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 /* lock's skip list pointers fix mode */
41 #define LDLM_JOIN_NONE 0
42 #define LDLM_MODE_JOIN_RIGHT 1
43 #define LDLM_MODE_JOIN_LEFT (1 << 1)
44 #define LDLM_POLICY_JOIN_RIGHT (1 << 2)
45 #define LDLM_POLICY_JOIN_LEFT (1 << 3)
48 char *ldlm_lockname[] = {
59 char *ldlm_typename[] = {
66 char *ldlm_it2str(int it)
73 case (IT_OPEN | IT_CREAT):
86 CERROR("Unknown intent %d\n", it);
91 extern cfs_mem_cache_t *ldlm_lock_slab;
93 static ldlm_processing_policy ldlm_processing_policy_table[] = {
94 [LDLM_PLAIN] ldlm_process_plain_lock,
95 [LDLM_EXTENT] ldlm_process_extent_lock,
97 [LDLM_FLOCK] ldlm_process_flock_lock,
99 [LDLM_IBITS] ldlm_process_inodebits_lock,
102 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
104 return ldlm_processing_policy_table[res->lr_type];
107 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
113 * REFCOUNTED LOCK OBJECTS
118 * Lock refcounts, during creation:
119 * - one special one for allocation, dec'd only once in destroy
120 * - one for being a lock that's in-use
121 * - one for the addref associated with a new lock
123 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
125 atomic_inc(&lock->l_refc);
129 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
131 LASSERT(size == sizeof(*lock));
132 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
135 void ldlm_lock_put(struct ldlm_lock *lock)
139 LASSERT(lock->l_resource != LP_POISON);
140 LASSERT(atomic_read(&lock->l_refc) > 0);
141 if (atomic_dec_and_test(&lock->l_refc)) {
142 struct ldlm_resource *res;
145 "final lock_put on destroyed lock, freeing it.");
147 res = lock->l_resource;
148 LASSERT(lock->l_destroyed);
149 LASSERT(list_empty(&lock->l_res_link));
150 LASSERT(list_empty(&lock->l_pending_chain));
152 atomic_dec(&res->lr_namespace->ns_locks);
153 ldlm_resource_putref(res);
154 lock->l_resource = NULL;
155 if (lock->l_export) {
156 class_export_put(lock->l_export);
157 lock->l_export = NULL;
160 if (lock->l_lvb_data != NULL)
161 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
163 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
170 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
173 if (!list_empty(&lock->l_lru)) {
174 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
175 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
176 list_del_init(&lock->l_lru);
178 LASSERT(ns->ns_nr_unused >= 0);
184 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
186 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
189 spin_lock(&ns->ns_unused_lock);
190 rc = ldlm_lock_remove_from_lru_nolock(lock);
191 spin_unlock(&ns->ns_unused_lock);
196 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
198 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
199 lock->l_last_used = cfs_time_current();
200 LASSERT(list_empty(&lock->l_lru));
201 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
202 LASSERT(ns->ns_nr_unused >= 0);
206 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
208 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
210 spin_lock(&ns->ns_unused_lock);
211 ldlm_lock_add_to_lru_nolock(lock);
212 spin_unlock(&ns->ns_unused_lock);
216 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
218 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
220 spin_lock(&ns->ns_unused_lock);
221 if (!list_empty(&lock->l_lru)) {
222 ldlm_lock_remove_from_lru_nolock(lock);
223 ldlm_lock_add_to_lru_nolock(lock);
225 spin_unlock(&ns->ns_unused_lock);
229 /* This used to have a 'strict' flag, which recovery would use to mark an
230 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
231 * shall explain why it's gone: with the new hash table scheme, once you call
232 * ldlm_lock_destroy, you can never drop your final references on this lock.
233 * Because it's not in the hash table anymore. -phil */
234 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
238 if (lock->l_readers || lock->l_writers) {
239 LDLM_ERROR(lock, "lock still has references");
240 ldlm_lock_dump(D_ERROR, lock, 0);
244 if (!list_empty(&lock->l_res_link)) {
245 LDLM_ERROR(lock, "lock still on resource");
246 ldlm_lock_dump(D_ERROR, lock, 0);
250 if (lock->l_destroyed) {
251 LASSERT(list_empty(&lock->l_lru));
255 lock->l_destroyed = 1;
258 spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
259 list_del_init(&lock->l_export_chain);
261 spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
263 ldlm_lock_remove_from_lru(lock);
264 class_handle_unhash(&lock->l_handle);
267 /* Wake anyone waiting for this lock */
268 /* FIXME: I should probably add yet another flag, instead of using
269 * l_export to only call this on clients */
271 class_export_put(lock->l_export);
272 lock->l_export = NULL;
273 if (lock->l_export && lock->l_completion_ast)
274 lock->l_completion_ast(lock, 0);
280 void ldlm_lock_destroy(struct ldlm_lock *lock)
284 lock_res_and_lock(lock);
285 first = ldlm_lock_destroy_internal(lock);
286 unlock_res_and_lock(lock);
288 /* drop reference from hashtable only for first destroy */
294 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
298 first = ldlm_lock_destroy_internal(lock);
299 /* drop reference from hashtable only for first destroy */
305 /* this is called by portals_handle2object with the handle lock taken */
306 static void lock_handle_addref(void *lock)
308 LDLM_LOCK_GET((struct ldlm_lock *)lock);
312 * usage: pass in a resource on which you have done ldlm_resource_get
313 * pass in a parent lock on which you have done a ldlm_lock_get
314 * after return, ldlm_*_put the resource and parent
315 * returns: lock with refcount 2 - one for current caller and one for remote
317 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
319 struct ldlm_lock *lock;
322 if (resource == NULL)
325 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, CFS_ALLOC_IO, sizeof(*lock));
329 spin_lock_init(&lock->l_lock);
330 lock->l_resource = ldlm_resource_getref(resource);
332 atomic_set(&lock->l_refc, 2);
333 CFS_INIT_LIST_HEAD(&lock->l_res_link);
334 CFS_INIT_LIST_HEAD(&lock->l_lru);
335 CFS_INIT_LIST_HEAD(&lock->l_export_chain);
336 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
337 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
338 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
339 cfs_waitq_init(&lock->l_waitq);
340 lock->l_blocking_lock = NULL;
341 lock->l_sl_mode.prev = NULL;
342 lock->l_sl_mode.next = NULL;
343 lock->l_sl_policy.prev = NULL;
344 lock->l_sl_policy.next = NULL;
346 atomic_inc(&resource->lr_namespace->ns_locks);
347 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
348 class_handle_hash(&lock->l_handle, lock_handle_addref);
353 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
354 const struct ldlm_res_id *new_resid)
356 struct ldlm_resource *oldres = lock->l_resource;
357 struct ldlm_resource *newres;
361 LASSERT(ns->ns_client != 0);
363 lock_res_and_lock(lock);
364 if (memcmp(new_resid, &lock->l_resource->lr_name,
365 sizeof(lock->l_resource->lr_name)) == 0) {
367 unlock_res_and_lock(lock);
371 LASSERT(new_resid->name[0] != 0);
373 /* This function assumes that the lock isn't on any lists */
374 LASSERT(list_empty(&lock->l_res_link));
376 type = oldres->lr_type;
377 unlock_res_and_lock(lock);
379 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
380 if (newres == NULL) {
385 lock_res_and_lock(lock);
386 LASSERT(memcmp(new_resid, &lock->l_resource->lr_name,
387 sizeof(lock->l_resource->lr_name)) != 0);
389 lock->l_resource = newres;
391 unlock_res_and_lock(lock);
393 /* ...and the flowers are still standing! */
394 ldlm_resource_putref(oldres);
403 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
405 lockh->cookie = lock->l_handle.h_cookie;
408 /* if flags: atomically get the lock and set the flags.
