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):
141 CERROR("Unknown intent %d\n", it);
146 extern cfs_mem_cache_t *ldlm_lock_slab;
148 static ldlm_processing_policy ldlm_processing_policy_table[] = {
149 [LDLM_PLAIN] ldlm_process_plain_lock,
150 [LDLM_EXTENT] ldlm_process_extent_lock,
152 [LDLM_FLOCK] ldlm_process_flock_lock,
154 [LDLM_IBITS] ldlm_process_inodebits_lock,
157 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
159 return ldlm_processing_policy_table[res->lr_type];
162 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
168 * REFCOUNTED LOCK OBJECTS
173 * Lock refcounts, during creation:
174 * - one special one for allocation, dec'd only once in destroy
175 * - one for being a lock that's in-use
176 * - one for the addref associated with a new lock
178 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
180 cfs_atomic_inc(&lock->l_refc);
184 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
186 LASSERT(size == sizeof(*lock));
187 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
190 void ldlm_lock_put(struct ldlm_lock *lock)
194 LASSERT(lock->l_resource != LP_POISON);
195 LASSERT(cfs_atomic_read(&lock->l_refc) > 0);
196 if (cfs_atomic_dec_and_test(&lock->l_refc)) {
197 struct ldlm_resource *res;
200 "final lock_put on destroyed lock, freeing it.");
202 res = lock->l_resource;
203 LASSERT(lock->l_destroyed);
204 LASSERT(cfs_list_empty(&lock->l_res_link));
205 LASSERT(cfs_list_empty(&lock->l_pending_chain));
207 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
209 lu_ref_del(&res->lr_reference, "lock", lock);
210 ldlm_resource_putref(res);
211 lock->l_resource = NULL;
212 if (lock->l_export) {
213 class_export_lock_put(lock->l_export, lock);
214 lock->l_export = NULL;
217 if (lock->l_lvb_data != NULL)
218 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
220 ldlm_interval_free(ldlm_interval_detach(lock));
221 lu_ref_fini(&lock->l_reference);
222 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
229 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
232 if (!cfs_list_empty(&lock->l_lru)) {
233 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
235 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
236 cfs_list_del_init(&lock->l_lru);
237 if (lock->l_flags & LDLM_FL_SKIPPED)
238 lock->l_flags &= ~LDLM_FL_SKIPPED;
239 LASSERT(ns->ns_nr_unused > 0);
246 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
248 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
252 if (lock->l_ns_srv) {
253 LASSERT(cfs_list_empty(&lock->l_lru));
257 cfs_spin_lock(&ns->ns_lock);
258 rc = ldlm_lock_remove_from_lru_nolock(lock);
259 cfs_spin_unlock(&ns->ns_lock);
264 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
266 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
268 lock->l_last_used = cfs_time_current();
269 LASSERT(cfs_list_empty(&lock->l_lru));
270 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
271 cfs_list_add_tail(&lock->l_lru, &ns->ns_unused_list);
272 LASSERT(ns->ns_nr_unused >= 0);
276 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
278 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
281 cfs_spin_lock(&ns->ns_lock);
282 ldlm_lock_add_to_lru_nolock(lock);
283 cfs_spin_unlock(&ns->ns_lock);
287 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
289 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
292 if (lock->l_ns_srv) {
293 LASSERT(cfs_list_empty(&lock->l_lru));
298 cfs_spin_lock(&ns->ns_lock);
299 if (!cfs_list_empty(&lock->l_lru)) {
300 ldlm_lock_remove_from_lru_nolock(lock);
301 ldlm_lock_add_to_lru_nolock(lock);
303 cfs_spin_unlock(&ns->ns_lock);
307 /* This used to have a 'strict' flag, which recovery would use to mark an
308 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
309 * shall explain why it's gone: with the new hash table scheme, once you call
310 * ldlm_lock_destroy, you can never drop your final references on this lock.
311 * Because it's not in the hash table anymore. -phil */
312 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
316 if (lock->l_readers || lock->l_writers) {
317 LDLM_ERROR(lock, "lock still has references");
318 ldlm_lock_dump(D_ERROR, lock, 0);
322 if (!cfs_list_empty(&lock->l_res_link)) {
323 LDLM_ERROR(lock, "lock still on resource");
324 ldlm_lock_dump(D_ERROR, lock, 0);
328 if (lock->l_destroyed) {
329 LASSERT(cfs_list_empty(&lock->l_lru));
333 lock->l_destroyed = 1;
335 if (lock->l_export && lock->l_export->exp_lock_hash &&
336 !cfs_hlist_unhashed(&lock->l_exp_hash))
337 cfs_hash_del(lock->l_export->exp_lock_hash,
338 &lock->l_remote_handle, &lock->l_exp_hash);
340 ldlm_lock_remove_from_lru(lock);
341 class_handle_unhash(&lock->l_handle);
344 /* Wake anyone waiting for this lock */
345 /* FIXME: I should probably add yet another flag, instead of using
346 * l_export to only call this on clients */
348 class_export_put(lock->l_export);
349 lock->l_export = NULL;
350 if (lock->l_export && lock->l_completion_ast)
351 lock->l_completion_ast(lock, 0);
357 void ldlm_lock_destroy(struct ldlm_lock *lock)
361 lock_res_and_lock(lock);
362 first = ldlm_lock_destroy_internal(lock);
363 unlock_res_and_lock(lock);
365 /* drop reference from hashtable only for first destroy */
367 lu_ref_del(&lock->l_reference, "hash", lock);
368 LDLM_LOCK_RELEASE(lock);
373 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
377 first = ldlm_lock_destroy_internal(lock);
378 /* drop reference from hashtable only for first destroy */
380 lu_ref_del(&lock->l_reference, "hash", lock);
381 LDLM_LOCK_RELEASE(lock);
386 /* this is called by portals_handle2object with the handle lock taken */
387 static void lock_handle_addref(void *lock)
389 LDLM_LOCK_GET((struct ldlm_lock *)lock);
393 * usage: pass in a resource on which you have done ldlm_resource_get
394 * new lock will take over the refcount.
