4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
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 # include <linux/lustre_intent.h>
48 # include <liblustre.h>
51 #include <obd_class.h>
52 #include "ldlm_internal.h"
55 char *ldlm_lockname[] = {
67 char *ldlm_typename[] = {
74 static ldlm_policy_wire_to_local_t ldlm_policy_wire18_to_local[] = {
75 [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_wire_to_local,
76 [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_wire_to_local,
77 [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_wire18_to_local,
78 [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_wire_to_local,
81 static ldlm_policy_wire_to_local_t ldlm_policy_wire21_to_local[] = {
82 [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_wire_to_local,
83 [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_wire_to_local,
84 [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_wire21_to_local,
85 [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_wire_to_local,
88 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
89 [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_local_to_wire,
90 [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_local_to_wire,
91 [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_local_to_wire,
92 [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_local_to_wire,
96 * Converts lock policy from local format to on the wire lock_desc format
98 void ldlm_convert_policy_to_wire(ldlm_type_t type,
99 const ldlm_policy_data_t *lpolicy,
100 ldlm_wire_policy_data_t *wpolicy)
102 ldlm_policy_local_to_wire_t convert;
104 convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
106 convert(lpolicy, wpolicy);
110 * Converts lock policy from on the wire lock_desc format to local format
112 void ldlm_convert_policy_to_local(struct obd_export *exp, ldlm_type_t type,
113 const ldlm_wire_policy_data_t *wpolicy,
114 ldlm_policy_data_t *lpolicy)
116 ldlm_policy_wire_to_local_t convert;
119 /** some badnes for 2.0.0 clients, but 2.0.0 isn't supported */
120 new_client = (exp->exp_connect_flags & OBD_CONNECT_FULL20) != 0;
122 convert = ldlm_policy_wire21_to_local[type - LDLM_MIN_TYPE];
124 convert = ldlm_policy_wire18_to_local[type - LDLM_MIN_TYPE];
126 convert(wpolicy, lpolicy);
129 char *ldlm_it2str(int it)
136 case (IT_OPEN | IT_CREAT):
151 CERROR("Unknown intent %d\n", it);
156 extern cfs_mem_cache_t *ldlm_lock_slab;
158 #ifdef HAVE_SERVER_SUPPORT
159 static ldlm_processing_policy ldlm_processing_policy_table[] = {
160 [LDLM_PLAIN] ldlm_process_plain_lock,
161 [LDLM_EXTENT] ldlm_process_extent_lock,
163 [LDLM_FLOCK] ldlm_process_flock_lock,
165 [LDLM_IBITS] ldlm_process_inodebits_lock,
168 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
170 return ldlm_processing_policy_table[res->lr_type];
172 #endif /* HAVE_SERVER_SUPPORT */
174 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
180 * REFCOUNTED LOCK OBJECTS
185 * Lock refcounts, during creation:
186 * - one special one for allocation, dec'd only once in destroy
187 * - one for being a lock that's in-use
188 * - one for the addref associated with a new lock
190 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
192 cfs_atomic_inc(&lock->l_refc);
196 void ldlm_lock_put(struct ldlm_lock *lock)
200 LASSERT(lock->l_resource != LP_POISON);
201 LASSERT(cfs_atomic_read(&lock->l_refc) > 0);
202 if (cfs_atomic_dec_and_test(&lock->l_refc)) {
203 struct ldlm_resource *res;
206 "final lock_put on destroyed lock, freeing it.");
208 res = lock->l_resource;
209 LASSERT(lock->l_destroyed);
210 LASSERT(cfs_list_empty(&lock->l_res_link));
211 LASSERT(cfs_list_empty(&lock->l_pending_chain));
213 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
215 lu_ref_del(&res->lr_reference, "lock", lock);
216 ldlm_resource_putref(res);
217 lock->l_resource = NULL;
218 if (lock->l_export) {
219 class_export_lock_put(lock->l_export, lock);
220 lock->l_export = NULL;
223 if (lock->l_lvb_data != NULL)
224 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
226 ldlm_interval_free(ldlm_interval_detach(lock));
227 lu_ref_fini(&lock->l_reference);
228 OBD_FREE_RCU(lock, sizeof(*lock), &lock->l_handle);
234 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
237 if (!cfs_list_empty(&lock->l_lru)) {
238 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
240 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
241 cfs_list_del_init(&lock->l_lru);
242 if (lock->l_flags & LDLM_FL_SKIPPED)
243 lock->l_flags &= ~LDLM_FL_SKIPPED;
244 LASSERT(ns->ns_nr_unused > 0);
251 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
253 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
257 if (lock->l_ns_srv) {
258 LASSERT(cfs_list_empty(&lock->l_lru));
262 cfs_spin_lock(&ns->ns_lock);
263 rc = ldlm_lock_remove_from_lru_nolock(lock);
264 cfs_spin_unlock(&ns->ns_lock);
269 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
271 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
273 lock->l_last_used = cfs_time_current();
274 LASSERT(cfs_list_empty(&lock->l_lru));
275 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
276 cfs_list_add_tail(&lock->l_lru, &ns->ns_unused_list);
277 LASSERT(ns->ns_nr_unused >= 0);
281 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
283 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
286 cfs_spin_lock(&ns->ns_lock);
287 ldlm_lock_add_to_lru_nolock(lock);
288 cfs_spin_unlock(&ns->ns_lock);
292 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
294 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
297 if (lock->l_ns_srv) {
298 LASSERT(cfs_list_empty(&lock->l_lru));
303 cfs_spin_lock(&ns->ns_lock);
304 if (!cfs_list_empty(&lock->l_lru)) {
305 ldlm_lock_remove_from_lru_nolock(lock);
306 ldlm_lock_add_to_lru_nolock(lock);
308 cfs_spin_unlock(&ns->ns_lock);
312 /* This used to have a 'strict' flag, which recovery would use to mark an
313 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
314 * shall explain why it's gone: with the new hash table scheme, once you call
315 * ldlm_lock_destroy, you can never drop your final references on this lock.
