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 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
166 list_del_init(&lock->l_lru);
167 lock->l_resource->lr_namespace->ns_nr_unused--;
168 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
170 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
174 /* This used to have a 'strict' flact, which recovery would use to mark an
175 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
176 * shall explain why it's gone: with the new hash table scheme, once you call
177 * ldlm_lock_destroy, you can never drop your final references on this lock.
178 * Because it's not in the hash table anymore. -phil */
179 void ldlm_lock_destroy(struct ldlm_lock *lock)
182 l_lock(&lock->l_resource->lr_namespace->ns_lock);
184 if (!list_empty(&lock->l_children)) {
185 LDLM_ERROR(lock, "still has children (%p)!",
186 lock->l_children.next);
187 ldlm_lock_dump(D_ERROR, lock, 0);
190 if (lock->l_readers || lock->l_writers) {
191 LDLM_ERROR(lock, "lock still has references");
192 ldlm_lock_dump(D_ERROR, lock, 0);
196 if (!list_empty(&lock->l_res_link)) {
197 LDLM_ERROR(lock, "lock still on resource");
198 ldlm_lock_dump(D_ERROR, lock, 0);
202 if (lock->l_destroyed) {
203 LASSERT(list_empty(&lock->l_lru));
204 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
208 lock->l_destroyed = 1;
210 list_del_init(&lock->l_export_chain);
211 ldlm_lock_remove_from_lru(lock);
212 class_handle_unhash(&lock->l_handle);
215 /* Wake anyone waiting for this lock */
216 /* FIXME: I should probably add yet another flag, instead of using
217 * l_export to only call this on clients */
219 class_export_put(lock->l_export);
220 lock->l_export = NULL;
221 if (lock->l_export && lock->l_completion_ast)
222 lock->l_completion_ast(lock, 0);
225 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
230 /* this is called by portals_handle2object with the handle lock taken */
231 static void lock_handle_addref(void *lock)
233 LDLM_LOCK_GET((struct ldlm_lock *)lock);
237 * usage: pass in a resource on which you have done ldlm_resource_get
238 * pass in a parent lock on which you have done a ldlm_lock_get
239 * after return, ldlm_*_put the resource and parent
240 * returns: lock with refcount 2 - one for current caller and one for remote
242 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
243 struct ldlm_resource *resource)
245 struct ldlm_lock *lock;
248 if (resource == NULL)
251 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, SLAB_NOFS, sizeof(*lock));
255 lock->l_resource = ldlm_resource_getref(resource);
257 atomic_set(&lock->l_refc, 2);
258 INIT_LIST_HEAD(&lock->l_children);
259 INIT_LIST_HEAD(&lock->l_res_link);
260 INIT_LIST_HEAD(&lock->l_lru);
261 INIT_LIST_HEAD(&lock->l_export_chain);
262 INIT_LIST_HEAD(&lock->l_pending_chain);
263 INIT_LIST_HEAD(&lock->l_tmp);
264 init_waitqueue_head(&lock->l_waitq);
266 spin_lock(&resource->lr_namespace->ns_counter_lock);
267 resource->lr_namespace->ns_locks++;
268 spin_unlock(&resource->lr_namespace->ns_counter_lock);
270 if (parent != NULL) {
271 l_lock(&parent->l_resource->lr_namespace->ns_lock);
272 lock->l_parent = LDLM_LOCK_GET(parent);
273 list_add(&lock->l_childof, &parent->l_children);
274 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
277 INIT_LIST_HEAD(&lock->l_handle.h_link);
278 class_handle_hash(&lock->l_handle, lock_handle_addref);
283 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
284 struct ldlm_res_id new_resid)
286 struct ldlm_resource *oldres = lock->l_resource;
289 l_lock(&ns->ns_lock);
290 if (memcmp(&new_resid, &lock->l_resource->lr_name,
291 sizeof(lock->l_resource->lr_name)) == 0) {
293 l_unlock(&ns->ns_lock);
297 LASSERT(new_resid.name[0] != 0);
299 /* This function assumes that the lock isn't on any lists */
300 LASSERT(list_empty(&lock->l_res_link));
302 lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
303 lock->l_resource->lr_type,
305 if (lock->l_resource == NULL) {
310 /* ...and the flowers are still standing! */
311 ldlm_resource_putref(oldres);
313 l_unlock(&ns->ns_lock);
321 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
323 lockh->cookie = lock->l_handle.h_cookie;
326 /* if flags: atomically get the lock and set the flags.
