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 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
172 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
175 if (!list_empty(&lock->l_lru)) {
176 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
177 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
178 list_del_init(&lock->l_lru);
180 LASSERT(ns->ns_nr_unused >= 0);
186 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
188 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
191 spin_lock(&ns->ns_unused_lock);
192 rc = ldlm_lock_remove_from_lru_nolock(lock);
193 spin_unlock(&ns->ns_unused_lock);
198 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
200 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
201 lock->l_last_used = cfs_time_current();
202 LASSERT(list_empty(&lock->l_lru));
203 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
204 LASSERT(ns->ns_nr_unused >= 0);
208 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
210 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
212 spin_lock(&ns->ns_unused_lock);
213 ldlm_lock_add_to_lru_nolock(lock);
214 spin_unlock(&ns->ns_unused_lock);
218 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
220 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
222 spin_lock(&ns->ns_unused_lock);
223 if (!list_empty(&lock->l_lru)) {
224 ldlm_lock_remove_from_lru_nolock(lock);
225 ldlm_lock_add_to_lru_nolock(lock);
227 spin_unlock(&ns->ns_unused_lock);
231 /* This used to have a 'strict' flag, which recovery would use to mark an
232 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
233 * shall explain why it's gone: with the new hash table scheme, once you call
234 * ldlm_lock_destroy, you can never drop your final references on this lock.
235 * Because it's not in the hash table anymore. -phil */
236 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
240 if (lock->l_readers || lock->l_writers) {
241 LDLM_ERROR(lock, "lock still has references");
242 ldlm_lock_dump(D_ERROR, lock, 0);
246 if (!list_empty(&lock->l_res_link)) {
247 LDLM_ERROR(lock, "lock still on resource");
248 ldlm_lock_dump(D_ERROR, lock, 0);
252 if (lock->l_destroyed) {
253 LASSERT(list_empty(&lock->l_lru));
257 lock->l_destroyed = 1;
260 spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
261 list_del_init(&lock->l_export_chain);
263 spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
265 ldlm_lock_remove_from_lru(lock);
266 class_handle_unhash(&lock->l_handle);
269 /* Wake anyone waiting for this lock */
270 /* FIXME: I should probably add yet another flag, instead of using
271 * l_export to only call this on clients */
273 class_export_put(lock->l_export);
274 lock->l_export = NULL;
275 if (lock->l_export && lock->l_completion_ast)
276 lock->l_completion_ast(lock, 0);
282 void ldlm_lock_destroy(struct ldlm_lock *lock)
286 lock_res_and_lock(lock);
287 first = ldlm_lock_destroy_internal(lock);
288 unlock_res_and_lock(lock);
290 /* drop reference from hashtable only for first destroy */
296 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
300 first = ldlm_lock_destroy_internal(lock);
301 /* drop reference from hashtable only for first destroy */
307 /* this is called by portals_handle2object with the handle lock taken */
308 static void lock_handle_addref(void *lock)
310 LDLM_LOCK_GET((struct ldlm_lock *)lock);
314 * usage: pass in a resource on which you have done ldlm_resource_get
315 * pass in a parent lock on which you have done a ldlm_lock_get
316 * after return, ldlm_*_put the resource and parent
317 * returns: lock with refcount 2 - one for current caller and one for remote
319 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
321 struct ldlm_lock *lock;
324 if (resource == NULL)
327 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, CFS_ALLOC_IO, sizeof(*lock));
331 spin_lock_init(&lock->l_lock);
332 lock->l_resource = ldlm_resource_getref(resource);
334 atomic_set(&lock->l_refc, 2);
335 CFS_INIT_LIST_HEAD(&lock->l_res_link);
336 CFS_INIT_LIST_HEAD(&lock->l_lru);
337 CFS_INIT_LIST_HEAD(&lock->l_export_chain);
338 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
339 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
340 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
341 cfs_waitq_init(&lock->l_waitq);
342 lock->l_blocking_lock = NULL;
343 lock->l_sl_mode.prev = NULL;
344 lock->l_sl_mode.next = NULL;
345 lock->l_sl_policy.prev = NULL;
346 lock->l_sl_policy.next = NULL;
348 atomic_inc(&resource->lr_namespace->ns_locks);
349 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
350 class_handle_hash(&lock->l_handle, lock_handle_addref);
355 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
356 const struct ldlm_res_id *new_resid)
358 struct ldlm_resource *oldres = lock->l_resource;
359 struct ldlm_resource *newres;
363 LASSERT(ns_is_client(ns));
365 lock_res_and_lock(lock);
366 if (memcmp(new_resid, &lock->l_resource->lr_name,
367 sizeof(lock->l_resource->lr_name)) == 0) {
369 unlock_res_and_lock(lock);
373 LASSERT(new_resid->name[0] != 0);
375 /* This function assumes that the lock isn't on any lists */
376 LASSERT(list_empty(&lock->l_res_link));
378 type = oldres->lr_type;
379 unlock_res_and_lock(lock);
381 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
382 if (newres == NULL) {
387 lock_res_and_lock(lock);
388 LASSERT(memcmp(new_resid, &lock->l_resource->lr_name,
389 sizeof(lock->l_resource->lr_name)) != 0);
391 lock->l_resource = newres;
393 unlock_res_and_lock(lock);
395 /* ...and the flowers are still standing! */
396 ldlm_resource_putref(oldres);
405 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
407 lockh->cookie = lock->l_handle.h_cookie;
410 /* if flags: atomically get the lock and set the flags.
