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;
89 static ldlm_processing_policy ldlm_processing_policy_table[] = {
90 [LDLM_PLAIN] ldlm_process_plain_lock,
91 [LDLM_EXTENT] ldlm_process_extent_lock,
93 [LDLM_FLOCK] ldlm_process_flock_lock,
95 [LDLM_IBITS] ldlm_process_inodebits_lock,
98 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
100 return ldlm_processing_policy_table[res->lr_type];
103 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
109 * REFCOUNTED LOCK OBJECTS
114 * Lock refcounts, during creation:
115 * - one special one for allocation, dec'd only once in destroy
116 * - one for being a lock that's in-use
117 * - one for the addref associated with a new lock
119 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
121 atomic_inc(&lock->l_refc);
125 void ldlm_lock_put(struct ldlm_lock *lock)
129 LASSERT(lock->l_resource != LP_POISON);
130 LASSERT(atomic_read(&lock->l_refc) > 0);
131 if (atomic_dec_and_test(&lock->l_refc)) {
132 struct ldlm_resource *res;
134 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
136 lock_res_and_lock(lock);
137 res = lock->l_resource;
138 LASSERT(lock->l_destroyed);
139 LASSERT(list_empty(&lock->l_res_link));
142 LDLM_LOCK_PUT(lock->l_parent);
143 unlock_res_and_lock(lock);
145 atomic_dec(&res->lr_namespace->ns_locks);
146 ldlm_resource_putref(res);
147 lock->l_resource = NULL;
149 class_export_put(lock->l_export);
151 if (lock->l_lvb_data != NULL)
152 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
154 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
160 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
163 spin_lock(&lock->l_resource->lr_namespace->ns_unused_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 spin_unlock(&lock->l_resource->lr_namespace->ns_unused_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)
183 lock_res_and_lock(lock);
185 if (!list_empty(&lock->l_children)) {
186 LDLM_ERROR(lock, "still has children (%p)!",
187 lock->l_children.next);
188 ldlm_lock_dump(D_ERROR, lock, 0);
191 if (lock->l_readers || lock->l_writers) {
192 LDLM_ERROR(lock, "lock still has references");
193 ldlm_lock_dump(D_ERROR, lock, 0);
197 if (!list_empty(&lock->l_res_link)) {
198 LDLM_ERROR(lock, "lock still on resource");
199 ldlm_lock_dump(D_ERROR, lock, 0);
203 if (lock->l_destroyed) {
204 LASSERT(list_empty(&lock->l_lru));
205 unlock_res_and_lock(lock);
209 lock->l_destroyed = 1;
211 if (lock->l_export) {
212 spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
213 if (!list_empty(&lock->l_export_chain))
214 list_del_init(&lock->l_export_chain);
215 spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
217 LASSERT(list_empty(&lock->l_export_chain));
220 ldlm_lock_remove_from_lru(lock);
221 class_handle_unhash(&lock->l_handle);
224 /* Wake anyone waiting for this lock */
225 /* FIXME: I should probably add yet another flag, instead of using
226 * l_export to only call this on clients */
228 class_export_put(lock->l_export);
229 lock->l_export = NULL;
230 if (lock->l_export && lock->l_completion_ast)
231 lock->l_completion_ast(lock, 0);
234 unlock_res_and_lock(lock);
239 /* this is called by portals_handle2object with the handle lock taken */
240 static void lock_handle_addref(void *lock)
242 LDLM_LOCK_GET((struct ldlm_lock *)lock);
246 * usage: pass in a resource on which you have done ldlm_resource_get
247 * pass in a parent lock on which you have done a ldlm_lock_get
248 * after return, ldlm_*_put the resource and parent
249 * returns: lock with refcount 2 - one for current caller and one for remote
251 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
252 struct ldlm_resource *resource)
254 struct ldlm_lock *lock;
257 if (resource == NULL)
260 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, SLAB_NOFS, sizeof(*lock));
264 lock->l_resource = ldlm_resource_getref(resource);
266 atomic_set(&lock->l_refc, 2);
267 INIT_LIST_HEAD(&lock->l_children);
268 INIT_LIST_HEAD(&lock->l_res_link);
269 INIT_LIST_HEAD(&lock->l_lru);
270 INIT_LIST_HEAD(&lock->l_export_chain);
271 INIT_LIST_HEAD(&lock->l_pending_chain);
