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 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
164 lu_ref_del(&res->lr_reference, "lock", lock);
165 ldlm_resource_putref(res);
166 lock->l_resource = NULL;
167 if (lock->l_export) {
168 class_export_lock_put(lock->l_export, lock);
169 lock->l_export = NULL;
172 if (lock->l_lvb_data != NULL)
173 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
175 ldlm_interval_free(ldlm_interval_detach(lock));
176 lu_ref_fini(&lock->l_reference);
177 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
184 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
187 if (!cfs_list_empty(&lock->l_lru)) {
188 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
190 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
191 cfs_list_del_init(&lock->l_lru);
192 if (lock->l_flags & LDLM_FL_SKIPPED)
193 lock->l_flags &= ~LDLM_FL_SKIPPED;
194 LASSERT(ns->ns_nr_unused > 0);
201 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
203 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
207 if (lock->l_ns_srv) {
208 LASSERT(cfs_list_empty(&lock->l_lru));
212 cfs_spin_lock(&ns->ns_lock);
213 rc = ldlm_lock_remove_from_lru_nolock(lock);
214 cfs_spin_unlock(&ns->ns_lock);
219 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
221 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
223 lock->l_last_used = cfs_time_current();
224 LASSERT(cfs_list_empty(&lock->l_lru));
225 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
226 cfs_list_add_tail(&lock->l_lru, &ns->ns_unused_list);
227 LASSERT(ns->ns_nr_unused >= 0);
231 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
233 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
236 cfs_spin_lock(&ns->ns_lock);
237 ldlm_lock_add_to_lru_nolock(lock);
238 cfs_spin_unlock(&ns->ns_lock);
242 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
244 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
247 if (lock->l_ns_srv) {
248 LASSERT(cfs_list_empty(&lock->l_lru));
253 cfs_spin_lock(&ns->ns_lock);
254 if (!cfs_list_empty(&lock->l_lru)) {
255 ldlm_lock_remove_from_lru_nolock(lock);
256 ldlm_lock_add_to_lru_nolock(lock);
258 cfs_spin_unlock(&ns->ns_lock);
262 /* This used to have a 'strict' flag, which recovery would use to mark an
263 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
264 * shall explain why it's gone: with the new hash table scheme, once you call
265 * ldlm_lock_destroy, you can never drop your final references on this lock.
266 * Because it's not in the hash table anymore. -phil */
267 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
271 if (lock->l_readers || lock->l_writers) {
272 LDLM_ERROR(lock, "lock still has references");
273 ldlm_lock_dump(D_ERROR, lock, 0);
277 if (!cfs_list_empty(&lock->l_res_link)) {
278 LDLM_ERROR(lock, "lock still on resource");
279 ldlm_lock_dump(D_ERROR, lock, 0);
283 if (lock->l_destroyed) {
284 LASSERT(cfs_list_empty(&lock->l_lru));
288 lock->l_destroyed = 1;
290 if (lock->l_export && lock->l_export->exp_lock_hash &&
291 !cfs_hlist_unhashed(&lock->l_exp_hash))
292 cfs_hash_del(lock->l_export->exp_lock_hash,
293 &lock->l_remote_handle, &lock->l_exp_hash);
295 ldlm_lock_remove_from_lru(lock);
296 class_handle_unhash(&lock->l_handle);
299 /* Wake anyone waiting for this lock */
300 /* FIXME: I should probably add yet another flag, instead of using
301 * l_export to only call this on clients */
303 class_export_put(lock->l_export);
304 lock->l_export = NULL;
305 if (lock->l_export && lock->l_completion_ast)
306 lock->l_completion_ast(lock, 0);
312 void ldlm_lock_destroy(struct ldlm_lock *lock)
316 lock_res_and_lock(lock);
317 first = ldlm_lock_destroy_internal(lock);
318 unlock_res_and_lock(lock);
320 /* drop reference from hashtable only for first destroy */
322 lu_ref_del(&lock->l_reference, "hash", lock);
323 LDLM_LOCK_RELEASE(lock);
328 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
332 first = ldlm_lock_destroy_internal(lock);
333 /* drop reference from hashtable only for first destroy */
335 lu_ref_del(&lock->l_reference, "hash", lock);
336 LDLM_LOCK_RELEASE(lock);
341 /* this is called by portals_handle2object with the handle lock taken */
342 static void lock_handle_addref(void *lock)
344 LDLM_LOCK_GET((struct ldlm_lock *)lock);
348 * usage: pass in a resource on which you have done ldlm_resource_get
349 * new lock will take over the refcount.
