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 the Lustre file system, http://www.lustre.org
9 * Lustre is a trademark of Cluster File Systems, Inc.
11 * You may have signed or agreed to another license before downloading
12 * this software. If so, you are bound by the terms and conditions
13 * of that agreement, and the following does not apply to you. See the
14 * LICENSE file included with this distribution for more information.
16 * If you did not agree to a different license, then this copy of Lustre
17 * is open source software; you can redistribute it and/or modify it
18 * under the terms of version 2 of the GNU General Public License as
19 * published by the Free Software Foundation.
21 * In either case, Lustre is distributed in the hope that it will be
22 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
23 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
24 * license text for more details.
27 #define DEBUG_SUBSYSTEM S_LDLM
30 # include <libcfs/libcfs.h>
31 # include <linux/lustre_intent.h>
33 # include <liblustre.h>
34 # include <libcfs/kp30.h>
37 #include <obd_class.h>
38 #include "ldlm_internal.h"
40 //struct lustre_lock ldlm_everything_lock;
42 /* lock's skip list pointers fix mode */
43 #define LDLM_JOIN_NONE 0
44 #define LDLM_MODE_JOIN_RIGHT 1
45 #define LDLM_MODE_JOIN_LEFT (1 << 1)
46 #define LDLM_POLICY_JOIN_RIGHT (1 << 2)
47 #define LDLM_POLICY_JOIN_LEFT (1 << 3)
50 char *ldlm_lockname[] = {
61 char *ldlm_typename[] = {
68 char *ldlm_it2str(int it)
75 case (IT_OPEN | IT_CREAT):
88 CERROR("Unknown intent %d\n", it);
93 extern cfs_mem_cache_t *ldlm_lock_slab;
95 static ldlm_processing_policy ldlm_processing_policy_table[] = {
96 [LDLM_PLAIN] ldlm_process_plain_lock,
97 [LDLM_EXTENT] ldlm_process_extent_lock,
99 [LDLM_FLOCK] ldlm_process_flock_lock,
101 [LDLM_IBITS] ldlm_process_inodebits_lock,
104 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
106 return ldlm_processing_policy_table[res->lr_type];
109 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
115 * REFCOUNTED LOCK OBJECTS
120 * Lock refcounts, during creation:
121 * - one special one for allocation, dec'd only once in destroy
122 * - one for being a lock that's in-use
123 * - one for the addref associated with a new lock
125 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
127 atomic_inc(&lock->l_refc);
131 void ldlm_lock_put(struct ldlm_lock *lock)
135 LASSERT(lock->l_resource != LP_POISON);
136 LASSERT(atomic_read(&lock->l_refc) > 0);
137 if (atomic_dec_and_test(&lock->l_refc)) {
138 struct ldlm_resource *res;
140 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing it.");
142 res = lock->l_resource;
143 LASSERT(lock->l_destroyed);
144 LASSERT(list_empty(&lock->l_res_link));
145 LASSERT(list_empty(&lock->l_pending_chain));
148 LDLM_LOCK_PUT(lock->l_parent);
150 atomic_dec(&res->lr_namespace->ns_locks);
151 ldlm_resource_putref(res);
152 lock->l_resource = NULL;
154 class_export_put(lock->l_export);
156 if (lock->l_lvb_data != NULL)
157 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
159 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
165 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
168 if (!list_empty(&lock->l_lru)) {
169 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
170 list_del_init(&lock->l_lru);
171 lock->l_resource->lr_namespace->ns_nr_unused--;
172 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
178 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
182 spin_lock(&lock->l_resource->lr_namespace->ns_unused_lock);
183 rc = ldlm_lock_remove_from_lru_nolock(lock);
184 spin_unlock(&lock->l_resource->lr_namespace->ns_unused_lock);
189 /* This used to have a 'strict' flag, which recovery would use to mark an
190 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
191 * shall explain why it's gone: with the new hash table scheme, once you call
192 * ldlm_lock_destroy, you can never drop your final references on this lock.
193 * Because it's not in the hash table anymore. -phil */
194 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
198 if (!list_empty(&lock->l_children)) {
199 LDLM_ERROR(lock, "still has children (%p)!",
200 lock->l_children.next);
201 ldlm_lock_dump(D_ERROR, lock, 0);
204 if (lock->l_readers || lock->l_writers) {
205 LDLM_ERROR(lock, "lock still has references");
206 ldlm_lock_dump(D_ERROR, lock, 0);
210 if (!list_empty(&lock->l_res_link)) {
211 LDLM_ERROR(lock, "lock still on resource");
212 ldlm_lock_dump(D_ERROR, lock, 0);
216 if (lock->l_destroyed) {
217 LASSERT(list_empty(&lock->l_lru));
221 lock->l_destroyed = 1;
224 spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
225 list_del_init(&lock->l_export_chain);
227 spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
229 ldlm_lock_remove_from_lru(lock);
230 class_handle_unhash(&lock->l_handle);
233 /* Wake anyone waiting for this lock */
234 /* FIXME: I should probably add yet another flag, instead of using
235 * l_export to only call this on clients */
237 class_export_put(lock->l_export);
238 lock->l_export = NULL;
239 if (lock->l_export && lock->l_completion_ast)
240 lock->l_completion_ast(lock, 0);
246 void ldlm_lock_destroy(struct ldlm_lock *lock)
250 lock_res_and_lock(lock);
251 first = ldlm_lock_destroy_internal(lock);
252 unlock_res_and_lock(lock);
254 /* drop reference from hashtable only for first destroy */
260 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
264 first = ldlm_lock_destroy_internal(lock);
265 /* drop reference from hashtable only for first destroy */
271 /* this is called by portals_handle2object with the handle lock taken */
272 static void lock_handle_addref(void *lock)
274 LDLM_LOCK_GET((struct ldlm_lock *)lock);
278 * usage: pass in a resource on which you have done ldlm_resource_get
279 * pass in a parent lock on which you have done a ldlm_lock_get
280 * after return, ldlm_*_put the resource and parent
