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);
194 * Release lock reference.
196 * Also frees the lock if it was last reference.
198 void ldlm_lock_put(struct ldlm_lock *lock)
202 LASSERT(lock->l_resource != LP_POISON);
203 LASSERT(refcount_read(&lock->l_handle.h_ref) > 0);
204 if (refcount_dec_and_test(&lock->l_handle.h_ref)) {
205 struct ldlm_resource *res;
208 "final lock_put on destroyed lock, freeing it.");
210 res = lock->l_resource;
211 LASSERT(ldlm_is_destroyed(lock));
212 LASSERT(list_empty(&lock->l_exp_list));
213 LASSERT(list_empty(&lock->l_res_link));
214 LASSERT(list_empty(&lock->l_pending_chain));
216 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
218 lu_ref_del(&res->lr_reference, "lock", lock);
219 if (lock->l_export) {
220 class_export_lock_put(lock->l_export, lock);
221 lock->l_export = NULL;
224 if (lock->l_lvb_data != NULL)
225 OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
227 if (res->lr_type == LDLM_EXTENT) {
228 ldlm_interval_free(ldlm_interval_detach(lock));
229 } else if (res->lr_type == LDLM_IBITS) {
230 if (lock->l_ibits_node != NULL)
231 OBD_SLAB_FREE_PTR(lock->l_ibits_node,
232 ldlm_inodebits_slab);
234 ldlm_resource_putref(res);
235 lock->l_resource = NULL;
236 lu_ref_fini(&lock->l_reference);
237 OBD_FREE_RCU(lock, sizeof(*lock), &lock->l_handle);
242 EXPORT_SYMBOL(ldlm_lock_put);
245 * Removes LDLM lock \a lock from LRU. Assumes LRU is already locked.
247 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
250 if (!list_empty(&lock->l_lru)) {
251 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
253 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
254 if (ns->ns_last_pos == &lock->l_lru)
255 ns->ns_last_pos = lock->l_lru.prev;
256 list_del_init(&lock->l_lru);
257 LASSERT(ns->ns_nr_unused > 0);
265 * Removes LDLM lock \a lock from LRU. Obtains the LRU lock first.
267 * If \a last_use is non-zero, it will remove the lock from LRU only if
268 * it matches lock's l_last_used.
270 * \retval 0 if \a last_use is set, the lock is not in LRU list or \a last_use
271 * doesn't match lock's l_last_used;
272 * otherwise, the lock hasn't been in the LRU list.
273 * \retval 1 the lock was in LRU list and removed.
275 int ldlm_lock_remove_from_lru_check(struct ldlm_lock *lock, ktime_t last_use)
277 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
281 if (ldlm_is_ns_srv(lock)) {
282 LASSERT(list_empty(&lock->l_lru));
286 spin_lock(&ns->ns_lock);
287 if (!ktime_compare(last_use, ktime_set(0, 0)) ||
288 !ktime_compare(last_use, lock->l_last_used))
289 rc = ldlm_lock_remove_from_lru_nolock(lock);
290 spin_unlock(&ns->ns_lock);
296 * Adds LDLM lock \a lock to namespace LRU. Assumes LRU is already locked.
298 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
300 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
302 lock->l_last_used = ktime_get();
303 LASSERT(list_empty(&lock->l_lru));
304 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
305 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
306 LASSERT(ns->ns_nr_unused >= 0);
311 * Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks
314 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
316 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
319 spin_lock(&ns->ns_lock);
320 ldlm_lock_add_to_lru_nolock(lock);
321 spin_unlock(&ns->ns_lock);
326 * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
327 * the LRU. Performs necessary LRU locking
329 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
331 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
334 if (ldlm_is_ns_srv(lock)) {
335 LASSERT(list_empty(&lock->l_lru));
340 spin_lock(&ns->ns_lock);
341 if (!list_empty(&lock->l_lru)) {
342 ldlm_lock_remove_from_lru_nolock(lock);
343 ldlm_lock_add_to_lru_nolock(lock);
345 spin_unlock(&ns->ns_lock);
350 * Helper to destroy a locked lock.
352 * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
353 * Must be called with l_lock and lr_lock held.
355 * Does not actually free the lock data, but rather marks the lock as
356 * destroyed by setting l_destroyed field in the lock to 1. Destroys a
357 * handle->lock association too, so that the lock can no longer be found
358 * and removes the lock from LRU list. Actual lock freeing occurs when
359 * last lock reference goes away.
361 * Original comment (of some historical value):
362 * This used to have a 'strict' flag, which recovery would use to mark an
363 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
364 * shall explain why it's gone: with the new hash table scheme, once you call
365 * ldlm_lock_destroy, you can never drop your final references on this lock.
366 * Because it's not in the hash table anymore. -phil
368 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
372 if (lock->l_readers || lock->l_writers) {
373 LDLM_ERROR(lock, "lock still has references");
377 if (!list_empty(&lock->l_res_link)) {
378 LDLM_ERROR(lock, "lock still on resource");
382 if (ldlm_is_destroyed(lock)) {
383 LASSERT(list_empty(&lock->l_lru));
387 ldlm_set_destroyed(lock);
389 if (lock->l_export && lock->l_export->exp_lock_hash) {
390 /* NB: it's safe to call cfs_hash_del() even lock isn't
391 * in exp_lock_hash. */
392 /* In the function below, .hs_keycmp resolves to
393 * ldlm_export_lock_keycmp() */
394 /* coverity[overrun-buffer-val] */
395 cfs_hash_del(lock->l_export->exp_lock_hash,
396 &lock->l_remote_handle, &lock->l_exp_hash);
399 ldlm_lock_remove_from_lru(lock);
400 class_handle_unhash(&lock->l_handle);
407 * Destroys a LDLM lock \a lock. Performs necessary locking first.
409 void ldlm_lock_destroy(struct ldlm_lock *lock)
413 lock_res_and_lock(lock);
414 first = ldlm_lock_destroy_internal(lock);
415 unlock_res_and_lock(lock);
417 /* drop reference from hashtable only for first destroy */
419 lu_ref_del(&lock->l_reference, "hash", lock);
420 LDLM_LOCK_RELEASE(lock);
426 * Destroys a LDLM lock \a lock that is already locked.
428 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
432 first = ldlm_lock_destroy_internal(lock);
433 /* drop reference from hashtable only for first destroy */
435 lu_ref_del(&lock->l_reference, "hash", lock);
436 LDLM_LOCK_RELEASE(lock);
441 static void lock_handle_free(void *lock, int size)
443 LASSERT(size == sizeof(struct ldlm_lock));
444 OBD_SLAB_FREE(lock, ldlm_lock_slab, size);
447 static struct portals_handle_ops lock_handle_ops = {
448 .hop_free = lock_handle_free,
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_LIST_HEAD_RCU(&lock->l_handle.h_link);
493 class_handle_hash(&lock->l_handle, &lock_handle_ops);
495 lu_ref_init(&lock->l_reference);
496 lu_ref_add(&lock->l_reference, "hash", lock);
497 lock->l_callback_timeout = 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 = lock->l_resource;
519 struct ldlm_resource *newres;
523 LASSERT(ns_is_client(ns));
525 lock_res_and_lock(lock);
526 if (memcmp(new_resid, &lock->l_resource->lr_name,
527 sizeof(lock->l_resource->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 lock = class_handle2object(handle->cookie, &lock_handle_ops);
607 if (lock->l_export != NULL && lock->l_export->exp_failed) {
608 CDEBUG(D_INFO, "lock export failed: lock %p, exp %p\n",
609 lock, lock->l_export);
614 /* It's unlikely but possible that someone marked the lock as
615 * destroyed after we did handle2object on it */
616 if ((flags == 0) && !ldlm_is_destroyed(lock)) {
617 lu_ref_add(&lock->l_reference, "handle", current);
621 lock_res_and_lock(lock);
623 LASSERT(lock->l_resource != NULL);
625 lu_ref_add_atomic(&lock->l_reference, "handle", current);
626 if (unlikely(ldlm_is_destroyed(lock))) {
627 unlock_res_and_lock(lock);
628 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
633 /* If we're setting flags, make sure none of them are already set. */
635 if ((lock->l_flags & flags) != 0) {
636 unlock_res_and_lock(lock);
641 lock->l_flags |= flags;
644 unlock_res_and_lock(lock);
647 EXPORT_SYMBOL(__ldlm_handle2lock);
648 /** @} ldlm_handles */
651 * Fill in "on the wire" representation for given LDLM lock into supplied
652 * lock descriptor \a desc structure.
