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 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_lock.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_LDLM
45 # include <libcfs/libcfs.h>
46 # ifndef HAVE_VFS_INTENT_PATCHES
47 # include <linux/lustre_intent.h>
50 # include <liblustre.h>
53 #include <obd_class.h>
54 #include "ldlm_internal.h"
57 char *ldlm_lockname[] = {
68 char *ldlm_typename[] = {
75 char *ldlm_it2str(int it)
82 case (IT_OPEN | IT_CREAT):
95 CERROR("Unknown intent %d\n", it);
100 extern cfs_mem_cache_t *ldlm_lock_slab;
102 static ldlm_processing_policy ldlm_processing_policy_table[] = {
103 [LDLM_PLAIN] ldlm_process_plain_lock,
104 [LDLM_EXTENT] ldlm_process_extent_lock,
106 [LDLM_FLOCK] ldlm_process_flock_lock,
108 [LDLM_IBITS] ldlm_process_inodebits_lock,
111 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
113 return ldlm_processing_policy_table[res->lr_type];
116 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
122 * REFCOUNTED LOCK OBJECTS
127 * Lock refcounts, during creation:
128 * - one special one for allocation, dec'd only once in destroy
129 * - one for being a lock that's in-use
130 * - one for the addref associated with a new lock
132 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
134 atomic_inc(&lock->l_refc);
138 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
140 LASSERT(size == sizeof(*lock));
141 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
144 void ldlm_lock_put(struct ldlm_lock *lock)
148 LASSERT(lock->l_resource != LP_POISON);
149 LASSERT(atomic_read(&lock->l_refc) > 0);
150 if (atomic_dec_and_test(&lock->l_refc)) {
151 struct ldlm_resource *res;
154 "final lock_put on destroyed lock, freeing it.");
156 res = lock->l_resource;
157 LASSERT(lock->l_destroyed);
158 LASSERT(list_empty(&lock->l_res_link));
159 LASSERT(list_empty(&lock->l_pending_chain));
161 atomic_dec(&res->lr_namespace->ns_locks);
162 ldlm_resource_putref(res);
163 lock->l_resource = NULL;
164 if (lock->l_export) {
165 class_export_put(lock->l_export);
166 lock->l_export = NULL;
169 if (lock->l_lvb_data != NULL)
170 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
172 ldlm_interval_free(ldlm_interval_detach(lock));
173 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
180 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
183 if (!list_empty(&lock->l_lru)) {
184 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
185 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
186 list_del_init(&lock->l_lru);
188 LASSERT(ns->ns_nr_unused >= 0);
194 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
196 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
199 spin_lock(&ns->ns_unused_lock);
200 rc = ldlm_lock_remove_from_lru_nolock(lock);
201 spin_unlock(&ns->ns_unused_lock);
206 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
208 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
209 lock->l_last_used = cfs_time_current();
210 LASSERT(list_empty(&lock->l_lru));
211 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
212 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
213 LASSERT(ns->ns_nr_unused >= 0);
217 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
219 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
221 spin_lock(&ns->ns_unused_lock);
222 ldlm_lock_add_to_lru_nolock(lock);
223 spin_unlock(&ns->ns_unused_lock);
227 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
229 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
231 spin_lock(&ns->ns_unused_lock);
232 if (!list_empty(&lock->l_lru)) {
233 ldlm_lock_remove_from_lru_nolock(lock);
234 ldlm_lock_add_to_lru_nolock(lock);
236 spin_unlock(&ns->ns_unused_lock);
240 /* This used to have a 'strict' flag, which recovery would use to mark an
241 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
242 * shall explain why it's gone: with the new hash table scheme, once you call
243 * ldlm_lock_destroy, you can never drop your final references on this lock.
244 * Because it's not in the hash table anymore. -phil */
245 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
249 if (lock->l_readers || lock->l_writers) {
250 LDLM_ERROR(lock, "lock still has references");
251 ldlm_lock_dump(D_ERROR, lock, 0);
255 if (!list_empty(&lock->l_res_link)) {
256 LDLM_ERROR(lock, "lock still on resource");
257 ldlm_lock_dump(D_ERROR, lock, 0);
261 if (lock->l_destroyed) {
262 LASSERT(list_empty(&lock->l_lru));
266 lock->l_destroyed = 1;
268 if (lock->l_export && lock->l_export->exp_lock_hash)
269 lustre_hash_del(lock->l_export->exp_lock_hash,
270 &lock->l_remote_handle, &lock->l_exp_hash);
272 ldlm_lock_remove_from_lru(lock);
273 class_handle_unhash(&lock->l_handle);
276 /* Wake anyone waiting for this lock */
277 /* FIXME: I should probably add yet another flag, instead of using
278 * l_export to only call this on clients */
280 class_export_put(lock->l_export);
281 lock->l_export = NULL;
282 if (lock->l_export && lock->l_completion_ast)
283 lock->l_completion_ast(lock, 0);
289 void ldlm_lock_destroy(struct ldlm_lock *lock)
293 lock_res_and_lock(lock);
294 first = ldlm_lock_destroy_internal(lock);
295 unlock_res_and_lock(lock);
297 /* drop reference from hashtable only for first destroy */
303 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
307 first = ldlm_lock_destroy_internal(lock);
308 /* drop reference from hashtable only for first destroy */
314 /* this is called by portals_handle2object with the handle lock taken */
315 static void lock_handle_addref(void *lock)
317 LDLM_LOCK_GET((struct ldlm_lock *)lock);
321 * usage: pass in a resource on which you have done ldlm_resource_get
322 * pass in a parent lock on which you have done a ldlm_lock_get
323 * after return, ldlm_*_put the resource and parent
324 * returns: lock with refcount 2 - one for current caller and one for remote
326 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
328 struct ldlm_lock *lock;
331 if (resource == NULL)
334 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, CFS_ALLOC_IO, sizeof(*lock));
338 spin_lock_init(&lock->l_lock);
339 lock->l_resource = ldlm_resource_getref(resource);
341 atomic_set(&lock->l_refc, 2);
342 CFS_INIT_LIST_HEAD(&lock->l_res_link);
343 CFS_INIT_LIST_HEAD(&lock->l_lru);
344 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
345 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
346 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