272 INIT_LIST_HEAD(&lock->l_tmp);
273 INIT_LIST_HEAD(&lock->l_bl_ast);
274 INIT_LIST_HEAD(&lock->l_cp_ast);
275 init_waitqueue_head(&lock->l_waitq);
276 lock->l_blocking_lock = NULL;
279 atomic_inc(&resource->lr_namespace->ns_locks);
281 if (parent != NULL) {
282 spin_lock(&resource->lr_namespace->ns_hash_lock);
283 lock->l_parent = LDLM_LOCK_GET(parent);
284 list_add(&lock->l_childof, &parent->l_children);
285 spin_unlock(&resource->lr_namespace->ns_hash_lock);
288 INIT_LIST_HEAD(&lock->l_handle.h_link);
289 class_handle_hash(&lock->l_handle, lock_handle_addref);
294 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
295 struct ldlm_res_id new_resid)
297 struct ldlm_resource *oldres = lock->l_resource;
298 struct ldlm_resource *newres;
302 LASSERT(ns->ns_client != 0);
304 lock_res_and_lock(lock);
305 if (memcmp(&new_resid, &lock->l_resource->lr_name,
306 sizeof(lock->l_resource->lr_name)) == 0) {
308 unlock_res_and_lock(lock);
312 LASSERT(new_resid.name[0] != 0);
314 /* This function assumes that the lock isn't on any lists */
315 LASSERT(list_empty(&lock->l_res_link));
317 type = oldres->lr_type;
318 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
319 if (newres == NULL) {
325 lock->l_resource = newres;
328 unlock_bitlock(lock);
330 /* ...and the flowers are still standing! */
331 ldlm_resource_putref(oldres);
340 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
342 lockh->cookie = lock->l_handle.h_cookie;
345 /* if flags: atomically get the lock and set the flags.
346 * Return NULL if flag already set
349 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
351 struct ldlm_namespace *ns;
352 struct ldlm_lock *lock = NULL, *retval = NULL;
357 lock = class_handle2object(handle->cookie);
361 LASSERT(lock->l_resource != NULL);
362 ns = lock->l_resource->lr_namespace;
365 lock_res_and_lock(lock);
367 /* It's unlikely but possible that someone marked the lock as
368 * destroyed after we did handle2object on it */
369 if (lock->l_destroyed) {
370 unlock_res_and_lock(lock);
371 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
376 if (flags && (lock->l_flags & flags)) {
377 unlock_res_and_lock(lock);
383 lock->l_flags |= flags;
385 unlock_res_and_lock(lock);
392 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
393 struct lustre_handle *handle)
395 struct ldlm_lock *retval = NULL;
396 retval = __ldlm_handle2lock(handle, 0);
400 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
402 ldlm_res2desc(lock->l_resource, &desc->l_resource);
403 desc->l_req_mode = lock->l_req_mode;
404 desc->l_granted_mode = lock->l_granted_mode;
405 memcpy(&desc->l_policy_data, &lock->l_policy_data,
406 sizeof(desc->l_policy_data));
409 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
410 struct list_head *work_list)
412 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
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;
419 LASSERT(list_empty(&lock->l_bl_ast));
420 list_add(&lock->l_bl_ast, work_list);
422 LASSERT(lock->l_blocking_lock == NULL);
423 lock->l_blocking_lock = LDLM_LOCK_GET(new);
427 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
429 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
430 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
431 lock->l_flags |= LDLM_FL_CP_REQD;
432 LASSERT(list_empty(&lock->l_cp_ast));
433 list_add(&lock->l_cp_ast, work_list);
438 /* must be called with lr_lock held */
439 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
440 struct list_head *work_list)
443 check_res_locked(lock->l_resource);
445 ldlm_add_bl_work_item(lock, new, work_list);
447 ldlm_add_cp_work_item(lock, work_list);
451 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
453 struct ldlm_lock *lock;
455 lock = ldlm_handle2lock(lockh);
456 ldlm_lock_addref_internal(lock, mode);
460 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
462 ldlm_lock_remove_from_lru(lock);
463 if (mode & (LCK_NL | LCK_CR | LCK_PR))
465 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
467 lock->l_last_used = jiffies;
469 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
