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 LASSERT(atomic_read(&lock->l_export->exp_locks_count) > 0);
168 atomic_dec(&lock->l_export->exp_locks_count);
169 class_export_put(lock->l_export);
170 lock->l_export = NULL;
173 if (lock->l_lvb_data != NULL)
174 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
176 ldlm_interval_free(ldlm_interval_detach(lock));
177 lu_ref_fini(&lock->l_reference);
178 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
185 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
188 if (!list_empty(&lock->l_lru)) {
189 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
190 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
191 list_del_init(&lock->l_lru);
192 LASSERT(ns->ns_nr_unused > 0);
199 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
201 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
204 spin_lock(&ns->ns_unused_lock);
205 rc = ldlm_lock_remove_from_lru_nolock(lock);
206 spin_unlock(&ns->ns_unused_lock);
211 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
213 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
214 lock->l_last_used = cfs_time_current();
215 LASSERT(list_empty(&lock->l_lru));
216 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
217 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
218 LASSERT(ns->ns_nr_unused >= 0);
222 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
224 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
226 spin_lock(&ns->ns_unused_lock);
227 ldlm_lock_add_to_lru_nolock(lock);
228 spin_unlock(&ns->ns_unused_lock);
232 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
234 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
236 spin_lock(&ns->ns_unused_lock);
237 if (!list_empty(&lock->l_lru)) {
238 ldlm_lock_remove_from_lru_nolock(lock);
239 ldlm_lock_add_to_lru_nolock(lock);
241 spin_unlock(&ns->ns_unused_lock);
245 /* This used to have a 'strict' flag, which recovery would use to mark an
246 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
247 * shall explain why it's gone: with the new hash table scheme, once you call
248 * ldlm_lock_destroy, you can never drop your final references on this lock.
249 * Because it's not in the hash table anymore. -phil */
250 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
254 if (lock->l_readers || lock->l_writers) {
255 LDLM_ERROR(lock, "lock still has references");
256 ldlm_lock_dump(D_ERROR, lock, 0);
260 if (!list_empty(&lock->l_res_link)) {
261 LDLM_ERROR(lock, "lock still on resource");
262 ldlm_lock_dump(D_ERROR, lock, 0);
266 if (lock->l_destroyed) {
267 LASSERT(list_empty(&lock->l_lru));
271 lock->l_destroyed = 1;
273 if (lock->l_export && lock->l_export->exp_lock_hash &&
274 !hlist_unhashed(&lock->l_exp_hash))
275 lustre_hash_del(lock->l_export->exp_lock_hash,
276 &lock->l_remote_handle, &lock->l_exp_hash);
278 ldlm_lock_remove_from_lru(lock);
279 class_handle_unhash(&lock->l_handle);
282 /* Wake anyone waiting for this lock */
283 /* FIXME: I should probably add yet another flag, instead of using
284 * l_export to only call this on clients */
286 class_export_put(lock->l_export);
287 lock->l_export = NULL;
288 if (lock->l_export && lock->l_completion_ast)
289 lock->l_completion_ast(lock, 0);
295 void ldlm_lock_destroy(struct ldlm_lock *lock)
299 lock_res_and_lock(lock);
300 first = ldlm_lock_destroy_internal(lock);
301 unlock_res_and_lock(lock);
303 /* drop reference from hashtable only for first destroy */
305 lu_ref_del(&lock->l_reference, "hash", lock);
306 LDLM_LOCK_RELEASE(lock);
311 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
315 first = ldlm_lock_destroy_internal(lock);
316 /* drop reference from hashtable only for first destroy */
318 lu_ref_del(&lock->l_reference, "hash", lock);
319 LDLM_LOCK_RELEASE(lock);
324 /* this is called by portals_handle2object with the handle lock taken */
325 static void lock_handle_addref(void *lock)
327 LDLM_LOCK_GET((struct ldlm_lock *)lock);
331 * usage: pass in a resource on which you have done ldlm_resource_get
332 * pass in a parent lock on which you have done a ldlm_lock_get
333 * after return, ldlm_*_put the resource and parent
334 * returns: lock with refcount 2 - one for current caller and one for remote
336 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
338 struct ldlm_lock *lock;
341 if (resource == NULL)
344 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, CFS_ALLOC_IO);
348 spin_lock_init(&lock->l_lock);
349 lock->l_resource = ldlm_resource_getref(resource);
350 lu_ref_add(&resource->lr_reference, "lock", lock);
352 atomic_set(&lock->l_refc, 2);
353 CFS_INIT_LIST_HEAD(&lock->l_res_link);
354 CFS_INIT_LIST_HEAD(&lock->l_lru);
355 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
356 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
357 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
358 CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
359 cfs_waitq_init(&lock->l_waitq);
360 lock->l_blocking_lock = NULL;
361 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
362 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
363 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
365 atomic_inc(&resource->lr_namespace->ns_locks);
366 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
367 class_handle_hash(&lock->l_handle, lock_handle_addref);
369 CFS_INIT_LIST_HEAD(&lock->l_extents_list);
370 spin_lock_init(&lock->l_extents_list_lock);
371 CFS_INIT_LIST_HEAD(&lock->l_cache_locks_list);
372 lu_ref_init(&lock->l_reference);
373 lu_ref_add(&lock->l_reference, "hash", lock);
374 lock->l_callback_timeout = 0;
379 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
380 const struct ldlm_res_id *new_resid)
382 struct ldlm_resource *oldres = lock->l_resource;
383 struct ldlm_resource *newres;
387 LASSERT(ns_is_client(ns));
389 lock_res_and_lock(lock);
390 if (memcmp(new_resid, &lock->l_resource->lr_name,
391 sizeof(lock->l_resource->lr_name)) == 0) {
393 unlock_res_and_lock(lock);
397 LASSERT(new_resid->name[0] != 0);
399 /* This function assumes that the lock isn't on any lists */
400 LASSERT(list_empty(&lock->l_res_link));
402 type = oldres->lr_type;
403 unlock_res_and_lock(lock);
405 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
406 lu_ref_add(&newres->lr_reference, "lock", lock);
410 * To flip the lock from the old to the new resource, lock, oldres and
411 * newres have to be locked. Resource spin-locks are nested within
412 * lock->l_lock, and are taken in the memory address order to avoid
415 spin_lock(&lock->l_lock);
416 oldres = lock->l_resource;
417 if (oldres < newres) {
419 lock_res_nested(newres, LRT_NEW);
422 lock_res_nested(oldres, LRT_NEW);
424 LASSERT(memcmp(new_resid, &oldres->lr_name,
425 sizeof oldres->lr_name) != 0);
426 lock->l_resource = newres;
428 unlock_res_and_lock(lock);
430 /* ...and the flowers are still standing! */
431 lu_ref_del(&oldres->lr_reference, "lock", lock);
432 ldlm_resource_putref(oldres);
441 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
443 lockh->cookie = lock->l_handle.h_cookie;
446 /* if flags: atomically get the lock and set the flags.
