1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_lock.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_LDLM
45 # include <libcfs/libcfs.h>
46 # ifndef HAVE_VFS_INTENT_PATCHES
47 # include <linux/lustre_intent.h>
50 # include <liblustre.h>
53 #include <obd_class.h>
54 #include "ldlm_internal.h"
57 char *ldlm_lockname[] = {
69 char *ldlm_typename[] = {
76 char *ldlm_it2str(int it)
83 case (IT_OPEN | IT_CREAT):
96 CERROR("Unknown intent %d\n", it);
101 extern cfs_mem_cache_t *ldlm_lock_slab;
103 static ldlm_processing_policy ldlm_processing_policy_table[] = {
104 [LDLM_PLAIN] ldlm_process_plain_lock,
105 [LDLM_EXTENT] ldlm_process_extent_lock,
107 [LDLM_FLOCK] ldlm_process_flock_lock,
109 [LDLM_IBITS] ldlm_process_inodebits_lock,
112 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
114 return ldlm_processing_policy_table[res->lr_type];
117 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
123 * REFCOUNTED LOCK OBJECTS
128 * Lock refcounts, during creation:
129 * - one special one for allocation, dec'd only once in destroy
130 * - one for being a lock that's in-use
131 * - one for the addref associated with a new lock
133 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
135 atomic_inc(&lock->l_refc);
139 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
141 LASSERT(size == sizeof(*lock));
142 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
145 void ldlm_lock_put(struct ldlm_lock *lock)
149 LASSERT(lock->l_resource != LP_POISON);
150 LASSERT(atomic_read(&lock->l_refc) > 0);
151 if (atomic_dec_and_test(&lock->l_refc)) {
152 struct ldlm_resource *res;
155 "final lock_put on destroyed lock, freeing it.");
157 res = lock->l_resource;
158 LASSERT(lock->l_destroyed);
159 LASSERT(list_empty(&lock->l_res_link));
160 LASSERT(list_empty(&lock->l_pending_chain));
162 atomic_dec(&res->lr_namespace->ns_locks);
163 lu_ref_del(&res->lr_reference, "lock", lock);
164 ldlm_resource_putref(res);
165 lock->l_resource = NULL;
166 if (lock->l_export) {
167 class_export_put(lock->l_export);
168 lock->l_export = NULL;
171 if (lock->l_lvb_data != NULL)
172 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
174 ldlm_interval_free(ldlm_interval_detach(lock));
175 lu_ref_fini(&lock->l_reference);
176 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
183 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
186 if (!list_empty(&lock->l_lru)) {
187 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
188 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
189 list_del_init(&lock->l_lru);
191 LASSERT(ns->ns_nr_unused >= 0);
197 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
199 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
202 spin_lock(&ns->ns_unused_lock);
203 rc = ldlm_lock_remove_from_lru_nolock(lock);
204 spin_unlock(&ns->ns_unused_lock);
209 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
211 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
212 lock->l_last_used = cfs_time_current();
213 LASSERT(list_empty(&lock->l_lru));
214 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
215 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
216 LASSERT(ns->ns_nr_unused >= 0);
220 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
222 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
224 spin_lock(&ns->ns_unused_lock);
225 ldlm_lock_add_to_lru_nolock(lock);
226 spin_unlock(&ns->ns_unused_lock);
230 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
232 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
234 spin_lock(&ns->ns_unused_lock);
235 if (!list_empty(&lock->l_lru)) {
236 ldlm_lock_remove_from_lru_nolock(lock);
237 ldlm_lock_add_to_lru_nolock(lock);
239 spin_unlock(&ns->ns_unused_lock);
243 /* This used to have a 'strict' flag, which recovery would use to mark an
244 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
245 * shall explain why it's gone: with the new hash table scheme, once you call
246 * ldlm_lock_destroy, you can never drop your final references on this lock.
247 * Because it's not in the hash table anymore. -phil */
248 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
252 if (lock->l_readers || lock->l_writers) {
253 LDLM_ERROR(lock, "lock still has references");
254 ldlm_lock_dump(D_ERROR, lock, 0);
258 if (!list_empty(&lock->l_res_link)) {
259 LDLM_ERROR(lock, "lock still on resource");
260 ldlm_lock_dump(D_ERROR, lock, 0);
264 if (lock->l_destroyed) {
265 LASSERT(list_empty(&lock->l_lru));
269 lock->l_destroyed = 1;
271 if (lock->l_export && lock->l_export->exp_lock_hash &&
272 !hlist_unhashed(&lock->l_exp_hash))
273 lustre_hash_del(lock->l_export->exp_lock_hash,
274 &lock->l_remote_handle, &lock->l_exp_hash);
276 ldlm_lock_remove_from_lru(lock);
277 class_handle_unhash(&lock->l_handle);
280 /* Wake anyone waiting for this lock */
281 /* FIXME: I should probably add yet another flag, instead of using
282 * l_export to only call this on clients */
284 class_export_put(lock->l_export);
285 lock->l_export = NULL;
286 if (lock->l_export && lock->l_completion_ast)
287 lock->l_completion_ast(lock, 0);
293 void ldlm_lock_destroy(struct ldlm_lock *lock)
297 lock_res_and_lock(lock);
298 first = ldlm_lock_destroy_internal(lock);
299 unlock_res_and_lock(lock);
301 /* drop reference from hashtable only for first destroy */
303 lu_ref_del(&lock->l_reference, "hash", lock);
304 LDLM_LOCK_RELEASE(lock);
309 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
313 first = ldlm_lock_destroy_internal(lock);
314 /* drop reference from hashtable only for first destroy */
316 lu_ref_del(&lock->l_reference, "hash", lock);
317 LDLM_LOCK_RELEASE(lock);
322 /* this is called by portals_handle2object with the handle lock taken */
323 static void lock_handle_addref(void *lock)
325 LDLM_LOCK_GET((struct ldlm_lock *)lock);
329 * usage: pass in a resource on which you have done ldlm_resource_get
330 * pass in a parent lock on which you have done a ldlm_lock_get
331 * after return, ldlm_*_put the resource and parent
332 * returns: lock with refcount 2 - one for current caller and one for remote
334 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
336 struct ldlm_lock *lock;
339 if (resource == NULL)
342 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, CFS_ALLOC_IO, sizeof(*lock));
346 spin_lock_init(&lock->l_lock);
347 lock->l_resource = ldlm_resource_getref(resource);
348 lu_ref_add(&resource->lr_reference, "lock", lock);
350 atomic_set(&lock->l_refc, 2);
351 CFS_INIT_LIST_HEAD(&lock->l_res_link);
352 CFS_INIT_LIST_HEAD(&lock->l_lru);
353 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
354 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