327 * Return NULL if flag already set
330 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
332 struct ldlm_namespace *ns;
333 struct ldlm_lock *lock = NULL, *retval = NULL;
338 lock = class_handle2object(handle->cookie);
342 LASSERT(lock->l_resource != NULL);
343 ns = lock->l_resource->lr_namespace;
346 l_lock(&ns->ns_lock);
348 /* It's unlikely but possible that someone marked the lock as
349 * destroyed after we did handle2object on it */
350 if (lock->l_destroyed) {
351 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
356 if (flags && (lock->l_flags & flags)) {
362 lock->l_flags |= flags;
367 l_unlock(&ns->ns_lock);
371 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
372 struct lustre_handle *handle)
374 struct ldlm_lock *retval = NULL;
376 l_lock(&ns->ns_lock);
377 retval = __ldlm_handle2lock(handle, 0);
378 l_unlock(&ns->ns_lock);
383 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
385 ldlm_res2desc(lock->l_resource, &desc->l_resource);
386 desc->l_req_mode = lock->l_req_mode;
387 desc->l_granted_mode = lock->l_granted_mode;
388 memcpy(&desc->l_policy_data, &lock->l_policy_data,
389 sizeof(desc->l_policy_data));
392 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
393 void *data, int datalen)
395 struct ldlm_ast_work *w;
398 l_lock(&lock->l_resource->lr_namespace->ns_lock);
399 if (new && (lock->l_flags & LDLM_FL_AST_SENT))
402 CDEBUG(D_OTHER, "lock %p incompatible; sending blocking AST.\n", lock);
404 OBD_ALLOC(w, sizeof(*w));
411 w->w_datalen = datalen;
413 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
414 lock->l_flags |= LDLM_FL_AST_SENT;
415 /* If the enqueuing client said so, tell the AST recipient to
416 * discard dirty data, rather than writing back. */
417 if (new->l_flags & LDLM_AST_DISCARD_DATA)
418 lock->l_flags |= LDLM_FL_DISCARD_DATA;
420 ldlm_lock2desc(new, &w->w_desc);
423 w->w_lock = LDLM_LOCK_GET(lock);
424 list_add(&w->w_list, lock->l_resource->lr_tmp);
427 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
430 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
432 struct ldlm_lock *lock;
434 lock = ldlm_handle2lock(lockh);
435 ldlm_lock_addref_internal(lock, mode);
439 /* only called for local locks */
440 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
442 l_lock(&lock->l_resource->lr_namespace->ns_lock);
443 ldlm_lock_remove_from_lru(lock);
444 if (mode & (LCK_NL | LCK_CR | LCK_PR))
446 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
448 lock->l_last_used = jiffies;
450 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
451 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
454 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
456 struct ldlm_namespace *ns;
459 ns = lock->l_resource->lr_namespace;
461 l_lock(&ns->ns_lock);
462 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
463 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
464 LASSERT(lock->l_readers > 0);
467 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
468 LASSERT(lock->l_writers > 0);
472 if (lock->l_flags & LDLM_FL_LOCAL &&
473 !lock->l_readers && !lock->l_writers) {
474 /* If this is a local lock on a server namespace and this was
475 * the last reference, cancel the lock. */
476 CDEBUG(D_INFO, "forcing cancel of local lock\n");
477 lock->l_flags |= LDLM_FL_CBPENDING;
480 if (!lock->l_readers && !lock->l_writers &&
481 (lock->l_flags & LDLM_FL_CBPENDING)) {
482 /* If we received a blocked AST and this was the last reference,
483 * run the callback. */
484 if (ns->ns_client == LDLM_NAMESPACE_SERVER && lock->l_export)
485 CERROR("FL_CBPENDING set on non-local lock--just a "
488 LDLM_DEBUG(lock, "final decref done on cbpending lock");
490 LDLM_LOCK_GET(lock); /* dropped by bl thread */
491 ldlm_lock_remove_from_lru(lock);
492 l_unlock(&ns->ns_lock);
493 if (ldlm_bl_to_thread(ns, NULL, lock) != 0)
494 ldlm_handle_bl_callback(ns, NULL, lock);