411 * Return NULL if flag already set
414 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
417 struct ldlm_namespace *ns;
418 struct ldlm_lock *lock, *retval = NULL;
423 lock = class_handle2object(handle->cookie);
427 LASSERT(lock->l_resource != NULL);
428 ns = lock->l_resource->lr_namespace;
431 lock_res_and_lock(lock);
433 /* It's unlikely but possible that someone marked the lock as
434 * destroyed after we did handle2object on it */
435 if (lock->l_destroyed) {
436 unlock_res_and_lock(lock);
437 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
442 if (flags && (lock->l_flags & flags)) {
443 unlock_res_and_lock(lock);
449 lock->l_flags |= flags;
451 unlock_res_and_lock(lock);
458 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
459 const struct lustre_handle *handle)
461 struct ldlm_lock *retval = NULL;
462 retval = __ldlm_handle2lock(handle, 0);
466 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
468 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
469 /* INODEBITS_INTEROP: If the other side does not support
470 * inodebits, reply with a plain lock descriptor.
472 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
473 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
474 struct ldlm_resource res = *lock->l_resource;
476 /* Make sure all the right bits are set in this lock we
477 are going to pass to client */
478 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
479 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
480 "Inappropriate inode lock bits during "
481 "conversion " LPU64 "\n",
482 lock->l_policy_data.l_inodebits.bits);
483 res.lr_type = LDLM_PLAIN;
484 ldlm_res2desc(&res, &desc->l_resource);
485 /* Convert "new" lock mode to something old client can
487 if ((lock->l_req_mode == LCK_CR) ||
488 (lock->l_req_mode == LCK_CW))
489 desc->l_req_mode = LCK_PR;
491 desc->l_req_mode = lock->l_req_mode;
492 if ((lock->l_granted_mode == LCK_CR) ||
493 (lock->l_granted_mode == LCK_CW)) {
494 desc->l_granted_mode = LCK_PR;
496 /* We never grant PW/EX locks to clients */
497 LASSERT((lock->l_granted_mode != LCK_PW) &&
498 (lock->l_granted_mode != LCK_EX));
499 desc->l_granted_mode = lock->l_granted_mode;
502 /* We do not copy policy here, because there is no
503 policy for plain locks */
505 ldlm_res2desc(lock->l_resource, &desc->l_resource);
506 desc->l_req_mode = lock->l_req_mode;
507 desc->l_granted_mode = lock->l_granted_mode;
508 desc->l_policy_data = lock->l_policy_data;
512 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
513 struct list_head *work_list)
515 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
516 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
517 lock->l_flags |= LDLM_FL_AST_SENT;
518 /* If the enqueuing client said so, tell the AST recipient to
519 * discard dirty data, rather than writing back. */
520 if (new->l_flags & LDLM_AST_DISCARD_DATA)
521 lock->l_flags |= LDLM_FL_DISCARD_DATA;
522 LASSERT(list_empty(&lock->l_bl_ast));
523 list_add(&lock->l_bl_ast, work_list);
525 LASSERT(lock->l_blocking_lock == NULL);
526 lock->l_blocking_lock = LDLM_LOCK_GET(new);
530 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
532 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
533 lock->l_flags |= LDLM_FL_CP_REQD;
534 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
535 LASSERT(list_empty(&lock->l_cp_ast));
536 list_add(&lock->l_cp_ast, work_list);
541 /* must be called with lr_lock held */
542 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
543 struct list_head *work_list)
546 check_res_locked(lock->l_resource);
548 ldlm_add_bl_work_item(lock, new, work_list);
550 ldlm_add_cp_work_item(lock, work_list);
554 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
556 struct ldlm_lock *lock;
558 lock = ldlm_handle2lock(lockh);
559 LASSERT(lock != NULL);
560 ldlm_lock_addref_internal(lock, mode);
564 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
566 ldlm_lock_remove_from_lru(lock);
567 if (mode & (LCK_NL | LCK_CR | LCK_PR))
569 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
572 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
575 /* only called for local locks */
576 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
578 lock_res_and_lock(lock);
579 ldlm_lock_addref_internal_nolock(lock, mode);
580 unlock_res_and_lock(lock);
583 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
585 struct ldlm_namespace *ns;
