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, 2012, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
38 * lustre/ldlm/ldlm_lock.c
40 * Author: Peter Braam <braam@clusterfs.com>
41 * Author: Phil Schwan <phil@clusterfs.com>
44 #define DEBUG_SUBSYSTEM S_LDLM
47 # include <libcfs/libcfs.h>
48 # include <linux/lustre_intent.h>
50 # include <liblustre.h>
53 #include <obd_class.h>
54 #include "ldlm_internal.h"
57 char *ldlm_lockname[] = {
69 char *ldlm_typename[] = {
76 static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = {
77 [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_wire_to_local,
78 [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_wire_to_local,
79 [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_wire_to_local,
80 [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_wire_to_local,
83 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
84 [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_local_to_wire,
85 [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_local_to_wire,
86 [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_local_to_wire,
87 [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_local_to_wire,
91 * Converts lock policy from local format to on the wire lock_desc format
93 void ldlm_convert_policy_to_wire(ldlm_type_t type,
94 const ldlm_policy_data_t *lpolicy,
95 ldlm_wire_policy_data_t *wpolicy)
97 ldlm_policy_local_to_wire_t convert;
99 convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
101 convert(lpolicy, wpolicy);
105 * Converts lock policy from on the wire lock_desc format to local format
107 void ldlm_convert_policy_to_local(ldlm_type_t type,
108 const ldlm_wire_policy_data_t *wpolicy,
109 ldlm_policy_data_t *lpolicy)
111 ldlm_policy_wire_to_local_t convert;
113 convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE];
115 convert(wpolicy, lpolicy);
118 char *ldlm_it2str(int it)
125 case (IT_OPEN | IT_CREAT):
140 CERROR("Unknown intent %d\n", it);
145 extern cfs_mem_cache_t *ldlm_lock_slab;
147 static ldlm_processing_policy ldlm_processing_policy_table[] = {
148 [LDLM_PLAIN] ldlm_process_plain_lock,
149 [LDLM_EXTENT] ldlm_process_extent_lock,
151 [LDLM_FLOCK] ldlm_process_flock_lock,
153 [LDLM_IBITS] ldlm_process_inodebits_lock,
156 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
158 return ldlm_processing_policy_table[res->lr_type];
161 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
167 * REFCOUNTED LOCK OBJECTS
172 * Lock refcounts, during creation:
173 * - one special one for allocation, dec'd only once in destroy
174 * - one for being a lock that's in-use
175 * - one for the addref associated with a new lock
177 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
179 cfs_atomic_inc(&lock->l_refc);
183 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
185 LASSERT(size == sizeof(*lock));
186 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
189 void ldlm_lock_put(struct ldlm_lock *lock)
193 LASSERT(lock->l_resource != LP_POISON);
194 LASSERT(cfs_atomic_read(&lock->l_refc) > 0);
195 if (cfs_atomic_dec_and_test(&lock->l_refc)) {
196 struct ldlm_resource *res;
199 "final lock_put on destroyed lock, freeing it.");
201 res = lock->l_resource;
202 LASSERT(lock->l_destroyed);
203 LASSERT(cfs_list_empty(&lock->l_res_link));
204 LASSERT(cfs_list_empty(&lock->l_pending_chain));
206 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
208 lu_ref_del(&res->lr_reference, "lock", lock);
209 ldlm_resource_putref(res);
210 lock->l_resource = NULL;
211 if (lock->l_export) {
212 class_export_lock_put(lock->l_export, lock);
213 lock->l_export = NULL;
216 if (lock->l_lvb_data != NULL)
217 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
219 ldlm_interval_free(ldlm_interval_detach(lock));
220 lu_ref_fini(&lock->l_reference);
221 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
228 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
231 if (!cfs_list_empty(&lock->l_lru)) {
232 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
234 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
235 cfs_list_del_init(&lock->l_lru);
236 if (lock->l_flags & LDLM_FL_SKIPPED)
237 lock->l_flags &= ~LDLM_FL_SKIPPED;
238 LASSERT(ns->ns_nr_unused > 0);
245 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
247 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
251 if (lock->l_ns_srv) {
252 LASSERT(cfs_list_empty(&lock->l_lru));
256 cfs_spin_lock(&ns->ns_lock);
257 rc = ldlm_lock_remove_from_lru_nolock(lock);
258 cfs_spin_unlock(&ns->ns_lock);
263 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
265 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
267 lock->l_last_used = cfs_time_current();
268 LASSERT(cfs_list_empty(&lock->l_lru));
269 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
270 cfs_list_add_tail(&lock->l_lru, &ns->ns_unused_list);
271 LASSERT(ns->ns_nr_unused >= 0);
275 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
277 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
280 cfs_spin_lock(&ns->ns_lock);
281 ldlm_lock_add_to_lru_nolock(lock);
282 cfs_spin_unlock(&ns->ns_lock);
286 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
288 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
291 if (lock->l_ns_srv) {
292 LASSERT(cfs_list_empty(&lock->l_lru));
297 cfs_spin_lock(&ns->ns_lock);
298 if (!cfs_list_empty(&lock->l_lru)) {
299 ldlm_lock_remove_from_lru_nolock(lock);
300 ldlm_lock_add_to_lru_nolock(lock);
302 cfs_spin_unlock(&ns->ns_lock);
306 /* This used to have a 'strict' flag, which recovery would use to mark an
307 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
308 * shall explain why it's gone: with the new hash table scheme, once you call
309 * ldlm_lock_destroy, you can never drop your final references on this lock.
310 * Because it's not in the hash table anymore. -phil */
311 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
315 if (lock->l_readers || lock->l_writers) {
316 LDLM_ERROR(lock, "lock still has references");
317 ldlm_lock_dump(D_ERROR, lock, 0);
321 if (!cfs_list_empty(&lock->l_res_link)) {
322 LDLM_ERROR(lock, "lock still on resource");
323 ldlm_lock_dump(D_ERROR, lock, 0);
327 if (lock->l_destroyed) {
328 LASSERT(cfs_list_empty(&lock->l_lru));
332 lock->l_destroyed = 1;
334 if (lock->l_export && lock->l_export->exp_lock_hash &&
335 !cfs_hlist_unhashed(&lock->l_exp_hash))
336 cfs_hash_del(lock->l_export->exp_lock_hash,
337 &lock->l_remote_handle, &lock->l_exp_hash);
339 ldlm_lock_remove_from_lru(lock);
340 class_handle_unhash(&lock->l_handle);
343 /* Wake anyone waiting for this lock */
344 /* FIXME: I should probably add yet another flag, instead of using
345 * l_export to only call this on clients */
347 class_export_put(lock->l_export);
348 lock->l_export = NULL;
349 if (lock->l_export && lock->l_completion_ast)
350 lock->l_completion_ast(lock, 0);
356 void ldlm_lock_destroy(struct ldlm_lock *lock)
360 lock_res_and_lock(lock);
361 first = ldlm_lock_destroy_internal(lock);
362 unlock_res_and_lock(lock);
364 /* drop reference from hashtable only for first destroy */
366 lu_ref_del(&lock->l_reference, "hash", lock);
367 LDLM_LOCK_RELEASE(lock);
372 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
376 first = ldlm_lock_destroy_internal(lock);
377 /* drop reference from hashtable only for first destroy */
379 lu_ref_del(&lock->l_reference, "hash", lock);
380 LDLM_LOCK_RELEASE(lock);
385 /* this is called by portals_handle2object with the handle lock taken */
386 static void lock_handle_addref(void *lock)
388 LDLM_LOCK_GET((struct ldlm_lock *)lock);
392 * usage: pass in a resource on which you have done ldlm_resource_get
393 * new lock will take over the refcount.
