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 # include <linux/lustre_intent.h>
48 # include <liblustre.h>
49 # include <libcfs/kp30.h>
52 #include <obd_class.h>
53 #include "ldlm_internal.h"
55 //struct lustre_lock ldlm_everything_lock;
58 char *ldlm_lockname[] = {
69 char *ldlm_typename[] = {
76 char *ldlm_it2str(int it)
83 case (IT_OPEN | IT_CREAT):
96 CERROR("Unknown intent %d\n", it);
101 extern cfs_mem_cache_t *ldlm_lock_slab;
103 static ldlm_processing_policy ldlm_processing_policy_table[] = {
104 [LDLM_PLAIN] ldlm_process_plain_lock,
105 [LDLM_EXTENT] ldlm_process_extent_lock,
107 [LDLM_FLOCK] ldlm_process_flock_lock,
109 [LDLM_IBITS] ldlm_process_inodebits_lock,
112 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
114 return ldlm_processing_policy_table[res->lr_type];
117 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
123 * REFCOUNTED LOCK OBJECTS
128 * Lock refcounts, during creation:
129 * - one special one for allocation, dec'd only once in destroy
130 * - one for being a lock that's in-use
131 * - one for the addref associated with a new lock
133 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
135 atomic_inc(&lock->l_refc);
139 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
141 LASSERT(size == sizeof(*lock));
142 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
145 void ldlm_lock_put(struct ldlm_lock *lock)
149 LASSERT(lock->l_resource != LP_POISON);
150 LASSERT(atomic_read(&lock->l_refc) > 0);
151 if (atomic_dec_and_test(&lock->l_refc)) {
152 struct ldlm_resource *res;
154 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing it.");
156 res = lock->l_resource;
157 LASSERT(lock->l_destroyed);
158 LASSERT(list_empty(&lock->l_res_link));
159 LASSERT(list_empty(&lock->l_pending_chain));
161 atomic_dec(&res->lr_namespace->ns_locks);
162 ldlm_resource_putref(res);
163 lock->l_resource = NULL;
164 if (lock->l_export) {
165 class_export_put(lock->l_export);
166 lock->l_export = NULL;
169 if (lock->l_lvb_data != NULL)
170 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
172 ldlm_interval_free(ldlm_interval_detach(lock));
173 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
180 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
183 if (!list_empty(&lock->l_lru)) {
184 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
185 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
186 list_del_init(&lock->l_lru);
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_waitq_init(&lock->l_waitq);
348 lock->l_blocking_lock = NULL;
349 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
350 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
351 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
353 atomic_inc(&resource->lr_namespace->ns_locks);
354 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
355 class_handle_hash(&lock->l_handle, lock_handle_addref);
357 CFS_INIT_LIST_HEAD(&lock->l_extents_list);
358 spin_lock_init(&lock->l_extents_list_lock);
359 CFS_INIT_LIST_HEAD(&lock->l_cache_locks_list);
360 lock->l_callback_timeout = 0;
365 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
366 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(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(struct lustre_handle *handle, int flags)
424 struct ldlm_namespace *ns;
425 struct ldlm_lock *lock = NULL, *retval = NULL;
430 lock = class_handle2object(handle->cookie);
434 LASSERT(lock->l_resource != NULL);
435 ns = lock->l_resource->lr_namespace;
438 lock_res_and_lock(lock);
440 /* It's unlikely but possible that someone marked the lock as
441 * destroyed after we did handle2object on it */
442 if (lock->l_destroyed) {
443 unlock_res_and_lock(lock);
444 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
449 if (flags && (lock->l_flags & flags)) {
450 unlock_res_and_lock(lock);
456 lock->l_flags |= flags;
458 unlock_res_and_lock(lock);
465 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
466 struct lustre_handle *handle)
468 struct ldlm_lock *retval = NULL;
469 retval = __ldlm_handle2lock(handle, 0);
473 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
475 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
476 /* INODEBITS_INTEROP: If the other side does not support
477 * inodebits, reply with a plain lock descriptor.
