1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Phil Schwan <phil@clusterfs.com>
8 * This file is part of Lustre, http://www.lustre.org.
10 * Lustre is free software; you can redistribute it and/or
11 * modify it under the terms of version 2 of the GNU General Public
12 * License as published by the Free Software Foundation.
14 * Lustre is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with Lustre; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 #define DEBUG_SUBSYSTEM S_LDLM
27 # include <linux/slab.h>
28 # include <linux/module.h>
29 # include <linux/lustre_dlm.h>
31 # include <liblustre.h>
32 # include <linux/kp30.h>
35 #include <linux/obd_class.h>
36 #include "ldlm_internal.h"
38 //struct lustre_lock ldlm_everything_lock;
41 char *ldlm_lockname[] = {
51 char *ldlm_typename[] = {
57 char *ldlm_it2str(int it)
64 case (IT_OPEN | IT_CREAT):
77 CERROR("Unknown intent %d\n", it);
82 extern kmem_cache_t *ldlm_lock_slab;
83 struct lustre_lock ldlm_handle_lock;
85 static ldlm_processing_policy ldlm_processing_policy_table[] = {
86 [LDLM_PLAIN] ldlm_process_plain_lock,
87 [LDLM_EXTENT] ldlm_process_extent_lock,
89 [LDLM_FLOCK] ldlm_process_flock_lock,
93 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
95 return ldlm_processing_policy_table[res->lr_type];
98 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
104 * REFCOUNTED LOCK OBJECTS
109 * Lock refcounts, during creation:
110 * - one special one for allocation, dec'd only once in destroy
111 * - one for being a lock that's in-use
112 * - one for the addref associated with a new lock
114 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
116 atomic_inc(&lock->l_refc);
120 void ldlm_lock_put(struct ldlm_lock *lock)
124 if (atomic_dec_and_test(&lock->l_refc)) {
125 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
127 l_lock(&ns->ns_lock);
128 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
129 LASSERT(lock->l_destroyed);
130 LASSERT(list_empty(&lock->l_res_link));
132 spin_lock(&ns->ns_counter_lock);
134 spin_unlock(&ns->ns_counter_lock);
136 ldlm_resource_putref(lock->l_resource);
137 lock->l_resource = NULL;
139 class_export_put(lock->l_export);
142 LDLM_LOCK_PUT(lock->l_parent);
144 if (lock->l_lvb_data != NULL)
145 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
147 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
148 l_unlock(&ns->ns_lock);
154 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
157 l_lock(&lock->l_resource->lr_namespace->ns_lock);
158 if (!list_empty(&lock->l_lru)) {
159 list_del_init(&lock->l_lru);
160 lock->l_resource->lr_namespace->ns_nr_unused--;
161 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
163 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
167 /* This used to have a 'strict' flact, which recovery would use to mark an
168 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
169 * shall explain why it's gone: with the new hash table scheme, once you call
170 * ldlm_lock_destroy, you can never drop your final references on this lock.
171 * Because it's not in the hash table anymore. -phil */
172 void ldlm_lock_destroy(struct ldlm_lock *lock)
175 l_lock(&lock->l_resource->lr_namespace->ns_lock);
177 if (!list_empty(&lock->l_children)) {
178 LDLM_ERROR(lock, "still has children (%p)!",
179 lock->l_children.next);
180 ldlm_lock_dump(D_ERROR, lock, 0);
183 if (lock->l_readers || lock->l_writers) {
184 LDLM_ERROR(lock, "lock still has references");
185 ldlm_lock_dump(D_ERROR, lock, 0);
189 if (!list_empty(&lock->l_res_link)) {
190 LDLM_ERROR(lock, "lock still on resource");
191 ldlm_lock_dump(D_ERROR, lock, 0);
195 if (lock->l_destroyed) {
196 LASSERT(list_empty(&lock->l_lru));
197 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
201 lock->l_destroyed = 1;
203 list_del_init(&lock->l_export_chain);
204 ldlm_lock_remove_from_lru(lock);
205 class_handle_unhash(&lock->l_handle);
208 /* Wake anyone waiting for this lock */
209 /* FIXME: I should probably add yet another flag, instead of using
210 * l_export to only call this on clients */
212 class_export_put(lock->l_export);
213 lock->l_export = NULL;
214 if (lock->l_export && lock->l_completion_ast)
215 lock->l_completion_ast(lock, 0);
218 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