409 * Return NULL if flag already set
412 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
415 struct ldlm_namespace *ns;
416 struct ldlm_lock *lock, *retval = NULL;
421 lock = class_handle2object(handle->cookie);
425 LASSERT(lock->l_resource != NULL);
426 ns = lock->l_resource->lr_namespace;
429 lock_res_and_lock(lock);
431 /* It's unlikely but possible that someone marked the lock as
432 * destroyed after we did handle2object on it */
433 if (lock->l_destroyed) {
434 unlock_res_and_lock(lock);
435 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
440 if (flags && (lock->l_flags & flags)) {
441 unlock_res_and_lock(lock);
447 lock->l_flags |= flags;
449 unlock_res_and_lock(lock);
456 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
457 const struct lustre_handle *handle)
459 struct ldlm_lock *retval = NULL;
460 retval = __ldlm_handle2lock(handle, 0);
464 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
466 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
467 /* INODEBITS_INTEROP: If the other side does not support
468 * inodebits, reply with a plain lock descriptor.
470 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
471 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
472 struct ldlm_resource res = *lock->l_resource;
474 /* Make sure all the right bits are set in this lock we
475 are going to pass to client */
476 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
477 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
478 "Inappropriate inode lock bits during "
479 "conversion " LPU64 "\n",
480 lock->l_policy_data.l_inodebits.bits);
481 res.lr_type = LDLM_PLAIN;
482 ldlm_res2desc(&res, &desc->l_resource);
483 /* Convert "new" lock mode to something old client can
485 if ((lock->l_req_mode == LCK_CR) ||
486 (lock->l_req_mode == LCK_CW))
487 desc->l_req_mode = LCK_PR;
489 desc->l_req_mode = lock->l_req_mode;
490 if ((lock->l_granted_mode == LCK_CR) ||
491 (lock->l_granted_mode == LCK_CW)) {
492 desc->l_granted_mode = LCK_PR;
494 /* We never grant PW/EX locks to clients */
495 LASSERT((lock->l_granted_mode != LCK_PW) &&
496 (lock->l_granted_mode != LCK_EX));
497 desc->l_granted_mode = lock->l_granted_mode;
500 /* We do not copy policy here, because there is no
501 policy for plain locks */
503 ldlm_res2desc(lock->l_resource, &desc->l_resource);
504 desc->l_req_mode = lock->l_req_mode;
505 desc->l_granted_mode = lock->l_granted_mode;
506 desc->l_policy_data = lock->l_policy_data;
510 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
511 struct list_head *work_list)
513 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
514 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
515 lock->l_flags |= LDLM_FL_AST_SENT;
516 /* If the enqueuing client said so, tell the AST recipient to
517 * discard dirty data, rather than writing back. */
518 if (new->l_flags & LDLM_AST_DISCARD_DATA)
519 lock->l_flags |= LDLM_FL_DISCARD_DATA;
520 LASSERT(list_empty(&lock->l_bl_ast));
521 list_add(&lock->l_bl_ast, work_list);
523 LASSERT(lock->l_blocking_lock == NULL);
524 lock->l_blocking_lock = LDLM_LOCK_GET(new);
528 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
530 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
531 lock->l_flags |= LDLM_FL_CP_REQD;
532 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
533 LASSERT(list_empty(&lock->l_cp_ast));
534 list_add(&lock->l_cp_ast, work_list);
539 /* must be called with lr_lock held */
540 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
541 struct list_head *work_list)
544 check_res_locked(lock->l_resource);
546 ldlm_add_bl_work_item(lock, new, work_list);
548 ldlm_add_cp_work_item(lock, work_list);
552 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
554 struct ldlm_lock *lock;
556 lock = ldlm_handle2lock(lockh);
557 LASSERT(lock != NULL);
558 ldlm_lock_addref_internal(lock, mode);
562 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
564 ldlm_lock_remove_from_lru(lock);
565 if (mode & (LCK_NL | LCK_CR | LCK_PR))
567 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
570 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
573 /* only called for local locks */
574 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
576 lock_res_and_lock(lock);
577 ldlm_lock_addref_internal_nolock(lock, mode);
578 unlock_res_and_lock(lock);
581 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
583 struct ldlm_namespace *ns;
586 lock_res_and_lock(lock);
588 ns = lock->l_resource->lr_namespace;
590 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
591 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
592 LASSERT(lock->l_readers > 0);
595 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
596 LASSERT(lock->l_writers > 0);
600 if (lock->l_flags & LDLM_FL_LOCAL &&
601 !lock->l_readers && !lock->l_writers) {
602 /* If this is a local lock on a server namespace and this was
603 * the last reference, cancel the lock. */
604 CDEBUG(D_INFO, "forcing cancel of local lock\n");
605 lock->l_flags |= LDLM_FL_CBPENDING;
608 if (!lock->l_readers && !lock->l_writers &&
609 (lock->l_flags & LDLM_FL_CBPENDING)) {
610 /* If we received a blocked AST and this was the last reference,