394 * returns: lock with refcount 2 - one for current caller and one for remote
396 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
398 struct ldlm_lock *lock;
401 if (resource == NULL)
404 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, CFS_ALLOC_IO);
408 cfs_spin_lock_init(&lock->l_lock);
409 lock->l_resource = resource;
410 lu_ref_add(&resource->lr_reference, "lock", lock);
412 cfs_atomic_set(&lock->l_refc, 2);
413 CFS_INIT_LIST_HEAD(&lock->l_res_link);
414 CFS_INIT_LIST_HEAD(&lock->l_lru);
415 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
416 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
417 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
418 CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
419 cfs_waitq_init(&lock->l_waitq);
420 lock->l_blocking_lock = NULL;
421 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
422 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
423 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
425 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
427 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
428 class_handle_hash(&lock->l_handle, lock_handle_addref);
430 lu_ref_init(&lock->l_reference);
431 lu_ref_add(&lock->l_reference, "hash", lock);
432 lock->l_callback_timeout = 0;
434 #if LUSTRE_TRACKS_LOCK_EXP_REFS
435 CFS_INIT_LIST_HEAD(&lock->l_exp_refs_link);
436 lock->l_exp_refs_nr = 0;
437 lock->l_exp_refs_target = NULL;
439 CFS_INIT_LIST_HEAD(&lock->l_exp_list);
444 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
445 const struct ldlm_res_id *new_resid)
447 struct ldlm_resource *oldres = lock->l_resource;
448 struct ldlm_resource *newres;
452 LASSERT(ns_is_client(ns));
454 lock_res_and_lock(lock);
455 if (memcmp(new_resid, &lock->l_resource->lr_name,
456 sizeof(lock->l_resource->lr_name)) == 0) {
458 unlock_res_and_lock(lock);
462 LASSERT(new_resid->name[0] != 0);
464 /* This function assumes that the lock isn't on any lists */
465 LASSERT(cfs_list_empty(&lock->l_res_link));
467 type = oldres->lr_type;
468 unlock_res_and_lock(lock);
470 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
474 lu_ref_add(&newres->lr_reference, "lock", lock);
476 * To flip the lock from the old to the new resource, lock, oldres and
477 * newres have to be locked. Resource spin-locks are nested within
478 * lock->l_lock, and are taken in the memory address order to avoid
481 cfs_spin_lock(&lock->l_lock);
482 oldres = lock->l_resource;
483 if (oldres < newres) {
485 lock_res_nested(newres, LRT_NEW);
488 lock_res_nested(oldres, LRT_NEW);
490 LASSERT(memcmp(new_resid, &oldres->lr_name,
491 sizeof oldres->lr_name) != 0);
492 lock->l_resource = newres;
494 unlock_res_and_lock(lock);
496 /* ...and the flowers are still standing! */
497 lu_ref_del(&oldres->lr_reference, "lock", lock);
498 ldlm_resource_putref(oldres);
507 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
509 lockh->cookie = lock->l_handle.h_cookie;
512 /* if flags: atomically get the lock and set the flags.
513 * Return NULL if flag already set
516 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
519 struct ldlm_lock *lock;
524 lock = class_handle2object(handle->cookie);
528 /* It's unlikely but possible that someone marked the lock as
529 * destroyed after we did handle2object on it */
530 if (flags == 0 && !lock->l_destroyed) {
531 lu_ref_add(&lock->l_reference, "handle", cfs_current());
535 lock_res_and_lock(lock);
537 LASSERT(lock->l_resource != NULL);
539 lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
540 if (unlikely(lock->l_destroyed)) {
541 unlock_res_and_lock(lock);
542 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
547 if (flags && (lock->l_flags & flags)) {
548 unlock_res_and_lock(lock);
554 lock->l_flags |= flags;
556 unlock_res_and_lock(lock);
560 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
562 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
563 /* INODEBITS_INTEROP: If the other side does not support
564 * inodebits, reply with a plain lock descriptor.
566 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
567 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
568 /* Make sure all the right bits are set in this lock we
569 are going to pass to client */
570 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
571 (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
572 MDS_INODELOCK_LAYOUT),
573 "Inappropriate inode lock bits during "
574 "conversion " LPU64 "\n",
575 lock->l_policy_data.l_inodebits.bits);
577 ldlm_res2desc(lock->l_resource, &desc->l_resource);
578 desc->l_resource.lr_type = LDLM_PLAIN;
580 /* Convert "new" lock mode to something old client can
582 if ((lock->l_req_mode == LCK_CR) ||
583 (lock->l_req_mode == LCK_CW))
584 desc->l_req_mode = LCK_PR;
586 desc->l_req_mode = lock->l_req_mode;
587 if ((lock->l_granted_mode == LCK_CR) ||
588 (lock->l_granted_mode == LCK_CW)) {
589 desc->l_granted_mode = LCK_PR;
591 /* We never grant PW/EX locks to clients */
592 LASSERT((lock->l_granted_mode != LCK_PW) &&
593 (lock->l_granted_mode != LCK_EX));
594 desc->l_granted_mode = lock->l_granted_mode;
597 /* We do not copy policy here, because there is no
598 policy for plain locks */
600 ldlm_res2desc(lock->l_resource, &desc->l_resource);
601 desc->l_req_mode = lock->l_req_mode;
602 desc->l_granted_mode = lock->l_granted_mode;
603 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
604 &lock->l_policy_data,
605 &desc->l_policy_data);
609 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
610 cfs_list_t *work_list)
612 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
613 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
614 lock->l_flags |= LDLM_FL_AST_SENT;
615 /* If the enqueuing client said so, tell the AST recipient to
616 * discard dirty data, rather than writing back. */
617 if (new->l_flags & LDLM_AST_DISCARD_DATA)
618 lock->l_flags |= LDLM_FL_DISCARD_DATA;
619 LASSERT(cfs_list_empty(&lock->l_bl_ast));
620 cfs_list_add(&lock->l_bl_ast, work_list);
622 LASSERT(lock->l_blocking_lock == NULL);
623 lock->l_blocking_lock = LDLM_LOCK_GET(new);
627 void ldlm_add_cp_work_item(struct ldlm_lock *lock, cfs_list_t *work_list)
629 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
630 lock->l_flags |= LDLM_FL_CP_REQD;
631 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
632 LASSERT(cfs_list_empty(&lock->l_cp_ast));
633 cfs_list_add(&lock->l_cp_ast, work_list);
638 /* must be called with lr_lock held */
639 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
640 cfs_list_t *work_list)
643 check_res_locked(lock->l_resource);
645 ldlm_add_bl_work_item(lock, new, work_list);
647 ldlm_add_cp_work_item(lock, work_list);
651 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
653 struct ldlm_lock *lock;
655 lock = ldlm_handle2lock(lockh);
656 LASSERT(lock != NULL);
657 ldlm_lock_addref_internal(lock, mode);
661 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
663 ldlm_lock_remove_from_lru(lock);
664 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
666 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
668 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
670 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
673 lu_ref_add_atomic(&lock->l_reference, "user", lock);
674 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
678 * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
681 * \retval 0 success, lock was addref-ed
683 * \retval -EAGAIN lock is being canceled.
