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 2008 Sun Microsystems, Inc. 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 # ifndef HAVE_VFS_INTENT_PATCHES
47 # include <linux/lustre_intent.h>
50 # include <liblustre.h>
53 #include <obd_class.h>
54 #include "ldlm_internal.h"
57 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;
155 "final lock_put on destroyed lock, freeing it.");
157 res = lock->l_resource;
158 LASSERT(lock->l_destroyed);
159 LASSERT(list_empty(&lock->l_res_link));
160 LASSERT(list_empty(&lock->l_pending_chain));
162 atomic_dec(&res->lr_namespace->ns_locks);
163 lu_ref_del(&res->lr_reference, "lock", lock);
164 ldlm_resource_putref(res);
165 lock->l_resource = NULL;
166 if (lock->l_export) {
167 class_export_put(lock->l_export);
168 lock->l_export = NULL;
171 if (lock->l_lvb_data != NULL)
172 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
174 ldlm_interval_free(ldlm_interval_detach(lock));
175 lu_ref_fini(&lock->l_reference);
176 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
183 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
186 if (!list_empty(&lock->l_lru)) {
187 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
188 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
189 list_del_init(&lock->l_lru);
191 LASSERT(ns->ns_nr_unused >= 0);
197 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
199 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
202 spin_lock(&ns->ns_unused_lock);
203 rc = ldlm_lock_remove_from_lru_nolock(lock);
204 spin_unlock(&ns->ns_unused_lock);
209 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
211 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
212 lock->l_last_used = cfs_time_current();
213 LASSERT(list_empty(&lock->l_lru));
214 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
215 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
216 LASSERT(ns->ns_nr_unused >= 0);
220 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
222 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
224 spin_lock(&ns->ns_unused_lock);
225 ldlm_lock_add_to_lru_nolock(lock);
226 spin_unlock(&ns->ns_unused_lock);
230 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
232 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
234 spin_lock(&ns->ns_unused_lock);
235 if (!list_empty(&lock->l_lru)) {
236 ldlm_lock_remove_from_lru_nolock(lock);
237 ldlm_lock_add_to_lru_nolock(lock);
239 spin_unlock(&ns->ns_unused_lock);
243 /* This used to have a 'strict' flag, which recovery would use to mark an
244 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
245 * shall explain why it's gone: with the new hash table scheme, once you call
246 * ldlm_lock_destroy, you can never drop your final references on this lock.
247 * Because it's not in the hash table anymore. -phil */
248 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
252 if (lock->l_readers || lock->l_writers) {
253 LDLM_ERROR(lock, "lock still has references");
254 ldlm_lock_dump(D_ERROR, lock, 0);
258 if (!list_empty(&lock->l_res_link)) {
259 LDLM_ERROR(lock, "lock still on resource");
260 ldlm_lock_dump(D_ERROR, lock, 0);
264 if (lock->l_destroyed) {
265 LASSERT(list_empty(&lock->l_lru));
269 lock->l_destroyed = 1;
271 if (lock->l_export && lock->l_export->exp_lock_hash)
272 lustre_hash_del(lock->l_export->exp_lock_hash,
273 &lock->l_remote_handle, &lock->l_exp_hash);
275 ldlm_lock_remove_from_lru(lock);
276 class_handle_unhash(&lock->l_handle);
279 /* Wake anyone waiting for this lock */
280 /* FIXME: I should probably add yet another flag, instead of using
281 * l_export to only call this on clients */
283 class_export_put(lock->l_export);
284 lock->l_export = NULL;
285 if (lock->l_export && lock->l_completion_ast)
286 lock->l_completion_ast(lock, 0);
292 void ldlm_lock_destroy(struct ldlm_lock *lock)
296 lock_res_and_lock(lock);
297 first = ldlm_lock_destroy_internal(lock);
298 unlock_res_and_lock(lock);
300 /* drop reference from hashtable only for first destroy */
302 lu_ref_del(&lock->l_reference, "hash", lock);
303 LDLM_LOCK_RELEASE(lock);
308 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
312 first = ldlm_lock_destroy_internal(lock);
313 /* drop reference from hashtable only for first destroy */
315 lu_ref_del(&lock->l_reference, "hash", lock);
316 LDLM_LOCK_RELEASE(lock);
321 /* this is called by portals_handle2object with the handle lock taken */
322 static void lock_handle_addref(void *lock)
324 LDLM_LOCK_GET((struct ldlm_lock *)lock);
328 * usage: pass in a resource on which you have done ldlm_resource_get
329 * pass in a parent lock on which you have done a ldlm_lock_get
330 * after return, ldlm_*_put the resource and parent
331 * returns: lock with refcount 2 - one for current caller and one for remote
333 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
335 struct ldlm_lock *lock;
338 if (resource == NULL)
341 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, CFS_ALLOC_IO, sizeof(*lock));
345 spin_lock_init(&lock->l_lock);
346 lock->l_resource = ldlm_resource_getref(resource);
347 lu_ref_add(&resource->lr_reference, "lock", lock);
349 atomic_set(&lock->l_refc, 2);
350 CFS_INIT_LIST_HEAD(&lock->l_res_link);
351 CFS_INIT_LIST_HEAD(&lock->l_lru);
352 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
353 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
354 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
355 CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
356 cfs_waitq_init(&lock->l_waitq);
357 lock->l_blocking_lock = NULL;
358 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
359 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
360 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
362 atomic_inc(&resource->lr_namespace->ns_locks);
363 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
364 class_handle_hash(&lock->l_handle, lock_handle_addref);
366 CFS_INIT_LIST_HEAD(&lock->l_extents_list);
367 spin_lock_init(&lock->l_extents_list_lock);
368 CFS_INIT_LIST_HEAD(&lock->l_cache_locks_list);
369 lu_ref_init(&lock->l_reference);
370 lu_ref_add(&lock->l_reference, "hash", lock);
375 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
376 const struct ldlm_res_id *new_resid)
378 struct ldlm_resource *oldres = lock->l_resource;
379 struct ldlm_resource *newres;
383 LASSERT(ns_is_client(ns));
385 lock_res_and_lock(lock);
386 if (memcmp(new_resid, &lock->l_resource->lr_name,
387 sizeof(lock->l_resource->lr_name)) == 0) {
389 unlock_res_and_lock(lock);
393 LASSERT(new_resid->name[0] != 0);
395 /* This function assumes that the lock isn't on any lists */
396 LASSERT(list_empty(&lock->l_res_link));
398 type = oldres->lr_type;
399 unlock_res_and_lock(lock);
401 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
402 lu_ref_add(&newres->lr_reference, "lock", lock);
406 * To flip the lock from the old to the new resource, lock, oldres and
407 * newres have to be locked. Resource spin-locks are nested within
408 * lock->l_lock, and are taken in the memory address order to avoid
411 spin_lock(&lock->l_lock);
412 oldres = lock->l_resource;
413 if (oldres < newres) {
415 lock_res_nested(newres, LRT_NEW);
418 lock_res_nested(oldres, LRT_NEW);
420 LASSERT(memcmp(new_resid, &oldres->lr_name,
421 sizeof oldres->lr_name) != 0);
422 lock->l_resource = newres;
424 unlock_res_and_lock(lock);
426 /* ...and the flowers are still standing! */
427 lu_ref_del(&oldres->lr_reference, "lock", lock);
428 ldlm_resource_putref(oldres);
437 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
439 lockh->cookie = lock->l_handle.h_cookie;
442 /* if flags: atomically get the lock and set the flags.
