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/dcache.h>
29 # include <linux/namei.h>
30 # include <linux/module.h>
31 # include <linux/lustre_dlm.h>
33 # include <liblustre.h>
34 # include <libcfs/kp30.h>
37 #include <linux/obd_class.h>
38 #include "ldlm_internal.h"
40 //struct lustre_lock ldlm_everything_lock;
43 char *ldlm_lockname[] = {
53 char *ldlm_typename[] = {
60 char *ldlm_it2str(int it)
67 case (IT_OPEN | IT_CREAT):
82 CERROR("Unknown intent %d\n", it);
87 extern kmem_cache_t *ldlm_lock_slab;
88 struct lustre_lock ldlm_handle_lock;
90 static ldlm_processing_policy ldlm_processing_policy_table[] = {
91 [LDLM_PLAIN] ldlm_process_plain_lock,
92 [LDLM_EXTENT] ldlm_process_extent_lock,
94 [LDLM_FLOCK] ldlm_process_flock_lock,
96 [LDLM_IBITS] ldlm_process_inodebits_lock,
99 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
101 return ldlm_processing_policy_table[res->lr_type];
104 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
110 * REFCOUNTED LOCK OBJECTS
115 * Lock refcounts, during creation:
116 * - one special one for allocation, dec'd only once in destroy
117 * - one for being a lock that's in-use
118 * - one for the addref associated with a new lock
120 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
122 atomic_inc(&lock->l_refc);
126 void ldlm_lock_put(struct ldlm_lock *lock)
130 if (atomic_dec_and_test(&lock->l_refc)) {
131 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
133 l_lock(&ns->ns_lock);
134 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
135 LASSERT(lock->l_destroyed);
136 LASSERT(list_empty(&lock->l_res_link));
138 spin_lock(&ns->ns_counter_lock);
140 spin_unlock(&ns->ns_counter_lock);
142 ldlm_resource_putref(lock->l_resource);
143 lock->l_resource = NULL;
145 class_export_put(lock->l_export);
148 LDLM_LOCK_PUT(lock->l_parent);
150 if (lock->l_lvb_data != NULL)
151 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
153 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
154 l_unlock(&ns->ns_lock);
160 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
163 l_lock(&lock->l_resource->lr_namespace->ns_lock);
164 if (!list_empty(&lock->l_lru)) {
165 list_del_init(&lock->l_lru);
166 lock->l_resource->lr_namespace->ns_nr_unused--;
167 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
169 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
173 /* This used to have a 'strict' flact, which recovery would use to mark an
174 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
175 * shall explain why it's gone: with the new hash table scheme, once you call
176 * ldlm_lock_destroy, you can never drop your final references on this lock.
177 * Because it's not in the hash table anymore. -phil */
178 void ldlm_lock_destroy(struct ldlm_lock *lock)
181 l_lock(&lock->l_resource->lr_namespace->ns_lock);
183 if (!list_empty(&lock->l_children)) {
184 LDLM_ERROR(lock, "still has children (%p)!",
185 lock->l_children.next);
186 ldlm_lock_dump(D_ERROR, lock, 0);
189 if (lock->l_readers || lock->l_writers) {
190 LDLM_ERROR(lock, "lock still has references");
191 ldlm_lock_dump(D_ERROR, lock, 0);
