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, 1);
304 if (lock->l_resource == NULL) {
309 /* ...and the flowers are still standing! */
310 ldlm_resource_putref(oldres);
312 l_unlock(&ns->ns_lock);
320 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
322 lockh->cookie = lock->l_handle.h_cookie;
325 /* if flags: atomically get the lock and set the flags.
326 * Return NULL if flag already set
329 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
331 struct ldlm_namespace *ns;
332 struct ldlm_lock *lock = NULL, *retval = NULL;
337 lock = class_handle2object(handle->cookie);
341 LASSERT(lock->l_resource != NULL);
342 ns = lock->l_resource->lr_namespace;
345 l_lock(&ns->ns_lock);
347 /* It's unlikely but possible that someone marked the lock as
348 * destroyed after we did handle2object on it */
349 if (lock->l_destroyed) {
350 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
355 if (flags && (lock->l_flags & flags)) {
361 lock->l_flags |= flags;
366 l_unlock(&ns->ns_lock);
370 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
371 struct lustre_handle *handle)
373 struct ldlm_lock *retval = NULL;
375 l_lock(&ns->ns_lock);
376 retval = __ldlm_handle2lock(handle, 0);
377 l_unlock(&ns->ns_lock);
382 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
384 ldlm_res2desc(lock->l_resource, &desc->l_resource);
385 desc->l_req_mode = lock->l_req_mode;
386 desc->l_granted_mode = lock->l_granted_mode;
387 memcpy(&desc->l_policy_data, &lock->l_policy_data,
388 sizeof(desc->l_policy_data));
391 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
392 void *data, int datalen)
394 struct ldlm_ast_work *w;
397 l_lock(&lock->l_resource->lr_namespace->ns_lock);
398 if (new && (lock->l_flags & LDLM_FL_AST_SENT))
401 CDEBUG(D_OTHER, "lock %p incompatible; sending blocking AST.\n", lock);
403 OBD_ALLOC(w, sizeof(*w));
410 w->w_datalen = datalen;
412 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
413 lock->l_flags |= LDLM_FL_AST_SENT;
414 /* If the enqueuing client said so, tell the AST recipient to
415 * discard dirty data, rather than writing back. */
416 if (new->l_flags & LDLM_AST_DISCARD_DATA)
417 lock->l_flags |= LDLM_FL_DISCARD_DATA;
419 ldlm_lock2desc(new, &w->w_desc);
422 w->w_lock = LDLM_LOCK_GET(lock);
423 list_add(&w->w_list, lock->l_resource->lr_tmp);
426 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
429 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
431 struct ldlm_lock *lock;
433 lock = ldlm_handle2lock(lockh);
434 ldlm_lock_addref_internal(lock, mode);
438 /* only called for local locks */
439 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
441 l_lock(&lock->l_resource->lr_namespace->ns_lock);
442 ldlm_lock_remove_from_lru(lock);
443 if (mode & (LCK_NL | LCK_CR | LCK_PR))
445 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
447 lock->l_last_used = jiffies;
449 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
450 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
453 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
455 struct ldlm_namespace *ns;
458 ns = lock->l_resource->lr_namespace;
460 l_lock(&ns->ns_lock);
461 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
462 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
463 LASSERT(lock->l_readers > 0);
466 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
467 LASSERT(lock->l_writers > 0);
471 if (lock->l_flags & LDLM_FL_LOCAL &&
472 !lock->l_readers && !lock->l_writers) {
473 /* If this is a local lock on a server namespace and this was
474 * the last reference, cancel the lock. */
475 CDEBUG(D_INFO, "forcing cancel of local lock\n");
476 lock->l_flags |= LDLM_FL_CBPENDING;
479 if (!lock->l_readers && !lock->l_writers &&
480 (lock->l_flags & LDLM_FL_CBPENDING)) {
481 /* If we received a blocked AST and this was the last reference,
482 * run the callback. */
483 if (ns->ns_client == LDLM_NAMESPACE_SERVER && lock->l_export)
484 CERROR("FL_CBPENDING set on non-local lock--just a "
487 LDLM_DEBUG(lock, "final decref done on cbpending lock");