395 * returns: lock with refcount 2 - one for current caller and one for remote
397 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
399 struct ldlm_lock *lock;
402 if (resource == NULL)
405 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, CFS_ALLOC_IO);
409 cfs_spin_lock_init(&lock->l_lock);
410 lock->l_resource = resource;
411 lu_ref_add(&resource->lr_reference, "lock", lock);
413 cfs_atomic_set(&lock->l_refc, 2);
414 CFS_INIT_LIST_HEAD(&lock->l_res_link);
415 CFS_INIT_LIST_HEAD(&lock->l_lru);
416 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
417 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
418 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
419 CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
420 cfs_waitq_init(&lock->l_waitq);
421 lock->l_blocking_lock = NULL;
422 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
423 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
424 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
426 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
428 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
429 class_handle_hash(&lock->l_handle, lock_handle_addref);
431 lu_ref_init(&lock->l_reference);
432 lu_ref_add(&lock->l_reference, "hash", lock);
433 lock->l_callback_timeout = 0;
435 #if LUSTRE_TRACKS_LOCK_EXP_REFS
436 CFS_INIT_LIST_HEAD(&lock->l_exp_refs_link);
437 lock->l_exp_refs_nr = 0;
438 lock->l_exp_refs_target = NULL;
440 CFS_INIT_LIST_HEAD(&lock->l_exp_list);
445 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
446 const struct ldlm_res_id *new_resid)
448 struct ldlm_resource *oldres = lock->l_resource;
449 struct ldlm_resource *newres;
453 LASSERT(ns_is_client(ns));
455 lock_res_and_lock(lock);
456 if (memcmp(new_resid, &lock->l_resource->lr_name,
457 sizeof(lock->l_resource->lr_name)) == 0) {
459 unlock_res_and_lock(lock);
463 LASSERT(new_resid->name[0] != 0);
465 /* This function assumes that the lock isn't on any lists */
466 LASSERT(cfs_list_empty(&lock->l_res_link));
468 type = oldres->lr_type;
469 unlock_res_and_lock(lock);
471 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
475 lu_ref_add(&newres->lr_reference, "lock", lock);
477 * To flip the lock from the old to the new resource, lock, oldres and
478 * newres have to be locked. Resource spin-locks are nested within
479 * lock->l_lock, and are taken in the memory address order to avoid
482 cfs_spin_lock(&lock->l_lock);
483 oldres = lock->l_resource;
484 if (oldres < newres) {
486 lock_res_nested(newres, LRT_NEW);
489 lock_res_nested(oldres, LRT_NEW);
491 LASSERT(memcmp(new_resid, &oldres->lr_name,
492 sizeof oldres->lr_name) != 0);
493 lock->l_resource = newres;
495 unlock_res_and_lock(lock);
497 /* ...and the flowers are still standing! */
498 lu_ref_del(&oldres->lr_reference, "lock", lock);
499 ldlm_resource_putref(oldres);
508 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
510 lockh->cookie = lock->l_handle.h_cookie;
513 /* if flags: atomically get the lock and set the flags.
514 * Return NULL if flag already set
517 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
520 struct ldlm_lock *lock;
525 lock = class_handle2object(handle->cookie);
529 /* It's unlikely but possible that someone marked the lock as
530 * destroyed after we did handle2object on it */
531 if (flags == 0 && !lock->l_destroyed) {
532 lu_ref_add(&lock->l_reference, "handle", cfs_current());
536 lock_res_and_lock(lock);
538 LASSERT(lock->l_resource != NULL);
540 lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
541 if (unlikely(lock->l_destroyed)) {
542 unlock_res_and_lock(lock);
543 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
548 if (flags && (lock->l_flags & flags)) {
549 unlock_res_and_lock(lock);
555 lock->l_flags |= flags;
557 unlock_res_and_lock(lock);
561 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
563 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
564 /* INODEBITS_INTEROP: If the other side does not support
565 * inodebits, reply with a plain lock descriptor.
567 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
568 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
569 /* Make sure all the right bits are set in this lock we
570 are going to pass to client */
571 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
572 (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
573 MDS_INODELOCK_LAYOUT),
574 "Inappropriate inode lock bits during "
575 "conversion " LPU64 "\n",
576 lock->l_policy_data.l_inodebits.bits);
578 ldlm_res2desc(lock->l_resource, &desc->l_resource);
579 desc->l_resource.lr_type = LDLM_PLAIN;
581 /* Convert "new" lock mode to something old client can
583 if ((lock->l_req_mode == LCK_CR) ||
584 (lock->l_req_mode == LCK_CW))
585 desc->l_req_mode = LCK_PR;
587 desc->l_req_mode = lock->l_req_mode;
588 if ((lock->l_granted_mode == LCK_CR) ||
589 (lock->l_granted_mode == LCK_CW)) {
590 desc->l_granted_mode = LCK_PR;
592 /* We never grant PW/EX locks to clients */
593 LASSERT((lock->l_granted_mode != LCK_PW) &&
594 (lock->l_granted_mode != LCK_EX));
595 desc->l_granted_mode = lock->l_granted_mode;
598 /* We do not copy policy here, because there is no
599 policy for plain locks */
601 ldlm_res2desc(lock->l_resource, &desc->l_resource);
602 desc->l_req_mode = lock->l_req_mode;
603 desc->l_granted_mode = lock->l_granted_mode;
604 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
605 &lock->l_policy_data,
606 &desc->l_policy_data);
610 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
611 cfs_list_t *work_list)
613 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
614 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
615 lock->l_flags |= LDLM_FL_AST_SENT;
616 /* If the enqueuing client said so, tell the AST recipient to
617 * discard dirty data, rather than writing back. */
618 if (new->l_flags & LDLM_AST_DISCARD_DATA)
619 lock->l_flags |= LDLM_FL_DISCARD_DATA;
620 LASSERT(cfs_list_empty(&lock->l_bl_ast));
621 cfs_list_add(&lock->l_bl_ast, work_list);
623 LASSERT(lock->l_blocking_lock == NULL);
624 lock->l_blocking_lock = LDLM_LOCK_GET(new);
628 void ldlm_add_cp_work_item(struct ldlm_lock *lock, cfs_list_t *work_list)
630 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
631 lock->l_flags |= LDLM_FL_CP_REQD;
632 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
633 LASSERT(cfs_list_empty(&lock->l_cp_ast));
634 cfs_list_add(&lock->l_cp_ast, work_list);
639 /* must be called with lr_lock held */
640 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
641 cfs_list_t *work_list)
644 check_res_locked(lock->l_resource);
646 ldlm_add_bl_work_item(lock, new, work_list);
648 ldlm_add_cp_work_item(lock, work_list);
652 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
654 struct ldlm_lock *lock;
656 lock = ldlm_handle2lock(lockh);
657 LASSERT(lock != NULL);
658 ldlm_lock_addref_internal(lock, mode);
662 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
664 ldlm_lock_remove_from_lru(lock);
665 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
667 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
669 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
671 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
674 lu_ref_add_atomic(&lock->l_reference, "user", lock);
675 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
679 * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
682 * \retval 0 success, lock was addref-ed
684 * \retval -EAGAIN lock is being canceled.