316 * Because it's not in the hash table anymore. -phil */
317 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
321 if (lock->l_readers || lock->l_writers) {
322 LDLM_ERROR(lock, "lock still has references");
326 if (!cfs_list_empty(&lock->l_res_link)) {
327 LDLM_ERROR(lock, "lock still on resource");
331 if (lock->l_destroyed) {
332 LASSERT(cfs_list_empty(&lock->l_lru));
336 lock->l_destroyed = 1;
338 if (lock->l_export && lock->l_export->exp_lock_hash) {
339 /* NB: it's safe to call cfs_hash_del() even lock isn't
340 * in exp_lock_hash. */
341 cfs_hash_del(lock->l_export->exp_lock_hash,
342 &lock->l_remote_handle, &lock->l_exp_hash);
345 ldlm_lock_remove_from_lru(lock);
346 class_handle_unhash(&lock->l_handle);
349 /* Wake anyone waiting for this lock */
350 /* FIXME: I should probably add yet another flag, instead of using
351 * l_export to only call this on clients */
353 class_export_put(lock->l_export);
354 lock->l_export = NULL;
355 if (lock->l_export && lock->l_completion_ast)
356 lock->l_completion_ast(lock, 0);
362 void ldlm_lock_destroy(struct ldlm_lock *lock)
366 lock_res_and_lock(lock);
367 first = ldlm_lock_destroy_internal(lock);
368 unlock_res_and_lock(lock);
370 /* drop reference from hashtable only for first destroy */
372 lu_ref_del(&lock->l_reference, "hash", lock);
373 LDLM_LOCK_RELEASE(lock);
378 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
382 first = ldlm_lock_destroy_internal(lock);
383 /* drop reference from hashtable only for first destroy */
385 lu_ref_del(&lock->l_reference, "hash", lock);
386 LDLM_LOCK_RELEASE(lock);
391 /* this is called by portals_handle2object with the handle lock taken */
392 static void lock_handle_addref(void *lock)
394 LDLM_LOCK_GET((struct ldlm_lock *)lock);
397 static void lock_handle_free(void *lock, int size)
399 LASSERT(size == sizeof(struct ldlm_lock));
400 OBD_SLAB_FREE(lock, ldlm_lock_slab, size);
403 struct portals_handle_ops lock_handle_ops = {
404 .hop_addref = lock_handle_addref,
405 .hop_free = lock_handle_free,
409 * usage: pass in a resource on which you have done ldlm_resource_get
410 * new lock will take over the refcount.
411 * returns: lock with refcount 2 - one for current caller and one for remote
413 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
415 struct ldlm_lock *lock;
418 if (resource == NULL)
421 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, CFS_ALLOC_IO);
425 cfs_spin_lock_init(&lock->l_lock);
426 lock->l_resource = resource;
427 lu_ref_add(&resource->lr_reference, "lock", lock);
429 cfs_atomic_set(&lock->l_refc, 2);
430 CFS_INIT_LIST_HEAD(&lock->l_res_link);
431 CFS_INIT_LIST_HEAD(&lock->l_lru);
432 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
433 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
434 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
435 CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
436 cfs_waitq_init(&lock->l_waitq);
437 lock->l_blocking_lock = NULL;
438 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
439 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
440 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
441 CFS_INIT_HLIST_NODE(&lock->l_exp_flock_hash);
443 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
445 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
446 class_handle_hash(&lock->l_handle, &lock_handle_ops);
448 lu_ref_init(&lock->l_reference);
449 lu_ref_add(&lock->l_reference, "hash", lock);
450 lock->l_callback_timeout = 0;
452 #if LUSTRE_TRACKS_LOCK_EXP_REFS
453 CFS_INIT_LIST_HEAD(&lock->l_exp_refs_link);
454 lock->l_exp_refs_nr = 0;
455 lock->l_exp_refs_target = NULL;
457 CFS_INIT_LIST_HEAD(&lock->l_exp_list);
462 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
463 const struct ldlm_res_id *new_resid)
465 struct ldlm_resource *oldres = lock->l_resource;
466 struct ldlm_resource *newres;
470 LASSERT(ns_is_client(ns));
472 lock_res_and_lock(lock);
473 if (memcmp(new_resid, &lock->l_resource->lr_name,
474 sizeof(lock->l_resource->lr_name)) == 0) {
476 unlock_res_and_lock(lock);
480 LASSERT(new_resid->name[0] != 0);
482 /* This function assumes that the lock isn't on any lists */
483 LASSERT(cfs_list_empty(&lock->l_res_link));
485 type = oldres->lr_type;
486 unlock_res_and_lock(lock);
488 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
492 lu_ref_add(&newres->lr_reference, "lock", lock);
494 * To flip the lock from the old to the new resource, lock, oldres and
495 * newres have to be locked. Resource spin-locks are nested within
496 * lock->l_lock, and are taken in the memory address order to avoid
499 cfs_spin_lock(&lock->l_lock);
500 oldres = lock->l_resource;
501 if (oldres < newres) {
503 lock_res_nested(newres, LRT_NEW);
506 lock_res_nested(oldres, LRT_NEW);
508 LASSERT(memcmp(new_resid, &oldres->lr_name,
509 sizeof oldres->lr_name) != 0);
510 lock->l_resource = newres;
512 unlock_res_and_lock(lock);
514 /* ...and the flowers are still standing! */
515 lu_ref_del(&oldres->lr_reference, "lock", lock);
516 ldlm_resource_putref(oldres);
525 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
527 lockh->cookie = lock->l_handle.h_cookie;
530 /* if flags: atomically get the lock and set the flags.
531 * Return NULL if flag already set
534 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
537 struct ldlm_lock *lock;
542 lock = class_handle2object(handle->cookie);
546 /* It's unlikely but possible that someone marked the lock as
547 * destroyed after we did handle2object on it */
548 if (flags == 0 && !lock->l_destroyed) {
549 lu_ref_add(&lock->l_reference, "handle", cfs_current());
553 lock_res_and_lock(lock);
555 LASSERT(lock->l_resource != NULL);
557 lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
558 if (unlikely(lock->l_destroyed)) {
559 unlock_res_and_lock(lock);
560 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
565 if (flags && (lock->l_flags & flags)) {
566 unlock_res_and_lock(lock);
572 lock->l_flags |= flags;
574 unlock_res_and_lock(lock);
578 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
580 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
581 /* INODEBITS_INTEROP: If the other side does not support
582 * inodebits, reply with a plain lock descriptor.
