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 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
480 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
481 /* INODEBITS_INTEROP: If the other side does not support
482 * inodebits, reply with a plain lock descriptor.
484 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
485 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
486 struct ldlm_resource res = *lock->l_resource;
488 /* Make sure all the right bits are set in this lock we
489 are going to pass to client */
490 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
491 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
492 "Inappropriate inode lock bits during "
493 "conversion " LPU64 "\n",
494 lock->l_policy_data.l_inodebits.bits);
495 res.lr_type = LDLM_PLAIN;
496 ldlm_res2desc(&res, &desc->l_resource);
497 /* Convert "new" lock mode to something old client can
499 if ((lock->l_req_mode == LCK_CR) ||
500 (lock->l_req_mode == LCK_CW))
501 desc->l_req_mode = LCK_PR;
503 desc->l_req_mode = lock->l_req_mode;
504 if ((lock->l_granted_mode == LCK_CR) ||
505 (lock->l_granted_mode == LCK_CW)) {
506 desc->l_granted_mode = LCK_PR;
508 /* We never grant PW/EX locks to clients */
509 LASSERT((lock->l_granted_mode != LCK_PW) &&
510 (lock->l_granted_mode != LCK_EX));
511 desc->l_granted_mode = lock->l_granted_mode;
514 /* We do not copy policy here, because there is no
515 policy for plain locks */
517 ldlm_res2desc(lock->l_resource, &desc->l_resource);
518 desc->l_req_mode = lock->l_req_mode;
519 desc->l_granted_mode = lock->l_granted_mode;
520 desc->l_policy_data = lock->l_policy_data;
524 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
525 struct list_head *work_list)
527 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
528 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
529 lock->l_flags |= LDLM_FL_AST_SENT;
530 /* If the enqueuing client said so, tell the AST recipient to
531 * discard dirty data, rather than writing back. */
532 if (new->l_flags & LDLM_AST_DISCARD_DATA)
533 lock->l_flags |= LDLM_FL_DISCARD_DATA;
534 LASSERT(list_empty(&lock->l_bl_ast));
535 list_add(&lock->l_bl_ast, work_list);
537 LASSERT(lock->l_blocking_lock == NULL);
538 lock->l_blocking_lock = LDLM_LOCK_GET(new);
542 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
544 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
545 lock->l_flags |= LDLM_FL_CP_REQD;
546 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
547 LASSERT(list_empty(&lock->l_cp_ast));
548 list_add(&lock->l_cp_ast, work_list);
553 /* must be called with lr_lock held */
554 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
555 struct list_head *work_list)
558 check_res_locked(lock->l_resource);
560 ldlm_add_bl_work_item(lock, new, work_list);
562 ldlm_add_cp_work_item(lock, work_list);
566 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
568 struct ldlm_lock *lock;
570 lock = ldlm_handle2lock(lockh);
571 LASSERT(lock != NULL);
572 ldlm_lock_addref_internal(lock, mode);
576 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
578 ldlm_lock_remove_from_lru(lock);
579 if (mode & (LCK_NL | LCK_CR | LCK_PR))
581 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
584 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
587 /* only called for local locks */
588 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
590 lock_res_and_lock(lock);
591 ldlm_lock_addref_internal_nolock(lock, mode);
592 unlock_res_and_lock(lock);
595 /* only called in ldlm_flock_destroy and for local locks.
596 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
597 * * ldlm_lock_remove_from_lru() does nothing, it is safe
598 * * for ldlm_flock_destroy usage by dropping some code */
599 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
601 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
602 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
603 LASSERT(lock->l_readers > 0);
606 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
607 LASSERT(lock->l_writers > 0);
611 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
614 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
616 struct ldlm_namespace *ns;
619 lock_res_and_lock(lock);
621 ns = lock->l_resource->lr_namespace;
623 ldlm_lock_decref_internal_nolock(lock, mode);
625 if (lock->l_flags & LDLM_FL_LOCAL &&
626 !lock->l_readers && !lock->l_writers) {
627 /* If this is a local lock on a server namespace and this was
628 * the last reference, cancel the lock. */
629 CDEBUG(D_INFO, "forcing cancel of local lock\n");
630 lock->l_flags |= LDLM_FL_CBPENDING;
633 if (!lock->l_readers && !lock->l_writers &&
634 (lock->l_flags & LDLM_FL_CBPENDING)) {
635 /* If we received a blocked AST and this was the last reference,
636 * run the callback. */
637 if (ns_is_server(ns) && lock->l_export)
638 CERROR("FL_CBPENDING set on non-local lock--just a "
641 LDLM_DEBUG(lock, "final decref done on cbpending lock");
643 LDLM_LOCK_GET(lock); /* dropped by bl thread */
644 ldlm_lock_remove_from_lru(lock);
645 unlock_res_and_lock(lock);
646 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
647 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
648 ldlm_handle_bl_callback(ns, NULL, lock);
649 } else if (ns_is_client(ns) &&
650 !lock->l_readers && !lock->l_writers &&
651 !(lock->l_flags & LDLM_FL_BL_AST)) {
652 /* If this is a client-side namespace and this was the last
653 * reference, put it on the LRU. */
654 ldlm_lock_add_to_lru(lock);
655 unlock_res_and_lock(lock);
656 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
657 * are not supported by the server, otherwise, it is done on
659 if (!exp_connect_cancelset(lock->l_conn_export) &&
660 !ns_connect_lru_resize(ns))
661 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
663 unlock_res_and_lock(lock);
669 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
671 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
672 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
673 ldlm_lock_decref_internal(lock, mode);
677 /* This will drop a lock reference and mark it for destruction, but will not
678 * necessarily cancel the lock before returning. */
679 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
681 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
684 LASSERT(lock != NULL);
686 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
687 lock_res_and_lock(lock);
688 lock->l_flags |= LDLM_FL_CBPENDING;
689 unlock_res_and_lock(lock);
690 ldlm_lock_decref_internal(lock, mode);
694 struct sl_insert_point {
695 struct list_head *res_link;
696 struct list_head *mode_link;
697 struct list_head *policy_link;
701 * search_granted_lock
704 * Finds a position to insert the new lock.
