1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_lock.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_LDLM
45 # include <libcfs/libcfs.h>
46 # ifndef HAVE_VFS_INTENT_PATCHES
47 # include <linux/lustre_intent.h>
50 # include <liblustre.h>
53 #include <obd_class.h>
54 #include "ldlm_internal.h"
57 char *ldlm_lockname[] = {
68 char *ldlm_typename[] = {
75 char *ldlm_it2str(int it)
82 case (IT_OPEN | IT_CREAT):
95 CERROR("Unknown intent %d\n", it);
100 extern cfs_mem_cache_t *ldlm_lock_slab;
102 static ldlm_processing_policy ldlm_processing_policy_table[] = {
103 [LDLM_PLAIN] ldlm_process_plain_lock,
104 [LDLM_EXTENT] ldlm_process_extent_lock,
106 [LDLM_FLOCK] ldlm_process_flock_lock,
108 [LDLM_IBITS] ldlm_process_inodebits_lock,
111 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
113 return ldlm_processing_policy_table[res->lr_type];
116 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
122 * REFCOUNTED LOCK OBJECTS
127 * Lock refcounts, during creation:
128 * - one special one for allocation, dec'd only once in destroy
129 * - one for being a lock that's in-use
130 * - one for the addref associated with a new lock
132 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
134 atomic_inc(&lock->l_refc);
138 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
140 LASSERT(size == sizeof(*lock));
141 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
144 void ldlm_lock_put(struct ldlm_lock *lock)
148 LASSERT(lock->l_resource != LP_POISON);
149 LASSERT(atomic_read(&lock->l_refc) > 0);
150 if (atomic_dec_and_test(&lock->l_refc)) {
151 struct ldlm_resource *res;
154 "final lock_put on destroyed lock, freeing it.");
156 res = lock->l_resource;
157 LASSERT(lock->l_destroyed);
158 LASSERT(list_empty(&lock->l_res_link));
159 LASSERT(list_empty(&lock->l_pending_chain));
161 atomic_dec(&res->lr_namespace->ns_locks);
162 lu_ref_del(&res->lr_reference, "lock", lock);
163 ldlm_resource_putref(res);
164 lock->l_resource = NULL;
165 if (lock->l_export) {
166 class_export_put(lock->l_export);
167 lock->l_export = NULL;
170 if (lock->l_lvb_data != NULL)
171 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
173 ldlm_interval_free(ldlm_interval_detach(lock));
174 lu_ref_fini(&lock->l_reference);
175 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
182 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
185 if (!list_empty(&lock->l_lru)) {
186 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
187 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
188 list_del_init(&lock->l_lru);
190 LASSERT(ns->ns_nr_unused >= 0);
196 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
198 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
201 spin_lock(&ns->ns_unused_lock);
202 rc = ldlm_lock_remove_from_lru_nolock(lock);
203 spin_unlock(&ns->ns_unused_lock);
208 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
210 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
211 lock->l_last_used = cfs_time_current();
212 LASSERT(list_empty(&lock->l_lru));
213 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
214 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
215 LASSERT(ns->ns_nr_unused >= 0);
219 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
221 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
223 spin_lock(&ns->ns_unused_lock);
224 ldlm_lock_add_to_lru_nolock(lock);
225 spin_unlock(&ns->ns_unused_lock);
229 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
231 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
233 spin_lock(&ns->ns_unused_lock);
234 if (!list_empty(&lock->l_lru)) {
235 ldlm_lock_remove_from_lru_nolock(lock);
236 ldlm_lock_add_to_lru_nolock(lock);
238 spin_unlock(&ns->ns_unused_lock);
242 /* This used to have a 'strict' flag, which recovery would use to mark an
243 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
244 * shall explain why it's gone: with the new hash table scheme, once you call
245 * ldlm_lock_destroy, you can never drop your final references on this lock.
246 * Because it's not in the hash table anymore. -phil */
247 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
251 if (lock->l_readers || lock->l_writers) {
252 LDLM_ERROR(lock, "lock still has references");
253 ldlm_lock_dump(D_ERROR, lock, 0);
257 if (!list_empty(&lock->l_res_link)) {
258 LDLM_ERROR(lock, "lock still on resource");
259 ldlm_lock_dump(D_ERROR, lock, 0);
263 if (lock->l_destroyed) {
264 LASSERT(list_empty(&lock->l_lru));
268 lock->l_destroyed = 1;
270 if (lock->l_export && lock->l_export->exp_lock_hash)
271 lustre_hash_del(lock->l_export->exp_lock_hash,
272 &lock->l_remote_handle, &lock->l_exp_hash);
274 ldlm_lock_remove_from_lru(lock);
275 class_handle_unhash(&lock->l_handle);
278 /* Wake anyone waiting for this lock */
279 /* FIXME: I should probably add yet another flag, instead of using
280 * l_export to only call this on clients */
282 class_export_put(lock->l_export);
283 lock->l_export = NULL;
284 if (lock->l_export && lock->l_completion_ast)
285 lock->l_completion_ast(lock, 0);
291 void ldlm_lock_destroy(struct ldlm_lock *lock)
295 lock_res_and_lock(lock);
296 first = ldlm_lock_destroy_internal(lock);
297 unlock_res_and_lock(lock);
299 /* drop reference from hashtable only for first destroy */
301 lu_ref_del(&lock->l_reference, "hash", lock);
302 LDLM_LOCK_RELEASE(lock);
307 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
311 first = ldlm_lock_destroy_internal(lock);
312 /* drop reference from hashtable only for first destroy */
314 lu_ref_del(&lock->l_reference, "hash", lock);
315 LDLM_LOCK_RELEASE(lock);
320 /* this is called by portals_handle2object with the handle lock taken */
321 static void lock_handle_addref(void *lock)
323 LDLM_LOCK_GET((struct ldlm_lock *)lock);
327 * usage: pass in a resource on which you have done ldlm_resource_get
328 * pass in a parent lock on which you have done a ldlm_lock_get
329 * after return, ldlm_*_put the resource and parent
330 * returns: lock with refcount 2 - one for current caller and one for remote
332 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
334 struct ldlm_lock *lock;
337 if (resource == NULL)
340 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, CFS_ALLOC_IO, sizeof(*lock));
344 spin_lock_init(&lock->l_lock);
345 lock->l_resource = ldlm_resource_getref(resource);
346 lu_ref_add(&resource->lr_reference, "lock", lock);
348 atomic_set(&lock->l_refc, 2);
349 CFS_INIT_LIST_HEAD(&lock->l_res_link);
350 CFS_INIT_LIST_HEAD(&lock->l_lru);
351 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
352 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