584 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
585 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
586 /* Make sure all the right bits are set in this lock we
587 are going to pass to client */
588 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
589 (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
590 MDS_INODELOCK_LAYOUT),
591 "Inappropriate inode lock bits during "
592 "conversion " LPU64 "\n",
593 lock->l_policy_data.l_inodebits.bits);
595 ldlm_res2desc(lock->l_resource, &desc->l_resource);
596 desc->l_resource.lr_type = LDLM_PLAIN;
598 /* Convert "new" lock mode to something old client can
600 if ((lock->l_req_mode == LCK_CR) ||
601 (lock->l_req_mode == LCK_CW))
602 desc->l_req_mode = LCK_PR;
604 desc->l_req_mode = lock->l_req_mode;
605 if ((lock->l_granted_mode == LCK_CR) ||
606 (lock->l_granted_mode == LCK_CW)) {
607 desc->l_granted_mode = LCK_PR;
609 /* We never grant PW/EX locks to clients */
610 LASSERT((lock->l_granted_mode != LCK_PW) &&
611 (lock->l_granted_mode != LCK_EX));
612 desc->l_granted_mode = lock->l_granted_mode;
615 /* We do not copy policy here, because there is no
616 policy for plain locks */
618 ldlm_res2desc(lock->l_resource, &desc->l_resource);
619 desc->l_req_mode = lock->l_req_mode;
620 desc->l_granted_mode = lock->l_granted_mode;
621 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
622 &lock->l_policy_data,
623 &desc->l_policy_data);
627 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
628 cfs_list_t *work_list)
630 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
631 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
632 lock->l_flags |= LDLM_FL_AST_SENT;
633 /* If the enqueuing client said so, tell the AST recipient to
634 * discard dirty data, rather than writing back. */
635 if (new->l_flags & LDLM_AST_DISCARD_DATA)
636 lock->l_flags |= LDLM_FL_DISCARD_DATA;
637 LASSERT(cfs_list_empty(&lock->l_bl_ast));
638 cfs_list_add(&lock->l_bl_ast, work_list);
640 LASSERT(lock->l_blocking_lock == NULL);
641 lock->l_blocking_lock = LDLM_LOCK_GET(new);
645 void ldlm_add_cp_work_item(struct ldlm_lock *lock, cfs_list_t *work_list)
647 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
648 lock->l_flags |= LDLM_FL_CP_REQD;
649 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
650 LASSERT(cfs_list_empty(&lock->l_cp_ast));
651 cfs_list_add(&lock->l_cp_ast, work_list);
656 /* must be called with lr_lock held */
657 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
658 cfs_list_t *work_list)
661 check_res_locked(lock->l_resource);
663 ldlm_add_bl_work_item(lock, new, work_list);
665 ldlm_add_cp_work_item(lock, work_list);
669 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
671 struct ldlm_lock *lock;
673 lock = ldlm_handle2lock(lockh);
674 LASSERT(lock != NULL);
675 ldlm_lock_addref_internal(lock, mode);
679 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
681 ldlm_lock_remove_from_lru(lock);
682 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
684 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
686 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
688 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
691 lu_ref_add_atomic(&lock->l_reference, "user", lock);
692 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
696 * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
699 * \retval 0 success, lock was addref-ed
701 * \retval -EAGAIN lock is being canceled.
703 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
705 struct ldlm_lock *lock;
709 lock = ldlm_handle2lock(lockh);
711 lock_res_and_lock(lock);
712 if (lock->l_readers != 0 || lock->l_writers != 0 ||
713 !(lock->l_flags & LDLM_FL_CBPENDING)) {
714 ldlm_lock_addref_internal_nolock(lock, mode);
717 unlock_res_and_lock(lock);
723 /* only called for local locks */
724 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
726 lock_res_and_lock(lock);
727 ldlm_lock_addref_internal_nolock(lock, mode);
728 unlock_res_and_lock(lock);
731 /* only called in ldlm_flock_destroy and for local locks.
732 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
733 * * ldlm_lock_remove_from_lru() does nothing, it is safe
734 * * for ldlm_flock_destroy usage by dropping some code */
735 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
737 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
738 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
739 LASSERT(lock->l_readers > 0);
740 lu_ref_del(&lock->l_reference, "reader", lock);
743 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
744 LASSERT(lock->l_writers > 0);
745 lu_ref_del(&lock->l_reference, "writer", lock);
749 lu_ref_del(&lock->l_reference, "user", lock);
750 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
753 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
755 struct ldlm_namespace *ns;
758 lock_res_and_lock(lock);
760 ns = ldlm_lock_to_ns(lock);
762 ldlm_lock_decref_internal_nolock(lock, mode);
764 if (lock->l_flags & LDLM_FL_LOCAL &&
765 !lock->l_readers && !lock->l_writers) {
766 /* If this is a local lock on a server namespace and this was
767 * the last reference, cancel the lock. */
768 CDEBUG(D_INFO, "forcing cancel of local lock\n");
769 lock->l_flags |= LDLM_FL_CBPENDING;
772 if (!lock->l_readers && !lock->l_writers &&
773 (lock->l_flags & LDLM_FL_CBPENDING)) {
774 /* If we received a blocked AST and this was the last reference,
775 * run the callback. */
776 if (lock->l_ns_srv && lock->l_export)
777 CERROR("FL_CBPENDING set on non-local lock--just a "
780 LDLM_DEBUG(lock, "final decref done on cbpending lock");
782 LDLM_LOCK_GET(lock); /* dropped by bl thread */
783 ldlm_lock_remove_from_lru(lock);
784 unlock_res_and_lock(lock);
786 if (lock->l_flags & LDLM_FL_FAIL_LOC)
787 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
789 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
790 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
791 ldlm_handle_bl_callback(ns, NULL, lock);
792 } else if (ns_is_client(ns) &&
793 !lock->l_readers && !lock->l_writers &&
794 !(lock->l_flags & LDLM_FL_NO_LRU) &&
795 !(lock->l_flags & LDLM_FL_BL_AST)) {
797 LDLM_DEBUG(lock, "add lock into lru list");
799 /* If this is a client-side namespace and this was the last
800 * reference, put it on the LRU. */
801 ldlm_lock_add_to_lru(lock);
802 unlock_res_and_lock(lock);
804 if (lock->l_flags & LDLM_FL_FAIL_LOC)
805 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
807 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
808 * are not supported by the server, otherwise, it is done on
810 if (!exp_connect_cancelset(lock->l_conn_export) &&
811 !ns_connect_lru_resize(ns))
812 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
814 LDLM_DEBUG(lock, "do not add lock into lru list");
815 unlock_res_and_lock(lock);
821 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
823 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
824 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
825 ldlm_lock_decref_internal(lock, mode);
829 /* This will drop a lock reference and mark it for destruction, but will not
830 * necessarily cancel the lock before returning. */
831 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
833 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
836 LASSERT(lock != NULL);
838 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
839 lock_res_and_lock(lock);
840 lock->l_flags |= LDLM_FL_CBPENDING;
841 unlock_res_and_lock(lock);
842 ldlm_lock_decref_internal(lock, mode);
846 struct sl_insert_point {
847 cfs_list_t *res_link;
848 cfs_list_t *mode_link;
849 cfs_list_t *policy_link;
853 * search_granted_lock
856 * Finds a position to insert the new lock.