281 * returns: lock with refcount 2 - one for current caller and one for remote
283 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
284 struct ldlm_resource *resource)
286 struct ldlm_lock *lock;
289 if (resource == NULL)
292 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, CFS_ALLOC_IO, sizeof(*lock));
296 lock->l_resource = ldlm_resource_getref(resource);
298 atomic_set(&lock->l_refc, 2);
299 CFS_INIT_LIST_HEAD(&lock->l_children);
300 CFS_INIT_LIST_HEAD(&lock->l_res_link);
301 CFS_INIT_LIST_HEAD(&lock->l_lru);
302 CFS_INIT_LIST_HEAD(&lock->l_export_chain);
303 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
304 CFS_INIT_LIST_HEAD(&lock->l_tmp);
305 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
306 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
307 cfs_waitq_init(&lock->l_waitq);
308 lock->l_blocking_lock = NULL;
310 lock->l_sl_mode.prev = NULL;
311 lock->l_sl_mode.next = NULL;
312 lock->l_sl_policy.prev = NULL;
313 lock->l_sl_policy.next = NULL;
315 atomic_inc(&resource->lr_namespace->ns_locks);
317 if (parent != NULL) {
318 spin_lock(&resource->lr_namespace->ns_hash_lock);
319 lock->l_parent = LDLM_LOCK_GET(parent);
320 list_add(&lock->l_childof, &parent->l_children);
321 spin_unlock(&resource->lr_namespace->ns_hash_lock);
324 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
325 class_handle_hash(&lock->l_handle, lock_handle_addref);
330 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
331 struct ldlm_res_id new_resid)
333 struct ldlm_resource *oldres = lock->l_resource;
334 struct ldlm_resource *newres;
338 LASSERT(ns->ns_client != 0);
340 lock_res_and_lock(lock);
341 if (memcmp(&new_resid, &lock->l_resource->lr_name,
342 sizeof(lock->l_resource->lr_name)) == 0) {
344 unlock_res_and_lock(lock);
348 LASSERT(new_resid.name[0] != 0);
350 /* This function assumes that the lock isn't on any lists */
351 LASSERT(list_empty(&lock->l_res_link));
353 type = oldres->lr_type;
354 unlock_res_and_lock(lock);
356 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
357 if (newres == NULL) {
362 lock_res_and_lock(lock);
363 LASSERT(memcmp(&new_resid, &lock->l_resource->lr_name,
364 sizeof(lock->l_resource->lr_name)) != 0);
366 lock->l_resource = newres;
369 unlock_bitlock(lock);
371 /* ...and the flowers are still standing! */
372 ldlm_resource_putref(oldres);
381 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
383 lockh->cookie = lock->l_handle.h_cookie;
386 /* if flags: atomically get the lock and set the flags.
387 * Return NULL if flag already set
390 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
392 struct ldlm_namespace *ns;
393 struct ldlm_lock *lock = NULL, *retval = NULL;
398 lock = class_handle2object(handle->cookie);
402 LASSERT(lock->l_resource != NULL);
403 ns = lock->l_resource->lr_namespace;
406 lock_res_and_lock(lock);
408 /* It's unlikely but possible that someone marked the lock as
409 * destroyed after we did handle2object on it */
410 if (lock->l_destroyed) {
411 unlock_res_and_lock(lock);
412 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
417 if (flags && (lock->l_flags & flags)) {
418 unlock_res_and_lock(lock);
424 lock->l_flags |= flags;
426 unlock_res_and_lock(lock);
433 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
434 struct lustre_handle *handle)
436 struct ldlm_lock *retval = NULL;
437 retval = __ldlm_handle2lock(handle, 0);
441 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
443 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
444 /* INODEBITS_INTEROP: If the other side does not support
445 * inodebits, reply with a plain lock descriptor.
447 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
448 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
449 struct ldlm_resource res = *lock->l_resource;
451 /* Make sure all the right bits are set in this lock we
452 are going to pass to client */
453 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
454 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
455 "Inappropriate inode lock bits during "
456 "conversion " LPU64 "\n",
457 lock->l_policy_data.l_inodebits.bits);
458 res.lr_type = LDLM_PLAIN;
459 ldlm_res2desc(&res, &desc->l_resource);
460 /* Convert "new" lock mode to something old client can
462 if ((lock->l_req_mode == LCK_CR) ||
463 (lock->l_req_mode == LCK_CW))
464 desc->l_req_mode = LCK_PR;
466 desc->l_req_mode = lock->l_req_mode;
467 if ((lock->l_granted_mode == LCK_CR) ||
468 (lock->l_granted_mode == LCK_CW)) {
469 desc->l_granted_mode = LCK_PR;
471 /* We never grant PW/EX locks to clients */
472 LASSERT((lock->l_granted_mode != LCK_PW) &&
473 (lock->l_granted_mode != LCK_EX));
474 desc->l_granted_mode = lock->l_granted_mode;
477 /* We do not copy policy here, because there is no
478 policy for plain locks */
480 ldlm_res2desc(lock->l_resource, &desc->l_resource);
481 desc->l_req_mode = lock->l_req_mode;
482 desc->l_granted_mode = lock->l_granted_mode;
483 desc->l_policy_data = lock->l_policy_data;
487 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
488 struct list_head *work_list)
490 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
491 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
492 lock->l_flags |= LDLM_FL_AST_SENT;
493 /* If the enqueuing client said so, tell the AST recipient to
494 * discard dirty data, rather than writing back. */
495 if (new->l_flags & LDLM_AST_DISCARD_DATA)
496 lock->l_flags |= LDLM_FL_DISCARD_DATA;
497 LASSERT(list_empty(&lock->l_bl_ast));
498 list_add(&lock->l_bl_ast, work_list);
500 LASSERT(lock->l_blocking_lock == NULL);
501 lock->l_blocking_lock = LDLM_LOCK_GET(new);
505 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
507 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
508 lock->l_flags |= LDLM_FL_CP_REQD;
509 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
510 LASSERT(list_empty(&lock->l_cp_ast));
511 list_add(&lock->l_cp_ast, work_list);
516 /* must be called with lr_lock held */
