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 if (!lustre_handle_is_used(handle))
606 lock = class_handle2object(handle->cookie, &lock_handle_ops);
611 if (lock->l_export != NULL && lock->l_export->exp_failed) {
612 CDEBUG(D_INFO, "lock export failed: lock %p, exp %p\n",
613 lock, lock->l_export);
618 /* It's unlikely but possible that someone marked the lock as
619 * destroyed after we did handle2object on it */
620 if ((flags == 0) && !ldlm_is_destroyed(lock)) {
621 lu_ref_add(&lock->l_reference, "handle", current);
625 lock_res_and_lock(lock);
627 LASSERT(lock->l_resource != NULL);
629 lu_ref_add_atomic(&lock->l_reference, "handle", current);
630 if (unlikely(ldlm_is_destroyed(lock))) {
631 unlock_res_and_lock(lock);
632 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
637 /* If we're setting flags, make sure none of them are already set. */
639 if ((lock->l_flags & flags) != 0) {
640 unlock_res_and_lock(lock);
645 lock->l_flags |= flags;
648 unlock_res_and_lock(lock);
651 EXPORT_SYMBOL(__ldlm_handle2lock);
652 /** @} ldlm_handles */
655 * Fill in "on the wire" representation for given LDLM lock into supplied
656 * lock descriptor \a desc structure.
658 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
660 ldlm_res2desc(lock->l_resource, &desc->l_resource);
661 desc->l_req_mode = lock->l_req_mode;
662 desc->l_granted_mode = lock->l_granted_mode;
663 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
664 &lock->l_policy_data,
665 &desc->l_policy_data);
669 * Add a lock to list of conflicting locks to send AST to.
671 * Only add if we have not sent a blocking AST to the lock yet.
673 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
674 struct list_head *work_list)
676 if (!ldlm_is_ast_sent(lock)) {
677 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
678 ldlm_set_ast_sent(lock);
679 /* If the enqueuing client said so, tell the AST recipient to
680 * discard dirty data, rather than writing back. */
681 if (ldlm_is_ast_discard_data(new))
682 ldlm_set_discard_data(lock);
684 /* Lock can be converted from a blocking state back to granted
685 * after lock convert or COS downgrade but still be in an
686 * older bl_list because it is controlled only by
687 * ldlm_work_bl_ast_lock(), let it be processed there.
689 if (list_empty(&lock->l_bl_ast)) {
690 list_add(&lock->l_bl_ast, work_list);
693 LASSERT(lock->l_blocking_lock == NULL);
694 lock->l_blocking_lock = LDLM_LOCK_GET(new);
699 * Add a lock to list of just granted locks to send completion AST to.
701 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
702 struct list_head *work_list)
704 if (!ldlm_is_cp_reqd(lock)) {
705 ldlm_set_cp_reqd(lock);
706 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
707 LASSERT(list_empty(&lock->l_cp_ast));
708 list_add(&lock->l_cp_ast, work_list);
714 * Aggregator function to add AST work items into a list. Determines
715 * what sort of an AST work needs to be done and calls the proper
717 * Must be called with lr_lock held.
719 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
720 struct list_head *work_list)
723 check_res_locked(lock->l_resource);
725 ldlm_add_bl_work_item(lock, new, work_list);
727 ldlm_add_cp_work_item(lock, work_list);
732 * Add specified reader/writer reference to LDLM lock with handle \a lockh.
733 * r/w reference type is determined by \a mode
734 * Calls ldlm_lock_addref_internal.
736 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
738 struct ldlm_lock *lock;
740 lock = ldlm_handle2lock(lockh);
741 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
742 ldlm_lock_addref_internal(lock, mode);
745 EXPORT_SYMBOL(ldlm_lock_addref);
749 * Add specified reader/writer reference to LDLM lock \a lock.
750 * r/w reference type is determined by \a mode
751 * Removes lock from LRU if it is there.
752 * Assumes the LDLM lock is already locked.
754 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
757 ldlm_lock_remove_from_lru(lock);
758 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
760 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
762 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
764 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
767 lu_ref_add_atomic(&lock->l_reference, "user", lock);
768 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
772 * Attempts to add reader/writer reference to a lock with handle \a lockh, and
773 * fails if lock is already LDLM_FL_CBPENDING or destroyed.
775 * \retval 0 success, lock was addref-ed
777 * \retval -EAGAIN lock is being canceled.
779 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
781 struct ldlm_lock *lock;
785 lock = ldlm_handle2lock(lockh);
787 lock_res_and_lock(lock);
788 if (lock->l_readers != 0 || lock->l_writers != 0 ||
789 !ldlm_is_cbpending(lock)) {
790 ldlm_lock_addref_internal_nolock(lock, mode);
793 unlock_res_and_lock(lock);
798 EXPORT_SYMBOL(ldlm_lock_addref_try);
801 * Add specified reader/writer reference to LDLM lock \a lock.
802 * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
803 * Only called for local locks.
805 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
807 lock_res_and_lock(lock);
808 ldlm_lock_addref_internal_nolock(lock, mode);
809 unlock_res_and_lock(lock);
813 * Removes reader/writer reference for LDLM lock \a lock.
814 * Assumes LDLM lock is already locked.
815 * only called in ldlm_flock_destroy and for local locks.
816 * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
817 * that cannot be placed in LRU.
819 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
822 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
823 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
824 LASSERT(lock->l_readers > 0);
825 lu_ref_del(&lock->l_reference, "reader", lock);
828 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
829 LASSERT(lock->l_writers > 0);
830 lu_ref_del(&lock->l_reference, "writer", lock);
834 lu_ref_del(&lock->l_reference, "user", lock);
835 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
839 * Removes reader/writer reference for LDLM lock \a lock.
840 * Locks LDLM lock first.
841 * If the lock is determined to be client lock on a client and r/w refcount
842 * drops to zero and the lock is not blocked, the lock is added to LRU lock
844 * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
846 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
848 struct ldlm_namespace *ns;
851 lock_res_and_lock(lock);
853 ns = ldlm_lock_to_ns(lock);
855 ldlm_lock_decref_internal_nolock(lock, mode);
857 if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
858 !lock->l_readers && !lock->l_writers) {
859 /* If this is a local lock on a server namespace and this was
860 * the last reference, cancel the lock.
862 * Group locks are special:
863 * They must not go in LRU, but they are not called back
864 * like non-group locks, instead they are manually released.