685 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
687 struct ldlm_lock *lock;
691 lock = ldlm_handle2lock(lockh);
693 lock_res_and_lock(lock);
694 if (lock->l_readers != 0 || lock->l_writers != 0 ||
695 !(lock->l_flags & LDLM_FL_CBPENDING)) {
696 ldlm_lock_addref_internal_nolock(lock, mode);
699 unlock_res_and_lock(lock);
705 /* only called for local locks */
706 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
708 lock_res_and_lock(lock);
709 ldlm_lock_addref_internal_nolock(lock, mode);
710 unlock_res_and_lock(lock);
713 /* only called in ldlm_flock_destroy and for local locks.
714 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
715 * * ldlm_lock_remove_from_lru() does nothing, it is safe
716 * * for ldlm_flock_destroy usage by dropping some code */
717 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
719 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
720 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
721 LASSERT(lock->l_readers > 0);
722 lu_ref_del(&lock->l_reference, "reader", lock);
725 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
726 LASSERT(lock->l_writers > 0);
727 lu_ref_del(&lock->l_reference, "writer", lock);
731 lu_ref_del(&lock->l_reference, "user", lock);
732 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
735 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
737 struct ldlm_namespace *ns;
740 lock_res_and_lock(lock);
742 ns = ldlm_lock_to_ns(lock);
744 ldlm_lock_decref_internal_nolock(lock, mode);
746 if (lock->l_flags & LDLM_FL_LOCAL &&
747 !lock->l_readers && !lock->l_writers) {
748 /* If this is a local lock on a server namespace and this was
749 * the last reference, cancel the lock. */
750 CDEBUG(D_INFO, "forcing cancel of local lock\n");
751 lock->l_flags |= LDLM_FL_CBPENDING;
754 if (!lock->l_readers && !lock->l_writers &&
755 (lock->l_flags & LDLM_FL_CBPENDING)) {
756 /* If we received a blocked AST and this was the last reference,
757 * run the callback. */
758 if (lock->l_ns_srv && lock->l_export)
759 CERROR("FL_CBPENDING set on non-local lock--just a "
762 LDLM_DEBUG(lock, "final decref done on cbpending lock");
764 LDLM_LOCK_GET(lock); /* dropped by bl thread */
765 ldlm_lock_remove_from_lru(lock);
766 unlock_res_and_lock(lock);
768 if (lock->l_flags & LDLM_FL_FAIL_LOC)
769 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
771 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
772 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
773 ldlm_handle_bl_callback(ns, NULL, lock);
774 } else if (ns_is_client(ns) &&
775 !lock->l_readers && !lock->l_writers &&
776 !(lock->l_flags & LDLM_FL_NO_LRU) &&
777 !(lock->l_flags & LDLM_FL_BL_AST)) {
779 LDLM_DEBUG(lock, "add lock into lru list");
781 /* If this is a client-side namespace and this was the last
782 * reference, put it on the LRU. */
783 ldlm_lock_add_to_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 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
790 * are not supported by the server, otherwise, it is done on
792 if (!exp_connect_cancelset(lock->l_conn_export) &&
793 !ns_connect_lru_resize(ns))
794 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
796 LDLM_DEBUG(lock, "do not add lock into lru list");
797 unlock_res_and_lock(lock);
803 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
805 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
806 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
807 ldlm_lock_decref_internal(lock, mode);
811 /* This will drop a lock reference and mark it for destruction, but will not
812 * necessarily cancel the lock before returning. */
813 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
815 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
818 LASSERT(lock != NULL);
820 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
821 lock_res_and_lock(lock);
822 lock->l_flags |= LDLM_FL_CBPENDING;
823 unlock_res_and_lock(lock);
824 ldlm_lock_decref_internal(lock, mode);
828 struct sl_insert_point {
829 cfs_list_t *res_link;
830 cfs_list_t *mode_link;
831 cfs_list_t *policy_link;
835 * search_granted_lock
838 * Finds a position to insert the new lock.
