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 (c) 2002, 2010, Oracle and/or its affiliates. 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 cfs_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(cfs_atomic_read(&lock->l_refc) > 0);
151 if (cfs_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(cfs_list_empty(&lock->l_res_link));
160 LASSERT(cfs_list_empty(&lock->l_pending_chain));
162 cfs_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_lock_put(lock->l_export, lock);
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 (!cfs_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 cfs_list_del_init(&lock->l_lru);
190 if (lock->l_flags & LDLM_FL_SKIPPED)
191 lock->l_flags &= ~LDLM_FL_SKIPPED;
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 cfs_spin_lock(&ns->ns_unused_lock);
205 rc = ldlm_lock_remove_from_lru_nolock(lock);
206 cfs_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(cfs_list_empty(&lock->l_lru));
216 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
217 cfs_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 cfs_spin_lock(&ns->ns_unused_lock);
227 ldlm_lock_add_to_lru_nolock(lock);
228 cfs_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 cfs_spin_lock(&ns->ns_unused_lock);
237 if (!cfs_list_empty(&lock->l_lru)) {
238 ldlm_lock_remove_from_lru_nolock(lock);
239 ldlm_lock_add_to_lru_nolock(lock);
241 cfs_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 (!cfs_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(cfs_list_empty(&lock->l_lru));
271 lock->l_destroyed = 1;
273 if (lock->l_export && lock->l_export->exp_lock_hash &&
274 !cfs_hlist_unhashed(&lock->l_exp_hash))
275 cfs_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 cfs_spin_lock_init(&lock->l_lock);
349 lock->l_resource = ldlm_resource_getref(resource);
350 lu_ref_add(&resource->lr_reference, "lock", lock);
352 cfs_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 cfs_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 cfs_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;
376 #if LUSTRE_TRACKS_LOCK_EXP_REFS
377 CFS_INIT_LIST_HEAD(&lock->l_exp_refs_link);
378 lock->l_exp_refs_nr = 0;
379 lock->l_exp_refs_target = NULL;
385 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
386 const struct ldlm_res_id *new_resid)
388 struct ldlm_resource *oldres = lock->l_resource;
389 struct ldlm_resource *newres;
393 LASSERT(ns_is_client(ns));
395 lock_res_and_lock(lock);
396 if (memcmp(new_resid, &lock->l_resource->lr_name,
397 sizeof(lock->l_resource->lr_name)) == 0) {
399 unlock_res_and_lock(lock);
403 LASSERT(new_resid->name[0] != 0);
405 /* This function assumes that the lock isn't on any lists */
406 LASSERT(cfs_list_empty(&lock->l_res_link));
408 type = oldres->lr_type;
409 unlock_res_and_lock(lock);
411 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
412 lu_ref_add(&newres->lr_reference, "lock", lock);
416 * To flip the lock from the old to the new resource, lock, oldres and
417 * newres have to be locked. Resource spin-locks are nested within
418 * lock->l_lock, and are taken in the memory address order to avoid
421 cfs_spin_lock(&lock->l_lock);
422 oldres = lock->l_resource;
423 if (oldres < newres) {
425 lock_res_nested(newres, LRT_NEW);
428 lock_res_nested(oldres, LRT_NEW);
430 LASSERT(memcmp(new_resid, &oldres->lr_name,
431 sizeof oldres->lr_name) != 0);
432 lock->l_resource = newres;
434 unlock_res_and_lock(lock);
436 /* ...and the flowers are still standing! */
437 lu_ref_del(&oldres->lr_reference, "lock", lock);
438 ldlm_resource_putref(oldres);
447 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
449 lockh->cookie = lock->l_handle.h_cookie;
452 /* if flags: atomically get the lock and set the flags.
453 * Return NULL if flag already set
456 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
459 struct ldlm_namespace *ns;
460 struct ldlm_lock *lock, *retval = NULL;
465 lock = class_handle2object(handle->cookie);
469 LASSERT(lock->l_resource != NULL);
470 ns = lock->l_resource->lr_namespace;
473 lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
474 lock_res_and_lock(lock);
476 /* It's unlikely but possible that someone marked the lock as
477 * destroyed after we did handle2object on it */
478 if (lock->l_destroyed) {
479 unlock_res_and_lock(lock);
480 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
485 if (flags && (lock->l_flags & flags)) {
486 unlock_res_and_lock(lock);
492 lock->l_flags |= flags;
494 unlock_res_and_lock(lock);
501 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
503 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
504 /* INODEBITS_INTEROP: If the other side does not support
505 * inodebits, reply with a plain lock descriptor.
