1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * standalone DLM module
8 * Copyright (C) 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
28 #include <linux/module.h>
30 #include <linux/types.h>
31 #include <linux/slab.h>
32 #include <linux/highmem.h>
33 #include <linux/init.h>
34 #include <linux/sysctl.h>
35 #include <linux/random.h>
36 #include <linux/blkdev.h>
37 #include <linux/socket.h>
38 #include <linux/inet.h>
39 #include <linux/spinlock.h>
40 #include <linux/delay.h>
43 #include "cluster/heartbeat.h"
44 #include "cluster/nodemanager.h"
45 #include "cluster/tcp.h"
48 #include "dlmcommon.h"
49 #include "dlmdomain.h"
52 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
53 #include "cluster/masklog.h"
55 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
56 struct dlm_master_list_entry *mle,
57 struct o2nm_node *node,
59 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
60 struct dlm_master_list_entry *mle,
61 struct o2nm_node *node,
64 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
65 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
66 struct dlm_lock_resource *res,
67 void *nodemap, u32 flags);
68 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
70 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
71 struct dlm_master_list_entry *mle,
78 if (namelen != mle->mnamelen ||
79 memcmp(name, mle->mname, namelen) != 0)
85 static struct kmem_cache *dlm_lockres_cache = NULL;
86 static struct kmem_cache *dlm_lockname_cache = NULL;
87 static struct kmem_cache *dlm_mle_cache = NULL;
89 static void dlm_mle_release(struct kref *kref);
90 static void dlm_init_mle(struct dlm_master_list_entry *mle,
91 enum dlm_mle_type type,
93 struct dlm_lock_resource *res,
95 unsigned int namelen);
96 static void dlm_put_mle(struct dlm_master_list_entry *mle);
97 static void __dlm_put_mle(struct dlm_master_list_entry *mle);
98 static int dlm_find_mle(struct dlm_ctxt *dlm,
99 struct dlm_master_list_entry **mle,
100 char *name, unsigned int namelen);
102 static int dlm_do_master_request(struct dlm_lock_resource *res,
103 struct dlm_master_list_entry *mle, int to);
106 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
107 struct dlm_lock_resource *res,
108 struct dlm_master_list_entry *mle,
110 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
111 struct dlm_lock_resource *res,
112 struct dlm_master_list_entry *mle,
114 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
115 struct dlm_lock_resource *res,
116 struct dlm_master_list_entry *mle,
117 struct dlm_master_list_entry **oldmle,
118 const char *name, unsigned int namelen,
119 u8 new_master, u8 master);
121 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
122 struct dlm_lock_resource *res);
123 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
124 struct dlm_lock_resource *res);
125 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
126 struct dlm_lock_resource *res,
128 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
129 struct dlm_lock_resource *res);
132 int dlm_is_host_down(int errno)
149 case -EINVAL: /* if returned from our tcp code,
150 this means there is no socket */
158 * MASTER LIST FUNCTIONS
163 * regarding master list entries and heartbeat callbacks:
165 * in order to avoid sleeping and allocation that occurs in
166 * heartbeat, master list entries are simply attached to the
167 * dlm's established heartbeat callbacks. the mle is attached
168 * when it is created, and since the dlm->spinlock is held at
169 * that time, any heartbeat event will be properly discovered
170 * by the mle. the mle needs to be detached from the
171 * dlm->mle_hb_events list as soon as heartbeat events are no
172 * longer useful to the mle, and before the mle is freed.
174 * as a general rule, heartbeat events are no longer needed by
175 * the mle once an "answer" regarding the lock master has been
178 static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
179 struct dlm_master_list_entry *mle)
181 assert_spin_locked(&dlm->spinlock);
183 list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
187 static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
188 struct dlm_master_list_entry *mle)
190 if (!list_empty(&mle->hb_events))
191 list_del_init(&mle->hb_events);
195 static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
196 struct dlm_master_list_entry *mle)
198 spin_lock(&dlm->spinlock);
199 __dlm_mle_detach_hb_events(dlm, mle);
200 spin_unlock(&dlm->spinlock);
203 static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
205 struct dlm_ctxt *dlm;
208 assert_spin_locked(&dlm->spinlock);
209 assert_spin_locked(&dlm->master_lock);
211 kref_get(&mle->mle_refs);
214 static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
216 struct dlm_ctxt *dlm;
219 spin_lock(&dlm->spinlock);
220 spin_lock(&dlm->master_lock);
223 spin_unlock(&dlm->master_lock);
224 spin_unlock(&dlm->spinlock);
228 /* remove from list and free */
229 static void __dlm_put_mle(struct dlm_master_list_entry *mle)
231 struct dlm_ctxt *dlm;
234 assert_spin_locked(&dlm->spinlock);
235 assert_spin_locked(&dlm->master_lock);
236 if (!atomic_read(&mle->mle_refs.refcount)) {
237 /* this may or may not crash, but who cares.
