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[] = {
68 char *ldlm_typename[] = {
75 char *ldlm_it2str(int it)
82 case (IT_OPEN | IT_CREAT):
95 CERROR("Unknown intent %d\n", it);
100 extern cfs_mem_cache_t *ldlm_lock_slab;
102 static ldlm_processing_policy ldlm_processing_policy_table[] = {
103 [LDLM_PLAIN] ldlm_process_plain_lock,
104 [LDLM_EXTENT] ldlm_process_extent_lock,
106 [LDLM_FLOCK] ldlm_process_flock_lock,
108 [LDLM_IBITS] ldlm_process_inodebits_lock,
111 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
113 return ldlm_processing_policy_table[res->lr_type];
116 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
122 * REFCOUNTED LOCK OBJECTS
127 * Lock refcounts, during creation:
128 * - one special one for allocation, dec'd only once in destroy
129 * - one for being a lock that's in-use
130 * - one for the addref associated with a new lock
132 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
134 atomic_inc(&lock->l_refc);
138 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
140 LASSERT(size == sizeof(*lock));
141 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
144 void ldlm_lock_put(struct ldlm_lock *lock)
148 LASSERT(lock->l_resource != LP_POISON);
149 LASSERT(atomic_read(&lock->l_refc) > 0);
150 if (atomic_dec_and_test(&lock->l_refc)) {
151 struct ldlm_resource *res;
154 "final lock_put on destroyed lock, freeing it.");
156 res = lock->l_resource;
157 LASSERT(lock->l_destroyed);
158 LASSERT(list_empty(&lock->l_res_link));
159 LASSERT(list_empty(&lock->l_pending_chain));
161 atomic_dec(&res->lr_namespace->ns_locks);
162 ldlm_resource_putref(res);
163 lock->l_resource = NULL;
164 if (lock->l_export) {
165 class_export_put(lock->l_export);
166 lock->l_export = NULL;
169 if (lock->l_lvb_data != NULL)
170 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
172 ldlm_interval_free(ldlm_interval_detach(lock));
173 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
180 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
183 if (!list_empty(&lock->l_lru)) {
184 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
185 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
186 list_del_init(&lock->l_lru);
188 LASSERT(ns->ns_nr_unused >= 0);
194 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
196 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
199 spin_lock(&ns->ns_unused_lock);
200 rc = ldlm_lock_remove_from_lru_nolock(lock);
201 spin_unlock(&ns->ns_unused_lock);
206 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
208 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
209 lock->l_last_used = cfs_time_current();
210 LASSERT(list_empty(&lock->l_lru));
211 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
212 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
213 LASSERT(ns->ns_nr_unused >= 0);
217 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
219 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
221 spin_lock(&ns->ns_unused_lock);
222 ldlm_lock_add_to_lru_nolock(lock);
223 spin_unlock(&ns->ns_unused_lock);
227 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
229 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
231 spin_lock(&ns->ns_unused_lock);
232 if (!list_empty(&lock->l_lru)) {
233 ldlm_lock_remove_from_lru_nolock(lock);
234 ldlm_lock_add_to_lru_nolock(lock);
236 spin_unlock(&ns->ns_unused_lock);
240 /* This used to have a 'strict' flag, which recovery would use to mark an
241 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
242 * shall explain why it's gone: with the new hash table scheme, once you call
243 * ldlm_lock_destroy, you can never drop your final references on this lock.
244 * Because it's not in the hash table anymore. -phil */
245 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
249 if (lock->l_readers || lock->l_writers) {
250 LDLM_ERROR(lock, "lock still has references");
251 ldlm_lock_dump(D_ERROR, lock, 0);
255 if (!list_empty(&lock->l_res_link)) {
256 LDLM_ERROR(lock, "lock still on resource");
257 ldlm_lock_dump(D_ERROR, lock, 0);
261 if (lock->l_destroyed) {
262 LASSERT(list_empty(&lock->l_lru));
266 lock->l_destroyed = 1;
268 if (lock->l_export && lock->l_export->exp_lock_hash)
269 lustre_hash_del(lock->l_export->exp_lock_hash,
270 &lock->l_remote_handle, &lock->l_exp_hash);
272 ldlm_lock_remove_from_lru(lock);
273 class_handle_unhash(&lock->l_handle);
276 /* Wake anyone waiting for this lock */
277 /* FIXME: I should probably add yet another flag, instead of using
278 * l_export to only call this on clients */
280 class_export_put(lock->l_export);
281 lock->l_export = NULL;
282 if (lock->l_export && lock->l_completion_ast)
283 lock->l_completion_ast(lock, 0);
289 void ldlm_lock_destroy(struct ldlm_lock *lock)
293 lock_res_and_lock(lock);
294 first = ldlm_lock_destroy_internal(lock);
295 unlock_res_and_lock(lock);
297 /* drop reference from hashtable only for first destroy */
303 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
307 first = ldlm_lock_destroy_internal(lock);
308 /* drop reference from hashtable only for first destroy */
314 /* this is called by portals_handle2object with the handle lock taken */
315 static void lock_handle_addref(void *lock)
317 LDLM_LOCK_GET((struct ldlm_lock *)lock);
321 * usage: pass in a resource on which you have done ldlm_resource_get
322 * pass in a parent lock on which you have done a ldlm_lock_get
323 * after return, ldlm_*_put the resource and parent
324 * returns: lock with refcount 2 - one for current caller and one for remote
326 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
328 struct ldlm_lock *lock;
331 if (resource == NULL)
334 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, CFS_ALLOC_IO, sizeof(*lock));
338 spin_lock_init(&lock->l_lock);
339 lock->l_resource = ldlm_resource_getref(resource);
341 atomic_set(&lock->l_refc, 2);
342 CFS_INIT_LIST_HEAD(&lock->l_res_link);
343 CFS_INIT_LIST_HEAD(&lock->l_lru);
344 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
345 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
346 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
347 CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
348 cfs_waitq_init(&lock->l_waitq);
349 lock->l_blocking_lock = NULL;
350 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
351 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
352 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
354 atomic_inc(&resource->lr_namespace->ns_locks);
355 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
356 class_handle_hash(&lock->l_handle, lock_handle_addref);
358 CFS_INIT_LIST_HEAD(&lock->l_extents_list);
359 spin_lock_init(&lock->l_extents_list_lock);
360 CFS_INIT_LIST_HEAD(&lock->l_cache_locks_list);
365 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
366 const struct ldlm_res_id *new_resid)
368 struct ldlm_resource *oldres = lock->l_resource;
369 struct ldlm_resource *newres;
373 LASSERT(ns_is_client(ns));
375 lock_res_and_lock(lock);
376 if (memcmp(new_resid, &lock->l_resource->lr_name,
377 sizeof(lock->l_resource->lr_name)) == 0) {
379 unlock_res_and_lock(lock);
383 LASSERT(new_resid->name[0] != 0);
385 /* This function assumes that the lock isn't on any lists */
386 LASSERT(list_empty(&lock->l_res_link));
388 type = oldres->lr_type;
389 unlock_res_and_lock(lock);
391 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
395 lock_res_and_lock(lock);
396 LASSERT(memcmp(new_resid, &lock->l_resource->lr_name,
397 sizeof(lock->l_resource->lr_name)) != 0);
399 lock->l_resource = newres;
401 unlock_res_and_lock(lock);
403 /* ...and the flowers are still standing! */
404 ldlm_resource_putref(oldres);
413 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
415 lockh->cookie = lock->l_handle.h_cookie;
418 /* if flags: atomically get the lock and set the flags.
