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_lock *lock;
479 lock = class_handle2object(handle->cookie);
483 /* It's unlikely but possible that someone marked the lock as
484 * destroyed after we did handle2object on it */
485 if (flags == 0 && !lock->l_destroyed) {
486 lu_ref_add(&lock->l_reference, "handle", cfs_current());
490 lock_res_and_lock(lock);
492 LASSERT(lock->l_resource != NULL);
494 lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
495 if (unlikely(lock->l_destroyed)) {
496 unlock_res_and_lock(lock);
497 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
502 if (flags && (lock->l_flags & flags)) {
503 unlock_res_and_lock(lock);
509 lock->l_flags |= flags;
511 unlock_res_and_lock(lock);
515 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
517 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
518 /* INODEBITS_INTEROP: If the other side does not support
519 * inodebits, reply with a plain lock descriptor.
521 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
522 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
523 /* Make sure all the right bits are set in this lock we
524 are going to pass to client */
525 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
526 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
527 "Inappropriate inode lock bits during "
528 "conversion " LPU64 "\n",
529 lock->l_policy_data.l_inodebits.bits);
531 ldlm_res2desc(lock->l_resource, &desc->l_resource);
532 desc->l_resource.lr_type = LDLM_PLAIN;
534 /* Convert "new" lock mode to something old client can
536 if ((lock->l_req_mode == LCK_CR) ||
537 (lock->l_req_mode == LCK_CW))
538 desc->l_req_mode = LCK_PR;
540 desc->l_req_mode = lock->l_req_mode;
541 if ((lock->l_granted_mode == LCK_CR) ||
542 (lock->l_granted_mode == LCK_CW)) {
543 desc->l_granted_mode = LCK_PR;
545 /* We never grant PW/EX locks to clients */
546 LASSERT((lock->l_granted_mode != LCK_PW) &&
547 (lock->l_granted_mode != LCK_EX));
548 desc->l_granted_mode = lock->l_granted_mode;
551 /* We do not copy policy here, because there is no
552 policy for plain locks */
554 ldlm_res2desc(lock->l_resource, &desc->l_resource);
555 desc->l_req_mode = lock->l_req_mode;
556 desc->l_granted_mode = lock->l_granted_mode;
557 desc->l_policy_data = lock->l_policy_data;
561 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
562 cfs_list_t *work_list)
564 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
565 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
566 lock->l_flags |= LDLM_FL_AST_SENT;
567 /* If the enqueuing client said so, tell the AST recipient to
568 * discard dirty data, rather than writing back. */
569 if (new->l_flags & LDLM_AST_DISCARD_DATA)
570 lock->l_flags |= LDLM_FL_DISCARD_DATA;
571 LASSERT(cfs_list_empty(&lock->l_bl_ast));
572 cfs_list_add(&lock->l_bl_ast, work_list);
574 LASSERT(lock->l_blocking_lock == NULL);
575 lock->l_blocking_lock = LDLM_LOCK_GET(new);
579 void ldlm_add_cp_work_item(struct ldlm_lock *lock, cfs_list_t *work_list)
581 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
582 lock->l_flags |= LDLM_FL_CP_REQD;
583 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
584 LASSERT(cfs_list_empty(&lock->l_cp_ast));
585 cfs_list_add(&lock->l_cp_ast, work_list);
590 /* must be called with lr_lock held */
591 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
592 cfs_list_t *work_list)
595 check_res_locked(lock->l_resource);
597 ldlm_add_bl_work_item(lock, new, work_list);
599 ldlm_add_cp_work_item(lock, work_list);
603 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
605 struct ldlm_lock *lock;
607 lock = ldlm_handle2lock(lockh);
608 LASSERT(lock != NULL);
609 ldlm_lock_addref_internal(lock, mode);
613 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
615 ldlm_lock_remove_from_lru(lock);
616 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
618 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
620 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
622 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
625 lu_ref_add_atomic(&lock->l_reference, "user", lock);
626 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
630 * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
633 * \retval 0 success, lock was addref-ed
635 * \retval -EAGAIN lock is being canceled.