858 * queue [input]: the granted list where search acts on;
859 * req [input]: the lock whose position to be located;
860 * prev [output]: positions within 3 lists to insert @req to
864 * - ldlm_grant_lock_with_skiplist
866 static void search_granted_lock(cfs_list_t *queue,
867 struct ldlm_lock *req,
868 struct sl_insert_point *prev)
871 struct ldlm_lock *lock, *mode_end, *policy_end;
874 cfs_list_for_each(tmp, queue) {
875 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
877 mode_end = cfs_list_entry(lock->l_sl_mode.prev,
878 struct ldlm_lock, l_sl_mode);
880 if (lock->l_req_mode != req->l_req_mode) {
881 /* jump to last lock of mode group */
882 tmp = &mode_end->l_res_link;
886 /* suitable mode group is found */
887 if (lock->l_resource->lr_type == LDLM_PLAIN) {
888 /* insert point is last lock of the mode group */
889 prev->res_link = &mode_end->l_res_link;
890 prev->mode_link = &mode_end->l_sl_mode;
891 prev->policy_link = &req->l_sl_policy;
894 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
897 cfs_list_entry(lock->l_sl_policy.prev,
901 if (lock->l_policy_data.l_inodebits.bits ==
902 req->l_policy_data.l_inodebits.bits) {
903 /* insert point is last lock of
904 * the policy group */
906 &policy_end->l_res_link;
908 &policy_end->l_sl_mode;
910 &policy_end->l_sl_policy;
915 if (policy_end == mode_end)
916 /* done with mode group */
919 /* go to next policy group within mode group */
920 tmp = policy_end->l_res_link.next;
921 lock = cfs_list_entry(tmp, struct ldlm_lock,
923 } /* loop over policy groups within the mode group */
925 /* insert point is last lock of the mode group,
926 * new policy group is started */
927 prev->res_link = &mode_end->l_res_link;
928 prev->mode_link = &mode_end->l_sl_mode;
929 prev->policy_link = &req->l_sl_policy;
933 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
938 /* insert point is last lock on the queue,
939 * new mode group and new policy group are started */
940 prev->res_link = queue->prev;
941 prev->mode_link = &req->l_sl_mode;
942 prev->policy_link = &req->l_sl_policy;
947 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
948 struct sl_insert_point *prev)
950 struct ldlm_resource *res = lock->l_resource;
953 check_res_locked(res);
955 ldlm_resource_dump(D_INFO, res);
956 LDLM_DEBUG(lock, "About to add lock:");
958 if (lock->l_destroyed) {
959 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
963 LASSERT(cfs_list_empty(&lock->l_res_link));
964 LASSERT(cfs_list_empty(&lock->l_sl_mode));
965 LASSERT(cfs_list_empty(&lock->l_sl_policy));
967 cfs_list_add(&lock->l_res_link, prev->res_link);
968 cfs_list_add(&lock->l_sl_mode, prev->mode_link);
969 cfs_list_add(&lock->l_sl_policy, prev->policy_link);
974 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
976 struct sl_insert_point prev;
979 LASSERT(lock->l_req_mode == lock->l_granted_mode);
981 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
982 ldlm_granted_list_add_lock(lock, &prev);
987 * - ldlm_lock_enqueue
988 * - ldlm_reprocess_queue
989 * - ldlm_lock_convert
991 * must be called with lr_lock held
993 void ldlm_grant_lock(struct ldlm_lock *lock, cfs_list_t *work_list)
995 struct ldlm_resource *res = lock->l_resource;
998 check_res_locked(res);
1000 lock->l_granted_mode = lock->l_req_mode;
1001 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1002 ldlm_grant_lock_with_skiplist(lock);
1003 else if (res->lr_type == LDLM_EXTENT)
1004 ldlm_extent_add_lock(res, lock);
1006 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1008 if (lock->l_granted_mode < res->lr_most_restr)
1009 res->lr_most_restr = lock->l_granted_mode;
1011 if (work_list && lock->l_completion_ast != NULL)
1012 ldlm_add_ast_work_item(lock, NULL, work_list);
1014 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1018 /* returns a referenced lock or NULL. See the flag descriptions below, in the
1019 * comment above ldlm_lock_match */
1020 static struct ldlm_lock *search_queue(cfs_list_t *queue,
1022 ldlm_policy_data_t *policy,
1023 struct ldlm_lock *old_lock,
1024 int flags, int unref)
1026 struct ldlm_lock *lock;
1029 cfs_list_for_each(tmp, queue) {
1032 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1034 if (lock == old_lock)
1037 /* llite sometimes wants to match locks that will be
1038 * canceled when their users drop, but we allow it to match
1039 * if it passes in CBPENDING and the lock still has users.
1040 * this is generally only going to be used by children
1041 * whose parents already hold a lock so forward progress
1042 * can still happen. */
1043 if (lock->l_flags & LDLM_FL_CBPENDING &&
1044 !(flags & LDLM_FL_CBPENDING))
1046 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
1047 lock->l_readers == 0 && lock->l_writers == 0)
1050 if (!(lock->l_req_mode & *mode))
1052 match = lock->l_req_mode;
1054 if (lock->l_resource->lr_type == LDLM_EXTENT &&
1055 (lock->l_policy_data.l_extent.start >
1056 policy->l_extent.start ||
1057 lock->l_policy_data.l_extent.end < policy->l_extent.end))
1060 if (unlikely(match == LCK_GROUP) &&
1061 lock->l_resource->lr_type == LDLM_EXTENT &&
1062 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
1065 /* We match if we have existing lock with same or wider set
1067 if (lock->l_resource->lr_type == LDLM_IBITS &&
1068 ((lock->l_policy_data.l_inodebits.bits &
1069 policy->l_inodebits.bits) !=
1070 policy->l_inodebits.bits))
1074 (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED ||
1078 if ((flags & LDLM_FL_LOCAL_ONLY) &&
1079 !(lock->l_flags & LDLM_FL_LOCAL))
1082 if (flags & LDLM_FL_TEST_LOCK) {
1083 LDLM_LOCK_GET(lock);
1084 ldlm_lock_touch_in_lru(lock);
1086 ldlm_lock_addref_internal_nolock(lock, match);
1095 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1097 if (!lock->l_failed) {
1099 cfs_waitq_broadcast(&lock->l_waitq);
1102 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1104 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1106 lock_res_and_lock(lock);
1107 ldlm_lock_fail_match_locked(lock);
1108 unlock_res_and_lock(lock);
1110 EXPORT_SYMBOL(ldlm_lock_fail_match);
1112 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1114 lock->l_flags |= LDLM_FL_LVB_READY;
1115 cfs_waitq_broadcast(&lock->l_waitq);
1118 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1120 lock_res_and_lock(lock);
1121 ldlm_lock_allow_match_locked(lock);
1122 unlock_res_and_lock(lock);
1125 /* Can be called in two ways:
1127 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1128 * for a duplicate of.
1130 * Otherwise, all of the fields must be filled in, to match against.
1132 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1133 * server (ie, connh is NULL)
1134 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1135 * list will be considered
1136 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1137 * to be canceled can still be matched as long as they still have reader
1138 * or writer refernces
1139 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1140 * just tell us if we would have matched.