195 if (!list_empty(&lock->l_res_link)) {
196 LDLM_ERROR(lock, "lock still on resource");
197 ldlm_lock_dump(D_ERROR, lock, 0);
201 if (lock->l_destroyed) {
202 LASSERT(list_empty(&lock->l_lru));
203 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
207 lock->l_destroyed = 1;
209 list_del_init(&lock->l_export_chain);
210 ldlm_lock_remove_from_lru(lock);
211 class_handle_unhash(&lock->l_handle);
214 /* Wake anyone waiting for this lock */
215 /* FIXME: I should probably add yet another flag, instead of using
216 * l_export to only call this on clients */
218 class_export_put(lock->l_export);
219 lock->l_export = NULL;
220 if (lock->l_export && lock->l_completion_ast)
221 lock->l_completion_ast(lock, 0);
224 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
229 /* this is called by portals_handle2object with the handle lock taken */
230 static void lock_handle_addref(void *lock)
232 LDLM_LOCK_GET((struct ldlm_lock *)lock);
236 * usage: pass in a resource on which you have done ldlm_resource_get
237 * pass in a parent lock on which you have done a ldlm_lock_get
238 * after return, ldlm_*_put the resource and parent
239 * returns: lock with refcount 2 - one for current caller and one for remote
241 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
242 struct ldlm_resource *resource)
244 struct ldlm_lock *lock;
247 if (resource == NULL)
250 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, SLAB_NOFS, sizeof(*lock));
254 lock->l_resource = ldlm_resource_getref(resource);
256 atomic_set(&lock->l_refc, 2);
257 INIT_LIST_HEAD(&lock->l_children);
258 INIT_LIST_HEAD(&lock->l_res_link);
259 INIT_LIST_HEAD(&lock->l_lru);
260 INIT_LIST_HEAD(&lock->l_export_chain);
261 INIT_LIST_HEAD(&lock->l_pending_chain);
262 init_waitqueue_head(&lock->l_waitq);
264 spin_lock(&resource->lr_namespace->ns_counter_lock);
265 resource->lr_namespace->ns_locks++;
266 spin_unlock(&resource->lr_namespace->ns_counter_lock);
268 if (parent != NULL) {
269 l_lock(&parent->l_resource->lr_namespace->ns_lock);
270 lock->l_parent = LDLM_LOCK_GET(parent);
271 list_add(&lock->l_childof, &parent->l_children);
272 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
275 INIT_LIST_HEAD(&lock->l_handle.h_link);
276 class_handle_hash(&lock->l_handle, lock_handle_addref);
281 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
282 struct ldlm_res_id new_resid)
284 struct ldlm_resource *oldres = lock->l_resource;
287 l_lock(&ns->ns_lock);
288 if (memcmp(&new_resid, &lock->l_resource->lr_name,
289 sizeof(lock->l_resource->lr_name)) == 0) {
291 l_unlock(&ns->ns_lock);
295 LASSERT(new_resid.name[0] != 0);
297 /* This function assumes that the lock isn't on any lists */
298 LASSERT(list_empty(&lock->l_res_link));
300 lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
301 lock->l_resource->lr_type, 1);
302 if (lock->l_resource == NULL) {
307 /* ...and the flowers are still standing! */
308 ldlm_resource_putref(oldres);
310 l_unlock(&ns->ns_lock);
318 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
320 lockh->cookie = lock->l_handle.h_cookie;
323 /* if flags: atomically get the lock and set the flags.