495 } else if (ns->ns_client == LDLM_NAMESPACE_CLIENT &&
496 !lock->l_readers && !lock->l_writers) {
497 /* If this is a client-side namespace and this was the last
498 * reference, put it on the LRU. */
499 LASSERT(list_empty(&lock->l_lru));
500 LASSERT(ns->ns_nr_unused >= 0);
501 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
503 l_unlock(&ns->ns_lock);
504 ldlm_cancel_lru(ns, LDLM_ASYNC);
506 l_unlock(&ns->ns_lock);
509 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
514 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
516 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
517 LASSERT(lock != NULL);
518 ldlm_lock_decref_internal(lock, mode);
522 /* This will drop a lock reference and mark it for destruction, but will not
523 * necessarily cancel the lock before returning. */
524 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
526 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
529 LASSERT(lock != NULL);
531 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
532 l_lock(&lock->l_resource->lr_namespace->ns_lock);
533 lock->l_flags |= LDLM_FL_CBPENDING;
534 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
535 ldlm_lock_decref_internal(lock, mode);
540 * - ldlm_lock_enqueue
541 * - ldlm_reprocess_queue
542 * - ldlm_lock_convert
544 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen,
547 struct ldlm_resource *res = lock->l_resource;
550 l_lock(&lock->l_resource->lr_namespace->ns_lock);
551 lock->l_granted_mode = lock->l_req_mode;
552 ldlm_resource_add_lock(res, &res->lr_granted, lock);
554 if (lock->l_granted_mode < res->lr_most_restr)
555 res->lr_most_restr = lock->l_granted_mode;
557 if (run_ast && lock->l_completion_ast != NULL)
558 ldlm_add_ast_work_item(lock, NULL, data, datalen);
560 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
564 /* returns a referenced lock or NULL. See the flag descriptions below, in the
565 * comment above ldlm_lock_match */
566 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
567 ldlm_policy_data_t *policy,
568 struct ldlm_lock *old_lock, int flags)
570 struct ldlm_lock *lock;
571 struct list_head *tmp;
573 list_for_each(tmp, queue) {
574 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
576 if (lock == old_lock)
579 /* llite sometimes wants to match locks that will be
580 * canceled when their users drop, but we allow it to match
581 * if it passes in CBPENDING and the lock still has users.
582 * this is generally only going to be used by children
583 * whose parents already hold a lock so forward progress
584 * can still happen. */
585 if (lock->l_flags & LDLM_FL_CBPENDING &&
586 !(flags & LDLM_FL_CBPENDING))
588 if (lock->l_flags & LDLM_FL_CBPENDING &&
589 lock->l_readers == 0 && lock->l_writers == 0)
592 if (!(lock->l_req_mode & mode))
595 if (lock->l_resource->lr_type == LDLM_EXTENT &&
596 (lock->l_policy_data.l_extent.start >
597 policy->l_extent.start ||
598 lock->l_policy_data.l_extent.end < policy->l_extent.end))
601 if (lock->l_resource->lr_type == LDLM_EXTENT &&
603 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
606 /* We match if we have existing lock with same or wider set
608 if (lock->l_resource->lr_type == LDLM_IBITS &&
609 ((lock->l_policy_data.l_inodebits.bits &
610 policy->l_inodebits.bits) !=
611 policy->l_inodebits.bits))
614 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
617 if ((flags & LDLM_FL_LOCAL_ONLY) &&
618 !(lock->l_flags & LDLM_FL_LOCAL))
621 if (flags & LDLM_FL_TEST_LOCK)
624 ldlm_lock_addref_internal(lock, mode);
631 void ldlm_lock_allow_match(struct ldlm_lock *lock)
633 l_lock(&lock->l_resource->lr_namespace->ns_lock);
634 lock->l_flags |= LDLM_FL_CAN_MATCH;
635 wake_up(&lock->l_waitq);
636 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
639 /* Can be called in two ways:
641 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
642 * for a duplicate of.
