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.
32 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * lustre/ldlm/ldlm_lock.c
41 * Author: Peter Braam <braam@clusterfs.com>
42 * Author: Phil Schwan <phil@clusterfs.com>
45 #define DEBUG_SUBSYSTEM S_LDLM
48 # include <libcfs/libcfs.h>
49 # include <linux/lustre_intent.h>
51 # include <liblustre.h>
54 #include <obd_class.h>
55 #include "ldlm_internal.h"
58 char *ldlm_lockname[] = {
70 char *ldlm_typename[] = {
77 static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = {
78 [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_wire_to_local,
79 [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_wire_to_local,
80 [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_wire_to_local,
81 [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_wire_to_local,
84 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
85 [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_local_to_wire,
86 [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_local_to_wire,
87 [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_local_to_wire,
88 [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_local_to_wire,
92 * Converts lock policy from local format to on the wire lock_desc format
94 void ldlm_convert_policy_to_wire(ldlm_type_t type,
95 const ldlm_policy_data_t *lpolicy,
96 ldlm_wire_policy_data_t *wpolicy)
98 ldlm_policy_local_to_wire_t convert;
100 convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
102 convert(lpolicy, wpolicy);
106 * Converts lock policy from on the wire lock_desc format to local format
108 void ldlm_convert_policy_to_local(ldlm_type_t type,
109 const ldlm_wire_policy_data_t *wpolicy,
110 ldlm_policy_data_t *lpolicy)
112 ldlm_policy_wire_to_local_t convert;
114 convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE];
116 convert(wpolicy, lpolicy);
119 char *ldlm_it2str(int it)
126 case (IT_OPEN | IT_CREAT):
139 CERROR("Unknown intent %d\n", it);
144 extern cfs_mem_cache_t *ldlm_lock_slab;
146 static ldlm_processing_policy ldlm_processing_policy_table[] = {
147 [LDLM_PLAIN] ldlm_process_plain_lock,
148 [LDLM_EXTENT] ldlm_process_extent_lock,
150 [LDLM_FLOCK] ldlm_process_flock_lock,
152 [LDLM_IBITS] ldlm_process_inodebits_lock,
155 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
157 return ldlm_processing_policy_table[res->lr_type];
160 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
166 * REFCOUNTED LOCK OBJECTS
171 * Lock refcounts, during creation:
172 * - one special one for allocation, dec'd only once in destroy
173 * - one for being a lock that's in-use
174 * - one for the addref associated with a new lock
176 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
178 cfs_atomic_inc(&lock->l_refc);
182 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
184 LASSERT(size == sizeof(*lock));
185 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
188 void ldlm_lock_put(struct ldlm_lock *lock)
192 LASSERT(lock->l_resource != LP_POISON);
193 LASSERT(cfs_atomic_read(&lock->l_refc) > 0);
194 if (cfs_atomic_dec_and_test(&lock->l_refc)) {
195 struct ldlm_resource *res;
198 "final lock_put on destroyed lock, freeing it.");
200 res = lock->l_resource;
201 LASSERT(lock->l_destroyed);
202 LASSERT(cfs_list_empty(&lock->l_res_link));
203 LASSERT(cfs_list_empty(&lock->l_pending_chain));
205 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
207 lu_ref_del(&res->lr_reference, "lock", lock);
208 ldlm_resource_putref(res);
209 lock->l_resource = NULL;
210 if (lock->l_export) {
211 class_export_lock_put(lock->l_export, lock);
212 lock->l_export = NULL;
215 if (lock->l_lvb_data != NULL)
216 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
218 ldlm_interval_free(ldlm_interval_detach(lock));
219 lu_ref_fini(&lock->l_reference);
220 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
227 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
230 if (!cfs_list_empty(&lock->l_lru)) {
231 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
233 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
234 cfs_list_del_init(&lock->l_lru);
235 if (lock->l_flags & LDLM_FL_SKIPPED)
236 lock->l_flags &= ~LDLM_FL_SKIPPED;
237 LASSERT(ns->ns_nr_unused > 0);
244 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
246 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
250 if (lock->l_ns_srv) {
251 LASSERT(cfs_list_empty(&lock->l_lru));
255 cfs_spin_lock(&ns->ns_lock);
256 rc = ldlm_lock_remove_from_lru_nolock(lock);
257 cfs_spin_unlock(&ns->ns_lock);
262 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
264 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
266 lock->l_last_used = cfs_time_current();
267 LASSERT(cfs_list_empty(&lock->l_lru));
268 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
269 cfs_list_add_tail(&lock->l_lru, &ns->ns_unused_list);
270 LASSERT(ns->ns_nr_unused >= 0);
274 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
276 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
279 cfs_spin_lock(&ns->ns_lock);
280 ldlm_lock_add_to_lru_nolock(lock);
281 cfs_spin_unlock(&ns->ns_lock);
285 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
287 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
290 if (lock->l_ns_srv) {
291 LASSERT(cfs_list_empty(&lock->l_lru));
296 cfs_spin_lock(&ns->ns_lock);
297 if (!cfs_list_empty(&lock->l_lru)) {
298 ldlm_lock_remove_from_lru_nolock(lock);
299 ldlm_lock_add_to_lru_nolock(lock);
301 cfs_spin_unlock(&ns->ns_lock);
305 /* This used to have a 'strict' flag, which recovery would use to mark an
306 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
307 * shall explain why it's gone: with the new hash table scheme, once you call
308 * ldlm_lock_destroy, you can never drop your final references on this lock.