706 * queue [input]: the granted list where search acts on;
707 * req [input]: the lock whose position to be located;
708 * prev [output]: positions within 3 lists to insert @req to
712 * - ldlm_grant_lock_with_skiplist
714 static void search_granted_lock(struct list_head *queue,
715 struct ldlm_lock *req,
716 struct sl_insert_point *prev)
718 struct list_head *tmp;
719 struct ldlm_lock *lock, *mode_end, *policy_end;
722 list_for_each(tmp, queue) {
723 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
725 mode_end = list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
728 if (lock->l_req_mode != req->l_req_mode) {
729 /* jump to last lock of mode group */
730 tmp = &mode_end->l_res_link;
734 /* suitable mode group is found */
735 if (lock->l_resource->lr_type == LDLM_PLAIN) {
736 /* insert point is last lock of the mode group */
737 prev->res_link = &mode_end->l_res_link;
738 prev->mode_link = &mode_end->l_sl_mode;
739 prev->policy_link = &req->l_sl_policy;
742 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
744 policy_end = list_entry(lock->l_sl_policy.prev,
748 if (lock->l_policy_data.l_inodebits.bits ==
749 req->l_policy_data.l_inodebits.bits) {
750 /* insert point is last lock of
751 * the policy group */
753 &policy_end->l_res_link;
755 &policy_end->l_sl_mode;
757 &policy_end->l_sl_policy;
762 if (policy_end == mode_end)
763 /* done with mode group */
766 /* jump to next policy group within the mode group */
767 tmp = policy_end->l_res_link.next;
768 lock = list_entry(tmp, struct ldlm_lock,
770 } /* loop over policy groups within the mode group */
772 /* insert point is last lock of the mode group,
773 * new policy group is started */
774 prev->res_link = &mode_end->l_res_link;
775 prev->mode_link = &mode_end->l_sl_mode;
776 prev->policy_link = &req->l_sl_policy;
780 LDLM_ERROR(lock, "is not LDLM_PLAIN or LDLM_IBITS lock");
785 /* insert point is last lock on the queue,
786 * new mode group and new policy group are started */
787 prev->res_link = queue->prev;
788 prev->mode_link = &req->l_sl_mode;
789 prev->policy_link = &req->l_sl_policy;
794 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
795 struct sl_insert_point *prev)
797 struct ldlm_resource *res = lock->l_resource;
800 check_res_locked(res);
802 ldlm_resource_dump(D_INFO, res);
803 CDEBUG(D_OTHER, "About to add this lock:\n");
804 ldlm_lock_dump(D_OTHER, lock, 0);
806 if (lock->l_destroyed) {
807 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
811 LASSERT(list_empty(&lock->l_res_link));
812 LASSERT(list_empty(&lock->l_sl_mode));
813 LASSERT(list_empty(&lock->l_sl_policy));
815 list_add(&lock->l_res_link, prev->res_link);
816 list_add(&lock->l_sl_mode, prev->mode_link);
817 list_add(&lock->l_sl_policy, prev->policy_link);
822 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
824 struct sl_insert_point prev;
827 LASSERT(lock->l_req_mode == lock->l_granted_mode);
829 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
830 ldlm_granted_list_add_lock(lock, &prev);
835 * - ldlm_lock_enqueue
836 * - ldlm_reprocess_queue
837 * - ldlm_lock_convert
839 * must be called with lr_lock held
841 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
843 struct ldlm_resource *res = lock->l_resource;
846 check_res_locked(res);
848 lock->l_granted_mode = lock->l_req_mode;
849 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
850 ldlm_grant_lock_with_skiplist(lock);
851 else if (res->lr_type == LDLM_EXTENT)
852 ldlm_extent_add_lock(res, lock);
854 ldlm_resource_add_lock(res, &res->lr_granted, lock);
856 if (lock->l_granted_mode < res->lr_most_restr)
857 res->lr_most_restr = lock->l_granted_mode;
859 if (work_list && lock->l_completion_ast != NULL)
860 ldlm_add_ast_work_item(lock, NULL, work_list);
862 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
866 /* returns a referenced lock or NULL. See the flag descriptions below, in the
867 * comment above ldlm_lock_match */
868 static struct ldlm_lock *search_queue(struct list_head *queue,
870 ldlm_policy_data_t *policy,
871 struct ldlm_lock *old_lock, int flags)
873 struct ldlm_lock *lock;
874 struct list_head *tmp;
876 list_for_each(tmp, queue) {
879 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
881 if (lock == old_lock)
884 /* llite sometimes wants to match locks that will be
885 * canceled when their users drop, but we allow it to match
886 * if it passes in CBPENDING and the lock still has users.