840 * queue [input]: the granted list where search acts on;
841 * req [input]: the lock whose position to be located;
842 * prev [output]: positions within 3 lists to insert @req to
846 * - ldlm_grant_lock_with_skiplist
848 static void search_granted_lock(cfs_list_t *queue,
849 struct ldlm_lock *req,
850 struct sl_insert_point *prev)
853 struct ldlm_lock *lock, *mode_end, *policy_end;
856 cfs_list_for_each(tmp, queue) {
857 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
859 mode_end = cfs_list_entry(lock->l_sl_mode.prev,
860 struct ldlm_lock, l_sl_mode);
862 if (lock->l_req_mode != req->l_req_mode) {
863 /* jump to last lock of mode group */
864 tmp = &mode_end->l_res_link;
868 /* suitable mode group is found */
869 if (lock->l_resource->lr_type == LDLM_PLAIN) {
870 /* insert point is last lock of the mode group */
871 prev->res_link = &mode_end->l_res_link;
872 prev->mode_link = &mode_end->l_sl_mode;
873 prev->policy_link = &req->l_sl_policy;
876 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
879 cfs_list_entry(lock->l_sl_policy.prev,
883 if (lock->l_policy_data.l_inodebits.bits ==
884 req->l_policy_data.l_inodebits.bits) {
885 /* insert point is last lock of
886 * the policy group */
888 &policy_end->l_res_link;
890 &policy_end->l_sl_mode;
892 &policy_end->l_sl_policy;
897 if (policy_end == mode_end)
898 /* done with mode group */
901 /* go to next policy group within mode group */
902 tmp = policy_end->l_res_link.next;
903 lock = cfs_list_entry(tmp, struct ldlm_lock,
905 } /* loop over policy groups within the mode group */
907 /* insert point is last lock of the mode group,
908 * new policy group is started */
909 prev->res_link = &mode_end->l_res_link;
910 prev->mode_link = &mode_end->l_sl_mode;
911 prev->policy_link = &req->l_sl_policy;
915 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
920 /* insert point is last lock on the queue,
921 * new mode group and new policy group are started */
922 prev->res_link = queue->prev;
923 prev->mode_link = &req->l_sl_mode;
924 prev->policy_link = &req->l_sl_policy;
929 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
930 struct sl_insert_point *prev)
932 struct ldlm_resource *res = lock->l_resource;
935 check_res_locked(res);
937 ldlm_resource_dump(D_INFO, res);
938 CDEBUG(D_OTHER, "About to add this lock:\n");
939 ldlm_lock_dump(D_OTHER, lock, 0);
941 if (lock->l_destroyed) {
942 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
946 LASSERT(cfs_list_empty(&lock->l_res_link));
947 LASSERT(cfs_list_empty(&lock->l_sl_mode));
948 LASSERT(cfs_list_empty(&lock->l_sl_policy));
950 cfs_list_add(&lock->l_res_link, prev->res_link);
951 cfs_list_add(&lock->l_sl_mode, prev->mode_link);
952 cfs_list_add(&lock->l_sl_policy, prev->policy_link);
957 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
959 struct sl_insert_point prev;
962 LASSERT(lock->l_req_mode == lock->l_granted_mode);
964 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
965 ldlm_granted_list_add_lock(lock, &prev);
970 * - ldlm_lock_enqueue
971 * - ldlm_reprocess_queue
972 * - ldlm_lock_convert
974 * must be called with lr_lock held
976 void ldlm_grant_lock(struct ldlm_lock *lock, cfs_list_t *work_list)
978 struct ldlm_resource *res = lock->l_resource;
981 check_res_locked(res);
983 lock->l_granted_mode = lock->l_req_mode;
984 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
985 ldlm_grant_lock_with_skiplist(lock);
986 else if (res->lr_type == LDLM_EXTENT)
987 ldlm_extent_add_lock(res, lock);
989 ldlm_resource_add_lock(res, &res->lr_granted, lock);
991 if (lock->l_granted_mode < res->lr_most_restr)
992 res->lr_most_restr = lock->l_granted_mode;
994 if (work_list && lock->l_completion_ast != NULL)
995 ldlm_add_ast_work_item(lock, NULL, work_list);
997 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1001 /* returns a referenced lock or NULL. See the flag descriptions below, in the
1002 * comment above ldlm_lock_match */
1003 static struct ldlm_lock *search_queue(cfs_list_t *queue,
1005 ldlm_policy_data_t *policy,
1006 struct ldlm_lock *old_lock,
1007 int flags, int unref)
1009 struct ldlm_lock *lock;
1012 cfs_list_for_each(tmp, queue) {
1015 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1017 if (lock == old_lock)
1020 /* llite sometimes wants to match locks that will be
1021 * canceled when their users drop, but we allow it to match
1022 * if it passes in CBPENDING and the lock still has users.
1023 * this is generally only going to be used by children
1024 * whose parents already hold a lock so forward progress
1025 * can still happen. */
1026 if (lock->l_flags & LDLM_FL_CBPENDING &&
1027 !(flags & LDLM_FL_CBPENDING))
1029 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
1030 lock->l_readers == 0 && lock->l_writers == 0)
1033 if (!(lock->l_req_mode & *mode))
1035 match = lock->l_req_mode;
1037 if (lock->l_resource->lr_type == LDLM_EXTENT &&
1038 (lock->l_policy_data.l_extent.start >
1039 policy->l_extent.start ||
1040 lock->l_policy_data.l_extent.end < policy->l_extent.end))
1043 if (unlikely(match == LCK_GROUP) &&
1044 lock->l_resource->lr_type == LDLM_EXTENT &&
1045 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
1048 /* We match if we have existing lock with same or wider set
1050 if (lock->l_resource->lr_type == LDLM_IBITS &&
1051 ((lock->l_policy_data.l_inodebits.bits &
1052 policy->l_inodebits.bits) !=
1053 policy->l_inodebits.bits))
1057 (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED ||
1061 if ((flags & LDLM_FL_LOCAL_ONLY) &&
1062 !(lock->l_flags & LDLM_FL_LOCAL))
1065 if (flags & LDLM_FL_TEST_LOCK) {
1066 LDLM_LOCK_GET(lock);
1067 ldlm_lock_touch_in_lru(lock);
1069 ldlm_lock_addref_internal_nolock(lock, match);
1078 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1080 if (!lock->l_failed) {
1082 cfs_waitq_broadcast(&lock->l_waitq);
1085 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1087 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1089 lock_res_and_lock(lock);
1090 ldlm_lock_fail_match_locked(lock);
1091 unlock_res_and_lock(lock);
1093 EXPORT_SYMBOL(ldlm_lock_fail_match);
1095 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1097 lock->l_flags |= LDLM_FL_LVB_READY;
1098 cfs_waitq_broadcast(&lock->l_waitq);
1101 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1103 lock_res_and_lock(lock);
1104 ldlm_lock_allow_match_locked(lock);
1105 unlock_res_and_lock(lock);
1108 /* Can be called in two ways:
1110 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1111 * for a duplicate of.
1113 * Otherwise, all of the fields must be filled in, to match against.
1115 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1116 * server (ie, connh is NULL)
1117 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1118 * list will be considered
1119 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1120 * to be canceled can still be matched as long as they still have reader
1121 * or writer refernces
1122 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1123 * just tell us if we would have matched.
