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_readers != 0 || lock->l_writers != 0 ||
623 !(lock->l_flags & LDLM_FL_CBPENDING)) {
624 ldlm_lock_addref_internal_nolock(lock, mode);
627 unlock_res_and_lock(lock);
633 /* only called for local locks */
634 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
636 lock_res_and_lock(lock);
637 ldlm_lock_addref_internal_nolock(lock, mode);
638 unlock_res_and_lock(lock);
641 /* only called in ldlm_flock_destroy and for local locks.
642 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
643 * * ldlm_lock_remove_from_lru() does nothing, it is safe
644 * * for ldlm_flock_destroy usage by dropping some code */
645 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
647 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
648 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
649 LASSERT(lock->l_readers > 0);
650 lu_ref_del(&lock->l_reference, "reader", lock);
653 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
654 LASSERT(lock->l_writers > 0);
655 lu_ref_del(&lock->l_reference, "writer", lock);
659 lu_ref_del(&lock->l_reference, "user", lock);
660 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
663 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
665 struct ldlm_namespace *ns;
668 lock_res_and_lock(lock);
670 ns = lock->l_resource->lr_namespace;
672 ldlm_lock_decref_internal_nolock(lock, mode);
674 if (lock->l_flags & LDLM_FL_LOCAL &&
675 !lock->l_readers && !lock->l_writers) {
676 /* If this is a local lock on a server namespace and this was
677 * the last reference, cancel the lock. */
678 CDEBUG(D_INFO, "forcing cancel of local lock\n");
679 lock->l_flags |= LDLM_FL_CBPENDING;
682 if (!lock->l_readers && !lock->l_writers &&
683 (lock->l_flags & LDLM_FL_CBPENDING)) {
684 /* If we received a blocked AST and this was the last reference,
685 * run the callback. */
686 if (ns_is_server(ns) && lock->l_export)
687 CERROR("FL_CBPENDING set on non-local lock--just a "
690 LDLM_DEBUG(lock, "final decref done on cbpending lock");
692 LDLM_LOCK_GET(lock); /* dropped by bl thread */
693 ldlm_lock_remove_from_lru(lock);
694 unlock_res_and_lock(lock);
695 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
696 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
697 ldlm_handle_bl_callback(ns, NULL, lock);
698 } else if (ns_is_client(ns) &&
699 !lock->l_readers && !lock->l_writers &&
700 !(lock->l_flags & LDLM_FL_BL_AST)) {
701 /* If this is a client-side namespace and this was the last
702 * reference, put it on the LRU. */
703 ldlm_lock_add_to_lru(lock);
704 unlock_res_and_lock(lock);
705 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
706 * are not supported by the server, otherwise, it is done on
708 if (!exp_connect_cancelset(lock->l_conn_export) &&
709 !ns_connect_lru_resize(ns))
710 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
712 unlock_res_and_lock(lock);
718 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
720 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
721 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
722 ldlm_lock_decref_internal(lock, mode);
726 /* This will drop a lock reference and mark it for destruction, but will not
727 * necessarily cancel the lock before returning. */
728 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
730 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
733 LASSERT(lock != NULL);
735 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
736 lock_res_and_lock(lock);
737 lock->l_flags |= LDLM_FL_CBPENDING;
738 unlock_res_and_lock(lock);
739 ldlm_lock_decref_internal(lock, mode);
743 struct sl_insert_point {
744 struct list_head *res_link;
745 struct list_head *mode_link;
746 struct list_head *policy_link;
750 * search_granted_lock
753 * Finds a position to insert the new lock.