887 * this is generally only going to be used by children
888 * whose parents already hold a lock so forward progress
889 * can still happen. */
890 if (lock->l_flags & LDLM_FL_CBPENDING &&
891 !(flags & LDLM_FL_CBPENDING))
893 if (lock->l_flags & LDLM_FL_CBPENDING &&
894 lock->l_readers == 0 && lock->l_writers == 0)
897 if (!(lock->l_req_mode & *mode))
899 match = lock->l_req_mode;
901 if (lock->l_resource->lr_type == LDLM_EXTENT &&
902 (lock->l_policy_data.l_extent.start >
903 policy->l_extent.start ||
904 lock->l_policy_data.l_extent.end < policy->l_extent.end))
907 if (unlikely(match == LCK_GROUP) &&
908 lock->l_resource->lr_type == LDLM_EXTENT &&
909 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
912 /* We match if we have existing lock with same or wider set
914 if (lock->l_resource->lr_type == LDLM_IBITS &&
915 ((lock->l_policy_data.l_inodebits.bits &
916 policy->l_inodebits.bits) !=
917 policy->l_inodebits.bits))
920 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
923 if ((flags & LDLM_FL_LOCAL_ONLY) &&
924 !(lock->l_flags & LDLM_FL_LOCAL))
927 if (flags & LDLM_FL_TEST_LOCK) {
929 ldlm_lock_touch_in_lru(lock);
931 ldlm_lock_addref_internal_nolock(lock, match);
940 void ldlm_lock_allow_match(struct ldlm_lock *lock)
942 lock_res_and_lock(lock);
943 lock->l_flags |= LDLM_FL_LVB_READY;
944 cfs_waitq_signal(&lock->l_waitq);
945 unlock_res_and_lock(lock);
949 * Checks if requested extent lock is compatible with another owned lock.
951 * Checks if \a lock is compatible with a read or write lock
952 * (specified by \a rw) for an extent [\a start , \a end].
954 * \param lock the already owned lock
955 * \param rw OBD_BRW_READ if requested for reading,
956 * OBD_BRW_WRITE if requested for writing
957 * \param start start of the requested extent
958 * \param end end of the requested extent
959 * \param cookie transparent parameter for passing locking context
961 * \post result == 1, *cookie == context, appropriate lock is referenced
963 * \retval 1 owned lock is reused for the request
964 * \retval 0 no lock reused for the request
966 * \see ldlm_lock_fast_release
968 int ldlm_lock_fast_match(struct ldlm_lock *lock, int rw,
969 obd_off start, obd_off end,
972 LASSERT(rw == OBD_BRW_READ || rw == OBD_BRW_WRITE);
977 lock_res_and_lock(lock);
978 /* check if granted mode is compatible */
979 if (rw == OBD_BRW_WRITE &&
980 !(lock->l_granted_mode & (LCK_PW|LCK_GROUP)))
983 /* does the lock cover the region we would like to access? */
984 if ((lock->l_policy_data.l_extent.start > start) ||
985 (lock->l_policy_data.l_extent.end < end))
988 /* if we received a blocking callback and the lock is no longer
989 * referenced, don't use it */
990 if ((lock->l_flags & LDLM_FL_CBPENDING) &&
991 !lock->l_writers && !lock->l_readers)
994 ldlm_lock_addref_internal_nolock(lock, rw == OBD_BRW_WRITE ? LCK_PW : LCK_PR);
995 unlock_res_and_lock(lock);
996 *cookie = (void *)lock;
997 return 1; /* avoid using rc for stack relief */
1000 unlock_res_and_lock(lock);
1005 * Releases a reference to a lock taken in a "fast" way.
1007 * Releases a read or write (specified by \a rw) lock
1008 * referenced by \a cookie.
1010 * \param rw OBD_BRW_READ if requested for reading,
1011 * OBD_BRW_WRITE if requested for writing
1012 * \param cookie transparent parameter for passing locking context
1014 * \post appropriate lock is dereferenced
1016 * \see ldlm_lock_fast_lock
1018 void ldlm_lock_fast_release(void *cookie, int rw)
1020 struct ldlm_lock *lock = (struct ldlm_lock *)cookie;
1022 LASSERT(lock != NULL);
1023 LASSERT(rw == OBD_BRW_READ || rw == OBD_BRW_WRITE);
1024 LASSERT(rw == OBD_BRW_READ ||
1025 (lock->l_granted_mode & (LCK_PW | LCK_GROUP)));
1026 ldlm_lock_decref_internal(lock, rw == OBD_BRW_WRITE ? LCK_PW : LCK_PR);
1029 /* Can be called in two ways:
1031 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1032 * for a duplicate of.
1034 * Otherwise, all of the fields must be filled in, to match against.