324 * Return NULL if flag already set
327 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
329 struct ldlm_namespace *ns;
330 struct ldlm_lock *lock = NULL, *retval = NULL;
335 lock = class_handle2object(handle->cookie);
339 LASSERT(lock->l_resource != NULL);
340 ns = lock->l_resource->lr_namespace;
343 l_lock(&ns->ns_lock);
345 /* It's unlikely but possible that someone marked the lock as
346 * destroyed after we did handle2object on it */
347 if (lock->l_destroyed) {
348 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
353 if (flags && (lock->l_flags & flags)) {
359 lock->l_flags |= flags;
364 l_unlock(&ns->ns_lock);
368 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
369 struct lustre_handle *handle)
371 struct ldlm_lock *retval = NULL;
373 l_lock(&ns->ns_lock);
374 retval = __ldlm_handle2lock(handle, 0);
375 l_unlock(&ns->ns_lock);
380 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
382 ldlm_res2desc(lock->l_resource, &desc->l_resource);
383 desc->l_req_mode = lock->l_req_mode;
384 desc->l_granted_mode = lock->l_granted_mode;
385 memcpy(&desc->l_policy_data, &lock->l_policy_data,
386 sizeof(desc->l_policy_data));
389 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
390 void *data, int datalen)
392 struct ldlm_ast_work *w;
395 l_lock(&lock->l_resource->lr_namespace->ns_lock);
396 if (new && (lock->l_flags & LDLM_FL_AST_SENT))
399 CDEBUG(D_OTHER, "lock %p incompatible; sending blocking AST.\n", lock);
401 OBD_ALLOC(w, sizeof(*w));
408 w->w_datalen = datalen;
410 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
411 lock->l_flags |= LDLM_FL_AST_SENT;
412 /* If the enqueuing client said so, tell the AST recipient to
413 * discard dirty data, rather than writing back. */
414 if (new->l_flags & LDLM_AST_DISCARD_DATA)
415 lock->l_flags |= LDLM_FL_DISCARD_DATA;
417 ldlm_lock2desc(new, &w->w_desc);
420 w->w_lock = LDLM_LOCK_GET(lock);
421 list_add(&w->w_list, lock->l_resource->lr_tmp);
424 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
427 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
429 struct ldlm_lock *lock;
431 lock = ldlm_handle2lock(lockh);
432 ldlm_lock_addref_internal(lock, mode);
436 /* only called for local locks */
437 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
439 l_lock(&lock->l_resource->lr_namespace->ns_lock);
440 ldlm_lock_remove_from_lru(lock);
441 if (mode & (LCK_NL | LCK_CR | LCK_PR))
443 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
445 lock->l_last_used = jiffies;
446 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
448 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
451 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
453 struct ldlm_namespace *ns;
456 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
457 ns = lock->l_resource->lr_namespace;
458 l_lock(&ns->ns_lock);
459 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
460 LASSERT(lock->l_readers > 0);
463 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
464 LASSERT(lock->l_writers > 0);
468 if (lock->l_flags & LDLM_FL_LOCAL &&
469 !lock->l_readers && !lock->l_writers) {
470 /* If this is a local lock on a server namespace and this was
471 * the last reference, cancel the lock. */
472 CDEBUG(D_INFO, "forcing cancel of local lock\n");
473 lock->l_flags |= LDLM_FL_CBPENDING;
476 if (!lock->l_readers && !lock->l_writers &&
477 (lock->l_flags & LDLM_FL_CBPENDING)) {
478 /* If we received a blocked AST and this was the last reference,
479 * run the callback. */
480 if (ns->ns_client == LDLM_NAMESPACE_SERVER && lock->l_export)
481 CERROR("FL_CBPENDING set on non-local lock--just a "
484 LDLM_DEBUG(lock, "final decref done on cbpending lock");
486 LDLM_LOCK_GET(lock); /* dropped by bl thread */
487 ldlm_lock_remove_from_lru(lock);
488 l_unlock(&ns->ns_lock);
489 if (ldlm_bl_to_thread(ns, NULL, lock) != 0)
490 ldlm_handle_bl_callback(ns, NULL, lock);
491 } else if (ns->ns_client == LDLM_NAMESPACE_CLIENT &&
492 !lock->l_readers && !lock->l_writers) {
493 /* If this is a client-side namespace and this was the last
494 * reference, put it on the LRU. */
495 LASSERT(list_empty(&lock->l_lru));