443 * Return NULL if flag already set
446 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
449 struct ldlm_namespace *ns;
450 struct ldlm_lock *lock, *retval = NULL;
455 lock = class_handle2object(handle->cookie);
459 LASSERT(lock->l_resource != NULL);
460 ns = lock->l_resource->lr_namespace;
463 lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
464 lock_res_and_lock(lock);
466 /* It's unlikely but possible that someone marked the lock as
467 * destroyed after we did handle2object on it */
468 if (lock->l_destroyed) {
469 unlock_res_and_lock(lock);
470 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
475 if (flags && (lock->l_flags & flags)) {
476 unlock_res_and_lock(lock);
482 lock->l_flags |= flags;
484 unlock_res_and_lock(lock);
491 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
493 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
494 /* INODEBITS_INTEROP: If the other side does not support
495 * inodebits, reply with a plain lock descriptor.
497 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
498 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
499 struct ldlm_resource res = *lock->l_resource;
501 /* Make sure all the right bits are set in this lock we
502 are going to pass to client */
503 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
504 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
505 "Inappropriate inode lock bits during "
506 "conversion " LPU64 "\n",
507 lock->l_policy_data.l_inodebits.bits);
508 res.lr_type = LDLM_PLAIN;
509 ldlm_res2desc(&res, &desc->l_resource);
510 /* Convert "new" lock mode to something old client can
512 if ((lock->l_req_mode == LCK_CR) ||
513 (lock->l_req_mode == LCK_CW))
514 desc->l_req_mode = LCK_PR;
516 desc->l_req_mode = lock->l_req_mode;
517 if ((lock->l_granted_mode == LCK_CR) ||
518 (lock->l_granted_mode == LCK_CW)) {
519 desc->l_granted_mode = LCK_PR;
521 /* We never grant PW/EX locks to clients */
522 LASSERT((lock->l_granted_mode != LCK_PW) &&
523 (lock->l_granted_mode != LCK_EX));
524 desc->l_granted_mode = lock->l_granted_mode;
527 /* We do not copy policy here, because there is no
528 policy for plain locks */
530 ldlm_res2desc(lock->l_resource, &desc->l_resource);
531 desc->l_req_mode = lock->l_req_mode;
532 desc->l_granted_mode = lock->l_granted_mode;
533 desc->l_policy_data = lock->l_policy_data;
537 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
538 struct list_head *work_list)
540 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
541 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
542 lock->l_flags |= LDLM_FL_AST_SENT;
543 /* If the enqueuing client said so, tell the AST recipient to
544 * discard dirty data, rather than writing back. */
545 if (new->l_flags & LDLM_AST_DISCARD_DATA)
546 lock->l_flags |= LDLM_FL_DISCARD_DATA;
547 LASSERT(list_empty(&lock->l_bl_ast));
548 list_add(&lock->l_bl_ast, work_list);
550 LASSERT(lock->l_blocking_lock == NULL);
551 lock->l_blocking_lock = LDLM_LOCK_GET(new);
555 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
557 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
558 lock->l_flags |= LDLM_FL_CP_REQD;
559 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
560 LASSERT(list_empty(&lock->l_cp_ast));
561 list_add(&lock->l_cp_ast, work_list);
566 /* must be called with lr_lock held */
567 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
568 struct list_head *work_list)
571 check_res_locked(lock->l_resource);
573 ldlm_add_bl_work_item(lock, new, work_list);
575 ldlm_add_cp_work_item(lock, work_list);
579 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
581 struct ldlm_lock *lock;
583 lock = ldlm_handle2lock(lockh);
584 LASSERT(lock != NULL);
585 ldlm_lock_addref_internal(lock, mode);
589 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
591 ldlm_lock_remove_from_lru(lock);
592 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
594 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
596 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
598 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
601 lu_ref_add_atomic(&lock->l_reference, "user", lock);
602 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
606 * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
609 * \retval 0 success, lock was addref-ed
611 * \retval -EAGAIN lock is being canceled.
613 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
615 struct ldlm_lock *lock;
619 lock = ldlm_handle2lock(lockh);
621 lock_res_and_lock(lock);
622 if (!(lock->l_flags & LDLM_FL_CBPENDING)) {
623 ldlm_lock_addref_internal_nolock(lock, mode);
626 unlock_res_and_lock(lock);
632 /* only called for local locks */
633 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
635 lock_res_and_lock(lock);
636 ldlm_lock_addref_internal_nolock(lock, mode);
637 unlock_res_and_lock(lock);
640 /* only called in ldlm_flock_destroy and for local locks.