479 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
480 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
481 struct ldlm_resource res = *lock->l_resource;
483 /* Make sure all the right bits are set in this lock we
484 are going to pass to client */
485 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
486 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
487 "Inappropriate inode lock bits during "
488 "conversion " LPU64 "\n",
489 lock->l_policy_data.l_inodebits.bits);
490 res.lr_type = LDLM_PLAIN;
491 ldlm_res2desc(&res, &desc->l_resource);
492 /* Convert "new" lock mode to something old client can
494 if ((lock->l_req_mode == LCK_CR) ||
495 (lock->l_req_mode == LCK_CW))
496 desc->l_req_mode = LCK_PR;
498 desc->l_req_mode = lock->l_req_mode;
499 if ((lock->l_granted_mode == LCK_CR) ||
500 (lock->l_granted_mode == LCK_CW)) {
501 desc->l_granted_mode = LCK_PR;
503 /* We never grant PW/EX locks to clients */
504 LASSERT((lock->l_granted_mode != LCK_PW) &&
505 (lock->l_granted_mode != LCK_EX));
506 desc->l_granted_mode = lock->l_granted_mode;
509 /* We do not copy policy here, because there is no
510 policy for plain locks */
512 ldlm_res2desc(lock->l_resource, &desc->l_resource);
513 desc->l_req_mode = lock->l_req_mode;
514 desc->l_granted_mode = lock->l_granted_mode;
515 desc->l_policy_data = lock->l_policy_data;
519 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
520 struct list_head *work_list)
522 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
523 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
524 lock->l_flags |= LDLM_FL_AST_SENT;
525 /* If the enqueuing client said so, tell the AST recipient to
526 * discard dirty data, rather than writing back. */
527 if (new->l_flags & LDLM_AST_DISCARD_DATA)
528 lock->l_flags |= LDLM_FL_DISCARD_DATA;
529 LASSERT(list_empty(&lock->l_bl_ast));
530 list_add(&lock->l_bl_ast, work_list);
532 LASSERT(lock->l_blocking_lock == NULL);
533 lock->l_blocking_lock = LDLM_LOCK_GET(new);
537 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
539 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
540 lock->l_flags |= LDLM_FL_CP_REQD;
541 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
542 LASSERT(list_empty(&lock->l_cp_ast));
543 list_add(&lock->l_cp_ast, work_list);
548 /* must be called with lr_lock held */
549 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
550 struct list_head *work_list)
553 check_res_locked(lock->l_resource);
555 ldlm_add_bl_work_item(lock, new, work_list);
557 ldlm_add_cp_work_item(lock, work_list);
561 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
563 struct ldlm_lock *lock;
565 lock = ldlm_handle2lock(lockh);
566 LASSERT(lock != NULL);
567 ldlm_lock_addref_internal(lock, mode);
571 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
573 ldlm_lock_remove_from_lru(lock);
574 if (mode & (LCK_NL | LCK_CR | LCK_PR))
576 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
579 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
582 /* only called for local locks */
583 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
585 lock_res_and_lock(lock);
586 ldlm_lock_addref_internal_nolock(lock, mode);
587 unlock_res_and_lock(lock);
590 /* only called in ldlm_flock_destroy and for local locks.
591 * for LDLM_FLOCK type locks, l_blocking_ast is null, and
592 * ldlm_lock_remove_from_lru() does nothing, it is safe
593 * for ldlm_flock_destroy usage by dropping some code */
594 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
596 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
597 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
598 LASSERT(lock->l_readers > 0);
601 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
602 LASSERT(lock->l_writers > 0);
606 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
609 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
611 struct ldlm_namespace *ns;
614 lock_res_and_lock(lock);
616 ns = lock->l_resource->lr_namespace;
618 ldlm_lock_decref_internal_nolock(lock, mode);
620 if (lock->l_flags & LDLM_FL_LOCAL &&
621 !lock->l_readers && !lock->l_writers) {
622 /* If this is a local lock on a server namespace and this was
623 * the last reference, cancel the lock. */
624 CDEBUG(D_INFO, "forcing cancel of local lock\n");
625 lock->l_flags |= LDLM_FL_CBPENDING;
628 if (!lock->l_readers && !lock->l_writers &&
629 (lock->l_flags & LDLM_FL_CBPENDING)) {
630 /* If we received a blocked AST and this was the last reference,
631 * run the callback. */
632 if (ns_is_server(ns) && lock->l_export)
633 CERROR("FL_CBPENDING set on non-local lock--just a "
636 LDLM_DEBUG(lock, "final decref done on cbpending lock");
638 LDLM_LOCK_GET(lock); /* dropped by bl thread */
639 ldlm_lock_remove_from_lru(lock);
640 unlock_res_and_lock(lock);
642 if (lock->l_flags & LDLM_FL_FAIL_LOC)
643 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
645 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
646 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
647 ldlm_handle_bl_callback(ns, NULL, lock);
648 } else if (ns_is_client(ns) &&
649 !lock->l_readers && !lock->l_writers &&
650 !(lock->l_flags & LDLM_FL_NO_LRU) &&
651 !(lock->l_flags & LDLM_FL_BL_AST)) {
652 /* If this is a client-side namespace and this was the last
653 * reference, put it on the LRU. */
654 ldlm_lock_add_to_lru(lock);
655 unlock_res_and_lock(lock);
656 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
657 * are not supported by the server, otherwise, it is done on
659 if (!exp_connect_cancelset(lock->l_conn_export) &&
660 !ns_connect_lru_resize(ns))
661 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
663 unlock_res_and_lock(lock);
669 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
671 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
672 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
673 ldlm_lock_decref_internal(lock, mode);
677 /* This will drop a lock reference and mark it for destruction, but will not
678 * necessarily cancel the lock before returning. */
679 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
681 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
684 LASSERT(lock != NULL);
686 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
687 lock_res_and_lock(lock);
688 lock->l_flags |= LDLM_FL_CBPENDING;
689 unlock_res_and_lock(lock);
690 ldlm_lock_decref_internal(lock, mode);
694 struct sl_insert_point {
695 struct list_head *res_link;
696 struct list_head *mode_link;
697 struct list_head *policy_link;
701 * search_granted_lock
704 * Finds a position to insert the new lock.