355 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
356 CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
357 cfs_waitq_init(&lock->l_waitq);
358 lock->l_blocking_lock = NULL;
359 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
360 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
361 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
363 atomic_inc(&resource->lr_namespace->ns_locks);
364 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
365 class_handle_hash(&lock->l_handle, lock_handle_addref);
367 CFS_INIT_LIST_HEAD(&lock->l_extents_list);
368 spin_lock_init(&lock->l_extents_list_lock);
369 CFS_INIT_LIST_HEAD(&lock->l_cache_locks_list);
370 lu_ref_init(&lock->l_reference);
371 lu_ref_add(&lock->l_reference, "hash", lock);
376 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
377 const struct ldlm_res_id *new_resid)
379 struct ldlm_resource *oldres = lock->l_resource;
380 struct ldlm_resource *newres;
384 LASSERT(ns_is_client(ns));
386 lock_res_and_lock(lock);
387 if (memcmp(new_resid, &lock->l_resource->lr_name,
388 sizeof(lock->l_resource->lr_name)) == 0) {
390 unlock_res_and_lock(lock);
394 LASSERT(new_resid->name[0] != 0);
396 /* This function assumes that the lock isn't on any lists */
397 LASSERT(list_empty(&lock->l_res_link));
399 type = oldres->lr_type;
400 unlock_res_and_lock(lock);
402 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
403 lu_ref_add(&newres->lr_reference, "lock", lock);
407 * To flip the lock from the old to the new resource, lock, oldres and
408 * newres have to be locked. Resource spin-locks are nested within
409 * lock->l_lock, and are taken in the memory address order to avoid
412 spin_lock(&lock->l_lock);
413 oldres = lock->l_resource;
414 if (oldres < newres) {
416 lock_res_nested(newres, LRT_NEW);
419 lock_res_nested(oldres, LRT_NEW);
421 LASSERT(memcmp(new_resid, &oldres->lr_name,
422 sizeof oldres->lr_name) != 0);
423 lock->l_resource = newres;
425 unlock_res_and_lock(lock);
427 /* ...and the flowers are still standing! */
428 lu_ref_del(&oldres->lr_reference, "lock", lock);
429 ldlm_resource_putref(oldres);
438 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
440 lockh->cookie = lock->l_handle.h_cookie;
443 /* if flags: atomically get the lock and set the flags.
444 * Return NULL if flag already set
447 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
450 struct ldlm_namespace *ns;
451 struct ldlm_lock *lock, *retval = NULL;
456 lock = class_handle2object(handle->cookie);
460 LASSERT(lock->l_resource != NULL);
461 ns = lock->l_resource->lr_namespace;
464 lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
465 lock_res_and_lock(lock);
467 /* It's unlikely but possible that someone marked the lock as
468 * destroyed after we did handle2object on it */
469 if (lock->l_destroyed) {
470 unlock_res_and_lock(lock);
471 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
476 if (flags && (lock->l_flags & flags)) {
477 unlock_res_and_lock(lock);
483 lock->l_flags |= flags;
485 unlock_res_and_lock(lock);
492 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
494 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
495 /* INODEBITS_INTEROP: If the other side does not support
496 * inodebits, reply with a plain lock descriptor.
498 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
499 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
500 struct ldlm_resource res = *lock->l_resource;
502 /* Make sure all the right bits are set in this lock we
503 are going to pass to client */
504 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
505 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
506 "Inappropriate inode lock bits during "
507 "conversion " LPU64 "\n",
508 lock->l_policy_data.l_inodebits.bits);
509 res.lr_type = LDLM_PLAIN;
510 ldlm_res2desc(&res, &desc->l_resource);
511 /* Convert "new" lock mode to something old client can
513 if ((lock->l_req_mode == LCK_CR) ||
514 (lock->l_req_mode == LCK_CW))
515 desc->l_req_mode = LCK_PR;
517 desc->l_req_mode = lock->l_req_mode;
518 if ((lock->l_granted_mode == LCK_CR) ||
519 (lock->l_granted_mode == LCK_CW)) {
520 desc->l_granted_mode = LCK_PR;
522 /* We never grant PW/EX locks to clients */
523 LASSERT((lock->l_granted_mode != LCK_PW) &&
524 (lock->l_granted_mode != LCK_EX));
525 desc->l_granted_mode = lock->l_granted_mode;
528 /* We do not copy policy here, because there is no
529 policy for plain locks */
531 ldlm_res2desc(lock->l_resource, &desc->l_resource);
532 desc->l_req_mode = lock->l_req_mode;
533 desc->l_granted_mode = lock->l_granted_mode;
534 desc->l_policy_data = lock->l_policy_data;
538 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
539 struct list_head *work_list)
541 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
542 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
543 lock->l_flags |= LDLM_FL_AST_SENT;
544 /* If the enqueuing client said so, tell the AST recipient to
545 * discard dirty data, rather than writing back. */
546 if (new->l_flags & LDLM_AST_DISCARD_DATA)
547 lock->l_flags |= LDLM_FL_DISCARD_DATA;
548 LASSERT(list_empty(&lock->l_bl_ast));
549 list_add(&lock->l_bl_ast, work_list);
551 LASSERT(lock->l_blocking_lock == NULL);
552 lock->l_blocking_lock = LDLM_LOCK_GET(new);
556 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
558 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
559 lock->l_flags |= LDLM_FL_CP_REQD;
560 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
561 LASSERT(list_empty(&lock->l_cp_ast));
562 list_add(&lock->l_cp_ast, work_list);
567 /* must be called with lr_lock held */
568 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
569 struct list_head *work_list)
572 check_res_locked(lock->l_resource);
574 ldlm_add_bl_work_item(lock, new, work_list);
576 ldlm_add_cp_work_item(lock, work_list);
580 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
582 struct ldlm_lock *lock;
584 lock = ldlm_handle2lock(lockh);
585 LASSERT(lock != NULL);
586 ldlm_lock_addref_internal(lock, mode);
590 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
592 ldlm_lock_remove_from_lru(lock);
593 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
595 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
597 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
599 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
602 lu_ref_add_atomic(&lock->l_reference, "user", lock);
603 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
607 * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
610 * \retval 0 success, lock was addref-ed
612 * \retval -EAGAIN lock is being canceled.
