1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_lock.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_LDLM
45 # include <libcfs/libcfs.h>
46 # include <linux/lustre_intent.h>
48 # include <liblustre.h>
49 # include <libcfs/kp30.h>
52 #include <obd_class.h>
53 #include "ldlm_internal.h"
55 //struct lustre_lock ldlm_everything_lock;
58 char *ldlm_lockname[] = {
69 char *ldlm_typename[] = {
76 char *ldlm_it2str(int it)
83 case (IT_OPEN | IT_CREAT):
96 CERROR("Unknown intent %d\n", it);
101 extern cfs_mem_cache_t *ldlm_lock_slab;
103 static ldlm_processing_policy ldlm_processing_policy_table[] = {
104 [LDLM_PLAIN] ldlm_process_plain_lock,
105 [LDLM_EXTENT] ldlm_process_extent_lock,
107 [LDLM_FLOCK] ldlm_process_flock_lock,
109 [LDLM_IBITS] ldlm_process_inodebits_lock,
112 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
114 return ldlm_processing_policy_table[res->lr_type];
117 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
123 * REFCOUNTED LOCK OBJECTS
128 * Lock refcounts, during creation:
129 * - one special one for allocation, dec'd only once in destroy
130 * - one for being a lock that's in-use
131 * - one for the addref associated with a new lock
133 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
135 atomic_inc(&lock->l_refc);
139 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
141 LASSERT(size == sizeof(*lock));
142 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
145 void ldlm_lock_put(struct ldlm_lock *lock)
149 LASSERT(lock->l_resource != LP_POISON);
150 LASSERT(atomic_read(&lock->l_refc) > 0);
151 if (atomic_dec_and_test(&lock->l_refc)) {
152 struct ldlm_resource *res;
154 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing it.");
156 res = lock->l_resource;
157 LASSERT(lock->l_destroyed);
158 LASSERT(list_empty(&lock->l_res_link));
159 LASSERT(list_empty(&lock->l_pending_chain));
161 atomic_dec(&res->lr_namespace->ns_locks);
162 ldlm_resource_putref(res);
163 lock->l_resource = NULL;
164 if (lock->l_export) {
165 class_export_put(lock->l_export);
166 lock->l_export = NULL;
169 if (lock->l_lvb_data != NULL)
170 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
172 ldlm_interval_free(ldlm_interval_detach(lock));
173 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
180 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
183 if (!list_empty(&lock->l_lru)) {
184 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
185 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
186 list_del_init(&lock->l_lru);
187 if (lock->l_flags & LDLM_FL_SKIPPED)
188 lock->l_flags &= ~LDLM_FL_SKIPPED;
189 LASSERT(ns->ns_nr_unused > 0);
196 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
198 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
201 spin_lock(&ns->ns_unused_lock);
202 rc = ldlm_lock_remove_from_lru_nolock(lock);
203 spin_unlock(&ns->ns_unused_lock);
208 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
210 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
211 lock->l_last_used = cfs_time_current();
212 LASSERT(list_empty(&lock->l_lru));
213 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
214 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
215 LASSERT(ns->ns_nr_unused >= 0);
219 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
221 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
223 spin_lock(&ns->ns_unused_lock);
224 ldlm_lock_add_to_lru_nolock(lock);
225 spin_unlock(&ns->ns_unused_lock);
229 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
231 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
233 spin_lock(&ns->ns_unused_lock);
234 if (!list_empty(&lock->l_lru)) {
235 ldlm_lock_remove_from_lru_nolock(lock);
236 ldlm_lock_add_to_lru_nolock(lock);
238 spin_unlock(&ns->ns_unused_lock);
242 /* This used to have a 'strict' flag, which recovery would use to mark an
243 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
244 * shall explain why it's gone: with the new hash table scheme, once you call
245 * ldlm_lock_destroy, you can never drop your final references on this lock.
246 * Because it's not in the hash table anymore. -phil */
247 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
251 if (lock->l_readers || lock->l_writers) {
252 LDLM_ERROR(lock, "lock still has references");
253 ldlm_lock_dump(D_ERROR, lock, 0);
257 if (!list_empty(&lock->l_res_link)) {
258 LDLM_ERROR(lock, "lock still on resource");
259 ldlm_lock_dump(D_ERROR, lock, 0);
263 if (lock->l_destroyed) {
264 LASSERT(list_empty(&lock->l_lru));
268 lock->l_destroyed = 1;
270 if (lock->l_export && lock->l_export->exp_lock_hash)
271 lustre_hash_del(lock->l_export->exp_lock_hash,
272 &lock->l_remote_handle, &lock->l_exp_hash);
274 ldlm_lock_remove_from_lru(lock);
275 class_handle_unhash(&lock->l_handle);
278 /* Wake anyone waiting for this lock */
279 /* FIXME: I should probably add yet another flag, instead of using
280 * l_export to only call this on clients */
282 class_export_put(lock->l_export);
283 lock->l_export = NULL;
284 if (lock->l_export && lock->l_completion_ast)
285 lock->l_completion_ast(lock, 0);
291 void ldlm_lock_destroy(struct ldlm_lock *lock)
295 lock_res_and_lock(lock);
296 first = ldlm_lock_destroy_internal(lock);
297 unlock_res_and_lock(lock);
299 /* drop reference from hashtable only for first destroy */
305 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
309 first = ldlm_lock_destroy_internal(lock);
310 /* drop reference from hashtable only for first destroy */
316 /* this is called by portals_handle2object with the handle lock taken */
317 static void lock_handle_addref(void *lock)
319 LDLM_LOCK_GET((struct ldlm_lock *)lock);
323 * usage: pass in a resource on which you have done ldlm_resource_get
324 * pass in a parent lock on which you have done a ldlm_lock_get
325 * after return, ldlm_*_put the resource and parent
326 * returns: lock with refcount 2 - one for current caller and one for remote
328 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
330 struct ldlm_lock *lock;
333 if (resource == NULL)
336 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, CFS_ALLOC_IO, sizeof(*lock));
340 spin_lock_init(&lock->l_lock);
341 lock->l_resource = ldlm_resource_getref(resource);
343 atomic_set(&lock->l_refc, 2);