644 * Otherwise, all of the fields must be filled in, to match against.
646 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
647 * server (ie, connh is NULL)
648 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
649 * list will be considered
650 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
651 * to be canceled can still be matched as long as they still have reader
652 * or writer refernces
653 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
654 * just tell us if we would have matched.
656 * Returns 1 if it finds an already-existing lock that is compatible; in this
657 * case, lockh is filled in with a addref()ed lock
659 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
660 struct ldlm_res_id *res_id, __u32 type,
661 ldlm_policy_data_t *policy, ldlm_mode_t mode,
662 struct lustre_handle *lockh)
664 struct ldlm_resource *res;
665 struct ldlm_lock *lock, *old_lock = NULL;
670 old_lock = ldlm_handle2lock(lockh);
673 ns = old_lock->l_resource->lr_namespace;
674 res_id = &old_lock->l_resource->lr_name;
675 type = old_lock->l_resource->lr_type;
676 mode = old_lock->l_req_mode;
679 res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
681 LASSERT(old_lock == NULL);
685 l_lock(&ns->ns_lock);
687 lock = search_queue(&res->lr_granted, mode, policy, old_lock, flags);
690 if (flags & LDLM_FL_BLOCK_GRANTED)
692 lock = search_queue(&res->lr_converting, mode, policy, old_lock, flags);
695 lock = search_queue(&res->lr_waiting, mode, policy, old_lock, flags);
701 ldlm_resource_putref(res);
702 l_unlock(&ns->ns_lock);
705 ldlm_lock2handle(lock, lockh);
706 if (!(lock->l_flags & LDLM_FL_CAN_MATCH)) {
707 struct l_wait_info lwi;
708 if (lock->l_completion_ast) {
709 int err = lock->l_completion_ast(lock,
710 LDLM_FL_WAIT_NOREPROC,
718 lwi = LWI_TIMEOUT_INTR(obd_timeout*HZ, NULL,NULL,NULL);
720 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
721 l_wait_event(lock->l_waitq,
722 (lock->l_flags & LDLM_FL_CAN_MATCH), &lwi);
728 l_lock(&ns->ns_lock);
729 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
730 type == LDLM_PLAIN ? res_id->name[2] :
731 policy->l_extent.start,
732 type == LDLM_PLAIN ? res_id->name[3] :
733 policy->l_extent.end);
734 l_unlock(&ns->ns_lock);
735 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/* less verbose for test-only */
736 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
737 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
738 type, mode, res_id->name[0], res_id->name[1],
739 type == LDLM_PLAIN ? res_id->name[2] :
740 policy->l_extent.start,
741 type == LDLM_PLAIN ? res_id->name[3] :
742 policy->l_extent.end);
746 LDLM_LOCK_PUT(old_lock);
747 if (flags & LDLM_FL_TEST_LOCK && rc)
753 /* Returns a referenced lock */
754 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
755 struct lustre_handle *parent_lock_handle,
756 struct ldlm_res_id res_id, __u32 type,
758 ldlm_blocking_callback blocking,
759 ldlm_completion_callback completion,
760 ldlm_glimpse_callback glimpse,
761 void *data, __u32 lvb_len)
763 struct ldlm_resource *res, *parent_res = NULL;
764 struct ldlm_lock *lock, *parent_lock = NULL;
767 if (parent_lock_handle) {
768 parent_lock = ldlm_handle2lock(parent_lock_handle);
770 parent_res = parent_lock->l_resource;
773 res = ldlm_resource_get(ns, parent_res, res_id,
778 lock = ldlm_lock_new(parent_lock, res);
779 ldlm_resource_putref(res);
780 if (parent_lock != NULL)
781 LDLM_LOCK_PUT(parent_lock);
786 lock->l_req_mode = mode;
787 lock->l_ast_data = data;
788 lock->l_blocking_ast = blocking;
789 lock->l_completion_ast = completion;
790 lock->l_glimpse_ast = glimpse;
791 lock->l_pid = current->pid;
794 lock->l_lvb_len = lvb_len;
795 OBD_ALLOC(lock->l_lvb_data, lvb_len);
796 if (lock->l_lvb_data == NULL) {
797 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
805 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