614 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
616 struct ldlm_lock *lock;
620 lock = ldlm_handle2lock(lockh);
622 lock_res_and_lock(lock);
623 if (lock->l_readers != 0 || lock->l_writers != 0 ||
624 !(lock->l_flags & LDLM_FL_CBPENDING)) {
625 ldlm_lock_addref_internal_nolock(lock, mode);
628 unlock_res_and_lock(lock);
634 /* only called for local locks */
635 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
637 lock_res_and_lock(lock);
638 ldlm_lock_addref_internal_nolock(lock, mode);
639 unlock_res_and_lock(lock);
642 /* only called in ldlm_flock_destroy and for local locks.
643 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
644 * * ldlm_lock_remove_from_lru() does nothing, it is safe
645 * * for ldlm_flock_destroy usage by dropping some code */
646 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
648 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
649 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
650 LASSERT(lock->l_readers > 0);
651 lu_ref_del(&lock->l_reference, "reader", lock);
654 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
655 LASSERT(lock->l_writers > 0);
656 lu_ref_del(&lock->l_reference, "writer", lock);
660 lu_ref_del(&lock->l_reference, "user", lock);
661 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
664 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
666 struct ldlm_namespace *ns;
669 lock_res_and_lock(lock);
671 ns = lock->l_resource->lr_namespace;
673 ldlm_lock_decref_internal_nolock(lock, mode);
675 if (lock->l_flags & LDLM_FL_LOCAL &&
676 !lock->l_readers && !lock->l_writers) {
677 /* If this is a local lock on a server namespace and this was
678 * the last reference, cancel the lock. */
679 CDEBUG(D_INFO, "forcing cancel of local lock\n");
680 lock->l_flags |= LDLM_FL_CBPENDING;
683 if (!lock->l_readers && !lock->l_writers &&
684 (lock->l_flags & LDLM_FL_CBPENDING)) {
685 /* If we received a blocked AST and this was the last reference,
686 * run the callback. */
687 if (ns_is_server(ns) && lock->l_export)
688 CERROR("FL_CBPENDING set on non-local lock--just a "
691 LDLM_DEBUG(lock, "final decref done on cbpending lock");
693 LDLM_LOCK_GET(lock); /* dropped by bl thread */
694 ldlm_lock_remove_from_lru(lock);
695 unlock_res_and_lock(lock);
696 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
697 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
698 ldlm_handle_bl_callback(ns, NULL, lock);
699 } else if (ns_is_client(ns) &&
700 !lock->l_readers && !lock->l_writers &&
701 !(lock->l_flags & LDLM_FL_BL_AST)) {
702 /* If this is a client-side namespace and this was the last
703 * reference, put it on the LRU. */
704 ldlm_lock_add_to_lru(lock);
705 unlock_res_and_lock(lock);
706 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
707 * are not supported by the server, otherwise, it is done on
709 if (!exp_connect_cancelset(lock->l_conn_export) &&
710 !ns_connect_lru_resize(ns))
711 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
713 unlock_res_and_lock(lock);
719 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
721 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
722 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
723 ldlm_lock_decref_internal(lock, mode);
727 /* This will drop a lock reference and mark it for destruction, but will not
728 * necessarily cancel the lock before returning. */
729 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
731 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
734 LASSERT(lock != NULL);
736 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
737 lock_res_and_lock(lock);
738 lock->l_flags |= LDLM_FL_CBPENDING;
739 unlock_res_and_lock(lock);
740 ldlm_lock_decref_internal(lock, mode);
744 struct sl_insert_point {
745 struct list_head *res_link;
746 struct list_head *mode_link;
747 struct list_head *policy_link;
751 * search_granted_lock
754 * Finds a position to insert the new lock.
756 * queue [input]: the granted list where search acts on;
757 * req [input]: the lock whose position to be located;
758 * prev [output]: positions within 3 lists to insert @req to
762 * - ldlm_grant_lock_with_skiplist
764 static void search_granted_lock(struct list_head *queue,
765 struct ldlm_lock *req,
766 struct sl_insert_point *prev)
768 struct list_head *tmp;
769 struct ldlm_lock *lock, *mode_end, *policy_end;
772 list_for_each(tmp, queue) {
773 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
775 mode_end = list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
778 if (lock->l_req_mode != req->l_req_mode) {
779 /* jump to last lock of mode group */
780 tmp = &mode_end->l_res_link;
784 /* suitable mode group is found */
785 if (lock->l_resource->lr_type == LDLM_PLAIN) {
786 /* insert point is last lock of the mode group */
787 prev->res_link = &mode_end->l_res_link;
788 prev->mode_link = &mode_end->l_sl_mode;
789 prev->policy_link = &req->l_sl_policy;
792 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
794 policy_end = list_entry(lock->l_sl_policy.prev,
798 if (lock->l_policy_data.l_inodebits.bits ==
799 req->l_policy_data.l_inodebits.bits) {
800 /* insert point is last lock of
801 * the policy group */
803 &policy_end->l_res_link;