472 /* only called for local locks */
473 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
475 lock_res_and_lock(lock);
476 ldlm_lock_addref_internal_nolock(lock, mode);
477 unlock_res_and_lock(lock);
480 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
482 struct ldlm_namespace *ns;
485 lock_res_and_lock(lock);
487 ns = lock->l_resource->lr_namespace;
489 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
490 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
491 LASSERT(lock->l_readers > 0);
494 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
495 LASSERT(lock->l_writers > 0);
499 if (((lock->l_flags & LDLM_FL_LOCAL) || (mode == LCK_GROUP)) &&
500 !lock->l_readers && !lock->l_writers) {
501 /* If this is a local lock on a server namespace or a group
502 * lock and this was the last reference, cancel the lock. */
503 CDEBUG(D_INFO, "forcing cancel of local or group lock\n");
504 lock->l_flags |= LDLM_FL_CBPENDING;
507 if (!lock->l_readers && !lock->l_writers &&
508 (lock->l_flags & LDLM_FL_CBPENDING)) {
509 /* If we received a blocked AST and this was the last reference,
510 * run the callback. */
511 if (ns->ns_client == LDLM_NAMESPACE_SERVER && lock->l_export)
512 CERROR("FL_CBPENDING set on non-local lock--just a "
515 LDLM_DEBUG(lock, "final decref done on cbpending lock");
517 LDLM_LOCK_GET(lock); /* dropped by bl thread */
518 ldlm_lock_remove_from_lru(lock);
519 unlock_res_and_lock(lock);
520 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
521 ldlm_bl_to_thread(ns, NULL, lock) != 0)
522 ldlm_handle_bl_callback(ns, NULL, lock);
523 } else if (ns->ns_client == LDLM_NAMESPACE_CLIENT &&
524 !lock->l_readers && !lock->l_writers) {
525 /* If this is a client-side namespace and this was the last
526 * reference, put it on the LRU. */
527 LASSERT(list_empty(&lock->l_lru));
528 LASSERT(ns->ns_nr_unused >= 0);
529 spin_lock(&ns->ns_unused_lock);
530 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
532 spin_unlock(&ns->ns_unused_lock);
533 unlock_res_and_lock(lock);
534 ldlm_cancel_lru(ns, LDLM_ASYNC);
536 unlock_res_and_lock(lock);
539 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
544 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
546 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
547 LASSERTF(lock != NULL, "can't find lockh "LPX64"\n", lockh->cookie);
548 ldlm_lock_decref_internal(lock, mode);
552 /* This will drop a lock reference and mark it for destruction, but will not
553 * necessarily cancel the lock before returning. */
554 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
556 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
559 LASSERT(lock != NULL);
561 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
562 lock_res_and_lock(lock);
563 lock->l_flags |= LDLM_FL_CBPENDING;
564 unlock_res_and_lock(lock);
565 ldlm_lock_decref_internal(lock, mode);
570 * - ldlm_lock_enqueue
571 * - ldlm_reprocess_queue
572 * - ldlm_lock_convert
574 * must be called with lr_lock held
576 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
578 struct ldlm_resource *res = lock->l_resource;
581 check_res_locked(res);
583 lock->l_granted_mode = lock->l_req_mode;
584 ldlm_resource_add_lock(res, &res->lr_granted, lock);
586 if (lock->l_granted_mode < res->lr_most_restr)
587 res->lr_most_restr = lock->l_granted_mode;
589 if (work_list && lock->l_completion_ast != NULL)
590 ldlm_add_ast_work_item(lock, NULL, work_list);
595 /* returns a referenced lock or NULL. See the flag descriptions below, in the
596 * comment above ldlm_lock_match */
597 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
598 ldlm_policy_data_t *policy,
599 struct ldlm_lock *old_lock, int flags)
601 struct ldlm_lock *lock;
602 struct list_head *tmp;
604 list_for_each(tmp, queue) {
605 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
607 if (lock == old_lock)
610 /* llite sometimes wants to match locks that will be
611 * canceled when their users drop, but we allow it to match
612 * if it passes in CBPENDING and the lock still has users.