1036 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1037 * server (ie, connh is NULL)
1038 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1039 * list will be considered
1040 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1041 * to be canceled can still be matched as long as they still have reader
1042 * or writer refernces
1043 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1044 * just tell us if we would have matched.
1046 * Returns 1 if it finds an already-existing lock that is compatible; in this
1047 * case, lockh is filled in with a addref()ed lock
1049 * we also check security context, if that failed we simply return 0 (to keep
1050 * caller code unchanged), the context failure will be discovered by caller
1053 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1054 const struct ldlm_res_id *res_id, ldlm_type_t type,
1055 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1056 struct lustre_handle *lockh)
1058 struct ldlm_resource *res;
1059 struct ldlm_lock *lock, *old_lock = NULL;
1064 old_lock = ldlm_handle2lock(lockh);
1067 ns = old_lock->l_resource->lr_namespace;
1068 res_id = &old_lock->l_resource->lr_name;
1069 type = old_lock->l_resource->lr_type;
1070 mode = old_lock->l_req_mode;
1073 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1075 LASSERT(old_lock == NULL);
1081 lock = search_queue(&res->lr_granted, &mode, policy, old_lock, flags);
1084 if (flags & LDLM_FL_BLOCK_GRANTED)
1086 lock = search_queue(&res->lr_converting, &mode, policy, old_lock, flags);
1089 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock, flags);
1096 ldlm_resource_putref(res);
1099 ldlm_lock2handle(lock, lockh);
1100 if ((flags & LDLM_FL_LVB_READY) &&
1101 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1102 struct l_wait_info lwi;
1103 if (lock->l_completion_ast) {
1104 int err = lock->l_completion_ast(lock,
1105 LDLM_FL_WAIT_NOREPROC,
1108 if (flags & LDLM_FL_TEST_LOCK)
1109 LDLM_LOCK_PUT(lock);
1111 ldlm_lock_decref_internal(lock, mode);
1117 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), NULL,
1118 LWI_ON_SIGNAL_NOOP, NULL);
1120 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1121 l_wait_event(lock->l_waitq,
1122 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1127 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1128 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1129 res_id->name[2] : policy->l_extent.start,
1130 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1131 res_id->name[3] : policy->l_extent.end);
1133 /* check user's security context */
1134 if (lock->l_conn_export &&
1135 sptlrpc_import_check_ctx(
1136 class_exp2cliimp(lock->l_conn_export))) {
1137 if (!(flags & LDLM_FL_TEST_LOCK))
1138 ldlm_lock_decref_internal(lock, mode);
1142 if (flags & LDLM_FL_TEST_LOCK)
1143 LDLM_LOCK_PUT(lock);
1145 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1146 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1147 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1148 type, mode, res_id->name[0], res_id->name[1],
1149 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1150 res_id->name[2] :policy->l_extent.start,
1151 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1152 res_id->name[3] : policy->l_extent.end);
1155 LDLM_LOCK_PUT(old_lock);
1157 return rc ? mode : 0;
1160 /* Returns a referenced lock */
1161 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1162 const struct ldlm_res_id *res_id,
1165 const struct ldlm_callback_suite *cbs,
1166 void *data, __u32 lvb_len)
1168 struct ldlm_lock *lock;
1169 struct ldlm_resource *res;
1172 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1176 lock = ldlm_lock_new(res);
1177 ldlm_resource_putref(res);
1182 lock->l_req_mode = mode;
1183 lock->l_ast_data = data;
1184 lock->l_pid = cfs_curproc_pid();
1186 lock->l_blocking_ast = cbs->lcs_blocking;
1187 lock->l_completion_ast = cbs->lcs_completion;
1188 lock->l_glimpse_ast = cbs->lcs_glimpse;
1189 lock->l_weigh_ast = cbs->lcs_weigh;
1192 lock->l_tree_node = NULL;
1193 /* if this is the extent lock, allocate the interval tree node */
1194 if (type == LDLM_EXTENT) {
1195 if (ldlm_interval_alloc(lock) == NULL)
1200 lock->l_lvb_len = lvb_len;
1201 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1202 if (lock->l_lvb_data == NULL)
1209 if (lock->l_lvb_data)
1210 OBD_FREE(lock->l_lvb_data, lvb_len);
1211 ldlm_interval_free(ldlm_interval_detach(lock));
1212 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1216 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1217 struct ldlm_lock **lockp,
1218 void *cookie, int *flags)
1220 struct ldlm_lock *lock = *lockp;
1221 struct ldlm_resource *res = lock->l_resource;
1222 int local = ns_is_client(res->lr_namespace);
1223 ldlm_processing_policy policy;
1224 ldlm_error_t rc = ELDLM_OK;
1225 struct ldlm_interval *node = NULL;
1228 do_gettimeofday(&lock->l_enqueued_time);
1229 /* policies are not executed on the client or during replay */
1230 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1231 && !local && ns->ns_policy) {
1232 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1234 if (rc == ELDLM_LOCK_REPLACED) {
1235 /* The lock that was returned has already been granted,
1236 * and placed into lockp. If it's not the same as the
1237 * one we passed in, then destroy the old one and our
1238 * work here is done. */
1239 if (lock != *lockp) {
1240 ldlm_lock_destroy(lock);
1241 LDLM_LOCK_PUT(lock);
1243 *flags |= LDLM_FL_LOCK_CHANGED;
1245 } else if (rc != ELDLM_OK ||
1246 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1247 ldlm_lock_destroy(lock);
1252 /* For a replaying lock, it might be already in granted list. So
1253 * unlinking the lock will cause the interval node to be freed, we
1254 * have to allocate the interval node early otherwise we can't regrant
1255 * this lock in the future. - jay */
1256 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1257 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO,
1260 lock_res_and_lock(lock);
1261 if (local && lock->l_req_mode == lock->l_granted_mode) {
1262 /* The server returned a blocked lock, but it was granted
1263 * before we got a chance to actually enqueue it. We don't
1264 * need to do anything else. */
1265 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1266 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1267 GOTO(out, ELDLM_OK);
1270 ldlm_resource_unlink_lock(lock);
1271 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1273 ldlm_lock_destroy_nolock(lock);
1274 GOTO(out, rc = -ENOMEM);
1277 CFS_INIT_LIST_HEAD(&node->li_group);
1278 ldlm_interval_attach(node, lock);
1282 /* Some flags from the enqueue want to make it into the AST, via the
1283 * lock's l_flags. */
1284 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1286 /* This distinction between local lock trees is very important; a client
1287 * namespace only has information about locks taken by that client, and
1288 * thus doesn't have enough information to decide for itself if it can
1289 * be granted (below). In this case, we do exactly what the server
1290 * tells us to do, as dictated by the 'flags'.