517 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
518 struct list_head *work_list)
521 check_res_locked(lock->l_resource);
523 ldlm_add_bl_work_item(lock, new, work_list);
525 ldlm_add_cp_work_item(lock, work_list);
529 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
531 struct ldlm_lock *lock;
533 lock = ldlm_handle2lock(lockh);
534 LASSERT(lock != NULL);
535 ldlm_lock_addref_internal(lock, mode);
539 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
541 ldlm_lock_remove_from_lru(lock);
542 if (mode & (LCK_NL | LCK_CR | LCK_PR))
544 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
546 lock->l_last_used = cfs_time_current();
548 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
551 /* only called for local locks */
552 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
554 lock_res_and_lock(lock);
555 ldlm_lock_addref_internal_nolock(lock, mode);
556 unlock_res_and_lock(lock);
559 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
561 struct ldlm_namespace *ns;
564 lock_res_and_lock(lock);
566 ns = lock->l_resource->lr_namespace;
568 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
569 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
570 LASSERT(lock->l_readers > 0);
573 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
574 LASSERT(lock->l_writers > 0);
578 if (lock->l_flags & LDLM_FL_LOCAL &&
579 !lock->l_readers && !lock->l_writers) {
580 /* If this is a local lock on a server namespace and this was
581 * the last reference, cancel the lock. */
582 CDEBUG(D_INFO, "forcing cancel of local lock\n");
583 lock->l_flags |= LDLM_FL_CBPENDING;
586 if (!lock->l_readers && !lock->l_writers &&
587 (lock->l_flags & LDLM_FL_CBPENDING)) {
588 /* If we received a blocked AST and this was the last reference,
589 * run the callback. */
590 if (ns->ns_client == LDLM_NAMESPACE_SERVER && lock->l_export)
591 CERROR("FL_CBPENDING set on non-local lock--just a "
594 LDLM_DEBUG(lock, "final decref done on cbpending lock");
596 LDLM_LOCK_GET(lock); /* dropped by bl thread */
597 ldlm_lock_remove_from_lru(lock);
598 unlock_res_and_lock(lock);
599 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
600 ldlm_bl_to_thread(ns, NULL, lock) != 0)
601 ldlm_handle_bl_callback(ns, NULL, lock);
602 } else if (ns->ns_client == LDLM_NAMESPACE_CLIENT &&
603 !lock->l_readers && !lock->l_writers &&
604 !(lock->l_flags & LDLM_FL_NO_LRU)) {
605 /* If this is a client-side namespace and this was the last
606 * reference, put it on the LRU. */
607 LASSERT(list_empty(&lock->l_lru));
608 LASSERT(ns->ns_nr_unused >= 0);
609 spin_lock(&ns->ns_unused_lock);
610 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
612 spin_unlock(&ns->ns_unused_lock);
613 unlock_res_and_lock(lock);
614 ldlm_cancel_lru(ns, LDLM_ASYNC);
616 unlock_res_and_lock(lock);
619 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
624 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
626 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
627 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
628 ldlm_lock_decref_internal(lock, mode);
632 /* This will drop a lock reference and mark it for destruction, but will not
633 * necessarily cancel the lock before returning. */
634 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
636 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
639 LASSERT(lock != NULL);
641 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
642 lock_res_and_lock(lock);
643 lock->l_flags |= LDLM_FL_CBPENDING;
644 unlock_res_and_lock(lock);
645 ldlm_lock_decref_internal(lock, mode);
650 * search_granted_lock
653 * Finds a position to insert the new lock.
655 * queue [input]: the granted list where search acts on;
656 * req [input]: the lock whose position to be located;
657 * lockp [output]: the position where the lock should be inserted before, or
658 * NULL indicating @req should be appended to @queue.
660 * Bit-masks combination of following values indicating in which way the
661 * lock need to be inserted.
662 * - LDLM_JOIN_NONE: noting about skip list needs to be fixed;
663 * - LDLM_MODE_JOIN_RIGHT: @req needs join right becoming the head of a
665 * - LDLM_POLICY_JOIN_RIGHT: @req needs join right becoming the head of
668 * - ldlm_grant_lock_with_skiplist
670 static int search_granted_lock(struct list_head *queue,
671 struct ldlm_lock *req,
672 struct ldlm_lock **lockp)
674 struct list_head *tmp, *tmp_tail;
675 struct ldlm_lock *lock, *mode_head_lock;
676 int rc = LDLM_JOIN_NONE;
679 list_for_each(tmp, queue) {
680 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
682 if (lock->l_req_mode != req->l_req_mode) {
683 if (LDLM_SL_HEAD(&lock->l_sl_mode))
684 tmp = &list_entry(lock->l_sl_mode.next,
686 l_sl_mode)->l_res_link;
690 /* found the same mode group */
691 if (lock->l_resource->lr_type == LDLM_PLAIN) {
693 rc = LDLM_MODE_JOIN_RIGHT;
695 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
697 if (LDLM_SL_HEAD(&lock->l_sl_mode))
698 tmp_tail = &list_entry(lock->l_sl_mode.next,
700 l_sl_mode)->l_res_link;
701 mode_head_lock = lock;
703 if (lock->l_policy_data.l_inodebits.bits ==
704 req->l_policy_data.l_inodebits.bits) {
705 /* matched policy lock is found */
707 rc |= LDLM_POLICY_JOIN_RIGHT;
709 /* if the policy group head is also a
710 * mode group head or a single mode
712 if (LDLM_SL_HEAD(&lock->l_sl_mode) ||
714 LDLM_SL_EMPTY(&lock->l_sl_mode)))
715 rc |= LDLM_MODE_JOIN_RIGHT;
719 if (LDLM_SL_HEAD(&lock->l_sl_policy))
720 tmp = &list_entry(lock->l_sl_policy.next,
722 l_sl_policy)->l_res_link;
728 lock = list_entry(tmp, struct ldlm_lock,
730 } /* for all locks in the matched mode group */
732 /* no matched policy group is found, insert before
733 * the mode group head lock */
734 *lockp = mode_head_lock;
735 rc = LDLM_MODE_JOIN_RIGHT;
738 LDLM_ERROR(lock, "is not LDLM_PLAIN or LDLM_IBITS lock");
743 /* no matched mode group is found, append to the end */
751 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