686 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
688 struct ldlm_lock *lock;
692 lock = ldlm_handle2lock(lockh);
694 lock_res_and_lock(lock);
695 if (lock->l_readers != 0 || lock->l_writers != 0 ||
696 !(lock->l_flags & LDLM_FL_CBPENDING)) {
697 ldlm_lock_addref_internal_nolock(lock, mode);
700 unlock_res_and_lock(lock);
706 /* only called for local locks */
707 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
709 lock_res_and_lock(lock);
710 ldlm_lock_addref_internal_nolock(lock, mode);
711 unlock_res_and_lock(lock);
714 /* only called in ldlm_flock_destroy and for local locks.
715 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
716 * * ldlm_lock_remove_from_lru() does nothing, it is safe
717 * * for ldlm_flock_destroy usage by dropping some code */
718 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
720 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
721 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
722 LASSERT(lock->l_readers > 0);
723 lu_ref_del(&lock->l_reference, "reader", lock);
726 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
727 LASSERT(lock->l_writers > 0);
728 lu_ref_del(&lock->l_reference, "writer", lock);
732 lu_ref_del(&lock->l_reference, "user", lock);
733 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
736 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
738 struct ldlm_namespace *ns;
741 lock_res_and_lock(lock);
743 ns = ldlm_lock_to_ns(lock);
745 ldlm_lock_decref_internal_nolock(lock, mode);
747 if (lock->l_flags & LDLM_FL_LOCAL &&
748 !lock->l_readers && !lock->l_writers) {
749 /* If this is a local lock on a server namespace and this was
750 * the last reference, cancel the lock. */
751 CDEBUG(D_INFO, "forcing cancel of local lock\n");
752 lock->l_flags |= LDLM_FL_CBPENDING;
755 if (!lock->l_readers && !lock->l_writers &&
756 (lock->l_flags & LDLM_FL_CBPENDING)) {
757 /* If we received a blocked AST and this was the last reference,
758 * run the callback. */
759 if (lock->l_ns_srv && lock->l_export)
760 CERROR("FL_CBPENDING set on non-local lock--just a "
763 LDLM_DEBUG(lock, "final decref done on cbpending lock");
765 LDLM_LOCK_GET(lock); /* dropped by bl thread */
766 ldlm_lock_remove_from_lru(lock);
767 unlock_res_and_lock(lock);
769 if (lock->l_flags & LDLM_FL_FAIL_LOC)
770 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
772 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
773 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
774 ldlm_handle_bl_callback(ns, NULL, lock);
775 } else if (ns_is_client(ns) &&
776 !lock->l_readers && !lock->l_writers &&
777 !(lock->l_flags & LDLM_FL_NO_LRU) &&
778 !(lock->l_flags & LDLM_FL_BL_AST)) {
780 LDLM_DEBUG(lock, "add lock into lru list");
782 /* If this is a client-side namespace and this was the last
783 * reference, put it on the LRU. */
784 ldlm_lock_add_to_lru(lock);
785 unlock_res_and_lock(lock);
787 if (lock->l_flags & LDLM_FL_FAIL_LOC)
788 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
790 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
791 * are not supported by the server, otherwise, it is done on
793 if (!exp_connect_cancelset(lock->l_conn_export) &&
794 !ns_connect_lru_resize(ns))
795 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
797 LDLM_DEBUG(lock, "do not add lock into lru list");
798 unlock_res_and_lock(lock);
804 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
806 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
807 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
808 ldlm_lock_decref_internal(lock, mode);
812 /* This will drop a lock reference and mark it for destruction, but will not
813 * necessarily cancel the lock before returning. */
814 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
816 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
819 LASSERT(lock != NULL);
821 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
822 lock_res_and_lock(lock);
823 lock->l_flags |= LDLM_FL_CBPENDING;
824 unlock_res_and_lock(lock);
825 ldlm_lock_decref_internal(lock, mode);
829 struct sl_insert_point {
830 cfs_list_t *res_link;
831 cfs_list_t *mode_link;
832 cfs_list_t *policy_link;
836 * search_granted_lock
839 * Finds a position to insert the new lock.
841 * queue [input]: the granted list where search acts on;
842 * req [input]: the lock whose position to be located;
843 * prev [output]: positions within 3 lists to insert @req to
847 * - ldlm_grant_lock_with_skiplist
849 static void search_granted_lock(cfs_list_t *queue,
850 struct ldlm_lock *req,
851 struct sl_insert_point *prev)
854 struct ldlm_lock *lock, *mode_end, *policy_end;
857 cfs_list_for_each(tmp, queue) {
858 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
860 mode_end = cfs_list_entry(lock->l_sl_mode.prev,
861 struct ldlm_lock, l_sl_mode);
863 if (lock->l_req_mode != req->l_req_mode) {
864 /* jump to last lock of mode group */
865 tmp = &mode_end->l_res_link;
869 /* suitable mode group is found */
870 if (lock->l_resource->lr_type == LDLM_PLAIN) {
871 /* insert point is last lock of the mode group */
872 prev->res_link = &mode_end->l_res_link;
873 prev->mode_link = &mode_end->l_sl_mode;
874 prev->policy_link = &req->l_sl_policy;
877 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
880 cfs_list_entry(lock->l_sl_policy.prev,
884 if (lock->l_policy_data.l_inodebits.bits ==
885 req->l_policy_data.l_inodebits.bits) {
886 /* insert point is last lock of
887 * the policy group */
889 &policy_end->l_res_link;
891 &policy_end->l_sl_mode;
893 &policy_end->l_sl_policy;
898 if (policy_end == mode_end)
899 /* done with mode group */
902 /* go to next policy group within mode group */
903 tmp = policy_end->l_res_link.next;