637 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
639 struct ldlm_lock *lock;
643 lock = ldlm_handle2lock(lockh);
645 lock_res_and_lock(lock);
646 if (lock->l_readers != 0 || lock->l_writers != 0 ||
647 !(lock->l_flags & LDLM_FL_CBPENDING)) {
648 ldlm_lock_addref_internal_nolock(lock, mode);
651 unlock_res_and_lock(lock);
657 /* only called for local locks */
658 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
660 lock_res_and_lock(lock);
661 ldlm_lock_addref_internal_nolock(lock, mode);
662 unlock_res_and_lock(lock);
665 /* only called in ldlm_flock_destroy and for local locks.
666 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
667 * * ldlm_lock_remove_from_lru() does nothing, it is safe
668 * * for ldlm_flock_destroy usage by dropping some code */
669 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
671 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
672 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
673 LASSERT(lock->l_readers > 0);
674 lu_ref_del(&lock->l_reference, "reader", lock);
677 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
678 LASSERT(lock->l_writers > 0);
679 lu_ref_del(&lock->l_reference, "writer", lock);
683 lu_ref_del(&lock->l_reference, "user", lock);
684 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
687 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
689 struct ldlm_namespace *ns;
692 lock_res_and_lock(lock);
694 ns = ldlm_lock_to_ns(lock);
696 ldlm_lock_decref_internal_nolock(lock, mode);
698 if (lock->l_flags & LDLM_FL_LOCAL &&
699 !lock->l_readers && !lock->l_writers) {
700 /* If this is a local lock on a server namespace and this was
701 * the last reference, cancel the lock. */
702 CDEBUG(D_INFO, "forcing cancel of local lock\n");
703 lock->l_flags |= LDLM_FL_CBPENDING;
706 if (!lock->l_readers && !lock->l_writers &&
707 (lock->l_flags & LDLM_FL_CBPENDING)) {
708 /* If we received a blocked AST and this was the last reference,
709 * run the callback. */
710 if (lock->l_ns_srv && lock->l_export)
711 CERROR("FL_CBPENDING set on non-local lock--just a "
714 LDLM_DEBUG(lock, "final decref done on cbpending lock");
716 LDLM_LOCK_GET(lock); /* dropped by bl thread */
717 ldlm_lock_remove_from_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 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
724 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
725 ldlm_handle_bl_callback(ns, NULL, lock);
726 } else if (ns_is_client(ns) &&
727 !lock->l_readers && !lock->l_writers &&
728 !(lock->l_flags & LDLM_FL_BL_AST)) {
729 /* If this is a client-side namespace and this was the last
730 * reference, put it on the LRU. */
731 ldlm_lock_add_to_lru(lock);
732 unlock_res_and_lock(lock);
734 if (lock->l_flags & LDLM_FL_FAIL_LOC)
735 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
737 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
738 * are not supported by the server, otherwise, it is done on
740 if (!exp_connect_cancelset(lock->l_conn_export) &&
741 !ns_connect_lru_resize(ns))
742 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
744 unlock_res_and_lock(lock);
750 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
752 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
753 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
754 ldlm_lock_decref_internal(lock, mode);
758 /* This will drop a lock reference and mark it for destruction, but will not
759 * necessarily cancel the lock before returning. */
760 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
762 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
765 LASSERT(lock != NULL);
767 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
768 lock_res_and_lock(lock);
769 lock->l_flags |= LDLM_FL_CBPENDING;
770 unlock_res_and_lock(lock);
771 ldlm_lock_decref_internal(lock, mode);
775 struct sl_insert_point {
776 cfs_list_t *res_link;
777 cfs_list_t *mode_link;
778 cfs_list_t *policy_link;
782 * search_granted_lock
785 * Finds a position to insert the new lock.