447 * Return NULL if flag already set
450 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
453 struct ldlm_namespace *ns;
454 struct ldlm_lock *lock, *retval = NULL;
459 lock = class_handle2object(handle->cookie);
463 LASSERT(lock->l_resource != NULL);
464 ns = lock->l_resource->lr_namespace;
467 lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
468 lock_res_and_lock(lock);
470 /* It's unlikely but possible that someone marked the lock as
471 * destroyed after we did handle2object on it */
472 if (lock->l_destroyed) {
473 unlock_res_and_lock(lock);
474 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
479 if (flags && (lock->l_flags & flags)) {
480 unlock_res_and_lock(lock);
486 lock->l_flags |= flags;
488 unlock_res_and_lock(lock);
495 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
497 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
498 /* INODEBITS_INTEROP: If the other side does not support
499 * inodebits, reply with a plain lock descriptor.
501 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
502 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
503 struct ldlm_resource res = *lock->l_resource;
505 /* Make sure all the right bits are set in this lock we
506 are going to pass to client */
507 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
508 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
509 "Inappropriate inode lock bits during "
510 "conversion " LPU64 "\n",
511 lock->l_policy_data.l_inodebits.bits);
512 res.lr_type = LDLM_PLAIN;
513 ldlm_res2desc(&res, &desc->l_resource);
514 /* Convert "new" lock mode to something old client can
516 if ((lock->l_req_mode == LCK_CR) ||
517 (lock->l_req_mode == LCK_CW))
518 desc->l_req_mode = LCK_PR;
520 desc->l_req_mode = lock->l_req_mode;
521 if ((lock->l_granted_mode == LCK_CR) ||
522 (lock->l_granted_mode == LCK_CW)) {
523 desc->l_granted_mode = LCK_PR;
525 /* We never grant PW/EX locks to clients */
526 LASSERT((lock->l_granted_mode != LCK_PW) &&
527 (lock->l_granted_mode != LCK_EX));
528 desc->l_granted_mode = lock->l_granted_mode;
531 /* We do not copy policy here, because there is no
532 policy for plain locks */
534 ldlm_res2desc(lock->l_resource, &desc->l_resource);
535 desc->l_req_mode = lock->l_req_mode;
536 desc->l_granted_mode = lock->l_granted_mode;
537 desc->l_policy_data = lock->l_policy_data;
541 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
542 struct list_head *work_list)
544 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
545 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
546 lock->l_flags |= LDLM_FL_AST_SENT;
547 /* If the enqueuing client said so, tell the AST recipient to
548 * discard dirty data, rather than writing back. */
549 if (new->l_flags & LDLM_AST_DISCARD_DATA)
550 lock->l_flags |= LDLM_FL_DISCARD_DATA;
551 LASSERT(list_empty(&lock->l_bl_ast));
552 list_add(&lock->l_bl_ast, work_list);
554 LASSERT(lock->l_blocking_lock == NULL);
555 lock->l_blocking_lock = LDLM_LOCK_GET(new);
559 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
561 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
562 lock->l_flags |= LDLM_FL_CP_REQD;
563 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
564 LASSERT(list_empty(&lock->l_cp_ast));
565 list_add(&lock->l_cp_ast, work_list);
570 /* must be called with lr_lock held */
571 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
572 struct list_head *work_list)
575 check_res_locked(lock->l_resource);
577 ldlm_add_bl_work_item(lock, new, work_list);
579 ldlm_add_cp_work_item(lock, work_list);
583 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
585 struct ldlm_lock *lock;
587 lock = ldlm_handle2lock(lockh);
588 LASSERT(lock != NULL);
589 ldlm_lock_addref_internal(lock, mode);
593 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
595 ldlm_lock_remove_from_lru(lock);
596 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
598 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
600 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
602 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
605 lu_ref_add_atomic(&lock->l_reference, "user", lock);
606 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
610 * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
613 * \retval 0 success, lock was addref-ed
615 * \retval -EAGAIN lock is being canceled.
617 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
619 struct ldlm_lock *lock;
623 lock = ldlm_handle2lock(lockh);
625 lock_res_and_lock(lock);
626 if (lock->l_readers != 0 || lock->l_writers != 0 ||
627 !(lock->l_flags & LDLM_FL_CBPENDING)) {
628 ldlm_lock_addref_internal_nolock(lock, mode);
631 unlock_res_and_lock(lock);
637 /* only called for local locks */
638 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
640 lock_res_and_lock(lock);
641 ldlm_lock_addref_internal_nolock(lock, mode);
642 unlock_res_and_lock(lock);
645 /* only called in ldlm_flock_destroy and for local locks.