865 * They have an l_writers reference which they keep until
866 * they are manually released, so we remove them when they have
867 * no more reader or writer references. - LU-6368 */
868 ldlm_set_cbpending(lock);
871 if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
872 /* If we received a blocked AST and this was the last reference,
873 * run the callback. */
874 if (ldlm_is_ns_srv(lock) && lock->l_export)
875 CERROR("FL_CBPENDING set on non-local lock--just a "
878 LDLM_DEBUG(lock, "final decref done on cbpending lock");
880 LDLM_LOCK_GET(lock); /* dropped by bl thread */
881 ldlm_lock_remove_from_lru(lock);
882 unlock_res_and_lock(lock);
884 if (ldlm_is_fail_loc(lock))
885 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
887 if (ldlm_is_atomic_cb(lock) ||
888 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
889 ldlm_handle_bl_callback(ns, NULL, lock);
890 } else if (ns_is_client(ns) &&
891 !lock->l_readers && !lock->l_writers &&
892 !ldlm_is_no_lru(lock) &&
893 !ldlm_is_bl_ast(lock) &&
894 !ldlm_is_converting(lock)) {
896 LDLM_DEBUG(lock, "add lock into lru list");
898 /* If this is a client-side namespace and this was the last
899 * reference, put it on the LRU. */
900 ldlm_lock_add_to_lru(lock);
901 unlock_res_and_lock(lock);
903 if (ldlm_is_fail_loc(lock))
904 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
906 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
907 * are not supported by the server, otherwise, it is done on
909 if (!exp_connect_cancelset(lock->l_conn_export) &&
910 !ns_connect_lru_resize(ns))
911 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
913 LDLM_DEBUG(lock, "do not add lock into lru list");
914 unlock_res_and_lock(lock);
921 * Decrease reader/writer refcount for LDLM lock with handle \a lockh
923 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
925 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
926 LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
927 ldlm_lock_decref_internal(lock, mode);
930 EXPORT_SYMBOL(ldlm_lock_decref);
933 * Decrease reader/writer refcount for LDLM lock with handle
934 * \a lockh and mark it for subsequent cancellation once r/w refcount
935 * drops to zero instead of putting into LRU.
938 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
941 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
944 LASSERT(lock != NULL);
946 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
947 lock_res_and_lock(lock);
948 ldlm_set_cbpending(lock);
949 unlock_res_and_lock(lock);
950 ldlm_lock_decref_internal(lock, mode);
953 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
955 struct sl_insert_point {
956 struct list_head *res_link;
957 struct list_head *mode_link;
958 struct list_head *policy_link;
962 * Finds a position to insert the new lock into granted lock list.
964 * Used for locks eligible for skiplist optimization.
967 * queue [input]: the granted list where search acts on;
968 * req [input]: the lock whose position to be located;
969 * prev [output]: positions within 3 lists to insert @req to
973 * - ldlm_grant_lock_with_skiplist
975 static void search_granted_lock(struct list_head *queue,
976 struct ldlm_lock *req,
977 struct sl_insert_point *prev)
979 struct list_head *tmp;
980 struct ldlm_lock *lock, *mode_end, *policy_end;
983 list_for_each(tmp, queue) {
984 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
986 mode_end = list_entry(lock->l_sl_mode.prev,
987 struct ldlm_lock, l_sl_mode);
989 if (lock->l_req_mode != req->l_req_mode) {
990 /* jump to last lock of mode group */
991 tmp = &mode_end->l_res_link;
995 /* suitable mode group is found */
996 if (lock->l_resource->lr_type == LDLM_PLAIN) {
997 /* insert point is last lock of the mode group */
998 prev->res_link = &mode_end->l_res_link;
999 prev->mode_link = &mode_end->l_sl_mode;
1000 prev->policy_link = &req->l_sl_policy;
1003 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
1006 list_entry(lock->l_sl_policy.prev,
1010 if (lock->l_policy_data.l_inodebits.bits ==
1011 req->l_policy_data.l_inodebits.bits) {
1012 /* insert point is last lock of
1013 * the policy group */
1015 &policy_end->l_res_link;
1017 &policy_end->l_sl_mode;
1019 &policy_end->l_sl_policy;
1024 if (policy_end == mode_end)
1025 /* done with mode group */
1028 /* go to next policy group within mode group */
1029 tmp = policy_end->l_res_link.next;
1030 lock = list_entry(tmp, struct ldlm_lock,
1032 } /* loop over policy groups within the mode group */
1034 /* insert point is last lock of the mode group,
1035 * new policy group is started */
1036 prev->res_link = &mode_end->l_res_link;
1037 prev->mode_link = &mode_end->l_sl_mode;
1038 prev->policy_link = &req->l_sl_policy;
1042 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1047 /* insert point is last lock on the queue,
1048 * new mode group and new policy group are started */
1049 prev->res_link = queue->prev;
1050 prev->mode_link = &req->l_sl_mode;
1051 prev->policy_link = &req->l_sl_policy;
1056 * Add a lock into resource granted list after a position described by
1059 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1060 struct sl_insert_point *prev)
1062 struct ldlm_resource *res = lock->l_resource;
1065 check_res_locked(res);
1067 ldlm_resource_dump(D_INFO, res);
1068 LDLM_DEBUG(lock, "About to add lock:");
1070 if (ldlm_is_destroyed(lock)) {
1071 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1075 LASSERT(list_empty(&lock->l_res_link));
1076 LASSERT(list_empty(&lock->l_sl_mode));
1077 LASSERT(list_empty(&lock->l_sl_policy));
1080 * lock->link == prev->link means lock is first starting the group.
1081 * Don't re-add to itself to suppress kernel warnings.
1083 if (&lock->l_res_link != prev->res_link)
1084 list_add(&lock->l_res_link, prev->res_link);
1085 if (&lock->l_sl_mode != prev->mode_link)
1086 list_add(&lock->l_sl_mode, prev->mode_link);
1087 if (&lock->l_sl_policy != prev->policy_link)
1088 list_add(&lock->l_sl_policy, prev->policy_link);
1094 * Add a lock to granted list on a resource maintaining skiplist
1097 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1099 struct sl_insert_point prev;
1101 LASSERT(ldlm_is_granted(lock));
1103 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1104 ldlm_granted_list_add_lock(lock, &prev);
1108 * Perform lock granting bookkeeping.
1110 * Includes putting the lock into granted list and updating lock mode.
1112 * - ldlm_lock_enqueue
1113 * - ldlm_reprocess_queue
1115 * must be called with lr_lock held
1117 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1119 struct ldlm_resource *res = lock->l_resource;
1122 check_res_locked(res);
1124 lock->l_granted_mode = lock->l_req_mode;
1126 if (work_list && lock->l_completion_ast != NULL)
1127 ldlm_add_ast_work_item(lock, NULL, work_list);
1129 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1130 ldlm_grant_lock_with_skiplist(lock);
1131 else if (res->lr_type == LDLM_EXTENT)
1132 ldlm_extent_add_lock(res, lock);
1133 else if (res->lr_type == LDLM_FLOCK) {
1134 /* We should not add locks to granted list in the following
1136 * - this is an UNLOCK but not a real lock;
1137 * - this is a TEST lock;
1138 * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1139 * - this is a deadlock (flock cannot be granted) */
1140 if (lock->l_req_mode == 0 ||
1141 lock->l_req_mode == LCK_NL ||
1142 ldlm_is_test_lock(lock) ||
1143 ldlm_is_flock_deadlock(lock))
1145 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1150 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1155 * Check if the given @lock meets the criteria for a match.