613 * this is generally only going to be used by children
614 * whose parents already hold a lock so forward progress
615 * can still happen. */
616 if (lock->l_flags & LDLM_FL_CBPENDING &&
617 !(flags & LDLM_FL_CBPENDING))
619 if (lock->l_flags & LDLM_FL_CBPENDING &&
620 lock->l_readers == 0 && lock->l_writers == 0)
623 if (!(lock->l_req_mode & mode))
626 if (lock->l_resource->lr_type == LDLM_EXTENT) {
627 if (mode == LCK_GROUP) {
628 if (lock->l_policy_data.l_extent.gid !=
629 policy->l_extent.gid)
632 if ((lock->l_policy_data.l_extent.start >
633 policy->l_extent.start) ||
634 (lock->l_policy_data.l_extent.end <
635 policy->l_extent.end))
640 /* We match if we have existing lock with same or wider set
642 if (lock->l_resource->lr_type == LDLM_IBITS &&
643 ((lock->l_policy_data.l_inodebits.bits &
644 policy->l_inodebits.bits) !=
645 policy->l_inodebits.bits))
648 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
651 if ((flags & LDLM_FL_LOCAL_ONLY) &&
652 !(lock->l_flags & LDLM_FL_LOCAL))
655 if (flags & LDLM_FL_TEST_LOCK)
658 ldlm_lock_addref_internal_nolock(lock, mode);
665 void ldlm_lock_allow_match(struct ldlm_lock *lock)
667 lock_res_and_lock(lock);
668 lock->l_flags |= LDLM_FL_CAN_MATCH;
669 wake_up(&lock->l_waitq);
670 unlock_res_and_lock(lock);
673 /* Can be called in two ways:
675 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
676 * for a duplicate of.
678 * Otherwise, all of the fields must be filled in, to match against.
680 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
681 * server (ie, connh is NULL)
682 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
683 * list will be considered
684 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
685 * to be canceled can still be matched as long as they still have reader
686 * or writer refernces
687 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
688 * just tell us if we would have matched.