507 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
508 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
509 /* Make sure all the right bits are set in this lock we
510 are going to pass to client */
511 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
512 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
513 "Inappropriate inode lock bits during "
514 "conversion " LPU64 "\n",
515 lock->l_policy_data.l_inodebits.bits);
517 ldlm_res2desc(lock->l_resource, &desc->l_resource);
518 desc->l_resource.lr_type = LDLM_PLAIN;
520 /* Convert "new" lock mode to something old client can
522 if ((lock->l_req_mode == LCK_CR) ||
523 (lock->l_req_mode == LCK_CW))
524 desc->l_req_mode = LCK_PR;
526 desc->l_req_mode = lock->l_req_mode;
527 if ((lock->l_granted_mode == LCK_CR) ||
528 (lock->l_granted_mode == LCK_CW)) {
529 desc->l_granted_mode = LCK_PR;
531 /* We never grant PW/EX locks to clients */
532 LASSERT((lock->l_granted_mode != LCK_PW) &&
533 (lock->l_granted_mode != LCK_EX));
534 desc->l_granted_mode = lock->l_granted_mode;
537 /* We do not copy policy here, because there is no
538 policy for plain locks */
540 ldlm_res2desc(lock->l_resource, &desc->l_resource);
541 desc->l_req_mode = lock->l_req_mode;
542 desc->l_granted_mode = lock->l_granted_mode;
543 desc->l_policy_data = lock->l_policy_data;
547 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
548 cfs_list_t *work_list)
550 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
551 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
552 lock->l_flags |= LDLM_FL_AST_SENT;
553 /* If the enqueuing client said so, tell the AST recipient to
554 * discard dirty data, rather than writing back. */
555 if (new->l_flags & LDLM_AST_DISCARD_DATA)
556 lock->l_flags |= LDLM_FL_DISCARD_DATA;
557 LASSERT(cfs_list_empty(&lock->l_bl_ast));
558 cfs_list_add(&lock->l_bl_ast, work_list);
560 LASSERT(lock->l_blocking_lock == NULL);
561 lock->l_blocking_lock = LDLM_LOCK_GET(new);
565 void ldlm_add_cp_work_item(struct ldlm_lock *lock, cfs_list_t *work_list)
567 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
568 lock->l_flags |= LDLM_FL_CP_REQD;
569 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
570 LASSERT(cfs_list_empty(&lock->l_cp_ast));
571 cfs_list_add(&lock->l_cp_ast, work_list);
576 /* must be called with lr_lock held */
577 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
578 cfs_list_t *work_list)
581 check_res_locked(lock->l_resource);
583 ldlm_add_bl_work_item(lock, new, work_list);
585 ldlm_add_cp_work_item(lock, work_list);
589 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
591 struct ldlm_lock *lock;
593 lock = ldlm_handle2lock(lockh);
594 LASSERT(lock != NULL);
595 ldlm_lock_addref_internal(lock, mode);
599 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
601 ldlm_lock_remove_from_lru(lock);
602 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
604 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
606 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
608 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
611 lu_ref_add_atomic(&lock->l_reference, "user", lock);
612 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
616 * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
619 * \retval 0 success, lock was addref-ed
621 * \retval -EAGAIN lock is being canceled.
623 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
625 struct ldlm_lock *lock;
629 lock = ldlm_handle2lock(lockh);
631 lock_res_and_lock(lock);
632 if (lock->l_readers != 0 || lock->l_writers != 0 ||
633 !(lock->l_flags & LDLM_FL_CBPENDING)) {
634 ldlm_lock_addref_internal_nolock(lock, mode);
637 unlock_res_and_lock(lock);
643 /* only called for local locks */
644 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
646 lock_res_and_lock(lock);
647 ldlm_lock_addref_internal_nolock(lock, mode);
648 unlock_res_and_lock(lock);
651 /* only called in ldlm_flock_destroy and for local locks.
652 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
653 * * ldlm_lock_remove_from_lru() does nothing, it is safe
654 * * for ldlm_flock_destroy usage by dropping some code */
655 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
657 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
658 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
659 LASSERT(lock->l_readers > 0);
660 lu_ref_del(&lock->l_reference, "reader", lock);
663 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
664 LASSERT(lock->l_writers > 0);
665 lu_ref_del(&lock->l_reference, "writer", lock);
669 lu_ref_del(&lock->l_reference, "user", lock);
670 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
673 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
675 struct ldlm_namespace *ns;
678 lock_res_and_lock(lock);
680 ns = lock->l_resource->lr_namespace;
682 ldlm_lock_decref_internal_nolock(lock, mode);
684 if (lock->l_flags & LDLM_FL_LOCAL &&
685 !lock->l_readers && !lock->l_writers) {
686 /* If this is a local lock on a server namespace and this was
687 * the last reference, cancel the lock. */
688 CDEBUG(D_INFO, "forcing cancel of local lock\n");
689 lock->l_flags |= LDLM_FL_CBPENDING;
692 if (!lock->l_readers && !lock->l_writers &&
693 (lock->l_flags & LDLM_FL_CBPENDING)) {
694 /* If we received a blocked AST and this was the last reference,
695 * run the callback. */
696 if (ns_is_server(ns) && lock->l_export)
697 CERROR("FL_CBPENDING set on non-local lock--just a "
700 LDLM_DEBUG(lock, "final decref done on cbpending lock");
702 LDLM_LOCK_GET(lock); /* dropped by bl thread */
703 ldlm_lock_remove_from_lru(lock);
704 unlock_res_and_lock(lock);
706 if (lock->l_flags & LDLM_FL_FAIL_LOC)
707 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
709 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
710 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
711 ldlm_handle_bl_callback(ns, NULL, lock);
712 } else if (ns_is_client(ns) &&
713 !lock->l_readers && !lock->l_writers &&
714 !(lock->l_flags & LDLM_FL_BL_AST)) {
715 /* If this is a client-side namespace and this was the last
716 * reference, put it on the LRU. */
717 ldlm_lock_add_to_lru(lock);
718 unlock_res_and_lock(lock);
720 if (lock->l_flags & LDLM_FL_FAIL_LOC)
721 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
723 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
724 * are not supported by the server, otherwise, it is done on
726 if (!exp_connect_cancelset(lock->l_conn_export) &&
727 !ns_connect_lru_resize(ns))
728 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
730 unlock_res_and_lock(lock);
736 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
738 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
739 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
740 ldlm_lock_decref_internal(lock, mode);
744 /* This will drop a lock reference and mark it for destruction, but will not
745 * necessarily cancel the lock before returning. */
746 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
748 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
751 LASSERT(lock != NULL);
753 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
754 lock_res_and_lock(lock);
755 lock->l_flags |= LDLM_FL_CBPENDING;
756 unlock_res_and_lock(lock);
757 ldlm_lock_decref_internal(lock, mode);
761 struct sl_insert_point {
762 cfs_list_t *res_link;
763 cfs_list_t *mode_link;
764 cfs_list_t *policy_link;
768 * search_granted_lock
771 * Finds a position to insert the new lock.