347 CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
348 cfs_waitq_init(&lock->l_waitq);
349 lock->l_blocking_lock = NULL;
350 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
351 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
352 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
354 atomic_inc(&resource->lr_namespace->ns_locks);
355 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
356 class_handle_hash(&lock->l_handle, lock_handle_addref);
358 CFS_INIT_LIST_HEAD(&lock->l_extents_list);
359 spin_lock_init(&lock->l_extents_list_lock);
360 CFS_INIT_LIST_HEAD(&lock->l_cache_locks_list);
365 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
366 const struct ldlm_res_id *new_resid)
368 struct ldlm_resource *oldres = lock->l_resource;
369 struct ldlm_resource *newres;
373 LASSERT(ns_is_client(ns));
375 lock_res_and_lock(lock);
376 if (memcmp(new_resid, &lock->l_resource->lr_name,
377 sizeof(lock->l_resource->lr_name)) == 0) {
379 unlock_res_and_lock(lock);
383 LASSERT(new_resid->name[0] != 0);
385 /* This function assumes that the lock isn't on any lists */
386 LASSERT(list_empty(&lock->l_res_link));
388 type = oldres->lr_type;
389 unlock_res_and_lock(lock);
391 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
395 * To flip the lock from the old to the new resource, lock, oldres and
396 * newres have to be locked. Resource spin-locks are nested within
397 * lock->l_lock, and are taken in the memory address order to avoid
400 spin_lock(&lock->l_lock);
401 oldres = lock->l_resource;
402 if (oldres < newres) {
404 lock_res_nested(newres, LRT_NEW);
407 lock_res_nested(oldres, LRT_NEW);
409 LASSERT(memcmp(new_resid, &oldres->lr_name,
410 sizeof oldres->lr_name) != 0);
411 lock->l_resource = newres;
413 unlock_res_and_lock(lock);
415 /* ...and the flowers are still standing! */
416 ldlm_resource_putref(oldres);
425 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
427 lockh->cookie = lock->l_handle.h_cookie;
430 /* if flags: atomically get the lock and set the flags.
431 * Return NULL if flag already set
434 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
437 struct ldlm_namespace *ns;
438 struct ldlm_lock *lock, *retval = NULL;
443 lock = class_handle2object(handle->cookie);
447 LASSERT(lock->l_resource != NULL);
448 ns = lock->l_resource->lr_namespace;
451 lock_res_and_lock(lock);
453 /* It's unlikely but possible that someone marked the lock as
454 * destroyed after we did handle2object on it */
455 if (lock->l_destroyed) {
456 unlock_res_and_lock(lock);
457 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
462 if (flags && (lock->l_flags & flags)) {
463 unlock_res_and_lock(lock);
469 lock->l_flags |= flags;
471 unlock_res_and_lock(lock);
478 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
479 const struct lustre_handle *handle)
481 struct ldlm_lock *retval = NULL;
482 retval = __ldlm_handle2lock(handle, 0);
486 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
488 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
489 /* INODEBITS_INTEROP: If the other side does not support
490 * inodebits, reply with a plain lock descriptor.
492 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
493 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
494 struct ldlm_resource res = *lock->l_resource;
496 /* Make sure all the right bits are set in this lock we
497 are going to pass to client */
498 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
499 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
500 "Inappropriate inode lock bits during "
501 "conversion " LPU64 "\n",
502 lock->l_policy_data.l_inodebits.bits);
503 res.lr_type = LDLM_PLAIN;
504 ldlm_res2desc(&res, &desc->l_resource);
505 /* Convert "new" lock mode to something old client can
507 if ((lock->l_req_mode == LCK_CR) ||
508 (lock->l_req_mode == LCK_CW))
509 desc->l_req_mode = LCK_PR;
511 desc->l_req_mode = lock->l_req_mode;
512 if ((lock->l_granted_mode == LCK_CR) ||
513 (lock->l_granted_mode == LCK_CW)) {
514 desc->l_granted_mode = LCK_PR;
516 /* We never grant PW/EX locks to clients */
517 LASSERT((lock->l_granted_mode != LCK_PW) &&
518 (lock->l_granted_mode != LCK_EX));
519 desc->l_granted_mode = lock->l_granted_mode;
522 /* We do not copy policy here, because there is no
523 policy for plain locks */
525 ldlm_res2desc(lock->l_resource, &desc->l_resource);
526 desc->l_req_mode = lock->l_req_mode;
527 desc->l_granted_mode = lock->l_granted_mode;
528 desc->l_policy_data = lock->l_policy_data;
532 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
533 struct list_head *work_list)
535 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
536 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
537 lock->l_flags |= LDLM_FL_AST_SENT;
538 /* If the enqueuing client said so, tell the AST recipient to
539 * discard dirty data, rather than writing back. */
540 if (new->l_flags & LDLM_AST_DISCARD_DATA)
541 lock->l_flags |= LDLM_FL_DISCARD_DATA;
542 LASSERT(list_empty(&lock->l_bl_ast));
543 list_add(&lock->l_bl_ast, work_list);
545 LASSERT(lock->l_blocking_lock == NULL);
546 lock->l_blocking_lock = LDLM_LOCK_GET(new);
550 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
552 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
553 lock->l_flags |= LDLM_FL_CP_REQD;
554 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
555 LASSERT(list_empty(&lock->l_cp_ast));
556 list_add(&lock->l_cp_ast, work_list);
561 /* must be called with lr_lock held */
562 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
563 struct list_head *work_list)
566 check_res_locked(lock->l_resource);
568 ldlm_add_bl_work_item(lock, new, work_list);
570 ldlm_add_cp_work_item(lock, work_list);
574 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
576 struct ldlm_lock *lock;
578 lock = ldlm_handle2lock(lockh);
579 LASSERT(lock != NULL);
580 ldlm_lock_addref_internal(lock, mode);
584 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
586 ldlm_lock_remove_from_lru(lock);
587 if (mode & (LCK_NL | LCK_CR | LCK_PR))
589 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
592 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
595 /* only called for local locks */
596 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
598 lock_res_and_lock(lock);
599 ldlm_lock_addref_internal_nolock(lock, mode);
600 unlock_res_and_lock(lock);
603 /* only called in ldlm_flock_destroy and for local locks.