309 * Because it's not in the hash table anymore. -phil */
310 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
314 if (lock->l_readers || lock->l_writers) {
315 LDLM_ERROR(lock, "lock still has references");
316 ldlm_lock_dump(D_ERROR, lock, 0);
320 if (!cfs_list_empty(&lock->l_res_link)) {
321 LDLM_ERROR(lock, "lock still on resource");
322 ldlm_lock_dump(D_ERROR, lock, 0);
326 if (lock->l_destroyed) {
327 LASSERT(cfs_list_empty(&lock->l_lru));
331 lock->l_destroyed = 1;
333 if (lock->l_export && lock->l_export->exp_lock_hash &&
334 !cfs_hlist_unhashed(&lock->l_exp_hash))
335 cfs_hash_del(lock->l_export->exp_lock_hash,
336 &lock->l_remote_handle, &lock->l_exp_hash);
338 ldlm_lock_remove_from_lru(lock);
339 class_handle_unhash(&lock->l_handle);
342 /* Wake anyone waiting for this lock */
343 /* FIXME: I should probably add yet another flag, instead of using
344 * l_export to only call this on clients */
346 class_export_put(lock->l_export);
347 lock->l_export = NULL;
348 if (lock->l_export && lock->l_completion_ast)
349 lock->l_completion_ast(lock, 0);
355 void ldlm_lock_destroy(struct ldlm_lock *lock)
359 lock_res_and_lock(lock);
360 first = ldlm_lock_destroy_internal(lock);
361 unlock_res_and_lock(lock);
363 /* drop reference from hashtable only for first destroy */
365 lu_ref_del(&lock->l_reference, "hash", lock);
366 LDLM_LOCK_RELEASE(lock);
371 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
375 first = ldlm_lock_destroy_internal(lock);
376 /* drop reference from hashtable only for first destroy */
378 lu_ref_del(&lock->l_reference, "hash", lock);
379 LDLM_LOCK_RELEASE(lock);
384 /* this is called by portals_handle2object with the handle lock taken */
385 static void lock_handle_addref(void *lock)
387 LDLM_LOCK_GET((struct ldlm_lock *)lock);
391 * usage: pass in a resource on which you have done ldlm_resource_get
392 * new lock will take over the refcount.
393 * returns: lock with refcount 2 - one for current caller and one for remote
395 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
397 struct ldlm_lock *lock;
400 if (resource == NULL)
403 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, CFS_ALLOC_IO);
407 cfs_spin_lock_init(&lock->l_lock);
408 lock->l_resource = resource;
409 lu_ref_add(&resource->lr_reference, "lock", lock);
411 cfs_atomic_set(&lock->l_refc, 2);
412 CFS_INIT_LIST_HEAD(&lock->l_res_link);
413 CFS_INIT_LIST_HEAD(&lock->l_lru);
414 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
415 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
416 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
417 CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
418 cfs_waitq_init(&lock->l_waitq);
419 lock->l_blocking_lock = NULL;
420 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
421 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
422 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
424 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
426 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
427 class_handle_hash(&lock->l_handle, lock_handle_addref);
429 lu_ref_init(&lock->l_reference);
430 lu_ref_add(&lock->l_reference, "hash", lock);
431 lock->l_callback_timeout = 0;
433 #if LUSTRE_TRACKS_LOCK_EXP_REFS
434 CFS_INIT_LIST_HEAD(&lock->l_exp_refs_link);
435 lock->l_exp_refs_nr = 0;
436 lock->l_exp_refs_target = NULL;
442 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
443 const struct ldlm_res_id *new_resid)
445 struct ldlm_resource *oldres = lock->l_resource;
446 struct ldlm_resource *newres;
450 LASSERT(ns_is_client(ns));
452 lock_res_and_lock(lock);
453 if (memcmp(new_resid, &lock->l_resource->lr_name,
454 sizeof(lock->l_resource->lr_name)) == 0) {
456 unlock_res_and_lock(lock);
460 LASSERT(new_resid->name[0] != 0);
462 /* This function assumes that the lock isn't on any lists */
463 LASSERT(cfs_list_empty(&lock->l_res_link));
465 type = oldres->lr_type;
466 unlock_res_and_lock(lock);
468 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
472 lu_ref_add(&newres->lr_reference, "lock", lock);
474 * To flip the lock from the old to the new resource, lock, oldres and
475 * newres have to be locked. Resource spin-locks are nested within
476 * lock->l_lock, and are taken in the memory address order to avoid
479 cfs_spin_lock(&lock->l_lock);
480 oldres = lock->l_resource;
481 if (oldres < newres) {
483 lock_res_nested(newres, LRT_NEW);
486 lock_res_nested(oldres, LRT_NEW);
488 LASSERT(memcmp(new_resid, &oldres->lr_name,
489 sizeof oldres->lr_name) != 0);
490 lock->l_resource = newres;
492 unlock_res_and_lock(lock);
494 /* ...and the flowers are still standing! */
495 lu_ref_del(&oldres->lr_reference, "lock", lock);
496 ldlm_resource_putref(oldres);
505 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
507 lockh->cookie = lock->l_handle.h_cookie;
510 /* if flags: atomically get the lock and set the flags.
511 * Return NULL if flag already set
514 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
517 struct ldlm_lock *lock;
522 lock = class_handle2object(handle->cookie);
526 /* It's unlikely but possible that someone marked the lock as
527 * destroyed after we did handle2object on it */
528 if (flags == 0 && !lock->l_destroyed) {
529 lu_ref_add(&lock->l_reference, "handle", cfs_current());
533 lock_res_and_lock(lock);
535 LASSERT(lock->l_resource != NULL);
537 lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
538 if (unlikely(lock->l_destroyed)) {
539 unlock_res_and_lock(lock);
540 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
545 if (flags && (lock->l_flags & flags)) {
546 unlock_res_and_lock(lock);
552 lock->l_flags |= flags;
554 unlock_res_and_lock(lock);
558 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
560 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
561 /* INODEBITS_INTEROP: If the other side does not support
562 * inodebits, reply with a plain lock descriptor.