690 * Returns 1 if it finds an already-existing lock that is compatible; in this
691 * case, lockh is filled in with a addref()ed lock
693 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
694 struct ldlm_res_id *res_id, __u32 type,
695 ldlm_policy_data_t *policy, ldlm_mode_t mode,
696 struct lustre_handle *lockh)
698 struct ldlm_resource *res;
699 struct ldlm_lock *lock, *old_lock = NULL;
704 old_lock = ldlm_handle2lock(lockh);
707 ns = old_lock->l_resource->lr_namespace;
708 res_id = &old_lock->l_resource->lr_name;
709 type = old_lock->l_resource->lr_type;
710 mode = old_lock->l_req_mode;
713 res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
715 LASSERT(old_lock == NULL);
721 lock = search_queue(&res->lr_granted, mode, policy, old_lock, flags);
724 if (flags & LDLM_FL_BLOCK_GRANTED)
726 lock = search_queue(&res->lr_converting, mode, policy, old_lock, flags);
729 lock = search_queue(&res->lr_waiting, mode, policy, old_lock, flags);
736 ldlm_resource_putref(res);
739 ldlm_lock2handle(lock, lockh);
740 if (!(lock->l_flags & LDLM_FL_CAN_MATCH)) {
741 struct l_wait_info lwi;
742 if (lock->l_completion_ast) {
743 int err = lock->l_completion_ast(lock,
744 LDLM_FL_WAIT_NOREPROC,
752 lwi = LWI_TIMEOUT_INTR(obd_timeout*HZ, NULL,NULL,NULL);
754 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
755 l_wait_event(lock->l_waitq,
756 (lock->l_flags & LDLM_FL_CAN_MATCH), &lwi);
762 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
763 type == LDLM_PLAIN ? res_id->name[2] :
764 policy->l_extent.start,
765 type == LDLM_PLAIN ? res_id->name[3] :
766 policy->l_extent.end);
767 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/* less verbose for test-only */
768 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
769 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
770 type, mode, res_id->name[0], res_id->name[1],
771 type == LDLM_PLAIN ? res_id->name[2] :
772 policy->l_extent.start,
773 type == LDLM_PLAIN ? res_id->name[3] :
774 policy->l_extent.end);
778 LDLM_LOCK_PUT(old_lock);
779 if (flags & LDLM_FL_TEST_LOCK && rc)
785 /* Returns a referenced lock */
786 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
787 struct lustre_handle *parent_lock_handle,
788 struct ldlm_res_id res_id, __u32 type,
790 ldlm_blocking_callback blocking,
791 ldlm_completion_callback completion,
792 ldlm_glimpse_callback glimpse,
793 void *data, __u32 lvb_len)
795 struct ldlm_resource *res, *parent_res = NULL;
796 struct ldlm_lock *lock, *parent_lock = NULL;
799 if (parent_lock_handle) {
800 parent_lock = ldlm_handle2lock(parent_lock_handle);
802 parent_res = parent_lock->l_resource;
805 res = ldlm_resource_get(ns, parent_res, res_id,
810 lock = ldlm_lock_new(parent_lock, res);
811 ldlm_resource_putref(res);
812 if (parent_lock != NULL)
813 LDLM_LOCK_PUT(parent_lock);
818 lock->l_req_mode = mode;
819 lock->l_ast_data = data;
820 lock->l_blocking_ast = blocking;
821 lock->l_completion_ast = completion;
822 lock->l_glimpse_ast = glimpse;
823 lock->l_pid = current->pid;
826 lock->l_lvb_len = lvb_len;
827 OBD_ALLOC(lock->l_lvb_data, lvb_len);
828 if (lock->l_lvb_data == NULL) {
829 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
837 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
838 struct ldlm_lock **lockp,
839 void *cookie, int *flags)
841 struct ldlm_lock *lock = *lockp;
842 struct ldlm_resource *res = lock->l_resource;
843 int local = res->lr_namespace->ns_client;
844 ldlm_processing_policy policy;
845 ldlm_error_t rc = ELDLM_OK;
848 /* policies are not executed on the client or during replay */
849 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
850 && !local && ns->ns_policy) {
851 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
853 if (rc == ELDLM_LOCK_REPLACED) {
854 /* The lock that was returned has already been granted,
855 * and placed into lockp. If it's not the same as the
856 * one we passed in, then destroy the old one and our
857 * work here is done. */
858 if (lock != *lockp) {
859 ldlm_lock_destroy(lock);
862 *flags |= LDLM_FL_LOCK_CHANGED;
864 } else if (rc == ELDLM_LOCK_ABORTED ||
865 (rc == 0 && (*flags & LDLM_FL_INTENT_ONLY))) {
866 ldlm_lock_destroy(lock);
869 LASSERT(rc == ELDLM_OK);
872 lock_res_and_lock(lock);
873 if (local && lock->l_req_mode == lock->l_granted_mode) {
874 /* The server returned a blocked lock, but it was granted before
875 * we got a chance to actually enqueue it. We don't need to do
877 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
878 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
882 /* Some flags from the enqueue want to make it into the AST, via the
884 lock->l_flags |= (*flags & LDLM_AST_DISCARD_DATA);
886 /* This distinction between local lock trees is very important; a client
887 * namespace only has information about locks taken by that client, and
888 * thus doesn't have enough information to decide for itself if it can
889 * be granted (below). In this case, we do exactly what the server
890 * tells us to do, as dictated by the 'flags'.