344 CFS_INIT_LIST_HEAD(&lock->l_res_link);
345 CFS_INIT_LIST_HEAD(&lock->l_lru);
346 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
347 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
348 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
349 cfs_waitq_init(&lock->l_waitq);
350 lock->l_blocking_lock = NULL;
351 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
352 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
353 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
355 atomic_inc(&resource->lr_namespace->ns_locks);
356 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
357 class_handle_hash(&lock->l_handle, lock_handle_addref);
359 CFS_INIT_LIST_HEAD(&lock->l_extents_list);
360 spin_lock_init(&lock->l_extents_list_lock);
361 CFS_INIT_LIST_HEAD(&lock->l_cache_locks_list);
362 lock->l_callback_timeout = 0;
367 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
368 struct ldlm_res_id new_resid)
370 struct ldlm_resource *oldres = lock->l_resource;
371 struct ldlm_resource *newres;
375 LASSERT(ns_is_client(ns));
377 lock_res_and_lock(lock);
378 if (memcmp(&new_resid, &lock->l_resource->lr_name,
379 sizeof(lock->l_resource->lr_name)) == 0) {
381 unlock_res_and_lock(lock);
385 LASSERT(new_resid.name[0] != 0);
387 /* This function assumes that the lock isn't on any lists */
388 LASSERT(list_empty(&lock->l_res_link));
390 type = oldres->lr_type;
391 unlock_res_and_lock(lock);
393 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
397 lock_res_and_lock(lock);
398 LASSERT(memcmp(&new_resid, &lock->l_resource->lr_name,
399 sizeof(lock->l_resource->lr_name)) != 0);
401 lock->l_resource = newres;
403 unlock_res_and_lock(lock);
405 /* ...and the flowers are still standing! */
406 ldlm_resource_putref(oldres);
415 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
417 lockh->cookie = lock->l_handle.h_cookie;
420 /* if flags: atomically get the lock and set the flags.
421 * Return NULL if flag already set
424 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
426 struct ldlm_namespace *ns;
427 struct ldlm_lock *lock = NULL, *retval = NULL;
432 lock = class_handle2object(handle->cookie);
436 LASSERT(lock->l_resource != NULL);
437 ns = lock->l_resource->lr_namespace;
440 lock_res_and_lock(lock);
442 /* It's unlikely but possible that someone marked the lock as
443 * destroyed after we did handle2object on it */
444 if (lock->l_destroyed) {
445 unlock_res_and_lock(lock);
446 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
451 if (flags && (lock->l_flags & flags)) {
452 unlock_res_and_lock(lock);
458 lock->l_flags |= flags;
460 unlock_res_and_lock(lock);
467 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
468 struct lustre_handle *handle)
470 struct ldlm_lock *retval = NULL;
471 retval = __ldlm_handle2lock(handle, 0);
475 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
477 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
478 /* INODEBITS_INTEROP: If the other side does not support
479 * inodebits, reply with a plain lock descriptor.
481 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
482 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
484 /* Make sure all the right bits are set in this lock we
485 are going to pass to client */
486 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
487 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
488 "Inappropriate inode lock bits during "
489 "conversion " LPU64 "\n",
490 lock->l_policy_data.l_inodebits.bits);
491 ldlm_res2desc(lock->l_resource, &desc->l_resource);
492 desc->l_resource.lr_type = LDLM_PLAIN;
493 /* Convert "new" lock mode to something old client can
495 if ((lock->l_req_mode == LCK_CR) ||
496 (lock->l_req_mode == LCK_CW))
497 desc->l_req_mode = LCK_PR;
499 desc->l_req_mode = lock->l_req_mode;
500 if ((lock->l_granted_mode == LCK_CR) ||
501 (lock->l_granted_mode == LCK_CW)) {
502 desc->l_granted_mode = LCK_PR;
504 /* We never grant PW/EX locks to clients */
505 LASSERT((lock->l_granted_mode != LCK_PW) &&
506 (lock->l_granted_mode != LCK_EX));
507 desc->l_granted_mode = lock->l_granted_mode;
510 /* We do not copy policy here, because there is no
511 policy for plain locks */
513 ldlm_res2desc(lock->l_resource, &desc->l_resource);
514 desc->l_req_mode = lock->l_req_mode;
515 desc->l_granted_mode = lock->l_granted_mode;
516 desc->l_policy_data = lock->l_policy_data;
520 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
521 struct list_head *work_list)
523 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
524 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
525 lock->l_flags |= LDLM_FL_AST_SENT;
526 /* If the enqueuing client said so, tell the AST recipient to
527 * discard dirty data, rather than writing back. */
528 if (new->l_flags & LDLM_AST_DISCARD_DATA)
529 lock->l_flags |= LDLM_FL_DISCARD_DATA;
530 LASSERT(list_empty(&lock->l_bl_ast));
531 list_add(&lock->l_bl_ast, work_list);
533 LASSERT(lock->l_blocking_lock == NULL);
534 lock->l_blocking_lock = LDLM_LOCK_GET(new);
538 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
540 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
541 lock->l_flags |= LDLM_FL_CP_REQD;
542 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
543 LASSERT(list_empty(&lock->l_cp_ast));
544 list_add(&lock->l_cp_ast, work_list);
549 /* must be called with lr_lock held */
550 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
551 struct list_head *work_list)
554 check_res_locked(lock->l_resource);
556 ldlm_add_bl_work_item(lock, new, work_list);
558 ldlm_add_cp_work_item(lock, work_list);
562 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
564 struct ldlm_lock *lock;
566 lock = ldlm_handle2lock(lockh);
567 LASSERT(lock != NULL);
568 ldlm_lock_addref_internal(lock, mode);
572 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
574 ldlm_lock_remove_from_lru(lock);
575 if (mode & (LCK_NL | LCK_CR | LCK_PR))
577 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
580 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
583 /* only called for local locks */
584 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
586 lock_res_and_lock(lock);
587 ldlm_lock_addref_internal_nolock(lock, mode);
588 unlock_res_and_lock(lock);
591 /* only called in ldlm_flock_destroy and for local locks.