350 * returns: lock with refcount 2 - one for current caller and one for remote
352 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
354 struct ldlm_lock *lock;
357 if (resource == NULL)
360 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, CFS_ALLOC_IO);
364 cfs_spin_lock_init(&lock->l_lock);
365 lock->l_resource = resource;
366 lu_ref_add(&resource->lr_reference, "lock", lock);
368 cfs_atomic_set(&lock->l_refc, 2);
369 CFS_INIT_LIST_HEAD(&lock->l_res_link);
370 CFS_INIT_LIST_HEAD(&lock->l_lru);
371 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
372 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
373 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
374 CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
375 cfs_waitq_init(&lock->l_waitq);
376 lock->l_blocking_lock = NULL;
377 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
378 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
379 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
381 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
383 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
384 class_handle_hash(&lock->l_handle, lock_handle_addref);
386 lu_ref_init(&lock->l_reference);
387 lu_ref_add(&lock->l_reference, "hash", lock);
388 lock->l_callback_timeout = 0;
390 #if LUSTRE_TRACKS_LOCK_EXP_REFS
391 CFS_INIT_LIST_HEAD(&lock->l_exp_refs_link);
392 lock->l_exp_refs_nr = 0;
393 lock->l_exp_refs_target = NULL;
399 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
400 const struct ldlm_res_id *new_resid)
402 struct ldlm_resource *oldres = lock->l_resource;
403 struct ldlm_resource *newres;
407 LASSERT(ns_is_client(ns));
409 lock_res_and_lock(lock);
410 if (memcmp(new_resid, &lock->l_resource->lr_name,
411 sizeof(lock->l_resource->lr_name)) == 0) {
413 unlock_res_and_lock(lock);
417 LASSERT(new_resid->name[0] != 0);
419 /* This function assumes that the lock isn't on any lists */
420 LASSERT(cfs_list_empty(&lock->l_res_link));
422 type = oldres->lr_type;
423 unlock_res_and_lock(lock);
425 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
429 lu_ref_add(&newres->lr_reference, "lock", lock);
431 * To flip the lock from the old to the new resource, lock, oldres and
432 * newres have to be locked. Resource spin-locks are nested within
433 * lock->l_lock, and are taken in the memory address order to avoid
436 cfs_spin_lock(&lock->l_lock);
437 oldres = lock->l_resource;
438 if (oldres < newres) {
440 lock_res_nested(newres, LRT_NEW);
443 lock_res_nested(oldres, LRT_NEW);
445 LASSERT(memcmp(new_resid, &oldres->lr_name,
446 sizeof oldres->lr_name) != 0);
447 lock->l_resource = newres;
449 unlock_res_and_lock(lock);
451 /* ...and the flowers are still standing! */
452 lu_ref_del(&oldres->lr_reference, "lock", lock);
453 ldlm_resource_putref(oldres);
462 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
464 lockh->cookie = lock->l_handle.h_cookie;
467 /* if flags: atomically get the lock and set the flags.
468 * Return NULL if flag already set
471 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
474 struct ldlm_namespace *ns;
475 struct ldlm_lock *lock, *retval = NULL;
480 lock = class_handle2object(handle->cookie);
484 LASSERT(lock->l_resource != NULL);
485 ns = ldlm_lock_to_ns(lock);
488 lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
489 lock_res_and_lock(lock);
491 /* It's unlikely but possible that someone marked the lock as
492 * destroyed after we did handle2object on it */
493 if (lock->l_destroyed) {
494 unlock_res_and_lock(lock);
495 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
500 if (flags && (lock->l_flags & flags)) {
501 unlock_res_and_lock(lock);
507 lock->l_flags |= flags;
509 unlock_res_and_lock(lock);
516 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
518 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
519 /* INODEBITS_INTEROP: If the other side does not support
520 * inodebits, reply with a plain lock descriptor.
522 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
523 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
524 /* Make sure all the right bits are set in this lock we
525 are going to pass to client */
526 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
527 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
528 "Inappropriate inode lock bits during "
529 "conversion " LPU64 "\n",
530 lock->l_policy_data.l_inodebits.bits);
532 ldlm_res2desc(lock->l_resource, &desc->l_resource);
533 desc->l_resource.lr_type = LDLM_PLAIN;
535 /* Convert "new" lock mode to something old client can
537 if ((lock->l_req_mode == LCK_CR) ||
538 (lock->l_req_mode == LCK_CW))
539 desc->l_req_mode = LCK_PR;
541 desc->l_req_mode = lock->l_req_mode;
542 if ((lock->l_granted_mode == LCK_CR) ||
543 (lock->l_granted_mode == LCK_CW)) {
544 desc->l_granted_mode = LCK_PR;
546 /* We never grant PW/EX locks to clients */
547 LASSERT((lock->l_granted_mode != LCK_PW) &&
548 (lock->l_granted_mode != LCK_EX));
549 desc->l_granted_mode = lock->l_granted_mode;
552 /* We do not copy policy here, because there is no
553 policy for plain locks */
555 ldlm_res2desc(lock->l_resource, &desc->l_resource);
556 desc->l_req_mode = lock->l_req_mode;
557 desc->l_granted_mode = lock->l_granted_mode;
558 desc->l_policy_data = lock->l_policy_data;
562 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
563 cfs_list_t *work_list)
565 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
566 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
567 lock->l_flags |= LDLM_FL_AST_SENT;
568 /* If the enqueuing client said so, tell the AST recipient to
569 * discard dirty data, rather than writing back. */
570 if (new->l_flags & LDLM_AST_DISCARD_DATA)
571 lock->l_flags |= LDLM_FL_DISCARD_DATA;
572 LASSERT(cfs_list_empty(&lock->l_bl_ast));
573 cfs_list_add(&lock->l_bl_ast, work_list);
575 LASSERT(lock->l_blocking_lock == NULL);
576 lock->l_blocking_lock = LDLM_LOCK_GET(new);
580 void ldlm_add_cp_work_item(struct ldlm_lock *lock, cfs_list_t *work_list)
582 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
583 lock->l_flags |= LDLM_FL_CP_REQD;
584 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
585 LASSERT(cfs_list_empty(&lock->l_cp_ast));
586 cfs_list_add(&lock->l_cp_ast, work_list);
591 /* must be called with lr_lock held */
592 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
593 cfs_list_t *work_list)
596 check_res_locked(lock->l_resource);
598 ldlm_add_bl_work_item(lock, new, work_list);
600 ldlm_add_cp_work_item(lock, work_list);
604 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
606 struct ldlm_lock *lock;
608 lock = ldlm_handle2lock(lockh);
609 LASSERT(lock != NULL);
610 ldlm_lock_addref_internal(lock, mode);
614 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
616 ldlm_lock_remove_from_lru(lock);
617 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
619 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
621 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
623 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
626 lu_ref_add_atomic(&lock->l_reference, "user", lock);
627 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
631 * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
634 * \retval 0 success, lock was addref-ed
636 * \retval -EAGAIN lock is being canceled.