604 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
605 * * ldlm_lock_remove_from_lru() does nothing, it is safe
606 * * for ldlm_flock_destroy usage by dropping some code */
607 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
609 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
610 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
611 LASSERT(lock->l_readers > 0);
614 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
615 LASSERT(lock->l_writers > 0);
619 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
622 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
624 struct ldlm_namespace *ns;
627 lock_res_and_lock(lock);
629 ns = lock->l_resource->lr_namespace;
631 ldlm_lock_decref_internal_nolock(lock, mode);
633 if (lock->l_flags & LDLM_FL_LOCAL &&
634 !lock->l_readers && !lock->l_writers) {
635 /* If this is a local lock on a server namespace and this was
636 * the last reference, cancel the lock. */
637 CDEBUG(D_INFO, "forcing cancel of local lock\n");
638 lock->l_flags |= LDLM_FL_CBPENDING;
641 if (!lock->l_readers && !lock->l_writers &&
642 (lock->l_flags & LDLM_FL_CBPENDING)) {
643 /* If we received a blocked AST and this was the last reference,
644 * run the callback. */
645 if (ns_is_server(ns) && lock->l_export)
646 CERROR("FL_CBPENDING set on non-local lock--just a "
649 LDLM_DEBUG(lock, "final decref done on cbpending lock");
651 LDLM_LOCK_GET(lock); /* dropped by bl thread */
652 ldlm_lock_remove_from_lru(lock);
653 unlock_res_and_lock(lock);
654 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
655 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
656 ldlm_handle_bl_callback(ns, NULL, lock);
657 } else if (ns_is_client(ns) &&
658 !lock->l_readers && !lock->l_writers &&
659 !(lock->l_flags & LDLM_FL_BL_AST)) {
660 /* If this is a client-side namespace and this was the last
661 * reference, put it on the LRU. */
662 ldlm_lock_add_to_lru(lock);
663 unlock_res_and_lock(lock);
664 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
665 * are not supported by the server, otherwise, it is done on
667 if (!exp_connect_cancelset(lock->l_conn_export) &&
668 !ns_connect_lru_resize(ns))
669 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
671 unlock_res_and_lock(lock);
677 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
679 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
680 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
681 ldlm_lock_decref_internal(lock, mode);
685 /* This will drop a lock reference and mark it for destruction, but will not
686 * necessarily cancel the lock before returning. */
687 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
689 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
692 LASSERT(lock != NULL);
694 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
695 lock_res_and_lock(lock);
696 lock->l_flags |= LDLM_FL_CBPENDING;
697 unlock_res_and_lock(lock);
698 ldlm_lock_decref_internal(lock, mode);
702 struct sl_insert_point {
703 struct list_head *res_link;
704 struct list_head *mode_link;
705 struct list_head *policy_link;
709 * search_granted_lock
712 * Finds a position to insert the new lock.
714 * queue [input]: the granted list where search acts on;
715 * req [input]: the lock whose position to be located;
716 * prev [output]: positions within 3 lists to insert @req to
720 * - ldlm_grant_lock_with_skiplist
722 static void search_granted_lock(struct list_head *queue,
723 struct ldlm_lock *req,
724 struct sl_insert_point *prev)
726 struct list_head *tmp;
727 struct ldlm_lock *lock, *mode_end, *policy_end;
730 list_for_each(tmp, queue) {
731 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
733 mode_end = list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
736 if (lock->l_req_mode != req->l_req_mode) {
737 /* jump to last lock of mode group */
738 tmp = &mode_end->l_res_link;
742 /* suitable mode group is found */
743 if (lock->l_resource->lr_type == LDLM_PLAIN) {
744 /* insert point is last lock of the mode group */
745 prev->res_link = &mode_end->l_res_link;
746 prev->mode_link = &mode_end->l_sl_mode;
747 prev->policy_link = &req->l_sl_policy;
750 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
752 policy_end = list_entry(lock->l_sl_policy.prev,
756 if (lock->l_policy_data.l_inodebits.bits ==
757 req->l_policy_data.l_inodebits.bits) {
758 /* insert point is last lock of
759 * the policy group */
761 &policy_end->l_res_link;
763 &policy_end->l_sl_mode;
765 &policy_end->l_sl_policy;
770 if (policy_end == mode_end)
771 /* done with mode group */
774 /* jump to next policy group within the mode group */
775 tmp = policy_end->l_res_link.next;
776 lock = list_entry(tmp, struct ldlm_lock,
778 } /* loop over policy groups within the mode group */
780 /* insert point is last lock of the mode group,
781 * new policy group is started */
782 prev->res_link = &mode_end->l_res_link;
783 prev->mode_link = &mode_end->l_sl_mode;
784 prev->policy_link = &req->l_sl_policy;
788 LDLM_ERROR(lock, "is not LDLM_PLAIN or LDLM_IBITS lock");
793 /* insert point is last lock on the queue,
794 * new mode group and new policy group are started */
795 prev->res_link = queue->prev;
796 prev->mode_link = &req->l_sl_mode;
797 prev->policy_link = &req->l_sl_policy;
802 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
803 struct sl_insert_point *prev)
805 struct ldlm_resource *res = lock->l_resource;
808 check_res_locked(res);
810 ldlm_resource_dump(D_INFO, res);
811 CDEBUG(D_OTHER, "About to add this lock:\n");
812 ldlm_lock_dump(D_OTHER, lock, 0);
814 if (lock->l_destroyed) {
815 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
819 LASSERT(list_empty(&lock->l_res_link));
820 LASSERT(list_empty(&lock->l_sl_mode));
821 LASSERT(list_empty(&lock->l_sl_policy));
823 list_add(&lock->l_res_link, prev->res_link);
824 list_add(&lock->l_sl_mode, prev->mode_link);
825 list_add(&lock->l_sl_policy, prev->policy_link);
830 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
832 struct sl_insert_point prev;
835 LASSERT(lock->l_req_mode == lock->l_granted_mode);
837 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
838 ldlm_granted_list_add_lock(lock, &prev);
843 * - ldlm_lock_enqueue
844 * - ldlm_reprocess_queue
845 * - ldlm_lock_convert
847 * must be called with lr_lock held
849 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
851 struct ldlm_resource *res = lock->l_resource;
854 check_res_locked(res);
856 lock->l_granted_mode = lock->l_req_mode;
857 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
858 ldlm_grant_lock_with_skiplist(lock);
859 else if (res->lr_type == LDLM_EXTENT)
860 ldlm_extent_add_lock(res, lock);
862 ldlm_resource_add_lock(res, &res->lr_granted, lock);
864 if (lock->l_granted_mode < res->lr_most_restr)
865 res->lr_most_restr = lock->l_granted_mode;
867 if (work_list && lock->l_completion_ast != NULL)
868 ldlm_add_ast_work_item(lock, NULL, work_list);
870 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
874 /* returns a referenced lock or NULL. See the flag descriptions below, in the
875 * comment above ldlm_lock_match */
876 static struct ldlm_lock *search_queue(struct list_head *queue,
878 ldlm_policy_data_t *policy,
879 struct ldlm_lock *old_lock, int flags)
881 struct ldlm_lock *lock;
882 struct list_head *tmp;
884 list_for_each(tmp, queue) {
887 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
889 if (lock == old_lock)
892 /* llite sometimes wants to match locks that will be
893 * canceled when their users drop, but we allow it to match
894 * if it passes in CBPENDING and the lock still has users.