773 * queue [input]: the granted list where search acts on;
774 * req [input]: the lock whose position to be located;
775 * prev [output]: positions within 3 lists to insert @req to
779 * - ldlm_grant_lock_with_skiplist
781 static void search_granted_lock(cfs_list_t *queue,
782 struct ldlm_lock *req,
783 struct sl_insert_point *prev)
786 struct ldlm_lock *lock, *mode_end, *policy_end;
789 cfs_list_for_each(tmp, queue) {
790 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
792 mode_end = cfs_list_entry(lock->l_sl_mode.prev,
793 struct ldlm_lock, l_sl_mode);
795 if (lock->l_req_mode != req->l_req_mode) {
796 /* jump to last lock of mode group */
797 tmp = &mode_end->l_res_link;
801 /* suitable mode group is found */
802 if (lock->l_resource->lr_type == LDLM_PLAIN) {
803 /* insert point is last lock of the mode group */
804 prev->res_link = &mode_end->l_res_link;
805 prev->mode_link = &mode_end->l_sl_mode;
806 prev->policy_link = &req->l_sl_policy;
809 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
812 cfs_list_entry(lock->l_sl_policy.prev,
816 if (lock->l_policy_data.l_inodebits.bits ==
817 req->l_policy_data.l_inodebits.bits) {
818 /* insert point is last lock of
819 * the policy group */
821 &policy_end->l_res_link;
823 &policy_end->l_sl_mode;
825 &policy_end->l_sl_policy;
830 if (policy_end == mode_end)
831 /* done with mode group */
834 /* go to next policy group within mode group */
835 tmp = policy_end->l_res_link.next;
836 lock = cfs_list_entry(tmp, struct ldlm_lock,
838 } /* loop over policy groups within the mode group */
840 /* insert point is last lock of the mode group,
841 * new policy group is started */
842 prev->res_link = &mode_end->l_res_link;
843 prev->mode_link = &mode_end->l_sl_mode;
844 prev->policy_link = &req->l_sl_policy;
848 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
853 /* insert point is last lock on the queue,
854 * new mode group and new policy group are started */
855 prev->res_link = queue->prev;
856 prev->mode_link = &req->l_sl_mode;
857 prev->policy_link = &req->l_sl_policy;
862 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
863 struct sl_insert_point *prev)
865 struct ldlm_resource *res = lock->l_resource;
868 check_res_locked(res);
870 ldlm_resource_dump(D_INFO, res);
871 CDEBUG(D_OTHER, "About to add this lock:\n");
872 ldlm_lock_dump(D_OTHER, lock, 0);
874 if (lock->l_destroyed) {
875 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
879 LASSERT(cfs_list_empty(&lock->l_res_link));
880 LASSERT(cfs_list_empty(&lock->l_sl_mode));
881 LASSERT(cfs_list_empty(&lock->l_sl_policy));
883 cfs_list_add(&lock->l_res_link, prev->res_link);
884 cfs_list_add(&lock->l_sl_mode, prev->mode_link);
885 cfs_list_add(&lock->l_sl_policy, prev->policy_link);
890 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
892 struct sl_insert_point prev;
895 LASSERT(lock->l_req_mode == lock->l_granted_mode);
897 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
898 ldlm_granted_list_add_lock(lock, &prev);
903 * - ldlm_lock_enqueue
904 * - ldlm_reprocess_queue
905 * - ldlm_lock_convert
907 * must be called with lr_lock held
909 void ldlm_grant_lock(struct ldlm_lock *lock, cfs_list_t *work_list)
911 struct ldlm_resource *res = lock->l_resource;
914 check_res_locked(res);
916 lock->l_granted_mode = lock->l_req_mode;
917 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
918 ldlm_grant_lock_with_skiplist(lock);
919 else if (res->lr_type == LDLM_EXTENT)
920 ldlm_extent_add_lock(res, lock);
922 ldlm_resource_add_lock(res, &res->lr_granted, lock);
924 if (lock->l_granted_mode < res->lr_most_restr)
925 res->lr_most_restr = lock->l_granted_mode;
927 if (work_list && lock->l_completion_ast != NULL)
928 ldlm_add_ast_work_item(lock, NULL, work_list);
930 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
934 /* returns a referenced lock or NULL. See the flag descriptions below, in the
935 * comment above ldlm_lock_match */
936 static struct ldlm_lock *search_queue(cfs_list_t *queue,
938 ldlm_policy_data_t *policy,
939 struct ldlm_lock *old_lock,
940 int flags, int unref)
942 struct ldlm_lock *lock;
945 cfs_list_for_each(tmp, queue) {
948 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
950 if (lock == old_lock)
953 /* llite sometimes wants to match locks that will be
954 * canceled when their users drop, but we allow it to match
955 * if it passes in CBPENDING and the lock still has users.
