wire_to_local,
72 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_wire_to_local,
73 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_wire_to_local,
76 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
77 [LDLM_PLAIN - LDLM_MIN_TYPE] = ldlm_plain_policy_local_to_wire,
78 [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_local_to_wire,
79 [LDLM_FLOCK - LDLM_MIN_TYPE] = ldlm_flock_policy_local_to_wire,
80 [LDLM_IBITS - LDLM_MIN_TYPE] = ldlm_ibits_policy_local_to_wire,
84 * Converts lock policy from local format to on the wire lock_desc format
86 void ldlm_convert_policy_to_wire(enum ldlm_type type,
87 const union ldlm_policy_data *lpolicy,
88 union ldlm_wire_policy_data *wpolicy)
90 ldlm_policy_local_to_wire_t convert;
92 convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
94 convert(lpolicy, wpolicy);
98 * Converts lock policy from on the wire lock_desc format to local format
100 void ldlm_convert_policy_to_local(struct obd_export *exp, enum ldlm_type type,
101 const union ldlm_wire_policy_data *wpolicy,
102 union ldlm_policy_data *lpolicy)
104 ldlm_policy_wire_to_local_t convert;
106 convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE];
108 convert(wpolicy, lpolicy);
111 const char *ldlm_it2str(enum ldlm_intent_flags it)
118 case (IT_OPEN | IT_CREAT):
133 CERROR("Unknown intent 0x%08x\n", it);
137 EXPORT_SYMBOL(ldlm_it2str);
139 extern struct kmem_cache *ldlm_lock_slab;
141 #ifdef HAVE_SERVER_SUPPORT
142 static ldlm_processing_policy ldlm_processing_policy_table[] = {
143 [LDLM_PLAIN] = ldlm_process_plain_lock,
144 [LDLM_EXTENT] = ldlm_process_extent_lock,
145 [LDLM_FLOCK] = ldlm_process_flock_lock,
146 [LDLM_IBITS] = ldlm_process_inodebits_lock,
149 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
151 return ldlm_processing_policy_table[res->lr_type];
153 EXPORT_SYMBOL(ldlm_get_processing_policy);
154 #endif /* HAVE_SERVER_SUPPORT */
156 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
160 EXPORT_SYMBOL(ldlm_register_intent);
163 * REFCOUNTED LOCK OBJECTS
168 * Get a reference on a lock.
170 * Lock refcounts, during creation:
171 * - one special one for allocation, dec'd only once in destroy
172 * - one for being a lock that's in-use
173 * - one for the addref associated with a new lock
175 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
177 atomic_inc(&lock->l_refc);
180 EXPORT_SYMBOL(ldlm_lock_get);
183 * Release lock reference.
185 * Also frees the lock if it was last reference.
187 void ldlm_lock_put(struct ldlm_lock *lock)
191 LASSERT(lock->l_resource != LP_POISON);
192 LASSERT(atomic_read(&lock->l_refc) > 0);
193 if (atomic_dec_and_test(&lock->l_refc)) {
194 struct ldlm_resource *res;
197 "final lock_put on destroyed lock, freeing it.");
199 res = lock->l_resource;
200 LASSERT(ldlm_is_destroyed(lock));
201 LASSERT(list_empty(&lock->l_exp_list));
202 LASSERT(list_empty(&lock->l_res_link));
203 LASSERT(list_empty(&lock->l_pending_chain));
205 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
207 lu_ref_del(&res->lr_reference, "lock", lock);
208 ldlm_resource_putref(res);
209 lock->l_resource = NULL;
210 if (lock->l_export) {
211 class_export_lock_put(lock->l_export, lock);
212 lock->l_export = NULL;
215 if (lock->l_lvb_data != NULL)
216 OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
218 ldlm_interval_free(ldlm_interval_detach(lock));
219 lu_ref_fini(&lock->l_reference);
220 OBD_FREE_RCU(lock, sizeof(*lock), &lock->l_handle);
225 EXPORT_SYMBOL(ldlm_lock_put);
228 * Removes LDLM lock \a lock from LRU. Assumes LRU is already locked.
230 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
233 if (!list_empty(&lock->l_lru)) {
234 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
236 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
237 list_del_init(&lock->l_lru);
238 LASSERT(ns->ns_nr_unused > 0);
246 * Removes LDLM lock \a lock from LRU. Obtains the LRU lock first.
248 * If \a last_use is non-zero, it will remove the lock from LRU only if
249 * it matches lock's l_last_used.
251 * \retval 0 if \a last_use is set, the lock is not in LRU list or \a last_use
252 * doesn't match lock's l_last_used;
253 * otherwise, the lock hasn't been in the LRU list.
254 * \retval 1 the lock was in LRU list and removed.
256 int ldlm_lock_remove_from_lru_check(struct ldlm_lock *lock, cfs_time_t last_use)
258 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
262 if (ldlm_is_ns_srv(lock)) {
263 LASSERT(list_empty(&lock->l_lru));
267 spin_lock(&ns->ns_lock);
268 if (last_use == 0 || last_use == lock->l_last_used)
269 rc = ldlm_lock_remove_from_lru_nolock(lock);
270 spin_unlock(&ns->ns_lock);
276 * Adds LDLM lock \a lock to namespace LRU. Assumes LRU is already locked.
278 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
280 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
282 lock->l_last_used = cfs_time_current();
283 LASSERT(list_empty(&lock->l_lru));
284 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
285 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
286 ldlm_clear_skipped(lock);
287 LASSERT(ns->ns_nr_unused >= 0);
292 * Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks
295 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
297 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
300 spin_lock(&ns->ns_lock);
301 ldlm_lock_add_to_lru_nolock(lock);
302 spin_unlock(&ns->ns_lock);
307 * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
308 * the LRU. Performs necessary LRU locking
310 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
312 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
315 if (ldlm_is_ns_srv(lock)) {
316 LASSERT(list_empty(&lock->l_lru));
321 spin_lock(&ns->ns_lock);
322 if (!list_empty(&lock->l_lru)) {
323 ldlm_lock_remove_from_lru_nolock(lock);
324 ldlm_lock_add_to_lru_nolock(lock);
326 spin_unlock(&ns->ns_lock);
331 * Helper to destroy a locked lock.
333 * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
334 * Must be called with l_lock and lr_lock held.
336 * Does not actually free the lock data, but rather marks the lock as
337 * destroyed by setting l_destroyed field in the lock to 1. Destroys a
338 * handle->lock association too, so that the lock can no longer be found
339 * and removes the lock from LRU list. Actual lock freeing occurs when
340 * last lock reference goes away.
342 * Original comment (of some historical value):
343 * This used to have a 'strict' flag, which recovery would use to mark an
344 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
345 * shall explain why it's gone: with the new hash table scheme, once you call
346 * ldlm_lock_destroy, you can never drop your final references on this lock.
347 * Because it's not in the hash table anymore. -phil
349 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
353 if (lock->l_readers || lock->l_writers) {
354 LDLM_ERROR(lock, "lock still has references");
358 if (!list_empty(&lock->l_res_link)) {
359 LDLM_ERROR(lock, "lock still on resource");
363 if (ldlm_is_destroyed(lock)) {
364 LASSERT(list_empty(&lock->l_lru));
368 ldlm_set_destroyed(lock);
370 if (lock->l_export && lock->l_export->exp_lock_hash) {
371 /* NB: it's safe to call cfs_hash_del() even lock isn't
372 * in exp_lock_hash. */
373 /* In the function below, .hs_keycmp resolves to
374 * ldlm_export_lock_keycmp() */
375 /* coverity[overrun-buffer-val] */
376 cfs_hash_del(lock->l_export->exp_lock_hash,
377 &lock->l_remote_handle, &lock->l_exp_hash);
380 ldlm_lock_remove_from_lru(lock);
381 class_handle_unhash(&lock->l_handle);
384 /* Wake anyone waiting for this lock */
385 /* FIXME: I should probably add yet another flag, instead of using
386 * l_export to only call this on clients */
388 class_export_put(lock->l_export);
389 lock->l_export = NULL;
390 if (lock->l_export && lock->l_completion_ast)
391 lock->l_completion_ast(lock, 0);
398 * Destroys a LDLM lock \a lock. Performs necessary locking first.
400 void ldlm_lock_destroy(struct ldlm_lock *lock)
404 lock_res_and_lock(lock);
405 first = ldlm_lock_destroy_internal(lock);
406 unlock_res_and_lock(lock);
408 /* drop reference from hashtable only for first destroy */
410 lu_ref_del(&lock->l_reference, "hash", lock);
411 LDLM_LOCK_RELEASE(lock);
417 * Destroys a LDLM lock \a lock that is already locked.
