1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Phil Schwan <phil@clusterfs.com>
8 * This file is part of Lustre, http://www.lustre.org.
10 * Lustre is free software; you can redistribute it and/or
11 * modify it under the terms of version 2 of the GNU General Public
12 * License as published by the Free Software Foundation.
14 * Lustre is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with Lustre; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 #define DEBUG_SUBSYSTEM S_LDLM
27 # include <linux/slab.h>
28 # include <linux/module.h>
29 # include <linux/lustre_dlm.h>
31 # include <liblustre.h>
32 # include <linux/kp30.h>
35 #include <linux/obd_class.h>
36 #include "ldlm_internal.h"
38 //struct lustre_lock ldlm_everything_lock;
41 char *ldlm_lockname[] = {
50 char *ldlm_typename[] = {
56 char *ldlm_it2str(int it)
63 case (IT_OPEN | IT_CREAT):
76 CERROR("Unknown intent %d\n", it);
81 extern kmem_cache_t *ldlm_lock_slab;
82 struct lustre_lock ldlm_handle_lock;
84 static ldlm_processing_policy ldlm_processing_policy_table[] = {
85 [LDLM_PLAIN] ldlm_process_plain_lock,
86 [LDLM_EXTENT] ldlm_process_extent_lock,
88 [LDLM_FLOCK] ldlm_process_flock_lock,
92 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
94 return ldlm_processing_policy_table[res->lr_type];
97 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
103 * REFCOUNTED LOCK OBJECTS
108 * Lock refcounts, during creation:
109 * - one special one for allocation, dec'd only once in destroy
110 * - one for being a lock that's in-use
111 * - one for the addref associated with a new lock
113 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
115 atomic_inc(&lock->l_refc);
119 void ldlm_lock_put(struct ldlm_lock *lock)
123 if (atomic_dec_and_test(&lock->l_refc)) {
124 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
126 l_lock(&ns->ns_lock);
127 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
128 LASSERT(lock->l_destroyed);
129 LASSERT(list_empty(&lock->l_res_link));
131 spin_lock(&ns->ns_counter_lock);
133 spin_unlock(&ns->ns_counter_lock);
135 ldlm_resource_putref(lock->l_resource);
136 lock->l_resource = NULL;
138 class_export_put(lock->l_export);
141 LDLM_LOCK_PUT(lock->l_parent);
143 if (lock->l_lvb_data != NULL)
144 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
146 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
147 l_unlock(&ns->ns_lock);
153 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
156 l_lock(&lock->l_resource->lr_namespace->ns_lock);
157 if (!list_empty(&lock->l_lru)) {
158 list_del_init(&lock->l_lru);
159 lock->l_resource->lr_namespace->ns_nr_unused--;
160 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
162 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
166 /* This used to have a 'strict' flact, which recovery would use to mark an
167 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
168 * shall explain why it's gone: with the new hash table scheme, once you call
169 * ldlm_lock_destroy, you can never drop your final references on this lock.
170 * Because it's not in the hash table anymore. -phil */
171 void ldlm_lock_destroy(struct ldlm_lock *lock)
174 l_lock(&lock->l_resource->lr_namespace->ns_lock);
176 if (!list_empty(&lock->l_children)) {
177 LDLM_ERROR(lock, "still has children (%p)!",
178 lock->l_children.next);
179 ldlm_lock_dump(D_ERROR, lock, 0);
182 if (lock->l_readers || lock->l_writers) {
183 LDLM_ERROR(lock, "lock still has references");
184 ldlm_lock_dump(D_ERROR, lock, 0);
188 if (!list_empty(&lock->l_res_link)) {
189 ldlm_lock_dump(D_ERROR, lock, 0);
193 if (lock->l_destroyed) {
194 LASSERT(list_empty(&lock->l_lru));
195 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
199 lock->l_destroyed = 1;
201 list_del_init(&lock->l_export_chain);
202 ldlm_lock_remove_from_lru(lock);
203 class_handle_unhash(&lock->l_handle);
206 /* Wake anyone waiting for this lock */
207 /* FIXME: I should probably add yet another flag, instead of using
208 * l_export to only call this on clients */
210 class_export_put(lock->l_export);
211 lock->l_export = NULL;
212 if (lock->l_export && lock->l_completion_ast)
213 lock->l_completion_ast(lock, 0);
216 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