956 * this is generally only going to be used by children
957 * whose parents already hold a lock so forward progress
958 * can still happen. */
959 if (lock->l_flags & LDLM_FL_CBPENDING &&
960 !(flags & LDLM_FL_CBPENDING))
962 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
963 lock->l_readers == 0 && lock->l_writers == 0)
966 if (!(lock->l_req_mode & *mode))
968 match = lock->l_req_mode;
970 if (lock->l_resource->lr_type == LDLM_EXTENT &&
971 (lock->l_policy_data.l_extent.start >
972 policy->l_extent.start ||
973 lock->l_policy_data.l_extent.end < policy->l_extent.end))
976 if (unlikely(match == LCK_GROUP) &&
977 lock->l_resource->lr_type == LDLM_EXTENT &&
978 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
981 /* We match if we have existing lock with same or wider set
983 if (lock->l_resource->lr_type == LDLM_IBITS &&
984 ((lock->l_policy_data.l_inodebits.bits &
985 policy->l_inodebits.bits) !=
986 policy->l_inodebits.bits))
990 (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED)))
993 if ((flags & LDLM_FL_LOCAL_ONLY) &&
994 !(lock->l_flags & LDLM_FL_LOCAL))
997 if (flags & LDLM_FL_TEST_LOCK) {
999 ldlm_lock_touch_in_lru(lock);
1001 ldlm_lock_addref_internal_nolock(lock, match);
1010 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1012 lock->l_flags |= LDLM_FL_LVB_READY;
1013 cfs_waitq_signal(&lock->l_waitq);
1016 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1018 lock_res_and_lock(lock);
1019 ldlm_lock_allow_match_locked(lock);
1020 unlock_res_and_lock(lock);
1023 /* Can be called in two ways:
1025 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1026 * for a duplicate of.
1028 * Otherwise, all of the fields must be filled in, to match against.
1030 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1031 * server (ie, connh is NULL)
1032 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1033 * list will be considered
1034 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1035 * to be canceled can still be matched as long as they still have reader
1036 * or writer refernces
1037 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1038 * just tell us if we would have matched.
1040 * Returns 1 if it finds an already-existing lock that is compatible; in this
1041 * case, lockh is filled in with a addref()ed lock
1043 * we also check security context, if that failed we simply return 0 (to keep
1044 * caller code unchanged), the context failure will be discovered by caller
1047 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1048 const struct ldlm_res_id *res_id, ldlm_type_t type,
1049 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1050 struct lustre_handle *lockh, int unref)
1052 struct ldlm_resource *res;
1053 struct ldlm_lock *lock, *old_lock = NULL;
1058 old_lock = ldlm_handle2lock(lockh);
1061 ns = old_lock->l_resource->lr_namespace;
1062 res_id = &old_lock->l_resource->lr_name;
1063 type = old_lock->l_resource->lr_type;
1064 mode = old_lock->l_req_mode;
1067 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1069 LASSERT(old_lock == NULL);
1073 LDLM_RESOURCE_ADDREF(res);
1076 lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1080 if (flags & LDLM_FL_BLOCK_GRANTED)
1082 lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1086 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1094 LDLM_RESOURCE_DELREF(res);
1095 ldlm_resource_putref(res);
1098 ldlm_lock2handle(lock, lockh);
1099 if ((flags & LDLM_FL_LVB_READY) &&
1100 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1101 struct l_wait_info lwi;
1102 if (lock->l_completion_ast) {
1103 int err = lock->l_completion_ast(lock,
1104 LDLM_FL_WAIT_NOREPROC,
1107 if (flags & LDLM_FL_TEST_LOCK)
1108 LDLM_LOCK_RELEASE(lock);
1110 ldlm_lock_decref_internal(lock,
1117 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1118 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1120 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1121 l_wait_event(lock->l_waitq,
1122 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1127 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1128 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1129 res_id->name[2] : policy->l_extent.start,
1130 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1131 res_id->name[3] : policy->l_extent.end);
1133 /* check user's security context */
1134 if (lock->l_conn_export &&
1135 sptlrpc_import_check_ctx(
1136 class_exp2cliimp(lock->l_conn_export))) {
1137 if (!(flags & LDLM_FL_TEST_LOCK))
1138 ldlm_lock_decref_internal(lock, mode);
1142 if (flags & LDLM_FL_TEST_LOCK)
1143 LDLM_LOCK_RELEASE(lock);
1145 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1146 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1147 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1148 type, mode, res_id->name[0], res_id->name[1],
1149 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1150 res_id->name[2] :policy->l_extent.start,
1151 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1152 res_id->name[3] : policy->l_extent.end);
1155 LDLM_LOCK_PUT(old_lock);
1157 return rc ? mode : 0;
1160 /* Returns a referenced lock */
1161 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1162 const struct ldlm_res_id *res_id,
1165 const struct ldlm_callback_suite *cbs,
1166 void *data, __u32 lvb_len)
1168 struct ldlm_lock *lock;
1169 struct ldlm_resource *res;
1172 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1176 lock = ldlm_lock_new(res);
1177 ldlm_resource_putref(res);
1182 lock->l_req_mode = mode;
1183 lock->l_ast_data = data;
1184 lock->l_pid = cfs_curproc_pid();
1186 lock->l_blocking_ast = cbs->lcs_blocking;
1187 lock->l_completion_ast = cbs->lcs_completion;
1188 lock->l_glimpse_ast = cbs->lcs_glimpse;
1189 lock->l_weigh_ast = cbs->lcs_weigh;
1192 lock->l_tree_node = NULL;
1193 /* if this is the extent lock, allocate the interval tree node */
1194 if (type == LDLM_EXTENT) {
1195 if (ldlm_interval_alloc(lock) == NULL)
1200 lock->l_lvb_len = lvb_len;
1201 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1202 if (lock->l_lvb_data == NULL)
1206 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1212 ldlm_lock_destroy(lock);
1213 LDLM_LOCK_RELEASE(lock);