564 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
565 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
566 /* Make sure all the right bits are set in this lock we
567 are going to pass to client */
568 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
569 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
570 "Inappropriate inode lock bits during "
571 "conversion " LPU64 "\n",
572 lock->l_policy_data.l_inodebits.bits);
574 ldlm_res2desc(lock->l_resource, &desc->l_resource);
575 desc->l_resource.lr_type = LDLM_PLAIN;
577 /* Convert "new" lock mode to something old client can
579 if ((lock->l_req_mode == LCK_CR) ||
580 (lock->l_req_mode == LCK_CW))
581 desc->l_req_mode = LCK_PR;
583 desc->l_req_mode = lock->l_req_mode;
584 if ((lock->l_granted_mode == LCK_CR) ||
585 (lock->l_granted_mode == LCK_CW)) {
586 desc->l_granted_mode = LCK_PR;
588 /* We never grant PW/EX locks to clients */
589 LASSERT((lock->l_granted_mode != LCK_PW) &&
590 (lock->l_granted_mode != LCK_EX));
591 desc->l_granted_mode = lock->l_granted_mode;
594 /* We do not copy policy here, because there is no
595 policy for plain locks */
597 ldlm_res2desc(lock->l_resource, &desc->l_resource);
598 desc->l_req_mode = lock->l_req_mode;
599 desc->l_granted_mode = lock->l_granted_mode;
600 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
601 &lock->l_policy_data,
602 &desc->l_policy_data);
606 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
607 cfs_list_t *work_list)
609 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
610 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
611 lock->l_flags |= LDLM_FL_AST_SENT;
612 /* If the enqueuing client said so, tell the AST recipient to
613 * discard dirty data, rather than writing back. */
614 if (new->l_flags & LDLM_AST_DISCARD_DATA)
615 lock->l_flags |= LDLM_FL_DISCARD_DATA;
616 LASSERT(cfs_list_empty(&lock->l_bl_ast));
617 cfs_list_add(&lock->l_bl_ast, work_list);
619 LASSERT(lock->l_blocking_lock == NULL);
620 lock->l_blocking_lock = LDLM_LOCK_GET(new);
624 void ldlm_add_cp_work_item(struct ldlm_lock *lock, cfs_list_t *work_list)
626 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
627 lock->l_flags |= LDLM_FL_CP_REQD;
628 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
629 LASSERT(cfs_list_empty(&lock->l_cp_ast));
630 cfs_list_add(&lock->l_cp_ast, work_list);
635 /* must be called with lr_lock held */
636 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
637 cfs_list_t *work_list)
640 check_res_locked(lock->l_resource);
642 ldlm_add_bl_work_item(lock, new, work_list);
644 ldlm_add_cp_work_item(lock, work_list);
648 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
650 struct ldlm_lock *lock;
652 lock = ldlm_handle2lock(lockh);
653 LASSERT(lock != NULL);
654 ldlm_lock_addref_internal(lock, mode);
658 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
660 ldlm_lock_remove_from_lru(lock);
661 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
663 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
665 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
667 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
670 lu_ref_add_atomic(&lock->l_reference, "user", lock);
671 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
675 * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
678 * \retval 0 success, lock was addref-ed
680 * \retval -EAGAIN lock is being canceled.
682 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
684 struct ldlm_lock *lock;
688 lock = ldlm_handle2lock(lockh);
690 lock_res_and_lock(lock);
691 if (lock->l_readers != 0 || lock->l_writers != 0 ||
692 !(lock->l_flags & LDLM_FL_CBPENDING)) {
693 ldlm_lock_addref_internal_nolock(lock, mode);
696 unlock_res_and_lock(lock);
702 /* only called for local locks */
703 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
705 lock_res_and_lock(lock);
706 ldlm_lock_addref_internal_nolock(lock, mode);
707 unlock_res_and_lock(lock);
710 /* only called in ldlm_flock_destroy and for local locks.
711 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
712 * * ldlm_lock_remove_from_lru() does nothing, it is safe
713 * * for ldlm_flock_destroy usage by dropping some code */
714 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
716 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
717 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
718 LASSERT(lock->l_readers > 0);
719 lu_ref_del(&lock->l_reference, "reader", lock);
722 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
723 LASSERT(lock->l_writers > 0);
724 lu_ref_del(&lock->l_reference, "writer", lock);
728 lu_ref_del(&lock->l_reference, "user", lock);
729 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
732 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
734 struct ldlm_namespace *ns;
737 lock_res_and_lock(lock);
739 ns = ldlm_lock_to_ns(lock);
741 ldlm_lock_decref_internal_nolock(lock, mode);
743 if (lock->l_flags & LDLM_FL_LOCAL &&
744 !lock->l_readers && !lock->l_writers) {
745 /* If this is a local lock on a server namespace and this was
746 * the last reference, cancel the lock. */
747 CDEBUG(D_INFO, "forcing cancel of local lock\n");
748 lock->l_flags |= LDLM_FL_CBPENDING;
751 if (!lock->l_readers && !lock->l_writers &&
752 (lock->l_flags & LDLM_FL_CBPENDING)) {
753 /* If we received a blocked AST and this was the last reference,
754 * run the callback. */
755 if (lock->l_ns_srv && lock->l_export)
756 CERROR("FL_CBPENDING set on non-local lock--just a "
759 LDLM_DEBUG(lock, "final decref done on cbpending lock");
761 LDLM_LOCK_GET(lock); /* dropped by bl thread */
762 ldlm_lock_remove_from_lru(lock);
763 unlock_res_and_lock(lock);
765 if (lock->l_flags & LDLM_FL_FAIL_LOC)
766 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
768 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
769 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
770 ldlm_handle_bl_callback(ns, NULL, lock);
771 } else if (ns_is_client(ns) &&
772 !lock->l_readers && !lock->l_writers &&
773 !(lock->l_flags & LDLM_FL_BL_AST)) {
774 /* If this is a client-side namespace and this was the last
775 * reference, put it on the LRU. */
776 ldlm_lock_add_to_lru(lock);
777 unlock_res_and_lock(lock);
779 if (lock->l_flags & LDLM_FL_FAIL_LOC)
780 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
782 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
783 * are not supported by the server, otherwise, it is done on
785 if (!exp_connect_cancelset(lock->l_conn_export) &&
786 !ns_connect_lru_resize(ns))
787 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
789 unlock_res_and_lock(lock);
795 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
797 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
798 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
799 ldlm_lock_decref_internal(lock, mode);
803 /* This will drop a lock reference and mark it for destruction, but will not
804 * necessarily cancel the lock before returning. */
805 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
807 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
810 LASSERT(lock != NULL);
812 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
813 lock_res_and_lock(lock);
814 lock->l_flags |= LDLM_FL_CBPENDING;
815 unlock_res_and_lock(lock);
816 ldlm_lock_decref_internal(lock, mode);
820 struct sl_insert_point {
821 cfs_list_t *res_link;
822 cfs_list_t *mode_link;
823 cfs_list_t *policy_link;
827 * search_granted_lock
830 * Finds a position to insert the new lock.