638 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
640 struct ldlm_lock *lock;
644 lock = ldlm_handle2lock(lockh);
646 lock_res_and_lock(lock);
647 if (lock->l_readers != 0 || lock->l_writers != 0 ||
648 !(lock->l_flags & LDLM_FL_CBPENDING)) {
649 ldlm_lock_addref_internal_nolock(lock, mode);
652 unlock_res_and_lock(lock);
658 /* only called for local locks */
659 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
661 lock_res_and_lock(lock);
662 ldlm_lock_addref_internal_nolock(lock, mode);
663 unlock_res_and_lock(lock);
666 /* only called in ldlm_flock_destroy and for local locks.
667 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
668 * * ldlm_lock_remove_from_lru() does nothing, it is safe
669 * * for ldlm_flock_destroy usage by dropping some code */
670 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
672 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
673 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
674 LASSERT(lock->l_readers > 0);
675 lu_ref_del(&lock->l_reference, "reader", lock);
678 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
679 LASSERT(lock->l_writers > 0);
680 lu_ref_del(&lock->l_reference, "writer", lock);
684 lu_ref_del(&lock->l_reference, "user", lock);
685 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
688 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
690 struct ldlm_namespace *ns;
693 lock_res_and_lock(lock);
695 ns = ldlm_lock_to_ns(lock);
697 ldlm_lock_decref_internal_nolock(lock, mode);
699 if (lock->l_flags & LDLM_FL_LOCAL &&
700 !lock->l_readers && !lock->l_writers) {
701 /* If this is a local lock on a server namespace and this was
702 * the last reference, cancel the lock. */
703 CDEBUG(D_INFO, "forcing cancel of local lock\n");
704 lock->l_flags |= LDLM_FL_CBPENDING;
707 if (!lock->l_readers && !lock->l_writers &&
708 (lock->l_flags & LDLM_FL_CBPENDING)) {
709 /* If we received a blocked AST and this was the last reference,
710 * run the callback. */
711 if (lock->l_ns_srv && lock->l_export)
712 CERROR("FL_CBPENDING set on non-local lock--just a "
715 LDLM_DEBUG(lock, "final decref done on cbpending lock");
717 LDLM_LOCK_GET(lock); /* dropped by bl thread */
718 ldlm_lock_remove_from_lru(lock);
719 unlock_res_and_lock(lock);
721 if (lock->l_flags & LDLM_FL_FAIL_LOC)
722 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
724 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
725 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
726 ldlm_handle_bl_callback(ns, NULL, lock);
727 } else if (ns_is_client(ns) &&
728 !lock->l_readers && !lock->l_writers &&
729 !(lock->l_flags & LDLM_FL_BL_AST)) {
730 /* If this is a client-side namespace and this was the last
731 * reference, put it on the LRU. */
732 ldlm_lock_add_to_lru(lock);
733 unlock_res_and_lock(lock);
735 if (lock->l_flags & LDLM_FL_FAIL_LOC)
736 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
738 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
739 * are not supported by the server, otherwise, it is done on
741 if (!exp_connect_cancelset(lock->l_conn_export) &&
742 !ns_connect_lru_resize(ns))
743 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
745 unlock_res_and_lock(lock);
751 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
753 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
754 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
755 ldlm_lock_decref_internal(lock, mode);
759 /* This will drop a lock reference and mark it for destruction, but will not
760 * necessarily cancel the lock before returning. */
761 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
763 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
766 LASSERT(lock != NULL);
768 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
769 lock_res_and_lock(lock);
770 lock->l_flags |= LDLM_FL_CBPENDING;
771 unlock_res_and_lock(lock);
772 ldlm_lock_decref_internal(lock, mode);
776 struct sl_insert_point {
777 cfs_list_t *res_link;
778 cfs_list_t *mode_link;
779 cfs_list_t *policy_link;
783 * search_granted_lock
786 * Finds a position to insert the new lock.
788 * queue [input]: the granted list where search acts on;
789 * req [input]: the lock whose position to be located;
790 * prev [output]: positions within 3 lists to insert @req to
794 * - ldlm_grant_lock_with_skiplist
796 static void search_granted_lock(cfs_list_t *queue,
797 struct ldlm_lock *req,
798 struct sl_insert_point *prev)
801 struct ldlm_lock *lock, *mode_end, *policy_end;
804 cfs_list_for_each(tmp, queue) {
805 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
807 mode_end = cfs_list_entry(lock->l_sl_mode.prev,
808 struct ldlm_lock, l_sl_mode);
810 if (lock->l_req_mode != req->l_req_mode) {
811 /* jump to last lock of mode group */
812 tmp = &mode_end->l_res_link;
816 /* suitable mode group is found */
817 if (lock->l_resource->lr_type == LDLM_PLAIN) {
818 /* insert point is last lock of the mode group */
819 prev->res_link = &mode_end->l_res_link;
820 prev->mode_link = &mode_end->l_sl_mode;
821 prev->policy_link = &req->l_sl_policy;
824 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
827 cfs_list_entry(lock->l_sl_policy.prev,
831 if (lock->l_policy_data.l_inodebits.bits ==
832 req->l_policy_data.l_inodebits.bits) {
833 /* insert point is last lock of
834 * the policy group */
836 &policy_end->l_res_link;
838 &policy_end->l_sl_mode;