654 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
656 ldlm_res2desc(lock->l_resource, &desc->l_resource);
657 desc->l_req_mode = lock->l_req_mode;
658 desc->l_granted_mode = lock->l_granted_mode;
659 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
660 &lock->l_policy_data,
661 &desc->l_policy_data);
665 * Add a lock to list of conflicting locks to send AST to.
667 * Only add if we have not sent a blocking AST to the lock yet.
669 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
670 struct list_head *work_list)
672 if (!ldlm_is_ast_sent(lock)) {
673 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
674 ldlm_set_ast_sent(lock);
675 /* If the enqueuing client said so, tell the AST recipient to
676 * discard dirty data, rather than writing back. */
677 if (ldlm_is_ast_discard_data(new))
678 ldlm_set_discard_data(lock);
680 /* Lock can be converted from a blocking state back to granted
681 * after lock convert or COS downgrade but still be in an
682 * older bl_list because it is controlled only by
683 * ldlm_work_bl_ast_lock(), let it be processed there.
685 if (list_empty(&lock->l_bl_ast)) {
686 list_add(&lock->l_bl_ast, work_list);
689 LASSERT(lock->l_blocking_lock == NULL);
690 lock->l_blocking_lock = LDLM_LOCK_GET(new);
695 * Add a lock to list of just granted locks to send completion AST to.
697 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
698 struct list_head *work_list)
700 if (!ldlm_is_cp_reqd(lock)) {
701 ldlm_set_cp_reqd(lock);
702 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
703 LASSERT(list_empty(&lock->l_cp_ast));
704 list_add(&lock->l_cp_ast, work_list);
710 * Aggregator function to add AST work items into a list. Determines
711 * what sort of an AST work needs to be done and calls the proper
713 * Must be called with lr_lock held.
715 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
716 struct list_head *work_list)
719 check_res_locked(lock->l_resource);
721 ldlm_add_bl_work_item(lock, new, work_list);
723 ldlm_add_cp_work_item(lock, work_list);
728 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
729 * r/w reference type is determined by \a mode
730 * Calls ldlm_lock_addref_internal.
732 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
734 struct ldlm_lock *lock;
736 lock = ldlm_handle2lock(lockh);
737 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
738 ldlm_lock_addref_internal(lock, mode);
741 EXPORT_SYMBOL(ldlm_lock_addref);
745 * Add specified reader/writer reference to LDLM lock \a lock.
746 * r/w reference type is determined by \a mode
747 * Removes lock from LRU if it is there.
748 * Assumes the LDLM lock is already locked.
750 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
753 ldlm_lock_remove_from_lru(lock);
754 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
756 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
758 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
760 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
763 lu_ref_add_atomic(&lock->l_reference, "user", lock);
764 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
768 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
769 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
771 * \retval 0 success, lock was addref-ed
773 * \retval -EAGAIN lock is being canceled.
775 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
777 struct ldlm_lock *lock;
781 lock = ldlm_handle2lock(lockh);
783 lock_res_and_lock(lock);
784 if (lock->l_readers != 0 || lock->l_writers != 0 ||
785 !ldlm_is_cbpending(lock)) {
786 ldlm_lock_addref_internal_nolock(lock, mode);
789 unlock_res_and_lock(lock);
794 EXPORT_SYMBOL(ldlm_lock_addref_try);
797 * Add specified reader/writer reference to LDLM lock \a lock.
798 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
799 * Only called for local locks.
801 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
803 lock_res_and_lock(lock);
804 ldlm_lock_addref_internal_nolock(lock, mode);
805 unlock_res_and_lock(lock);
809 * Removes reader/writer reference for LDLM lock \a lock.
810 * Assumes LDLM lock is already locked.
811 * only called in ldlm_flock_destroy and for local locks.
812 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
813 * that cannot be placed in LRU.
815 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
818 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
819 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
820 LASSERT(lock->l_readers > 0);
821 lu_ref_del(&lock->l_reference, "reader", lock);
824 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
825 LASSERT(lock->l_writers > 0);
826 lu_ref_del(&lock->l_reference, "writer", lock);
830 lu_ref_del(&lock->l_reference, "user", lock);
831 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
835 * Removes reader/writer reference for LDLM lock \a lock.
836 * Locks LDLM lock first.
837 * If the lock is determined to be client lock on a client and r/w refcount
838 * drops to zero and the lock is not blocked, the lock is added to LRU lock
840 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
842 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
844 struct ldlm_namespace *ns;
847 lock_res_and_lock(lock);
849 ns = ldlm_lock_to_ns(lock);
851 ldlm_lock_decref_internal_nolock(lock, mode);
853 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
854 !lock->l_readers && !lock->l_writers) {
855 /* If this is a local lock on a server namespace and this was
856 * the last reference, cancel the lock.
858 * Group locks are special:
859 * They must not go in LRU, but they are not called back
860 * like non-group locks, instead they are manually released.
861 * They have an l_writers reference which they keep until
862 * they are manually released, so we remove them when they have
863 * no more reader or writer references. - LU-6368 */
864 ldlm_set_cbpending(lock);
867 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
868 /* If we received a blocked AST and this was the last reference,
869 * run the callback. */
870 if (ldlm_is_ns_srv(lock) && lock->l_export)
871 CERROR("FL_CBPENDING set on non-local lock--just a "
874 LDLM_DEBUG(lock, "final decref done on cbpending lock");
876 LDLM_LOCK_GET(lock); /* dropped by bl thread */
877 ldlm_lock_remove_from_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 if (ldlm_is_atomic_cb(lock) ||
884 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
885 ldlm_handle_bl_callback(ns, NULL, lock);
886 } else if (ns_is_client(ns) &&
887 !lock->l_readers && !lock->l_writers &&
888 !ldlm_is_no_lru(lock) &&
889 !ldlm_is_bl_ast(lock) &&
890 !ldlm_is_converting(lock)) {
892 LDLM_DEBUG(lock, "add lock into lru list");
894 /* If this is a client-side namespace and this was the last
895 * reference, put it on the LRU. */
896 ldlm_lock_add_to_lru(lock);
897 unlock_res_and_lock(lock);
899 if (ldlm_is_fail_loc(lock))
900 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
902 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
903 * are not supported by the server, otherwise, it is done on
905 if (!exp_connect_cancelset(lock->l_conn_export) &&
906 !ns_connect_lru_resize(ns))
907 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
909 LDLM_DEBUG(lock, "do not add lock into lru list");
910 unlock_res_and_lock(lock);
917 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
919 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
921 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
922 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
923 ldlm_lock_decref_internal(lock, mode);
926 EXPORT_SYMBOL(ldlm_lock_decref);
929 * Decrease reader/writer refcount for LDLM lock with handle
930 * \a lockh and mark it for subsequent cancellation once r/w refcount
931 * drops to zero instead of putting into LRU.
934 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
937 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
940 LASSERT(lock != NULL);
942 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
943 lock_res_and_lock(lock);
944 ldlm_set_cbpending(lock);
945 unlock_res_and_lock(lock);
946 ldlm_lock_decref_internal(lock, mode);
949 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
951 struct sl_insert_point {
952 struct list_head *res_link;
953 struct list_head *mode_link;
954 struct list_head *policy_link;
958 * Finds a position to insert the new lock into granted lock list.
960 * Used for locks eligible for skiplist optimization.