787 * queue [input]: the granted list where search acts on;
788 * req [input]: the lock whose position to be located;
789 * prev [output]: positions within 3 lists to insert @req to
793 * - ldlm_grant_lock_with_skiplist
795 static void search_granted_lock(cfs_list_t *queue,
796 struct ldlm_lock *req,
797 struct sl_insert_point *prev)
800 struct ldlm_lock *lock, *mode_end, *policy_end;
803 cfs_list_for_each(tmp, queue) {
804 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
806 mode_end = cfs_list_entry(lock->l_sl_mode.prev,
807 struct ldlm_lock, l_sl_mode);
809 if (lock->l_req_mode != req->l_req_mode) {
810 /* jump to last lock of mode group */
811 tmp = &mode_end->l_res_link;
815 /* suitable mode group is found */
816 if (lock->l_resource->lr_type == LDLM_PLAIN) {
817 /* insert point is last lock of the mode group */
818 prev->res_link = &mode_end->l_res_link;
819 prev->mode_link = &mode_end->l_sl_mode;
820 prev->policy_link = &req->l_sl_policy;
823 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
826 cfs_list_entry(lock->l_sl_policy.prev,
830 if (lock->l_policy_data.l_inodebits.bits ==
831 req->l_policy_data.l_inodebits.bits) {
832 /* insert point is last lock of
833 * the policy group */
835 &policy_end->l_res_link;
837 &policy_end->l_sl_mode;
839 &policy_end->l_sl_policy;
844 if (policy_end == mode_end)
845 /* done with mode group */
848 /* go to next policy group within mode group */
849 tmp = policy_end->l_res_link.next;
850 lock = cfs_list_entry(tmp, struct ldlm_lock,
852 } /* loop over policy groups within the mode group */
854 /* insert point is last lock of the mode group,
855 * new policy group is started */
856 prev->res_link = &mode_end->l_res_link;
857 prev->mode_link = &mode_end->l_sl_mode;
858 prev->policy_link = &req->l_sl_policy;
862 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
867 /* insert point is last lock on the queue,
868 * new mode group and new policy group are started */
869 prev->res_link = queue->prev;
870 prev->mode_link = &req->l_sl_mode;
871 prev->policy_link = &req->l_sl_policy;
876 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
877 struct sl_insert_point *prev)
879 struct ldlm_resource *res = lock->l_resource;
882 check_res_locked(res);
884 ldlm_resource_dump(D_INFO, res);
885 CDEBUG(D_OTHER, "About to add this lock:\n");
886 ldlm_lock_dump(D_OTHER, lock, 0);
888 if (lock->l_destroyed) {
889 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
893 LASSERT(cfs_list_empty(&lock->l_res_link));
894 LASSERT(cfs_list_empty(&lock->l_sl_mode));
895 LASSERT(cfs_list_empty(&lock->l_sl_policy));
897 cfs_list_add(&lock->l_res_link, prev->res_link);
898 cfs_list_add(&lock->l_sl_mode, prev->mode_link);
899 cfs_list_add(&lock->l_sl_policy, prev->policy_link);
904 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
906 struct sl_insert_point prev;
909 LASSERT(lock->l_req_mode == lock->l_granted_mode);
911 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
912 ldlm_granted_list_add_lock(lock, &prev);
917 * - ldlm_lock_enqueue
918 * - ldlm_reprocess_queue
919 * - ldlm_lock_convert
921 * must be called with lr_lock held
923 void ldlm_grant_lock(struct ldlm_lock *lock, cfs_list_t *work_list)
925 struct ldlm_resource *res = lock->l_resource;
928 check_res_locked(res);
930 lock->l_granted_mode = lock->l_req_mode;
931 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
932 ldlm_grant_lock_with_skiplist(lock);
933 else if (res->lr_type == LDLM_EXTENT)
934 ldlm_extent_add_lock(res, lock);
936 ldlm_resource_add_lock(res, &res->lr_granted, lock);
938 if (lock->l_granted_mode < res->lr_most_restr)
939 res->lr_most_restr = lock->l_granted_mode;
941 if (work_list && lock->l_completion_ast != NULL)
942 ldlm_add_ast_work_item(lock, NULL, work_list);
944 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
948 /* returns a referenced lock or NULL. See the flag descriptions below, in the
949 * comment above ldlm_lock_match */
950 static struct ldlm_lock *search_queue(cfs_list_t *queue,
952 ldlm_policy_data_t *policy,
953 struct ldlm_lock *old_lock,
954 int flags, int unref)
956 struct ldlm_lock *lock;
959 cfs_list_for_each(tmp, queue) {
962 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
964 if (lock == old_lock)
967 /* llite sometimes wants to match locks that will be
968 * canceled when their users drop, but we allow it to match
969 * if it passes in CBPENDING and the lock still has users.
