1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_lock.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_LDLM
45 # include <libcfs/libcfs.h>
46 # ifndef HAVE_VFS_INTENT_PATCHES
47 # include <linux/lustre_intent.h>
50 # include <liblustre.h>
53 #include <obd_class.h>
54 #include "ldlm_internal.h"
57 char *ldlm_lockname[] = {
69 char *ldlm_typename[] = {
76 char *ldlm_it2str(int it)
83 case (IT_OPEN | IT_CREAT):
96 CERROR("Unknown intent %d\n", it);
101 extern cfs_mem_cache_t *ldlm_lock_slab;
103 static ldlm_processing_policy ldlm_processing_policy_table[] = {
104 [LDLM_PLAIN] ldlm_process_plain_lock,
105 [LDLM_EXTENT] ldlm_process_extent_lock,
107 [LDLM_FLOCK] ldlm_process_flock_lock,
109 [LDLM_IBITS] ldlm_process_inodebits_lock,
112 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
114 return ldlm_processing_policy_table[res->lr_type];
117 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
123 * REFCOUNTED LOCK OBJECTS
128 * Lock refcounts, during creation:
129 * - one special one for allocation, dec'd only once in destroy
130 * - one for being a lock that's in-use
131 * - one for the addref associated with a new lock
133 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
135 atomic_inc(&lock->l_refc);
139 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
141 LASSERT(size == sizeof(*lock));
142 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
145 void ldlm_lock_put(struct ldlm_lock *lock)
149 LASSERT(lock->l_resource != LP_POISON);
150 LASSERT(atomic_read(&lock->l_refc) > 0);
151 if (atomic_dec_and_test(&lock->l_refc)) {
152 struct ldlm_resource *res;
155 "final lock_put on destroyed lock, freeing it.");
157 res = lock->l_resource;
158 LASSERT(lock->l_destroyed);
159 LASSERT(list_empty(&lock->l_res_link));
160 LASSERT(list_empty(&lock->l_pending_chain));
162 atomic_dec(&res->lr_namespace->ns_locks);
163 lu_ref_del(&res->lr_reference, "lock", lock);
164 ldlm_resource_putref(res);
165 lock->l_resource = NULL;
166 if (lock->l_export) {
167 class_export_put(lock->l_export);
168 lock->l_export = NULL;
171 if (lock->l_lvb_data != NULL)
172 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
174 ldlm_interval_free(ldlm_interval_detach(lock));
175 lu_ref_fini(&lock->l_reference);
176 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
183 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
186 if (!list_empty(&lock->l_lru)) {
187 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
188 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
189 list_del_init(&lock->l_lru);
190 LASSERT(ns->ns_nr_unused > 0);
197 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
199 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
202 spin_lock(&ns->ns_unused_lock);
203 rc = ldlm_lock_remove_from_lru_nolock(lock);
204 spin_unlock(&ns->ns_unused_lock);
209 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
211 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
212 lock->l_last_used = cfs_time_current();
213 LASSERT(list_empty(&lock->l_lru));
214 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
215 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
216 LASSERT(ns->ns_nr_unused >= 0);
220 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
222 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
224 spin_lock(&ns->ns_unused_lock);
225 ldlm_lock_add_to_lru_nolock(lock);
226 spin_unlock(&ns->ns_unused_lock);
230 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
232 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
234 spin_lock(&ns->ns_unused_lock);
235 if (!list_empty(&lock->l_lru)) {
236 ldlm_lock_remove_from_lru_nolock(lock);
237 ldlm_lock_add_to_lru_nolock(lock);
239 spin_unlock(&ns->ns_unused_lock);
243 /* This used to have a 'strict' flag, which recovery would use to mark an
244 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
245 * shall explain why it's gone: with the new hash table scheme, once you call
246 * ldlm_lock_destroy, you can never drop your final references on this lock.
247 * Because it's not in the hash table anymore. -phil */
248 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
252 if (lock->l_readers || lock->l_writers) {
253 LDLM_ERROR(lock, "lock still has references");
254 ldlm_lock_dump(D_ERROR, lock, 0);
258 if (!list_empty(&lock->l_res_link)) {
259 LDLM_ERROR(lock, "lock still on resource");
260 ldlm_lock_dump(D_ERROR, lock, 0);
264 if (lock->l_destroyed) {
265 LASSERT(list_empty(&lock->l_lru));
269 lock->l_destroyed = 1;
271 if (lock->l_export && lock->l_export->exp_lock_hash &&
272 !hlist_unhashed(&lock->l_exp_hash))
273 lustre_hash_del(lock->l_export->exp_lock_hash,
274 &lock->l_remote_handle, &lock->l_exp_hash);
276 ldlm_lock_remove_from_lru(lock);
277 class_handle_unhash(&lock->l_handle);
280 /* Wake anyone waiting for this lock */
281 /* FIXME: I should probably add yet another flag, instead of using
282 * l_export to only call this on clients */
284 class_export_put(lock->l_export);
285 lock->l_export = NULL;
286 if (lock->l_export && lock->l_completion_ast)
287 lock->l_completion_ast(lock, 0);
293 void ldlm_lock_destroy(struct ldlm_lock *lock)
297 lock_res_and_lock(lock);
298 first = ldlm_lock_destroy_internal(lock);
299 unlock_res_and_lock(lock);
301 /* drop reference from hashtable only for first destroy */
303 lu_ref_del(&lock->l_reference, "hash", lock);
304 LDLM_LOCK_RELEASE(lock);
309 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
313 first = ldlm_lock_destroy_internal(lock);
314 /* drop reference from hashtable only for first destroy */
316 lu_ref_del(&lock->l_reference, "hash", lock);
317 LDLM_LOCK_RELEASE(lock);
322 /* this is called by portals_handle2object with the handle lock taken */
323 static void lock_handle_addref(void *lock)
325 LDLM_LOCK_GET((struct ldlm_lock *)lock);
329 * usage: pass in a resource on which you have done ldlm_resource_get
330 * pass in a parent lock on which you have done a ldlm_lock_get
331 * after return, ldlm_*_put the resource and parent
332 * returns: lock with refcount 2 - one for current caller and one for remote
334 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
336 struct ldlm_lock *lock;
339 if (resource == NULL)
342 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, CFS_ALLOC_IO);
346 spin_lock_init(&lock->l_lock);
347 lock->l_resource = ldlm_resource_getref(resource);
348 lu_ref_add(&resource->lr_reference, "lock", lock);
350 atomic_set(&lock->l_refc, 2);
351 CFS_INIT_LIST_HEAD(&lock->l_res_link);
352 CFS_INIT_LIST_HEAD(&lock->l_lru);
353 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
354 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