611 * run the callback. */
612 if (ns->ns_client == LDLM_NAMESPACE_SERVER && lock->l_export)
613 CERROR("FL_CBPENDING set on non-local lock--just a "
616 LDLM_DEBUG(lock, "final decref done on cbpending lock");
618 LDLM_LOCK_GET(lock); /* dropped by bl thread */
619 ldlm_lock_remove_from_lru(lock);
620 unlock_res_and_lock(lock);
621 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
622 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
623 ldlm_handle_bl_callback(ns, NULL, lock);
624 } else if (ns->ns_client == LDLM_NAMESPACE_CLIENT &&
625 !lock->l_readers && !lock->l_writers &&
626 !(lock->l_flags & LDLM_FL_NO_LRU)) {
627 /* If this is a client-side namespace and this was the last
628 * reference, put it on the LRU. */
629 ldlm_lock_add_to_lru(lock);
630 unlock_res_and_lock(lock);
631 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
632 * are not supported by the server, otherwise, it is done on
634 if (!exp_connect_cancelset(lock->l_conn_export) &&
635 !exp_connect_lru_resize(lock->l_conn_export))
636 ldlm_cancel_lru(ns, 0, LDLM_ASYNC);
638 unlock_res_and_lock(lock);
641 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
646 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
648 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
649 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
650 ldlm_lock_decref_internal(lock, mode);
654 /* This will drop a lock reference and mark it for destruction, but will not
655 * necessarily cancel the lock before returning. */
656 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
658 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
661 LASSERT(lock != NULL);
663 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
664 lock_res_and_lock(lock);
665 lock->l_flags |= LDLM_FL_CBPENDING;
666 unlock_res_and_lock(lock);
667 ldlm_lock_decref_internal(lock, mode);
672 * search_granted_lock
675 * Finds a position to insert the new lock.
677 * queue [input]: the granted list where search acts on;
678 * req [input]: the lock whose position to be located;
679 * lockp [output]: the position where the lock should be inserted before, or
680 * NULL indicating @req should be appended to @queue.
682 * Bit-masks combination of following values indicating in which way the
683 * lock need to be inserted.
684 * - LDLM_JOIN_NONE: noting about skip list needs to be fixed;
685 * - LDLM_MODE_JOIN_RIGHT: @req needs join right becoming the head of a
687 * - LDLM_POLICY_JOIN_RIGHT: @req needs join right becoming the head of
690 * - ldlm_grant_lock_with_skiplist
692 static int search_granted_lock(struct list_head *queue,
693 struct ldlm_lock *req,
694 struct ldlm_lock **lockp)
696 struct list_head *tmp, *tmp_tail;
697 struct ldlm_lock *lock, *mode_head_lock;
698 int rc = LDLM_JOIN_NONE;
701 list_for_each(tmp, queue) {
702 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
704 if (lock->l_req_mode != req->l_req_mode) {
705 if (LDLM_SL_HEAD(&lock->l_sl_mode))
706 tmp = &list_entry(lock->l_sl_mode.next,
708 l_sl_mode)->l_res_link;
712 /* found the same mode group */
713 if (lock->l_resource->lr_type == LDLM_PLAIN) {
715 rc = LDLM_MODE_JOIN_RIGHT;
717 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
719 if (LDLM_SL_HEAD(&lock->l_sl_mode))
720 tmp_tail = &list_entry(lock->l_sl_mode.next,
722 l_sl_mode)->l_res_link;
723 mode_head_lock = lock;
725 if (lock->l_policy_data.l_inodebits.bits ==
726 req->l_policy_data.l_inodebits.bits) {
727 /* matched policy lock is found */
729 rc |= LDLM_POLICY_JOIN_RIGHT;
731 /* if the policy group head is also a
732 * mode group head or a single mode
734 if (LDLM_SL_HEAD(&lock->l_sl_mode) ||
736 LDLM_SL_EMPTY(&lock->l_sl_mode)))
737 rc |= LDLM_MODE_JOIN_RIGHT;
741 if (LDLM_SL_HEAD(&lock->l_sl_policy))
742 tmp = &list_entry(lock->l_sl_policy.next,
744 l_sl_policy)->l_res_link;
750 lock = list_entry(tmp, struct ldlm_lock,
752 } /* for all locks in the matched mode group */
754 /* no matched policy group is found, insert before
755 * the mode group head lock */
756 *lockp = mode_head_lock;
757 rc = LDLM_MODE_JOIN_RIGHT;
760 LDLM_ERROR(lock, "is not LDLM_PLAIN or LDLM_IBITS lock");
765 /* no matched mode group is found, append to the end */
773 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
774 struct ldlm_lock *lockp,
777 struct ldlm_resource *res = lock->l_resource;
780 LASSERT(lockp || join == LDLM_JOIN_NONE);
782 check_res_locked(res);
784 ldlm_resource_dump(D_OTHER, res);
785 CDEBUG(D_OTHER, "About to add this lock:\n");
786 ldlm_lock_dump(D_OTHER, lock, 0);
788 if (lock->l_destroyed) {
789 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
793 LASSERT(list_empty(&lock->l_res_link));
796 list_add_tail(&lock->l_res_link, &lock->l_resource->lr_granted);
797 else if ((join & LDLM_MODE_JOIN_LEFT) || (join & LDLM_POLICY_JOIN_LEFT))
798 list_add(&lock->l_res_link, &lockp->l_res_link);
800 list_add_tail(&lock->l_res_link, &lockp->l_res_link);
803 if (join & LDLM_MODE_JOIN_RIGHT) {
804 LASSERT(! LDLM_SL_TAIL(&lockp->l_sl_mode));
805 if (LDLM_SL_EMPTY(&lockp->l_sl_mode)) {
806 lock->l_sl_mode.next = &lockp->l_sl_mode;
807 lockp->l_sl_mode.prev = &lock->l_sl_mode;