1142 * Returns 1 if it finds an already-existing lock that is compatible; in this
1143 * case, lockh is filled in with a addref()ed lock
1145 * we also check security context, if that failed we simply return 0 (to keep
1146 * caller code unchanged), the context failure will be discovered by caller
1149 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1150 const struct ldlm_res_id *res_id, ldlm_type_t type,
1151 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1152 struct lustre_handle *lockh, int unref)
1154 struct ldlm_resource *res;
1155 struct ldlm_lock *lock, *old_lock = NULL;
1160 old_lock = ldlm_handle2lock(lockh);
1163 ns = ldlm_lock_to_ns(old_lock);
1164 res_id = &old_lock->l_resource->lr_name;
1165 type = old_lock->l_resource->lr_type;
1166 mode = old_lock->l_req_mode;
1169 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1171 LASSERT(old_lock == NULL);
1175 LDLM_RESOURCE_ADDREF(res);
1178 lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1182 if (flags & LDLM_FL_BLOCK_GRANTED)
1184 lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1188 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1196 LDLM_RESOURCE_DELREF(res);
1197 ldlm_resource_putref(res);
1200 ldlm_lock2handle(lock, lockh);
1201 if ((flags & LDLM_FL_LVB_READY) &&
1202 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1203 struct l_wait_info lwi;
1204 if (lock->l_completion_ast) {
1205 int err = lock->l_completion_ast(lock,
1206 LDLM_FL_WAIT_NOREPROC,
1209 if (flags & LDLM_FL_TEST_LOCK)
1210 LDLM_LOCK_RELEASE(lock);
1212 ldlm_lock_decref_internal(lock,
1219 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1220 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1222 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1223 l_wait_event(lock->l_waitq,
1224 lock->l_flags & LDLM_FL_LVB_READY ||
1227 if (!(lock->l_flags & LDLM_FL_LVB_READY)) {
1228 if (flags & LDLM_FL_TEST_LOCK)
1229 LDLM_LOCK_RELEASE(lock);
1231 ldlm_lock_decref_internal(lock, mode);
1238 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1239 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1240 res_id->name[2] : policy->l_extent.start,
1241 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1242 res_id->name[3] : policy->l_extent.end);
1244 /* check user's security context */
1245 if (lock->l_conn_export &&
1246 sptlrpc_import_check_ctx(
1247 class_exp2cliimp(lock->l_conn_export))) {
1248 if (!(flags & LDLM_FL_TEST_LOCK))
1249 ldlm_lock_decref_internal(lock, mode);
1253 if (flags & LDLM_FL_TEST_LOCK)
1254 LDLM_LOCK_RELEASE(lock);
1256 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1257 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1258 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1259 type, mode, res_id->name[0], res_id->name[1],
1260 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1261 res_id->name[2] :policy->l_extent.start,
1262 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1263 res_id->name[3] : policy->l_extent.end);
1266 LDLM_LOCK_PUT(old_lock);
1268 return rc ? mode : 0;
1271 ldlm_mode_t ldlm_revalidate_lock_handle(struct lustre_handle *lockh,
1274 struct ldlm_lock *lock;
1275 ldlm_mode_t mode = 0;
1278 lock = ldlm_handle2lock(lockh);
1280 lock_res_and_lock(lock);
1281 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED ||
1285 if (lock->l_flags & LDLM_FL_CBPENDING &&
1286 lock->l_readers == 0 && lock->l_writers == 0)
1290 *bits = lock->l_policy_data.l_inodebits.bits;
1291 mode = lock->l_granted_mode;
1292 ldlm_lock_addref_internal_nolock(lock, mode);
1299 unlock_res_and_lock(lock);
1300 LDLM_LOCK_PUT(lock);
1304 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1306 /* Returns a referenced lock */
1307 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1308 const struct ldlm_res_id *res_id,
1311 const struct ldlm_callback_suite *cbs,
1312 void *data, __u32 lvb_len)
1314 struct ldlm_lock *lock;
1315 struct ldlm_resource *res;
1318 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1322 lock = ldlm_lock_new(res);
1327 lock->l_req_mode = mode;
1328 lock->l_ast_data = data;
1329 lock->l_pid = cfs_curproc_pid();
1330 lock->l_ns_srv = !!ns_is_server(ns);
1332 lock->l_blocking_ast = cbs->lcs_blocking;
1333 lock->l_completion_ast = cbs->lcs_completion;
1334 lock->l_glimpse_ast = cbs->lcs_glimpse;
1335 lock->l_weigh_ast = cbs->lcs_weigh;
1338 lock->l_tree_node = NULL;
1339 /* if this is the extent lock, allocate the interval tree node */
1340 if (type == LDLM_EXTENT) {
1341 if (ldlm_interval_alloc(lock) == NULL)
1346 lock->l_lvb_len = lvb_len;
1347 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1348 if (lock->l_lvb_data == NULL)
1352 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1358 ldlm_lock_destroy(lock);
1359 LDLM_LOCK_RELEASE(lock);
1363 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1364 struct ldlm_lock **lockp,
1365 void *cookie, int *flags)
1367 struct ldlm_lock *lock = *lockp;
1368 struct ldlm_resource *res = lock->l_resource;
1369 int local = ns_is_client(ldlm_res_to_ns(res));
1370 #ifdef HAVE_SERVER_SUPPORT
1371 ldlm_processing_policy policy;
1373 ldlm_error_t rc = ELDLM_OK;
1374 struct ldlm_interval *node = NULL;
1377 lock->l_last_activity = cfs_time_current_sec();
1378 /* policies are not executed on the client or during replay */
1379 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1380 && !local && ns->ns_policy) {
1381 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1383 if (rc == ELDLM_LOCK_REPLACED) {
1384 /* The lock that was returned has already been granted,
1385 * and placed into lockp. If it's not the same as the
1386 * one we passed in, then destroy the old one and our
1387 * work here is done. */
1388 if (lock != *lockp) {
1389 ldlm_lock_destroy(lock);
1390 LDLM_LOCK_RELEASE(lock);
1392 *flags |= LDLM_FL_LOCK_CHANGED;
1394 } else if (rc != ELDLM_OK ||
1395 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1396 ldlm_lock_destroy(lock);
1401 /* For a replaying lock, it might be already in granted list. So
1402 * unlinking the lock will cause the interval node to be freed, we
1403 * have to allocate the interval node early otherwise we can't regrant
1404 * this lock in the future. - jay */
1405 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1406 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1408 lock_res_and_lock(lock);
1409 if (local && lock->l_req_mode == lock->l_granted_mode) {
1410 /* The server returned a blocked lock, but it was granted
1411 * before we got a chance to actually enqueue it. We don't
1412 * need to do anything else. */
1413 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1414 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1415 GOTO(out, ELDLM_OK);
1418 ldlm_resource_unlink_lock(lock);
1419 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1421 ldlm_lock_destroy_nolock(lock);
1422 GOTO(out, rc = -ENOMEM);
1425 CFS_INIT_LIST_HEAD(&node->li_group);
1426 ldlm_interval_attach(node, lock);
1430 /* Some flags from the enqueue want to make it into the AST, via the
1431 * lock's l_flags. */
1432 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1434 /* This distinction between local lock trees is very important; a client
1435 * namespace only has information about locks taken by that client, and
1436 * thus doesn't have enough information to decide for itself if it can
1437 * be granted (below). In this case, we do exactly what the server
1438 * tells us to do, as dictated by the 'flags'.