1292 * We do exactly the same thing during recovery, when the server is
1293 * more or less trusting the clients not to lie.
1295 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1296 * granted/converting queues. */
1298 if (*flags & LDLM_FL_BLOCK_CONV)
1299 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1300 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1301 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1303 ldlm_grant_lock(lock, NULL);
1304 GOTO(out, ELDLM_OK);
1305 } else if (*flags & LDLM_FL_REPLAY) {
1306 if (*flags & LDLM_FL_BLOCK_CONV) {
1307 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1308 GOTO(out, ELDLM_OK);
1309 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1310 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1311 GOTO(out, ELDLM_OK);
1312 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1313 ldlm_grant_lock(lock, NULL);
1314 GOTO(out, ELDLM_OK);
1316 /* If no flags, fall through to normal enqueue path. */
1319 policy = ldlm_processing_policy_table[res->lr_type];
1320 policy(lock, flags, 1, &rc, NULL);
1323 unlock_res_and_lock(lock);
1325 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1329 /* Must be called with namespace taken: queue is waiting or converting. */
1330 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1331 struct list_head *work_list)
1333 struct list_head *tmp, *pos;
1334 ldlm_processing_policy policy;
1336 int rc = LDLM_ITER_CONTINUE;
1340 check_res_locked(res);
1342 policy = ldlm_processing_policy_table[res->lr_type];
1345 list_for_each_safe(tmp, pos, queue) {
1346 struct ldlm_lock *pending;
1347 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1349 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1352 rc = policy(pending, &flags, 0, &err, work_list);
1353 if (rc != LDLM_ITER_CONTINUE)
1360 /* Helper function for ldlm_run_ast_work().
1362 * Send an existing rpc set specified by @arg->set and then
1363 * destroy it. Create new one if @do_create flag is set. */
1365 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1369 ptlrpc_set_wait(arg->set);
1370 if (arg->type == LDLM_BL_CALLBACK)
1371 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1372 ptlrpc_set_destroy(arg->set);
1375 arg->set = ptlrpc_prep_set();
1381 ldlm_work_bl_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1383 struct ldlm_lock_desc d;
1384 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_bl_ast);
1387 /* nobody should touch l_bl_ast */
1388 lock_res_and_lock(lock);
1389 list_del_init(&lock->l_bl_ast);
1391 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1392 LASSERT(lock->l_bl_ast_run == 0);
1393 LASSERT(lock->l_blocking_lock);
1394 lock->l_bl_ast_run++;
1395 unlock_res_and_lock(lock);
1397 ldlm_lock2desc(lock->l_blocking_lock, &d);
1399 LDLM_LOCK_PUT(lock->l_blocking_lock);
1400 lock->l_blocking_lock = NULL;
1401 lock->l_blocking_ast(lock, &d, (void *)arg,
1403 LDLM_LOCK_PUT(lock);
1409 ldlm_work_cp_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1411 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_cp_ast);
1412 ldlm_completion_callback completion_callback;
1416 /* It's possible to receive a completion AST before we've set
1417 * the l_completion_ast pointer: either because the AST arrived
1418 * before the reply, or simply because there's a small race
1419 * window between receiving the reply and finishing the local
1420 * enqueue. (bug 842)
1422 * This can't happen with the blocking_ast, however, because we
1423 * will never call the local blocking_ast until we drop our
1424 * reader/writer reference, which we won't do until we get the
1425 * reply and finish enqueueing. */
1427 /* nobody should touch l_cp_ast */
1428 lock_res_and_lock(lock);
1429 list_del_init(&lock->l_cp_ast);
1430 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1431 /* save l_completion_ast since it can be changed by
1432 * mds_intent_policy(), see bug 14225 */
1433 completion_callback = lock->l_completion_ast;
1434 lock->l_flags &= ~LDLM_FL_CP_REQD;
1435 unlock_res_and_lock(lock);
1437 if (completion_callback != NULL) {
1438 completion_callback(lock, 0, (void *)arg);
1441 LDLM_LOCK_PUT(lock);
1447 ldlm_work_revoke_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1449 struct ldlm_lock_desc desc;
1450 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_rk_ast);
1453 list_del_init(&lock->l_rk_ast);
1455 /* the desc just pretend to exclusive */
1456 ldlm_lock2desc(lock, &desc);
1457 desc.l_req_mode = LCK_EX;
1458 desc.l_granted_mode = 0;
1460 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1461 LDLM_LOCK_PUT(lock);
1466 int ldlm_run_ast_work(struct list_head *rpc_list, ldlm_desc_ast_t ast_type)