895 * this is generally only going to be used by children
896 * whose parents already hold a lock so forward progress
897 * can still happen. */
898 if (lock->l_flags & LDLM_FL_CBPENDING &&
899 !(flags & LDLM_FL_CBPENDING))
901 if (lock->l_flags & LDLM_FL_CBPENDING &&
902 lock->l_readers == 0 && lock->l_writers == 0)
905 if (!(lock->l_req_mode & *mode))
907 match = lock->l_req_mode;
909 if (lock->l_resource->lr_type == LDLM_EXTENT &&
910 (lock->l_policy_data.l_extent.start >
911 policy->l_extent.start ||
912 lock->l_policy_data.l_extent.end < policy->l_extent.end))
915 if (unlikely(match == LCK_GROUP) &&
916 lock->l_resource->lr_type == LDLM_EXTENT &&
917 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
920 /* We match if we have existing lock with same or wider set
922 if (lock->l_resource->lr_type == LDLM_IBITS &&
923 ((lock->l_policy_data.l_inodebits.bits &
924 policy->l_inodebits.bits) !=
925 policy->l_inodebits.bits))
928 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
931 if ((flags & LDLM_FL_LOCAL_ONLY) &&
932 !(lock->l_flags & LDLM_FL_LOCAL))
935 if (flags & LDLM_FL_TEST_LOCK) {
937 ldlm_lock_touch_in_lru(lock);
939 ldlm_lock_addref_internal_nolock(lock, match);
948 void ldlm_lock_allow_match(struct ldlm_lock *lock)
950 lock_res_and_lock(lock);
951 lock->l_flags |= LDLM_FL_LVB_READY;
952 cfs_waitq_signal(&lock->l_waitq);
953 unlock_res_and_lock(lock);
957 * Checks if requested extent lock is compatible with another owned lock.
959 * Checks if \a lock is compatible with a read or write lock
960 * (specified by \a rw) for an extent [\a start , \a end].
962 * \param lock the already owned lock
963 * \param rw OBD_BRW_READ if requested for reading,
964 * OBD_BRW_WRITE if requested for writing
965 * \param start start of the requested extent
966 * \param end end of the requested extent
967 * \param cookie transparent parameter for passing locking context
969 * \post result == 1, *cookie == context, appropriate lock is referenced
971 * \retval 1 owned lock is reused for the request
972 * \retval 0 no lock reused for the request
974 * \see ldlm_lock_fast_release
976 int ldlm_lock_fast_match(struct ldlm_lock *lock, int rw,
977 obd_off start, obd_off end,
980 LASSERT(rw == OBD_BRW_READ || rw == OBD_BRW_WRITE);
985 lock_res_and_lock(lock);
986 /* check if granted mode is compatible */
987 if (rw == OBD_BRW_WRITE &&
988 !(lock->l_granted_mode & (LCK_PW|LCK_GROUP)))
991 /* does the lock cover the region we would like to access? */
992 if ((lock->l_policy_data.l_extent.start > start) ||
993 (lock->l_policy_data.l_extent.end < end))
996 /* if we received a blocking callback and the lock is no longer
997 * referenced, don't use it */
998 if ((lock->l_flags & LDLM_FL_CBPENDING) &&
999 !lock->l_writers && !lock->l_readers)
1002 ldlm_lock_addref_internal_nolock(lock, rw == OBD_BRW_WRITE ? LCK_PW : LCK_PR);
1003 unlock_res_and_lock(lock);
1004 *cookie = (void *)lock;
1005 return 1; /* avoid using rc for stack relief */
1008 unlock_res_and_lock(lock);
1013 * Releases a reference to a lock taken in a "fast" way.
1015 * Releases a read or write (specified by \a rw) lock
1016 * referenced by \a cookie.
1018 * \param rw OBD_BRW_READ if requested for reading,
1019 * OBD_BRW_WRITE if requested for writing
1020 * \param cookie transparent parameter for passing locking context
1022 * \post appropriate lock is dereferenced
1024 * \see ldlm_lock_fast_lock
1026 void ldlm_lock_fast_release(void *cookie, int rw)
1028 struct ldlm_lock *lock = (struct ldlm_lock *)cookie;
1030 LASSERT(lock != NULL);
1031 LASSERT(rw == OBD_BRW_READ || rw == OBD_BRW_WRITE);
1032 LASSERT(rw == OBD_BRW_READ ||
1033 (lock->l_granted_mode & (LCK_PW | LCK_GROUP)));
1034 ldlm_lock_decref_internal(lock, rw == OBD_BRW_WRITE ? LCK_PW : LCK_PR);
1037 /* Can be called in two ways:
1039 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1040 * for a duplicate of.
1042 * Otherwise, all of the fields must be filled in, to match against.
1044 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1045 * server (ie, connh is NULL)
1046 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1047 * list will be considered
1048 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1049 * to be canceled can still be matched as long as they still have reader
1050 * or writer refernces
1051 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1052 * just tell us if we would have matched.