755 * queue [input]: the granted list where search acts on;
756 * req [input]: the lock whose position to be located;
757 * prev [output]: positions within 3 lists to insert @req to
761 * - ldlm_grant_lock_with_skiplist
763 static void search_granted_lock(struct list_head *queue,
764 struct ldlm_lock *req,
765 struct sl_insert_point *prev)
767 struct list_head *tmp;
768 struct ldlm_lock *lock, *mode_end, *policy_end;
771 list_for_each(tmp, queue) {
772 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
774 mode_end = list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
777 if (lock->l_req_mode != req->l_req_mode) {
778 /* jump to last lock of mode group */
779 tmp = &mode_end->l_res_link;
783 /* suitable mode group is found */
784 if (lock->l_resource->lr_type == LDLM_PLAIN) {
785 /* insert point is last lock of the mode group */
786 prev->res_link = &mode_end->l_res_link;
787 prev->mode_link = &mode_end->l_sl_mode;
788 prev->policy_link = &req->l_sl_policy;
791 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
793 policy_end = list_entry(lock->l_sl_policy.prev,
797 if (lock->l_policy_data.l_inodebits.bits ==
798 req->l_policy_data.l_inodebits.bits) {
799 /* insert point is last lock of
800 * the policy group */
802 &policy_end->l_res_link;
804 &policy_end->l_sl_mode;
806 &policy_end->l_sl_policy;
811 if (policy_end == mode_end)
812 /* done with mode group */
815 /* go to next policy group within mode group */
816 tmp = policy_end->l_res_link.next;
817 lock = list_entry(tmp, struct ldlm_lock,
819 } /* loop over policy groups within the mode group */
821 /* insert point is last lock of the mode group,
822 * new policy group is started */
823 prev->res_link = &mode_end->l_res_link;
824 prev->mode_link = &mode_end->l_sl_mode;
825 prev->policy_link = &req->l_sl_policy;
829 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
834 /* insert point is last lock on the queue,
835 * new mode group and new policy group are started */
836 prev->res_link = queue->prev;
837 prev->mode_link = &req->l_sl_mode;
838 prev->policy_link = &req->l_sl_policy;
843 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
844 struct sl_insert_point *prev)
846 struct ldlm_resource *res = lock->l_resource;
849 check_res_locked(res);
851 ldlm_resource_dump(D_INFO, res);
852 CDEBUG(D_OTHER, "About to add this lock:\n");
853 ldlm_lock_dump(D_OTHER, lock, 0);
855 if (lock->l_destroyed) {
856 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
860 LASSERT(list_empty(&lock->l_res_link));
861 LASSERT(list_empty(&lock->l_sl_mode));
862 LASSERT(list_empty(&lock->l_sl_policy));
864 list_add(&lock->l_res_link, prev->res_link);
865 list_add(&lock->l_sl_mode, prev->mode_link);
866 list_add(&lock->l_sl_policy, prev->policy_link);
871 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
873 struct sl_insert_point prev;
876 LASSERT(lock->l_req_mode == lock->l_granted_mode);
878 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
879 ldlm_granted_list_add_lock(lock, &prev);
884 * - ldlm_lock_enqueue
885 * - ldlm_reprocess_queue
886 * - ldlm_lock_convert
888 * must be called with lr_lock held
890 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
892 struct ldlm_resource *res = lock->l_resource;
895 check_res_locked(res);
897 lock->l_granted_mode = lock->l_req_mode;
898 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
899 ldlm_grant_lock_with_skiplist(lock);
900 else if (res->lr_type == LDLM_EXTENT)
901 ldlm_extent_add_lock(res, lock);
903 ldlm_resource_add_lock(res, &res->lr_granted, lock);
905 if (lock->l_granted_mode < res->lr_most_restr)
906 res->lr_most_restr = lock->l_granted_mode;
908 if (work_list && lock->l_completion_ast != NULL)
909 ldlm_add_ast_work_item(lock, NULL, work_list);
911 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
915 /* returns a referenced lock or NULL. See the flag descriptions below, in the
916 * comment above ldlm_lock_match */
917 static struct ldlm_lock *search_queue(struct list_head *queue,
919 ldlm_policy_data_t *policy,
920 struct ldlm_lock *old_lock,
921 int flags, int unref)
923 struct ldlm_lock *lock;
924 struct list_head *tmp;
926 list_for_each(tmp, queue) {
929 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
931 if (lock == old_lock)
934 /* llite sometimes wants to match locks that will be
935 * canceled when their users drop, but we allow it to match
936 * if it passes in CBPENDING and the lock still has users.
937 * this is generally only going to be used by children
938 * whose parents already hold a lock so forward progress
939 * can still happen. */
940 if (lock->l_flags & LDLM_FL_CBPENDING &&
941 !(flags & LDLM_FL_CBPENDING))
943 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
944 lock->l_readers == 0 && lock->l_writers == 0)
947 if (!(lock->l_req_mode & *mode))
949 match = lock->l_req_mode;
951 if (lock->l_resource->lr_type == LDLM_EXTENT &&
952 (lock->l_policy_data.l_extent.start >
953 policy->l_extent.start ||
954 lock->l_policy_data.l_extent.end < policy->l_extent.end))
957 if (unlikely(match == LCK_GROUP) &&
958 lock->l_resource->lr_type == LDLM_EXTENT &&
959 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
962 /* We match if we have existing lock with same or wider set
964 if (lock->l_resource->lr_type == LDLM_IBITS &&
965 ((lock->l_policy_data.l_inodebits.bits &
966 policy->l_inodebits.bits) !=
967 policy->l_inodebits.bits))
971 (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED)))
974 if ((flags & LDLM_FL_LOCAL_ONLY) &&
975 !(lock->l_flags & LDLM_FL_LOCAL))
978 if (flags & LDLM_FL_TEST_LOCK) {
980 ldlm_lock_touch_in_lru(lock);
982 ldlm_lock_addref_internal_nolock(lock, match);
991 void ldlm_lock_allow_match(struct ldlm_lock *lock)
993 lock_res_and_lock(lock);
994 lock->l_flags |= LDLM_FL_LVB_READY;
995 cfs_waitq_signal(&lock->l_waitq);
996 unlock_res_and_lock(lock);
999 /* Can be called in two ways:
1001 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1002 * for a duplicate of.