221 /* this is called by portals_handle2object with the handle lock taken */
222 static void lock_handle_addref(void *lock)
224 LDLM_LOCK_GET((struct ldlm_lock *)lock);
228 * usage: pass in a resource on which you have done ldlm_resource_get
229 * pass in a parent lock on which you have done a ldlm_lock_get
230 * after return, ldlm_*_put the resource and parent
231 * returns: lock with refcount 1
233 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
234 struct ldlm_resource *resource)
236 struct ldlm_lock *lock;
239 if (resource == NULL)
242 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, SLAB_NOFS, sizeof(*lock));
246 lock->l_resource = ldlm_resource_getref(resource);
248 atomic_set(&lock->l_refc, 2);
249 INIT_LIST_HEAD(&lock->l_children);
250 INIT_LIST_HEAD(&lock->l_res_link);
251 INIT_LIST_HEAD(&lock->l_lru);
252 INIT_LIST_HEAD(&lock->l_export_chain);
253 INIT_LIST_HEAD(&lock->l_pending_chain);
254 init_waitqueue_head(&lock->l_waitq);
256 spin_lock(&resource->lr_namespace->ns_counter_lock);
257 resource->lr_namespace->ns_locks++;
258 spin_unlock(&resource->lr_namespace->ns_counter_lock);
260 if (parent != NULL) {
261 l_lock(&parent->l_resource->lr_namespace->ns_lock);
262 lock->l_parent = LDLM_LOCK_GET(parent);
263 list_add(&lock->l_childof, &parent->l_children);
264 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
267 INIT_LIST_HEAD(&lock->l_handle.h_link);
268 class_handle_hash(&lock->l_handle, lock_handle_addref);
273 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
274 struct ldlm_res_id new_resid)
276 struct ldlm_resource *oldres = lock->l_resource;
279 l_lock(&ns->ns_lock);
280 if (memcmp(&new_resid, &lock->l_resource->lr_name,
281 sizeof(lock->l_resource->lr_name)) == 0) {
283 l_unlock(&ns->ns_lock);
287 LASSERT(new_resid.name[0] != 0);
289 /* This function assumes that the lock isn't on any lists */
290 LASSERT(list_empty(&lock->l_res_link));
292 lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
293 lock->l_resource->lr_type, 1);
294 if (lock->l_resource == NULL) {
299 /* ...and the flowers are still standing! */
300 ldlm_resource_putref(oldres);
302 l_unlock(&ns->ns_lock);
310 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
312 lockh->cookie = lock->l_handle.h_cookie;
315 /* if flags: atomically get the lock and set the flags.
316 * Return NULL if flag already set
319 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
321 struct ldlm_namespace *ns;
322 struct ldlm_lock *lock = NULL, *retval = NULL;
327 lock = class_handle2object(handle->cookie);
331 LASSERT(lock->l_resource != NULL);
332 ns = lock->l_resource->lr_namespace;
335 l_lock(&ns->ns_lock);
337 /* It's unlikely but possible that someone marked the lock as
338 * destroyed after we did handle2object on it */
339 if (lock->l_destroyed) {
340 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
345 if (flags && (lock->l_flags & flags)) {
351 lock->l_flags |= flags;
356 l_unlock(&ns->ns_lock);
360 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
361 struct lustre_handle *handle)
363 struct ldlm_lock *retval = NULL;
365 l_lock(&ns->ns_lock);
366 retval = __ldlm_handle2lock(handle, 0);
367 l_unlock(&ns->ns_lock);
372 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
374 ldlm_res2desc(lock->l_resource, &desc->l_resource);
375 desc->l_req_mode = lock->l_req_mode;
376 desc->l_granted_mode = lock->l_granted_mode;
377 memcpy(&desc->l_policy_data, &lock->l_policy_data,
378 sizeof(desc->l_policy_data));
381 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
382 void *data, int datalen)
384 struct ldlm_ast_work *w;
387 l_lock(&lock->l_resource->lr_namespace->ns_lock);
388 if (new && (lock->l_flags & LDLM_FL_AST_SENT))
391 CDEBUG(D_OTHER, "lock %p incompatible; sending blocking AST.\n", lock);
393 OBD_ALLOC(w, sizeof(*w));
400 w->w_datalen = datalen;
402 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
403 lock->l_flags |= LDLM_FL_AST_SENT;
404 /* If the enqueuing client said so, tell the AST recipient to
405 * discard dirty data, rather than writing back. */
406 if (new->l_flags & LDLM_AST_DISCARD_DATA)
407 lock->l_flags |= LDLM_FL_DISCARD_DATA;
409 ldlm_lock2desc(new, &w->w_desc);
412 w->w_lock = LDLM_LOCK_GET(lock);
413 list_add(&w->w_list, lock->l_resource->lr_tmp);