223 /* this is called by portals_handle2object with the handle lock taken */
224 static void lock_handle_addref(void *lock)
226 LDLM_LOCK_GET((struct ldlm_lock *)lock);
230 * usage: pass in a resource on which you have done ldlm_resource_get
231 * pass in a parent lock on which you have done a ldlm_lock_get
232 * after return, ldlm_*_put the resource and parent
233 * returns: lock with refcount 1
235 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
236 struct ldlm_resource *resource)
238 struct ldlm_lock *lock;
241 if (resource == NULL)
244 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, SLAB_NOFS, sizeof(*lock));
248 lock->l_resource = ldlm_resource_getref(resource);
250 atomic_set(&lock->l_refc, 2);
251 INIT_LIST_HEAD(&lock->l_children);
252 INIT_LIST_HEAD(&lock->l_res_link);
253 INIT_LIST_HEAD(&lock->l_lru);
254 INIT_LIST_HEAD(&lock->l_export_chain);
255 INIT_LIST_HEAD(&lock->l_pending_chain);
256 init_waitqueue_head(&lock->l_waitq);
258 spin_lock(&resource->lr_namespace->ns_counter_lock);
259 resource->lr_namespace->ns_locks++;
260 spin_unlock(&resource->lr_namespace->ns_counter_lock);
262 if (parent != NULL) {
263 l_lock(&parent->l_resource->lr_namespace->ns_lock);
264 lock->l_parent = LDLM_LOCK_GET(parent);
265 list_add(&lock->l_childof, &parent->l_children);
266 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
269 INIT_LIST_HEAD(&lock->l_handle.h_link);
270 class_handle_hash(&lock->l_handle, lock_handle_addref);
275 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
276 struct ldlm_res_id new_resid)
278 struct ldlm_resource *oldres = lock->l_resource;
281 l_lock(&ns->ns_lock);
282 if (memcmp(&new_resid, &lock->l_resource->lr_name,
283 sizeof(lock->l_resource->lr_name)) == 0) {
285 l_unlock(&ns->ns_lock);
289 LASSERT(new_resid.name[0] != 0);
291 /* This function assumes that the lock isn't on any lists */
292 LASSERT(list_empty(&lock->l_res_link));
294 lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
295 lock->l_resource->lr_type, 1);
296 if (lock->l_resource == NULL) {
301 /* ...and the flowers are still standing! */
302 ldlm_resource_putref(oldres);
304 l_unlock(&ns->ns_lock);
312 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
314 lockh->cookie = lock->l_handle.h_cookie;
317 /* if flags: atomically get the lock and set the flags.
318 * Return NULL if flag already set
321 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
323 struct ldlm_namespace *ns;
324 struct ldlm_lock *lock = NULL, *retval = NULL;
329 lock = class_handle2object(handle->cookie);
333 LASSERT(lock->l_resource != NULL);
334 ns = lock->l_resource->lr_namespace;
337 l_lock(&ns->ns_lock);
339 /* It's unlikely but possible that someone marked the lock as
340 * destroyed after we did handle2object on it */
341 if (lock->l_destroyed) {
342 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
347 if (flags && (lock->l_flags & flags)) {
353 lock->l_flags |= flags;
358 l_unlock(&ns->ns_lock);
362 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
363 struct lustre_handle *handle)
365 struct ldlm_lock *retval = NULL;
367 l_lock(&ns->ns_lock);
368 retval = __ldlm_handle2lock(handle, 0);
369 l_unlock(&ns->ns_lock);
374 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
376 ldlm_res2desc(lock->l_resource, &desc->l_resource);
377 desc->l_req_mode = lock->l_req_mode;
378 desc->l_granted_mode = lock->l_granted_mode;
379 memcpy(&desc->l_policy_data, &lock->l_policy_data,
380 sizeof(desc->l_policy_data));
383 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
384 void *data, int datalen)
386 struct ldlm_ast_work *w;
389 l_lock(&lock->l_resource->lr_namespace->ns_lock);
390 if (new && (lock->l_flags & LDLM_FL_AST_SENT))
393 CDEBUG(D_OTHER, "lock %p incompatible; sending blocking AST.\n", lock);
395 OBD_ALLOC(w, sizeof(*w));
402 w->w_datalen = datalen;
404 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
405 lock->l_flags |= LDLM_FL_AST_SENT;
406 /* If the enqueuing client said so, tell the AST recipient to
407 * discard dirty data, rather than writing back. */
408 if (new->l_flags & LDLM_AST_DISCARD_DATA)
409 lock->l_flags |= LDLM_FL_DISCARD_DATA;
411 ldlm_lock2desc(new, &w->w_desc);
414 w->w_lock = LDLM_LOCK_GET(lock);
415 list_add(&w->w_list, lock->l_resource->lr_tmp);
418 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