1004 * Otherwise, all of the fields must be filled in, to match against.
1006 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1007 * server (ie, connh is NULL)
1008 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1009 * list will be considered
1010 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1011 * to be canceled can still be matched as long as they still have reader
1012 * or writer refernces
1013 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1014 * just tell us if we would have matched.
1016 * Returns 1 if it finds an already-existing lock that is compatible; in this
1017 * case, lockh is filled in with a addref()ed lock
1019 * we also check security context, if that failed we simply return 0 (to keep
1020 * caller code unchanged), the context failure will be discovered by caller
1023 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1024 const struct ldlm_res_id *res_id, ldlm_type_t type,
1025 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1026 struct lustre_handle *lockh, int unref)
1028 struct ldlm_resource *res;
1029 struct ldlm_lock *lock, *old_lock = NULL;
1034 old_lock = ldlm_handle2lock(lockh);
1037 ns = old_lock->l_resource->lr_namespace;
1038 res_id = &old_lock->l_resource->lr_name;
1039 type = old_lock->l_resource->lr_type;
1040 mode = old_lock->l_req_mode;
1043 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1045 LASSERT(old_lock == NULL);
1049 LDLM_RESOURCE_ADDREF(res);
1052 lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1056 if (flags & LDLM_FL_BLOCK_GRANTED)
1058 lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1062 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1070 LDLM_RESOURCE_DELREF(res);
1071 ldlm_resource_putref(res);
1074 ldlm_lock2handle(lock, lockh);
1075 if ((flags & LDLM_FL_LVB_READY) &&
1076 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1077 struct l_wait_info lwi;
1078 if (lock->l_completion_ast) {
1079 int err = lock->l_completion_ast(lock,
1080 LDLM_FL_WAIT_NOREPROC,
1083 if (flags & LDLM_FL_TEST_LOCK)
1084 LDLM_LOCK_RELEASE(lock);
1086 ldlm_lock_decref_internal(lock,
1093 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1094 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1096 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1097 l_wait_event(lock->l_waitq,
1098 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1103 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1104 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1105 res_id->name[2] : policy->l_extent.start,
1106 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1107 res_id->name[3] : policy->l_extent.end);
1109 /* check user's security context */
1110 if (lock->l_conn_export &&
1111 sptlrpc_import_check_ctx(
1112 class_exp2cliimp(lock->l_conn_export))) {
1113 if (!(flags & LDLM_FL_TEST_LOCK))
1114 ldlm_lock_decref_internal(lock, mode);
1118 if (flags & LDLM_FL_TEST_LOCK)
1119 LDLM_LOCK_RELEASE(lock);
1121 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1122 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1123 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1124 type, mode, res_id->name[0], res_id->name[1],
1125 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1126 res_id->name[2] :policy->l_extent.start,
1127 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1128 res_id->name[3] : policy->l_extent.end);
1131 LDLM_LOCK_PUT(old_lock);
1133 return rc ? mode : 0;
1136 /* Returns a referenced lock */
1137 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1138 const struct ldlm_res_id *res_id,
1141 const struct ldlm_callback_suite *cbs,
1142 void *data, __u32 lvb_len)
1144 struct ldlm_lock *lock;
1145 struct ldlm_resource *res;
1148 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1152 lock = ldlm_lock_new(res);
1153 ldlm_resource_putref(res);
1158 lock->l_req_mode = mode;
1159 lock->l_ast_data = data;
1160 lock->l_pid = cfs_curproc_pid();
1162 lock->l_blocking_ast = cbs->lcs_blocking;
1163 lock->l_completion_ast = cbs->lcs_completion;
1164 lock->l_glimpse_ast = cbs->lcs_glimpse;
1165 lock->l_weigh_ast = cbs->lcs_weigh;
1168 lock->l_tree_node = NULL;
1169 /* if this is the extent lock, allocate the interval tree node */
1170 if (type == LDLM_EXTENT) {
1171 if (ldlm_interval_alloc(lock) == NULL)
1176 lock->l_lvb_len = lvb_len;
1177 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1178 if (lock->l_lvb_data == NULL)
1185 if (lock->l_lvb_data)
1186 OBD_FREE(lock->l_lvb_data, lvb_len);
1187 ldlm_interval_free(ldlm_interval_detach(lock));