1217 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1218 struct ldlm_lock **lockp,
1219 void *cookie, int *flags)
1221 struct ldlm_lock *lock = *lockp;
1222 struct ldlm_resource *res = lock->l_resource;
1223 int local = ns_is_client(res->lr_namespace);
1224 ldlm_processing_policy policy;
1225 ldlm_error_t rc = ELDLM_OK;
1226 struct ldlm_interval *node = NULL;
1229 lock->l_last_activity = cfs_time_current_sec();
1230 /* policies are not executed on the client or during replay */
1231 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1232 && !local && ns->ns_policy) {
1233 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1235 if (rc == ELDLM_LOCK_REPLACED) {
1236 /* The lock that was returned has already been granted,
1237 * and placed into lockp. If it's not the same as the
1238 * one we passed in, then destroy the old one and our
1239 * work here is done. */
1240 if (lock != *lockp) {
1241 ldlm_lock_destroy(lock);
1242 LDLM_LOCK_RELEASE(lock);
1244 *flags |= LDLM_FL_LOCK_CHANGED;
1246 } else if (rc != ELDLM_OK ||
1247 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1248 ldlm_lock_destroy(lock);
1253 /* For a replaying lock, it might be already in granted list. So
1254 * unlinking the lock will cause the interval node to be freed, we
1255 * have to allocate the interval node early otherwise we can't regrant
1256 * this lock in the future. - jay */
1257 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1258 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1260 lock_res_and_lock(lock);
1261 if (local && lock->l_req_mode == lock->l_granted_mode) {
1262 /* The server returned a blocked lock, but it was granted
1263 * before we got a chance to actually enqueue it. We don't
1264 * need to do anything else. */
1265 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1266 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1267 GOTO(out, ELDLM_OK);
1270 ldlm_resource_unlink_lock(lock);
1271 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1273 ldlm_lock_destroy_nolock(lock);
1274 GOTO(out, rc = -ENOMEM);
1277 CFS_INIT_LIST_HEAD(&node->li_group);
1278 ldlm_interval_attach(node, lock);
1282 /* Some flags from the enqueue want to make it into the AST, via the
1283 * lock's l_flags. */
1284 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1286 /* This distinction between local lock trees is very important; a client
1287 * namespace only has information about locks taken by that client, and
1288 * thus doesn't have enough information to decide for itself if it can
1289 * be granted (below). In this case, we do exactly what the server
1290 * tells us to do, as dictated by the 'flags'.
1292 * We do exactly the same thing during recovery, when the server is
1293 * more or less trusting the clients not to lie.
1295 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1296 * granted/converting queues. */
1298 if (*flags & LDLM_FL_BLOCK_CONV)
1299 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1300 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1301 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1303 ldlm_grant_lock(lock, NULL);
1304 GOTO(out, ELDLM_OK);
1305 } else if (*flags & LDLM_FL_REPLAY) {
1306 if (*flags & LDLM_FL_BLOCK_CONV) {
1307 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1308 GOTO(out, ELDLM_OK);
1309 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1310 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1311 GOTO(out, ELDLM_OK);
1312 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1313 ldlm_grant_lock(lock, NULL);
1314 GOTO(out, ELDLM_OK);
1316 /* If no flags, fall through to normal enqueue path. */
1319 policy = ldlm_processing_policy_table[res->lr_type];
1320 policy(lock, flags, 1, &rc, NULL);
1323 unlock_res_and_lock(lock);
1325 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1329 /* Must be called with namespace taken: queue is waiting or converting. */
1330 int ldlm_reprocess_queue(struct ldlm_resource *res, cfs_list_t *queue,
1331 cfs_list_t *work_list)
1333 cfs_list_t *tmp, *pos;
1334 ldlm_processing_policy policy;
1336 int rc = LDLM_ITER_CONTINUE;
1340 check_res_locked(res);
1342 policy = ldlm_processing_policy_table[res->lr_type];
1345 cfs_list_for_each_safe(tmp, pos, queue) {
1346 struct ldlm_lock *pending;
1347 pending = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1349 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1352 rc = policy(pending, &flags, 0, &err, work_list);
1353 if (rc != LDLM_ITER_CONTINUE)
1360 /* Helper function for ldlm_run_ast_work().
1362 * Send an existing rpc set specified by @arg->set and then
1363 * destroy it. Create new one if @do_create flag is set. */
1365 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1369 ptlrpc_set_wait(arg->set);
1370 if (arg->type == LDLM_BL_CALLBACK)
1371 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1372 ptlrpc_set_destroy(arg->set);
1375 arg->set = ptlrpc_prep_set();
1381 ldlm_work_bl_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1383 struct ldlm_lock_desc d;
1384 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1388 /* nobody should touch l_bl_ast */
1389 lock_res_and_lock(lock);
1390 cfs_list_del_init(&lock->l_bl_ast);
1392 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1393 LASSERT(lock->l_bl_ast_run == 0);
1394 LASSERT(lock->l_blocking_lock);
1395 lock->l_bl_ast_run++;
1396 unlock_res_and_lock(lock);
1398 ldlm_lock2desc(lock->l_blocking_lock, &d);
1400 lock->l_blocking_ast(lock, &d, (void *)arg,
1402 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1403 lock->l_blocking_lock = NULL;
1404 LDLM_LOCK_RELEASE(lock);
1410 ldlm_work_cp_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1412 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock, l_cp_ast);
1413 ldlm_completion_callback completion_callback;
1417 /* It's possible to receive a completion AST before we've set
1418 * the l_completion_ast pointer: either because the AST arrived
1419 * before the reply, or simply because there's a small race
1420 * window between receiving the reply and finishing the local
1421 * enqueue. (bug 842)
1423 * This can't happen with the blocking_ast, however, because we