706 * queue [input]: the granted list where search acts on;
707 * req [input]: the lock whose position to be located;
708 * prev [output]: positions within 3 lists to insert @req to
712 * - ldlm_grant_lock_with_skiplist
714 static void search_granted_lock(struct list_head *queue,
715 struct ldlm_lock *req,
716 struct sl_insert_point *prev)
718 struct list_head *tmp;
719 struct ldlm_lock *lock, *mode_end, *policy_end;
722 list_for_each(tmp, queue) {
723 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
725 mode_end = list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
728 if (lock->l_req_mode != req->l_req_mode) {
729 /* jump to last lock of mode group */
730 tmp = &mode_end->l_res_link;
734 /* suitable mode group is found */
735 if (lock->l_resource->lr_type == LDLM_PLAIN) {
736 /* insert point is last lock of the mode group */
737 prev->res_link = &mode_end->l_res_link;
738 prev->mode_link = &mode_end->l_sl_mode;
739 prev->policy_link = &req->l_sl_policy;
742 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
744 policy_end = list_entry(lock->l_sl_policy.prev,
748 if (lock->l_policy_data.l_inodebits.bits ==
749 req->l_policy_data.l_inodebits.bits) {
750 /* insert point is last lock of
751 * the policy group */
753 &policy_end->l_res_link;
755 &policy_end->l_sl_mode;
757 &policy_end->l_sl_policy;
762 if (policy_end == mode_end)
763 /* done with mode group */
766 /* jump to next policy group within the mode group */
767 tmp = policy_end->l_res_link.next;
768 lock = list_entry(tmp, struct ldlm_lock,
770 } /* loop over policy groups within the mode group */
772 /* insert point is last lock of the mode group,
773 * new policy group is started */
774 prev->res_link = &mode_end->l_res_link;
775 prev->mode_link = &mode_end->l_sl_mode;
776 prev->policy_link = &req->l_sl_policy;
780 LDLM_ERROR(lock, "is not LDLM_PLAIN or LDLM_IBITS lock");
785 /* insert point is last lock on the queue,
786 * new mode group and new policy group are started */
787 prev->res_link = queue->prev;
788 prev->mode_link = &req->l_sl_mode;
789 prev->policy_link = &req->l_sl_policy;
794 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
795 struct sl_insert_point *prev)
797 struct ldlm_resource *res = lock->l_resource;
800 check_res_locked(res);
802 ldlm_resource_dump(D_INFO, res);
803 CDEBUG(D_OTHER, "About to add this lock:\n");
804 ldlm_lock_dump(D_OTHER, lock, 0);
806 if (lock->l_destroyed) {
807 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
811 LASSERT(list_empty(&lock->l_res_link));
812 LASSERT(list_empty(&lock->l_sl_mode));
813 LASSERT(list_empty(&lock->l_sl_policy));
815 list_add(&lock->l_res_link, prev->res_link);
816 list_add(&lock->l_sl_mode, prev->mode_link);
817 list_add(&lock->l_sl_policy, prev->policy_link);
822 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
824 struct sl_insert_point prev;
827 LASSERT(lock->l_req_mode == lock->l_granted_mode);
829 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
830 ldlm_granted_list_add_lock(lock, &prev);
835 * - ldlm_lock_enqueue
836 * - ldlm_reprocess_queue
837 * - ldlm_lock_convert
839 * must be called with lr_lock held
841 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
843 struct ldlm_resource *res = lock->l_resource;
846 check_res_locked(res);
848 lock->l_granted_mode = lock->l_req_mode;
849 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
850 ldlm_grant_lock_with_skiplist(lock);
851 else if (res->lr_type == LDLM_EXTENT)
852 ldlm_extent_add_lock(res, lock);
854 ldlm_resource_add_lock(res, &res->lr_granted, lock);
856 if (lock->l_granted_mode < res->lr_most_restr)
857 res->lr_most_restr = lock->l_granted_mode;
859 if (work_list && lock->l_completion_ast != NULL)
860 ldlm_add_ast_work_item(lock, NULL, work_list);
862 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
866 /* returns a referenced lock or NULL. See the flag descriptions below, in the
867 * comment above ldlm_lock_match */
868 static struct ldlm_lock *search_queue(struct list_head *queue,
870 ldlm_policy_data_t *policy,
871 struct ldlm_lock *old_lock, int flags)
873 struct ldlm_lock *lock;
874 struct list_head *tmp;
876 list_for_each(tmp, queue) {
879 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
881 if (lock == old_lock)
884 /* llite sometimes wants to match locks that will be
885 * canceled when their users drop, but we allow it to match
886 * if it passes in CBPENDING and the lock still has users.
887 * this is generally only going to be used by children
888 * whose parents already hold a lock so forward progress
889 * can still happen. */
890 if (lock->l_flags & LDLM_FL_CBPENDING &&
891 !(flags & LDLM_FL_CBPENDING))
893 if (lock->l_flags & LDLM_FL_CBPENDING &&
894 lock->l_readers == 0 && lock->l_writers == 0)
897 if (!(lock->l_req_mode & *mode))
900 match = lock->l_req_mode;
901 if (lock->l_resource->lr_type == LDLM_EXTENT &&
902 (lock->l_policy_data.l_extent.start >
903 policy->l_extent.start ||
904 lock->l_policy_data.l_extent.end < policy->l_extent.end))
907 if (unlikely(match == LCK_GROUP) &&
908 lock->l_resource->lr_type == LDLM_EXTENT &&
909 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
912 /* We match if we have existing lock with same or wider set
914 if (lock->l_resource->lr_type == LDLM_IBITS &&
915 ((lock->l_policy_data.l_inodebits.bits &