419 * Return NULL if flag already set
422 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
425 struct ldlm_namespace *ns;
426 struct ldlm_lock *lock, *retval = NULL;
431 lock = class_handle2object(handle->cookie);
435 LASSERT(lock->l_resource != NULL);
436 ns = lock->l_resource->lr_namespace;
439 lock_res_and_lock(lock);
441 /* It's unlikely but possible that someone marked the lock as
442 * destroyed after we did handle2object on it */
443 if (lock->l_destroyed) {
444 unlock_res_and_lock(lock);
445 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
450 if (flags && (lock->l_flags & flags)) {
451 unlock_res_and_lock(lock);
457 lock->l_flags |= flags;
459 unlock_res_and_lock(lock);
466 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
467 const struct lustre_handle *handle)
469 struct ldlm_lock *retval = NULL;
470 retval = __ldlm_handle2lock(handle, 0);
474 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
476 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
477 /* INODEBITS_INTEROP: If the other side does not support
478 * inodebits, reply with a plain lock descriptor.
480 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
481 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
482 struct ldlm_resource res = *lock->l_resource;
484 /* Make sure all the right bits are set in this lock we
485 are going to pass to client */
486 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
487 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
488 "Inappropriate inode lock bits during "
489 "conversion " LPU64 "\n",
490 lock->l_policy_data.l_inodebits.bits);
491 res.lr_type = LDLM_PLAIN;
492 ldlm_res2desc(&res, &desc->l_resource);
493 /* Convert "new" lock mode to something old client can
495 if ((lock->l_req_mode == LCK_CR) ||
496 (lock->l_req_mode == LCK_CW))
497 desc->l_req_mode = LCK_PR;
499 desc->l_req_mode = lock->l_req_mode;
500 if ((lock->l_granted_mode == LCK_CR) ||
501 (lock->l_granted_mode == LCK_CW)) {
502 desc->l_granted_mode = LCK_PR;
504 /* We never grant PW/EX locks to clients */
505 LASSERT((lock->l_granted_mode != LCK_PW) &&
506 (lock->l_granted_mode != LCK_EX));
507 desc->l_granted_mode = lock->l_granted_mode;
510 /* We do not copy policy here, because there is no
511 policy for plain locks */
513 ldlm_res2desc(lock->l_resource, &desc->l_resource);
514 desc->l_req_mode = lock->l_req_mode;
515 desc->l_granted_mode = lock->l_granted_mode;
516 desc->l_policy_data = lock->l_policy_data;
520 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
521 struct list_head *work_list)
523 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
524 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
525 lock->l_flags |= LDLM_FL_AST_SENT;
526 /* If the enqueuing client said so, tell the AST recipient to
527 * discard dirty data, rather than writing back. */
528 if (new->l_flags & LDLM_AST_DISCARD_DATA)
529 lock->l_flags |= LDLM_FL_DISCARD_DATA;
530 LASSERT(list_empty(&lock->l_bl_ast));
531 list_add(&lock->l_bl_ast, work_list);
533 LASSERT(lock->l_blocking_lock == NULL);
534 lock->l_blocking_lock = LDLM_LOCK_GET(new);
538 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
540 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
541 lock->l_flags |= LDLM_FL_CP_REQD;
542 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
543 LASSERT(list_empty(&lock->l_cp_ast));
544 list_add(&lock->l_cp_ast, work_list);
549 /* must be called with lr_lock held */
550 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
551 struct list_head *work_list)
554 check_res_locked(lock->l_resource);
556 ldlm_add_bl_work_item(lock, new, work_list);
558 ldlm_add_cp_work_item(lock, work_list);
562 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
564 struct ldlm_lock *lock;
566 lock = ldlm_handle2lock(lockh);
567 LASSERT(lock != NULL);
568 ldlm_lock_addref_internal(lock, mode);
572 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
574 ldlm_lock_remove_from_lru(lock);
575 if (mode & (LCK_NL | LCK_CR | LCK_PR))
577 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
580 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
583 /* only called for local locks */
584 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
586 lock_res_and_lock(lock);
587 ldlm_lock_addref_internal_nolock(lock, mode);
588 unlock_res_and_lock(lock);
591 /* only called in ldlm_flock_destroy and for local locks.