1054 * Returns 1 if it finds an already-existing lock that is compatible; in this
1055 * case, lockh is filled in with a addref()ed lock
1057 * we also check security context, if that failed we simply return 0 (to keep
1058 * caller code unchanged), the context failure will be discovered by caller
1061 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1062 const struct ldlm_res_id *res_id, ldlm_type_t type,
1063 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1064 struct lustre_handle *lockh)
1066 struct ldlm_resource *res;
1067 struct ldlm_lock *lock, *old_lock = NULL;
1072 old_lock = ldlm_handle2lock(lockh);
1075 ns = old_lock->l_resource->lr_namespace;
1076 res_id = &old_lock->l_resource->lr_name;
1077 type = old_lock->l_resource->lr_type;
1078 mode = old_lock->l_req_mode;
1081 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1083 LASSERT(old_lock == NULL);
1089 lock = search_queue(&res->lr_granted, &mode, policy, old_lock, flags);
1092 if (flags & LDLM_FL_BLOCK_GRANTED)
1094 lock = search_queue(&res->lr_converting, &mode, policy, old_lock, flags);
1097 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock, flags);
1104 ldlm_resource_putref(res);
1107 ldlm_lock2handle(lock, lockh);
1108 if ((flags & LDLM_FL_LVB_READY) &&
1109 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1110 struct l_wait_info lwi;
1111 if (lock->l_completion_ast) {
1112 int err = lock->l_completion_ast(lock,
1113 LDLM_FL_WAIT_NOREPROC,
1116 if (flags & LDLM_FL_TEST_LOCK)
1117 LDLM_LOCK_PUT(lock);
1119 ldlm_lock_decref_internal(lock, mode);
1125 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), NULL,
1126 LWI_ON_SIGNAL_NOOP, NULL);
1128 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1129 l_wait_event(lock->l_waitq,
1130 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1135 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1136 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1137 res_id->name[2] : policy->l_extent.start,
1138 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1139 res_id->name[3] : policy->l_extent.end);
1141 /* check user's security context */
1142 if (lock->l_conn_export &&
1143 sptlrpc_import_check_ctx(
1144 class_exp2cliimp(lock->l_conn_export))) {
1145 if (!(flags & LDLM_FL_TEST_LOCK))
1146 ldlm_lock_decref_internal(lock, mode);
1150 if (flags & LDLM_FL_TEST_LOCK)
1151 LDLM_LOCK_PUT(lock);
1153 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1154 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1155 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1156 type, mode, res_id->name[0], res_id->name[1],
1157 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1158 res_id->name[2] :policy->l_extent.start,
1159 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1160 res_id->name[3] : policy->l_extent.end);
1163 LDLM_LOCK_PUT(old_lock);
1165 return rc ? mode : 0;
1168 /* Returns a referenced lock */
1169 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1170 const struct ldlm_res_id *res_id,
1173 const struct ldlm_callback_suite *cbs,
1174 void *data, __u32 lvb_len)
1176 struct ldlm_lock *lock;
1177 struct ldlm_resource *res;
1180 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1184 lock = ldlm_lock_new(res);
1185 ldlm_resource_putref(res);
1190 lock->l_req_mode = mode;
1191 lock->l_ast_data = data;
1192 lock->l_pid = cfs_curproc_pid();
1194 lock->l_blocking_ast = cbs->lcs_blocking;
1195 lock->l_completion_ast = cbs->lcs_completion;
1196 lock->l_glimpse_ast = cbs->lcs_glimpse;
1197 lock->l_weigh_ast = cbs->lcs_weigh;
1200 lock->l_tree_node = NULL;
1201 /* if this is the extent lock, allocate the interval tree node */
1202 if (type == LDLM_EXTENT) {
1203 if (ldlm_interval_alloc(lock) == NULL)
1208 lock->l_lvb_len = lvb_len;
1209 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1210 if (lock->l_lvb_data == NULL)
1217 if (lock->l_lvb_data)
1218 OBD_FREE(lock->l_lvb_data, lvb_len);
1219 ldlm_interval_free(ldlm_interval_detach(lock));
1220 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1224 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1225 struct ldlm_lock **lockp,
1226 void *cookie, int *flags)
1228 struct ldlm_lock *lock = *lockp;
1229 struct ldlm_resource *res = lock->l_resource;
1230 int local = ns_is_client(res->lr_namespace);
1231 ldlm_processing_policy policy;
1232 ldlm_error_t rc = ELDLM_OK;
1233 struct ldlm_interval *node = NULL;
1236 do_gettimeofday(&lock->l_enqueued_time);
1237 /* policies are not executed on the client or during replay */
1238 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1239 && !local && ns->ns_policy) {
1240 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1242 if (rc == ELDLM_LOCK_REPLACED) {
1243 /* The lock that was returned has already been granted,
1244 * and placed into lockp. If it's not the same as the
1245 * one we passed in, then destroy the old one and our
1246 * work here is done. */
1247 if (lock != *lockp) {
1248 ldlm_lock_destroy(lock);
1249 LDLM_LOCK_PUT(lock);
1251 *flags |= LDLM_FL_LOCK_CHANGED;
1253 } else if (rc != ELDLM_OK ||
1254 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1255 ldlm_lock_destroy(lock);
1260 /* For a replaying lock, it might be already in granted list. So
1261 * unlinking the lock will cause the interval node to be freed, we
1262 * have to allocate the interval node early otherwise we can't regrant
1263 * this lock in the future. - jay */
1264 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1265 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO,
1268 lock_res_and_lock(lock);
1269 if (local && lock->l_req_mode == lock->l_granted_mode) {
1270 /* The server returned a blocked lock, but it was granted
1271 * before we got a chance to actually enqueue it. We don't
1272 * need to do anything else. */
1273 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1274 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1275 GOTO(out, ELDLM_OK);
1278 ldlm_resource_unlink_lock(lock);
1279 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1281 ldlm_lock_destroy_nolock(lock);
1282 GOTO(out, rc = -ENOMEM);
1285 CFS_INIT_LIST_HEAD(&node->li_group);
1286 ldlm_interval_attach(node, lock);
1290 /* Some flags from the enqueue want to make it into the AST, via the
1291 * lock's l_flags. */
1292 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1294 /* This distinction between local lock trees is very important; a client
1295 * namespace only has information about locks taken by that client, and
1296 * thus doesn't have enough information to decide for itself if it can
1297 * be granted (below). In this case, we do exactly what the server
1298 * tells us to do, as dictated by the 'flags'.
1300 * We do exactly the same thing during recovery, when the server is
1301 * more or less trusting the clients not to lie.