1188 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1192 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1193 struct ldlm_lock **lockp,
1194 void *cookie, int *flags)
1196 struct ldlm_lock *lock = *lockp;
1197 struct ldlm_resource *res = lock->l_resource;
1198 int local = ns_is_client(res->lr_namespace);
1199 ldlm_processing_policy policy;
1200 ldlm_error_t rc = ELDLM_OK;
1201 struct ldlm_interval *node = NULL;
1204 do_gettimeofday(&lock->l_enqueued_time);
1205 /* policies are not executed on the client or during replay */
1206 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1207 && !local && ns->ns_policy) {
1208 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1210 if (rc == ELDLM_LOCK_REPLACED) {
1211 /* The lock that was returned has already been granted,
1212 * and placed into lockp. If it's not the same as the
1213 * one we passed in, then destroy the old one and our
1214 * work here is done. */
1215 if (lock != *lockp) {
1216 ldlm_lock_destroy(lock);
1217 LDLM_LOCK_RELEASE(lock);
1219 *flags |= LDLM_FL_LOCK_CHANGED;
1221 } else if (rc != ELDLM_OK ||
1222 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1223 ldlm_lock_destroy(lock);
1228 /* For a replaying lock, it might be already in granted list. So
1229 * unlinking the lock will cause the interval node to be freed, we
1230 * have to allocate the interval node early otherwise we can't regrant
1231 * this lock in the future. - jay */
1232 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1233 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO,
1236 lock_res_and_lock(lock);
1237 if (local && lock->l_req_mode == lock->l_granted_mode) {
1238 /* The server returned a blocked lock, but it was granted
1239 * before we got a chance to actually enqueue it. We don't
1240 * need to do anything else. */
1241 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1242 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1243 GOTO(out, ELDLM_OK);
1246 ldlm_resource_unlink_lock(lock);
1247 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1249 ldlm_lock_destroy_nolock(lock);
1250 GOTO(out, rc = -ENOMEM);
1253 CFS_INIT_LIST_HEAD(&node->li_group);
1254 ldlm_interval_attach(node, lock);
1258 /* Some flags from the enqueue want to make it into the AST, via the
1259 * lock's l_flags. */
1260 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1262 /* This distinction between local lock trees is very important; a client
1263 * namespace only has information about locks taken by that client, and
1264 * thus doesn't have enough information to decide for itself if it can
1265 * be granted (below). In this case, we do exactly what the server
1266 * tells us to do, as dictated by the 'flags'.
1268 * We do exactly the same thing during recovery, when the server is
1269 * more or less trusting the clients not to lie.
1271 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1272 * granted/converting queues. */
1274 if (*flags & LDLM_FL_BLOCK_CONV)
1275 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1276 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1277 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1279 ldlm_grant_lock(lock, NULL);
1280 GOTO(out, ELDLM_OK);
1281 } else if (*flags & LDLM_FL_REPLAY) {
1282 if (*flags & LDLM_FL_BLOCK_CONV) {
1283 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1284 GOTO(out, ELDLM_OK);
1285 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1286 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1287 GOTO(out, ELDLM_OK);
1288 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1289 ldlm_grant_lock(lock, NULL);
1290 GOTO(out, ELDLM_OK);
1292 /* If no flags, fall through to normal enqueue path. */
1295 policy = ldlm_processing_policy_table[res->lr_type];
1296 policy(lock, flags, 1, &rc, NULL);
1299 unlock_res_and_lock(lock);
1301 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1305 /* Must be called with namespace taken: queue is waiting or converting. */
1306 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1307 struct list_head *work_list)
1309 struct list_head *tmp, *pos;
1310 ldlm_processing_policy policy;
1312 int rc = LDLM_ITER_CONTINUE;
1316 check_res_locked(res);
1318 policy = ldlm_processing_policy_table[res->lr_type];
1321 list_for_each_safe(tmp, pos, queue) {
1322 struct ldlm_lock *pending;
1323 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1325 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1328 rc = policy(pending, &flags, 0, &err, work_list);
1329 if (rc != LDLM_ITER_CONTINUE)
1336 /* Helper function for ldlm_run_ast_work().