892 * We do exactly the same thing during recovery, when the server is
893 * more or less trusting the clients not to lie.
895 * FIXME (bug 268): Detect obvious lies by checking compatibility in
896 * granted/converting queues. */
897 ldlm_resource_unlink_lock(lock);
899 if (*flags & LDLM_FL_BLOCK_CONV)
900 ldlm_resource_add_lock(res, &res->lr_converting, lock);
901 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
902 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
904 ldlm_grant_lock(lock, NULL);
906 } else if (*flags & LDLM_FL_REPLAY) {
907 if (*flags & LDLM_FL_BLOCK_CONV) {
908 ldlm_resource_add_lock(res, &res->lr_converting, lock);
910 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
911 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
913 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
914 ldlm_grant_lock(lock, NULL);
917 /* If no flags, fall through to normal enqueue path. */
920 policy = ldlm_processing_policy_table[res->lr_type];
921 policy(lock, flags, 1, &rc, NULL);
924 unlock_res_and_lock(lock);
928 /* Must be called with namespace taken: queue is waiting or converting. */
929 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
930 struct list_head *work_list)
932 struct list_head *tmp, *pos;
933 ldlm_processing_policy policy;
935 int rc = LDLM_ITER_CONTINUE;
939 check_res_locked(res);
941 policy = ldlm_processing_policy_table[res->lr_type];
944 list_for_each_safe(tmp, pos, queue) {
945 struct ldlm_lock *pending;
946 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
948 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
951 rc = policy(pending, &flags, 0, &err, work_list);
952 if (rc != LDLM_ITER_CONTINUE)
959 int ldlm_run_bl_ast_work(struct list_head *rpc_list)
961 struct list_head *tmp, *pos;
962 struct ldlm_lock_desc d;
963 int rc = 0, retval = 0;
966 list_for_each_safe(tmp, pos, rpc_list) {
967 struct ldlm_lock *lock =
968 list_entry(tmp, struct ldlm_lock, l_bl_ast);
970 /* nobody should touch l_bl_ast */
971 lock_res_and_lock(lock);
972 list_del_init(&lock->l_bl_ast);
974 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
975 LASSERT(lock->l_bl_ast_run == 0);
976 LASSERT(lock->l_blocking_lock);
977 lock->l_bl_ast_run++;
978 unlock_res_and_lock(lock);
980 ldlm_lock2desc(lock->l_blocking_lock, &d);
982 LDLM_LOCK_PUT(lock->l_blocking_lock);
983 lock->l_blocking_lock = NULL;
984 rc = lock->l_blocking_ast(lock, &d, NULL, LDLM_CB_BLOCKING);
989 CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
990 "disconnect client\n");
996 int ldlm_run_cp_ast_work(struct list_head *rpc_list)
998 struct list_head *tmp, *pos;
999 int rc = 0, retval = 0;
1002 /* It's possible to receive a completion AST before we've set
1003 * the l_completion_ast pointer: either because the AST arrived
1004 * before the reply, or simply because there's a small race
1005 * window between receiving the reply and finishing the local
1006 * enqueue. (bug 842)
1008 * This can't happen with the blocking_ast, however, because we
1009 * will never call the local blocking_ast until we drop our
1010 * reader/writer reference, which we won't do until we get the
1011 * reply and finish enqueueing. */
1013 list_for_each_safe(tmp, pos, rpc_list) {
1014 struct ldlm_lock *lock =
1015 list_entry(tmp, struct ldlm_lock, l_cp_ast);
1017 /* nobody should touch l_cp_ast */
1018 lock_res_and_lock(lock);
1019 list_del_init(&lock->l_cp_ast);
1020 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1021 lock->l_flags &= ~LDLM_FL_CP_REQD;
1022 unlock_res_and_lock(lock);