419 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
423 first = ldlm_lock_destroy_internal(lock);
424 /* drop reference from hashtable only for first destroy */
426 lu_ref_del(&lock->l_reference, "hash", lock);
427 LDLM_LOCK_RELEASE(lock);
432 /* this is called by portals_handle2object with the handle lock taken */
433 static void lock_handle_addref(void *lock)
435 LDLM_LOCK_GET((struct ldlm_lock *)lock);
438 static void lock_handle_free(void *lock, int size)
440 LASSERT(size == sizeof(struct ldlm_lock));
441 OBD_SLAB_FREE(lock, ldlm_lock_slab, size);
444 static struct portals_handle_ops lock_handle_ops = {
445 .hop_addref = lock_handle_addref,
446 .hop_free = lock_handle_free,
451 * Allocate and initialize new lock structure.
453 * usage: pass in a resource on which you have done ldlm_resource_get
454 * new lock will take over the refcount.
455 * returns: lock with refcount 2 - one for current caller and one for remote
457 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
459 struct ldlm_lock *lock;
462 if (resource == NULL)
465 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, GFP_NOFS);
469 spin_lock_init(&lock->l_lock);
470 lock->l_resource = resource;
471 lu_ref_add(&resource->lr_reference, "lock", lock);
473 atomic_set(&lock->l_refc, 2);
474 INIT_LIST_HEAD(&lock->l_res_link);
475 INIT_LIST_HEAD(&lock->l_lru);
476 INIT_LIST_HEAD(&lock->l_pending_chain);
477 INIT_LIST_HEAD(&lock->l_bl_ast);
478 INIT_LIST_HEAD(&lock->l_cp_ast);
479 INIT_LIST_HEAD(&lock->l_rk_ast);
480 init_waitqueue_head(&lock->l_waitq);
481 lock->l_blocking_lock = NULL;
482 INIT_LIST_HEAD(&lock->l_sl_mode);
483 INIT_LIST_HEAD(&lock->l_sl_policy);
484 INIT_HLIST_NODE(&lock->l_exp_hash);
485 INIT_HLIST_NODE(&lock->l_exp_flock_hash);
487 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
489 INIT_LIST_HEAD(&lock->l_handle.h_link);
490 class_handle_hash(&lock->l_handle, &lock_handle_ops);
492 lu_ref_init(&lock->l_reference);
493 lu_ref_add(&lock->l_reference, "hash", lock);
494 lock->l_callback_timeout = 0;
496 #if LUSTRE_TRACKS_LOCK_EXP_REFS
497 INIT_LIST_HEAD(&lock->l_exp_refs_link);
498 lock->l_exp_refs_nr = 0;
499 lock->l_exp_refs_target = NULL;
501 INIT_LIST_HEAD(&lock->l_exp_list);
507 * Moves LDLM lock \a lock to another resource.
508 * This is used on client when server returns some other lock than requested
509 * (typically as a result of intent operation)
511 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
512 const struct ldlm_res_id *new_resid)
514 struct ldlm_resource *oldres = lock->l_resource;
515 struct ldlm_resource *newres;
519 LASSERT(ns_is_client(ns));
521 lock_res_and_lock(lock);
522 if (memcmp(new_resid, &lock->l_resource->lr_name,
523 sizeof(lock->l_resource->lr_name)) == 0) {
525 unlock_res_and_lock(lock);
529 LASSERT(new_resid->name[0] != 0);
531 /* This function assumes that the lock isn't on any lists */
532 LASSERT(list_empty(&lock->l_res_link));
534 type = oldres->lr_type;
535 unlock_res_and_lock(lock);
537 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
539 RETURN(PTR_ERR(newres));
541 lu_ref_add(&newres->lr_reference, "lock", lock);
543 * To flip the lock from the old to the new resource, lock, oldres and
544 * newres have to be locked. Resource spin-locks are nested within
545 * lock->l_lock, and are taken in the memory address order to avoid
548 spin_lock(&lock->l_lock);
549 oldres = lock->l_resource;
550 if (oldres < newres) {
552 lock_res_nested(newres, LRT_NEW);
555 lock_res_nested(oldres, LRT_NEW);
557 LASSERT(memcmp(new_resid, &oldres->lr_name,
558 sizeof oldres->lr_name) != 0);
559 lock->l_resource = newres;
561 unlock_res_and_lock(lock);
563 /* ...and the flowers are still standing! */
564 lu_ref_del(&oldres->lr_reference, "lock", lock);
565 ldlm_resource_putref(oldres);
570 /** \defgroup ldlm_handles LDLM HANDLES
571 * Ways to get hold of locks without any addresses.
576 * Fills in handle for LDLM lock \a lock into supplied \a lockh
577 * Does not take any references.
579 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
581 lockh->cookie = lock->l_handle.h_cookie;
583 EXPORT_SYMBOL(ldlm_lock2handle);
586 * Obtain a lock reference by handle.
588 * if \a flags: atomically get the lock and set the flags.
589 * Return NULL if flag already set
591 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
594 struct ldlm_lock *lock;
599 lock = class_handle2object(handle->cookie, NULL);
603 /* It's unlikely but possible that someone marked the lock as
604 * destroyed after we did handle2object on it */
605 if ((flags == 0) && !ldlm_is_destroyed(lock)) {
606 lu_ref_add(&lock->l_reference, "handle", current);
610 lock_res_and_lock(lock);
612 LASSERT(lock->l_resource != NULL);
614 lu_ref_add_atomic(&lock->l_reference, "handle", current);
615 if (unlikely(ldlm_is_destroyed(lock))) {
616 unlock_res_and_lock(lock);
617 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
622 /* If we're setting flags, make sure none of them are already set. */
624 if ((lock->l_flags & flags) != 0) {
625 unlock_res_and_lock(lock);
630 lock->l_flags |= flags;
633 unlock_res_and_lock(lock);
636 EXPORT_SYMBOL(__ldlm_handle2lock);
637 /** @} ldlm_handles */
640 * Fill in "on the wire" representation for given LDLM lock into supplied
641 * lock descriptor \a desc structure.
643 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
645 ldlm_res2desc(lock->l_resource, &desc->l_resource);
646 desc->l_req_mode = lock->l_req_mode;
647 desc->l_granted_mode = lock->l_granted_mode;
648 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
649 &lock->l_policy_data,
650 &desc->l_policy_data);
654 * Add a lock to list of conflicting locks to send AST to.
656 * Only add if we have not sent a blocking AST to the lock yet.
658 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
659 struct list_head *work_list)
661 if (!ldlm_is_ast_sent(lock)) {
662 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
663 ldlm_set_ast_sent(lock);
664 /* If the enqueuing client said so, tell the AST recipient to
665 * discard dirty data, rather than writing back. */
666 if (ldlm_is_ast_discard_data(new))
667 ldlm_set_discard_data(lock);
668 LASSERT(list_empty(&lock->l_bl_ast));
669 list_add(&lock->l_bl_ast, work_list);
671 LASSERT(lock->l_blocking_lock == NULL);
672 lock->l_blocking_lock = LDLM_LOCK_GET(new);
677 * Add a lock to list of just granted locks to send completion AST to.
679 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
680 struct list_head *work_list)
682 if (!ldlm_is_cp_reqd(lock)) {
683 ldlm_set_cp_reqd(lock);
684 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
685 LASSERT(list_empty(&lock->l_cp_ast));
686 list_add(&lock->l_cp_ast, work_list);
692 * Aggregator function to add AST work items into a list. Determines
693 * what sort of an AST work needs to be done and calls the proper
695 * Must be called with lr_lock held.
697 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
698 struct list_head *work_list)
701 check_res_locked(lock->l_resource);
703 ldlm_add_bl_work_item(lock, new, work_list);
705 ldlm_add_cp_work_item(lock, work_list);
710 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
711 * r/w reference type is determined by \a mode
712 * Calls ldlm_lock_addref_internal.
714 void ldlm_lock_addref(struct lustre_handle *lockh, enum ldlm_mode mode)
716 struct ldlm_lock *lock;
718 lock = ldlm_handle2lock(lockh);
719 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
720 ldlm_lock_addref_internal(lock, mode);
723 EXPORT_SYMBOL(ldlm_lock_addref);
727 * Add specified reader/writer reference to LDLM lock \a lock.
728 * r/w reference type is determined by \a mode
729 * Removes lock from LRU if it is there.
730 * Assumes the LDLM lock is already locked.
732 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
735 ldlm_lock_remove_from_lru(lock);
736 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
738 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
740 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
742 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
745 lu_ref_add_atomic(&lock->l_reference, "user", lock);
746 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
750 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
751 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
753 * \retval 0 success, lock was addref-ed
755 * \retval -EAGAIN lock is being canceled.