840 &policy_end->l_sl_policy;
845 if (policy_end == mode_end)
846 /* done with mode group */
849 /* go to next policy group within mode group */
850 tmp = policy_end->l_res_link.next;
851 lock = cfs_list_entry(tmp, struct ldlm_lock,
853 } /* loop over policy groups within the mode group */
855 /* insert point is last lock of the mode group,
856 * new policy group is started */
857 prev->res_link = &mode_end->l_res_link;
858 prev->mode_link = &mode_end->l_sl_mode;
859 prev->policy_link = &req->l_sl_policy;
863 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
868 /* insert point is last lock on the queue,
869 * new mode group and new policy group are started */
870 prev->res_link = queue->prev;
871 prev->mode_link = &req->l_sl_mode;
872 prev->policy_link = &req->l_sl_policy;
877 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
878 struct sl_insert_point *prev)
880 struct ldlm_resource *res = lock->l_resource;
883 check_res_locked(res);
885 ldlm_resource_dump(D_INFO, res);
886 CDEBUG(D_OTHER, "About to add this lock:\n");
887 ldlm_lock_dump(D_OTHER, lock, 0);
889 if (lock->l_destroyed) {
890 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
894 LASSERT(cfs_list_empty(&lock->l_res_link));
895 LASSERT(cfs_list_empty(&lock->l_sl_mode));
896 LASSERT(cfs_list_empty(&lock->l_sl_policy));
898 cfs_list_add(&lock->l_res_link, prev->res_link);
899 cfs_list_add(&lock->l_sl_mode, prev->mode_link);
900 cfs_list_add(&lock->l_sl_policy, prev->policy_link);
905 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
907 struct sl_insert_point prev;
910 LASSERT(lock->l_req_mode == lock->l_granted_mode);
912 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
913 ldlm_granted_list_add_lock(lock, &prev);
918 * - ldlm_lock_enqueue
919 * - ldlm_reprocess_queue
920 * - ldlm_lock_convert
922 * must be called with lr_lock held
924 void ldlm_grant_lock(struct ldlm_lock *lock, cfs_list_t *work_list)
926 struct ldlm_resource *res = lock->l_resource;
929 check_res_locked(res);
931 lock->l_granted_mode = lock->l_req_mode;
932 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
933 ldlm_grant_lock_with_skiplist(lock);
934 else if (res->lr_type == LDLM_EXTENT)
935 ldlm_extent_add_lock(res, lock);
937 ldlm_resource_add_lock(res, &res->lr_granted, lock);
939 if (lock->l_granted_mode < res->lr_most_restr)
940 res->lr_most_restr = lock->l_granted_mode;
942 if (work_list && lock->l_completion_ast != NULL)
943 ldlm_add_ast_work_item(lock, NULL, work_list);
945 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
949 /* returns a referenced lock or NULL. See the flag descriptions below, in the
950 * comment above ldlm_lock_match */
951 static struct ldlm_lock *search_queue(cfs_list_t *queue,
953 ldlm_policy_data_t *policy,
954 struct ldlm_lock *old_lock,
955 int flags, int unref)
957 struct ldlm_lock *lock;
960 cfs_list_for_each(tmp, queue) {
963 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
965 if (lock == old_lock)
968 /* llite sometimes wants to match locks that will be
969 * canceled when their users drop, but we allow it to match
970 * if it passes in CBPENDING and the lock still has users.
971 * this is generally only going to be used by children
972 * whose parents already hold a lock so forward progress
973 * can still happen. */
974 if (lock->l_flags & LDLM_FL_CBPENDING &&
975 !(flags & LDLM_FL_CBPENDING))
977 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
978 lock->l_readers == 0 && lock->l_writers == 0)
981 if (!(lock->l_req_mode & *mode))
983 match = lock->l_req_mode;
985 if (lock->l_resource->lr_type == LDLM_EXTENT &&
986 (lock->l_policy_data.l_extent.start >
987 policy->l_extent.start ||
988 lock->l_policy_data.l_extent.end < policy->l_extent.end))
991 if (unlikely(match == LCK_GROUP) &&
992 lock->l_resource->lr_type == LDLM_EXTENT &&
993 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
996 /* We match if we have existing lock with same or wider set
998 if (lock->l_resource->lr_type == LDLM_IBITS &&
999 ((lock->l_policy_data.l_inodebits.bits &
1000 policy->l_inodebits.bits) !=
1001 policy->l_inodebits.bits))
1005 (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED)))
1008 if ((flags & LDLM_FL_LOCAL_ONLY) &&
1009 !(lock->l_flags & LDLM_FL_LOCAL))
1012 if (flags & LDLM_FL_TEST_LOCK) {
1013 LDLM_LOCK_GET(lock);
1014 ldlm_lock_touch_in_lru(lock);
1016 ldlm_lock_addref_internal_nolock(lock, match);
1025 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1027 lock->l_flags |= LDLM_FL_LVB_READY;
1028 cfs_waitq_signal(&lock->l_waitq);
1031 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1033 lock_res_and_lock(lock);
1034 ldlm_lock_allow_match_locked(lock);
1035 unlock_res_and_lock(lock);
1038 /* Can be called in two ways:
1040 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1041 * for a duplicate of.
1043 * Otherwise, all of the fields must be filled in, to match against.
1045 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1046 * server (ie, connh is NULL)
1047 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1048 * list will be considered
1049 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1050 * to be canceled can still be matched as long as they still have reader
1051 * or writer refernces
1052 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1053 * just tell us if we would have matched.