353 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
354 CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
355 cfs_waitq_init(&lock->l_waitq);
356 lock->l_blocking_lock = NULL;
357 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
358 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
359 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
361 atomic_inc(&resource->lr_namespace->ns_locks);
362 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
363 class_handle_hash(&lock->l_handle, lock_handle_addref);
365 CFS_INIT_LIST_HEAD(&lock->l_extents_list);
366 spin_lock_init(&lock->l_extents_list_lock);
367 CFS_INIT_LIST_HEAD(&lock->l_cache_locks_list);
368 lu_ref_init(&lock->l_reference);
369 lu_ref_add(&lock->l_reference, "hash", lock);
374 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
375 const struct ldlm_res_id *new_resid)
377 struct ldlm_resource *oldres = lock->l_resource;
378 struct ldlm_resource *newres;
382 LASSERT(ns_is_client(ns));
384 lock_res_and_lock(lock);
385 if (memcmp(new_resid, &lock->l_resource->lr_name,
386 sizeof(lock->l_resource->lr_name)) == 0) {
388 unlock_res_and_lock(lock);
392 LASSERT(new_resid->name[0] != 0);
394 /* This function assumes that the lock isn't on any lists */
395 LASSERT(list_empty(&lock->l_res_link));
397 type = oldres->lr_type;
398 unlock_res_and_lock(lock);
400 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
401 lu_ref_add(&newres->lr_reference, "lock", lock);
405 * To flip the lock from the old to the new resource, lock, oldres and
406 * newres have to be locked. Resource spin-locks are nested within
407 * lock->l_lock, and are taken in the memory address order to avoid
410 spin_lock(&lock->l_lock);
411 oldres = lock->l_resource;
412 if (oldres < newres) {
414 lock_res_nested(newres, LRT_NEW);
417 lock_res_nested(oldres, LRT_NEW);
419 LASSERT(memcmp(new_resid, &oldres->lr_name,
420 sizeof oldres->lr_name) != 0);
421 lock->l_resource = newres;
423 unlock_res_and_lock(lock);
425 /* ...and the flowers are still standing! */
426 lu_ref_del(&oldres->lr_reference, "lock", lock);
427 ldlm_resource_putref(oldres);
436 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
438 lockh->cookie = lock->l_handle.h_cookie;
441 /* if flags: atomically get the lock and set the flags.
442 * Return NULL if flag already set
445 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
448 struct ldlm_namespace *ns;
449 struct ldlm_lock *lock, *retval = NULL;
454 lock = class_handle2object(handle->cookie);
458 LASSERT(lock->l_resource != NULL);
459 ns = lock->l_resource->lr_namespace;
462 lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
463 lock_res_and_lock(lock);
465 /* It's unlikely but possible that someone marked the lock as
466 * destroyed after we did handle2object on it */
467 if (lock->l_destroyed) {
468 unlock_res_and_lock(lock);
469 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
474 if (flags && (lock->l_flags & flags)) {
475 unlock_res_and_lock(lock);
481 lock->l_flags |= flags;
483 unlock_res_and_lock(lock);
490 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
492 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
493 /* INODEBITS_INTEROP: If the other side does not support
494 * inodebits, reply with a plain lock descriptor.
496 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
497 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
498 struct ldlm_resource res = *lock->l_resource;
500 /* Make sure all the right bits are set in this lock we
501 are going to pass to client */
502 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
503 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
504 "Inappropriate inode lock bits during "
505 "conversion " LPU64 "\n",
506 lock->l_policy_data.l_inodebits.bits);
507 res.lr_type = LDLM_PLAIN;
508 ldlm_res2desc(&res, &desc->l_resource);
509 /* Convert "new" lock mode to something old client can
511 if ((lock->l_req_mode == LCK_CR) ||
512 (lock->l_req_mode == LCK_CW))
513 desc->l_req_mode = LCK_PR;
515 desc->l_req_mode = lock->l_req_mode;
516 if ((lock->l_granted_mode == LCK_CR) ||
517 (lock->l_granted_mode == LCK_CW)) {
518 desc->l_granted_mode = LCK_PR;
520 /* We never grant PW/EX locks to clients */
521 LASSERT((lock->l_granted_mode != LCK_PW) &&
522 (lock->l_granted_mode != LCK_EX));
523 desc->l_granted_mode = lock->l_granted_mode;
526 /* We do not copy policy here, because there is no
527 policy for plain locks */
529 ldlm_res2desc(lock->l_resource, &desc->l_resource);
530 desc->l_req_mode = lock->l_req_mode;
531 desc->l_granted_mode = lock->l_granted_mode;
532 desc->l_policy_data = lock->l_policy_data;
536 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
537 struct list_head *work_list)
539 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
540 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
541 lock->l_flags |= LDLM_FL_AST_SENT;
542 /* If the enqueuing client said so, tell the AST recipient to
543 * discard dirty data, rather than writing back. */
544 if (new->l_flags & LDLM_AST_DISCARD_DATA)
545 lock->l_flags |= LDLM_FL_DISCARD_DATA;
546 LASSERT(list_empty(&lock->l_bl_ast));
547 list_add(&lock->l_bl_ast, work_list);
549 LASSERT(lock->l_blocking_lock == NULL);
550 lock->l_blocking_lock = LDLM_LOCK_GET(new);
554 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
556 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
557 lock->l_flags |= LDLM_FL_CP_REQD;
558 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
559 LASSERT(list_empty(&lock->l_cp_ast));
560 list_add(&lock->l_cp_ast, work_list);
565 /* must be called with lr_lock held */
566 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
567 struct list_head *work_list)
570 check_res_locked(lock->l_resource);
572 ldlm_add_bl_work_item(lock, new, work_list);
574 ldlm_add_cp_work_item(lock, work_list);
578 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
580 struct ldlm_lock *lock;
582 lock = ldlm_handle2lock(lockh);
583 LASSERT(lock != NULL);
584 ldlm_lock_addref_internal(lock, mode);
588 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
590 ldlm_lock_remove_from_lru(lock);
591 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
593 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
595 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
597 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
600 lu_ref_add_atomic(&lock->l_reference, "user", lock);
601 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
605 * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
608 * \retval 0 success, lock was addref-ed
610 * \retval -EAGAIN lock is being canceled.