757 int ldlm_lock_addref_try(struct lustre_handle *lockh, enum ldlm_mode mode)
759 struct ldlm_lock *lock;
763 lock = ldlm_handle2lock(lockh);
765 lock_res_and_lock(lock);
766 if (lock->l_readers != 0 || lock->l_writers != 0 ||
767 !ldlm_is_cbpending(lock)) {
768 ldlm_lock_addref_internal_nolock(lock, mode);
771 unlock_res_and_lock(lock);
776 EXPORT_SYMBOL(ldlm_lock_addref_try);
779 * Add specified reader/writer reference to LDLM lock \a lock.
780 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
781 * Only called for local locks.
783 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
785 lock_res_and_lock(lock);
786 ldlm_lock_addref_internal_nolock(lock, mode);
787 unlock_res_and_lock(lock);
791 * Removes reader/writer reference for LDLM lock \a lock.
792 * Assumes LDLM lock is already locked.
793 * only called in ldlm_flock_destroy and for local locks.
794 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
795 * that cannot be placed in LRU.
797 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
800 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
801 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
802 LASSERT(lock->l_readers > 0);
803 lu_ref_del(&lock->l_reference, "reader", lock);
806 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
807 LASSERT(lock->l_writers > 0);
808 lu_ref_del(&lock->l_reference, "writer", lock);
812 lu_ref_del(&lock->l_reference, "user", lock);
813 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
817 * Removes reader/writer reference for LDLM lock \a lock.
818 * Locks LDLM lock first.
819 * If the lock is determined to be client lock on a client and r/w refcount
820 * drops to zero and the lock is not blocked, the lock is added to LRU lock
822 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
824 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
826 struct ldlm_namespace *ns;
829 lock_res_and_lock(lock);
831 ns = ldlm_lock_to_ns(lock);
833 ldlm_lock_decref_internal_nolock(lock, mode);
835 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
836 !lock->l_readers && !lock->l_writers) {
837 /* If this is a local lock on a server namespace and this was
838 * the last reference, cancel the lock.
840 * Group locks are special:
841 * They must not go in LRU, but they are not called back
842 * like non-group locks, instead they are manually released.
843 * They have an l_writers reference which they keep until
844 * they are manually released, so we remove them when they have
845 * no more reader or writer references. - LU-6368 */
846 ldlm_set_cbpending(lock);
849 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
850 /* If we received a blocked AST and this was the last reference,
851 * run the callback. */
852 if (ldlm_is_ns_srv(lock) && lock->l_export)
853 CERROR("FL_CBPENDING set on non-local lock--just a "
856 LDLM_DEBUG(lock, "final decref done on cbpending lock");
858 LDLM_LOCK_GET(lock); /* dropped by bl thread */
859 ldlm_lock_remove_from_lru(lock);
860 unlock_res_and_lock(lock);
862 if (ldlm_is_fail_loc(lock))
863 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
865 if (ldlm_is_atomic_cb(lock) ||
866 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
867 ldlm_handle_bl_callback(ns, NULL, lock);
868 } else if (ns_is_client(ns) &&
869 !lock->l_readers && !lock->l_writers &&
870 !ldlm_is_no_lru(lock) &&
871 !ldlm_is_bl_ast(lock)) {
873 LDLM_DEBUG(lock, "add lock into lru list");
875 /* If this is a client-side namespace and this was the last
876 * reference, put it on the LRU. */
877 ldlm_lock_add_to_lru(lock);
878 unlock_res_and_lock(lock);
880 if (ldlm_is_fail_loc(lock))
881 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
883 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
884 * are not supported by the server, otherwise, it is done on
886 if (!exp_connect_cancelset(lock->l_conn_export) &&
887 !ns_connect_lru_resize(ns))
888 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
890 LDLM_DEBUG(lock, "do not add lock into lru list");
891 unlock_res_and_lock(lock);
898 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
900 void ldlm_lock_decref(struct lustre_handle *lockh, enum ldlm_mode mode)
902 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
903 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
904 ldlm_lock_decref_internal(lock, mode);
907 EXPORT_SYMBOL(ldlm_lock_decref);
910 * Decrease reader/writer refcount for LDLM lock with handle
911 * \a lockh and mark it for subsequent cancellation once r/w refcount
912 * drops to zero instead of putting into LRU.
915 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh,
918 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
921 LASSERT(lock != NULL);
923 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
924 lock_res_and_lock(lock);
925 ldlm_set_cbpending(lock);
926 unlock_res_and_lock(lock);
927 ldlm_lock_decref_internal(lock, mode);
930 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
932 struct sl_insert_point {
933 struct list_head *res_link;
934 struct list_head *mode_link;
935 struct list_head *policy_link;
939 * Finds a position to insert the new lock into granted lock list.
941 * Used for locks eligible for skiplist optimization.
944 * queue [input]: the granted list where search acts on;
945 * req [input]: the lock whose position to be located;
946 * prev [output]: positions within 3 lists to insert @req to
950 * - ldlm_grant_lock_with_skiplist
952 static void search_granted_lock(struct list_head *queue,
953 struct ldlm_lock *req,
954 struct sl_insert_point *prev)
956 struct list_head *tmp;
957 struct ldlm_lock *lock, *mode_end, *policy_end;
960 list_for_each(tmp, queue) {
961 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
963 mode_end = list_entry(lock->l_sl_mode.prev,
964 struct ldlm_lock, l_sl_mode);
966 if (lock->l_req_mode != req->l_req_mode) {
967 /* jump to last lock of mode group */
968 tmp = &mode_end->l_res_link;
972 /* suitable mode group is found */
973 if (lock->l_resource->lr_type == LDLM_PLAIN) {
974 /* insert point is last lock of the mode group */
975 prev->res_link = &mode_end->l_res_link;
976 prev->mode_link = &mode_end->l_sl_mode;
977 prev->policy_link = &req->l_sl_policy;
980 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
983 list_entry(lock->l_sl_policy.prev,
987 if (lock->l_policy_data.l_inodebits.bits ==
988 req->l_policy_data.l_inodebits.bits) {
989 /* insert point is last lock of
990 * the policy group */
992 &policy_end->l_res_link;
994 &policy_end->l_sl_mode;
996 &policy_end->l_sl_policy;
1001 if (policy_end == mode_end)
1002 /* done with mode group */
1005 /* go to next policy group within mode group */
1006 tmp = policy_end->l_res_link.next;
1007 lock = list_entry(tmp, struct ldlm_lock,
1009 } /* loop over policy groups within the mode group */
1011 /* insert point is last lock of the mode group,
1012 * new policy group is started */
1013 prev->res_link = &mode_end->l_res_link;
1014 prev->mode_link = &mode_end->l_sl_mode;
1015 prev->policy_link = &req->l_sl_policy;
1019 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1024 /* insert point is last lock on the queue,
1025 * new mode group and new policy group are started */
1026 prev->res_link = queue->prev;
1027 prev->mode_link = &req->l_sl_mode;
1028 prev->policy_link = &req->l_sl_policy;
1034 * Add a lock into resource granted list after a position described by
1037 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1038 struct sl_insert_point *prev)
1040 struct ldlm_resource *res = lock->l_resource;
1043 check_res_locked(res);
1045 ldlm_resource_dump(D_INFO, res);
1046 LDLM_DEBUG(lock, "About to add lock:");
1048 if (ldlm_is_destroyed(lock)) {
1049 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1053 LASSERT(list_empty(&lock->l_res_link));
1054 LASSERT(list_empty(&lock->l_sl_mode));
1055 LASSERT(list_empty(&lock->l_sl_policy));
1058 * lock->link == prev->link means lock is first starting the group.
1059 * Don't re-add to itself to suppress kernel warnings.
1061 if (&lock->l_res_link != prev->res_link)
1062 list_add(&lock->l_res_link, prev->res_link);
1063 if (&lock->l_sl_mode != prev->mode_link)
1064 list_add(&lock->l_sl_mode, prev->mode_link);
1065 if (&lock->l_sl_policy != prev->policy_link)
1066 list_add(&lock->l_sl_policy, prev->policy_link);
1072 * Add a lock to granted list on a resource maintaining skiplist
1075 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1077 struct sl_insert_point prev;
1080 LASSERT(lock->l_req_mode == lock->l_granted_mode);
1082 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1083 ldlm_granted_list_add_lock(lock, &prev);
1088 * Perform lock granting bookkeeping.
1090 * Includes putting the lock into granted list and updating lock mode.