1055 * Returns 1 if it finds an already-existing lock that is compatible; in this
1056 * case, lockh is filled in with a addref()ed lock
1058 * we also check security context, if that failed we simply return 0 (to keep
1059 * caller code unchanged), the context failure will be discovered by caller
1062 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1063 const struct ldlm_res_id *res_id, ldlm_type_t type,
1064 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1065 struct lustre_handle *lockh, int unref)
1067 struct ldlm_resource *res;
1068 struct ldlm_lock *lock, *old_lock = NULL;
1073 old_lock = ldlm_handle2lock(lockh);
1076 ns = ldlm_lock_to_ns(old_lock);
1077 res_id = &old_lock->l_resource->lr_name;
1078 type = old_lock->l_resource->lr_type;
1079 mode = old_lock->l_req_mode;
1082 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1084 LASSERT(old_lock == NULL);
1088 LDLM_RESOURCE_ADDREF(res);
1091 lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1095 if (flags & LDLM_FL_BLOCK_GRANTED)
1097 lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1101 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1109 LDLM_RESOURCE_DELREF(res);
1110 ldlm_resource_putref(res);
1113 ldlm_lock2handle(lock, lockh);
1114 if ((flags & LDLM_FL_LVB_READY) &&
1115 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1116 struct l_wait_info lwi;
1117 if (lock->l_completion_ast) {
1118 int err = lock->l_completion_ast(lock,
1119 LDLM_FL_WAIT_NOREPROC,
1122 if (flags & LDLM_FL_TEST_LOCK)
1123 LDLM_LOCK_RELEASE(lock);
1125 ldlm_lock_decref_internal(lock,
1132 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1133 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1135 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1136 l_wait_event(lock->l_waitq,
1137 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1142 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1143 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1144 res_id->name[2] : policy->l_extent.start,
1145 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1146 res_id->name[3] : policy->l_extent.end);
1148 /* check user's security context */
1149 if (lock->l_conn_export &&
1150 sptlrpc_import_check_ctx(
1151 class_exp2cliimp(lock->l_conn_export))) {
1152 if (!(flags & LDLM_FL_TEST_LOCK))
1153 ldlm_lock_decref_internal(lock, mode);
1157 if (flags & LDLM_FL_TEST_LOCK)
1158 LDLM_LOCK_RELEASE(lock);
1160 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1161 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1162 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1163 type, mode, res_id->name[0], res_id->name[1],
1164 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1165 res_id->name[2] :policy->l_extent.start,
1166 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1167 res_id->name[3] : policy->l_extent.end);
1170 LDLM_LOCK_PUT(old_lock);
1172 return rc ? mode : 0;
1175 /* Returns a referenced lock */
1176 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1177 const struct ldlm_res_id *res_id,
1180 const struct ldlm_callback_suite *cbs,
1181 void *data, __u32 lvb_len)
1183 struct ldlm_lock *lock;
1184 struct ldlm_resource *res;
1187 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1191 lock = ldlm_lock_new(res);
1196 lock->l_req_mode = mode;
1197 lock->l_ast_data = data;
1198 lock->l_pid = cfs_curproc_pid();
1199 lock->l_ns_srv = ns_is_server(ns);
1201 lock->l_blocking_ast = cbs->lcs_blocking;
1202 lock->l_completion_ast = cbs->lcs_completion;
1203 lock->l_glimpse_ast = cbs->lcs_glimpse;
1204 lock->l_weigh_ast = cbs->lcs_weigh;
1207 lock->l_tree_node = NULL;
1208 /* if this is the extent lock, allocate the interval tree node */
1209 if (type == LDLM_EXTENT) {
1210 if (ldlm_interval_alloc(lock) == NULL)
1215 lock->l_lvb_len = lvb_len;
1216 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1217 if (lock->l_lvb_data == NULL)
1221 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1227 ldlm_lock_destroy(lock);
1228 LDLM_LOCK_RELEASE(lock);
1232 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1233 struct ldlm_lock **lockp,
1234 void *cookie, int *flags)
1236 struct ldlm_lock *lock = *lockp;
1237 struct ldlm_resource *res = lock->l_resource;
1238 int local = ns_is_client(ldlm_res_to_ns(res));
1239 ldlm_processing_policy policy;
1240 ldlm_error_t rc = ELDLM_OK;
1241 struct ldlm_interval *node = NULL;
1244 lock->l_last_activity = cfs_time_current_sec();
1245 /* policies are not executed on the client or during replay */
1246 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1247 && !local && ns->ns_policy) {
1248 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1250 if (rc == ELDLM_LOCK_REPLACED) {
1251 /* The lock that was returned has already been granted,
1252 * and placed into lockp. If it's not the same as the
1253 * one we passed in, then destroy the old one and our
1254 * work here is done. */
1255 if (lock != *lockp) {
1256 ldlm_lock_destroy(lock);
1257 LDLM_LOCK_RELEASE(lock);
1259 *flags |= LDLM_FL_LOCK_CHANGED;
1261 } else if (rc != ELDLM_OK ||
1262 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1263 ldlm_lock_destroy(lock);
1268 /* For a replaying lock, it might be already in granted list. So
1269 * unlinking the lock will cause the interval node to be freed, we
1270 * have to allocate the interval node early otherwise we can't regrant
1271 * this lock in the future. - jay */
1272 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1273 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1275 lock_res_and_lock(lock);
1276 if (local && lock->l_req_mode == lock->l_granted_mode) {
1277 /* The server returned a blocked lock, but it was granted
1278 * before we got a chance to actually enqueue it. We don't
1279 * need to do anything else. */
1280 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1281 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1282 GOTO(out, ELDLM_OK);
1285 ldlm_resource_unlink_lock(lock);
1286 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1288 ldlm_lock_destroy_nolock(lock);
1289 GOTO(out, rc = -ENOMEM);
1292 CFS_INIT_LIST_HEAD(&node->li_group);
1293 ldlm_interval_attach(node, lock);
1297 /* Some flags from the enqueue want to make it into the AST, via the
1298 * lock's l_flags. */
1299 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1301 /* This distinction between local lock trees is very important; a client
1302 * namespace only has information about locks taken by that client, and
1303 * thus doesn't have enough information to decide for itself if it can
1304 * be granted (below). In this case, we do exactly what the server
1305 * tells us to do, as dictated by the 'flags'.
1307 * We do exactly the same thing during recovery, when the server is
1308 * more or less trusting the clients not to lie.