806 struct ldlm_lock **lockp,
807 void *cookie, int *flags)
809 struct ldlm_lock *lock = *lockp;
810 struct ldlm_resource *res = lock->l_resource;
811 int local = res->lr_namespace->ns_client;
812 ldlm_processing_policy policy;
813 ldlm_error_t rc = ELDLM_OK;
816 /* policies are not executed on the client or during replay */
817 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
818 && !local && ns->ns_policy) {
819 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
821 if (rc == ELDLM_LOCK_REPLACED) {
822 /* The lock that was returned has already been granted,
823 * and placed into lockp. If it's not the same as the
824 * one we passed in, then destroy the old one and our
825 * work here is done. */
826 if (lock != *lockp) {
827 ldlm_lock_destroy(lock);
830 *flags |= LDLM_FL_LOCK_CHANGED;
832 } else if (rc == ELDLM_LOCK_ABORTED ||
833 (rc == 0 && (*flags & LDLM_FL_INTENT_ONLY))) {
834 ldlm_lock_destroy(lock);
837 LASSERT(rc == ELDLM_OK);
840 l_lock(&ns->ns_lock);
841 if (local && lock->l_req_mode == lock->l_granted_mode) {
842 /* The server returned a blocked lock, but it was granted before
843 * we got a chance to actually enqueue it. We don't need to do
845 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
846 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
850 /* Some flags from the enqueue want to make it into the AST, via the
852 lock->l_flags |= (*flags & LDLM_AST_DISCARD_DATA);
854 /* This distinction between local lock trees is very important; a client
855 * namespace only has information about locks taken by that client, and
856 * thus doesn't have enough information to decide for itself if it can
857 * be granted (below). In this case, we do exactly what the server
858 * tells us to do, as dictated by the 'flags'.
860 * We do exactly the same thing during recovery, when the server is
861 * more or less trusting the clients not to lie.
863 * FIXME (bug 268): Detect obvious lies by checking compatibility in
864 * granted/converting queues. */
865 ldlm_resource_unlink_lock(lock);
867 if (*flags & LDLM_FL_BLOCK_CONV)
868 ldlm_resource_add_lock(res, &res->lr_converting, lock);
869 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
870 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
872 ldlm_grant_lock(lock, NULL, 0, 0);
874 } else if (*flags & LDLM_FL_REPLAY) {
875 if (*flags & LDLM_FL_BLOCK_CONV) {
876 ldlm_resource_add_lock(res, &res->lr_converting, lock);
878 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
879 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
881 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
882 ldlm_grant_lock(lock, NULL, 0, 0);
885 /* If no flags, fall through to normal enqueue path. */
888 policy = ldlm_processing_policy_table[res->lr_type];
889 policy(lock, flags, 1, &rc);
892 l_unlock(&ns->ns_lock);
896 /* Must be called with namespace taken: queue is waiting or converting. */
897 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue)
899 struct list_head *tmp, *pos;
900 ldlm_processing_policy policy;
902 int rc = LDLM_ITER_CONTINUE;
906 policy = ldlm_processing_policy_table[res->lr_type];
909 list_for_each_safe(tmp, pos, queue) {
910 struct ldlm_lock *pending;
911 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
913 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
916 rc = policy(pending, &flags, 0, &err);
917 if (rc != LDLM_ITER_CONTINUE)
924 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list)
926 struct list_head *tmp, *pos;
930 l_check_no_ns_lock(ns);
932 list_for_each_safe(tmp, pos, rpc_list) {
933 struct ldlm_ast_work *w =
934 list_entry(tmp, struct ldlm_ast_work, w_list);
936 /* It's possible to receive a completion AST before we've set
937 * the l_completion_ast pointer: either because the AST arrived
938 * before the reply, or simply because there's a small race
