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 # ifndef HAVE_VFS_INTENT_PATCHES
32 # include <linux/lustre_intent.h>
35 # include <liblustre.h>
36 # include <libcfs/kp30.h>
39 #include <obd_class.h>
40 #include "ldlm_internal.h"
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 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
133 LASSERT(size == sizeof(*lock));
134 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
137 void ldlm_lock_put(struct ldlm_lock *lock)
141 LASSERT(lock->l_resource != LP_POISON);
142 LASSERT(atomic_read(&lock->l_refc) > 0);
143 if (atomic_dec_and_test(&lock->l_refc)) {
144 struct ldlm_resource *res;
147 "final lock_put on destroyed lock, freeing it.");
149 res = lock->l_resource;
150 LASSERT(lock->l_destroyed);
151 LASSERT(list_empty(&lock->l_res_link));
152 LASSERT(list_empty(&lock->l_pending_chain));
154 atomic_dec(&res->lr_namespace->ns_locks);
155 ldlm_resource_putref(res);
156 lock->l_resource = NULL;
157 if (lock->l_export) {
158 class_export_put(lock->l_export);
159 lock->l_export = NULL;
162 if (lock->l_lvb_data != NULL)
163 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
165 ldlm_interval_free(ldlm_interval_detach(lock));
166 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
173 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
176 if (!list_empty(&lock->l_lru)) {
177 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
178 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
179 list_del_init(&lock->l_lru);
181 LASSERT(ns->ns_nr_unused >= 0);
187 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
189 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
192 spin_lock(&ns->ns_unused_lock);
193 rc = ldlm_lock_remove_from_lru_nolock(lock);
194 spin_unlock(&ns->ns_unused_lock);
199 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
201 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
202 lock->l_last_used = cfs_time_current();
203 LASSERT(list_empty(&lock->l_lru));
204 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
205 LASSERT(ns->ns_nr_unused >= 0);
209 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
211 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
213 spin_lock(&ns->ns_unused_lock);
214 ldlm_lock_add_to_lru_nolock(lock);
215 spin_unlock(&ns->ns_unused_lock);
219 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
221 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
223 spin_lock(&ns->ns_unused_lock);
224 if (!list_empty(&lock->l_lru)) {
225 ldlm_lock_remove_from_lru_nolock(lock);
226 ldlm_lock_add_to_lru_nolock(lock);
228 spin_unlock(&ns->ns_unused_lock);
232 /* This used to have a 'strict' flag, which recovery would use to mark an
233 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
234 * shall explain why it's gone: with the new hash table scheme, once you call
235 * ldlm_lock_destroy, you can never drop your final references on this lock.
236 * Because it's not in the hash table anymore. -phil */
237 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
241 if (lock->l_readers || lock->l_writers) {
242 LDLM_ERROR(lock, "lock still has references");
243 ldlm_lock_dump(D_ERROR, lock, 0);
247 if (!list_empty(&lock->l_res_link)) {
248 LDLM_ERROR(lock, "lock still on resource");
249 ldlm_lock_dump(D_ERROR, lock, 0);
253 if (lock->l_destroyed) {
254 LASSERT(list_empty(&lock->l_lru));
258 lock->l_destroyed = 1;
261 spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
262 list_del_init(&lock->l_export_chain);
264 spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
266 ldlm_lock_remove_from_lru(lock);
267 class_handle_unhash(&lock->l_handle);
270 /* Wake anyone waiting for this lock */
271 /* FIXME: I should probably add yet another flag, instead of using
272 * l_export to only call this on clients */
274 class_export_put(lock->l_export);
275 lock->l_export = NULL;
276 if (lock->l_export && lock->l_completion_ast)
277 lock->l_completion_ast(lock, 0);
283 void ldlm_lock_destroy(struct ldlm_lock *lock)
287 lock_res_and_lock(lock);
288 first = ldlm_lock_destroy_internal(lock);
289 unlock_res_and_lock(lock);
291 /* drop reference from hashtable only for first destroy */
297 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
301 first = ldlm_lock_destroy_internal(lock);
302 /* drop reference from hashtable only for first destroy */
308 /* this is called by portals_handle2object with the handle lock taken */
309 static void lock_handle_addref(void *lock)
311 LDLM_LOCK_GET((struct ldlm_lock *)lock);
315 * usage: pass in a resource on which you have done ldlm_resource_get
316 * pass in a parent lock on which you have done a ldlm_lock_get
317 * after return, ldlm_*_put the resource and parent
318 * returns: lock with refcount 2 - one for current caller and one for remote
320 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
322 struct ldlm_lock *lock;
325 if (resource == NULL)
328 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, CFS_ALLOC_IO, sizeof(*lock));
332 spin_lock_init(&lock->l_lock);
333 lock->l_resource = ldlm_resource_getref(resource);
335 atomic_set(&lock->l_refc, 2);
336 CFS_INIT_LIST_HEAD(&lock->l_res_link);
337 CFS_INIT_LIST_HEAD(&lock->l_lru);
338 CFS_INIT_LIST_HEAD(&lock->l_export_chain);
339 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
340 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
341 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
342 cfs_waitq_init(&lock->l_waitq);
343 lock->l_blocking_lock = NULL;
344 lock->l_sl_mode.prev = NULL;
345 lock->l_sl_mode.next = NULL;
346 lock->l_sl_policy.prev = NULL;
347 lock->l_sl_policy.next = NULL;
349 atomic_inc(&resource->lr_namespace->ns_locks);
350 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
351 class_handle_hash(&lock->l_handle, lock_handle_addref);
356 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
357 const struct ldlm_res_id *new_resid)
359 struct ldlm_resource *oldres = lock->l_resource;
360 struct ldlm_resource *newres;
364 LASSERT(ns_is_client(ns));
366 lock_res_and_lock(lock);
367 if (memcmp(new_resid, &lock->l_resource->lr_name,
368 sizeof(lock->l_resource->lr_name)) == 0) {
370 unlock_res_and_lock(lock);
374 LASSERT(new_resid->name[0] != 0);
376 /* This function assumes that the lock isn't on any lists */
377 LASSERT(list_empty(&lock->l_res_link));
379 type = oldres->lr_type;
380 unlock_res_and_lock(lock);
382 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
386 lock_res_and_lock(lock);
387 LASSERT(memcmp(new_resid, &lock->l_resource->lr_name,
388 sizeof(lock->l_resource->lr_name)) != 0);
390 lock->l_resource = newres;
392 unlock_res_and_lock(lock);
394 /* ...and the flowers are still standing! */
395 ldlm_resource_putref(oldres);
404 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
406 lockh->cookie = lock->l_handle.h_cookie;
409 /* if flags: atomically get the lock and set the flags.