808 } else if (LDLM_SL_HEAD(&lockp->l_sl_mode)) {
809 lock->l_sl_mode.next = lockp->l_sl_mode.next;
810 lockp->l_sl_mode.next = NULL;
811 lock->l_sl_mode.next->prev = &lock->l_sl_mode;
813 } else if (join & LDLM_MODE_JOIN_LEFT) {
814 LASSERT(! LDLM_SL_HEAD(&lockp->l_sl_mode));
815 if (LDLM_SL_EMPTY(&lockp->l_sl_mode)) {
816 lock->l_sl_mode.prev = &lockp->l_sl_mode;
817 lockp->l_sl_mode.next = &lock->l_sl_mode;
818 } else if (LDLM_SL_TAIL(&lockp->l_sl_mode)) {
819 lock->l_sl_mode.prev = lockp->l_sl_mode.prev;
820 lockp->l_sl_mode.prev = NULL;
821 lock->l_sl_mode.prev->next = &lock->l_sl_mode;
825 if (join & LDLM_POLICY_JOIN_RIGHT) {
826 LASSERT(! LDLM_SL_TAIL(&lockp->l_sl_policy));
827 if (LDLM_SL_EMPTY(&lockp->l_sl_policy)) {
828 lock->l_sl_policy.next = &lockp->l_sl_policy;
829 lockp->l_sl_policy.prev = &lock->l_sl_policy;
830 } else if (LDLM_SL_HEAD(&lockp->l_sl_policy)) {
831 lock->l_sl_policy.next = lockp->l_sl_policy.next;
832 lockp->l_sl_policy.next = NULL;
833 lock->l_sl_policy.next->prev = &lock->l_sl_policy;
835 } else if (join & LDLM_POLICY_JOIN_LEFT) {
836 LASSERT(! LDLM_SL_HEAD(&lockp->l_sl_policy));
837 if (LDLM_SL_EMPTY(&lockp->l_sl_policy)) {
838 lock->l_sl_policy.prev = &lockp->l_sl_policy;
839 lockp->l_sl_policy.next = &lock->l_sl_policy;
840 } else if (LDLM_SL_TAIL(&lockp->l_sl_policy)) {
841 lock->l_sl_policy.prev = lockp->l_sl_policy.prev;
842 lockp->l_sl_policy.prev = NULL;
843 lock->l_sl_policy.prev->next = &lock->l_sl_policy;
850 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
852 int join = LDLM_JOIN_NONE;
853 struct ldlm_lock *lockp = NULL;
856 LASSERT(lock->l_req_mode == lock->l_granted_mode);
858 join = search_granted_lock(&lock->l_resource->lr_granted, lock, &lockp);
859 ldlm_granted_list_add_lock(lock, lockp, join);
864 * - ldlm_lock_enqueue
865 * - ldlm_reprocess_queue
866 * - ldlm_lock_convert
868 * must be called with lr_lock held
870 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
872 struct ldlm_resource *res = lock->l_resource;
875 check_res_locked(res);
877 lock->l_granted_mode = lock->l_req_mode;
878 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
879 ldlm_grant_lock_with_skiplist(lock);
881 ldlm_resource_add_lock(res, &res->lr_granted, lock);
883 if (lock->l_granted_mode < res->lr_most_restr)
884 res->lr_most_restr = lock->l_granted_mode;
886 if (work_list && lock->l_completion_ast != NULL)
887 ldlm_add_ast_work_item(lock, NULL, work_list);
889 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
893 /* returns a referenced lock or NULL. See the flag descriptions below, in the
894 * comment above ldlm_lock_match */
895 static struct ldlm_lock *search_queue(struct list_head *queue,
897 ldlm_policy_data_t *policy,
898 struct ldlm_lock *old_lock, int flags)
900 struct ldlm_lock *lock;
901 struct list_head *tmp;
903 list_for_each(tmp, queue) {
906 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
908 if (lock == old_lock)
911 /* llite sometimes wants to match locks that will be
912 * canceled when their users drop, but we allow it to match
913 * if it passes in CBPENDING and the lock still has users.
914 * this is generally only going to be used by children
915 * whose parents already hold a lock so forward progress
916 * can still happen. */
917 if (lock->l_flags & LDLM_FL_CBPENDING &&
918 !(flags & LDLM_FL_CBPENDING))
920 if (lock->l_flags & LDLM_FL_CBPENDING &&
921 lock->l_readers == 0 && lock->l_writers == 0)
924 if (!(lock->l_req_mode & *mode))
926 match = lock->l_req_mode;
928 if (lock->l_resource->lr_type == LDLM_EXTENT &&
929 (lock->l_policy_data.l_extent.start >
930 policy->l_extent.start ||
931 lock->l_policy_data.l_extent.end < policy->l_extent.end))
934 if (unlikely(match == LCK_GROUP) &&
935 lock->l_resource->lr_type == LDLM_EXTENT &&
936 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
939 /* We match if we have existing lock with same or wider set
941 if (lock->l_resource->lr_type == LDLM_IBITS &&
942 ((lock->l_policy_data.l_inodebits.bits &
943 policy->l_inodebits.bits) !=
944 policy->l_inodebits.bits))
947 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
950 if ((flags & LDLM_FL_LOCAL_ONLY) &&
951 !(lock->l_flags & LDLM_FL_LOCAL))
954 if (flags & LDLM_FL_TEST_LOCK) {
956 ldlm_lock_touch_in_lru(lock);
958 ldlm_lock_addref_internal_nolock(lock, match);
967 void ldlm_lock_allow_match(struct ldlm_lock *lock)
969 lock_res_and_lock(lock);
970 lock->l_flags |= LDLM_FL_LVB_READY;
971 cfs_waitq_signal(&lock->l_waitq);
972 unlock_res_and_lock(lock);
975 /* Can be called in two ways:
977 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
978 * for a duplicate of.
980 * Otherwise, all of the fields must be filled in, to match against.
982 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
983 * server (ie, connh is NULL)
984 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
985 * list will be considered
986 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
987 * to be canceled can still be matched as long as they still have reader
988 * or writer refernces
989 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
990 * just tell us if we would have matched.