1024 if (lock->l_completion_ast != NULL)
1025 rc = lock->l_completion_ast(lock, 0, 0);
1027 if (rc == -ERESTART)
1030 CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
1031 "disconnect client\n");
1032 LDLM_LOCK_PUT(lock);
1037 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1039 ldlm_reprocess_all(res);
1040 return LDLM_ITER_CONTINUE;
1043 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1045 struct list_head *tmp;
1048 spin_lock(&ns->ns_hash_lock);
1049 for (i = 0; i < RES_HASH_SIZE; i++) {
1050 tmp = ns->ns_hash[i].next;
1051 while (tmp != &(ns->ns_hash[i])) {
1052 struct ldlm_resource *res =
1053 list_entry(tmp, struct ldlm_resource, lr_hash);
1055 ldlm_resource_getref(res);
1056 spin_unlock(&ns->ns_hash_lock);
1058 rc = reprocess_one_queue(res, NULL);
1060 spin_lock(&ns->ns_hash_lock);
1062 ldlm_resource_putref_locked(res);
1064 if (rc == LDLM_ITER_STOP)
1069 spin_unlock(&ns->ns_hash_lock);
1073 void ldlm_reprocess_all(struct ldlm_resource *res)
1075 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1079 /* Local lock trees don't get reprocessed. */
1080 if (res->lr_namespace->ns_client) {
1087 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1088 if (rc == LDLM_ITER_CONTINUE)
1089 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1092 rc = ldlm_run_cp_ast_work(&rpc_list);
1093 if (rc == -ERESTART) {
1094 LASSERT(list_empty(&rpc_list));
1100 void ldlm_cancel_callback(struct ldlm_lock *lock)
1102 check_res_locked(lock->l_resource);
1104 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1105 lock->l_flags |= LDLM_FL_CANCEL;
1106 if (lock->l_blocking_ast) {
1107 unlock_res_and_lock(lock);
1108 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1110 lock_res_and_lock(lock);
1112 LDLM_DEBUG(lock, "no blocking ast");
1117 void ldlm_lock_cancel(struct ldlm_lock *lock)
1119 struct ldlm_resource *res;
1120 struct ldlm_namespace *ns;
1123 ldlm_del_waiting_lock(lock);
1124 lock_res_and_lock(lock);
1126 res = lock->l_resource;
1127 ns = res->lr_namespace;
1129 /* Please do not, no matter how tempting, remove this LBUG without
1130 * talking to me first. -phik */
1131 if (lock->l_readers || lock->l_writers) {
1132 LDLM_ERROR(lock, "lock still has references");
1136 ldlm_cancel_callback(lock);
1138 ldlm_resource_unlink_lock(lock);
1139 unlock_res_and_lock(lock);
1141 ldlm_lock_destroy(lock);
1146 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1148 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1154 lock->l_ast_data = data;
1155 LDLM_LOCK_PUT(lock);
1159 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1161 struct ldlm_lock *lock;
1162 struct ldlm_resource *res;
1164 spin_lock(&exp->exp_ldlm_data.led_lock);
1165 while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1166 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1167 struct ldlm_lock, l_export_chain);
1168 res = ldlm_resource_getref(lock->l_resource);
1169 LDLM_LOCK_GET(lock);
1170 spin_unlock(&exp->exp_ldlm_data.led_lock);
1172 LDLM_DEBUG(lock, "export %p", exp);
1173 ldlm_lock_cancel(lock);
1174 ldlm_reprocess_all(res);
1176 ldlm_resource_putref(res);
1177 LDLM_LOCK_PUT(lock);
1178 spin_lock(&exp->exp_ldlm_data.led_lock);
1180 spin_unlock(&exp->exp_ldlm_data.led_lock);
1183 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1186 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1187 struct ldlm_resource *res;
1188 struct ldlm_namespace *ns;
1194 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1195 *flags |= LDLM_FL_BLOCK_GRANTED;
1196 RETURN(lock->l_resource);