410 * Return NULL if flag already set
413 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
416 struct ldlm_namespace *ns;
417 struct ldlm_lock *lock, *retval = NULL;
422 lock = class_handle2object(handle->cookie);
426 LASSERT(lock->l_resource != NULL);
427 ns = lock->l_resource->lr_namespace;
430 lock_res_and_lock(lock);
432 /* It's unlikely but possible that someone marked the lock as
433 * destroyed after we did handle2object on it */
434 if (lock->l_destroyed) {
435 unlock_res_and_lock(lock);
436 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
441 if (flags && (lock->l_flags & flags)) {
442 unlock_res_and_lock(lock);
448 lock->l_flags |= flags;
450 unlock_res_and_lock(lock);
457 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
458 const struct lustre_handle *handle)
460 struct ldlm_lock *retval = NULL;
461 retval = __ldlm_handle2lock(handle, 0);
465 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
467 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
468 /* INODEBITS_INTEROP: If the other side does not support
469 * inodebits, reply with a plain lock descriptor.
471 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
472 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
473 struct ldlm_resource res = *lock->l_resource;
475 /* Make sure all the right bits are set in this lock we
476 are going to pass to client */
477 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
478 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
479 "Inappropriate inode lock bits during "
480 "conversion " LPU64 "\n",
481 lock->l_policy_data.l_inodebits.bits);
482 res.lr_type = LDLM_PLAIN;
483 ldlm_res2desc(&res, &desc->l_resource);
484 /* Convert "new" lock mode to something old client can
486 if ((lock->l_req_mode == LCK_CR) ||
487 (lock->l_req_mode == LCK_CW))
488 desc->l_req_mode = LCK_PR;
490 desc->l_req_mode = lock->l_req_mode;
491 if ((lock->l_granted_mode == LCK_CR) ||
492 (lock->l_granted_mode == LCK_CW)) {
493 desc->l_granted_mode = LCK_PR;
495 /* We never grant PW/EX locks to clients */
496 LASSERT((lock->l_granted_mode != LCK_PW) &&
497 (lock->l_granted_mode != LCK_EX));
498 desc->l_granted_mode = lock->l_granted_mode;
501 /* We do not copy policy here, because there is no
502 policy for plain locks */
504 ldlm_res2desc(lock->l_resource, &desc->l_resource);
505 desc->l_req_mode = lock->l_req_mode;
506 desc->l_granted_mode = lock->l_granted_mode;
507 desc->l_policy_data = lock->l_policy_data;
511 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
512 struct list_head *work_list)
514 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
515 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
516 lock->l_flags |= LDLM_FL_AST_SENT;
517 /* If the enqueuing client said so, tell the AST recipient to
518 * discard dirty data, rather than writing back. */
519 if (new->l_flags & LDLM_AST_DISCARD_DATA)
520 lock->l_flags |= LDLM_FL_DISCARD_DATA;
521 LASSERT(list_empty(&lock->l_bl_ast));
522 list_add(&lock->l_bl_ast, work_list);
524 LASSERT(lock->l_blocking_lock == NULL);
525 lock->l_blocking_lock = LDLM_LOCK_GET(new);
529 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
531 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
532 lock->l_flags |= LDLM_FL_CP_REQD;
533 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
534 LASSERT(list_empty(&lock->l_cp_ast));
535 list_add(&lock->l_cp_ast, work_list);
540 /* must be called with lr_lock held */
541 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
542 struct list_head *work_list)
545 check_res_locked(lock->l_resource);
547 ldlm_add_bl_work_item(lock, new, work_list);
549 ldlm_add_cp_work_item(lock, work_list);
553 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
555 struct ldlm_lock *lock;
557 lock = ldlm_handle2lock(lockh);
558 LASSERT(lock != NULL);
559 ldlm_lock_addref_internal(lock, mode);
563 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
565 ldlm_lock_remove_from_lru(lock);
566 if (mode & (LCK_NL | LCK_CR | LCK_PR))
568 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
571 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
574 /* only called for local locks */
575 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
577 lock_res_and_lock(lock);
578 ldlm_lock_addref_internal_nolock(lock, mode);
579 unlock_res_and_lock(lock);
582 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
584 struct ldlm_namespace *ns;
587 lock_res_and_lock(lock);
589 ns = lock->l_resource->lr_namespace;
591 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
592 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
593 LASSERT(lock->l_readers > 0);
596 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
597 LASSERT(lock->l_writers > 0);
601 if (lock->l_flags & LDLM_FL_LOCAL &&
602 !lock->l_readers && !lock->l_writers) {
603 /* If this is a local lock on a server namespace and this was
604 * the last reference, cancel the lock. */
605 CDEBUG(D_INFO, "forcing cancel of local lock\n");
606 lock->l_flags |= LDLM_FL_CBPENDING;
609 if (!lock->l_readers && !lock->l_writers &&
610 (lock->l_flags & LDLM_FL_CBPENDING)) {
611 /* If we received a blocked AST and this was the last reference,
612 * run the callback. */
613 if (ns_is_server(ns) && lock->l_export)
614 CERROR("FL_CBPENDING set on non-local lock--just a "
617 LDLM_DEBUG(lock, "final decref done on cbpending lock");
619 LDLM_LOCK_GET(lock); /* dropped by bl thread */
620 ldlm_lock_remove_from_lru(lock);
621 unlock_res_and_lock(lock);
622 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
623 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
624 ldlm_handle_bl_callback(ns, NULL, lock);
625 } else if (ns_is_client(ns) &&
626 !lock->l_readers && !lock->l_writers &&
627 !(lock->l_flags & LDLM_FL_NO_LRU) &&
628 !(lock->l_flags & LDLM_FL_BL_AST)) {
629 /* If this is a client-side namespace and this was the last
630 * reference, put it on the LRU. */
631 ldlm_lock_add_to_lru(lock);
632 unlock_res_and_lock(lock);
633 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
634 * are not supported by the server, otherwise, it is done on
636 if (!exp_connect_cancelset(lock->l_conn_export) &&
637 !ns_connect_lru_resize(ns))
638 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
640 unlock_res_and_lock(lock);
643 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
648 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
650 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
651 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
652 ldlm_lock_decref_internal(lock, mode);
656 /* This will drop a lock reference and mark it for destruction, but will not
657 * necessarily cancel the lock before returning. */
658 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
660 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
663 LASSERT(lock != NULL);
665 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
666 lock_res_and_lock(lock);
667 lock->l_flags |= LDLM_FL_CBPENDING;
668 unlock_res_and_lock(lock);
669 ldlm_lock_decref_internal(lock, mode);
674 * search_granted_lock
677 * Finds a position to insert the new lock.