963 * queue [input]: the granted list where search acts on;
964 * req [input]: the lock whose position to be located;
965 * prev [output]: positions within 3 lists to insert @req to
969 * - ldlm_grant_lock_with_skiplist
971 static void search_granted_lock(struct list_head *queue,
972 struct ldlm_lock *req,
973 struct sl_insert_point *prev)
975 struct list_head *tmp;
976 struct ldlm_lock *lock, *mode_end, *policy_end;
979 list_for_each(tmp, queue) {
980 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
982 mode_end = list_entry(lock->l_sl_mode.prev,
983 struct ldlm_lock, l_sl_mode);
985 if (lock->l_req_mode != req->l_req_mode) {
986 /* jump to last lock of mode group */
987 tmp = &mode_end->l_res_link;
991 /* suitable mode group is found */
992 if (lock->l_resource->lr_type == LDLM_PLAIN) {
993 /* insert point is last lock of the mode group */
994 prev->res_link = &mode_end->l_res_link;
995 prev->mode_link = &mode_end->l_sl_mode;
996 prev->policy_link = &req->l_sl_policy;
999 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
1002 list_entry(lock->l_sl_policy.prev,
1006 if (lock->l_policy_data.l_inodebits.bits ==
1007 req->l_policy_data.l_inodebits.bits) {
1008 /* insert point is last lock of
1009 * the policy group */
1011 &policy_end->l_res_link;
1013 &policy_end->l_sl_mode;
1015 &policy_end->l_sl_policy;
1020 if (policy_end == mode_end)
1021 /* done with mode group */
1024 /* go to next policy group within mode group */
1025 tmp = policy_end->l_res_link.next;
1026 lock = list_entry(tmp, struct ldlm_lock,
1028 } /* loop over policy groups within the mode group */
1030 /* insert point is last lock of the mode group,
1031 * new policy group is started */
1032 prev->res_link = &mode_end->l_res_link;
1033 prev->mode_link = &mode_end->l_sl_mode;
1034 prev->policy_link = &req->l_sl_policy;
1038 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1043 /* insert point is last lock on the queue,
1044 * new mode group and new policy group are started */
1045 prev->res_link = queue->prev;
1046 prev->mode_link = &req->l_sl_mode;
1047 prev->policy_link = &req->l_sl_policy;
1052 * Add a lock into resource granted list after a position described by
1055 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1056 struct sl_insert_point *prev)
1058 struct ldlm_resource *res = lock->l_resource;
1061 check_res_locked(res);
1063 ldlm_resource_dump(D_INFO, res);
1064 LDLM_DEBUG(lock, "About to add lock:");
1066 if (ldlm_is_destroyed(lock)) {
1067 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1071 LASSERT(list_empty(&lock->l_res_link));
1072 LASSERT(list_empty(&lock->l_sl_mode));
1073 LASSERT(list_empty(&lock->l_sl_policy));
1076 * lock->link == prev->link means lock is first starting the group.
1077 * Don't re-add to itself to suppress kernel warnings.
1079 if (&lock->l_res_link != prev->res_link)
1080 list_add(&lock->l_res_link, prev->res_link);
1081 if (&lock->l_sl_mode != prev->mode_link)
1082 list_add(&lock->l_sl_mode, prev->mode_link);
1083 if (&lock->l_sl_policy != prev->policy_link)
1084 list_add(&lock->l_sl_policy, prev->policy_link);
1090 * Add a lock to granted list on a resource maintaining skiplist
1093 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1095 struct sl_insert_point prev;
1097 LASSERT(ldlm_is_granted(lock));
1099 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1100 ldlm_granted_list_add_lock(lock, &prev);
1104 * Perform lock granting bookkeeping.
1106 * Includes putting the lock into granted list and updating lock mode.
1108 * - ldlm_lock_enqueue
1109 * - ldlm_reprocess_queue
1111 * must be called with lr_lock held
1113 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1115 struct ldlm_resource *res = lock->l_resource;
1118 check_res_locked(res);
1120 lock->l_granted_mode = lock->l_req_mode;
1122 if (work_list && lock->l_completion_ast != NULL)
1123 ldlm_add_ast_work_item(lock, NULL, work_list);
1125 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1126 ldlm_grant_lock_with_skiplist(lock);
1127 else if (res->lr_type == LDLM_EXTENT)
1128 ldlm_extent_add_lock(res, lock);
1129 else if (res->lr_type == LDLM_FLOCK) {
1130 /* We should not add locks to granted list in the following
1132 * - this is an UNLOCK but not a real lock;
1133 * - this is a TEST lock;
1134 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1135 * - this is a deadlock (flock cannot be granted) */
1136 if (lock->l_req_mode == 0 ||
1137 lock->l_req_mode == LCK_NL ||
1138 ldlm_is_test_lock(lock) ||
1139 ldlm_is_flock_deadlock(lock))
1141 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1146 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1151 * Check if the given @lock meets the criteria for a match.
1152 * A reference on the lock is taken if matched.
1154 * \param lock test-against this lock
1155 * \param data parameters
1157 static int lock_matches(struct ldlm_lock *lock, struct ldlm_match_data *data)
1159 union ldlm_policy_data *lpol = &lock->l_policy_data;
1160 enum ldlm_mode match = LCK_MINMODE;
1162 if (lock == data->lmd_old)
1163 return INTERVAL_ITER_STOP;
1165 /* Check if this lock can be matched.
1166 * Used by LU-2919(exclusive open) for open lease lock */
1167 if (ldlm_is_excl(lock))
1168 return INTERVAL_ITER_CONT;
1170 /* llite sometimes wants to match locks that will be
1171 * canceled when their users drop, but we allow it to match
1172 * if it passes in CBPENDING and the lock still has users.
1173 * this is generally only going to be used by children
1174 * whose parents already hold a lock so forward progress
1175 * can still happen. */
1176 if (ldlm_is_cbpending(lock) &&
1177 !(data->lmd_flags & LDLM_FL_CBPENDING))
1178 return INTERVAL_ITER_CONT;
1179 if (!data->lmd_unref && ldlm_is_cbpending(lock) &&
1180 lock->l_readers == 0 && lock->l_writers == 0)
1181 return INTERVAL_ITER_CONT;
1183 if (!(lock->l_req_mode & *data->lmd_mode))
1184 return INTERVAL_ITER_CONT;
1186 /* When we search for ast_data, we are not doing a traditional match,
1187 * so we don't worry about IBITS or extent matching.
1189 if (data->lmd_has_ast_data) {
1190 if (!lock->l_ast_data)
1191 return INTERVAL_ITER_CONT;
1196 match = lock->l_req_mode;
1198 switch (lock->l_resource->lr_type) {
1200 if (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1201 lpol->l_extent.end < data->lmd_policy->l_extent.end)
1202 return INTERVAL_ITER_CONT;
1204 if (unlikely(match == LCK_GROUP) &&
1205 data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1206 lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1207 return INTERVAL_ITER_CONT;
1210 /* We match if we have existing lock with same or wider set
1212 if ((lpol->l_inodebits.bits &
1213 data->lmd_policy->l_inodebits.bits) !=
1214 data->lmd_policy->l_inodebits.bits)
1215 return INTERVAL_ITER_CONT;
1221 /* We match if we have existing lock with same or wider set
1223 if (!data->lmd_unref && LDLM_HAVE_MASK(lock, GONE))
1224 return INTERVAL_ITER_CONT;
1226 if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1227 return INTERVAL_ITER_CONT;
1229 /* Filter locks by skipping flags */
1230 if (data->lmd_skip_flags & lock->l_flags)
1231 return INTERVAL_ITER_CONT;
1234 if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1235 LDLM_LOCK_GET(lock);
1236 ldlm_lock_touch_in_lru(lock);
1238 ldlm_lock_addref_internal_nolock(lock, match);
1241 *data->lmd_mode = match;
1242 data->lmd_lock = lock;
1244 return INTERVAL_ITER_STOP;
1247 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1249 struct ldlm_interval *node = to_ldlm_interval(in);
1250 struct ldlm_match_data *data = args;
1251 struct ldlm_lock *lock;
1254 list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1255 rc = lock_matches(lock, data);
1256 if (rc == INTERVAL_ITER_STOP)
1257 return INTERVAL_ITER_STOP;
1259 return INTERVAL_ITER_CONT;
1263 * Search for a lock with given parameters in interval trees.
1265 * \param res search for a lock in this resource
1266 * \param data parameters
1268 * \retval a referenced lock or NULL.
1270 struct ldlm_lock *search_itree(struct ldlm_resource *res,
1271 struct ldlm_match_data *data)
1273 struct interval_node_extent ext = {
1274 .start = data->lmd_policy->l_extent.start,
1275 .end = data->lmd_policy->l_extent.end
1279 data->lmd_lock = NULL;
1281 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1282 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1284 if (tree->lit_root == NULL)
1287 if (!(tree->lit_mode & *data->lmd_mode))
1290 interval_search(tree->lit_root, &ext,
1291 itree_overlap_cb, data);
1293 return data->lmd_lock;
1298 EXPORT_SYMBOL(search_itree);
1302 * Search for a lock with given properties in a queue.