1092 * - ldlm_lock_enqueue
1093 * - ldlm_reprocess_queue
1094 * - ldlm_lock_convert
1096 * must be called with lr_lock held
1098 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1100 struct ldlm_resource *res = lock->l_resource;
1103 check_res_locked(res);
1105 lock->l_granted_mode = lock->l_req_mode;
1107 if (work_list && lock->l_completion_ast != NULL)
1108 ldlm_add_ast_work_item(lock, NULL, work_list);
1110 /* We should not add locks to granted list in the following cases:
1111 * - this is an UNLOCK but not a real lock;
1112 * - this is a TEST lock;
1113 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1114 * - this is a deadlock (flock cannot be granted) */
1115 if (lock->l_req_mode == 0 ||
1116 lock->l_req_mode == LCK_NL ||
1117 ldlm_is_test_lock(lock) ||
1118 ldlm_is_flock_deadlock(lock))
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);
1126 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1128 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1133 * Describe the overlap between two locks. itree_overlap_cb data.
1135 struct lock_match_data {
1136 struct ldlm_lock *lmd_old;
1137 struct ldlm_lock *lmd_lock;
1138 enum ldlm_mode *lmd_mode;
1139 union ldlm_policy_data *lmd_policy;
1145 * Check if the given @lock meets the criteria for a match.
1146 * A reference on the lock is taken if matched.
1148 * \param lock test-against this lock
1149 * \param data parameters
1151 static int lock_matches(struct ldlm_lock *lock, struct lock_match_data *data)
1153 union ldlm_policy_data *lpol = &lock->l_policy_data;
1154 enum ldlm_mode match;
1156 if (lock == data->lmd_old)
1157 return INTERVAL_ITER_STOP;
1159 /* Check if this lock can be matched.
1160 * Used by LU-2919(exclusive open) for open lease lock */
1161 if (ldlm_is_excl(lock))
1162 return INTERVAL_ITER_CONT;
1164 /* llite sometimes wants to match locks that will be
1165 * canceled when their users drop, but we allow it to match
1166 * if it passes in CBPENDING and the lock still has users.
1167 * this is generally only going to be used by children
1168 * whose parents already hold a lock so forward progress
1169 * can still happen. */
1170 if (ldlm_is_cbpending(lock) &&
1171 !(data->lmd_flags & LDLM_FL_CBPENDING))
1172 return INTERVAL_ITER_CONT;
1173 if (!data->lmd_unref && ldlm_is_cbpending(lock) &&
1174 lock->l_readers == 0 && lock->l_writers == 0)
1175 return INTERVAL_ITER_CONT;
1177 if (!(lock->l_req_mode & *data->lmd_mode))
1178 return INTERVAL_ITER_CONT;
1179 match = lock->l_req_mode;
1181 switch (lock->l_resource->lr_type) {
1183 if (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1184 lpol->l_extent.end < data->lmd_policy->l_extent.end)
1185 return INTERVAL_ITER_CONT;
1187 if (unlikely(match == LCK_GROUP) &&
1188 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1189 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1190 return INTERVAL_ITER_CONT;
1193 /* We match if we have existing lock with same or wider set
1195 if ((lpol->l_inodebits.bits &
1196 data->lmd_policy->l_inodebits.bits) !=
1197 data->lmd_policy->l_inodebits.bits)
1198 return INTERVAL_ITER_CONT;
1204 /* We match if we have existing lock with same or wider set
1206 if (!data->lmd_unref && LDLM_HAVE_MASK(lock, GONE))
1207 return INTERVAL_ITER_CONT;
1209 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1210 return INTERVAL_ITER_CONT;
1212 if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1213 LDLM_LOCK_GET(lock);
1214 ldlm_lock_touch_in_lru(lock);
1216 ldlm_lock_addref_internal_nolock(lock, match);
1219 *data->lmd_mode = match;
1220 data->lmd_lock = lock;
1222 return INTERVAL_ITER_STOP;
1225 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1227 struct ldlm_interval *node = to_ldlm_interval(in);
1228 struct lock_match_data *data = args;
1229 struct ldlm_lock *lock;
1232 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1233 rc = lock_matches(lock, data);
1234 if (rc == INTERVAL_ITER_STOP)
1235 return INTERVAL_ITER_STOP;
1237 return INTERVAL_ITER_CONT;
1241 * Search for a lock with given parameters in interval trees.
1243 * \param res search for a lock in this resource
1244 * \param data parameters
1246 * \retval a referenced lock or NULL.
1248 static struct ldlm_lock *search_itree(struct ldlm_resource *res,
1249 struct lock_match_data *data)
1251 struct interval_node_extent ext = {
1252 .start = data->lmd_policy->l_extent.start,
1253 .end = data->lmd_policy->l_extent.end
1257 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1258 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1260 if (tree->lit_root == NULL)
1263 if (!(tree->lit_mode & *data->lmd_mode))
1266 interval_search(tree->lit_root, &ext,
1267 itree_overlap_cb, data);
1269 return data->lmd_lock;
1274 * Search for a lock with given properties in a queue.
1276 * \param queue search for a lock in this queue
1277 * \param data parameters
1279 * \retval a referenced lock or NULL.
1281 static struct ldlm_lock *search_queue(struct list_head *queue,
1282 struct lock_match_data *data)
1284 struct ldlm_lock *lock;
1287 list_for_each_entry(lock, queue, l_res_link) {
1288 rc = lock_matches(lock, data);
1289 if (rc == INTERVAL_ITER_STOP)
1290 return data->lmd_lock;
1295 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1297 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1298 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1299 wake_up_all(&lock->l_waitq);
1302 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1304 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1306 lock_res_and_lock(lock);
1307 ldlm_lock_fail_match_locked(lock);
1308 unlock_res_and_lock(lock);
1312 * Mark lock as "matchable" by OST.
1314 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1316 * Assumes LDLM lock is already locked.
1318 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1320 ldlm_set_lvb_ready(lock);
1321 wake_up_all(&lock->l_waitq);
1323 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1326 * Mark lock as "matchable" by OST.
1327 * Locks the lock and then \see ldlm_lock_allow_match_locked
1329 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1331 lock_res_and_lock(lock);
1332 ldlm_lock_allow_match_locked(lock);
1333 unlock_res_and_lock(lock);
1335 EXPORT_SYMBOL(ldlm_lock_allow_match);
1338 * Attempt to find a lock with specified properties.
1340 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1343 * Can be called in two ways:
1345 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1346 * for a duplicate of.
1348 * Otherwise, all of the fields must be filled in, to match against.
1350 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1351 * server (ie, connh is NULL)
1352 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1353 * list will be considered
1354 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1355 * to be canceled can still be matched as long as they still have reader
1356 * or writer refernces
1357 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1358 * just tell us if we would have matched.
1360 * \retval 1 if it finds an already-existing lock that is compatible; in this
1361 * case, lockh is filled in with a addref()ed lock
1363 * We also check security context, and if that fails we simply return 0 (to
1364 * keep caller code unchanged), the context failure will be discovered by
1365 * caller sometime later.