679 * queue [input]: the granted list where search acts on;
680 * req [input]: the lock whose position to be located;
681 * lockp [output]: the position where the lock should be inserted before, or
682 * NULL indicating @req should be appended to @queue.
684 * Bit-masks combination of following values indicating in which way the
685 * lock need to be inserted.
686 * - LDLM_JOIN_NONE: noting about skip list needs to be fixed;
687 * - LDLM_MODE_JOIN_RIGHT: @req needs join right becoming the head of a
689 * - LDLM_POLICY_JOIN_RIGHT: @req needs join right becoming the head of
692 * - ldlm_grant_lock_with_skiplist
694 static int search_granted_lock(struct list_head *queue,
695 struct ldlm_lock *req,
696 struct ldlm_lock **lockp)
698 struct list_head *tmp, *tmp_tail;
699 struct ldlm_lock *lock, *mode_head_lock;
700 int rc = LDLM_JOIN_NONE;
703 list_for_each(tmp, queue) {
704 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
706 if (lock->l_req_mode != req->l_req_mode) {
707 if (LDLM_SL_HEAD(&lock->l_sl_mode))
708 tmp = &list_entry(lock->l_sl_mode.next,
710 l_sl_mode)->l_res_link;
714 /* found the same mode group */
715 if (lock->l_resource->lr_type == LDLM_PLAIN) {
717 rc = LDLM_MODE_JOIN_RIGHT;
719 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
721 if (LDLM_SL_HEAD(&lock->l_sl_mode))
722 tmp_tail = &list_entry(lock->l_sl_mode.next,
724 l_sl_mode)->l_res_link;
725 mode_head_lock = lock;
727 if (lock->l_policy_data.l_inodebits.bits ==
728 req->l_policy_data.l_inodebits.bits) {
729 /* matched policy lock is found */
731 rc |= LDLM_POLICY_JOIN_RIGHT;
733 /* if the policy group head is also a
734 * mode group head or a single mode
736 if (LDLM_SL_HEAD(&lock->l_sl_mode) ||
738 LDLM_SL_EMPTY(&lock->l_sl_mode)))
739 rc |= LDLM_MODE_JOIN_RIGHT;
743 if (LDLM_SL_HEAD(&lock->l_sl_policy))
744 tmp = &list_entry(lock->l_sl_policy.next,
746 l_sl_policy)->l_res_link;
752 lock = list_entry(tmp, struct ldlm_lock,
754 } /* for all locks in the matched mode group */
756 /* no matched policy group is found, insert before
757 * the mode group head lock */
758 *lockp = mode_head_lock;
759 rc = LDLM_MODE_JOIN_RIGHT;
762 LDLM_ERROR(lock, "is not LDLM_PLAIN or LDLM_IBITS lock");
767 /* no matched mode group is found, append to the end */
775 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
776 struct ldlm_lock *lockp,
779 struct ldlm_resource *res = lock->l_resource;
782 LASSERT(lockp || join == LDLM_JOIN_NONE);
784 check_res_locked(res);
786 ldlm_resource_dump(D_OTHER, res);
787 CDEBUG(D_OTHER, "About to add this lock:\n");
788 ldlm_lock_dump(D_OTHER, lock, 0);
790 if (lock->l_destroyed) {
791 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
795 LASSERT(list_empty(&lock->l_res_link));
798 list_add_tail(&lock->l_res_link, &lock->l_resource->lr_granted);
799 else if ((join & LDLM_MODE_JOIN_LEFT) || (join & LDLM_POLICY_JOIN_LEFT))
800 list_add(&lock->l_res_link, &lockp->l_res_link);
802 list_add_tail(&lock->l_res_link, &lockp->l_res_link);
805 if (join & LDLM_MODE_JOIN_RIGHT) {
806 LASSERT(! LDLM_SL_TAIL(&lockp->l_sl_mode));
807 if (LDLM_SL_EMPTY(&lockp->l_sl_mode)) {
808 lock->l_sl_mode.next = &lockp->l_sl_mode;
809 lockp->l_sl_mode.prev = &lock->l_sl_mode;
810 } else if (LDLM_SL_HEAD(&lockp->l_sl_mode)) {
811 lock->l_sl_mode.next = lockp->l_sl_mode.next;
812 lockp->l_sl_mode.next = NULL;
813 lock->l_sl_mode.next->prev = &lock->l_sl_mode;
815 } else if (join & LDLM_MODE_JOIN_LEFT) {
816 LASSERT(! LDLM_SL_HEAD(&lockp->l_sl_mode));
817 if (LDLM_SL_EMPTY(&lockp->l_sl_mode)) {
818 lock->l_sl_mode.prev = &lockp->l_sl_mode;
819 lockp->l_sl_mode.next = &lock->l_sl_mode;
820 } else if (LDLM_SL_TAIL(&lockp->l_sl_mode)) {
821 lock->l_sl_mode.prev = lockp->l_sl_mode.prev;
822 lockp->l_sl_mode.prev = NULL;
823 lock->l_sl_mode.prev->next = &lock->l_sl_mode;
827 if (join & LDLM_POLICY_JOIN_RIGHT) {
828 LASSERT(! LDLM_SL_TAIL(&lockp->l_sl_policy));
829 if (LDLM_SL_EMPTY(&lockp->l_sl_policy)) {
830 lock->l_sl_policy.next = &lockp->l_sl_policy;
831 lockp->l_sl_policy.prev = &lock->l_sl_policy;
832 } else if (LDLM_SL_HEAD(&lockp->l_sl_policy)) {
833 lock->l_sl_policy.next = lockp->l_sl_policy.next;
834 lockp->l_sl_policy.next = NULL;
835 lock->l_sl_policy.next->prev = &lock->l_sl_policy;
837 } else if (join & LDLM_POLICY_JOIN_LEFT) {
838 LASSERT(! LDLM_SL_HEAD(&lockp->l_sl_policy));
839 if (LDLM_SL_EMPTY(&lockp->l_sl_policy)) {
840 lock->l_sl_policy.prev = &lockp->l_sl_policy;
841 lockp->l_sl_policy.next = &lock->l_sl_policy;
842 } else if (LDLM_SL_TAIL(&lockp->l_sl_policy)) {
843 lock->l_sl_policy.prev = lockp->l_sl_policy.prev;
844 lockp->l_sl_policy.prev = NULL;
845 lock->l_sl_policy.prev->next = &lock->l_sl_policy;
852 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
854 int join = LDLM_JOIN_NONE;
855 struct ldlm_lock *lockp = NULL;
858 LASSERT(lock->l_req_mode == lock->l_granted_mode);
860 join = search_granted_lock(&lock->l_resource->lr_granted, lock, &lockp);
861 ldlm_granted_list_add_lock(lock, lockp, join);
866 * - ldlm_lock_enqueue
867 * - ldlm_reprocess_queue
868 * - ldlm_lock_convert
870 * must be called with lr_lock held
872 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
874 struct ldlm_resource *res = lock->l_resource;
877 check_res_locked(res);
879 lock->l_granted_mode = lock->l_req_mode;
880 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
881 ldlm_grant_lock_with_skiplist(lock);
882 else if (res->lr_type == LDLM_EXTENT)
883 ldlm_extent_add_lock(res, lock);
885 ldlm_resource_add_lock(res, &res->lr_granted, lock);
887 if (lock->l_granted_mode < res->lr_most_restr)
888 res->lr_most_restr = lock->l_granted_mode;
890 if (work_list && lock->l_completion_ast != NULL)
891 ldlm_add_ast_work_item(lock, NULL, work_list);
893 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
897 /* returns a referenced lock or NULL. See the flag descriptions below, in the
898 * comment above ldlm_lock_match */
899 static struct ldlm_lock *search_queue(struct list_head *queue,
901 ldlm_policy_data_t *policy,
902 struct ldlm_lock *old_lock, int flags)
904 struct ldlm_lock *lock;
905 struct list_head *tmp;
907 list_for_each(tmp, queue) {
910 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
912 if (lock == old_lock)
915 /* llite sometimes wants to match locks that will be
916 * canceled when their users drop, but we allow it to match
917 * if it passes in CBPENDING and the lock still has users.