489 LDLM_LOCK_GET(lock); /* dropped by bl thread */
490 ldlm_lock_remove_from_lru(lock);
491 l_unlock(&ns->ns_lock);
492 if (ldlm_bl_to_thread(ns, NULL, lock) != 0)
493 ldlm_handle_bl_callback(ns, NULL, lock);
494 } else if (ns->ns_client == LDLM_NAMESPACE_CLIENT &&
495 !lock->l_readers && !lock->l_writers) {
496 /* If this is a client-side namespace and this was the last
497 * reference, put it on the LRU. */
498 LASSERT(list_empty(&lock->l_lru));
499 LASSERT(ns->ns_nr_unused >= 0);
500 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
502 l_unlock(&ns->ns_lock);
503 ldlm_cancel_lru(ns, LDLM_ASYNC);
505 l_unlock(&ns->ns_lock);
508 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
513 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
515 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
516 LASSERT(lock != NULL);
517 ldlm_lock_decref_internal(lock, mode);
521 /* This will drop a lock reference and mark it for destruction, but will not
522 * necessarily cancel the lock before returning. */
523 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
525 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
528 LASSERT(lock != NULL);
530 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
531 l_lock(&lock->l_resource->lr_namespace->ns_lock);
532 lock->l_flags |= LDLM_FL_CBPENDING;
533 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
534 ldlm_lock_decref_internal(lock, mode);
539 * - ldlm_lock_enqueue
540 * - ldlm_reprocess_queue
541 * - ldlm_lock_convert
543 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen,
546 struct ldlm_resource *res = lock->l_resource;
549 l_lock(&lock->l_resource->lr_namespace->ns_lock);
550 lock->l_granted_mode = lock->l_req_mode;
551 ldlm_resource_add_lock(res, &res->lr_granted, lock);
553 if (lock->l_granted_mode < res->lr_most_restr)
554 res->lr_most_restr = lock->l_granted_mode;
556 if (run_ast && lock->l_completion_ast != NULL)
557 ldlm_add_ast_work_item(lock, NULL, data, datalen);
559 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
563 /* returns a referenced lock or NULL. See the flag descriptions below, in the
564 * comment above ldlm_lock_match */
565 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
566 ldlm_policy_data_t *policy,
567 struct ldlm_lock *old_lock, int flags)
569 struct ldlm_lock *lock;
570 struct list_head *tmp;
572 list_for_each(tmp, queue) {
573 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
575 if (lock == old_lock)
578 /* llite sometimes wants to match locks that will be
579 * canceled when their users drop, but we allow it to match
580 * if it passes in CBPENDING and the lock still has users.
581 * this is generally only going to be used by children
582 * whose parents already hold a lock so forward progress
583 * can still happen. */
584 if (lock->l_flags & LDLM_FL_CBPENDING &&
585 !(flags & LDLM_FL_CBPENDING))
587 if (lock->l_flags & LDLM_FL_CBPENDING &&
588 lock->l_readers == 0 && lock->l_writers == 0)
591 if (!(lock->l_req_mode & mode))
594 if (lock->l_resource->lr_type == LDLM_EXTENT &&
595 (lock->l_policy_data.l_extent.start >
596 policy->l_extent.start ||
597 lock->l_policy_data.l_extent.end < policy->l_extent.end))
600 if (lock->l_resource->lr_type == LDLM_EXTENT &&
602 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
605 /* We match if we have existing lock with same or wider set
607 if (lock->l_resource->lr_type == LDLM_IBITS &&
608 ((lock->l_policy_data.l_inodebits.bits &
609 policy->l_inodebits.bits) !=
610 policy->l_inodebits.bits))
613 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
616 if ((flags & LDLM_FL_LOCAL_ONLY) &&
617 !(lock->l_flags & LDLM_FL_LOCAL))
620 if (flags & LDLM_FL_TEST_LOCK)
623 ldlm_lock_addref_internal(lock, mode);
630 void ldlm_lock_allow_match(struct ldlm_lock *lock)
632 l_lock(&lock->l_resource->lr_namespace->ns_lock);
633 lock->l_flags |= LDLM_FL_CAN_MATCH;
634 wake_up(&lock->l_waitq);
635 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
638 /* Can be called in two ways:
640 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
641 * for a duplicate of.