1367 enum ldlm_mode ldlm_lock_match(struct ldlm_namespace *ns, __u64 flags,
1368 const struct ldlm_res_id *res_id,
1369 enum ldlm_type type,
1370 union ldlm_policy_data *policy,
1371 enum ldlm_mode mode,
1372 struct lustre_handle *lockh, int unref)
1374 struct lock_match_data data = {
1378 .lmd_policy = policy,
1382 struct ldlm_resource *res;
1383 struct ldlm_lock *lock;
1388 data.lmd_old = ldlm_handle2lock(lockh);
1389 LASSERT(data.lmd_old != NULL);
1391 ns = ldlm_lock_to_ns(data.lmd_old);
1392 res_id = &data.lmd_old->l_resource->lr_name;
1393 type = data.lmd_old->l_resource->lr_type;
1394 *data.lmd_mode = data.lmd_old->l_req_mode;
1397 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1399 LASSERT(data.lmd_old == NULL);
1403 LDLM_RESOURCE_ADDREF(res);
1406 if (res->lr_type == LDLM_EXTENT)
1407 lock = search_itree(res, &data);
1409 lock = search_queue(&res->lr_granted, &data);
1412 if (flags & LDLM_FL_BLOCK_GRANTED)
1414 lock = search_queue(&res->lr_converting, &data);
1417 lock = search_queue(&res->lr_waiting, &data);
1424 LDLM_RESOURCE_DELREF(res);
1425 ldlm_resource_putref(res);
1428 ldlm_lock2handle(lock, lockh);
1429 if ((flags & LDLM_FL_LVB_READY) &&
1430 (!ldlm_is_lvb_ready(lock))) {
1431 __u64 wait_flags = LDLM_FL_LVB_READY |
1432 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1433 struct l_wait_info lwi;
1434 if (lock->l_completion_ast) {
1435 int err = lock->l_completion_ast(lock,
1436 LDLM_FL_WAIT_NOREPROC,
1439 if (flags & LDLM_FL_TEST_LOCK)
1440 LDLM_LOCK_RELEASE(lock);
1442 ldlm_lock_decref_internal(lock,
1449 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1450 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1452 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1453 l_wait_event(lock->l_waitq,
1454 lock->l_flags & wait_flags,
1456 if (!ldlm_is_lvb_ready(lock)) {
1457 if (flags & LDLM_FL_TEST_LOCK)
1458 LDLM_LOCK_RELEASE(lock);
1460 ldlm_lock_decref_internal(lock, mode);
1467 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1468 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1469 res_id->name[2] : policy->l_extent.start,
1470 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1471 res_id->name[3] : policy->l_extent.end);
1473 /* check user's security context */
1474 if (lock->l_conn_export &&
1475 sptlrpc_import_check_ctx(
1476 class_exp2cliimp(lock->l_conn_export))) {
1477 if (!(flags & LDLM_FL_TEST_LOCK))
1478 ldlm_lock_decref_internal(lock, mode);
1482 if (flags & LDLM_FL_TEST_LOCK)
1483 LDLM_LOCK_RELEASE(lock);
1485 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1486 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1487 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1488 type, mode, res_id->name[0], res_id->name[1],
1489 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1490 res_id->name[2] :policy->l_extent.start,
1491 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1492 res_id->name[3] : policy->l_extent.end);
1494 if (data.lmd_old != NULL)
1495 LDLM_LOCK_PUT(data.lmd_old);
1497 return rc ? mode : 0;
1499 EXPORT_SYMBOL(ldlm_lock_match);
1501 enum ldlm_mode ldlm_revalidate_lock_handle(struct lustre_handle *lockh,
1504 struct ldlm_lock *lock;
1505 enum ldlm_mode mode = 0;
1508 lock = ldlm_handle2lock(lockh);
1510 lock_res_and_lock(lock);
1511 if (LDLM_HAVE_MASK(lock, GONE))
1514 if (ldlm_is_cbpending(lock) &&
1515 lock->l_readers == 0 && lock->l_writers == 0)
1519 *bits = lock->l_policy_data.l_inodebits.bits;
1520 mode = lock->l_granted_mode;
1521 ldlm_lock_addref_internal_nolock(lock, mode);
1528 unlock_res_and_lock(lock);
1529 LDLM_LOCK_PUT(lock);
1533 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1535 /** The caller must guarantee that the buffer is large enough. */
1536 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1537 enum req_location loc, void *data, int size)
1542 LASSERT(data != NULL);
1545 switch (lock->l_lvb_type) {
1547 if (size == sizeof(struct ost_lvb)) {
1548 if (loc == RCL_CLIENT)
1549 lvb = req_capsule_client_swab_get(pill,
1551 lustre_swab_ost_lvb);
1553 lvb = req_capsule_server_swab_get(pill,
1555 lustre_swab_ost_lvb);
1556 if (unlikely(lvb == NULL)) {
1557 LDLM_ERROR(lock, "no LVB");
1561 memcpy(data, lvb, size);
1562 } else if (size == sizeof(struct ost_lvb_v1)) {
1563 struct ost_lvb *olvb = data;
1565 if (loc == RCL_CLIENT)
1566 lvb = req_capsule_client_swab_get(pill,
1568 lustre_swab_ost_lvb_v1);
1570 lvb = req_capsule_server_sized_swab_get(pill,
1572 lustre_swab_ost_lvb_v1);
1573 if (unlikely(lvb == NULL)) {
1574 LDLM_ERROR(lock, "no LVB");
1578 memcpy(data, lvb, size);
1579 olvb->lvb_mtime_ns = 0;
1580 olvb->lvb_atime_ns = 0;
1581 olvb->lvb_ctime_ns = 0;
1583 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1589 if (size == sizeof(struct lquota_lvb)) {
1590 if (loc == RCL_CLIENT)
1591 lvb = req_capsule_client_swab_get(pill,
1593 lustre_swab_lquota_lvb);
1595 lvb = req_capsule_server_swab_get(pill,
1597 lustre_swab_lquota_lvb);
1598 if (unlikely(lvb == NULL)) {
1599 LDLM_ERROR(lock, "no LVB");
1603 memcpy(data, lvb, size);
1605 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1614 if (loc == RCL_CLIENT)
1615 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1617 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1618 if (unlikely(lvb == NULL)) {
1619 LDLM_ERROR(lock, "no LVB");
1623 memcpy(data, lvb, size);
1626 LDLM_ERROR(lock, "Unknown LVB type: %d\n", lock->l_lvb_type);
1627 libcfs_debug_dumpstack(NULL);
1635 * Create and fill in new LDLM lock with specified properties.
1636 * Returns a referenced lock
1638 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1639 const struct ldlm_res_id *res_id,
1640 enum ldlm_type type,
1641 enum ldlm_mode mode,
1642 const struct ldlm_callback_suite *cbs,
1643 void *data, __u32 lvb_len,
1644 enum lvb_type lvb_type)
1646 struct ldlm_lock *lock;
1647 struct ldlm_resource *res;
1651 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1653 RETURN(ERR_CAST(res));
1655 lock = ldlm_lock_new(res);
1657 RETURN(ERR_PTR(-ENOMEM));
1659 lock->l_req_mode = mode;
1660 lock->l_ast_data = data;
1661 lock->l_pid = current_pid();
1662 if (ns_is_server(ns))
1663 ldlm_set_ns_srv(lock);
1665 lock->l_blocking_ast = cbs->lcs_blocking;
1666 lock->l_completion_ast = cbs->lcs_completion;
1667 lock->l_glimpse_ast = cbs->lcs_glimpse;
1670 lock->l_tree_node = NULL;
1671 /* if this is the extent lock, allocate the interval tree node */
1672 if (type == LDLM_EXTENT)
1673 if (ldlm_interval_alloc(lock) == NULL)
1674 GOTO(out, rc = -ENOMEM);
1677 lock->l_lvb_len = lvb_len;
1678 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1679 if (lock->l_lvb_data == NULL)
1680 GOTO(out, rc = -ENOMEM);
1683 lock->l_lvb_type = lvb_type;
1684 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1685 GOTO(out, rc = -ENOENT);
1690 ldlm_lock_destroy(lock);
1691 LDLM_LOCK_RELEASE(lock);
1692 RETURN(ERR_PTR(rc));
1696 * Enqueue (request) a lock.
1698 * Does not block. As a result of enqueue the lock would be put
1699 * into granted or waiting list.
1701 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1702 * set, skip all the enqueueing and delegate lock processing to intent policy
1705 enum ldlm_error ldlm_lock_enqueue(struct ldlm_namespace *ns,
1706 struct ldlm_lock **lockp,
1707 void *cookie, __u64 *flags)
1709 struct ldlm_lock *lock = *lockp;
1710 struct ldlm_resource *res = lock->l_resource;
1711 int local = ns_is_client(ldlm_res_to_ns(res));
1712 #ifdef HAVE_SERVER_SUPPORT
1713 ldlm_processing_policy policy;
1715 enum ldlm_error rc = ELDLM_OK;
1716 struct ldlm_interval *node = NULL;
1719 /* policies are not executed on the client or during replay */
1720 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1721 && !local && ns->ns_policy) {
1722 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1724 if (rc == ELDLM_LOCK_REPLACED) {
1725 /* The lock that was returned has already been granted,
1726 * and placed into lockp. If it's not the same as the
1727 * one we passed in, then destroy the old one and our
1728 * work here is done. */
1729 if (lock != *lockp) {
1730 ldlm_lock_destroy(lock);
1731 LDLM_LOCK_RELEASE(lock);
1733 *flags |= LDLM_FL_LOCK_CHANGED;
1735 } else if (rc != ELDLM_OK ||
1736 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1737 ldlm_lock_destroy(lock);
1742 if (*flags & LDLM_FL_RESENT) {
1743 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1744 * Set LOCK_CHANGED always.
1745 * Check if the lock is granted for BLOCK_GRANTED.