752 struct ldlm_lock *lockp,
755 struct ldlm_resource *res = lock->l_resource;
758 LASSERT(lockp || join == LDLM_JOIN_NONE);
760 check_res_locked(res);
762 ldlm_resource_dump(D_OTHER, res);
763 CDEBUG(D_OTHER, "About to add this lock:\n");
764 ldlm_lock_dump(D_OTHER, lock, 0);
766 if (lock->l_destroyed) {
767 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
771 LASSERT(list_empty(&lock->l_res_link));
774 list_add_tail(&lock->l_res_link, &lock->l_resource->lr_granted);
775 else if ((join & LDLM_MODE_JOIN_LEFT) || (join & LDLM_POLICY_JOIN_LEFT))
776 list_add(&lock->l_res_link, &lockp->l_res_link);
778 list_add_tail(&lock->l_res_link, &lockp->l_res_link);
781 if (join & LDLM_MODE_JOIN_RIGHT) {
782 LASSERT(! LDLM_SL_TAIL(&lockp->l_sl_mode));
783 if (LDLM_SL_EMPTY(&lockp->l_sl_mode)) {
784 lock->l_sl_mode.next = &lockp->l_sl_mode;
785 lockp->l_sl_mode.prev = &lock->l_sl_mode;
786 } else if (LDLM_SL_HEAD(&lockp->l_sl_mode)) {
787 lock->l_sl_mode.next = lockp->l_sl_mode.next;
788 lockp->l_sl_mode.next = NULL;
789 lock->l_sl_mode.next->prev = &lock->l_sl_mode;
791 } else if (join & LDLM_MODE_JOIN_LEFT) {
792 LASSERT(! LDLM_SL_HEAD(&lockp->l_sl_mode));
793 if (LDLM_SL_EMPTY(&lockp->l_sl_mode)) {
794 lock->l_sl_mode.prev = &lockp->l_sl_mode;
795 lockp->l_sl_mode.next = &lock->l_sl_mode;
796 } else if (LDLM_SL_TAIL(&lockp->l_sl_mode)) {
797 lock->l_sl_mode.prev = lockp->l_sl_mode.prev;
798 lockp->l_sl_mode.prev = NULL;
799 lock->l_sl_mode.prev->next = &lock->l_sl_mode;
803 if (join & LDLM_POLICY_JOIN_RIGHT) {
804 LASSERT(! LDLM_SL_TAIL(&lockp->l_sl_policy));
805 if (LDLM_SL_EMPTY(&lockp->l_sl_policy)) {
806 lock->l_sl_policy.next = &lockp->l_sl_policy;
807 lockp->l_sl_policy.prev = &lock->l_sl_policy;
808 } else if (LDLM_SL_HEAD(&lockp->l_sl_policy)) {
809 lock->l_sl_policy.next = lockp->l_sl_policy.next;
810 lockp->l_sl_policy.next = NULL;
811 lock->l_sl_policy.next->prev = &lock->l_sl_policy;
813 } else if (join & LDLM_POLICY_JOIN_LEFT) {
814 LASSERT(! LDLM_SL_HEAD(&lockp->l_sl_policy));
815 if (LDLM_SL_EMPTY(&lockp->l_sl_policy)) {
816 lock->l_sl_policy.prev = &lockp->l_sl_policy;
817 lockp->l_sl_policy.next = &lock->l_sl_policy;
818 } else if (LDLM_SL_TAIL(&lockp->l_sl_policy)) {
819 lock->l_sl_policy.prev = lockp->l_sl_policy.prev;
820 lockp->l_sl_policy.prev = NULL;
821 lock->l_sl_policy.prev->next = &lock->l_sl_policy;
828 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
830 int join = LDLM_JOIN_NONE;
831 struct ldlm_lock *lockp = NULL;
834 LASSERT(lock->l_req_mode == lock->l_granted_mode);
836 join = search_granted_lock(&lock->l_resource->lr_granted, lock, &lockp);
837 ldlm_granted_list_add_lock(lock, lockp, join);
842 * - ldlm_lock_enqueue
843 * - ldlm_reprocess_queue
844 * - ldlm_lock_convert
846 * must be called with lr_lock held
848 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
850 struct ldlm_resource *res = lock->l_resource;
853 check_res_locked(res);
855 lock->l_granted_mode = lock->l_req_mode;
856 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
857 ldlm_grant_lock_with_skiplist(lock);
859 ldlm_resource_add_lock(res, &res->lr_granted, lock);
861 if (lock->l_granted_mode < res->lr_most_restr)
862 res->lr_most_restr = lock->l_granted_mode;
864 if (work_list && lock->l_completion_ast != NULL)
865 ldlm_add_ast_work_item(lock, NULL, work_list);
870 /* returns a referenced lock or NULL. See the flag descriptions below, in the
871 * comment above ldlm_lock_match */
872 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
873 ldlm_policy_data_t *policy,
874 struct ldlm_lock *old_lock, int flags)
876 struct ldlm_lock *lock;
877 struct list_head *tmp;
879 list_for_each(tmp, queue) {
880 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
882 if (lock == old_lock)
885 /* llite sometimes wants to match locks that will be
886 * canceled when their users drop, but we allow it to match
887 * if it passes in CBPENDING and the lock still has users.
888 * this is generally only going to be used by children
889 * whose parents already hold a lock so forward progress
890 * can still happen. */
891 if (lock->l_flags & LDLM_FL_CBPENDING &&
892 !(flags & LDLM_FL_CBPENDING))
894 if (lock->l_flags & LDLM_FL_CBPENDING &&
895 lock->l_readers == 0 && lock->l_writers == 0)
898 if (!(lock->l_req_mode & mode))
901 if (lock->l_resource->lr_type == LDLM_EXTENT &&
902 (lock->l_policy_data.l_extent.start >
903 policy->l_extent.start ||
904 lock->l_policy_data.l_extent.end < policy->l_extent.end))
907 if (unlikely(mode == LCK_GROUP) &&
908 lock->l_resource->lr_type == LDLM_EXTENT &&
909 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
912 /* We match if we have existing lock with same or wider set
914 if (lock->l_resource->lr_type == LDLM_IBITS &&
915 ((lock->l_policy_data.l_inodebits.bits &
916 policy->l_inodebits.bits) !=
917 policy->l_inodebits.bits))
920 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
923 if ((flags & LDLM_FL_LOCAL_ONLY) &&
924 !(lock->l_flags & LDLM_FL_LOCAL))
927 if (flags & LDLM_FL_TEST_LOCK)
930 ldlm_lock_addref_internal_nolock(lock, mode);
937 void ldlm_lock_allow_match(struct ldlm_lock *lock)
939 lock_res_and_lock(lock);
940 lock->l_flags |= LDLM_FL_LVB_READY;
941 cfs_waitq_signal(&lock->l_waitq);
942 unlock_res_and_lock(lock);
945 /* Can be called in two ways:
947 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
948 * for a duplicate of.
950 * Otherwise, all of the fields must be filled in, to match against.
952 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
953 * server (ie, connh is NULL)
954 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
955 * list will be considered
956 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
957 * to be canceled can still be matched as long as they still have reader
958 * or writer refernces
959 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
960 * just tell us if we would have matched.
