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 void unlock_bitlock(struct ldlm_lock *lock);
296 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
297 struct ldlm_res_id new_resid)
299 struct ldlm_resource *oldres = lock->l_resource;
300 struct ldlm_resource *newres;
304 LASSERT(ns->ns_client != 0);
306 lock_res_and_lock(lock);
307 if (memcmp(&new_resid, &lock->l_resource->lr_name,
308 sizeof(lock->l_resource->lr_name)) == 0) {
310 unlock_res_and_lock(lock);
314 LASSERT(new_resid.name[0] != 0);
316 /* This function assumes that the lock isn't on any lists */
317 LASSERT(list_empty(&lock->l_res_link));
319 type = oldres->lr_type;
320 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
321 if (newres == NULL) {
327 lock->l_resource = newres;
330 unlock_bitlock(lock);
332 /* ...and the flowers are still standing! */
333 ldlm_resource_putref(oldres);
342 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
344 lockh->cookie = lock->l_handle.h_cookie;
347 /* if flags: atomically get the lock and set the flags.
348 * Return NULL if flag already set
351 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
353 struct ldlm_namespace *ns;
354 struct ldlm_lock *lock = NULL, *retval = NULL;
359 lock = class_handle2object(handle->cookie);
363 LASSERT(lock->l_resource != NULL);
364 ns = lock->l_resource->lr_namespace;
367 lock_res_and_lock(lock);
369 /* It's unlikely but possible that someone marked the lock as
370 * destroyed after we did handle2object on it */
371 if (lock->l_destroyed) {
372 unlock_res_and_lock(lock);
373 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
378 if (flags && (lock->l_flags & flags)) {
379 unlock_res_and_lock(lock);
385 lock->l_flags |= flags;
387 unlock_res_and_lock(lock);
394 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
395 struct lustre_handle *handle)
397 struct ldlm_lock *retval = NULL;
398 retval = __ldlm_handle2lock(handle, 0);
402 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
404 ldlm_res2desc(lock->l_resource, &desc->l_resource);
405 desc->l_req_mode = lock->l_req_mode;
406 desc->l_granted_mode = lock->l_granted_mode;
407 memcpy(&desc->l_policy_data, &lock->l_policy_data,
408 sizeof(desc->l_policy_data));
411 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
412 struct list_head *work_list)
414 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
415 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
416 lock->l_flags |= LDLM_FL_AST_SENT;
417 /* If the enqueuing client said so, tell the AST recipient to
418 * discard dirty data, rather than writing back. */
419 if (new->l_flags & LDLM_AST_DISCARD_DATA)
420 lock->l_flags |= LDLM_FL_DISCARD_DATA;
421 LASSERT(list_empty(&lock->l_bl_ast));
422 list_add(&lock->l_bl_ast, work_list);
424 LASSERT(lock->l_blocking_lock == NULL);
425 lock->l_blocking_lock = LDLM_LOCK_GET(new);
429 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
431 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
432 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
433 lock->l_flags |= LDLM_FL_CP_REQD;
434 LASSERT(list_empty(&lock->l_cp_ast));
435 list_add(&lock->l_cp_ast, work_list);
440 /* must be called with lr_lock held */
441 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
442 struct list_head *work_list)
445 check_res_locked(lock->l_resource);
447 ldlm_add_bl_work_item(lock, new, work_list);
449 ldlm_add_cp_work_item(lock, work_list);
453 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
455 struct ldlm_lock *lock;
457 lock = ldlm_handle2lock(lockh);
458 ldlm_lock_addref_internal(lock, mode);
462 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
464 ldlm_lock_remove_from_lru(lock);
465 if (mode & (LCK_NL | LCK_CR | LCK_PR))
467 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
469 lock->l_last_used = jiffies;
471 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
474 /* only called for local locks */
475 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
477 lock_res_and_lock(lock);
478 ldlm_lock_addref_internal_nolock(lock, mode);
479 unlock_res_and_lock(lock);
482 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
484 struct ldlm_namespace *ns;
487 lock_res_and_lock(lock);
489 ns = lock->l_resource->lr_namespace;
491 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
492 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
493 LASSERT(lock->l_readers > 0);
496 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
497 LASSERT(lock->l_writers > 0);
501 if (lock->l_flags & LDLM_FL_LOCAL &&
502 !lock->l_readers && !lock->l_writers) {
503 /* If this is a local lock on a server namespace and this was