1303 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1304 * granted/converting queues. */
1306 if (*flags & LDLM_FL_BLOCK_CONV)
1307 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1308 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1309 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1311 ldlm_grant_lock(lock, NULL);
1312 GOTO(out, ELDLM_OK);
1313 } else if (*flags & LDLM_FL_REPLAY) {
1314 if (*flags & LDLM_FL_BLOCK_CONV) {
1315 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1316 GOTO(out, ELDLM_OK);
1317 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1318 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1319 GOTO(out, ELDLM_OK);
1320 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1321 ldlm_grant_lock(lock, NULL);
1322 GOTO(out, ELDLM_OK);
1324 /* If no flags, fall through to normal enqueue path. */
1327 policy = ldlm_processing_policy_table[res->lr_type];
1328 policy(lock, flags, 1, &rc, NULL);
1331 unlock_res_and_lock(lock);
1333 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1337 /* Must be called with namespace taken: queue is waiting or converting. */
1338 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1339 struct list_head *work_list)
1341 struct list_head *tmp, *pos;
1342 ldlm_processing_policy policy;
1344 int rc = LDLM_ITER_CONTINUE;
1348 check_res_locked(res);
1350 policy = ldlm_processing_policy_table[res->lr_type];
1353 list_for_each_safe(tmp, pos, queue) {
1354 struct ldlm_lock *pending;
1355 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1357 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1360 rc = policy(pending, &flags, 0, &err, work_list);
1361 if (rc != LDLM_ITER_CONTINUE)
1368 /* Helper function for ldlm_run_ast_work().
1370 * Send an existing rpc set specified by @arg->set and then
1371 * destroy it. Create new one if @do_create flag is set. */
1373 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1377 ptlrpc_set_wait(arg->set);
1378 if (arg->type == LDLM_BL_CALLBACK)
1379 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1380 ptlrpc_set_destroy(arg->set);
1383 arg->set = ptlrpc_prep_set();
1389 ldlm_work_bl_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1391 struct ldlm_lock_desc d;
1392 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_bl_ast);
1395 /* nobody should touch l_bl_ast */
1396 lock_res_and_lock(lock);
1397 list_del_init(&lock->l_bl_ast);
1399 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1400 LASSERT(lock->l_bl_ast_run == 0);
1401 LASSERT(lock->l_blocking_lock);
1402 lock->l_bl_ast_run++;
1403 unlock_res_and_lock(lock);
1405 ldlm_lock2desc(lock->l_blocking_lock, &d);
1407 LDLM_LOCK_PUT(lock->l_blocking_lock);
1408 lock->l_blocking_lock = NULL;
1409 lock->l_blocking_ast(lock, &d, (void *)arg,
1411 LDLM_LOCK_PUT(lock);
1417 ldlm_work_cp_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1419 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_cp_ast);
1420 ldlm_completion_callback completion_callback;
1424 /* It's possible to receive a completion AST before we've set
1425 * the l_completion_ast pointer: either because the AST arrived
1426 * before the reply, or simply because there's a small race
1427 * window between receiving the reply and finishing the local
1428 * enqueue. (bug 842)
1430 * This can't happen with the blocking_ast, however, because we
1431 * will never call the local blocking_ast until we drop our
1432 * reader/writer reference, which we won't do until we get the
1433 * reply and finish enqueueing. */
1435 /* nobody should touch l_cp_ast */
1436 lock_res_and_lock(lock);
1437 list_del_init(&lock->l_cp_ast);
1438 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1439 /* save l_completion_ast since it can be changed by
1440 * mds_intent_policy(), see bug 14225 */
1441 completion_callback = lock->l_completion_ast;
1442 lock->l_flags &= ~LDLM_FL_CP_REQD;
1443 unlock_res_and_lock(lock);
1445 if (completion_callback != NULL) {
1446 completion_callback(lock, 0, (void *)arg);
1449 LDLM_LOCK_PUT(lock);
1455 ldlm_work_revoke_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1457 struct ldlm_lock_desc desc;
1458 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_rk_ast);
1461 list_del_init(&lock->l_rk_ast);
1463 /* the desc just pretend to exclusive */
1464 ldlm_lock2desc(lock, &desc);
1465 desc.l_req_mode = LCK_EX;
1466 desc.l_granted_mode = 0;
1468 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1469 LDLM_LOCK_PUT(lock);
1474 int ldlm_run_ast_work(struct list_head *rpc_list, ldlm_desc_ast_t ast_type)
1476 struct ldlm_cb_set_arg arg;
1477 struct list_head *tmp, *pos;
1478 int (*work_ast_lock)(struct list_head *tmp, struct ldlm_cb_set_arg *arg);
1482 arg.set = ptlrpc_prep_set();
1483 atomic_set(&arg.restart, 0);
1485 case LDLM_WORK_BL_AST:
1486 arg.type = LDLM_BL_CALLBACK;
1487 work_ast_lock = ldlm_work_bl_ast_lock;
1489 case LDLM_WORK_CP_AST:
1490 arg.type = LDLM_CP_CALLBACK;
1491 work_ast_lock = ldlm_work_cp_ast_lock;
1493 case LDLM_WORK_REVOKE_AST:
1494 arg.type = LDLM_BL_CALLBACK;
1495 work_ast_lock = ldlm_work_revoke_ast_lock;
1502 list_for_each_safe(tmp, pos, rpc_list) {
1503 ast_count += work_ast_lock(tmp, &arg);
1505 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1506 * and create a new set for requests that remained in
1508 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1509 ldlm_send_and_maybe_create_set(&arg, 1);
1515 ldlm_send_and_maybe_create_set(&arg, 0);
1517 /* In case when number of ASTs is multiply of
1518 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1519 * @arg.set must be destroyed here, otherwise we get
1520 * write memory leaking. */
1521 ptlrpc_set_destroy(arg.set);
1523 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1526 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1528 ldlm_reprocess_all(res);
1529 return LDLM_ITER_CONTINUE;
1532 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1534 struct list_head *tmp;
1541 spin_lock(&ns->ns_hash_lock);
1542 for (i = 0; i < RES_HASH_SIZE; i++) {
1543 tmp = ns->ns_hash[i].next;
1544 while (tmp != &(ns->ns_hash[i])) {
1545 struct ldlm_resource *res =
1546 list_entry(tmp, struct ldlm_resource, lr_hash);