646 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
647 * * ldlm_lock_remove_from_lru() does nothing, it is safe
648 * * for ldlm_flock_destroy usage by dropping some code */
649 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
651 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
652 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
653 LASSERT(lock->l_readers > 0);
654 lu_ref_del(&lock->l_reference, "reader", lock);
657 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
658 LASSERT(lock->l_writers > 0);
659 lu_ref_del(&lock->l_reference, "writer", lock);
663 lu_ref_del(&lock->l_reference, "user", lock);
664 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
667 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
669 struct ldlm_namespace *ns;
672 lock_res_and_lock(lock);
674 ns = lock->l_resource->lr_namespace;
676 ldlm_lock_decref_internal_nolock(lock, mode);
678 if (lock->l_flags & LDLM_FL_LOCAL &&
679 !lock->l_readers && !lock->l_writers) {
680 /* If this is a local lock on a server namespace and this was
681 * the last reference, cancel the lock. */
682 CDEBUG(D_INFO, "forcing cancel of local lock\n");
683 lock->l_flags |= LDLM_FL_CBPENDING;
686 if (!lock->l_readers && !lock->l_writers &&
687 (lock->l_flags & LDLM_FL_CBPENDING)) {
688 /* If we received a blocked AST and this was the last reference,
689 * run the callback. */
690 if (ns_is_server(ns) && lock->l_export)
691 CERROR("FL_CBPENDING set on non-local lock--just a "
694 LDLM_DEBUG(lock, "final decref done on cbpending lock");
696 LDLM_LOCK_GET(lock); /* dropped by bl thread */
697 ldlm_lock_remove_from_lru(lock);
698 unlock_res_and_lock(lock);
700 if (lock->l_flags & LDLM_FL_FAIL_LOC)
701 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
703 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
704 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
705 ldlm_handle_bl_callback(ns, NULL, lock);
706 } else if (ns_is_client(ns) &&
707 !lock->l_readers && !lock->l_writers &&
708 !(lock->l_flags & LDLM_FL_BL_AST)) {
709 /* If this is a client-side namespace and this was the last
710 * reference, put it on the LRU. */
711 ldlm_lock_add_to_lru(lock);
712 unlock_res_and_lock(lock);
714 if (lock->l_flags & LDLM_FL_FAIL_LOC)
715 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
717 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
718 * are not supported by the server, otherwise, it is done on
720 if (!exp_connect_cancelset(lock->l_conn_export) &&
721 !ns_connect_lru_resize(ns))
722 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
724 unlock_res_and_lock(lock);
730 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
732 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
733 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
734 ldlm_lock_decref_internal(lock, mode);
738 /* This will drop a lock reference and mark it for destruction, but will not
739 * necessarily cancel the lock before returning. */
740 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
742 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
745 LASSERT(lock != NULL);
747 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
748 lock_res_and_lock(lock);
749 lock->l_flags |= LDLM_FL_CBPENDING;
750 unlock_res_and_lock(lock);
751 ldlm_lock_decref_internal(lock, mode);
755 struct sl_insert_point {
756 struct list_head *res_link;
757 struct list_head *mode_link;
758 struct list_head *policy_link;
762 * search_granted_lock
765 * Finds a position to insert the new lock.
767 * queue [input]: the granted list where search acts on;
768 * req [input]: the lock whose position to be located;
769 * prev [output]: positions within 3 lists to insert @req to
773 * - ldlm_grant_lock_with_skiplist
775 static void search_granted_lock(struct list_head *queue,
776 struct ldlm_lock *req,
777 struct sl_insert_point *prev)
779 struct list_head *tmp;
780 struct ldlm_lock *lock, *mode_end, *policy_end;
783 list_for_each(tmp, queue) {
784 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
786 mode_end = list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
789 if (lock->l_req_mode != req->l_req_mode) {
790 /* jump to last lock of mode group */
791 tmp = &mode_end->l_res_link;
795 /* suitable mode group is found */
796 if (lock->l_resource->lr_type == LDLM_PLAIN) {
797 /* insert point is last lock of the mode group */
798 prev->res_link = &mode_end->l_res_link;
799 prev->mode_link = &mode_end->l_sl_mode;
800 prev->policy_link = &req->l_sl_policy;
803 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
805 policy_end = list_entry(lock->l_sl_policy.prev,
809 if (lock->l_policy_data.l_inodebits.bits ==
810 req->l_policy_data.l_inodebits.bits) {
811 /* insert point is last lock of
812 * the policy group */
814 &policy_end->l_res_link;
816 &policy_end->l_sl_mode;
818 &policy_end->l_sl_policy;
823 if (policy_end == mode_end)
824 /* done with mode group */
827 /* go to next policy group within mode group */
828 tmp = policy_end->l_res_link.next;
829 lock = list_entry(tmp, struct ldlm_lock,
831 } /* loop over policy groups within the mode group */
833 /* insert point is last lock of the mode group,
834 * new policy group is started */
835 prev->res_link = &mode_end->l_res_link;
836 prev->mode_link = &mode_end->l_sl_mode;
837 prev->policy_link = &req->l_sl_policy;
841 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
846 /* insert point is last lock on the queue,
847 * new mode group and new policy group are started */
848 prev->res_link = queue->prev;
849 prev->mode_link = &req->l_sl_mode;
850 prev->policy_link = &req->l_sl_policy;
855 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
856 struct sl_insert_point *prev)
858 struct ldlm_resource *res = lock->l_resource;
861 check_res_locked(res);
863 ldlm_resource_dump(D_INFO, res);
864 CDEBUG(D_OTHER, "About to add this lock:\n");
865 ldlm_lock_dump(D_OTHER, lock, 0);
867 if (lock->l_destroyed) {
868 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
872 LASSERT(list_empty(&lock->l_res_link));
873 LASSERT(list_empty(&lock->l_sl_mode));
874 LASSERT(list_empty(&lock->l_sl_policy));
876 list_add(&lock->l_res_link, prev->res_link);
877 list_add(&lock->l_sl_mode, prev->mode_link);
878 list_add(&lock->l_sl_policy, prev->policy_link);
883 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
885 struct sl_insert_point prev;
888 LASSERT(lock->l_req_mode == lock->l_granted_mode);
890 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
891 ldlm_granted_list_add_lock(lock, &prev);
896 * - ldlm_lock_enqueue
897 * - ldlm_reprocess_queue
898 * - ldlm_lock_convert
900 * must be called with lr_lock held
902 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
904 struct ldlm_resource *res = lock->l_resource;
907 check_res_locked(res);
909 lock->l_granted_mode = lock->l_req_mode;
910 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
911 ldlm_grant_lock_with_skiplist(lock);
912 else if (res->lr_type == LDLM_EXTENT)
913 ldlm_extent_add_lock(res, lock);
915 ldlm_resource_add_lock(res, &res->lr_granted, lock);
917 if (lock->l_granted_mode < res->lr_most_restr)
918 res->lr_most_restr = lock->l_granted_mode;
920 if (work_list && lock->l_completion_ast != NULL)
921 ldlm_add_ast_work_item(lock, NULL, work_list);
923 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
927 /* returns a referenced lock or NULL. See the flag descriptions below, in the
928 * comment above ldlm_lock_match */
929 static struct ldlm_lock *search_queue(struct list_head *queue,
931 ldlm_policy_data_t *policy,
932 struct ldlm_lock *old_lock,
933 int flags, int unref)
935 struct ldlm_lock *lock;
936 struct list_head *tmp;
938 list_for_each(tmp, queue) {
941 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
943 if (lock == old_lock)
946 /* llite sometimes wants to match locks that will be
947 * canceled when their users drop, but we allow it to match
948 * if it passes in CBPENDING and the lock still has users.