416 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
419 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
421 struct ldlm_lock *lock;
423 lock = ldlm_handle2lock(lockh);
424 ldlm_lock_addref_internal(lock, mode);
428 /* only called for local locks */
429 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
431 l_lock(&lock->l_resource->lr_namespace->ns_lock);
432 ldlm_lock_remove_from_lru(lock);
433 if (mode & (LCK_NL | LCK_CR | LCK_PR))
435 if (mode & (LCK_EX | LCK_CW | LCK_PW))
437 lock->l_last_used = jiffies;
438 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
440 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
443 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
445 struct ldlm_namespace *ns;
448 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
449 ns = lock->l_resource->lr_namespace;
450 l_lock(&ns->ns_lock);
451 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
452 LASSERT(lock->l_readers > 0);
455 if (mode & (LCK_EX | LCK_CW | LCK_PW)) {
456 LASSERT(lock->l_writers > 0);
460 if (lock->l_flags & LDLM_FL_LOCAL &&
461 !lock->l_readers && !lock->l_writers) {
462 /* If this is a local lock on a server namespace and this was
463 * the last reference, cancel the lock. */
464 CDEBUG(D_INFO, "forcing cancel of local lock\n");
465 lock->l_flags |= LDLM_FL_CBPENDING;
468 if (!lock->l_readers && !lock->l_writers &&
469 (lock->l_flags & LDLM_FL_CBPENDING)) {
470 /* If we received a blocked AST and this was the last reference,
471 * run the callback. */
472 if (ns->ns_client == LDLM_NAMESPACE_SERVER && lock->l_export)
473 CERROR("FL_CBPENDING set on non-local lock--just a "
476 LDLM_DEBUG(lock, "final decref done on cbpending lock");
478 LDLM_LOCK_GET(lock); /* dropped by bl thread */
479 ldlm_lock_remove_from_lru(lock);
480 ldlm_bl_to_thread(ns, NULL, lock);
481 l_unlock(&ns->ns_lock);
482 } else if (ns->ns_client == LDLM_NAMESPACE_CLIENT &&
483 !lock->l_readers && !lock->l_writers) {
484 /* If this is a client-side namespace and this was the last
485 * reference, put it on the LRU. */
486 LASSERT(list_empty(&lock->l_lru));
487 LASSERT(ns->ns_nr_unused >= 0);
488 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
490 l_unlock(&ns->ns_lock);
493 l_unlock(&ns->ns_lock);
496 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
501 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
503 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
504 LASSERT(lock != NULL);
505 ldlm_lock_decref_internal(lock, mode);
509 /* This will drop a lock reference and mark it for destruction, but will not
510 * necessarily cancel the lock before returning. */
511 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
513 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
516 LASSERT(lock != NULL);
518 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
519 l_lock(&lock->l_resource->lr_namespace->ns_lock);
520 lock->l_flags |= LDLM_FL_CBPENDING;
521 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
522 ldlm_lock_decref_internal(lock, mode);
527 * - ldlm_lock_enqueue
528 * - ldlm_reprocess_queue
529 * - ldlm_lock_convert
531 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen,
534 struct ldlm_resource *res = lock->l_resource;
537 l_lock(&lock->l_resource->lr_namespace->ns_lock);
538 lock->l_granted_mode = lock->l_req_mode;
539 ldlm_resource_add_lock(res, &res->lr_granted, lock);
541 if (lock->l_granted_mode < res->lr_most_restr)
542 res->lr_most_restr = lock->l_granted_mode;
544 if (run_ast && lock->l_completion_ast != NULL)
545 ldlm_add_ast_work_item(lock, NULL, data, datalen);
547 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
551 /* returns a referenced lock or NULL. See the flag descriptions below, in the
552 * comment above ldlm_lock_match */
553 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
554 ldlm_policy_data_t *policy,
555 struct ldlm_lock *old_lock, int flags)
557 struct ldlm_lock *lock;
558 struct list_head *tmp;
560 list_for_each(tmp, queue) {
561 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
563 if (lock == old_lock)
566 /* llite sometimes wants to match locks that will be
567 * canceled when their users drop, but we allow it to match
568 * if it passes in CBPENDING and the lock still has users.