643 * Otherwise, all of the fields must be filled in, to match against.
645 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
646 * server (ie, connh is NULL)
647 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
648 * list will be considered
649 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
650 * to be canceled can still be matched as long as they still have reader
651 * or writer refernces
652 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
653 * just tell us if we would have matched.
655 * Returns 1 if it finds an already-existing lock that is compatible; in this
656 * case, lockh is filled in with a addref()ed lock
658 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
659 struct ldlm_res_id *res_id, __u32 type,
660 ldlm_policy_data_t *policy, ldlm_mode_t mode,
661 struct lustre_handle *lockh)
663 struct ldlm_resource *res;
664 struct ldlm_lock *lock, *old_lock = NULL;
669 old_lock = ldlm_handle2lock(lockh);
672 ns = old_lock->l_resource->lr_namespace;
673 res_id = &old_lock->l_resource->lr_name;
674 type = old_lock->l_resource->lr_type;
675 mode = old_lock->l_req_mode;
678 res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
680 LASSERT(old_lock == NULL);
684 l_lock(&ns->ns_lock);
686 lock = search_queue(&res->lr_granted, mode, policy, old_lock, flags);
689 if (flags & LDLM_FL_BLOCK_GRANTED)
691 lock = search_queue(&res->lr_converting, mode, policy, old_lock, flags);
694 lock = search_queue(&res->lr_waiting, mode, policy, old_lock, flags);
700 ldlm_resource_putref(res);
701 l_unlock(&ns->ns_lock);
704 ldlm_lock2handle(lock, lockh);
705 if (!(lock->l_flags & LDLM_FL_CAN_MATCH)) {
706 struct l_wait_info lwi;
707 if (lock->l_completion_ast) {
708 int err = lock->l_completion_ast(lock,
709 LDLM_FL_WAIT_NOREPROC,
717 lwi = LWI_TIMEOUT_INTR(obd_timeout*HZ, NULL,NULL,NULL);
719 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
720 l_wait_event(lock->l_waitq,
721 (lock->l_flags & LDLM_FL_CAN_MATCH), &lwi);
727 l_lock(&ns->ns_lock);
728 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
729 type == LDLM_PLAIN ? res_id->name[2] :
730 policy->l_extent.start,
731 type == LDLM_PLAIN ? res_id->name[3] :
732 policy->l_extent.end);
733 l_unlock(&ns->ns_lock);
734 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/* less verbose for test-only */
735 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
736 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
737 type, mode, res_id->name[0], res_id->name[1],
738 type == LDLM_PLAIN ? res_id->name[2] :
739 policy->l_extent.start,
740 type == LDLM_PLAIN ? res_id->name[3] :
741 policy->l_extent.end);
745 LDLM_LOCK_PUT(old_lock);
746 if (flags & LDLM_FL_TEST_LOCK && rc)
752 /* Returns a referenced lock */
753 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
754 struct lustre_handle *parent_lock_handle,
755 struct ldlm_res_id res_id, __u32 type,
757 ldlm_blocking_callback blocking,
758 ldlm_completion_callback completion,
759 ldlm_glimpse_callback glimpse,
760 void *data, __u32 lvb_len)
762 struct ldlm_resource *res, *parent_res = NULL;
763 struct ldlm_lock *lock, *parent_lock = NULL;
766 if (parent_lock_handle) {
767 parent_lock = ldlm_handle2lock(parent_lock_handle);
769 parent_res = parent_lock->l_resource;
772 res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
776 lock = ldlm_lock_new(parent_lock, res);
777 ldlm_resource_putref(res);
778 if (parent_lock != NULL)
779 LDLM_LOCK_PUT(parent_lock);
784 lock->l_req_mode = mode;
785 lock->l_ast_data = data;
786 lock->l_blocking_ast = blocking;
787 lock->l_completion_ast = completion;
788 lock->l_glimpse_ast = glimpse;
789 lock->l_pid = current->pid;
792 lock->l_lvb_len = lvb_len;
793 OBD_ALLOC(lock->l_lvb_data, lvb_len);
794 if (lock->l_lvb_data == NULL) {
795 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
803 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