592 * for LDLM_FLOCK type locks, l_blocking_ast is null, and
593 * ldlm_lock_remove_from_lru() does nothing, it is safe
594 * for ldlm_flock_destroy usage by dropping some code */
595 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
597 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
598 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
599 LASSERT(lock->l_readers > 0);
602 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
603 LASSERT(lock->l_writers > 0);
607 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
610 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
612 struct ldlm_namespace *ns;
615 lock_res_and_lock(lock);
617 ns = lock->l_resource->lr_namespace;
619 ldlm_lock_decref_internal_nolock(lock, mode);
621 if (lock->l_flags & LDLM_FL_LOCAL &&
622 !lock->l_readers && !lock->l_writers) {
623 /* If this is a local lock on a server namespace and this was
624 * the last reference, cancel the lock. */
625 CDEBUG(D_INFO, "forcing cancel of local lock\n");
626 lock->l_flags |= LDLM_FL_CBPENDING;
629 if (!lock->l_readers && !lock->l_writers &&
630 (lock->l_flags & LDLM_FL_CBPENDING)) {
631 /* If we received a blocked AST and this was the last reference,
632 * run the callback. */
633 if (ns_is_server(ns) && lock->l_export)
634 CERROR("FL_CBPENDING set on non-local lock--just a "
637 LDLM_DEBUG(lock, "final decref done on cbpending lock");
639 LDLM_LOCK_GET(lock); /* dropped by bl thread */
640 ldlm_lock_remove_from_lru(lock);
641 unlock_res_and_lock(lock);
643 if (lock->l_flags & LDLM_FL_FAIL_LOC)
644 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
646 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
647 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
648 ldlm_handle_bl_callback(ns, NULL, lock);
649 } else if (ns_is_client(ns) &&
650 !lock->l_readers && !lock->l_writers &&
651 !(lock->l_flags & LDLM_FL_NO_LRU) &&
652 !(lock->l_flags & LDLM_FL_BL_AST)) {
653 /* If this is a client-side namespace and this was the last
654 * reference, put it on the LRU. */
655 ldlm_lock_add_to_lru(lock);
656 unlock_res_and_lock(lock);
658 if (lock->l_flags & LDLM_FL_FAIL_LOC)
659 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
661 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
662 * are not supported by the server, otherwise, it is done on
664 if (!exp_connect_cancelset(lock->l_conn_export) &&
665 !ns_connect_lru_resize(ns))
666 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
668 unlock_res_and_lock(lock);
674 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
676 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
677 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
678 ldlm_lock_decref_internal(lock, mode);
682 /* This will drop a lock reference and mark it for destruction, but will not
683 * necessarily cancel the lock before returning. */
684 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
686 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
689 LASSERT(lock != NULL);
691 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
692 lock_res_and_lock(lock);
693 lock->l_flags |= LDLM_FL_CBPENDING;
694 unlock_res_and_lock(lock);
695 ldlm_lock_decref_internal(lock, mode);
699 struct sl_insert_point {
700 struct list_head *res_link;
701 struct list_head *mode_link;
702 struct list_head *policy_link;
706 * search_granted_lock
709 * Finds a position to insert the new lock.
711 * queue [input]: the granted list where search acts on;
712 * req [input]: the lock whose position to be located;
713 * prev [output]: positions within 3 lists to insert @req to
717 * - ldlm_grant_lock_with_skiplist
719 static void search_granted_lock(struct list_head *queue,
720 struct ldlm_lock *req,
721 struct sl_insert_point *prev)
723 struct list_head *tmp;
724 struct ldlm_lock *lock, *mode_end, *policy_end;
727 list_for_each(tmp, queue) {
728 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
730 mode_end = list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
733 if (lock->l_req_mode != req->l_req_mode) {
734 /* jump to last lock of mode group */
735 tmp = &mode_end->l_res_link;
739 /* suitable mode group is found */
740 if (lock->l_resource->lr_type == LDLM_PLAIN) {
741 /* insert point is last lock of the mode group */
742 prev->res_link = &mode_end->l_res_link;
743 prev->mode_link = &mode_end->l_sl_mode;
744 prev->policy_link = &req->l_sl_policy;
747 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
749 policy_end = list_entry(lock->l_sl_policy.prev,
753 if (lock->l_policy_data.l_inodebits.bits ==
754 req->l_policy_data.l_inodebits.bits) {
755 /* insert point is last lock of
756 * the policy group */
758 &policy_end->l_res_link;
760 &policy_end->l_sl_mode;
762 &policy_end->l_sl_policy;
767 if (policy_end == mode_end)
768 /* done with mode group */
771 /* jump to next policy group within the mode group */
772 tmp = policy_end->l_res_link.next;
773 lock = list_entry(tmp, struct ldlm_lock,
775 } /* loop over policy groups within the mode group */
777 /* insert point is last lock of the mode group,
778 * new policy group is started */
779 prev->res_link = &mode_end->l_res_link;
780 prev->mode_link = &mode_end->l_sl_mode;
781 prev->policy_link = &req->l_sl_policy;
785 LDLM_ERROR(lock, "is not LDLM_PLAIN or LDLM_IBITS lock");
790 /* insert point is last lock on the queue,
791 * new mode group and new policy group are started */
792 prev->res_link = queue->prev;
793 prev->mode_link = &req->l_sl_mode;
794 prev->policy_link = &req->l_sl_policy;
799 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
800 struct sl_insert_point *prev)
802 struct ldlm_resource *res = lock->l_resource;
805 check_res_locked(res);
807 ldlm_resource_dump(D_INFO, res);
808 CDEBUG(D_OTHER, "About to add this lock:\n");
809 ldlm_lock_dump(D_OTHER, lock, 0);
811 if (lock->l_destroyed) {
812 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
816 LASSERT(list_empty(&lock->l_res_link));
817 LASSERT(list_empty(&lock->l_sl_mode));
818 LASSERT(list_empty(&lock->l_sl_policy));
820 list_add(&lock->l_res_link, prev->res_link);
821 list_add(&lock->l_sl_mode, prev->mode_link);
822 list_add(&lock->l_sl_policy, prev->policy_link);
827 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
829 struct sl_insert_point prev;
832 LASSERT(lock->l_req_mode == lock->l_granted_mode);
834 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
835 ldlm_granted_list_add_lock(lock, &prev);
840 * - ldlm_lock_enqueue
841 * - ldlm_reprocess_queue
842 * - ldlm_lock_convert
844 * must be called with lr_lock held
846 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
848 struct ldlm_resource *res = lock->l_resource;
851 check_res_locked(res);
853 lock->l_granted_mode = lock->l_req_mode;
854 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
855 ldlm_grant_lock_with_skiplist(lock);
856 else if (res->lr_type == LDLM_EXTENT)
857 ldlm_extent_add_lock(res, lock);
859 ldlm_resource_add_lock(res, &res->lr_granted, lock);
861 if (lock->l_granted_mode < res->lr_most_restr)
862 res->lr_most_restr = lock->l_granted_mode;
864 if (work_list && lock->l_completion_ast != NULL)
865 ldlm_add_ast_work_item(lock, NULL, work_list);
867 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
871 /* returns a referenced lock or NULL. See the flag descriptions below, in the
872 * comment above ldlm_lock_match */
873 static struct ldlm_lock *search_queue(struct list_head *queue,
875 ldlm_policy_data_t *policy,
876 struct ldlm_lock *old_lock, int flags)
878 struct ldlm_lock *lock;
879 struct list_head *tmp;
881 list_for_each(tmp, queue) {
884 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
886 if (lock == old_lock)
889 /* llite sometimes wants to match locks that will be
890 * canceled when their users drop, but we allow it to match
891 * if it passes in CBPENDING and the lock still has users.