496 LASSERT(ns->ns_nr_unused >= 0);
497 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
499 l_unlock(&ns->ns_lock);
500 ldlm_cancel_lru(ns, LDLM_ASYNC);
502 l_unlock(&ns->ns_lock);
505 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
510 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
512 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
513 LASSERT(lock != NULL);
514 ldlm_lock_decref_internal(lock, mode);
518 /* This will drop a lock reference and mark it for destruction, but will not
519 * necessarily cancel the lock before returning. */
520 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
522 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
525 LASSERT(lock != NULL);
527 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
528 l_lock(&lock->l_resource->lr_namespace->ns_lock);
529 lock->l_flags |= LDLM_FL_CBPENDING;
530 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
531 ldlm_lock_decref_internal(lock, mode);
536 * - ldlm_lock_enqueue
537 * - ldlm_reprocess_queue
538 * - ldlm_lock_convert
540 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen,
543 struct ldlm_resource *res = lock->l_resource;
546 l_lock(&lock->l_resource->lr_namespace->ns_lock);
547 lock->l_granted_mode = lock->l_req_mode;
548 ldlm_resource_add_lock(res, &res->lr_granted, lock);
550 if (lock->l_granted_mode < res->lr_most_restr)
551 res->lr_most_restr = lock->l_granted_mode;
553 if (run_ast && lock->l_completion_ast != NULL)
554 ldlm_add_ast_work_item(lock, NULL, data, datalen);
556 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
560 /* returns a referenced lock or NULL. See the flag descriptions below, in the
561 * comment above ldlm_lock_match */
562 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
563 ldlm_policy_data_t *policy,
564 struct ldlm_lock *old_lock, int flags)
566 struct ldlm_lock *lock;
567 struct list_head *tmp;
569 list_for_each(tmp, queue) {
570 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
572 if (lock == old_lock)
575 /* llite sometimes wants to match locks that will be
576 * canceled when their users drop, but we allow it to match
577 * if it passes in CBPENDING and the lock still has users.
578 * this is generally only going to be used by children
579 * whose parents already hold a lock so forward progress
580 * can still happen. */
581 if (lock->l_flags & LDLM_FL_CBPENDING &&
582 !(flags & LDLM_FL_CBPENDING))
584 if (lock->l_flags & LDLM_FL_CBPENDING &&
585 lock->l_readers == 0 && lock->l_writers == 0)
588 if (!(lock->l_req_mode & mode))
591 if (lock->l_resource->lr_type == LDLM_EXTENT &&
592 (lock->l_policy_data.l_extent.start >
593 policy->l_extent.start ||
594 lock->l_policy_data.l_extent.end < policy->l_extent.end))
597 if (lock->l_resource->lr_type == LDLM_EXTENT &&
599 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
602 /* We match if we have existing lock with same or wider set
604 if (lock->l_resource->lr_type == LDLM_IBITS &&
605 ((lock->l_policy_data.l_inodebits.bits &
606 policy->l_inodebits.bits) !=
607 policy->l_inodebits.bits))
610 if (lock->l_destroyed)
613 if ((flags & LDLM_FL_LOCAL_ONLY) &&
614 !(lock->l_flags & LDLM_FL_LOCAL))
617 if (flags & LDLM_FL_TEST_LOCK)
620 ldlm_lock_addref_internal(lock, mode);
627 void ldlm_lock_allow_match(struct ldlm_lock *lock)
629 l_lock(&lock->l_resource->lr_namespace->ns_lock);
630 lock->l_flags |= LDLM_FL_CAN_MATCH;
631 wake_up(&lock->l_waitq);
632 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
635 /* Can be called in two ways:
637 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
638 * for a duplicate of.
640 * Otherwise, all of the fields must be filled in, to match against.
642 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
643 * server (ie, connh is NULL)
644 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
645 * list will be considered
646 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
647 * to be canceled can still be matched as long as they still have reader
648 * or writer refernces
649 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
650 * just tell us if we would have matched.