949 * this is generally only going to be used by children
950 * whose parents already hold a lock so forward progress
951 * can still happen. */
952 if (lock->l_flags & LDLM_FL_CBPENDING &&
953 !(flags & LDLM_FL_CBPENDING))
955 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
956 lock->l_readers == 0 && lock->l_writers == 0)
959 if (!(lock->l_req_mode & *mode))
961 match = lock->l_req_mode;
963 if (lock->l_resource->lr_type == LDLM_EXTENT &&
964 (lock->l_policy_data.l_extent.start >
965 policy->l_extent.start ||
966 lock->l_policy_data.l_extent.end < policy->l_extent.end))
969 if (unlikely(match == LCK_GROUP) &&
970 lock->l_resource->lr_type == LDLM_EXTENT &&
971 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
974 /* We match if we have existing lock with same or wider set
976 if (lock->l_resource->lr_type == LDLM_IBITS &&
977 ((lock->l_policy_data.l_inodebits.bits &
978 policy->l_inodebits.bits) !=
979 policy->l_inodebits.bits))
983 (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED)))
986 if ((flags & LDLM_FL_LOCAL_ONLY) &&
987 !(lock->l_flags & LDLM_FL_LOCAL))
990 if (flags & LDLM_FL_TEST_LOCK) {
992 ldlm_lock_touch_in_lru(lock);
994 ldlm_lock_addref_internal_nolock(lock, match);
1003 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1005 lock->l_flags |= LDLM_FL_LVB_READY;
1006 cfs_waitq_signal(&lock->l_waitq);
1009 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1011 lock_res_and_lock(lock);
1012 ldlm_lock_allow_match_locked(lock);
1013 unlock_res_and_lock(lock);
1016 /* Can be called in two ways:
1018 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1019 * for a duplicate of.
1021 * Otherwise, all of the fields must be filled in, to match against.
1023 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1024 * server (ie, connh is NULL)
1025 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1026 * list will be considered
1027 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1028 * to be canceled can still be matched as long as they still have reader
1029 * or writer refernces
1030 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1031 * just tell us if we would have matched.
1033 * Returns 1 if it finds an already-existing lock that is compatible; in this
1034 * case, lockh is filled in with a addref()ed lock
1036 * we also check security context, if that failed we simply return 0 (to keep
1037 * caller code unchanged), the context failure will be discovered by caller
1040 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1041 const struct ldlm_res_id *res_id, ldlm_type_t type,
1042 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1043 struct lustre_handle *lockh, int unref)
1045 struct ldlm_resource *res;
1046 struct ldlm_lock *lock, *old_lock = NULL;
1051 old_lock = ldlm_handle2lock(lockh);
1054 ns = old_lock->l_resource->lr_namespace;
1055 res_id = &old_lock->l_resource->lr_name;
1056 type = old_lock->l_resource->lr_type;
1057 mode = old_lock->l_req_mode;
1060 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1062 LASSERT(old_lock == NULL);
1066 LDLM_RESOURCE_ADDREF(res);
1069 lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1073 if (flags & LDLM_FL_BLOCK_GRANTED)
1075 lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1079 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1087 LDLM_RESOURCE_DELREF(res);
1088 ldlm_resource_putref(res);
1091 ldlm_lock2handle(lock, lockh);
1092 if ((flags & LDLM_FL_LVB_READY) &&
1093 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1094 struct l_wait_info lwi;
1095 if (lock->l_completion_ast) {
1096 int err = lock->l_completion_ast(lock,
1097 LDLM_FL_WAIT_NOREPROC,
1100 if (flags & LDLM_FL_TEST_LOCK)
1101 LDLM_LOCK_RELEASE(lock);
1103 ldlm_lock_decref_internal(lock,
1110 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1111 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1113 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1114 l_wait_event(lock->l_waitq,
1115 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1120 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1121 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1122 res_id->name[2] : policy->l_extent.start,
1123 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1124 res_id->name[3] : policy->l_extent.end);
1126 /* check user's security context */
1127 if (lock->l_conn_export &&
1128 sptlrpc_import_check_ctx(
1129 class_exp2cliimp(lock->l_conn_export))) {
1130 if (!(flags & LDLM_FL_TEST_LOCK))
1131 ldlm_lock_decref_internal(lock, mode);
1135 if (flags & LDLM_FL_TEST_LOCK)
1136 LDLM_LOCK_RELEASE(lock);
1138 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1139 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1140 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1141 type, mode, res_id->name[0], res_id->name[1],
1142 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1143 res_id->name[2] :policy->l_extent.start,
1144 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1145 res_id->name[3] : policy->l_extent.end);
1148 LDLM_LOCK_PUT(old_lock);
1150 return rc ? mode : 0;
1153 /* Returns a referenced lock */
1154 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1155 const struct ldlm_res_id *res_id,
1158 const struct ldlm_callback_suite *cbs,
1159 void *data, __u32 lvb_len)
1161 struct ldlm_lock *lock;
1162 struct ldlm_resource *res;