1746 * Take NO_TIMEOUT from the lock as it is inherited through
1747 * LDLM_FL_INHERIT_MASK */
1748 *flags |= LDLM_FL_LOCK_CHANGED;
1749 if (lock->l_req_mode != lock->l_granted_mode)
1750 *flags |= LDLM_FL_BLOCK_GRANTED;
1751 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1755 /* For a replaying lock, it might be already in granted list. So
1756 * unlinking the lock will cause the interval node to be freed, we
1757 * have to allocate the interval node early otherwise we can't regrant
1758 * this lock in the future. - jay */
1759 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1760 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1762 lock_res_and_lock(lock);
1763 if (local && lock->l_req_mode == lock->l_granted_mode) {
1764 /* The server returned a blocked lock, but it was granted
1765 * before we got a chance to actually enqueue it. We don't
1766 * need to do anything else. */
1767 *flags &= ~LDLM_FL_BLOCKED_MASK;
1768 GOTO(out, rc = ELDLM_OK);
1771 ldlm_resource_unlink_lock(lock);
1772 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1774 ldlm_lock_destroy_nolock(lock);
1775 GOTO(out, rc = -ENOMEM);
1778 INIT_LIST_HEAD(&node->li_group);
1779 ldlm_interval_attach(node, lock);
1783 /* Some flags from the enqueue want to make it into the AST, via the
1784 * lock's l_flags. */
1785 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1786 ldlm_set_ast_discard_data(lock);
1787 if (*flags & LDLM_FL_TEST_LOCK)
1788 ldlm_set_test_lock(lock);
1790 /* This distinction between local lock trees is very important; a client
1791 * namespace only has information about locks taken by that client, and
1792 * thus doesn't have enough information to decide for itself if it can
1793 * be granted (below). In this case, we do exactly what the server
1794 * tells us to do, as dictated by the 'flags'.
1796 * We do exactly the same thing during recovery, when the server is
1797 * more or less trusting the clients not to lie.
1799 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1800 * granted/converting queues. */
1802 if (*flags & LDLM_FL_BLOCK_CONV)
1803 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1804 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1805 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1807 ldlm_grant_lock(lock, NULL);
1808 GOTO(out, rc = ELDLM_OK);
1809 #ifdef HAVE_SERVER_SUPPORT
1810 } else if (*flags & LDLM_FL_REPLAY) {
1811 if (*flags & LDLM_FL_BLOCK_CONV) {
1812 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1813 GOTO(out, rc = ELDLM_OK);
1814 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1815 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1816 GOTO(out, rc = ELDLM_OK);
1817 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1818 ldlm_grant_lock(lock, NULL);
1819 GOTO(out, rc = ELDLM_OK);
1821 /* If no flags, fall through to normal enqueue path. */
1824 policy = ldlm_processing_policy_table[res->lr_type];
1825 policy(lock, flags, 1, &rc, NULL);
1829 CERROR("This is client-side-only module, cannot handle "
1830 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1836 unlock_res_and_lock(lock);
1838 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1842 #ifdef HAVE_SERVER_SUPPORT
1844 * Iterate through all waiting locks on a given resource queue and attempt to
1847 * Must be called with resource lock held.
1849 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1850 struct list_head *work_list)
1852 struct list_head *tmp, *pos;
1853 ldlm_processing_policy policy;
1855 int rc = LDLM_ITER_CONTINUE;
1856 enum ldlm_error err;
1859 check_res_locked(res);
1861 policy = ldlm_processing_policy_table[res->lr_type];
1864 list_for_each_safe(tmp, pos, queue) {
1865 struct ldlm_lock *pending;
1867 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1869 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1872 rc = policy(pending, &flags, 0, &err, work_list);
1873 if (rc != LDLM_ITER_CONTINUE)
1882 * Process a call to blocking AST callback for a lock in ast_work list
1885 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
1887 struct ldlm_cb_set_arg *arg = opaq;
1888 struct ldlm_lock_desc d;
1890 struct ldlm_lock *lock;
1893 if (list_empty(arg->list))
1896 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
1898 /* nobody should touch l_bl_ast */
1899 lock_res_and_lock(lock);
1900 list_del_init(&lock->l_bl_ast);
1902 LASSERT(ldlm_is_ast_sent(lock));
1903 LASSERT(lock->l_bl_ast_run == 0);
1904 LASSERT(lock->l_blocking_lock);
1905 lock->l_bl_ast_run++;
1906 unlock_res_and_lock(lock);
1908 ldlm_lock2desc(lock->l_blocking_lock, &d);
1910 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
1911 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1912 lock->l_blocking_lock = NULL;
1913 LDLM_LOCK_RELEASE(lock);
1919 * Process a call to completion AST callback for a lock in ast_work list
1922 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
1924 struct ldlm_cb_set_arg *arg = opaq;
1926 struct ldlm_lock *lock;
1927 ldlm_completion_callback completion_callback;
1930 if (list_empty(arg->list))
1933 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
1935 /* It's possible to receive a completion AST before we've set
1936 * the l_completion_ast pointer: either because the AST arrived
1937 * before the reply, or simply because there's a small race
1938 * window between receiving the reply and finishing the local
1939 * enqueue. (bug 842)
1941 * This can't happen with the blocking_ast, however, because we
1942 * will never call the local blocking_ast until we drop our
1943 * reader/writer reference, which we won't do until we get the
1944 * reply and finish enqueueing. */
1946 /* nobody should touch l_cp_ast */
1947 lock_res_and_lock(lock);
1948 list_del_init(&lock->l_cp_ast);
1949 LASSERT(ldlm_is_cp_reqd(lock));
1950 /* save l_completion_ast since it can be changed by
1951 * mds_intent_policy(), see bug 14225 */
1952 completion_callback = lock->l_completion_ast;
1953 ldlm_clear_cp_reqd(lock);
1954 unlock_res_and_lock(lock);
1956 if (completion_callback != NULL)
1957 rc = completion_callback(lock, 0, (void *)arg);
1958 LDLM_LOCK_RELEASE(lock);
1964 * Process a call to revocation AST callback for a lock in ast_work list
1967 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
1969 struct ldlm_cb_set_arg *arg = opaq;
1970 struct ldlm_lock_desc desc;
1972 struct ldlm_lock *lock;
1975 if (list_empty(arg->list))
1978 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
1979 list_del_init(&lock->l_rk_ast);
1981 /* the desc just pretend to exclusive */
1982 ldlm_lock2desc(lock, &desc);
1983 desc.l_req_mode = LCK_EX;
1984 desc.l_granted_mode = 0;
1986 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1987 LDLM_LOCK_RELEASE(lock);
1993 * Process a call to glimpse AST callback for a lock in ast_work list
1995 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
1997 struct ldlm_cb_set_arg *arg = opaq;
1998 struct ldlm_glimpse_work *gl_work;
1999 struct ldlm_lock *lock;
2003 if (list_empty(arg->list))
2006 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2008 list_del_init(&gl_work->gl_list);
2010 lock = gl_work->gl_lock;
2012 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2013 arg->gl_desc = gl_work->gl_desc;
2015 /* invoke the actual glimpse callback */
2016 if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
2019 LDLM_LOCK_RELEASE(lock);
2021 if ((gl_work->gl_flags & LDLM_GL_WORK_NOFREE) == 0)
2022 OBD_FREE_PTR(gl_work);
2028 * Process list of locks in need of ASTs being sent.
2030 * Used on server to send multiple ASTs together instead of sending one by
2033 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2034 ldlm_desc_ast_t ast_type)
2036 struct ldlm_cb_set_arg *arg;
2037 set_producer_func work_ast_lock;
2040 if (list_empty(rpc_list))
2047 atomic_set(&arg->restart, 0);
2048 arg->list = rpc_list;
2051 case LDLM_WORK_BL_AST:
2052 arg->type = LDLM_BL_CALLBACK;
2053 work_ast_lock = ldlm_work_bl_ast_lock;
2055 case LDLM_WORK_CP_AST:
2056 arg->type = LDLM_CP_CALLBACK;
2057 work_ast_lock = ldlm_work_cp_ast_lock;
2059 case LDLM_WORK_REVOKE_AST:
2060 arg->type = LDLM_BL_CALLBACK;
2061 work_ast_lock = ldlm_work_revoke_ast_lock;
2063 case LDLM_WORK_GL_AST:
2064 arg->type = LDLM_GL_CALLBACK;
2065 work_ast_lock = ldlm_work_gl_ast_lock;
2071 /* We create a ptlrpc request set with flow control extension.
2072 * This request set will use the work_ast_lock function to produce new
2073 * requests and will send a new request each time one completes in order
2074 * to keep the number of requests in flight to ns_max_parallel_ast */
2075 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2076 work_ast_lock, arg);
2077 if (arg->set == NULL)
2078 GOTO(out, rc = -ENOMEM);
2080 ptlrpc_set_wait(arg->set);
2081 ptlrpc_set_destroy(arg->set);
2083 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2090 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
2092 ldlm_reprocess_all(res);
2093 return LDLM_ITER_CONTINUE;
2096 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2097 struct hlist_node *hnode, void *arg)
2099 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2102 rc = reprocess_one_queue(res, arg);
2104 return rc == LDLM_ITER_STOP;
2108 * Iterate through all resources on a namespace attempting to grant waiting
2111 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
2116 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2117 ldlm_reprocess_res, NULL, 0);
2123 * Try to grant all waiting locks on a resource.