805 &policy_end->l_sl_mode;
807 &policy_end->l_sl_policy;
812 if (policy_end == mode_end)
813 /* done with mode group */
816 /* go to next policy group within mode group */
817 tmp = policy_end->l_res_link.next;
818 lock = list_entry(tmp, struct ldlm_lock,
820 } /* loop over policy groups within the mode group */
822 /* insert point is last lock of the mode group,
823 * new policy group is started */
824 prev->res_link = &mode_end->l_res_link;
825 prev->mode_link = &mode_end->l_sl_mode;
826 prev->policy_link = &req->l_sl_policy;
830 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
835 /* insert point is last lock on the queue,
836 * new mode group and new policy group are started */
837 prev->res_link = queue->prev;
838 prev->mode_link = &req->l_sl_mode;
839 prev->policy_link = &req->l_sl_policy;
844 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
845 struct sl_insert_point *prev)
847 struct ldlm_resource *res = lock->l_resource;
850 check_res_locked(res);
852 ldlm_resource_dump(D_INFO, res);
853 CDEBUG(D_OTHER, "About to add this lock:\n");
854 ldlm_lock_dump(D_OTHER, lock, 0);
856 if (lock->l_destroyed) {
857 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
861 LASSERT(list_empty(&lock->l_res_link));
862 LASSERT(list_empty(&lock->l_sl_mode));
863 LASSERT(list_empty(&lock->l_sl_policy));
865 list_add(&lock->l_res_link, prev->res_link);
866 list_add(&lock->l_sl_mode, prev->mode_link);
867 list_add(&lock->l_sl_policy, prev->policy_link);
872 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
874 struct sl_insert_point prev;
877 LASSERT(lock->l_req_mode == lock->l_granted_mode);
879 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
880 ldlm_granted_list_add_lock(lock, &prev);
885 * - ldlm_lock_enqueue
886 * - ldlm_reprocess_queue
887 * - ldlm_lock_convert
889 * must be called with lr_lock held
891 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
893 struct ldlm_resource *res = lock->l_resource;
896 check_res_locked(res);
898 lock->l_granted_mode = lock->l_req_mode;
899 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
900 ldlm_grant_lock_with_skiplist(lock);
901 else if (res->lr_type == LDLM_EXTENT)
902 ldlm_extent_add_lock(res, lock);
904 ldlm_resource_add_lock(res, &res->lr_granted, lock);
906 if (lock->l_granted_mode < res->lr_most_restr)
907 res->lr_most_restr = lock->l_granted_mode;
909 if (work_list && lock->l_completion_ast != NULL)
910 ldlm_add_ast_work_item(lock, NULL, work_list);
912 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
916 /* returns a referenced lock or NULL. See the flag descriptions below, in the
917 * comment above ldlm_lock_match */
918 static struct ldlm_lock *search_queue(struct list_head *queue,
920 ldlm_policy_data_t *policy,
921 struct ldlm_lock *old_lock,
922 int flags, int unref)
924 struct ldlm_lock *lock;
925 struct list_head *tmp;
927 list_for_each(tmp, queue) {
930 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
932 if (lock == old_lock)
935 /* llite sometimes wants to match locks that will be
936 * canceled when their users drop, but we allow it to match
937 * if it passes in CBPENDING and the lock still has users.
938 * this is generally only going to be used by children
939 * whose parents already hold a lock so forward progress
940 * can still happen. */
941 if (lock->l_flags & LDLM_FL_CBPENDING &&
942 !(flags & LDLM_FL_CBPENDING))
944 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
945 lock->l_readers == 0 && lock->l_writers == 0)
948 if (!(lock->l_req_mode & *mode))
950 match = lock->l_req_mode;
952 if (lock->l_resource->lr_type == LDLM_EXTENT &&
953 (lock->l_policy_data.l_extent.start >
954 policy->l_extent.start ||
955 lock->l_policy_data.l_extent.end < policy->l_extent.end))
958 if (unlikely(match == LCK_GROUP) &&
959 lock->l_resource->lr_type == LDLM_EXTENT &&
960 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
963 /* We match if we have existing lock with same or wider set
965 if (lock->l_resource->lr_type == LDLM_IBITS &&
966 ((lock->l_policy_data.l_inodebits.bits &
967 policy->l_inodebits.bits) !=
968 policy->l_inodebits.bits))
972 (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED)))
975 if ((flags & LDLM_FL_LOCAL_ONLY) &&
976 !(lock->l_flags & LDLM_FL_LOCAL))
979 if (flags & LDLM_FL_TEST_LOCK) {
981 ldlm_lock_touch_in_lru(lock);
983 ldlm_lock_addref_internal_nolock(lock, match);
992 void ldlm_lock_allow_match(struct ldlm_lock *lock)
994 lock_res_and_lock(lock);
995 lock->l_flags |= LDLM_FL_LVB_READY;
996 cfs_waitq_signal(&lock->l_waitq);
997 unlock_res_and_lock(lock);
1000 /* Can be called in two ways:
1002 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1003 * for a duplicate of.
1005 * Otherwise, all of the fields must be filled in, to match against.
1007 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1008 * server (ie, connh is NULL)
1009 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1010 * list will be considered
1011 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1012 * to be canceled can still be matched as long as they still have reader
1013 * or writer refernces
1014 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1015 * just tell us if we would have matched.