421 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
423 struct ldlm_lock *lock;
425 lock = ldlm_handle2lock(lockh);
426 ldlm_lock_addref_internal(lock, mode);
430 /* only called for local locks */
431 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
433 l_lock(&lock->l_resource->lr_namespace->ns_lock);
434 ldlm_lock_remove_from_lru(lock);
435 if (mode & (LCK_NL | LCK_CR | LCK_PR))
437 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
439 lock->l_last_used = jiffies;
440 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
442 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
445 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
447 struct ldlm_namespace *ns;
450 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
451 ns = lock->l_resource->lr_namespace;
452 l_lock(&ns->ns_lock);
453 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
454 LASSERT(lock->l_readers > 0);
457 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
458 LASSERT(lock->l_writers > 0);
462 if (lock->l_flags & LDLM_FL_LOCAL &&
463 !lock->l_readers && !lock->l_writers) {
464 /* If this is a local lock on a server namespace and this was
465 * the last reference, cancel the lock. */
466 CDEBUG(D_INFO, "forcing cancel of local lock\n");
467 lock->l_flags |= LDLM_FL_CBPENDING;
470 if (!lock->l_readers && !lock->l_writers &&
471 (lock->l_flags & LDLM_FL_CBPENDING)) {
472 /* If we received a blocked AST and this was the last reference,
473 * run the callback. */
474 if (ns->ns_client == LDLM_NAMESPACE_SERVER && lock->l_export)
475 CERROR("FL_CBPENDING set on non-local lock--just a "
478 LDLM_DEBUG(lock, "final decref done on cbpending lock");
480 LDLM_LOCK_GET(lock); /* dropped by bl thread */
481 ldlm_lock_remove_from_lru(lock);
483 ldlm_bl_to_thread(ns, NULL, lock);
484 l_unlock(&ns->ns_lock);
486 l_unlock(&ns->ns_lock);
487 ldlm_handle_bl_callback(ns, NULL, lock);
489 } else if (ns->ns_client == LDLM_NAMESPACE_CLIENT &&
490 !lock->l_readers && !lock->l_writers) {
491 /* If this is a client-side namespace and this was the last
492 * reference, put it on the LRU. */
493 LASSERT(list_empty(&lock->l_lru));
494 LASSERT(ns->ns_nr_unused >= 0);
495 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
497 l_unlock(&ns->ns_lock);
498 ldlm_cancel_lru(ns, LDLM_ASYNC);
500 l_unlock(&ns->ns_lock);
503 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
508 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
510 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
511 LASSERT(lock != NULL);
512 ldlm_lock_decref_internal(lock, mode);
516 /* This will drop a lock reference and mark it for destruction, but will not
517 * necessarily cancel the lock before returning. */
518 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
520 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
523 LASSERT(lock != NULL);
525 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
526 l_lock(&lock->l_resource->lr_namespace->ns_lock);
527 lock->l_flags |= LDLM_FL_CBPENDING;
528 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
529 ldlm_lock_decref_internal(lock, mode);
534 * - ldlm_lock_enqueue
535 * - ldlm_reprocess_queue
536 * - ldlm_lock_convert
538 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen,
541 struct ldlm_resource *res = lock->l_resource;
544 l_lock(&lock->l_resource->lr_namespace->ns_lock);
545 lock->l_granted_mode = lock->l_req_mode;
546 ldlm_resource_add_lock(res, &res->lr_granted, lock);
548 if (lock->l_granted_mode < res->lr_most_restr)
549 res->lr_most_restr = lock->l_granted_mode;
551 if (run_ast && lock->l_completion_ast != NULL)
552 ldlm_add_ast_work_item(lock, NULL, data, datalen);
554 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
558 /* returns a referenced lock or NULL. See the flag descriptions below, in the
559 * comment above ldlm_lock_match */
560 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
561 ldlm_policy_data_t *policy,
562 struct ldlm_lock *old_lock, int flags)
564 struct ldlm_lock *lock;
565 struct list_head *tmp;
567 list_for_each(tmp, queue) {
568 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
570 if (lock == old_lock)
573 /* llite sometimes wants to match locks that will be
574 * canceled when their users drop, but we allow it to match
575 * if it passes in CBPENDING and the lock still has users.