355 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
356 CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
357 cfs_waitq_init(&lock->l_waitq);
358 lock->l_blocking_lock = NULL;
359 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
360 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
361 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
363 atomic_inc(&resource->lr_namespace->ns_locks);
364 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
365 class_handle_hash(&lock->l_handle, lock_handle_addref);
367 CFS_INIT_LIST_HEAD(&lock->l_extents_list);
368 spin_lock_init(&lock->l_extents_list_lock);
369 CFS_INIT_LIST_HEAD(&lock->l_cache_locks_list);
370 lu_ref_init(&lock->l_reference);
371 lu_ref_add(&lock->l_reference, "hash", lock);
372 lock->l_callback_timeout = 0;
377 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
378 const struct ldlm_res_id *new_resid)
380 struct ldlm_resource *oldres = lock->l_resource;
381 struct ldlm_resource *newres;
385 LASSERT(ns_is_client(ns));
387 lock_res_and_lock(lock);
388 if (memcmp(new_resid, &lock->l_resource->lr_name,
389 sizeof(lock->l_resource->lr_name)) == 0) {
391 unlock_res_and_lock(lock);
395 LASSERT(new_resid->name[0] != 0);
397 /* This function assumes that the lock isn't on any lists */
398 LASSERT(list_empty(&lock->l_res_link));
400 type = oldres->lr_type;
401 unlock_res_and_lock(lock);
403 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
404 lu_ref_add(&newres->lr_reference, "lock", lock);
408 * To flip the lock from the old to the new resource, lock, oldres and
409 * newres have to be locked. Resource spin-locks are nested within
410 * lock->l_lock, and are taken in the memory address order to avoid
413 spin_lock(&lock->l_lock);
414 oldres = lock->l_resource;
415 if (oldres < newres) {
417 lock_res_nested(newres, LRT_NEW);
420 lock_res_nested(oldres, LRT_NEW);
422 LASSERT(memcmp(new_resid, &oldres->lr_name,
423 sizeof oldres->lr_name) != 0);
424 lock->l_resource = newres;
426 unlock_res_and_lock(lock);
428 /* ...and the flowers are still standing! */
429 lu_ref_del(&oldres->lr_reference, "lock", lock);
430 ldlm_resource_putref(oldres);
439 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
441 lockh->cookie = lock->l_handle.h_cookie;
444 /* if flags: atomically get the lock and set the flags.
445 * Return NULL if flag already set
448 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
451 struct ldlm_namespace *ns;
452 struct ldlm_lock *lock, *retval = NULL;
457 lock = class_handle2object(handle->cookie);
461 LASSERT(lock->l_resource != NULL);
462 ns = lock->l_resource->lr_namespace;
465 lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
466 lock_res_and_lock(lock);
468 /* It's unlikely but possible that someone marked the lock as
469 * destroyed after we did handle2object on it */
470 if (lock->l_destroyed) {
471 unlock_res_and_lock(lock);
472 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
477 if (flags && (lock->l_flags & flags)) {
478 unlock_res_and_lock(lock);
484 lock->l_flags |= flags;
486 unlock_res_and_lock(lock);
493 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
495 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
496 /* INODEBITS_INTEROP: If the other side does not support
497 * inodebits, reply with a plain lock descriptor.
499 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
500 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
501 struct ldlm_resource res = *lock->l_resource;
503 /* Make sure all the right bits are set in this lock we
504 are going to pass to client */
505 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
506 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
507 "Inappropriate inode lock bits during "
508 "conversion " LPU64 "\n",
509 lock->l_policy_data.l_inodebits.bits);
510 res.lr_type = LDLM_PLAIN;
511 ldlm_res2desc(&res, &desc->l_resource);
512 /* Convert "new" lock mode to something old client can
514 if ((lock->l_req_mode == LCK_CR) ||
515 (lock->l_req_mode == LCK_CW))
516 desc->l_req_mode = LCK_PR;
518 desc->l_req_mode = lock->l_req_mode;
519 if ((lock->l_granted_mode == LCK_CR) ||
520 (lock->l_granted_mode == LCK_CW)) {
521 desc->l_granted_mode = LCK_PR;
523 /* We never grant PW/EX locks to clients */
524 LASSERT((lock->l_granted_mode != LCK_PW) &&
525 (lock->l_granted_mode != LCK_EX));
526 desc->l_granted_mode = lock->l_granted_mode;
529 /* We do not copy policy here, because there is no
530 policy for plain locks */
532 ldlm_res2desc(lock->l_resource, &desc->l_resource);
533 desc->l_req_mode = lock->l_req_mode;
534 desc->l_granted_mode = lock->l_granted_mode;
535 desc->l_policy_data = lock->l_policy_data;
539 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
540 struct list_head *work_list)
542 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
543 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
544 lock->l_flags |= LDLM_FL_AST_SENT;
545 /* If the enqueuing client said so, tell the AST recipient to
546 * discard dirty data, rather than writing back. */
547 if (new->l_flags & LDLM_AST_DISCARD_DATA)
548 lock->l_flags |= LDLM_FL_DISCARD_DATA;
549 LASSERT(list_empty(&lock->l_bl_ast));
550 list_add(&lock->l_bl_ast, work_list);
552 LASSERT(lock->l_blocking_lock == NULL);
553 lock->l_blocking_lock = LDLM_LOCK_GET(new);
557 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
559 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
560 lock->l_flags |= LDLM_FL_CP_REQD;
561 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
562 LASSERT(list_empty(&lock->l_cp_ast));
563 list_add(&lock->l_cp_ast, work_list);
568 /* must be called with lr_lock held */
569 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
570 struct list_head *work_list)
573 check_res_locked(lock->l_resource);
575 ldlm_add_bl_work_item(lock, new, work_list);
577 ldlm_add_cp_work_item(lock, work_list);
581 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
583 struct ldlm_lock *lock;
585 lock = ldlm_handle2lock(lockh);
586 LASSERT(lock != NULL);
587 ldlm_lock_addref_internal(lock, mode);
591 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
593 ldlm_lock_remove_from_lru(lock);
594 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
596 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
598 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
600 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
603 lu_ref_add_atomic(&lock->l_reference, "user", lock);
604 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
608 * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
611 * \retval 0 success, lock was addref-ed
613 * \retval -EAGAIN lock is being canceled.
