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;
850 lock_res_and_lock(lock);
852 ns = ldlm_lock_to_ns(lock);
854 ldlm_lock_decref_internal_nolock(lock, mode);
856 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
857 !lock->l_readers && !lock->l_writers) {
858 /* If this is a local lock on a server namespace and this was
859 * the last reference, cancel the lock.
861 * Group locks are special:
862 * They must not go in LRU, but they are not called back
863 * like non-group locks, instead they are manually released.
864 * They have an l_writers reference which they keep until
865 * they are manually released, so we remove them when they have
866 * no more reader or writer references. - LU-6368 */
867 ldlm_set_cbpending(lock);
870 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
871 /* If we received a blocked AST and this was the last reference,
872 * run the callback. */
873 if (ldlm_is_ns_srv(lock) && lock->l_export)
874 CERROR("FL_CBPENDING set on non-local lock--just a "
877 LDLM_DEBUG(lock, "final decref done on cbpending lock");
879 LDLM_LOCK_GET(lock); /* dropped by bl thread */
880 ldlm_lock_remove_from_lru(lock);
881 unlock_res_and_lock(lock);
883 if (ldlm_is_fail_loc(lock))
884 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
886 if (ldlm_is_atomic_cb(lock) ||
887 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
888 ldlm_handle_bl_callback(ns, NULL, lock);
889 } else if (ns_is_client(ns) &&
890 !lock->l_readers && !lock->l_writers &&
891 !ldlm_is_no_lru(lock) &&
892 !ldlm_is_bl_ast(lock) &&
893 !ldlm_is_converting(lock)) {
895 LDLM_DEBUG(lock, "add lock into lru list");
897 /* If this is a client-side namespace and this was the last
898 * reference, put it on the LRU. */
899 ldlm_lock_add_to_lru(lock);
900 unlock_res_and_lock(lock);
902 if (ldlm_is_fail_loc(lock))
903 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
905 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
906 * are not supported by the server, otherwise, it is done on
908 if (!exp_connect_cancelset(lock->l_conn_export) &&
909 !ns_connect_lru_resize(ns))
910 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
912 LDLM_DEBUG(lock, "do not add lock into lru list");
913 unlock_res_and_lock(lock);
920 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
922 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
924 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
925 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
926 ldlm_lock_decref_internal(lock, mode);
929 EXPORT_SYMBOL(ldlm_lock_decref);
932 * Decrease reader/writer refcount for LDLM lock with handle
933 * \a lockh and mark it for subsequent cancellation once r/w refcount
934 * drops to zero instead of putting into LRU.
937 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
940 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
943 LASSERT(lock != NULL);
945 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
946 lock_res_and_lock(lock);
947 ldlm_set_cbpending(lock);
948 unlock_res_and_lock(lock);
949 ldlm_lock_decref_internal(lock, mode);
952 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
954 struct sl_insert_point {
955 struct list_head *res_link;
956 struct list_head *mode_link;
957 struct list_head *policy_link;
961 * Finds a position to insert the new lock into granted lock list.
963 * Used for locks eligible for skiplist optimization.
966 * queue [input]: the granted list where search acts on;
967 * req [input]: the lock whose position to be located;
968 * prev [output]: positions within 3 lists to insert @req to
972 * - ldlm_grant_lock_with_skiplist
974 static void search_granted_lock(struct list_head *queue,
975 struct ldlm_lock *req,
976 struct sl_insert_point *prev)
978 struct list_head *tmp;
979 struct ldlm_lock *lock, *mode_end, *policy_end;
982 list_for_each(tmp, queue) {
983 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
985 mode_end = list_entry(lock->l_sl_mode.prev,
986 struct ldlm_lock, l_sl_mode);
988 if (lock->l_req_mode != req->l_req_mode) {
989 /* jump to last lock of mode group */
990 tmp = &mode_end->l_res_link;
994 /* suitable mode group is found */
995 if (lock->l_resource->lr_type == LDLM_PLAIN) {
996 /* insert point is last lock of the mode group */
997 prev->res_link = &mode_end->l_res_link;
998 prev->mode_link = &mode_end->l_sl_mode;
999 prev->policy_link = &req->l_sl_policy;
1002 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
1005 list_entry(lock->l_sl_policy.prev,
1009 if (lock->l_policy_data.l_inodebits.bits ==
1010 req->l_policy_data.l_inodebits.bits) {
1011 /* insert point is last lock of
1012 * the policy group */
1014 &policy_end->l_res_link;
1016 &policy_end->l_sl_mode;
1018 &policy_end->l_sl_policy;
1023 if (policy_end == mode_end)
1024 /* done with mode group */
1027 /* go to next policy group within mode group */
1028 tmp = policy_end->l_res_link.next;
1029 lock = list_entry(tmp, struct ldlm_lock,
1031 } /* loop over policy groups within the mode group */
1033 /* insert point is last lock of the mode group,
1034 * new policy group is started */
1035 prev->res_link = &mode_end->l_res_link;
1036 prev->mode_link = &mode_end->l_sl_mode;
1037 prev->policy_link = &req->l_sl_policy;
1041 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1046 /* insert point is last lock on the queue,
1047 * new mode group and new policy group are started */
1048 prev->res_link = queue->prev;
1049 prev->mode_link = &req->l_sl_mode;
1050 prev->policy_link = &req->l_sl_policy;
1055 * Add a lock into resource granted list after a position described by
1058 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1059 struct sl_insert_point *prev)
1061 struct ldlm_resource *res = lock->l_resource;
1064 check_res_locked(res);
1066 ldlm_resource_dump(D_INFO, res);
1067 LDLM_DEBUG(lock, "About to add lock:");
1069 if (ldlm_is_destroyed(lock)) {
1070 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1074 LASSERT(list_empty(&lock->l_res_link));
1075 LASSERT(list_empty(&lock->l_sl_mode));
1076 LASSERT(list_empty(&lock->l_sl_policy));
1079 * lock->link == prev->link means lock is first starting the group.