592 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
593 * * ldlm_lock_remove_from_lru() does nothing, it is safe
594 * * for ldlm_flock_destroy usage by dropping some code */
595 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
597 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
598 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
599 LASSERT(lock->l_readers > 0);
602 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
603 LASSERT(lock->l_writers > 0);
607 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
610 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
612 struct ldlm_namespace *ns;
615 lock_res_and_lock(lock);
617 ns = lock->l_resource->lr_namespace;
619 ldlm_lock_decref_internal_nolock(lock, mode);
621 if (lock->l_flags & LDLM_FL_LOCAL &&
622 !lock->l_readers && !lock->l_writers) {
623 /* If this is a local lock on a server namespace and this was
624 * the last reference, cancel the lock. */
625 CDEBUG(D_INFO, "forcing cancel of local lock\n");
626 lock->l_flags |= LDLM_FL_CBPENDING;
629 if (!lock->l_readers && !lock->l_writers &&
630 (lock->l_flags & LDLM_FL_CBPENDING)) {
631 /* If we received a blocked AST and this was the last reference,
632 * run the callback. */
633 if (ns_is_server(ns) && lock->l_export)
634 CERROR("FL_CBPENDING set on non-local lock--just a "
637 LDLM_DEBUG(lock, "final decref done on cbpending lock");
639 LDLM_LOCK_GET(lock); /* dropped by bl thread */
640 ldlm_lock_remove_from_lru(lock);
641 unlock_res_and_lock(lock);
642 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
643 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
644 ldlm_handle_bl_callback(ns, NULL, lock);
645 } else if (ns_is_client(ns) &&
646 !lock->l_readers && !lock->l_writers &&
647 !(lock->l_flags & LDLM_FL_NO_LRU) &&
648 !(lock->l_flags & LDLM_FL_BL_AST)) {
649 /* If this is a client-side namespace and this was the last
650 * reference, put it on the LRU. */
651 ldlm_lock_add_to_lru(lock);
652 unlock_res_and_lock(lock);
653 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
654 * are not supported by the server, otherwise, it is done on
656 if (!exp_connect_cancelset(lock->l_conn_export) &&
657 !ns_connect_lru_resize(ns))
658 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
660 unlock_res_and_lock(lock);
666 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
668 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
669 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
670 ldlm_lock_decref_internal(lock, mode);
674 /* This will drop a lock reference and mark it for destruction, but will not
675 * necessarily cancel the lock before returning. */
676 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
678 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
681 LASSERT(lock != NULL);
683 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
684 lock_res_and_lock(lock);
685 lock->l_flags |= LDLM_FL_CBPENDING;
686 unlock_res_and_lock(lock);
687 ldlm_lock_decref_internal(lock, mode);
691 struct sl_insert_point {
692 struct list_head *res_link;
693 struct list_head *mode_link;
694 struct list_head *policy_link;
698 * search_granted_lock
701 * Finds a position to insert the new lock.
703 * queue [input]: the granted list where search acts on;
704 * req [input]: the lock whose position to be located;
705 * prev [output]: positions within 3 lists to insert @req to
709 * - ldlm_grant_lock_with_skiplist
711 static void search_granted_lock(struct list_head *queue,
712 struct ldlm_lock *req,
713 struct sl_insert_point *prev)
715 struct list_head *tmp;
716 struct ldlm_lock *lock, *mode_end, *policy_end;
719 list_for_each(tmp, queue) {
720 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
722 mode_end = list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
725 if (lock->l_req_mode != req->l_req_mode) {
726 /* jump to last lock of mode group */
727 tmp = &mode_end->l_res_link;
731 /* suitable mode group is found */
732 if (lock->l_resource->lr_type == LDLM_PLAIN) {
733 /* insert point is last lock of the mode group */
734 prev->res_link = &mode_end->l_res_link;
735 prev->mode_link = &mode_end->l_sl_mode;
736 prev->policy_link = &req->l_sl_policy;
739 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
741 policy_end = list_entry(lock->l_sl_policy.prev,
745 if (lock->l_policy_data.l_inodebits.bits ==
746 req->l_policy_data.l_inodebits.bits) {
747 /* insert point is last lock of
748 * the policy group */
750 &policy_end->l_res_link;
752 &policy_end->l_sl_mode;
754 &policy_end->l_sl_policy;
759 if (policy_end == mode_end)
760 /* done with mode group */
763 /* jump to next policy group within the mode group */
764 tmp = policy_end->l_res_link.next;
765 lock = list_entry(tmp, struct ldlm_lock,
767 } /* loop over policy groups within the mode group */
769 /* insert point is last lock of the mode group,
770 * new policy group is started */
771 prev->res_link = &mode_end->l_res_link;
772 prev->mode_link = &mode_end->l_sl_mode;
773 prev->policy_link = &req->l_sl_policy;
777 LDLM_ERROR(lock, "is not LDLM_PLAIN or LDLM_IBITS lock");
782 /* insert point is last lock on the queue,
783 * new mode group and new policy group are started */
784 prev->res_link = queue->prev;
785 prev->mode_link = &req->l_sl_mode;
786 prev->policy_link = &req->l_sl_policy;
791 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
792 struct sl_insert_point *prev)
794 struct ldlm_resource *res = lock->l_resource;
797 check_res_locked(res);
799 ldlm_resource_dump(D_INFO, res);
800 CDEBUG(D_OTHER, "About to add this lock:\n");
801 ldlm_lock_dump(D_OTHER, lock, 0);
803 if (lock->l_destroyed) {
804 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
808 LASSERT(list_empty(&lock->l_res_link));
809 LASSERT(list_empty(&lock->l_sl_mode));
810 LASSERT(list_empty(&lock->l_sl_policy));
812 list_add(&lock->l_res_link, prev->res_link);
813 list_add(&lock->l_sl_mode, prev->mode_link);
814 list_add(&lock->l_sl_policy, prev->policy_link);
819 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
821 struct sl_insert_point prev;
824 LASSERT(lock->l_req_mode == lock->l_granted_mode);
826 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
827 ldlm_granted_list_add_lock(lock, &prev);
832 * - ldlm_lock_enqueue
833 * - ldlm_reprocess_queue
834 * - ldlm_lock_convert
836 * must be called with lr_lock held
838 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
840 struct ldlm_resource *res = lock->l_resource;
843 check_res_locked(res);
845 lock->l_granted_mode = lock->l_req_mode;
846 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
847 ldlm_grant_lock_with_skiplist(lock);
848 else if (res->lr_type == LDLM_EXTENT)
849 ldlm_extent_add_lock(res, lock);
851 ldlm_resource_add_lock(res, &res->lr_granted, lock);
853 if (lock->l_granted_mode < res->lr_most_restr)
854 res->lr_most_restr = lock->l_granted_mode;
856 if (work_list && lock->l_completion_ast != NULL)
857 ldlm_add_ast_work_item(lock, NULL, work_list);
859 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
863 /* returns a referenced lock or NULL. See the flag descriptions below, in the
864 * comment above ldlm_lock_match */
865 static struct ldlm_lock *search_queue(struct list_head *queue,
867 ldlm_policy_data_t *policy,
868 struct ldlm_lock *old_lock, int flags)
870 struct ldlm_lock *lock;
871 struct list_head *tmp;
873 list_for_each(tmp, queue) {
876 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
878 if (lock == old_lock)
881 /* llite sometimes wants to match locks that will be
882 * canceled when their users drop, but we allow it to match
883 * if it passes in CBPENDING and the lock still has users.