612 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
614 struct ldlm_lock *lock;
618 lock = ldlm_handle2lock(lockh);
620 lock_res_and_lock(lock);
621 if (!(lock->l_flags & LDLM_FL_CBPENDING)) {
622 ldlm_lock_addref_internal_nolock(lock, mode);
625 unlock_res_and_lock(lock);
631 /* only called for local locks */
632 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
634 lock_res_and_lock(lock);
635 ldlm_lock_addref_internal_nolock(lock, mode);
636 unlock_res_and_lock(lock);
639 /* only called in ldlm_flock_destroy and for local locks.
640 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
641 * * ldlm_lock_remove_from_lru() does nothing, it is safe
642 * * for ldlm_flock_destroy usage by dropping some code */
643 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
645 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
646 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
647 LASSERT(lock->l_readers > 0);
648 lu_ref_del(&lock->l_reference, "reader", lock);
651 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
652 LASSERT(lock->l_writers > 0);
653 lu_ref_del(&lock->l_reference, "writer", lock);
657 lu_ref_del(&lock->l_reference, "user", lock);
658 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
661 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
663 struct ldlm_namespace *ns;
666 lock_res_and_lock(lock);
668 ns = lock->l_resource->lr_namespace;
670 ldlm_lock_decref_internal_nolock(lock, mode);
672 if (lock->l_flags & LDLM_FL_LOCAL &&
673 !lock->l_readers && !lock->l_writers) {
674 /* If this is a local lock on a server namespace and this was
675 * the last reference, cancel the lock. */
676 CDEBUG(D_INFO, "forcing cancel of local lock\n");
677 lock->l_flags |= LDLM_FL_CBPENDING;
680 if (!lock->l_readers && !lock->l_writers &&
681 (lock->l_flags & LDLM_FL_CBPENDING)) {
682 /* If we received a blocked AST and this was the last reference,
683 * run the callback. */
684 if (ns_is_server(ns) && lock->l_export)
685 CERROR("FL_CBPENDING set on non-local lock--just a "
688 LDLM_DEBUG(lock, "final decref done on cbpending lock");
690 LDLM_LOCK_GET(lock); /* dropped by bl thread */
691 ldlm_lock_remove_from_lru(lock);
692 unlock_res_and_lock(lock);
693 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
694 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
695 ldlm_handle_bl_callback(ns, NULL, lock);
696 } else if (ns_is_client(ns) &&
697 !lock->l_readers && !lock->l_writers &&
698 !(lock->l_flags & LDLM_FL_BL_AST)) {
699 /* If this is a client-side namespace and this was the last
700 * reference, put it on the LRU. */
701 ldlm_lock_add_to_lru(lock);
702 unlock_res_and_lock(lock);
703 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
704 * are not supported by the server, otherwise, it is done on
706 if (!exp_connect_cancelset(lock->l_conn_export) &&
707 !ns_connect_lru_resize(ns))
708 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
710 unlock_res_and_lock(lock);
716 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
718 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
719 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
720 ldlm_lock_decref_internal(lock, mode);
724 /* This will drop a lock reference and mark it for destruction, but will not
725 * necessarily cancel the lock before returning. */
726 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
728 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
731 LASSERT(lock != NULL);
733 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
734 lock_res_and_lock(lock);
735 lock->l_flags |= LDLM_FL_CBPENDING;
736 unlock_res_and_lock(lock);
737 ldlm_lock_decref_internal(lock, mode);
741 struct sl_insert_point {
742 struct list_head *res_link;
743 struct list_head *mode_link;
744 struct list_head *policy_link;
748 * search_granted_lock
751 * Finds a position to insert the new lock.
753 * queue [input]: the granted list where search acts on;
754 * req [input]: the lock whose position to be located;
755 * prev [output]: positions within 3 lists to insert @req to
759 * - ldlm_grant_lock_with_skiplist
761 static void search_granted_lock(struct list_head *queue,
762 struct ldlm_lock *req,
763 struct sl_insert_point *prev)
765 struct list_head *tmp;
766 struct ldlm_lock *lock, *mode_end, *policy_end;
769 list_for_each(tmp, queue) {
770 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
772 mode_end = list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
775 if (lock->l_req_mode != req->l_req_mode) {
776 /* jump to last lock of mode group */
777 tmp = &mode_end->l_res_link;
781 /* suitable mode group is found */
782 if (lock->l_resource->lr_type == LDLM_PLAIN) {
783 /* insert point is last lock of the mode group */
784 prev->res_link = &mode_end->l_res_link;
785 prev->mode_link = &mode_end->l_sl_mode;
786 prev->policy_link = &req->l_sl_policy;
789 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
791 policy_end = list_entry(lock->l_sl_policy.prev,
795 if (lock->l_policy_data.l_inodebits.bits ==
796 req->l_policy_data.l_inodebits.bits) {
797 /* insert point is last lock of
798 * the policy group */
800 &policy_end->l_res_link;
802 &policy_end->l_sl_mode;
804 &policy_end->l_sl_policy;
809 if (policy_end == mode_end)
810 /* done with mode group */
813 /* jump to next policy group within the mode group */
814 tmp = policy_end->l_res_link.next;
815 lock = list_entry(tmp, struct ldlm_lock,
817 } /* loop over policy groups within the mode group */
819 /* insert point is last lock of the mode group,
820 * new policy group is started */
821 prev->res_link = &mode_end->l_res_link;
822 prev->mode_link = &mode_end->l_sl_mode;
823 prev->policy_link = &req->l_sl_policy;
827 LDLM_ERROR(lock, "is not LDLM_PLAIN or LDLM_IBITS lock");
832 /* insert point is last lock on the queue,
833 * new mode group and new policy group are started */
834 prev->res_link = queue->prev;
835 prev->mode_link = &req->l_sl_mode;
836 prev->policy_link = &req->l_sl_policy;
841 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
842 struct sl_insert_point *prev)
844 struct ldlm_resource *res = lock->l_resource;
847 check_res_locked(res);
849 ldlm_resource_dump(D_INFO, res);
850 CDEBUG(D_OTHER, "About to add this lock:\n");
851 ldlm_lock_dump(D_OTHER, lock, 0);
853 if (lock->l_destroyed) {
854 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
858 LASSERT(list_empty(&lock->l_res_link));
859 LASSERT(list_empty(&lock->l_sl_mode));
860 LASSERT(list_empty(&lock->l_sl_policy));
862 list_add(&lock->l_res_link, prev->res_link);
863 list_add(&lock->l_sl_mode, prev->mode_link);
864 list_add(&lock->l_sl_policy, prev->policy_link);
869 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
871 struct sl_insert_point prev;
874 LASSERT(lock->l_req_mode == lock->l_granted_mode);
876 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
877 ldlm_granted_list_add_lock(lock, &prev);
882 * - ldlm_lock_enqueue
883 * - ldlm_reprocess_queue
884 * - ldlm_lock_convert
886 * must be called with lr_lock held
888 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
890 struct ldlm_resource *res = lock->l_resource;
893 check_res_locked(res);
895 lock->l_granted_mode = lock->l_req_mode;
896 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
897 ldlm_grant_lock_with_skiplist(lock);
898 else if (res->lr_type == LDLM_EXTENT)
899 ldlm_extent_add_lock(res, lock);
901 ldlm_resource_add_lock(res, &res->lr_granted, lock);
903 if (lock->l_granted_mode < res->lr_most_restr)
904 res->lr_most_restr = lock->l_granted_mode;
906 if (work_list && lock->l_completion_ast != NULL)
907 ldlm_add_ast_work_item(lock, NULL, work_list);
909 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
913 /* returns a referenced lock or NULL. See the flag descriptions below, in the
914 * comment above ldlm_lock_match */
915 static struct ldlm_lock *search_queue(struct list_head *queue,
917 ldlm_policy_data_t *policy,
918 struct ldlm_lock *old_lock, int flags)
920 struct ldlm_lock *lock;
921 struct list_head *tmp;
923 list_for_each(tmp, queue) {
926 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
928 if (lock == old_lock)
931 /* llite sometimes wants to match locks that will be
932 * canceled when their users drop, but we allow it to match
933 * if it passes in CBPENDING and the lock still has users.