970 * this is generally only going to be used by children
971 * whose parents already hold a lock so forward progress
972 * can still happen. */
973 if (lock->l_flags & LDLM_FL_CBPENDING &&
974 !(flags & LDLM_FL_CBPENDING))
976 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
977 lock->l_readers == 0 && lock->l_writers == 0)
980 if (!(lock->l_req_mode & *mode))
982 match = lock->l_req_mode;
984 if (lock->l_resource->lr_type == LDLM_EXTENT &&
985 (lock->l_policy_data.l_extent.start >
986 policy->l_extent.start ||
987 lock->l_policy_data.l_extent.end < policy->l_extent.end))
990 if (unlikely(match == LCK_GROUP) &&
991 lock->l_resource->lr_type == LDLM_EXTENT &&
992 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
995 /* We match if we have existing lock with same or wider set
997 if (lock->l_resource->lr_type == LDLM_IBITS &&
998 ((lock->l_policy_data.l_inodebits.bits &
999 policy->l_inodebits.bits) !=
1000 policy->l_inodebits.bits))
1004 (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED)))
1007 if ((flags & LDLM_FL_LOCAL_ONLY) &&
1008 !(lock->l_flags & LDLM_FL_LOCAL))
1011 if (flags & LDLM_FL_TEST_LOCK) {
1012 LDLM_LOCK_GET(lock);
1013 ldlm_lock_touch_in_lru(lock);
1015 ldlm_lock_addref_internal_nolock(lock, match);
1024 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1026 lock->l_flags |= LDLM_FL_LVB_READY;
1027 cfs_waitq_signal(&lock->l_waitq);
1030 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1032 lock_res_and_lock(lock);
1033 ldlm_lock_allow_match_locked(lock);
1034 unlock_res_and_lock(lock);
1037 /* Can be called in two ways:
1039 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1040 * for a duplicate of.
1042 * Otherwise, all of the fields must be filled in, to match against.
1044 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1045 * server (ie, connh is NULL)
1046 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1047 * list will be considered
1048 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1049 * to be canceled can still be matched as long as they still have reader
1050 * or writer refernces
1051 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1052 * just tell us if we would have matched.
1054 * Returns 1 if it finds an already-existing lock that is compatible; in this
1055 * case, lockh is filled in with a addref()ed lock
1057 * we also check security context, if that failed we simply return 0 (to keep
1058 * caller code unchanged), the context failure will be discovered by caller
1061 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1062 const struct ldlm_res_id *res_id, ldlm_type_t type,
1063 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1064 struct lustre_handle *lockh, int unref)
1066 struct ldlm_resource *res;
1067 struct ldlm_lock *lock, *old_lock = NULL;
1072 old_lock = ldlm_handle2lock(lockh);
1075 ns = ldlm_lock_to_ns(old_lock);
1076 res_id = &old_lock->l_resource->lr_name;
1077 type = old_lock->l_resource->lr_type;
1078 mode = old_lock->l_req_mode;
1081 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1083 LASSERT(old_lock == NULL);
1087 LDLM_RESOURCE_ADDREF(res);
1090 lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1094 if (flags & LDLM_FL_BLOCK_GRANTED)
1096 lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1100 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1108 LDLM_RESOURCE_DELREF(res);
1109 ldlm_resource_putref(res);
1112 ldlm_lock2handle(lock, lockh);
1113 if ((flags & LDLM_FL_LVB_READY) &&
1114 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1115 struct l_wait_info lwi;
1116 if (lock->l_completion_ast) {
1117 int err = lock->l_completion_ast(lock,
1118 LDLM_FL_WAIT_NOREPROC,
1121 if (flags & LDLM_FL_TEST_LOCK)
1122 LDLM_LOCK_RELEASE(lock);
1124 ldlm_lock_decref_internal(lock,
1131 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1132 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1134 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1135 l_wait_event(lock->l_waitq,
1136 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1141 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1142 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1143 res_id->name[2] : policy->l_extent.start,
1144 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1145 res_id->name[3] : policy->l_extent.end);
1147 /* check user's security context */
1148 if (lock->l_conn_export &&
1149 sptlrpc_import_check_ctx(