832 * queue [input]: the granted list where search acts on;
833 * req [input]: the lock whose position to be located;
834 * prev [output]: positions within 3 lists to insert @req to
838 * - ldlm_grant_lock_with_skiplist
840 static void search_granted_lock(cfs_list_t *queue,
841 struct ldlm_lock *req,
842 struct sl_insert_point *prev)
845 struct ldlm_lock *lock, *mode_end, *policy_end;
848 cfs_list_for_each(tmp, queue) {
849 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
851 mode_end = cfs_list_entry(lock->l_sl_mode.prev,
852 struct ldlm_lock, l_sl_mode);
854 if (lock->l_req_mode != req->l_req_mode) {
855 /* jump to last lock of mode group */
856 tmp = &mode_end->l_res_link;
860 /* suitable mode group is found */
861 if (lock->l_resource->lr_type == LDLM_PLAIN) {
862 /* insert point is last lock of the mode group */
863 prev->res_link = &mode_end->l_res_link;
864 prev->mode_link = &mode_end->l_sl_mode;
865 prev->policy_link = &req->l_sl_policy;
868 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
871 cfs_list_entry(lock->l_sl_policy.prev,
875 if (lock->l_policy_data.l_inodebits.bits ==
876 req->l_policy_data.l_inodebits.bits) {
877 /* insert point is last lock of
878 * the policy group */
880 &policy_end->l_res_link;
882 &policy_end->l_sl_mode;
884 &policy_end->l_sl_policy;
889 if (policy_end == mode_end)
890 /* done with mode group */
893 /* go to next policy group within mode group */
894 tmp = policy_end->l_res_link.next;
895 lock = cfs_list_entry(tmp, struct ldlm_lock,
897 } /* loop over policy groups within the mode group */
899 /* insert point is last lock of the mode group,
900 * new policy group is started */
901 prev->res_link = &mode_end->l_res_link;
902 prev->mode_link = &mode_end->l_sl_mode;
903 prev->policy_link = &req->l_sl_policy;
907 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
912 /* insert point is last lock on the queue,
913 * new mode group and new policy group are started */
914 prev->res_link = queue->prev;
915 prev->mode_link = &req->l_sl_mode;
916 prev->policy_link = &req->l_sl_policy;
921 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
922 struct sl_insert_point *prev)
924 struct ldlm_resource *res = lock->l_resource;
927 check_res_locked(res);
929 ldlm_resource_dump(D_INFO, res);
930 CDEBUG(D_OTHER, "About to add this lock:\n");
931 ldlm_lock_dump(D_OTHER, lock, 0);
933 if (lock->l_destroyed) {
934 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
938 LASSERT(cfs_list_empty(&lock->l_res_link));
939 LASSERT(cfs_list_empty(&lock->l_sl_mode));
940 LASSERT(cfs_list_empty(&lock->l_sl_policy));
942 cfs_list_add(&lock->l_res_link, prev->res_link);
943 cfs_list_add(&lock->l_sl_mode, prev->mode_link);
944 cfs_list_add(&lock->l_sl_policy, prev->policy_link);
949 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
951 struct sl_insert_point prev;
954 LASSERT(lock->l_req_mode == lock->l_granted_mode);
956 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
957 ldlm_granted_list_add_lock(lock, &prev);
962 * - ldlm_lock_enqueue
963 * - ldlm_reprocess_queue
964 * - ldlm_lock_convert
966 * must be called with lr_lock held
968 void ldlm_grant_lock(struct ldlm_lock *lock, cfs_list_t *work_list)
970 struct ldlm_resource *res = lock->l_resource;
973 check_res_locked(res);
975 lock->l_granted_mode = lock->l_req_mode;
976 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
977 ldlm_grant_lock_with_skiplist(lock);
978 else if (res->lr_type == LDLM_EXTENT)
979 ldlm_extent_add_lock(res, lock);
981 ldlm_resource_add_lock(res, &res->lr_granted, lock);
983 if (lock->l_granted_mode < res->lr_most_restr)
984 res->lr_most_restr = lock->l_granted_mode;
986 if (work_list && lock->l_completion_ast != NULL)
987 ldlm_add_ast_work_item(lock, NULL, work_list);
989 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
993 /* returns a referenced lock or NULL. See the flag descriptions below, in the
994 * comment above ldlm_lock_match */
995 static struct ldlm_lock *search_queue(cfs_list_t *queue,
997 ldlm_policy_data_t *policy,
998 struct ldlm_lock *old_lock,
999 int flags, int unref)
1001 struct ldlm_lock *lock;
1004 cfs_list_for_each(tmp, queue) {
1007 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1009 if (lock == old_lock)
1012 /* llite sometimes wants to match locks that will be
1013 * canceled when their users drop, but we allow it to match
1014 * if it passes in CBPENDING and the lock still has users.
1015 * this is generally only going to be used by children
1016 * whose parents already hold a lock so forward progress
1017 * can still happen. */
1018 if (lock->l_flags & LDLM_FL_CBPENDING &&
1019 !(flags & LDLM_FL_CBPENDING))
1021 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
1022 lock->l_readers == 0 && lock->l_writers == 0)
1025 if (!(lock->l_req_mode & *mode))
1027 match = lock->l_req_mode;
1029 if (lock->l_resource->lr_type == LDLM_EXTENT &&
1030 (lock->l_policy_data.l_extent.start >
1031 policy->l_extent.start ||
1032 lock->l_policy_data.l_extent.end < policy->l_extent.end))
1035 if (unlikely(match == LCK_GROUP) &&
1036 lock->l_resource->lr_type == LDLM_EXTENT &&
1037 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
1040 /* We match if we have existing lock with same or wider set
1042 if (lock->l_resource->lr_type == LDLM_IBITS &&
1043 ((lock->l_policy_data.l_inodebits.bits &
1044 policy->l_inodebits.bits) !=
1045 policy->l_inodebits.bits))
1049 (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED)))
1052 if ((flags & LDLM_FL_LOCAL_ONLY) &&
1053 !(lock->l_flags & LDLM_FL_LOCAL))
1056 if (flags & LDLM_FL_TEST_LOCK) {
1057 LDLM_LOCK_GET(lock);
1058 ldlm_lock_touch_in_lru(lock);
1060 ldlm_lock_addref_internal_nolock(lock, match);
1069 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1071 lock->l_flags |= LDLM_FL_LVB_READY;
1072 cfs_waitq_signal(&lock->l_waitq);
1075 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1077 lock_res_and_lock(lock);
1078 ldlm_lock_allow_match_locked(lock);
1079 unlock_res_and_lock(lock);
1082 /* Can be called in two ways:
1084 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1085 * for a duplicate of.
1087 * Otherwise, all of the fields must be filled in, to match against.
1089 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1090 * server (ie, connh is NULL)
1091 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1092 * list will be considered
1093 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1094 * to be canceled can still be matched as long as they still have reader
1095 * or writer refernces
1096 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1097 * just tell us if we would have matched.