962 * Returns 1 if it finds an already-existing lock that is compatible; in this
963 * case, lockh is filled in with a addref()ed lock
965 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
966 struct ldlm_res_id *res_id, ldlm_type_t type,
967 ldlm_policy_data_t *policy, ldlm_mode_t mode,
968 struct lustre_handle *lockh)
970 struct ldlm_resource *res;
971 struct ldlm_lock *lock, *old_lock = NULL;
976 old_lock = ldlm_handle2lock(lockh);
979 ns = old_lock->l_resource->lr_namespace;
980 res_id = &old_lock->l_resource->lr_name;
981 type = old_lock->l_resource->lr_type;
982 mode = old_lock->l_req_mode;
985 res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
987 LASSERT(old_lock == NULL);
993 lock = search_queue(&res->lr_granted, mode, policy, old_lock, flags);
996 if (flags & LDLM_FL_BLOCK_GRANTED)
998 lock = search_queue(&res->lr_converting, mode, policy, old_lock, flags);
1001 lock = search_queue(&res->lr_waiting, mode, policy, old_lock, flags);
1008 ldlm_resource_putref(res);
1011 ldlm_lock2handle(lock, lockh);
1012 if ((flags & LDLM_FL_LVB_READY) && (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1013 struct l_wait_info lwi;
1014 if (lock->l_completion_ast) {
1015 int err = lock->l_completion_ast(lock,
1016 LDLM_FL_WAIT_NOREPROC,
1019 if (flags & LDLM_FL_TEST_LOCK)
1020 LDLM_LOCK_PUT(lock);
1022 ldlm_lock_decref_internal(lock, mode);
1028 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), NULL,
1029 LWI_ON_SIGNAL_NOOP, NULL);
1031 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1032 l_wait_event(lock->l_waitq,
1033 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1038 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1039 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1040 res_id->name[2] : policy->l_extent.start,
1041 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1042 res_id->name[3] : policy->l_extent.end);
1043 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1044 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1045 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1046 type, mode, res_id->name[0], res_id->name[1],
1047 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1048 res_id->name[2] :policy->l_extent.start,
1049 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1050 res_id->name[3] : policy->l_extent.end);
1053 LDLM_LOCK_PUT(old_lock);
1054 if (flags & LDLM_FL_TEST_LOCK && rc)
1055 LDLM_LOCK_PUT(lock);
1060 /* Returns a referenced lock */
1061 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1062 struct lustre_handle *parent_lock_handle,
1063 struct ldlm_res_id res_id, ldlm_type_t type,
1065 ldlm_blocking_callback blocking,
1066 ldlm_completion_callback completion,
1067 ldlm_glimpse_callback glimpse,
1068 void *data, __u32 lvb_len)
1070 struct ldlm_resource *res, *parent_res = NULL;
1071 struct ldlm_lock *lock, *parent_lock = NULL;
1074 if (parent_lock_handle) {
1075 parent_lock = ldlm_handle2lock(parent_lock_handle);
1077 parent_res = parent_lock->l_resource;
1080 res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
1084 lock = ldlm_lock_new(parent_lock, res);
1085 ldlm_resource_putref(res);
1086 if (parent_lock != NULL)
1087 LDLM_LOCK_PUT(parent_lock);
1092 lock->l_req_mode = mode;
1093 lock->l_ast_data = data;
1094 lock->l_blocking_ast = blocking;
1095 lock->l_completion_ast = completion;
1096 lock->l_glimpse_ast = glimpse;
1097 lock->l_pid = cfs_curproc_pid();
1100 lock->l_lvb_len = lvb_len;
1101 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1102 if (lock->l_lvb_data == NULL) {
1103 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1111 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1112 struct ldlm_lock **lockp,
1113 void *cookie, int *flags)
1115 struct ldlm_lock *lock = *lockp;
1116 struct ldlm_resource *res = lock->l_resource;
1117 int local = res->lr_namespace->ns_client;
1118 ldlm_processing_policy policy;
1119 ldlm_error_t rc = ELDLM_OK;
1122 do_gettimeofday(&lock->l_enqueued_time);
1123 /* policies are not executed on the client or during replay */
1124 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1125 && !local && ns->ns_policy) {
1126 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1128 if (rc == ELDLM_LOCK_REPLACED) {
1129 /* The lock that was returned has already been granted,
1130 * and placed into lockp. If it's not the same as the
1131 * one we passed in, then destroy the old one and our
1132 * work here is done. */
1133 if (lock != *lockp) {
1134 ldlm_lock_destroy(lock);
1135 LDLM_LOCK_PUT(lock);
1137 *flags |= LDLM_FL_LOCK_CHANGED;
1139 } else if (rc != ELDLM_OK ||
1140 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1141 ldlm_lock_destroy(lock);
1146 lock_res_and_lock(lock);
1147 if (local && lock->l_req_mode == lock->l_granted_mode) {
1148 /* The server returned a blocked lock, but it was granted before
1149 * we got a chance to actually enqueue it. We don't need to do
1151 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1152 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1153 GOTO(out, ELDLM_OK);
1156 /* Some flags from the enqueue want to make it into the AST, via the
1157 * lock's l_flags. */
1158 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1160 /* This distinction between local lock trees is very important; a client
1161 * namespace only has information about locks taken by that client, and
1162 * thus doesn't have enough information to decide for itself if it can
1163 * be granted (below). In this case, we do exactly what the server
1164 * tells us to do, as dictated by the 'flags'.
1166 * We do exactly the same thing during recovery, when the server is
1167 * more or less trusting the clients not to lie.