641 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
642 * * ldlm_lock_remove_from_lru() does nothing, it is safe
643 * * for ldlm_flock_destroy usage by dropping some code */
644 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
646 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
647 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
648 LASSERT(lock->l_readers > 0);
649 lu_ref_del(&lock->l_reference, "reader", lock);
652 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
653 LASSERT(lock->l_writers > 0);
654 lu_ref_del(&lock->l_reference, "writer", lock);
658 lu_ref_del(&lock->l_reference, "user", lock);
659 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
662 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
664 struct ldlm_namespace *ns;
667 lock_res_and_lock(lock);
669 ns = lock->l_resource->lr_namespace;
671 ldlm_lock_decref_internal_nolock(lock, mode);
673 if (lock->l_flags & LDLM_FL_LOCAL &&
674 !lock->l_readers && !lock->l_writers) {
675 /* If this is a local lock on a server namespace and this was
676 * the last reference, cancel the lock. */
677 CDEBUG(D_INFO, "forcing cancel of local lock\n");
678 lock->l_flags |= LDLM_FL_CBPENDING;
681 if (!lock->l_readers && !lock->l_writers &&
682 (lock->l_flags & LDLM_FL_CBPENDING)) {
683 /* If we received a blocked AST and this was the last reference,
684 * run the callback. */
685 if (ns_is_server(ns) && lock->l_export)
686 CERROR("FL_CBPENDING set on non-local lock--just a "
689 LDLM_DEBUG(lock, "final decref done on cbpending lock");
691 LDLM_LOCK_GET(lock); /* dropped by bl thread */
692 ldlm_lock_remove_from_lru(lock);
693 unlock_res_and_lock(lock);
694 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
695 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
696 ldlm_handle_bl_callback(ns, NULL, lock);
697 } else if (ns_is_client(ns) &&
698 !lock->l_readers && !lock->l_writers &&
699 !(lock->l_flags & LDLM_FL_BL_AST)) {
700 /* If this is a client-side namespace and this was the last
701 * reference, put it on the LRU. */
702 ldlm_lock_add_to_lru(lock);
703 unlock_res_and_lock(lock);
704 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
705 * are not supported by the server, otherwise, it is done on
707 if (!exp_connect_cancelset(lock->l_conn_export) &&
708 !ns_connect_lru_resize(ns))
709 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
711 unlock_res_and_lock(lock);
717 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
719 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
720 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
721 ldlm_lock_decref_internal(lock, mode);
725 /* This will drop a lock reference and mark it for destruction, but will not
726 * necessarily cancel the lock before returning. */
727 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
729 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
732 LASSERT(lock != NULL);
734 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
735 lock_res_and_lock(lock);
736 lock->l_flags |= LDLM_FL_CBPENDING;
737 unlock_res_and_lock(lock);
738 ldlm_lock_decref_internal(lock, mode);
742 struct sl_insert_point {
743 struct list_head *res_link;
744 struct list_head *mode_link;
745 struct list_head *policy_link;
749 * search_granted_lock
752 * Finds a position to insert the new lock.
754 * queue [input]: the granted list where search acts on;
755 * req [input]: the lock whose position to be located;
756 * prev [output]: positions within 3 lists to insert @req to
760 * - ldlm_grant_lock_with_skiplist
762 static void search_granted_lock(struct list_head *queue,
763 struct ldlm_lock *req,
764 struct sl_insert_point *prev)
766 struct list_head *tmp;
767 struct ldlm_lock *lock, *mode_end, *policy_end;
770 list_for_each(tmp, queue) {
771 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
773 mode_end = list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
776 if (lock->l_req_mode != req->l_req_mode) {
777 /* jump to last lock of mode group */
778 tmp = &mode_end->l_res_link;
782 /* suitable mode group is found */
783 if (lock->l_resource->lr_type == LDLM_PLAIN) {
784 /* insert point is last lock of the mode group */
785 prev->res_link = &mode_end->l_res_link;
786 prev->mode_link = &mode_end->l_sl_mode;
787 prev->policy_link = &req->l_sl_policy;
790 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
792 policy_end = list_entry(lock->l_sl_policy.prev,
796 if (lock->l_policy_data.l_inodebits.bits ==
797 req->l_policy_data.l_inodebits.bits) {
798 /* insert point is last lock of
799 * the policy group */
801 &policy_end->l_res_link;
803 &policy_end->l_sl_mode;
805 &policy_end->l_sl_policy;
810 if (policy_end == mode_end)
811 /* done with mode group */
814 /* go to next policy group within mode group */
815 tmp = policy_end->l_res_link.next;
816 lock = list_entry(tmp, struct ldlm_lock,
818 } /* loop over policy groups within the mode group */
820 /* insert point is last lock of the mode group,
821 * new policy group is started */
822 prev->res_link = &mode_end->l_res_link;
823 prev->mode_link = &mode_end->l_sl_mode;
824 prev->policy_link = &req->l_sl_policy;
828 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
833 /* insert point is last lock on the queue,
834 * new mode group and new policy group are started */
835 prev->res_link = queue->prev;
836 prev->mode_link = &req->l_sl_mode;
837 prev->policy_link = &req->l_sl_policy;
842 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
843 struct sl_insert_point *prev)
845 struct ldlm_resource *res = lock->l_resource;
848 check_res_locked(res);
850 ldlm_resource_dump(D_INFO, res);
851 CDEBUG(D_OTHER, "About to add this lock:\n");
852 ldlm_lock_dump(D_OTHER, lock, 0);
854 if (lock->l_destroyed) {
855 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
859 LASSERT(list_empty(&lock->l_res_link));
860 LASSERT(list_empty(&lock->l_sl_mode));
861 LASSERT(list_empty(&lock->l_sl_policy));
863 list_add(&lock->l_res_link, prev->res_link);
864 list_add(&lock->l_sl_mode, prev->mode_link);
865 list_add(&lock->l_sl_policy, prev->policy_link);
870 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
872 struct sl_insert_point prev;
875 LASSERT(lock->l_req_mode == lock->l_granted_mode);
877 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
878 ldlm_granted_list_add_lock(lock, &prev);
883 * - ldlm_lock_enqueue
884 * - ldlm_reprocess_queue
885 * - ldlm_lock_convert
887 * must be called with lr_lock held
889 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
891 struct ldlm_resource *res = lock->l_resource;
894 check_res_locked(res);
896 lock->l_granted_mode = lock->l_req_mode;
897 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
898 ldlm_grant_lock_with_skiplist(lock);
899 else if (res->lr_type == LDLM_EXTENT)
900 ldlm_extent_add_lock(res, lock);
902 ldlm_resource_add_lock(res, &res->lr_granted, lock);
904 if (lock->l_granted_mode < res->lr_most_restr)
905 res->lr_most_restr = lock->l_granted_mode;
907 if (work_list && lock->l_completion_ast != NULL)
908 ldlm_add_ast_work_item(lock, NULL, work_list);
910 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
914 /* returns a referenced lock or NULL. See the flag descriptions below, in the
915 * comment above ldlm_lock_match */
916 static struct ldlm_lock *search_queue(struct list_head *queue,
918 ldlm_policy_data_t *policy,
919 struct ldlm_lock *old_lock, int flags)
921 struct ldlm_lock *lock;
922 struct list_head *tmp;
924 list_for_each(tmp, queue) {
927 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
929 if (lock == old_lock)
932 /* llite sometimes wants to match locks that will be
933 * canceled when their users drop, but we allow it to match
934 * if it passes in CBPENDING and the lock still has users.