892 * this is generally only going to be used by children
893 * whose parents already hold a lock so forward progress
894 * can still happen. */
895 if (lock->l_flags & LDLM_FL_CBPENDING &&
896 !(flags & LDLM_FL_CBPENDING))
898 if (lock->l_flags & LDLM_FL_CBPENDING &&
899 lock->l_readers == 0 && lock->l_writers == 0)
902 if (!(lock->l_req_mode & *mode))
905 match = lock->l_req_mode;
906 if (lock->l_resource->lr_type == LDLM_EXTENT &&
907 (lock->l_policy_data.l_extent.start >
908 policy->l_extent.start ||
909 lock->l_policy_data.l_extent.end < policy->l_extent.end))
912 if (unlikely(match == LCK_GROUP) &&
913 lock->l_resource->lr_type == LDLM_EXTENT &&
914 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
917 /* We match if we have existing lock with same or wider set
919 if (lock->l_resource->lr_type == LDLM_IBITS &&
920 ((lock->l_policy_data.l_inodebits.bits &
921 policy->l_inodebits.bits) !=
922 policy->l_inodebits.bits))
925 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
928 if ((flags & LDLM_FL_LOCAL_ONLY) &&
929 !(lock->l_flags & LDLM_FL_LOCAL))
932 if (flags & LDLM_FL_TEST_LOCK) {
934 ldlm_lock_touch_in_lru(lock);
936 ldlm_lock_addref_internal_nolock(lock, match);
945 void ldlm_lock_allow_match(struct ldlm_lock *lock)
947 lock_res_and_lock(lock);
948 lock->l_flags |= LDLM_FL_LVB_READY;
949 cfs_waitq_signal(&lock->l_waitq);
950 unlock_res_and_lock(lock);
953 int ldlm_lock_fast_match(struct ldlm_lock *lock, int rw,
954 obd_off start, obd_off end,
955 struct lustre_handle *lockh)
957 LASSERT(rw == OBD_BRW_READ || rw == OBD_BRW_WRITE);
958 LASSERT(lockh != NULL);
963 lock_res_and_lock(lock);
964 /* check if granted mode is compatible */
965 if (rw == OBD_BRW_WRITE &&
966 !(lock->l_granted_mode & (LCK_PW|LCK_GROUP)))
969 /* does the lock cover the region we would like to access? */
970 if ((lock->l_policy_data.l_extent.start > start) ||
971 (lock->l_policy_data.l_extent.end < end))
974 /* if we received a blocking callback and the lock is no longer
975 * referenced, don't use it */
976 if ((lock->l_flags & LDLM_FL_CBPENDING) &&
977 !lock->l_writers && !lock->l_readers)
980 ldlm_lock_addref_internal_nolock(lock,
981 rw == OBD_BRW_WRITE ? LCK_PW : LCK_PR);
982 unlock_res_and_lock(lock);
983 ldlm_lock2handle(lock, lockh);
984 return 1; /* avoid using rc for stack relief */
987 unlock_res_and_lock(lock);
991 /* Can be called in two ways:
993 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
994 * for a duplicate of.
996 * Otherwise, all of the fields must be filled in, to match against.
998 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
999 * server (ie, connh is NULL)
1000 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1001 * list will be considered
1002 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1003 * to be canceled can still be matched as long as they still have reader
1004 * or writer refernces
1005 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1006 * just tell us if we would have matched.