576 * this is generally only going to be used by children
577 * whose parents already hold a lock so forward progress
578 * can still happen. */
579 if (lock->l_flags & LDLM_FL_CBPENDING &&
580 !(flags & LDLM_FL_CBPENDING))
582 if (lock->l_flags & LDLM_FL_CBPENDING &&
583 lock->l_readers == 0 && lock->l_writers == 0)
586 if (!(lock->l_req_mode & mode))
589 if (lock->l_resource->lr_type == LDLM_EXTENT &&
590 (lock->l_policy_data.l_extent.start >
591 policy->l_extent.start ||
592 lock->l_policy_data.l_extent.end < policy->l_extent.end))
595 if (lock->l_resource->lr_type == LDLM_EXTENT &&
597 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
600 if (lock->l_destroyed)
603 if ((flags & LDLM_FL_LOCAL_ONLY) &&
604 !(lock->l_flags & LDLM_FL_LOCAL))
607 if (flags & LDLM_FL_TEST_LOCK)
610 ldlm_lock_addref_internal(lock, mode);
617 void ldlm_lock_allow_match(struct ldlm_lock *lock)
619 l_lock(&lock->l_resource->lr_namespace->ns_lock);
620 lock->l_flags |= LDLM_FL_CAN_MATCH;
621 wake_up(&lock->l_waitq);
622 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
625 /* Can be called in two ways:
627 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
628 * for a duplicate of.
630 * Otherwise, all of the fields must be filled in, to match against.
632 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
633 * server (ie, connh is NULL)
634 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
635 * list will be considered
636 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
637 * to be canceled can still be matched as long as they still have reader
638 * or writer refernces
639 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
640 * just tell us if we would have matched.
642 * Returns 1 if it finds an already-existing lock that is compatible; in this
643 * case, lockh is filled in with a addref()ed lock
645 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
646 struct ldlm_res_id *res_id, __u32 type,
647 ldlm_policy_data_t *policy, ldlm_mode_t mode,
648 struct lustre_handle *lockh)
650 struct ldlm_resource *res;
651 struct ldlm_lock *lock, *old_lock = NULL;
656 old_lock = ldlm_handle2lock(lockh);
659 ns = old_lock->l_resource->lr_namespace;
660 res_id = &old_lock->l_resource->lr_name;
661 type = old_lock->l_resource->lr_type;
662 mode = old_lock->l_req_mode;
665 res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
667 LASSERT(old_lock == NULL);
671 l_lock(&ns->ns_lock);
673 lock = search_queue(&res->lr_granted, mode, policy, old_lock, flags);
676 if (flags & LDLM_FL_BLOCK_GRANTED)
678 lock = search_queue(&res->lr_converting, mode, policy, old_lock, flags);
681 lock = search_queue(&res->lr_waiting, mode, policy, old_lock, flags);
687 ldlm_resource_putref(res);
688 l_unlock(&ns->ns_lock);
691 ldlm_lock2handle(lock, lockh);
692 if (!(lock->l_flags & LDLM_FL_CAN_MATCH)) {
693 struct l_wait_info lwi;
694 if (lock->l_completion_ast)
695 lock->l_completion_ast(lock,
696 LDLM_FL_WAIT_NOREPROC,
699 lwi = LWI_TIMEOUT_INTR(obd_timeout*HZ, NULL,NULL,NULL);
701 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
702 l_wait_event(lock->l_waitq,
703 (lock->l_flags & LDLM_FL_CAN_MATCH), &lwi);
707 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
708 type == LDLM_PLAIN ? res_id->name[2] :
709 policy->l_extent.start,
710 type == LDLM_PLAIN ? res_id->name[3] :
711 policy->l_extent.end);
712 else if (!(flags & LDLM_FL_TEST_LOCK)) /* less verbose for test-only */