935 * this is generally only going to be used by children
936 * whose parents already hold a lock so forward progress
937 * can still happen. */
938 if (lock->l_flags & LDLM_FL_CBPENDING &&
939 !(flags & LDLM_FL_CBPENDING))
941 if (lock->l_flags & LDLM_FL_CBPENDING &&
942 lock->l_readers == 0 && lock->l_writers == 0)
945 if (!(lock->l_req_mode & *mode))
947 match = lock->l_req_mode;
949 if (lock->l_resource->lr_type == LDLM_EXTENT &&
950 (lock->l_policy_data.l_extent.start >
951 policy->l_extent.start ||
952 lock->l_policy_data.l_extent.end < policy->l_extent.end))
955 if (unlikely(match == LCK_GROUP) &&
956 lock->l_resource->lr_type == LDLM_EXTENT &&
957 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
960 /* We match if we have existing lock with same or wider set
962 if (lock->l_resource->lr_type == LDLM_IBITS &&
963 ((lock->l_policy_data.l_inodebits.bits &
964 policy->l_inodebits.bits) !=
965 policy->l_inodebits.bits))
968 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
971 if ((flags & LDLM_FL_LOCAL_ONLY) &&
972 !(lock->l_flags & LDLM_FL_LOCAL))
975 if (flags & LDLM_FL_TEST_LOCK) {
977 ldlm_lock_touch_in_lru(lock);
979 ldlm_lock_addref_internal_nolock(lock, match);
988 void ldlm_lock_allow_match(struct ldlm_lock *lock)
990 lock_res_and_lock(lock);
991 lock->l_flags |= LDLM_FL_LVB_READY;
992 cfs_waitq_signal(&lock->l_waitq);
993 unlock_res_and_lock(lock);
997 * Checks if requested extent lock is compatible with another owned lock.
999 * Checks if \a lock is compatible with a read or write lock
1000 * (specified by \a rw) for an extent [\a start , \a end].
1002 * \param lock the already owned lock
1003 * \param rw OBD_BRW_READ if requested for reading,
1004 * OBD_BRW_WRITE if requested for writing
1005 * \param start start of the requested extent
1006 * \param end end of the requested extent
1007 * \param cookie transparent parameter for passing locking context
1009 * \post result == 1, *cookie == context, appropriate lock is referenced
1011 * \retval 1 owned lock is reused for the request
1012 * \retval 0 no lock reused for the request
1014 * \see ldlm_lock_fast_release
1016 int ldlm_lock_fast_match(struct ldlm_lock *lock, int rw,
1017 obd_off start, obd_off end,
1020 LASSERT(rw == OBD_BRW_READ || rw == OBD_BRW_WRITE);
1025 lock_res_and_lock(lock);
1026 /* check if granted mode is compatible */
1027 if (rw == OBD_BRW_WRITE &&
1028 !(lock->l_granted_mode & (LCK_PW|LCK_GROUP)))
1031 /* does the lock cover the region we would like to access? */
1032 if ((lock->l_policy_data.l_extent.start > start) ||
1033 (lock->l_policy_data.l_extent.end < end))
1036 /* if we received a blocking callback and the lock is no longer
1037 * referenced, don't use it */
1038 if ((lock->l_flags & LDLM_FL_CBPENDING) &&
1039 !lock->l_writers && !lock->l_readers)
1042 ldlm_lock_addref_internal_nolock(lock, rw == OBD_BRW_WRITE ?
1044 unlock_res_and_lock(lock);
1045 *cookie = (void *)lock;
1046 return 1; /* avoid using rc for stack relief */
1049 unlock_res_and_lock(lock);
1054 * Releases a reference to a lock taken in a "fast" way.
1056 * Releases a read or write (specified by \a rw) lock
1057 * referenced by \a cookie.
1059 * \param rw OBD_BRW_READ if requested for reading,
1060 * OBD_BRW_WRITE if requested for writing
1061 * \param cookie transparent parameter for passing locking context
1063 * \post appropriate lock is dereferenced
1065 * \see ldlm_lock_fast_lock
1067 void ldlm_lock_fast_release(void *cookie, int rw)
1069 struct ldlm_lock *lock = (struct ldlm_lock *)cookie;
1071 LASSERT(lock != NULL);
1072 LASSERT(rw == OBD_BRW_READ || rw == OBD_BRW_WRITE);
1073 LASSERT(rw == OBD_BRW_READ ||
1074 (lock->l_granted_mode & (LCK_PW | LCK_GROUP)));
1075 ldlm_lock_decref_internal(lock, rw == OBD_BRW_WRITE ? LCK_PW : LCK_PR);
1078 /* Can be called in two ways:
1080 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1081 * for a duplicate of.
1083 * Otherwise, all of the fields must be filled in, to match against.
1085 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1086 * server (ie, connh is NULL)
1087 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1088 * list will be considered
1089 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1090 * to be canceled can still be matched as long as they still have reader
1091 * or writer refernces
1092 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1093 * just tell us if we would have matched.