239 mlog(ML_ERROR, "bad mle: %p\n", mle);
240 dlm_print_one_mle(mle);
243 kref_put(&mle->mle_refs, dlm_mle_release);
247 /* must not have any spinlocks coming in */
248 static void dlm_put_mle(struct dlm_master_list_entry *mle)
250 struct dlm_ctxt *dlm;
253 spin_lock(&dlm->spinlock);
254 spin_lock(&dlm->master_lock);
256 spin_unlock(&dlm->master_lock);
257 spin_unlock(&dlm->spinlock);
260 static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
262 kref_get(&mle->mle_refs);
265 static void dlm_init_mle(struct dlm_master_list_entry *mle,
266 enum dlm_mle_type type,
267 struct dlm_ctxt *dlm,
268 struct dlm_lock_resource *res,
270 unsigned int namelen)
272 assert_spin_locked(&dlm->spinlock);
276 INIT_HLIST_NODE(&mle->master_hash_node);
277 INIT_LIST_HEAD(&mle->hb_events);
278 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
279 spin_lock_init(&mle->spinlock);
280 init_waitqueue_head(&mle->wq);
281 atomic_set(&mle->woken, 0);
282 kref_init(&mle->mle_refs);
283 memset(mle->response_map, 0, sizeof(mle->response_map));
284 mle->master = O2NM_MAX_NODES;
285 mle->new_master = O2NM_MAX_NODES;
288 BUG_ON(mle->type != DLM_MLE_BLOCK &&
289 mle->type != DLM_MLE_MASTER &&
290 mle->type != DLM_MLE_MIGRATION);
292 if (mle->type == DLM_MLE_MASTER) {
295 memcpy(mle->mname, res->lockname.name, res->lockname.len);
296 mle->mnamelen = res->lockname.len;
297 mle->mnamehash = res->lockname.hash;
301 memcpy(mle->mname, name, namelen);
302 mle->mnamelen = namelen;
303 mle->mnamehash = dlm_lockid_hash(name, namelen);
306 atomic_inc(&dlm->mle_tot_count[mle->type]);
307 atomic_inc(&dlm->mle_cur_count[mle->type]);
309 /* copy off the node_map and register hb callbacks on our copy */
310 memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
311 memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
312 clear_bit(dlm->node_num, mle->vote_map);
313 clear_bit(dlm->node_num, mle->node_map);
315 /* attach the mle to the domain node up/down events */
316 __dlm_mle_attach_hb_events(dlm, mle);
319 void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
321 assert_spin_locked(&dlm->spinlock);
322 assert_spin_locked(&dlm->master_lock);
324 if (!hlist_unhashed(&mle->master_hash_node))
325 hlist_del_init(&mle->master_hash_node);
328 void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
330 struct hlist_head *bucket;
332 assert_spin_locked(&dlm->master_lock);
334 bucket = dlm_master_hash(dlm, mle->mnamehash);
335 hlist_add_head(&mle->master_hash_node, bucket);
338 /* returns 1 if found, 0 if not */
339 static int dlm_find_mle(struct dlm_ctxt *dlm,
340 struct dlm_master_list_entry **mle,
341 char *name, unsigned int namelen)
343 struct dlm_master_list_entry *tmpmle;
344 struct hlist_head *bucket;
345 struct hlist_node *list;
348 assert_spin_locked(&dlm->master_lock);
350 hash = dlm_lockid_hash(name, namelen);
351 bucket = dlm_master_hash(dlm, hash);
352 hlist_for_each(list, bucket) {
353 tmpmle = hlist_entry(list, struct dlm_master_list_entry,
355 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
364 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
366 struct dlm_master_list_entry *mle;
368 assert_spin_locked(&dlm->spinlock);
370 list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
372 dlm_mle_node_up(dlm, mle, NULL, idx);
374 dlm_mle_node_down(dlm, mle, NULL, idx);
378 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
379 struct dlm_master_list_entry *mle,
380 struct o2nm_node *node, int idx)
382 spin_lock(&mle->spinlock);
384 if (!test_bit(idx, mle->node_map))
385 mlog(0, "node %u already removed from nodemap!\n", idx);
387 clear_bit(idx, mle->node_map);
389 spin_unlock(&mle->spinlock);
392 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
393 struct dlm_master_list_entry *mle,
394 struct o2nm_node *node, int idx)
396 spin_lock(&mle->spinlock);
398 if (test_bit(idx, mle->node_map))
399 mlog(0, "node %u already in node map!\n", idx);
401 set_bit(idx, mle->node_map);
403 spin_unlock(&mle->spinlock);
407 int dlm_init_mle_cache(void)
409 dlm_mle_cache = kmem_cache_create("o2dlm_mle",
410 sizeof(struct dlm_master_list_entry),
411 0, SLAB_HWCACHE_ALIGN,
413 if (dlm_mle_cache == NULL)
418 void dlm_destroy_mle_cache(void)
421 kmem_cache_destroy(dlm_mle_cache);
424 static void dlm_mle_release(struct kref *kref)
426 struct dlm_master_list_entry *mle;
427 struct dlm_ctxt *dlm;
429 mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
432 assert_spin_locked(&dlm->spinlock);
433 assert_spin_locked(&dlm->master_lock);
435 mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
438 /* remove from list if not already */
439 __dlm_unlink_mle(dlm, mle);
441 /* detach the mle from the domain node up/down events */
442 __dlm_mle_detach_hb_events(dlm, mle);
444 atomic_dec(&dlm->mle_cur_count[mle->type]);
446 /* NOTE: kfree under spinlock here.