652 * Returns 1 if it finds an already-existing lock that is compatible; in this
653 * case, lockh is filled in with a addref()ed lock
655 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
656 struct ldlm_res_id *res_id, __u32 type,
657 ldlm_policy_data_t *policy, ldlm_mode_t mode,
658 struct lustre_handle *lockh)
660 struct ldlm_resource *res;
661 struct ldlm_lock *lock, *old_lock = NULL;
666 old_lock = ldlm_handle2lock(lockh);
669 ns = old_lock->l_resource->lr_namespace;
670 res_id = &old_lock->l_resource->lr_name;
671 type = old_lock->l_resource->lr_type;
672 mode = old_lock->l_req_mode;
675 res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
677 LASSERT(old_lock == NULL);
681 l_lock(&ns->ns_lock);
683 lock = search_queue(&res->lr_granted, mode, policy, old_lock, flags);
686 if (flags & LDLM_FL_BLOCK_GRANTED)
688 lock = search_queue(&res->lr_converting, mode, policy, old_lock, flags);
691 lock = search_queue(&res->lr_waiting, mode, policy, old_lock, flags);
697 ldlm_resource_putref(res);
698 l_unlock(&ns->ns_lock);
701 ldlm_lock2handle(lock, lockh);
702 if (!(lock->l_flags & LDLM_FL_CAN_MATCH)) {
703 struct l_wait_info lwi;
704 if (lock->l_completion_ast)
705 lock->l_completion_ast(lock,
706 LDLM_FL_WAIT_NOREPROC,
709 lwi = LWI_TIMEOUT_INTR(obd_timeout*HZ, NULL,NULL,NULL);
711 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
712 l_wait_event(lock->l_waitq,
713 (lock->l_flags & LDLM_FL_CAN_MATCH), &lwi);
717 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
718 type == LDLM_PLAIN ? res_id->name[2] :
719 policy->l_extent.start,
720 type == LDLM_PLAIN ? res_id->name[3] :
721 policy->l_extent.end);
722 else if (!(flags & LDLM_FL_TEST_LOCK)) /* less verbose for test-only */
723 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
724 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
725 type, mode, res_id->name[0], res_id->name[1],
726 type == LDLM_PLAIN ? res_id->name[2] :
727 policy->l_extent.start,
728 type == LDLM_PLAIN ? res_id->name[3] :
729 policy->l_extent.end);
732 LDLM_LOCK_PUT(old_lock);
733 if (flags & LDLM_FL_TEST_LOCK && rc)
739 /* Returns a referenced lock */
740 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
741 struct lustre_handle *parent_lock_handle,
742 struct ldlm_res_id res_id, __u32 type,
744 ldlm_blocking_callback blocking,
745 ldlm_completion_callback completion,
746 ldlm_glimpse_callback glimpse,
747 void *data, __u32 lvb_len)
749 struct ldlm_resource *res, *parent_res = NULL;
750 struct ldlm_lock *lock, *parent_lock = NULL;
753 if (parent_lock_handle) {
754 parent_lock = ldlm_handle2lock(parent_lock_handle);
756 parent_res = parent_lock->l_resource;
759 res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
763 lock = ldlm_lock_new(parent_lock, res);
764 ldlm_resource_putref(res);
765 if (parent_lock != NULL)
766 LDLM_LOCK_PUT(parent_lock);
771 lock->l_req_mode = mode;
772 lock->l_ast_data = data;
773 lock->l_blocking_ast = blocking;
774 lock->l_completion_ast = completion;
775 lock->l_glimpse_ast = glimpse;
778 lock->l_lvb_len = lvb_len;
779 OBD_ALLOC(lock->l_lvb_data, lvb_len);
780 if (lock->l_lvb_data == NULL) {
781 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
789 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
790 struct ldlm_lock **lockp,
791 void *cookie, int *flags)
793 struct ldlm_lock *lock = *lockp;
794 struct ldlm_resource *res = lock->l_resource;
795 int local = res->lr_namespace->ns_client;
796 ldlm_processing_policy policy;
797 ldlm_error_t rc = ELDLM_OK;
800 /* policies are not executed on the client or during replay */
801 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
802 && !local && ns->ns_policy) {
803 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
805 if (rc == ELDLM_LOCK_REPLACED) {