504 * the last reference, cancel the lock. */
505 CDEBUG(D_INFO, "forcing cancel of local lock\n");
506 lock->l_flags |= LDLM_FL_CBPENDING;
509 if (!lock->l_readers && !lock->l_writers &&
510 (lock->l_flags & LDLM_FL_CBPENDING)) {
511 /* If we received a blocked AST and this was the last reference,
512 * run the callback. */
513 if (ns->ns_client == LDLM_NAMESPACE_SERVER && lock->l_export)
514 CERROR("FL_CBPENDING set on non-local lock--just a "
517 LDLM_DEBUG(lock, "final decref done on cbpending lock");
519 LDLM_LOCK_GET(lock); /* dropped by bl thread */
520 ldlm_lock_remove_from_lru(lock);
521 unlock_res_and_lock(lock);
522 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
523 ldlm_bl_to_thread(ns, NULL, lock) != 0)
524 ldlm_handle_bl_callback(ns, NULL, lock);
525 } else if (ns->ns_client == LDLM_NAMESPACE_CLIENT &&
526 !lock->l_readers && !lock->l_writers) {
527 /* If this is a client-side namespace and this was the last
528 * reference, put it on the LRU. */
529 LASSERT(list_empty(&lock->l_lru));
530 LASSERT(ns->ns_nr_unused >= 0);
531 spin_lock(&ns->ns_unused_lock);
532 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
534 spin_unlock(&ns->ns_unused_lock);
535 unlock_res_and_lock(lock);
536 ldlm_cancel_lru(ns, LDLM_ASYNC);
538 unlock_res_and_lock(lock);
541 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
546 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
548 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
549 LASSERTF(lock != NULL, "can't find lockh "LPX64"\n", lockh->cookie);
550 ldlm_lock_decref_internal(lock, mode);
554 /* This will drop a lock reference and mark it for destruction, but will not
555 * necessarily cancel the lock before returning. */
556 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
558 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
561 LASSERT(lock != NULL);
563 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
564 lock_res_and_lock(lock);
565 lock->l_flags |= LDLM_FL_CBPENDING;
566 unlock_res_and_lock(lock);
567 ldlm_lock_decref_internal(lock, mode);
572 * - ldlm_lock_enqueue
573 * - ldlm_reprocess_queue
574 * - ldlm_lock_convert
576 * must be called with lr_lock held
578 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
580 struct ldlm_resource *res = lock->l_resource;
583 check_res_locked(res);
585 lock->l_granted_mode = lock->l_req_mode;
586 ldlm_resource_add_lock(res, &res->lr_granted, lock);
588 if (lock->l_granted_mode < res->lr_most_restr)
589 res->lr_most_restr = lock->l_granted_mode;
591 if (work_list && lock->l_completion_ast != NULL)
592 ldlm_add_ast_work_item(lock, NULL, work_list);
597 /* returns a referenced lock or NULL. See the flag descriptions below, in the
598 * comment above ldlm_lock_match */
599 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
600 ldlm_policy_data_t *policy,
601 struct ldlm_lock *old_lock, int flags)
603 struct ldlm_lock *lock;
604 struct list_head *tmp;
606 list_for_each(tmp, queue) {
607 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
609 if (lock == old_lock)
612 /* llite sometimes wants to match locks that will be
613 * canceled when their users drop, but we allow it to match
614 * if it passes in CBPENDING and the lock still has users.
615 * this is generally only going to be used by children
616 * whose parents already hold a lock so forward progress
617 * can still happen. */
618 if (lock->l_flags & LDLM_FL_CBPENDING &&
619 !(flags & LDLM_FL_CBPENDING))
621 if (lock->l_flags & LDLM_FL_CBPENDING &&
622 lock->l_readers == 0 && lock->l_writers == 0)
625 if (!(lock->l_req_mode & mode))
628 if (lock->l_resource->lr_type == LDLM_EXTENT &&
629 (lock->l_policy_data.l_extent.start >
630 policy->l_extent.start ||
631 lock->l_policy_data.l_extent.end < policy->l_extent.end))
634 if (lock->l_resource->lr_type == LDLM_EXTENT &&
636 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
639 /* We match if we have existing lock with same or wider set
641 if (lock->l_resource->lr_type == LDLM_IBITS &&
642 ((lock->l_policy_data.l_inodebits.bits &
643 policy->l_inodebits.bits) !=
644 policy->l_inodebits.bits))
647 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
650 if ((flags & LDLM_FL_LOCAL_ONLY) &&
651 !(lock->l_flags & LDLM_FL_LOCAL))
654 if (flags & LDLM_FL_TEST_LOCK)
657 ldlm_lock_addref_internal_nolock(lock, mode);
664 void ldlm_lock_allow_match(struct ldlm_lock *lock)
666 lock_res_and_lock(lock);
667 lock->l_flags |= LDLM_FL_CAN_MATCH;
668 wake_up(&lock->l_waitq);
669 unlock_res_and_lock(lock);
672 /* Can be called in two ways:
674 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
675 * for a duplicate of.