918 * this is generally only going to be used by children
919 * whose parents already hold a lock so forward progress
920 * can still happen. */
921 if (lock->l_flags & LDLM_FL_CBPENDING &&
922 !(flags & LDLM_FL_CBPENDING))
924 if (lock->l_flags & LDLM_FL_CBPENDING &&
925 lock->l_readers == 0 && lock->l_writers == 0)
928 if (!(lock->l_req_mode & *mode))
930 match = lock->l_req_mode;
932 if (lock->l_resource->lr_type == LDLM_EXTENT &&
933 (lock->l_policy_data.l_extent.start >
934 policy->l_extent.start ||
935 lock->l_policy_data.l_extent.end < policy->l_extent.end))
938 if (unlikely(match == LCK_GROUP) &&
939 lock->l_resource->lr_type == LDLM_EXTENT &&
940 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
943 /* We match if we have existing lock with same or wider set
945 if (lock->l_resource->lr_type == LDLM_IBITS &&
946 ((lock->l_policy_data.l_inodebits.bits &
947 policy->l_inodebits.bits) !=
948 policy->l_inodebits.bits))
951 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
954 if ((flags & LDLM_FL_LOCAL_ONLY) &&
955 !(lock->l_flags & LDLM_FL_LOCAL))
958 if (flags & LDLM_FL_TEST_LOCK) {
960 ldlm_lock_touch_in_lru(lock);
962 ldlm_lock_addref_internal_nolock(lock, match);
971 void ldlm_lock_allow_match(struct ldlm_lock *lock)
973 lock_res_and_lock(lock);
974 lock->l_flags |= LDLM_FL_LVB_READY;
975 cfs_waitq_signal(&lock->l_waitq);
976 unlock_res_and_lock(lock);
979 /* Can be called in two ways:
981 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
982 * for a duplicate of.
984 * Otherwise, all of the fields must be filled in, to match against.
986 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
987 * server (ie, connh is NULL)
988 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
989 * list will be considered
990 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
991 * to be canceled can still be matched as long as they still have reader
992 * or writer refernces
993 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
994 * just tell us if we would have matched.
996 * Returns 1 if it finds an already-existing lock that is compatible; in this
997 * case, lockh is filled in with a addref()ed lock
999 * we also check security context, if that failed we simply return 0 (to keep
1000 * caller code unchanged), the context failure will be discovered by caller
1003 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1004 const struct ldlm_res_id *res_id, ldlm_type_t type,
1005 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1006 struct lustre_handle *lockh)
1008 struct ldlm_resource *res;
1009 struct ldlm_lock *lock, *old_lock = NULL;
1014 old_lock = ldlm_handle2lock(lockh);
1017 ns = old_lock->l_resource->lr_namespace;
1018 res_id = &old_lock->l_resource->lr_name;
1019 type = old_lock->l_resource->lr_type;
1020 mode = old_lock->l_req_mode;
1023 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1025 LASSERT(old_lock == NULL);
1031 lock = search_queue(&res->lr_granted, &mode, policy, old_lock, flags);
1034 if (flags & LDLM_FL_BLOCK_GRANTED)
1036 lock = search_queue(&res->lr_converting, &mode, policy, old_lock, flags);
1039 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock, flags);
1046 ldlm_resource_putref(res);
1049 ldlm_lock2handle(lock, lockh);
1050 if ((flags & LDLM_FL_LVB_READY) &&
1051 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1052 struct l_wait_info lwi;
1053 if (lock->l_completion_ast) {
1054 int err = lock->l_completion_ast(lock,
1055 LDLM_FL_WAIT_NOREPROC,
1058 if (flags & LDLM_FL_TEST_LOCK)
1059 LDLM_LOCK_PUT(lock);
1061 ldlm_lock_decref_internal(lock, mode);
1067 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), NULL,
1068 LWI_ON_SIGNAL_NOOP, NULL);
1070 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1071 l_wait_event(lock->l_waitq,
1072 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1077 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1078 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1079 res_id->name[2] : policy->l_extent.start,
1080 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1081 res_id->name[3] : policy->l_extent.end);
1083 /* check user's security context */
1084 if (lock->l_conn_export &&
1085 sptlrpc_import_check_ctx(
1086 class_exp2cliimp(lock->l_conn_export))) {
1087 if (!(flags & LDLM_FL_TEST_LOCK))
1088 ldlm_lock_decref_internal(lock, mode);
1092 if (flags & LDLM_FL_TEST_LOCK)
1093 LDLM_LOCK_PUT(lock);
1095 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1096 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1097 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1098 type, mode, res_id->name[0], res_id->name[1],
1099 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1100 res_id->name[2] :policy->l_extent.start,
1101 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1102 res_id->name[3] : policy->l_extent.end);
1105 LDLM_LOCK_PUT(old_lock);
1107 return rc ? mode : 0;
1110 /* Returns a referenced lock */
1111 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1112 const struct ldlm_res_id *res_id,
1115 ldlm_blocking_callback blocking,
1116 ldlm_completion_callback completion,
1117 ldlm_glimpse_callback glimpse,
1118 void *data, __u32 lvb_len)
1120 struct ldlm_lock *lock;
1121 struct ldlm_resource *res;
1124 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1128 lock = ldlm_lock_new(res);