569 * this is generally only going to be used by children
570 * whose parents already hold a lock so forward progress
571 * can still happen. */
572 if (lock->l_flags & LDLM_FL_CBPENDING &&
573 !(flags & LDLM_FL_CBPENDING))
575 if (lock->l_flags & LDLM_FL_CBPENDING &&
576 lock->l_readers == 0 && lock->l_writers == 0)
579 if (!(lock->l_req_mode & mode))
582 if (lock->l_resource->lr_type == LDLM_EXTENT &&
583 (lock->l_policy_data.l_extent.start >
584 policy->l_extent.start ||
585 lock->l_policy_data.l_extent.end < policy->l_extent.end))
588 if (lock->l_destroyed)
591 if ((flags & LDLM_FL_LOCAL_ONLY) &&
592 !(lock->l_flags & LDLM_FL_LOCAL))
595 if (flags & LDLM_FL_TEST_LOCK)
598 ldlm_lock_addref_internal(lock, mode);
605 void ldlm_lock_allow_match(struct ldlm_lock *lock)
607 l_lock(&lock->l_resource->lr_namespace->ns_lock);
608 lock->l_flags |= LDLM_FL_CAN_MATCH;
609 wake_up(&lock->l_waitq);
610 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
613 /* Can be called in two ways:
615 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
616 * for a duplicate of.
618 * Otherwise, all of the fields must be filled in, to match against.
620 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
621 * server (ie, connh is NULL)
622 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
623 * list will be considered
624 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
625 * to be canceled can still be matched as long as they still have reader
626 * or writer refernces
627 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
628 * just tell us if we would have matched.
630 * Returns 1 if it finds an already-existing lock that is compatible; in this
631 * case, lockh is filled in with a addref()ed lock
633 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
634 struct ldlm_res_id *res_id, __u32 type,
635 ldlm_policy_data_t *policy, ldlm_mode_t mode,
636 struct lustre_handle *lockh)
638 struct ldlm_resource *res;
639 struct ldlm_lock *lock, *old_lock = NULL;
644 old_lock = ldlm_handle2lock(lockh);
647 ns = old_lock->l_resource->lr_namespace;
648 res_id = &old_lock->l_resource->lr_name;
649 type = old_lock->l_resource->lr_type;
650 mode = old_lock->l_req_mode;
653 res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
655 LASSERT(old_lock == NULL);
659 l_lock(&ns->ns_lock);
661 lock = search_queue(&res->lr_granted, mode, policy, old_lock, flags);
664 if (flags & LDLM_FL_BLOCK_GRANTED)
666 lock = search_queue(&res->lr_converting, mode, policy, old_lock, flags);
669 lock = search_queue(&res->lr_waiting, mode, policy, old_lock, flags);
675 ldlm_resource_putref(res);
676 l_unlock(&ns->ns_lock);
679 ldlm_lock2handle(lock, lockh);
680 if (!(lock->l_flags & LDLM_FL_CAN_MATCH)) {
681 struct l_wait_info lwi;
682 if (lock->l_completion_ast)
683 lock->l_completion_ast(lock,
684 LDLM_FL_WAIT_NOREPROC,
687 lwi = LWI_TIMEOUT_INTR(obd_timeout*HZ, NULL,NULL,NULL);
689 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
690 l_wait_event(lock->l_waitq,
691 (lock->l_flags & LDLM_FL_CAN_MATCH), &lwi);
695 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
696 type == LDLM_PLAIN ? res_id->name[2] :
697 policy->l_extent.start,
698 type == LDLM_PLAIN ? res_id->name[3] :
699 policy->l_extent.end);
700 else if (!(flags & LDLM_FL_TEST_LOCK)) /* less verbose for test-only */
701 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
702 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
703 type, mode, res_id->name[0], res_id->name[1],
704 type == LDLM_PLAIN ? res_id->name[2] :
705 policy->l_extent.start,
706 type == LDLM_PLAIN ? res_id->name[3] :
707 policy->l_extent.end);
710 LDLM_LOCK_PUT(old_lock);
711 if (flags & LDLM_FL_TEST_LOCK && rc)
717 /* Returns a referenced lock */
718 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
719 struct lustre_handle *parent_lock_handle,
720 struct ldlm_res_id res_id, __u32 type,
722 ldlm_blocking_callback blocking,
723 ldlm_completion_callback completion,
724 ldlm_glimpse_callback glimpse,
725 void *data, __u32 lvb_len)
727 struct ldlm_resource *res, *parent_res = NULL;