992 * Returns 1 if it finds an already-existing lock that is compatible; in this
993 * case, lockh is filled in with a addref()ed lock
995 * we also check security context, if that failed we simply return 0 (to keep
996 * caller code unchanged), the context failure will be discovered by caller
999 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1000 const struct ldlm_res_id *res_id, ldlm_type_t type,
1001 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1002 struct lustre_handle *lockh)
1004 struct ldlm_resource *res;
1005 struct ldlm_lock *lock, *old_lock = NULL;
1010 old_lock = ldlm_handle2lock(lockh);
1013 ns = old_lock->l_resource->lr_namespace;
1014 res_id = &old_lock->l_resource->lr_name;
1015 type = old_lock->l_resource->lr_type;
1016 mode = old_lock->l_req_mode;
1019 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1021 LASSERT(old_lock == NULL);
1027 lock = search_queue(&res->lr_granted, &mode, policy, old_lock, flags);
1030 if (flags & LDLM_FL_BLOCK_GRANTED)
1032 lock = search_queue(&res->lr_converting, &mode, policy, old_lock, flags);
1035 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock, flags);
1042 ldlm_resource_putref(res);
1045 ldlm_lock2handle(lock, lockh);
1046 if ((flags & LDLM_FL_LVB_READY) &&
1047 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1048 struct l_wait_info lwi;
1049 if (lock->l_completion_ast) {
1050 int err = lock->l_completion_ast(lock,
1051 LDLM_FL_WAIT_NOREPROC,
1054 if (flags & LDLM_FL_TEST_LOCK)
1055 LDLM_LOCK_PUT(lock);
1057 ldlm_lock_decref_internal(lock, mode);
1063 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), NULL,
1064 LWI_ON_SIGNAL_NOOP, NULL);
1066 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1067 l_wait_event(lock->l_waitq,
1068 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1073 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1074 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1075 res_id->name[2] : policy->l_extent.start,
1076 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1077 res_id->name[3] : policy->l_extent.end);
1079 /* check user's security context */
1080 if (lock->l_conn_export &&
1081 sptlrpc_import_check_ctx(
1082 class_exp2cliimp(lock->l_conn_export))) {
1083 if (!(flags & LDLM_FL_TEST_LOCK))
1084 ldlm_lock_decref_internal(lock, mode);
1088 if (flags & LDLM_FL_TEST_LOCK)
1089 LDLM_LOCK_PUT(lock);
1091 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1092 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1093 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1094 type, mode, res_id->name[0], res_id->name[1],
1095 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1096 res_id->name[2] :policy->l_extent.start,
1097 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1098 res_id->name[3] : policy->l_extent.end);
1101 LDLM_LOCK_PUT(old_lock);
1103 return rc ? mode : 0;
1106 /* Returns a referenced lock */
1107 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1108 const struct ldlm_res_id *res_id,
1111 ldlm_blocking_callback blocking,
1112 ldlm_completion_callback completion,
1113 ldlm_glimpse_callback glimpse,
1114 void *data, __u32 lvb_len)
1116 struct ldlm_lock *lock;
1117 struct ldlm_resource *res;
1120 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1124 lock = ldlm_lock_new(res);
1125 ldlm_resource_putref(res);
1130 lock->l_req_mode = mode;
1131 lock->l_ast_data = data;
1132 lock->l_blocking_ast = blocking;
1133 lock->l_completion_ast = completion;
1134 lock->l_glimpse_ast = glimpse;
1135 lock->l_pid = cfs_curproc_pid();
1138 lock->l_lvb_len = lvb_len;
1139 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1140 if (lock->l_lvb_data == NULL) {
1141 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1149 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1150 struct ldlm_lock **lockp,
1151 void *cookie, int *flags)
1153 struct ldlm_lock *lock = *lockp;
1154 struct ldlm_resource *res = lock->l_resource;
1155 int local = res->lr_namespace->ns_client;
1156 ldlm_processing_policy policy;
1157 ldlm_error_t rc = ELDLM_OK;
1160 do_gettimeofday(&lock->l_enqueued_time);
1161 /* policies are not executed on the client or during replay */
1162 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1163 && !local && ns->ns_policy) {
1164 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1166 if (rc == ELDLM_LOCK_REPLACED) {
1167 /* The lock that was returned has already been granted,
1168 * and placed into lockp. If it's not the same as the
1169 * one we passed in, then destroy the old one and our
1170 * work here is done. */
1171 if (lock != *lockp) {
1172 ldlm_lock_destroy(lock);
1173 LDLM_LOCK_PUT(lock);
1175 *flags |= LDLM_FL_LOCK_CHANGED;
1177 } else if (rc != ELDLM_OK ||
1178 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1179 ldlm_lock_destroy(lock);
1184 lock_res_and_lock(lock);
1185 if (local && lock->l_req_mode == lock->l_granted_mode) {
1186 /* The server returned a blocked lock, but it was granted before
1187 * we got a chance to actually enqueue it. We don't need to do
1189 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1190 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1191 GOTO(out, ELDLM_OK);
1194 /* Some flags from the enqueue want to make it into the AST, via the
1195 * lock's l_flags. */
1196 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1198 /* This distinction between local lock trees is very important; a client
1199 * namespace only has information about locks taken by that client, and
1200 * thus doesn't have enough information to decide for itself if it can
1201 * be granted (below). In this case, we do exactly what the server
1202 * tells us to do, as dictated by the 'flags'.
1204 * We do exactly the same thing during recovery, when the server is
1205 * more or less trusting the clients not to lie.