1468 struct ldlm_cb_set_arg arg;
1469 struct list_head *tmp, *pos;
1470 int (*work_ast_lock)(struct list_head *tmp, struct ldlm_cb_set_arg *arg);
1474 arg.set = ptlrpc_prep_set();
1475 atomic_set(&arg.restart, 0);
1477 case LDLM_WORK_BL_AST:
1478 arg.type = LDLM_BL_CALLBACK;
1479 work_ast_lock = ldlm_work_bl_ast_lock;
1481 case LDLM_WORK_CP_AST:
1482 arg.type = LDLM_CP_CALLBACK;
1483 work_ast_lock = ldlm_work_cp_ast_lock;
1485 case LDLM_WORK_REVOKE_AST:
1486 arg.type = LDLM_BL_CALLBACK;
1487 work_ast_lock = ldlm_work_revoke_ast_lock;
1494 list_for_each_safe(tmp, pos, rpc_list) {
1495 ast_count += work_ast_lock(tmp, &arg);
1497 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1498 * and create a new set for requests that remained in
1500 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1501 ldlm_send_and_maybe_create_set(&arg, 1);
1507 ldlm_send_and_maybe_create_set(&arg, 0);
1509 /* In case when number of ASTs is multiply of
1510 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1511 * @arg.set must be destroyed here, otherwise we get
1512 * write memory leaking. */
1513 ptlrpc_set_destroy(arg.set);
1515 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1518 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1520 ldlm_reprocess_all(res);
1521 return LDLM_ITER_CONTINUE;
1524 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1526 struct list_head *tmp;
1533 spin_lock(&ns->ns_hash_lock);
1534 for (i = 0; i < RES_HASH_SIZE; i++) {
1535 tmp = ns->ns_hash[i].next;
1536 while (tmp != &(ns->ns_hash[i])) {
1537 struct ldlm_resource *res =
1538 list_entry(tmp, struct ldlm_resource, lr_hash);
1540 ldlm_resource_getref(res);
1541 spin_unlock(&ns->ns_hash_lock);
1543 rc = reprocess_one_queue(res, NULL);
1545 spin_lock(&ns->ns_hash_lock);
1547 ldlm_resource_putref_locked(res);
1549 if (rc == LDLM_ITER_STOP)
1554 spin_unlock(&ns->ns_hash_lock);
1558 void ldlm_reprocess_all(struct ldlm_resource *res)
1560 CFS_LIST_HEAD(rpc_list);
1564 /* Local lock trees don't get reprocessed. */
1565 if (ns_is_client(res->lr_namespace)) {
1572 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1573 if (rc == LDLM_ITER_CONTINUE)
1574 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1577 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1578 if (rc == -ERESTART) {
1579 LASSERT(list_empty(&rpc_list));
1585 void ldlm_cancel_callback(struct ldlm_lock *lock)
1587 check_res_locked(lock->l_resource);
1588 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1589 lock->l_flags |= LDLM_FL_CANCEL;
1590 if (lock->l_blocking_ast) {
1591 // l_check_no_ns_lock(ns);
1592 unlock_res_and_lock(lock);
1593 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1595 lock_res_and_lock(lock);
1597 LDLM_DEBUG(lock, "no blocking ast");
1600 lock->l_flags |= LDLM_FL_BL_DONE;
1603 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1605 if (req->l_resource->lr_type != LDLM_PLAIN &&
1606 req->l_resource->lr_type != LDLM_IBITS)
1609 list_del_init(&req->l_sl_policy);
1610 list_del_init(&req->l_sl_mode);
1613 void ldlm_lock_cancel(struct ldlm_lock *lock)
1615 struct ldlm_resource *res;
1616 struct ldlm_namespace *ns;
1619 lock_res_and_lock(lock);
1621 res = lock->l_resource;
1622 ns = res->lr_namespace;
1624 /* Please do not, no matter how tempting, remove this LBUG without
1625 * talking to me first. -phik */
1626 if (lock->l_readers || lock->l_writers) {
1627 LDLM_ERROR(lock, "lock still has references");
1631 ldlm_del_waiting_lock(lock);
1633 /* Releases cancel callback. */
1634 ldlm_cancel_callback(lock);
1636 /* Yes, second time, just in case it was added again while we were
1637 running with no res lock in ldlm_cancel_callback */
1638 ldlm_del_waiting_lock(lock);
1639 ldlm_resource_unlink_lock(lock);
1640 ldlm_lock_destroy_nolock(lock);
1642 if (lock->l_granted_mode == lock->l_req_mode)
1643 ldlm_pool_del(&ns->ns_pool, lock);
1645 /* Make sure we will not be called again for same lock what is possible
1646 * if not to zero out lock->l_granted_mode */
1647 lock->l_granted_mode = LCK_MINMODE;
1648 unlock_res_and_lock(lock);
1653 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1655 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1661 lock->l_ast_data = data;
1662 LDLM_LOCK_PUT(lock);
1666 void ldlm_cancel_locks_for_export_cb(void *obj, void *data)
1668 struct obd_export *exp = data;
1669 struct ldlm_lock *lock = obj;
1670 struct ldlm_resource *res;
1672 res = ldlm_resource_getref(lock->l_resource);
1673 LDLM_LOCK_GET(lock);
1675 LDLM_DEBUG(lock, "export %p", exp);
1676 ldlm_res_lvbo_update(res, NULL, 0, 1);
1677 ldlm_lock_cancel(lock);
1678 ldlm_reprocess_all(res);
1679 ldlm_resource_putref(res);