1165 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1169 lock = ldlm_lock_new(res);
1170 ldlm_resource_putref(res);
1175 lock->l_req_mode = mode;
1176 lock->l_ast_data = data;
1177 lock->l_pid = cfs_curproc_pid();
1179 lock->l_blocking_ast = cbs->lcs_blocking;
1180 lock->l_completion_ast = cbs->lcs_completion;
1181 lock->l_glimpse_ast = cbs->lcs_glimpse;
1182 lock->l_weigh_ast = cbs->lcs_weigh;
1185 lock->l_tree_node = NULL;
1186 /* if this is the extent lock, allocate the interval tree node */
1187 if (type == LDLM_EXTENT) {
1188 if (ldlm_interval_alloc(lock) == NULL)
1193 lock->l_lvb_len = lvb_len;
1194 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1195 if (lock->l_lvb_data == NULL)
1202 if (lock->l_lvb_data)
1203 OBD_FREE(lock->l_lvb_data, lvb_len);
1204 ldlm_interval_free(ldlm_interval_detach(lock));
1205 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1209 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1210 struct ldlm_lock **lockp,
1211 void *cookie, int *flags)
1213 struct ldlm_lock *lock = *lockp;
1214 struct ldlm_resource *res = lock->l_resource;
1215 int local = ns_is_client(res->lr_namespace);
1216 ldlm_processing_policy policy;
1217 ldlm_error_t rc = ELDLM_OK;
1218 struct ldlm_interval *node = NULL;
1221 lock->l_last_activity = cfs_time_current_sec();
1222 /* policies are not executed on the client or during replay */
1223 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1224 && !local && ns->ns_policy) {
1225 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1227 if (rc == ELDLM_LOCK_REPLACED) {
1228 /* The lock that was returned has already been granted,
1229 * and placed into lockp. If it's not the same as the
1230 * one we passed in, then destroy the old one and our
1231 * work here is done. */
1232 if (lock != *lockp) {
1233 ldlm_lock_destroy(lock);
1234 LDLM_LOCK_RELEASE(lock);
1236 *flags |= LDLM_FL_LOCK_CHANGED;
1238 } else if (rc != ELDLM_OK ||
1239 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1240 ldlm_lock_destroy(lock);
1245 /* For a replaying lock, it might be already in granted list. So
1246 * unlinking the lock will cause the interval node to be freed, we
1247 * have to allocate the interval node early otherwise we can't regrant
1248 * this lock in the future. - jay */
1249 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1250 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1252 lock_res_and_lock(lock);
1253 if (local && lock->l_req_mode == lock->l_granted_mode) {
1254 /* The server returned a blocked lock, but it was granted
1255 * before we got a chance to actually enqueue it. We don't
1256 * need to do anything else. */
1257 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1258 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1259 GOTO(out, ELDLM_OK);
1262 ldlm_resource_unlink_lock(lock);
1263 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1265 ldlm_lock_destroy_nolock(lock);
1266 GOTO(out, rc = -ENOMEM);
1269 CFS_INIT_LIST_HEAD(&node->li_group);
1270 ldlm_interval_attach(node, lock);
1274 /* Some flags from the enqueue want to make it into the AST, via the
1275 * lock's l_flags. */
1276 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1278 /* This distinction between local lock trees is very important; a client
1279 * namespace only has information about locks taken by that client, and
1280 * thus doesn't have enough information to decide for itself if it can
1281 * be granted (below). In this case, we do exactly what the server
1282 * tells us to do, as dictated by the 'flags'.
1284 * We do exactly the same thing during recovery, when the server is
1285 * more or less trusting the clients not to lie.
1287 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1288 * granted/converting queues. */
1290 if (*flags & LDLM_FL_BLOCK_CONV)
1291 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1292 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1293 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1295 ldlm_grant_lock(lock, NULL);
1296 GOTO(out, ELDLM_OK);
1297 } else if (*flags & LDLM_FL_REPLAY) {
1298 if (*flags & LDLM_FL_BLOCK_CONV) {
1299 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1300 GOTO(out, ELDLM_OK);
1301 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1302 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1303 GOTO(out, ELDLM_OK);
1304 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1305 ldlm_grant_lock(lock, NULL);
1306 GOTO(out, ELDLM_OK);
1308 /* If no flags, fall through to normal enqueue path. */
1311 policy = ldlm_processing_policy_table[res->lr_type];
1312 policy(lock, flags, 1, &rc, NULL);
1315 unlock_res_and_lock(lock);
1317 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1321 /* Must be called with namespace taken: queue is waiting or converting. */
1322 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1323 struct list_head *work_list)
1325 struct list_head *tmp, *pos;
1326 ldlm_processing_policy policy;
1328 int rc = LDLM_ITER_CONTINUE;
1332 check_res_locked(res);
1334 policy = ldlm_processing_policy_table[res->lr_type];
1337 list_for_each_safe(tmp, pos, queue) {
1338 struct ldlm_lock *pending;
1339 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1341 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1344 rc = policy(pending, &flags, 0, &err, work_list);
1345 if (rc != LDLM_ITER_CONTINUE)
1352 /* Helper function for ldlm_run_ast_work().