904 lock = cfs_list_entry(tmp, struct ldlm_lock,
906 } /* loop over policy groups within the mode group */
908 /* insert point is last lock of the mode group,
909 * new policy group is started */
910 prev->res_link = &mode_end->l_res_link;
911 prev->mode_link = &mode_end->l_sl_mode;
912 prev->policy_link = &req->l_sl_policy;
916 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
921 /* insert point is last lock on the queue,
922 * new mode group and new policy group are started */
923 prev->res_link = queue->prev;
924 prev->mode_link = &req->l_sl_mode;
925 prev->policy_link = &req->l_sl_policy;
930 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
931 struct sl_insert_point *prev)
933 struct ldlm_resource *res = lock->l_resource;
936 check_res_locked(res);
938 ldlm_resource_dump(D_INFO, res);
939 CDEBUG(D_OTHER, "About to add this lock:\n");
940 ldlm_lock_dump(D_OTHER, lock, 0);
942 if (lock->l_destroyed) {
943 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
947 LASSERT(cfs_list_empty(&lock->l_res_link));
948 LASSERT(cfs_list_empty(&lock->l_sl_mode));
949 LASSERT(cfs_list_empty(&lock->l_sl_policy));
951 cfs_list_add(&lock->l_res_link, prev->res_link);
952 cfs_list_add(&lock->l_sl_mode, prev->mode_link);
953 cfs_list_add(&lock->l_sl_policy, prev->policy_link);
958 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
960 struct sl_insert_point prev;
963 LASSERT(lock->l_req_mode == lock->l_granted_mode);
965 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
966 ldlm_granted_list_add_lock(lock, &prev);
971 * - ldlm_lock_enqueue
972 * - ldlm_reprocess_queue
973 * - ldlm_lock_convert
975 * must be called with lr_lock held
977 void ldlm_grant_lock(struct ldlm_lock *lock, cfs_list_t *work_list)
979 struct ldlm_resource *res = lock->l_resource;
982 check_res_locked(res);
984 lock->l_granted_mode = lock->l_req_mode;
985 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
986 ldlm_grant_lock_with_skiplist(lock);
987 else if (res->lr_type == LDLM_EXTENT)
988 ldlm_extent_add_lock(res, lock);
990 ldlm_resource_add_lock(res, &res->lr_granted, lock);
992 if (lock->l_granted_mode < res->lr_most_restr)
993 res->lr_most_restr = lock->l_granted_mode;
995 if (work_list && lock->l_completion_ast != NULL)
996 ldlm_add_ast_work_item(lock, NULL, work_list);
998 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1002 /* returns a referenced lock or NULL. See the flag descriptions below, in the
1003 * comment above ldlm_lock_match */
1004 static struct ldlm_lock *search_queue(cfs_list_t *queue,
1006 ldlm_policy_data_t *policy,
1007 struct ldlm_lock *old_lock,
1008 int flags, int unref)
1010 struct ldlm_lock *lock;
1013 cfs_list_for_each(tmp, queue) {
1016 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1018 if (lock == old_lock)
1021 /* llite sometimes wants to match locks that will be
1022 * canceled when their users drop, but we allow it to match
1023 * if it passes in CBPENDING and the lock still has users.
1024 * this is generally only going to be used by children
1025 * whose parents already hold a lock so forward progress
1026 * can still happen. */
1027 if (lock->l_flags & LDLM_FL_CBPENDING &&
1028 !(flags & LDLM_FL_CBPENDING))
1030 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
1031 lock->l_readers == 0 && lock->l_writers == 0)
1034 if (!(lock->l_req_mode & *mode))
1036 match = lock->l_req_mode;
1038 if (lock->l_resource->lr_type == LDLM_EXTENT &&
1039 (lock->l_policy_data.l_extent.start >
1040 policy->l_extent.start ||
1041 lock->l_policy_data.l_extent.end < policy->l_extent.end))
1044 if (unlikely(match == LCK_GROUP) &&
1045 lock->l_resource->lr_type == LDLM_EXTENT &&
1046 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
1049 /* We match if we have existing lock with same or wider set
1051 if (lock->l_resource->lr_type == LDLM_IBITS &&
1052 ((lock->l_policy_data.l_inodebits.bits &
1053 policy->l_inodebits.bits) !=
1054 policy->l_inodebits.bits))
1058 (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED ||
1059 lock->l_fail_value != 0))
1062 if ((flags & LDLM_FL_LOCAL_ONLY) &&
1063 !(lock->l_flags & LDLM_FL_LOCAL))
1066 if (flags & LDLM_FL_TEST_LOCK) {
1067 LDLM_LOCK_GET(lock);
1068 ldlm_lock_touch_in_lru(lock);
1070 ldlm_lock_addref_internal_nolock(lock, match);
1079 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock, int rc)
1081 if (lock->l_fail_value == 0) {
1082 lock->l_fail_value = rc;
1083 cfs_waitq_signal(&lock->l_waitq);
1086 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1088 void ldlm_lock_fail_match(struct ldlm_lock *lock, int rc)
1090 lock_res_and_lock(lock);
1091 ldlm_lock_fail_match_locked(lock, rc);
1092 unlock_res_and_lock(lock);
1094 EXPORT_SYMBOL(ldlm_lock_fail_match);
1096 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1098 lock->l_flags |= LDLM_FL_LVB_READY;
1099 cfs_waitq_signal(&lock->l_waitq);
1102 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1104 lock_res_and_lock(lock);
1105 ldlm_lock_allow_match_locked(lock);
1106 unlock_res_and_lock(lock);
1109 /* Can be called in two ways:
1111 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1112 * for a duplicate of.
1114 * Otherwise, all of the fields must be filled in, to match against.
1116 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1117 * server (ie, connh is NULL)
1118 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1119 * list will be considered
1120 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1121 * to be canceled can still be matched as long as they still have reader
1122 * or writer refernces
1123 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1124 * just tell us if we would have matched.