1099 * Returns 1 if it finds an already-existing lock that is compatible; in this
1100 * case, lockh is filled in with a addref()ed lock
1102 * we also check security context, if that failed we simply return 0 (to keep
1103 * caller code unchanged), the context failure will be discovered by caller
1106 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1107 const struct ldlm_res_id *res_id, ldlm_type_t type,
1108 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1109 struct lustre_handle *lockh, int unref)
1111 struct ldlm_resource *res;
1112 struct ldlm_lock *lock, *old_lock = NULL;
1117 old_lock = ldlm_handle2lock(lockh);
1120 ns = ldlm_lock_to_ns(old_lock);
1121 res_id = &old_lock->l_resource->lr_name;
1122 type = old_lock->l_resource->lr_type;
1123 mode = old_lock->l_req_mode;
1126 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1128 LASSERT(old_lock == NULL);
1132 LDLM_RESOURCE_ADDREF(res);
1135 lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1139 if (flags & LDLM_FL_BLOCK_GRANTED)
1141 lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1145 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1153 LDLM_RESOURCE_DELREF(res);
1154 ldlm_resource_putref(res);
1157 ldlm_lock2handle(lock, lockh);
1158 if ((flags & LDLM_FL_LVB_READY) &&
1159 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1160 struct l_wait_info lwi;
1161 if (lock->l_completion_ast) {
1162 int err = lock->l_completion_ast(lock,
1163 LDLM_FL_WAIT_NOREPROC,
1166 if (flags & LDLM_FL_TEST_LOCK)
1167 LDLM_LOCK_RELEASE(lock);
1169 ldlm_lock_decref_internal(lock,
1176 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1177 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1179 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1180 l_wait_event(lock->l_waitq,
1181 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1186 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1187 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1188 res_id->name[2] : policy->l_extent.start,
1189 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1190 res_id->name[3] : policy->l_extent.end);
1192 /* check user's security context */
1193 if (lock->l_conn_export &&
1194 sptlrpc_import_check_ctx(
1195 class_exp2cliimp(lock->l_conn_export))) {
1196 if (!(flags & LDLM_FL_TEST_LOCK))
1197 ldlm_lock_decref_internal(lock, mode);
1201 if (flags & LDLM_FL_TEST_LOCK)
1202 LDLM_LOCK_RELEASE(lock);
1204 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1205 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1206 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1207 type, mode, res_id->name[0], res_id->name[1],
1208 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1209 res_id->name[2] :policy->l_extent.start,
1210 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1211 res_id->name[3] : policy->l_extent.end);
1214 LDLM_LOCK_PUT(old_lock);
1216 return rc ? mode : 0;
1219 /* Returns a referenced lock */
1220 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1221 const struct ldlm_res_id *res_id,
1224 const struct ldlm_callback_suite *cbs,
1225 void *data, __u32 lvb_len)
1227 struct ldlm_lock *lock;
1228 struct ldlm_resource *res;
1231 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1235 lock = ldlm_lock_new(res);
1240 lock->l_req_mode = mode;
1241 lock->l_ast_data = data;
1242 lock->l_pid = cfs_curproc_pid();
1243 lock->l_ns_srv = ns_is_server(ns);
1245 lock->l_blocking_ast = cbs->lcs_blocking;
1246 lock->l_completion_ast = cbs->lcs_completion;
1247 lock->l_glimpse_ast = cbs->lcs_glimpse;
1248 lock->l_weigh_ast = cbs->lcs_weigh;
1251 lock->l_tree_node = NULL;
1252 /* if this is the extent lock, allocate the interval tree node */
1253 if (type == LDLM_EXTENT) {
1254 if (ldlm_interval_alloc(lock) == NULL)
1259 lock->l_lvb_len = lvb_len;
1260 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1261 if (lock->l_lvb_data == NULL)
1265 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1271 ldlm_lock_destroy(lock);
1272 LDLM_LOCK_RELEASE(lock);
1276 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1277 struct ldlm_lock **lockp,
1278 void *cookie, int *flags)
1280 struct ldlm_lock *lock = *lockp;
1281 struct ldlm_resource *res = lock->l_resource;
1282 int local = ns_is_client(ldlm_res_to_ns(res));
1283 ldlm_processing_policy policy;
1284 ldlm_error_t rc = ELDLM_OK;
1285 struct ldlm_interval *node = NULL;
1288 lock->l_last_activity = cfs_time_current_sec();
1289 /* policies are not executed on the client or during replay */
1290 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1291 && !local && ns->ns_policy) {
1292 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1294 if (rc == ELDLM_LOCK_REPLACED) {
1295 /* The lock that was returned has already been granted,
1296 * and placed into lockp. If it's not the same as the
1297 * one we passed in, then destroy the old one and our
1298 * work here is done. */
1299 if (lock != *lockp) {
1300 ldlm_lock_destroy(lock);
1301 LDLM_LOCK_RELEASE(lock);
1303 *flags |= LDLM_FL_LOCK_CHANGED;
1305 } else if (rc != ELDLM_OK ||
1306 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1307 ldlm_lock_destroy(lock);
1312 /* For a replaying lock, it might be already in granted list. So
1313 * unlinking the lock will cause the interval node to be freed, we
1314 * have to allocate the interval node early otherwise we can't regrant
1315 * this lock in the future. - jay */
1316 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1317 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1319 lock_res_and_lock(lock);
1320 if (local && lock->l_req_mode == lock->l_granted_mode) {
1321 /* The server returned a blocked lock, but it was granted
1322 * before we got a chance to actually enqueue it. We don't
1323 * need to do anything else. */
1324 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1325 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1326 GOTO(out, ELDLM_OK);
1329 ldlm_resource_unlink_lock(lock);
1330 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1332 ldlm_lock_destroy_nolock(lock);
1333 GOTO(out, rc = -ENOMEM);
1336 CFS_INIT_LIST_HEAD(&node->li_group);
1337 ldlm_interval_attach(node, lock);
1341 /* Some flags from the enqueue want to make it into the AST, via the
1342 * lock's l_flags. */
1343 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1345 /* This distinction between local lock trees is very important; a client
1346 * namespace only has information about locks taken by that client, and
1347 * thus doesn't have enough information to decide for itself if it can
1348 * be granted (below). In this case, we do exactly what the server
1349 * tells us to do, as dictated by the 'flags'.
1351 * We do exactly the same thing during recovery, when the server is
1352 * more or less trusting the clients not to lie.