1156 * A reference on the lock is taken if matched.
1158 * \param lock test-against this lock
1159 * \param data parameters
1161 static int lock_matches(struct ldlm_lock *lock, struct ldlm_match_data *data)
1163 union ldlm_policy_data *lpol = &lock->l_policy_data;
1164 enum ldlm_mode match = LCK_MINMODE;
1166 if (lock == data->lmd_old)
1167 return INTERVAL_ITER_STOP;
1169 /* Check if this lock can be matched.
1170 * Used by LU-2919(exclusive open) for open lease lock */
1171 if (ldlm_is_excl(lock))
1172 return INTERVAL_ITER_CONT;
1174 /* llite sometimes wants to match locks that will be
1175 * canceled when their users drop, but we allow it to match
1176 * if it passes in CBPENDING and the lock still has users.
1177 * this is generally only going to be used by children
1178 * whose parents already hold a lock so forward progress
1179 * can still happen. */
1180 if (ldlm_is_cbpending(lock) &&
1181 !(data->lmd_flags & LDLM_FL_CBPENDING))
1182 return INTERVAL_ITER_CONT;
1183 if (!data->lmd_unref && ldlm_is_cbpending(lock) &&
1184 lock->l_readers == 0 && lock->l_writers == 0)
1185 return INTERVAL_ITER_CONT;
1187 if (!(lock->l_req_mode & *data->lmd_mode))
1188 return INTERVAL_ITER_CONT;
1190 /* When we search for ast_data, we are not doing a traditional match,
1191 * so we don't worry about IBITS or extent matching.
1193 if (data->lmd_has_ast_data) {
1194 if (!lock->l_ast_data)
1195 return INTERVAL_ITER_CONT;
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_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, int unref)
1410 struct ldlm_match_data data = {
1414 .lmd_policy = policy,
1416 .lmd_skip_flags = skip_flags,
1418 .lmd_has_ast_data = false,
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;
1461 struct l_wait_info lwi;
1463 if (lock->l_completion_ast) {
1464 int err = lock->l_completion_ast(lock,
1465 LDLM_FL_WAIT_NOREPROC,
1468 GOTO(out_fail_match, matched = 0);
1471 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1472 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1474 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1475 l_wait_event(lock->l_waitq, lock->l_flags & wait_flags,
1477 if (!ldlm_is_lvb_ready(lock))
1478 GOTO(out_fail_match, matched = 0);
1481 /* check user's security context */
1482 if (lock->l_conn_export &&
1483 sptlrpc_import_check_ctx(
1484 class_exp2cliimp(lock->l_conn_export)))
1485 GOTO(out_fail_match, matched = 0);
1487 LDLM_DEBUG(lock, "matched (%llu %llu)",
1488 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1489 res_id->name[2] : policy->l_extent.start,
1490 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1491 res_id->name[3] : policy->l_extent.end);
1494 if (flags & LDLM_FL_TEST_LOCK)
1495 LDLM_LOCK_RELEASE(lock);
1497 ldlm_lock_decref_internal(lock, mode);
1500 /* less verbose for test-only */
1501 if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
1502 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1503 "%llu/%llu (%llu %llu)", ns,
1504 type, mode, res_id->name[0], res_id->name[1],
1505 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1506 res_id->name[2] : policy->l_extent.start,
1507 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1508 res_id->name[3] : policy->l_extent.end);
1510 if (data.lmd_old != NULL)
1511 LDLM_LOCK_PUT(data.lmd_old);
1515 EXPORT_SYMBOL(ldlm_lock_match_with_skip);
1517 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1520 struct ldlm_lock *lock;
1521 enum ldlm_mode mode = 0;
1524 lock = ldlm_handle2lock(lockh);
1526 lock_res_and_lock(lock);
1527 if (LDLM_HAVE_MASK(lock, GONE))
1530 if (ldlm_is_cbpending(lock) &&
1531 lock->l_readers == 0 && lock->l_writers == 0)
1535 *bits = lock->l_policy_data.l_inodebits.bits;
1536 mode = lock->l_granted_mode;
1537 ldlm_lock_addref_internal_nolock(lock, mode);
1544 unlock_res_and_lock(lock);
1545 LDLM_LOCK_PUT(lock);
1549 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1551 /** The caller must guarantee that the buffer is large enough. */
1552 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1553 enum req_location loc, void *data, int size)
1558 LASSERT(data != NULL);
1561 switch (lock->l_lvb_type) {
1563 if (size == sizeof(struct ost_lvb)) {
1564 if (loc == RCL_CLIENT)
1565 lvb = req_capsule_client_swab_get(pill,
1567 lustre_swab_ost_lvb);
1569 lvb = req_capsule_server_swab_get(pill,
1571 lustre_swab_ost_lvb);
1572 if (unlikely(lvb == NULL)) {
1573 LDLM_ERROR(lock, "no LVB");
1577 memcpy(data, lvb, size);
1578 } else if (size == sizeof(struct ost_lvb_v1)) {
1579 struct ost_lvb *olvb = data;
1581 if (loc == RCL_CLIENT)
1582 lvb = req_capsule_client_swab_get(pill,
1584 lustre_swab_ost_lvb_v1);
1586 lvb = req_capsule_server_sized_swab_get(pill,
1588 lustre_swab_ost_lvb_v1);
1589 if (unlikely(lvb == NULL)) {
1590 LDLM_ERROR(lock, "no LVB");
1594 memcpy(data, lvb, size);
1595 olvb->lvb_mtime_ns = 0;
1596 olvb->lvb_atime_ns = 0;
1597 olvb->lvb_ctime_ns = 0;
1599 LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1605 if (size == sizeof(struct lquota_lvb)) {
1606 if (loc == RCL_CLIENT)
1607 lvb = req_capsule_client_swab_get(pill,
1609 lustre_swab_lquota_lvb);
1611 lvb = req_capsule_server_swab_get(pill,
1613 lustre_swab_lquota_lvb);
1614 if (unlikely(lvb == NULL)) {
1615 LDLM_ERROR(lock, "no LVB");
1619 memcpy(data, lvb, size);
1621 LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1630 if (loc == RCL_CLIENT)
1631 lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1633 lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1634 if (unlikely(lvb == NULL)) {
1635 LDLM_ERROR(lock, "no LVB");
1639 memcpy(data, lvb, size);
1642 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1643 libcfs_debug_dumpstack(NULL);
1651 * Create and fill in new LDLM lock with specified properties.