1080 * Don't re-add to itself to suppress kernel warnings.
1082 if (&lock->l_res_link != prev->res_link)
1083 list_add(&lock->l_res_link, prev->res_link);
1084 if (&lock->l_sl_mode != prev->mode_link)
1085 list_add(&lock->l_sl_mode, prev->mode_link);
1086 if (&lock->l_sl_policy != prev->policy_link)
1087 list_add(&lock->l_sl_policy, prev->policy_link);
1093 * Add a lock to granted list on a resource maintaining skiplist
1096 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1098 struct sl_insert_point prev;
1100 LASSERT(ldlm_is_granted(lock));
1102 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1103 ldlm_granted_list_add_lock(lock, &prev);
1107 * Perform lock granting bookkeeping.
1109 * Includes putting the lock into granted list and updating lock mode.
1111 * - ldlm_lock_enqueue
1112 * - ldlm_reprocess_queue
1114 * must be called with lr_lock held
1116 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1118 struct ldlm_resource *res = lock->l_resource;
1121 check_res_locked(res);
1123 lock->l_granted_mode = lock->l_req_mode;
1125 if (work_list && lock->l_completion_ast != NULL)
1126 ldlm_add_ast_work_item(lock, NULL, work_list);
1128 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1129 ldlm_grant_lock_with_skiplist(lock);
1130 else if (res->lr_type == LDLM_EXTENT)
1131 ldlm_extent_add_lock(res, lock);
1132 else if (res->lr_type == LDLM_FLOCK) {
1133 /* We should not add locks to granted list in the following
1135 * - this is an UNLOCK but not a real lock;
1136 * - this is a TEST lock;
1137 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1138 * - this is a deadlock (flock cannot be granted) */
1139 if (lock->l_req_mode == 0 ||
1140 lock->l_req_mode == LCK_NL ||
1141 ldlm_is_test_lock(lock) ||
1142 ldlm_is_flock_deadlock(lock))
1144 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1149 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1154 * Check if the given @lock meets the criteria for a match.
1155 * A reference on the lock is taken if matched.
1157 * \param lock test-against this lock
1158 * \param data parameters
1160 static int lock_matches(struct ldlm_lock *lock, struct ldlm_match_data *data)
1162 union ldlm_policy_data *lpol = &lock->l_policy_data;
1163 enum ldlm_mode match = LCK_MINMODE;
1165 if (lock == data->lmd_old)
1166 return INTERVAL_ITER_STOP;
1168 /* Check if this lock can be matched.
1169 * Used by LU-2919(exclusive open) for open lease lock */
1170 if (ldlm_is_excl(lock))
1171 return INTERVAL_ITER_CONT;
1173 /* llite sometimes wants to match locks that will be
1174 * canceled when their users drop, but we allow it to match
1175 * if it passes in CBPENDING and the lock still has users.
1176 * this is generally only going to be used by children
1177 * whose parents already hold a lock so forward progress
1178 * can still happen. */
1179 if (ldlm_is_cbpending(lock) &&
1180 !(data->lmd_flags & LDLM_FL_CBPENDING))
1181 return INTERVAL_ITER_CONT;
1182 if (!(data->lmd_match & LDLM_MATCH_UNREF) && ldlm_is_cbpending(lock) &&
1183 lock->l_readers == 0 && lock->l_writers == 0)
1184 return INTERVAL_ITER_CONT;
1186 if (!(lock->l_req_mode & *data->lmd_mode))
1187 return INTERVAL_ITER_CONT;
1189 /* When we search for ast_data, we are not doing a traditional match,
1190 * so we don't worry about IBITS or extent matching.
1192 if (data->lmd_match & (LDLM_MATCH_AST | LDLM_MATCH_AST_ANY)) {
1193 if (!lock->l_ast_data)
1194 return INTERVAL_ITER_CONT;
1196 if (data->lmd_match & LDLM_MATCH_AST_ANY)
1200 match = lock->l_req_mode;
1202 switch (lock->l_resource->lr_type) {
1204 if (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1205 lpol->l_extent.end < data->lmd_policy->l_extent.end)
1206 return INTERVAL_ITER_CONT;
1208 if (unlikely(match == LCK_GROUP) &&
1209 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1210 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1211 return INTERVAL_ITER_CONT;
1214 /* We match if we have existing lock with same or wider set
1216 if ((lpol->l_inodebits.bits &
1217 data->lmd_policy->l_inodebits.bits) !=
1218 data->lmd_policy->l_inodebits.bits)
1219 return INTERVAL_ITER_CONT;
1225 /* We match if we have existing lock with same or wider set
1227 if (!(data->lmd_match & LDLM_MATCH_UNREF) && LDLM_HAVE_MASK(lock, GONE))
1228 return INTERVAL_ITER_CONT;
1230 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1231 return INTERVAL_ITER_CONT;
1233 /* Filter locks by skipping flags */
1234 if (data->lmd_skip_flags & lock->l_flags)
1235 return INTERVAL_ITER_CONT;
1238 if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1239 LDLM_LOCK_GET(lock);
1240 ldlm_lock_touch_in_lru(lock);
1242 ldlm_lock_addref_internal_nolock(lock, match);
1245 *data->lmd_mode = match;
1246 data->lmd_lock = lock;
1248 return INTERVAL_ITER_STOP;
1251 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1253 struct ldlm_interval *node = to_ldlm_interval(in);
1254 struct ldlm_match_data *data = args;
1255 struct ldlm_lock *lock;
1258 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1259 rc = lock_matches(lock, data);
1260 if (rc == INTERVAL_ITER_STOP)
1261 return INTERVAL_ITER_STOP;
1263 return INTERVAL_ITER_CONT;
1267 * Search for a lock with given parameters in interval trees.