1304 * \param queue search for a lock in this queue
1305 * \param data parameters
1307 * \retval a referenced lock or NULL.
1309 static struct ldlm_lock *search_queue(struct list_head *queue,
1310 struct ldlm_match_data *data)
1312 struct ldlm_lock *lock;
1315 data->lmd_lock = NULL;
1317 list_for_each_entry(lock, queue, l_res_link) {
1318 rc = lock_matches(lock, data);
1319 if (rc == INTERVAL_ITER_STOP)
1320 return data->lmd_lock;
1326 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1328 if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1329 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1330 wake_up_all(&lock->l_waitq);
1333 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1335 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1337 lock_res_and_lock(lock);
1338 ldlm_lock_fail_match_locked(lock);
1339 unlock_res_and_lock(lock);
1343 * Mark lock as "matchable" by OST.
1345 * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1347 * Assumes LDLM lock is already locked.
1349 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1351 ldlm_set_lvb_ready(lock);
1352 wake_up_all(&lock->l_waitq);
1354 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1357 * Mark lock as "matchable" by OST.
1358 * Locks the lock and then \see ldlm_lock_allow_match_locked
1360 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1362 lock_res_and_lock(lock);
1363 ldlm_lock_allow_match_locked(lock);
1364 unlock_res_and_lock(lock);
1366 EXPORT_SYMBOL(ldlm_lock_allow_match);
1369 * Attempt to find a lock with specified properties.
1371 * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1374 * Can be called in two ways:
1376 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1377 * for a duplicate of.
1379 * Otherwise, all of the fields must be filled in, to match against.
1381 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1382 * server (ie, connh is NULL)
1383 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1384 * list will be considered
1385 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1386 * to be canceled can still be matched as long as they still have reader
1387 * or writer refernces
1388 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1389 * just tell us if we would have matched.
1391 * \retval 1 if it finds an already-existing lock that is compatible; in this
1392 * case, lockh is filled in with a addref()ed lock
1394 * We also check security context, and if that fails we simply return 0 (to
1395 * keep caller code unchanged), the context failure will be discovered by
1396 * caller sometime later.
1398 enum ldlm_mode ldlm_lock_match_with_skip(struct ldlm_namespace *ns,
1399 __u64 flags, __u64 skip_flags,
1400 const struct ldlm_res_id *res_id,
1401 enum ldlm_type type,
1402 union ldlm_policy_data *policy,
1403 enum ldlm_mode mode,
1404 struct lustre_handle *lockh, int unref)
1406 struct ldlm_match_data data = {
1410 .lmd_policy = policy,
1412 .lmd_skip_flags = skip_flags,
1414 .lmd_has_ast_data = false,
1416 struct ldlm_resource *res;
1417 struct ldlm_lock *lock;
1423 data.lmd_old = ldlm_handle2lock(lockh);
1424 LASSERT(data.lmd_old != NULL);
1426 ns = ldlm_lock_to_ns(data.lmd_old);
1427 res_id = &data.lmd_old->l_resource->lr_name;
1428 type = data.lmd_old->l_resource->lr_type;
1429 *data.lmd_mode = data.lmd_old->l_req_mode;
1432 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1434 LASSERT(data.lmd_old == NULL);
1438 LDLM_RESOURCE_ADDREF(res);
1440 if (res->lr_type == LDLM_EXTENT)
1441 lock = search_itree(res, &data);
1443 lock = search_queue(&res->lr_granted, &data);
1444 if (!lock && !(flags & LDLM_FL_BLOCK_GRANTED))
1445 lock = search_queue(&res->lr_waiting, &data);
1446 matched = lock ? mode : 0;
1448 LDLM_RESOURCE_DELREF(res);
1449 ldlm_resource_putref(res);
1452 ldlm_lock2handle(lock, lockh);
1453 if ((flags & LDLM_FL_LVB_READY) &&
1454 (!ldlm_is_lvb_ready(lock))) {
1455 __u64 wait_flags = LDLM_FL_LVB_READY |
1456 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1457 struct l_wait_info lwi;
1459 if (lock->l_completion_ast) {
1460 int err = lock->l_completion_ast(lock,
1461 LDLM_FL_WAIT_NOREPROC,
1464 GOTO(out_fail_match, matched = 0);
1467 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1468 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1470 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1471 l_wait_event(lock->l_waitq, lock->l_flags & wait_flags,
1473 if (!ldlm_is_lvb_ready(lock))
1474 GOTO(out_fail_match, matched = 0);
1477 /* check user's security context */
1478 if (lock->l_conn_export &&
1479 sptlrpc_import_check_ctx(
1480 class_exp2cliimp(lock->l_conn_export)))
1481 GOTO(out_fail_match, matched = 0);
1483 LDLM_DEBUG(lock, "matched (%llu %llu)",
1484 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1485 res_id->name[2] : policy->l_extent.start,
1486 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1487 res_id->name[3] : policy->l_extent.end);
1490 if (flags & LDLM_FL_TEST_LOCK)
1491 LDLM_LOCK_RELEASE(lock);
1493 ldlm_lock_decref_internal(lock, mode);
1496 /* less verbose for test-only */
1497 if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
1498 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1499 "%llu/%llu (%llu %llu)", ns,
1500 type, mode, res_id->name[0], res_id->name[1],
1501 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1502 res_id->name[2] : policy->l_extent.start,
1503 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1504 res_id->name[3] : policy->l_extent.end);
1506 if (data.lmd_old != NULL)
1507 LDLM_LOCK_PUT(data.lmd_old);
1511 EXPORT_SYMBOL(ldlm_lock_match_with_skip);
1513 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1516 struct ldlm_lock *lock;
1517 enum ldlm_mode mode = 0;
1520 lock = ldlm_handle2lock(lockh);
1522 lock_res_and_lock(lock);
1523 if (LDLM_HAVE_MASK(lock, GONE))
1526 if (ldlm_is_cbpending(lock) &&
1527 lock->l_readers == 0 && lock->l_writers == 0)
1531 *bits = lock->l_policy_data.l_inodebits.bits;
1532 mode = lock->l_granted_mode;
1533 ldlm_lock_addref_internal_nolock(lock, mode);
1540 unlock_res_and_lock(lock);
1541 LDLM_LOCK_PUT(lock);
1545 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1547 /** The caller must guarantee that the buffer is large enough. */
1548 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1549 enum req_location loc, void *data, int size)
1554 LASSERT(data != NULL);
1557 switch (lock->l_lvb_type) {
1559 if (size == sizeof(struct ost_lvb)) {
1560 if (loc == RCL_CLIENT)
1561 lvb = req_capsule_client_swab_get(pill,
1563 lustre_swab_ost_lvb);
1565 lvb = req_capsule_server_swab_get(pill,
1567 lustre_swab_ost_lvb);
1568 if (unlikely(lvb == NULL)) {
1569 LDLM_ERROR(lock, "no LVB");
1573 memcpy(data, lvb, size);
1574 } else if (size == sizeof(struct ost_lvb_v1)) {
1575 struct ost_lvb *olvb = data;
1577 if (loc == RCL_CLIENT)
1578 lvb = req_capsule_client_swab_get(pill,
1580 lustre_swab_ost_lvb_v1);
1582 lvb = req_capsule_server_sized_swab_get(pill,
1584 lustre_swab_ost_lvb_v1);
1585 if (unlikely(lvb == NULL)) {
1586 LDLM_ERROR(lock, "no LVB");
1590 memcpy(data, lvb, size);
1591 olvb->lvb_mtime_ns = 0;
1592 olvb->lvb_atime_ns = 0;
1593 olvb->lvb_ctime_ns = 0;
1595 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1601 if (size == sizeof(struct lquota_lvb)) {
1602 if (loc == RCL_CLIENT)
1603 lvb = req_capsule_client_swab_get(pill,
1605 lustre_swab_lquota_lvb);
1607 lvb = req_capsule_server_swab_get(pill,
1609 lustre_swab_lquota_lvb);
1610 if (unlikely(lvb == NULL)) {
1611 LDLM_ERROR(lock, "no LVB");
1615 memcpy(data, lvb, size);
1617 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1626 if (loc == RCL_CLIENT)
1627 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1629 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1630 if (unlikely(lvb == NULL)) {
1631 LDLM_ERROR(lock, "no LVB");
1635 memcpy(data, lvb, size);
1638 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1639 libcfs_debug_dumpstack(NULL);
1647 * Create and fill in new LDLM lock with specified properties.