1548 ldlm_resource_getref(res);
1549 spin_unlock(&ns->ns_hash_lock);
1551 rc = reprocess_one_queue(res, NULL);
1553 spin_lock(&ns->ns_hash_lock);
1555 ldlm_resource_putref_locked(res);
1557 if (rc == LDLM_ITER_STOP)
1562 spin_unlock(&ns->ns_hash_lock);
1566 void ldlm_reprocess_all(struct ldlm_resource *res)
1568 CFS_LIST_HEAD(rpc_list);
1572 /* Local lock trees don't get reprocessed. */
1573 if (ns_is_client(res->lr_namespace)) {
1580 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1581 if (rc == LDLM_ITER_CONTINUE)
1582 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1585 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1586 if (rc == -ERESTART) {
1587 LASSERT(list_empty(&rpc_list));
1593 void ldlm_cancel_callback(struct ldlm_lock *lock)
1595 check_res_locked(lock->l_resource);
1596 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1597 lock->l_flags |= LDLM_FL_CANCEL;
1598 if (lock->l_blocking_ast) {
1599 // l_check_no_ns_lock(ns);
1600 unlock_res_and_lock(lock);
1601 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1603 lock_res_and_lock(lock);
1605 LDLM_DEBUG(lock, "no blocking ast");
1608 lock->l_flags |= LDLM_FL_BL_DONE;
1611 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1613 if (req->l_resource->lr_type != LDLM_PLAIN &&
1614 req->l_resource->lr_type != LDLM_IBITS)
1617 list_del_init(&req->l_sl_policy);
1618 list_del_init(&req->l_sl_mode);
1621 void ldlm_lock_cancel(struct ldlm_lock *lock)
1623 struct ldlm_resource *res;
1624 struct ldlm_namespace *ns;
1627 lock_res_and_lock(lock);
1629 res = lock->l_resource;
1630 ns = res->lr_namespace;
1632 /* Please do not, no matter how tempting, remove this LBUG without
1633 * talking to me first. -phik */
1634 if (lock->l_readers || lock->l_writers) {
1635 LDLM_ERROR(lock, "lock still has references");
1639 ldlm_del_waiting_lock(lock);
1641 /* Releases cancel callback. */
1642 ldlm_cancel_callback(lock);
1644 /* Yes, second time, just in case it was added again while we were
1645 running with no res lock in ldlm_cancel_callback */
1646 ldlm_del_waiting_lock(lock);
1647 ldlm_resource_unlink_lock(lock);
1648 ldlm_lock_destroy_nolock(lock);
1650 if (lock->l_granted_mode == lock->l_req_mode)
1651 ldlm_pool_del(&ns->ns_pool, lock);
1653 /* Make sure we will not be called again for same lock what is possible
1654 * if not to zero out lock->l_granted_mode */
1655 lock->l_granted_mode = LCK_MINMODE;
1656 unlock_res_and_lock(lock);
1661 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1663 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1669 lock->l_ast_data = data;
1670 LDLM_LOCK_PUT(lock);
1674 void ldlm_cancel_locks_for_export_cb(void *obj, void *data)
1676 struct obd_export *exp = data;
1677 struct ldlm_lock *lock = obj;
1678 struct ldlm_resource *res;
1680 res = ldlm_resource_getref(lock->l_resource);
1681 LDLM_LOCK_GET(lock);
1683 LDLM_DEBUG(lock, "export %p", exp);
1684 ldlm_res_lvbo_update(res, NULL, 0, 1);
1685 ldlm_lock_cancel(lock);
1686 ldlm_reprocess_all(res);
1687 ldlm_resource_putref(res);
1688 LDLM_LOCK_PUT(lock);
1691 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1693 lustre_hash_for_each_empty(exp->exp_lock_hash,
1694 ldlm_cancel_locks_for_export_cb, exp);
1697 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1700 CFS_LIST_HEAD(rpc_list);
1701 struct ldlm_resource *res;
1702 struct ldlm_namespace *ns;
1705 struct sl_insert_point prev;
1707 struct ldlm_interval *node;
1710 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1711 *flags |= LDLM_FL_BLOCK_GRANTED;
1712 RETURN(lock->l_resource);
1715 /* I can't check the type of lock here because the bitlock of lock
1716 * is not held here, so do the allocation blindly. -jay */
1717 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO, sizeof(*node));
1718 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1721 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1722 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1724 lock_res_and_lock(lock);
1726 res = lock->l_resource;
1727 ns = res->lr_namespace;
1729 old_mode = lock->l_req_mode;
1730 lock->l_req_mode = new_mode;
1731 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1732 /* remember the lock position where the lock might be
1733 * added back to the granted list later and also
1734 * remember the join mode for skiplist fixing. */
1735 prev.res_link = lock->l_res_link.prev;
1736 prev.mode_link = lock->l_sl_mode.prev;
1737 prev.policy_link = lock->l_sl_policy.prev;
1738 ldlm_resource_unlink_lock(lock);
1740 ldlm_resource_unlink_lock(lock);
1741 if (res->lr_type == LDLM_EXTENT) {
1742 /* FIXME: ugly code, I have to attach the lock to a
1743 * interval node again since perhaps it will be granted
1745 CFS_INIT_LIST_HEAD(&node->li_group);
1746 ldlm_interval_attach(node, lock);
1751 /* If this is a local resource, put it on the appropriate list. */
1752 if (ns_is_client(res->lr_namespace)) {
1753 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1754 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1756 /* This should never happen, because of the way the
1757 * server handles conversions. */
1758 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1762 ldlm_grant_lock(lock, &rpc_list);
1764 /* FIXME: completion handling not with ns_lock held ! */
1765 if (lock->l_completion_ast)
1766 lock->l_completion_ast(lock, 0, NULL);
1770 ldlm_processing_policy policy;
1771 policy = ldlm_processing_policy_table[res->lr_type];
1772 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1773 if (rc == LDLM_ITER_STOP) {
1774 lock->l_req_mode = old_mode;
1775 if (res->lr_type == LDLM_EXTENT)
1776 ldlm_extent_add_lock(res, lock);
1778 ldlm_granted_list_add_lock(lock, &prev);
1782 *flags |= LDLM_FL_BLOCK_GRANTED;
1786 unlock_res_and_lock(lock);
1789 ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1791 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1795 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1797 struct obd_device *obd = NULL;
1799 if (!((libcfs_debug | D_ERROR) & level))
1803 CDEBUG(level, " NULL LDLM lock\n");
1807 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1808 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1810 if (lock->l_conn_export != NULL)