1652 * Returns a referenced lock
1654 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1655 const struct ldlm_res_id *res_id,
1656 enum ldlm_type type,
1657 enum ldlm_mode mode,
1658 const struct ldlm_callback_suite *cbs,
1659 void *data, __u32 lvb_len,
1660 enum lvb_type lvb_type)
1662 struct ldlm_lock *lock;
1663 struct ldlm_resource *res;
1667 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1669 RETURN(ERR_CAST(res));
1671 lock = ldlm_lock_new(res);
1673 RETURN(ERR_PTR(-ENOMEM));
1675 lock->l_req_mode = mode;
1676 lock->l_ast_data = data;
1677 lock->l_pid = current_pid();
1678 if (ns_is_server(ns))
1679 ldlm_set_ns_srv(lock);
1681 lock->l_blocking_ast = cbs->lcs_blocking;
1682 lock->l_completion_ast = cbs->lcs_completion;
1683 lock->l_glimpse_ast = cbs->lcs_glimpse;
1688 rc = ldlm_extent_alloc_lock(lock);
1691 rc = ldlm_inodebits_alloc_lock(lock);
1700 lock->l_lvb_len = lvb_len;
1701 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1702 if (lock->l_lvb_data == NULL)
1703 GOTO(out, rc = -ENOMEM);
1706 lock->l_lvb_type = lvb_type;
1707 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1708 GOTO(out, rc = -ENOENT);
1713 ldlm_lock_destroy(lock);
1714 LDLM_LOCK_RELEASE(lock);
1715 RETURN(ERR_PTR(rc));
1718 #ifdef HAVE_SERVER_SUPPORT
1719 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1722 struct ldlm_resource *res = lock->l_resource;
1723 enum ldlm_error rc = ELDLM_OK;
1724 LIST_HEAD(rpc_list);
1725 ldlm_processing_policy policy;
1729 policy = ldlm_get_processing_policy(res);
1731 policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1732 if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1733 res->lr_type != LDLM_FLOCK) {
1734 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1735 if (rc == -ERESTART)
1739 if (!list_empty(&rpc_list))
1740 ldlm_discard_bl_list(&rpc_list);
1747 * Enqueue (request) a lock.
1749 * Does not block. As a result of enqueue the lock would be put
1750 * into granted or waiting list.
1752 * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1753 * set, skip all the enqueueing and delegate lock processing to intent policy
1756 enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
1757 struct ldlm_namespace *ns,
1758 struct ldlm_lock **lockp,
1759 void *cookie, __u64 *flags)
1761 struct ldlm_lock *lock = *lockp;
1762 struct ldlm_resource *res = lock->l_resource;
1763 int local = ns_is_client(ldlm_res_to_ns(res));
1764 enum ldlm_error rc = ELDLM_OK;
1765 struct ldlm_interval *node = NULL;
1766 #ifdef HAVE_SERVER_SUPPORT
1767 bool reconstruct = false;
1771 /* policies are not executed on the client or during replay */
1772 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1773 && !local && ns->ns_policy) {
1774 rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
1776 if (rc == ELDLM_LOCK_REPLACED) {
1777 /* The lock that was returned has already been granted,
1778 * and placed into lockp. If it's not the same as the
1779 * one we passed in, then destroy the old one and our
1780 * work here is done. */
1781 if (lock != *lockp) {
1782 ldlm_lock_destroy(lock);
1783 LDLM_LOCK_RELEASE(lock);
1785 *flags |= LDLM_FL_LOCK_CHANGED;
1787 } else if (rc != ELDLM_OK &&
1788 ldlm_is_granted(lock)) {
1789 LASSERT(*flags & LDLM_FL_RESENT);
1790 /* It may happen that ns_policy returns an error in
1791 * resend case, object may be unlinked or just some
1792 * error occurs. It is unclear if lock reached the
1793 * client in the original reply, just leave the lock on
1794 * server, not returning it again to client. Due to
1795 * LU-6529, the server will not OOM. */
1797 } else if (rc != ELDLM_OK ||
1798 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1799 ldlm_lock_destroy(lock);
1804 if (*flags & LDLM_FL_RESENT) {
1805 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1806 * Set LOCK_CHANGED always.
1807 * Check if the lock is granted for BLOCK_GRANTED.
1808 * Take NO_TIMEOUT from the lock as it is inherited through
1809 * LDLM_FL_INHERIT_MASK */
1810 *flags |= LDLM_FL_LOCK_CHANGED;
1811 if (!ldlm_is_granted(lock))
1812 *flags |= LDLM_FL_BLOCK_GRANTED;
1813 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1817 /* For a replaying lock, it might be already in granted list. So
1818 * unlinking the lock will cause the interval node to be freed, we
1819 * have to allocate the interval node early otherwise we can't regrant
1820 * this lock in the future. - jay */
1821 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1822 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1824 #ifdef HAVE_SERVER_SUPPORT
1825 reconstruct = !local && res->lr_type == LDLM_FLOCK &&
1826 !(*flags & LDLM_FL_TEST_LOCK);
1828 rc = req_can_reconstruct(cookie, NULL);
1837 lock_res_and_lock(lock);
1838 if (local && ldlm_is_granted(lock)) {
1839 /* The server returned a blocked lock, but it was granted
1840 * before we got a chance to actually enqueue it. We don't
1841 * need to do anything else. */
1842 *flags &= ~LDLM_FL_BLOCKED_MASK;
1843 GOTO(out, rc = ELDLM_OK);
1846 ldlm_resource_unlink_lock(lock);
1847 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1849 ldlm_lock_destroy_nolock(lock);
1850 GOTO(out, rc = -ENOMEM);
1853 INIT_LIST_HEAD(&node->li_group);
1854 ldlm_interval_attach(node, lock);
1858 /* Some flags from the enqueue want to make it into the AST, via the
1859 * lock's l_flags. */
1860 if (*flags & LDLM_FL_AST_DISCARD_DATA)
1861 ldlm_set_ast_discard_data(lock);
1862 if (*flags & LDLM_FL_TEST_LOCK)
1863 ldlm_set_test_lock(lock);
1864 if (*flags & LDLM_FL_COS_INCOMPAT)
1865 ldlm_set_cos_incompat(lock);
1866 if (*flags & LDLM_FL_COS_ENABLED)
1867 ldlm_set_cos_enabled(lock);
1869 /* This distinction between local lock trees is very important; a client
1870 * namespace only has information about locks taken by that client, and
1871 * thus doesn't have enough information to decide for itself if it can
1872 * be granted (below). In this case, we do exactly what the server
1873 * tells us to do, as dictated by the 'flags'.