1126 * Returns 1 if it finds an already-existing lock that is compatible; in this
1127 * case, lockh is filled in with a addref()ed lock
1129 * we also check security context, if that failed we simply return 0 (to keep
1130 * caller code unchanged), the context failure will be discovered by caller
1133 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1134 const struct ldlm_res_id *res_id, ldlm_type_t type,
1135 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1136 struct lustre_handle *lockh, int unref)
1138 struct ldlm_resource *res;
1139 struct ldlm_lock *lock, *old_lock = NULL;
1144 old_lock = ldlm_handle2lock(lockh);
1147 ns = ldlm_lock_to_ns(old_lock);
1148 res_id = &old_lock->l_resource->lr_name;
1149 type = old_lock->l_resource->lr_type;
1150 mode = old_lock->l_req_mode;
1153 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1155 LASSERT(old_lock == NULL);
1159 LDLM_RESOURCE_ADDREF(res);
1162 lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1166 if (flags & LDLM_FL_BLOCK_GRANTED)
1168 lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1172 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1180 LDLM_RESOURCE_DELREF(res);
1181 ldlm_resource_putref(res);
1184 ldlm_lock2handle(lock, lockh);
1185 if ((flags & LDLM_FL_LVB_READY) &&
1186 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1187 struct l_wait_info lwi;
1188 if (lock->l_completion_ast) {
1189 int err = lock->l_completion_ast(lock,
1190 LDLM_FL_WAIT_NOREPROC,
1193 if (flags & LDLM_FL_TEST_LOCK)
1194 LDLM_LOCK_RELEASE(lock);
1196 ldlm_lock_decref_internal(lock,
1203 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1204 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1206 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1207 l_wait_event(lock->l_waitq,
1208 lock->l_flags & LDLM_FL_LVB_READY ||
1209 lock->l_fail_value != 0,
1211 if (!(lock->l_flags & LDLM_FL_LVB_READY)) {
1212 if (flags & LDLM_FL_TEST_LOCK)
1213 LDLM_LOCK_RELEASE(lock);
1215 ldlm_lock_decref_internal(lock, mode);
1222 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1223 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1224 res_id->name[2] : policy->l_extent.start,
1225 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1226 res_id->name[3] : policy->l_extent.end);
1228 /* check user's security context */
1229 if (lock->l_conn_export &&
1230 sptlrpc_import_check_ctx(
1231 class_exp2cliimp(lock->l_conn_export))) {
1232 if (!(flags & LDLM_FL_TEST_LOCK))
1233 ldlm_lock_decref_internal(lock, mode);
1237 if (flags & LDLM_FL_TEST_LOCK)
1238 LDLM_LOCK_RELEASE(lock);
1240 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1241 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1242 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1243 type, mode, res_id->name[0], res_id->name[1],
1244 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1245 res_id->name[2] :policy->l_extent.start,
1246 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1247 res_id->name[3] : policy->l_extent.end);
1250 LDLM_LOCK_PUT(old_lock);
1252 return rc ? mode : 0;
1255 ldlm_mode_t ldlm_revalidate_lock_handle(struct lustre_handle *lockh,
1258 struct ldlm_lock *lock;
1259 ldlm_mode_t mode = 0;
1262 lock = ldlm_handle2lock(lockh);
1264 lock_res_and_lock(lock);
1265 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED ||
1266 lock->l_fail_value != 0)
1269 if (lock->l_flags & LDLM_FL_CBPENDING &&
1270 lock->l_readers == 0 && lock->l_writers == 0)
1274 *bits = lock->l_policy_data.l_inodebits.bits;
1275 mode = lock->l_granted_mode;
1276 ldlm_lock_addref_internal_nolock(lock, mode);
1283 unlock_res_and_lock(lock);
1284 LDLM_LOCK_PUT(lock);
1288 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1290 /* Returns a referenced lock */
1291 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1292 const struct ldlm_res_id *res_id,
1295 const struct ldlm_callback_suite *cbs,
1296 void *data, __u32 lvb_len)
1298 struct ldlm_lock *lock;
1299 struct ldlm_resource *res;
1302 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1306 lock = ldlm_lock_new(res);
1311 lock->l_req_mode = mode;
1312 lock->l_ast_data = data;
1313 lock->l_pid = cfs_curproc_pid();
1314 lock->l_ns_srv = ns_is_server(ns);
1316 lock->l_blocking_ast = cbs->lcs_blocking;
1317 lock->l_completion_ast = cbs->lcs_completion;
1318 lock->l_glimpse_ast = cbs->lcs_glimpse;
1319 lock->l_weigh_ast = cbs->lcs_weigh;
1322 lock->l_tree_node = NULL;
1323 /* if this is the extent lock, allocate the interval tree node */
1324 if (type == LDLM_EXTENT) {
1325 if (ldlm_interval_alloc(lock) == NULL)
1330 lock->l_lvb_len = lvb_len;
1331 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1332 if (lock->l_lvb_data == NULL)
1336 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1342 ldlm_lock_destroy(lock);
1343 LDLM_LOCK_RELEASE(lock);
1347 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1348 struct ldlm_lock **lockp,
1349 void *cookie, int *flags)
1351 struct ldlm_lock *lock = *lockp;
1352 struct ldlm_resource *res = lock->l_resource;
1353 int local = ns_is_client(ldlm_res_to_ns(res));
1354 ldlm_processing_policy policy;
1355 ldlm_error_t rc = ELDLM_OK;
1356 struct ldlm_interval *node = NULL;
1359 lock->l_last_activity = cfs_time_current_sec();
1360 /* policies are not executed on the client or during replay */
1361 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1362 && !local && ns->ns_policy) {
1363 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1365 if (rc == ELDLM_LOCK_REPLACED) {
1366 /* The lock that was returned has already been granted,
1367 * and placed into lockp. If it's not the same as the
1368 * one we passed in, then destroy the old one and our
1369 * work here is done. */
1370 if (lock != *lockp) {
1371 ldlm_lock_destroy(lock);
1372 LDLM_LOCK_RELEASE(lock);
1374 *flags |= LDLM_FL_LOCK_CHANGED;
1376 } else if (rc != ELDLM_OK ||
1377 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1378 ldlm_lock_destroy(lock);
1383 /* For a replaying lock, it might be already in granted list. So
1384 * unlinking the lock will cause the interval node to be freed, we
1385 * have to allocate the interval node early otherwise we can't regrant
1386 * this lock in the future. - jay */
1387 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1388 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1390 lock_res_and_lock(lock);
1391 if (local && lock->l_req_mode == lock->l_granted_mode) {
1392 /* The server returned a blocked lock, but it was granted
1393 * before we got a chance to actually enqueue it. We don't
1394 * need to do anything else. */
1395 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1396 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1397 GOTO(out, ELDLM_OK);
1400 ldlm_resource_unlink_lock(lock);
1401 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1403 ldlm_lock_destroy_nolock(lock);
1404 GOTO(out, rc = -ENOMEM);
1407 CFS_INIT_LIST_HEAD(&node->li_group);
1408 ldlm_interval_attach(node, lock);
1412 /* Some flags from the enqueue want to make it into the AST, via the
1413 * lock's l_flags. */
1414 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1416 /* This distinction between local lock trees is very important; a client
1417 * namespace only has information about locks taken by that client, and
1418 * thus doesn't have enough information to decide for itself if it can
1419 * be granted (below). In this case, we do exactly what the server
1420 * tells us to do, as dictated by the 'flags'.
1422 * We do exactly the same thing during recovery, when the server is
1423 * more or less trusting the clients not to lie.