1008 * Returns 1 if it finds an already-existing lock that is compatible; in this
1009 * case, lockh is filled in with a addref()ed lock
1011 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1012 struct ldlm_res_id *res_id, ldlm_type_t type,
1013 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1014 struct lustre_handle *lockh)
1016 struct ldlm_resource *res;
1017 struct ldlm_lock *lock, *old_lock = NULL;
1022 old_lock = ldlm_handle2lock(lockh);
1025 ns = old_lock->l_resource->lr_namespace;
1026 res_id = &old_lock->l_resource->lr_name;
1027 type = old_lock->l_resource->lr_type;
1028 mode = old_lock->l_req_mode;
1031 res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
1033 LASSERT(old_lock == NULL);
1039 lock = search_queue(&res->lr_granted, &mode, policy, old_lock, flags);
1042 if (flags & LDLM_FL_BLOCK_GRANTED)
1044 lock = search_queue(&res->lr_converting, &mode, policy, old_lock, flags);
1047 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock, flags);
1054 ldlm_resource_putref(res);
1057 ldlm_lock2handle(lock, lockh);
1058 if ((flags & LDLM_FL_LVB_READY) && (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1059 struct l_wait_info lwi;
1060 if (lock->l_completion_ast) {
1061 int err = lock->l_completion_ast(lock,
1062 LDLM_FL_WAIT_NOREPROC,
1065 if (flags & LDLM_FL_TEST_LOCK)
1066 LDLM_LOCK_PUT(lock);
1068 ldlm_lock_decref_internal(lock, mode);
1074 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), NULL,
1075 LWI_ON_SIGNAL_NOOP, NULL);
1077 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1078 l_wait_event(lock->l_waitq,
1079 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1084 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1085 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1086 res_id->name[2] : policy->l_extent.start,
1087 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1088 res_id->name[3] : policy->l_extent.end);
1089 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1090 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1091 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1092 type, mode, res_id->name[0], res_id->name[1],
1093 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1094 res_id->name[2] :policy->l_extent.start,
1095 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1096 res_id->name[3] : policy->l_extent.end);
1099 LDLM_LOCK_PUT(old_lock);
1100 if (flags & LDLM_FL_TEST_LOCK && rc)
1101 LDLM_LOCK_PUT(lock);
1103 return rc ? mode : 0;
1106 /* Returns a referenced lock */
1107 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1108 struct ldlm_res_id res_id, ldlm_type_t type,
1110 ldlm_blocking_callback blocking,
1111 ldlm_completion_callback completion,
1112 ldlm_glimpse_callback glimpse,
1113 void *data, __u32 lvb_len)
1115 struct ldlm_lock *lock;
1116 struct ldlm_resource *res;
1119 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1123 lock = ldlm_lock_new(res);
1124 ldlm_resource_putref(res);
1129 lock->l_req_mode = mode;
1130 lock->l_ast_data = data;
1131 lock->l_blocking_ast = blocking;
1132 lock->l_completion_ast = completion;
1133 lock->l_glimpse_ast = glimpse;
1134 lock->l_pid = cfs_curproc_pid();
1136 lock->l_tree_node = NULL;
1137 /* if this is the extent lock, allocate the interval tree node */
1138 if (type == LDLM_EXTENT) {
1139 if (ldlm_interval_alloc(lock) == NULL)
1144 lock->l_lvb_len = lvb_len;
1145 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1146 if (lock->l_lvb_data == NULL)
1153 if (lock->l_lvb_data)
1154 OBD_FREE(lock->l_lvb_data, lvb_len);
1155 ldlm_interval_free(ldlm_interval_detach(lock));
1156 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1160 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1161 struct ldlm_lock **lockp,
1162 void *cookie, int *flags)
1164 struct ldlm_lock *lock = *lockp;
1165 struct ldlm_resource *res = lock->l_resource;
1166 int local = ns_is_client(res->lr_namespace);
1167 ldlm_processing_policy policy;
1168 ldlm_error_t rc = ELDLM_OK;
1169 struct ldlm_interval *node = NULL;
1172 lock->l_last_activity = cfs_time_current_sec();
1173 /* policies are not executed on the client or during replay */
1174 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1175 && !local && ns->ns_policy) {
1176 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1178 if (rc == ELDLM_LOCK_REPLACED) {
1179 /* The lock that was returned has already been granted,
1180 * and placed into lockp. If it's not the same as the
1181 * one we passed in, then destroy the old one and our
1182 * work here is done. */
1183 if (lock != *lockp) {
1184 ldlm_lock_destroy(lock);
1185 LDLM_LOCK_PUT(lock);
1187 *flags |= LDLM_FL_LOCK_CHANGED;
1189 } else if (rc != ELDLM_OK ||
1190 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1191 ldlm_lock_destroy(lock);
1196 /* For a replaying lock, it might be already in granted list. So
1197 * unlinking the lock will cause the interval node to be freed, we
1198 * have to allocate the interval node early otherwise we can't regrant
1199 * this lock in the future. - jay */
1200 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1201 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO,
1203 if(res->lr_type == LDLM_EXTENT)
1204 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_LOCAL, 30);
1206 lock_res_and_lock(lock);
1207 if (lock->l_destroyed)
1208 GOTO(out, rc = -EAGAIN);
1210 if (local && lock->l_req_mode == lock->l_granted_mode) {
1211 /* The server returned a blocked lock, but it was granted
1212 * before we got a chance to actually enqueue it. We don't
1213 * need to do anything else. */
1214 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1215 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1216 GOTO(out, ELDLM_OK);
1219 ldlm_resource_unlink_lock(lock);
1220 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1221 LASSERT(!local && (*flags & LDLM_FL_REPLAY));
1223 ldlm_lock_destroy_nolock(lock);
1224 GOTO(out, rc = -ENOMEM);
1227 CFS_INIT_LIST_HEAD(&node->li_group);
1228 ldlm_interval_attach(node, lock);
1232 /* Some flags from the enqueue want to make it into the AST, via the
1233 * lock's l_flags. */
1234 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1236 /* This distinction between local lock trees is very important; a client
1237 * namespace only has information about locks taken by that client, and
1238 * thus doesn't have enough information to decide for itself if it can
1239 * be granted (below). In this case, we do exactly what the server
1240 * tells us to do, as dictated by the 'flags'.
1242 * We do exactly the same thing during recovery, when the server is
1243 * more or less trusting the clients not to lie.
1245 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1246 * granted/converting queues. */
1248 if (*flags & LDLM_FL_BLOCK_CONV)
1249 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1250 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1251 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1253 ldlm_grant_lock(lock, NULL);
1254 GOTO(out, ELDLM_OK);
1255 } else if (*flags & LDLM_FL_REPLAY) {
1256 if (*flags & LDLM_FL_BLOCK_CONV) {
1257 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1258 GOTO(out, ELDLM_OK);
1259 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1260 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1261 GOTO(out, ELDLM_OK);
1262 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1263 ldlm_grant_lock(lock, NULL);
1264 GOTO(out, ELDLM_OK);
1266 /* If no flags, fall through to normal enqueue path. */
1269 policy = ldlm_processing_policy_table[res->lr_type];
1270 policy(lock, flags, 1, &rc, NULL);
1273 unlock_res_and_lock(lock);
1275 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1279 /* Must be called with namespace taken: queue is waiting or converting. */
1280 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1281 struct list_head *work_list)
1283 struct list_head *tmp, *pos;
1284 ldlm_processing_policy policy;
1286 int rc = LDLM_ITER_CONTINUE;
1290 check_res_locked(res);
1292 policy = ldlm_processing_policy_table[res->lr_type];
1295 list_for_each_safe(tmp, pos, queue) {
1296 struct ldlm_lock *pending;
1297 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1299 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1302 rc = policy(pending, &flags, 0, &err, work_list);
1303 if (rc != LDLM_ITER_CONTINUE)
1310 /* Helper function for pair ldlm_run_{bl,cp}_ast_work().