1095 * Returns 1 if it finds an already-existing lock that is compatible; in this
1096 * case, lockh is filled in with a addref()ed lock
1098 * we also check security context, if that failed we simply return 0 (to keep
1099 * caller code unchanged), the context failure will be discovered by caller
1102 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1103 const struct ldlm_res_id *res_id, ldlm_type_t type,
1104 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1105 struct lustre_handle *lockh)
1107 struct ldlm_resource *res;
1108 struct ldlm_lock *lock, *old_lock = NULL;
1113 old_lock = ldlm_handle2lock(lockh);
1116 ns = old_lock->l_resource->lr_namespace;
1117 res_id = &old_lock->l_resource->lr_name;
1118 type = old_lock->l_resource->lr_type;
1119 mode = old_lock->l_req_mode;
1122 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1124 LASSERT(old_lock == NULL);
1128 LDLM_RESOURCE_ADDREF(res);
1131 lock = search_queue(&res->lr_granted, &mode, policy, old_lock, flags);
1134 if (flags & LDLM_FL_BLOCK_GRANTED)
1136 lock = search_queue(&res->lr_converting, &mode, policy, old_lock,flags);
1139 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock, flags);
1146 LDLM_RESOURCE_DELREF(res);
1147 ldlm_resource_putref(res);
1150 ldlm_lock2handle(lock, lockh);
1151 if ((flags & LDLM_FL_LVB_READY) &&
1152 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1153 struct l_wait_info lwi;
1154 if (lock->l_completion_ast) {
1155 int err = lock->l_completion_ast(lock,
1156 LDLM_FL_WAIT_NOREPROC,
1159 if (flags & LDLM_FL_TEST_LOCK)
1160 LDLM_LOCK_RELEASE(lock);
1162 ldlm_lock_decref_internal(lock,
1169 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1170 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1172 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1173 l_wait_event(lock->l_waitq,
1174 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1179 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1180 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1181 res_id->name[2] : policy->l_extent.start,
1182 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1183 res_id->name[3] : policy->l_extent.end);
1185 /* check user's security context */
1186 if (lock->l_conn_export &&
1187 sptlrpc_import_check_ctx(
1188 class_exp2cliimp(lock->l_conn_export))) {
1189 if (!(flags & LDLM_FL_TEST_LOCK))
1190 ldlm_lock_decref_internal(lock, mode);
1194 if (flags & LDLM_FL_TEST_LOCK)
1195 LDLM_LOCK_RELEASE(lock);
1197 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1198 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1199 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1200 type, mode, res_id->name[0], res_id->name[1],
1201 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1202 res_id->name[2] :policy->l_extent.start,
1203 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1204 res_id->name[3] : policy->l_extent.end);
1207 LDLM_LOCK_PUT(old_lock);
1209 return rc ? mode : 0;
1212 /* Returns a referenced lock */
1213 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1214 const struct ldlm_res_id *res_id,
1217 const struct ldlm_callback_suite *cbs,
1218 void *data, __u32 lvb_len)
1220 struct ldlm_lock *lock;
1221 struct ldlm_resource *res;
1224 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1228 lock = ldlm_lock_new(res);
1229 ldlm_resource_putref(res);
1234 lock->l_req_mode = mode;
1235 lock->l_ast_data = data;
1236 lock->l_pid = cfs_curproc_pid();
1238 lock->l_blocking_ast = cbs->lcs_blocking;
1239 lock->l_completion_ast = cbs->lcs_completion;
1240 lock->l_glimpse_ast = cbs->lcs_glimpse;
1241 lock->l_weigh_ast = cbs->lcs_weigh;
1244 lock->l_tree_node = NULL;
1245 /* if this is the extent lock, allocate the interval tree node */
1246 if (type == LDLM_EXTENT) {
1247 if (ldlm_interval_alloc(lock) == NULL)
1252 lock->l_lvb_len = lvb_len;
1253 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1254 if (lock->l_lvb_data == NULL)
1261 if (lock->l_lvb_data)
1262 OBD_FREE(lock->l_lvb_data, lvb_len);
1263 ldlm_interval_free(ldlm_interval_detach(lock));
1264 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1268 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1269 struct ldlm_lock **lockp,
1270 void *cookie, int *flags)
1272 struct ldlm_lock *lock = *lockp;
1273 struct ldlm_resource *res = lock->l_resource;
1274 int local = ns_is_client(res->lr_namespace);
1275 ldlm_processing_policy policy;
1276 ldlm_error_t rc = ELDLM_OK;
1277 struct ldlm_interval *node = NULL;
1280 do_gettimeofday(&lock->l_enqueued_time);
1281 /* policies are not executed on the client or during replay */
1282 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1283 && !local && ns->ns_policy) {
1284 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1286 if (rc == ELDLM_LOCK_REPLACED) {
1287 /* The lock that was returned has already been granted,
1288 * and placed into lockp. If it's not the same as the
1289 * one we passed in, then destroy the old one and our
1290 * work here is done. */
1291 if (lock != *lockp) {
1292 ldlm_lock_destroy(lock);
1293 LDLM_LOCK_RELEASE(lock);
1295 *flags |= LDLM_FL_LOCK_CHANGED;
1297 } else if (rc != ELDLM_OK ||
1298 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1299 ldlm_lock_destroy(lock);
1304 /* For a replaying lock, it might be already in granted list. So
1305 * unlinking the lock will cause the interval node to be freed, we
1306 * have to allocate the interval node early otherwise we can't regrant
1307 * this lock in the future. - jay */
1308 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1309 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO,
1312 lock_res_and_lock(lock);
1313 if (local && lock->l_req_mode == lock->l_granted_mode) {
1314 /* The server returned a blocked lock, but it was granted
1315 * before we got a chance to actually enqueue it. We don't
1316 * need to do anything else. */
1317 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1318 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1319 GOTO(out, ELDLM_OK);
1322 ldlm_resource_unlink_lock(lock);
1323 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1325 ldlm_lock_destroy_nolock(lock);
1326 GOTO(out, rc = -ENOMEM);
1329 CFS_INIT_LIST_HEAD(&node->li_group);
1330 ldlm_interval_attach(node, lock);
1334 /* Some flags from the enqueue want to make it into the AST, via the
1335 * lock's l_flags. */
1336 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1338 /* This distinction between local lock trees is very important; a client
1339 * namespace only has information about locks taken by that client, and
1340 * thus doesn't have enough information to decide for itself if it can
1341 * be granted (below). In this case, we do exactly what the server
1342 * tells us to do, as dictated by the 'flags'.
1344 * We do exactly the same thing during recovery, when the server is
1345 * more or less trusting the clients not to lie.