615 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
617 struct ldlm_lock *lock;
621 lock = ldlm_handle2lock(lockh);
623 lock_res_and_lock(lock);
624 if (lock->l_readers != 0 || lock->l_writers != 0 ||
625 !(lock->l_flags & LDLM_FL_CBPENDING)) {
626 ldlm_lock_addref_internal_nolock(lock, mode);
629 unlock_res_and_lock(lock);
635 /* only called for local locks */
636 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
638 lock_res_and_lock(lock);
639 ldlm_lock_addref_internal_nolock(lock, mode);
640 unlock_res_and_lock(lock);
643 /* only called in ldlm_flock_destroy and for local locks.
644 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
645 * * ldlm_lock_remove_from_lru() does nothing, it is safe
646 * * for ldlm_flock_destroy usage by dropping some code */
647 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
649 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
650 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
651 LASSERT(lock->l_readers > 0);
652 lu_ref_del(&lock->l_reference, "reader", lock);
655 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
656 LASSERT(lock->l_writers > 0);
657 lu_ref_del(&lock->l_reference, "writer", lock);
661 lu_ref_del(&lock->l_reference, "user", lock);
662 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
665 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
667 struct ldlm_namespace *ns;
670 lock_res_and_lock(lock);
672 ns = lock->l_resource->lr_namespace;
674 ldlm_lock_decref_internal_nolock(lock, mode);
676 if (lock->l_flags & LDLM_FL_LOCAL &&
677 !lock->l_readers && !lock->l_writers) {
678 /* If this is a local lock on a server namespace and this was
679 * the last reference, cancel the lock. */
680 CDEBUG(D_INFO, "forcing cancel of local lock\n");
681 lock->l_flags |= LDLM_FL_CBPENDING;
684 if (!lock->l_readers && !lock->l_writers &&
685 (lock->l_flags & LDLM_FL_CBPENDING)) {
686 /* If we received a blocked AST and this was the last reference,
687 * run the callback. */
688 if (ns_is_server(ns) && lock->l_export)
689 CERROR("FL_CBPENDING set on non-local lock--just a "
692 LDLM_DEBUG(lock, "final decref done on cbpending lock");
694 LDLM_LOCK_GET(lock); /* dropped by bl thread */
695 ldlm_lock_remove_from_lru(lock);
696 unlock_res_and_lock(lock);
698 if (lock->l_flags & LDLM_FL_FAIL_LOC)
699 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
701 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
702 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
703 ldlm_handle_bl_callback(ns, NULL, lock);
704 } else if (ns_is_client(ns) &&
705 !lock->l_readers && !lock->l_writers &&
706 !(lock->l_flags & LDLM_FL_BL_AST)) {
707 /* If this is a client-side namespace and this was the last
708 * reference, put it on the LRU. */
709 ldlm_lock_add_to_lru(lock);
710 unlock_res_and_lock(lock);
712 if (lock->l_flags & LDLM_FL_FAIL_LOC)
713 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
715 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
716 * are not supported by the server, otherwise, it is done on
718 if (!exp_connect_cancelset(lock->l_conn_export) &&
719 !ns_connect_lru_resize(ns))
720 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
722 unlock_res_and_lock(lock);
728 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
730 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
731 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
732 ldlm_lock_decref_internal(lock, mode);
736 /* This will drop a lock reference and mark it for destruction, but will not
737 * necessarily cancel the lock before returning. */
738 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
740 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
743 LASSERT(lock != NULL);
745 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
746 lock_res_and_lock(lock);
747 lock->l_flags |= LDLM_FL_CBPENDING;
748 unlock_res_and_lock(lock);
749 ldlm_lock_decref_internal(lock, mode);
753 struct sl_insert_point {
754 struct list_head *res_link;
755 struct list_head *mode_link;
756 struct list_head *policy_link;
760 * search_granted_lock
763 * Finds a position to insert the new lock.
765 * queue [input]: the granted list where search acts on;
766 * req [input]: the lock whose position to be located;
767 * prev [output]: positions within 3 lists to insert @req to
771 * - ldlm_grant_lock_with_skiplist
773 static void search_granted_lock(struct list_head *queue,
774 struct ldlm_lock *req,
775 struct sl_insert_point *prev)
777 struct list_head *tmp;
778 struct ldlm_lock *lock, *mode_end, *policy_end;
781 list_for_each(tmp, queue) {
782 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
784 mode_end = list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
787 if (lock->l_req_mode != req->l_req_mode) {
788 /* jump to last lock of mode group */
789 tmp = &mode_end->l_res_link;
793 /* suitable mode group is found */
794 if (lock->l_resource->lr_type == LDLM_PLAIN) {
795 /* insert point is last lock of the mode group */
796 prev->res_link = &mode_end->l_res_link;
797 prev->mode_link = &mode_end->l_sl_mode;
798 prev->policy_link = &req->l_sl_policy;
801 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
803 policy_end = list_entry(lock->l_sl_policy.prev,
807 if (lock->l_policy_data.l_inodebits.bits ==
808 req->l_policy_data.l_inodebits.bits) {
809 /* insert point is last lock of
810 * the policy group */
812 &policy_end->l_res_link;
814 &policy_end->l_sl_mode;
816 &policy_end->l_sl_policy;
821 if (policy_end == mode_end)
822 /* done with mode group */
825 /* go to next policy group within mode group */
826 tmp = policy_end->l_res_link.next;
827 lock = list_entry(tmp, struct ldlm_lock,
829 } /* loop over policy groups within the mode group */
831 /* insert point is last lock of the mode group,
832 * new policy group is started */
833 prev->res_link = &mode_end->l_res_link;
834 prev->mode_link = &mode_end->l_sl_mode;
835 prev->policy_link = &req->l_sl_policy;
839 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
844 /* insert point is last lock on the queue,
845 * new mode group and new policy group are started */
846 prev->res_link = queue->prev;
847 prev->mode_link = &req->l_sl_mode;
848 prev->policy_link = &req->l_sl_policy;
853 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
854 struct sl_insert_point *prev)
856 struct ldlm_resource *res = lock->l_resource;
859 check_res_locked(res);
861 ldlm_resource_dump(D_INFO, res);