1648 * Returns a referenced lock
1650 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1651 const struct ldlm_res_id *res_id,
1652 enum ldlm_type type,
1653 enum ldlm_mode mode,
1654 const struct ldlm_callback_suite *cbs,
1655 void *data, __u32 lvb_len,
1656 enum lvb_type lvb_type)
1658 struct ldlm_lock *lock;
1659 struct ldlm_resource *res;
1663 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1665 RETURN(ERR_CAST(res));
1667 lock = ldlm_lock_new(res);
1669 RETURN(ERR_PTR(-ENOMEM));
1671 lock->l_req_mode = mode;
1672 lock->l_ast_data = data;
1673 lock->l_pid = current_pid();
1674 if (ns_is_server(ns))
1675 ldlm_set_ns_srv(lock);
1677 lock->l_blocking_ast = cbs->lcs_blocking;
1678 lock->l_completion_ast = cbs->lcs_completion;
1679 lock->l_glimpse_ast = cbs->lcs_glimpse;
1684 rc = ldlm_extent_alloc_lock(lock);
1687 rc = ldlm_inodebits_alloc_lock(lock);
1696 lock->l_lvb_len = lvb_len;
1697 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1698 if (lock->l_lvb_data == NULL)
1699 GOTO(out, rc = -ENOMEM);
1702 lock->l_lvb_type = lvb_type;
1703 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1704 GOTO(out, rc = -ENOENT);
1709 ldlm_lock_destroy(lock);
1710 LDLM_LOCK_RELEASE(lock);
1711 RETURN(ERR_PTR(rc));
1714 #ifdef HAVE_SERVER_SUPPORT
1715 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1718 struct ldlm_resource *res = lock->l_resource;
1719 enum ldlm_error rc = ELDLM_OK;
1720 LIST_HEAD(rpc_list);
1721 ldlm_processing_policy policy;
1725 policy = ldlm_get_processing_policy(res);
1727 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1728 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1729 res->lr_type != LDLM_FLOCK) {
1730 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1731 if (rc == -ERESTART)
1735 if (!list_empty(&rpc_list))
1736 ldlm_discard_bl_list(&rpc_list);
1743 * Enqueue (request) a lock.
1745 * Does not block. As a result of enqueue the lock would be put
1746 * into granted or waiting list.
1748 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1749 * set, skip all the enqueueing and delegate lock processing to intent policy
1752 enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
1753 struct ldlm_namespace *ns,
1754 struct ldlm_lock **lockp,
1755 void *cookie, __u64 *flags)
1757 struct ldlm_lock *lock = *lockp;
1758 struct ldlm_resource *res = lock->l_resource;
1759 int local = ns_is_client(ldlm_res_to_ns(res));
1760 enum ldlm_error rc = ELDLM_OK;
1761 struct ldlm_interval *node = NULL;
1764 /* policies are not executed on the client or during replay */
1765 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1766 && !local && ns->ns_policy) {
1767 rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
1769 if (rc == ELDLM_LOCK_REPLACED) {
1770 /* The lock that was returned has already been granted,
1771 * and placed into lockp. If it's not the same as the
1772 * one we passed in, then destroy the old one and our
1773 * work here is done. */
1774 if (lock != *lockp) {
1775 ldlm_lock_destroy(lock);
1776 LDLM_LOCK_RELEASE(lock);
1778 *flags |= LDLM_FL_LOCK_CHANGED;
1780 } else if (rc != ELDLM_OK &&
1781 ldlm_is_granted(lock)) {
1782 LASSERT(*flags & LDLM_FL_RESENT);
1783 /* It may happen that ns_policy returns an error in
1784 * resend case, object may be unlinked or just some
1785 * error occurs. It is unclear if lock reached the
1786 * client in the original reply, just leave the lock on
1787 * server, not returning it again to client. Due to
1788 * LU-6529, the server will not OOM. */
1790 } else if (rc != ELDLM_OK ||
1791 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1792 ldlm_lock_destroy(lock);
1797 if (*flags & LDLM_FL_RESENT) {
1798 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1799 * Set LOCK_CHANGED always.
1800 * Check if the lock is granted for BLOCK_GRANTED.
1801 * Take NO_TIMEOUT from the lock as it is inherited through
1802 * LDLM_FL_INHERIT_MASK */
1803 *flags |= LDLM_FL_LOCK_CHANGED;
1804 if (!ldlm_is_granted(lock))
1805 *flags |= LDLM_FL_BLOCK_GRANTED;
1806 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1810 /* For a replaying lock, it might be already in granted list. So
1811 * unlinking the lock will cause the interval node to be freed, we
1812 * have to allocate the interval node early otherwise we can't regrant
1813 * this lock in the future. - jay */
1814 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1815 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1817 lock_res_and_lock(lock);
1818 if (local && ldlm_is_granted(lock)) {
1819 /* The server returned a blocked lock, but it was granted
1820 * before we got a chance to actually enqueue it. We don't
1821 * need to do anything else. */
1822 *flags &= ~LDLM_FL_BLOCKED_MASK;
1823 GOTO(out, rc = ELDLM_OK);
1826 ldlm_resource_unlink_lock(lock);
1827 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1829 ldlm_lock_destroy_nolock(lock);
1830 GOTO(out, rc = -ENOMEM);
1833 INIT_LIST_HEAD(&node->li_group);
1834 ldlm_interval_attach(node, lock);
1838 /* Some flags from the enqueue want to make it into the AST, via the
1839 * lock's l_flags. */
1840 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1841 ldlm_set_ast_discard_data(lock);
1842 if (*flags & LDLM_FL_TEST_LOCK)
1843 ldlm_set_test_lock(lock);
1844 if (*flags & LDLM_FL_COS_INCOMPAT)
1845 ldlm_set_cos_incompat(lock);
1846 if (*flags & LDLM_FL_COS_ENABLED)
1847 ldlm_set_cos_enabled(lock);
1849 /* This distinction between local lock trees is very important; a client
1850 * namespace only has information about locks taken by that client, and
1851 * thus doesn't have enough information to decide for itself if it can
1852 * be granted (below). In this case, we do exactly what the server
1853 * tells us to do, as dictated by the 'flags'.
1855 * We do exactly the same thing during recovery, when the server is
1856 * more or less trusting the clients not to lie.
1858 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1861 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1862 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1864 ldlm_grant_lock(lock, NULL);
1865 GOTO(out, rc = ELDLM_OK);
1866 #ifdef HAVE_SERVER_SUPPORT
1867 } else if (*flags & LDLM_FL_REPLAY) {
1868 if (*flags & LDLM_FL_BLOCK_WAIT) {
1869 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1870 GOTO(out, rc = ELDLM_OK);
1871 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1872 ldlm_grant_lock(lock, NULL);
1873 GOTO(out, rc = ELDLM_OK);
1875 /* If no flags, fall through to normal enqueue path. */
1878 rc = ldlm_lock_enqueue_helper(lock, flags);
1882 CERROR("This is client-side-only module, cannot handle "
1883 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1889 unlock_res_and_lock(lock);
1891 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1895 #ifdef HAVE_SERVER_SUPPORT
1897 * Iterate through all waiting locks on a given resource queue and attempt to
1900 * Must be called with resource lock held.