804 struct ldlm_lock **lockp,
805 void *cookie, int *flags)
807 struct ldlm_lock *lock = *lockp;
808 struct ldlm_resource *res = lock->l_resource;
809 int local = res->lr_namespace->ns_client;
810 ldlm_processing_policy policy;
811 ldlm_error_t rc = ELDLM_OK;
814 /* policies are not executed on the client or during replay */
815 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
816 && !local && ns->ns_policy) {
817 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
819 if (rc == ELDLM_LOCK_REPLACED) {
820 /* The lock that was returned has already been granted,
821 * and placed into lockp. If it's not the same as the
822 * one we passed in, then destroy the old one and our
823 * work here is done. */
824 if (lock != *lockp) {
825 ldlm_lock_destroy(lock);
828 *flags |= LDLM_FL_LOCK_CHANGED;
830 } else if (rc == ELDLM_LOCK_ABORTED ||
831 (rc == 0 && (*flags & LDLM_FL_INTENT_ONLY))) {
832 ldlm_lock_destroy(lock);
835 LASSERT(rc == ELDLM_OK);
838 l_lock(&ns->ns_lock);
839 if (local && lock->l_req_mode == lock->l_granted_mode) {
840 /* The server returned a blocked lock, but it was granted before
841 * we got a chance to actually enqueue it. We don't need to do
843 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
844 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
848 /* Some flags from the enqueue want to make it into the AST, via the
850 lock->l_flags |= (*flags & LDLM_AST_DISCARD_DATA);
852 /* This distinction between local lock trees is very important; a client
853 * namespace only has information about locks taken by that client, and
854 * thus doesn't have enough information to decide for itself if it can
855 * be granted (below). In this case, we do exactly what the server
856 * tells us to do, as dictated by the 'flags'.
858 * We do exactly the same thing during recovery, when the server is
859 * more or less trusting the clients not to lie.
861 * FIXME (bug 268): Detect obvious lies by checking compatibility in
862 * granted/converting queues. */
863 ldlm_resource_unlink_lock(lock);
865 if (*flags & LDLM_FL_BLOCK_CONV)
866 ldlm_resource_add_lock(res, &res->lr_converting, lock);
867 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
868 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
870 ldlm_grant_lock(lock, NULL, 0, 0);
872 } else if (*flags & LDLM_FL_REPLAY) {
873 if (*flags & LDLM_FL_BLOCK_CONV) {
874 ldlm_resource_add_lock(res, &res->lr_converting, lock);
876 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
877 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
879 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
880 ldlm_grant_lock(lock, NULL, 0, 0);
883 /* If no flags, fall through to normal enqueue path. */
886 policy = ldlm_processing_policy_table[res->lr_type];
887 policy(lock, flags, 1, &rc);
890 l_unlock(&ns->ns_lock);
894 /* Must be called with namespace taken: queue is waiting or converting. */
895 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue)
897 struct list_head *tmp, *pos;
898 ldlm_processing_policy policy;
900 int rc = LDLM_ITER_CONTINUE;
904 policy = ldlm_processing_policy_table[res->lr_type];
907 list_for_each_safe(tmp, pos, queue) {
908 struct ldlm_lock *pending;
909 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
911 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
914 rc = policy(pending, &flags, 0, &err);
915 if (rc != LDLM_ITER_CONTINUE)
922 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list)
924 struct list_head *tmp, *pos;
928 l_check_no_ns_lock(ns);
930 list_for_each_safe(tmp, pos, rpc_list) {
931 struct ldlm_ast_work *w =
932 list_entry(tmp, struct ldlm_ast_work, w_list);
934 /* It's possible to receive a completion AST before we've set
935 * the l_completion_ast pointer: either because the AST arrived
936 * before the reply, or simply because there's a small race