916 policy->l_inodebits.bits) !=
917 policy->l_inodebits.bits))
920 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
923 if ((flags & LDLM_FL_LOCAL_ONLY) &&
924 !(lock->l_flags & LDLM_FL_LOCAL))
927 if (flags & LDLM_FL_TEST_LOCK) {
929 ldlm_lock_touch_in_lru(lock);
931 ldlm_lock_addref_internal_nolock(lock, match);
940 void ldlm_lock_allow_match(struct ldlm_lock *lock)
942 lock_res_and_lock(lock);
943 lock->l_flags |= LDLM_FL_LVB_READY;
944 cfs_waitq_signal(&lock->l_waitq);
945 unlock_res_and_lock(lock);
948 int ldlm_lock_fast_match(struct ldlm_lock *lock, int rw,
949 obd_off start, obd_off end,
952 LASSERT(rw == OBD_BRW_READ || rw == OBD_BRW_WRITE);
957 lock_res_and_lock(lock);
958 /* check if granted mode is compatible */
959 if (rw == OBD_BRW_WRITE &&
960 !(lock->l_granted_mode & (LCK_PW|LCK_GROUP)))
963 /* does the lock cover the region we would like to access? */
964 if ((lock->l_policy_data.l_extent.start > start) ||
965 (lock->l_policy_data.l_extent.end < end))
968 /* if we received a blocking callback and the lock is no longer
969 * referenced, don't use it */
970 if ((lock->l_flags & LDLM_FL_CBPENDING) &&
971 !lock->l_writers && !lock->l_readers)
974 ldlm_lock_addref_internal_nolock(lock, rw == OBD_BRW_WRITE ? LCK_PW : LCK_PR);
975 unlock_res_and_lock(lock);
976 *cookie = (void *)lock;
977 return 1; /* avoid using rc for stack relief */
980 unlock_res_and_lock(lock);
984 void ldlm_lock_fast_release(void *cookie, int rw)
986 struct ldlm_lock *lock = (struct ldlm_lock *)cookie;
988 LASSERT(lock != NULL);
989 LASSERT(rw == OBD_BRW_READ || rw == OBD_BRW_WRITE);
990 LASSERT(rw == OBD_BRW_READ || (lock->l_granted_mode & (LCK_PW | LCK_GROUP)));
991 ldlm_lock_decref_internal(lock, rw == OBD_BRW_WRITE ? LCK_PW : LCK_PR);
994 /* Can be called in two ways:
996 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
997 * for a duplicate of.
999 * Otherwise, all of the fields must be filled in, to match against.
1001 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1002 * server (ie, connh is NULL)
1003 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1004 * list will be considered
1005 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1006 * to be canceled can still be matched as long as they still have reader
1007 * or writer refernces
1008 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1009 * just tell us if we would have matched.
1011 * Returns 1 if it finds an already-existing lock that is compatible; in this
1012 * case, lockh is filled in with a addref()ed lock
1014 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1015 struct ldlm_res_id *res_id, ldlm_type_t type,
1016 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1017 struct lustre_handle *lockh)
1019 struct ldlm_resource *res;
1020 struct ldlm_lock *lock, *old_lock = NULL;
1025 old_lock = ldlm_handle2lock(lockh);
1028 ns = old_lock->l_resource->lr_namespace;
1029 res_id = &old_lock->l_resource->lr_name;
1030 type = old_lock->l_resource->lr_type;
1031 mode = old_lock->l_req_mode;
1034 res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
1036 LASSERT(old_lock == NULL);
1042 lock = search_queue(&res->lr_granted, &mode, policy, old_lock, flags);
1045 if (flags & LDLM_FL_BLOCK_GRANTED)
1047 lock = search_queue(&res->lr_converting, &mode, policy, old_lock, flags);
1050 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock, flags);
1057 ldlm_resource_putref(res);
1060 ldlm_lock2handle(lock, lockh);
1061 if ((flags & LDLM_FL_LVB_READY) && (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1062 struct l_wait_info lwi;
1063 if (lock->l_completion_ast) {
1064 int err = lock->l_completion_ast(lock,
1065 LDLM_FL_WAIT_NOREPROC,
1068 if (flags & LDLM_FL_TEST_LOCK)
1069 LDLM_LOCK_PUT(lock);
1071 ldlm_lock_decref_internal(lock, mode);
1077 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), NULL,
1078 LWI_ON_SIGNAL_NOOP, NULL);
1080 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1081 l_wait_event(lock->l_waitq,
1082 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1087 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1088 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1089 res_id->name[2] : policy->l_extent.start,
1090 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1091 res_id->name[3] : policy->l_extent.end);
1092 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1093 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1094 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1095 type, mode, res_id->name[0], res_id->name[1],
1096 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1097 res_id->name[2] :policy->l_extent.start,
1098 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1099 res_id->name[3] : policy->l_extent.end);
1102 LDLM_LOCK_PUT(old_lock);
1103 if (flags & LDLM_FL_TEST_LOCK && rc)
1104 LDLM_LOCK_PUT(lock);
1106 return rc ? mode : 0;
1109 /* Returns a referenced lock */
1110 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1111 struct ldlm_res_id res_id, ldlm_type_t type,
1113 ldlm_blocking_callback blocking,
1114 ldlm_completion_callback completion,
1115 ldlm_glimpse_callback glimpse,
1116 void *data, __u32 lvb_len)
1118 struct ldlm_lock *lock;
1119 struct ldlm_resource *res;
1122 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1126 lock = ldlm_lock_new(res);