1902 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1903 struct list_head *work_list,
1904 enum ldlm_process_intention intention,
1905 struct ldlm_lock *hint)
1907 struct list_head *tmp, *pos;
1908 ldlm_processing_policy policy;
1910 int rc = LDLM_ITER_CONTINUE;
1911 enum ldlm_error err;
1912 LIST_HEAD(bl_ast_list);
1916 check_res_locked(res);
1918 policy = ldlm_get_processing_policy(res);
1920 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1921 intention == LDLM_PROCESS_RECOVERY);
1924 list_for_each_safe(tmp, pos, queue) {
1925 struct ldlm_lock *pending;
1926 LIST_HEAD(rpc_list);
1928 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1930 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1933 rc = policy(pending, &flags, intention, &err, &rpc_list);
1934 if (pending->l_granted_mode == pending->l_req_mode ||
1935 res->lr_type == LDLM_FLOCK) {
1936 list_splice(&rpc_list, work_list);
1938 list_splice(&rpc_list, &bl_ast_list);
1941 * When this is called from recovery done, we always want
1942 * to scan the whole list no matter what 'rc' is returned.
1944 if (rc != LDLM_ITER_CONTINUE &&
1945 intention == LDLM_PROCESS_RESCAN)
1949 if (!list_empty(&bl_ast_list)) {
1952 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
1956 if (rc == -ERESTART)
1960 if (!list_empty(&bl_ast_list))
1961 ldlm_discard_bl_list(&bl_ast_list);
1963 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
1967 * Conflicting locks are detected for a lock to be enqueued, add the lock
1968 * into waiting list and send blocking ASTs to the conflicting locks.
1970 * \param[in] lock The lock to be enqueued.
1971 * \param[out] flags Lock flags for the lock to be enqueued.
1972 * \param[in] rpc_list Conflicting locks list.
1974 * \retval -ERESTART: Some lock was instantly canceled while sending
1975 * blocking ASTs, caller needs to re-check conflicting
1977 * \retval -EAGAIN: Lock was destroyed, caller should return error.
1978 * \reval 0: Lock is successfully added in waiting list.
1980 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
1981 struct list_head *rpc_list)
1983 struct ldlm_resource *res = lock->l_resource;
1987 check_res_locked(res);
1989 /* If either of the compat_queue()s returned failure, then we
1990 * have ASTs to send and must go onto the waiting list.
1992 * bug 2322: we used to unlink and re-add here, which was a
1993 * terrible folly -- if we goto restart, we could get
1994 * re-ordered! Causes deadlock, because ASTs aren't sent! */
1995 if (list_empty(&lock->l_res_link))
1996 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1999 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
2002 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
2003 !ns_is_client(ldlm_res_to_ns(res)))
2004 class_fail_export(lock->l_export);
2007 if (rc == -ERESTART) {
2008 /* 15715: The lock was granted and destroyed after
2009 * resource lock was dropped. Interval node was freed
2010 * in ldlm_lock_destroy. Anyway, this always happens
2011 * when a client is being evicted. So it would be
2012 * ok to return an error. -jay */
2013 if (ldlm_is_destroyed(lock))
2016 /* lock was granted while resource was unlocked. */
2017 if (ldlm_is_granted(lock)) {
2018 /* bug 11300: if the lock has been granted,
2019 * break earlier because otherwise, we will go
2020 * to restart and ldlm_resource_unlink will be
2021 * called and it causes the interval node to be
2022 * freed. Then we will fail at
2023 * ldlm_extent_add_lock() */
2024 *flags &= ~LDLM_FL_BLOCKED_MASK;
2030 *flags |= LDLM_FL_BLOCK_GRANTED;
2036 * Discard all AST work items from list.
2038 * If for whatever reason we do not want to send ASTs to conflicting locks
2039 * anymore, disassemble the list with this function.
2041 void ldlm_discard_bl_list(struct list_head *bl_list)
2043 struct ldlm_lock *lock, *tmp;
2047 list_for_each_entry_safe(lock, tmp, bl_list, l_bl_ast) {
2048 LASSERT(!list_empty(&lock->l_bl_ast));
2049 list_del_init(&lock->l_bl_ast);
2050 ldlm_clear_ast_sent(lock);
2051 LASSERT(lock->l_bl_ast_run == 0);
2052 ldlm_clear_blocking_lock(lock);
2053 LDLM_LOCK_RELEASE(lock);
2059 * Process a call to blocking AST callback for a lock in ast_work list
2062 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2064 struct ldlm_cb_set_arg *arg = opaq;
2065 struct ldlm_lock *lock;
2066 struct ldlm_lock_desc d;
2067 struct ldlm_bl_desc bld;
2072 if (list_empty(arg->list))
2075 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2077 /* nobody should touch l_bl_ast but some locks in the list may become
2078 * granted after lock convert or COS downgrade, these locks should be
2079 * just skipped here and removed from the list.
2081 lock_res_and_lock(lock);
2082 list_del_init(&lock->l_bl_ast);
2084 /* lock is not blocking lock anymore, but was kept in the list because
2085 * it can managed only here.
2087 if (!ldlm_is_ast_sent(lock)) {
2088 unlock_res_and_lock(lock);
2089 LDLM_LOCK_RELEASE(lock);
2093 LASSERT(lock->l_blocking_lock);
2094 ldlm_lock2desc(lock->l_blocking_lock, &d);
2095 /* copy blocking lock ibits in cancel_bits as well,
2096 * new client may use them for lock convert and it is
2097 * important to use new field to convert locks from
2100 d.l_policy_data.l_inodebits.cancel_bits =
2101 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2103 /* Blocking lock is being destroyed here but some information about it
2104 * may be needed inside l_blocking_ast() function below,
2105 * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
2107 bld.bl_same_client = lock->l_client_cookie ==
2108 lock->l_blocking_lock->l_client_cookie;
2109 bld.bl_cos_incompat = ldlm_is_cos_incompat(lock->l_blocking_lock);
2110 arg->bl_desc = &bld;
2112 LASSERT(ldlm_is_ast_sent(lock));
2113 LASSERT(lock->l_bl_ast_run == 0);
2114 lock->l_bl_ast_run++;
2115 ldlm_clear_blocking_lock(lock);
2116 unlock_res_and_lock(lock);
2118 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2120 LDLM_LOCK_RELEASE(lock);
2126 * Process a call to revocation AST callback for a lock in ast_work list
2129 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2131 struct ldlm_cb_set_arg *arg = opaq;
2132 struct ldlm_lock_desc desc;
2134 struct ldlm_lock *lock;
2137 if (list_empty(arg->list))
2140 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2141 list_del_init(&lock->l_rk_ast);
2143 /* the desc just pretend to exclusive */
2144 ldlm_lock2desc(lock, &desc);
2145 desc.l_req_mode = LCK_EX;
2146 desc.l_granted_mode = 0;
2148 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2149 LDLM_LOCK_RELEASE(lock);
2155 * Process a call to glimpse AST callback for a lock in ast_work list
2157 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2159 struct ldlm_cb_set_arg *arg = opaq;
2160 struct ldlm_glimpse_work *gl_work;
2161 struct ldlm_lock *lock;
2165 if (list_empty(arg->list))
2168 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2170 list_del_init(&gl_work->gl_list);
2172 lock = gl_work->gl_lock;
2174 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2175 arg->gl_desc = gl_work->gl_desc;
2176 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2177 arg->gl_interpret_data = gl_work->gl_interpret_data;
2179 /* invoke the actual glimpse callback */
2180 if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
2183 LDLM_LOCK_RELEASE(lock);
2184 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2185 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2187 OBD_FREE_PTR(gl_work);
2194 * Process a call to completion AST callback for a lock in ast_work list
2197 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2199 struct ldlm_cb_set_arg *arg = opaq;
2200 struct ldlm_lock *lock;
2201 ldlm_completion_callback completion_callback;
2206 if (list_empty(arg->list))
2209 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2211 /* It's possible to receive a completion AST before we've set
2212 * the l_completion_ast pointer: either because the AST arrived
2213 * before the reply, or simply because there's a small race
2214 * window between receiving the reply and finishing the local
2215 * enqueue. (bug 842)
2217 * This can't happen with the blocking_ast, however, because we
2218 * will never call the local blocking_ast until we drop our
2219 * reader/writer reference, which we won't do until we get the
2220 * reply and finish enqueueing. */
2222 /* nobody should touch l_cp_ast */
2223 lock_res_and_lock(lock);
2224 list_del_init(&lock->l_cp_ast);
2225 LASSERT(ldlm_is_cp_reqd(lock));
2226 /* save l_completion_ast since it can be changed by
2227 * mds_intent_policy(), see bug 14225 */
2228 completion_callback = lock->l_completion_ast;
2229 ldlm_clear_cp_reqd(lock);
2230 unlock_res_and_lock(lock);
2232 if (completion_callback != NULL)
2233 rc = completion_callback(lock, 0, (void *)arg);
2234 LDLM_LOCK_RELEASE(lock);
2240 * Process list of locks in need of ASTs being sent.