884 * this is generally only going to be used by children
885 * whose parents already hold a lock so forward progress
886 * can still happen. */
887 if (lock->l_flags & LDLM_FL_CBPENDING &&
888 !(flags & LDLM_FL_CBPENDING))
890 if (lock->l_flags & LDLM_FL_CBPENDING &&
891 lock->l_readers == 0 && lock->l_writers == 0)
894 if (!(lock->l_req_mode & *mode))
896 match = lock->l_req_mode;
898 if (lock->l_resource->lr_type == LDLM_EXTENT &&
899 (lock->l_policy_data.l_extent.start >
900 policy->l_extent.start ||
901 lock->l_policy_data.l_extent.end < policy->l_extent.end))
904 if (unlikely(match == LCK_GROUP) &&
905 lock->l_resource->lr_type == LDLM_EXTENT &&
906 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
909 /* We match if we have existing lock with same or wider set
911 if (lock->l_resource->lr_type == LDLM_IBITS &&
912 ((lock->l_policy_data.l_inodebits.bits &
913 policy->l_inodebits.bits) !=
914 policy->l_inodebits.bits))
917 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
920 if ((flags & LDLM_FL_LOCAL_ONLY) &&
921 !(lock->l_flags & LDLM_FL_LOCAL))
924 if (flags & LDLM_FL_TEST_LOCK) {
926 ldlm_lock_touch_in_lru(lock);
928 ldlm_lock_addref_internal_nolock(lock, match);
937 void ldlm_lock_allow_match(struct ldlm_lock *lock)
939 lock_res_and_lock(lock);
940 lock->l_flags |= LDLM_FL_LVB_READY;
941 cfs_waitq_signal(&lock->l_waitq);
942 unlock_res_and_lock(lock);
946 * Checks if requested extent lock is compatible with another owned lock.
948 * Checks if \a lock is compatible with a read or write lock
949 * (specified by \a rw) for an extent [\a start , \a end].
951 * \param lock the already owned lock
952 * \param rw OBD_BRW_READ if requested for reading,
953 * OBD_BRW_WRITE if requested for writing
954 * \param start start of the requested extent
955 * \param end end of the requested extent
956 * \param cookie transparent parameter for passing locking context
958 * \post result == 1, *cookie == context, appropriate lock is referenced
960 * \retval 1 owned lock is reused for the request
961 * \retval 0 no lock reused for the request
963 * \see ldlm_lock_fast_release
965 int ldlm_lock_fast_match(struct ldlm_lock *lock, int rw,
966 obd_off start, obd_off end,
969 LASSERT(rw == OBD_BRW_READ || rw == OBD_BRW_WRITE);
974 lock_res_and_lock(lock);
975 /* check if granted mode is compatible */
976 if (rw == OBD_BRW_WRITE &&
977 !(lock->l_granted_mode & (LCK_PW|LCK_GROUP)))
980 /* does the lock cover the region we would like to access? */
981 if ((lock->l_policy_data.l_extent.start > start) ||
982 (lock->l_policy_data.l_extent.end < end))
985 /* if we received a blocking callback and the lock is no longer
986 * referenced, don't use it */
987 if ((lock->l_flags & LDLM_FL_CBPENDING) &&
988 !lock->l_writers && !lock->l_readers)
991 ldlm_lock_addref_internal_nolock(lock, rw == OBD_BRW_WRITE ? LCK_PW : LCK_PR);
992 unlock_res_and_lock(lock);
993 *cookie = (void *)lock;
994 return 1; /* avoid using rc for stack relief */
997 unlock_res_and_lock(lock);
1002 * Releases a reference to a lock taken in a "fast" way.
1004 * Releases a read or write (specified by \a rw) lock
1005 * referenced by \a cookie.
1007 * \param rw OBD_BRW_READ if requested for reading,
1008 * OBD_BRW_WRITE if requested for writing
1009 * \param cookie transparent parameter for passing locking context
1011 * \post appropriate lock is dereferenced
1013 * \see ldlm_lock_fast_lock
1015 void ldlm_lock_fast_release(void *cookie, int rw)
1017 struct ldlm_lock *lock = (struct ldlm_lock *)cookie;
1019 LASSERT(lock != NULL);
1020 LASSERT(rw == OBD_BRW_READ || rw == OBD_BRW_WRITE);
1021 LASSERT(rw == OBD_BRW_READ ||
1022 (lock->l_granted_mode & (LCK_PW | LCK_GROUP)));
1023 ldlm_lock_decref_internal(lock, rw == OBD_BRW_WRITE ? LCK_PW : LCK_PR);
1026 /* Can be called in two ways:
1028 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1029 * for a duplicate of.
1031 * Otherwise, all of the fields must be filled in, to match against.
1033 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1034 * server (ie, connh is NULL)
1035 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1036 * list will be considered
1037 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1038 * to be canceled can still be matched as long as they still have reader
1039 * or writer refernces
1040 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1041 * just tell us if we would have matched.