713 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
714 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
715 type, mode, res_id->name[0], res_id->name[1],
716 type == LDLM_PLAIN ? res_id->name[2] :
717 policy->l_extent.start,
718 type == LDLM_PLAIN ? res_id->name[3] :
719 policy->l_extent.end);
722 LDLM_LOCK_PUT(old_lock);
723 if (flags & LDLM_FL_TEST_LOCK && rc)
729 /* Returns a referenced lock */
730 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
731 struct lustre_handle *parent_lock_handle,
732 struct ldlm_res_id res_id, __u32 type,
734 ldlm_blocking_callback blocking,
735 ldlm_completion_callback completion,
736 ldlm_glimpse_callback glimpse,
737 void *data, __u32 lvb_len)
739 struct ldlm_resource *res, *parent_res = NULL;
740 struct ldlm_lock *lock, *parent_lock = NULL;
743 if (parent_lock_handle) {
744 parent_lock = ldlm_handle2lock(parent_lock_handle);
746 parent_res = parent_lock->l_resource;
749 res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
753 lock = ldlm_lock_new(parent_lock, res);
754 ldlm_resource_putref(res);
755 if (parent_lock != NULL)
756 LDLM_LOCK_PUT(parent_lock);
761 lock->l_req_mode = mode;
762 lock->l_ast_data = data;
763 lock->l_blocking_ast = blocking;
764 lock->l_completion_ast = completion;
765 lock->l_glimpse_ast = glimpse;
768 lock->l_lvb_len = lvb_len;
769 OBD_ALLOC(lock->l_lvb_data, lvb_len);
770 if (lock->l_lvb_data == NULL) {
771 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
779 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
780 struct ldlm_lock **lockp,
781 void *cookie, int *flags)
783 struct ldlm_lock *lock = *lockp;
784 struct ldlm_resource *res = lock->l_resource;
785 int local = res->lr_namespace->ns_client;
786 ldlm_processing_policy policy;
787 ldlm_error_t rc = ELDLM_OK;
790 /* policies are not executed on the client or during replay */
791 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
792 && !local && ns->ns_policy) {
793 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
795 if (rc == ELDLM_LOCK_REPLACED) {
796 /* The lock that was returned has already been granted,
797 * and placed into lockp. If it's not the same as the
798 * one we passed in, then destroy the old one and our
799 * work here is done. */
800 if (lock != *lockp) {
801 ldlm_lock_destroy(lock);
804 *flags |= LDLM_FL_LOCK_CHANGED;
806 } else if (rc == ELDLM_LOCK_ABORTED ||
807 (rc == 0 && (*flags & LDLM_FL_INTENT_ONLY))) {
808 ldlm_lock_destroy(lock);
811 LASSERT(rc == ELDLM_OK);
814 l_lock(&ns->ns_lock);
815 if (local && lock->l_req_mode == lock->l_granted_mode) {
816 /* The server returned a blocked lock, but it was granted before
817 * we got a chance to actually enqueue it. We don't need to do
819 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
820 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
824 /* Some flags from the enqueue want to make it into the AST, via the
826 lock->l_flags |= (*flags & LDLM_AST_DISCARD_DATA);
828 /* This distinction between local lock trees is very important; a client
829 * namespace only has information about locks taken by that client, and
830 * thus doesn't have enough information to decide for itself if it can
831 * be granted (below). In this case, we do exactly what the server
832 * tells us to do, as dictated by the 'flags'.
834 * We do exactly the same thing during recovery, when the server is
835 * more or less trusting the clients not to lie.