1425 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1426 * granted/converting queues. */
1428 if (*flags & LDLM_FL_BLOCK_CONV)
1429 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1430 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1431 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1433 ldlm_grant_lock(lock, NULL);
1434 GOTO(out, ELDLM_OK);
1435 } else if (*flags & LDLM_FL_REPLAY) {
1436 if (*flags & LDLM_FL_BLOCK_CONV) {
1437 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1438 GOTO(out, ELDLM_OK);
1439 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1440 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1441 GOTO(out, ELDLM_OK);
1442 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1443 ldlm_grant_lock(lock, NULL);
1444 GOTO(out, ELDLM_OK);
1446 /* If no flags, fall through to normal enqueue path. */
1449 policy = ldlm_processing_policy_table[res->lr_type];
1450 policy(lock, flags, 1, &rc, NULL);
1453 unlock_res_and_lock(lock);
1455 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1459 /* Must be called with namespace taken: queue is waiting or converting. */
1460 int ldlm_reprocess_queue(struct ldlm_resource *res, cfs_list_t *queue,
1461 cfs_list_t *work_list)
1463 cfs_list_t *tmp, *pos;
1464 ldlm_processing_policy policy;
1466 int rc = LDLM_ITER_CONTINUE;
1470 check_res_locked(res);
1472 policy = ldlm_processing_policy_table[res->lr_type];
1475 cfs_list_for_each_safe(tmp, pos, queue) {
1476 struct ldlm_lock *pending;
1477 pending = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1479 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1482 rc = policy(pending, &flags, 0, &err, work_list);
1483 if (rc != LDLM_ITER_CONTINUE)
1491 ldlm_work_bl_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1493 struct ldlm_lock_desc d;
1494 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1499 /* nobody should touch l_bl_ast */
1500 lock_res_and_lock(lock);
1501 cfs_list_del_init(&lock->l_bl_ast);
1503 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1504 LASSERT(lock->l_bl_ast_run == 0);
1505 LASSERT(lock->l_blocking_lock);
1506 lock->l_bl_ast_run++;
1507 unlock_res_and_lock(lock);
1509 ldlm_lock2desc(lock->l_blocking_lock, &d);
1511 rc = lock->l_blocking_ast(lock, &d, (void *)arg,
1513 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1514 lock->l_blocking_lock = NULL;
1515 LDLM_LOCK_RELEASE(lock);
1521 ldlm_work_cp_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1523 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock, l_cp_ast);
1524 ldlm_completion_callback completion_callback;
1528 /* It's possible to receive a completion AST before we've set
1529 * the l_completion_ast pointer: either because the AST arrived
1530 * before the reply, or simply because there's a small race
1531 * window between receiving the reply and finishing the local
1532 * enqueue. (bug 842)
1534 * This can't happen with the blocking_ast, however, because we
1535 * will never call the local blocking_ast until we drop our
1536 * reader/writer reference, which we won't do until we get the
1537 * reply and finish enqueueing. */
1539 /* nobody should touch l_cp_ast */
1540 lock_res_and_lock(lock);
1541 cfs_list_del_init(&lock->l_cp_ast);
1542 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1543 /* save l_completion_ast since it can be changed by
1544 * mds_intent_policy(), see bug 14225 */
1545 completion_callback = lock->l_completion_ast;
1546 lock->l_flags &= ~LDLM_FL_CP_REQD;
1547 unlock_res_and_lock(lock);
1549 if (completion_callback != NULL)
1550 rc = completion_callback(lock, 0, (void *)arg);
1551 LDLM_LOCK_RELEASE(lock);
1557 ldlm_work_revoke_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1559 struct ldlm_lock_desc desc;
1560 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1565 cfs_list_del_init(&lock->l_rk_ast);
1567 /* the desc just pretend to exclusive */
1568 ldlm_lock2desc(lock, &desc);
1569 desc.l_req_mode = LCK_EX;
1570 desc.l_granted_mode = 0;
1572 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1573 LDLM_LOCK_RELEASE(lock);
1578 int ldlm_run_ast_work(struct ldlm_namespace *ns, cfs_list_t *rpc_list,
1579 ldlm_desc_ast_t ast_type)
1581 struct l_wait_info lwi = { 0 };
1582 struct ldlm_cb_set_arg *arg;
1583 cfs_list_t *tmp, *pos;
1584 int (*work_ast_lock)(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg);
1585 unsigned int max_ast_count;
1589 if (cfs_list_empty(rpc_list))
1596 cfs_atomic_set(&arg->restart, 0);
1597 cfs_atomic_set(&arg->rpcs, 0);
1598 cfs_atomic_set(&arg->refcount, 1);
1599 cfs_waitq_init(&arg->waitq);
1602 case LDLM_WORK_BL_AST:
1603 arg->type = LDLM_BL_CALLBACK;
1604 work_ast_lock = ldlm_work_bl_ast_lock;
1606 case LDLM_WORK_CP_AST:
1607 arg->type = LDLM_CP_CALLBACK;
1608 work_ast_lock = ldlm_work_cp_ast_lock;
1610 case LDLM_WORK_REVOKE_AST:
1611 arg->type = LDLM_BL_CALLBACK;
1612 work_ast_lock = ldlm_work_revoke_ast_lock;
1618 max_ast_count = ns->ns_max_parallel_ast ? : UINT_MAX;
1619 arg->threshold = max_ast_count;
1621 cfs_list_for_each_safe(tmp, pos, rpc_list) {
1622 (void)work_ast_lock(tmp, arg);
1623 if (cfs_atomic_read(&arg->rpcs) < max_ast_count)
1626 l_wait_event(arg->waitq,
1627 cfs_atomic_read(&arg->rpcs) < arg->threshold,
1632 l_wait_event(arg->waitq, cfs_atomic_read(&arg->rpcs) == 0, &lwi);
1634 rc = cfs_atomic_read(&arg->restart) ? -ERESTART : 0;
1639 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1641 ldlm_reprocess_all(res);
1642 return LDLM_ITER_CONTINUE;
1645 static int ldlm_reprocess_res(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1646 cfs_hlist_node_t *hnode, void *arg)
1648 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1651 rc = reprocess_one_queue(res, arg);
1653 return rc == LDLM_ITER_STOP;
1656 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1661 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1662 ldlm_reprocess_res, NULL);
1667 void ldlm_reprocess_all(struct ldlm_resource *res)
1669 CFS_LIST_HEAD(rpc_list);
1673 /* Local lock trees don't get reprocessed. */
1674 if (ns_is_client(ldlm_res_to_ns(res))) {
1681 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1682 if (rc == LDLM_ITER_CONTINUE)
1683 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1686 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
1688 if (rc == -ERESTART) {
1689 LASSERT(cfs_list_empty(&rpc_list));
1695 void ldlm_cancel_callback(struct ldlm_lock *lock)
1697 check_res_locked(lock->l_resource);