939 * window between receiving the reply and finishing the local
942 * This can't happen with the blocking_ast, however, because we
943 * will never call the local blocking_ast until we drop our
944 * reader/writer reference, which we won't do until we get the
945 * reply and finish enqueueing. */
946 LASSERT(w->w_lock != NULL);
948 LASSERT(w->w_lock->l_blocking_ast != NULL);
949 rc = w->w_lock->l_blocking_ast
950 (w->w_lock, &w->w_desc, w->w_data,
952 } else if (w->w_lock->l_completion_ast != NULL) {
953 LASSERT(w->w_lock->l_completion_ast != NULL);
954 rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags,
962 CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
963 "disconnect client\n");
964 LDLM_LOCK_PUT(w->w_lock);
965 list_del(&w->w_list);
966 OBD_FREE(w, sizeof(*w));
971 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
973 ldlm_reprocess_all(res);
974 return LDLM_ITER_CONTINUE;
977 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
981 l_lock(&ns->ns_lock);
982 for (i = 0; i < RES_HASH_SIZE; i++) {
983 struct list_head *tmp, *next;
984 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
985 struct ldlm_resource *res =
986 list_entry(tmp, struct ldlm_resource, lr_hash);
988 ldlm_resource_getref(res);
989 l_unlock(&ns->ns_lock);
990 rc = reprocess_one_queue(res, NULL);
991 l_lock(&ns->ns_lock);
993 ldlm_resource_putref(res);
994 if (rc == LDLM_ITER_STOP)
999 l_unlock(&ns->ns_lock);
1003 void ldlm_reprocess_all(struct ldlm_resource *res)
1005 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1009 /* Local lock trees don't get reprocessed. */
1010 if (res->lr_namespace->ns_client) {
1016 l_lock(&res->lr_namespace->ns_lock);
1017 res->lr_tmp = &rpc_list;
1019 rc = ldlm_reprocess_queue(res, &res->lr_converting);
1020 if (rc == LDLM_ITER_CONTINUE)
1021 ldlm_reprocess_queue(res, &res->lr_waiting);
1024 l_unlock(&res->lr_namespace->ns_lock);
1026 rc = ldlm_run_ast_work(res->lr_namespace, &rpc_list);
1027 if (rc == -ERESTART) {
1028 LASSERT(list_empty(&rpc_list));
1034 void ldlm_cancel_callback(struct ldlm_lock *lock)
1036 l_lock(&lock->l_resource->lr_namespace->ns_lock);
1037 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1038 lock->l_flags |= LDLM_FL_CANCEL;
1039 if (lock->l_blocking_ast) {
1040 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1041 // l_check_no_ns_lock(lock->l_resource->lr_namespace);
1042 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1046 LDLM_DEBUG(lock, "no blocking ast");
1049 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1052 void ldlm_lock_cancel(struct ldlm_lock *lock)
1054 struct ldlm_resource *res;
1055 struct ldlm_namespace *ns;
1058 res = lock->l_resource;
1059 ns = res->lr_namespace;
1061 l_lock(&ns->ns_lock);
1062 ldlm_del_waiting_lock(lock);
1064 /* Please do not, no matter how tempting, remove this LBUG without
1065 * talking to me first. -phik */
1066 if (lock->l_readers || lock->l_writers) {
1067 LDLM_ERROR(lock, "lock still has references");
1071 ldlm_cancel_callback(lock);
1073 ldlm_resource_unlink_lock(lock);
1074 ldlm_lock_destroy(lock);
1075 l_unlock(&ns->ns_lock);
1079 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1081 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1087 lock->l_ast_data = data;
1088 LDLM_LOCK_PUT(lock);
1092 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1094 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1095 struct ldlm_lock *lock;
1096 struct ldlm_resource *res;
1098 l_lock(&ns->ns_lock);
1099 while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1100 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1101 struct ldlm_lock, l_export_chain);
1102 res = ldlm_resource_getref(lock->l_resource);
1103 LDLM_DEBUG(lock, "export %p", exp);
1104 ldlm_lock_cancel(lock);
1105 l_unlock(&ns->ns_lock);
1106 ldlm_reprocess_all(res);
1107 ldlm_resource_putref(res);
1108 l_lock(&ns->ns_lock);