1347 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1348 * granted/converting queues. */
1350 if (*flags & LDLM_FL_BLOCK_CONV)
1351 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1352 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1353 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1355 ldlm_grant_lock(lock, NULL);
1356 GOTO(out, ELDLM_OK);
1357 } else if (*flags & LDLM_FL_REPLAY) {
1358 if (*flags & LDLM_FL_BLOCK_CONV) {
1359 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1360 GOTO(out, ELDLM_OK);
1361 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1362 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1363 GOTO(out, ELDLM_OK);
1364 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1365 ldlm_grant_lock(lock, NULL);
1366 GOTO(out, ELDLM_OK);
1368 /* If no flags, fall through to normal enqueue path. */
1371 policy = ldlm_processing_policy_table[res->lr_type];
1372 policy(lock, flags, 1, &rc, NULL);
1375 unlock_res_and_lock(lock);
1377 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1381 /* Must be called with namespace taken: queue is waiting or converting. */
1382 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1383 struct list_head *work_list)
1385 struct list_head *tmp, *pos;
1386 ldlm_processing_policy policy;
1388 int rc = LDLM_ITER_CONTINUE;
1392 check_res_locked(res);
1394 policy = ldlm_processing_policy_table[res->lr_type];
1397 list_for_each_safe(tmp, pos, queue) {
1398 struct ldlm_lock *pending;
1399 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1401 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1404 rc = policy(pending, &flags, 0, &err, work_list);
1405 if (rc != LDLM_ITER_CONTINUE)
1412 /* Helper function for ldlm_run_ast_work().
1414 * Send an existing rpc set specified by @arg->set and then
1415 * destroy it. Create new one if @do_create flag is set. */
1417 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1421 ptlrpc_set_wait(arg->set);
1422 if (arg->type == LDLM_BL_CALLBACK)
1423 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1424 ptlrpc_set_destroy(arg->set);
1427 arg->set = ptlrpc_prep_set();
1433 ldlm_work_bl_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1435 struct ldlm_lock_desc d;
1436 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_bl_ast);
1439 /* nobody should touch l_bl_ast */
1440 lock_res_and_lock(lock);
1441 list_del_init(&lock->l_bl_ast);
1443 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1444 LASSERT(lock->l_bl_ast_run == 0);
1445 LASSERT(lock->l_blocking_lock);
1446 lock->l_bl_ast_run++;
1447 unlock_res_and_lock(lock);
1449 ldlm_lock2desc(lock->l_blocking_lock, &d);
1451 lock->l_blocking_ast(lock, &d, (void *)arg,
1453 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1454 lock->l_blocking_lock = NULL;
1455 LDLM_LOCK_RELEASE(lock);
1461 ldlm_work_cp_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1463 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_cp_ast);
1464 ldlm_completion_callback completion_callback;
1468 /* It's possible to receive a completion AST before we've set
1469 * the l_completion_ast pointer: either because the AST arrived
1470 * before the reply, or simply because there's a small race
1471 * window between receiving the reply and finishing the local
1472 * enqueue. (bug 842)
1474 * This can't happen with the blocking_ast, however, because we
1475 * will never call the local blocking_ast until we drop our
1476 * reader/writer reference, which we won't do until we get the
1477 * reply and finish enqueueing. */
1479 /* nobody should touch l_cp_ast */
1480 lock_res_and_lock(lock);
1481 list_del_init(&lock->l_cp_ast);
1482 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1483 /* save l_completion_ast since it can be changed by
1484 * mds_intent_policy(), see bug 14225 */
1485 completion_callback = lock->l_completion_ast;
1486 lock->l_flags &= ~LDLM_FL_CP_REQD;
1487 unlock_res_and_lock(lock);
1489 if (completion_callback != NULL) {
1490 completion_callback(lock, 0, (void *)arg);
1493 LDLM_LOCK_RELEASE(lock);
1499 ldlm_work_revoke_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1501 struct ldlm_lock_desc desc;
1502 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_rk_ast);
1505 list_del_init(&lock->l_rk_ast);
1507 /* the desc just pretend to exclusive */
1508 ldlm_lock2desc(lock, &desc);
1509 desc.l_req_mode = LCK_EX;
1510 desc.l_granted_mode = 0;
1512 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1513 LDLM_LOCK_RELEASE(lock);
1518 int ldlm_run_ast_work(struct list_head *rpc_list, ldlm_desc_ast_t ast_type)
1520 struct ldlm_cb_set_arg arg;
1521 struct list_head *tmp, *pos;
1522 int (*work_ast_lock)(struct list_head *tmp,struct ldlm_cb_set_arg *arg);
1526 arg.set = ptlrpc_prep_set();
1527 atomic_set(&arg.restart, 0);
1529 case LDLM_WORK_BL_AST:
1530 arg.type = LDLM_BL_CALLBACK;
1531 work_ast_lock = ldlm_work_bl_ast_lock;
1533 case LDLM_WORK_CP_AST:
1534 arg.type = LDLM_CP_CALLBACK;
1535 work_ast_lock = ldlm_work_cp_ast_lock;
1537 case LDLM_WORK_REVOKE_AST:
1538 arg.type = LDLM_BL_CALLBACK;
1539 work_ast_lock = ldlm_work_revoke_ast_lock;
1546 list_for_each_safe(tmp, pos, rpc_list) {
1547 ast_count += work_ast_lock(tmp, &arg);
1549 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1550 * and create a new set for requests that remained in
1552 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1553 ldlm_send_and_maybe_create_set(&arg, 1);
1559 ldlm_send_and_maybe_create_set(&arg, 0);
1561 /* In case when number of ASTs is multiply of
1562 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1563 * @arg.set must be destroyed here, otherwise we get
1564 * write memory leaking. */
1565 ptlrpc_set_destroy(arg.set);
1567 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1570 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1572 ldlm_reprocess_all(res);
1573 return LDLM_ITER_CONTINUE;
1576 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1578 struct list_head *tmp;
1585 spin_lock(&ns->ns_hash_lock);
1586 for (i = 0; i < RES_HASH_SIZE; i++) {
1587 tmp = ns->ns_hash[i].next;
1588 while (tmp != &(ns->ns_hash[i])) {
1589 struct ldlm_resource *res =
1590 list_entry(tmp, struct ldlm_resource, lr_hash);
1592 ldlm_resource_getref(res);
1593 spin_unlock(&ns->ns_hash_lock);
1594 LDLM_RESOURCE_ADDREF(res);
1596 rc = reprocess_one_queue(res, NULL);
1598 LDLM_RESOURCE_DELREF(res);
1599 spin_lock(&ns->ns_hash_lock);