1440 * We do exactly the same thing during recovery, when the server is
1441 * more or less trusting the clients not to lie.
1443 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1444 * granted/converting queues. */
1446 if (*flags & LDLM_FL_BLOCK_CONV)
1447 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1448 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1449 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1451 ldlm_grant_lock(lock, NULL);
1452 GOTO(out, ELDLM_OK);
1453 #ifdef HAVE_SERVER_SUPPORT
1454 } else if (*flags & LDLM_FL_REPLAY) {
1455 if (*flags & LDLM_FL_BLOCK_CONV) {
1456 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1457 GOTO(out, ELDLM_OK);
1458 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1459 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1460 GOTO(out, ELDLM_OK);
1461 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1462 ldlm_grant_lock(lock, NULL);
1463 GOTO(out, ELDLM_OK);
1465 /* If no flags, fall through to normal enqueue path. */
1468 policy = ldlm_processing_policy_table[res->lr_type];
1469 policy(lock, flags, 1, &rc, NULL);
1473 CERROR("This is client-side-only module, cannot handle "
1474 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1480 unlock_res_and_lock(lock);
1482 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1486 #ifdef HAVE_SERVER_SUPPORT
1487 /* Must be called with namespace taken: queue is waiting or converting. */
1488 int ldlm_reprocess_queue(struct ldlm_resource *res, cfs_list_t *queue,
1489 cfs_list_t *work_list)
1491 cfs_list_t *tmp, *pos;
1492 ldlm_processing_policy policy;
1494 int rc = LDLM_ITER_CONTINUE;
1498 check_res_locked(res);
1500 policy = ldlm_processing_policy_table[res->lr_type];
1503 cfs_list_for_each_safe(tmp, pos, queue) {
1504 struct ldlm_lock *pending;
1505 pending = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1507 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1510 rc = policy(pending, &flags, 0, &err, work_list);
1511 if (rc != LDLM_ITER_CONTINUE)
1520 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
1522 struct ldlm_cb_set_arg *arg = opaq;
1523 struct ldlm_lock_desc d;
1525 struct ldlm_lock *lock;
1528 if (cfs_list_empty(arg->list))
1531 lock = cfs_list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
1533 /* nobody should touch l_bl_ast */
1534 lock_res_and_lock(lock);
1535 cfs_list_del_init(&lock->l_bl_ast);
1537 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1538 LASSERT(lock->l_bl_ast_run == 0);
1539 LASSERT(lock->l_blocking_lock);
1540 lock->l_bl_ast_run++;
1541 unlock_res_and_lock(lock);
1543 ldlm_lock2desc(lock->l_blocking_lock, &d);
1545 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
1546 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1547 lock->l_blocking_lock = NULL;
1548 LDLM_LOCK_RELEASE(lock);
1554 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
1556 struct ldlm_cb_set_arg *arg = opaq;
1558 struct ldlm_lock *lock;
1559 ldlm_completion_callback completion_callback;
1562 if (cfs_list_empty(arg->list))
1565 lock = cfs_list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
1567 /* It's possible to receive a completion AST before we've set
1568 * the l_completion_ast pointer: either because the AST arrived
1569 * before the reply, or simply because there's a small race
1570 * window between receiving the reply and finishing the local
1571 * enqueue. (bug 842)
1573 * This can't happen with the blocking_ast, however, because we
1574 * will never call the local blocking_ast until we drop our
1575 * reader/writer reference, which we won't do until we get the
1576 * reply and finish enqueueing. */
1578 /* nobody should touch l_cp_ast */
1579 lock_res_and_lock(lock);
1580 cfs_list_del_init(&lock->l_cp_ast);
1581 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1582 /* save l_completion_ast since it can be changed by
1583 * mds_intent_policy(), see bug 14225 */
1584 completion_callback = lock->l_completion_ast;
1585 lock->l_flags &= ~LDLM_FL_CP_REQD;
1586 unlock_res_and_lock(lock);
1588 if (completion_callback != NULL)
1589 rc = completion_callback(lock, 0, (void *)arg);
1590 LDLM_LOCK_RELEASE(lock);
1596 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
1598 struct ldlm_cb_set_arg *arg = opaq;
1599 struct ldlm_lock_desc desc;
1601 struct ldlm_lock *lock;
1604 if (cfs_list_empty(arg->list))
1607 lock = cfs_list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
1608 cfs_list_del_init(&lock->l_rk_ast);
1610 /* the desc just pretend to exclusive */
1611 ldlm_lock2desc(lock, &desc);
1612 desc.l_req_mode = LCK_EX;
1613 desc.l_granted_mode = 0;
1615 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1616 LDLM_LOCK_RELEASE(lock);
1621 int ldlm_run_ast_work(struct ldlm_namespace *ns, cfs_list_t *rpc_list,
1622 ldlm_desc_ast_t ast_type)
1624 struct ldlm_cb_set_arg *arg;
1625 set_producer_func work_ast_lock;
1628 if (cfs_list_empty(rpc_list))
1635 cfs_atomic_set(&arg->restart, 0);
1636 arg->list = rpc_list;
1639 case LDLM_WORK_BL_AST:
1640 arg->type = LDLM_BL_CALLBACK;
1641 work_ast_lock = ldlm_work_bl_ast_lock;
1643 case LDLM_WORK_CP_AST:
1644 arg->type = LDLM_CP_CALLBACK;
1645 work_ast_lock = ldlm_work_cp_ast_lock;
1647 case LDLM_WORK_REVOKE_AST:
1648 arg->type = LDLM_BL_CALLBACK;
1649 work_ast_lock = ldlm_work_revoke_ast_lock;
1655 /* We create a ptlrpc request set with flow control extension.