1354 * Send an existing rpc set specified by @arg->set and then
1355 * destroy it. Create new one if @do_create flag is set. */
1357 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1361 ptlrpc_set_wait(arg->set);
1362 if (arg->type == LDLM_BL_CALLBACK)
1363 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1364 ptlrpc_set_destroy(arg->set);
1367 arg->set = ptlrpc_prep_set();
1373 ldlm_work_bl_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1375 struct ldlm_lock_desc d;
1376 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_bl_ast);
1379 /* nobody should touch l_bl_ast */
1380 lock_res_and_lock(lock);
1381 list_del_init(&lock->l_bl_ast);
1383 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1384 LASSERT(lock->l_bl_ast_run == 0);
1385 LASSERT(lock->l_blocking_lock);
1386 lock->l_bl_ast_run++;
1387 unlock_res_and_lock(lock);
1389 ldlm_lock2desc(lock->l_blocking_lock, &d);
1391 lock->l_blocking_ast(lock, &d, (void *)arg,
1393 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1394 lock->l_blocking_lock = NULL;
1395 LDLM_LOCK_RELEASE(lock);
1401 ldlm_work_cp_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1403 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_cp_ast);
1404 ldlm_completion_callback completion_callback;
1408 /* It's possible to receive a completion AST before we've set
1409 * the l_completion_ast pointer: either because the AST arrived
1410 * before the reply, or simply because there's a small race
1411 * window between receiving the reply and finishing the local
1412 * enqueue. (bug 842)
1414 * This can't happen with the blocking_ast, however, because we
1415 * will never call the local blocking_ast until we drop our
1416 * reader/writer reference, which we won't do until we get the
1417 * reply and finish enqueueing. */
1419 /* nobody should touch l_cp_ast */
1420 lock_res_and_lock(lock);
1421 list_del_init(&lock->l_cp_ast);
1422 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1423 /* save l_completion_ast since it can be changed by
1424 * mds_intent_policy(), see bug 14225 */
1425 completion_callback = lock->l_completion_ast;
1426 lock->l_flags &= ~LDLM_FL_CP_REQD;
1427 unlock_res_and_lock(lock);
1429 if (completion_callback != NULL) {
1430 completion_callback(lock, 0, (void *)arg);
1433 LDLM_LOCK_RELEASE(lock);
1439 ldlm_work_revoke_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1441 struct ldlm_lock_desc desc;
1442 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_rk_ast);
1445 list_del_init(&lock->l_rk_ast);
1447 /* the desc just pretend to exclusive */
1448 ldlm_lock2desc(lock, &desc);
1449 desc.l_req_mode = LCK_EX;
1450 desc.l_granted_mode = 0;
1452 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1453 LDLM_LOCK_RELEASE(lock);
1458 int ldlm_run_ast_work(struct list_head *rpc_list, ldlm_desc_ast_t ast_type)
1460 struct ldlm_cb_set_arg arg;
1461 struct list_head *tmp, *pos;
1462 int (*work_ast_lock)(struct list_head *tmp,struct ldlm_cb_set_arg *arg);
1466 arg.set = ptlrpc_prep_set();
1467 if (NULL == arg.set)
1469 atomic_set(&arg.restart, 0);
1471 case LDLM_WORK_BL_AST:
1472 arg.type = LDLM_BL_CALLBACK;
1473 work_ast_lock = ldlm_work_bl_ast_lock;
1475 case LDLM_WORK_CP_AST:
1476 arg.type = LDLM_CP_CALLBACK;
1477 work_ast_lock = ldlm_work_cp_ast_lock;
1479 case LDLM_WORK_REVOKE_AST:
1480 arg.type = LDLM_BL_CALLBACK;
1481 work_ast_lock = ldlm_work_revoke_ast_lock;
1488 list_for_each_safe(tmp, pos, rpc_list) {
1489 ast_count += work_ast_lock(tmp, &arg);
1491 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1492 * and create a new set for requests that remained in
1494 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1495 ldlm_send_and_maybe_create_set(&arg, 1);
1501 ldlm_send_and_maybe_create_set(&arg, 0);
1503 /* In case when number of ASTs is multiply of
1504 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1505 * @arg.set must be destroyed here, otherwise we get
1506 * write memory leaking. */
1507 ptlrpc_set_destroy(arg.set);
1509 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1512 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1514 ldlm_reprocess_all(res);
1515 return LDLM_ITER_CONTINUE;
1518 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1520 struct list_head *tmp;
1527 spin_lock(&ns->ns_hash_lock);
1528 for (i = 0; i < RES_HASH_SIZE; i++) {
1529 tmp = ns->ns_hash[i].next;
1530 while (tmp != &(ns->ns_hash[i])) {
1531 struct ldlm_resource *res =
1532 list_entry(tmp, struct ldlm_resource, lr_hash);
1534 ldlm_resource_getref(res);
1535 spin_unlock(&ns->ns_hash_lock);
1536 LDLM_RESOURCE_ADDREF(res);
1538 rc = reprocess_one_queue(res, NULL);
1540 LDLM_RESOURCE_DELREF(res);
1541 spin_lock(&ns->ns_hash_lock);
1543 ldlm_resource_putref_locked(res);
1545 if (rc == LDLM_ITER_STOP)
1550 spin_unlock(&ns->ns_hash_lock);
1554 void ldlm_reprocess_all(struct ldlm_resource *res)
1556 CFS_LIST_HEAD(rpc_list);
1560 /* Local lock trees don't get reprocessed. */
1561 if (ns_is_client(res->lr_namespace)) {
1568 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1569 if (rc == LDLM_ITER_CONTINUE)
1570 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1573 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1574 if (rc == -ERESTART) {
1575 LASSERT(list_empty(&rpc_list));
1581 void ldlm_cancel_callback(struct ldlm_lock *lock)
1583 check_res_locked(lock->l_resource);
1584 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1585 lock->l_flags |= LDLM_FL_CANCEL;
1586 if (lock->l_blocking_ast) {
1587 // l_check_no_ns_lock(ns);
1588 unlock_res_and_lock(lock);
1589 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1591 lock_res_and_lock(lock);
1593 LDLM_DEBUG(lock, "no blocking ast");
1596 lock->l_flags |= LDLM_FL_BL_DONE;
1599 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1601 if (req->l_resource->lr_type != LDLM_PLAIN &&
1602 req->l_resource->lr_type != LDLM_IBITS)
1605 list_del_init(&req->l_sl_policy);
1606 list_del_init(&req->l_sl_mode);
1609 void ldlm_lock_cancel(struct ldlm_lock *lock)
1611 struct ldlm_resource *res;
1612 struct ldlm_namespace *ns;
1615 lock_res_and_lock(lock);
1617 res = lock->l_resource;
1618 ns = res->lr_namespace;
1620 /* Please do not, no matter how tempting, remove this LBUG without
1621 * talking to me first. -phik */
1622 if (lock->l_readers || lock->l_writers) {
1623 LDLM_ERROR(lock, "lock still has references");
1627 ldlm_del_waiting_lock(lock);
1629 /* Releases cancel callback. */
1630 ldlm_cancel_callback(lock);
1632 /* Yes, second time, just in case it was added again while we were
1633 running with no res lock in ldlm_cancel_callback */
1634 ldlm_del_waiting_lock(lock);
1635 ldlm_resource_unlink_lock(lock);
1636 ldlm_lock_destroy_nolock(lock);
1638 if (lock->l_granted_mode == lock->l_req_mode)
1639 ldlm_pool_del(&ns->ns_pool, lock);
1641 /* Make sure we will not be called again for same lock what is possible
1642 * if not to zero out lock->l_granted_mode */
1643 lock->l_granted_mode = LCK_MINMODE;
1644 unlock_res_and_lock(lock);
1649 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1651 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1657 lock->l_ast_data = data;
1658 LDLM_LOCK_PUT(lock);
1662 void ldlm_cancel_locks_for_export_cb(void *obj, void *data)
1664 struct obd_export *exp = data;
1665 struct ldlm_lock *lock = obj;
1666 struct ldlm_resource *res;
1668 res = ldlm_resource_getref(lock->l_resource);
1669 LDLM_LOCK_GET(lock);
1671 LDLM_DEBUG(lock, "export %p", exp);
1672 ldlm_res_lvbo_update(res, NULL, 0, 1);
1673 ldlm_lock_cancel(lock);
1674 ldlm_reprocess_all(res);
1675 ldlm_resource_putref(res);
1676 LDLM_LOCK_RELEASE(lock);
1679 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1681 lustre_hash_for_each_empty(exp->exp_lock_hash,
1682 ldlm_cancel_locks_for_export_cb, exp);
1686 * Downgrade an exclusive lock.