1169 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1170 * granted/converting queues. */
1171 ldlm_resource_unlink_lock(lock);
1173 if (*flags & LDLM_FL_BLOCK_CONV)
1174 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1175 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1176 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1178 ldlm_grant_lock(lock, NULL);
1179 GOTO(out, ELDLM_OK);
1180 } else if (*flags & LDLM_FL_REPLAY) {
1181 if (*flags & LDLM_FL_BLOCK_CONV) {
1182 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1183 GOTO(out, ELDLM_OK);
1184 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1185 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1186 GOTO(out, ELDLM_OK);
1187 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1188 ldlm_grant_lock(lock, NULL);
1189 GOTO(out, ELDLM_OK);
1191 /* If no flags, fall through to normal enqueue path. */
1194 policy = ldlm_processing_policy_table[res->lr_type];
1195 policy(lock, flags, 1, &rc, NULL);
1198 unlock_res_and_lock(lock);
1202 /* Must be called with namespace taken: queue is waiting or converting. */
1203 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1204 struct list_head *work_list)
1206 struct list_head *tmp, *pos;
1207 ldlm_processing_policy policy;
1209 int rc = LDLM_ITER_CONTINUE;
1213 check_res_locked(res);
1215 policy = ldlm_processing_policy_table[res->lr_type];
1218 list_for_each_safe(tmp, pos, queue) {
1219 struct ldlm_lock *pending;
1220 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1222 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1225 rc = policy(pending, &flags, 0, &err, work_list);
1226 if (rc != LDLM_ITER_CONTINUE)
1233 int ldlm_run_bl_ast_work(struct list_head *rpc_list)
1235 struct list_head *tmp, *pos;
1236 struct ldlm_lock_desc d;
1237 int rc = 0, retval = 0;
1240 list_for_each_safe(tmp, pos, rpc_list) {
1241 struct ldlm_lock *lock =
1242 list_entry(tmp, struct ldlm_lock, l_bl_ast);
1244 /* nobody should touch l_bl_ast */
1245 lock_res_and_lock(lock);
1246 list_del_init(&lock->l_bl_ast);
1248 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1249 LASSERT(lock->l_bl_ast_run == 0);
1250 LASSERT(lock->l_blocking_lock);
1251 lock->l_bl_ast_run++;
1252 unlock_res_and_lock(lock);
1254 ldlm_lock2desc(lock->l_blocking_lock, &d);
1256 LDLM_LOCK_PUT(lock->l_blocking_lock);
1257 lock->l_blocking_lock = NULL;
1258 rc = lock->l_blocking_ast(lock, &d, NULL, LDLM_CB_BLOCKING);
1260 if (rc == -ERESTART)
1263 CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
1264 "disconnect client\n");
1265 LDLM_LOCK_PUT(lock);
1270 int ldlm_run_cp_ast_work(struct list_head *rpc_list)
1272 struct list_head *tmp, *pos;
1273 int rc = 0, retval = 0;
1276 /* It's possible to receive a completion AST before we've set
1277 * the l_completion_ast pointer: either because the AST arrived
1278 * before the reply, or simply because there's a small race
1279 * window between receiving the reply and finishing the local
1280 * enqueue. (bug 842)
1282 * This can't happen with the blocking_ast, however, because we
1283 * will never call the local blocking_ast until we drop our
1284 * reader/writer reference, which we won't do until we get the
1285 * reply and finish enqueueing. */
1287 list_for_each_safe(tmp, pos, rpc_list) {
1288 struct ldlm_lock *lock =
1289 list_entry(tmp, struct ldlm_lock, l_cp_ast);
1291 /* nobody should touch l_cp_ast */
1292 lock_res_and_lock(lock);
1293 list_del_init(&lock->l_cp_ast);
1294 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1295 lock->l_flags &= ~LDLM_FL_CP_REQD;
1296 unlock_res_and_lock(lock);
1298 if (lock->l_completion_ast != NULL)
1299 rc = lock->l_completion_ast(lock, 0, 0);
1300 if (rc == -ERESTART)
1303 CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
1304 "disconnect client\n");
1305 LDLM_LOCK_PUT(lock);
1310 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1312 ldlm_reprocess_all(res);
1313 return LDLM_ITER_CONTINUE;
1316 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1318 struct list_head *tmp;
1322 spin_lock(&ns->ns_hash_lock);
1323 for (i = 0; i < RES_HASH_SIZE; i++) {
1324 tmp = ns->ns_hash[i].next;
1325 while (tmp != &(ns->ns_hash[i])) {
1326 struct ldlm_resource *res =
1327 list_entry(tmp, struct ldlm_resource, lr_hash);
1329 ldlm_resource_getref(res);
1330 spin_unlock(&ns->ns_hash_lock);
1332 rc = reprocess_one_queue(res, NULL);
1334 spin_lock(&ns->ns_hash_lock);
1336 ldlm_resource_putref_locked(res);
1338 if (rc == LDLM_ITER_STOP)
1343 spin_unlock(&ns->ns_hash_lock);
1347 void ldlm_reprocess_all(struct ldlm_resource *res)
1349 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1353 /* Local lock trees don't get reprocessed. */
1354 if (res->lr_namespace->ns_client) {
1361 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1362 if (rc == LDLM_ITER_CONTINUE)
1363 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1366 rc = ldlm_run_cp_ast_work(&rpc_list);
1367 if (rc == -ERESTART) {
1368 LASSERT(list_empty(&rpc_list));
1374 void ldlm_cancel_callback(struct ldlm_lock *lock)
1376 check_res_locked(lock->l_resource);
1377 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1378 lock->l_flags |= LDLM_FL_CANCEL;
1379 if (lock->l_blocking_ast) {
1380 // l_check_no_ns_lock(ns);
1381 unlock_res_and_lock(lock);
1382 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1384 lock_res_and_lock(lock);
1386 LDLM_DEBUG(lock, "no blocking ast");
1391 static void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1393 struct ldlm_lock *lock;
1395 if (LDLM_SL_HEAD(&req->l_sl_mode)) {
1396 lock = list_entry(req->l_res_link.next, struct ldlm_lock,
1398 if (req->l_sl_mode.next == &lock->l_sl_mode) {
1399 lock->l_sl_mode.prev = NULL;
1401 lock->l_sl_mode.next = req->l_sl_mode.next;
1402 lock->l_sl_mode.next->prev = &lock->l_sl_mode;
1404 req->l_sl_mode.next = NULL;
1405 } else if (LDLM_SL_TAIL(&req->l_sl_mode)) {
1406 lock = list_entry(req->l_res_link.prev, struct ldlm_lock,
1408 if (req->l_sl_mode.prev == &lock->l_sl_mode) {