1811 obd = lock->l_conn_export->exp_obd;
1812 if (lock->l_export && lock->l_export->exp_connection) {
1813 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1814 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1815 lock->l_remote_handle.cookie);
1816 } else if (obd == NULL) {
1817 CDEBUG(level, " Node: local\n");
1819 struct obd_import *imp = obd->u.cli.cl_import;
1820 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1821 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1822 lock->l_remote_handle.cookie);
1824 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1826 lock->l_resource->lr_name.name[0],
1827 lock->l_resource->lr_name.name[1],
1828 lock->l_resource->lr_name.name[2]);
1829 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1830 "write: %d flags: %#x\n", ldlm_lockname[lock->l_req_mode],
1831 ldlm_lockname[lock->l_granted_mode],
1832 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1834 if (lock->l_resource->lr_type == LDLM_EXTENT)
1835 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1836 " (req "LPU64"-"LPU64")\n",
1837 lock->l_policy_data.l_extent.start,
1838 lock->l_policy_data.l_extent.end,
1839 lock->l_req_extent.start, lock->l_req_extent.end);
1840 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1841 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1842 lock->l_policy_data.l_flock.pid,
1843 lock->l_policy_data.l_flock.start,
1844 lock->l_policy_data.l_flock.end);
1845 else if (lock->l_resource->lr_type == LDLM_IBITS)
1846 CDEBUG(level, " Bits: "LPX64"\n",
1847 lock->l_policy_data.l_inodebits.bits);
1850 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1852 struct ldlm_lock *lock;
1854 if (!((libcfs_debug | D_ERROR) & level))
1857 lock = ldlm_handle2lock(lockh);
1861 ldlm_lock_dump(D_OTHER, lock, 0);
1863 LDLM_LOCK_PUT(lock);
1866 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1867 struct libcfs_debug_msg_data *data, const char *fmt,
1871 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1873 va_start(args, fmt);
1875 if (lock->l_resource == NULL) {
1876 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1877 data->msg_fn, data->msg_line, fmt, args,
1878 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1879 "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "
1880 LPX64" expref: %d pid: %u\n", lock,
1881 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1882 lock->l_readers, lock->l_writers,
1883 ldlm_lockname[lock->l_granted_mode],
1884 ldlm_lockname[lock->l_req_mode],
1885 lock->l_flags, lock->l_remote_handle.cookie,
1887 atomic_read(&lock->l_export->exp_refcount) : -99,
1893 switch (lock->l_resource->lr_type) {
1895 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1896 data->msg_fn, data->msg_line, fmt, args,
1897 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1898 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1899 "] (req "LPU64"->"LPU64") flags: %x remote: "LPX64
1900 " expref: %d pid: %u\n",
1901 lock->l_resource->lr_namespace->ns_name, lock,
1902 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1903 lock->l_readers, lock->l_writers,
1904 ldlm_lockname[lock->l_granted_mode],
1905 ldlm_lockname[lock->l_req_mode],
1906 lock->l_resource->lr_name.name[0],
1907 lock->l_resource->lr_name.name[1],
1908 atomic_read(&lock->l_resource->lr_refcount),
1909 ldlm_typename[lock->l_resource->lr_type],
1910 lock->l_policy_data.l_extent.start,
1911 lock->l_policy_data.l_extent.end,
1912 lock->l_req_extent.start, lock->l_req_extent.end,
1913 lock->l_flags, lock->l_remote_handle.cookie,
1915 atomic_read(&lock->l_export->exp_refcount) : -99,
1920 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1921 data->msg_fn, data->msg_line, fmt, args,
1922 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1923 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1924 "["LPU64"->"LPU64"] flags: %x remote: "LPX64
1925 " expref: %d pid: %u\n",
1926 lock->l_resource->lr_namespace->ns_name, lock,
1927 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1928 lock->l_readers, lock->l_writers,
1929 ldlm_lockname[lock->l_granted_mode],
1930 ldlm_lockname[lock->l_req_mode],
1931 lock->l_resource->lr_name.name[0],
1932 lock->l_resource->lr_name.name[1],
1933 atomic_read(&lock->l_resource->lr_refcount),
1934 ldlm_typename[lock->l_resource->lr_type],
1935 lock->l_policy_data.l_flock.pid,
1936 lock->l_policy_data.l_flock.start,
1937 lock->l_policy_data.l_flock.end,
1938 lock->l_flags, lock->l_remote_handle.cookie,
1940 atomic_read(&lock->l_export->exp_refcount) : -99,
1945 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1946 data->msg_fn, data->msg_line, fmt, args,
1947 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1948 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1949 "flags: %x remote: "LPX64" expref: %d "
1951 lock->l_resource->lr_namespace->ns_name,
1952 lock, lock->l_handle.h_cookie,
1953 atomic_read (&lock->l_refc),
1954 lock->l_readers, lock->l_writers,
1955 ldlm_lockname[lock->l_granted_mode],
1956 ldlm_lockname[lock->l_req_mode],
1957 lock->l_resource->lr_name.name[0],
1958 lock->l_resource->lr_name.name[1],
1959 lock->l_policy_data.l_inodebits.bits,
1960 atomic_read(&lock->l_resource->lr_refcount),
1961 ldlm_typename[lock->l_resource->lr_type],
1962 lock->l_flags, lock->l_remote_handle.cookie,
1964 atomic_read(&lock->l_export->exp_refcount) : -99,
1969 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1970 data->msg_fn, data->msg_line, fmt, args,
1971 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1972 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: %x "
1973 "remote: "LPX64" expref: %d pid: %u\n",
1974 lock->l_resource->lr_namespace->ns_name,
1975 lock, lock->l_handle.h_cookie,
1976 atomic_read (&lock->l_refc),
1977 lock->l_readers, lock->l_writers,
1978 ldlm_lockname[lock->l_granted_mode],
1979 ldlm_lockname[lock->l_req_mode],
1980 lock->l_resource->lr_name.name[0],
1981 lock->l_resource->lr_name.name[1],
1982 atomic_read(&lock->l_resource->lr_refcount),
1983 ldlm_typename[lock->l_resource->lr_type],
1984 lock->l_flags, lock->l_remote_handle.cookie,
1986 atomic_read(&lock->l_export->exp_refcount) : -99,
1992 EXPORT_SYMBOL(_ldlm_lock_debug);