1150 class_exp2cliimp(lock->l_conn_export))) {
1151 if (!(flags & LDLM_FL_TEST_LOCK))
1152 ldlm_lock_decref_internal(lock, mode);
1156 if (flags & LDLM_FL_TEST_LOCK)
1157 LDLM_LOCK_RELEASE(lock);
1159 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1160 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1161 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1162 type, mode, res_id->name[0], res_id->name[1],
1163 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1164 res_id->name[2] :policy->l_extent.start,
1165 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1166 res_id->name[3] : policy->l_extent.end);
1169 LDLM_LOCK_PUT(old_lock);
1171 return rc ? mode : 0;
1174 /* Returns a referenced lock */
1175 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1176 const struct ldlm_res_id *res_id,
1179 const struct ldlm_callback_suite *cbs,
1180 void *data, __u32 lvb_len)
1182 struct ldlm_lock *lock;
1183 struct ldlm_resource *res;
1186 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1190 lock = ldlm_lock_new(res);
1195 lock->l_req_mode = mode;
1196 lock->l_ast_data = data;
1197 lock->l_pid = cfs_curproc_pid();
1198 lock->l_ns_srv = ns_is_server(ns);
1200 lock->l_blocking_ast = cbs->lcs_blocking;
1201 lock->l_completion_ast = cbs->lcs_completion;
1202 lock->l_glimpse_ast = cbs->lcs_glimpse;
1203 lock->l_weigh_ast = cbs->lcs_weigh;
1206 lock->l_tree_node = NULL;
1207 /* if this is the extent lock, allocate the interval tree node */
1208 if (type == LDLM_EXTENT) {
1209 if (ldlm_interval_alloc(lock) == NULL)
1214 lock->l_lvb_len = lvb_len;
1215 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1216 if (lock->l_lvb_data == NULL)
1220 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1226 ldlm_lock_destroy(lock);
1227 LDLM_LOCK_RELEASE(lock);
1231 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1232 struct ldlm_lock **lockp,
1233 void *cookie, int *flags)
1235 struct ldlm_lock *lock = *lockp;
1236 struct ldlm_resource *res = lock->l_resource;
1237 int local = ns_is_client(ldlm_res_to_ns(res));
1238 ldlm_processing_policy policy;
1239 ldlm_error_t rc = ELDLM_OK;
1240 struct ldlm_interval *node = NULL;
1243 lock->l_last_activity = cfs_time_current_sec();
1244 /* policies are not executed on the client or during replay */
1245 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1246 && !local && ns->ns_policy) {
1247 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1249 if (rc == ELDLM_LOCK_REPLACED) {
1250 /* The lock that was returned has already been granted,
1251 * and placed into lockp. If it's not the same as the
1252 * one we passed in, then destroy the old one and our
1253 * work here is done. */
1254 if (lock != *lockp) {
1255 ldlm_lock_destroy(lock);
1256 LDLM_LOCK_RELEASE(lock);
1258 *flags |= LDLM_FL_LOCK_CHANGED;
1260 } else if (rc != ELDLM_OK ||
1261 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1262 ldlm_lock_destroy(lock);
1267 /* For a replaying lock, it might be already in granted list. So
1268 * unlinking the lock will cause the interval node to be freed, we
1269 * have to allocate the interval node early otherwise we can't regrant
1270 * this lock in the future. - jay */
1271 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1272 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1274 lock_res_and_lock(lock);
1275 if (local && lock->l_req_mode == lock->l_granted_mode) {
1276 /* The server returned a blocked lock, but it was granted
1277 * before we got a chance to actually enqueue it. We don't
1278 * need to do anything else. */
1279 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1280 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1281 GOTO(out, ELDLM_OK);
1284 ldlm_resource_unlink_lock(lock);
1285 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1287 ldlm_lock_destroy_nolock(lock);
1288 GOTO(out, rc = -ENOMEM);
1291 CFS_INIT_LIST_HEAD(&node->li_group);
1292 ldlm_interval_attach(node, lock);
1296 /* Some flags from the enqueue want to make it into the AST, via the
1297 * lock's l_flags. */
1298 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1300 /* This distinction between local lock trees is very important; a client
1301 * namespace only has information about locks taken by that client, and
1302 * thus doesn't have enough information to decide for itself if it can
1303 * be granted (below). In this case, we do exactly what the server
1304 * tells us to do, as dictated by the 'flags'.