862 CDEBUG(D_OTHER, "About to add this lock:\n");
863 ldlm_lock_dump(D_OTHER, lock, 0);
865 if (lock->l_destroyed) {
866 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
870 LASSERT(list_empty(&lock->l_res_link));
871 LASSERT(list_empty(&lock->l_sl_mode));
872 LASSERT(list_empty(&lock->l_sl_policy));
874 list_add(&lock->l_res_link, prev->res_link);
875 list_add(&lock->l_sl_mode, prev->mode_link);
876 list_add(&lock->l_sl_policy, prev->policy_link);
881 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
883 struct sl_insert_point prev;
886 LASSERT(lock->l_req_mode == lock->l_granted_mode);
888 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
889 ldlm_granted_list_add_lock(lock, &prev);
894 * - ldlm_lock_enqueue
895 * - ldlm_reprocess_queue
896 * - ldlm_lock_convert
898 * must be called with lr_lock held
900 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
902 struct ldlm_resource *res = lock->l_resource;
905 check_res_locked(res);
907 lock->l_granted_mode = lock->l_req_mode;
908 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
909 ldlm_grant_lock_with_skiplist(lock);
910 else if (res->lr_type == LDLM_EXTENT)
911 ldlm_extent_add_lock(res, lock);
913 ldlm_resource_add_lock(res, &res->lr_granted, lock);
915 if (lock->l_granted_mode < res->lr_most_restr)
916 res->lr_most_restr = lock->l_granted_mode;
918 if (work_list && lock->l_completion_ast != NULL)
919 ldlm_add_ast_work_item(lock, NULL, work_list);
921 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
925 /* returns a referenced lock or NULL. See the flag descriptions below, in the
926 * comment above ldlm_lock_match */
927 static struct ldlm_lock *search_queue(struct list_head *queue,
929 ldlm_policy_data_t *policy,
930 struct ldlm_lock *old_lock,
931 int flags, int unref)
933 struct ldlm_lock *lock;
934 struct list_head *tmp;
936 list_for_each(tmp, queue) {
939 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
941 if (lock == old_lock)
944 /* llite sometimes wants to match locks that will be
945 * canceled when their users drop, but we allow it to match
946 * if it passes in CBPENDING and the lock still has users.
947 * this is generally only going to be used by children
948 * whose parents already hold a lock so forward progress
949 * can still happen. */
950 if (lock->l_flags & LDLM_FL_CBPENDING &&
951 !(flags & LDLM_FL_CBPENDING))
953 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
954 lock->l_readers == 0 && lock->l_writers == 0)
957 if (!(lock->l_req_mode & *mode))
959 match = lock->l_req_mode;
961 if (lock->l_resource->lr_type == LDLM_EXTENT &&
962 (lock->l_policy_data.l_extent.start >
963 policy->l_extent.start ||
964 lock->l_policy_data.l_extent.end < policy->l_extent.end))
967 if (unlikely(match == LCK_GROUP) &&
968 lock->l_resource->lr_type == LDLM_EXTENT &&
969 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
972 /* We match if we have existing lock with same or wider set
974 if (lock->l_resource->lr_type == LDLM_IBITS &&
975 ((lock->l_policy_data.l_inodebits.bits &
976 policy->l_inodebits.bits) !=
977 policy->l_inodebits.bits))
981 (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED)))
984 if ((flags & LDLM_FL_LOCAL_ONLY) &&
985 !(lock->l_flags & LDLM_FL_LOCAL))
988 if (flags & LDLM_FL_TEST_LOCK) {
990 ldlm_lock_touch_in_lru(lock);
992 ldlm_lock_addref_internal_nolock(lock, match);
1001 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1003 lock->l_flags |= LDLM_FL_LVB_READY;
1004 cfs_waitq_signal(&lock->l_waitq);
1007 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1009 lock_res_and_lock(lock);
1010 ldlm_lock_allow_match_locked(lock);
1011 unlock_res_and_lock(lock);
1014 /* Can be called in two ways:
1016 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1017 * for a duplicate of.
1019 * Otherwise, all of the fields must be filled in, to match against.
1021 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1022 * server (ie, connh is NULL)
1023 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1024 * list will be considered
1025 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1026 * to be canceled can still be matched as long as they still have reader
1027 * or writer refernces
1028 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1029 * just tell us if we would have matched.
1031 * Returns 1 if it finds an already-existing lock that is compatible; in this
1032 * case, lockh is filled in with a addref()ed lock
1034 * we also check security context, if that failed we simply return 0 (to keep
1035 * caller code unchanged), the context failure will be discovered by caller
1038 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1039 const struct ldlm_res_id *res_id, ldlm_type_t type,
1040 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1041 struct lustre_handle *lockh, int unref)
1043 struct ldlm_resource *res;
1044 struct ldlm_lock *lock, *old_lock = NULL;
1049 old_lock = ldlm_handle2lock(lockh);
1052 ns = old_lock->l_resource->lr_namespace;
1053 res_id = &old_lock->l_resource->lr_name;
1054 type = old_lock->l_resource->lr_type;
1055 mode = old_lock->l_req_mode;
1058 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1060 LASSERT(old_lock == NULL);
1064 LDLM_RESOURCE_ADDREF(res);
1067 lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1071 if (flags & LDLM_FL_BLOCK_GRANTED)
1073 lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1077 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1085 LDLM_RESOURCE_DELREF(res);
1086 ldlm_resource_putref(res);
1089 ldlm_lock2handle(lock, lockh);
1090 if ((flags & LDLM_FL_LVB_READY) &&
1091 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1092 struct l_wait_info lwi;
1093 if (lock->l_completion_ast) {
1094 int err = lock->l_completion_ast(lock,
1095 LDLM_FL_WAIT_NOREPROC,
1098 if (flags & LDLM_FL_TEST_LOCK)
1099 LDLM_LOCK_RELEASE(lock);
1101 ldlm_lock_decref_internal(lock,
1108 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1109 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1111 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1112 l_wait_event(lock->l_waitq,
1113 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1118 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1119 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1120 res_id->name[2] : policy->l_extent.start,
1121 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1122 res_id->name[3] : policy->l_extent.end);
1124 /* check user's security context */