588 lock_res_and_lock(lock);
590 ns = lock->l_resource->lr_namespace;
592 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
593 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
594 LASSERT(lock->l_readers > 0);
597 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
598 LASSERT(lock->l_writers > 0);
602 if (lock->l_flags & LDLM_FL_LOCAL &&
603 !lock->l_readers && !lock->l_writers) {
604 /* If this is a local lock on a server namespace and this was
605 * the last reference, cancel the lock. */
606 CDEBUG(D_INFO, "forcing cancel of local lock\n");
607 lock->l_flags |= LDLM_FL_CBPENDING;
610 if (!lock->l_readers && !lock->l_writers &&
611 (lock->l_flags & LDLM_FL_CBPENDING)) {
612 /* If we received a blocked AST and this was the last reference,
613 * run the callback. */
614 if (ns_is_server(ns) && lock->l_export)
615 CERROR("FL_CBPENDING set on non-local lock--just a "
618 LDLM_DEBUG(lock, "final decref done on cbpending lock");
620 LDLM_LOCK_GET(lock); /* dropped by bl thread */
621 ldlm_lock_remove_from_lru(lock);
622 unlock_res_and_lock(lock);
623 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
624 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
625 ldlm_handle_bl_callback(ns, NULL, lock);
626 } else if (ns_is_client(ns) &&
627 !lock->l_readers && !lock->l_writers &&
628 !(lock->l_flags & LDLM_FL_NO_LRU)) {
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);
883 ldlm_resource_add_lock(res, &res->lr_granted, lock);
885 if (lock->l_granted_mode < res->lr_most_restr)
886 res->lr_most_restr = lock->l_granted_mode;
888 if (work_list && lock->l_completion_ast != NULL)
889 ldlm_add_ast_work_item(lock, NULL, work_list);
891 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
895 /* returns a referenced lock or NULL. See the flag descriptions below, in the
896 * comment above ldlm_lock_match */
897 static struct ldlm_lock *search_queue(struct list_head *queue,
899 ldlm_policy_data_t *policy,
900 struct ldlm_lock *old_lock, int flags)
902 struct ldlm_lock *lock;
903 struct list_head *tmp;
905 list_for_each(tmp, queue) {
908 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
910 if (lock == old_lock)
913 /* llite sometimes wants to match locks that will be
914 * canceled when their users drop, but we allow it to match
915 * if it passes in CBPENDING and the lock still has users.
916 * this is generally only going to be used by children
917 * whose parents already hold a lock so forward progress
918 * can still happen. */
919 if (lock->l_flags & LDLM_FL_CBPENDING &&
920 !(flags & LDLM_FL_CBPENDING))
922 if (lock->l_flags & LDLM_FL_CBPENDING &&
923 lock->l_readers == 0 && lock->l_writers == 0)
926 if (!(lock->l_req_mode & *mode))
928 match = lock->l_req_mode;
930 if (lock->l_resource->lr_type == LDLM_EXTENT &&
931 (lock->l_policy_data.l_extent.start >
932 policy->l_extent.start ||
933 lock->l_policy_data.l_extent.end < policy->l_extent.end))
936 if (unlikely(match == LCK_GROUP) &&
937 lock->l_resource->lr_type == LDLM_EXTENT &&
938 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
941 /* We match if we have existing lock with same or wider set
943 if (lock->l_resource->lr_type == LDLM_IBITS &&
944 ((lock->l_policy_data.l_inodebits.bits &
945 policy->l_inodebits.bits) !=
946 policy->l_inodebits.bits))
949 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
952 if ((flags & LDLM_FL_LOCAL_ONLY) &&
953 !(lock->l_flags & LDLM_FL_LOCAL))
956 if (flags & LDLM_FL_TEST_LOCK) {
958 ldlm_lock_touch_in_lru(lock);
960 ldlm_lock_addref_internal_nolock(lock, match);
969 void ldlm_lock_allow_match(struct ldlm_lock *lock)
971 lock_res_and_lock(lock);
972 lock->l_flags |= LDLM_FL_LVB_READY;
973 cfs_waitq_signal(&lock->l_waitq);
974 unlock_res_and_lock(lock);
977 /* Can be called in two ways:
979 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
980 * for a duplicate of.
982 * Otherwise, all of the fields must be filled in, to match against.
984 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
985 * server (ie, connh is NULL)
986 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
987 * list will be considered
988 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
989 * to be canceled can still be matched as long as they still have reader
990 * or writer refernces
991 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
992 * just tell us if we would have matched.