837 * FIXME (bug 268): Detect obvious lies by checking compatibility in
838 * granted/converting queues. */
839 ldlm_resource_unlink_lock(lock);
841 if (*flags & LDLM_FL_BLOCK_CONV)
842 ldlm_resource_add_lock(res, &res->lr_converting, lock);
843 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
844 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
846 ldlm_grant_lock(lock, NULL, 0, 0);
848 } else if (*flags & LDLM_FL_REPLAY) {
849 if (*flags & LDLM_FL_BLOCK_CONV) {
850 ldlm_resource_add_lock(res, &res->lr_converting, lock);
852 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
853 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
855 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
856 ldlm_grant_lock(lock, NULL, 0, 0);
859 /* If no flags, fall through to normal enqueue path. */
862 policy = ldlm_processing_policy_table[res->lr_type];
863 policy(lock, flags, 1, &rc);
866 l_unlock(&ns->ns_lock);
870 /* Must be called with namespace taken: queue is waiting or converting. */
871 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue)
873 struct list_head *tmp, *pos;
874 ldlm_processing_policy policy;
876 int rc = LDLM_ITER_CONTINUE;
880 policy = ldlm_processing_policy_table[res->lr_type];
883 list_for_each_safe(tmp, pos, queue) {
884 struct ldlm_lock *pending;
885 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
887 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
890 rc = policy(pending, &flags, 0, &err);
891 if (rc != LDLM_ITER_CONTINUE)
898 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list)
900 struct list_head *tmp, *pos;
904 l_check_no_ns_lock(ns);
906 list_for_each_safe(tmp, pos, rpc_list) {
907 struct ldlm_ast_work *w =
908 list_entry(tmp, struct ldlm_ast_work, w_list);
910 /* It's possible to receive a completion AST before we've set
911 * the l_completion_ast pointer: either because the AST arrived
912 * before the reply, or simply because there's a small race
913 * window between receiving the reply and finishing the local
916 * This can't happen with the blocking_ast, however, because we
917 * will never call the local blocking_ast until we drop our
918 * reader/writer reference, which we won't do until we get the
919 * reply and finish enqueueing. */
920 LASSERT(w->w_lock != NULL);
922 LASSERT(w->w_lock->l_blocking_ast != NULL);
923 rc = w->w_lock->l_blocking_ast
924 (w->w_lock, &w->w_desc, w->w_data,
926 } else if (w->w_lock->l_completion_ast != NULL) {
927 LASSERT(w->w_lock->l_completion_ast != NULL);
928 rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags,
936 CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
937 "disconnect client\n");
938 LDLM_LOCK_PUT(w->w_lock);
939 list_del(&w->w_list);
940 OBD_FREE(w, sizeof(*w));
945 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
947 ldlm_reprocess_all(res);
948 return LDLM_ITER_CONTINUE;
951 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
955 l_lock(&ns->ns_lock);
956 for (i = 0; i < RES_HASH_SIZE; i++) {
957 struct list_head *tmp, *next;
958 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
959 struct ldlm_resource *res =
960 list_entry(tmp, struct ldlm_resource, lr_hash);
962 ldlm_resource_getref(res);
963 l_unlock(&ns->ns_lock);
964 rc = reprocess_one_queue(res, NULL);
965 l_lock(&ns->ns_lock);
967 ldlm_resource_putref(res);
968 if (rc == LDLM_ITER_STOP)
973 l_unlock(&ns->ns_lock);
977 void ldlm_reprocess_all(struct ldlm_resource *res)
979 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
983 /* Local lock trees don't get reprocessed. */
984 if (res->lr_namespace->ns_client) {
990 l_lock(&res->lr_namespace->ns_lock);
991 res->lr_tmp = &rpc_list;
993 rc = ldlm_reprocess_queue(res, &res->lr_converting);
994 if (rc == LDLM_ITER_CONTINUE)
995 ldlm_reprocess_queue(res, &res->lr_waiting);
998 l_unlock(&res->lr_namespace->ns_lock);
1000 rc = ldlm_run_ast_work(res->lr_namespace, &rpc_list);
1001 if (rc == -ERESTART) {
1002 LASSERT(list_empty(&rpc_list));
1008 void ldlm_cancel_callback(struct ldlm_lock *lock)
1010 l_lock(&lock->l_resource->lr_namespace->ns_lock);
1011 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1012 lock->l_flags |= LDLM_FL_CANCEL;
1013 if (lock->l_blocking_ast) {
1014 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1015 // l_check_no_ns_lock(lock->l_resource->lr_namespace);
1016 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1020 LDLM_DEBUG(lock, "no blocking ast");
1023 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1026 void ldlm_lock_cancel(struct ldlm_lock *lock)
1028 struct ldlm_resource *res;
1029 struct ldlm_namespace *ns;
1032 /* There's no race between calling this and taking the ns lock below;
1033 * a lock can only be put on the waiting list once, because it can only
1034 * issue a blocking AST once. */