934 * this is generally only going to be used by children
935 * whose parents already hold a lock so forward progress
936 * can still happen. */
937 if (lock->l_flags & LDLM_FL_CBPENDING &&
938 !(flags & LDLM_FL_CBPENDING))
940 if (lock->l_flags & LDLM_FL_CBPENDING &&
941 lock->l_readers == 0 && lock->l_writers == 0)
944 if (!(lock->l_req_mode & *mode))
946 match = lock->l_req_mode;
948 if (lock->l_resource->lr_type == LDLM_EXTENT &&
949 (lock->l_policy_data.l_extent.start >
950 policy->l_extent.start ||
951 lock->l_policy_data.l_extent.end < policy->l_extent.end))
954 if (unlikely(match == LCK_GROUP) &&
955 lock->l_resource->lr_type == LDLM_EXTENT &&
956 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
959 /* We match if we have existing lock with same or wider set
961 if (lock->l_resource->lr_type == LDLM_IBITS &&
962 ((lock->l_policy_data.l_inodebits.bits &
963 policy->l_inodebits.bits) !=
964 policy->l_inodebits.bits))
967 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
970 if ((flags & LDLM_FL_LOCAL_ONLY) &&
971 !(lock->l_flags & LDLM_FL_LOCAL))
974 if (flags & LDLM_FL_TEST_LOCK) {
976 ldlm_lock_touch_in_lru(lock);
978 ldlm_lock_addref_internal_nolock(lock, match);
987 void ldlm_lock_allow_match(struct ldlm_lock *lock)
989 lock_res_and_lock(lock);
990 lock->l_flags |= LDLM_FL_LVB_READY;
991 cfs_waitq_signal(&lock->l_waitq);
992 unlock_res_and_lock(lock);
996 * Checks if requested extent lock is compatible with another owned lock.
998 * Checks if \a lock is compatible with a read or write lock
999 * (specified by \a rw) for an extent [\a start , \a end].
1001 * \param lock the already owned lock
1002 * \param rw OBD_BRW_READ if requested for reading,
1003 * OBD_BRW_WRITE if requested for writing
1004 * \param start start of the requested extent
1005 * \param end end of the requested extent
1006 * \param cookie transparent parameter for passing locking context
1008 * \post result == 1, *cookie == context, appropriate lock is referenced
1010 * \retval 1 owned lock is reused for the request
1011 * \retval 0 no lock reused for the request
1013 * \see ldlm_lock_fast_release
1015 int ldlm_lock_fast_match(struct ldlm_lock *lock, int rw,
1016 obd_off start, obd_off end,
1019 LASSERT(rw == OBD_BRW_READ || rw == OBD_BRW_WRITE);
1024 lock_res_and_lock(lock);
1025 /* check if granted mode is compatible */
1026 if (rw == OBD_BRW_WRITE &&
1027 !(lock->l_granted_mode & (LCK_PW|LCK_GROUP)))
1030 /* does the lock cover the region we would like to access? */
1031 if ((lock->l_policy_data.l_extent.start > start) ||
1032 (lock->l_policy_data.l_extent.end < end))
1035 /* if we received a blocking callback and the lock is no longer
1036 * referenced, don't use it */
1037 if ((lock->l_flags & LDLM_FL_CBPENDING) &&
1038 !lock->l_writers && !lock->l_readers)
1041 ldlm_lock_addref_internal_nolock(lock, rw == OBD_BRW_WRITE ? LCK_PW : LCK_PR);
1042 unlock_res_and_lock(lock);
1043 *cookie = (void *)lock;
1044 return 1; /* avoid using rc for stack relief */
1047 unlock_res_and_lock(lock);
1052 * Releases a reference to a lock taken in a "fast" way.
1054 * Releases a read or write (specified by \a rw) lock
1055 * referenced by \a cookie.
1057 * \param rw OBD_BRW_READ if requested for reading,
1058 * OBD_BRW_WRITE if requested for writing
1059 * \param cookie transparent parameter for passing locking context
1061 * \post appropriate lock is dereferenced
1063 * \see ldlm_lock_fast_lock
1065 void ldlm_lock_fast_release(void *cookie, int rw)
1067 struct ldlm_lock *lock = (struct ldlm_lock *)cookie;
1069 LASSERT(lock != NULL);
1070 LASSERT(rw == OBD_BRW_READ || rw == OBD_BRW_WRITE);
1071 LASSERT(rw == OBD_BRW_READ ||
1072 (lock->l_granted_mode & (LCK_PW | LCK_GROUP)));
1073 ldlm_lock_decref_internal(lock, rw == OBD_BRW_WRITE ? LCK_PW : LCK_PR);
1076 /* Can be called in two ways:
1078 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1079 * for a duplicate of.
1081 * Otherwise, all of the fields must be filled in, to match against.