994 * Returns 1 if it finds an already-existing lock that is compatible; in this
995 * case, lockh is filled in with a addref()ed lock
997 * we also check security context, if that failed we simply return 0 (to keep
998 * caller code unchanged), the context failure will be discovered by caller
1001 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1002 const struct ldlm_res_id *res_id, ldlm_type_t type,
1003 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1004 struct lustre_handle *lockh)
1006 struct ldlm_resource *res;
1007 struct ldlm_lock *lock, *old_lock = NULL;
1012 old_lock = ldlm_handle2lock(lockh);
1015 ns = old_lock->l_resource->lr_namespace;
1016 res_id = &old_lock->l_resource->lr_name;
1017 type = old_lock->l_resource->lr_type;
1018 mode = old_lock->l_req_mode;
1021 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1023 LASSERT(old_lock == NULL);
1029 lock = search_queue(&res->lr_granted, &mode, policy, old_lock, flags);
1032 if (flags & LDLM_FL_BLOCK_GRANTED)
1034 lock = search_queue(&res->lr_converting, &mode, policy, old_lock, flags);
1037 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock, flags);
1044 ldlm_resource_putref(res);
1047 ldlm_lock2handle(lock, lockh);
1048 if ((flags & LDLM_FL_LVB_READY) &&
1049 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1050 struct l_wait_info lwi;
1051 if (lock->l_completion_ast) {
1052 int err = lock->l_completion_ast(lock,
1053 LDLM_FL_WAIT_NOREPROC,
1056 if (flags & LDLM_FL_TEST_LOCK)
1057 LDLM_LOCK_PUT(lock);
1059 ldlm_lock_decref_internal(lock, mode);
1065 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), NULL,
1066 LWI_ON_SIGNAL_NOOP, NULL);
1068 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1069 l_wait_event(lock->l_waitq,
1070 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1075 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1076 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1077 res_id->name[2] : policy->l_extent.start,
1078 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1079 res_id->name[3] : policy->l_extent.end);
1081 /* check user's security context */
1082 if (lock->l_conn_export &&
1083 sptlrpc_import_check_ctx(
1084 class_exp2cliimp(lock->l_conn_export))) {
1085 if (!(flags & LDLM_FL_TEST_LOCK))
1086 ldlm_lock_decref_internal(lock, mode);
1090 if (flags & LDLM_FL_TEST_LOCK)
1091 LDLM_LOCK_PUT(lock);
1093 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1094 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1095 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1096 type, mode, res_id->name[0], res_id->name[1],
1097 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1098 res_id->name[2] :policy->l_extent.start,
1099 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1100 res_id->name[3] : policy->l_extent.end);
1103 LDLM_LOCK_PUT(old_lock);
1105 return rc ? mode : 0;
1108 /* Returns a referenced lock */
1109 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1110 const struct ldlm_res_id *res_id,
1113 ldlm_blocking_callback blocking,
1114 ldlm_completion_callback completion,
1115 ldlm_glimpse_callback glimpse,
1116 void *data, __u32 lvb_len)
1118 struct ldlm_lock *lock;
1119 struct ldlm_resource *res;
1122 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1126 lock = ldlm_lock_new(res);
1127 ldlm_resource_putref(res);
1132 lock->l_req_mode = mode;
1133 lock->l_ast_data = data;
1134 lock->l_blocking_ast = blocking;
1135 lock->l_completion_ast = completion;
1136 lock->l_glimpse_ast = glimpse;
1137 lock->l_pid = cfs_curproc_pid();
1140 lock->l_lvb_len = lvb_len;
1141 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1142 if (lock->l_lvb_data == NULL) {
1143 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1151 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1152 struct ldlm_lock **lockp,
1153 void *cookie, int *flags)
1155 struct ldlm_lock *lock = *lockp;
1156 struct ldlm_resource *res = lock->l_resource;
1157 int local = ns_is_client(res->lr_namespace);
1158 ldlm_processing_policy policy;
1159 ldlm_error_t rc = ELDLM_OK;
1162 do_gettimeofday(&lock->l_enqueued_time);
1163 /* policies are not executed on the client or during replay */
1164 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1165 && !local && ns->ns_policy) {
1166 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1168 if (rc == ELDLM_LOCK_REPLACED) {
1169 /* The lock that was returned has already been granted,
1170 * and placed into lockp. If it's not the same as the
1171 * one we passed in, then destroy the old one and our
1172 * work here is done. */
1173 if (lock != *lockp) {
1174 ldlm_lock_destroy(lock);
1175 LDLM_LOCK_PUT(lock);
1177 *flags |= LDLM_FL_LOCK_CHANGED;
1179 } else if (rc != ELDLM_OK ||
1180 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1181 ldlm_lock_destroy(lock);
1186 lock_res_and_lock(lock);
1187 if (local && lock->l_req_mode == lock->l_granted_mode) {
1188 /* The server returned a blocked lock, but it was granted before
1189 * we got a chance to actually enqueue it. We don't need to do
1191 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1192 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1193 GOTO(out, ELDLM_OK);
1196 /* Some flags from the enqueue want to make it into the AST, via the
1197 * lock's l_flags. */
1198 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1200 /* This distinction between local lock trees is very important; a client
1201 * namespace only has information about locks taken by that client, and
1202 * thus doesn't have enough information to decide for itself if it can
1203 * be granted (below). In this case, we do exactly what the server
1204 * tells us to do, as dictated by the 'flags'.
1206 * We do exactly the same thing during recovery, when the server is
1207 * more or less trusting the clients not to lie.