1338 * Send an existing rpc set specified by @arg->set and then
1339 * destroy it. Create new one if @do_create flag is set. */
1341 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1345 ptlrpc_set_wait(arg->set);
1346 if (arg->type == LDLM_BL_CALLBACK)
1347 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1348 ptlrpc_set_destroy(arg->set);
1351 arg->set = ptlrpc_prep_set();
1357 ldlm_work_bl_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1359 struct ldlm_lock_desc d;
1360 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_bl_ast);
1363 /* nobody should touch l_bl_ast */
1364 lock_res_and_lock(lock);
1365 list_del_init(&lock->l_bl_ast);
1367 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1368 LASSERT(lock->l_bl_ast_run == 0);
1369 LASSERT(lock->l_blocking_lock);
1370 lock->l_bl_ast_run++;
1371 unlock_res_and_lock(lock);
1373 ldlm_lock2desc(lock->l_blocking_lock, &d);
1375 lock->l_blocking_ast(lock, &d, (void *)arg,
1377 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1378 lock->l_blocking_lock = NULL;
1379 LDLM_LOCK_RELEASE(lock);
1385 ldlm_work_cp_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1387 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_cp_ast);
1388 ldlm_completion_callback completion_callback;
1392 /* It's possible to receive a completion AST before we've set
1393 * the l_completion_ast pointer: either because the AST arrived
1394 * before the reply, or simply because there's a small race
1395 * window between receiving the reply and finishing the local
1396 * enqueue. (bug 842)
1398 * This can't happen with the blocking_ast, however, because we
1399 * will never call the local blocking_ast until we drop our
1400 * reader/writer reference, which we won't do until we get the
1401 * reply and finish enqueueing. */
1403 /* nobody should touch l_cp_ast */
1404 lock_res_and_lock(lock);
1405 list_del_init(&lock->l_cp_ast);
1406 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1407 /* save l_completion_ast since it can be changed by
1408 * mds_intent_policy(), see bug 14225 */
1409 completion_callback = lock->l_completion_ast;
1410 lock->l_flags &= ~LDLM_FL_CP_REQD;
1411 unlock_res_and_lock(lock);
1413 if (completion_callback != NULL) {
1414 completion_callback(lock, 0, (void *)arg);
1417 LDLM_LOCK_RELEASE(lock);
1423 ldlm_work_revoke_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1425 struct ldlm_lock_desc desc;
1426 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_rk_ast);
1429 list_del_init(&lock->l_rk_ast);
1431 /* the desc just pretend to exclusive */
1432 ldlm_lock2desc(lock, &desc);
1433 desc.l_req_mode = LCK_EX;
1434 desc.l_granted_mode = 0;
1436 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1437 LDLM_LOCK_RELEASE(lock);
1442 int ldlm_run_ast_work(struct list_head *rpc_list, ldlm_desc_ast_t ast_type)
1444 struct ldlm_cb_set_arg arg;
1445 struct list_head *tmp, *pos;
1446 int (*work_ast_lock)(struct list_head *tmp,struct ldlm_cb_set_arg *arg);
1450 arg.set = ptlrpc_prep_set();
1451 atomic_set(&arg.restart, 0);
1453 case LDLM_WORK_BL_AST:
1454 arg.type = LDLM_BL_CALLBACK;
1455 work_ast_lock = ldlm_work_bl_ast_lock;
1457 case LDLM_WORK_CP_AST:
1458 arg.type = LDLM_CP_CALLBACK;
1459 work_ast_lock = ldlm_work_cp_ast_lock;
1461 case LDLM_WORK_REVOKE_AST:
1462 arg.type = LDLM_BL_CALLBACK;
1463 work_ast_lock = ldlm_work_revoke_ast_lock;
1470 list_for_each_safe(tmp, pos, rpc_list) {
1471 ast_count += work_ast_lock(tmp, &arg);
1473 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1474 * and create a new set for requests that remained in
1476 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1477 ldlm_send_and_maybe_create_set(&arg, 1);
1483 ldlm_send_and_maybe_create_set(&arg, 0);
1485 /* In case when number of ASTs is multiply of
1486 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1487 * @arg.set must be destroyed here, otherwise we get
1488 * write memory leaking. */
1489 ptlrpc_set_destroy(arg.set);