1269 * \param res search for a lock in this resource
1270 * \param data parameters
1272 * \retval a referenced lock or NULL.
1274 struct ldlm_lock *search_itree(struct ldlm_resource *res,
1275 struct ldlm_match_data *data)
1277 struct interval_node_extent ext = {
1278 .start = data->lmd_policy->l_extent.start,
1279 .end = data->lmd_policy->l_extent.end
1283 data->lmd_lock = NULL;
1285 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1286 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1288 if (tree->lit_root == NULL)
1291 if (!(tree->lit_mode & *data->lmd_mode))
1294 interval_search(tree->lit_root, &ext,
1295 itree_overlap_cb, data);
1297 return data->lmd_lock;
1302 EXPORT_SYMBOL(search_itree);
1306 * Search for a lock with given properties in a queue.
1308 * \param queue search for a lock in this queue
1309 * \param data parameters
1311 * \retval a referenced lock or NULL.
1313 static struct ldlm_lock *search_queue(struct list_head *queue,
1314 struct ldlm_match_data *data)
1316 struct ldlm_lock *lock;
1319 data->lmd_lock = NULL;
1321 list_for_each_entry(lock, queue, l_res_link) {
1322 rc = lock_matches(lock, data);
1323 if (rc == INTERVAL_ITER_STOP)
1324 return data->lmd_lock;
1330 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1332 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1333 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1334 wake_up_all(&lock->l_waitq);
1337 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1339 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1341 lock_res_and_lock(lock);
1342 ldlm_lock_fail_match_locked(lock);
1343 unlock_res_and_lock(lock);
1347 * Mark lock as "matchable" by OST.
1349 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1351 * Assumes LDLM lock is already locked.
1353 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1355 ldlm_set_lvb_ready(lock);
1356 wake_up_all(&lock->l_waitq);
1358 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1361 * Mark lock as "matchable" by OST.
1362 * Locks the lock and then \see ldlm_lock_allow_match_locked
1364 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1366 lock_res_and_lock(lock);
1367 ldlm_lock_allow_match_locked(lock);
1368 unlock_res_and_lock(lock);
1370 EXPORT_SYMBOL(ldlm_lock_allow_match);
1373 * Attempt to find a lock with specified properties.
1375 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1378 * Can be called in two ways:
1380 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1381 * for a duplicate of.
1383 * Otherwise, all of the fields must be filled in, to match against.
1385 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1386 * server (ie, connh is NULL)
1387 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1388 * list will be considered
1389 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1390 * to be canceled can still be matched as long as they still have reader
1391 * or writer refernces
1392 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1393 * just tell us if we would have matched.
1395 * \retval 1 if it finds an already-existing lock that is compatible; in this
1396 * case, lockh is filled in with a addref()ed lock
1398 * We also check security context, and if that fails we simply return 0 (to
1399 * keep caller code unchanged), the context failure will be discovered by
1400 * caller sometime later.
1402 enum ldlm_mode ldlm_lock_match_with_skip(struct ldlm_namespace *ns,
1403 __u64 flags, __u64 skip_flags,
1404 const struct ldlm_res_id *res_id,
1405 enum ldlm_type type,
1406 union ldlm_policy_data *policy,
1407 enum ldlm_mode mode,
1408 struct lustre_handle *lockh,
1409 enum ldlm_match_flags match_flags)
1411 struct ldlm_match_data data = {
1415 .lmd_policy = policy,
1417 .lmd_skip_flags = skip_flags,
1418 .lmd_match = match_flags,
1420 struct ldlm_resource *res;
1421 struct ldlm_lock *lock;
1427 data.lmd_old = ldlm_handle2lock(lockh);
1428 LASSERT(data.lmd_old != NULL);
1430 ns = ldlm_lock_to_ns(data.lmd_old);
1431 res_id = &data.lmd_old->l_resource->lr_name;
1432 type = data.lmd_old->l_resource->lr_type;
1433 *data.lmd_mode = data.lmd_old->l_req_mode;
1436 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1438 LASSERT(data.lmd_old == NULL);
1442 LDLM_RESOURCE_ADDREF(res);
1444 if (res->lr_type == LDLM_EXTENT)
1445 lock = search_itree(res, &data);
1447 lock = search_queue(&res->lr_granted, &data);
1448 if (!lock && !(flags & LDLM_FL_BLOCK_GRANTED))
1449 lock = search_queue(&res->lr_waiting, &data);
1450 matched = lock ? mode : 0;
1452 LDLM_RESOURCE_DELREF(res);
1453 ldlm_resource_putref(res);
1456 ldlm_lock2handle(lock, lockh);
1457 if ((flags & LDLM_FL_LVB_READY) &&
1458 (!ldlm_is_lvb_ready(lock))) {
1459 __u64 wait_flags = LDLM_FL_LVB_READY |
1460 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1462 if (lock->l_completion_ast) {
1463 int err = lock->l_completion_ast(lock,
1464 LDLM_FL_WAIT_NOREPROC,
1467 GOTO(out_fail_match, matched = 0);
1470 wait_event_idle_timeout(
1472 lock->l_flags & wait_flags,
1473 cfs_time_seconds(obd_timeout));
1475 if (!ldlm_is_lvb_ready(lock))
1476 GOTO(out_fail_match, matched = 0);
1479 /* check user's security context */
1480 if (lock->l_conn_export &&
1481 sptlrpc_import_check_ctx(
1482 class_exp2cliimp(lock->l_conn_export)))
1483 GOTO(out_fail_match, matched = 0);
1485 LDLM_DEBUG(lock, "matched (%llu %llu)",
1486 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1487 res_id->name[2] : policy->l_extent.start,
1488 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1489 res_id->name[3] : policy->l_extent.end);
1492 if (flags & LDLM_FL_TEST_LOCK)
1493 LDLM_LOCK_RELEASE(lock);