1083 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1084 * server (ie, connh is NULL)
1085 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1086 * list will be considered
1087 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1088 * to be canceled can still be matched as long as they still have reader
1089 * or writer refernces
1090 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1091 * just tell us if we would have matched.
1093 * Returns 1 if it finds an already-existing lock that is compatible; in this
1094 * case, lockh is filled in with a addref()ed lock
1096 * we also check security context, if that failed we simply return 0 (to keep
1097 * caller code unchanged), the context failure will be discovered by caller
1100 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1101 const struct ldlm_res_id *res_id, ldlm_type_t type,
1102 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1103 struct lustre_handle *lockh)
1105 struct ldlm_resource *res;
1106 struct ldlm_lock *lock, *old_lock = NULL;
1111 old_lock = ldlm_handle2lock(lockh);
1114 ns = old_lock->l_resource->lr_namespace;
1115 res_id = &old_lock->l_resource->lr_name;
1116 type = old_lock->l_resource->lr_type;
1117 mode = old_lock->l_req_mode;
1120 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1122 LASSERT(old_lock == NULL);
1126 LDLM_RESOURCE_ADDREF(res);
1129 lock = search_queue(&res->lr_granted, &mode, policy, old_lock, flags);
1132 if (flags & LDLM_FL_BLOCK_GRANTED)
1134 lock = search_queue(&res->lr_converting, &mode, policy, old_lock, flags);
1137 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock, flags);
1144 LDLM_RESOURCE_DELREF(res);
1145 ldlm_resource_putref(res);
1148 ldlm_lock2handle(lock, lockh);
1149 if ((flags & LDLM_FL_LVB_READY) &&
1150 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1151 struct l_wait_info lwi;
1152 if (lock->l_completion_ast) {
1153 int err = lock->l_completion_ast(lock,
1154 LDLM_FL_WAIT_NOREPROC,
1157 if (flags & LDLM_FL_TEST_LOCK)
1158 LDLM_LOCK_RELEASE(lock);
1160 ldlm_lock_decref_internal(lock, mode);
1166 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), NULL,
1167 LWI_ON_SIGNAL_NOOP, NULL);
1169 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1170 l_wait_event(lock->l_waitq,
1171 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1176 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1177 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1178 res_id->name[2] : policy->l_extent.start,
1179 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1180 res_id->name[3] : policy->l_extent.end);
1182 /* check user's security context */
1183 if (lock->l_conn_export &&
1184 sptlrpc_import_check_ctx(
1185 class_exp2cliimp(lock->l_conn_export))) {
1186 if (!(flags & LDLM_FL_TEST_LOCK))
1187 ldlm_lock_decref_internal(lock, mode);
1191 if (flags & LDLM_FL_TEST_LOCK)
1192 LDLM_LOCK_RELEASE(lock);
1194 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1195 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1196 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1197 type, mode, res_id->name[0], res_id->name[1],
1198 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1199 res_id->name[2] :policy->l_extent.start,
1200 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1201 res_id->name[3] : policy->l_extent.end);
1204 LDLM_LOCK_PUT(old_lock);
1206 return rc ? mode : 0;
1209 /* Returns a referenced lock */
1210 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1211 const struct ldlm_res_id *res_id,
1214 const struct ldlm_callback_suite *cbs,
1215 void *data, __u32 lvb_len)
1217 struct ldlm_lock *lock;
1218 struct ldlm_resource *res;
1221 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1225 lock = ldlm_lock_new(res);
1226 ldlm_resource_putref(res);
1231 lock->l_req_mode = mode;
1232 lock->l_ast_data = data;
1233 lock->l_pid = cfs_curproc_pid();
1235 lock->l_blocking_ast = cbs->lcs_blocking;
1236 lock->l_completion_ast = cbs->lcs_completion;
1237 lock->l_glimpse_ast = cbs->lcs_glimpse;
1238 lock->l_weigh_ast = cbs->lcs_weigh;
1241 lock->l_tree_node = NULL;
1242 /* if this is the extent lock, allocate the interval tree node */
1243 if (type == LDLM_EXTENT) {
1244 if (ldlm_interval_alloc(lock) == NULL)
1249 lock->l_lvb_len = lvb_len;
1250 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1251 if (lock->l_lvb_data == NULL)
1258 if (lock->l_lvb_data)
1259 OBD_FREE(lock->l_lvb_data, lvb_len);
1260 ldlm_interval_free(ldlm_interval_detach(lock));
1261 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1265 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1266 struct ldlm_lock **lockp,
1267 void *cookie, int *flags)
1269 struct ldlm_lock *lock = *lockp;
1270 struct ldlm_resource *res = lock->l_resource;
1271 int local = ns_is_client(res->lr_namespace);
1272 ldlm_processing_policy policy;
1273 ldlm_error_t rc = ELDLM_OK;
1274 struct ldlm_interval *node = NULL;
1277 do_gettimeofday(&lock->l_enqueued_time);
1278 /* policies are not executed on the client or during replay */
1279 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1280 && !local && ns->ns_policy) {
1281 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1283 if (rc == ELDLM_LOCK_REPLACED) {
1284 /* The lock that was returned has already been granted,
1285 * and placed into lockp. If it's not the same as the
1286 * one we passed in, then destroy the old one and our
1287 * work here is done. */
1288 if (lock != *lockp) {
1289 ldlm_lock_destroy(lock);
1290 LDLM_LOCK_RELEASE(lock);
1292 *flags |= LDLM_FL_LOCK_CHANGED;
1294 } else if (rc != ELDLM_OK ||
1295 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1296 ldlm_lock_destroy(lock);
1301 /* For a replaying lock, it might be already in granted list. So
1302 * unlinking the lock will cause the interval node to be freed, we
1303 * have to allocate the interval node early otherwise we can't regrant
1304 * this lock in the future. - jay */
1305 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1306 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO,
1309 lock_res_and_lock(lock);
1310 if (local && lock->l_req_mode == lock->l_granted_mode) {
1311 /* The server returned a blocked lock, but it was granted
1312 * before we got a chance to actually enqueue it. We don't
1313 * need to do anything else. */
1314 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1315 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1316 GOTO(out, ELDLM_OK);
1319 ldlm_resource_unlink_lock(lock);
1320 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1322 ldlm_lock_destroy_nolock(lock);
1323 GOTO(out, rc = -ENOMEM);
1326 CFS_INIT_LIST_HEAD(&node->li_group);
1327 ldlm_interval_attach(node, lock);
1331 /* Some flags from the enqueue want to make it into the AST, via the
1332 * lock's l_flags. */
1333 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1335 /* This distinction between local lock trees is very important; a client
1336 * namespace only has information about locks taken by that client, and
1337 * thus doesn't have enough information to decide for itself if it can
1338 * be granted (below). In this case, we do exactly what the server
1339 * tells us to do, as dictated by the 'flags'.