1310 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1311 * granted/converting queues. */
1313 if (*flags & LDLM_FL_BLOCK_CONV)
1314 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1315 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1316 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1318 ldlm_grant_lock(lock, NULL);
1319 GOTO(out, ELDLM_OK);
1320 } else if (*flags & LDLM_FL_REPLAY) {
1321 if (*flags & LDLM_FL_BLOCK_CONV) {
1322 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1323 GOTO(out, ELDLM_OK);
1324 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1325 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1326 GOTO(out, ELDLM_OK);
1327 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1328 ldlm_grant_lock(lock, NULL);
1329 GOTO(out, ELDLM_OK);
1331 /* If no flags, fall through to normal enqueue path. */
1334 policy = ldlm_processing_policy_table[res->lr_type];
1335 policy(lock, flags, 1, &rc, NULL);
1338 unlock_res_and_lock(lock);
1340 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1344 /* Must be called with namespace taken: queue is waiting or converting. */
1345 int ldlm_reprocess_queue(struct ldlm_resource *res, cfs_list_t *queue,
1346 cfs_list_t *work_list)
1348 cfs_list_t *tmp, *pos;
1349 ldlm_processing_policy policy;
1351 int rc = LDLM_ITER_CONTINUE;
1355 check_res_locked(res);
1357 policy = ldlm_processing_policy_table[res->lr_type];
1360 cfs_list_for_each_safe(tmp, pos, queue) {
1361 struct ldlm_lock *pending;
1362 pending = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1364 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1367 rc = policy(pending, &flags, 0, &err, work_list);
1368 if (rc != LDLM_ITER_CONTINUE)
1375 /* Helper function for ldlm_run_ast_work().
1377 * Send an existing rpc set specified by @arg->set and then
1378 * destroy it. Create new one if @do_create flag is set. */
1380 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1384 ptlrpc_set_wait(arg->set);
1385 if (arg->type == LDLM_BL_CALLBACK)
1386 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1387 ptlrpc_set_destroy(arg->set);
1390 arg->set = ptlrpc_prep_set();
1396 ldlm_work_bl_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1398 struct ldlm_lock_desc d;
1399 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1403 /* nobody should touch l_bl_ast */
1404 lock_res_and_lock(lock);
1405 cfs_list_del_init(&lock->l_bl_ast);
1407 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1408 LASSERT(lock->l_bl_ast_run == 0);
1409 LASSERT(lock->l_blocking_lock);
1410 lock->l_bl_ast_run++;
1411 unlock_res_and_lock(lock);
1413 ldlm_lock2desc(lock->l_blocking_lock, &d);
1415 lock->l_blocking_ast(lock, &d, (void *)arg,
1417 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1418 lock->l_blocking_lock = NULL;
1419 LDLM_LOCK_RELEASE(lock);
1425 ldlm_work_cp_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1427 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock, l_cp_ast);
1428 ldlm_completion_callback completion_callback;
1432 /* It's possible to receive a completion AST before we've set
1433 * the l_completion_ast pointer: either because the AST arrived
1434 * before the reply, or simply because there's a small race
1435 * window between receiving the reply and finishing the local
1436 * enqueue. (bug 842)
1438 * This can't happen with the blocking_ast, however, because we
1439 * will never call the local blocking_ast until we drop our
1440 * reader/writer reference, which we won't do until we get the
1441 * reply and finish enqueueing. */
1443 /* nobody should touch l_cp_ast */
1444 lock_res_and_lock(lock);
1445 cfs_list_del_init(&lock->l_cp_ast);
1446 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1447 /* save l_completion_ast since it can be changed by
1448 * mds_intent_policy(), see bug 14225 */
1449 completion_callback = lock->l_completion_ast;
1450 lock->l_flags &= ~LDLM_FL_CP_REQD;
1451 unlock_res_and_lock(lock);
1453 if (completion_callback != NULL) {
1454 completion_callback(lock, 0, (void *)arg);
1457 LDLM_LOCK_RELEASE(lock);
1463 ldlm_work_revoke_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1465 struct ldlm_lock_desc desc;
1466 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1470 cfs_list_del_init(&lock->l_rk_ast);
1472 /* the desc just pretend to exclusive */
1473 ldlm_lock2desc(lock, &desc);
1474 desc.l_req_mode = LCK_EX;
1475 desc.l_granted_mode = 0;
1477 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1478 LDLM_LOCK_RELEASE(lock);
1483 int ldlm_run_ast_work(cfs_list_t *rpc_list, ldlm_desc_ast_t ast_type)
1485 struct ldlm_cb_set_arg arg;
1486 cfs_list_t *tmp, *pos;
1487 int (*work_ast_lock)(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg);
1491 if (cfs_list_empty(rpc_list))
1494 arg.set = ptlrpc_prep_set();
1495 if (NULL == arg.set)
1497 cfs_atomic_set(&arg.restart, 0);
1499 case LDLM_WORK_BL_AST:
1500 arg.type = LDLM_BL_CALLBACK;
1501 work_ast_lock = ldlm_work_bl_ast_lock;
1503 case LDLM_WORK_CP_AST:
1504 arg.type = LDLM_CP_CALLBACK;
1505 work_ast_lock = ldlm_work_cp_ast_lock;
1507 case LDLM_WORK_REVOKE_AST:
1508 arg.type = LDLM_BL_CALLBACK;
1509 work_ast_lock = ldlm_work_revoke_ast_lock;
1516 cfs_list_for_each_safe(tmp, pos, rpc_list) {
1517 ast_count += work_ast_lock(tmp, &arg);
1519 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1520 * and create a new set for requests that remained in
1522 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1523 ldlm_send_and_maybe_create_set(&arg, 1);
1529 ldlm_send_and_maybe_create_set(&arg, 0);
1531 /* In case when number of ASTs is multiply of
1532 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1533 * @arg.set must be destroyed here, otherwise we get
1534 * write memory leaking. */
1535 ptlrpc_set_destroy(arg.set);
1537 RETURN(cfs_atomic_read(&arg.restart) ? -ERESTART : 0);
1540 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1542 ldlm_reprocess_all(res);
1543 return LDLM_ITER_CONTINUE;
1546 static int ldlm_reprocess_res(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1547 cfs_hlist_node_t *hnode, void *arg)
1549 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1552 rc = reprocess_one_queue(res, arg);
1554 return rc == LDLM_ITER_STOP;
1557 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1562 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1563 ldlm_reprocess_res, NULL);
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(ldlm_res_to_ns(res))) {
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 = ldlm_res_to_ns(res);
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 int ldlm_cancel_locks_for_export_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1677 cfs_hlist_node_t *hnode, void *data)
1680 struct obd_export *exp = data;
1681 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
1682 struct ldlm_resource *res;
1684 res = ldlm_resource_getref(lock->l_resource);
1685 LDLM_LOCK_GET(lock);
1687 LDLM_DEBUG(lock, "export %p", exp);
1688 ldlm_res_lvbo_update(res, NULL, 1);
1689 ldlm_lock_cancel(lock);
1690 ldlm_reprocess_all(res);
1691 ldlm_resource_putref(res);
1692 LDLM_LOCK_RELEASE(lock);
1696 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1698 cfs_hash_for_each_empty(exp->exp_lock_hash,
1699 ldlm_cancel_locks_for_export_cb, exp);
1703 * Downgrade an exclusive lock.