1125 if (lock->l_conn_export &&
1126 sptlrpc_import_check_ctx(
1127 class_exp2cliimp(lock->l_conn_export))) {
1128 if (!(flags & LDLM_FL_TEST_LOCK))
1129 ldlm_lock_decref_internal(lock, mode);
1133 if (flags & LDLM_FL_TEST_LOCK)
1134 LDLM_LOCK_RELEASE(lock);
1136 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1137 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1138 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1139 type, mode, res_id->name[0], res_id->name[1],
1140 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1141 res_id->name[2] :policy->l_extent.start,
1142 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1143 res_id->name[3] : policy->l_extent.end);
1146 LDLM_LOCK_PUT(old_lock);
1148 return rc ? mode : 0;
1151 /* Returns a referenced lock */
1152 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1153 const struct ldlm_res_id *res_id,
1156 const struct ldlm_callback_suite *cbs,
1157 void *data, __u32 lvb_len)
1159 struct ldlm_lock *lock;
1160 struct ldlm_resource *res;
1163 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1167 lock = ldlm_lock_new(res);
1168 ldlm_resource_putref(res);
1173 lock->l_req_mode = mode;
1174 lock->l_ast_data = data;
1175 lock->l_pid = cfs_curproc_pid();
1177 lock->l_blocking_ast = cbs->lcs_blocking;
1178 lock->l_completion_ast = cbs->lcs_completion;
1179 lock->l_glimpse_ast = cbs->lcs_glimpse;
1180 lock->l_weigh_ast = cbs->lcs_weigh;
1183 lock->l_tree_node = NULL;
1184 /* if this is the extent lock, allocate the interval tree node */
1185 if (type == LDLM_EXTENT) {
1186 if (ldlm_interval_alloc(lock) == NULL)
1191 lock->l_lvb_len = lvb_len;
1192 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1193 if (lock->l_lvb_data == NULL)
1200 if (lock->l_lvb_data)
1201 OBD_FREE(lock->l_lvb_data, lvb_len);
1202 ldlm_interval_free(ldlm_interval_detach(lock));
1203 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1207 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1208 struct ldlm_lock **lockp,
1209 void *cookie, int *flags)
1211 struct ldlm_lock *lock = *lockp;
1212 struct ldlm_resource *res = lock->l_resource;
1213 int local = ns_is_client(res->lr_namespace);
1214 ldlm_processing_policy policy;
1215 ldlm_error_t rc = ELDLM_OK;
1216 struct ldlm_interval *node = NULL;
1219 lock->l_last_activity = cfs_time_current_sec();
1220 /* policies are not executed on the client or during replay */
1221 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1222 && !local && ns->ns_policy) {
1223 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1225 if (rc == ELDLM_LOCK_REPLACED) {
1226 /* The lock that was returned has already been granted,
1227 * and placed into lockp. If it's not the same as the
1228 * one we passed in, then destroy the old one and our
1229 * work here is done. */
1230 if (lock != *lockp) {
1231 ldlm_lock_destroy(lock);
1232 LDLM_LOCK_RELEASE(lock);
1234 *flags |= LDLM_FL_LOCK_CHANGED;
1236 } else if (rc != ELDLM_OK ||
1237 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1238 ldlm_lock_destroy(lock);
1243 /* For a replaying lock, it might be already in granted list. So
1244 * unlinking the lock will cause the interval node to be freed, we
1245 * have to allocate the interval node early otherwise we can't regrant
1246 * this lock in the future. - jay */
1247 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1248 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1250 lock_res_and_lock(lock);
1251 if (local && lock->l_req_mode == lock->l_granted_mode) {
1252 /* The server returned a blocked lock, but it was granted
1253 * before we got a chance to actually enqueue it. We don't
1254 * need to do anything else. */
1255 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1256 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1257 GOTO(out, ELDLM_OK);
1260 ldlm_resource_unlink_lock(lock);
1261 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1263 ldlm_lock_destroy_nolock(lock);
1264 GOTO(out, rc = -ENOMEM);
1267 CFS_INIT_LIST_HEAD(&node->li_group);
1268 ldlm_interval_attach(node, lock);
1272 /* Some flags from the enqueue want to make it into the AST, via the
1273 * lock's l_flags. */
1274 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1276 /* This distinction between local lock trees is very important; a client
1277 * namespace only has information about locks taken by that client, and
1278 * thus doesn't have enough information to decide for itself if it can
1279 * be granted (below). In this case, we do exactly what the server
1280 * tells us to do, as dictated by the 'flags'.
1282 * We do exactly the same thing during recovery, when the server is
1283 * more or less trusting the clients not to lie.
1285 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1286 * granted/converting queues. */
1288 if (*flags & LDLM_FL_BLOCK_CONV)
1289 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1290 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1291 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1293 ldlm_grant_lock(lock, NULL);
1294 GOTO(out, ELDLM_OK);
1295 } else if (*flags & LDLM_FL_REPLAY) {
1296 if (*flags & LDLM_FL_BLOCK_CONV) {
1297 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1298 GOTO(out, ELDLM_OK);
1299 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1300 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1301 GOTO(out, ELDLM_OK);
1302 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1303 ldlm_grant_lock(lock, NULL);
1304 GOTO(out, ELDLM_OK);
1306 /* If no flags, fall through to normal enqueue path. */
1309 policy = ldlm_processing_policy_table[res->lr_type];
1310 policy(lock, flags, 1, &rc, NULL);
1313 unlock_res_and_lock(lock);
1315 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1319 /* Must be called with namespace taken: queue is waiting or converting. */
1320 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1321 struct list_head *work_list)
1323 struct list_head *tmp, *pos;
1324 ldlm_processing_policy policy;
1326 int rc = LDLM_ITER_CONTINUE;
1330 check_res_locked(res);
1332 policy = ldlm_processing_policy_table[res->lr_type];
1335 list_for_each_safe(tmp, pos, queue) {
1336 struct ldlm_lock *pending;
1337 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1339 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1342 rc = policy(pending, &flags, 0, &err, work_list);
1343 if (rc != LDLM_ITER_CONTINUE)
1350 /* Helper function for ldlm_run_ast_work().