1127 ldlm_resource_putref(res);
1132 lock->l_req_mode = mode;
1133 lock->l_ast_data = data;
1134 lock->l_blocking_ast = blocking;
1135 lock->l_completion_ast = completion;
1136 lock->l_glimpse_ast = glimpse;
1137 lock->l_pid = cfs_curproc_pid();
1139 lock->l_tree_node = NULL;
1140 /* if this is the extent lock, allocate the interval tree node */
1141 if (type == LDLM_EXTENT) {
1142 if (ldlm_interval_alloc(lock) == NULL)
1147 lock->l_lvb_len = lvb_len;
1148 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1149 if (lock->l_lvb_data == NULL)
1156 if (lock->l_lvb_data)
1157 OBD_FREE(lock->l_lvb_data, lvb_len);
1158 ldlm_interval_free(ldlm_interval_detach(lock));
1159 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1163 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1164 struct ldlm_lock **lockp,
1165 void *cookie, int *flags)
1167 struct ldlm_lock *lock = *lockp;
1168 struct ldlm_resource *res = lock->l_resource;
1169 int local = ns_is_client(res->lr_namespace);
1170 ldlm_processing_policy policy;
1171 ldlm_error_t rc = ELDLM_OK;
1172 struct ldlm_interval *node = NULL;
1175 do_gettimeofday(&lock->l_enqueued_time);
1176 /* policies are not executed on the client or during replay */
1177 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1178 && !local && ns->ns_policy) {
1179 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1181 if (rc == ELDLM_LOCK_REPLACED) {
1182 /* The lock that was returned has already been granted,
1183 * and placed into lockp. If it's not the same as the
1184 * one we passed in, then destroy the old one and our
1185 * work here is done. */
1186 if (lock != *lockp) {
1187 ldlm_lock_destroy(lock);
1188 LDLM_LOCK_PUT(lock);
1190 *flags |= LDLM_FL_LOCK_CHANGED;
1192 } else if (rc != ELDLM_OK ||
1193 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1194 ldlm_lock_destroy(lock);
1199 /* For a replaying lock, it might be already in granted list. So
1200 * unlinking the lock will cause the interval node to be freed, we
1201 * have to allocate the interval node early otherwise we can't regrant
1202 * this lock in the future. - jay */
1203 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1204 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO,
1207 lock_res_and_lock(lock);
1208 if (local && lock->l_req_mode == lock->l_granted_mode) {
1209 /* The server returned a blocked lock, but it was granted
1210 * before we got a chance to actually enqueue it. We don't
1211 * need to do anything else. */
1212 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1213 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1214 GOTO(out, ELDLM_OK);
1217 ldlm_resource_unlink_lock(lock);
1218 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1219 LASSERT(!local && (*flags & LDLM_FL_REPLAY));
1221 ldlm_lock_destroy_nolock(lock);
1222 GOTO(out, rc = -ENOMEM);
1225 CFS_INIT_LIST_HEAD(&node->li_group);
1226 ldlm_interval_attach(node, lock);
1230 /* Some flags from the enqueue want to make it into the AST, via the
1231 * lock's l_flags. */
1232 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1234 /* This distinction between local lock trees is very important; a client
1235 * namespace only has information about locks taken by that client, and
1236 * thus doesn't have enough information to decide for itself if it can
1237 * be granted (below). In this case, we do exactly what the server
1238 * tells us to do, as dictated by the 'flags'.
1240 * We do exactly the same thing during recovery, when the server is
1241 * more or less trusting the clients not to lie.
1243 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1244 * granted/converting queues. */
1246 if (*flags & LDLM_FL_BLOCK_CONV)
1247 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1248 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1249 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1251 ldlm_grant_lock(lock, NULL);
1252 GOTO(out, ELDLM_OK);
1253 } else if (*flags & LDLM_FL_REPLAY) {
1254 if (*flags & LDLM_FL_BLOCK_CONV) {
1255 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1256 GOTO(out, ELDLM_OK);
1257 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1258 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1259 GOTO(out, ELDLM_OK);
1260 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1261 ldlm_grant_lock(lock, NULL);
1262 GOTO(out, ELDLM_OK);
1264 /* If no flags, fall through to normal enqueue path. */
1267 policy = ldlm_processing_policy_table[res->lr_type];
1268 policy(lock, flags, 1, &rc, NULL);
1271 unlock_res_and_lock(lock);
1273 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1277 /* Must be called with namespace taken: queue is waiting or converting. */
1278 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1279 struct list_head *work_list)
1281 struct list_head *tmp, *pos;
1282 ldlm_processing_policy policy;
1284 int rc = LDLM_ITER_CONTINUE;
1288 check_res_locked(res);
1290 policy = ldlm_processing_policy_table[res->lr_type];
1293 list_for_each_safe(tmp, pos, queue) {
1294 struct ldlm_lock *pending;
1295 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1297 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1300 rc = policy(pending, &flags, 0, &err, work_list);
1301 if (rc != LDLM_ITER_CONTINUE)
1308 /* Helper function for pair ldlm_run_{bl,cp}_ast_work().