1129 ldlm_resource_putref(res);
1134 lock->l_req_mode = mode;
1135 lock->l_ast_data = data;
1136 lock->l_blocking_ast = blocking;
1137 lock->l_completion_ast = completion;
1138 lock->l_glimpse_ast = glimpse;
1139 lock->l_pid = cfs_curproc_pid();
1141 lock->l_tree_node = NULL;
1142 /* if this is the extent lock, allocate the interval tree node */
1143 if (type == LDLM_EXTENT) {
1144 if (ldlm_interval_alloc(lock) == NULL)
1149 lock->l_lvb_len = lvb_len;
1150 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1151 if (lock->l_lvb_data == NULL)
1158 if (lock->l_lvb_data)
1159 OBD_FREE(lock->l_lvb_data, lvb_len);
1160 ldlm_interval_free(ldlm_interval_detach(lock));
1161 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1165 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1166 struct ldlm_lock **lockp,
1167 void *cookie, int *flags)
1169 struct ldlm_lock *lock = *lockp;
1170 struct ldlm_resource *res = lock->l_resource;
1171 int local = ns_is_client(res->lr_namespace);
1172 ldlm_processing_policy policy;
1173 ldlm_error_t rc = ELDLM_OK;
1174 struct ldlm_interval *node = NULL;
1177 do_gettimeofday(&lock->l_enqueued_time);
1178 /* policies are not executed on the client or during replay */
1179 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1180 && !local && ns->ns_policy) {
1181 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1183 if (rc == ELDLM_LOCK_REPLACED) {
1184 /* The lock that was returned has already been granted,
1185 * and placed into lockp. If it's not the same as the
1186 * one we passed in, then destroy the old one and our
1187 * work here is done. */
1188 if (lock != *lockp) {
1189 ldlm_lock_destroy(lock);
1190 LDLM_LOCK_PUT(lock);
1192 *flags |= LDLM_FL_LOCK_CHANGED;
1194 } else if (rc != ELDLM_OK ||
1195 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1196 ldlm_lock_destroy(lock);
1201 /* For a replaying lock, it might be already in granted list. So
1202 * unlinking the lock will cause the interval node to be freed, we
1203 * have to allocate the interval node early otherwise we can't regrant
1204 * this lock in the future. - jay */
1205 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1206 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO,
1209 lock_res_and_lock(lock);
1210 if (local && lock->l_req_mode == lock->l_granted_mode) {
1211 /* The server returned a blocked lock, but it was granted
1212 * before we got a chance to actually enqueue it. We don't
1213 * need to do anything else. */
1214 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1215 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1216 GOTO(out, ELDLM_OK);
1219 ldlm_resource_unlink_lock(lock);
1220 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1222 ldlm_lock_destroy_nolock(lock);
1223 GOTO(out, rc = -ENOMEM);
1226 CFS_INIT_LIST_HEAD(&node->li_group);
1227 ldlm_interval_attach(node, lock);
1231 /* Some flags from the enqueue want to make it into the AST, via the
1232 * lock's l_flags. */
1233 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1235 /* This distinction between local lock trees is very important; a client
1236 * namespace only has information about locks taken by that client, and
1237 * thus doesn't have enough information to decide for itself if it can
1238 * be granted (below). In this case, we do exactly what the server
1239 * tells us to do, as dictated by the 'flags'.
1241 * We do exactly the same thing during recovery, when the server is
1242 * more or less trusting the clients not to lie.
1244 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1245 * granted/converting queues. */
1247 if (*flags & LDLM_FL_BLOCK_CONV)
1248 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1249 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1250 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1252 ldlm_grant_lock(lock, NULL);
1253 GOTO(out, ELDLM_OK);
1254 } else if (*flags & LDLM_FL_REPLAY) {
1255 if (*flags & LDLM_FL_BLOCK_CONV) {
1256 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1257 GOTO(out, ELDLM_OK);
1258 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1259 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1260 GOTO(out, ELDLM_OK);
1261 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1262 ldlm_grant_lock(lock, NULL);
1263 GOTO(out, ELDLM_OK);
1265 /* If no flags, fall through to normal enqueue path. */
1268 policy = ldlm_processing_policy_table[res->lr_type];
1269 policy(lock, flags, 1, &rc, NULL);
1272 unlock_res_and_lock(lock);
1274 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1278 /* Must be called with namespace taken: queue is waiting or converting. */
1279 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1280 struct list_head *work_list)
1282 struct list_head *tmp, *pos;
1283 ldlm_processing_policy policy;
1285 int rc = LDLM_ITER_CONTINUE;
1289 check_res_locked(res);
1291 policy = ldlm_processing_policy_table[res->lr_type];
1294 list_for_each_safe(tmp, pos, queue) {
1295 struct ldlm_lock *pending;
1296 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1298 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1301 rc = policy(pending, &flags, 0, &err, work_list);
1302 if (rc != LDLM_ITER_CONTINUE)
1309 /* Helper function for ldlm_run_ast_work().