677 * Otherwise, all of the fields must be filled in, to match against.
679 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
680 * server (ie, connh is NULL)
681 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
682 * list will be considered
683 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
684 * to be canceled can still be matched as long as they still have reader
685 * or writer refernces
686 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
687 * just tell us if we would have matched.
689 * Returns 1 if it finds an already-existing lock that is compatible; in this
690 * case, lockh is filled in with a addref()ed lock
692 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
693 struct ldlm_res_id *res_id, __u32 type,
694 ldlm_policy_data_t *policy, ldlm_mode_t mode,
695 struct lustre_handle *lockh)
697 struct ldlm_resource *res;
698 struct ldlm_lock *lock, *old_lock = NULL;
703 old_lock = ldlm_handle2lock(lockh);
706 ns = old_lock->l_resource->lr_namespace;
707 res_id = &old_lock->l_resource->lr_name;
708 type = old_lock->l_resource->lr_type;
709 mode = old_lock->l_req_mode;
712 res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
714 LASSERT(old_lock == NULL);
720 lock = search_queue(&res->lr_granted, mode, policy, old_lock, flags);
723 if (flags & LDLM_FL_BLOCK_GRANTED)
725 lock = search_queue(&res->lr_converting, mode, policy, old_lock, flags);
728 lock = search_queue(&res->lr_waiting, mode, policy, old_lock, flags);
735 ldlm_resource_putref(res);
738 ldlm_lock2handle(lock, lockh);
739 if (!(lock->l_flags & LDLM_FL_CAN_MATCH)) {
740 struct l_wait_info lwi;
741 if (lock->l_completion_ast) {
742 int err = lock->l_completion_ast(lock,
743 LDLM_FL_WAIT_NOREPROC,
751 lwi = LWI_TIMEOUT_INTR(obd_timeout*HZ, NULL,NULL,NULL);
753 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
754 l_wait_event(lock->l_waitq,
755 (lock->l_flags & LDLM_FL_CAN_MATCH), &lwi);
761 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
762 type == LDLM_PLAIN ? res_id->name[2] :
763 policy->l_extent.start,
764 type == LDLM_PLAIN ? res_id->name[3] :
765 policy->l_extent.end);
766 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/* less verbose for test-only */
767 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
768 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
769 type, mode, res_id->name[0], res_id->name[1],
770 type == LDLM_PLAIN ? res_id->name[2] :
771 policy->l_extent.start,
772 type == LDLM_PLAIN ? res_id->name[3] :
773 policy->l_extent.end);
777 LDLM_LOCK_PUT(old_lock);
778 if (flags & LDLM_FL_TEST_LOCK && rc)
784 /* Returns a referenced lock */
785 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
786 struct lustre_handle *parent_lock_handle,
787 struct ldlm_res_id res_id, __u32 type,
789 ldlm_blocking_callback blocking,
790 ldlm_completion_callback completion,
791 ldlm_glimpse_callback glimpse,
792 void *data, __u32 lvb_len)
794 struct ldlm_resource *res, *parent_res = NULL;
795 struct ldlm_lock *lock, *parent_lock = NULL;
798 if (parent_lock_handle) {
799 parent_lock = ldlm_handle2lock(parent_lock_handle);
801 parent_res = parent_lock->l_resource;
804 res = ldlm_resource_get(ns, parent_res, res_id,
809 lock = ldlm_lock_new(parent_lock, res);
810 ldlm_resource_putref(res);
811 if (parent_lock != NULL)
812 LDLM_LOCK_PUT(parent_lock);
817 lock->l_req_mode = mode;
818 lock->l_ast_data = data;
819 lock->l_blocking_ast = blocking;
820 lock->l_completion_ast = completion;
821 lock->l_glimpse_ast = glimpse;
822 lock->l_pid = current->pid;
825 lock->l_lvb_len = lvb_len;
826 OBD_ALLOC(lock->l_lvb_data, lvb_len);
827 if (lock->l_lvb_data == NULL) {
828 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
836 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
837 struct ldlm_lock **lockp,
838 void *cookie, int *flags)
840 struct ldlm_lock *lock = *lockp;
841 struct ldlm_resource *res = lock->l_resource;
842 int local = res->lr_namespace->ns_client;
843 ldlm_processing_policy policy;
844 ldlm_error_t rc = ELDLM_OK;