1601 ldlm_resource_putref_locked(res);
1603 if (rc == LDLM_ITER_STOP)
1608 spin_unlock(&ns->ns_hash_lock);
1612 void ldlm_reprocess_all(struct ldlm_resource *res)
1614 CFS_LIST_HEAD(rpc_list);
1618 /* Local lock trees don't get reprocessed. */
1619 if (ns_is_client(res->lr_namespace)) {
1626 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1627 if (rc == LDLM_ITER_CONTINUE)
1628 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1631 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1632 if (rc == -ERESTART) {
1633 LASSERT(list_empty(&rpc_list));
1639 void ldlm_cancel_callback(struct ldlm_lock *lock)
1641 check_res_locked(lock->l_resource);
1642 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1643 lock->l_flags |= LDLM_FL_CANCEL;
1644 if (lock->l_blocking_ast) {
1645 // l_check_no_ns_lock(ns);
1646 unlock_res_and_lock(lock);
1647 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1649 lock_res_and_lock(lock);
1651 LDLM_DEBUG(lock, "no blocking ast");
1654 lock->l_flags |= LDLM_FL_BL_DONE;
1657 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1659 if (req->l_resource->lr_type != LDLM_PLAIN &&
1660 req->l_resource->lr_type != LDLM_IBITS)
1663 list_del_init(&req->l_sl_policy);
1664 list_del_init(&req->l_sl_mode);
1667 void ldlm_lock_cancel(struct ldlm_lock *lock)
1669 struct ldlm_resource *res;
1670 struct ldlm_namespace *ns;
1673 lock_res_and_lock(lock);
1675 res = lock->l_resource;
1676 ns = res->lr_namespace;
1678 /* Please do not, no matter how tempting, remove this LBUG without
1679 * talking to me first. -phik */
1680 if (lock->l_readers || lock->l_writers) {
1681 LDLM_ERROR(lock, "lock still has references");
1685 ldlm_del_waiting_lock(lock);
1687 /* Releases cancel callback. */
1688 ldlm_cancel_callback(lock);
1690 /* Yes, second time, just in case it was added again while we were
1691 running with no res lock in ldlm_cancel_callback */
1692 ldlm_del_waiting_lock(lock);
1693 ldlm_resource_unlink_lock(lock);
1694 ldlm_lock_destroy_nolock(lock);
1696 if (lock->l_granted_mode == lock->l_req_mode)
1697 ldlm_pool_del(&ns->ns_pool, lock);
1699 /* Make sure we will not be called again for same lock what is possible
1700 * if not to zero out lock->l_granted_mode */
1701 lock->l_granted_mode = LCK_MINMODE;
1702 unlock_res_and_lock(lock);
1707 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1709 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1715 lock->l_ast_data = data;
1716 LDLM_LOCK_PUT(lock);
1720 void ldlm_cancel_locks_for_export_cb(void *obj, void *data)
1722 struct obd_export *exp = data;
1723 struct ldlm_lock *lock = obj;
1724 struct ldlm_resource *res;
1726 res = ldlm_resource_getref(lock->l_resource);
1727 LDLM_LOCK_GET(lock);
1729 LDLM_DEBUG(lock, "export %p", exp);
1730 ldlm_res_lvbo_update(res, NULL, 0, 1);
1731 ldlm_lock_cancel(lock);
1732 ldlm_reprocess_all(res);
1733 ldlm_resource_putref(res);
1734 LDLM_LOCK_RELEASE(lock);
1737 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1739 lustre_hash_for_each_empty(exp->exp_lock_hash,
1740 ldlm_cancel_locks_for_export_cb, exp);
1744 * Downgrade an exclusive lock.
1746 * A fast variant of ldlm_lock_convert for convertion of exclusive
1747 * locks. The convertion is always successful.
1749 * \param lock A lock to convert
1750 * \param new_mode new lock mode
1752 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1756 LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1757 LASSERT(new_mode == LCK_COS);
1759 lock_res_and_lock(lock);
1760 ldlm_resource_unlink_lock(lock);
1761 lock->l_req_mode = new_mode;
1762 ldlm_grant_lock(lock, NULL);
1763 unlock_res_and_lock(lock);
1764 ldlm_reprocess_all(lock->l_resource);
1769 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1772 CFS_LIST_HEAD(rpc_list);
1773 struct ldlm_resource *res;
1774 struct ldlm_namespace *ns;
1777 struct sl_insert_point prev;
1779 struct ldlm_interval *node;
1782 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1783 *flags |= LDLM_FL_BLOCK_GRANTED;
1784 RETURN(lock->l_resource);
1787 /* I can't check the type of lock here because the bitlock of lock
1788 * is not held here, so do the allocation blindly. -jay */
1789 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO, sizeof(*node));
1790 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1793 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1794 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1796 lock_res_and_lock(lock);
1798 res = lock->l_resource;
1799 ns = res->lr_namespace;
1801 old_mode = lock->l_req_mode;
1802 lock->l_req_mode = new_mode;
1803 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1804 /* remember the lock position where the lock might be
1805 * added back to the granted list later and also
1806 * remember the join mode for skiplist fixing. */
1807 prev.res_link = lock->l_res_link.prev;
1808 prev.mode_link = lock->l_sl_mode.prev;
1809 prev.policy_link = lock->l_sl_policy.prev;
1810 ldlm_resource_unlink_lock(lock);
1812 ldlm_resource_unlink_lock(lock);
1813 if (res->lr_type == LDLM_EXTENT) {
1814 /* FIXME: ugly code, I have to attach the lock to a
1815 * interval node again since perhaps it will be granted
1817 CFS_INIT_LIST_HEAD(&node->li_group);
1818 ldlm_interval_attach(node, lock);
1823 /* If this is a local resource, put it on the appropriate list. */
1824 if (ns_is_client(res->lr_namespace)) {
1825 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1826 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1828 /* This should never happen, because of the way the
1829 * server handles conversions. */
1830 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1834 ldlm_grant_lock(lock, &rpc_list);
1836 /* FIXME: completion handling not with ns_lock held ! */
1837 if (lock->l_completion_ast)
1838 lock->l_completion_ast(lock, 0, NULL);
1842 ldlm_processing_policy policy;
1843 policy = ldlm_processing_policy_table[res->lr_type];
1844 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1845 if (rc == LDLM_ITER_STOP) {
1846 lock->l_req_mode = old_mode;
1847 if (res->lr_type == LDLM_EXTENT)
1848 ldlm_extent_add_lock(res, lock);
1850 ldlm_granted_list_add_lock(lock, &prev);
1854 *flags |= LDLM_FL_BLOCK_GRANTED;
1858 unlock_res_and_lock(lock);
1861 ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1863 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1867 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1869 struct obd_device *obd = NULL;