1207 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1208 * granted/converting queues. */
1209 ldlm_resource_unlink_lock(lock);
1211 if (*flags & LDLM_FL_BLOCK_CONV)
1212 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1213 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1214 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1216 ldlm_grant_lock(lock, NULL);
1217 GOTO(out, ELDLM_OK);
1218 } else if (*flags & LDLM_FL_REPLAY) {
1219 if (*flags & LDLM_FL_BLOCK_CONV) {
1220 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1221 GOTO(out, ELDLM_OK);
1222 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1223 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1224 GOTO(out, ELDLM_OK);
1225 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1226 ldlm_grant_lock(lock, NULL);
1227 GOTO(out, ELDLM_OK);
1229 /* If no flags, fall through to normal enqueue path. */
1232 policy = ldlm_processing_policy_table[res->lr_type];
1233 policy(lock, flags, 1, &rc, NULL);
1236 unlock_res_and_lock(lock);
1240 /* Must be called with namespace taken: queue is waiting or converting. */
1241 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1242 struct list_head *work_list)
1244 struct list_head *tmp, *pos;
1245 ldlm_processing_policy policy;
1247 int rc = LDLM_ITER_CONTINUE;
1251 check_res_locked(res);
1253 policy = ldlm_processing_policy_table[res->lr_type];
1256 list_for_each_safe(tmp, pos, queue) {
1257 struct ldlm_lock *pending;
1258 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1260 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1263 rc = policy(pending, &flags, 0, &err, work_list);
1264 if (rc != LDLM_ITER_CONTINUE)
1271 int ldlm_run_bl_ast_work(struct list_head *rpc_list)
1273 struct list_head *tmp, *pos;
1274 struct ldlm_lock_desc d;
1275 int rc = 0, retval = 0;
1278 list_for_each_safe(tmp, pos, rpc_list) {
1279 struct ldlm_lock *lock =
1280 list_entry(tmp, struct ldlm_lock, l_bl_ast);
1282 /* nobody should touch l_bl_ast */
1283 lock_res_and_lock(lock);
1284 list_del_init(&lock->l_bl_ast);
1286 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1287 LASSERT(lock->l_bl_ast_run == 0);
1288 LASSERT(lock->l_blocking_lock);
1289 lock->l_bl_ast_run++;
1290 unlock_res_and_lock(lock);
1292 ldlm_lock2desc(lock->l_blocking_lock, &d);
1294 LDLM_LOCK_PUT(lock->l_blocking_lock);
1295 lock->l_blocking_lock = NULL;
1296 rc = lock->l_blocking_ast(lock, &d, NULL, LDLM_CB_BLOCKING);
1298 if (rc == -ERESTART)
1301 CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
1302 "disconnect client\n");
1303 LDLM_LOCK_PUT(lock);
1308 int ldlm_run_cp_ast_work(struct list_head *rpc_list)
1310 struct list_head *tmp, *pos;
1311 int rc = 0, retval = 0;
1314 /* It's possible to receive a completion AST before we've set
1315 * the l_completion_ast pointer: either because the AST arrived
1316 * before the reply, or simply because there's a small race
1317 * window between receiving the reply and finishing the local
1318 * enqueue. (bug 842)
1320 * This can't happen with the blocking_ast, however, because we
1321 * will never call the local blocking_ast until we drop our
1322 * reader/writer reference, which we won't do until we get the
1323 * reply and finish enqueueing. */
1325 list_for_each_safe(tmp, pos, rpc_list) {
1326 struct ldlm_lock *lock =
1327 list_entry(tmp, struct ldlm_lock, l_cp_ast);
1329 /* nobody should touch l_cp_ast */
1330 lock_res_and_lock(lock);
1331 list_del_init(&lock->l_cp_ast);
1332 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1333 lock->l_flags &= ~LDLM_FL_CP_REQD;
1334 unlock_res_and_lock(lock);
1336 if (lock->l_completion_ast != NULL)
1337 rc = lock->l_completion_ast(lock, 0, 0);
1338 if (rc == -ERESTART)
1341 CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
1342 "disconnect client\n");
1343 LDLM_LOCK_PUT(lock);
1348 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1350 ldlm_reprocess_all(res);
1351 return LDLM_ITER_CONTINUE;
1354 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1356 struct list_head *tmp;
1363 spin_lock(&ns->ns_hash_lock);
1364 for (i = 0; i < RES_HASH_SIZE; i++) {
1365 tmp = ns->ns_hash[i].next;
1366 while (tmp != &(ns->ns_hash[i])) {
1367 struct ldlm_resource *res =
1368 list_entry(tmp, struct ldlm_resource, lr_hash);
1370 ldlm_resource_getref(res);
1371 spin_unlock(&ns->ns_hash_lock);
1373 rc = reprocess_one_queue(res, NULL);
1375 spin_lock(&ns->ns_hash_lock);
1377 ldlm_resource_putref_locked(res);
1379 if (rc == LDLM_ITER_STOP)
1384 spin_unlock(&ns->ns_hash_lock);
1388 void ldlm_reprocess_all(struct ldlm_resource *res)
1390 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1394 /* Local lock trees don't get reprocessed. */
1395 if (res->lr_namespace->ns_client) {
1402 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1403 if (rc == LDLM_ITER_CONTINUE)
1404 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1407 rc = ldlm_run_cp_ast_work(&rpc_list);
1408 if (rc == -ERESTART) {
1409 LASSERT(list_empty(&rpc_list));
1415 void ldlm_cancel_callback(struct ldlm_lock *lock)
1417 check_res_locked(lock->l_resource);
1418 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1419 lock->l_flags |= LDLM_FL_CANCEL;
1420 if (lock->l_blocking_ast) {
1421 // l_check_no_ns_lock(ns);
1422 unlock_res_and_lock(lock);
1423 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1425 lock_res_and_lock(lock);
1427 LDLM_DEBUG(lock, "no blocking ast");
1430 lock->l_flags |= LDLM_FL_BL_DONE;
1433 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1435 struct ldlm_lock *lock;
1437 if (req->l_resource->lr_type != LDLM_PLAIN &&
1438 req->l_resource->lr_type != LDLM_IBITS)
1441 if (LDLM_SL_HEAD(&req->l_sl_mode)) {
1442 lock = list_entry(req->l_res_link.next, struct ldlm_lock,
1444 if (req->l_sl_mode.next == &lock->l_sl_mode) {
1445 lock->l_sl_mode.prev = NULL;
1447 lock->l_sl_mode.next = req->l_sl_mode.next;
1448 lock->l_sl_mode.next->prev = &lock->l_sl_mode;
1450 req->l_sl_mode.next = NULL;
1451 } else if (LDLM_SL_TAIL(&req->l_sl_mode)) {
1452 lock = list_entry(req->l_res_link.prev, struct ldlm_lock,