1680 LDLM_LOCK_PUT(lock);
1683 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1685 lustre_hash_for_each_empty(exp->exp_lock_hash,
1686 ldlm_cancel_locks_for_export_cb, exp);
1689 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1692 CFS_LIST_HEAD(rpc_list);
1693 struct ldlm_resource *res;
1694 struct ldlm_namespace *ns;
1697 struct sl_insert_point prev;
1699 struct ldlm_interval *node;
1702 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1703 *flags |= LDLM_FL_BLOCK_GRANTED;
1704 RETURN(lock->l_resource);
1707 /* I can't check the type of lock here because the bitlock of lock
1708 * is not held here, so do the allocation blindly. -jay */
1709 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO, sizeof(*node));
1710 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1713 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1714 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1716 lock_res_and_lock(lock);
1718 res = lock->l_resource;
1719 ns = res->lr_namespace;
1721 old_mode = lock->l_req_mode;
1722 lock->l_req_mode = new_mode;
1723 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1724 /* remember the lock position where the lock might be
1725 * added back to the granted list later and also
1726 * remember the join mode for skiplist fixing. */
1727 prev.res_link = lock->l_res_link.prev;
1728 prev.mode_link = lock->l_sl_mode.prev;
1729 prev.policy_link = lock->l_sl_policy.prev;
1730 ldlm_resource_unlink_lock(lock);
1732 ldlm_resource_unlink_lock(lock);
1733 if (res->lr_type == LDLM_EXTENT) {
1734 /* FIXME: ugly code, I have to attach the lock to a
1735 * interval node again since perhaps it will be granted
1737 CFS_INIT_LIST_HEAD(&node->li_group);
1738 ldlm_interval_attach(node, lock);
1743 /* If this is a local resource, put it on the appropriate list. */
1744 if (ns_is_client(res->lr_namespace)) {
1745 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1746 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1748 /* This should never happen, because of the way the
1749 * server handles conversions. */
1750 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1754 ldlm_grant_lock(lock, &rpc_list);
1756 /* FIXME: completion handling not with ns_lock held ! */
1757 if (lock->l_completion_ast)
1758 lock->l_completion_ast(lock, 0, NULL);
1762 ldlm_processing_policy policy;
1763 policy = ldlm_processing_policy_table[res->lr_type];
1764 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1765 if (rc == LDLM_ITER_STOP) {
1766 lock->l_req_mode = old_mode;
1767 if (res->lr_type == LDLM_EXTENT)
1768 ldlm_extent_add_lock(res, lock);
1770 ldlm_granted_list_add_lock(lock, &prev);
1774 *flags |= LDLM_FL_BLOCK_GRANTED;
1778 unlock_res_and_lock(lock);
1781 ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1783 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1787 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1789 struct obd_device *obd = NULL;
1791 if (!((libcfs_debug | D_ERROR) & level))
1795 CDEBUG(level, " NULL LDLM lock\n");
1799 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1800 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1802 if (lock->l_conn_export != NULL)
1803 obd = lock->l_conn_export->exp_obd;
1804 if (lock->l_export && lock->l_export->exp_connection) {
1805 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1806 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1807 lock->l_remote_handle.cookie);
1808 } else if (obd == NULL) {
1809 CDEBUG(level, " Node: local\n");
1811 struct obd_import *imp = obd->u.cli.cl_import;
1812 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1813 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1814 lock->l_remote_handle.cookie);
1816 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1818 lock->l_resource->lr_name.name[0],
1819 lock->l_resource->lr_name.name[1],
1820 lock->l_resource->lr_name.name[2]);
1821 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1822 "write: %d flags: %#x\n", ldlm_lockname[lock->l_req_mode],
1823 ldlm_lockname[lock->l_granted_mode],
1824 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1826 if (lock->l_resource->lr_type == LDLM_EXTENT)
1827 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1828 " (req "LPU64"-"LPU64")\n",
1829 lock->l_policy_data.l_extent.start,
1830 lock->l_policy_data.l_extent.end,
1831 lock->l_req_extent.start, lock->l_req_extent.end);
1832 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1833 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1834 lock->l_policy_data.l_flock.pid,
1835 lock->l_policy_data.l_flock.start,
1836 lock->l_policy_data.l_flock.end);
1837 else if (lock->l_resource->lr_type == LDLM_IBITS)