1312 * Send an existing rpc set specified by @arg->set and then
1313 * destroy it. Create new one if @do_create flag is set. */
1315 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1319 rc = ptlrpc_set_wait(arg->set);
1320 if (arg->type == LDLM_BL_CALLBACK)
1321 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1322 ptlrpc_set_destroy(arg->set);
1325 arg->set = ptlrpc_prep_set();
1328 int ldlm_run_bl_ast_work(struct list_head *rpc_list)
1330 struct ldlm_cb_set_arg arg;
1331 struct list_head *tmp, *pos;
1332 struct ldlm_lock_desc d;
1337 if (list_empty(rpc_list))
1340 arg.set = ptlrpc_prep_set();
1341 if (NULL == arg.set)
1343 atomic_set(&arg.restart, 0);
1344 arg.type = LDLM_BL_CALLBACK;
1347 list_for_each_safe(tmp, pos, rpc_list) {
1348 struct ldlm_lock *lock =
1349 list_entry(tmp, struct ldlm_lock, l_bl_ast);
1351 /* nobody should touch l_bl_ast */
1352 lock_res_and_lock(lock);
1353 list_del_init(&lock->l_bl_ast);
1355 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1356 LASSERT(lock->l_bl_ast_run == 0);
1357 LASSERT(lock->l_blocking_lock);
1358 lock->l_bl_ast_run++;
1359 unlock_res_and_lock(lock);
1361 ldlm_lock2desc(lock->l_blocking_lock, &d);
1363 LDLM_LOCK_PUT(lock->l_blocking_lock);
1364 lock->l_blocking_lock = NULL;
1365 rc = lock->l_blocking_ast(lock, &d, (void *)&arg,
1367 LDLM_LOCK_PUT(lock);
1370 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1371 * and create a new set for requests that remained in
1373 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1374 ldlm_send_and_maybe_create_set(&arg, 1);
1380 ldlm_send_and_maybe_create_set(&arg, 0);
1382 /* In case when number of ASTs is multiply of
1383 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1384 * @arg.set must be destroyed here, otherwise we get
1385 * write memory leaking. */
1386 ptlrpc_set_destroy(arg.set);
1388 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1391 int ldlm_run_cp_ast_work(struct list_head *rpc_list)
1393 struct ldlm_cb_set_arg arg;
1394 struct list_head *tmp, *pos;
1399 if (list_empty(rpc_list))
1402 arg.set = ptlrpc_prep_set();
1403 if (NULL == arg.set)
1405 atomic_set(&arg.restart, 0);
1406 arg.type = LDLM_CP_CALLBACK;
1408 /* It's possible to receive a completion AST before we've set
1409 * the l_completion_ast pointer: either because the AST arrived
1410 * before the reply, or simply because there's a small race
1411 * window between receiving the reply and finishing the local
1412 * enqueue. (bug 842)
1414 * This can't happen with the blocking_ast, however, because we
1415 * will never call the local blocking_ast until we drop our
1416 * reader/writer reference, which we won't do until we get the
1417 * reply and finish enqueueing. */
1420 list_for_each_safe(tmp, pos, rpc_list) {
1421 struct ldlm_lock *lock =
1422 list_entry(tmp, struct ldlm_lock, l_cp_ast);
1423 ldlm_completion_callback completion_callback;
1425 /* nobody should touch l_cp_ast */
1426 lock_res_and_lock(lock);
1427 list_del_init(&lock->l_cp_ast);
1428 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1429 /* save l_completion_ast since it can be changed by
1430 * mds_intent_policy(), see bug 14225 */
1431 completion_callback = lock->l_completion_ast;
1432 lock->l_flags &= ~LDLM_FL_CP_REQD;
1433 unlock_res_and_lock(lock);
1435 if (completion_callback != NULL) {
1436 rc = completion_callback(lock, 0, (void *)&arg);
1439 LDLM_LOCK_PUT(lock);
1441 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1442 * and create a new set for requests that remained in
1444 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1445 ldlm_send_and_maybe_create_set(&arg, 1);
1451 ldlm_send_and_maybe_create_set(&arg, 0);
1453 /* In case when number of ASTs is multiply of
1454 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1455 * @arg.set must be destroyed here, otherwise we get
1456 * write memory leaking. */
1457 ptlrpc_set_destroy(arg.set);
1459 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1462 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1464 ldlm_reprocess_all(res);
1465 return LDLM_ITER_CONTINUE;
1468 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1470 struct list_head *tmp;
1474 spin_lock(&ns->ns_hash_lock);
1475 for (i = 0; i < RES_HASH_SIZE; i++) {
1476 tmp = ns->ns_hash[i].next;
1477 while (tmp != &(ns->ns_hash[i])) {
1478 struct ldlm_resource *res =
1479 list_entry(tmp, struct ldlm_resource, lr_hash);
1481 ldlm_resource_getref(res);
1482 spin_unlock(&ns->ns_hash_lock);
1484 rc = reprocess_one_queue(res, NULL);
1486 spin_lock(&ns->ns_hash_lock);
1488 ldlm_resource_putref_locked(res);
1490 if (rc == LDLM_ITER_STOP)
1495 spin_unlock(&ns->ns_hash_lock);
1499 void ldlm_reprocess_all(struct ldlm_resource *res)
1501 CFS_LIST_HEAD(rpc_list);
1505 /* Local lock trees don't get reprocessed. */
1506 if (ns_is_client(res->lr_namespace)) {
1513 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1514 if (rc == LDLM_ITER_CONTINUE)
1515 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1518 rc = ldlm_run_cp_ast_work(&rpc_list);
1519 if (rc == -ERESTART) {
1520 LASSERT(list_empty(&rpc_list));
1526 void ldlm_cancel_callback(struct ldlm_lock *lock)
1528 check_res_locked(lock->l_resource);
1529 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1530 lock->l_flags |= LDLM_FL_CANCEL;
1531 if (lock->l_blocking_ast) {
1532 // l_check_no_ns_lock(ns);
1533 unlock_res_and_lock(lock);
1534 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1536 lock_res_and_lock(lock);
1538 LDLM_DEBUG(lock, "no blocking ast");
1541 lock->l_flags |= LDLM_FL_BL_DONE;
1544 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1546 if (req->l_resource->lr_type != LDLM_PLAIN &&
1547 req->l_resource->lr_type != LDLM_IBITS)