1705 * A fast variant of ldlm_lock_convert for convertion of exclusive
1706 * locks. The convertion is always successful.
1708 * \param lock A lock to convert
1709 * \param new_mode new lock mode
1711 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1715 LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1716 LASSERT(new_mode == LCK_COS);
1718 lock_res_and_lock(lock);
1719 ldlm_resource_unlink_lock(lock);
1721 * Remove the lock from pool as it will be added again in
1722 * ldlm_grant_lock() called below.
1724 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
1726 lock->l_req_mode = new_mode;
1727 ldlm_grant_lock(lock, NULL);
1728 unlock_res_and_lock(lock);
1729 ldlm_reprocess_all(lock->l_resource);
1734 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1737 CFS_LIST_HEAD(rpc_list);
1738 struct ldlm_resource *res;
1739 struct ldlm_namespace *ns;
1742 struct sl_insert_point prev;
1744 struct ldlm_interval *node;
1747 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1748 *flags |= LDLM_FL_BLOCK_GRANTED;
1749 RETURN(lock->l_resource);
1752 /* I can't check the type of lock here because the bitlock of lock
1753 * is not held here, so do the allocation blindly. -jay */
1754 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1755 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1758 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1759 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1761 lock_res_and_lock(lock);
1763 res = lock->l_resource;
1764 ns = ldlm_res_to_ns(res);
1766 old_mode = lock->l_req_mode;
1767 lock->l_req_mode = new_mode;
1768 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1769 /* remember the lock position where the lock might be
1770 * added back to the granted list later and also
1771 * remember the join mode for skiplist fixing. */
1772 prev.res_link = lock->l_res_link.prev;
1773 prev.mode_link = lock->l_sl_mode.prev;
1774 prev.policy_link = lock->l_sl_policy.prev;
1775 ldlm_resource_unlink_lock(lock);
1777 ldlm_resource_unlink_lock(lock);
1778 if (res->lr_type == LDLM_EXTENT) {
1779 /* FIXME: ugly code, I have to attach the lock to a
1780 * interval node again since perhaps it will be granted
1782 CFS_INIT_LIST_HEAD(&node->li_group);
1783 ldlm_interval_attach(node, lock);
1789 * Remove old lock from the pool before adding the lock with new
1790 * mode below in ->policy()
1792 ldlm_pool_del(&ns->ns_pool, lock);
1794 /* If this is a local resource, put it on the appropriate list. */
1795 if (ns_is_client(ldlm_res_to_ns(res))) {
1796 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1797 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1799 /* This should never happen, because of the way the
1800 * server handles conversions. */
1801 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1805 ldlm_grant_lock(lock, &rpc_list);
1807 /* FIXME: completion handling not with lr_lock held ! */
1808 if (lock->l_completion_ast)
1809 lock->l_completion_ast(lock, 0, NULL);
1813 ldlm_processing_policy policy;
1814 policy = ldlm_processing_policy_table[res->lr_type];
1815 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1816 if (rc == LDLM_ITER_STOP) {
1817 lock->l_req_mode = old_mode;
1818 if (res->lr_type == LDLM_EXTENT)
1819 ldlm_extent_add_lock(res, lock);
1821 ldlm_granted_list_add_lock(lock, &prev);
1825 *flags |= LDLM_FL_BLOCK_GRANTED;
1829 unlock_res_and_lock(lock);
1832 ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1834 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1838 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1840 struct obd_device *obd = NULL;
1842 if (!((libcfs_debug | D_ERROR) & level))
1846 CDEBUG(level, " NULL LDLM lock\n");
1850 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1851 lock, lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1853 if (lock->l_conn_export != NULL)
1854 obd = lock->l_conn_export->exp_obd;
1855 if (lock->l_export && lock->l_export->exp_connection) {
1856 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1857 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1858 lock->l_remote_handle.cookie);
1859 } else if (obd == NULL) {
1860 CDEBUG(level, " Node: local\n");
1862 struct obd_import *imp = obd->u.cli.cl_import;
1863 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1864 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1865 lock->l_remote_handle.cookie);
1867 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1869 lock->l_resource->lr_name.name[0],
1870 lock->l_resource->lr_name.name[1],
1871 lock->l_resource->lr_name.name[2]);
1872 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1873 "write: %d flags: "LPX64"\n", ldlm_lockname[lock->l_req_mode],
1874 ldlm_lockname[lock->l_granted_mode],
1875 cfs_atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1877 if (lock->l_resource->lr_type == LDLM_EXTENT)
1878 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1879 " (req "LPU64"-"LPU64")\n",
1880 lock->l_policy_data.l_extent.start,
1881 lock->l_policy_data.l_extent.end,
1882 lock->l_req_extent.start, lock->l_req_extent.end);
1883 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1884 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1885 lock->l_policy_data.l_flock.pid,