2242 * Used on server to send multiple ASTs together instead of sending one by
2245 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2246 ldlm_desc_ast_t ast_type)
2248 struct ldlm_cb_set_arg *arg;
2249 set_producer_func work_ast_lock;
2252 if (list_empty(rpc_list))
2259 atomic_set(&arg->restart, 0);
2260 arg->list = rpc_list;
2263 case LDLM_WORK_CP_AST:
2264 arg->type = LDLM_CP_CALLBACK;
2265 work_ast_lock = ldlm_work_cp_ast_lock;
2267 #ifdef HAVE_SERVER_SUPPORT
2268 case LDLM_WORK_BL_AST:
2269 arg->type = LDLM_BL_CALLBACK;
2270 work_ast_lock = ldlm_work_bl_ast_lock;
2272 case LDLM_WORK_REVOKE_AST:
2273 arg->type = LDLM_BL_CALLBACK;
2274 work_ast_lock = ldlm_work_revoke_ast_lock;
2276 case LDLM_WORK_GL_AST:
2277 arg->type = LDLM_GL_CALLBACK;
2278 work_ast_lock = ldlm_work_gl_ast_lock;
2285 /* We create a ptlrpc request set with flow control extension.
2286 * This request set will use the work_ast_lock function to produce new
2287 * requests and will send a new request each time one completes in order
2288 * to keep the number of requests in flight to ns_max_parallel_ast */
2289 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2290 work_ast_lock, arg);
2291 if (arg->set == NULL)
2292 GOTO(out, rc = -ENOMEM);
2294 ptlrpc_set_wait(NULL, arg->set);
2295 ptlrpc_set_destroy(arg->set);
2297 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2305 * Try to grant all waiting locks on a resource.
2307 * Calls ldlm_reprocess_queue on waiting queue.
2309 * Typically called after some resource locks are cancelled to see
2310 * if anything could be granted as a result of the cancellation.
2312 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2313 enum ldlm_process_intention intention,
2314 struct ldlm_lock *hint)
2316 struct list_head rpc_list;
2317 #ifdef HAVE_SERVER_SUPPORT
2318 ldlm_reprocessing_policy reprocess;
2319 struct obd_device *obd;
2324 INIT_LIST_HEAD(&rpc_list);
2325 /* Local lock trees don't get reprocessed. */
2326 if (ns_is_client(ldlm_res_to_ns(res))) {
2331 /* Disable reprocess during lock replay stage but allow during
2332 * request replay stage.
2334 obd = ldlm_res_to_ns(res)->ns_obd;
2335 if (obd->obd_recovering &&
2336 atomic_read(&obd->obd_req_replay_clients) == 0)
2340 reprocess = ldlm_get_reprocessing_policy(res);
2341 reprocess(res, &res->lr_waiting, &rpc_list, intention, hint);
2344 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2346 if (rc == -ERESTART) {
2347 LASSERT(list_empty(&rpc_list));
2353 INIT_LIST_HEAD(&rpc_list);
2354 if (!ns_is_client(ldlm_res_to_ns(res))) {
2355 CERROR("This is client-side-only module, cannot handle "
2356 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2363 void ldlm_reprocess_all(struct ldlm_resource *res, struct ldlm_lock *hint)
2365 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN, hint);
2367 EXPORT_SYMBOL(ldlm_reprocess_all);
2369 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2370 struct hlist_node *hnode, void *arg)
2372 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2374 /* This is only called once after recovery done. LU-8306. */
2375 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY, NULL);
2380 * Iterate through all resources on a namespace attempting to grant waiting
2383 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2388 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2389 ldlm_reprocess_res, NULL, 0);
2395 * Helper function to call blocking AST for LDLM lock \a lock in a
2396 * "cancelling" mode.
2398 void ldlm_cancel_callback(struct ldlm_lock *lock)
2400 check_res_locked(lock->l_resource);
2401 if (!ldlm_is_cancel(lock)) {
2402 ldlm_set_cancel(lock);
2403 if (lock->l_blocking_ast) {
2404 unlock_res_and_lock(lock);
2405 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2407 lock_res_and_lock(lock);
2409 LDLM_DEBUG(lock, "no blocking ast");
2412 /* only canceller can set bl_done bit */
2413 ldlm_set_bl_done(lock);
2414 wake_up_all(&lock->l_waitq);
2415 } else if (!ldlm_is_bl_done(lock)) {
2416 /* The lock is guaranteed to have been canceled once
2417 * returning from this function. */
2418 unlock_res_and_lock(lock);
2419 wait_event_idle(lock->l_waitq, is_bl_done(lock));
2420 lock_res_and_lock(lock);
2425 * Remove skiplist-enabled LDLM lock \a req from granted list
2427 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2429 if (req->l_resource->lr_type != LDLM_PLAIN &&
2430 req->l_resource->lr_type != LDLM_IBITS)
2433 list_del_init(&req->l_sl_policy);
2434 list_del_init(&req->l_sl_mode);
2438 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2440 void ldlm_lock_cancel(struct ldlm_lock *lock)
2442 struct ldlm_resource *res;
2443 struct ldlm_namespace *ns;
2446 lock_res_and_lock(lock);
2448 res = lock->l_resource;
2449 ns = ldlm_res_to_ns(res);
2451 /* Please do not, no matter how tempting, remove this LBUG without
2452 * talking to me first. -phik */
2453 if (lock->l_readers || lock->l_writers) {
2454 LDLM_ERROR(lock, "lock still has references");
2455 unlock_res_and_lock(lock);
2459 if (ldlm_is_waited(lock))
2460 ldlm_del_waiting_lock(lock);
2462 /* Releases cancel callback. */
2463 ldlm_cancel_callback(lock);
2465 /* Yes, second time, just in case it was added again while we were
2466 * running with no res lock in ldlm_cancel_callback */
2467 if (ldlm_is_waited(lock))
2468 ldlm_del_waiting_lock(lock);
2470 ldlm_resource_unlink_lock(lock);
2471 ldlm_lock_destroy_nolock(lock);
2473 if (ldlm_is_granted(lock))
2474 ldlm_pool_del(&ns->ns_pool, lock);
2476 /* Make sure we will not be called again for same lock what is possible
2477 * if not to zero out lock->l_granted_mode */
2478 lock->l_granted_mode = LCK_MINMODE;
2479 unlock_res_and_lock(lock);
2483 EXPORT_SYMBOL(ldlm_lock_cancel);
2486 * Set opaque data into the lock that only makes sense to upper layer.
2488 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2490 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2495 if (lock->l_ast_data == NULL)
2496 lock->l_ast_data = data;
2497 if (lock->l_ast_data == data)
2499 LDLM_LOCK_PUT(lock);
2503 EXPORT_SYMBOL(ldlm_lock_set_data);
2505 struct export_cl_data {
2506 const struct lu_env *ecl_env;
2507 struct obd_export *ecl_exp;
2511 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2512 struct ldlm_lock *lock,
2513 struct export_cl_data *ecl)
2515 struct ldlm_resource *res;
2517 res = ldlm_resource_getref(lock->l_resource);
2519 ldlm_lvbo_update(res, lock, NULL, 1);
2520 ldlm_lock_cancel(lock);
2521 if (!exp->exp_obd->obd_stopping)
2522 ldlm_reprocess_all(res, lock);
2523 ldlm_resource_putref(res);
2526 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2527 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2528 exp, ecl->ecl_loop);
2533 * Iterator function for ldlm_export_cancel_locks.
2534 * Cancels passed locks.
2537 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2538 struct hlist_node *hnode, void *data)
2541 struct export_cl_data *ecl = (struct export_cl_data *)data;
2542 struct obd_export *exp = ecl->ecl_exp;
2543 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2545 LDLM_LOCK_GET(lock);
2546 ldlm_cancel_lock_for_export(exp, lock, ecl);
2547 LDLM_LOCK_RELEASE(lock);
2553 * Cancel all blocked locks for given export.