1310 * Send an existing rpc set specified by @arg->set and then
1311 * destroy it. Create new one if @do_create flag is set. */
1313 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1317 rc = ptlrpc_set_wait(arg->set);
1318 if (arg->type == LDLM_BL_CALLBACK)
1319 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1320 ptlrpc_set_destroy(arg->set);
1323 arg->set = ptlrpc_prep_set();
1326 int ldlm_run_bl_ast_work(struct list_head *rpc_list)
1328 struct ldlm_cb_set_arg arg;
1329 struct list_head *tmp, *pos;
1330 struct ldlm_lock_desc d;
1335 arg.set = ptlrpc_prep_set();
1336 atomic_set(&arg.restart, 0);
1337 arg.type = LDLM_BL_CALLBACK;
1340 list_for_each_safe(tmp, pos, rpc_list) {
1341 struct ldlm_lock *lock =
1342 list_entry(tmp, struct ldlm_lock, l_bl_ast);
1344 /* nobody should touch l_bl_ast */
1345 lock_res_and_lock(lock);
1346 list_del_init(&lock->l_bl_ast);
1348 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1349 LASSERT(lock->l_bl_ast_run == 0);
1350 LASSERT(lock->l_blocking_lock);
1351 lock->l_bl_ast_run++;
1352 unlock_res_and_lock(lock);
1354 ldlm_lock2desc(lock->l_blocking_lock, &d);
1356 LDLM_LOCK_PUT(lock->l_blocking_lock);
1357 lock->l_blocking_lock = NULL;
1358 rc = lock->l_blocking_ast(lock, &d, (void *)&arg,
1360 LDLM_LOCK_PUT(lock);
1363 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1364 * and create a new set for requests that remained in
1366 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1367 ldlm_send_and_maybe_create_set(&arg, 1);
1373 ldlm_send_and_maybe_create_set(&arg, 0);
1375 /* In case when number of ASTs is multiply of
1376 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1377 * @arg.set must be destroyed here, otherwise we get
1378 * write memory leaking. */
1379 ptlrpc_set_destroy(arg.set);
1381 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1384 int ldlm_run_cp_ast_work(struct list_head *rpc_list)
1386 struct ldlm_cb_set_arg arg;
1387 struct list_head *tmp, *pos;
1392 arg.set = ptlrpc_prep_set();
1393 atomic_set(&arg.restart, 0);
1394 arg.type = LDLM_CP_CALLBACK;
1396 /* It's possible to receive a completion AST before we've set
1397 * the l_completion_ast pointer: either because the AST arrived
1398 * before the reply, or simply because there's a small race
1399 * window between receiving the reply and finishing the local
1400 * enqueue. (bug 842)
1402 * This can't happen with the blocking_ast, however, because we
1403 * will never call the local blocking_ast until we drop our
1404 * reader/writer reference, which we won't do until we get the
1405 * reply and finish enqueueing. */
1408 list_for_each_safe(tmp, pos, rpc_list) {
1409 struct ldlm_lock *lock =
1410 list_entry(tmp, struct ldlm_lock, l_cp_ast);
1411 ldlm_completion_callback completion_callback;
1413 /* nobody should touch l_cp_ast */
1414 lock_res_and_lock(lock);
1415 list_del_init(&lock->l_cp_ast);
1416 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1417 /* save l_completion_ast since it can be changed by
1418 * mds_intent_policy(), see bug 14225 */
1419 completion_callback = lock->l_completion_ast;
1420 lock->l_flags &= ~LDLM_FL_CP_REQD;
1421 unlock_res_and_lock(lock);
1423 if (completion_callback != NULL) {
1424 rc = completion_callback(lock, 0, (void *)&arg);
1427 LDLM_LOCK_PUT(lock);
1429 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1430 * and create a new set for requests that remained in
1432 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1433 ldlm_send_and_maybe_create_set(&arg, 1);
1439 ldlm_send_and_maybe_create_set(&arg, 0);
1441 /* In case when number of ASTs is multiply of
1442 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1443 * @arg.set must be destroyed here, otherwise we get
1444 * write memory leaking. */
1445 ptlrpc_set_destroy(arg.set);
1447 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1450 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1452 ldlm_reprocess_all(res);
1453 return LDLM_ITER_CONTINUE;
1456 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1458 struct list_head *tmp;
1462 spin_lock(&ns->ns_hash_lock);
1463 for (i = 0; i < RES_HASH_SIZE; i++) {
1464 tmp = ns->ns_hash[i].next;
1465 while (tmp != &(ns->ns_hash[i])) {
1466 struct ldlm_resource *res =
1467 list_entry(tmp, struct ldlm_resource, lr_hash);
1469 ldlm_resource_getref(res);
1470 spin_unlock(&ns->ns_hash_lock);
1472 rc = reprocess_one_queue(res, NULL);
1474 spin_lock(&ns->ns_hash_lock);
1476 ldlm_resource_putref_locked(res);
1478 if (rc == LDLM_ITER_STOP)
1483 spin_unlock(&ns->ns_hash_lock);
1487 void ldlm_reprocess_all(struct ldlm_resource *res)
1489 CFS_LIST_HEAD(rpc_list);
1493 /* Local lock trees don't get reprocessed. */
1494 if (ns_is_client(res->lr_namespace)) {
1501 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1502 if (rc == LDLM_ITER_CONTINUE)
1503 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1506 rc = ldlm_run_cp_ast_work(&rpc_list);
1507 if (rc == -ERESTART) {
1508 LASSERT(list_empty(&rpc_list));
1514 void ldlm_cancel_callback(struct ldlm_lock *lock)
1516 check_res_locked(lock->l_resource);
1517 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1518 lock->l_flags |= LDLM_FL_CANCEL;
1519 if (lock->l_blocking_ast) {
1520 // l_check_no_ns_lock(ns);
1521 unlock_res_and_lock(lock);
1522 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1524 lock_res_and_lock(lock);
1526 LDLM_DEBUG(lock, "no blocking ast");
1529 lock->l_flags |= LDLM_FL_BL_DONE;
1532 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1534 if (req->l_resource->lr_type != LDLM_PLAIN &&
1535 req->l_resource->lr_type != LDLM_IBITS)
1538 list_del_init(&req->l_sl_policy);