1209 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1210 * granted/converting queues. */
1211 ldlm_resource_unlink_lock(lock);
1213 if (*flags & LDLM_FL_BLOCK_CONV)
1214 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1215 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1216 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1218 ldlm_grant_lock(lock, NULL);
1219 GOTO(out, ELDLM_OK);
1220 } else if (*flags & LDLM_FL_REPLAY) {
1221 if (*flags & LDLM_FL_BLOCK_CONV) {
1222 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1223 GOTO(out, ELDLM_OK);
1224 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1225 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1226 GOTO(out, ELDLM_OK);
1227 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1228 ldlm_grant_lock(lock, NULL);
1229 GOTO(out, ELDLM_OK);
1231 /* If no flags, fall through to normal enqueue path. */
1234 policy = ldlm_processing_policy_table[res->lr_type];
1235 policy(lock, flags, 1, &rc, NULL);
1238 unlock_res_and_lock(lock);
1242 /* Must be called with namespace taken: queue is waiting or converting. */
1243 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1244 struct list_head *work_list)
1246 struct list_head *tmp, *pos;
1247 ldlm_processing_policy policy;
1249 int rc = LDLM_ITER_CONTINUE;
1253 check_res_locked(res);
1255 policy = ldlm_processing_policy_table[res->lr_type];
1258 list_for_each_safe(tmp, pos, queue) {
1259 struct ldlm_lock *pending;
1260 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1262 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1265 rc = policy(pending, &flags, 0, &err, work_list);
1266 if (rc != LDLM_ITER_CONTINUE)
1273 /* Helper function for pair ldlm_run_{bl,cp}_ast_work().
1275 * Send an existing rpc set specified by @arg->set and then
1276 * destroy it. Create new one if @do_create flag is set. */
1278 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1282 rc = ptlrpc_set_wait(arg->set);
1283 if (arg->type == LDLM_BL_CALLBACK)
1284 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1285 ptlrpc_set_destroy(arg->set);
1288 arg->set = ptlrpc_prep_set();
1291 int ldlm_run_bl_ast_work(struct list_head *rpc_list)
1293 struct ldlm_cb_set_arg arg;
1294 struct list_head *tmp, *pos;
1295 struct ldlm_lock_desc d;
1300 arg.set = ptlrpc_prep_set();
1301 atomic_set(&arg.restart, 0);
1302 arg.type = LDLM_BL_CALLBACK;
1305 list_for_each_safe(tmp, pos, rpc_list) {
1306 struct ldlm_lock *lock =
1307 list_entry(tmp, struct ldlm_lock, l_bl_ast);
1309 /* nobody should touch l_bl_ast */
1310 lock_res_and_lock(lock);
1311 list_del_init(&lock->l_bl_ast);
1313 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1314 LASSERT(lock->l_bl_ast_run == 0);
1315 LASSERT(lock->l_blocking_lock);
1316 lock->l_bl_ast_run++;
1317 unlock_res_and_lock(lock);
1319 ldlm_lock2desc(lock->l_blocking_lock, &d);
1321 LDLM_LOCK_PUT(lock->l_blocking_lock);
1322 lock->l_blocking_lock = NULL;
1323 rc = lock->l_blocking_ast(lock, &d, (void *)&arg,
1325 LDLM_LOCK_PUT(lock);
1328 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1329 * and create a new set for requests that remained in
1331 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1332 ldlm_send_and_maybe_create_set(&arg, 1);
1338 ldlm_send_and_maybe_create_set(&arg, 0);
1340 /* In case when number of ASTs is multiply of
1341 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1342 * @arg.set must be destroyed here, otherwise we get
1343 * write memory leaking. */
1344 ptlrpc_set_destroy(arg.set);
1346 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1349 int ldlm_run_cp_ast_work(struct list_head *rpc_list)
1351 struct ldlm_cb_set_arg arg;
1352 struct list_head *tmp, *pos;
1357 arg.set = ptlrpc_prep_set();
1358 atomic_set(&arg.restart, 0);
1359 arg.type = LDLM_CP_CALLBACK;
1361 /* It's possible to receive a completion AST before we've set
1362 * the l_completion_ast pointer: either because the AST arrived
1363 * before the reply, or simply because there's a small race
1364 * window between receiving the reply and finishing the local
1365 * enqueue. (bug 842)
1367 * This can't happen with the blocking_ast, however, because we
1368 * will never call the local blocking_ast until we drop our
1369 * reader/writer reference, which we won't do until we get the
1370 * reply and finish enqueueing. */
1373 list_for_each_safe(tmp, pos, rpc_list) {
1374 struct ldlm_lock *lock =
1375 list_entry(tmp, struct ldlm_lock, l_cp_ast);
1376 ldlm_completion_callback completion_callback;
1378 /* nobody should touch l_cp_ast */
1379 lock_res_and_lock(lock);
1380 list_del_init(&lock->l_cp_ast);
1381 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1382 /* save l_completion_ast since it can be changed by
1383 * mds_intent_policy(), see bug 14225 */
1384 completion_callback = lock->l_completion_ast;
1385 lock->l_flags &= ~LDLM_FL_CP_REQD;