1875 * We do exactly the same thing during recovery, when the server is
1876 * more or less trusting the clients not to lie.
1878 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1881 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1882 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1884 ldlm_grant_lock(lock, NULL);
1885 GOTO(out, rc = ELDLM_OK);
1886 #ifdef HAVE_SERVER_SUPPORT
1887 } else if (*flags & LDLM_FL_REPLAY) {
1888 if (*flags & LDLM_FL_BLOCK_WAIT) {
1889 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1890 GOTO(out, rc = ELDLM_OK);
1891 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1892 ldlm_grant_lock(lock, NULL);
1893 GOTO(out, rc = ELDLM_OK);
1895 /* If no flags, fall through to normal enqueue path. */
1898 rc = ldlm_lock_enqueue_helper(lock, flags);
1902 CERROR("This is client-side-only module, cannot handle "
1903 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1909 unlock_res_and_lock(lock);
1911 #ifdef HAVE_SERVER_SUPPORT
1913 struct ptlrpc_request *req = cookie;
1915 tgt_mk_reply_data(NULL, NULL,
1916 &req->rq_export->exp_target_data,
1917 req, 0, NULL, false, 0);
1921 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1925 #ifdef HAVE_SERVER_SUPPORT
1927 * Iterate through all waiting locks on a given resource queue and attempt to
1930 * Must be called with resource lock held.
1932 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1933 struct list_head *work_list,
1934 enum ldlm_process_intention intention,
1935 struct ldlm_lock *hint)
1937 struct list_head *tmp, *pos;
1938 ldlm_processing_policy policy;
1940 int rc = LDLM_ITER_CONTINUE;
1941 enum ldlm_error err;
1942 LIST_HEAD(bl_ast_list);
1946 check_res_locked(res);
1948 policy = ldlm_get_processing_policy(res);
1950 LASSERT(intention == LDLM_PROCESS_RESCAN ||
1951 intention == LDLM_PROCESS_RECOVERY);
1954 list_for_each_safe(tmp, pos, queue) {
1955 struct ldlm_lock *pending;
1956 LIST_HEAD(rpc_list);
1958 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1960 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1963 rc = policy(pending, &flags, intention, &err, &rpc_list);
1964 if (pending->l_granted_mode == pending->l_req_mode ||
1965 res->lr_type == LDLM_FLOCK) {
1966 list_splice(&rpc_list, work_list);
1968 list_splice(&rpc_list, &bl_ast_list);
1971 * When this is called from recovery done, we always want
1972 * to scan the whole list no matter what 'rc' is returned.
1974 if (rc != LDLM_ITER_CONTINUE &&
1975 intention == LDLM_PROCESS_RESCAN)
1979 if (!list_empty(&bl_ast_list)) {
1982 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
1986 if (rc == -ERESTART)
1990 if (!list_empty(&bl_ast_list))
1991 ldlm_discard_bl_list(&bl_ast_list);
1993 RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
1997 * Conflicting locks are detected for a lock to be enqueued, add the lock
1998 * into waiting list and send blocking ASTs to the conflicting locks.
2000 * \param[in] lock The lock to be enqueued.
2001 * \param[out] flags Lock flags for the lock to be enqueued.
2002 * \param[in] rpc_list Conflicting locks list.
2004 * \retval -ERESTART: Some lock was instantly canceled while sending
2005 * blocking ASTs, caller needs to re-check conflicting
2007 * \retval -EAGAIN: Lock was destroyed, caller should return error.
2008 * \reval 0: Lock is successfully added in waiting list.
2010 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
2011 struct list_head *rpc_list)
2013 struct ldlm_resource *res = lock->l_resource;
2017 check_res_locked(res);
2019 /* If either of the compat_queue()s returned failure, then we
2020 * have ASTs to send and must go onto the waiting list.
2022 * bug 2322: we used to unlink and re-add here, which was a
2023 * terrible folly -- if we goto restart, we could get
2024 * re-ordered! Causes deadlock, because ASTs aren't sent! */
2025 if (list_empty(&lock->l_res_link))
2026 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
2029 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
2032 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
2033 !ns_is_client(ldlm_res_to_ns(res)))
2034 class_fail_export(lock->l_export);
2037 if (rc == -ERESTART) {
2038 /* 15715: The lock was granted and destroyed after
2039 * resource lock was dropped. Interval node was freed
2040 * in ldlm_lock_destroy. Anyway, this always happens
2041 * when a client is being evicted. So it would be
2042 * ok to return an error. -jay */
2043 if (ldlm_is_destroyed(lock))
2046 /* lock was granted while resource was unlocked. */
2047 if (ldlm_is_granted(lock)) {
2048 /* bug 11300: if the lock has been granted,
2049 * break earlier because otherwise, we will go
2050 * to restart and ldlm_resource_unlink will be
2051 * called and it causes the interval node to be
2052 * freed. Then we will fail at
2053 * ldlm_extent_add_lock() */
2054 *flags &= ~LDLM_FL_BLOCKED_MASK;
2060 *flags |= LDLM_FL_BLOCK_GRANTED;
2066 * Discard all AST work items from list.
2068 * If for whatever reason we do not want to send ASTs to conflicting locks
2069 * anymore, disassemble the list with this function.
2071 void ldlm_discard_bl_list(struct list_head *bl_list)
2073 struct ldlm_lock *lock, *tmp;
2077 list_for_each_entry_safe(lock, tmp, bl_list, l_bl_ast) {
2078 LASSERT(!list_empty(&lock->l_bl_ast));
2079 list_del_init(&lock->l_bl_ast);
2080 ldlm_clear_ast_sent(lock);
2081 LASSERT(lock->l_bl_ast_run == 0);
2082 ldlm_clear_blocking_lock(lock);
2083 LDLM_LOCK_RELEASE(lock);
2089 * Process a call to blocking AST callback for a lock in ast_work list
2092 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2094 struct ldlm_cb_set_arg *arg = opaq;
2095 struct ldlm_lock *lock;
2096 struct ldlm_lock_desc d;
2097 struct ldlm_bl_desc bld;
2102 if (list_empty(arg->list))
2105 lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2107 /* nobody should touch l_bl_ast but some locks in the list may become
2108 * granted after lock convert or COS downgrade, these locks should be
2109 * just skipped here and removed from the list.
2111 lock_res_and_lock(lock);
2112 list_del_init(&lock->l_bl_ast);
2114 /* lock is not blocking lock anymore, but was kept in the list because
2115 * it can managed only here.