1125 * Returns 1 if it finds an already-existing lock that is compatible; in this
1126 * case, lockh is filled in with a addref()ed lock
1128 * we also check security context, if that failed we simply return 0 (to keep
1129 * caller code unchanged), the context failure will be discovered by caller
1132 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1133 const struct ldlm_res_id *res_id, ldlm_type_t type,
1134 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1135 struct lustre_handle *lockh, int unref)
1137 struct ldlm_resource *res;
1138 struct ldlm_lock *lock, *old_lock = NULL;
1143 old_lock = ldlm_handle2lock(lockh);
1146 ns = ldlm_lock_to_ns(old_lock);
1147 res_id = &old_lock->l_resource->lr_name;
1148 type = old_lock->l_resource->lr_type;
1149 mode = old_lock->l_req_mode;
1152 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1154 LASSERT(old_lock == NULL);
1158 LDLM_RESOURCE_ADDREF(res);
1161 lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1165 if (flags & LDLM_FL_BLOCK_GRANTED)
1167 lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1171 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1179 LDLM_RESOURCE_DELREF(res);
1180 ldlm_resource_putref(res);
1183 ldlm_lock2handle(lock, lockh);
1184 if ((flags & LDLM_FL_LVB_READY) &&
1185 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1186 struct l_wait_info lwi;
1187 if (lock->l_completion_ast) {
1188 int err = lock->l_completion_ast(lock,
1189 LDLM_FL_WAIT_NOREPROC,
1192 if (flags & LDLM_FL_TEST_LOCK)
1193 LDLM_LOCK_RELEASE(lock);
1195 ldlm_lock_decref_internal(lock,
1202 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1203 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1205 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1206 l_wait_event(lock->l_waitq,
1207 lock->l_flags & LDLM_FL_LVB_READY ||
1210 if (!(lock->l_flags & LDLM_FL_LVB_READY)) {
1211 if (flags & LDLM_FL_TEST_LOCK)
1212 LDLM_LOCK_RELEASE(lock);
1214 ldlm_lock_decref_internal(lock, mode);
1221 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1222 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1223 res_id->name[2] : policy->l_extent.start,
1224 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1225 res_id->name[3] : policy->l_extent.end);
1227 /* check user's security context */
1228 if (lock->l_conn_export &&
1229 sptlrpc_import_check_ctx(
1230 class_exp2cliimp(lock->l_conn_export))) {
1231 if (!(flags & LDLM_FL_TEST_LOCK))
1232 ldlm_lock_decref_internal(lock, mode);
1236 if (flags & LDLM_FL_TEST_LOCK)
1237 LDLM_LOCK_RELEASE(lock);
1239 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1240 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1241 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1242 type, mode, res_id->name[0], res_id->name[1],
1243 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1244 res_id->name[2] :policy->l_extent.start,
1245 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1246 res_id->name[3] : policy->l_extent.end);
1249 LDLM_LOCK_PUT(old_lock);
1251 return rc ? mode : 0;
1254 ldlm_mode_t ldlm_revalidate_lock_handle(struct lustre_handle *lockh,
1257 struct ldlm_lock *lock;
1258 ldlm_mode_t mode = 0;
1261 lock = ldlm_handle2lock(lockh);
1263 lock_res_and_lock(lock);
1264 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED ||
1268 if (lock->l_flags & LDLM_FL_CBPENDING &&
1269 lock->l_readers == 0 && lock->l_writers == 0)
1273 *bits = lock->l_policy_data.l_inodebits.bits;
1274 mode = lock->l_granted_mode;
1275 ldlm_lock_addref_internal_nolock(lock, mode);
1282 unlock_res_and_lock(lock);
1283 LDLM_LOCK_PUT(lock);
1287 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1289 /* Returns a referenced lock */
1290 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1291 const struct ldlm_res_id *res_id,
1294 const struct ldlm_callback_suite *cbs,
1295 void *data, __u32 lvb_len)
1297 struct ldlm_lock *lock;
1298 struct ldlm_resource *res;
1301 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1305 lock = ldlm_lock_new(res);
1310 lock->l_req_mode = mode;
1311 lock->l_ast_data = data;
1312 lock->l_pid = cfs_curproc_pid();
1313 lock->l_ns_srv = !!ns_is_server(ns);
1315 lock->l_blocking_ast = cbs->lcs_blocking;
1316 lock->l_completion_ast = cbs->lcs_completion;
1317 lock->l_glimpse_ast = cbs->lcs_glimpse;
1318 lock->l_weigh_ast = cbs->lcs_weigh;
1321 lock->l_tree_node = NULL;
1322 /* if this is the extent lock, allocate the interval tree node */
1323 if (type == LDLM_EXTENT) {
1324 if (ldlm_interval_alloc(lock) == NULL)
1329 lock->l_lvb_len = lvb_len;
1330 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1331 if (lock->l_lvb_data == NULL)
1335 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1341 ldlm_lock_destroy(lock);
1342 LDLM_LOCK_RELEASE(lock);
1346 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1347 struct ldlm_lock **lockp,
1348 void *cookie, int *flags)
1350 struct ldlm_lock *lock = *lockp;
1351 struct ldlm_resource *res = lock->l_resource;
1352 int local = ns_is_client(ldlm_res_to_ns(res));
1353 ldlm_processing_policy policy;
1354 ldlm_error_t rc = ELDLM_OK;
1355 struct ldlm_interval *node = NULL;
1358 lock->l_last_activity = cfs_time_current_sec();
1359 /* policies are not executed on the client or during replay */
1360 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1361 && !local && ns->ns_policy) {
1362 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1364 if (rc == ELDLM_LOCK_REPLACED) {
1365 /* The lock that was returned has already been granted,
1366 * and placed into lockp. If it's not the same as the
1367 * one we passed in, then destroy the old one and our
1368 * work here is done. */
1369 if (lock != *lockp) {
1370 ldlm_lock_destroy(lock);
1371 LDLM_LOCK_RELEASE(lock);
1373 *flags |= LDLM_FL_LOCK_CHANGED;
1375 } else if (rc != ELDLM_OK ||
1376 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1377 ldlm_lock_destroy(lock);
1382 /* For a replaying lock, it might be already in granted list. So
1383 * unlinking the lock will cause the interval node to be freed, we
1384 * have to allocate the interval node early otherwise we can't regrant
1385 * this lock in the future. - jay */
1386 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1387 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1389 lock_res_and_lock(lock);
1390 if (local && lock->l_req_mode == lock->l_granted_mode) {
1391 /* The server returned a blocked lock, but it was granted
1392 * before we got a chance to actually enqueue it. We don't
1393 * need to do anything else. */
1394 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1395 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1396 GOTO(out, ELDLM_OK);
1399 ldlm_resource_unlink_lock(lock);
1400 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1402 ldlm_lock_destroy_nolock(lock);
1403 GOTO(out, rc = -ENOMEM);
1406 CFS_INIT_LIST_HEAD(&node->li_group);
1407 ldlm_interval_attach(node, lock);
1411 /* Some flags from the enqueue want to make it into the AST, via the
1412 * lock's l_flags. */
1413 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1415 /* This distinction between local lock trees is very important; a client
1416 * namespace only has information about locks taken by that client, and
1417 * thus doesn't have enough information to decide for itself if it can
1418 * be granted (below). In this case, we do exactly what the server
1419 * tells us to do, as dictated by the 'flags'.
1421 * We do exactly the same thing during recovery, when the server is
1422 * more or less trusting the clients not to lie.