937 * window between receiving the reply and finishing the local
940 * This can't happen with the blocking_ast, however, because we
941 * will never call the local blocking_ast until we drop our
942 * reader/writer reference, which we won't do until we get the
943 * reply and finish enqueueing. */
944 LASSERT(w->w_lock != NULL);
946 LASSERT(w->w_lock->l_blocking_ast != NULL);
947 rc = w->w_lock->l_blocking_ast
948 (w->w_lock, &w->w_desc, w->w_data,
950 } else if (w->w_lock->l_completion_ast != NULL) {
951 LASSERT(w->w_lock->l_completion_ast != NULL);
952 rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags,
960 CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
961 "disconnect client\n");
962 LDLM_LOCK_PUT(w->w_lock);
963 list_del(&w->w_list);
964 OBD_FREE(w, sizeof(*w));
969 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
971 ldlm_reprocess_all(res);
972 return LDLM_ITER_CONTINUE;
975 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
979 l_lock(&ns->ns_lock);
980 for (i = 0; i < RES_HASH_SIZE; i++) {
981 struct list_head *tmp, *next;
982 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
983 struct ldlm_resource *res =
984 list_entry(tmp, struct ldlm_resource, lr_hash);
986 ldlm_resource_getref(res);
987 l_unlock(&ns->ns_lock);
988 rc = reprocess_one_queue(res, NULL);
989 l_lock(&ns->ns_lock);
991 ldlm_resource_putref(res);
992 if (rc == LDLM_ITER_STOP)
997 l_unlock(&ns->ns_lock);
1001 void ldlm_reprocess_all(struct ldlm_resource *res)
1003 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1007 /* Local lock trees don't get reprocessed. */
1008 if (res->lr_namespace->ns_client) {
1014 l_lock(&res->lr_namespace->ns_lock);
1015 res->lr_tmp = &rpc_list;
1017 rc = ldlm_reprocess_queue(res, &res->lr_converting);
1018 if (rc == LDLM_ITER_CONTINUE)
1019 ldlm_reprocess_queue(res, &res->lr_waiting);
1022 l_unlock(&res->lr_namespace->ns_lock);
1024 rc = ldlm_run_ast_work(res->lr_namespace, &rpc_list);
1025 if (rc == -ERESTART) {
1026 LASSERT(list_empty(&rpc_list));
1032 void ldlm_cancel_callback(struct ldlm_lock *lock)
1034 l_lock(&lock->l_resource->lr_namespace->ns_lock);
1035 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1036 lock->l_flags |= LDLM_FL_CANCEL;
1037 if (lock->l_blocking_ast) {
1038 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1039 // l_check_no_ns_lock(lock->l_resource->lr_namespace);
1040 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1044 LDLM_DEBUG(lock, "no blocking ast");
1047 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1050 void ldlm_lock_cancel(struct ldlm_lock *lock)
1052 struct ldlm_resource *res;
1053 struct ldlm_namespace *ns;
1056 res = lock->l_resource;
1057 ns = res->lr_namespace;
1059 l_lock(&ns->ns_lock);
1060 ldlm_del_waiting_lock(lock);
1062 /* Please do not, no matter how tempting, remove this LBUG without
1063 * talking to me first. -phik */
1064 if (lock->l_readers || lock->l_writers) {
1065 LDLM_ERROR(lock, "lock still has references");
1069 ldlm_cancel_callback(lock);
1071 ldlm_resource_unlink_lock(lock);
1072 ldlm_lock_destroy(lock);
1073 l_unlock(&ns->ns_lock);
1077 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1079 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1085 lock->l_ast_data = data;
1086 LDLM_LOCK_PUT(lock);
1090 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1092 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1093 struct ldlm_lock *lock;
1094 struct ldlm_resource *res;
1096 l_lock(&ns->ns_lock);
1097 while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1098 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1099 struct ldlm_lock, l_export_chain);
1100 res = ldlm_resource_getref(lock->l_resource);
1101 LDLM_DEBUG(lock, "export %p", exp);
1102 ldlm_lock_cancel(lock);
1103 l_unlock(&ns->ns_lock);
1104 ldlm_reprocess_all(res);
1105 ldlm_resource_putref(res);
1106 l_lock(&ns->ns_lock);