1550 list_del_init(&req->l_sl_policy);
1551 list_del_init(&req->l_sl_mode);
1554 void ldlm_lock_cancel(struct ldlm_lock *lock)
1556 struct ldlm_resource *res;
1557 struct ldlm_namespace *ns;
1560 lock_res_and_lock(lock);
1562 res = lock->l_resource;
1563 ns = res->lr_namespace;
1565 /* Please do not, no matter how tempting, remove this LBUG without
1566 * talking to me first. -phik */
1567 if (lock->l_readers || lock->l_writers) {
1568 LDLM_ERROR(lock, "lock still has references");
1572 ldlm_del_waiting_lock(lock);
1573 /* Releases res lock */
1574 ldlm_cancel_callback(lock);
1576 /* Yes, second time, just in case it was added again while we were
1577 running with no res lock in ldlm_cancel_callback */
1578 ldlm_del_waiting_lock(lock);
1580 ldlm_resource_unlink_lock(lock);
1581 ldlm_lock_destroy_nolock(lock);
1583 if (lock->l_granted_mode == lock->l_req_mode)
1584 ldlm_pool_del(&ns->ns_pool, lock);
1586 /* Make sure we will not be called again for same lock what is possible
1587 * if not to zero out lock->l_granted_mode */
1588 lock->l_granted_mode = 0;
1589 unlock_res_and_lock(lock);
1594 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1596 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1602 lock->l_ast_data = data;
1603 LDLM_LOCK_PUT(lock);
1607 void ldlm_cancel_locks_for_export_cb(void *obj, void *data)
1609 struct obd_export *exp = data;
1610 struct ldlm_lock *lock = obj;
1611 struct ldlm_resource *res;
1613 res = ldlm_resource_getref(lock->l_resource);
1614 LDLM_LOCK_GET(lock);
1616 LDLM_DEBUG(lock, "export %p", exp);
1617 ldlm_res_lvbo_update(res, NULL, 0, 1);
1619 ldlm_lock_cancel(lock);
1620 ldlm_reprocess_all(res);
1622 ldlm_resource_putref(res);
1623 LDLM_LOCK_PUT(lock);
1626 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1628 lustre_hash_for_each_empty(exp->exp_lock_hash,
1629 ldlm_cancel_locks_for_export_cb, exp);
1632 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1635 CFS_LIST_HEAD(rpc_list);
1636 struct ldlm_resource *res;
1637 struct ldlm_namespace *ns;
1640 struct sl_insert_point prev;
1642 struct ldlm_interval *node;
1645 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1646 *flags |= LDLM_FL_BLOCK_GRANTED;
1647 RETURN(lock->l_resource);
1650 /* I can't check the type of lock here because the bitlock of lock
1651 * is not held here, so do the allocation blindly. -jay */
1652 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO, sizeof(*node));
1653 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1656 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1657 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1659 lock_res_and_lock(lock);
1660 if (unlikely(lock->l_destroyed != 0)) {
1661 unlock_res_and_lock(lock);
1665 res = lock->l_resource;
1666 ns = res->lr_namespace;
1668 old_mode = lock->l_req_mode;
1669 lock->l_req_mode = new_mode;
1670 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1671 /* remember the lock position where the lock might be
1672 * added back to the granted list later and also
1673 * remember the join mode for skiplist fixing. */
1674 prev.res_link = lock->l_res_link.prev;
1675 prev.mode_link = lock->l_sl_mode.prev;
1676 prev.policy_link = lock->l_sl_policy.prev;
1677 ldlm_resource_unlink_lock(lock);
1679 ldlm_resource_unlink_lock(lock);
1680 if (res->lr_type == LDLM_EXTENT) {
1681 /* FIXME: ugly code, I have to attach the lock to a
1682 * interval node again since perhaps it will be granted
1684 CFS_INIT_LIST_HEAD(&node->li_group);
1685 ldlm_interval_attach(node, lock);
1690 /* If this is a local resource, put it on the appropriate list. */
1691 if (ns_is_client(res->lr_namespace)) {
1692 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1693 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1695 /* This should never happen, because of the way the
1696 * server handles conversions. */
1697 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1701 ldlm_grant_lock(lock, &rpc_list);
1703 /* FIXME: completion handling not with ns_lock held ! */
1704 if (lock->l_completion_ast)
1705 lock->l_completion_ast(lock, 0, NULL);
1709 ldlm_processing_policy policy;
1710 policy = ldlm_processing_policy_table[res->lr_type];
1711 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1712 if (rc == LDLM_ITER_STOP) {
1713 lock->l_req_mode = old_mode;
1714 if (res->lr_type == LDLM_EXTENT)
1715 ldlm_extent_add_lock(res, lock);
1717 ldlm_granted_list_add_lock(lock, &prev);
1721 *flags |= LDLM_FL_BLOCK_GRANTED;
1725 unlock_res_and_lock(lock);
1728 ldlm_run_cp_ast_work(&rpc_list);
1730 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1734 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1736 struct obd_device *obd = NULL;
1738 if (!((libcfs_debug | D_ERROR) & level))
1742 CDEBUG(level, " NULL LDLM lock\n");
1746 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1747 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1749 if (lock->l_conn_export != NULL)
1750 obd = lock->l_conn_export->exp_obd;
1751 if (lock->l_export && lock->l_export->exp_connection) {
1752 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1753 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1754 lock->l_remote_handle.cookie);
1755 } else if (obd == NULL) {
1756 CDEBUG(level, " Node: local\n");
1758 struct obd_import *imp = obd->u.cli.cl_import;
1759 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1760 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1761 lock->l_remote_handle.cookie);
1763 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource,
1764 lock->l_resource->lr_name.name[0],
1765 lock->l_resource->lr_name.name[1]);
1766 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1767 "write: %d flags: "LPX64"\n", ldlm_lockname[lock->l_req_mode],