1306 * We do exactly the same thing during recovery, when the server is
1307 * more or less trusting the clients not to lie.
1309 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1310 * granted/converting queues. */
1312 if (*flags & LDLM_FL_BLOCK_CONV)
1313 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1314 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1315 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1317 ldlm_grant_lock(lock, NULL);
1318 GOTO(out, ELDLM_OK);
1319 } else if (*flags & LDLM_FL_REPLAY) {
1320 if (*flags & LDLM_FL_BLOCK_CONV) {
1321 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1322 GOTO(out, ELDLM_OK);
1323 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1324 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1325 GOTO(out, ELDLM_OK);
1326 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1327 ldlm_grant_lock(lock, NULL);
1328 GOTO(out, ELDLM_OK);
1330 /* If no flags, fall through to normal enqueue path. */
1333 policy = ldlm_processing_policy_table[res->lr_type];
1334 policy(lock, flags, 1, &rc, NULL);
1337 unlock_res_and_lock(lock);
1339 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1343 /* Must be called with namespace taken: queue is waiting or converting. */
1344 int ldlm_reprocess_queue(struct ldlm_resource *res, cfs_list_t *queue,
1345 cfs_list_t *work_list)
1347 cfs_list_t *tmp, *pos;
1348 ldlm_processing_policy policy;
1350 int rc = LDLM_ITER_CONTINUE;
1354 check_res_locked(res);
1356 policy = ldlm_processing_policy_table[res->lr_type];
1359 cfs_list_for_each_safe(tmp, pos, queue) {
1360 struct ldlm_lock *pending;
1361 pending = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1363 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1366 rc = policy(pending, &flags, 0, &err, work_list);
1367 if (rc != LDLM_ITER_CONTINUE)
1374 /* Helper function for ldlm_run_ast_work().