728 struct ldlm_lock *lock, *parent_lock = NULL;
731 if (parent_lock_handle) {
732 parent_lock = ldlm_handle2lock(parent_lock_handle);
734 parent_res = parent_lock->l_resource;
737 res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
741 lock = ldlm_lock_new(parent_lock, res);
742 ldlm_resource_putref(res);
743 if (parent_lock != NULL)
744 LDLM_LOCK_PUT(parent_lock);
749 lock->l_req_mode = mode;
750 lock->l_ast_data = data;
751 lock->l_blocking_ast = blocking;
752 lock->l_completion_ast = completion;
753 lock->l_glimpse_ast = glimpse;
756 lock->l_lvb_len = lvb_len;
757 OBD_ALLOC(lock->l_lvb_data, lvb_len);
758 if (lock->l_lvb_data == NULL) {
759 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
767 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
768 struct ldlm_lock **lockp,
769 void *cookie, int *flags)
771 struct ldlm_lock *lock = *lockp;
772 struct ldlm_resource *res = lock->l_resource;
773 int local = res->lr_namespace->ns_client;
774 ldlm_processing_policy policy;
775 ldlm_error_t rc = ELDLM_OK;
778 /* policies are not executed on the client or during replay */
779 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
780 && !local && ns->ns_policy) {
781 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
783 if (rc == ELDLM_LOCK_REPLACED) {
784 /* The lock that was returned has already been granted,
785 * and placed into lockp. If it's not the same as the
786 * one we passed in, then destroy the old one and our
787 * work here is done. */
788 if (lock != *lockp) {
789 ldlm_lock_destroy(lock);
792 *flags |= LDLM_FL_LOCK_CHANGED;
794 } else if (rc == ELDLM_LOCK_ABORTED ||
795 (rc == 0 && (*flags & LDLM_FL_INTENT_ONLY))) {
796 ldlm_lock_destroy(lock);
799 LASSERT(rc == ELDLM_OK);
802 l_lock(&ns->ns_lock);
803 if (local && lock->l_req_mode == lock->l_granted_mode) {
804 /* The server returned a blocked lock, but it was granted before
805 * we got a chance to actually enqueue it. We don't need to do
807 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
808 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
812 /* Some flags from the enqueue want to make it into the AST, via the
814 lock->l_flags |= (*flags & LDLM_AST_DISCARD_DATA);
816 /* This distinction between local lock trees is very important; a client
817 * namespace only has information about locks taken by that client, and
818 * thus doesn't have enough information to decide for itself if it can
819 * be granted (below). In this case, we do exactly what the server
820 * tells us to do, as dictated by the 'flags'.
822 * We do exactly the same thing during recovery, when the server is
823 * more or less trusting the clients not to lie.
825 * FIXME (bug 268): Detect obvious lies by checking compatibility in
826 * granted/converting queues. */
827 ldlm_resource_unlink_lock(lock);
829 if (*flags & LDLM_FL_BLOCK_CONV)
830 ldlm_resource_add_lock(res, &res->lr_converting, lock);
831 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
832 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
834 ldlm_grant_lock(lock, NULL, 0, 0);
836 } else if (*flags & LDLM_FL_REPLAY) {
837 if (*flags & LDLM_FL_BLOCK_CONV) {
838 ldlm_resource_add_lock(res, &res->lr_converting, lock);
840 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
841 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
843 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
844 ldlm_grant_lock(lock, NULL, 0, 0);
847 /* If no flags, fall through to normal enqueue path. */
850 policy = ldlm_processing_policy_table[res->lr_type];
851 policy(lock, flags, 1, &rc);
854 l_unlock(&ns->ns_lock);
858 /* Must be called with namespace taken: queue is waiting or converting. */
859 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue)
861 struct list_head *tmp, *pos;
862 ldlm_processing_policy policy;
864 int rc = LDLM_ITER_CONTINUE;
868 policy = ldlm_processing_policy_table[res->lr_type];
871 list_for_each_safe(tmp, pos, queue) {
872 struct ldlm_lock *pending;
873 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
875 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