1768 ldlm_lockname[lock->l_granted_mode],
1769 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1771 if (lock->l_resource->lr_type == LDLM_EXTENT)
1772 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1773 " (req "LPU64"-"LPU64")\n",
1774 lock->l_policy_data.l_extent.start,
1775 lock->l_policy_data.l_extent.end,
1776 lock->l_req_extent.start, lock->l_req_extent.end);
1777 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1778 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1779 lock->l_policy_data.l_flock.pid,
1780 lock->l_policy_data.l_flock.start,
1781 lock->l_policy_data.l_flock.end);
1782 else if (lock->l_resource->lr_type == LDLM_IBITS)
1783 CDEBUG(level, " Bits: "LPX64"\n",
1784 lock->l_policy_data.l_inodebits.bits);
1787 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1789 struct ldlm_lock *lock;
1791 if (!((libcfs_debug | D_ERROR) & level))
1794 lock = ldlm_handle2lock(lockh);
1798 ldlm_lock_dump(D_OTHER, lock, 0);
1800 LDLM_LOCK_PUT(lock);
1803 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1804 struct libcfs_debug_msg_data *data, const char *fmt,
1808 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1810 va_start(args, fmt);
1811 if (lock->l_resource == NULL) {
1812 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1813 data->msg_fn, data->msg_line, fmt, args,
1814 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1815 "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" remote: "
1816 LPX64" expref: %d pid: %u timeout: %lu\n", lock,
1817 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1818 lock->l_readers, lock->l_writers,
1819 ldlm_lockname[lock->l_granted_mode],
1820 ldlm_lockname[lock->l_req_mode],
1821 lock->l_flags, lock->l_remote_handle.cookie,
1823 atomic_read(&lock->l_export->exp_refcount) : -99,
1824 lock->l_pid, lock->l_callback_timeout);
1829 switch (lock->l_resource->lr_type) {
1831 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1832 data->msg_fn, data->msg_line, fmt, args,
1833 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1834 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1835 "] (req "LPU64"->"LPU64") flags: "LPX64" remote: "LPX64
1836 " expref: %d pid: %u timeout %lu\n",
1837 lock->l_resource->lr_namespace->ns_name, lock,
1838 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1839 lock->l_readers, lock->l_writers,
1840 ldlm_lockname[lock->l_granted_mode],
1841 ldlm_lockname[lock->l_req_mode],
1842 lock->l_resource->lr_name.name[0],
1843 lock->l_resource->lr_name.name[1],
1844 atomic_read(&lock->l_resource->lr_refcount),
1845 ldlm_typename[lock->l_resource->lr_type],
1846 lock->l_policy_data.l_extent.start,
1847 lock->l_policy_data.l_extent.end,
1848 lock->l_req_extent.start, lock->l_req_extent.end,
1849 lock->l_flags, lock->l_remote_handle.cookie,
1851 atomic_read(&lock->l_export->exp_refcount) : -99,
1852 lock->l_pid, lock->l_callback_timeout);
1855 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1856 data->msg_fn, data->msg_line, fmt, args,
1857 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1858 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1859 "["LPU64"->"LPU64"] flags: "LPX64" remote: "LPX64
1860 " expref: %d pid: %u timeout: %lu\n",
1861 lock->l_resource->lr_namespace->ns_name, lock,
1862 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1863 lock->l_readers, lock->l_writers,
1864 ldlm_lockname[lock->l_granted_mode],
1865 ldlm_lockname[lock->l_req_mode],
1866 lock->l_resource->lr_name.name[0],
1867 lock->l_resource->lr_name.name[1],
1868 atomic_read(&lock->l_resource->lr_refcount),
1869 ldlm_typename[lock->l_resource->lr_type],
1870 lock->l_policy_data.l_flock.pid,
1871 lock->l_policy_data.l_flock.start,
1872 lock->l_policy_data.l_flock.end,
1873 lock->l_flags, lock->l_remote_handle.cookie,
1875 atomic_read(&lock->l_export->exp_refcount) : -99,
1876 lock->l_pid, lock->l_callback_timeout);
1879 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1880 data->msg_fn, data->msg_line, fmt, args,
1881 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1882 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1883 "flags: "LPX64" remote: "LPX64" expref: %d "
1884 "pid: %u timeout: %lu\n",
1885 lock->l_resource->lr_namespace->ns_name,
1886 lock, lock->l_handle.h_cookie,
1887 atomic_read (&lock->l_refc),
1888 lock->l_readers, lock->l_writers,
1889 ldlm_lockname[lock->l_granted_mode],
1890 ldlm_lockname[lock->l_req_mode],
1891 lock->l_resource->lr_name.name[0],
1892 lock->l_resource->lr_name.name[1],
1893 lock->l_policy_data.l_inodebits.bits,
1894 atomic_read(&lock->l_resource->lr_refcount),
1895 ldlm_typename[lock->l_resource->lr_type],
1896 lock->l_flags, lock->l_remote_handle.cookie,
1898 atomic_read(&lock->l_export->exp_refcount) : -99,
1899 lock->l_pid, lock->l_callback_timeout);
1902 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1903 data->msg_fn, data->msg_line, fmt, args,
1904 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1905 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" "
1906 "remote: "LPX64" expref: %d pid: %u timeout %lu\n",
1907 lock->l_resource->lr_namespace->ns_name,
1908 lock, lock->l_handle.h_cookie,
1909 atomic_read (&lock->l_refc),
1910 lock->l_readers, lock->l_writers,
1911 ldlm_lockname[lock->l_granted_mode],
1912 ldlm_lockname[lock->l_req_mode],
1913 lock->l_resource->lr_name.name[0],
1914 lock->l_resource->lr_name.name[1],
1915 atomic_read(&lock->l_resource->lr_refcount),
1916 ldlm_typename[lock->l_resource->lr_type],
1917 lock->l_flags, lock->l_remote_handle.cookie,
1919 atomic_read(&lock->l_export->exp_refcount) : -99,
1920 lock->l_pid, lock->l_callback_timeout);
1925 EXPORT_SYMBOL(_ldlm_lock_debug);