1017 * Returns 1 if it finds an already-existing lock that is compatible; in this
1018 * case, lockh is filled in with a addref()ed lock
1020 * we also check security context, if that failed we simply return 0 (to keep
1021 * caller code unchanged), the context failure will be discovered by caller
1024 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1025 const struct ldlm_res_id *res_id, ldlm_type_t type,
1026 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1027 struct lustre_handle *lockh, int unref)
1029 struct ldlm_resource *res;
1030 struct ldlm_lock *lock, *old_lock = NULL;
1035 old_lock = ldlm_handle2lock(lockh);
1038 ns = old_lock->l_resource->lr_namespace;
1039 res_id = &old_lock->l_resource->lr_name;
1040 type = old_lock->l_resource->lr_type;
1041 mode = old_lock->l_req_mode;
1044 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1046 LASSERT(old_lock == NULL);
1050 LDLM_RESOURCE_ADDREF(res);
1053 lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1057 if (flags & LDLM_FL_BLOCK_GRANTED)
1059 lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1063 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1071 LDLM_RESOURCE_DELREF(res);
1072 ldlm_resource_putref(res);
1075 ldlm_lock2handle(lock, lockh);
1076 if ((flags & LDLM_FL_LVB_READY) &&
1077 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1078 struct l_wait_info lwi;
1079 if (lock->l_completion_ast) {
1080 int err = lock->l_completion_ast(lock,
1081 LDLM_FL_WAIT_NOREPROC,
1084 if (flags & LDLM_FL_TEST_LOCK)
1085 LDLM_LOCK_RELEASE(lock);
1087 ldlm_lock_decref_internal(lock,
1094 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1095 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1097 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1098 l_wait_event(lock->l_waitq,
1099 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1104 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1105 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1106 res_id->name[2] : policy->l_extent.start,
1107 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1108 res_id->name[3] : policy->l_extent.end);
1110 /* check user's security context */
1111 if (lock->l_conn_export &&
1112 sptlrpc_import_check_ctx(
1113 class_exp2cliimp(lock->l_conn_export))) {
1114 if (!(flags & LDLM_FL_TEST_LOCK))
1115 ldlm_lock_decref_internal(lock, mode);
1119 if (flags & LDLM_FL_TEST_LOCK)
1120 LDLM_LOCK_RELEASE(lock);
1122 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1123 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1124 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1125 type, mode, res_id->name[0], res_id->name[1],
1126 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1127 res_id->name[2] :policy->l_extent.start,
1128 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1129 res_id->name[3] : policy->l_extent.end);
1132 LDLM_LOCK_PUT(old_lock);
1134 return rc ? mode : 0;
1137 /* Returns a referenced lock */
1138 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1139 const struct ldlm_res_id *res_id,
1142 const struct ldlm_callback_suite *cbs,
1143 void *data, __u32 lvb_len)
1145 struct ldlm_lock *lock;
1146 struct ldlm_resource *res;
1149 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1153 lock = ldlm_lock_new(res);
1154 ldlm_resource_putref(res);
1159 lock->l_req_mode = mode;
1160 lock->l_ast_data = data;
1161 lock->l_pid = cfs_curproc_pid();
1163 lock->l_blocking_ast = cbs->lcs_blocking;
1164 lock->l_completion_ast = cbs->lcs_completion;
1165 lock->l_glimpse_ast = cbs->lcs_glimpse;
1166 lock->l_weigh_ast = cbs->lcs_weigh;
1169 lock->l_tree_node = NULL;
1170 /* if this is the extent lock, allocate the interval tree node */
1171 if (type == LDLM_EXTENT) {
1172 if (ldlm_interval_alloc(lock) == NULL)
1177 lock->l_lvb_len = lvb_len;
1178 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1179 if (lock->l_lvb_data == NULL)
1186 if (lock->l_lvb_data)
1187 OBD_FREE(lock->l_lvb_data, lvb_len);
1188 ldlm_interval_free(ldlm_interval_detach(lock));
1189 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1193 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1194 struct ldlm_lock **lockp,
1195 void *cookie, int *flags)
1197 struct ldlm_lock *lock = *lockp;
1198 struct ldlm_resource *res = lock->l_resource;
1199 int local = ns_is_client(res->lr_namespace);
1200 ldlm_processing_policy policy;
1201 ldlm_error_t rc = ELDLM_OK;
1202 struct ldlm_interval *node = NULL;
1205 do_gettimeofday(&lock->l_enqueued_time);
1206 /* policies are not executed on the client or during replay */
1207 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1208 && !local && ns->ns_policy) {
1209 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1211 if (rc == ELDLM_LOCK_REPLACED) {
1212 /* The lock that was returned has already been granted,
1213 * and placed into lockp. If it's not the same as the
1214 * one we passed in, then destroy the old one and our
1215 * work here is done. */
1216 if (lock != *lockp) {
1217 ldlm_lock_destroy(lock);
1218 LDLM_LOCK_RELEASE(lock);
1220 *flags |= LDLM_FL_LOCK_CHANGED;
1222 } else if (rc != ELDLM_OK ||
1223 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1224 ldlm_lock_destroy(lock);
1229 /* For a replaying lock, it might be already in granted list. So
1230 * unlinking the lock will cause the interval node to be freed, we
1231 * have to allocate the interval node early otherwise we can't regrant
1232 * this lock in the future. - jay */
1233 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1234 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO,
1237 lock_res_and_lock(lock);
1238 if (local && lock->l_req_mode == lock->l_granted_mode) {
1239 /* The server returned a blocked lock, but it was granted
1240 * before we got a chance to actually enqueue it. We don't
1241 * need to do anything else. */
1242 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1243 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1244 GOTO(out, ELDLM_OK);
1247 ldlm_resource_unlink_lock(lock);
1248 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1250 ldlm_lock_destroy_nolock(lock);
1251 GOTO(out, rc = -ENOMEM);
1254 CFS_INIT_LIST_HEAD(&node->li_group);
1255 ldlm_interval_attach(node, lock);
1259 /* Some flags from the enqueue want to make it into the AST, via the
1260 * lock's l_flags. */
1261 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1263 /* This distinction between local lock trees is very important; a client
1264 * namespace only has information about locks taken by that client, and
1265 * thus doesn't have enough information to decide for itself if it can
1266 * be granted (below). In this case, we do exactly what the server
1267 * tells us to do, as dictated by the 'flags'.
1269 * We do exactly the same thing during recovery, when the server is
1270 * more or less trusting the clients not to lie.
1272 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1273 * granted/converting queues. */
1275 if (*flags & LDLM_FL_BLOCK_CONV)
1276 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1277 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1278 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1280 ldlm_grant_lock(lock, NULL);
1281 GOTO(out, ELDLM_OK);
1282 } else if (*flags & LDLM_FL_REPLAY) {
1283 if (*flags & LDLM_FL_BLOCK_CONV) {
1284 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1285 GOTO(out, ELDLM_OK);
1286 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1287 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1288 GOTO(out, ELDLM_OK);
1289 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1290 ldlm_grant_lock(lock, NULL);
1291 GOTO(out, ELDLM_OK);
1293 /* If no flags, fall through to normal enqueue path. */
1296 policy = ldlm_processing_policy_table[res->lr_type];
1297 policy(lock, flags, 1, &rc, NULL);
1300 unlock_res_and_lock(lock);
1302 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1306 /* Must be called with namespace taken: queue is waiting or converting. */
1307 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1308 struct list_head *work_list)
1310 struct list_head *tmp, *pos;
1311 ldlm_processing_policy policy;
1313 int rc = LDLM_ITER_CONTINUE;
1317 check_res_locked(res);
1319 policy = ldlm_processing_policy_table[res->lr_type];
1322 list_for_each_safe(tmp, pos, queue) {
1323 struct ldlm_lock *pending;
1324 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1326 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1329 rc = policy(pending, &flags, 0, &err, work_list);
1330 if (rc != LDLM_ITER_CONTINUE)
1337 /* Helper function for ldlm_run_ast_work().