878 rc = policy(pending, &flags, 0, &err);
879 if (rc != LDLM_ITER_CONTINUE)
886 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list)
888 struct list_head *tmp, *pos;
892 l_check_no_ns_lock(ns);
894 list_for_each_safe(tmp, pos, rpc_list) {
895 struct ldlm_ast_work *w =
896 list_entry(tmp, struct ldlm_ast_work, w_list);
898 /* It's possible to receive a completion AST before we've set
899 * the l_completion_ast pointer: either because the AST arrived
900 * before the reply, or simply because there's a small race
901 * window between receiving the reply and finishing the local
904 * This can't happen with the blocking_ast, however, because we
905 * will never call the local blocking_ast until we drop our
906 * reader/writer reference, which we won't do until we get the
907 * reply and finish enqueueing. */
908 LASSERT(w->w_lock != NULL);
910 LASSERT(w->w_lock->l_blocking_ast != NULL);
911 rc = w->w_lock->l_blocking_ast
912 (w->w_lock, &w->w_desc, w->w_data,
914 } else if (w->w_lock->l_completion_ast != NULL) {
915 LASSERT(w->w_lock->l_completion_ast != NULL);
916 rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags,
924 CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
925 "disconnect client\n");
926 LDLM_LOCK_PUT(w->w_lock);
927 list_del(&w->w_list);
928 OBD_FREE(w, sizeof(*w));
933 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
935 ldlm_reprocess_all(res);
936 return LDLM_ITER_CONTINUE;
939 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
943 l_lock(&ns->ns_lock);
944 for (i = 0; i < RES_HASH_SIZE; i++) {
945 struct list_head *tmp, *next;
946 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
947 struct ldlm_resource *res =
948 list_entry(tmp, struct ldlm_resource, lr_hash);
950 ldlm_resource_getref(res);
951 l_unlock(&ns->ns_lock);
952 rc = reprocess_one_queue(res, NULL);
953 l_lock(&ns->ns_lock);
955 ldlm_resource_putref(res);
956 if (rc == LDLM_ITER_STOP)
961 l_unlock(&ns->ns_lock);
965 void ldlm_reprocess_all(struct ldlm_resource *res)
967 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
971 /* Local lock trees don't get reprocessed. */
972 if (res->lr_namespace->ns_client) {
978 l_lock(&res->lr_namespace->ns_lock);
979 res->lr_tmp = &rpc_list;
981 rc = ldlm_reprocess_queue(res, &res->lr_converting);
982 if (rc == LDLM_ITER_CONTINUE)
983 ldlm_reprocess_queue(res, &res->lr_waiting);
986 l_unlock(&res->lr_namespace->ns_lock);
988 rc = ldlm_run_ast_work(res->lr_namespace, &rpc_list);
989 if (rc == -ERESTART) {
990 LASSERT(list_empty(&rpc_list));
996 void ldlm_cancel_callback(struct ldlm_lock *lock)
998 l_lock(&lock->l_resource->lr_namespace->ns_lock);
999 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1000 lock->l_flags |= LDLM_FL_CANCEL;
1001 if (lock->l_blocking_ast) {
1002 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1003 // l_check_no_ns_lock(lock->l_resource->lr_namespace);
1004 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1008 LDLM_DEBUG(lock, "no blocking ast");
1011 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1014 void ldlm_lock_cancel(struct ldlm_lock *lock)
1016 struct ldlm_resource *res;
1017 struct ldlm_namespace *ns;
1020 /* There's no race between calling this and taking the ns lock below;
1021 * a lock can only be put on the waiting list once, because it can only
1022 * issue a blocking AST once. */
1023 ldlm_del_waiting_lock(lock);
1025 res = lock->l_resource;
1026 ns = res->lr_namespace;
1028 l_lock(&ns->ns_lock);
1029 /* Please do not, no matter how tempting, remove this LBUG without
1030 * talking to me first. -phik */
1031 if (lock->l_readers || lock->l_writers) {
1032 LDLM_DEBUG(lock, "lock still has references");
1033 ldlm_lock_dump(D_OTHER, lock, 0);
1037 ldlm_cancel_callback(lock);
1039 ldlm_resource_unlink_lock(lock);
1040 ldlm_lock_destroy(lock);
1041 l_unlock(&ns->ns_lock);
1045 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1047 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1053 lock->l_ast_data = data;
1054 LDLM_LOCK_PUT(lock);