1491 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1494 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1496 ldlm_reprocess_all(res);
1497 return LDLM_ITER_CONTINUE;
1500 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1502 struct list_head *tmp;
1509 spin_lock(&ns->ns_hash_lock);
1510 for (i = 0; i < RES_HASH_SIZE; i++) {
1511 tmp = ns->ns_hash[i].next;
1512 while (tmp != &(ns->ns_hash[i])) {
1513 struct ldlm_resource *res =
1514 list_entry(tmp, struct ldlm_resource, lr_hash);
1516 ldlm_resource_getref(res);
1517 spin_unlock(&ns->ns_hash_lock);
1518 LDLM_RESOURCE_ADDREF(res);
1520 rc = reprocess_one_queue(res, NULL);
1522 LDLM_RESOURCE_DELREF(res);
1523 spin_lock(&ns->ns_hash_lock);
1525 ldlm_resource_putref_locked(res);
1527 if (rc == LDLM_ITER_STOP)
1532 spin_unlock(&ns->ns_hash_lock);
1536 void ldlm_reprocess_all(struct ldlm_resource *res)
1538 CFS_LIST_HEAD(rpc_list);
1542 /* Local lock trees don't get reprocessed. */
1543 if (ns_is_client(res->lr_namespace)) {
1550 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1551 if (rc == LDLM_ITER_CONTINUE)
1552 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1555 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1556 if (rc == -ERESTART) {
1557 LASSERT(list_empty(&rpc_list));
1563 void ldlm_cancel_callback(struct ldlm_lock *lock)
1565 check_res_locked(lock->l_resource);
1566 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1567 lock->l_flags |= LDLM_FL_CANCEL;
1568 if (lock->l_blocking_ast) {
1569 // l_check_no_ns_lock(ns);
1570 unlock_res_and_lock(lock);
1571 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1573 lock_res_and_lock(lock);
1575 LDLM_DEBUG(lock, "no blocking ast");
1578 lock->l_flags |= LDLM_FL_BL_DONE;
1581 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1583 if (req->l_resource->lr_type != LDLM_PLAIN &&
1584 req->l_resource->lr_type != LDLM_IBITS)
1587 list_del_init(&req->l_sl_policy);
1588 list_del_init(&req->l_sl_mode);
1591 void ldlm_lock_cancel(struct ldlm_lock *lock)
1593 struct ldlm_resource *res;
1594 struct ldlm_namespace *ns;
1597 lock_res_and_lock(lock);
1599 res = lock->l_resource;
1600 ns = res->lr_namespace;
1602 /* Please do not, no matter how tempting, remove this LBUG without
1603 * talking to me first. -phik */
1604 if (lock->l_readers || lock->l_writers) {
1605 LDLM_ERROR(lock, "lock still has references");
1609 ldlm_del_waiting_lock(lock);
1611 /* Releases cancel callback. */
1612 ldlm_cancel_callback(lock);
1614 /* Yes, second time, just in case it was added again while we were
1615 running with no res lock in ldlm_cancel_callback */
1616 ldlm_del_waiting_lock(lock);
1617 ldlm_resource_unlink_lock(lock);
1618 ldlm_lock_destroy_nolock(lock);
1620 if (lock->l_granted_mode == lock->l_req_mode)
1621 ldlm_pool_del(&ns->ns_pool, lock);
1623 /* Make sure we will not be called again for same lock what is possible
1624 * if not to zero out lock->l_granted_mode */
1625 lock->l_granted_mode = LCK_MINMODE;
1626 unlock_res_and_lock(lock);
1631 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1633 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1639 lock->l_ast_data = data;
1640 LDLM_LOCK_PUT(lock);
1644 void ldlm_cancel_locks_for_export_cb(void *obj, void *data)
1646 struct obd_export *exp = data;
1647 struct ldlm_lock *lock = obj;
1648 struct ldlm_resource *res;
1650 res = ldlm_resource_getref(lock->l_resource);
1651 LDLM_LOCK_GET(lock);
1653 LDLM_DEBUG(lock, "export %p", exp);
1654 ldlm_res_lvbo_update(res, NULL, 0, 1);
1655 ldlm_lock_cancel(lock);
1656 ldlm_reprocess_all(res);
1657 ldlm_resource_putref(res);
1658 LDLM_LOCK_RELEASE(lock);
1661 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1663 lustre_hash_for_each_empty(exp->exp_lock_hash,
1664 ldlm_cancel_locks_for_export_cb, exp);
1668 * Downgrade an exclusive lock.
1670 * A fast variant of ldlm_lock_convert for convertion of exclusive
1671 * locks. The convertion is always successful.