1035 ldlm_del_waiting_lock(lock);
1037 res = lock->l_resource;
1038 ns = res->lr_namespace;
1040 l_lock(&ns->ns_lock);
1041 /* Please do not, no matter how tempting, remove this LBUG without
1042 * talking to me first. -phik */
1043 if (lock->l_readers || lock->l_writers) {
1044 LDLM_ERROR(lock, "lock still has references");
1048 ldlm_cancel_callback(lock);
1050 ldlm_resource_unlink_lock(lock);
1051 ldlm_lock_destroy(lock);
1052 l_unlock(&ns->ns_lock);
1056 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1058 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1064 lock->l_ast_data = data;
1065 LDLM_LOCK_PUT(lock);
1069 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1071 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1072 struct ldlm_lock *lock;
1073 struct ldlm_resource *res;
1075 l_lock(&ns->ns_lock);
1076 while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1077 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1078 struct ldlm_lock, l_export_chain);
1079 res = ldlm_resource_getref(lock->l_resource);
1080 LDLM_DEBUG(lock, "export %p", exp);
1081 ldlm_lock_cancel(lock);
1082 l_unlock(&ns->ns_lock);
1083 ldlm_reprocess_all(res);
1084 ldlm_resource_putref(res);
1085 l_lock(&ns->ns_lock);
1087 l_unlock(&ns->ns_lock);
1090 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1093 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1094 struct ldlm_resource *res;
1095 struct ldlm_namespace *ns;
1101 res = lock->l_resource;
1102 ns = res->lr_namespace;
1104 l_lock(&ns->ns_lock);
1106 lock->l_req_mode = new_mode;
1107 ldlm_resource_unlink_lock(lock);
1109 /* If this is a local resource, put it on the appropriate list. */
1110 if (res->lr_namespace->ns_client) {
1111 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1112 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1114 /* This should never happen, because of the way the
1115 * server handles conversions. */
1118 res->lr_tmp = &rpc_list;
1119 ldlm_grant_lock(lock, NULL, 0, 0);
1122 /* FIXME: completion handling not with ns_lock held ! */
1123 if (lock->l_completion_ast)
1124 lock->l_completion_ast(lock, 0, NULL);
1127 /* FIXME: We should try the conversion right away and possibly
1128 * return success without the need for an extra AST */
1129 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1130 *flags |= LDLM_FL_BLOCK_CONV;
1133 l_unlock(&ns->ns_lock);
1136 ldlm_run_ast_work(ns, &rpc_list);
1140 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1142 char str[PTL_NALFMT_SIZE];
1143 struct obd_device *obd = NULL;
1145 if (!((portal_debug | D_ERROR) & level))
1149 CDEBUG(level, " NULL LDLM lock\n");
1153 CDEBUG(level, " -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d)\n",
1154 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1156 if (lock->l_conn_export != NULL)
1157 obd = lock->l_conn_export->exp_obd;
1158 if (lock->l_export && lock->l_export->exp_connection) {
1159 CDEBUG(level, " Node: NID "LPX64" (%s) on %s (rhandle: "LPX64")\n",
1160 lock->l_export->exp_connection->c_peer.peer_nid,
1161 portals_nid2str(lock->l_export->exp_connection->c_peer.peer_ni->pni_number,
1162 lock->l_export->exp_connection->c_peer.peer_nid, str),
1163 lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1164 lock->l_remote_handle.cookie);
1165 } else if (obd == NULL) {
1166 CDEBUG(level, " Node: local\n");
1168 struct obd_import *imp = obd->u.cli.cl_import;
1169 CDEBUG(level, " Node: NID "LPX64" (%s) on %s (rhandle: "LPX64")\n",
1170 imp->imp_connection->c_peer.peer_nid,
1171 portals_nid2str(imp->imp_connection->c_peer.peer_ni->pni_number,
1172 imp->imp_connection->c_peer.peer_nid, str),
1173 imp->imp_connection->c_peer.peer_ni->pni_name,
1174 lock->l_remote_handle.cookie);
1176 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource,
1177 lock->l_resource->lr_name.name[0],
1178 lock->l_resource->lr_name.name[1]);
1179 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1180 "write: %d\n", ldlm_lockname[lock->l_req_mode],
1181 ldlm_lockname[lock->l_granted_mode],
1182 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers);
1183 if (lock->l_resource->lr_type == LDLM_EXTENT)
1184 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1185 " (req "LPU64"-"LPU64")\n",
1186 lock->l_policy_data.l_extent.start,
1187 lock->l_policy_data.l_extent.end,
1188 lock->l_req_extent.start, lock->l_req_extent.end);
1189 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1190 CDEBUG(level, " Pid: "LPU64" Extent: "LPU64" -> "LPU64"\n",
1191 lock->l_policy_data.l_flock.pid,
1192 lock->l_policy_data.l_flock.start,
1193 lock->l_policy_data.l_flock.end);
1196 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1198 struct ldlm_lock *lock;
1200 lock = ldlm_handle2lock(lockh);
1204 ldlm_lock_dump(D_OTHER, lock, 0);
1206 LDLM_LOCK_PUT(lock);