1376 * Send an existing rpc set specified by @arg->set and then
1377 * destroy it. Create new one if @do_create flag is set. */
1379 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1383 ptlrpc_set_wait(arg->set);
1384 if (arg->type == LDLM_BL_CALLBACK)
1385 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1386 ptlrpc_set_destroy(arg->set);
1389 arg->set = ptlrpc_prep_set();
1395 ldlm_work_bl_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1397 struct ldlm_lock_desc d;
1398 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1402 /* nobody should touch l_bl_ast */
1403 lock_res_and_lock(lock);
1404 cfs_list_del_init(&lock->l_bl_ast);
1406 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1407 LASSERT(lock->l_bl_ast_run == 0);
1408 LASSERT(lock->l_blocking_lock);
1409 lock->l_bl_ast_run++;
1410 unlock_res_and_lock(lock);
1412 ldlm_lock2desc(lock->l_blocking_lock, &d);
1414 lock->l_blocking_ast(lock, &d, (void *)arg,
1416 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1417 lock->l_blocking_lock = NULL;
1418 LDLM_LOCK_RELEASE(lock);
1424 ldlm_work_cp_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1426 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock, l_cp_ast);
1427 ldlm_completion_callback completion_callback;
1431 /* It's possible to receive a completion AST before we've set
1432 * the l_completion_ast pointer: either because the AST arrived
1433 * before the reply, or simply because there's a small race
1434 * window between receiving the reply and finishing the local
1435 * enqueue. (bug 842)
1437 * This can't happen with the blocking_ast, however, because we
1438 * will never call the local blocking_ast until we drop our
1439 * reader/writer reference, which we won't do until we get the
1440 * reply and finish enqueueing. */
1442 /* nobody should touch l_cp_ast */
1443 lock_res_and_lock(lock);
1444 cfs_list_del_init(&lock->l_cp_ast);
1445 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1446 /* save l_completion_ast since it can be changed by
1447 * mds_intent_policy(), see bug 14225 */
1448 completion_callback = lock->l_completion_ast;
1449 lock->l_flags &= ~LDLM_FL_CP_REQD;
1450 unlock_res_and_lock(lock);
1452 if (completion_callback != NULL) {
1453 completion_callback(lock, 0, (void *)arg);
1456 LDLM_LOCK_RELEASE(lock);
1462 ldlm_work_revoke_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1464 struct ldlm_lock_desc desc;
1465 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1469 cfs_list_del_init(&lock->l_rk_ast);
1471 /* the desc just pretend to exclusive */
1472 ldlm_lock2desc(lock, &desc);
1473 desc.l_req_mode = LCK_EX;
1474 desc.l_granted_mode = 0;
1476 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1477 LDLM_LOCK_RELEASE(lock);
1482 int ldlm_run_ast_work(cfs_list_t *rpc_list, ldlm_desc_ast_t ast_type)
1484 struct ldlm_cb_set_arg arg;
1485 cfs_list_t *tmp, *pos;
1486 int (*work_ast_lock)(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg);
1490 if (cfs_list_empty(rpc_list))
1493 arg.set = ptlrpc_prep_set();
1494 if (NULL == arg.set)
1496 cfs_atomic_set(&arg.restart, 0);
1498 case LDLM_WORK_BL_AST:
1499 arg.type = LDLM_BL_CALLBACK;
1500 work_ast_lock = ldlm_work_bl_ast_lock;
1502 case LDLM_WORK_CP_AST:
1503 arg.type = LDLM_CP_CALLBACK;
1504 work_ast_lock = ldlm_work_cp_ast_lock;
1506 case LDLM_WORK_REVOKE_AST:
1507 arg.type = LDLM_BL_CALLBACK;
1508 work_ast_lock = ldlm_work_revoke_ast_lock;
1515 cfs_list_for_each_safe(tmp, pos, rpc_list) {
1516 ast_count += work_ast_lock(tmp, &arg);
1518 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1519 * and create a new set for requests that remained in
1521 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1522 ldlm_send_and_maybe_create_set(&arg, 1);
1528 ldlm_send_and_maybe_create_set(&arg, 0);
1530 /* In case when number of ASTs is multiply of
1531 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1532 * @arg.set must be destroyed here, otherwise we get
1533 * write memory leaking. */
1534 ptlrpc_set_destroy(arg.set);
1536 RETURN(cfs_atomic_read(&arg.restart) ? -ERESTART : 0);
1539 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1541 ldlm_reprocess_all(res);
1542 return LDLM_ITER_CONTINUE;
1545 static int ldlm_reprocess_res(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1546 cfs_hlist_node_t *hnode, void *arg)
1548 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1551 rc = reprocess_one_queue(res, arg);