1495 ldlm_lock_decref_internal(lock, mode);
1498 /* less verbose for test-only */
1499 if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
1500 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1501 "%llu/%llu (%llu %llu)", ns,
1502 type, mode, res_id->name[0], res_id->name[1],
1503 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1504 res_id->name[2] : policy->l_extent.start,
1505 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1506 res_id->name[3] : policy->l_extent.end);
1508 if (data.lmd_old != NULL)
1509 LDLM_LOCK_PUT(data.lmd_old);
1513 EXPORT_SYMBOL(ldlm_lock_match_with_skip);
1515 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1518 struct ldlm_lock *lock;
1519 enum ldlm_mode mode = 0;
1522 lock = ldlm_handle2lock(lockh);
1524 lock_res_and_lock(lock);
1525 if (LDLM_HAVE_MASK(lock, GONE))
1528 if (ldlm_is_cbpending(lock) &&
1529 lock->l_readers == 0 && lock->l_writers == 0)
1533 *bits = lock->l_policy_data.l_inodebits.bits;
1534 mode = lock->l_granted_mode;
1535 ldlm_lock_addref_internal_nolock(lock, mode);
1542 unlock_res_and_lock(lock);
1543 LDLM_LOCK_PUT(lock);
1547 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1549 /** The caller must guarantee that the buffer is large enough. */
1550 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1551 enum req_location loc, void *data, int size)
1556 LASSERT(data != NULL);
1559 switch (lock->l_lvb_type) {
1561 if (size == sizeof(struct ost_lvb)) {
1562 if (loc == RCL_CLIENT)
1563 lvb = req_capsule_client_swab_get(pill,
1565 lustre_swab_ost_lvb);
1567 lvb = req_capsule_server_swab_get(pill,
1569 lustre_swab_ost_lvb);
1570 if (unlikely(lvb == NULL)) {
1571 LDLM_ERROR(lock, "no LVB");
1575 memcpy(data, lvb, size);
1576 } else if (size == sizeof(struct ost_lvb_v1)) {
1577 struct ost_lvb *olvb = data;
1579 if (loc == RCL_CLIENT)
1580 lvb = req_capsule_client_swab_get(pill,
1582 lustre_swab_ost_lvb_v1);
1584 lvb = req_capsule_server_sized_swab_get(pill,
1586 lustre_swab_ost_lvb_v1);
1587 if (unlikely(lvb == NULL)) {
1588 LDLM_ERROR(lock, "no LVB");
1592 memcpy(data, lvb, size);
1593 olvb->lvb_mtime_ns = 0;
1594 olvb->lvb_atime_ns = 0;
1595 olvb->lvb_ctime_ns = 0;
1597 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1603 if (size == sizeof(struct lquota_lvb)) {
1604 if (loc == RCL_CLIENT)
1605 lvb = req_capsule_client_swab_get(pill,
1607 lustre_swab_lquota_lvb);
1609 lvb = req_capsule_server_swab_get(pill,
1611 lustre_swab_lquota_lvb);
1612 if (unlikely(lvb == NULL)) {
1613 LDLM_ERROR(lock, "no LVB");
1617 memcpy(data, lvb, size);
1619 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1628 if (loc == RCL_CLIENT)
1629 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1631 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1632 if (unlikely(lvb == NULL)) {
1633 LDLM_ERROR(lock, "no LVB");
1637 memcpy(data, lvb, size);
1640 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1641 libcfs_debug_dumpstack(NULL);
1649 * Create and fill in new LDLM lock with specified properties.
1650 * Returns a referenced lock
1652 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1653 const struct ldlm_res_id *res_id,
1654 enum ldlm_type type,
1655 enum ldlm_mode mode,
1656 const struct ldlm_callback_suite *cbs,
1657 void *data, __u32 lvb_len,
1658 enum lvb_type lvb_type)
1660 struct ldlm_lock *lock;
1661 struct ldlm_resource *res;
1665 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1667 RETURN(ERR_CAST(res));
1669 lock = ldlm_lock_new(res);
1671 RETURN(ERR_PTR(-ENOMEM));
1673 lock->l_req_mode = mode;
1674 lock->l_ast_data = data;
1675 lock->l_pid = current->pid;
1676 if (ns_is_server(ns))
1677 ldlm_set_ns_srv(lock);
1679 lock->l_blocking_ast = cbs->lcs_blocking;
1680 lock->l_completion_ast = cbs->lcs_completion;
1681 lock->l_glimpse_ast = cbs->lcs_glimpse;
1686 rc = ldlm_extent_alloc_lock(lock);
1689 rc = ldlm_inodebits_alloc_lock(lock);
1698 lock->l_lvb_len = lvb_len;
1699 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1700 if (lock->l_lvb_data == NULL)
1701 GOTO(out, rc = -ENOMEM);
1704 lock->l_lvb_type = lvb_type;
1705 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1706 GOTO(out, rc = -ENOENT);
1711 ldlm_lock_destroy(lock);
1712 LDLM_LOCK_RELEASE(lock);
1713 RETURN(ERR_PTR(rc));
1716 #ifdef HAVE_SERVER_SUPPORT
1717 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1720 struct ldlm_resource *res = lock->l_resource;
1721 enum ldlm_error rc = ELDLM_OK;
1722 LIST_HEAD(rpc_list);
1723 ldlm_processing_policy policy;
1727 policy = ldlm_get_processing_policy(res);
1729 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1730 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1731 res->lr_type != LDLM_FLOCK) {
1732 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1733 if (rc == -ERESTART)
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);
2040 if (rc == -ERESTART) {
2041 /* 15715: The lock was granted and destroyed after
2042 * resource lock was dropped. Interval node was freed
2043 * in ldlm_lock_destroy. Anyway, this always happens
2044 * when a client is being evicted. So it would be
2045 * ok to return an error. -jay */
2046 if (ldlm_is_destroyed(lock))
2049 /* lock was granted while resource was unlocked. */
2050 if (ldlm_is_granted(lock)) {
2051 /* bug 11300: if the lock has been granted,
2052 * break earlier because otherwise, we will go
2053 * to restart and ldlm_resource_unlink will be
2054 * called and it causes the interval node to be
2055 * freed. Then we will fail at
2056 * ldlm_extent_add_lock() */
2057 *flags &= ~LDLM_FL_BLOCKED_MASK;
2063 *flags |= LDLM_FL_BLOCK_GRANTED;
2069 * Discard all AST work items from list.