1311 * Send an existing rpc set specified by @arg->set and then
1312 * destroy it. Create new one if @do_create flag is set. */
1314 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1318 ptlrpc_set_wait(arg->set);
1319 if (arg->type == LDLM_BL_CALLBACK)
1320 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1321 ptlrpc_set_destroy(arg->set);
1324 arg->set = ptlrpc_prep_set();
1330 ldlm_work_bl_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1332 struct ldlm_lock_desc d;
1333 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_bl_ast);
1336 /* nobody should touch l_bl_ast */
1337 lock_res_and_lock(lock);
1338 list_del_init(&lock->l_bl_ast);
1340 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1341 LASSERT(lock->l_bl_ast_run == 0);
1342 LASSERT(lock->l_blocking_lock);
1343 lock->l_bl_ast_run++;
1344 unlock_res_and_lock(lock);
1346 ldlm_lock2desc(lock->l_blocking_lock, &d);
1348 LDLM_LOCK_PUT(lock->l_blocking_lock);
1349 lock->l_blocking_lock = NULL;
1350 lock->l_blocking_ast(lock, &d, (void *)arg,
1352 LDLM_LOCK_PUT(lock);
1358 ldlm_work_cp_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1360 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_cp_ast);
1361 ldlm_completion_callback completion_callback;
1365 /* It's possible to receive a completion AST before we've set
1366 * the l_completion_ast pointer: either because the AST arrived
1367 * before the reply, or simply because there's a small race
1368 * window between receiving the reply and finishing the local
1369 * enqueue. (bug 842)
1371 * This can't happen with the blocking_ast, however, because we
1372 * will never call the local blocking_ast until we drop our
1373 * reader/writer reference, which we won't do until we get the
1374 * reply and finish enqueueing. */
1376 /* nobody should touch l_cp_ast */
1377 lock_res_and_lock(lock);
1378 list_del_init(&lock->l_cp_ast);
1379 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1380 /* save l_completion_ast since it can be changed by
1381 * mds_intent_policy(), see bug 14225 */
1382 completion_callback = lock->l_completion_ast;
1383 lock->l_flags &= ~LDLM_FL_CP_REQD;
1384 unlock_res_and_lock(lock);
1386 if (completion_callback != NULL) {
1387 completion_callback(lock, 0, (void *)arg);
1390 LDLM_LOCK_PUT(lock);
1396 ldlm_work_revoke_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1398 struct ldlm_lock_desc desc;
1399 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_export_chain);
1402 list_del_init(&lock->l_export_chain);
1404 /* the desc just pretend to exclusive */
1405 ldlm_lock2desc(lock, &desc);
1406 desc.l_req_mode = LCK_EX;
1407 desc.l_granted_mode = 0;
1409 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1410 LDLM_LOCK_PUT(lock);
1415 int ldlm_run_ast_work(struct list_head *rpc_list, ldlm_desc_ast_t ast_type)
1417 struct ldlm_cb_set_arg arg;
1418 struct list_head *tmp, *pos;
1419 int (*work_ast_lock)(struct list_head *tmp, struct ldlm_cb_set_arg *arg);
1423 arg.set = ptlrpc_prep_set();
1424 atomic_set(&arg.restart, 0);
1426 case LDLM_WORK_BL_AST:
1427 arg.type = LDLM_BL_CALLBACK;
1428 work_ast_lock = ldlm_work_bl_ast_lock;
1430 case LDLM_WORK_CP_AST:
1431 arg.type = LDLM_CP_CALLBACK;
1432 work_ast_lock = ldlm_work_cp_ast_lock;
1434 case LDLM_WORK_REVOKE_AST:
1435 arg.type = LDLM_BL_CALLBACK;
1436 work_ast_lock = ldlm_work_revoke_ast_lock;
1443 list_for_each_safe(tmp, pos, rpc_list) {
1444 ast_count += work_ast_lock(tmp, &arg);
1446 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1447 * and create a new set for requests that remained in
1449 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1450 ldlm_send_and_maybe_create_set(&arg, 1);
1456 ldlm_send_and_maybe_create_set(&arg, 0);
1458 /* In case when number of ASTs is multiply of
1459 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1460 * @arg.set must be destroyed here, otherwise we get
1461 * write memory leaking. */
1462 ptlrpc_set_destroy(arg.set);
1464 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1467 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1469 ldlm_reprocess_all(res);
1470 return LDLM_ITER_CONTINUE;
1473 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1475 struct list_head *tmp;
1482 spin_lock(&ns->ns_hash_lock);
1483 for (i = 0; i < RES_HASH_SIZE; i++) {
1484 tmp = ns->ns_hash[i].next;
1485 while (tmp != &(ns->ns_hash[i])) {
1486 struct ldlm_resource *res =
1487 list_entry(tmp, struct ldlm_resource, lr_hash);
1489 ldlm_resource_getref(res);
1490 spin_unlock(&ns->ns_hash_lock);
1492 rc = reprocess_one_queue(res, NULL);
1494 spin_lock(&ns->ns_hash_lock);
1496 ldlm_resource_putref_locked(res);
1498 if (rc == LDLM_ITER_STOP)
1503 spin_unlock(&ns->ns_hash_lock);
1507 void ldlm_reprocess_all(struct ldlm_resource *res)
1509 CFS_LIST_HEAD(rpc_list);
1513 /* Local lock trees don't get reprocessed. */
1514 if (ns_is_client(res->lr_namespace)) {
1521 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1522 if (rc == LDLM_ITER_CONTINUE)
1523 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1526 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1527 if (rc == -ERESTART) {
1528 LASSERT(list_empty(&rpc_list));
1534 void ldlm_cancel_callback(struct ldlm_lock *lock)
1536 check_res_locked(lock->l_resource);
1537 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1538 lock->l_flags |= LDLM_FL_CANCEL;
1539 if (lock->l_blocking_ast) {
1540 // l_check_no_ns_lock(ns);
1541 unlock_res_and_lock(lock);
1542 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1544 lock_res_and_lock(lock);
1546 LDLM_DEBUG(lock, "no blocking ast");
1549 lock->l_flags |= LDLM_FL_BL_DONE;
1552 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1554 struct ldlm_lock *lock;
1556 if (req->l_resource->lr_type != LDLM_PLAIN &&
1557 req->l_resource->lr_type != LDLM_IBITS)
1560 if (LDLM_SL_HEAD(&req->l_sl_mode)) {
1561 lock = list_entry(req->l_res_link.next, struct ldlm_lock,
1563 if (req->l_sl_mode.next == &lock->l_sl_mode) {
1564 lock->l_sl_mode.prev = NULL;
1566 lock->l_sl_mode.next = req->l_sl_mode.next;
1567 lock->l_sl_mode.next->prev = &lock->l_sl_mode;
1569 req->l_sl_mode.next = NULL;
1570 } else if (LDLM_SL_TAIL(&req->l_sl_mode)) {
1571 lock = list_entry(req->l_res_link.prev, struct ldlm_lock,
1573 if (req->l_sl_mode.prev == &lock->l_sl_mode) {
1574 lock->l_sl_mode.next = NULL;