1656 * This request set will use the work_ast_lock function to produce new
1657 * requests and will send a new request each time one completes in order
1658 * to keep the number of requests in flight to ns_max_parallel_ast */
1659 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
1660 work_ast_lock, arg);
1661 if (arg->set == NULL)
1662 GOTO(out, rc = -ENOMEM);
1664 ptlrpc_set_wait(arg->set);
1665 ptlrpc_set_destroy(arg->set);
1667 rc = cfs_atomic_read(&arg->restart) ? -ERESTART : 0;
1674 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1676 ldlm_reprocess_all(res);
1677 return LDLM_ITER_CONTINUE;
1680 static int ldlm_reprocess_res(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1681 cfs_hlist_node_t *hnode, void *arg)
1683 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1686 rc = reprocess_one_queue(res, arg);
1688 return rc == LDLM_ITER_STOP;
1691 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1696 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1697 ldlm_reprocess_res, NULL);
1702 void ldlm_reprocess_all(struct ldlm_resource *res)
1704 CFS_LIST_HEAD(rpc_list);
1706 #ifdef HAVE_SERVER_SUPPORT
1709 /* Local lock trees don't get reprocessed. */
1710 if (ns_is_client(ldlm_res_to_ns(res))) {
1717 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1718 if (rc == LDLM_ITER_CONTINUE)
1719 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1722 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
1724 if (rc == -ERESTART) {
1725 LASSERT(cfs_list_empty(&rpc_list));
1730 if (!ns_is_client(ldlm_res_to_ns(res))) {
1731 CERROR("This is client-side-only module, cannot handle "
1732 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1739 void ldlm_cancel_callback(struct ldlm_lock *lock)
1741 check_res_locked(lock->l_resource);
1742 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1743 lock->l_flags |= LDLM_FL_CANCEL;
1744 if (lock->l_blocking_ast) {
1745 // l_check_no_ns_lock(ns);
1746 unlock_res_and_lock(lock);
1747 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1749 lock_res_and_lock(lock);
1751 LDLM_DEBUG(lock, "no blocking ast");
1754 lock->l_flags |= LDLM_FL_BL_DONE;
1757 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1759 if (req->l_resource->lr_type != LDLM_PLAIN &&
1760 req->l_resource->lr_type != LDLM_IBITS)
1763 cfs_list_del_init(&req->l_sl_policy);
1764 cfs_list_del_init(&req->l_sl_mode);
1767 void ldlm_lock_cancel(struct ldlm_lock *lock)
1769 struct ldlm_resource *res;
1770 struct ldlm_namespace *ns;
1773 lock_res_and_lock(lock);
1775 res = lock->l_resource;
1776 ns = ldlm_res_to_ns(res);
1778 /* Please do not, no matter how tempting, remove this LBUG without
1779 * talking to me first. -phik */
1780 if (lock->l_readers || lock->l_writers) {
1781 LDLM_ERROR(lock, "lock still has references");
1785 ldlm_del_waiting_lock(lock);
1787 /* Releases cancel callback. */
1788 ldlm_cancel_callback(lock);
1790 /* Yes, second time, just in case it was added again while we were
1791 running with no res lock in ldlm_cancel_callback */
1792 ldlm_del_waiting_lock(lock);
1793 ldlm_resource_unlink_lock(lock);
1794 ldlm_lock_destroy_nolock(lock);
1796 if (lock->l_granted_mode == lock->l_req_mode)
1797 ldlm_pool_del(&ns->ns_pool, lock);
1799 /* Make sure we will not be called again for same lock what is possible
1800 * if not to zero out lock->l_granted_mode */
1801 lock->l_granted_mode = LCK_MINMODE;
1802 unlock_res_and_lock(lock);
1807 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1809 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1814 if (lock->l_ast_data == NULL)
1815 lock->l_ast_data = data;
1816 if (lock->l_ast_data == data)
1818 LDLM_LOCK_PUT(lock);
1822 EXPORT_SYMBOL(ldlm_lock_set_data);
1824 struct export_cl_data {
1825 struct obd_export *ecl_exp;
1829 int ldlm_cancel_locks_for_export_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1830 cfs_hlist_node_t *hnode, void *data)
1833 struct export_cl_data *ecl = (struct export_cl_data *)data;
1834 struct obd_export *exp = ecl->ecl_exp;
1835 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
1836 struct ldlm_resource *res;
1838 res = ldlm_resource_getref(lock->l_resource);
1839 LDLM_LOCK_GET(lock);
1841 LDLM_DEBUG(lock, "export %p", exp);
1842 ldlm_res_lvbo_update(res, NULL, 1);
1843 ldlm_lock_cancel(lock);
1844 ldlm_reprocess_all(res);
1845 ldlm_resource_putref(res);
1846 LDLM_LOCK_RELEASE(lock);
1849 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
1851 "Cancel lock %p for export %p (loop %d), still have "
1852 "%d locks left on hash table.\n",
1853 lock, exp, ecl->ecl_loop,
1854 cfs_atomic_read(&hs->hs_count));
1860 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1862 struct export_cl_data ecl = {
1867 cfs_hash_for_each_empty(exp->exp_lock_hash,
1868 ldlm_cancel_locks_for_export_cb, &ecl);
1872 * Downgrade an exclusive lock.
1874 * A fast variant of ldlm_lock_convert for convertion of exclusive
1875 * locks. The convertion is always successful.
1877 * \param lock A lock to convert
1878 * \param new_mode new lock mode
1880 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1884 LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1885 LASSERT(new_mode == LCK_COS);
1887 lock_res_and_lock(lock);
1888 ldlm_resource_unlink_lock(lock);
1890 * Remove the lock from pool as it will be added again in
1891 * ldlm_grant_lock() called below.
1893 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
1895 lock->l_req_mode = new_mode;
1896 ldlm_grant_lock(lock, NULL);
1897 unlock_res_and_lock(lock);
1898 ldlm_reprocess_all(lock->l_resource);
1903 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1906 CFS_LIST_HEAD(rpc_list);
1907 struct ldlm_resource *res;
1908 struct ldlm_namespace *ns;
1910 #ifdef HAVE_SERVER_SUPPORT
1912 struct sl_insert_point prev;
1914 struct ldlm_interval *node;
1917 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1918 *flags |= LDLM_FL_BLOCK_GRANTED;
1919 RETURN(lock->l_resource);
1922 /* I can't check the type of lock here because the bitlock of lock
1923 * is not held here, so do the allocation blindly. -jay */
1924 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1925 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1928 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1929 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1931 lock_res_and_lock(lock);
1933 res = lock->l_resource;
1934 ns = ldlm_res_to_ns(res);
1936 #ifdef HAVE_SERVER_SUPPORT
1937 old_mode = lock->l_req_mode;
1939 lock->l_req_mode = new_mode;
1940 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1941 #ifdef HAVE_SERVER_SUPPORT
1942 /* remember the lock position where the lock might be
1943 * added back to the granted list later and also
1944 * remember the join mode for skiplist fixing. */
1945 prev.res_link = lock->l_res_link.prev;
1946 prev.mode_link = lock->l_sl_mode.prev;
1947 prev.policy_link = lock->l_sl_policy.prev;
1949 ldlm_resource_unlink_lock(lock);
1951 ldlm_resource_unlink_lock(lock);
1952 if (res->lr_type == LDLM_EXTENT) {
1953 /* FIXME: ugly code, I have to attach the lock to a
1954 * interval node again since perhaps it will be granted
1956 CFS_INIT_LIST_HEAD(&node->li_group);
1957 ldlm_interval_attach(node, lock);
1963 * Remove old lock from the pool before adding the lock with new
1964 * mode below in ->policy()
1966 ldlm_pool_del(&ns->ns_pool, lock);
1968 /* If this is a local resource, put it on the appropriate list. */
1969 if (ns_is_client(ldlm_res_to_ns(res))) {
1970 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1971 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1973 /* This should never happen, because of the way the
1974 * server handles conversions. */
1975 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1979 ldlm_grant_lock(lock, &rpc_list);
1981 /* FIXME: completion handling not with lr_lock held ! */
1982 if (lock->l_completion_ast)
1983 lock->l_completion_ast(lock, 0, NULL);
1985 #ifdef HAVE_SERVER_SUPPORT
1990 ldlm_processing_policy policy;
1991 policy = ldlm_processing_policy_table[res->lr_type];
1992 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1993 if (rc == LDLM_ITER_STOP) {
1994 lock->l_req_mode = old_mode;
1995 if (res->lr_type == LDLM_EXTENT)
1996 ldlm_extent_add_lock(res, lock);
1998 ldlm_granted_list_add_lock(lock, &prev);
2002 *flags |= LDLM_FL_BLOCK_GRANTED;
2008 CERROR("This is client-side-only module, cannot handle "
2009 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2013 unlock_res_and_lock(lock);
2016 ldlm_run_ast_work(ns, &rpc_list, LDLM_WORK_CP_AST);
2018 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
2022 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
2024 struct ldlm_lock *lock;
2026 if (!((libcfs_debug | D_ERROR) & level))
2029 lock = ldlm_handle2lock(lockh);
2033 LDLM_DEBUG_LIMIT(level, lock, "###");
2035 LDLM_LOCK_PUT(lock);
2038 void _ldlm_lock_debug(struct ldlm_lock *lock,
2039 struct libcfs_debug_msg_data *msgdata,
2040 const char *fmt, ...)