1341 * We do exactly the same thing during recovery, when the server is
1342 * more or less trusting the clients not to lie.
1344 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1345 * granted/converting queues. */
1347 if (*flags & LDLM_FL_BLOCK_CONV)
1348 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1349 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1350 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1352 ldlm_grant_lock(lock, NULL);
1353 GOTO(out, ELDLM_OK);
1354 } else if (*flags & LDLM_FL_REPLAY) {
1355 if (*flags & LDLM_FL_BLOCK_CONV) {
1356 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1357 GOTO(out, ELDLM_OK);
1358 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1359 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1360 GOTO(out, ELDLM_OK);
1361 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1362 ldlm_grant_lock(lock, NULL);
1363 GOTO(out, ELDLM_OK);
1365 /* If no flags, fall through to normal enqueue path. */
1368 policy = ldlm_processing_policy_table[res->lr_type];
1369 policy(lock, flags, 1, &rc, NULL);
1372 unlock_res_and_lock(lock);
1374 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1378 /* Must be called with namespace taken: queue is waiting or converting. */
1379 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1380 struct list_head *work_list)
1382 struct list_head *tmp, *pos;
1383 ldlm_processing_policy policy;
1385 int rc = LDLM_ITER_CONTINUE;
1389 check_res_locked(res);
1391 policy = ldlm_processing_policy_table[res->lr_type];
1394 list_for_each_safe(tmp, pos, queue) {
1395 struct ldlm_lock *pending;
1396 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1398 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1401 rc = policy(pending, &flags, 0, &err, work_list);
1402 if (rc != LDLM_ITER_CONTINUE)
1409 /* Helper function for ldlm_run_ast_work().
1411 * Send an existing rpc set specified by @arg->set and then
1412 * destroy it. Create new one if @do_create flag is set. */
1414 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1418 ptlrpc_set_wait(arg->set);
1419 if (arg->type == LDLM_BL_CALLBACK)
1420 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1421 ptlrpc_set_destroy(arg->set);
1424 arg->set = ptlrpc_prep_set();
1430 ldlm_work_bl_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1432 struct ldlm_lock_desc d;
1433 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_bl_ast);
1436 /* nobody should touch l_bl_ast */
1437 lock_res_and_lock(lock);
1438 list_del_init(&lock->l_bl_ast);
1440 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1441 LASSERT(lock->l_bl_ast_run == 0);
1442 LASSERT(lock->l_blocking_lock);
1443 lock->l_bl_ast_run++;
1444 unlock_res_and_lock(lock);
1446 ldlm_lock2desc(lock->l_blocking_lock, &d);
1448 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1449 lock->l_blocking_lock = NULL;
1450 lock->l_blocking_ast(lock, &d, (void *)arg,
1452 LDLM_LOCK_RELEASE(lock);
1458 ldlm_work_cp_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1460 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_cp_ast);
1461 ldlm_completion_callback completion_callback;
1465 /* It's possible to receive a completion AST before we've set
1466 * the l_completion_ast pointer: either because the AST arrived
1467 * before the reply, or simply because there's a small race
1468 * window between receiving the reply and finishing the local
1469 * enqueue. (bug 842)
1471 * This can't happen with the blocking_ast, however, because we
1472 * will never call the local blocking_ast until we drop our
1473 * reader/writer reference, which we won't do until we get the
1474 * reply and finish enqueueing. */
1476 /* nobody should touch l_cp_ast */
1477 lock_res_and_lock(lock);
1478 list_del_init(&lock->l_cp_ast);
1479 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1480 /* save l_completion_ast since it can be changed by
1481 * mds_intent_policy(), see bug 14225 */
1482 completion_callback = lock->l_completion_ast;
1483 lock->l_flags &= ~LDLM_FL_CP_REQD;
1484 unlock_res_and_lock(lock);
1486 if (completion_callback != NULL) {
1487 completion_callback(lock, 0, (void *)arg);
1490 LDLM_LOCK_RELEASE(lock);
1496 ldlm_work_revoke_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1498 struct ldlm_lock_desc desc;
1499 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_rk_ast);
1502 list_del_init(&lock->l_rk_ast);
1504 /* the desc just pretend to exclusive */
1505 ldlm_lock2desc(lock, &desc);
1506 desc.l_req_mode = LCK_EX;
1507 desc.l_granted_mode = 0;
1509 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1510 LDLM_LOCK_RELEASE(lock);
1515 int ldlm_run_ast_work(struct list_head *rpc_list, ldlm_desc_ast_t ast_type)
1517 struct ldlm_cb_set_arg arg;
1518 struct list_head *tmp, *pos;
1519 int (*work_ast_lock)(struct list_head *tmp, struct ldlm_cb_set_arg *arg);
1523 arg.set = ptlrpc_prep_set();
1524 atomic_set(&arg.restart, 0);
1526 case LDLM_WORK_BL_AST:
1527 arg.type = LDLM_BL_CALLBACK;
1528 work_ast_lock = ldlm_work_bl_ast_lock;
1530 case LDLM_WORK_CP_AST:
1531 arg.type = LDLM_CP_CALLBACK;
1532 work_ast_lock = ldlm_work_cp_ast_lock;
1534 case LDLM_WORK_REVOKE_AST:
1535 arg.type = LDLM_BL_CALLBACK;
1536 work_ast_lock = ldlm_work_revoke_ast_lock;
1543 list_for_each_safe(tmp, pos, rpc_list) {
1544 ast_count += work_ast_lock(tmp, &arg);
1546 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1547 * and create a new set for requests that remained in
1549 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1550 ldlm_send_and_maybe_create_set(&arg, 1);
1556 ldlm_send_and_maybe_create_set(&arg, 0);
1558 /* In case when number of ASTs is multiply of
1559 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1560 * @arg.set must be destroyed here, otherwise we get
1561 * write memory leaking. */
1562 ptlrpc_set_destroy(arg.set);
1564 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1567 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1569 ldlm_reprocess_all(res);
1570 return LDLM_ITER_CONTINUE;
1573 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1575 struct list_head *tmp;
1582 spin_lock(&ns->ns_hash_lock);
1583 for (i = 0; i < RES_HASH_SIZE; i++) {
1584 tmp = ns->ns_hash[i].next;