1043 * Returns 1 if it finds an already-existing lock that is compatible; in this
1044 * case, lockh is filled in with a addref()ed lock
1046 * we also check security context, if that failed we simply return 0 (to keep
1047 * caller code unchanged), the context failure will be discovered by caller
1050 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1051 const struct ldlm_res_id *res_id, ldlm_type_t type,
1052 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1053 struct lustre_handle *lockh)
1055 struct ldlm_resource *res;
1056 struct ldlm_lock *lock, *old_lock = NULL;
1061 old_lock = ldlm_handle2lock(lockh);
1064 ns = old_lock->l_resource->lr_namespace;
1065 res_id = &old_lock->l_resource->lr_name;
1066 type = old_lock->l_resource->lr_type;
1067 mode = old_lock->l_req_mode;
1070 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1072 LASSERT(old_lock == NULL);
1078 lock = search_queue(&res->lr_granted, &mode, policy, old_lock, flags);
1081 if (flags & LDLM_FL_BLOCK_GRANTED)
1083 lock = search_queue(&res->lr_converting, &mode, policy, old_lock, flags);
1086 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock, flags);
1093 ldlm_resource_putref(res);
1096 ldlm_lock2handle(lock, lockh);
1097 if ((flags & LDLM_FL_LVB_READY) &&
1098 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1099 struct l_wait_info lwi;
1100 if (lock->l_completion_ast) {
1101 int err = lock->l_completion_ast(lock,
1102 LDLM_FL_WAIT_NOREPROC,
1105 if (flags & LDLM_FL_TEST_LOCK)
1106 LDLM_LOCK_PUT(lock);
1108 ldlm_lock_decref_internal(lock, mode);
1114 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), NULL,
1115 LWI_ON_SIGNAL_NOOP, NULL);
1117 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1118 l_wait_event(lock->l_waitq,
1119 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1124 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
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);
1130 /* check user's security context */
1131 if (lock->l_conn_export &&
1132 sptlrpc_import_check_ctx(
1133 class_exp2cliimp(lock->l_conn_export))) {
1134 if (!(flags & LDLM_FL_TEST_LOCK))
1135 ldlm_lock_decref_internal(lock, mode);
1139 if (flags & LDLM_FL_TEST_LOCK)
1140 LDLM_LOCK_PUT(lock);
1142 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1143 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1144 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1145 type, mode, res_id->name[0], res_id->name[1],
1146 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1147 res_id->name[2] :policy->l_extent.start,
1148 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1149 res_id->name[3] : policy->l_extent.end);
1152 LDLM_LOCK_PUT(old_lock);
1154 return rc ? mode : 0;
1157 /* Returns a referenced lock */
1158 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1159 const struct ldlm_res_id *res_id,
1162 ldlm_blocking_callback blocking,
1163 ldlm_completion_callback completion,
1164 ldlm_glimpse_callback glimpse,
1165 void *data, __u32 lvb_len)
1167 struct ldlm_lock *lock;
1168 struct ldlm_resource *res;
1171 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1175 lock = ldlm_lock_new(res);
1176 ldlm_resource_putref(res);
1181 lock->l_req_mode = mode;
1182 lock->l_ast_data = data;
1183 lock->l_blocking_ast = blocking;
1184 lock->l_completion_ast = completion;
1185 lock->l_glimpse_ast = glimpse;
1186 lock->l_pid = cfs_curproc_pid();
1188 lock->l_tree_node = NULL;
1189 /* if this is the extent lock, allocate the interval tree node */
1190 if (type == LDLM_EXTENT) {
1191 if (ldlm_interval_alloc(lock) == NULL)
1196 lock->l_lvb_len = lvb_len;
1197 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1198 if (lock->l_lvb_data == NULL)
1205 if (lock->l_lvb_data)
1206 OBD_FREE(lock->l_lvb_data, lvb_len);
1207 ldlm_interval_free(ldlm_interval_detach(lock));
1208 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1212 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1213 struct ldlm_lock **lockp,
1214 void *cookie, int *flags)
1216 struct ldlm_lock *lock = *lockp;
1217 struct ldlm_resource *res = lock->l_resource;
1218 int local = ns_is_client(res->lr_namespace);
1219 ldlm_processing_policy policy;
1220 ldlm_error_t rc = ELDLM_OK;
1221 struct ldlm_interval *node = NULL;
1224 do_gettimeofday(&lock->l_enqueued_time);
1225 /* policies are not executed on the client or during replay */
1226 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1227 && !local && ns->ns_policy) {
1228 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1230 if (rc == ELDLM_LOCK_REPLACED) {
1231 /* The lock that was returned has already been granted,
1232 * and placed into lockp. If it's not the same as the
1233 * one we passed in, then destroy the old one and our
1234 * work here is done. */
1235 if (lock != *lockp) {
1236 ldlm_lock_destroy(lock);
1237 LDLM_LOCK_PUT(lock);
1239 *flags |= LDLM_FL_LOCK_CHANGED;
1241 } else if (rc != ELDLM_OK ||
1242 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1243 ldlm_lock_destroy(lock);
1248 /* For a replaying lock, it might be already in granted list. So
1249 * unlinking the lock will cause the interval node to be freed, we
1250 * have to allocate the interval node early otherwise we can't regrant
1251 * this lock in the future. - jay */
1252 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1253 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO,
1256 lock_res_and_lock(lock);
1257 if (local && lock->l_req_mode == lock->l_granted_mode) {
1258 /* The server returned a blocked lock, but it was granted
1259 * before we got a chance to actually enqueue it. We don't
1260 * need to do anything else. */
1261 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1262 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1263 GOTO(out, ELDLM_OK);
1266 ldlm_resource_unlink_lock(lock);
1267 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1269 ldlm_lock_destroy_nolock(lock);
1270 GOTO(out, rc = -ENOMEM);
1273 CFS_INIT_LIST_HEAD(&node->li_group);
1274 ldlm_interval_attach(node, lock);
1278 /* Some flags from the enqueue want to make it into the AST, via the
1279 * lock's l_flags. */
1280 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1282 /* This distinction between local lock trees is very important; a client
1283 * namespace only has information about locks taken by that client, and
1284 * thus doesn't have enough information to decide for itself if it can
1285 * be granted (below). In this case, we do exactly what the server
1286 * tells us to do, as dictated by the 'flags'.
1288 * We do exactly the same thing during recovery, when the server is
1289 * more or less trusting the clients not to lie.