1576 lock->l_sl_mode.prev = req->l_sl_mode.prev;
1577 lock->l_sl_mode.prev->next = &lock->l_sl_mode;
1579 req->l_sl_mode.prev = NULL;
1582 if (LDLM_SL_HEAD(&req->l_sl_policy)) {
1583 lock = list_entry(req->l_res_link.next, struct ldlm_lock,
1585 if (req->l_sl_policy.next == &lock->l_sl_policy) {
1586 lock->l_sl_policy.prev = NULL;
1588 lock->l_sl_policy.next = req->l_sl_policy.next;
1589 lock->l_sl_policy.next->prev = &lock->l_sl_policy;
1591 req->l_sl_policy.next = NULL;
1592 } else if (LDLM_SL_TAIL(&req->l_sl_policy)) {
1593 lock = list_entry(req->l_res_link.prev, struct ldlm_lock,
1595 if (req->l_sl_policy.prev == &lock->l_sl_policy) {
1596 lock->l_sl_policy.next = NULL;
1598 lock->l_sl_policy.prev = req->l_sl_policy.prev;
1599 lock->l_sl_policy.prev->next = &lock->l_sl_policy;
1601 req->l_sl_policy.prev = NULL;
1605 void ldlm_lock_cancel(struct ldlm_lock *lock)
1607 struct ldlm_resource *res;
1608 struct ldlm_namespace *ns;
1611 lock_res_and_lock(lock);
1613 res = lock->l_resource;
1614 ns = res->lr_namespace;
1616 /* Please do not, no matter how tempting, remove this LBUG without
1617 * talking to me first. -phik */
1618 if (lock->l_readers || lock->l_writers) {
1619 LDLM_ERROR(lock, "lock still has references");
1623 ldlm_del_waiting_lock(lock);
1625 /* Releases cancel callback. */
1626 ldlm_cancel_callback(lock);
1628 /* Yes, second time, just in case it was added again while we were
1629 running with no res lock in ldlm_cancel_callback */
1630 ldlm_del_waiting_lock(lock);
1631 ldlm_resource_unlink_lock(lock);
1632 ldlm_lock_destroy_nolock(lock);
1634 if (lock->l_granted_mode == lock->l_req_mode)
1635 ldlm_pool_del(&ns->ns_pool, lock);
1637 /* Make sure we will not be called again for same lock what is possible
1638 * if not to zero out lock->l_granted_mode */
1639 lock->l_granted_mode = 0;
1640 unlock_res_and_lock(lock);
1645 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1647 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1653 lock->l_ast_data = data;
1654 LDLM_LOCK_PUT(lock);
1658 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1660 struct ldlm_lock *lock;
1661 struct ldlm_resource *res;
1663 spin_lock(&exp->exp_ldlm_data.led_lock);
1664 while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1665 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1666 struct ldlm_lock, l_export_chain);
1667 res = ldlm_resource_getref(lock->l_resource);
1668 LDLM_LOCK_GET(lock);
1669 spin_unlock(&exp->exp_ldlm_data.led_lock);
1671 LDLM_DEBUG(lock, "export %p", exp);
1672 ldlm_res_lvbo_update(res, NULL, 0, 1);
1674 ldlm_lock_cancel(lock);
1675 ldlm_reprocess_all(res);
1677 ldlm_resource_putref(res);
1678 LDLM_LOCK_PUT(lock);
1679 spin_lock(&exp->exp_ldlm_data.led_lock);
1681 spin_unlock(&exp->exp_ldlm_data.led_lock);
1684 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1687 CFS_LIST_HEAD(rpc_list);
1688 struct ldlm_resource *res;
1689 struct ldlm_namespace *ns;
1692 struct ldlm_lock *mark_lock = NULL;
1693 int join = LDLM_JOIN_NONE;
1695 struct ldlm_interval *node;
1698 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1699 *flags |= LDLM_FL_BLOCK_GRANTED;
1700 RETURN(lock->l_resource);
1703 /* I can't check the type of lock here because the bitlock of lock
1704 * is not held here, so do the allocation blindly. -jay */
1705 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO, sizeof(*node));
1706 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1709 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1710 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1712 lock_res_and_lock(lock);
1714 res = lock->l_resource;
1715 ns = res->lr_namespace;
1717 old_mode = lock->l_req_mode;
1718 lock->l_req_mode = new_mode;
1719 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1720 /* remember the lock position where the lock might be
1721 * added back to the granted list later and also
1722 * remember the join mode for skiplist fixing. */
1723 if (LDLM_SL_HEAD(&lock->l_sl_mode))
1724 join = LDLM_MODE_JOIN_RIGHT;
1725 else if (LDLM_SL_TAIL(&lock->l_sl_mode))
1726 join = LDLM_MODE_JOIN_LEFT;
1727 if (LDLM_SL_HEAD(&lock->l_sl_policy))
1728 join |= LDLM_POLICY_JOIN_RIGHT;
1729 else if (LDLM_SL_TAIL(&lock->l_sl_policy))
1730 join |= LDLM_POLICY_JOIN_LEFT;
1732 LASSERT(!((join & LDLM_MODE_JOIN_RIGHT) &&
1733 (join & LDLM_POLICY_JOIN_LEFT)));
1734 LASSERT(!((join & LDLM_MODE_JOIN_LEFT) &&
1735 (join & LDLM_POLICY_JOIN_RIGHT)));
1737 if ((join & LDLM_MODE_JOIN_LEFT) ||
1738 (join & LDLM_POLICY_JOIN_LEFT))
1739 mark_lock = list_entry(lock->l_res_link.prev,
1740 struct ldlm_lock, l_res_link);
1741 else if (lock->l_res_link.next != &res->lr_granted)
1742 mark_lock = list_entry(lock->l_res_link.next,
1743 struct ldlm_lock, l_res_link);
1745 ldlm_resource_unlink_lock(lock);
1746 if (res->lr_type == LDLM_EXTENT) {
1747 /* FIXME: ugly code, I have to attach the lock to a
1748 * interval node again since perhaps it will be granted
1750 CFS_INIT_LIST_HEAD(&node->li_group);
1751 ldlm_interval_attach(node, lock);
1756 /* If this is a local resource, put it on the appropriate list. */
1757 if (ns_is_client(res->lr_namespace)) {
1758 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1759 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1761 /* This should never happen, because of the way the
1762 * server handles conversions. */
1763 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1767 ldlm_grant_lock(lock, &rpc_list);
1769 /* FIXME: completion handling not with ns_lock held ! */
1770 if (lock->l_completion_ast)
1771 lock->l_completion_ast(lock, 0, NULL);
1775 ldlm_processing_policy policy;
1776 policy = ldlm_processing_policy_table[res->lr_type];
1777 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1778 if (rc == LDLM_ITER_STOP) {
1779 lock->l_req_mode = old_mode;
1780 if (res->lr_type == LDLM_EXTENT)
1781 ldlm_extent_add_lock(res, lock);
1783 ldlm_granted_list_add_lock(lock, mark_lock,
1787 *flags |= LDLM_FL_BLOCK_GRANTED;
1791 unlock_res_and_lock(lock);
1794 ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1796 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1800 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1802 struct obd_device *obd = NULL;
1804 if (!((libcfs_debug | D_ERROR) & level))
1808 CDEBUG(level, " NULL LDLM lock\n");