1424 * will never call the local blocking_ast until we drop our
1425 * reader/writer reference, which we won't do until we get the
1426 * reply and finish enqueueing. */
1428 /* nobody should touch l_cp_ast */
1429 lock_res_and_lock(lock);
1430 cfs_list_del_init(&lock->l_cp_ast);
1431 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1432 /* save l_completion_ast since it can be changed by
1433 * mds_intent_policy(), see bug 14225 */
1434 completion_callback = lock->l_completion_ast;
1435 lock->l_flags &= ~LDLM_FL_CP_REQD;
1436 unlock_res_and_lock(lock);
1438 if (completion_callback != NULL) {
1439 completion_callback(lock, 0, (void *)arg);
1442 LDLM_LOCK_RELEASE(lock);
1448 ldlm_work_revoke_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1450 struct ldlm_lock_desc desc;
1451 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1455 cfs_list_del_init(&lock->l_rk_ast);
1457 /* the desc just pretend to exclusive */
1458 ldlm_lock2desc(lock, &desc);
1459 desc.l_req_mode = LCK_EX;
1460 desc.l_granted_mode = 0;
1462 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1463 LDLM_LOCK_RELEASE(lock);
1468 int ldlm_run_ast_work(cfs_list_t *rpc_list, ldlm_desc_ast_t ast_type)
1470 struct ldlm_cb_set_arg arg;
1471 cfs_list_t *tmp, *pos;
1472 int (*work_ast_lock)(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg);
1476 if (cfs_list_empty(rpc_list))
1479 arg.set = ptlrpc_prep_set();
1480 if (NULL == arg.set)
1482 cfs_atomic_set(&arg.restart, 0);
1484 case LDLM_WORK_BL_AST:
1485 arg.type = LDLM_BL_CALLBACK;
1486 work_ast_lock = ldlm_work_bl_ast_lock;
1488 case LDLM_WORK_CP_AST:
1489 arg.type = LDLM_CP_CALLBACK;
1490 work_ast_lock = ldlm_work_cp_ast_lock;
1492 case LDLM_WORK_REVOKE_AST:
1493 arg.type = LDLM_BL_CALLBACK;
1494 work_ast_lock = ldlm_work_revoke_ast_lock;
1501 cfs_list_for_each_safe(tmp, pos, rpc_list) {
1502 ast_count += work_ast_lock(tmp, &arg);
1504 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1505 * and create a new set for requests that remained in
1507 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1508 ldlm_send_and_maybe_create_set(&arg, 1);
1514 ldlm_send_and_maybe_create_set(&arg, 0);
1516 /* In case when number of ASTs is multiply of
1517 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1518 * @arg.set must be destroyed here, otherwise we get
1519 * write memory leaking. */
1520 ptlrpc_set_destroy(arg.set);
1522 RETURN(cfs_atomic_read(&arg.restart) ? -ERESTART : 0);
1525 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1527 ldlm_reprocess_all(res);
1528 return LDLM_ITER_CONTINUE;
1531 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1540 cfs_spin_lock(&ns->ns_hash_lock);
1541 for (i = 0; i < RES_HASH_SIZE; i++) {
1542 tmp = ns->ns_hash[i].next;
1543 while (tmp != &(ns->ns_hash[i])) {
1544 struct ldlm_resource *res =
1545 cfs_list_entry(tmp, struct ldlm_resource,
1548 ldlm_resource_getref(res);
1549 cfs_spin_unlock(&ns->ns_hash_lock);
1550 LDLM_RESOURCE_ADDREF(res);
1552 rc = reprocess_one_queue(res, NULL);
1554 LDLM_RESOURCE_DELREF(res);
1555 cfs_spin_lock(&ns->ns_hash_lock);
1557 ldlm_resource_putref_locked(res);
1559 if (rc == LDLM_ITER_STOP)
1564 cfs_spin_unlock(&ns->ns_hash_lock);
1568 void ldlm_reprocess_all(struct ldlm_resource *res)
1570 CFS_LIST_HEAD(rpc_list);
1574 /* Local lock trees don't get reprocessed. */
1575 if (ns_is_client(res->lr_namespace)) {
1582 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1583 if (rc == LDLM_ITER_CONTINUE)
1584 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1587 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1588 if (rc == -ERESTART) {
1589 LASSERT(cfs_list_empty(&rpc_list));
1595 void ldlm_cancel_callback(struct ldlm_lock *lock)
1597 check_res_locked(lock->l_resource);
1598 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1599 lock->l_flags |= LDLM_FL_CANCEL;
1600 if (lock->l_blocking_ast) {
1601 // l_check_no_ns_lock(ns);
1602 unlock_res_and_lock(lock);
1603 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1605 lock_res_and_lock(lock);
1607 LDLM_DEBUG(lock, "no blocking ast");
1610 lock->l_flags |= LDLM_FL_BL_DONE;
1613 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1615 if (req->l_resource->lr_type != LDLM_PLAIN &&
1616 req->l_resource->lr_type != LDLM_IBITS)
1619 cfs_list_del_init(&req->l_sl_policy);
1620 cfs_list_del_init(&req->l_sl_mode);
1623 void ldlm_lock_cancel(struct ldlm_lock *lock)
1625 struct ldlm_resource *res;
1626 struct ldlm_namespace *ns;
1629 lock_res_and_lock(lock);
1631 res = lock->l_resource;
1632 ns = res->lr_namespace;
1634 /* Please do not, no matter how tempting, remove this LBUG without
1635 * talking to me first. -phik */
1636 if (lock->l_readers || lock->l_writers) {
1637 LDLM_ERROR(lock, "lock still has references");
1641 ldlm_del_waiting_lock(lock);
1643 /* Releases cancel callback. */
1644 ldlm_cancel_callback(lock);
1646 /* Yes, second time, just in case it was added again while we were
1647 running with no res lock in ldlm_cancel_callback */
1648 ldlm_del_waiting_lock(lock);
1649 ldlm_resource_unlink_lock(lock);
1650 ldlm_lock_destroy_nolock(lock);
1652 if (lock->l_granted_mode == lock->l_req_mode)
1653 ldlm_pool_del(&ns->ns_pool, lock);
1655 /* Make sure we will not be called again for same lock what is possible
1656 * if not to zero out lock->l_granted_mode */
1657 lock->l_granted_mode = LCK_MINMODE;
1658 unlock_res_and_lock(lock);
1663 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1665 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1671 lock->l_ast_data = data;
1672 LDLM_LOCK_PUT(lock);
1676 void ldlm_cancel_locks_for_export_cb(void *obj, void *data)
1678 struct obd_export *exp = data;
1679 struct ldlm_lock *lock = obj;
1680 struct ldlm_resource *res;
1682 res = ldlm_resource_getref(lock->l_resource);
1683 LDLM_LOCK_GET(lock);
1685 LDLM_DEBUG(lock, "export %p", exp);
1686 ldlm_res_lvbo_update(res, NULL, 1);
1687 ldlm_lock_cancel(lock);
1688 ldlm_reprocess_all(res);
1689 ldlm_resource_putref(res);
1690 LDLM_LOCK_RELEASE(lock);
1693 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1695 cfs_hash_for_each_empty(exp->exp_lock_hash,
1696 ldlm_cancel_locks_for_export_cb, exp);
1700 * Downgrade an exclusive lock.