1291 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1292 * granted/converting queues. */
1294 if (*flags & LDLM_FL_BLOCK_CONV)
1295 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1296 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1297 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1299 ldlm_grant_lock(lock, NULL);
1300 GOTO(out, ELDLM_OK);
1301 } else if (*flags & LDLM_FL_REPLAY) {
1302 if (*flags & LDLM_FL_BLOCK_CONV) {
1303 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1304 GOTO(out, ELDLM_OK);
1305 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1306 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1307 GOTO(out, ELDLM_OK);
1308 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1309 ldlm_grant_lock(lock, NULL);
1310 GOTO(out, ELDLM_OK);
1312 /* If no flags, fall through to normal enqueue path. */
1315 policy = ldlm_processing_policy_table[res->lr_type];
1316 policy(lock, flags, 1, &rc, NULL);
1319 unlock_res_and_lock(lock);
1321 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1325 /* Must be called with namespace taken: queue is waiting or converting. */
1326 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1327 struct list_head *work_list)
1329 struct list_head *tmp, *pos;
1330 ldlm_processing_policy policy;
1332 int rc = LDLM_ITER_CONTINUE;
1336 check_res_locked(res);
1338 policy = ldlm_processing_policy_table[res->lr_type];
1341 list_for_each_safe(tmp, pos, queue) {
1342 struct ldlm_lock *pending;
1343 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1345 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1348 rc = policy(pending, &flags, 0, &err, work_list);
1349 if (rc != LDLM_ITER_CONTINUE)
1356 /* Helper function for ldlm_run_ast_work().
1358 * Send an existing rpc set specified by @arg->set and then
1359 * destroy it. Create new one if @do_create flag is set. */
1361 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1365 ptlrpc_set_wait(arg->set);
1366 if (arg->type == LDLM_BL_CALLBACK)
1367 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1368 ptlrpc_set_destroy(arg->set);
1371 arg->set = ptlrpc_prep_set();
1377 ldlm_work_bl_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1379 struct ldlm_lock_desc d;
1380 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_bl_ast);
1383 /* nobody should touch l_bl_ast */
1384 lock_res_and_lock(lock);
1385 list_del_init(&lock->l_bl_ast);
1387 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1388 LASSERT(lock->l_bl_ast_run == 0);
1389 LASSERT(lock->l_blocking_lock);
1390 lock->l_bl_ast_run++;
1391 unlock_res_and_lock(lock);
1393 ldlm_lock2desc(lock->l_blocking_lock, &d);
1395 LDLM_LOCK_PUT(lock->l_blocking_lock);
1396 lock->l_blocking_lock = NULL;
1397 lock->l_blocking_ast(lock, &d, (void *)arg,
1399 LDLM_LOCK_PUT(lock);
1405 ldlm_work_cp_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1407 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_cp_ast);
1408 ldlm_completion_callback completion_callback;
1412 /* It's possible to receive a completion AST before we've set
1413 * the l_completion_ast pointer: either because the AST arrived
1414 * before the reply, or simply because there's a small race
1415 * window between receiving the reply and finishing the local
1416 * enqueue. (bug 842)
1418 * This can't happen with the blocking_ast, however, because we
1419 * will never call the local blocking_ast until we drop our
1420 * reader/writer reference, which we won't do until we get the
1421 * reply and finish enqueueing. */
1423 /* nobody should touch l_cp_ast */
1424 lock_res_and_lock(lock);
1425 list_del_init(&lock->l_cp_ast);
1426 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1427 /* save l_completion_ast since it can be changed by
1428 * mds_intent_policy(), see bug 14225 */
1429 completion_callback = lock->l_completion_ast;
1430 lock->l_flags &= ~LDLM_FL_CP_REQD;
1431 unlock_res_and_lock(lock);
1433 if (completion_callback != NULL) {
1434 completion_callback(lock, 0, (void *)arg);
1437 LDLM_LOCK_PUT(lock);
1443 ldlm_work_revoke_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1445 struct ldlm_lock_desc desc;
1446 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_rk_ast);
1449 list_del_init(&lock->l_rk_ast);
1451 /* the desc just pretend to exclusive */
1452 ldlm_lock2desc(lock, &desc);
1453 desc.l_req_mode = LCK_EX;
1454 desc.l_granted_mode = 0;
1456 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1457 LDLM_LOCK_PUT(lock);
1462 int ldlm_run_ast_work(struct list_head *rpc_list, ldlm_desc_ast_t ast_type)
1464 struct ldlm_cb_set_arg arg;
1465 struct list_head *tmp, *pos;
1466 int (*work_ast_lock)(struct list_head *tmp, struct ldlm_cb_set_arg *arg);
1470 arg.set = ptlrpc_prep_set();
1471 atomic_set(&arg.restart, 0);
1473 case LDLM_WORK_BL_AST:
1474 arg.type = LDLM_BL_CALLBACK;
1475 work_ast_lock = ldlm_work_bl_ast_lock;
1477 case LDLM_WORK_CP_AST:
1478 arg.type = LDLM_CP_CALLBACK;
1479 work_ast_lock = ldlm_work_cp_ast_lock;
1481 case LDLM_WORK_REVOKE_AST:
1482 arg.type = LDLM_BL_CALLBACK;
1483 work_ast_lock = ldlm_work_revoke_ast_lock;
1490 list_for_each_safe(tmp, pos, rpc_list) {
1491 ast_count += work_ast_lock(tmp, &arg);
1493 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1494 * and create a new set for requests that remained in
1496 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1497 ldlm_send_and_maybe_create_set(&arg, 1);
1503 ldlm_send_and_maybe_create_set(&arg, 0);
1505 /* In case when number of ASTs is multiply of
1506 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1507 * @arg.set must be destroyed here, otherwise we get
1508 * write memory leaking. */
1509 ptlrpc_set_destroy(arg.set);
1511 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1514 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1516 ldlm_reprocess_all(res);
1517 return LDLM_ITER_CONTINUE;
1520 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1522 struct list_head *tmp;
1529 spin_lock(&ns->ns_hash_lock);
1530 for (i = 0; i < RES_HASH_SIZE; i++) {
1531 tmp = ns->ns_hash[i].next;
1532 while (tmp != &(ns->ns_hash[i])) {
1533 struct ldlm_resource *res =
1534 list_entry(tmp, struct ldlm_resource, lr_hash);