2125 * Calls ldlm_reprocess_queue on converting and waiting queues.
2127 * Typically called after some resource locks are cancelled to see
2128 * if anything could be granted as a result of the cancellation.
2130 void ldlm_reprocess_all(struct ldlm_resource *res)
2132 struct list_head rpc_list;
2133 #ifdef HAVE_SERVER_SUPPORT
2137 INIT_LIST_HEAD(&rpc_list);
2138 /* Local lock trees don't get reprocessed. */
2139 if (ns_is_client(ldlm_res_to_ns(res))) {
2146 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
2147 if (rc == LDLM_ITER_CONTINUE)
2148 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
2151 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2153 if (rc == -ERESTART) {
2154 LASSERT(list_empty(&rpc_list));
2160 INIT_LIST_HEAD(&rpc_list);
2161 if (!ns_is_client(ldlm_res_to_ns(res))) {
2162 CERROR("This is client-side-only module, cannot handle "
2163 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2169 EXPORT_SYMBOL(ldlm_reprocess_all);
2171 static bool is_bl_done(struct ldlm_lock *lock)
2173 bool bl_done = true;
2175 if (!ldlm_is_bl_done(lock)) {
2176 lock_res_and_lock(lock);
2177 bl_done = ldlm_is_bl_done(lock);
2178 unlock_res_and_lock(lock);
2185 * Helper function to call blocking AST for LDLM lock \a lock in a
2186 * "cancelling" mode.
2188 void ldlm_cancel_callback(struct ldlm_lock *lock)
2190 check_res_locked(lock->l_resource);
2191 if (!ldlm_is_cancel(lock)) {
2192 ldlm_set_cancel(lock);
2193 if (lock->l_blocking_ast) {
2194 unlock_res_and_lock(lock);
2195 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2197 lock_res_and_lock(lock);
2199 LDLM_DEBUG(lock, "no blocking ast");
2202 /* only canceller can set bl_done bit */
2203 ldlm_set_bl_done(lock);
2204 wake_up_all(&lock->l_waitq);
2205 } else if (!ldlm_is_bl_done(lock)) {
2206 struct l_wait_info lwi = { 0 };
2208 /* The lock is guaranteed to have been canceled once
2209 * returning from this function. */
2210 unlock_res_and_lock(lock);
2211 l_wait_event(lock->l_waitq, is_bl_done(lock), &lwi);
2212 lock_res_and_lock(lock);
2217 * Remove skiplist-enabled LDLM lock \a req from granted list
2219 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2221 if (req->l_resource->lr_type != LDLM_PLAIN &&
2222 req->l_resource->lr_type != LDLM_IBITS)
2225 list_del_init(&req->l_sl_policy);
2226 list_del_init(&req->l_sl_mode);
2230 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2232 void ldlm_lock_cancel(struct ldlm_lock *lock)
2234 struct ldlm_resource *res;
2235 struct ldlm_namespace *ns;
2238 lock_res_and_lock(lock);
2240 res = lock->l_resource;
2241 ns = ldlm_res_to_ns(res);
2243 /* Please do not, no matter how tempting, remove this LBUG without
2244 * talking to me first. -phik */
2245 if (lock->l_readers || lock->l_writers) {
2246 LDLM_ERROR(lock, "lock still has references");
2250 if (ldlm_is_waited(lock))
2251 ldlm_del_waiting_lock(lock);
2253 /* Releases cancel callback. */
2254 ldlm_cancel_callback(lock);
2256 LASSERT(!ldlm_is_waited(lock));
2258 ldlm_resource_unlink_lock(lock);
2259 ldlm_lock_destroy_nolock(lock);
2261 if (lock->l_granted_mode == lock->l_req_mode)
2262 ldlm_pool_del(&ns->ns_pool, lock);
2264 /* Make sure we will not be called again for same lock what is possible
2265 * if not to zero out lock->l_granted_mode */
2266 lock->l_granted_mode = LCK_MINMODE;
2267 unlock_res_and_lock(lock);
2271 EXPORT_SYMBOL(ldlm_lock_cancel);
2274 * Set opaque data into the lock that only makes sense to upper layer.
2276 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
2278 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2283 if (lock->l_ast_data == NULL)
2284 lock->l_ast_data = data;
2285 if (lock->l_ast_data == data)
2287 LDLM_LOCK_PUT(lock);
2291 EXPORT_SYMBOL(ldlm_lock_set_data);
2293 struct export_cl_data {
2294 struct obd_export *ecl_exp;
2298 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2299 struct ldlm_lock *lock,
2300 struct export_cl_data *ecl)
2302 struct ldlm_resource *res;
2304 res = ldlm_resource_getref(lock->l_resource);
2306 ldlm_res_lvbo_update(res, NULL, 1);
2307 ldlm_lock_cancel(lock);
2308 if (!exp->exp_obd->obd_stopping)
2309 ldlm_reprocess_all(res);
2310 ldlm_resource_putref(res);
2313 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2314 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2315 exp, ecl->ecl_loop);
2320 * Iterator function for ldlm_export_cancel_locks.
2321 * Cancels passed locks.
2324 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2325 struct hlist_node *hnode, void *data)
2328 struct export_cl_data *ecl = (struct export_cl_data *)data;
2329 struct obd_export *exp = ecl->ecl_exp;
2330 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2332 LDLM_LOCK_GET(lock);
2333 ldlm_cancel_lock_for_export(exp, lock, ecl);
2334 LDLM_LOCK_RELEASE(lock);
2340 * Cancel all blocked locks for given export.
2342 * Typically called on client disconnection/eviction
2344 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2346 struct export_cl_data ecl = {
2351 while (!list_empty(&exp->exp_bl_list)) {
2352 struct ldlm_lock *lock;
2354 spin_lock_bh(&exp->exp_bl_list_lock);
2355 if (!list_empty(&exp->exp_bl_list)) {
2356 lock = list_entry(exp->exp_bl_list.next,
2357 struct ldlm_lock, l_exp_list);
2358 LDLM_LOCK_GET(lock);
2359 list_del_init(&lock->l_exp_list);
2363 spin_unlock_bh(&exp->exp_bl_list_lock);
2368 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2369 LDLM_LOCK_RELEASE(lock);
2372 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2373 "left on hash table %d.\n", exp, ecl.ecl_loop,
2374 atomic_read(&exp->exp_lock_hash->hs_count));
2376 return ecl.ecl_loop;
2380 * Cancel all locks for given export.
2382 * Typically called after client disconnection/eviction
2384 int ldlm_export_cancel_locks(struct obd_export *exp)
2386 struct export_cl_data ecl = {
2391 cfs_hash_for_each_empty(exp->exp_lock_hash,
2392 ldlm_cancel_locks_for_export_cb, &ecl);
2394 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2395 "left on hash table %d.\n", exp, ecl.ecl_loop,
2396 atomic_read(&exp->exp_lock_hash->hs_count));
2398 return ecl.ecl_loop;
2402 * Downgrade an exclusive lock.
2404 * A fast variant of ldlm_lock_convert for convertion of exclusive
2405 * locks. The convertion is always successful.
2406 * Used by Commit on Sharing (COS) code.
2408 * \param lock A lock to convert
2409 * \param new_mode new lock mode
2411 void ldlm_lock_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2415 LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
2416 LASSERT(new_mode == LCK_COS);
2418 lock_res_and_lock(lock);
2419 ldlm_resource_unlink_lock(lock);
2421 * Remove the lock from pool as it will be added again in
2422 * ldlm_grant_lock() called below.
2424 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2426 lock->l_req_mode = new_mode;
2427 ldlm_grant_lock(lock, NULL);
2428 unlock_res_and_lock(lock);
2429 ldlm_reprocess_all(lock->l_resource);
2433 EXPORT_SYMBOL(ldlm_lock_downgrade);
2436 * Attempt to convert already granted lock to a different mode.