1352 * Send an existing rpc set specified by @arg->set and then
1353 * destroy it. Create new one if @do_create flag is set. */
1355 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1359 ptlrpc_set_wait(arg->set);
1360 if (arg->type == LDLM_BL_CALLBACK)
1361 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1362 ptlrpc_set_destroy(arg->set);
1365 arg->set = ptlrpc_prep_set();
1371 ldlm_work_bl_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1373 struct ldlm_lock_desc d;
1374 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_bl_ast);
1377 /* nobody should touch l_bl_ast */
1378 lock_res_and_lock(lock);
1379 list_del_init(&lock->l_bl_ast);
1381 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1382 LASSERT(lock->l_bl_ast_run == 0);
1383 LASSERT(lock->l_blocking_lock);
1384 lock->l_bl_ast_run++;
1385 unlock_res_and_lock(lock);
1387 ldlm_lock2desc(lock->l_blocking_lock, &d);
1389 lock->l_blocking_ast(lock, &d, (void *)arg,
1391 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1392 lock->l_blocking_lock = NULL;
1393 LDLM_LOCK_RELEASE(lock);
1399 ldlm_work_cp_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1401 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_cp_ast);
1402 ldlm_completion_callback completion_callback;
1406 /* It's possible to receive a completion AST before we've set
1407 * the l_completion_ast pointer: either because the AST arrived
1408 * before the reply, or simply because there's a small race
1409 * window between receiving the reply and finishing the local
1410 * enqueue. (bug 842)
1412 * This can't happen with the blocking_ast, however, because we
1413 * will never call the local blocking_ast until we drop our
1414 * reader/writer reference, which we won't do until we get the
1415 * reply and finish enqueueing. */
1417 /* nobody should touch l_cp_ast */
1418 lock_res_and_lock(lock);
1419 list_del_init(&lock->l_cp_ast);
1420 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1421 /* save l_completion_ast since it can be changed by
1422 * mds_intent_policy(), see bug 14225 */
1423 completion_callback = lock->l_completion_ast;
1424 lock->l_flags &= ~LDLM_FL_CP_REQD;
1425 unlock_res_and_lock(lock);
1427 if (completion_callback != NULL) {
1428 completion_callback(lock, 0, (void *)arg);
1431 LDLM_LOCK_RELEASE(lock);
1437 ldlm_work_revoke_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1439 struct ldlm_lock_desc desc;
1440 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_rk_ast);
1443 list_del_init(&lock->l_rk_ast);
1445 /* the desc just pretend to exclusive */
1446 ldlm_lock2desc(lock, &desc);
1447 desc.l_req_mode = LCK_EX;
1448 desc.l_granted_mode = 0;
1450 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1451 LDLM_LOCK_RELEASE(lock);
1456 int ldlm_run_ast_work(struct list_head *rpc_list, ldlm_desc_ast_t ast_type)
1458 struct ldlm_cb_set_arg arg;
1459 struct list_head *tmp, *pos;
1460 int (*work_ast_lock)(struct list_head *tmp,struct ldlm_cb_set_arg *arg);
1464 arg.set = ptlrpc_prep_set();
1465 atomic_set(&arg.restart, 0);
1467 case LDLM_WORK_BL_AST:
1468 arg.type = LDLM_BL_CALLBACK;
1469 work_ast_lock = ldlm_work_bl_ast_lock;
1471 case LDLM_WORK_CP_AST:
1472 arg.type = LDLM_CP_CALLBACK;
1473 work_ast_lock = ldlm_work_cp_ast_lock;
1475 case LDLM_WORK_REVOKE_AST:
1476 arg.type = LDLM_BL_CALLBACK;
1477 work_ast_lock = ldlm_work_revoke_ast_lock;
1484 list_for_each_safe(tmp, pos, rpc_list) {
1485 ast_count += work_ast_lock(tmp, &arg);
1487 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1488 * and create a new set for requests that remained in
1490 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1491 ldlm_send_and_maybe_create_set(&arg, 1);
1497 ldlm_send_and_maybe_create_set(&arg, 0);
1499 /* In case when number of ASTs is multiply of
1500 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1501 * @arg.set must be destroyed here, otherwise we get
1502 * write memory leaking. */
1503 ptlrpc_set_destroy(arg.set);
1505 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1508 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1510 ldlm_reprocess_all(res);
1511 return LDLM_ITER_CONTINUE;
1514 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1516 struct list_head *tmp;
1523 spin_lock(&ns->ns_hash_lock);
1524 for (i = 0; i < RES_HASH_SIZE; i++) {
1525 tmp = ns->ns_hash[i].next;
1526 while (tmp != &(ns->ns_hash[i])) {
1527 struct ldlm_resource *res =
1528 list_entry(tmp, struct ldlm_resource, lr_hash);
1530 ldlm_resource_getref(res);
1531 spin_unlock(&ns->ns_hash_lock);
1532 LDLM_RESOURCE_ADDREF(res);
1534 rc = reprocess_one_queue(res, NULL);
1536 LDLM_RESOURCE_DELREF(res);
1537 spin_lock(&ns->ns_hash_lock);
1539 ldlm_resource_putref_locked(res);
1541 if (rc == LDLM_ITER_STOP)
1546 spin_unlock(&ns->ns_hash_lock);
1550 void ldlm_reprocess_all(struct ldlm_resource *res)
1552 CFS_LIST_HEAD(rpc_list);
1556 /* Local lock trees don't get reprocessed. */
1557 if (ns_is_client(res->lr_namespace)) {
1564 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1565 if (rc == LDLM_ITER_CONTINUE)
1566 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1569 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1570 if (rc == -ERESTART) {
1571 LASSERT(list_empty(&rpc_list));
1577 void ldlm_cancel_callback(struct ldlm_lock *lock)
1579 check_res_locked(lock->l_resource);
1580 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1581 lock->l_flags |= LDLM_FL_CANCEL;
1582 if (lock->l_blocking_ast) {
1583 // l_check_no_ns_lock(ns);
1584 unlock_res_and_lock(lock);
1585 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1587 lock_res_and_lock(lock);
1589 LDLM_DEBUG(lock, "no blocking ast");
1592 lock->l_flags |= LDLM_FL_BL_DONE;
1595 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1597 if (req->l_resource->lr_type != LDLM_PLAIN &&
1598 req->l_resource->lr_type != LDLM_IBITS)
1601 list_del_init(&req->l_sl_policy);
1602 list_del_init(&req->l_sl_mode);
1605 void ldlm_lock_cancel(struct ldlm_lock *lock)
1607 struct ldlm_resource *res;
1608 struct ldlm_namespace *ns;
1611 lock_res_and_lock(lock);