2071 * If for whatever reason we do not want to send ASTs to conflicting locks
2072 * anymore, disassemble the list with this function.
2074 void ldlm_discard_bl_list(struct list_head *bl_list)
2076 struct ldlm_lock *lock, *tmp;
2080 list_for_each_entry_safe(lock, tmp, bl_list, l_bl_ast) {
2081 LASSERT(!list_empty(&lock->l_bl_ast));
2082 list_del_init(&lock->l_bl_ast);
2083 ldlm_clear_ast_sent(lock);
2084 LASSERT(lock->l_bl_ast_run == 0);
2085 ldlm_clear_blocking_lock(lock);
2086 LDLM_LOCK_RELEASE(lock);
2092 * Process a call to blocking AST callback for a lock in ast_work list
2095 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2097 struct ldlm_cb_set_arg *arg = opaq;
2098 struct ldlm_lock *lock;
2099 struct ldlm_lock_desc d;
2100 struct ldlm_bl_desc bld;
2105 if (list_empty(arg->list))
2108 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2110 /* nobody should touch l_bl_ast but some locks in the list may become
2111 * granted after lock convert or COS downgrade, these locks should be
2112 * just skipped here and removed from the list.
2114 lock_res_and_lock(lock);
2115 list_del_init(&lock->l_bl_ast);
2117 /* lock is not blocking lock anymore, but was kept in the list because
2118 * it can managed only here.
2120 if (!ldlm_is_ast_sent(lock)) {
2121 unlock_res_and_lock(lock);
2122 LDLM_LOCK_RELEASE(lock);
2126 LASSERT(lock->l_blocking_lock);
2127 ldlm_lock2desc(lock->l_blocking_lock, &d);
2128 /* copy blocking lock ibits in cancel_bits as well,
2129 * new client may use them for lock convert and it is
2130 * important to use new field to convert locks from
2133 d.l_policy_data.l_inodebits.cancel_bits =
2134 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2136 /* Blocking lock is being destroyed here but some information about it
2137 * may be needed inside l_blocking_ast() function below,
2138 * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
2140 bld.bl_same_client = lock->l_client_cookie ==
2141 lock->l_blocking_lock->l_client_cookie;
2142 bld.bl_cos_incompat = ldlm_is_cos_incompat(lock->l_blocking_lock);
2143 arg->bl_desc = &bld;
2145 LASSERT(ldlm_is_ast_sent(lock));
2146 LASSERT(lock->l_bl_ast_run == 0);
2147 lock->l_bl_ast_run++;
2148 ldlm_clear_blocking_lock(lock);
2149 unlock_res_and_lock(lock);
2151 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2153 LDLM_LOCK_RELEASE(lock);
2159 * Process a call to revocation AST callback for a lock in ast_work list
2162 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2164 struct ldlm_cb_set_arg *arg = opaq;
2165 struct ldlm_lock_desc desc;
2167 struct ldlm_lock *lock;
2170 if (list_empty(arg->list))
2173 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2174 list_del_init(&lock->l_rk_ast);
2176 /* the desc just pretend to exclusive */
2177 ldlm_lock2desc(lock, &desc);
2178 desc.l_req_mode = LCK_EX;
2179 desc.l_granted_mode = 0;
2181 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2182 LDLM_LOCK_RELEASE(lock);
2188 * Process a call to glimpse AST callback for a lock in ast_work list
2190 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2192 struct ldlm_cb_set_arg *arg = opaq;
2193 struct ldlm_glimpse_work *gl_work;
2194 struct ldlm_lock *lock;
2198 if (list_empty(arg->list))
2201 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2203 list_del_init(&gl_work->gl_list);
2205 lock = gl_work->gl_lock;
2207 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2208 arg->gl_desc = gl_work->gl_desc;
2209 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2210 arg->gl_interpret_data = gl_work->gl_interpret_data;
2212 /* invoke the actual glimpse callback */
2213 rc = lock->l_glimpse_ast(lock, (void *)arg);
2215 rc = 1; /* update LVB if this is server lock */
2216 else if (rc == -ELDLM_NO_LOCK_DATA)
2217 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
2219 LDLM_LOCK_RELEASE(lock);
2220 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2221 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2223 OBD_FREE_PTR(gl_work);
2230 * Process a call to completion AST callback for a lock in ast_work list
2233 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2235 struct ldlm_cb_set_arg *arg = opaq;
2236 struct ldlm_lock *lock;
2237 ldlm_completion_callback completion_callback;
2242 if (list_empty(arg->list))
2245 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2247 /* It's possible to receive a completion AST before we've set
2248 * the l_completion_ast pointer: either because the AST arrived
2249 * before the reply, or simply because there's a small race
2250 * window between receiving the reply and finishing the local
2251 * enqueue. (bug 842)
2253 * This can't happen with the blocking_ast, however, because we
2254 * will never call the local blocking_ast until we drop our
2255 * reader/writer reference, which we won't do until we get the
2256 * reply and finish enqueueing. */
2258 /* nobody should touch l_cp_ast */
2259 lock_res_and_lock(lock);
2260 list_del_init(&lock->l_cp_ast);
2261 LASSERT(ldlm_is_cp_reqd(lock));
2262 /* save l_completion_ast since it can be changed by
2263 * mds_intent_policy(), see bug 14225 */
2264 completion_callback = lock->l_completion_ast;
2265 ldlm_clear_cp_reqd(lock);
2266 unlock_res_and_lock(lock);
2268 if (completion_callback != NULL)
2269 rc = completion_callback(lock, 0, (void *)arg);
2270 LDLM_LOCK_RELEASE(lock);
2276 * Process list of locks in need of ASTs being sent.