1812 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1813 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1815 if (lock->l_conn_export != NULL)
1816 obd = lock->l_conn_export->exp_obd;
1817 if (lock->l_export && lock->l_export->exp_connection) {
1818 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1819 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1820 lock->l_remote_handle.cookie);
1821 } else if (obd == NULL) {
1822 CDEBUG(level, " Node: local\n");
1824 struct obd_import *imp = obd->u.cli.cl_import;
1825 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1826 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1827 lock->l_remote_handle.cookie);
1829 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1831 lock->l_resource->lr_name.name[0],
1832 lock->l_resource->lr_name.name[1],
1833 lock->l_resource->lr_name.name[2]);
1834 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1835 "write: %d flags: %#x\n", ldlm_lockname[lock->l_req_mode],
1836 ldlm_lockname[lock->l_granted_mode],
1837 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1839 if (lock->l_resource->lr_type == LDLM_EXTENT)
1840 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1841 " (req "LPU64"-"LPU64")\n",
1842 lock->l_policy_data.l_extent.start,
1843 lock->l_policy_data.l_extent.end,
1844 lock->l_req_extent.start, lock->l_req_extent.end);
1845 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1846 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1847 lock->l_policy_data.l_flock.pid,
1848 lock->l_policy_data.l_flock.start,
1849 lock->l_policy_data.l_flock.end);
1850 else if (lock->l_resource->lr_type == LDLM_IBITS)
1851 CDEBUG(level, " Bits: "LPX64"\n",
1852 lock->l_policy_data.l_inodebits.bits);
1855 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1857 struct ldlm_lock *lock;
1859 if (!((libcfs_debug | D_ERROR) & level))
1862 lock = ldlm_handle2lock(lockh);
1866 ldlm_lock_dump(D_OTHER, lock, 0);
1868 LDLM_LOCK_PUT(lock);
1871 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1872 struct libcfs_debug_msg_data *data, const char *fmt,
1876 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1878 va_start(args, fmt);
1880 if (lock->l_resource == NULL) {
1881 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1882 data->msg_fn, data->msg_line, fmt, args,
1883 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1884 "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "
1885 LPX64" expref: %d pid: %u\n", lock,
1886 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1887 lock->l_readers, lock->l_writers,
1888 ldlm_lockname[lock->l_granted_mode],
1889 ldlm_lockname[lock->l_req_mode],
1890 lock->l_flags, lock->l_remote_handle.cookie,
1892 atomic_read(&lock->l_export->exp_refcount) : -99,
1898 switch (lock->l_resource->lr_type) {
1900 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1901 data->msg_fn, data->msg_line, fmt, args,
1902 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1903 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1904 "] (req "LPU64"->"LPU64") flags: %x remote: "LPX64
1905 " expref: %d pid: %u\n",
1906 lock->l_resource->lr_namespace->ns_name, lock,
1907 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1908 lock->l_readers, lock->l_writers,
1909 ldlm_lockname[lock->l_granted_mode],
1910 ldlm_lockname[lock->l_req_mode],
1911 lock->l_resource->lr_name.name[0],
1912 lock->l_resource->lr_name.name[1],
1913 atomic_read(&lock->l_resource->lr_refcount),
1914 ldlm_typename[lock->l_resource->lr_type],
1915 lock->l_policy_data.l_extent.start,
1916 lock->l_policy_data.l_extent.end,
1917 lock->l_req_extent.start, lock->l_req_extent.end,
1918 lock->l_flags, lock->l_remote_handle.cookie,
1920 atomic_read(&lock->l_export->exp_refcount) : -99,
1925 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1926 data->msg_fn, data->msg_line, fmt, args,
1927 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1928 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1929 "["LPU64"->"LPU64"] flags: %x remote: "LPX64
1930 " expref: %d pid: %u\n",
1931 lock->l_resource->lr_namespace->ns_name, lock,
1932 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1933 lock->l_readers, lock->l_writers,
1934 ldlm_lockname[lock->l_granted_mode],
1935 ldlm_lockname[lock->l_req_mode],
1936 lock->l_resource->lr_name.name[0],
1937 lock->l_resource->lr_name.name[1],
1938 atomic_read(&lock->l_resource->lr_refcount),
1939 ldlm_typename[lock->l_resource->lr_type],
1940 lock->l_policy_data.l_flock.pid,
1941 lock->l_policy_data.l_flock.start,
1942 lock->l_policy_data.l_flock.end,
1943 lock->l_flags, lock->l_remote_handle.cookie,
1945 atomic_read(&lock->l_export->exp_refcount) : -99,
1950 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1951 data->msg_fn, data->msg_line, fmt, args,
1952 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1953 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1954 "flags: %x remote: "LPX64" expref: %d "
1956 lock->l_resource->lr_namespace->ns_name,
1957 lock, lock->l_handle.h_cookie,
1958 atomic_read (&lock->l_refc),
1959 lock->l_readers, lock->l_writers,
1960 ldlm_lockname[lock->l_granted_mode],
1961 ldlm_lockname[lock->l_req_mode],
1962 lock->l_resource->lr_name.name[0],
1963 lock->l_resource->lr_name.name[1],
1964 lock->l_policy_data.l_inodebits.bits,
1965 atomic_read(&lock->l_resource->lr_refcount),
1966 ldlm_typename[lock->l_resource->lr_type],
1967 lock->l_flags, lock->l_remote_handle.cookie,
1969 atomic_read(&lock->l_export->exp_refcount) : -99,
1974 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1975 data->msg_fn, data->msg_line, fmt, args,
1976 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1977 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: %x "
1978 "remote: "LPX64" expref: %d pid: %u\n",
1979 lock->l_resource->lr_namespace->ns_name,
1980 lock, lock->l_handle.h_cookie,
1981 atomic_read (&lock->l_refc),
1982 lock->l_readers, lock->l_writers,
1983 ldlm_lockname[lock->l_granted_mode],
1984 ldlm_lockname[lock->l_req_mode],
1985 lock->l_resource->lr_name.name[0],
1986 lock->l_resource->lr_name.name[1],
1987 atomic_read(&lock->l_resource->lr_refcount),
1988 ldlm_typename[lock->l_resource->lr_type],
1989 lock->l_flags, lock->l_remote_handle.cookie,
1991 atomic_read(&lock->l_export->exp_refcount) : -99,
1997 EXPORT_SYMBOL(_ldlm_lock_debug);