1585 while (tmp != &(ns->ns_hash[i])) {
1586 struct ldlm_resource *res =
1587 list_entry(tmp, struct ldlm_resource, lr_hash);
1589 ldlm_resource_getref(res);
1590 spin_unlock(&ns->ns_hash_lock);
1591 LDLM_RESOURCE_ADDREF(res);
1593 rc = reprocess_one_queue(res, NULL);
1595 LDLM_RESOURCE_DELREF(res);
1596 spin_lock(&ns->ns_hash_lock);
1598 ldlm_resource_putref_locked(res);
1600 if (rc == LDLM_ITER_STOP)
1605 spin_unlock(&ns->ns_hash_lock);
1609 void ldlm_reprocess_all(struct ldlm_resource *res)
1611 CFS_LIST_HEAD(rpc_list);
1615 /* Local lock trees don't get reprocessed. */
1616 if (ns_is_client(res->lr_namespace)) {
1623 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1624 if (rc == LDLM_ITER_CONTINUE)
1625 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1628 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1629 if (rc == -ERESTART) {
1630 LASSERT(list_empty(&rpc_list));
1636 void ldlm_cancel_callback(struct ldlm_lock *lock)
1638 check_res_locked(lock->l_resource);
1639 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1640 lock->l_flags |= LDLM_FL_CANCEL;
1641 if (lock->l_blocking_ast) {
1642 // l_check_no_ns_lock(ns);
1643 unlock_res_and_lock(lock);
1644 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1646 lock_res_and_lock(lock);
1648 LDLM_DEBUG(lock, "no blocking ast");
1651 lock->l_flags |= LDLM_FL_BL_DONE;
1654 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1656 if (req->l_resource->lr_type != LDLM_PLAIN &&
1657 req->l_resource->lr_type != LDLM_IBITS)
1660 list_del_init(&req->l_sl_policy);
1661 list_del_init(&req->l_sl_mode);
1664 void ldlm_lock_cancel(struct ldlm_lock *lock)
1666 struct ldlm_resource *res;
1667 struct ldlm_namespace *ns;
1670 lock_res_and_lock(lock);
1672 res = lock->l_resource;
1673 ns = res->lr_namespace;
1675 /* Please do not, no matter how tempting, remove this LBUG without
1676 * talking to me first. -phik */
1677 if (lock->l_readers || lock->l_writers) {
1678 LDLM_ERROR(lock, "lock still has references");
1682 ldlm_del_waiting_lock(lock);
1684 /* Releases cancel callback. */
1685 ldlm_cancel_callback(lock);
1687 /* Yes, second time, just in case it was added again while we were
1688 running with no res lock in ldlm_cancel_callback */
1689 ldlm_del_waiting_lock(lock);
1690 ldlm_resource_unlink_lock(lock);
1691 ldlm_lock_destroy_nolock(lock);
1693 if (lock->l_granted_mode == lock->l_req_mode)
1694 ldlm_pool_del(&ns->ns_pool, lock);
1696 /* Make sure we will not be called again for same lock what is possible
1697 * if not to zero out lock->l_granted_mode */
1698 lock->l_granted_mode = LCK_MINMODE;
1699 unlock_res_and_lock(lock);
1704 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1706 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1712 lock->l_ast_data = data;
1713 LDLM_LOCK_PUT(lock);
1717 void ldlm_cancel_locks_for_export_cb(void *obj, void *data)
1719 struct obd_export *exp = data;
1720 struct ldlm_lock *lock = obj;
1721 struct ldlm_resource *res;
1723 res = ldlm_resource_getref(lock->l_resource);
1724 LDLM_LOCK_GET(lock);
1726 LDLM_DEBUG(lock, "export %p", exp);
1727 ldlm_res_lvbo_update(res, NULL, 0, 1);
1728 ldlm_lock_cancel(lock);
1729 ldlm_reprocess_all(res);
1730 ldlm_resource_putref(res);
1731 LDLM_LOCK_RELEASE(lock);
1734 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1736 lustre_hash_for_each_empty(exp->exp_lock_hash,
1737 ldlm_cancel_locks_for_export_cb, exp);
1740 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1743 CFS_LIST_HEAD(rpc_list);
1744 struct ldlm_resource *res;
1745 struct ldlm_namespace *ns;
1748 struct sl_insert_point prev;
1750 struct ldlm_interval *node;
1753 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1754 *flags |= LDLM_FL_BLOCK_GRANTED;
1755 RETURN(lock->l_resource);
1758 /* I can't check the type of lock here because the bitlock of lock
1759 * is not held here, so do the allocation blindly. -jay */
1760 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO, sizeof(*node));
1761 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1764 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1765 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1767 lock_res_and_lock(lock);
1769 res = lock->l_resource;
1770 ns = res->lr_namespace;
1772 old_mode = lock->l_req_mode;
1773 lock->l_req_mode = new_mode;
1774 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1775 /* remember the lock position where the lock might be
1776 * added back to the granted list later and also
1777 * remember the join mode for skiplist fixing. */
1778 prev.res_link = lock->l_res_link.prev;
1779 prev.mode_link = lock->l_sl_mode.prev;
1780 prev.policy_link = lock->l_sl_policy.prev;
1781 ldlm_resource_unlink_lock(lock);
1783 ldlm_resource_unlink_lock(lock);
1784 if (res->lr_type == LDLM_EXTENT) {
1785 /* FIXME: ugly code, I have to attach the lock to a
1786 * interval node again since perhaps it will be granted
1788 CFS_INIT_LIST_HEAD(&node->li_group);
1789 ldlm_interval_attach(node, lock);
1794 /* If this is a local resource, put it on the appropriate list. */
1795 if (ns_is_client(res->lr_namespace)) {
1796 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1797 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1799 /* This should never happen, because of the way the
1800 * server handles conversions. */
1801 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1805 ldlm_grant_lock(lock, &rpc_list);
1807 /* FIXME: completion handling not with ns_lock held ! */
1808 if (lock->l_completion_ast)
1809 lock->l_completion_ast(lock, 0, NULL);
1813 ldlm_processing_policy policy;
1814 policy = ldlm_processing_policy_table[res->lr_type];
1815 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1816 if (rc == LDLM_ITER_STOP) {
1817 lock->l_req_mode = old_mode;
1818 if (res->lr_type == LDLM_EXTENT)
1819 ldlm_extent_add_lock(res, lock);
1821 ldlm_granted_list_add_lock(lock, &prev);
1825 *flags |= LDLM_FL_BLOCK_GRANTED;
1829 unlock_res_and_lock(lock);
1832 ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1834 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1838 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1840 struct obd_device *obd = NULL;
1842 if (!((libcfs_debug | D_ERROR) & level))