1354 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1355 * granted/converting queues. */
1357 if (*flags & LDLM_FL_BLOCK_CONV)
1358 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1359 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1360 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1362 ldlm_grant_lock(lock, NULL);
1363 GOTO(out, ELDLM_OK);
1364 } else if (*flags & LDLM_FL_REPLAY) {
1365 if (*flags & LDLM_FL_BLOCK_CONV) {
1366 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1367 GOTO(out, ELDLM_OK);
1368 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1369 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1370 GOTO(out, ELDLM_OK);
1371 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1372 ldlm_grant_lock(lock, NULL);
1373 GOTO(out, ELDLM_OK);
1375 /* If no flags, fall through to normal enqueue path. */
1378 policy = ldlm_processing_policy_table[res->lr_type];
1379 policy(lock, flags, 1, &rc, NULL);
1382 unlock_res_and_lock(lock);
1384 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1388 /* Must be called with namespace taken: queue is waiting or converting. */
1389 int ldlm_reprocess_queue(struct ldlm_resource *res, cfs_list_t *queue,
1390 cfs_list_t *work_list)
1392 cfs_list_t *tmp, *pos;
1393 ldlm_processing_policy policy;
1395 int rc = LDLM_ITER_CONTINUE;
1399 check_res_locked(res);
1401 policy = ldlm_processing_policy_table[res->lr_type];
1404 cfs_list_for_each_safe(tmp, pos, queue) {
1405 struct ldlm_lock *pending;
1406 pending = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1408 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1411 rc = policy(pending, &flags, 0, &err, work_list);
1412 if (rc != LDLM_ITER_CONTINUE)
1419 /* Helper function for ldlm_run_ast_work().
1421 * Send an existing rpc set specified by @arg->set and then
1422 * destroy it. Create new one if @do_create flag is set. */
1424 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1428 ptlrpc_set_wait(arg->set);
1429 if (arg->type == LDLM_BL_CALLBACK)
1430 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1431 ptlrpc_set_destroy(arg->set);
1434 arg->set = ptlrpc_prep_set();
1440 ldlm_work_bl_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1442 struct ldlm_lock_desc d;
1443 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1447 /* nobody should touch l_bl_ast */
1448 lock_res_and_lock(lock);
1449 cfs_list_del_init(&lock->l_bl_ast);
1451 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1452 LASSERT(lock->l_bl_ast_run == 0);
1453 LASSERT(lock->l_blocking_lock);
1454 lock->l_bl_ast_run++;
1455 unlock_res_and_lock(lock);
1457 ldlm_lock2desc(lock->l_blocking_lock, &d);
1459 lock->l_blocking_ast(lock, &d, (void *)arg,
1461 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1462 lock->l_blocking_lock = NULL;
1463 LDLM_LOCK_RELEASE(lock);
1469 ldlm_work_cp_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1471 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock, l_cp_ast);
1472 ldlm_completion_callback completion_callback;
1476 /* It's possible to receive a completion AST before we've set
1477 * the l_completion_ast pointer: either because the AST arrived
1478 * before the reply, or simply because there's a small race
1479 * window between receiving the reply and finishing the local
1480 * enqueue. (bug 842)
1482 * This can't happen with the blocking_ast, however, because we
1483 * will never call the local blocking_ast until we drop our
1484 * reader/writer reference, which we won't do until we get the
1485 * reply and finish enqueueing. */
1487 /* nobody should touch l_cp_ast */
1488 lock_res_and_lock(lock);
1489 cfs_list_del_init(&lock->l_cp_ast);
1490 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1491 /* save l_completion_ast since it can be changed by
1492 * mds_intent_policy(), see bug 14225 */
1493 completion_callback = lock->l_completion_ast;
1494 lock->l_flags &= ~LDLM_FL_CP_REQD;
1495 unlock_res_and_lock(lock);
1497 if (completion_callback != NULL) {
1498 completion_callback(lock, 0, (void *)arg);
1501 LDLM_LOCK_RELEASE(lock);
1507 ldlm_work_revoke_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1509 struct ldlm_lock_desc desc;
1510 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1514 cfs_list_del_init(&lock->l_rk_ast);
1516 /* the desc just pretend to exclusive */
1517 ldlm_lock2desc(lock, &desc);
1518 desc.l_req_mode = LCK_EX;
1519 desc.l_granted_mode = 0;
1521 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1522 LDLM_LOCK_RELEASE(lock);
1527 int ldlm_run_ast_work(cfs_list_t *rpc_list, ldlm_desc_ast_t ast_type)
1529 struct ldlm_cb_set_arg arg;
1530 cfs_list_t *tmp, *pos;
1531 int (*work_ast_lock)(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg);
1535 if (cfs_list_empty(rpc_list))
1538 arg.set = ptlrpc_prep_set();
1539 if (NULL == arg.set)
1541 cfs_atomic_set(&arg.restart, 0);
1543 case LDLM_WORK_BL_AST:
1544 arg.type = LDLM_BL_CALLBACK;
1545 work_ast_lock = ldlm_work_bl_ast_lock;
1547 case LDLM_WORK_CP_AST:
1548 arg.type = LDLM_CP_CALLBACK;
1549 work_ast_lock = ldlm_work_cp_ast_lock;
1551 case LDLM_WORK_REVOKE_AST:
1552 arg.type = LDLM_BL_CALLBACK;
1553 work_ast_lock = ldlm_work_revoke_ast_lock;
1560 cfs_list_for_each_safe(tmp, pos, rpc_list) {
1561 ast_count += work_ast_lock(tmp, &arg);
1563 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1564 * and create a new set for requests that remained in
1566 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1567 ldlm_send_and_maybe_create_set(&arg, 1);
1573 ldlm_send_and_maybe_create_set(&arg, 0);
1575 /* In case when number of ASTs is multiply of
1576 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1577 * @arg.set must be destroyed here, otherwise we get
1578 * write memory leaking. */
1579 ptlrpc_set_destroy(arg.set);
1581 RETURN(cfs_atomic_read(&arg.restart) ? -ERESTART : 0);
1584 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1586 ldlm_reprocess_all(res);
1587 return LDLM_ITER_CONTINUE;
1590 static int ldlm_reprocess_res(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1591 cfs_hlist_node_t *hnode, void *arg)
1593 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1596 rc = reprocess_one_queue(res, arg);
1598 return rc == LDLM_ITER_STOP;
1601 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1606 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1607 ldlm_reprocess_res, NULL);
1612 void ldlm_reprocess_all(struct ldlm_resource *res)
1614 CFS_LIST_HEAD(rpc_list);
1618 /* Local lock trees don't get reprocessed. */
1619 if (ns_is_client(ldlm_res_to_ns(res))) {
1626 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1627 if (rc == LDLM_ITER_CONTINUE)