1536 ldlm_resource_getref(res);
1537 spin_unlock(&ns->ns_hash_lock);
1539 rc = reprocess_one_queue(res, NULL);
1541 spin_lock(&ns->ns_hash_lock);
1543 ldlm_resource_putref_locked(res);
1545 if (rc == LDLM_ITER_STOP)
1550 spin_unlock(&ns->ns_hash_lock);
1554 void ldlm_reprocess_all(struct ldlm_resource *res)
1556 CFS_LIST_HEAD(rpc_list);
1560 /* Local lock trees don't get reprocessed. */
1561 if (ns_is_client(res->lr_namespace)) {
1568 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1569 if (rc == LDLM_ITER_CONTINUE)
1570 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1573 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1574 if (rc == -ERESTART) {
1575 LASSERT(list_empty(&rpc_list));
1581 void ldlm_cancel_callback(struct ldlm_lock *lock)
1583 check_res_locked(lock->l_resource);
1584 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1585 lock->l_flags |= LDLM_FL_CANCEL;
1586 if (lock->l_blocking_ast) {
1587 // l_check_no_ns_lock(ns);
1588 unlock_res_and_lock(lock);
1589 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1591 lock_res_and_lock(lock);
1593 LDLM_DEBUG(lock, "no blocking ast");
1596 lock->l_flags |= LDLM_FL_BL_DONE;
1599 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1601 if (req->l_resource->lr_type != LDLM_PLAIN &&
1602 req->l_resource->lr_type != LDLM_IBITS)
1605 list_del_init(&req->l_sl_policy);
1606 list_del_init(&req->l_sl_mode);
1609 void ldlm_lock_cancel(struct ldlm_lock *lock)
1611 struct ldlm_resource *res;
1612 struct ldlm_namespace *ns;
1615 lock_res_and_lock(lock);
1617 res = lock->l_resource;
1618 ns = res->lr_namespace;
1620 /* Please do not, no matter how tempting, remove this LBUG without
1621 * talking to me first. -phik */
1622 if (lock->l_readers || lock->l_writers) {
1623 LDLM_ERROR(lock, "lock still has references");
1627 ldlm_del_waiting_lock(lock);
1629 /* Releases cancel callback. */
1630 ldlm_cancel_callback(lock);
1632 /* Yes, second time, just in case it was added again while we were
1633 running with no res lock in ldlm_cancel_callback */
1634 ldlm_del_waiting_lock(lock);
1635 ldlm_resource_unlink_lock(lock);
1636 ldlm_lock_destroy_nolock(lock);
1638 if (lock->l_granted_mode == lock->l_req_mode)
1639 ldlm_pool_del(&ns->ns_pool, lock);
1641 /* Make sure we will not be called again for same lock what is possible
1642 * if not to zero out lock->l_granted_mode */
1643 lock->l_granted_mode = LCK_MINMODE;
1644 unlock_res_and_lock(lock);
1649 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1651 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1657 lock->l_ast_data = data;
1658 LDLM_LOCK_PUT(lock);
1662 void ldlm_cancel_locks_for_export_cb(void *obj, void *data)
1664 struct obd_export *exp = data;
1665 struct ldlm_lock *lock = obj;
1666 struct ldlm_resource *res;
1668 res = ldlm_resource_getref(lock->l_resource);
1669 LDLM_LOCK_GET(lock);
1671 LDLM_DEBUG(lock, "export %p", exp);
1672 ldlm_res_lvbo_update(res, NULL, 0, 1);
1673 ldlm_lock_cancel(lock);
1674 ldlm_reprocess_all(res);
1675 ldlm_resource_putref(res);
1676 LDLM_LOCK_PUT(lock);
1679 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1681 lustre_hash_for_each_empty(exp->exp_lock_hash,
1682 ldlm_cancel_locks_for_export_cb, exp);
1685 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1688 CFS_LIST_HEAD(rpc_list);
1689 struct ldlm_resource *res;
1690 struct ldlm_namespace *ns;
1693 struct sl_insert_point prev;
1695 struct ldlm_interval *node;
1698 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1699 *flags |= LDLM_FL_BLOCK_GRANTED;
1700 RETURN(lock->l_resource);
1703 /* I can't check the type of lock here because the bitlock of lock
1704 * is not held here, so do the allocation blindly. -jay */
1705 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO, sizeof(*node));
1706 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1709 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1710 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1712 lock_res_and_lock(lock);
1714 res = lock->l_resource;
1715 ns = res->lr_namespace;
1717 old_mode = lock->l_req_mode;
1718 lock->l_req_mode = new_mode;
1719 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1720 /* remember the lock position where the lock might be
1721 * added back to the granted list later and also
1722 * remember the join mode for skiplist fixing. */
1723 prev.res_link = lock->l_res_link.prev;
1724 prev.mode_link = lock->l_sl_mode.prev;
1725 prev.policy_link = lock->l_sl_policy.prev;
1726 ldlm_resource_unlink_lock(lock);
1728 ldlm_resource_unlink_lock(lock);
1729 if (res->lr_type == LDLM_EXTENT) {
1730 /* FIXME: ugly code, I have to attach the lock to a
1731 * interval node again since perhaps it will be granted
1733 CFS_INIT_LIST_HEAD(&node->li_group);
1734 ldlm_interval_attach(node, lock);
1739 /* If this is a local resource, put it on the appropriate list. */
1740 if (ns_is_client(res->lr_namespace)) {
1741 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1742 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1744 /* This should never happen, because of the way the
1745 * server handles conversions. */
1746 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1750 ldlm_grant_lock(lock, &rpc_list);
1752 /* FIXME: completion handling not with ns_lock held ! */
1753 if (lock->l_completion_ast)
1754 lock->l_completion_ast(lock, 0, NULL);
1758 ldlm_processing_policy policy;
1759 policy = ldlm_processing_policy_table[res->lr_type];
1760 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1761 if (rc == LDLM_ITER_STOP) {
1762 lock->l_req_mode = old_mode;
1763 if (res->lr_type == LDLM_EXTENT)
1764 ldlm_extent_add_lock(res, lock);
1766 ldlm_granted_list_add_lock(lock, &prev);
1770 *flags |= LDLM_FL_BLOCK_GRANTED;
1774 unlock_res_and_lock(lock);
1777 ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1779 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1783 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1785 struct obd_device *obd = NULL;
1787 if (!((libcfs_debug | D_ERROR) & level))
1791 CDEBUG(level, " NULL LDLM lock\n");
1795 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1796 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1798 if (lock->l_conn_export != NULL)