1454 if (req->l_sl_mode.prev == &lock->l_sl_mode) {
1455 lock->l_sl_mode.next = NULL;
1457 lock->l_sl_mode.prev = req->l_sl_mode.prev;
1458 lock->l_sl_mode.prev->next = &lock->l_sl_mode;
1460 req->l_sl_mode.prev = NULL;
1463 if (LDLM_SL_HEAD(&req->l_sl_policy)) {
1464 lock = list_entry(req->l_res_link.next, struct ldlm_lock,
1466 if (req->l_sl_policy.next == &lock->l_sl_policy) {
1467 lock->l_sl_policy.prev = NULL;
1469 lock->l_sl_policy.next = req->l_sl_policy.next;
1470 lock->l_sl_policy.next->prev = &lock->l_sl_policy;
1472 req->l_sl_policy.next = NULL;
1473 } else if (LDLM_SL_TAIL(&req->l_sl_policy)) {
1474 lock = list_entry(req->l_res_link.prev, struct ldlm_lock,
1476 if (req->l_sl_policy.prev == &lock->l_sl_policy) {
1477 lock->l_sl_policy.next = NULL;
1479 lock->l_sl_policy.prev = req->l_sl_policy.prev;
1480 lock->l_sl_policy.prev->next = &lock->l_sl_policy;
1482 req->l_sl_policy.prev = NULL;
1486 void ldlm_lock_cancel(struct ldlm_lock *lock)
1488 struct ldlm_resource *res;
1489 struct ldlm_namespace *ns;
1492 lock_res_and_lock(lock);
1494 res = lock->l_resource;
1495 ns = res->lr_namespace;
1497 /* Please do not, no matter how tempting, remove this LBUG without
1498 * talking to me first. -phik */
1499 if (lock->l_readers || lock->l_writers) {
1500 LDLM_ERROR(lock, "lock still has references");
1504 ldlm_del_waiting_lock(lock);
1506 /* Releases cancel callback. */
1507 ldlm_cancel_callback(lock);
1509 /* Yes, second time, just in case it was added again while we were
1510 running with no res lock in ldlm_cancel_callback */
1511 ldlm_del_waiting_lock(lock);
1512 ldlm_resource_unlink_lock(lock);
1513 ldlm_lock_destroy_nolock(lock);
1515 if (lock->l_granted_mode == lock->l_req_mode)
1516 ldlm_pool_del(&ns->ns_pool, lock);
1518 /* Make sure we will not be called again for same lock what is possible
1519 * if not to zero out lock->l_granted_mode */
1520 lock->l_granted_mode = 0;
1521 unlock_res_and_lock(lock);
1526 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1528 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1534 lock->l_ast_data = data;
1535 LDLM_LOCK_PUT(lock);
1539 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1541 struct ldlm_lock *lock;
1542 struct ldlm_resource *res;
1544 spin_lock(&exp->exp_ldlm_data.led_lock);
1545 while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1546 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1547 struct ldlm_lock, l_export_chain);
1548 res = ldlm_resource_getref(lock->l_resource);
1549 LDLM_LOCK_GET(lock);
1550 spin_unlock(&exp->exp_ldlm_data.led_lock);
1552 LDLM_DEBUG(lock, "export %p", exp);
1553 ldlm_res_lvbo_update(res, NULL, 0, 1);
1555 ldlm_lock_cancel(lock);
1556 ldlm_reprocess_all(res);
1558 ldlm_resource_putref(res);
1559 LDLM_LOCK_PUT(lock);
1560 spin_lock(&exp->exp_ldlm_data.led_lock);
1562 spin_unlock(&exp->exp_ldlm_data.led_lock);
1565 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1568 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1569 struct ldlm_resource *res;
1570 struct ldlm_namespace *ns;
1573 struct ldlm_lock *mark_lock = NULL;
1574 int join = LDLM_JOIN_NONE;
1578 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1579 *flags |= LDLM_FL_BLOCK_GRANTED;
1580 RETURN(lock->l_resource);
1583 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1584 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1586 lock_res_and_lock(lock);
1588 res = lock->l_resource;
1589 ns = res->lr_namespace;
1591 old_mode = lock->l_req_mode;
1592 lock->l_req_mode = new_mode;
1593 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1594 /* remember the lock position where the lock might be
1595 * added back to the granted list later and also
1596 * remember the join mode for skiplist fixing. */
1597 if (LDLM_SL_HEAD(&lock->l_sl_mode))
1598 join = LDLM_MODE_JOIN_RIGHT;
1599 else if (LDLM_SL_TAIL(&lock->l_sl_mode))
1600 join = LDLM_MODE_JOIN_LEFT;
1601 if (LDLM_SL_HEAD(&lock->l_sl_policy))
1602 join |= LDLM_POLICY_JOIN_RIGHT;
1603 else if (LDLM_SL_TAIL(&lock->l_sl_policy))
1604 join |= LDLM_POLICY_JOIN_LEFT;
1606 LASSERT(!((join & LDLM_MODE_JOIN_RIGHT) &&
1607 (join & LDLM_POLICY_JOIN_LEFT)));
1608 LASSERT(!((join & LDLM_MODE_JOIN_LEFT) &&
1609 (join & LDLM_POLICY_JOIN_RIGHT)));
1611 if ((join & LDLM_MODE_JOIN_LEFT) ||
1612 (join & LDLM_POLICY_JOIN_LEFT))
1613 mark_lock = list_entry(lock->l_res_link.prev,
1614 struct ldlm_lock, l_res_link);
1615 else if (lock->l_res_link.next != &res->lr_granted)
1616 mark_lock = list_entry(lock->l_res_link.next,
1617 struct ldlm_lock, l_res_link);
1619 ldlm_resource_unlink_lock(lock);
1621 /* If this is a local resource, put it on the appropriate list. */
1622 if (res->lr_namespace->ns_client) {
1623 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1624 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1626 /* This should never happen, because of the way the
1627 * server handles conversions. */
1628 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1632 ldlm_grant_lock(lock, &rpc_list);
1634 /* FIXME: completion handling not with ns_lock held ! */
1635 if (lock->l_completion_ast)
1636 lock->l_completion_ast(lock, 0, NULL);
1640 ldlm_processing_policy policy;
1641 policy = ldlm_processing_policy_table[res->lr_type];
1642 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1643 if (rc == LDLM_ITER_STOP) {
1644 lock->l_req_mode = old_mode;
1645 ldlm_granted_list_add_lock(lock, mark_lock, join);
1648 *flags |= LDLM_FL_BLOCK_GRANTED;
1652 unlock_res_and_lock(lock);
1655 ldlm_run_cp_ast_work(&rpc_list);
1659 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1661 struct obd_device *obd = NULL;
1663 if (!((libcfs_debug | D_ERROR) & level))
1667 CDEBUG(level, " NULL LDLM lock\n");
1671 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1672 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1674 if (lock->l_conn_export != NULL)
1675 obd = lock->l_conn_export->exp_obd;
1676 if (lock->l_export && lock->l_export->exp_connection) {