1339 * Send an existing rpc set specified by @arg->set and then
1340 * destroy it. Create new one if @do_create flag is set. */
1342 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1346 ptlrpc_set_wait(arg->set);
1347 if (arg->type == LDLM_BL_CALLBACK)
1348 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1349 ptlrpc_set_destroy(arg->set);
1352 arg->set = ptlrpc_prep_set();
1358 ldlm_work_bl_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1360 struct ldlm_lock_desc d;
1361 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_bl_ast);
1364 /* nobody should touch l_bl_ast */
1365 lock_res_and_lock(lock);
1366 list_del_init(&lock->l_bl_ast);
1368 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1369 LASSERT(lock->l_bl_ast_run == 0);
1370 LASSERT(lock->l_blocking_lock);
1371 lock->l_bl_ast_run++;
1372 unlock_res_and_lock(lock);
1374 ldlm_lock2desc(lock->l_blocking_lock, &d);
1376 lock->l_blocking_ast(lock, &d, (void *)arg,
1378 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1379 lock->l_blocking_lock = NULL;
1380 LDLM_LOCK_RELEASE(lock);
1386 ldlm_work_cp_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1388 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_cp_ast);
1389 ldlm_completion_callback completion_callback;
1393 /* It's possible to receive a completion AST before we've set
1394 * the l_completion_ast pointer: either because the AST arrived
1395 * before the reply, or simply because there's a small race
1396 * window between receiving the reply and finishing the local
1397 * enqueue. (bug 842)
1399 * This can't happen with the blocking_ast, however, because we
1400 * will never call the local blocking_ast until we drop our
1401 * reader/writer reference, which we won't do until we get the
1402 * reply and finish enqueueing. */
1404 /* nobody should touch l_cp_ast */
1405 lock_res_and_lock(lock);
1406 list_del_init(&lock->l_cp_ast);
1407 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1408 /* save l_completion_ast since it can be changed by
1409 * mds_intent_policy(), see bug 14225 */
1410 completion_callback = lock->l_completion_ast;
1411 lock->l_flags &= ~LDLM_FL_CP_REQD;
1412 unlock_res_and_lock(lock);
1414 if (completion_callback != NULL) {
1415 completion_callback(lock, 0, (void *)arg);
1418 LDLM_LOCK_RELEASE(lock);
1424 ldlm_work_revoke_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1426 struct ldlm_lock_desc desc;
1427 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_rk_ast);
1430 list_del_init(&lock->l_rk_ast);
1432 /* the desc just pretend to exclusive */
1433 ldlm_lock2desc(lock, &desc);
1434 desc.l_req_mode = LCK_EX;
1435 desc.l_granted_mode = 0;
1437 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1438 LDLM_LOCK_RELEASE(lock);
1443 int ldlm_run_ast_work(struct list_head *rpc_list, ldlm_desc_ast_t ast_type)
1445 struct ldlm_cb_set_arg arg;
1446 struct list_head *tmp, *pos;
1447 int (*work_ast_lock)(struct list_head *tmp,struct ldlm_cb_set_arg *arg);
1451 arg.set = ptlrpc_prep_set();
1452 atomic_set(&arg.restart, 0);
1454 case LDLM_WORK_BL_AST:
1455 arg.type = LDLM_BL_CALLBACK;
1456 work_ast_lock = ldlm_work_bl_ast_lock;
1458 case LDLM_WORK_CP_AST:
1459 arg.type = LDLM_CP_CALLBACK;
1460 work_ast_lock = ldlm_work_cp_ast_lock;
1462 case LDLM_WORK_REVOKE_AST:
1463 arg.type = LDLM_BL_CALLBACK;
1464 work_ast_lock = ldlm_work_revoke_ast_lock;
1471 list_for_each_safe(tmp, pos, rpc_list) {
1472 ast_count += work_ast_lock(tmp, &arg);
1474 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1475 * and create a new set for requests that remained in
1477 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1478 ldlm_send_and_maybe_create_set(&arg, 1);
1484 ldlm_send_and_maybe_create_set(&arg, 0);
1486 /* In case when number of ASTs is multiply of
1487 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1488 * @arg.set must be destroyed here, otherwise we get
1489 * write memory leaking. */
1490 ptlrpc_set_destroy(arg.set);
1492 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1495 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1497 ldlm_reprocess_all(res);
1498 return LDLM_ITER_CONTINUE;
1501 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1503 struct list_head *tmp;
1510 spin_lock(&ns->ns_hash_lock);
1511 for (i = 0; i < RES_HASH_SIZE; i++) {
1512 tmp = ns->ns_hash[i].next;
1513 while (tmp != &(ns->ns_hash[i])) {
1514 struct ldlm_resource *res =
1515 list_entry(tmp, struct ldlm_resource, lr_hash);
1517 ldlm_resource_getref(res);
1518 spin_unlock(&ns->ns_hash_lock);
1519 LDLM_RESOURCE_ADDREF(res);
1521 rc = reprocess_one_queue(res, NULL);
1523 LDLM_RESOURCE_DELREF(res);
1524 spin_lock(&ns->ns_hash_lock);
1526 ldlm_resource_putref_locked(res);
1528 if (rc == LDLM_ITER_STOP)
1533 spin_unlock(&ns->ns_hash_lock);
1537 void ldlm_reprocess_all(struct ldlm_resource *res)
1539 CFS_LIST_HEAD(rpc_list);
1543 /* Local lock trees don't get reprocessed. */
1544 if (ns_is_client(res->lr_namespace)) {
1551 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1552 if (rc == LDLM_ITER_CONTINUE)
1553 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1556 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1557 if (rc == -ERESTART) {
1558 LASSERT(list_empty(&rpc_list));
1564 void ldlm_cancel_callback(struct ldlm_lock *lock)
1566 check_res_locked(lock->l_resource);
1567 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1568 lock->l_flags |= LDLM_FL_CANCEL;
1569 if (lock->l_blocking_ast) {
1570 // l_check_no_ns_lock(ns);
1571 unlock_res_and_lock(lock);
1572 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1574 lock_res_and_lock(lock);
1576 LDLM_DEBUG(lock, "no blocking ast");
1579 lock->l_flags |= LDLM_FL_BL_DONE;
1582 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1584 if (req->l_resource->lr_type != LDLM_PLAIN &&
1585 req->l_resource->lr_type != LDLM_IBITS)