1539 list_del_init(&req->l_sl_mode);
1542 void ldlm_lock_cancel(struct ldlm_lock *lock)
1544 struct ldlm_resource *res;
1545 struct ldlm_namespace *ns;
1548 lock_res_and_lock(lock);
1550 res = lock->l_resource;
1551 ns = res->lr_namespace;
1553 /* Please do not, no matter how tempting, remove this LBUG without
1554 * talking to me first. -phik */
1555 if (lock->l_readers || lock->l_writers) {
1556 LDLM_ERROR(lock, "lock still has references");
1560 ldlm_del_waiting_lock(lock);
1562 /* Releases res lock */
1563 ldlm_cancel_callback(lock);
1565 /* Yes, second time, just in case it was added again while we were
1566 running with no res lock in ldlm_cancel_callback */
1567 ldlm_del_waiting_lock(lock);
1568 ldlm_resource_unlink_lock(lock);
1569 ldlm_lock_destroy_nolock(lock);
1571 if (lock->l_granted_mode == lock->l_req_mode)
1572 ldlm_pool_del(&ns->ns_pool, lock);
1574 /* Make sure we will not be called again for same lock what is possible
1575 * if not to zero out lock->l_granted_mode */
1576 lock->l_granted_mode = 0;
1577 unlock_res_and_lock(lock);
1582 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1584 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1590 lock->l_ast_data = data;
1591 LDLM_LOCK_PUT(lock);
1595 void ldlm_cancel_locks_for_export_cb(void *obj, void *data)
1597 struct obd_export *exp = data;
1598 struct ldlm_lock *lock = obj;
1599 struct ldlm_resource *res;
1601 res = ldlm_resource_getref(lock->l_resource);
1602 LDLM_LOCK_GET(lock);
1604 LDLM_DEBUG(lock, "export %p", exp);
1605 ldlm_res_lvbo_update(res, NULL, 0, 1);
1607 ldlm_lock_cancel(lock);
1608 ldlm_reprocess_all(res);
1610 ldlm_resource_putref(res);
1611 LDLM_LOCK_PUT(lock);
1614 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1616 lustre_hash_for_each_empty(exp->exp_lock_hash,
1617 ldlm_cancel_locks_for_export_cb, exp);
1620 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1623 CFS_LIST_HEAD(rpc_list);
1624 struct ldlm_resource *res;
1625 struct ldlm_namespace *ns;
1628 struct sl_insert_point prev;
1630 struct ldlm_interval *node;
1633 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1634 *flags |= LDLM_FL_BLOCK_GRANTED;
1635 RETURN(lock->l_resource);
1638 /* I can't check the type of lock here because the bitlock of lock
1639 * is not held here, so do the allocation blindly. -jay */
1640 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO, sizeof(*node));
1641 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1644 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1645 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1647 lock_res_and_lock(lock);
1649 res = lock->l_resource;
1650 ns = res->lr_namespace;
1652 old_mode = lock->l_req_mode;
1653 lock->l_req_mode = new_mode;
1654 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1655 /* remember the lock position where the lock might be
1656 * added back to the granted list later and also
1657 * remember the join mode for skiplist fixing. */
1658 prev.res_link = lock->l_res_link.prev;
1659 prev.mode_link = lock->l_sl_mode.prev;
1660 prev.policy_link = lock->l_sl_policy.prev;
1661 ldlm_resource_unlink_lock(lock);
1663 ldlm_resource_unlink_lock(lock);
1664 if (res->lr_type == LDLM_EXTENT) {
1665 /* FIXME: ugly code, I have to attach the lock to a
1666 * interval node again since perhaps it will be granted
1668 CFS_INIT_LIST_HEAD(&node->li_group);
1669 ldlm_interval_attach(node, lock);
1674 /* If this is a local resource, put it on the appropriate list. */
1675 if (ns_is_client(res->lr_namespace)) {
1676 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1677 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1679 /* This should never happen, because of the way the
1680 * server handles conversions. */
1681 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1685 ldlm_grant_lock(lock, &rpc_list);
1687 /* FIXME: completion handling not with ns_lock held ! */
1688 if (lock->l_completion_ast)
1689 lock->l_completion_ast(lock, 0, NULL);
1693 ldlm_processing_policy policy;
1694 policy = ldlm_processing_policy_table[res->lr_type];
1695 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1696 if (rc == LDLM_ITER_STOP) {
1697 lock->l_req_mode = old_mode;
1698 if (res->lr_type == LDLM_EXTENT)
1699 ldlm_extent_add_lock(res, lock);
1701 ldlm_granted_list_add_lock(lock, &prev);
1705 *flags |= LDLM_FL_BLOCK_GRANTED;
1709 unlock_res_and_lock(lock);
1712 ldlm_run_cp_ast_work(&rpc_list);
1714 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1718 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1720 struct obd_device *obd = NULL;
1722 if (!((libcfs_debug | D_ERROR) & level))
1726 CDEBUG(level, " NULL LDLM lock\n");
1730 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1731 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1733 if (lock->l_conn_export != NULL)
1734 obd = lock->l_conn_export->exp_obd;
1735 if (lock->l_export && lock->l_export->exp_connection) {
1736 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1737 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1738 lock->l_remote_handle.cookie);
1739 } else if (obd == NULL) {
1740 CDEBUG(level, " Node: local\n");
1742 struct obd_import *imp = obd->u.cli.cl_import;
1743 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1744 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1745 lock->l_remote_handle.cookie);
1747 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource,
1748 lock->l_resource->lr_name.name[0],
1749 lock->l_resource->lr_name.name[1]);
1750 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1751 "write: %d flags: "LPX64"\n", ldlm_lockname[lock->l_req_mode],
1752 ldlm_lockname[lock->l_granted_mode],