1702 * A fast variant of ldlm_lock_convert for convertion of exclusive
1703 * locks. The convertion is always successful.
1705 * \param lock A lock to convert
1706 * \param new_mode new lock mode
1708 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1710 struct ldlm_namespace *ns;
1713 LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1714 LASSERT(new_mode == LCK_COS);
1716 lock_res_and_lock(lock);
1717 ldlm_resource_unlink_lock(lock);
1719 * Remove the lock from pool as it will be added again in
1720 * ldlm_grant_lock() called below.
1722 ns = lock->l_resource->lr_namespace;
1723 ldlm_pool_del(&ns->ns_pool, lock);
1725 lock->l_req_mode = new_mode;
1726 ldlm_grant_lock(lock, NULL);
1727 unlock_res_and_lock(lock);
1728 ldlm_reprocess_all(lock->l_resource);
1733 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1736 CFS_LIST_HEAD(rpc_list);
1737 struct ldlm_resource *res;
1738 struct ldlm_namespace *ns;
1741 struct sl_insert_point prev;
1743 struct ldlm_interval *node;
1746 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1747 *flags |= LDLM_FL_BLOCK_GRANTED;
1748 RETURN(lock->l_resource);
1751 /* I can't check the type of lock here because the bitlock of lock
1752 * is not held here, so do the allocation blindly. -jay */
1753 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1754 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1757 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1758 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1760 lock_res_and_lock(lock);
1762 res = lock->l_resource;
1763 ns = res->lr_namespace;
1765 old_mode = lock->l_req_mode;
1766 lock->l_req_mode = new_mode;
1767 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1768 /* remember the lock position where the lock might be
1769 * added back to the granted list later and also
1770 * remember the join mode for skiplist fixing. */
1771 prev.res_link = lock->l_res_link.prev;
1772 prev.mode_link = lock->l_sl_mode.prev;
1773 prev.policy_link = lock->l_sl_policy.prev;
1774 ldlm_resource_unlink_lock(lock);
1776 ldlm_resource_unlink_lock(lock);
1777 if (res->lr_type == LDLM_EXTENT) {
1778 /* FIXME: ugly code, I have to attach the lock to a
1779 * interval node again since perhaps it will be granted
1781 CFS_INIT_LIST_HEAD(&node->li_group);
1782 ldlm_interval_attach(node, lock);
1788 * Remove old lock from the pool before adding the lock with new
1789 * mode below in ->policy()
1791 ldlm_pool_del(&ns->ns_pool, lock);
1793 /* If this is a local resource, put it on the appropriate list. */
1794 if (ns_is_client(res->lr_namespace)) {
1795 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1796 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1798 /* This should never happen, because of the way the
1799 * server handles conversions. */
1800 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1804 ldlm_grant_lock(lock, &rpc_list);
1806 /* FIXME: completion handling not with ns_lock held ! */
1807 if (lock->l_completion_ast)
1808 lock->l_completion_ast(lock, 0, NULL);
1812 ldlm_processing_policy policy;
1813 policy = ldlm_processing_policy_table[res->lr_type];
1814 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1815 if (rc == LDLM_ITER_STOP) {
1816 lock->l_req_mode = old_mode;
1817 if (res->lr_type == LDLM_EXTENT)
1818 ldlm_extent_add_lock(res, lock);
1820 ldlm_granted_list_add_lock(lock, &prev);
1824 *flags |= LDLM_FL_BLOCK_GRANTED;
1828 unlock_res_and_lock(lock);
1831 ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1833 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1837 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1839 struct obd_device *obd = NULL;
1841 if (!((libcfs_debug | D_ERROR) & level))
1845 CDEBUG(level, " NULL LDLM lock\n");
1849 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1850 lock, lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1852 if (lock->l_conn_export != NULL)
1853 obd = lock->l_conn_export->exp_obd;
1854 if (lock->l_export && lock->l_export->exp_connection) {
1855 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1856 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1857 lock->l_remote_handle.cookie);
1858 } else if (obd == NULL) {
1859 CDEBUG(level, " Node: local\n");
1861 struct obd_import *imp = obd->u.cli.cl_import;
1862 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1863 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1864 lock->l_remote_handle.cookie);
1866 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1868 lock->l_resource->lr_name.name[0],
1869 lock->l_resource->lr_name.name[1],
1870 lock->l_resource->lr_name.name[2]);
1871 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1872 "write: %d flags: "LPX64"\n", ldlm_lockname[lock->l_req_mode],
1873 ldlm_lockname[lock->l_granted_mode],
1874 cfs_atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1876 if (lock->l_resource->lr_type == LDLM_EXTENT)
1877 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1878 " (req "LPU64"-"LPU64")\n",
1879 lock->l_policy_data.l_extent.start,
1880 lock->l_policy_data.l_extent.end,
1881 lock->l_req_extent.start, lock->l_req_extent.end);
1882 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1883 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1884 lock->l_policy_data.l_flock.pid,