1424 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1425 * granted/converting queues. */
1427 if (*flags & LDLM_FL_BLOCK_CONV)
1428 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1429 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1430 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1432 ldlm_grant_lock(lock, NULL);
1433 GOTO(out, ELDLM_OK);
1434 } else if (*flags & LDLM_FL_REPLAY) {
1435 if (*flags & LDLM_FL_BLOCK_CONV) {
1436 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1437 GOTO(out, ELDLM_OK);
1438 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1439 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1440 GOTO(out, ELDLM_OK);
1441 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1442 ldlm_grant_lock(lock, NULL);
1443 GOTO(out, ELDLM_OK);
1445 /* If no flags, fall through to normal enqueue path. */
1448 policy = ldlm_processing_policy_table[res->lr_type];
1449 policy(lock, flags, 1, &rc, NULL);
1452 unlock_res_and_lock(lock);
1454 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1458 /* Must be called with namespace taken: queue is waiting or converting. */
1459 int ldlm_reprocess_queue(struct ldlm_resource *res, cfs_list_t *queue,
1460 cfs_list_t *work_list)
1462 cfs_list_t *tmp, *pos;
1463 ldlm_processing_policy policy;
1465 int rc = LDLM_ITER_CONTINUE;
1469 check_res_locked(res);
1471 policy = ldlm_processing_policy_table[res->lr_type];
1474 cfs_list_for_each_safe(tmp, pos, queue) {
1475 struct ldlm_lock *pending;
1476 pending = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1478 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1481 rc = policy(pending, &flags, 0, &err, work_list);
1482 if (rc != LDLM_ITER_CONTINUE)
1490 ldlm_work_bl_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1492 struct ldlm_lock_desc d;
1493 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1498 /* nobody should touch l_bl_ast */
1499 lock_res_and_lock(lock);
1500 cfs_list_del_init(&lock->l_bl_ast);
1502 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1503 LASSERT(lock->l_bl_ast_run == 0);
1504 LASSERT(lock->l_blocking_lock);
1505 lock->l_bl_ast_run++;
1506 unlock_res_and_lock(lock);
1508 ldlm_lock2desc(lock->l_blocking_lock, &d);
1510 rc = lock->l_blocking_ast(lock, &d, (void *)arg,
1512 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1513 lock->l_blocking_lock = NULL;
1514 LDLM_LOCK_RELEASE(lock);
1520 ldlm_work_cp_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1522 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock, l_cp_ast);
1523 ldlm_completion_callback completion_callback;
1527 /* It's possible to receive a completion AST before we've set
1528 * the l_completion_ast pointer: either because the AST arrived
1529 * before the reply, or simply because there's a small race
1530 * window between receiving the reply and finishing the local
1531 * enqueue. (bug 842)
1533 * This can't happen with the blocking_ast, however, because we
1534 * will never call the local blocking_ast until we drop our
1535 * reader/writer reference, which we won't do until we get the
1536 * reply and finish enqueueing. */
1538 /* nobody should touch l_cp_ast */
1539 lock_res_and_lock(lock);
1540 cfs_list_del_init(&lock->l_cp_ast);
1541 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1542 /* save l_completion_ast since it can be changed by
1543 * mds_intent_policy(), see bug 14225 */
1544 completion_callback = lock->l_completion_ast;
1545 lock->l_flags &= ~LDLM_FL_CP_REQD;
1546 unlock_res_and_lock(lock);
1548 if (completion_callback != NULL)
1549 rc = completion_callback(lock, 0, (void *)arg);
1550 LDLM_LOCK_RELEASE(lock);
1556 ldlm_work_revoke_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1558 struct ldlm_lock_desc desc;
1559 struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1564 cfs_list_del_init(&lock->l_rk_ast);
1566 /* the desc just pretend to exclusive */
1567 ldlm_lock2desc(lock, &desc);
1568 desc.l_req_mode = LCK_EX;
1569 desc.l_granted_mode = 0;
1571 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1572 LDLM_LOCK_RELEASE(lock);
1577 int ldlm_run_ast_work(struct ldlm_namespace *ns, cfs_list_t *rpc_list,
1578 ldlm_desc_ast_t ast_type)
1580 struct l_wait_info lwi = { 0 };
1581 struct ldlm_cb_set_arg *arg;
1582 cfs_list_t *tmp, *pos;
1583 int (*work_ast_lock)(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg);
1584 unsigned int max_ast_count;
1588 if (cfs_list_empty(rpc_list))
1595 cfs_atomic_set(&arg->restart, 0);
1596 cfs_atomic_set(&arg->rpcs, 0);
1597 cfs_atomic_set(&arg->refcount, 1);
1598 cfs_waitq_init(&arg->waitq);
1601 case LDLM_WORK_BL_AST:
1602 arg->type = LDLM_BL_CALLBACK;
1603 work_ast_lock = ldlm_work_bl_ast_lock;
1605 case LDLM_WORK_CP_AST:
1606 arg->type = LDLM_CP_CALLBACK;
1607 work_ast_lock = ldlm_work_cp_ast_lock;
1609 case LDLM_WORK_REVOKE_AST:
1610 arg->type = LDLM_BL_CALLBACK;
1611 work_ast_lock = ldlm_work_revoke_ast_lock;
1617 max_ast_count = ns->ns_max_parallel_ast ? : UINT_MAX;
1618 arg->threshold = max_ast_count;
1620 cfs_list_for_each_safe(tmp, pos, rpc_list) {
1621 (void)work_ast_lock(tmp, arg);
1622 if (cfs_atomic_read(&arg->rpcs) < max_ast_count)
1625 l_wait_event(arg->waitq,
1626 cfs_atomic_read(&arg->rpcs) < arg->threshold,
1631 l_wait_event(arg->waitq, cfs_atomic_read(&arg->rpcs) == 0, &lwi);
1633 rc = cfs_atomic_read(&arg->restart) ? -ERESTART : 0;
1638 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1640 ldlm_reprocess_all(res);
1641 return LDLM_ITER_CONTINUE;
1644 static int ldlm_reprocess_res(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1645 cfs_hlist_node_t *hnode, void *arg)
1647 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1650 rc = reprocess_one_queue(res, arg);
1652 return rc == LDLM_ITER_STOP;
1655 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1660 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1661 ldlm_reprocess_res, NULL);
1666 void ldlm_reprocess_all(struct ldlm_resource *res)
1668 CFS_LIST_HEAD(rpc_list);
1672 /* Local lock trees don't get reprocessed. */
1673 if (ns_is_client(ldlm_res_to_ns(res))) {
1680 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1681 if (rc == LDLM_ITER_CONTINUE)
1682 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1685 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
1687 if (rc == -ERESTART) {
1688 LASSERT(cfs_list_empty(&rpc_list));
1694 void ldlm_cancel_callback(struct ldlm_lock *lock)
1696 check_res_locked(lock->l_resource);