1871 if (!((libcfs_debug | D_ERROR) & level))
1875 CDEBUG(level, " NULL LDLM lock\n");
1879 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1880 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1882 if (lock->l_conn_export != NULL)
1883 obd = lock->l_conn_export->exp_obd;
1884 if (lock->l_export && lock->l_export->exp_connection) {
1885 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1886 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1887 lock->l_remote_handle.cookie);
1888 } else if (obd == NULL) {
1889 CDEBUG(level, " Node: local\n");
1891 struct obd_import *imp = obd->u.cli.cl_import;
1892 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1893 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1894 lock->l_remote_handle.cookie);
1896 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1898 lock->l_resource->lr_name.name[0],
1899 lock->l_resource->lr_name.name[1],
1900 lock->l_resource->lr_name.name[2]);
1901 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1902 "write: %d flags: %#x\n", ldlm_lockname[lock->l_req_mode],
1903 ldlm_lockname[lock->l_granted_mode],
1904 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1906 if (lock->l_resource->lr_type == LDLM_EXTENT)
1907 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1908 " (req "LPU64"-"LPU64")\n",
1909 lock->l_policy_data.l_extent.start,
1910 lock->l_policy_data.l_extent.end,
1911 lock->l_req_extent.start, lock->l_req_extent.end);
1912 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1913 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1914 lock->l_policy_data.l_flock.pid,
1915 lock->l_policy_data.l_flock.start,
1916 lock->l_policy_data.l_flock.end);
1917 else if (lock->l_resource->lr_type == LDLM_IBITS)
1918 CDEBUG(level, " Bits: "LPX64"\n",
1919 lock->l_policy_data.l_inodebits.bits);
1922 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1924 struct ldlm_lock *lock;
1926 if (!((libcfs_debug | D_ERROR) & level))
1929 lock = ldlm_handle2lock(lockh);
1933 ldlm_lock_dump(D_OTHER, lock, 0);
1935 LDLM_LOCK_PUT(lock);
1938 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1939 struct libcfs_debug_msg_data *data, const char *fmt,
1943 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1945 va_start(args, fmt);
1947 if (lock->l_resource == NULL) {
1948 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1949 data->msg_fn, data->msg_line, fmt, args,
1950 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1951 "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "
1952 LPX64" expref: %d pid: %u\n", lock,
1953 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1954 lock->l_readers, lock->l_writers,
1955 ldlm_lockname[lock->l_granted_mode],
1956 ldlm_lockname[lock->l_req_mode],
1957 lock->l_flags, lock->l_remote_handle.cookie,
1959 atomic_read(&lock->l_export->exp_refcount) : -99,
1965 switch (lock->l_resource->lr_type) {
1967 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1968 data->msg_fn, data->msg_line, fmt, args,
1969 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1970 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1971 "] (req "LPU64"->"LPU64") flags: %x remote: "LPX64
1972 " expref: %d pid: %u\n",
1973 lock->l_resource->lr_namespace->ns_name, lock,
1974 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1975 lock->l_readers, lock->l_writers,
1976 ldlm_lockname[lock->l_granted_mode],
1977 ldlm_lockname[lock->l_req_mode],
1978 lock->l_resource->lr_name.name[0],
1979 lock->l_resource->lr_name.name[1],
1980 atomic_read(&lock->l_resource->lr_refcount),
1981 ldlm_typename[lock->l_resource->lr_type],
1982 lock->l_policy_data.l_extent.start,
1983 lock->l_policy_data.l_extent.end,
1984 lock->l_req_extent.start, lock->l_req_extent.end,
1985 lock->l_flags, lock->l_remote_handle.cookie,
1987 atomic_read(&lock->l_export->exp_refcount) : -99,
1992 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1993 data->msg_fn, data->msg_line, fmt, args,
1994 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1995 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1996 "["LPU64"->"LPU64"] flags: %x remote: "LPX64
1997 " expref: %d pid: %u\n",
1998 lock->l_resource->lr_namespace->ns_name, lock,
1999 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
2000 lock->l_readers, lock->l_writers,
2001 ldlm_lockname[lock->l_granted_mode],
2002 ldlm_lockname[lock->l_req_mode],
2003 lock->l_resource->lr_name.name[0],
2004 lock->l_resource->lr_name.name[1],
2005 atomic_read(&lock->l_resource->lr_refcount),
2006 ldlm_typename[lock->l_resource->lr_type],
2007 lock->l_policy_data.l_flock.pid,
2008 lock->l_policy_data.l_flock.start,
2009 lock->l_policy_data.l_flock.end,
2010 lock->l_flags, lock->l_remote_handle.cookie,
2012 atomic_read(&lock->l_export->exp_refcount) : -99,
2017 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2018 data->msg_fn, data->msg_line, fmt, args,
2019 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2020 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
2021 "flags: %x remote: "LPX64" expref: %d "
2023 lock->l_resource->lr_namespace->ns_name,
2024 lock, lock->l_handle.h_cookie,
2025 atomic_read (&lock->l_refc),
2026 lock->l_readers, lock->l_writers,
2027 ldlm_lockname[lock->l_granted_mode],
2028 ldlm_lockname[lock->l_req_mode],
2029 lock->l_resource->lr_name.name[0],
2030 lock->l_resource->lr_name.name[1],
2031 lock->l_policy_data.l_inodebits.bits,
2032 atomic_read(&lock->l_resource->lr_refcount),
2033 ldlm_typename[lock->l_resource->lr_type],
2034 lock->l_flags, lock->l_remote_handle.cookie,
2036 atomic_read(&lock->l_export->exp_refcount) : -99,
2041 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2042 data->msg_fn, data->msg_line, fmt, args,
2043 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2044 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: %x "
2045 "remote: "LPX64" expref: %d pid: %u\n",
2046 lock->l_resource->lr_namespace->ns_name,
2047 lock, lock->l_handle.h_cookie,
2048 atomic_read (&lock->l_refc),
2049 lock->l_readers, lock->l_writers,
2050 ldlm_lockname[lock->l_granted_mode],
2051 ldlm_lockname[lock->l_req_mode],
2052 lock->l_resource->lr_name.name[0],
2053 lock->l_resource->lr_name.name[1],
2054 atomic_read(&lock->l_resource->lr_refcount),
2055 ldlm_typename[lock->l_resource->lr_type],
2056 lock->l_flags, lock->l_remote_handle.cookie,
2058 atomic_read(&lock->l_export->exp_refcount) : -99,
2064 EXPORT_SYMBOL(_ldlm_lock_debug);