1613 res = lock->l_resource;
1614 ns = res->lr_namespace;
1616 /* Please do not, no matter how tempting, remove this LBUG without
1617 * talking to me first. -phik */
1618 if (lock->l_readers || lock->l_writers) {
1619 LDLM_ERROR(lock, "lock still has references");
1623 ldlm_del_waiting_lock(lock);
1625 /* Releases cancel callback. */
1626 ldlm_cancel_callback(lock);
1628 /* Yes, second time, just in case it was added again while we were
1629 running with no res lock in ldlm_cancel_callback */
1630 ldlm_del_waiting_lock(lock);
1631 ldlm_resource_unlink_lock(lock);
1632 ldlm_lock_destroy_nolock(lock);
1634 if (lock->l_granted_mode == lock->l_req_mode)
1635 ldlm_pool_del(&ns->ns_pool, lock);
1637 /* Make sure we will not be called again for same lock what is possible
1638 * if not to zero out lock->l_granted_mode */
1639 lock->l_granted_mode = LCK_MINMODE;
1640 unlock_res_and_lock(lock);
1645 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1647 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1653 lock->l_ast_data = data;
1654 LDLM_LOCK_PUT(lock);
1658 void ldlm_cancel_locks_for_export_cb(void *obj, void *data)
1660 struct obd_export *exp = data;
1661 struct ldlm_lock *lock = obj;
1662 struct ldlm_resource *res;
1664 res = ldlm_resource_getref(lock->l_resource);
1665 LDLM_LOCK_GET(lock);
1667 LDLM_DEBUG(lock, "export %p", exp);
1668 ldlm_res_lvbo_update(res, NULL, 0, 1);
1669 ldlm_lock_cancel(lock);
1670 ldlm_reprocess_all(res);
1671 ldlm_resource_putref(res);
1672 LDLM_LOCK_RELEASE(lock);
1675 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1677 lustre_hash_for_each_empty(exp->exp_lock_hash,
1678 ldlm_cancel_locks_for_export_cb, exp);
1682 * Downgrade an exclusive lock.
1684 * A fast variant of ldlm_lock_convert for convertion of exclusive
1685 * locks. The convertion is always successful.
1687 * \param lock A lock to convert
1688 * \param new_mode new lock mode
1690 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1692 struct ldlm_namespace *ns;
1695 LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1696 LASSERT(new_mode == LCK_COS);
1698 lock_res_and_lock(lock);
1699 ldlm_resource_unlink_lock(lock);
1701 * Remove the lock from pool as it will be added again in
1702 * ldlm_grant_lock() called below.
1704 ns = lock->l_resource->lr_namespace;
1705 ldlm_pool_del(&ns->ns_pool, lock);
1707 lock->l_req_mode = new_mode;
1708 ldlm_grant_lock(lock, NULL);
1709 unlock_res_and_lock(lock);
1710 ldlm_reprocess_all(lock->l_resource);
1715 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1718 CFS_LIST_HEAD(rpc_list);
1719 struct ldlm_resource *res;
1720 struct ldlm_namespace *ns;
1723 struct sl_insert_point prev;
1725 struct ldlm_interval *node;
1728 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1729 *flags |= LDLM_FL_BLOCK_GRANTED;
1730 RETURN(lock->l_resource);
1733 /* I can't check the type of lock here because the bitlock of lock
1734 * is not held here, so do the allocation blindly. -jay */
1735 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1736 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1739 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1740 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1742 lock_res_and_lock(lock);
1744 res = lock->l_resource;
1745 ns = res->lr_namespace;
1747 old_mode = lock->l_req_mode;
1748 lock->l_req_mode = new_mode;
1749 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1750 /* remember the lock position where the lock might be
1751 * added back to the granted list later and also
1752 * remember the join mode for skiplist fixing. */
1753 prev.res_link = lock->l_res_link.prev;
1754 prev.mode_link = lock->l_sl_mode.prev;
1755 prev.policy_link = lock->l_sl_policy.prev;
1756 ldlm_resource_unlink_lock(lock);
1758 ldlm_resource_unlink_lock(lock);
1759 if (res->lr_type == LDLM_EXTENT) {
1760 /* FIXME: ugly code, I have to attach the lock to a
1761 * interval node again since perhaps it will be granted
1763 CFS_INIT_LIST_HEAD(&node->li_group);
1764 ldlm_interval_attach(node, lock);
1770 * Remove old lock from the pool before adding the lock with new
1771 * mode below in ->policy()
1773 ldlm_pool_del(&ns->ns_pool, lock);
1775 /* If this is a local resource, put it on the appropriate list. */
1776 if (ns_is_client(res->lr_namespace)) {
1777 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1778 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1780 /* This should never happen, because of the way the
1781 * server handles conversions. */
1782 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1786 ldlm_grant_lock(lock, &rpc_list);
1788 /* FIXME: completion handling not with ns_lock held ! */
1789 if (lock->l_completion_ast)
1790 lock->l_completion_ast(lock, 0, NULL);
1794 ldlm_processing_policy policy;
1795 policy = ldlm_processing_policy_table[res->lr_type];
1796 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1797 if (rc == LDLM_ITER_STOP) {
1798 lock->l_req_mode = old_mode;
1799 if (res->lr_type == LDLM_EXTENT)
1800 ldlm_extent_add_lock(res, lock);
1802 ldlm_granted_list_add_lock(lock, &prev);
1806 *flags |= LDLM_FL_BLOCK_GRANTED;
1810 unlock_res_and_lock(lock);
1813 ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1815 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1819 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1821 struct obd_device *obd = NULL;
1823 if (!((libcfs_debug | D_ERROR) & level))
1827 CDEBUG(level, " NULL LDLM lock\n");
1831 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1832 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1834 if (lock->l_conn_export != NULL)
1835 obd = lock->l_conn_export->exp_obd;
1836 if (lock->l_export && lock->l_export->exp_connection) {
1837 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1838 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1839 lock->l_remote_handle.cookie);
1840 } else if (obd == NULL) {
1841 CDEBUG(level, " Node: local\n");
1843 struct obd_import *imp = obd->u.cli.cl_import;