2278 * Used on server to send multiple ASTs together instead of sending one by
2281 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2282 ldlm_desc_ast_t ast_type)
2284 struct ldlm_cb_set_arg *arg;
2285 set_producer_func work_ast_lock;
2288 if (list_empty(rpc_list))
2295 atomic_set(&arg->restart, 0);
2296 arg->list = rpc_list;
2299 case LDLM_WORK_CP_AST:
2300 arg->type = LDLM_CP_CALLBACK;
2301 work_ast_lock = ldlm_work_cp_ast_lock;
2303 #ifdef HAVE_SERVER_SUPPORT
2304 case LDLM_WORK_BL_AST:
2305 arg->type = LDLM_BL_CALLBACK;
2306 work_ast_lock = ldlm_work_bl_ast_lock;
2308 case LDLM_WORK_REVOKE_AST:
2309 arg->type = LDLM_BL_CALLBACK;
2310 work_ast_lock = ldlm_work_revoke_ast_lock;
2312 case LDLM_WORK_GL_AST:
2313 arg->type = LDLM_GL_CALLBACK;
2314 work_ast_lock = ldlm_work_gl_ast_lock;
2321 /* We create a ptlrpc request set with flow control extension.
2322 * This request set will use the work_ast_lock function to produce new
2323 * requests and will send a new request each time one completes in order
2324 * to keep the number of requests in flight to ns_max_parallel_ast */
2325 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2326 work_ast_lock, arg);
2327 if (arg->set == NULL)
2328 GOTO(out, rc = -ENOMEM);
2330 ptlrpc_set_wait(NULL, arg->set);
2331 ptlrpc_set_destroy(arg->set);
2333 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2341 * Try to grant all waiting locks on a resource.
2343 * Calls ldlm_reprocess_queue on waiting queue.
2345 * Typically called after some resource locks are cancelled to see
2346 * if anything could be granted as a result of the cancellation.
2348 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2349 enum ldlm_process_intention intention,
2350 struct ldlm_lock *hint)
2352 LIST_HEAD(rpc_list);
2353 #ifdef HAVE_SERVER_SUPPORT
2354 ldlm_reprocessing_policy reprocess;
2355 struct obd_device *obd;
2360 /* Local lock trees don't get reprocessed. */
2361 if (ns_is_client(ldlm_res_to_ns(res))) {
2366 /* Disable reprocess during lock replay stage but allow during
2367 * request replay stage.
2369 obd = ldlm_res_to_ns(res)->ns_obd;
2370 if (obd->obd_recovering &&
2371 atomic_read(&obd->obd_req_replay_clients) == 0)
2375 reprocess = ldlm_get_reprocessing_policy(res);
2376 reprocess(res, &res->lr_waiting, &rpc_list, intention, hint);
2379 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2381 if (rc == -ERESTART) {
2382 LASSERT(list_empty(&rpc_list));
2388 if (!ns_is_client(ldlm_res_to_ns(res))) {
2389 CERROR("This is client-side-only module, cannot handle "
2390 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2397 void ldlm_reprocess_all(struct ldlm_resource *res, struct ldlm_lock *hint)
2399 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN, hint);
2401 EXPORT_SYMBOL(ldlm_reprocess_all);
2403 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2404 struct hlist_node *hnode, void *arg)
2406 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2408 /* This is only called once after recovery done. LU-8306. */
2409 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY, NULL);
2414 * Iterate through all resources on a namespace attempting to grant waiting
2417 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2422 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2423 ldlm_reprocess_res, NULL, 0);
2429 * Helper function to call blocking AST for LDLM lock \a lock in a
2430 * "cancelling" mode.