847 /* policies are not executed on the client or during replay */
848 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
849 && !local && ns->ns_policy) {
850 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
852 if (rc == ELDLM_LOCK_REPLACED) {
853 /* The lock that was returned has already been granted,
854 * and placed into lockp. If it's not the same as the
855 * one we passed in, then destroy the old one and our
856 * work here is done. */
857 if (lock != *lockp) {
858 ldlm_lock_destroy(lock);
861 *flags |= LDLM_FL_LOCK_CHANGED;
863 } else if (rc == ELDLM_LOCK_ABORTED ||
864 (rc == 0 && (*flags & LDLM_FL_INTENT_ONLY))) {
865 ldlm_lock_destroy(lock);
868 LASSERT(rc == ELDLM_OK);
871 lock_res_and_lock(lock);
872 if (local && lock->l_req_mode == lock->l_granted_mode) {
873 /* The server returned a blocked lock, but it was granted before
874 * we got a chance to actually enqueue it. We don't need to do
876 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
877 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
881 /* Some flags from the enqueue want to make it into the AST, via the
883 lock->l_flags |= (*flags & LDLM_AST_DISCARD_DATA);
885 /* This distinction between local lock trees is very important; a client
886 * namespace only has information about locks taken by that client, and
887 * thus doesn't have enough information to decide for itself if it can
888 * be granted (below). In this case, we do exactly what the server
889 * tells us to do, as dictated by the 'flags'.
891 * We do exactly the same thing during recovery, when the server is
892 * more or less trusting the clients not to lie.
894 * FIXME (bug 268): Detect obvious lies by checking compatibility in
895 * granted/converting queues. */
896 ldlm_resource_unlink_lock(lock);
898 if (*flags & LDLM_FL_BLOCK_CONV)
899 ldlm_resource_add_lock(res, &res->lr_converting, lock);
900 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
901 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
903 ldlm_grant_lock(lock, NULL);
905 } else if (*flags & LDLM_FL_REPLAY) {
906 if (*flags & LDLM_FL_BLOCK_CONV) {
907 ldlm_resource_add_lock(res, &res->lr_converting, lock);
909 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
910 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
912 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
913 ldlm_grant_lock(lock, NULL);
916 /* If no flags, fall through to normal enqueue path. */
919 policy = ldlm_processing_policy_table[res->lr_type];
920 policy(lock, flags, 1, &rc, NULL);
923 unlock_res_and_lock(lock);
927 /* Must be called with namespace taken: queue is waiting or converting. */
928 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
929 struct list_head *work_list)
931 struct list_head *tmp, *pos;
932 ldlm_processing_policy policy;
934 int rc = LDLM_ITER_CONTINUE;
938 check_res_locked(res);
940 policy = ldlm_processing_policy_table[res->lr_type];
943 list_for_each_safe(tmp, pos, queue) {
944 struct ldlm_lock *pending;
945 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
947 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
950 rc = policy(pending, &flags, 0, &err, work_list);
951 if (rc != LDLM_ITER_CONTINUE)
958 int ldlm_run_bl_ast_work(struct list_head *rpc_list)
960 struct list_head *tmp, *pos;
961 struct ldlm_lock_desc d;
962 int rc = 0, retval = 0;
965 list_for_each_safe(tmp, pos, rpc_list) {
966 struct ldlm_lock *lock =
967 list_entry(tmp, struct ldlm_lock, l_bl_ast);
969 /* nobody should touch l_bl_ast */
970 lock_res_and_lock(lock);
971 list_del_init(&lock->l_bl_ast);
973 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
974 LASSERT(lock->l_bl_ast_run == 0);
975 LASSERT(lock->l_blocking_lock);
976 lock->l_bl_ast_run++;
977 unlock_res_and_lock(lock);
979 ldlm_lock2desc(lock->l_blocking_lock, &d);
981 LDLM_LOCK_PUT(lock->l_blocking_lock);
982 lock->l_blocking_lock = NULL;
983 rc = lock->l_blocking_ast(lock, &d, NULL, LDLM_CB_BLOCKING);
988 CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
989 "disconnect client\n");