1110 l_unlock(&ns->ns_lock);
1113 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1116 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1117 struct ldlm_resource *res;
1118 struct ldlm_namespace *ns;
1124 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1125 *flags |= LDLM_FL_BLOCK_GRANTED;
1126 RETURN(lock->l_resource);
1129 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1130 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1132 res = lock->l_resource;
1133 ns = res->lr_namespace;
1135 l_lock(&ns->ns_lock);
1137 old_mode = lock->l_req_mode;
1138 lock->l_req_mode = new_mode;
1139 ldlm_resource_unlink_lock(lock);
1141 /* If this is a local resource, put it on the appropriate list. */
1142 if (res->lr_namespace->ns_client) {
1143 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1144 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1146 /* This should never happen, because of the way the
1147 * server handles conversions. */
1148 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1152 res->lr_tmp = &rpc_list;
1153 ldlm_grant_lock(lock, NULL, 0, 0);
1156 /* FIXME: completion handling not with ns_lock held ! */
1157 if (lock->l_completion_ast)
1158 lock->l_completion_ast(lock, 0, NULL);
1162 ldlm_processing_policy policy;
1163 policy = ldlm_processing_policy_table[res->lr_type];
1164 res->lr_tmp = &rpc_list;
1165 rc = policy(lock, &pflags, 0, &err);
1167 if (rc == LDLM_ITER_STOP) {
1168 lock->l_req_mode = old_mode;
1169 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1172 *flags |= LDLM_FL_BLOCK_GRANTED;
1177 l_unlock(&ns->ns_lock);
1180 ldlm_run_ast_work(ns, &rpc_list);
1184 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1186 char str[PTL_NALFMT_SIZE];
1187 struct obd_device *obd = NULL;
1189 if (!((portal_debug | D_ERROR) & level))
1193 CDEBUG(level, " NULL LDLM lock\n");
1197 CDEBUG(level, " -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1198 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1200 if (lock->l_conn_export != NULL)
1201 obd = lock->l_conn_export->exp_obd;
1202 if (lock->l_export && lock->l_export->exp_connection) {
1203 CDEBUG(level, " Node: NID %s on %s (rhandle: "LPX64")\n",
1204 ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str),
1205 lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1206 lock->l_remote_handle.cookie);
1207 } else if (obd == NULL) {
1208 CDEBUG(level, " Node: local\n");
1210 struct obd_import *imp = obd->u.cli.cl_import;
1211 CDEBUG(level, " Node: NID %s on %s (rhandle: "LPX64")\n",
1212 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1213 imp->imp_connection->c_peer.peer_ni->pni_name,
1214 lock->l_remote_handle.cookie);
1216 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource,
1217 lock->l_resource->lr_name.name[0],
1218 lock->l_resource->lr_name.name[1]);
1219 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1220 "write: %d\n", ldlm_lockname[lock->l_req_mode],
1221 ldlm_lockname[lock->l_granted_mode],
1222 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers);
1223 if (lock->l_resource->lr_type == LDLM_EXTENT)
1224 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1225 " (req "LPU64"-"LPU64")\n",
1226 lock->l_policy_data.l_extent.start,
1227 lock->l_policy_data.l_extent.end,
1228 lock->l_req_extent.start, lock->l_req_extent.end);
1229 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1230 CDEBUG(level, " Pid: "LPU64" Extent: "LPU64" -> "LPU64"\n",
1231 lock->l_policy_data.l_flock.pid,
1232 lock->l_policy_data.l_flock.start,
1233 lock->l_policy_data.l_flock.end);
1234 else if (lock->l_resource->lr_type == LDLM_IBITS)
1235 CDEBUG(level, " Bits: "LPX64"\n",
1236 lock->l_policy_data.l_inodebits.bits);
1239 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1241 struct ldlm_lock *lock;
1243 lock = ldlm_handle2lock(lockh);
1247 ldlm_lock_dump(D_OTHER, lock, 0);
1249 LDLM_LOCK_PUT(lock);