2043 struct obd_export *exp = lock->l_export;
2044 struct ldlm_resource *resource = lock->l_resource;
2045 char *nid = "local";
2047 va_start(args, fmt);
2049 if (exp && exp->exp_connection) {
2050 nid = libcfs_nid2str(exp->exp_connection->c_peer.nid);
2051 } else if (exp && exp->exp_obd != NULL) {
2052 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2053 nid = libcfs_nid2str(imp->imp_connection->c_peer.nid);
2056 if (resource == NULL) {
2057 libcfs_debug_vmsg2(msgdata, fmt, args,
2058 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2059 "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" nid: %s "
2060 "remote: "LPX64" expref: %d pid: %u timeout: %lu\n",
2062 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2063 lock->l_readers, lock->l_writers,
2064 ldlm_lockname[lock->l_granted_mode],
2065 ldlm_lockname[lock->l_req_mode],
2066 lock->l_flags, nid, lock->l_remote_handle.cookie,
2067 exp ? cfs_atomic_read(&exp->exp_refcount) : -99,
2068 lock->l_pid, lock->l_callback_timeout);
2073 switch (resource->lr_type) {
2075 libcfs_debug_vmsg2(msgdata, fmt, args,
2076 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2077 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
2078 "] (req "LPU64"->"LPU64") flags: "LPX64" nid: %s remote:"
2079 " "LPX64" expref: %d pid: %u timeout %lu\n",
2080 ldlm_lock_to_ns_name(lock), lock,
2081 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2082 lock->l_readers, lock->l_writers,
2083 ldlm_lockname[lock->l_granted_mode],
2084 ldlm_lockname[lock->l_req_mode],
2085 resource->lr_name.name[0],
2086 resource->lr_name.name[1],
2087 cfs_atomic_read(&resource->lr_refcount),
2088 ldlm_typename[resource->lr_type],
2089 lock->l_policy_data.l_extent.start,
2090 lock->l_policy_data.l_extent.end,
2091 lock->l_req_extent.start, lock->l_req_extent.end,
2092 lock->l_flags, nid, lock->l_remote_handle.cookie,
2093 exp ? cfs_atomic_read(&exp->exp_refcount) : -99,
2094 lock->l_pid, lock->l_callback_timeout);
2098 libcfs_debug_vmsg2(msgdata, fmt, args,
2099 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2100 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
2101 "["LPU64"->"LPU64"] flags: "LPX64" nid: %s remote: "LPX64
2102 " expref: %d pid: %u timeout: %lu\n",
2103 ldlm_lock_to_ns_name(lock), lock,
2104 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2105 lock->l_readers, lock->l_writers,
2106 ldlm_lockname[lock->l_granted_mode],
2107 ldlm_lockname[lock->l_req_mode],
2108 resource->lr_name.name[0],
2109 resource->lr_name.name[1],
2110 cfs_atomic_read(&resource->lr_refcount),
2111 ldlm_typename[resource->lr_type],
2112 lock->l_policy_data.l_flock.pid,
2113 lock->l_policy_data.l_flock.start,
2114 lock->l_policy_data.l_flock.end,
2115 lock->l_flags, nid, lock->l_remote_handle.cookie,
2116 exp ? cfs_atomic_read(&exp->exp_refcount) : -99,
2117 lock->l_pid, lock->l_callback_timeout);
2121 libcfs_debug_vmsg2(msgdata, fmt, args,
2122 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2123 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
2124 "flags: "LPX64" nid: %s remote: "LPX64" expref: %d "
2125 "pid: %u timeout: %lu\n",
2126 ldlm_lock_to_ns_name(lock),
2127 lock, lock->l_handle.h_cookie,
2128 cfs_atomic_read (&lock->l_refc),
2129 lock->l_readers, lock->l_writers,
2130 ldlm_lockname[lock->l_granted_mode],
2131 ldlm_lockname[lock->l_req_mode],
2132 resource->lr_name.name[0],
2133 resource->lr_name.name[1],
2134 lock->l_policy_data.l_inodebits.bits,
2135 cfs_atomic_read(&resource->lr_refcount),
2136 ldlm_typename[resource->lr_type],
2137 lock->l_flags, nid, lock->l_remote_handle.cookie,
2138 exp ? cfs_atomic_read(&exp->exp_refcount) : -99,
2139 lock->l_pid, lock->l_callback_timeout);
2143 libcfs_debug_vmsg2(msgdata, fmt, args,
2144 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2145 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" "
2146 "nid: %s remote: "LPX64" expref: %d pid: %u timeout %lu"
2148 ldlm_lock_to_ns_name(lock),
2149 lock, lock->l_handle.h_cookie,
2150 cfs_atomic_read (&lock->l_refc),
2151 lock->l_readers, lock->l_writers,
2152 ldlm_lockname[lock->l_granted_mode],
2153 ldlm_lockname[lock->l_req_mode],
2154 resource->lr_name.name[0],
2155 resource->lr_name.name[1],
2156 cfs_atomic_read(&resource->lr_refcount),
2157 ldlm_typename[resource->lr_type],
2158 lock->l_flags, nid, lock->l_remote_handle.cookie,
2159 exp ? cfs_atomic_read(&exp->exp_refcount) : -99,
2160 lock->l_pid, lock->l_callback_timeout);
2165 EXPORT_SYMBOL(_ldlm_lock_debug);