1409 lock->l_sl_mode.next = NULL;
1411 lock->l_sl_mode.prev = req->l_sl_mode.prev;
1412 lock->l_sl_mode.prev->next = &lock->l_sl_mode;
1414 req->l_sl_mode.prev = NULL;
1417 if (LDLM_SL_HEAD(&req->l_sl_policy)) {
1418 lock = list_entry(req->l_res_link.next, struct ldlm_lock,
1420 if (req->l_sl_policy.next == &lock->l_sl_policy) {
1421 lock->l_sl_policy.prev = NULL;
1423 lock->l_sl_policy.next = req->l_sl_policy.next;
1424 lock->l_sl_policy.next->prev = &lock->l_sl_policy;
1426 req->l_sl_policy.next = NULL;
1427 } else if (LDLM_SL_TAIL(&req->l_sl_policy)) {
1428 lock = list_entry(req->l_res_link.prev, struct ldlm_lock,
1430 if (req->l_sl_policy.prev == &lock->l_sl_policy) {
1431 lock->l_sl_policy.next = NULL;
1433 lock->l_sl_policy.prev = req->l_sl_policy.prev;
1434 lock->l_sl_policy.prev->next = &lock->l_sl_policy;
1436 req->l_sl_policy.prev = NULL;
1440 void ldlm_lock_cancel(struct ldlm_lock *lock)
1442 struct ldlm_resource *res;
1443 struct ldlm_namespace *ns;
1446 lock_res_and_lock(lock);
1448 res = lock->l_resource;
1449 ns = res->lr_namespace;
1451 /* Please do not, no matter how tempting, remove this LBUG without
1452 * talking to me first. -phik */
1453 if (lock->l_readers || lock->l_writers) {
1454 LDLM_ERROR(lock, "lock still has references");
1458 ldlm_del_waiting_lock(lock);
1460 /* Releases res lock */
1461 ldlm_cancel_callback(lock);
1463 /* Yes, second time, just in case it was added again while we were
1464 running with no res lock in ldlm_cancel_callback */
1465 ldlm_del_waiting_lock(lock);
1466 if (!(LDLM_SL_EMPTY(&lock->l_sl_mode) &&
1467 LDLM_SL_EMPTY(&lock->l_sl_policy)) &&
1468 (lock->l_resource->lr_type == LDLM_PLAIN ||
1469 lock->l_resource->lr_type == LDLM_IBITS))
1470 ldlm_unlink_lock_skiplist(lock);
1471 ldlm_resource_unlink_lock(lock);
1472 ldlm_lock_destroy_nolock(lock);
1473 unlock_res_and_lock(lock);
1478 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1480 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1486 lock->l_ast_data = data;
1487 LDLM_LOCK_PUT(lock);
1491 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1493 struct ldlm_lock *lock;
1494 struct ldlm_resource *res;
1496 spin_lock(&exp->exp_ldlm_data.led_lock);
1497 while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1498 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1499 struct ldlm_lock, l_export_chain);
1500 res = ldlm_resource_getref(lock->l_resource);
1501 LDLM_LOCK_GET(lock);
1502 spin_unlock(&exp->exp_ldlm_data.led_lock);
1504 LDLM_DEBUG(lock, "export %p", exp);
1505 ldlm_lock_cancel(lock);
1506 ldlm_reprocess_all(res);
1508 ldlm_resource_putref(res);
1509 LDLM_LOCK_PUT(lock);
1510 spin_lock(&exp->exp_ldlm_data.led_lock);
1512 spin_unlock(&exp->exp_ldlm_data.led_lock);
1515 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1518 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1519 struct ldlm_resource *res;
1520 struct ldlm_namespace *ns;
1523 struct ldlm_lock *mark_lock = NULL;
1524 int join= LDLM_JOIN_NONE;
1528 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1529 *flags |= LDLM_FL_BLOCK_GRANTED;
1530 RETURN(lock->l_resource);
1533 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1534 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1536 lock_res_and_lock(lock);
1538 res = lock->l_resource;
1539 ns = res->lr_namespace;
1541 old_mode = lock->l_req_mode;
1542 lock->l_req_mode = new_mode;
1543 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1544 /* remember the lock position where the lock might be
1545 * added back to the granted list later and also
1546 * remember the join mode for skiplist fixing. */
1547 if (LDLM_SL_HEAD(&lock->l_sl_mode))
1548 join = LDLM_MODE_JOIN_RIGHT;
1549 else if (LDLM_SL_TAIL(&lock->l_sl_mode))
1550 join = LDLM_MODE_JOIN_LEFT;
1551 if (LDLM_SL_HEAD(&lock->l_sl_policy))
1552 join |= LDLM_POLICY_JOIN_RIGHT;
1553 else if (LDLM_SL_TAIL(&lock->l_sl_policy))
1554 join |= LDLM_POLICY_JOIN_LEFT;
1556 LASSERT(!((join & LDLM_MODE_JOIN_RIGHT) &&
1557 (join & LDLM_POLICY_JOIN_LEFT)));
1558 LASSERT(!((join & LDLM_MODE_JOIN_LEFT) &&
1559 (join & LDLM_POLICY_JOIN_RIGHT)));
1561 if ((join & LDLM_MODE_JOIN_LEFT) ||
1562 (join & LDLM_POLICY_JOIN_LEFT))
1563 mark_lock = list_entry(lock->l_res_link.prev,
1564 struct ldlm_lock, l_res_link);
1565 else if (lock->l_res_link.next != &res->lr_granted)
1566 mark_lock = list_entry(lock->l_res_link.next,
1567 struct ldlm_lock, l_res_link);
1568 if (join != LDLM_JOIN_NONE)
1569 ldlm_unlink_lock_skiplist(lock);
1571 ldlm_resource_unlink_lock(lock);
1573 /* If this is a local resource, put it on the appropriate list. */
1574 if (res->lr_namespace->ns_client) {
1575 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1576 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1578 /* This should never happen, because of the way the
1579 * server handles conversions. */
1580 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1584 ldlm_grant_lock(lock, &rpc_list);
1586 /* FIXME: completion handling not with ns_lock held ! */
1587 if (lock->l_completion_ast)
1588 lock->l_completion_ast(lock, 0, NULL);
1592 ldlm_processing_policy policy;
1593 policy = ldlm_processing_policy_table[res->lr_type];
1594 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1595 if (rc == LDLM_ITER_STOP) {
1596 lock->l_req_mode = old_mode;
1597 ldlm_granted_list_add_lock(lock, mark_lock, join);
1600 *flags |= LDLM_FL_BLOCK_GRANTED;
1604 unlock_res_and_lock(lock);
1607 ldlm_run_cp_ast_work(&rpc_list);
1611 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1613 struct obd_device *obd = NULL;
1615 if (!((libcfs_debug | D_ERROR) & level))
1619 CDEBUG(level, " NULL LDLM lock\n");
1623 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1624 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1626 if (lock->l_conn_export != NULL)
1627 obd = lock->l_conn_export->exp_obd;
1628 if (lock->l_export && lock->l_export->exp_connection) {