1058 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1060 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1061 struct ldlm_lock *lock;
1062 struct ldlm_resource *res;
1064 l_lock(&ns->ns_lock);
1065 while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1066 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1067 struct ldlm_lock, l_export_chain);
1068 res = ldlm_resource_getref(lock->l_resource);
1069 LDLM_DEBUG(lock, "export %p", exp);
1070 ldlm_lock_cancel(lock);
1071 l_unlock(&ns->ns_lock);
1072 ldlm_reprocess_all(res);
1073 ldlm_resource_putref(res);
1074 l_lock(&ns->ns_lock);
1076 l_unlock(&ns->ns_lock);
1079 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1082 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1083 struct ldlm_resource *res;
1084 struct ldlm_namespace *ns;
1090 res = lock->l_resource;
1091 ns = res->lr_namespace;
1093 l_lock(&ns->ns_lock);
1095 lock->l_req_mode = new_mode;
1096 ldlm_resource_unlink_lock(lock);
1098 /* If this is a local resource, put it on the appropriate list. */
1099 if (res->lr_namespace->ns_client) {
1100 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1101 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1103 /* This should never happen, because of the way the
1104 * server handles conversions. */
1107 res->lr_tmp = &rpc_list;
1108 ldlm_grant_lock(lock, NULL, 0, 0);
1111 /* FIXME: completion handling not with ns_lock held ! */
1112 if (lock->l_completion_ast)
1113 lock->l_completion_ast(lock, 0, NULL);
1116 /* FIXME: We should try the conversion right away and possibly
1117 * return success without the need for an extra AST */
1118 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1119 *flags |= LDLM_FL_BLOCK_CONV;
1122 l_unlock(&ns->ns_lock);
1125 ldlm_run_ast_work(ns, &rpc_list);
1129 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1131 char str[PTL_NALFMT_SIZE];
1132 struct obd_device *obd = NULL;
1134 if (!((portal_debug | D_ERROR) & level))
1138 CDEBUG(level, " NULL LDLM lock\n");
1142 CDEBUG(level, " -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d)\n",
1143 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1145 if (lock->l_conn_export != NULL)
1146 obd = lock->l_conn_export->exp_obd;
1147 if (lock->l_export && lock->l_export->exp_connection) {
1148 CDEBUG(level, " Node: NID "LPX64" (%s) on %s (rhandle: "LPX64")\n",
1149 lock->l_export->exp_connection->c_peer.peer_nid,
1150 portals_nid2str(lock->l_export->exp_connection->c_peer.peer_ni->pni_number,
1151 lock->l_export->exp_connection->c_peer.peer_nid, str),
1152 lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1153 lock->l_remote_handle.cookie);
1154 } else if (obd == NULL) {
1155 CDEBUG(level, " Node: local\n");
1157 struct obd_import *imp = obd->u.cli.cl_import;
1158 CDEBUG(level, " Node: NID "LPX64" (%s) on %s (rhandle: "LPX64")\n",
1159 imp->imp_connection->c_peer.peer_nid,
1160 portals_nid2str(imp->imp_connection->c_peer.peer_ni->pni_number,
1161 imp->imp_connection->c_peer.peer_nid, str),
1162 imp->imp_connection->c_peer.peer_ni->pni_name,
1163 lock->l_remote_handle.cookie);
1165 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource,
1166 lock->l_resource->lr_name.name[0],
1167 lock->l_resource->lr_name.name[1]);
1168 CDEBUG(level, " Req mode: %d, grant mode: %d, rc: %u, read: %d, "
1169 "write: %d\n", (int)lock->l_req_mode, (int)lock->l_granted_mode,
1170 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers);
1171 if (lock->l_resource->lr_type == LDLM_EXTENT)
1172 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1173 " (req "LPU64"-"LPU64")\n",
1174 lock->l_policy_data.l_extent.start,
1175 lock->l_policy_data.l_extent.end,
1176 lock->l_req_extent.start, lock->l_req_extent.end);
1177 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1178 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1179 lock->l_policy_data.l_flock.pid,
1180 lock->l_policy_data.l_flock.start,
1181 lock->l_policy_data.l_flock.end);
1184 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1186 struct ldlm_lock *lock;
1188 lock = ldlm_handle2lock(lockh);
1192 ldlm_lock_dump(D_OTHER, lock, 0);
1194 LDLM_LOCK_PUT(lock);