1628 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1631 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1632 if (rc == -ERESTART) {
1633 LASSERT(cfs_list_empty(&rpc_list));
1639 void ldlm_cancel_callback(struct ldlm_lock *lock)
1641 check_res_locked(lock->l_resource);
1642 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1643 lock->l_flags |= LDLM_FL_CANCEL;
1644 if (lock->l_blocking_ast) {
1645 // l_check_no_ns_lock(ns);
1646 unlock_res_and_lock(lock);
1647 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1649 lock_res_and_lock(lock);
1651 LDLM_DEBUG(lock, "no blocking ast");
1654 lock->l_flags |= LDLM_FL_BL_DONE;
1657 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1659 if (req->l_resource->lr_type != LDLM_PLAIN &&
1660 req->l_resource->lr_type != LDLM_IBITS)
1663 cfs_list_del_init(&req->l_sl_policy);
1664 cfs_list_del_init(&req->l_sl_mode);
1667 void ldlm_lock_cancel(struct ldlm_lock *lock)
1669 struct ldlm_resource *res;
1670 struct ldlm_namespace *ns;
1673 lock_res_and_lock(lock);
1675 res = lock->l_resource;
1676 ns = ldlm_res_to_ns(res);
1678 /* Please do not, no matter how tempting, remove this LBUG without
1679 * talking to me first. -phik */
1680 if (lock->l_readers || lock->l_writers) {
1681 LDLM_ERROR(lock, "lock still has references");
1685 ldlm_del_waiting_lock(lock);
1687 /* Releases cancel callback. */
1688 ldlm_cancel_callback(lock);
1690 /* Yes, second time, just in case it was added again while we were
1691 running with no res lock in ldlm_cancel_callback */
1692 ldlm_del_waiting_lock(lock);
1693 ldlm_resource_unlink_lock(lock);
1694 ldlm_lock_destroy_nolock(lock);
1696 if (lock->l_granted_mode == lock->l_req_mode)
1697 ldlm_pool_del(&ns->ns_pool, lock);
1699 /* Make sure we will not be called again for same lock what is possible
1700 * if not to zero out lock->l_granted_mode */
1701 lock->l_granted_mode = LCK_MINMODE;
1702 unlock_res_and_lock(lock);
1707 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1709 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1715 lock->l_ast_data = data;
1716 LDLM_LOCK_PUT(lock);
1720 int ldlm_cancel_locks_for_export_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1721 cfs_hlist_node_t *hnode, void *data)
1724 struct obd_export *exp = data;
1725 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
1726 struct ldlm_resource *res;
1728 res = ldlm_resource_getref(lock->l_resource);
1729 LDLM_LOCK_GET(lock);
1731 LDLM_DEBUG(lock, "export %p", exp);
1732 ldlm_res_lvbo_update(res, NULL, 1);
1733 ldlm_lock_cancel(lock);
1734 ldlm_reprocess_all(res);
1735 ldlm_resource_putref(res);
1736 LDLM_LOCK_RELEASE(lock);
1740 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1742 cfs_hash_for_each_empty(exp->exp_lock_hash,
1743 ldlm_cancel_locks_for_export_cb, exp);
1747 * Downgrade an exclusive lock.
1749 * A fast variant of ldlm_lock_convert for convertion of exclusive
1750 * locks. The convertion is always successful.
1752 * \param lock A lock to convert
1753 * \param new_mode new lock mode
1755 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1759 LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1760 LASSERT(new_mode == LCK_COS);
1762 lock_res_and_lock(lock);
1763 ldlm_resource_unlink_lock(lock);
1765 * Remove the lock from pool as it will be added again in
1766 * ldlm_grant_lock() called below.
1768 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
1770 lock->l_req_mode = new_mode;
1771 ldlm_grant_lock(lock, NULL);
1772 unlock_res_and_lock(lock);
1773 ldlm_reprocess_all(lock->l_resource);
1778 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1781 CFS_LIST_HEAD(rpc_list);
1782 struct ldlm_resource *res;
1783 struct ldlm_namespace *ns;
1786 struct sl_insert_point prev;
1788 struct ldlm_interval *node;
1791 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1792 *flags |= LDLM_FL_BLOCK_GRANTED;
1793 RETURN(lock->l_resource);
1796 /* I can't check the type of lock here because the bitlock of lock
1797 * is not held here, so do the allocation blindly. -jay */
1798 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1799 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1802 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1803 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1805 lock_res_and_lock(lock);
1807 res = lock->l_resource;
1808 ns = ldlm_res_to_ns(res);
1810 old_mode = lock->l_req_mode;
1811 lock->l_req_mode = new_mode;
1812 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1813 /* remember the lock position where the lock might be
1814 * added back to the granted list later and also
1815 * remember the join mode for skiplist fixing. */
1816 prev.res_link = lock->l_res_link.prev;
1817 prev.mode_link = lock->l_sl_mode.prev;
1818 prev.policy_link = lock->l_sl_policy.prev;
1819 ldlm_resource_unlink_lock(lock);
1821 ldlm_resource_unlink_lock(lock);
1822 if (res->lr_type == LDLM_EXTENT) {
1823 /* FIXME: ugly code, I have to attach the lock to a
1824 * interval node again since perhaps it will be granted
1826 CFS_INIT_LIST_HEAD(&node->li_group);
1827 ldlm_interval_attach(node, lock);
1833 * Remove old lock from the pool before adding the lock with new
1834 * mode below in ->policy()
1836 ldlm_pool_del(&ns->ns_pool, lock);
1838 /* If this is a local resource, put it on the appropriate list. */
1839 if (ns_is_client(ldlm_res_to_ns(res))) {
1840 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1841 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1843 /* This should never happen, because of the way the
1844 * server handles conversions. */
1845 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1849 ldlm_grant_lock(lock, &rpc_list);
1851 /* FIXME: completion handling not with lr_lock held ! */
1852 if (lock->l_completion_ast)
1853 lock->l_completion_ast(lock, 0, NULL);
1857 ldlm_processing_policy policy;
1858 policy = ldlm_processing_policy_table[res->lr_type];
1859 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1860 if (rc == LDLM_ITER_STOP) {
1861 lock->l_req_mode = old_mode;
1862 if (res->lr_type == LDLM_EXTENT)
1863 ldlm_extent_add_lock(res, lock);
1865 ldlm_granted_list_add_lock(lock, &prev);
1869 *flags |= LDLM_FL_BLOCK_GRANTED;
1873 unlock_res_and_lock(lock);
1876 ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1878 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1882 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1884 struct obd_device *obd = NULL;
1886 if (!((libcfs_debug | D_ERROR) & level))
1890 CDEBUG(level, " NULL LDLM lock\n");