1677 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1678 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1679 lock->l_remote_handle.cookie);
1680 } else if (obd == NULL) {
1681 CDEBUG(level, " Node: local\n");
1683 struct obd_import *imp = obd->u.cli.cl_import;
1684 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1685 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1686 lock->l_remote_handle.cookie);
1688 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1690 lock->l_resource->lr_name.name[0],
1691 lock->l_resource->lr_name.name[1],
1692 lock->l_resource->lr_name.name[2]);
1693 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1694 "write: %d flags: %#x\n", ldlm_lockname[lock->l_req_mode],
1695 ldlm_lockname[lock->l_granted_mode],
1696 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1698 if (lock->l_resource->lr_type == LDLM_EXTENT)
1699 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1700 " (req "LPU64"-"LPU64")\n",
1701 lock->l_policy_data.l_extent.start,
1702 lock->l_policy_data.l_extent.end,
1703 lock->l_req_extent.start, lock->l_req_extent.end);
1704 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1705 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1706 lock->l_policy_data.l_flock.pid,
1707 lock->l_policy_data.l_flock.start,
1708 lock->l_policy_data.l_flock.end);
1709 else if (lock->l_resource->lr_type == LDLM_IBITS)
1710 CDEBUG(level, " Bits: "LPX64"\n",
1711 lock->l_policy_data.l_inodebits.bits);
1714 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1716 struct ldlm_lock *lock;
1718 if (!((libcfs_debug | D_ERROR) & level))
1721 lock = ldlm_handle2lock(lockh);
1725 ldlm_lock_dump(D_OTHER, lock, 0);
1727 LDLM_LOCK_PUT(lock);
1730 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1731 struct libcfs_debug_msg_data *data, const char *fmt,
1735 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1737 va_start(args, fmt);
1739 if (lock->l_resource == NULL) {
1740 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1741 data->msg_fn, data->msg_line, fmt, args,
1742 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1743 "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "
1744 LPX64" expref: %d pid: %u\n", lock,
1745 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1746 lock->l_readers, lock->l_writers,
1747 ldlm_lockname[lock->l_granted_mode],
1748 ldlm_lockname[lock->l_req_mode],
1749 lock->l_flags, lock->l_remote_handle.cookie,
1751 atomic_read(&lock->l_export->exp_refcount) : -99,
1757 switch (lock->l_resource->lr_type) {
1759 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1760 data->msg_fn, data->msg_line, fmt, args,
1761 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1762 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1763 "] (req "LPU64"->"LPU64") flags: %x remote: "LPX64
1764 " expref: %d pid: %u\n",
1765 lock->l_resource->lr_namespace->ns_name, lock,
1766 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1767 lock->l_readers, lock->l_writers,
1768 ldlm_lockname[lock->l_granted_mode],
1769 ldlm_lockname[lock->l_req_mode],
1770 lock->l_resource->lr_name.name[0],
1771 lock->l_resource->lr_name.name[1],
1772 atomic_read(&lock->l_resource->lr_refcount),
1773 ldlm_typename[lock->l_resource->lr_type],
1774 lock->l_policy_data.l_extent.start,
1775 lock->l_policy_data.l_extent.end,
1776 lock->l_req_extent.start, lock->l_req_extent.end,
1777 lock->l_flags, lock->l_remote_handle.cookie,
1779 atomic_read(&lock->l_export->exp_refcount) : -99,
1784 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1785 data->msg_fn, data->msg_line, fmt, args,
1786 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1787 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1788 "["LPU64"->"LPU64"] flags: %x remote: "LPX64
1789 " expref: %d pid: %u\n",
1790 lock->l_resource->lr_namespace->ns_name, lock,
1791 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1792 lock->l_readers, lock->l_writers,
1793 ldlm_lockname[lock->l_granted_mode],
1794 ldlm_lockname[lock->l_req_mode],
1795 lock->l_resource->lr_name.name[0],
1796 lock->l_resource->lr_name.name[1],
1797 atomic_read(&lock->l_resource->lr_refcount),
1798 ldlm_typename[lock->l_resource->lr_type],
1799 lock->l_policy_data.l_flock.pid,
1800 lock->l_policy_data.l_flock.start,
1801 lock->l_policy_data.l_flock.end,
1802 lock->l_flags, lock->l_remote_handle.cookie,
1804 atomic_read(&lock->l_export->exp_refcount) : -99,
1809 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1810 data->msg_fn, data->msg_line, fmt, args,
1811 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1812 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1813 "flags: %x remote: "LPX64" expref: %d "
1815 lock->l_resource->lr_namespace->ns_name,
1816 lock, lock->l_handle.h_cookie,
1817 atomic_read (&lock->l_refc),
1818 lock->l_readers, lock->l_writers,
1819 ldlm_lockname[lock->l_granted_mode],
1820 ldlm_lockname[lock->l_req_mode],
1821 lock->l_resource->lr_name.name[0],
1822 lock->l_resource->lr_name.name[1],
1823 lock->l_policy_data.l_inodebits.bits,
1824 atomic_read(&lock->l_resource->lr_refcount),
1825 ldlm_typename[lock->l_resource->lr_type],
1826 lock->l_flags, lock->l_remote_handle.cookie,
1828 atomic_read(&lock->l_export->exp_refcount) : -99,
1833 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1834 data->msg_fn, data->msg_line, fmt, args,
1835 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1836 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: %x "
1837 "remote: "LPX64" expref: %d pid: %u\n",
1838 lock->l_resource->lr_namespace->ns_name,
1839 lock, lock->l_handle.h_cookie,
1840 atomic_read (&lock->l_refc),
1841 lock->l_readers, lock->l_writers,
1842 ldlm_lockname[lock->l_granted_mode],
1843 ldlm_lockname[lock->l_req_mode],
1844 lock->l_resource->lr_name.name[0],
1845 lock->l_resource->lr_name.name[1],
1846 atomic_read(&lock->l_resource->lr_refcount),
1847 ldlm_typename[lock->l_resource->lr_type],
1848 lock->l_flags, lock->l_remote_handle.cookie,
1850 atomic_read(&lock->l_export->exp_refcount) : -99,
1856 EXPORT_SYMBOL(_ldlm_lock_debug);