1838 CDEBUG(level, " Bits: "LPX64"\n",
1839 lock->l_policy_data.l_inodebits.bits);
1842 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1844 struct ldlm_lock *lock;
1846 if (!((libcfs_debug | D_ERROR) & level))
1849 lock = ldlm_handle2lock(lockh);
1853 ldlm_lock_dump(D_OTHER, lock, 0);
1855 LDLM_LOCK_PUT(lock);
1858 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1859 struct libcfs_debug_msg_data *data, const char *fmt,
1863 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1865 va_start(args, fmt);
1867 if (lock->l_resource == NULL) {
1868 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1869 data->msg_fn, data->msg_line, fmt, args,
1870 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1871 "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "
1872 LPX64" expref: %d pid: %u\n", lock,
1873 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1874 lock->l_readers, lock->l_writers,
1875 ldlm_lockname[lock->l_granted_mode],
1876 ldlm_lockname[lock->l_req_mode],
1877 lock->l_flags, lock->l_remote_handle.cookie,
1879 atomic_read(&lock->l_export->exp_refcount) : -99,
1885 switch (lock->l_resource->lr_type) {
1887 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1888 data->msg_fn, data->msg_line, fmt, args,
1889 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1890 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1891 "] (req "LPU64"->"LPU64") flags: %x remote: "LPX64
1892 " expref: %d pid: %u\n",
1893 lock->l_resource->lr_namespace->ns_name, lock,
1894 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1895 lock->l_readers, lock->l_writers,
1896 ldlm_lockname[lock->l_granted_mode],
1897 ldlm_lockname[lock->l_req_mode],
1898 lock->l_resource->lr_name.name[0],
1899 lock->l_resource->lr_name.name[1],
1900 atomic_read(&lock->l_resource->lr_refcount),
1901 ldlm_typename[lock->l_resource->lr_type],
1902 lock->l_policy_data.l_extent.start,
1903 lock->l_policy_data.l_extent.end,
1904 lock->l_req_extent.start, lock->l_req_extent.end,
1905 lock->l_flags, lock->l_remote_handle.cookie,
1907 atomic_read(&lock->l_export->exp_refcount) : -99,
1912 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1913 data->msg_fn, data->msg_line, fmt, args,
1914 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1915 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1916 "["LPU64"->"LPU64"] flags: %x remote: "LPX64
1917 " expref: %d pid: %u\n",
1918 lock->l_resource->lr_namespace->ns_name, lock,
1919 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1920 lock->l_readers, lock->l_writers,
1921 ldlm_lockname[lock->l_granted_mode],
1922 ldlm_lockname[lock->l_req_mode],
1923 lock->l_resource->lr_name.name[0],
1924 lock->l_resource->lr_name.name[1],
1925 atomic_read(&lock->l_resource->lr_refcount),
1926 ldlm_typename[lock->l_resource->lr_type],
1927 lock->l_policy_data.l_flock.pid,
1928 lock->l_policy_data.l_flock.start,
1929 lock->l_policy_data.l_flock.end,
1930 lock->l_flags, lock->l_remote_handle.cookie,
1932 atomic_read(&lock->l_export->exp_refcount) : -99,
1937 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1938 data->msg_fn, data->msg_line, fmt, args,
1939 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1940 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1941 "flags: %x remote: "LPX64" expref: %d "
1943 lock->l_resource->lr_namespace->ns_name,
1944 lock, lock->l_handle.h_cookie,
1945 atomic_read (&lock->l_refc),
1946 lock->l_readers, lock->l_writers,
1947 ldlm_lockname[lock->l_granted_mode],
1948 ldlm_lockname[lock->l_req_mode],
1949 lock->l_resource->lr_name.name[0],
1950 lock->l_resource->lr_name.name[1],
1951 lock->l_policy_data.l_inodebits.bits,
1952 atomic_read(&lock->l_resource->lr_refcount),
1953 ldlm_typename[lock->l_resource->lr_type],
1954 lock->l_flags, lock->l_remote_handle.cookie,
1956 atomic_read(&lock->l_export->exp_refcount) : -99,
1961 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1962 data->msg_fn, data->msg_line, fmt, args,
1963 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1964 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: %x "
1965 "remote: "LPX64" expref: %d pid: %u\n",
1966 lock->l_resource->lr_namespace->ns_name,
1967 lock, lock->l_handle.h_cookie,
1968 atomic_read (&lock->l_refc),
1969 lock->l_readers, lock->l_writers,
1970 ldlm_lockname[lock->l_granted_mode],
1971 ldlm_lockname[lock->l_req_mode],
1972 lock->l_resource->lr_name.name[0],
1973 lock->l_resource->lr_name.name[1],
1974 atomic_read(&lock->l_resource->lr_refcount),
1975 ldlm_typename[lock->l_resource->lr_type],
1976 lock->l_flags, lock->l_remote_handle.cookie,
1978 atomic_read(&lock->l_export->exp_refcount) : -99,
1984 EXPORT_SYMBOL(_ldlm_lock_debug);