1894 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1895 lock, lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1897 if (lock->l_conn_export != NULL)
1898 obd = lock->l_conn_export->exp_obd;
1899 if (lock->l_export && lock->l_export->exp_connection) {
1900 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1901 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1902 lock->l_remote_handle.cookie);
1903 } else if (obd == NULL) {
1904 CDEBUG(level, " Node: local\n");
1906 struct obd_import *imp = obd->u.cli.cl_import;
1907 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1908 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1909 lock->l_remote_handle.cookie);
1911 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1913 lock->l_resource->lr_name.name[0],
1914 lock->l_resource->lr_name.name[1],
1915 lock->l_resource->lr_name.name[2]);
1916 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1917 "write: %d flags: "LPX64"\n", ldlm_lockname[lock->l_req_mode],
1918 ldlm_lockname[lock->l_granted_mode],
1919 cfs_atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1921 if (lock->l_resource->lr_type == LDLM_EXTENT)
1922 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1923 " (req "LPU64"-"LPU64")\n",
1924 lock->l_policy_data.l_extent.start,
1925 lock->l_policy_data.l_extent.end,
1926 lock->l_req_extent.start, lock->l_req_extent.end);
1927 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1928 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1929 lock->l_policy_data.l_flock.pid,
1930 lock->l_policy_data.l_flock.start,
1931 lock->l_policy_data.l_flock.end);
1932 else if (lock->l_resource->lr_type == LDLM_IBITS)
1933 CDEBUG(level, " Bits: "LPX64"\n",
1934 lock->l_policy_data.l_inodebits.bits);
1937 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1939 struct ldlm_lock *lock;
1941 if (!((libcfs_debug | D_ERROR) & level))
1944 lock = ldlm_handle2lock(lockh);
1948 ldlm_lock_dump(D_OTHER, lock, 0);
1950 LDLM_LOCK_PUT(lock);
1953 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1954 struct libcfs_debug_msg_data *data, const char *fmt,
1958 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1960 va_start(args, fmt);
1962 if (lock->l_resource == NULL) {
1963 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1964 data->msg_fn, data->msg_line, fmt, args,
1965 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1966 "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" remote: "
1967 LPX64" expref: %d pid: %u timeout: %lu\n", lock,
1968 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1969 lock->l_readers, lock->l_writers,
1970 ldlm_lockname[lock->l_granted_mode],
1971 ldlm_lockname[lock->l_req_mode],
1972 lock->l_flags, lock->l_remote_handle.cookie,
1974 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
1975 lock->l_pid, lock->l_callback_timeout);
1980 switch (lock->l_resource->lr_type) {
1982 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
1983 data->msg_fn, data->msg_line, fmt, args,
1984 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1985 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1986 "] (req "LPU64"->"LPU64") flags: "LPX64" remote: "LPX64
1987 " expref: %d pid: %u timeout %lu\n",
1988 ldlm_lock_to_ns_name(lock), lock,
1989 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1990 lock->l_readers, lock->l_writers,
1991 ldlm_lockname[lock->l_granted_mode],
1992 ldlm_lockname[lock->l_req_mode],
1993 lock->l_resource->lr_name.name[0],
1994 lock->l_resource->lr_name.name[1],
1995 cfs_atomic_read(&lock->l_resource->lr_refcount),
1996 ldlm_typename[lock->l_resource->lr_type],
1997 lock->l_policy_data.l_extent.start,
1998 lock->l_policy_data.l_extent.end,
1999 lock->l_req_extent.start, lock->l_req_extent.end,
2000 lock->l_flags, lock->l_remote_handle.cookie,
2002 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2003 lock->l_pid, lock->l_callback_timeout);
2007 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2008 data->msg_fn, data->msg_line, fmt, args,
2009 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2010 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
2011 "["LPU64"->"LPU64"] flags: "LPX64" remote: "LPX64
2012 " expref: %d pid: %u timeout: %lu\n",
2013 ldlm_lock_to_ns_name(lock), lock,
2014 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2015 lock->l_readers, lock->l_writers,
2016 ldlm_lockname[lock->l_granted_mode],
2017 ldlm_lockname[lock->l_req_mode],
2018 lock->l_resource->lr_name.name[0],
2019 lock->l_resource->lr_name.name[1],
2020 cfs_atomic_read(&lock->l_resource->lr_refcount),
2021 ldlm_typename[lock->l_resource->lr_type],
2022 lock->l_policy_data.l_flock.pid,
2023 lock->l_policy_data.l_flock.start,
2024 lock->l_policy_data.l_flock.end,
2025 lock->l_flags, lock->l_remote_handle.cookie,
2027 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2028 lock->l_pid, lock->l_callback_timeout);
2032 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2033 data->msg_fn, data->msg_line, fmt, args,
2034 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2035 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
2036 "flags: "LPX64" remote: "LPX64" expref: %d "
2037 "pid: %u timeout: %lu\n",
2038 ldlm_lock_to_ns_name(lock),
2039 lock, lock->l_handle.h_cookie,
2040 cfs_atomic_read (&lock->l_refc),
2041 lock->l_readers, lock->l_writers,
2042 ldlm_lockname[lock->l_granted_mode],
2043 ldlm_lockname[lock->l_req_mode],
2044 lock->l_resource->lr_name.name[0],
2045 lock->l_resource->lr_name.name[1],
2046 lock->l_policy_data.l_inodebits.bits,
2047 cfs_atomic_read(&lock->l_resource->lr_refcount),
2048 ldlm_typename[lock->l_resource->lr_type],
2049 lock->l_flags, lock->l_remote_handle.cookie,
2051 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2052 lock->l_pid, lock->l_callback_timeout);
2056 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2057 data->msg_fn, data->msg_line, fmt, args,
2058 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2059 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" "
2060 "remote: "LPX64" expref: %d pid: %u timeout %lu\n",
2061 ldlm_lock_to_ns_name(lock),
2062 lock, lock->l_handle.h_cookie,
2063 cfs_atomic_read (&lock->l_refc),
2064 lock->l_readers, lock->l_writers,
2065 ldlm_lockname[lock->l_granted_mode],
2066 ldlm_lockname[lock->l_req_mode],
2067 lock->l_resource->lr_name.name[0],
2068 lock->l_resource->lr_name.name[1],
2069 cfs_atomic_read(&lock->l_resource->lr_refcount),
2070 ldlm_typename[lock->l_resource->lr_type],
2071 lock->l_flags, lock->l_remote_handle.cookie,
2073 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2074 lock->l_pid, lock->l_callback_timeout);
2079 EXPORT_SYMBOL(_ldlm_lock_debug);