1673 * \param lock A lock to convert
1674 * \param new_mode new lock mode
1676 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1680 LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1681 LASSERT(new_mode == LCK_COS);
1683 lock_res_and_lock(lock);
1684 ldlm_resource_unlink_lock(lock);
1685 lock->l_req_mode = new_mode;
1686 ldlm_grant_lock(lock, NULL);
1687 unlock_res_and_lock(lock);
1688 ldlm_reprocess_all(lock->l_resource);
1693 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1696 CFS_LIST_HEAD(rpc_list);
1697 struct ldlm_resource *res;
1698 struct ldlm_namespace *ns;
1701 struct sl_insert_point prev;
1703 struct ldlm_interval *node;
1706 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1707 *flags |= LDLM_FL_BLOCK_GRANTED;
1708 RETURN(lock->l_resource);
1711 /* I can't check the type of lock here because the bitlock of lock
1712 * is not held here, so do the allocation blindly. -jay */
1713 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO, sizeof(*node));
1714 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1717 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1718 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1720 lock_res_and_lock(lock);
1722 res = lock->l_resource;
1723 ns = res->lr_namespace;
1725 old_mode = lock->l_req_mode;
1726 lock->l_req_mode = new_mode;
1727 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1728 /* remember the lock position where the lock might be
1729 * added back to the granted list later and also
1730 * remember the join mode for skiplist fixing. */
1731 prev.res_link = lock->l_res_link.prev;
1732 prev.mode_link = lock->l_sl_mode.prev;
1733 prev.policy_link = lock->l_sl_policy.prev;
1734 ldlm_resource_unlink_lock(lock);
1736 ldlm_resource_unlink_lock(lock);
1737 if (res->lr_type == LDLM_EXTENT) {
1738 /* FIXME: ugly code, I have to attach the lock to a
1739 * interval node again since perhaps it will be granted
1741 CFS_INIT_LIST_HEAD(&node->li_group);
1742 ldlm_interval_attach(node, lock);
1747 /* If this is a local resource, put it on the appropriate list. */
1748 if (ns_is_client(res->lr_namespace)) {
1749 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1750 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1752 /* This should never happen, because of the way the
1753 * server handles conversions. */
1754 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1758 ldlm_grant_lock(lock, &rpc_list);
1760 /* FIXME: completion handling not with ns_lock held ! */
1761 if (lock->l_completion_ast)
1762 lock->l_completion_ast(lock, 0, NULL);
1766 ldlm_processing_policy policy;
1767 policy = ldlm_processing_policy_table[res->lr_type];
1768 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1769 if (rc == LDLM_ITER_STOP) {
1770 lock->l_req_mode = old_mode;
1771 if (res->lr_type == LDLM_EXTENT)
1772 ldlm_extent_add_lock(res, lock);
1774 ldlm_granted_list_add_lock(lock, &prev);
1778 *flags |= LDLM_FL_BLOCK_GRANTED;
1782 unlock_res_and_lock(lock);
1785 ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1787 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1791 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1793 struct obd_device *obd = NULL;
1795 if (!((libcfs_debug | D_ERROR) & level))
1799 CDEBUG(level, " NULL LDLM lock\n");
1803 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1804 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1806 if (lock->l_conn_export != NULL)
1807 obd = lock->l_conn_export->exp_obd;
1808 if (lock->l_export && lock->l_export->exp_connection) {
1809 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1810 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1811 lock->l_remote_handle.cookie);
1812 } else if (obd == NULL) {
1813 CDEBUG(level, " Node: local\n");
1815 struct obd_import *imp = obd->u.cli.cl_import;
1816 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1817 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1818 lock->l_remote_handle.cookie);
1820 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1822 lock->l_resource->lr_name.name[0],
1823 lock->l_resource->lr_name.name[1],
1824 lock->l_resource->lr_name.name[2]);
1825 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1826 "write: %d flags: %#x\n", ldlm_lockname[lock->l_req_mode],
1827 ldlm_lockname[lock->l_granted_mode],
1828 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1830 if (lock->l_resource->lr_type == LDLM_EXTENT)
1831 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1832 " (req "LPU64"-"LPU64")\n",
1833 lock->l_policy_data.l_extent.start,
1834 lock->l_policy_data.l_extent.end,
1835 lock->l_req_extent.start, lock->l_req_extent.end);
1836 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1837 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1838 lock->l_policy_data.l_flock.pid,