1799 obd = lock->l_conn_export->exp_obd;
1800 if (lock->l_export && lock->l_export->exp_connection) {
1801 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1802 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1803 lock->l_remote_handle.cookie);
1804 } else if (obd == NULL) {
1805 CDEBUG(level, " Node: local\n");
1807 struct obd_import *imp = obd->u.cli.cl_import;
1808 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1809 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1810 lock->l_remote_handle.cookie);
1812 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1814 lock->l_resource->lr_name.name[0],
1815 lock->l_resource->lr_name.name[1],
1816 lock->l_resource->lr_name.name[2]);
1817 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1818 "write: %d flags: %#x\n", ldlm_lockname[lock->l_req_mode],
1819 ldlm_lockname[lock->l_granted_mode],
1820 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1822 if (lock->l_resource->lr_type == LDLM_EXTENT)
1823 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1824 " (req "LPU64"-"LPU64")\n",
1825 lock->l_policy_data.l_extent.start,
1826 lock->l_policy_data.l_extent.end,
1827 lock->l_req_extent.start, lock->l_req_extent.end);
1828 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1829 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1830 lock->l_policy_data.l_flock.pid,
1831 lock->l_policy_data.l_flock.start,
1832 lock->l_policy_data.l_flock.end);
1833 else if (lock->l_resource->lr_type == LDLM_IBITS)
1834 CDEBUG(level, " Bits: "LPX64"\n",
1835 lock->l_policy_data.l_inodebits.bits);
1838 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1840 struct ldlm_lock *lock;
1842 if (!((libcfs_debug | D_ERROR) & level))
1845 lock = ldlm_handle2lock(lockh);
1849 ldlm_lock_dump(D_OTHER, lock, 0);
1851 LDLM_LOCK_PUT(lock);
1854 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1855 struct libcfs_debug_msg_data *data, const char *fmt,
1859 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1861 va_start(args, fmt);
1863 if (lock->l_resource == NULL) {
1864 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1865 data->msg_fn, data->msg_line, fmt, args,
1866 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1867 "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "
1868 LPX64" expref: %d pid: %u\n", lock,
1869 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1870 lock->l_readers, lock->l_writers,
1871 ldlm_lockname[lock->l_granted_mode],
1872 ldlm_lockname[lock->l_req_mode],
1873 lock->l_flags, lock->l_remote_handle.cookie,
1875 atomic_read(&lock->l_export->exp_refcount) : -99,
1881 switch (lock->l_resource->lr_type) {
1883 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1884 data->msg_fn, data->msg_line, fmt, args,
1885 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1886 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1887 "] (req "LPU64"->"LPU64") flags: %x remote: "LPX64
1888 " expref: %d pid: %u\n",
1889 lock->l_resource->lr_namespace->ns_name, lock,
1890 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1891 lock->l_readers, lock->l_writers,
1892 ldlm_lockname[lock->l_granted_mode],
1893 ldlm_lockname[lock->l_req_mode],
1894 lock->l_resource->lr_name.name[0],
1895 lock->l_resource->lr_name.name[1],
1896 atomic_read(&lock->l_resource->lr_refcount),
1897 ldlm_typename[lock->l_resource->lr_type],
1898 lock->l_policy_data.l_extent.start,
1899 lock->l_policy_data.l_extent.end,
1900 lock->l_req_extent.start, lock->l_req_extent.end,
1901 lock->l_flags, lock->l_remote_handle.cookie,
1903 atomic_read(&lock->l_export->exp_refcount) : -99,
1908 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1909 data->msg_fn, data->msg_line, fmt, args,
1910 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1911 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1912 "["LPU64"->"LPU64"] flags: %x remote: "LPX64
1913 " expref: %d pid: %u\n",
1914 lock->l_resource->lr_namespace->ns_name, lock,
1915 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1916 lock->l_readers, lock->l_writers,
1917 ldlm_lockname[lock->l_granted_mode],
1918 ldlm_lockname[lock->l_req_mode],
1919 lock->l_resource->lr_name.name[0],
1920 lock->l_resource->lr_name.name[1],
1921 atomic_read(&lock->l_resource->lr_refcount),
1922 ldlm_typename[lock->l_resource->lr_type],
1923 lock->l_policy_data.l_flock.pid,
1924 lock->l_policy_data.l_flock.start,
1925 lock->l_policy_data.l_flock.end,
1926 lock->l_flags, lock->l_remote_handle.cookie,
1928 atomic_read(&lock->l_export->exp_refcount) : -99,
1933 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1934 data->msg_fn, data->msg_line, fmt, args,
1935 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1936 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1937 "flags: %x remote: "LPX64" expref: %d "
1939 lock->l_resource->lr_namespace->ns_name,
1940 lock, lock->l_handle.h_cookie,
1941 atomic_read (&lock->l_refc),
1942 lock->l_readers, lock->l_writers,
1943 ldlm_lockname[lock->l_granted_mode],
1944 ldlm_lockname[lock->l_req_mode],
1945 lock->l_resource->lr_name.name[0],
1946 lock->l_resource->lr_name.name[1],
1947 lock->l_policy_data.l_inodebits.bits,
1948 atomic_read(&lock->l_resource->lr_refcount),
1949 ldlm_typename[lock->l_resource->lr_type],
1950 lock->l_flags, lock->l_remote_handle.cookie,
1952 atomic_read(&lock->l_export->exp_refcount) : -99,
1957 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1958 data->msg_fn, data->msg_line, fmt, args,
1959 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1960 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: %x "
1961 "remote: "LPX64" expref: %d pid: %u\n",
1962 lock->l_resource->lr_namespace->ns_name,
1963 lock, lock->l_handle.h_cookie,
1964 atomic_read (&lock->l_refc),
1965 lock->l_readers, lock->l_writers,
1966 ldlm_lockname[lock->l_granted_mode],
1967 ldlm_lockname[lock->l_req_mode],
1968 lock->l_resource->lr_name.name[0],
1969 lock->l_resource->lr_name.name[1],
1970 atomic_read(&lock->l_resource->lr_refcount),
1971 ldlm_typename[lock->l_resource->lr_type],
1972 lock->l_flags, lock->l_remote_handle.cookie,
1974 atomic_read(&lock->l_export->exp_refcount) : -99,
1980 EXPORT_SYMBOL(_ldlm_lock_debug);