1844 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1845 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1846 lock->l_remote_handle.cookie);
1848 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1850 lock->l_resource->lr_name.name[0],
1851 lock->l_resource->lr_name.name[1],
1852 lock->l_resource->lr_name.name[2]);
1853 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1854 "write: %d flags: "LPX64"\n", ldlm_lockname[lock->l_req_mode],
1855 ldlm_lockname[lock->l_granted_mode],
1856 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1858 if (lock->l_resource->lr_type == LDLM_EXTENT)
1859 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1860 " (req "LPU64"-"LPU64")\n",
1861 lock->l_policy_data.l_extent.start,
1862 lock->l_policy_data.l_extent.end,
1863 lock->l_req_extent.start, lock->l_req_extent.end);
1864 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1865 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1866 lock->l_policy_data.l_flock.pid,
1867 lock->l_policy_data.l_flock.start,
1868 lock->l_policy_data.l_flock.end);
1869 else if (lock->l_resource->lr_type == LDLM_IBITS)
1870 CDEBUG(level, " Bits: "LPX64"\n",
1871 lock->l_policy_data.l_inodebits.bits);
1874 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1876 struct ldlm_lock *lock;
1878 if (!((libcfs_debug | D_ERROR) & level))
1881 lock = ldlm_handle2lock(lockh);
1885 ldlm_lock_dump(D_OTHER, lock, 0);
1887 LDLM_LOCK_PUT(lock);
1890 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1891 struct libcfs_debug_msg_data *data, const char *fmt,
1895 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1897 va_start(args, fmt);
1899 if (lock->l_resource == NULL) {
1900 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1901 data->msg_fn, data->msg_line, fmt, args,
1902 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1903 "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" remote: "
1904 LPX64" expref: %d pid: %u timeout: %lu\n", lock,
1905 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1906 lock->l_readers, lock->l_writers,
1907 ldlm_lockname[lock->l_granted_mode],
1908 ldlm_lockname[lock->l_req_mode],
1909 lock->l_flags, lock->l_remote_handle.cookie,
1911 atomic_read(&lock->l_export->exp_refcount) : -99,
1912 lock->l_pid, lock->l_callback_timeout);
1917 switch (lock->l_resource->lr_type) {
1919 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1920 data->msg_fn, data->msg_line, fmt, args,
1921 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1922 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1923 "] (req "LPU64"->"LPU64") flags: "LPX64" remote: "LPX64
1924 " expref: %d pid: %u timeout %lu\n",
1925 lock->l_resource->lr_namespace->ns_name, lock,
1926 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1927 lock->l_readers, lock->l_writers,
1928 ldlm_lockname[lock->l_granted_mode],
1929 ldlm_lockname[lock->l_req_mode],
1930 lock->l_resource->lr_name.name[0],
1931 lock->l_resource->lr_name.name[1],
1932 atomic_read(&lock->l_resource->lr_refcount),
1933 ldlm_typename[lock->l_resource->lr_type],
1934 lock->l_policy_data.l_extent.start,
1935 lock->l_policy_data.l_extent.end,
1936 lock->l_req_extent.start, lock->l_req_extent.end,
1937 lock->l_flags, lock->l_remote_handle.cookie,
1939 atomic_read(&lock->l_export->exp_refcount) : -99,
1940 lock->l_pid, lock->l_callback_timeout);
1944 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1945 data->msg_fn, data->msg_line, fmt, args,
1946 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1947 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1948 "["LPU64"->"LPU64"] flags: "LPX64" remote: "LPX64
1949 " expref: %d pid: %u timeout: %lu\n",
1950 lock->l_resource->lr_namespace->ns_name, lock,
1951 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1952 lock->l_readers, lock->l_writers,
1953 ldlm_lockname[lock->l_granted_mode],
1954 ldlm_lockname[lock->l_req_mode],
1955 lock->l_resource->lr_name.name[0],
1956 lock->l_resource->lr_name.name[1],
1957 atomic_read(&lock->l_resource->lr_refcount),
1958 ldlm_typename[lock->l_resource->lr_type],
1959 lock->l_policy_data.l_flock.pid,
1960 lock->l_policy_data.l_flock.start,
1961 lock->l_policy_data.l_flock.end,
1962 lock->l_flags, lock->l_remote_handle.cookie,
1964 atomic_read(&lock->l_export->exp_refcount) : -99,
1965 lock->l_pid, lock->l_callback_timeout);
1969 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1970 data->msg_fn, data->msg_line, fmt, args,
1971 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1972 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1973 "flags: "LPX64" remote: "LPX64" expref: %d "
1974 "pid: %u timeout: %lu\n",
1975 lock->l_resource->lr_namespace->ns_name,
1976 lock, lock->l_handle.h_cookie,
1977 atomic_read (&lock->l_refc),
1978 lock->l_readers, lock->l_writers,
1979 ldlm_lockname[lock->l_granted_mode],
1980 ldlm_lockname[lock->l_req_mode],
1981 lock->l_resource->lr_name.name[0],
1982 lock->l_resource->lr_name.name[1],
1983 lock->l_policy_data.l_inodebits.bits,
1984 atomic_read(&lock->l_resource->lr_refcount),
1985 ldlm_typename[lock->l_resource->lr_type],
1986 lock->l_flags, lock->l_remote_handle.cookie,
1988 atomic_read(&lock->l_export->exp_refcount) : -99,
1989 lock->l_pid, lock->l_callback_timeout);
1993 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1994 data->msg_fn, data->msg_line, fmt, args,
1995 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1996 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" "
1997 "remote: "LPX64" expref: %d pid: %u timeout %lu\n",
1998 lock->l_resource->lr_namespace->ns_name,
1999 lock, lock->l_handle.h_cookie,
2000 atomic_read (&lock->l_refc),
2001 lock->l_readers, lock->l_writers,
2002 ldlm_lockname[lock->l_granted_mode],
2003 ldlm_lockname[lock->l_req_mode],
2004 lock->l_resource->lr_name.name[0],
2005 lock->l_resource->lr_name.name[1],
2006 atomic_read(&lock->l_resource->lr_refcount),
2007 ldlm_typename[lock->l_resource->lr_type],
2008 lock->l_flags, lock->l_remote_handle.cookie,
2010 atomic_read(&lock->l_export->exp_refcount) : -99,
2011 lock->l_pid, lock->l_callback_timeout);
2016 EXPORT_SYMBOL(_ldlm_lock_debug);