995 int ldlm_run_cp_ast_work(struct list_head *rpc_list)
997 struct list_head *tmp, *pos;
998 int rc = 0, retval = 0;
1001 /* It's possible to receive a completion AST before we've set
1002 * the l_completion_ast pointer: either because the AST arrived
1003 * before the reply, or simply because there's a small race
1004 * window between receiving the reply and finishing the local
1005 * enqueue. (bug 842)
1007 * This can't happen with the blocking_ast, however, because we
1008 * will never call the local blocking_ast until we drop our
1009 * reader/writer reference, which we won't do until we get the
1010 * reply and finish enqueueing. */
1012 list_for_each_safe(tmp, pos, rpc_list) {
1013 struct ldlm_lock *lock =
1014 list_entry(tmp, struct ldlm_lock, l_cp_ast);
1016 /* nobody should touch l_cp_ast */
1017 lock_res_and_lock(lock);
1018 list_del_init(&lock->l_cp_ast);
1019 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1020 lock->l_flags &= ~LDLM_FL_CP_REQD;
1021 unlock_res_and_lock(lock);
1023 if (lock->l_completion_ast != NULL)
1024 rc = lock->l_completion_ast(lock, 0, 0);
1026 if (rc == -ERESTART)
1029 CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
1030 "disconnect client\n");
1031 LDLM_LOCK_PUT(lock);
1036 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1038 ldlm_reprocess_all(res);
1039 return LDLM_ITER_CONTINUE;
1042 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1044 struct list_head *tmp;
1047 spin_lock(&ns->ns_hash_lock);
1048 for (i = 0; i < RES_HASH_SIZE; i++) {
1049 tmp = ns->ns_hash[i].next;
1050 while (tmp != &(ns->ns_hash[i])) {
1051 struct ldlm_resource *res =
1052 list_entry(tmp, struct ldlm_resource, lr_hash);
1054 ldlm_resource_getref(res);
1055 spin_unlock(&ns->ns_hash_lock);
1057 rc = reprocess_one_queue(res, NULL);
1059 spin_lock(&ns->ns_hash_lock);
1061 ldlm_resource_putref_locked(res);
1063 if (rc == LDLM_ITER_STOP)
1068 spin_unlock(&ns->ns_hash_lock);
1072 void ldlm_reprocess_all(struct ldlm_resource *res)
1074 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1078 /* Local lock trees don't get reprocessed. */
1079 if (res->lr_namespace->ns_client) {
1086 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1087 if (rc == LDLM_ITER_CONTINUE)
1088 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1091 rc = ldlm_run_cp_ast_work(&rpc_list);
1092 if (rc == -ERESTART) {
1093 LASSERT(list_empty(&rpc_list));
1099 void ldlm_cancel_callback(struct ldlm_lock *lock)
1101 check_res_locked(lock->l_resource);
1103 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1104 lock->l_flags |= LDLM_FL_CANCEL;
1105 if (lock->l_blocking_ast) {
1106 unlock_res_and_lock(lock);
1107 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1109 lock_res_and_lock(lock);
1111 LDLM_DEBUG(lock, "no blocking ast");
1116 void ldlm_lock_cancel(struct ldlm_lock *lock)
1118 struct ldlm_resource *res;
1119 struct ldlm_namespace *ns;
1122 ldlm_del_waiting_lock(lock);
1123 lock_res_and_lock(lock);
1125 res = lock->l_resource;
1126 ns = res->lr_namespace;
1128 /* Please do not, no matter how tempting, remove this LBUG without
1129 * talking to me first. -phik */
1130 if (lock->l_readers || lock->l_writers) {
1131 LDLM_ERROR(lock, "lock still has references");
1135 ldlm_cancel_callback(lock);
1137 ldlm_resource_unlink_lock(lock);
1138 unlock_res_and_lock(lock);
1140 ldlm_lock_destroy(lock);
1145 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1147 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1153 lock->l_ast_data = data;
1154 LDLM_LOCK_PUT(lock);
1158 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1160 struct ldlm_lock *lock;
1161 struct ldlm_resource *res;
1163 spin_lock(&exp->exp_ldlm_data.led_lock);
1164 while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1165 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1166 struct ldlm_lock, l_export_chain);
1167 res = ldlm_resource_getref(lock->l_resource);
1168 LDLM_LOCK_GET(lock);
1169 spin_unlock(&exp->exp_ldlm_data.led_lock);
1171 LDLM_DEBUG(lock, "export %p", exp);