1688 * A fast variant of ldlm_lock_convert for convertion of exclusive
1689 * locks. The convertion is always successful.
1691 * \param lock A lock to convert
1692 * \param new_mode new lock mode
1694 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1696 struct ldlm_namespace *ns;
1699 LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1700 LASSERT(new_mode == LCK_COS);
1702 lock_res_and_lock(lock);
1703 ldlm_resource_unlink_lock(lock);
1705 * Remove the lock from pool as it will be added again in
1706 * ldlm_grant_lock() called below.
1708 ns = lock->l_resource->lr_namespace;
1709 ldlm_pool_del(&ns->ns_pool, lock);
1711 lock->l_req_mode = new_mode;
1712 ldlm_grant_lock(lock, NULL);
1713 unlock_res_and_lock(lock);
1714 ldlm_reprocess_all(lock->l_resource);
1719 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1722 CFS_LIST_HEAD(rpc_list);
1723 struct ldlm_resource *res;
1724 struct ldlm_namespace *ns;
1727 struct sl_insert_point prev;
1729 struct ldlm_interval *node;
1732 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1733 *flags |= LDLM_FL_BLOCK_GRANTED;
1734 RETURN(lock->l_resource);
1737 /* I can't check the type of lock here because the bitlock of lock
1738 * is not held here, so do the allocation blindly. -jay */
1739 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1740 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1743 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1744 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1746 lock_res_and_lock(lock);
1748 res = lock->l_resource;
1749 ns = res->lr_namespace;
1751 old_mode = lock->l_req_mode;
1752 lock->l_req_mode = new_mode;
1753 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1754 /* remember the lock position where the lock might be
1755 * added back to the granted list later and also
1756 * remember the join mode for skiplist fixing. */
1757 prev.res_link = lock->l_res_link.prev;
1758 prev.mode_link = lock->l_sl_mode.prev;
1759 prev.policy_link = lock->l_sl_policy.prev;
1760 ldlm_resource_unlink_lock(lock);
1762 ldlm_resource_unlink_lock(lock);
1763 if (res->lr_type == LDLM_EXTENT) {
1764 /* FIXME: ugly code, I have to attach the lock to a
1765 * interval node again since perhaps it will be granted
1767 CFS_INIT_LIST_HEAD(&node->li_group);
1768 ldlm_interval_attach(node, lock);
1774 * Remove old lock from the pool before adding the lock with new
1775 * mode below in ->policy()
1777 ldlm_pool_del(&ns->ns_pool, lock);
1779 /* If this is a local resource, put it on the appropriate list. */
1780 if (ns_is_client(res->lr_namespace)) {
1781 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1782 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1784 /* This should never happen, because of the way the
1785 * server handles conversions. */
1786 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1790 ldlm_grant_lock(lock, &rpc_list);
1792 /* FIXME: completion handling not with ns_lock held ! */
1793 if (lock->l_completion_ast)
1794 lock->l_completion_ast(lock, 0, NULL);
1798 ldlm_processing_policy policy;
1799 policy = ldlm_processing_policy_table[res->lr_type];
1800 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1801 if (rc == LDLM_ITER_STOP) {
1802 lock->l_req_mode = old_mode;
1803 if (res->lr_type == LDLM_EXTENT)
1804 ldlm_extent_add_lock(res, lock);
1806 ldlm_granted_list_add_lock(lock, &prev);
1810 *flags |= LDLM_FL_BLOCK_GRANTED;
1814 unlock_res_and_lock(lock);
1817 ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1819 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1823 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1825 struct obd_device *obd = NULL;
1827 if (!((libcfs_debug | D_ERROR) & level))
1831 CDEBUG(level, " NULL LDLM lock\n");
1835 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1836 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1838 if (lock->l_conn_export != NULL)
1839 obd = lock->l_conn_export->exp_obd;
1840 if (lock->l_export && lock->l_export->exp_connection) {
1841 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1842 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1843 lock->l_remote_handle.cookie);
1844 } else if (obd == NULL) {
1845 CDEBUG(level, " Node: local\n");
1847 struct obd_import *imp = obd->u.cli.cl_import;
1848 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1849 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1850 lock->l_remote_handle.cookie);
1852 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1854 lock->l_resource->lr_name.name[0],
1855 lock->l_resource->lr_name.name[1],
1856 lock->l_resource->lr_name.name[2]);
1857 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1858 "write: %d flags: "LPX64"\n", ldlm_lockname[lock->l_req_mode],
1859 ldlm_lockname[lock->l_granted_mode],
1860 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1862 if (lock->l_resource->lr_type == LDLM_EXTENT)
1863 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1864 " (req "LPU64"-"LPU64")\n",
1865 lock->l_policy_data.l_extent.start,
1866 lock->l_policy_data.l_extent.end,
1867 lock->l_req_extent.start, lock->l_req_extent.end);
1868 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1869 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1870 lock->l_policy_data.l_flock.pid,