2438 * While lock conversion is not currently used, future client-side
2439 * optimizations could take advantage of it to avoid discarding cached
2442 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock,
2443 enum ldlm_mode new_mode, __u32 *flags)
2445 struct list_head rpc_list;
2446 struct ldlm_resource *res;
2447 struct ldlm_namespace *ns;
2449 #ifdef HAVE_SERVER_SUPPORT
2451 struct sl_insert_point prev;
2453 struct ldlm_interval *node;
2456 INIT_LIST_HEAD(&rpc_list);
2457 /* Just return if mode is unchanged. */
2458 if (new_mode == lock->l_granted_mode) {
2459 *flags |= LDLM_FL_BLOCK_GRANTED;
2460 RETURN(lock->l_resource);
2463 /* I can't check the type of lock here because the bitlock of lock
2464 * is not held here, so do the allocation blindly. -jay */
2465 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
2466 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
2469 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
2470 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
2472 lock_res_and_lock(lock);
2474 res = lock->l_resource;
2475 ns = ldlm_res_to_ns(res);
2477 #ifdef HAVE_SERVER_SUPPORT
2478 old_mode = lock->l_req_mode;
2480 lock->l_req_mode = new_mode;
2481 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
2482 #ifdef HAVE_SERVER_SUPPORT
2483 /* remember the lock position where the lock might be
2484 * added back to the granted list later and also
2485 * remember the join mode for skiplist fixing. */
2486 prev.res_link = lock->l_res_link.prev;
2487 prev.mode_link = lock->l_sl_mode.prev;
2488 prev.policy_link = lock->l_sl_policy.prev;
2490 ldlm_resource_unlink_lock(lock);
2492 ldlm_resource_unlink_lock(lock);
2493 if (res->lr_type == LDLM_EXTENT) {
2494 /* FIXME: ugly code, I have to attach the lock to a
2495 * interval node again since perhaps it will be granted
2497 INIT_LIST_HEAD(&node->li_group);
2498 ldlm_interval_attach(node, lock);
2504 * Remove old lock from the pool before adding the lock with new
2505 * mode below in ->policy()
2507 ldlm_pool_del(&ns->ns_pool, lock);
2509 /* If this is a local resource, put it on the appropriate list. */
2510 if (ns_is_client(ldlm_res_to_ns(res))) {
2511 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
2512 ldlm_resource_add_lock(res, &res->lr_converting, lock);
2514 /* This should never happen, because of the way the
2515 * server handles conversions. */
2516 LDLM_ERROR(lock, "Erroneous flags %x on local lock\n",
2520 ldlm_grant_lock(lock, &rpc_list);
2522 /* FIXME: completion handling not with lr_lock held ! */
2523 if (lock->l_completion_ast)
2524 lock->l_completion_ast(lock, 0, NULL);
2526 #ifdef HAVE_SERVER_SUPPORT
2529 enum ldlm_error err;
2531 ldlm_processing_policy policy;
2533 policy = ldlm_processing_policy_table[res->lr_type];
2534 rc = policy(lock, &pflags, 0, &err, &rpc_list);
2535 if (rc == LDLM_ITER_STOP) {
2536 lock->l_req_mode = old_mode;
2537 if (res->lr_type == LDLM_EXTENT)
2538 ldlm_extent_add_lock(res, lock);
2540 ldlm_granted_list_add_lock(lock, &prev);
2544 *flags |= LDLM_FL_BLOCK_GRANTED;
2550 CERROR("This is client-side-only module, cannot handle "
2551 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2555 unlock_res_and_lock(lock);
2558 ldlm_run_ast_work(ns, &rpc_list, LDLM_WORK_CP_AST);
2560 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
2565 * Print lock with lock handle \a lockh description into debug log.
2567 * Used when printing all locks on a resource for debug purposes.
2569 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
2571 struct ldlm_lock *lock;
2573 if (!((libcfs_debug | D_ERROR) & level))
2576 lock = ldlm_handle2lock(lockh);
2580 LDLM_DEBUG_LIMIT(level, lock, "###");
2582 LDLM_LOCK_PUT(lock);
2584 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2587 * Print lock information with custom message into debug log.
2590 void _ldlm_lock_debug(struct ldlm_lock *lock,
2591 struct libcfs_debug_msg_data *msgdata,
2592 const char *fmt, ...)
2595 struct obd_export *exp = lock->l_export;
2596 struct ldlm_resource *resource = lock->l_resource;
2597 char *nid = "local";
2599 va_start(args, fmt);
2601 if (exp && exp->exp_connection) {
2602 nid = libcfs_nid2str(exp->exp_connection->c_peer.nid);
2603 } else if (exp && exp->exp_obd != NULL) {
2604 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2605 nid = libcfs_nid2str(imp->imp_connection->c_peer.nid);
2608 if (resource == NULL) {
2609 libcfs_debug_vmsg2(msgdata, fmt, args,
2610 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2611 "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" nid: %s "
2612 "remote: "LPX64" expref: %d pid: %u timeout: %lu "
2615 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2616 lock->l_readers, lock->l_writers,
2617 ldlm_lockname[lock->l_granted_mode],
2618 ldlm_lockname[lock->l_req_mode],
2619 lock->l_flags, nid, lock->l_remote_handle.cookie,
2620 exp ? atomic_read(&exp->exp_refcount) : -99,
2621 lock->l_pid, lock->l_callback_timeout, lock->l_lvb_type);
2626 switch (resource->lr_type) {
2628 libcfs_debug_vmsg2(msgdata, fmt, args,
2629 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2630 "res: "DLDLMRES" rrc: %d type: %s ["LPU64"->"LPU64"] "
2631 "(req "LPU64"->"LPU64") flags: "LPX64" nid: %s remote: "
2632 LPX64" expref: %d pid: %u timeout: %lu lvb_type: %d\n",
2633 ldlm_lock_to_ns_name(lock), lock,
2634 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2635 lock->l_readers, lock->l_writers,
2636 ldlm_lockname[lock->l_granted_mode],
2637 ldlm_lockname[lock->l_req_mode],
2639 atomic_read(&resource->lr_refcount),
2640 ldlm_typename[resource->lr_type],
2641 lock->l_policy_data.l_extent.start,
2642 lock->l_policy_data.l_extent.end,
2643 lock->l_req_extent.start, lock->l_req_extent.end,
2644 lock->l_flags, nid, lock->l_remote_handle.cookie,
2645 exp ? atomic_read(&exp->exp_refcount) : -99,
2646 lock->l_pid, lock->l_callback_timeout,
2651 libcfs_debug_vmsg2(msgdata, fmt, args,
2652 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2653 "res: "DLDLMRES" rrc: %d type: %s pid: %d "
2654 "["LPU64"->"LPU64"] flags: "LPX64" nid: %s "
2655 "remote: "LPX64" expref: %d pid: %u timeout: %lu\n",
2656 ldlm_lock_to_ns_name(lock), lock,
2657 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2658 lock->l_readers, lock->l_writers,
2659 ldlm_lockname[lock->l_granted_mode],
2660 ldlm_lockname[lock->l_req_mode],
2662 atomic_read(&resource->lr_refcount),
2663 ldlm_typename[resource->lr_type],
2664 lock->l_policy_data.l_flock.pid,
2665 lock->l_policy_data.l_flock.start,
2666 lock->l_policy_data.l_flock.end,
2667 lock->l_flags, nid, lock->l_remote_handle.cookie,
2668 exp ? atomic_read(&exp->exp_refcount) : -99,
2669 lock->l_pid, lock->l_callback_timeout);
2673 libcfs_debug_vmsg2(msgdata, fmt, args,
2674 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2675 "res: "DLDLMRES" bits "LPX64" rrc: %d type: %s "
2676 "flags: "LPX64" nid: %s remote: "LPX64" expref: %d "
2677 "pid: %u timeout: %lu lvb_type: %d\n",
2678 ldlm_lock_to_ns_name(lock),
2679 lock, lock->l_handle.h_cookie,
2680 atomic_read(&lock->l_refc),
2681 lock->l_readers, lock->l_writers,
2682 ldlm_lockname[lock->l_granted_mode],
2683 ldlm_lockname[lock->l_req_mode],
2685 lock->l_policy_data.l_inodebits.bits,
2686 atomic_read(&resource->lr_refcount),
2687 ldlm_typename[resource->lr_type],
2688 lock->l_flags, nid, lock->l_remote_handle.cookie,
2689 exp ? atomic_read(&exp->exp_refcount) : -99,
2690 lock->l_pid, lock->l_callback_timeout,
2695 libcfs_debug_vmsg2(msgdata, fmt, args,
2696 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2697 "res: "DLDLMRES" rrc: %d type: %s flags: "LPX64" "
2698 "nid: %s remote: "LPX64" expref: %d pid: %u "
2699 "timeout: %lu lvb_type: %d\n",
2700 ldlm_lock_to_ns_name(lock),
2701 lock, lock->l_handle.h_cookie,
2702 atomic_read(&lock->l_refc),
2703 lock->l_readers, lock->l_writers,
2704 ldlm_lockname[lock->l_granted_mode],
2705 ldlm_lockname[lock->l_req_mode],
2707 atomic_read(&resource->lr_refcount),
2708 ldlm_typename[resource->lr_type],
2709 lock->l_flags, nid, lock->l_remote_handle.cookie,
2710 exp ? atomic_read(&exp->exp_refcount) : -99,
2711 lock->l_pid, lock->l_callback_timeout,
2717 EXPORT_SYMBOL(_ldlm_lock_debug);