1588 list_del_init(&req->l_sl_policy);
1589 list_del_init(&req->l_sl_mode);
1592 void ldlm_lock_cancel(struct ldlm_lock *lock)
1594 struct ldlm_resource *res;
1595 struct ldlm_namespace *ns;
1598 lock_res_and_lock(lock);
1600 res = lock->l_resource;
1601 ns = res->lr_namespace;
1603 /* Please do not, no matter how tempting, remove this LBUG without
1604 * talking to me first. -phik */
1605 if (lock->l_readers || lock->l_writers) {
1606 LDLM_ERROR(lock, "lock still has references");
1610 ldlm_del_waiting_lock(lock);
1612 /* Releases cancel callback. */
1613 ldlm_cancel_callback(lock);
1615 /* Yes, second time, just in case it was added again while we were
1616 running with no res lock in ldlm_cancel_callback */
1617 ldlm_del_waiting_lock(lock);
1618 ldlm_resource_unlink_lock(lock);
1619 ldlm_lock_destroy_nolock(lock);
1621 if (lock->l_granted_mode == lock->l_req_mode)
1622 ldlm_pool_del(&ns->ns_pool, lock);
1624 /* Make sure we will not be called again for same lock what is possible
1625 * if not to zero out lock->l_granted_mode */
1626 lock->l_granted_mode = LCK_MINMODE;
1627 unlock_res_and_lock(lock);
1632 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1634 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1640 lock->l_ast_data = data;
1641 LDLM_LOCK_PUT(lock);
1645 void ldlm_cancel_locks_for_export_cb(void *obj, void *data)
1647 struct obd_export *exp = data;
1648 struct ldlm_lock *lock = obj;
1649 struct ldlm_resource *res;
1651 res = ldlm_resource_getref(lock->l_resource);
1652 LDLM_LOCK_GET(lock);
1654 LDLM_DEBUG(lock, "export %p", exp);
1655 ldlm_res_lvbo_update(res, NULL, 0, 1);
1656 ldlm_lock_cancel(lock);
1657 ldlm_reprocess_all(res);
1658 ldlm_resource_putref(res);
1659 LDLM_LOCK_RELEASE(lock);
1662 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1664 lustre_hash_for_each_empty(exp->exp_lock_hash,
1665 ldlm_cancel_locks_for_export_cb, exp);
1669 * Downgrade an exclusive lock.
1671 * A fast variant of ldlm_lock_convert for convertion of exclusive
1672 * locks. The convertion is always successful.
1674 * \param lock A lock to convert
1675 * \param new_mode new lock mode
1677 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1681 LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1682 LASSERT(new_mode == LCK_COS);
1684 lock_res_and_lock(lock);
1685 ldlm_resource_unlink_lock(lock);
1686 lock->l_req_mode = new_mode;
1687 ldlm_grant_lock(lock, NULL);
1688 unlock_res_and_lock(lock);
1689 ldlm_reprocess_all(lock->l_resource);
1694 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1697 CFS_LIST_HEAD(rpc_list);
1698 struct ldlm_resource *res;
1699 struct ldlm_namespace *ns;
1702 struct sl_insert_point prev;
1704 struct ldlm_interval *node;
1707 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1708 *flags |= LDLM_FL_BLOCK_GRANTED;
1709 RETURN(lock->l_resource);
1712 /* I can't check the type of lock here because the bitlock of lock
1713 * is not held here, so do the allocation blindly. -jay */
1714 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO, sizeof(*node));
1715 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1718 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1719 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1721 lock_res_and_lock(lock);
1723 res = lock->l_resource;
1724 ns = res->lr_namespace;
1726 old_mode = lock->l_req_mode;
1727 lock->l_req_mode = new_mode;
1728 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1729 /* remember the lock position where the lock might be
1730 * added back to the granted list later and also
1731 * remember the join mode for skiplist fixing. */
1732 prev.res_link = lock->l_res_link.prev;
1733 prev.mode_link = lock->l_sl_mode.prev;
1734 prev.policy_link = lock->l_sl_policy.prev;
1735 ldlm_resource_unlink_lock(lock);
1737 ldlm_resource_unlink_lock(lock);
1738 if (res->lr_type == LDLM_EXTENT) {
1739 /* FIXME: ugly code, I have to attach the lock to a
1740 * interval node again since perhaps it will be granted
1742 CFS_INIT_LIST_HEAD(&node->li_group);
1743 ldlm_interval_attach(node, lock);
1748 /* If this is a local resource, put it on the appropriate list. */
1749 if (ns_is_client(res->lr_namespace)) {
1750 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1751 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1753 /* This should never happen, because of the way the
1754 * server handles conversions. */
1755 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1759 ldlm_grant_lock(lock, &rpc_list);
1761 /* FIXME: completion handling not with ns_lock held ! */
1762 if (lock->l_completion_ast)
1763 lock->l_completion_ast(lock, 0, NULL);
1767 ldlm_processing_policy policy;
1768 policy = ldlm_processing_policy_table[res->lr_type];
1769 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1770 if (rc == LDLM_ITER_STOP) {
1771 lock->l_req_mode = old_mode;
1772 if (res->lr_type == LDLM_EXTENT)
1773 ldlm_extent_add_lock(res, lock);
1775 ldlm_granted_list_add_lock(lock, &prev);
1779 *flags |= LDLM_FL_BLOCK_GRANTED;
1783 unlock_res_and_lock(lock);
1786 ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1788 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1792 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1794 struct obd_device *obd = NULL;
1796 if (!((libcfs_debug | D_ERROR) & level))
1800 CDEBUG(level, " NULL LDLM lock\n");
1804 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1805 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1807 if (lock->l_conn_export != NULL)
1808 obd = lock->l_conn_export->exp_obd;
1809 if (lock->l_export && lock->l_export->exp_connection) {
1810 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1811 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1812 lock->l_remote_handle.cookie);
1813 } else if (obd == NULL) {
1814 CDEBUG(level, " Node: local\n");