2117 if (!ldlm_is_ast_sent(lock)) {
2118 unlock_res_and_lock(lock);
2119 LDLM_LOCK_RELEASE(lock);
2123 LASSERT(lock->l_blocking_lock);
2124 ldlm_lock2desc(lock->l_blocking_lock, &d);
2125 /* copy blocking lock ibits in cancel_bits as well,
2126 * new client may use them for lock convert and it is
2127 * important to use new field to convert locks from
2130 d.l_policy_data.l_inodebits.cancel_bits =
2131 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2133 /* Blocking lock is being destroyed here but some information about it
2134 * may be needed inside l_blocking_ast() function below,
2135 * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
2137 bld.bl_same_client = lock->l_client_cookie ==
2138 lock->l_blocking_lock->l_client_cookie;
2139 bld.bl_cos_incompat = ldlm_is_cos_incompat(lock->l_blocking_lock);
2140 arg->bl_desc = &bld;
2142 LASSERT(ldlm_is_ast_sent(lock));
2143 LASSERT(lock->l_bl_ast_run == 0);
2144 lock->l_bl_ast_run++;
2145 ldlm_clear_blocking_lock(lock);
2146 unlock_res_and_lock(lock);
2148 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2150 LDLM_LOCK_RELEASE(lock);
2156 * Process a call to revocation AST callback for a lock in ast_work list
2159 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2161 struct ldlm_cb_set_arg *arg = opaq;
2162 struct ldlm_lock_desc desc;
2164 struct ldlm_lock *lock;
2167 if (list_empty(arg->list))
2170 lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2171 list_del_init(&lock->l_rk_ast);
2173 /* the desc just pretend to exclusive */
2174 ldlm_lock2desc(lock, &desc);
2175 desc.l_req_mode = LCK_EX;
2176 desc.l_granted_mode = 0;
2178 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2179 LDLM_LOCK_RELEASE(lock);
2185 * Process a call to glimpse AST callback for a lock in ast_work list
2187 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2189 struct ldlm_cb_set_arg *arg = opaq;
2190 struct ldlm_glimpse_work *gl_work;
2191 struct ldlm_lock *lock;
2195 if (list_empty(arg->list))
2198 gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2200 list_del_init(&gl_work->gl_list);
2202 lock = gl_work->gl_lock;
2204 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2205 arg->gl_desc = gl_work->gl_desc;
2206 arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2207 arg->gl_interpret_data = gl_work->gl_interpret_data;
2209 /* invoke the actual glimpse callback */
2210 if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
2213 LDLM_LOCK_RELEASE(lock);
2214 if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2215 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2217 OBD_FREE_PTR(gl_work);
2224 * Process a call to completion AST callback for a lock in ast_work list
2227 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2229 struct ldlm_cb_set_arg *arg = opaq;
2230 struct ldlm_lock *lock;
2231 ldlm_completion_callback completion_callback;
2236 if (list_empty(arg->list))
2239 lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2241 /* It's possible to receive a completion AST before we've set
2242 * the l_completion_ast pointer: either because the AST arrived
2243 * before the reply, or simply because there's a small race
2244 * window between receiving the reply and finishing the local
2245 * enqueue. (bug 842)
2247 * This can't happen with the blocking_ast, however, because we
2248 * will never call the local blocking_ast until we drop our
2249 * reader/writer reference, which we won't do until we get the
2250 * reply and finish enqueueing. */
2252 /* nobody should touch l_cp_ast */
2253 lock_res_and_lock(lock);
2254 list_del_init(&lock->l_cp_ast);
2255 LASSERT(ldlm_is_cp_reqd(lock));
2256 /* save l_completion_ast since it can be changed by
2257 * mds_intent_policy(), see bug 14225 */
2258 completion_callback = lock->l_completion_ast;
2259 ldlm_clear_cp_reqd(lock);
2260 unlock_res_and_lock(lock);
2262 if (completion_callback != NULL)
2263 rc = completion_callback(lock, 0, (void *)arg);
2264 LDLM_LOCK_RELEASE(lock);
2270 * Process list of locks in need of ASTs being sent.
2272 * Used on server to send multiple ASTs together instead of sending one by
2275 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2276 ldlm_desc_ast_t ast_type)
2278 struct ldlm_cb_set_arg *arg;
2279 set_producer_func work_ast_lock;
2282 if (list_empty(rpc_list))
2289 atomic_set(&arg->restart, 0);
2290 arg->list = rpc_list;
2293 case LDLM_WORK_CP_AST:
2294 arg->type = LDLM_CP_CALLBACK;
2295 work_ast_lock = ldlm_work_cp_ast_lock;
2297 #ifdef HAVE_SERVER_SUPPORT
2298 case LDLM_WORK_BL_AST:
2299 arg->type = LDLM_BL_CALLBACK;
2300 work_ast_lock = ldlm_work_bl_ast_lock;
2302 case LDLM_WORK_REVOKE_AST:
2303 arg->type = LDLM_BL_CALLBACK;
2304 work_ast_lock = ldlm_work_revoke_ast_lock;
2306 case LDLM_WORK_GL_AST:
2307 arg->type = LDLM_GL_CALLBACK;
2308 work_ast_lock = ldlm_work_gl_ast_lock;
2315 /* We create a ptlrpc request set with flow control extension.
2316 * This request set will use the work_ast_lock function to produce new
2317 * requests and will send a new request each time one completes in order
2318 * to keep the number of requests in flight to ns_max_parallel_ast */
2319 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2320 work_ast_lock, arg);
2321 if (arg->set == NULL)
2322 GOTO(out, rc = -ENOMEM);
2324 ptlrpc_set_wait(NULL, arg->set);
2325 ptlrpc_set_destroy(arg->set);
2327 rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2335 * Try to grant all waiting locks on a resource.
2337 * Calls ldlm_reprocess_queue on waiting queue.
2339 * Typically called after some resource locks are cancelled to see
2340 * if anything could be granted as a result of the cancellation.
2342 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2343 enum ldlm_process_intention intention,
2344 struct ldlm_lock *hint)
2346 struct list_head rpc_list;
2347 #ifdef HAVE_SERVER_SUPPORT
2348 ldlm_reprocessing_policy reprocess;
2349 struct obd_device *obd;
2354 INIT_LIST_HEAD(&rpc_list);
2355 /* Local lock trees don't get reprocessed. */
2356 if (ns_is_client(ldlm_res_to_ns(res))) {
2361 /* Disable reprocess during lock replay stage but allow during
2362 * request replay stage.