2432 void ldlm_cancel_callback(struct ldlm_lock *lock)
2434 check_res_locked(lock->l_resource);
2435 if (!ldlm_is_cancel(lock)) {
2436 ldlm_set_cancel(lock);
2437 if (lock->l_blocking_ast) {
2438 unlock_res_and_lock(lock);
2439 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2441 lock_res_and_lock(lock);
2443 LDLM_DEBUG(lock, "no blocking ast");
2446 /* only canceller can set bl_done bit */
2447 ldlm_set_bl_done(lock);
2448 wake_up_all(&lock->l_waitq);
2449 } else if (!ldlm_is_bl_done(lock)) {
2450 /* The lock is guaranteed to have been canceled once
2451 * returning from this function. */
2452 unlock_res_and_lock(lock);
2453 wait_event_idle(lock->l_waitq, is_bl_done(lock));
2454 lock_res_and_lock(lock);
2459 * Remove skiplist-enabled LDLM lock \a req from granted list
2461 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2463 if (req->l_resource->lr_type != LDLM_PLAIN &&
2464 req->l_resource->lr_type != LDLM_IBITS)
2467 list_del_init(&req->l_sl_policy);
2468 list_del_init(&req->l_sl_mode);
2472 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2474 void ldlm_lock_cancel(struct ldlm_lock *lock)
2476 struct ldlm_resource *res;
2477 struct ldlm_namespace *ns;
2480 lock_res_and_lock(lock);
2482 res = lock->l_resource;
2483 ns = ldlm_res_to_ns(res);
2485 /* Please do not, no matter how tempting, remove this LBUG without
2486 * talking to me first. -phik */
2487 if (lock->l_readers || lock->l_writers) {
2488 LDLM_ERROR(lock, "lock still has references");
2489 unlock_res_and_lock(lock);
2493 if (ldlm_is_waited(lock))
2494 ldlm_del_waiting_lock(lock);
2496 /* Releases cancel callback. */
2497 ldlm_cancel_callback(lock);
2499 /* Yes, second time, just in case it was added again while we were
2500 * running with no res lock in ldlm_cancel_callback */
2501 if (ldlm_is_waited(lock))
2502 ldlm_del_waiting_lock(lock);
2504 ldlm_resource_unlink_lock(lock);
2505 ldlm_lock_destroy_nolock(lock);
2507 if (ldlm_is_granted(lock))
2508 ldlm_pool_del(&ns->ns_pool, lock);
2510 /* Make sure we will not be called again for same lock what is possible
2511 * if not to zero out lock->l_granted_mode */
2512 lock->l_granted_mode = LCK_MINMODE;
2513 unlock_res_and_lock(lock);
2517 EXPORT_SYMBOL(ldlm_lock_cancel);
2520 * Set opaque data into the lock that only makes sense to upper layer.
2522 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2524 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2529 if (lock->l_ast_data == NULL)
2530 lock->l_ast_data = data;
2531 if (lock->l_ast_data == data)
2533 LDLM_LOCK_PUT(lock);
2537 EXPORT_SYMBOL(ldlm_lock_set_data);
2539 struct export_cl_data {
2540 const struct lu_env *ecl_env;
2541 struct obd_export *ecl_exp;
2545 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2546 struct ldlm_lock *lock,
2547 struct export_cl_data *ecl)
2549 struct ldlm_resource *res;
2551 res = ldlm_resource_getref(lock->l_resource);
2553 ldlm_lvbo_update(res, lock, NULL, 1);
2554 ldlm_lock_cancel(lock);
2555 if (!exp->exp_obd->obd_stopping)
2556 ldlm_reprocess_all(res, lock);
2557 ldlm_resource_putref(res);
2560 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2561 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2562 exp, ecl->ecl_loop);
2567 * Iterator function for ldlm_export_cancel_locks.
2568 * Cancels passed locks.
2571 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2572 struct hlist_node *hnode, void *data)
2575 struct export_cl_data *ecl = (struct export_cl_data *)data;
2576 struct obd_export *exp = ecl->ecl_exp;
2577 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2579 LDLM_LOCK_GET(lock);
2580 ldlm_cancel_lock_for_export(exp, lock, ecl);
2581 LDLM_LOCK_RELEASE(lock);
2587 * Cancel all blocked locks for given export.
2589 * Typically called on client disconnection/eviction
2591 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2594 struct export_cl_data ecl = {
2600 rc = lu_env_init(&env, LCT_DT_THREAD);
2605 while (!list_empty(&exp->exp_bl_list)) {
2606 struct ldlm_lock *lock;
2608 spin_lock_bh(&exp->exp_bl_list_lock);
2609 if (!list_empty(&exp->exp_bl_list)) {
2610 lock = list_entry(exp->exp_bl_list.next,
2611 struct ldlm_lock, l_exp_list);
2612 LDLM_LOCK_GET(lock);
2613 list_del_init(&lock->l_exp_list);
2617 spin_unlock_bh(&exp->exp_bl_list_lock);
2622 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2623 LDLM_LOCK_RELEASE(lock);
2628 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2629 "left on hash table %d.\n", exp, ecl.ecl_loop,
2630 atomic_read(&exp->exp_lock_hash->hs_count));
2632 return ecl.ecl_loop;
2636 * Cancel all locks for given export.
2638 * Typically called after client disconnection/eviction
2640 int ldlm_export_cancel_locks(struct obd_export *exp)
2642 struct export_cl_data ecl;
2646 rc = lu_env_init(&env, LCT_DT_THREAD);
2653 cfs_hash_for_each_empty(exp->exp_lock_hash,
2654 ldlm_cancel_locks_for_export_cb, &ecl);
2656 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2657 "left on hash table %d.\n", exp, ecl.ecl_loop,
2658 atomic_read(&exp->exp_lock_hash->hs_count));
2660 if (ecl.ecl_loop > 0 &&
2661 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2662 exp->exp_obd->obd_stopping)
2663 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2667 return ecl.ecl_loop;
2671 * Downgrade an PW/EX lock to COS | CR mode.
2673 * A lock mode convertion from PW/EX mode to less conflict mode. The
2674 * convertion may fail if lock was canceled before downgrade, but it doesn't
2675 * indicate any problem, because such lock has no reader or writer, and will
2678 * Used by Commit on Sharing (COS) code to force object changes commit in case
2679 * of conflict. Converted lock is considered as new lock and all blocking AST
2680 * things are cleared, so any pending or new blocked lock on that lock will
2681 * cause new call to blocking_ast and force resource object commit.
2683 * Also used by layout_change to replace EX lock to CR lock.
2685 * \param lock A lock to convert
2686 * \param new_mode new lock mode
2688 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2690 #ifdef HAVE_SERVER_SUPPORT
2693 LASSERT(new_mode == LCK_COS || new_mode == LCK_CR);
2695 lock_res_and_lock(lock);
2697 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2698 unlock_res_and_lock(lock);
2700 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2701 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2705 ldlm_resource_unlink_lock(lock);
2707 * Remove the lock from pool as it will be added again in
2708 * ldlm_grant_lock() called below.