1386 unlock_res_and_lock(lock);
1388 if (completion_callback != NULL) {
1389 rc = completion_callback(lock, 0, (void *)&arg);
1392 LDLM_LOCK_PUT(lock);
1394 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1395 * and create a new set for requests that remained in
1397 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1398 ldlm_send_and_maybe_create_set(&arg, 1);
1404 ldlm_send_and_maybe_create_set(&arg, 0);
1406 /* In case when number of ASTs is multiply of
1407 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1408 * @arg.set must be destroyed here, otherwise we get
1409 * write memory leaking. */
1410 ptlrpc_set_destroy(arg.set);
1412 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1415 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1417 ldlm_reprocess_all(res);
1418 return LDLM_ITER_CONTINUE;
1421 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1423 struct list_head *tmp;
1430 spin_lock(&ns->ns_hash_lock);
1431 for (i = 0; i < RES_HASH_SIZE; i++) {
1432 tmp = ns->ns_hash[i].next;
1433 while (tmp != &(ns->ns_hash[i])) {
1434 struct ldlm_resource *res =
1435 list_entry(tmp, struct ldlm_resource, lr_hash);
1437 ldlm_resource_getref(res);
1438 spin_unlock(&ns->ns_hash_lock);
1440 rc = reprocess_one_queue(res, NULL);
1442 spin_lock(&ns->ns_hash_lock);
1444 ldlm_resource_putref_locked(res);
1446 if (rc == LDLM_ITER_STOP)
1451 spin_unlock(&ns->ns_hash_lock);
1455 void ldlm_reprocess_all(struct ldlm_resource *res)
1457 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1461 /* Local lock trees don't get reprocessed. */
1462 if (ns_is_client(res->lr_namespace)) {
1469 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1470 if (rc == LDLM_ITER_CONTINUE)
1471 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1474 rc = ldlm_run_cp_ast_work(&rpc_list);
1475 if (rc == -ERESTART) {
1476 LASSERT(list_empty(&rpc_list));
1482 void ldlm_cancel_callback(struct ldlm_lock *lock)
1484 check_res_locked(lock->l_resource);
1485 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1486 lock->l_flags |= LDLM_FL_CANCEL;
1487 if (lock->l_blocking_ast) {
1488 // l_check_no_ns_lock(ns);
1489 unlock_res_and_lock(lock);
1490 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1492 lock_res_and_lock(lock);
1494 LDLM_DEBUG(lock, "no blocking ast");
1497 lock->l_flags |= LDLM_FL_BL_DONE;
1500 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1502 struct ldlm_lock *lock;
1504 if (req->l_resource->lr_type != LDLM_PLAIN &&
1505 req->l_resource->lr_type != LDLM_IBITS)
1508 if (LDLM_SL_HEAD(&req->l_sl_mode)) {
1509 lock = list_entry(req->l_res_link.next, struct ldlm_lock,
1511 if (req->l_sl_mode.next == &lock->l_sl_mode) {
1512 lock->l_sl_mode.prev = NULL;
1514 lock->l_sl_mode.next = req->l_sl_mode.next;
1515 lock->l_sl_mode.next->prev = &lock->l_sl_mode;
1517 req->l_sl_mode.next = NULL;
1518 } else if (LDLM_SL_TAIL(&req->l_sl_mode)) {
1519 lock = list_entry(req->l_res_link.prev, struct ldlm_lock,
1521 if (req->l_sl_mode.prev == &lock->l_sl_mode) {
1522 lock->l_sl_mode.next = NULL;
1524 lock->l_sl_mode.prev = req->l_sl_mode.prev;
1525 lock->l_sl_mode.prev->next = &lock->l_sl_mode;
1527 req->l_sl_mode.prev = NULL;
1530 if (LDLM_SL_HEAD(&req->l_sl_policy)) {
1531 lock = list_entry(req->l_res_link.next, struct ldlm_lock,
1533 if (req->l_sl_policy.next == &lock->l_sl_policy) {
1534 lock->l_sl_policy.prev = NULL;
1536 lock->l_sl_policy.next = req->l_sl_policy.next;
1537 lock->l_sl_policy.next->prev = &lock->l_sl_policy;
1539 req->l_sl_policy.next = NULL;
1540 } else if (LDLM_SL_TAIL(&req->l_sl_policy)) {
1541 lock = list_entry(req->l_res_link.prev, struct ldlm_lock,
1543 if (req->l_sl_policy.prev == &lock->l_sl_policy) {
1544 lock->l_sl_policy.next = NULL;
1546 lock->l_sl_policy.prev = req->l_sl_policy.prev;
1547 lock->l_sl_policy.prev->next = &lock->l_sl_policy;
1549 req->l_sl_policy.prev = NULL;
1553 void ldlm_lock_cancel(struct ldlm_lock *lock)
1555 struct ldlm_resource *res;
1556 struct ldlm_namespace *ns;
1559 lock_res_and_lock(lock);
1561 res = lock->l_resource;
1562 ns = res->lr_namespace;
1564 /* Please do not, no matter how tempting, remove this LBUG without
1565 * talking to me first. -phik */
1566 if (lock->l_readers || lock->l_writers) {
1567 LDLM_ERROR(lock, "lock still has references");
1571 ldlm_del_waiting_lock(lock);
1573 /* Releases cancel callback. */
1574 ldlm_cancel_callback(lock);
1576 /* Yes, second time, just in case it was added again while we were
1577 running with no res lock in ldlm_cancel_callback */
1578 ldlm_del_waiting_lock(lock);
1579 ldlm_resource_unlink_lock(lock);
1580 ldlm_lock_destroy_nolock(lock);
1582 if (lock->l_granted_mode == lock->l_req_mode)
1583 ldlm_pool_del(&ns->ns_pool, lock);
1585 /* Make sure we will not be called again for same lock what is possible
1586 * if not to zero out lock->l_granted_mode */