2364 obd = ldlm_res_to_ns(res)->ns_obd;
2365 if (obd->obd_recovering &&
2366 atomic_read(&obd->obd_req_replay_clients) == 0)
2370 reprocess = ldlm_get_reprocessing_policy(res);
2371 reprocess(res, &res->lr_waiting, &rpc_list, intention, hint);
2374 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2376 if (rc == -ERESTART) {
2377 LASSERT(list_empty(&rpc_list));
2383 INIT_LIST_HEAD(&rpc_list);
2384 if (!ns_is_client(ldlm_res_to_ns(res))) {
2385 CERROR("This is client-side-only module, cannot handle "
2386 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2393 void ldlm_reprocess_all(struct ldlm_resource *res, struct ldlm_lock *hint)
2395 __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN, hint);
2397 EXPORT_SYMBOL(ldlm_reprocess_all);
2399 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2400 struct hlist_node *hnode, void *arg)
2402 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2404 /* This is only called once after recovery done. LU-8306. */
2405 __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY, NULL);
2410 * Iterate through all resources on a namespace attempting to grant waiting
2413 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2418 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2419 ldlm_reprocess_res, NULL, 0);
2425 * Helper function to call blocking AST for LDLM lock \a lock in a
2426 * "cancelling" mode.
2428 void ldlm_cancel_callback(struct ldlm_lock *lock)
2430 check_res_locked(lock->l_resource);
2431 if (!ldlm_is_cancel(lock)) {
2432 ldlm_set_cancel(lock);
2433 if (lock->l_blocking_ast) {
2434 unlock_res_and_lock(lock);
2435 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2437 lock_res_and_lock(lock);
2439 LDLM_DEBUG(lock, "no blocking ast");
2442 /* only canceller can set bl_done bit */
2443 ldlm_set_bl_done(lock);
2444 wake_up_all(&lock->l_waitq);
2445 } else if (!ldlm_is_bl_done(lock)) {
2446 /* The lock is guaranteed to have been canceled once
2447 * returning from this function. */
2448 unlock_res_and_lock(lock);
2449 wait_event_idle(lock->l_waitq, is_bl_done(lock));
2450 lock_res_and_lock(lock);
2455 * Remove skiplist-enabled LDLM lock \a req from granted list
2457 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2459 if (req->l_resource->lr_type != LDLM_PLAIN &&
2460 req->l_resource->lr_type != LDLM_IBITS)
2463 list_del_init(&req->l_sl_policy);
2464 list_del_init(&req->l_sl_mode);
2468 * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2470 void ldlm_lock_cancel(struct ldlm_lock *lock)
2472 struct ldlm_resource *res;
2473 struct ldlm_namespace *ns;
2476 lock_res_and_lock(lock);
2478 res = lock->l_resource;
2479 ns = ldlm_res_to_ns(res);
2481 /* Please do not, no matter how tempting, remove this LBUG without
2482 * talking to me first. -phik */
2483 if (lock->l_readers || lock->l_writers) {
2484 LDLM_ERROR(lock, "lock still has references");
2485 unlock_res_and_lock(lock);
2489 if (ldlm_is_waited(lock))
2490 ldlm_del_waiting_lock(lock);
2492 /* Releases cancel callback. */
2493 ldlm_cancel_callback(lock);
2495 /* Yes, second time, just in case it was added again while we were
2496 * running with no res lock in ldlm_cancel_callback */
2497 if (ldlm_is_waited(lock))
2498 ldlm_del_waiting_lock(lock);
2500 ldlm_resource_unlink_lock(lock);
2501 ldlm_lock_destroy_nolock(lock);
2503 if (ldlm_is_granted(lock))
2504 ldlm_pool_del(&ns->ns_pool, lock);
2506 /* Make sure we will not be called again for same lock what is possible
2507 * if not to zero out lock->l_granted_mode */
2508 lock->l_granted_mode = LCK_MINMODE;
2509 unlock_res_and_lock(lock);
2513 EXPORT_SYMBOL(ldlm_lock_cancel);
2516 * Set opaque data into the lock that only makes sense to upper layer.
2518 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2520 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2525 if (lock->l_ast_data == NULL)
2526 lock->l_ast_data = data;
2527 if (lock->l_ast_data == data)
2529 LDLM_LOCK_PUT(lock);
2533 EXPORT_SYMBOL(ldlm_lock_set_data);
2535 struct export_cl_data {
2536 const struct lu_env *ecl_env;
2537 struct obd_export *ecl_exp;
2541 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2542 struct ldlm_lock *lock,
2543 struct export_cl_data *ecl)
2545 struct ldlm_resource *res;
2547 res = ldlm_resource_getref(lock->l_resource);
2549 ldlm_lvbo_update(res, lock, NULL, 1);
2550 ldlm_lock_cancel(lock);
2551 if (!exp->exp_obd->obd_stopping)
2552 ldlm_reprocess_all(res, lock);
2553 ldlm_resource_putref(res);
2556 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2557 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2558 exp, ecl->ecl_loop);
2563 * Iterator function for ldlm_export_cancel_locks.
2564 * Cancels passed locks.
2567 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2568 struct hlist_node *hnode, void *data)
2571 struct export_cl_data *ecl = (struct export_cl_data *)data;
2572 struct obd_export *exp = ecl->ecl_exp;
2573 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2575 LDLM_LOCK_GET(lock);
2576 ldlm_cancel_lock_for_export(exp, lock, ecl);
2577 LDLM_LOCK_RELEASE(lock);
2583 * Cancel all blocked locks for given export.
2585 * Typically called on client disconnection/eviction
2587 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2590 struct export_cl_data ecl = {
2596 rc = lu_env_init(&env, LCT_DT_THREAD);
2601 while (!list_empty(&exp->exp_bl_list)) {
2602 struct ldlm_lock *lock;
2604 spin_lock_bh(&exp->exp_bl_list_lock);
2605 if (!list_empty(&exp->exp_bl_list)) {
2606 lock = list_entry(exp->exp_bl_list.next,
2607 struct ldlm_lock, l_exp_list);
2608 LDLM_LOCK_GET(lock);
2609 list_del_init(&lock->l_exp_list);
2613 spin_unlock_bh(&exp->exp_bl_list_lock);
2618 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2619 LDLM_LOCK_RELEASE(lock);
2624 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2625 "left on hash table %d.\n", exp, ecl.ecl_loop,
2626 atomic_read(&exp->exp_lock_hash->hs_count));
2628 return ecl.ecl_loop;
2632 * Cancel all locks for given export.
2634 * Typically called after client disconnection/eviction
2636 int ldlm_export_cancel_locks(struct obd_export *exp)
2638 struct export_cl_data ecl;
2642 rc = lu_env_init(&env, LCT_DT_THREAD);
2649 cfs_hash_for_each_empty(exp->exp_lock_hash,
2650 ldlm_cancel_locks_for_export_cb, &ecl);
2652 CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2653 "left on hash table %d.\n", exp, ecl.ecl_loop,
2654 atomic_read(&exp->exp_lock_hash->hs_count));
2656 if (ecl.ecl_loop > 0 &&
2657 atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2658 exp->exp_obd->obd_stopping)
2659 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2663 return ecl.ecl_loop;
2667 * Downgrade an PW/EX lock to COS | CR mode.