2710 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2712 /* Consider downgraded lock as a new lock and clear all states
2713 * related to a previous blocking AST processing.
2715 ldlm_clear_blocking_data(lock);
2717 lock->l_req_mode = new_mode;
2718 ldlm_grant_lock(lock, NULL);
2719 unlock_res_and_lock(lock);
2721 ldlm_reprocess_all(lock->l_resource, lock);
2726 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2729 * Print lock with lock handle \a lockh description into debug log.
2731 * Used when printing all locks on a resource for debug purposes.
2733 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2735 struct ldlm_lock *lock;
2737 if (!((libcfs_debug | D_ERROR) & level))
2740 lock = ldlm_handle2lock(lockh);
2744 LDLM_DEBUG_LIMIT(level, lock, "###");
2746 LDLM_LOCK_PUT(lock);
2748 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2751 * Print lock information with custom message into debug log.
2754 void _ldlm_lock_debug(struct ldlm_lock *lock,
2755 struct libcfs_debug_msg_data *msgdata,
2756 const char *fmt, ...)
2759 struct obd_export *exp = lock->l_export;
2760 struct ldlm_resource *resource = NULL;
2761 struct va_format vaf;
2762 char *nid = "local";
2764 /* on server-side resource of lock doesn't change */
2765 if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2766 if (lock->l_resource != NULL)
2767 resource = ldlm_resource_getref(lock->l_resource);
2768 } else if (spin_trylock(&lock->l_lock)) {
2769 if (lock->l_resource != NULL)
2770 resource = ldlm_resource_getref(lock->l_resource);
2771 spin_unlock(&lock->l_lock);
2774 va_start(args, fmt);
2778 if (exp && exp->exp_connection) {
2779 nid = obd_export_nid2str(exp);
2780 } else if (exp && exp->exp_obd != NULL) {
2781 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2782 nid = obd_import_nid2str(imp);
2785 if (resource == NULL) {
2786 libcfs_debug_msg(msgdata,
2787 "%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",
2790 lock->l_handle.h_cookie,
2791 refcount_read(&lock->l_handle.h_ref),
2792 lock->l_readers, lock->l_writers,
2793 ldlm_lockname[lock->l_granted_mode],
2794 ldlm_lockname[lock->l_req_mode],
2796 lock->l_remote_handle.cookie,
2797 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2798 lock->l_pid, lock->l_callback_timestamp,
2804 switch (resource->lr_type) {
2806 libcfs_debug_msg(msgdata,
2807 "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s [%llu->%llu] (req %llu->%llu) flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2809 ldlm_lock_to_ns_name(lock), lock,
2810 lock->l_handle.h_cookie,
2811 refcount_read(&lock->l_handle.h_ref),
2812 lock->l_readers, lock->l_writers,
2813 ldlm_lockname[lock->l_granted_mode],
2814 ldlm_lockname[lock->l_req_mode],
2816 atomic_read(&resource->lr_refcount),
2817 ldlm_typename[resource->lr_type],
2818 lock->l_policy_data.l_extent.start,
2819 lock->l_policy_data.l_extent.end,
2820 lock->l_req_extent.start, lock->l_req_extent.end,
2822 lock->l_remote_handle.cookie,
2823 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2824 lock->l_pid, lock->l_callback_timestamp,
2829 libcfs_debug_msg(msgdata,
2830 "%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",
2832 ldlm_lock_to_ns_name(lock), lock,
2833 lock->l_handle.h_cookie,
2834 refcount_read(&lock->l_handle.h_ref),
2835 lock->l_readers, lock->l_writers,
2836 ldlm_lockname[lock->l_granted_mode],
2837 ldlm_lockname[lock->l_req_mode],
2839 atomic_read(&resource->lr_refcount),
2840 ldlm_typename[resource->lr_type],
2841 lock->l_policy_data.l_flock.pid,
2842 lock->l_policy_data.l_flock.start,
2843 lock->l_policy_data.l_flock.end,
2845 lock->l_remote_handle.cookie,
2846 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2847 lock->l_pid, lock->l_callback_timestamp);
2851 libcfs_debug_msg(msgdata,
2852 "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " bits %#llx/%#llx rrc: %d type: %s flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2854 ldlm_lock_to_ns_name(lock),
2855 lock, lock->l_handle.h_cookie,
2856 refcount_read(&lock->l_handle.h_ref),
2857 lock->l_readers, lock->l_writers,
2858 ldlm_lockname[lock->l_granted_mode],
2859 ldlm_lockname[lock->l_req_mode],
2861 lock->l_policy_data.l_inodebits.bits,
2862 lock->l_policy_data.l_inodebits.try_bits,
2863 atomic_read(&resource->lr_refcount),
2864 ldlm_typename[resource->lr_type],
2866 lock->l_remote_handle.cookie,
2867 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2868 lock->l_pid, lock->l_callback_timestamp,
2873 libcfs_debug_msg(msgdata,
2874 "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2876 ldlm_lock_to_ns_name(lock),
2877 lock, lock->l_handle.h_cookie,
2878 refcount_read(&lock->l_handle.h_ref),
2879 lock->l_readers, lock->l_writers,
2880 ldlm_lockname[lock->l_granted_mode],
2881 ldlm_lockname[lock->l_req_mode],
2883 atomic_read(&resource->lr_refcount),
2884 ldlm_typename[resource->lr_type],
2886 lock->l_remote_handle.cookie,
2887 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2888 lock->l_pid, lock->l_callback_timestamp,
2893 ldlm_resource_putref(resource);
2895 EXPORT_SYMBOL(_ldlm_lock_debug);