806 /* The lock that was returned has already been granted,
807 * and placed into lockp. If it's not the same as the
808 * one we passed in, then destroy the old one and our
809 * work here is done. */
810 if (lock != *lockp) {
811 ldlm_lock_destroy(lock);
814 *flags |= LDLM_FL_LOCK_CHANGED;
816 } else if (rc == ELDLM_LOCK_ABORTED ||
817 (rc == 0 && (*flags & LDLM_FL_INTENT_ONLY))) {
818 ldlm_lock_destroy(lock);
821 LASSERT(rc == ELDLM_OK);
824 l_lock(&ns->ns_lock);
825 if (local && lock->l_req_mode == lock->l_granted_mode) {
826 /* The server returned a blocked lock, but it was granted before
827 * we got a chance to actually enqueue it. We don't need to do
829 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
830 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
834 /* Some flags from the enqueue want to make it into the AST, via the
836 lock->l_flags |= (*flags & LDLM_AST_DISCARD_DATA);
838 /* This distinction between local lock trees is very important; a client
839 * namespace only has information about locks taken by that client, and
840 * thus doesn't have enough information to decide for itself if it can
841 * be granted (below). In this case, we do exactly what the server
842 * tells us to do, as dictated by the 'flags'.
844 * We do exactly the same thing during recovery, when the server is
845 * more or less trusting the clients not to lie.
847 * FIXME (bug 268): Detect obvious lies by checking compatibility in
848 * granted/converting queues. */
849 ldlm_resource_unlink_lock(lock);
851 if (*flags & LDLM_FL_BLOCK_CONV)
852 ldlm_resource_add_lock(res, &res->lr_converting, lock);
853 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
854 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
856 ldlm_grant_lock(lock, NULL, 0, 0);
858 } else if (*flags & LDLM_FL_REPLAY) {
859 if (*flags & LDLM_FL_BLOCK_CONV) {
860 ldlm_resource_add_lock(res, &res->lr_converting, lock);
862 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
863 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
865 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
866 ldlm_grant_lock(lock, NULL, 0, 0);
869 /* If no flags, fall through to normal enqueue path. */
872 policy = ldlm_processing_policy_table[res->lr_type];
873 policy(lock, flags, 1, &rc);
876 l_unlock(&ns->ns_lock);
880 /* Must be called with namespace taken: queue is waiting or converting. */
881 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue)
883 struct list_head *tmp, *pos;
884 ldlm_processing_policy policy;
886 int rc = LDLM_ITER_CONTINUE;
890 policy = ldlm_processing_policy_table[res->lr_type];
893 list_for_each_safe(tmp, pos, queue) {
894 struct ldlm_lock *pending;
895 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
897 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
900 rc = policy(pending, &flags, 0, &err);
901 if (rc != LDLM_ITER_CONTINUE)
908 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list)
910 struct list_head *tmp, *pos;
914 l_check_no_ns_lock(ns);
916 list_for_each_safe(tmp, pos, rpc_list) {
917 struct ldlm_ast_work *w =
918 list_entry(tmp, struct ldlm_ast_work, w_list);
920 /* It's possible to receive a completion AST before we've set
921 * the l_completion_ast pointer: either because the AST arrived
922 * before the reply, or simply because there's a small race
923 * window between receiving the reply and finishing the local
926 * This can't happen with the blocking_ast, however, because we
927 * will never call the local blocking_ast until we drop our
928 * reader/writer reference, which we won't do until we get the
929 * reply and finish enqueueing. */
930 LASSERT(w->w_lock != NULL);
932 LASSERT(w->w_lock->l_blocking_ast != NULL);
933 rc = w->w_lock->l_blocking_ast
934 (w->w_lock, &w->w_desc, w->w_data,