2669 * A lock mode convertion from PW/EX mode to less conflict mode. The
2670 * convertion may fail if lock was canceled before downgrade, but it doesn't
2671 * indicate any problem, because such lock has no reader or writer, and will
2674 * Used by Commit on Sharing (COS) code to force object changes commit in case
2675 * of conflict. Converted lock is considered as new lock and all blocking AST
2676 * things are cleared, so any pending or new blocked lock on that lock will
2677 * cause new call to blocking_ast and force resource object commit.
2679 * Also used by layout_change to replace EX lock to CR lock.
2681 * \param lock A lock to convert
2682 * \param new_mode new lock mode
2684 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2686 #ifdef HAVE_SERVER_SUPPORT
2689 LASSERT(new_mode == LCK_COS || new_mode == LCK_CR);
2691 lock_res_and_lock(lock);
2693 if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2694 unlock_res_and_lock(lock);
2696 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2697 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2701 ldlm_resource_unlink_lock(lock);
2703 * Remove the lock from pool as it will be added again in
2704 * ldlm_grant_lock() called below.
2706 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2708 /* Consider downgraded lock as a new lock and clear all states
2709 * related to a previous blocking AST processing.
2711 ldlm_clear_blocking_data(lock);
2713 lock->l_req_mode = new_mode;
2714 ldlm_grant_lock(lock, NULL);
2715 unlock_res_and_lock(lock);
2717 ldlm_reprocess_all(lock->l_resource, lock);
2722 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2725 * Print lock with lock handle \a lockh description into debug log.
2727 * Used when printing all locks on a resource for debug purposes.
2729 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2731 struct ldlm_lock *lock;
2733 if (!((libcfs_debug | D_ERROR) & level))
2736 lock = ldlm_handle2lock(lockh);
2740 LDLM_DEBUG_LIMIT(level, lock, "###");
2742 LDLM_LOCK_PUT(lock);
2744 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2747 * Print lock information with custom message into debug log.
2750 void _ldlm_lock_debug(struct ldlm_lock *lock,
2751 struct libcfs_debug_msg_data *msgdata,
2752 const char *fmt, ...)
2755 struct obd_export *exp = lock->l_export;
2756 struct ldlm_resource *resource = NULL;
2757 struct va_format vaf;
2758 char *nid = "local";
2760 /* on server-side resource of lock doesn't change */
2761 if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2762 if (lock->l_resource != NULL)
2763 resource = ldlm_resource_getref(lock->l_resource);
2764 } else if (spin_trylock(&lock->l_lock)) {
2765 if (lock->l_resource != NULL)
2766 resource = ldlm_resource_getref(lock->l_resource);
2767 spin_unlock(&lock->l_lock);
2770 va_start(args, fmt);
2774 if (exp && exp->exp_connection) {
2775 nid = obd_export_nid2str(exp);
2776 } else if (exp && exp->exp_obd != NULL) {
2777 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2778 nid = obd_import_nid2str(imp);
2781 if (resource == NULL) {
2782 libcfs_debug_msg(msgdata,
2783 "%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",
2786 lock->l_handle.h_cookie,
2787 refcount_read(&lock->l_handle.h_ref),
2788 lock->l_readers, lock->l_writers,
2789 ldlm_lockname[lock->l_granted_mode],
2790 ldlm_lockname[lock->l_req_mode],
2792 lock->l_remote_handle.cookie,
2793 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2794 lock->l_pid, lock->l_callback_timeout,
2800 switch (resource->lr_type) {
2802 libcfs_debug_msg(msgdata,
2803 "%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",
2805 ldlm_lock_to_ns_name(lock), lock,
2806 lock->l_handle.h_cookie,
2807 refcount_read(&lock->l_handle.h_ref),
2808 lock->l_readers, lock->l_writers,
2809 ldlm_lockname[lock->l_granted_mode],
2810 ldlm_lockname[lock->l_req_mode],
2812 atomic_read(&resource->lr_refcount),
2813 ldlm_typename[resource->lr_type],
2814 lock->l_policy_data.l_extent.start,
2815 lock->l_policy_data.l_extent.end,
2816 lock->l_req_extent.start, lock->l_req_extent.end,
2818 lock->l_remote_handle.cookie,
2819 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2820 lock->l_pid, lock->l_callback_timeout,
2825 libcfs_debug_msg(msgdata,
2826 "%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",
2828 ldlm_lock_to_ns_name(lock), lock,
2829 lock->l_handle.h_cookie,
2830 refcount_read(&lock->l_handle.h_ref),
2831 lock->l_readers, lock->l_writers,
2832 ldlm_lockname[lock->l_granted_mode],
2833 ldlm_lockname[lock->l_req_mode],
2835 atomic_read(&resource->lr_refcount),
2836 ldlm_typename[resource->lr_type],
2837 lock->l_policy_data.l_flock.pid,
2838 lock->l_policy_data.l_flock.start,
2839 lock->l_policy_data.l_flock.end,
2841 lock->l_remote_handle.cookie,
2842 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2843 lock->l_pid, lock->l_callback_timeout);
2847 libcfs_debug_msg(msgdata,
2848 "%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",
2850 ldlm_lock_to_ns_name(lock),
2851 lock, lock->l_handle.h_cookie,
2852 refcount_read(&lock->l_handle.h_ref),
2853 lock->l_readers, lock->l_writers,
2854 ldlm_lockname[lock->l_granted_mode],
2855 ldlm_lockname[lock->l_req_mode],
2857 lock->l_policy_data.l_inodebits.bits,
2858 lock->l_policy_data.l_inodebits.try_bits,
2859 atomic_read(&resource->lr_refcount),
2860 ldlm_typename[resource->lr_type],
2862 lock->l_remote_handle.cookie,
2863 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2864 lock->l_pid, lock->l_callback_timeout,
2869 libcfs_debug_msg(msgdata,
2870 "%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",
2872 ldlm_lock_to_ns_name(lock),
2873 lock, lock->l_handle.h_cookie,
2874 refcount_read(&lock->l_handle.h_ref),
2875 lock->l_readers, lock->l_writers,
2876 ldlm_lockname[lock->l_granted_mode],
2877 ldlm_lockname[lock->l_req_mode],
2879 atomic_read(&resource->lr_refcount),
2880 ldlm_typename[resource->lr_type],
2882 lock->l_remote_handle.cookie,
2883 exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2884 lock->l_pid, lock->l_callback_timeout,
2889 ldlm_resource_putref(resource);
2891 EXPORT_SYMBOL(_ldlm_lock_debug);