1885 lock->l_policy_data.l_flock.start,
1886 lock->l_policy_data.l_flock.end);
1887 else if (lock->l_resource->lr_type == LDLM_IBITS)
1888 CDEBUG(level, " Bits: "LPX64"\n",
1889 lock->l_policy_data.l_inodebits.bits);
1892 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1894 struct ldlm_lock *lock;
1896 if (!((libcfs_debug | D_ERROR) & level))
1899 lock = ldlm_handle2lock(lockh);
1903 ldlm_lock_dump(D_OTHER, lock, 0);
1905 LDLM_LOCK_PUT(lock);
1908 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1909 struct libcfs_debug_msg_data *data, const char *fmt,
1913 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1915 va_start(args, fmt);
1917 if (lock->l_resource == NULL) {
1918 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1919 data->msg_fn, data->msg_line, fmt, args,
1920 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1921 "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" remote: "
1922 LPX64" expref: %d pid: %u timeout: %lu\n", lock,
1923 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1924 lock->l_readers, lock->l_writers,
1925 ldlm_lockname[lock->l_granted_mode],
1926 ldlm_lockname[lock->l_req_mode],
1927 lock->l_flags, lock->l_remote_handle.cookie,
1929 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
1930 lock->l_pid, lock->l_callback_timeout);
1935 switch (lock->l_resource->lr_type) {
1937 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1938 data->msg_fn, data->msg_line, fmt, args,
1939 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1940 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1941 "] (req "LPU64"->"LPU64") flags: "LPX64" remote: "LPX64
1942 " expref: %d pid: %u timeout %lu\n",
1943 lock->l_resource->lr_namespace->ns_name, lock,
1944 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1945 lock->l_readers, lock->l_writers,
1946 ldlm_lockname[lock->l_granted_mode],
1947 ldlm_lockname[lock->l_req_mode],
1948 lock->l_resource->lr_name.name[0],
1949 lock->l_resource->lr_name.name[1],
1950 cfs_atomic_read(&lock->l_resource->lr_refcount),
1951 ldlm_typename[lock->l_resource->lr_type],
1952 lock->l_policy_data.l_extent.start,
1953 lock->l_policy_data.l_extent.end,
1954 lock->l_req_extent.start, lock->l_req_extent.end,
1955 lock->l_flags, lock->l_remote_handle.cookie,
1957 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
1958 lock->l_pid, lock->l_callback_timeout);
1962 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1963 data->msg_fn, data->msg_line, fmt, args,
1964 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1965 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1966 "["LPU64"->"LPU64"] flags: "LPX64" remote: "LPX64
1967 " expref: %d pid: %u timeout: %lu\n",
1968 lock->l_resource->lr_namespace->ns_name, lock,
1969 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1970 lock->l_readers, lock->l_writers,
1971 ldlm_lockname[lock->l_granted_mode],
1972 ldlm_lockname[lock->l_req_mode],
1973 lock->l_resource->lr_name.name[0],
1974 lock->l_resource->lr_name.name[1],
1975 cfs_atomic_read(&lock->l_resource->lr_refcount),
1976 ldlm_typename[lock->l_resource->lr_type],
1977 lock->l_policy_data.l_flock.pid,
1978 lock->l_policy_data.l_flock.start,
1979 lock->l_policy_data.l_flock.end,
1980 lock->l_flags, lock->l_remote_handle.cookie,
1982 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
1983 lock->l_pid, lock->l_callback_timeout);
1987 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1988 data->msg_fn, data->msg_line, fmt, args,
1989 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1990 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1991 "flags: "LPX64" remote: "LPX64" expref: %d "
1992 "pid: %u timeout: %lu\n",
1993 lock->l_resource->lr_namespace->ns_name,
1994 lock, lock->l_handle.h_cookie,
1995 cfs_atomic_read (&lock->l_refc),
1996 lock->l_readers, lock->l_writers,
1997 ldlm_lockname[lock->l_granted_mode],
1998 ldlm_lockname[lock->l_req_mode],
1999 lock->l_resource->lr_name.name[0],
2000 lock->l_resource->lr_name.name[1],
2001 lock->l_policy_data.l_inodebits.bits,
2002 cfs_atomic_read(&lock->l_resource->lr_refcount),
2003 ldlm_typename[lock->l_resource->lr_type],
2004 lock->l_flags, lock->l_remote_handle.cookie,
2006 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2007 lock->l_pid, lock->l_callback_timeout);
2011 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2012 data->msg_fn, data->msg_line, fmt, args,
2013 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2014 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" "
2015 "remote: "LPX64" expref: %d pid: %u timeout %lu\n",
2016 lock->l_resource->lr_namespace->ns_name,
2017 lock, lock->l_handle.h_cookie,
2018 cfs_atomic_read (&lock->l_refc),
2019 lock->l_readers, lock->l_writers,
2020 ldlm_lockname[lock->l_granted_mode],
2021 ldlm_lockname[lock->l_req_mode],
2022 lock->l_resource->lr_name.name[0],
2023 lock->l_resource->lr_name.name[1],
2024 cfs_atomic_read(&lock->l_resource->lr_refcount),
2025 ldlm_typename[lock->l_resource->lr_type],
2026 lock->l_flags, lock->l_remote_handle.cookie,
2028 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2029 lock->l_pid, lock->l_callback_timeout);
2034 EXPORT_SYMBOL(_ldlm_lock_debug);