1753 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1755 if (lock->l_resource->lr_type == LDLM_EXTENT)
1756 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1757 " (req "LPU64"-"LPU64")\n",
1758 lock->l_policy_data.l_extent.start,
1759 lock->l_policy_data.l_extent.end,
1760 lock->l_req_extent.start, lock->l_req_extent.end);
1761 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1762 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1763 lock->l_policy_data.l_flock.pid,
1764 lock->l_policy_data.l_flock.start,
1765 lock->l_policy_data.l_flock.end);
1766 else if (lock->l_resource->lr_type == LDLM_IBITS)
1767 CDEBUG(level, " Bits: "LPX64"\n",
1768 lock->l_policy_data.l_inodebits.bits);
1771 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1773 struct ldlm_lock *lock;
1775 if (!((libcfs_debug | D_ERROR) & level))
1778 lock = ldlm_handle2lock(lockh);
1782 ldlm_lock_dump(D_OTHER, lock, 0);
1784 LDLM_LOCK_PUT(lock);
1787 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1788 struct libcfs_debug_msg_data *data, const char *fmt,
1792 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1794 va_start(args, fmt);
1795 if (lock->l_resource == NULL) {
1796 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1797 data->msg_fn, data->msg_line, fmt, args,
1798 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1799 "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" remote: "
1800 LPX64" expref: %d pid: %u timeout: %lu\n", lock,
1801 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1802 lock->l_readers, lock->l_writers,
1803 ldlm_lockname[lock->l_granted_mode],
1804 ldlm_lockname[lock->l_req_mode],
1805 lock->l_flags, lock->l_remote_handle.cookie,
1807 atomic_read(&lock->l_export->exp_refcount) : -99,
1808 lock->l_pid, lock->l_callback_timeout);
1813 switch (lock->l_resource->lr_type) {
1815 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1816 data->msg_fn, data->msg_line, fmt, args,
1817 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1818 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1819 "] (req "LPU64"->"LPU64") flags: "LPX64" remote: "LPX64
1820 " expref: %d pid: %u timeout %lu\n",
1821 lock->l_resource->lr_namespace->ns_name, lock,
1822 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1823 lock->l_readers, lock->l_writers,
1824 ldlm_lockname[lock->l_granted_mode],
1825 ldlm_lockname[lock->l_req_mode],
1826 lock->l_resource->lr_name.name[0],
1827 lock->l_resource->lr_name.name[1],
1828 atomic_read(&lock->l_resource->lr_refcount),
1829 ldlm_typename[lock->l_resource->lr_type],
1830 lock->l_policy_data.l_extent.start,
1831 lock->l_policy_data.l_extent.end,
1832 lock->l_req_extent.start, lock->l_req_extent.end,
1833 lock->l_flags, lock->l_remote_handle.cookie,
1835 atomic_read(&lock->l_export->exp_refcount) : -99,
1836 lock->l_pid, lock->l_callback_timeout);
1839 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1840 data->msg_fn, data->msg_line, fmt, args,
1841 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1842 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1843 "["LPU64"->"LPU64"] flags: "LPX64" remote: "LPX64
1844 " expref: %d pid: %u timeout: %lu\n",
1845 lock->l_resource->lr_namespace->ns_name, lock,
1846 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1847 lock->l_readers, lock->l_writers,
1848 ldlm_lockname[lock->l_granted_mode],
1849 ldlm_lockname[lock->l_req_mode],
1850 lock->l_resource->lr_name.name[0],
1851 lock->l_resource->lr_name.name[1],
1852 atomic_read(&lock->l_resource->lr_refcount),
1853 ldlm_typename[lock->l_resource->lr_type],
1854 lock->l_policy_data.l_flock.pid,
1855 lock->l_policy_data.l_flock.start,
1856 lock->l_policy_data.l_flock.end,
1857 lock->l_flags, lock->l_remote_handle.cookie,
1859 atomic_read(&lock->l_export->exp_refcount) : -99,
1860 lock->l_pid, lock->l_callback_timeout);
1863 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1864 data->msg_fn, data->msg_line, fmt, args,
1865 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1866 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1867 "flags: "LPX64" remote: "LPX64" expref: %d "
1868 "pid: %u timeout: %lu\n",
1869 lock->l_resource->lr_namespace->ns_name,
1870 lock, lock->l_handle.h_cookie,
1871 atomic_read (&lock->l_refc),
1872 lock->l_readers, lock->l_writers,
1873 ldlm_lockname[lock->l_granted_mode],
1874 ldlm_lockname[lock->l_req_mode],
1875 lock->l_resource->lr_name.name[0],
1876 lock->l_resource->lr_name.name[1],
1877 lock->l_policy_data.l_inodebits.bits,
1878 atomic_read(&lock->l_resource->lr_refcount),
1879 ldlm_typename[lock->l_resource->lr_type],
1880 lock->l_flags, lock->l_remote_handle.cookie,
1882 atomic_read(&lock->l_export->exp_refcount) : -99,
1883 lock->l_pid, lock->l_callback_timeout);
1886 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1887 data->msg_fn, data->msg_line, fmt, args,
1888 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1889 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" "
1890 "remote: "LPX64" expref: %d pid: %u timeout %lu\n",
1891 lock->l_resource->lr_namespace->ns_name,
1892 lock, lock->l_handle.h_cookie,
1893 atomic_read (&lock->l_refc),
1894 lock->l_readers, lock->l_writers,
1895 ldlm_lockname[lock->l_granted_mode],
1896 ldlm_lockname[lock->l_req_mode],
1897 lock->l_resource->lr_name.name[0],
1898 lock->l_resource->lr_name.name[1],
1899 atomic_read(&lock->l_resource->lr_refcount),
1900 ldlm_typename[lock->l_resource->lr_type],
1901 lock->l_flags, lock->l_remote_handle.cookie,
1903 atomic_read(&lock->l_export->exp_refcount) : -99,
1904 lock->l_pid, lock->l_callback_timeout);
1909 EXPORT_SYMBOL(_ldlm_lock_debug);