1698 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1699 lock->l_flags |= LDLM_FL_CANCEL;
1700 if (lock->l_blocking_ast) {
1701 // l_check_no_ns_lock(ns);
1702 unlock_res_and_lock(lock);
1703 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1705 lock_res_and_lock(lock);
1707 LDLM_DEBUG(lock, "no blocking ast");
1710 lock->l_flags |= LDLM_FL_BL_DONE;
1713 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1715 if (req->l_resource->lr_type != LDLM_PLAIN &&
1716 req->l_resource->lr_type != LDLM_IBITS)
1719 cfs_list_del_init(&req->l_sl_policy);
1720 cfs_list_del_init(&req->l_sl_mode);
1723 void ldlm_lock_cancel(struct ldlm_lock *lock)
1725 struct ldlm_resource *res;
1726 struct ldlm_namespace *ns;
1729 lock_res_and_lock(lock);
1731 res = lock->l_resource;
1732 ns = ldlm_res_to_ns(res);
1734 /* Please do not, no matter how tempting, remove this LBUG without
1735 * talking to me first. -phik */
1736 if (lock->l_readers || lock->l_writers) {
1737 LDLM_ERROR(lock, "lock still has references");
1741 ldlm_del_waiting_lock(lock);
1743 /* Releases cancel callback. */
1744 ldlm_cancel_callback(lock);
1746 /* Yes, second time, just in case it was added again while we were
1747 running with no res lock in ldlm_cancel_callback */
1748 ldlm_del_waiting_lock(lock);
1749 ldlm_resource_unlink_lock(lock);
1750 ldlm_lock_destroy_nolock(lock);
1752 if (lock->l_granted_mode == lock->l_req_mode)
1753 ldlm_pool_del(&ns->ns_pool, lock);
1755 /* Make sure we will not be called again for same lock what is possible
1756 * if not to zero out lock->l_granted_mode */
1757 lock->l_granted_mode = LCK_MINMODE;
1758 unlock_res_and_lock(lock);
1763 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1765 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1770 if (lock->l_ast_data == NULL)
1771 lock->l_ast_data = data;
1772 if (lock->l_ast_data == data)
1774 LDLM_LOCK_PUT(lock);
1778 EXPORT_SYMBOL(ldlm_lock_set_data);
1780 int ldlm_cancel_locks_for_export_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1781 cfs_hlist_node_t *hnode, void *data)
1784 struct obd_export *exp = data;
1785 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
1786 struct ldlm_resource *res;
1788 res = ldlm_resource_getref(lock->l_resource);
1789 LDLM_LOCK_GET(lock);
1791 LDLM_DEBUG(lock, "export %p", exp);
1792 ldlm_res_lvbo_update(res, NULL, 1);
1793 ldlm_lock_cancel(lock);
1794 ldlm_reprocess_all(res);
1795 ldlm_resource_putref(res);
1796 LDLM_LOCK_RELEASE(lock);
1800 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1802 cfs_hash_for_each_empty(exp->exp_lock_hash,
1803 ldlm_cancel_locks_for_export_cb, exp);
1807 * Downgrade an exclusive lock.
1809 * A fast variant of ldlm_lock_convert for convertion of exclusive
1810 * locks. The convertion is always successful.
1812 * \param lock A lock to convert
1813 * \param new_mode new lock mode
1815 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1819 LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1820 LASSERT(new_mode == LCK_COS);
1822 lock_res_and_lock(lock);
1823 ldlm_resource_unlink_lock(lock);
1825 * Remove the lock from pool as it will be added again in
1826 * ldlm_grant_lock() called below.
1828 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
1830 lock->l_req_mode = new_mode;
1831 ldlm_grant_lock(lock, NULL);
1832 unlock_res_and_lock(lock);
1833 ldlm_reprocess_all(lock->l_resource);
1838 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1841 CFS_LIST_HEAD(rpc_list);
1842 struct ldlm_resource *res;
1843 struct ldlm_namespace *ns;
1846 struct sl_insert_point prev;
1848 struct ldlm_interval *node;
1851 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1852 *flags |= LDLM_FL_BLOCK_GRANTED;
1853 RETURN(lock->l_resource);
1856 /* I can't check the type of lock here because the bitlock of lock
1857 * is not held here, so do the allocation blindly. -jay */
1858 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1859 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1862 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1863 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1865 lock_res_and_lock(lock);
1867 res = lock->l_resource;
1868 ns = ldlm_res_to_ns(res);
1870 old_mode = lock->l_req_mode;
1871 lock->l_req_mode = new_mode;
1872 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1873 /* remember the lock position where the lock might be
1874 * added back to the granted list later and also
1875 * remember the join mode for skiplist fixing. */
1876 prev.res_link = lock->l_res_link.prev;
1877 prev.mode_link = lock->l_sl_mode.prev;
1878 prev.policy_link = lock->l_sl_policy.prev;
1879 ldlm_resource_unlink_lock(lock);
1881 ldlm_resource_unlink_lock(lock);
1882 if (res->lr_type == LDLM_EXTENT) {
1883 /* FIXME: ugly code, I have to attach the lock to a
1884 * interval node again since perhaps it will be granted
1886 CFS_INIT_LIST_HEAD(&node->li_group);
1887 ldlm_interval_attach(node, lock);
1893 * Remove old lock from the pool before adding the lock with new
1894 * mode below in ->policy()
1896 ldlm_pool_del(&ns->ns_pool, lock);
1898 /* If this is a local resource, put it on the appropriate list. */
1899 if (ns_is_client(ldlm_res_to_ns(res))) {
1900 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1901 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1903 /* This should never happen, because of the way the
1904 * server handles conversions. */
1905 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1909 ldlm_grant_lock(lock, &rpc_list);
1911 /* FIXME: completion handling not with lr_lock held ! */
1912 if (lock->l_completion_ast)
1913 lock->l_completion_ast(lock, 0, NULL);
1917 ldlm_processing_policy policy;
1918 policy = ldlm_processing_policy_table[res->lr_type];
1919 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1920 if (rc == LDLM_ITER_STOP) {
1921 lock->l_req_mode = old_mode;
1922 if (res->lr_type == LDLM_EXTENT)
1923 ldlm_extent_add_lock(res, lock);
1925 ldlm_granted_list_add_lock(lock, &prev);
1929 *flags |= LDLM_FL_BLOCK_GRANTED;
1933 unlock_res_and_lock(lock);
1936 ldlm_run_ast_work(ns, &rpc_list, LDLM_WORK_CP_AST);
1938 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1942 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1944 struct obd_device *obd = NULL;
1946 if (!((libcfs_debug | D_ERROR) & level))
1950 CDEBUG(level, " NULL LDLM lock\n");
1954 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1955 lock, lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1957 if (lock->l_conn_export != NULL)