1553 return rc == LDLM_ITER_STOP;
1556 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1561 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1562 ldlm_reprocess_res, NULL);
1567 void ldlm_reprocess_all(struct ldlm_resource *res)
1569 CFS_LIST_HEAD(rpc_list);
1573 /* Local lock trees don't get reprocessed. */
1574 if (ns_is_client(ldlm_res_to_ns(res))) {
1581 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1582 if (rc == LDLM_ITER_CONTINUE)
1583 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1586 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1587 if (rc == -ERESTART) {
1588 LASSERT(cfs_list_empty(&rpc_list));
1594 void ldlm_cancel_callback(struct ldlm_lock *lock)
1596 check_res_locked(lock->l_resource);
1597 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1598 lock->l_flags |= LDLM_FL_CANCEL;
1599 if (lock->l_blocking_ast) {
1600 // l_check_no_ns_lock(ns);
1601 unlock_res_and_lock(lock);
1602 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1604 lock_res_and_lock(lock);
1606 LDLM_DEBUG(lock, "no blocking ast");
1609 lock->l_flags |= LDLM_FL_BL_DONE;
1612 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1614 if (req->l_resource->lr_type != LDLM_PLAIN &&
1615 req->l_resource->lr_type != LDLM_IBITS)
1618 cfs_list_del_init(&req->l_sl_policy);
1619 cfs_list_del_init(&req->l_sl_mode);
1622 void ldlm_lock_cancel(struct ldlm_lock *lock)
1624 struct ldlm_resource *res;
1625 struct ldlm_namespace *ns;
1628 lock_res_and_lock(lock);
1630 res = lock->l_resource;
1631 ns = ldlm_res_to_ns(res);
1633 /* Please do not, no matter how tempting, remove this LBUG without
1634 * talking to me first. -phik */
1635 if (lock->l_readers || lock->l_writers) {
1636 LDLM_ERROR(lock, "lock still has references");
1640 ldlm_del_waiting_lock(lock);
1642 /* Releases cancel callback. */
1643 ldlm_cancel_callback(lock);
1645 /* Yes, second time, just in case it was added again while we were
1646 running with no res lock in ldlm_cancel_callback */
1647 ldlm_del_waiting_lock(lock);
1648 ldlm_resource_unlink_lock(lock);
1649 ldlm_lock_destroy_nolock(lock);
1651 if (lock->l_granted_mode == lock->l_req_mode)
1652 ldlm_pool_del(&ns->ns_pool, lock);
1654 /* Make sure we will not be called again for same lock what is possible
1655 * if not to zero out lock->l_granted_mode */
1656 lock->l_granted_mode = LCK_MINMODE;
1657 unlock_res_and_lock(lock);
1662 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1664 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1670 lock->l_ast_data = data;
1671 LDLM_LOCK_PUT(lock);
1675 int ldlm_cancel_locks_for_export_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1676 cfs_hlist_node_t *hnode, void *data)
1679 struct obd_export *exp = data;
1680 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
1681 struct ldlm_resource *res;
1683 res = ldlm_resource_getref(lock->l_resource);
1684 LDLM_LOCK_GET(lock);
1686 LDLM_DEBUG(lock, "export %p", exp);
1687 ldlm_res_lvbo_update(res, NULL, 1);
1688 ldlm_lock_cancel(lock);
1689 ldlm_reprocess_all(res);
1690 ldlm_resource_putref(res);
1691 LDLM_LOCK_RELEASE(lock);
1695 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1697 cfs_hash_for_each_empty(exp->exp_lock_hash,
1698 ldlm_cancel_locks_for_export_cb, exp);
1702 * Downgrade an exclusive lock.
1704 * A fast variant of ldlm_lock_convert for convertion of exclusive
1705 * locks. The convertion is always successful.
1707 * \param lock A lock to convert
1708 * \param new_mode new lock mode
1710 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1714 LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1715 LASSERT(new_mode == LCK_COS);
1717 lock_res_and_lock(lock);
1718 ldlm_resource_unlink_lock(lock);
1720 * Remove the lock from pool as it will be added again in
1721 * ldlm_grant_lock() called below.
1723 ldlm_pool_del(&ldlm_lock_to_ns(lock)->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 = ldlm_res_to_ns(res);
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(ldlm_res_to_ns(res))) {
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 lr_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 ldlm_lock_to_ns_name(lock), 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 ldlm_lock_to_ns_name(lock), 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 ldlm_lock_to_ns_name(lock),
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 ldlm_lock_to_ns_name(lock),
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);