1871 lock->l_policy_data.l_flock.start,
1872 lock->l_policy_data.l_flock.end);
1873 else if (lock->l_resource->lr_type == LDLM_IBITS)
1874 CDEBUG(level, " Bits: "LPX64"\n",
1875 lock->l_policy_data.l_inodebits.bits);
1878 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1880 struct ldlm_lock *lock;
1882 if (!((libcfs_debug | D_ERROR) & level))
1885 lock = ldlm_handle2lock(lockh);
1889 ldlm_lock_dump(D_OTHER, lock, 0);
1891 LDLM_LOCK_PUT(lock);
1894 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1895 struct libcfs_debug_msg_data *data, const char *fmt,
1899 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1901 va_start(args, fmt);
1903 if (lock->l_resource == NULL) {
1904 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1905 data->msg_fn, data->msg_line, fmt, args,
1906 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1907 "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" remote: "
1908 LPX64" expref: %d pid: %u timeout: %lu\n", lock,
1909 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1910 lock->l_readers, lock->l_writers,
1911 ldlm_lockname[lock->l_granted_mode],
1912 ldlm_lockname[lock->l_req_mode],
1913 lock->l_flags, lock->l_remote_handle.cookie,
1915 atomic_read(&lock->l_export->exp_refcount) : -99,
1916 lock->l_pid, lock->l_callback_timeout);
1921 switch (lock->l_resource->lr_type) {
1923 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1924 data->msg_fn, data->msg_line, fmt, args,
1925 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1926 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1927 "] (req "LPU64"->"LPU64") flags: "LPX64" remote: "LPX64
1928 " expref: %d pid: %u timeout %lu\n",
1929 lock->l_resource->lr_namespace->ns_name, lock,
1930 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1931 lock->l_readers, lock->l_writers,
1932 ldlm_lockname[lock->l_granted_mode],
1933 ldlm_lockname[lock->l_req_mode],
1934 lock->l_resource->lr_name.name[0],
1935 lock->l_resource->lr_name.name[1],
1936 atomic_read(&lock->l_resource->lr_refcount),
1937 ldlm_typename[lock->l_resource->lr_type],
1938 lock->l_policy_data.l_extent.start,
1939 lock->l_policy_data.l_extent.end,
1940 lock->l_req_extent.start, lock->l_req_extent.end,
1941 lock->l_flags, lock->l_remote_handle.cookie,
1943 atomic_read(&lock->l_export->exp_refcount) : -99,
1944 lock->l_pid, lock->l_callback_timeout);
1948 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1949 data->msg_fn, data->msg_line, fmt, args,
1950 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1951 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1952 "["LPU64"->"LPU64"] flags: "LPX64" remote: "LPX64
1953 " expref: %d pid: %u timeout: %lu\n",
1954 lock->l_resource->lr_namespace->ns_name, lock,
1955 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1956 lock->l_readers, lock->l_writers,
1957 ldlm_lockname[lock->l_granted_mode],
1958 ldlm_lockname[lock->l_req_mode],
1959 lock->l_resource->lr_name.name[0],
1960 lock->l_resource->lr_name.name[1],
1961 atomic_read(&lock->l_resource->lr_refcount),
1962 ldlm_typename[lock->l_resource->lr_type],
1963 lock->l_policy_data.l_flock.pid,
1964 lock->l_policy_data.l_flock.start,
1965 lock->l_policy_data.l_flock.end,
1966 lock->l_flags, lock->l_remote_handle.cookie,
1968 atomic_read(&lock->l_export->exp_refcount) : -99,
1969 lock->l_pid, lock->l_callback_timeout);
1973 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1974 data->msg_fn, data->msg_line, fmt, args,
1975 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1976 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1977 "flags: "LPX64" remote: "LPX64" expref: %d "
1978 "pid: %u timeout: %lu\n",
1979 lock->l_resource->lr_namespace->ns_name,
1980 lock, lock->l_handle.h_cookie,
1981 atomic_read (&lock->l_refc),
1982 lock->l_readers, lock->l_writers,
1983 ldlm_lockname[lock->l_granted_mode],
1984 ldlm_lockname[lock->l_req_mode],
1985 lock->l_resource->lr_name.name[0],
1986 lock->l_resource->lr_name.name[1],
1987 lock->l_policy_data.l_inodebits.bits,
1988 atomic_read(&lock->l_resource->lr_refcount),
1989 ldlm_typename[lock->l_resource->lr_type],
1990 lock->l_flags, lock->l_remote_handle.cookie,
1992 atomic_read(&lock->l_export->exp_refcount) : -99,
1993 lock->l_pid, lock->l_callback_timeout);
1997 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1998 data->msg_fn, data->msg_line, fmt, args,
1999 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2000 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" "
2001 "remote: "LPX64" expref: %d pid: %u timeout %lu\n",
2002 lock->l_resource->lr_namespace->ns_name,
2003 lock, lock->l_handle.h_cookie,
2004 atomic_read (&lock->l_refc),
2005 lock->l_readers, lock->l_writers,
2006 ldlm_lockname[lock->l_granted_mode],
2007 ldlm_lockname[lock->l_req_mode],
2008 lock->l_resource->lr_name.name[0],
2009 lock->l_resource->lr_name.name[1],
2010 atomic_read(&lock->l_resource->lr_refcount),
2011 ldlm_typename[lock->l_resource->lr_type],
2012 lock->l_flags, lock->l_remote_handle.cookie,
2014 atomic_read(&lock->l_export->exp_refcount) : -99,
2015 lock->l_pid, lock->l_callback_timeout);
2020 EXPORT_SYMBOL(_ldlm_lock_debug);