936 } else if (w->w_lock->l_completion_ast != NULL) {
937 LASSERT(w->w_lock->l_completion_ast != NULL);
938 rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags,
946 CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
947 "disconnect client\n");
948 LDLM_LOCK_PUT(w->w_lock);
949 list_del(&w->w_list);
950 OBD_FREE(w, sizeof(*w));
955 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
957 ldlm_reprocess_all(res);
958 return LDLM_ITER_CONTINUE;
961 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
965 l_lock(&ns->ns_lock);
966 for (i = 0; i < RES_HASH_SIZE; i++) {
967 struct list_head *tmp, *next;
968 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
969 struct ldlm_resource *res =
970 list_entry(tmp, struct ldlm_resource, lr_hash);
972 ldlm_resource_getref(res);
973 l_unlock(&ns->ns_lock);
974 rc = reprocess_one_queue(res, NULL);
975 l_lock(&ns->ns_lock);
977 ldlm_resource_putref(res);
978 if (rc == LDLM_ITER_STOP)
983 l_unlock(&ns->ns_lock);
987 void ldlm_reprocess_all(struct ldlm_resource *res)
989 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
993 /* Local lock trees don't get reprocessed. */
994 if (res->lr_namespace->ns_client) {
1000 l_lock(&res->lr_namespace->ns_lock);
1001 res->lr_tmp = &rpc_list;
1003 rc = ldlm_reprocess_queue(res, &res->lr_converting);
1004 if (rc == LDLM_ITER_CONTINUE)
1005 ldlm_reprocess_queue(res, &res->lr_waiting);
1008 l_unlock(&res->lr_namespace->ns_lock);
1010 rc = ldlm_run_ast_work(res->lr_namespace, &rpc_list);
1011 if (rc == -ERESTART) {
1012 LASSERT(list_empty(&rpc_list));
1018 void ldlm_cancel_callback(struct ldlm_lock *lock)
1020 l_lock(&lock->l_resource->lr_namespace->ns_lock);
1021 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1022 lock->l_flags |= LDLM_FL_CANCEL;
1023 if (lock->l_blocking_ast) {
1024 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1025 // l_check_no_ns_lock(lock->l_resource->lr_namespace);
1026 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1030 LDLM_DEBUG(lock, "no blocking ast");
1033 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1036 void ldlm_lock_cancel(struct ldlm_lock *lock)
1038 struct ldlm_resource *res;
1039 struct ldlm_namespace *ns;
1042 /* There's no race between calling this and taking the ns lock below;
1043 * a lock can only be put on the waiting list once, because it can only
1044 * issue a blocking AST once. */
1045 ldlm_del_waiting_lock(lock);
1047 res = lock->l_resource;
1048 ns = res->lr_namespace;
1050 l_lock(&ns->ns_lock);
1051 /* Please do not, no matter how tempting, remove this LBUG without
1052 * talking to me first. -phik */
1053 if (lock->l_readers || lock->l_writers) {
1054 LDLM_ERROR(lock, "lock still has references");
1058 ldlm_cancel_callback(lock);
1060 ldlm_resource_unlink_lock(lock);
1061 ldlm_lock_destroy(lock);
1062 l_unlock(&ns->ns_lock);
1066 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1068 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1074 lock->l_ast_data = data;
1075 LDLM_LOCK_PUT(lock);
1079 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1081 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1082 struct ldlm_lock *lock;
1083 struct ldlm_resource *res;
1085 l_lock(&ns->ns_lock);
1086 while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1087 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1088 struct ldlm_lock, l_export_chain);
1089 res = ldlm_resource_getref(lock->l_resource);
1090 LDLM_DEBUG(lock, "export %p", exp);
1091 ldlm_lock_cancel(lock);
1092 l_unlock(&ns->ns_lock);
1093 ldlm_reprocess_all(res);
1094 ldlm_resource_putref(res);
1095 l_lock(&ns->ns_lock);
1097 l_unlock(&ns->ns_lock);
1100 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1103 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1104 struct ldlm_resource *res;