2555 * Typically called on client disconnection/eviction
2557 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2560 struct export_cl_data ecl = {
2566 rc = lu_env_init(&env, LCT_DT_THREAD);
2571 while (!list_empty(&exp->exp_bl_list)) {
2572 struct ldlm_lock *lock;
2574 spin_lock_bh(&exp->exp_bl_list_lock);
2575 if (!list_empty(&exp->exp_bl_list)) {
2576 lock = list_entry(exp->exp_bl_list.next,
2577 struct ldlm_lock, l_exp_list);
2578 LDLM_LOCK_GET(lock);
2579 list_del_init(&lock->l_exp_list);
2583 spin_unlock_bh(&exp->exp_bl_list_lock);
2588 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2589 LDLM_LOCK_RELEASE(lock);
2594 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2595 "left on hash table %d.\n", exp, ecl.ecl_loop,
2596 atomic_read(&exp->exp_lock_hash->hs_count));
2598 return ecl.ecl_loop;
2602 * Cancel all locks for given export.
2604 * Typically called after client disconnection/eviction
2606 int ldlm_export_cancel_locks(struct obd_export *exp)
2608 struct export_cl_data ecl;
2612 rc = lu_env_init(&env, LCT_DT_THREAD);
2619 cfs_hash_for_each_empty(exp->exp_lock_hash,
2620 ldlm_cancel_locks_for_export_cb, &ecl);
2622 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2623 "left on hash table %d.\n", exp, ecl.ecl_loop,
2624 atomic_read(&exp->exp_lock_hash->hs_count));
2626 if (ecl.ecl_loop > 0 &&
2627 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2628 exp->exp_obd->obd_stopping)
2629 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2633 return ecl.ecl_loop;
2637 * Downgrade an PW/EX lock to COS | CR mode.
2639 * A lock mode convertion from PW/EX mode to less conflict mode. The
2640 * convertion may fail if lock was canceled before downgrade, but it doesn't
2641 * indicate any problem, because such lock has no reader or writer, and will
2644 * Used by Commit on Sharing (COS) code to force object changes commit in case
2645 * of conflict. Converted lock is considered as new lock and all blocking AST
2646 * things are cleared, so any pending or new blocked lock on that lock will
2647 * cause new call to blocking_ast and force resource object commit.
2649 * Also used by layout_change to replace EX lock to CR lock.
2651 * \param lock A lock to convert
2652 * \param new_mode new lock mode
2654 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2656 #ifdef HAVE_SERVER_SUPPORT
2659 LASSERT(new_mode == LCK_COS || new_mode == LCK_CR);
2661 lock_res_and_lock(lock);
2663 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2664 unlock_res_and_lock(lock);
2666 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2667 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2671 ldlm_resource_unlink_lock(lock);
2673 * Remove the lock from pool as it will be added again in
2674 * ldlm_grant_lock() called below.
2676 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2678 /* Consider downgraded lock as a new lock and clear all states
2679 * related to a previous blocking AST processing.
2681 ldlm_clear_blocking_data(lock);
2683 lock->l_req_mode = new_mode;
2684 ldlm_grant_lock(lock, NULL);
2685 unlock_res_and_lock(lock);
2687 ldlm_reprocess_all(lock->l_resource, lock);
2692 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2695 * Print lock with lock handle \a lockh description into debug log.
2697 * Used when printing all locks on a resource for debug purposes.
2699 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2701 struct ldlm_lock *lock;
2703 if (!((libcfs_debug | D_ERROR) & level))
2706 lock = ldlm_handle2lock(lockh);
2710 LDLM_DEBUG_LIMIT(level, lock, "###");
2712 LDLM_LOCK_PUT(lock);
2714 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2717 * Print lock information with custom message into debug log.
2720 void _ldlm_lock_debug(struct ldlm_lock *lock,
2721 struct libcfs_debug_msg_data *msgdata,
2722 const char *fmt, ...)
2725 struct obd_export *exp = lock->l_export;
2726 struct ldlm_resource *resource = NULL;
2727 struct va_format vaf;
2728 char *nid = "local";
2730 /* on server-side resource of lock doesn't change */
2731 if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2732 if (lock->l_resource != NULL)
2733 resource = ldlm_resource_getref(lock->l_resource);
2734 } else if (spin_trylock(&lock->l_lock)) {
2735 if (lock->l_resource != NULL)
2736 resource = ldlm_resource_getref(lock->l_resource);
2737 spin_unlock(&lock->l_lock);
2740 va_start(args, fmt);
2744 if (exp && exp->exp_connection) {
2745 nid = obd_export_nid2str(exp);
2746 } else if (exp && exp->exp_obd != NULL) {
2747 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2748 nid = obd_import_nid2str(imp);
2751 if (resource == NULL) {
2752 libcfs_debug_msg(msgdata,
2753 "%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",
2756 lock->l_handle.h_cookie,
2757 refcount_read(&lock->l_handle.h_ref),
2758 lock->l_readers, lock->l_writers,
2759 ldlm_lockname[lock->l_granted_mode],
2760 ldlm_lockname[lock->l_req_mode],
2762 lock->l_remote_handle.cookie,
2763 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2764 lock->l_pid, lock->l_callback_timeout,
2770 switch (resource->lr_type) {
2772 libcfs_debug_msg(msgdata,
2773 "%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",
2775 ldlm_lock_to_ns_name(lock), lock,
2776 lock->l_handle.h_cookie,
2777 refcount_read(&lock->l_handle.h_ref),
2778 lock->l_readers, lock->l_writers,
2779 ldlm_lockname[lock->l_granted_mode],
2780 ldlm_lockname[lock->l_req_mode],
2782 atomic_read(&resource->lr_refcount),
2783 ldlm_typename[resource->lr_type],
2784 lock->l_policy_data.l_extent.start,
2785 lock->l_policy_data.l_extent.end,
2786 lock->l_req_extent.start, lock->l_req_extent.end,
2788 lock->l_remote_handle.cookie,
2789 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2790 lock->l_pid, lock->l_callback_timeout,
2795 libcfs_debug_msg(msgdata,
2796 "%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",
2798 ldlm_lock_to_ns_name(lock), lock,
2799 lock->l_handle.h_cookie,
2800 refcount_read(&lock->l_handle.h_ref),
2801 lock->l_readers, lock->l_writers,
2802 ldlm_lockname[lock->l_granted_mode],
2803 ldlm_lockname[lock->l_req_mode],
2805 atomic_read(&resource->lr_refcount),
2806 ldlm_typename[resource->lr_type],
2807 lock->l_policy_data.l_flock.pid,
2808 lock->l_policy_data.l_flock.start,
2809 lock->l_policy_data.l_flock.end,
2811 lock->l_remote_handle.cookie,
2812 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2813 lock->l_pid, lock->l_callback_timeout);
2817 libcfs_debug_msg(msgdata,
2818 "%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",
2820 ldlm_lock_to_ns_name(lock),
2821 lock, lock->l_handle.h_cookie,
2822 refcount_read(&lock->l_handle.h_ref),
2823 lock->l_readers, lock->l_writers,
2824 ldlm_lockname[lock->l_granted_mode],
2825 ldlm_lockname[lock->l_req_mode],
2827 lock->l_policy_data.l_inodebits.bits,
2828 lock->l_policy_data.l_inodebits.try_bits,
2829 atomic_read(&resource->lr_refcount),
2830 ldlm_typename[resource->lr_type],
2832 lock->l_remote_handle.cookie,
2833 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2834 lock->l_pid, lock->l_callback_timeout,
2839 libcfs_debug_msg(msgdata,
2840 "%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",
2842 ldlm_lock_to_ns_name(lock),
2843 lock, lock->l_handle.h_cookie,
2844 refcount_read(&lock->l_handle.h_ref),
2845 lock->l_readers, lock->l_writers,
2846 ldlm_lockname[lock->l_granted_mode],
2847 ldlm_lockname[lock->l_req_mode],
2849 atomic_read(&resource->lr_refcount),
2850 ldlm_typename[resource->lr_type],
2852 lock->l_remote_handle.cookie,
2853 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2854 lock->l_pid, lock->l_callback_timeout,
2859 ldlm_resource_putref(resource);
2861 EXPORT_SYMBOL(_ldlm_lock_debug);