1697 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1698 lock->l_flags |= LDLM_FL_CANCEL;
1699 if (lock->l_blocking_ast) {
1700 // l_check_no_ns_lock(ns);
1701 unlock_res_and_lock(lock);
1702 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1704 lock_res_and_lock(lock);
1706 LDLM_DEBUG(lock, "no blocking ast");
1709 lock->l_flags |= LDLM_FL_BL_DONE;
1712 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1714 if (req->l_resource->lr_type != LDLM_PLAIN &&
1715 req->l_resource->lr_type != LDLM_IBITS)
1718 cfs_list_del_init(&req->l_sl_policy);
1719 cfs_list_del_init(&req->l_sl_mode);
1722 void ldlm_lock_cancel(struct ldlm_lock *lock)
1724 struct ldlm_resource *res;
1725 struct ldlm_namespace *ns;
1728 lock_res_and_lock(lock);
1730 res = lock->l_resource;
1731 ns = ldlm_res_to_ns(res);
1733 /* Please do not, no matter how tempting, remove this LBUG without
1734 * talking to me first. -phik */
1735 if (lock->l_readers || lock->l_writers) {
1736 LDLM_ERROR(lock, "lock still has references");
1740 ldlm_del_waiting_lock(lock);
1742 /* Releases cancel callback. */
1743 ldlm_cancel_callback(lock);
1745 /* Yes, second time, just in case it was added again while we were
1746 running with no res lock in ldlm_cancel_callback */
1747 ldlm_del_waiting_lock(lock);
1748 ldlm_resource_unlink_lock(lock);
1749 ldlm_lock_destroy_nolock(lock);
1751 if (lock->l_granted_mode == lock->l_req_mode)
1752 ldlm_pool_del(&ns->ns_pool, lock);
1754 /* Make sure we will not be called again for same lock what is possible
1755 * if not to zero out lock->l_granted_mode */
1756 lock->l_granted_mode = LCK_MINMODE;
1757 unlock_res_and_lock(lock);
1762 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1764 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1769 if (lock->l_ast_data == NULL)
1770 lock->l_ast_data = data;
1771 if (lock->l_ast_data == data)
1773 LDLM_LOCK_PUT(lock);
1777 EXPORT_SYMBOL(ldlm_lock_set_data);
1779 int ldlm_cancel_locks_for_export_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1780 cfs_hlist_node_t *hnode, void *data)
1783 struct obd_export *exp = data;
1784 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
1785 struct ldlm_resource *res;
1787 res = ldlm_resource_getref(lock->l_resource);
1788 LDLM_LOCK_GET(lock);
1790 LDLM_DEBUG(lock, "export %p", exp);
1791 ldlm_res_lvbo_update(res, NULL, 1);
1792 ldlm_lock_cancel(lock);
1793 ldlm_reprocess_all(res);
1794 ldlm_resource_putref(res);
1795 LDLM_LOCK_RELEASE(lock);
1799 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1801 cfs_hash_for_each_empty(exp->exp_lock_hash,
1802 ldlm_cancel_locks_for_export_cb, exp);
1806 * Downgrade an exclusive lock.
1808 * A fast variant of ldlm_lock_convert for convertion of exclusive
1809 * locks. The convertion is always successful.
1811 * \param lock A lock to convert
1812 * \param new_mode new lock mode
1814 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1818 LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1819 LASSERT(new_mode == LCK_COS);
1821 lock_res_and_lock(lock);
1822 ldlm_resource_unlink_lock(lock);
1824 * Remove the lock from pool as it will be added again in
1825 * ldlm_grant_lock() called below.
1827 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
1829 lock->l_req_mode = new_mode;
1830 ldlm_grant_lock(lock, NULL);
1831 unlock_res_and_lock(lock);
1832 ldlm_reprocess_all(lock->l_resource);
1837 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1840 CFS_LIST_HEAD(rpc_list);
1841 struct ldlm_resource *res;
1842 struct ldlm_namespace *ns;
1845 struct sl_insert_point prev;
1847 struct ldlm_interval *node;
1850 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1851 *flags |= LDLM_FL_BLOCK_GRANTED;
1852 RETURN(lock->l_resource);
1855 /* I can't check the type of lock here because the bitlock of lock
1856 * is not held here, so do the allocation blindly. -jay */
1857 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1858 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1861 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1862 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1864 lock_res_and_lock(lock);
1866 res = lock->l_resource;
1867 ns = ldlm_res_to_ns(res);
1869 old_mode = lock->l_req_mode;
1870 lock->l_req_mode = new_mode;
1871 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1872 /* remember the lock position where the lock might be
1873 * added back to the granted list later and also
1874 * remember the join mode for skiplist fixing. */
1875 prev.res_link = lock->l_res_link.prev;
1876 prev.mode_link = lock->l_sl_mode.prev;
1877 prev.policy_link = lock->l_sl_policy.prev;
1878 ldlm_resource_unlink_lock(lock);
1880 ldlm_resource_unlink_lock(lock);
1881 if (res->lr_type == LDLM_EXTENT) {
1882 /* FIXME: ugly code, I have to attach the lock to a
1883 * interval node again since perhaps it will be granted
1885 CFS_INIT_LIST_HEAD(&node->li_group);
1886 ldlm_interval_attach(node, lock);
1892 * Remove old lock from the pool before adding the lock with new
1893 * mode below in ->policy()
1895 ldlm_pool_del(&ns->ns_pool, lock);
1897 /* If this is a local resource, put it on the appropriate list. */
1898 if (ns_is_client(ldlm_res_to_ns(res))) {
1899 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1900 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1902 /* This should never happen, because of the way the
1903 * server handles conversions. */
1904 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1908 ldlm_grant_lock(lock, &rpc_list);
1910 /* FIXME: completion handling not with lr_lock held ! */
1911 if (lock->l_completion_ast)
1912 lock->l_completion_ast(lock, 0, NULL);
1916 ldlm_processing_policy policy;
1917 policy = ldlm_processing_policy_table[res->lr_type];
1918 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1919 if (rc == LDLM_ITER_STOP) {
1920 lock->l_req_mode = old_mode;
1921 if (res->lr_type == LDLM_EXTENT)
1922 ldlm_extent_add_lock(res, lock);
1924 ldlm_granted_list_add_lock(lock, &prev);
1928 *flags |= LDLM_FL_BLOCK_GRANTED;
1932 unlock_res_and_lock(lock);
1935 ldlm_run_ast_work(ns, &rpc_list, LDLM_WORK_CP_AST);
1937 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1941 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1943 struct obd_device *obd = NULL;
1945 if (!((libcfs_debug | D_ERROR) & level))
1949 CDEBUG(level, " NULL LDLM lock\n");
1953 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1954 lock, lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1956 if (lock->l_conn_export != NULL)