1105 struct ldlm_namespace *ns;
1111 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1112 *flags |= LDLM_FL_BLOCK_GRANTED;
1113 RETURN(lock->l_resource);
1116 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1117 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1119 res = lock->l_resource;
1120 ns = res->lr_namespace;
1122 l_lock(&ns->ns_lock);
1124 old_mode = lock->l_req_mode;
1125 lock->l_req_mode = new_mode;
1126 ldlm_resource_unlink_lock(lock);
1128 /* If this is a local resource, put it on the appropriate list. */
1129 if (res->lr_namespace->ns_client) {
1130 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1131 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1133 /* This should never happen, because of the way the
1134 * server handles conversions. */
1135 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1139 res->lr_tmp = &rpc_list;
1140 ldlm_grant_lock(lock, NULL, 0, 0);
1143 /* FIXME: completion handling not with ns_lock held ! */
1144 if (lock->l_completion_ast)
1145 lock->l_completion_ast(lock, 0, NULL);
1149 ldlm_processing_policy policy;
1150 policy = ldlm_processing_policy_table[res->lr_type];
1151 res->lr_tmp = &rpc_list;
1152 rc = policy(lock, &pflags, 0, &err);
1154 if (rc == LDLM_ITER_STOP) {
1155 lock->l_req_mode = old_mode;
1156 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1159 *flags |= LDLM_FL_BLOCK_GRANTED;
1164 l_unlock(&ns->ns_lock);
1167 ldlm_run_ast_work(ns, &rpc_list);
1171 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1173 char str[PTL_NALFMT_SIZE];
1174 struct obd_device *obd = NULL;
1176 if (!((portal_debug | D_ERROR) & level))
1180 CDEBUG(level, " NULL LDLM lock\n");
1184 CDEBUG(level, " -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d)\n",
1185 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1187 if (lock->l_conn_export != NULL)
1188 obd = lock->l_conn_export->exp_obd;
1189 if (lock->l_export && lock->l_export->exp_connection) {
1190 CDEBUG(level, " Node: NID %s on %s (rhandle: "LPX64")\n",
1191 ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str),
1192 lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1193 lock->l_remote_handle.cookie);
1194 } else if (obd == NULL) {
1195 CDEBUG(level, " Node: local\n");
1197 struct obd_import *imp = obd->u.cli.cl_import;
1198 CDEBUG(level, " Node: NID %s on %s (rhandle: "LPX64")\n",
1199 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1200 imp->imp_connection->c_peer.peer_ni->pni_name,
1201 lock->l_remote_handle.cookie);
1203 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource,
1204 lock->l_resource->lr_name.name[0],
1205 lock->l_resource->lr_name.name[1]);
1206 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1207 "write: %d\n", ldlm_lockname[lock->l_req_mode],
1208 ldlm_lockname[lock->l_granted_mode],
1209 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers);
1210 if (lock->l_resource->lr_type == LDLM_EXTENT)
1211 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1212 " (req "LPU64"-"LPU64")\n",
1213 lock->l_policy_data.l_extent.start,
1214 lock->l_policy_data.l_extent.end,
1215 lock->l_req_extent.start, lock->l_req_extent.end);
1216 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1217 CDEBUG(level, " Pid: "LPU64" Extent: "LPU64" -> "LPU64"\n",
1218 lock->l_policy_data.l_flock.pid,
1219 lock->l_policy_data.l_flock.start,
1220 lock->l_policy_data.l_flock.end);
1221 else if (lock->l_resource->lr_type == LDLM_IBITS)
1222 CDEBUG(level, " Bits: "LPX64"\n",
1223 lock->l_policy_data.l_inodebits.bits);
1226 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1228 struct ldlm_lock *lock;
1230 lock = ldlm_handle2lock(lockh);
1234 ldlm_lock_dump(D_OTHER, lock, 0);
1236 LDLM_LOCK_PUT(lock);