1958 obd = lock->l_conn_export->exp_obd;
1959 if (lock->l_export && lock->l_export->exp_connection) {
1960 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1961 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1962 lock->l_remote_handle.cookie);
1963 } else if (obd == NULL) {
1964 CDEBUG(level, " Node: local\n");
1966 struct obd_import *imp = obd->u.cli.cl_import;
1967 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1968 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1969 lock->l_remote_handle.cookie);
1971 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1973 lock->l_resource->lr_name.name[0],
1974 lock->l_resource->lr_name.name[1],
1975 lock->l_resource->lr_name.name[2]);
1976 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1977 "write: %d flags: "LPX64"\n", ldlm_lockname[lock->l_req_mode],
1978 ldlm_lockname[lock->l_granted_mode],
1979 cfs_atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1981 if (lock->l_resource->lr_type == LDLM_EXTENT)
1982 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1983 " (req "LPU64"-"LPU64")\n",
1984 lock->l_policy_data.l_extent.start,
1985 lock->l_policy_data.l_extent.end,
1986 lock->l_req_extent.start, lock->l_req_extent.end);
1987 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1988 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1989 lock->l_policy_data.l_flock.pid,
1990 lock->l_policy_data.l_flock.start,
1991 lock->l_policy_data.l_flock.end);
1992 else if (lock->l_resource->lr_type == LDLM_IBITS)
1993 CDEBUG(level, " Bits: "LPX64"\n",
1994 lock->l_policy_data.l_inodebits.bits);
1997 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1999 struct ldlm_lock *lock;
2001 if (!((libcfs_debug | D_ERROR) & level))
2004 lock = ldlm_handle2lock(lockh);
2008 ldlm_lock_dump(D_OTHER, lock, 0);
2010 LDLM_LOCK_PUT(lock);
2013 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
2014 struct libcfs_debug_msg_data *data, const char *fmt,
2018 cfs_debug_limit_state_t *cdls = data->msg_cdls;
2020 va_start(args, fmt);
2022 if (lock->l_resource == NULL) {
2023 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2024 data->msg_fn, data->msg_line, fmt, args,
2025 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2026 "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" remote: "
2027 LPX64" expref: %d pid: %u timeout: %lu\n", lock,
2028 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2029 lock->l_readers, lock->l_writers,
2030 ldlm_lockname[lock->l_granted_mode],
2031 ldlm_lockname[lock->l_req_mode],
2032 lock->l_flags, lock->l_remote_handle.cookie,
2034 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2035 lock->l_pid, lock->l_callback_timeout);
2040 switch (lock->l_resource->lr_type) {
2042 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2043 data->msg_fn, data->msg_line, fmt, args,
2044 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2045 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
2046 "] (req "LPU64"->"LPU64") flags: "LPX64" remote: "LPX64
2047 " expref: %d pid: %u timeout %lu\n",
2048 ldlm_lock_to_ns_name(lock), lock,
2049 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2050 lock->l_readers, lock->l_writers,
2051 ldlm_lockname[lock->l_granted_mode],
2052 ldlm_lockname[lock->l_req_mode],
2053 lock->l_resource->lr_name.name[0],
2054 lock->l_resource->lr_name.name[1],
2055 cfs_atomic_read(&lock->l_resource->lr_refcount),
2056 ldlm_typename[lock->l_resource->lr_type],
2057 lock->l_policy_data.l_extent.start,
2058 lock->l_policy_data.l_extent.end,
2059 lock->l_req_extent.start, lock->l_req_extent.end,
2060 lock->l_flags, lock->l_remote_handle.cookie,
2062 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2063 lock->l_pid, lock->l_callback_timeout);
2067 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2068 data->msg_fn, data->msg_line, fmt, args,
2069 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2070 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
2071 "["LPU64"->"LPU64"] flags: "LPX64" remote: "LPX64
2072 " expref: %d pid: %u timeout: %lu\n",
2073 ldlm_lock_to_ns_name(lock), lock,
2074 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2075 lock->l_readers, lock->l_writers,
2076 ldlm_lockname[lock->l_granted_mode],
2077 ldlm_lockname[lock->l_req_mode],
2078 lock->l_resource->lr_name.name[0],
2079 lock->l_resource->lr_name.name[1],
2080 cfs_atomic_read(&lock->l_resource->lr_refcount),
2081 ldlm_typename[lock->l_resource->lr_type],
2082 lock->l_policy_data.l_flock.pid,
2083 lock->l_policy_data.l_flock.start,
2084 lock->l_policy_data.l_flock.end,
2085 lock->l_flags, lock->l_remote_handle.cookie,
2087 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2088 lock->l_pid, lock->l_callback_timeout);
2092 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2093 data->msg_fn, data->msg_line, fmt, args,
2094 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2095 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
2096 "flags: "LPX64" remote: "LPX64" expref: %d "
2097 "pid: %u timeout: %lu\n",
2098 ldlm_lock_to_ns_name(lock),
2099 lock, lock->l_handle.h_cookie,
2100 cfs_atomic_read (&lock->l_refc),
2101 lock->l_readers, lock->l_writers,
2102 ldlm_lockname[lock->l_granted_mode],
2103 ldlm_lockname[lock->l_req_mode],
2104 lock->l_resource->lr_name.name[0],
2105 lock->l_resource->lr_name.name[1],
2106 lock->l_policy_data.l_inodebits.bits,
2107 cfs_atomic_read(&lock->l_resource->lr_refcount),
2108 ldlm_typename[lock->l_resource->lr_type],
2109 lock->l_flags, lock->l_remote_handle.cookie,
2111 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2112 lock->l_pid, lock->l_callback_timeout);
2116 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2117 data->msg_fn, data->msg_line, fmt, args,
2118 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2119 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" "
2120 "remote: "LPX64" expref: %d pid: %u timeout %lu\n",
2121 ldlm_lock_to_ns_name(lock),
2122 lock, lock->l_handle.h_cookie,
2123 cfs_atomic_read (&lock->l_refc),
2124 lock->l_readers, lock->l_writers,
2125 ldlm_lockname[lock->l_granted_mode],
2126 ldlm_lockname[lock->l_req_mode],
2127 lock->l_resource->lr_name.name[0],
2128 lock->l_resource->lr_name.name[1],
2129 cfs_atomic_read(&lock->l_resource->lr_refcount),
2130 ldlm_typename[lock->l_resource->lr_type],
2131 lock->l_flags, lock->l_remote_handle.cookie,
2133 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2134 lock->l_pid, lock->l_callback_timeout);
2139 EXPORT_SYMBOL(_ldlm_lock_debug);