1839 lock->l_policy_data.l_flock.start,
1840 lock->l_policy_data.l_flock.end);
1841 else if (lock->l_resource->lr_type == LDLM_IBITS)
1842 CDEBUG(level, " Bits: "LPX64"\n",
1843 lock->l_policy_data.l_inodebits.bits);
1846 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1848 struct ldlm_lock *lock;
1850 if (!((libcfs_debug | D_ERROR) & level))
1853 lock = ldlm_handle2lock(lockh);
1857 ldlm_lock_dump(D_OTHER, lock, 0);
1859 LDLM_LOCK_PUT(lock);
1862 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1863 struct libcfs_debug_msg_data *data, const char *fmt,
1867 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1869 va_start(args, fmt);
1871 if (lock->l_resource == NULL) {
1872 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1873 data->msg_fn, data->msg_line, fmt, args,
1874 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1875 "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "
1876 LPX64" expref: %d pid: %u\n", lock,
1877 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1878 lock->l_readers, lock->l_writers,
1879 ldlm_lockname[lock->l_granted_mode],
1880 ldlm_lockname[lock->l_req_mode],
1881 lock->l_flags, lock->l_remote_handle.cookie,
1883 atomic_read(&lock->l_export->exp_refcount) : -99,
1889 switch (lock->l_resource->lr_type) {
1891 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1892 data->msg_fn, data->msg_line, fmt, args,
1893 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1894 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1895 "] (req "LPU64"->"LPU64") flags: %x remote: "LPX64
1896 " expref: %d pid: %u\n",
1897 lock->l_resource->lr_namespace->ns_name, lock,
1898 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1899 lock->l_readers, lock->l_writers,
1900 ldlm_lockname[lock->l_granted_mode],
1901 ldlm_lockname[lock->l_req_mode],
1902 lock->l_resource->lr_name.name[0],
1903 lock->l_resource->lr_name.name[1],
1904 atomic_read(&lock->l_resource->lr_refcount),
1905 ldlm_typename[lock->l_resource->lr_type],
1906 lock->l_policy_data.l_extent.start,
1907 lock->l_policy_data.l_extent.end,
1908 lock->l_req_extent.start, lock->l_req_extent.end,
1909 lock->l_flags, lock->l_remote_handle.cookie,
1911 atomic_read(&lock->l_export->exp_refcount) : -99,
1916 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1917 data->msg_fn, data->msg_line, fmt, args,
1918 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1919 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1920 "["LPU64"->"LPU64"] flags: %x remote: "LPX64
1921 " expref: %d pid: %u\n",
1922 lock->l_resource->lr_namespace->ns_name, lock,
1923 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1924 lock->l_readers, lock->l_writers,
1925 ldlm_lockname[lock->l_granted_mode],
1926 ldlm_lockname[lock->l_req_mode],
1927 lock->l_resource->lr_name.name[0],
1928 lock->l_resource->lr_name.name[1],
1929 atomic_read(&lock->l_resource->lr_refcount),
1930 ldlm_typename[lock->l_resource->lr_type],
1931 lock->l_policy_data.l_flock.pid,
1932 lock->l_policy_data.l_flock.start,
1933 lock->l_policy_data.l_flock.end,
1934 lock->l_flags, lock->l_remote_handle.cookie,
1936 atomic_read(&lock->l_export->exp_refcount) : -99,
1941 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1942 data->msg_fn, data->msg_line, fmt, args,
1943 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1944 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1945 "flags: %x remote: "LPX64" expref: %d "
1947 lock->l_resource->lr_namespace->ns_name,
1948 lock, lock->l_handle.h_cookie,
1949 atomic_read (&lock->l_refc),
1950 lock->l_readers, lock->l_writers,
1951 ldlm_lockname[lock->l_granted_mode],
1952 ldlm_lockname[lock->l_req_mode],
1953 lock->l_resource->lr_name.name[0],
1954 lock->l_resource->lr_name.name[1],
1955 lock->l_policy_data.l_inodebits.bits,
1956 atomic_read(&lock->l_resource->lr_refcount),
1957 ldlm_typename[lock->l_resource->lr_type],
1958 lock->l_flags, lock->l_remote_handle.cookie,
1960 atomic_read(&lock->l_export->exp_refcount) : -99,
1965 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1966 data->msg_fn, data->msg_line, fmt, args,
1967 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1968 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: %x "
1969 "remote: "LPX64" expref: %d pid: %u\n",
1970 lock->l_resource->lr_namespace->ns_name,
1971 lock, lock->l_handle.h_cookie,
1972 atomic_read (&lock->l_refc),
1973 lock->l_readers, lock->l_writers,
1974 ldlm_lockname[lock->l_granted_mode],
1975 ldlm_lockname[lock->l_req_mode],
1976 lock->l_resource->lr_name.name[0],
1977 lock->l_resource->lr_name.name[1],
1978 atomic_read(&lock->l_resource->lr_refcount),
1979 ldlm_typename[lock->l_resource->lr_type],
1980 lock->l_flags, lock->l_remote_handle.cookie,
1982 atomic_read(&lock->l_export->exp_refcount) : -99,
1988 EXPORT_SYMBOL(_ldlm_lock_debug);