1846 CDEBUG(level, " NULL LDLM lock\n");
1850 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1851 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1853 if (lock->l_conn_export != NULL)
1854 obd = lock->l_conn_export->exp_obd;
1855 if (lock->l_export && lock->l_export->exp_connection) {
1856 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1857 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1858 lock->l_remote_handle.cookie);
1859 } else if (obd == NULL) {
1860 CDEBUG(level, " Node: local\n");
1862 struct obd_import *imp = obd->u.cli.cl_import;
1863 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1864 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1865 lock->l_remote_handle.cookie);
1867 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1869 lock->l_resource->lr_name.name[0],
1870 lock->l_resource->lr_name.name[1],
1871 lock->l_resource->lr_name.name[2]);
1872 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1873 "write: %d flags: %#x\n", ldlm_lockname[lock->l_req_mode],
1874 ldlm_lockname[lock->l_granted_mode],
1875 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1877 if (lock->l_resource->lr_type == LDLM_EXTENT)
1878 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1879 " (req "LPU64"-"LPU64")\n",
1880 lock->l_policy_data.l_extent.start,
1881 lock->l_policy_data.l_extent.end,
1882 lock->l_req_extent.start, lock->l_req_extent.end);
1883 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1884 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1885 lock->l_policy_data.l_flock.pid,
1886 lock->l_policy_data.l_flock.start,
1887 lock->l_policy_data.l_flock.end);
1888 else if (lock->l_resource->lr_type == LDLM_IBITS)
1889 CDEBUG(level, " Bits: "LPX64"\n",
1890 lock->l_policy_data.l_inodebits.bits);
1893 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1895 struct ldlm_lock *lock;
1897 if (!((libcfs_debug | D_ERROR) & level))
1900 lock = ldlm_handle2lock(lockh);
1904 ldlm_lock_dump(D_OTHER, lock, 0);
1906 LDLM_LOCK_PUT(lock);
1909 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1910 struct libcfs_debug_msg_data *data, const char *fmt,
1914 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1916 va_start(args, fmt);
1918 if (lock->l_resource == NULL) {
1919 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1920 data->msg_fn, data->msg_line, fmt, args,
1921 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1922 "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "
1923 LPX64" expref: %d pid: %u\n", 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_flags, lock->l_remote_handle.cookie,
1930 atomic_read(&lock->l_export->exp_refcount) : -99,
1936 switch (lock->l_resource->lr_type) {
1938 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1939 data->msg_fn, data->msg_line, fmt, args,
1940 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1941 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1942 "] (req "LPU64"->"LPU64") flags: %x remote: "LPX64
1943 " expref: %d pid: %u\n",
1944 lock->l_resource->lr_namespace->ns_name, lock,
1945 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1946 lock->l_readers, lock->l_writers,
1947 ldlm_lockname[lock->l_granted_mode],
1948 ldlm_lockname[lock->l_req_mode],
1949 lock->l_resource->lr_name.name[0],
1950 lock->l_resource->lr_name.name[1],
1951 atomic_read(&lock->l_resource->lr_refcount),
1952 ldlm_typename[lock->l_resource->lr_type],
1953 lock->l_policy_data.l_extent.start,
1954 lock->l_policy_data.l_extent.end,
1955 lock->l_req_extent.start, lock->l_req_extent.end,
1956 lock->l_flags, lock->l_remote_handle.cookie,
1958 atomic_read(&lock->l_export->exp_refcount) : -99,
1963 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1964 data->msg_fn, data->msg_line, fmt, args,
1965 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1966 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1967 "["LPU64"->"LPU64"] flags: %x remote: "LPX64
1968 " expref: %d pid: %u\n",
1969 lock->l_resource->lr_namespace->ns_name, lock,
1970 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1971 lock->l_readers, lock->l_writers,
1972 ldlm_lockname[lock->l_granted_mode],
1973 ldlm_lockname[lock->l_req_mode],
1974 lock->l_resource->lr_name.name[0],
1975 lock->l_resource->lr_name.name[1],
1976 atomic_read(&lock->l_resource->lr_refcount),
1977 ldlm_typename[lock->l_resource->lr_type],
1978 lock->l_policy_data.l_flock.pid,
1979 lock->l_policy_data.l_flock.start,
1980 lock->l_policy_data.l_flock.end,
1981 lock->l_flags, lock->l_remote_handle.cookie,
1983 atomic_read(&lock->l_export->exp_refcount) : -99,
1988 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1989 data->msg_fn, data->msg_line, fmt, args,
1990 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1991 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1992 "flags: %x remote: "LPX64" expref: %d "
1994 lock->l_resource->lr_namespace->ns_name,
1995 lock, lock->l_handle.h_cookie,
1996 atomic_read (&lock->l_refc),
1997 lock->l_readers, lock->l_writers,
1998 ldlm_lockname[lock->l_granted_mode],
1999 ldlm_lockname[lock->l_req_mode],
2000 lock->l_resource->lr_name.name[0],
2001 lock->l_resource->lr_name.name[1],
2002 lock->l_policy_data.l_inodebits.bits,
2003 atomic_read(&lock->l_resource->lr_refcount),
2004 ldlm_typename[lock->l_resource->lr_type],
2005 lock->l_flags, lock->l_remote_handle.cookie,
2007 atomic_read(&lock->l_export->exp_refcount) : -99,
2012 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
2013 data->msg_fn, data->msg_line, fmt, args,
2014 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2015 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: %x "
2016 "remote: "LPX64" expref: %d pid: %u\n",
2017 lock->l_resource->lr_namespace->ns_name,
2018 lock, lock->l_handle.h_cookie,
2019 atomic_read (&lock->l_refc),
2020 lock->l_readers, lock->l_writers,
2021 ldlm_lockname[lock->l_granted_mode],
2022 ldlm_lockname[lock->l_req_mode],
2023 lock->l_resource->lr_name.name[0],
2024 lock->l_resource->lr_name.name[1],
2025 atomic_read(&lock->l_resource->lr_refcount),
2026 ldlm_typename[lock->l_resource->lr_type],
2027 lock->l_flags, lock->l_remote_handle.cookie,
2029 atomic_read(&lock->l_export->exp_refcount) : -99,
2035 EXPORT_SYMBOL(_ldlm_lock_debug);