1587 lock->l_granted_mode = 0;
1588 unlock_res_and_lock(lock);
1593 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1595 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1601 lock->l_ast_data = data;
1602 LDLM_LOCK_PUT(lock);
1606 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1608 struct ldlm_lock *lock;
1609 struct ldlm_resource *res;
1611 spin_lock(&exp->exp_ldlm_data.led_lock);
1612 while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1613 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1614 struct ldlm_lock, l_export_chain);
1615 res = ldlm_resource_getref(lock->l_resource);
1616 LDLM_LOCK_GET(lock);
1617 spin_unlock(&exp->exp_ldlm_data.led_lock);
1619 LDLM_DEBUG(lock, "export %p", exp);
1620 ldlm_res_lvbo_update(res, NULL, 0, 1);
1622 ldlm_lock_cancel(lock);
1623 ldlm_reprocess_all(res);
1625 ldlm_resource_putref(res);
1626 LDLM_LOCK_PUT(lock);
1627 spin_lock(&exp->exp_ldlm_data.led_lock);
1629 spin_unlock(&exp->exp_ldlm_data.led_lock);
1632 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1635 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1636 struct ldlm_resource *res;
1637 struct ldlm_namespace *ns;
1640 struct ldlm_lock *mark_lock = NULL;
1641 int join = LDLM_JOIN_NONE;
1645 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1646 *flags |= LDLM_FL_BLOCK_GRANTED;
1647 RETURN(lock->l_resource);
1650 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1651 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1653 lock_res_and_lock(lock);
1655 res = lock->l_resource;
1656 ns = res->lr_namespace;
1658 old_mode = lock->l_req_mode;
1659 lock->l_req_mode = new_mode;
1660 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1661 /* remember the lock position where the lock might be
1662 * added back to the granted list later and also
1663 * remember the join mode for skiplist fixing. */
1664 if (LDLM_SL_HEAD(&lock->l_sl_mode))
1665 join = LDLM_MODE_JOIN_RIGHT;
1666 else if (LDLM_SL_TAIL(&lock->l_sl_mode))
1667 join = LDLM_MODE_JOIN_LEFT;
1668 if (LDLM_SL_HEAD(&lock->l_sl_policy))
1669 join |= LDLM_POLICY_JOIN_RIGHT;
1670 else if (LDLM_SL_TAIL(&lock->l_sl_policy))
1671 join |= LDLM_POLICY_JOIN_LEFT;
1673 LASSERT(!((join & LDLM_MODE_JOIN_RIGHT) &&
1674 (join & LDLM_POLICY_JOIN_LEFT)));
1675 LASSERT(!((join & LDLM_MODE_JOIN_LEFT) &&
1676 (join & LDLM_POLICY_JOIN_RIGHT)));
1678 if ((join & LDLM_MODE_JOIN_LEFT) ||
1679 (join & LDLM_POLICY_JOIN_LEFT))
1680 mark_lock = list_entry(lock->l_res_link.prev,
1681 struct ldlm_lock, l_res_link);
1682 else if (lock->l_res_link.next != &res->lr_granted)
1683 mark_lock = list_entry(lock->l_res_link.next,
1684 struct ldlm_lock, l_res_link);
1686 ldlm_resource_unlink_lock(lock);
1688 /* If this is a local resource, put it on the appropriate list. */
1689 if (ns_is_client(res->lr_namespace)) {
1690 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1691 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1693 /* This should never happen, because of the way the
1694 * server handles conversions. */
1695 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1699 ldlm_grant_lock(lock, &rpc_list);
1701 /* FIXME: completion handling not with ns_lock held ! */
1702 if (lock->l_completion_ast)
1703 lock->l_completion_ast(lock, 0, NULL);
1707 ldlm_processing_policy policy;
1708 policy = ldlm_processing_policy_table[res->lr_type];
1709 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1710 if (rc == LDLM_ITER_STOP) {
1711 lock->l_req_mode = old_mode;
1712 ldlm_granted_list_add_lock(lock, mark_lock, join);
1715 *flags |= LDLM_FL_BLOCK_GRANTED;
1719 unlock_res_and_lock(lock);
1722 ldlm_run_cp_ast_work(&rpc_list);
1726 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1728 struct obd_device *obd = NULL;
1730 if (!((libcfs_debug | D_ERROR) & level))
1734 CDEBUG(level, " NULL LDLM lock\n");
1738 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1739 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1741 if (lock->l_conn_export != NULL)
1742 obd = lock->l_conn_export->exp_obd;
1743 if (lock->l_export && lock->l_export->exp_connection) {
1744 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1745 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1746 lock->l_remote_handle.cookie);
1747 } else if (obd == NULL) {
1748 CDEBUG(level, " Node: local\n");
1750 struct obd_import *imp = obd->u.cli.cl_import;
1751 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1752 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1753 lock->l_remote_handle.cookie);
1755 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1757 lock->l_resource->lr_name.name[0],
1758 lock->l_resource->lr_name.name[1],
1759 lock->l_resource->lr_name.name[2]);
1760 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1761 "write: %d flags: %#x\n", ldlm_lockname[lock->l_req_mode],
1762 ldlm_lockname[lock->l_granted_mode],