1816 struct obd_import *imp = obd->u.cli.cl_import;
1817 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1818 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1819 lock->l_remote_handle.cookie);
1821 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1823 lock->l_resource->lr_name.name[0],
1824 lock->l_resource->lr_name.name[1],
1825 lock->l_resource->lr_name.name[2]);
1826 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1827 "write: %d flags: %#x\n", ldlm_lockname[lock->l_req_mode],
1828 ldlm_lockname[lock->l_granted_mode],
1829 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1831 if (lock->l_resource->lr_type == LDLM_EXTENT)
1832 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1833 " (req "LPU64"-"LPU64")\n",
1834 lock->l_policy_data.l_extent.start,
1835 lock->l_policy_data.l_extent.end,
1836 lock->l_req_extent.start, lock->l_req_extent.end);
1837 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1838 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1839 lock->l_policy_data.l_flock.pid,
1840 lock->l_policy_data.l_flock.start,
1841 lock->l_policy_data.l_flock.end);
1842 else if (lock->l_resource->lr_type == LDLM_IBITS)
1843 CDEBUG(level, " Bits: "LPX64"\n",
1844 lock->l_policy_data.l_inodebits.bits);
1847 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1849 struct ldlm_lock *lock;
1851 if (!((libcfs_debug | D_ERROR) & level))
1854 lock = ldlm_handle2lock(lockh);
1858 ldlm_lock_dump(D_OTHER, lock, 0);
1860 LDLM_LOCK_PUT(lock);
1863 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1864 struct libcfs_debug_msg_data *data, const char *fmt,
1868 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1870 va_start(args, fmt);
1872 if (lock->l_resource == NULL) {
1873 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1874 data->msg_fn, data->msg_line, fmt, args,
1875 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1876 "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "
1877 LPX64" expref: %d pid: %u\n", lock,
1878 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1879 lock->l_readers, lock->l_writers,
1880 ldlm_lockname[lock->l_granted_mode],
1881 ldlm_lockname[lock->l_req_mode],
1882 lock->l_flags, lock->l_remote_handle.cookie,
1884 atomic_read(&lock->l_export->exp_refcount) : -99,
1890 switch (lock->l_resource->lr_type) {
1892 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1893 data->msg_fn, data->msg_line, fmt, args,
1894 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1895 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1896 "] (req "LPU64"->"LPU64") flags: %x remote: "LPX64
1897 " expref: %d pid: %u\n",
1898 lock->l_resource->lr_namespace->ns_name, lock,
1899 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1900 lock->l_readers, lock->l_writers,
1901 ldlm_lockname[lock->l_granted_mode],
1902 ldlm_lockname[lock->l_req_mode],
1903 lock->l_resource->lr_name.name[0],
1904 lock->l_resource->lr_name.name[1],
1905 atomic_read(&lock->l_resource->lr_refcount),
1906 ldlm_typename[lock->l_resource->lr_type],
1907 lock->l_policy_data.l_extent.start,
1908 lock->l_policy_data.l_extent.end,
1909 lock->l_req_extent.start, lock->l_req_extent.end,
1910 lock->l_flags, lock->l_remote_handle.cookie,
1912 atomic_read(&lock->l_export->exp_refcount) : -99,
1917 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1918 data->msg_fn, data->msg_line, fmt, args,
1919 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1920 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1921 "["LPU64"->"LPU64"] flags: %x remote: "LPX64
1922 " expref: %d pid: %u\n",
1923 lock->l_resource->lr_namespace->ns_name, lock,
1924 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1925 lock->l_readers, lock->l_writers,
1926 ldlm_lockname[lock->l_granted_mode],
1927 ldlm_lockname[lock->l_req_mode],
1928 lock->l_resource->lr_name.name[0],
1929 lock->l_resource->lr_name.name[1],
1930 atomic_read(&lock->l_resource->lr_refcount),
1931 ldlm_typename[lock->l_resource->lr_type],
1932 lock->l_policy_data.l_flock.pid,
1933 lock->l_policy_data.l_flock.start,
1934 lock->l_policy_data.l_flock.end,
1935 lock->l_flags, lock->l_remote_handle.cookie,
1937 atomic_read(&lock->l_export->exp_refcount) : -99,
1942 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1943 data->msg_fn, data->msg_line, fmt, args,
1944 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1945 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1946 "flags: %x remote: "LPX64" expref: %d "
1948 lock->l_resource->lr_namespace->ns_name,
1949 lock, lock->l_handle.h_cookie,
1950 atomic_read (&lock->l_refc),
1951 lock->l_readers, lock->l_writers,
1952 ldlm_lockname[lock->l_granted_mode],
1953 ldlm_lockname[lock->l_req_mode],
1954 lock->l_resource->lr_name.name[0],
1955 lock->l_resource->lr_name.name[1],
1956 lock->l_policy_data.l_inodebits.bits,
1957 atomic_read(&lock->l_resource->lr_refcount),
1958 ldlm_typename[lock->l_resource->lr_type],
1959 lock->l_flags, lock->l_remote_handle.cookie,
1961 atomic_read(&lock->l_export->exp_refcount) : -99,
1966 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1967 data->msg_fn, data->msg_line, fmt, args,
1968 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1969 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: %x "
1970 "remote: "LPX64" expref: %d pid: %u\n",
1971 lock->l_resource->lr_namespace->ns_name,
1972 lock, lock->l_handle.h_cookie,
1973 atomic_read (&lock->l_refc),
1974 lock->l_readers, lock->l_writers,
1975 ldlm_lockname[lock->l_granted_mode],
1976 ldlm_lockname[lock->l_req_mode],
1977 lock->l_resource->lr_name.name[0],
1978 lock->l_resource->lr_name.name[1],
1979 atomic_read(&lock->l_resource->lr_refcount),
1980 ldlm_typename[lock->l_resource->lr_type],
1981 lock->l_flags, lock->l_remote_handle.cookie,
1983 atomic_read(&lock->l_export->exp_refcount) : -99,
1989 EXPORT_SYMBOL(_ldlm_lock_debug);