1172 ldlm_lock_cancel(lock);
1173 ldlm_reprocess_all(res);
1175 ldlm_resource_putref(res);
1176 LDLM_LOCK_PUT(lock);
1177 spin_lock(&exp->exp_ldlm_data.led_lock);
1179 spin_unlock(&exp->exp_ldlm_data.led_lock);
1182 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1185 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1186 struct ldlm_resource *res;
1187 struct ldlm_namespace *ns;
1193 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1194 *flags |= LDLM_FL_BLOCK_GRANTED;
1195 RETURN(lock->l_resource);
1198 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1199 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1201 lock_res_and_lock(lock);
1203 res = lock->l_resource;
1204 ns = res->lr_namespace;
1206 old_mode = lock->l_req_mode;
1207 lock->l_req_mode = new_mode;
1208 ldlm_resource_unlink_lock(lock);
1210 /* If this is a local resource, put it on the appropriate list. */
1211 if (res->lr_namespace->ns_client) {
1212 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1213 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1215 /* This should never happen, because of the way the
1216 * server handles conversions. */
1217 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1221 ldlm_grant_lock(lock, &rpc_list);
1223 /* FIXME: completion handling not with ns_lock held ! */
1224 if (lock->l_completion_ast)
1225 lock->l_completion_ast(lock, 0, NULL);
1229 ldlm_processing_policy policy;
1230 policy = ldlm_processing_policy_table[res->lr_type];
1231 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1232 if (rc == LDLM_ITER_STOP) {
1233 lock->l_req_mode = old_mode;
1234 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1237 *flags |= LDLM_FL_BLOCK_GRANTED;
1241 unlock_res_and_lock(lock);
1244 ldlm_run_cp_ast_work(&rpc_list);
1248 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1250 char str[PTL_NALFMT_SIZE];
1251 struct obd_device *obd = NULL;
1253 if (!((portal_debug | D_ERROR) & level))
1257 CDEBUG(level, " NULL LDLM lock\n");
1261 CDEBUG(level, " -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1262 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1264 if (lock->l_conn_export != NULL)
1265 obd = lock->l_conn_export->exp_obd;
1266 if (lock->l_export && lock->l_export->exp_connection) {
1267 CDEBUG(level, " Node: NID %s on %s (rhandle: "LPX64")\n",
1268 ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str),
1269 lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1270 lock->l_remote_handle.cookie);
1271 } else if (obd == NULL) {
1272 CDEBUG(level, " Node: local\n");
1274 struct obd_import *imp = obd->u.cli.cl_import;
1275 CDEBUG(level, " Node: NID %s on %s (rhandle: "LPX64")\n",
1276 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1277 imp->imp_connection->c_peer.peer_ni->pni_name,
1278 lock->l_remote_handle.cookie);
1280 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource,
1281 lock->l_resource->lr_name.name[0],
1282 lock->l_resource->lr_name.name[1]);
1283 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1284 "write: %d\n", ldlm_lockname[lock->l_req_mode],
1285 ldlm_lockname[lock->l_granted_mode],
1286 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers);
1287 if (lock->l_resource->lr_type == LDLM_EXTENT)
1288 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1289 " (req "LPU64"-"LPU64")\n",
1290 lock->l_policy_data.l_extent.start,
1291 lock->l_policy_data.l_extent.end,
1292 lock->l_req_extent.start, lock->l_req_extent.end);
1293 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1294 CDEBUG(level, " Pid: "LPU64" Extent: "LPU64" -> "LPU64"\n",
1295 lock->l_policy_data.l_flock.pid,
1296 lock->l_policy_data.l_flock.start,
1297 lock->l_policy_data.l_flock.end);
1298 else if (lock->l_resource->lr_type == LDLM_IBITS)
1299 CDEBUG(level, " Bits: "LPX64"\n",
1300 lock->l_policy_data.l_inodebits.bits);
1303 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1305 struct ldlm_lock *lock;
1307 lock = ldlm_handle2lock(lockh);
1311 ldlm_lock_dump(D_OTHER, lock, 0);
1313 LDLM_LOCK_PUT(lock);