1886 lock->l_policy_data.l_flock.start,
1887 lock->l_policy_data.l_flock.end);
1888 else if (lock->l_resource->lr_type == LDLM_IBITS)
1889 CDEBUG(level, " Bits: "LPX64"\n",
1890 lock->l_policy_data.l_inodebits.bits);
1893 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1895 struct ldlm_lock *lock;
1897 if (!((libcfs_debug | D_ERROR) & level))
1900 lock = ldlm_handle2lock(lockh);
1904 ldlm_lock_dump(D_OTHER, lock, 0);
1906 LDLM_LOCK_PUT(lock);
1909 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1910 struct libcfs_debug_msg_data *data, const char *fmt,
1914 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1916 va_start(args, fmt);
1918 if (lock->l_resource == NULL) {
1919 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1920 data->msg_fn, data->msg_line, fmt, args,
1921 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1922 "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" remote: "
1923 LPX64" expref: %d pid: %u timeout: %lu\n", lock,
1924 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1925 lock->l_readers, lock->l_writers,
1926 ldlm_lockname[lock->l_granted_mode],
1927 ldlm_lockname[lock->l_req_mode],
1928 lock->l_flags, lock->l_remote_handle.cookie,
1930 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
1931 lock->l_pid, lock->l_callback_timeout);
1936 switch (lock->l_resource->lr_type) {
1938 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1939 data->msg_fn, data->msg_line, fmt, args,
1940 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1941 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1942 "] (req "LPU64"->"LPU64") flags: "LPX64" remote: "LPX64
1943 " expref: %d pid: %u timeout %lu\n",
1944 ldlm_lock_to_ns_name(lock), lock,
1945 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1946 lock->l_readers, lock->l_writers,
1947 ldlm_lockname[lock->l_granted_mode],
1948 ldlm_lockname[lock->l_req_mode],
1949 lock->l_resource->lr_name.name[0],
1950 lock->l_resource->lr_name.name[1],
1951 cfs_atomic_read(&lock->l_resource->lr_refcount),
1952 ldlm_typename[lock->l_resource->lr_type],
1953 lock->l_policy_data.l_extent.start,
1954 lock->l_policy_data.l_extent.end,
1955 lock->l_req_extent.start, lock->l_req_extent.end,
1956 lock->l_flags, lock->l_remote_handle.cookie,
1958 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
1959 lock->l_pid, lock->l_callback_timeout);
1963 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1964 data->msg_fn, data->msg_line, fmt, args,
1965 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1966 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1967 "["LPU64"->"LPU64"] flags: "LPX64" remote: "LPX64
1968 " expref: %d pid: %u timeout: %lu\n",
1969 ldlm_lock_to_ns_name(lock), lock,
1970 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1971 lock->l_readers, lock->l_writers,
1972 ldlm_lockname[lock->l_granted_mode],
1973 ldlm_lockname[lock->l_req_mode],
1974 lock->l_resource->lr_name.name[0],
1975 lock->l_resource->lr_name.name[1],
1976 cfs_atomic_read(&lock->l_resource->lr_refcount),
1977 ldlm_typename[lock->l_resource->lr_type],
1978 lock->l_policy_data.l_flock.pid,
1979 lock->l_policy_data.l_flock.start,
1980 lock->l_policy_data.l_flock.end,
1981 lock->l_flags, lock->l_remote_handle.cookie,
1983 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
1984 lock->l_pid, lock->l_callback_timeout);
1988 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1989 data->msg_fn, data->msg_line, fmt, args,
1990 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1991 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1992 "flags: "LPX64" remote: "LPX64" expref: %d "
1993 "pid: %u timeout: %lu\n",
1994 ldlm_lock_to_ns_name(lock),
1995 lock, lock->l_handle.h_cookie,
1996 cfs_atomic_read (&lock->l_refc),
1997 lock->l_readers, lock->l_writers,
1998 ldlm_lockname[lock->l_granted_mode],
1999 ldlm_lockname[lock->l_req_mode],
2000 lock->l_resource->lr_name.name[0],
2001 lock->l_resource->lr_name.name[1],
2002 lock->l_policy_data.l_inodebits.bits,
2003 cfs_atomic_read(&lock->l_resource->lr_refcount),
2004 ldlm_typename[lock->l_resource->lr_type],
2005 lock->l_flags, lock->l_remote_handle.cookie,
2007 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2008 lock->l_pid, lock->l_callback_timeout);
2012 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2013 data->msg_fn, data->msg_line, fmt, args,
2014 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2015 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" "
2016 "remote: "LPX64" expref: %d pid: %u timeout %lu\n",
2017 ldlm_lock_to_ns_name(lock),
2018 lock, lock->l_handle.h_cookie,
2019 cfs_atomic_read (&lock->l_refc),
2020 lock->l_readers, lock->l_writers,
2021 ldlm_lockname[lock->l_granted_mode],
2022 ldlm_lockname[lock->l_req_mode],
2023 lock->l_resource->lr_name.name[0],
2024 lock->l_resource->lr_name.name[1],
2025 cfs_atomic_read(&lock->l_resource->lr_refcount),
2026 ldlm_typename[lock->l_resource->lr_type],
2027 lock->l_flags, lock->l_remote_handle.cookie,
2029 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2030 lock->l_pid, lock->l_callback_timeout);
2035 EXPORT_SYMBOL(_ldlm_lock_debug);