1108 l_unlock(&ns->ns_lock);
1111 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1114 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1115 struct ldlm_resource *res;
1116 struct ldlm_namespace *ns;
1122 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1123 *flags |= LDLM_FL_BLOCK_GRANTED;
1124 RETURN(lock->l_resource);
1127 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1128 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1130 res = lock->l_resource;
1131 ns = res->lr_namespace;
1133 l_lock(&ns->ns_lock);
1135 old_mode = lock->l_req_mode;
1136 lock->l_req_mode = new_mode;
1137 ldlm_resource_unlink_lock(lock);
1139 /* If this is a local resource, put it on the appropriate list. */
1140 if (res->lr_namespace->ns_client) {
1141 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1142 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1144 /* This should never happen, because of the way the
1145 * server handles conversions. */
1146 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1150 res->lr_tmp = &rpc_list;
1151 ldlm_grant_lock(lock, NULL, 0, 0);
1154 /* FIXME: completion handling not with ns_lock held ! */
1155 if (lock->l_completion_ast)
1156 lock->l_completion_ast(lock, 0, NULL);
1160 ldlm_processing_policy policy;
1161 policy = ldlm_processing_policy_table[res->lr_type];
1162 res->lr_tmp = &rpc_list;
1163 rc = policy(lock, &pflags, 0, &err);
1165 if (rc == LDLM_ITER_STOP) {
1166 lock->l_req_mode = old_mode;
1167 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1170 *flags |= LDLM_FL_BLOCK_GRANTED;
1175 l_unlock(&ns->ns_lock);
1178 ldlm_run_ast_work(ns, &rpc_list);
1182 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1184 char str[PTL_NALFMT_SIZE];
1185 struct obd_device *obd = NULL;
1187 if (!((portal_debug | D_ERROR) & level))
1191 CDEBUG(level, " NULL LDLM lock\n");
1195 CDEBUG(level, " -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1196 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1198 if (lock->l_conn_export != NULL)
1199 obd = lock->l_conn_export->exp_obd;
1200 if (lock->l_export && lock->l_export->exp_connection) {
1201 CDEBUG(level, " Node: NID %s on %s (rhandle: "LPX64")\n",
1202 ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str),
1203 lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1204 lock->l_remote_handle.cookie);
1205 } else if (obd == NULL) {
1206 CDEBUG(level, " Node: local\n");
1208 struct obd_import *imp = obd->u.cli.cl_import;
1209 CDEBUG(level, " Node: NID %s on %s (rhandle: "LPX64")\n",
1210 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1211 imp->imp_connection->c_peer.peer_ni->pni_name,
1212 lock->l_remote_handle.cookie);
1214 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource,
1215 lock->l_resource->lr_name.name[0],
1216 lock->l_resource->lr_name.name[1]);
1217 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1218 "write: %d\n", ldlm_lockname[lock->l_req_mode],
1219 ldlm_lockname[lock->l_granted_mode],
1220 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers);
1221 if (lock->l_resource->lr_type == LDLM_EXTENT)
1222 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1223 " (req "LPU64"-"LPU64")\n",
1224 lock->l_policy_data.l_extent.start,
1225 lock->l_policy_data.l_extent.end,
1226 lock->l_req_extent.start, lock->l_req_extent.end);
1227 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1228 CDEBUG(level, " Pid: "LPU64" Extent: "LPU64" -> "LPU64"\n",
1229 lock->l_policy_data.l_flock.pid,
1230 lock->l_policy_data.l_flock.start,
1231 lock->l_policy_data.l_flock.end);
1232 else if (lock->l_resource->lr_type == LDLM_IBITS)
1233 CDEBUG(level, " Bits: "LPX64"\n",
1234 lock->l_policy_data.l_inodebits.bits);
1237 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1239 struct ldlm_lock *lock;
1241 lock = ldlm_handle2lock(lockh);
1245 ldlm_lock_dump(D_OTHER, lock, 0);
1247 LDLM_LOCK_PUT(lock);