1763 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1765 if (lock->l_resource->lr_type == LDLM_EXTENT)
1766 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1767 " (req "LPU64"-"LPU64")\n",
1768 lock->l_policy_data.l_extent.start,
1769 lock->l_policy_data.l_extent.end,
1770 lock->l_req_extent.start, lock->l_req_extent.end);
1771 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1772 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1773 lock->l_policy_data.l_flock.pid,
1774 lock->l_policy_data.l_flock.start,
1775 lock->l_policy_data.l_flock.end);
1776 else if (lock->l_resource->lr_type == LDLM_IBITS)
1777 CDEBUG(level, " Bits: "LPX64"\n",
1778 lock->l_policy_data.l_inodebits.bits);
1781 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1783 struct ldlm_lock *lock;
1785 if (!((libcfs_debug | D_ERROR) & level))
1788 lock = ldlm_handle2lock(lockh);
1792 ldlm_lock_dump(D_OTHER, lock, 0);
1794 LDLM_LOCK_PUT(lock);
1797 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1798 struct libcfs_debug_msg_data *data, const char *fmt,
1802 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1804 va_start(args, fmt);
1806 if (lock->l_resource == NULL) {
1807 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1808 data->msg_fn, data->msg_line, fmt, args,
1809 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1810 "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "
1811 LPX64" expref: %d pid: %u\n", lock,
1812 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1813 lock->l_readers, lock->l_writers,
1814 ldlm_lockname[lock->l_granted_mode],
1815 ldlm_lockname[lock->l_req_mode],
1816 lock->l_flags, lock->l_remote_handle.cookie,
1818 atomic_read(&lock->l_export->exp_refcount) : -99,
1824 switch (lock->l_resource->lr_type) {
1826 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1827 data->msg_fn, data->msg_line, fmt, args,
1828 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1829 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1830 "] (req "LPU64"->"LPU64") flags: %x remote: "LPX64
1831 " expref: %d pid: %u\n",
1832 lock->l_resource->lr_namespace->ns_name, lock,
1833 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1834 lock->l_readers, lock->l_writers,
1835 ldlm_lockname[lock->l_granted_mode],
1836 ldlm_lockname[lock->l_req_mode],
1837 lock->l_resource->lr_name.name[0],
1838 lock->l_resource->lr_name.name[1],
1839 atomic_read(&lock->l_resource->lr_refcount),
1840 ldlm_typename[lock->l_resource->lr_type],
1841 lock->l_policy_data.l_extent.start,
1842 lock->l_policy_data.l_extent.end,
1843 lock->l_req_extent.start, lock->l_req_extent.end,
1844 lock->l_flags, lock->l_remote_handle.cookie,
1846 atomic_read(&lock->l_export->exp_refcount) : -99,
1851 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1852 data->msg_fn, data->msg_line, fmt, args,
1853 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1854 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1855 "["LPU64"->"LPU64"] flags: %x remote: "LPX64
1856 " expref: %d pid: %u\n",
1857 lock->l_resource->lr_namespace->ns_name, lock,
1858 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1859 lock->l_readers, lock->l_writers,
1860 ldlm_lockname[lock->l_granted_mode],
1861 ldlm_lockname[lock->l_req_mode],
1862 lock->l_resource->lr_name.name[0],
1863 lock->l_resource->lr_name.name[1],
1864 atomic_read(&lock->l_resource->lr_refcount),
1865 ldlm_typename[lock->l_resource->lr_type],
1866 lock->l_policy_data.l_flock.pid,
1867 lock->l_policy_data.l_flock.start,
1868 lock->l_policy_data.l_flock.end,
1869 lock->l_flags, lock->l_remote_handle.cookie,
1871 atomic_read(&lock->l_export->exp_refcount) : -99,
1876 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1877 data->msg_fn, data->msg_line, fmt, args,
1878 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1879 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1880 "flags: %x remote: "LPX64" expref: %d "
1882 lock->l_resource->lr_namespace->ns_name,
1883 lock, lock->l_handle.h_cookie,
1884 atomic_read (&lock->l_refc),
1885 lock->l_readers, lock->l_writers,
1886 ldlm_lockname[lock->l_granted_mode],
1887 ldlm_lockname[lock->l_req_mode],
1888 lock->l_resource->lr_name.name[0],
1889 lock->l_resource->lr_name.name[1],
1890 lock->l_policy_data.l_inodebits.bits,
1891 atomic_read(&lock->l_resource->lr_refcount),
1892 ldlm_typename[lock->l_resource->lr_type],
1893 lock->l_flags, lock->l_remote_handle.cookie,
1895 atomic_read(&lock->l_export->exp_refcount) : -99,
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 flags: %x "
1904 "remote: "LPX64" expref: %d pid: %u\n",
1905 lock->l_resource->lr_namespace->ns_name,
1906 lock, lock->l_handle.h_cookie,
1907 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_flags, lock->l_remote_handle.cookie,
1917 atomic_read(&lock->l_export->exp_refcount) : -99,
1923 EXPORT_SYMBOL(_ldlm_lock_debug);