1199 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1200 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1202 lock_res_and_lock(lock);
1204 res = lock->l_resource;
1205 ns = res->lr_namespace;
1207 old_mode = lock->l_req_mode;
1208 lock->l_req_mode = new_mode;
1209 ldlm_resource_unlink_lock(lock);
1211 /* If this is a local resource, put it on the appropriate list. */
1212 if (res->lr_namespace->ns_client) {
1213 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1214 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1216 /* This should never happen, because of the way the
1217 * server handles conversions. */
1218 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1222 ldlm_grant_lock(lock, &rpc_list);
1224 /* FIXME: completion handling not with ns_lock held ! */
1225 if (lock->l_completion_ast)
1226 lock->l_completion_ast(lock, 0, NULL);
1230 ldlm_processing_policy policy;
1231 policy = ldlm_processing_policy_table[res->lr_type];
1232 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1233 if (rc == LDLM_ITER_STOP) {
1234 lock->l_req_mode = old_mode;
1235 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1238 *flags |= LDLM_FL_BLOCK_GRANTED;
1242 unlock_res_and_lock(lock);
1245 ldlm_run_cp_ast_work(&rpc_list);
1249 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1251 char str[PTL_NALFMT_SIZE];
1252 struct obd_device *obd = NULL;
1254 if (!((portal_debug | D_ERROR) & level))
1258 CDEBUG(level, " NULL LDLM lock\n");
1262 CDEBUG(level, " -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1263 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1265 if (lock->l_conn_export != NULL)
1266 obd = lock->l_conn_export->exp_obd;
1267 if (lock->l_export && lock->l_export->exp_connection) {
1268 CDEBUG(level, " Node: NID %s on %s (rhandle: "LPX64")\n",
1269 ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str),
1270 lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1271 lock->l_remote_handle.cookie);
1272 } else if (obd == NULL) {
1273 CDEBUG(level, " Node: local\n");
1275 struct obd_import *imp = obd->u.cli.cl_import;
1276 CDEBUG(level, " Node: NID %s on %s (rhandle: "LPX64")\n",
1277 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1278 imp->imp_connection->c_peer.peer_ni->pni_name,
1279 lock->l_remote_handle.cookie);
1281 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource,
1282 lock->l_resource->lr_name.name[0],
1283 lock->l_resource->lr_name.name[1]);
1284 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1285 "write: %d\n", ldlm_lockname[lock->l_req_mode],
1286 ldlm_lockname[lock->l_granted_mode],
1287 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers);
1288 if (lock->l_resource->lr_type == LDLM_EXTENT)
1289 CDEBUG(level, " Extent: "LPU64" -> "LPU64" gid: "LPU64
1290 " (req "LPU64"-"LPU64" gid: "LPU64")\n",
1291 lock->l_policy_data.l_extent.start,
1292 lock->l_policy_data.l_extent.end,
1293 lock->l_policy_data.l_extent.gid,
1294 lock->l_req_extent.start, lock->l_req_extent.end,
1295 lock->l_req_extent.gid);
1296 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1297 CDEBUG(level, " Pid: "LPU64" Extent: "LPU64" -> "LPU64"\n",
1298 lock->l_policy_data.l_flock.pid,
1299 lock->l_policy_data.l_flock.start,
1300 lock->l_policy_data.l_flock.end);
1301 else if (lock->l_resource->lr_type == LDLM_IBITS)
1302 CDEBUG(level, " Bits: "LPX64"\n",
1303 lock->l_policy_data.l_inodebits.bits);
1306 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1308 struct ldlm_lock *lock;
1310 lock = ldlm_handle2lock(lockh);
1314 ldlm_lock_dump(D_OTHER, lock, 0);
1316 LDLM_LOCK_PUT(lock);