1629 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1630 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1631 lock->l_remote_handle.cookie);
1632 } else if (obd == NULL) {
1633 CDEBUG(level, " Node: local\n");
1635 struct obd_import *imp = obd->u.cli.cl_import;
1636 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1637 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1638 lock->l_remote_handle.cookie);
1640 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource,
1641 lock->l_resource->lr_name.name[0],
1642 lock->l_resource->lr_name.name[1]);
1643 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1644 "write: %d flags: %#x\n", ldlm_lockname[lock->l_req_mode],
1645 ldlm_lockname[lock->l_granted_mode],
1646 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1648 if (lock->l_resource->lr_type == LDLM_EXTENT)
1649 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1650 " (req "LPU64"-"LPU64")\n",
1651 lock->l_policy_data.l_extent.start,
1652 lock->l_policy_data.l_extent.end,
1653 lock->l_req_extent.start, lock->l_req_extent.end);
1654 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1655 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1656 lock->l_policy_data.l_flock.pid,
1657 lock->l_policy_data.l_flock.start,
1658 lock->l_policy_data.l_flock.end);
1659 else if (lock->l_resource->lr_type == LDLM_IBITS)
1660 CDEBUG(level, " Bits: "LPX64"\n",
1661 lock->l_policy_data.l_inodebits.bits);
1664 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1666 struct ldlm_lock *lock;
1668 if (!((libcfs_debug | D_ERROR) & level))
1671 lock = ldlm_handle2lock(lockh);
1675 ldlm_lock_dump(D_OTHER, lock, 0);
1677 LDLM_LOCK_PUT(lock);
1680 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1681 struct libcfs_debug_msg_data *data, const char *fmt,
1685 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1687 va_start(args, fmt);
1688 if (lock->l_resource == NULL) {
1689 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1690 data->msg_fn, data->msg_line, fmt, args,
1691 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1692 "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "
1693 LPX64" expref: %d pid: %u\n", lock,
1694 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1695 lock->l_readers, lock->l_writers,
1696 ldlm_lockname[lock->l_granted_mode],
1697 ldlm_lockname[lock->l_req_mode],
1698 lock->l_flags, lock->l_remote_handle.cookie,
1700 atomic_read(&lock->l_export->exp_refcount) : -99,
1706 switch (lock->l_resource->lr_type) {
1708 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1709 data->msg_fn, data->msg_line, fmt, args,
1710 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1711 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1712 "] (req "LPU64"->"LPU64") flags: %x remote: "LPX64
1713 " expref: %d pid: %u\n",
1714 lock->l_resource->lr_namespace->ns_name, lock,
1715 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1716 lock->l_readers, lock->l_writers,
1717 ldlm_lockname[lock->l_granted_mode],
1718 ldlm_lockname[lock->l_req_mode],
1719 lock->l_resource->lr_name.name[0],
1720 lock->l_resource->lr_name.name[1],
1721 atomic_read(&lock->l_resource->lr_refcount),
1722 ldlm_typename[lock->l_resource->lr_type],
1723 lock->l_policy_data.l_extent.start,
1724 lock->l_policy_data.l_extent.end,
1725 lock->l_req_extent.start, lock->l_req_extent.end,
1726 lock->l_flags, lock->l_remote_handle.cookie,
1728 atomic_read(&lock->l_export->exp_refcount) : -99,
1732 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1733 data->msg_fn, data->msg_line, fmt, args,
1734 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1735 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1736 "["LPU64"->"LPU64"] flags: %x remote: "LPX64
1737 " expref: %d pid: %u\n",
1738 lock->l_resource->lr_namespace->ns_name, lock,
1739 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1740 lock->l_readers, lock->l_writers,
1741 ldlm_lockname[lock->l_granted_mode],
1742 ldlm_lockname[lock->l_req_mode],
1743 lock->l_resource->lr_name.name[0],
1744 lock->l_resource->lr_name.name[1],
1745 atomic_read(&lock->l_resource->lr_refcount),
1746 ldlm_typename[lock->l_resource->lr_type],
1747 lock->l_policy_data.l_flock.pid,
1748 lock->l_policy_data.l_flock.start,
1749 lock->l_policy_data.l_flock.end,
1750 lock->l_flags, lock->l_remote_handle.cookie,
1752 atomic_read(&lock->l_export->exp_refcount) : -99,
1756 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1757 data->msg_fn, data->msg_line, fmt, args,
1758 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1759 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1760 "flags: %x remote: "LPX64" expref: %d "
1762 lock->l_resource->lr_namespace->ns_name,
1763 lock, lock->l_handle.h_cookie,
1764 atomic_read (&lock->l_refc),
1765 lock->l_readers, lock->l_writers,
1766 ldlm_lockname[lock->l_granted_mode],
1767 ldlm_lockname[lock->l_req_mode],
1768 lock->l_resource->lr_name.name[0],
1769 lock->l_resource->lr_name.name[1],
1770 lock->l_policy_data.l_inodebits.bits,
1771 atomic_read(&lock->l_resource->lr_refcount),
1772 ldlm_typename[lock->l_resource->lr_type],
1773 lock->l_flags, lock->l_remote_handle.cookie,
1775 atomic_read(&lock->l_export->exp_refcount) : -99,
1779 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1780 data->msg_fn, data->msg_line, fmt, args,
1781 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1782 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: %x "
1783 "remote: "LPX64" expref: %d pid: %u\n",
1784 lock->l_resource->lr_namespace->ns_name,
1785 lock, lock->l_handle.h_cookie,
1786 atomic_read (&lock->l_refc),
1787 lock->l_readers, lock->l_writers,
1788 ldlm_lockname[lock->l_granted_mode],
1789 ldlm_lockname[lock->l_req_mode],
1790 lock->l_resource->lr_name.name[0],
1791 lock->l_resource->lr_name.name[1],
1792 atomic_read(&lock->l_resource->lr_refcount),
1793 ldlm_typename[lock->l_resource->lr_type],
1794 lock->l_flags, lock->l_remote_handle.cookie,
1796 atomic_read(&lock->l_export->exp_refcount) : -99,
1802 EXPORT_SYMBOL(_ldlm_lock_debug);