1957 obd = lock->l_conn_export->exp_obd;
1958 if (lock->l_export && lock->l_export->exp_connection) {
1959 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1960 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1961 lock->l_remote_handle.cookie);
1962 } else if (obd == NULL) {
1963 CDEBUG(level, " Node: local\n");
1965 struct obd_import *imp = obd->u.cli.cl_import;
1966 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1967 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1968 lock->l_remote_handle.cookie);
1970 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1972 lock->l_resource->lr_name.name[0],
1973 lock->l_resource->lr_name.name[1],
1974 lock->l_resource->lr_name.name[2]);
1975 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1976 "write: %d flags: "LPX64"\n", ldlm_lockname[lock->l_req_mode],
1977 ldlm_lockname[lock->l_granted_mode],
1978 cfs_atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1980 if (lock->l_resource->lr_type == LDLM_EXTENT)
1981 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1982 " (req "LPU64"-"LPU64")\n",
1983 lock->l_policy_data.l_extent.start,
1984 lock->l_policy_data.l_extent.end,
1985 lock->l_req_extent.start, lock->l_req_extent.end);
1986 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1987 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1988 lock->l_policy_data.l_flock.pid,
1989 lock->l_policy_data.l_flock.start,
1990 lock->l_policy_data.l_flock.end);
1991 else if (lock->l_resource->lr_type == LDLM_IBITS)
1992 CDEBUG(level, " Bits: "LPX64"\n",
1993 lock->l_policy_data.l_inodebits.bits);
1996 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1998 struct ldlm_lock *lock;
2000 if (!((libcfs_debug | D_ERROR) & level))
2003 lock = ldlm_handle2lock(lockh);
2007 ldlm_lock_dump(D_OTHER, lock, 0);
2009 LDLM_LOCK_PUT(lock);
2012 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
2013 struct libcfs_debug_msg_data *data, const char *fmt,
2017 cfs_debug_limit_state_t *cdls = data->msg_cdls;
2019 va_start(args, fmt);
2021 if (lock->l_resource == NULL) {
2022 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2023 data->msg_fn, data->msg_line, fmt, args,
2024 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2025 "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" remote: "
2026 LPX64" expref: %d pid: %u timeout: %lu\n", lock,
2027 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2028 lock->l_readers, lock->l_writers,
2029 ldlm_lockname[lock->l_granted_mode],
2030 ldlm_lockname[lock->l_req_mode],
2031 lock->l_flags, lock->l_remote_handle.cookie,
2033 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2034 lock->l_pid, lock->l_callback_timeout);
2039 switch (lock->l_resource->lr_type) {
2041 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2042 data->msg_fn, data->msg_line, fmt, args,
2043 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2044 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
2045 "] (req "LPU64"->"LPU64") flags: "LPX64" remote: "LPX64
2046 " expref: %d pid: %u timeout %lu\n",
2047 ldlm_lock_to_ns_name(lock), lock,
2048 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2049 lock->l_readers, lock->l_writers,
2050 ldlm_lockname[lock->l_granted_mode],
2051 ldlm_lockname[lock->l_req_mode],
2052 lock->l_resource->lr_name.name[0],
2053 lock->l_resource->lr_name.name[1],
2054 cfs_atomic_read(&lock->l_resource->lr_refcount),
2055 ldlm_typename[lock->l_resource->lr_type],
2056 lock->l_policy_data.l_extent.start,
2057 lock->l_policy_data.l_extent.end,
2058 lock->l_req_extent.start, lock->l_req_extent.end,
2059 lock->l_flags, lock->l_remote_handle.cookie,
2061 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2062 lock->l_pid, lock->l_callback_timeout);
2066 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2067 data->msg_fn, data->msg_line, fmt, args,
2068 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2069 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
2070 "["LPU64"->"LPU64"] flags: "LPX64" remote: "LPX64
2071 " expref: %d pid: %u timeout: %lu\n",
2072 ldlm_lock_to_ns_name(lock), lock,
2073 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2074 lock->l_readers, lock->l_writers,
2075 ldlm_lockname[lock->l_granted_mode],
2076 ldlm_lockname[lock->l_req_mode],
2077 lock->l_resource->lr_name.name[0],
2078 lock->l_resource->lr_name.name[1],
2079 cfs_atomic_read(&lock->l_resource->lr_refcount),
2080 ldlm_typename[lock->l_resource->lr_type],
2081 lock->l_policy_data.l_flock.pid,
2082 lock->l_policy_data.l_flock.start,
2083 lock->l_policy_data.l_flock.end,
2084 lock->l_flags, lock->l_remote_handle.cookie,
2086 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2087 lock->l_pid, lock->l_callback_timeout);
2091 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2092 data->msg_fn, data->msg_line, fmt, args,
2093 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2094 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
2095 "flags: "LPX64" remote: "LPX64" expref: %d "
2096 "pid: %u timeout: %lu\n",
2097 ldlm_lock_to_ns_name(lock),
2098 lock, lock->l_handle.h_cookie,
2099 cfs_atomic_read (&lock->l_refc),
2100 lock->l_readers, lock->l_writers,
2101 ldlm_lockname[lock->l_granted_mode],
2102 ldlm_lockname[lock->l_req_mode],
2103 lock->l_resource->lr_name.name[0],
2104 lock->l_resource->lr_name.name[1],
2105 lock->l_policy_data.l_inodebits.bits,
2106 cfs_atomic_read(&lock->l_resource->lr_refcount),
2107 ldlm_typename[lock->l_resource->lr_type],
2108 lock->l_flags, lock->l_remote_handle.cookie,
2110 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2111 lock->l_pid, lock->l_callback_timeout);
2115 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2116 data->msg_fn, data->msg_line, fmt, args,
2117 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2118 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" "
2119 "remote: "LPX64" expref: %d pid: %u timeout %lu\n",
2120 ldlm_lock_to_ns_name(lock),
2121 lock, lock->l_handle.h_cookie,
2122 cfs_atomic_read (&lock->l_refc),
2123 lock->l_readers, lock->l_writers,
2124 ldlm_lockname[lock->l_granted_mode],
2125 ldlm_lockname[lock->l_req_mode],
2126 lock->l_resource->lr_name.name[0],
2127 lock->l_resource->lr_name.name[1],
2128 cfs_atomic_read(&lock->l_resource->lr_refcount),
2129 ldlm_typename[lock->l_resource->lr_type],
2130 lock->l_flags, lock->l_remote_handle.cookie,
2132 cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2133 lock->l_pid, lock->l_callback_timeout);
2138 EXPORT_SYMBOL(_ldlm_lock_debug);