447 * if this is bad, we can move this to a freelist. */
448 kmem_cache_free(dlm_mle_cache, mle);
453 * LOCK RESOURCE FUNCTIONS
456 int dlm_init_master_caches(void)
458 dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
459 sizeof(struct dlm_lock_resource),
460 0, SLAB_HWCACHE_ALIGN, NULL);
461 if (!dlm_lockres_cache)
464 dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
465 DLM_LOCKID_NAME_MAX, 0,
466 SLAB_HWCACHE_ALIGN, NULL);
467 if (!dlm_lockname_cache)
472 dlm_destroy_master_caches();
476 void dlm_destroy_master_caches(void)
478 if (dlm_lockname_cache)
479 kmem_cache_destroy(dlm_lockname_cache);
481 if (dlm_lockres_cache)
482 kmem_cache_destroy(dlm_lockres_cache);
485 static void dlm_lockres_release(struct kref *kref)
487 struct dlm_lock_resource *res;
488 struct dlm_ctxt *dlm;
490 res = container_of(kref, struct dlm_lock_resource, refs);
493 /* This should not happen -- all lockres' have a name
494 * associated with them at init time. */
495 BUG_ON(!res->lockname.name);
497 mlog(0, "destroying lockres %.*s\n", res->lockname.len,
500 spin_lock(&dlm->track_lock);
501 if (!list_empty(&res->tracking))
502 list_del_init(&res->tracking);
504 mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n",
505 res->lockname.len, res->lockname.name);
506 dlm_print_one_lock_resource(res);
508 spin_unlock(&dlm->track_lock);
510 atomic_dec(&dlm->res_cur_count);
512 if (!hlist_unhashed(&res->hash_node) ||
513 !list_empty(&res->granted) ||
514 !list_empty(&res->converting) ||
515 !list_empty(&res->blocked) ||
516 !list_empty(&res->dirty) ||
517 !list_empty(&res->recovering) ||
518 !list_empty(&res->purge)) {
520 "Going to BUG for resource %.*s."
521 " We're on a list! [%c%c%c%c%c%c%c]\n",
522 res->lockname.len, res->lockname.name,
523 !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
524 !list_empty(&res->granted) ? 'G' : ' ',
525 !list_empty(&res->converting) ? 'C' : ' ',
526 !list_empty(&res->blocked) ? 'B' : ' ',
527 !list_empty(&res->dirty) ? 'D' : ' ',
528 !list_empty(&res->recovering) ? 'R' : ' ',
529 !list_empty(&res->purge) ? 'P' : ' ');
531 dlm_print_one_lock_resource(res);
534 /* By the time we're ready to blow this guy away, we shouldn't
535 * be on any lists. */
536 BUG_ON(!hlist_unhashed(&res->hash_node));
537 BUG_ON(!list_empty(&res->granted));
538 BUG_ON(!list_empty(&res->converting));
539 BUG_ON(!list_empty(&res->blocked));
540 BUG_ON(!list_empty(&res->dirty));
541 BUG_ON(!list_empty(&res->recovering));
542 BUG_ON(!list_empty(&res->purge));
544 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
546 kmem_cache_free(dlm_lockres_cache, res);
549 void dlm_lockres_put(struct dlm_lock_resource *res)
551 kref_put(&res->refs, dlm_lockres_release);
554 static void dlm_init_lockres(struct dlm_ctxt *dlm,
555 struct dlm_lock_resource *res,
556 const char *name, unsigned int namelen)
560 /* If we memset here, we lose our reference to the kmalloc'd
561 * res->lockname.name, so be sure to init every field
564 qname = (char *) res->lockname.name;
565 memcpy(qname, name, namelen);
567 res->lockname.len = namelen;
568 res->lockname.hash = dlm_lockid_hash(name, namelen);
570 init_waitqueue_head(&res->wq);
571 spin_lock_init(&res->spinlock);
572 INIT_HLIST_NODE(&res->hash_node);
573 INIT_LIST_HEAD(&res->granted);
574 INIT_LIST_HEAD(&res->converting);
575 INIT_LIST_HEAD(&res->blocked);
576 INIT_LIST_HEAD(&res->dirty);
577 INIT_LIST_HEAD(&res->recovering);
578 INIT_LIST_HEAD(&res->purge);
579 INIT_LIST_HEAD(&res->tracking);
580 atomic_set(&res->asts_reserved, 0);
581 res->migration_pending = 0;
582 res->inflight_locks = 0;
586 kref_init(&res->refs);
588 atomic_inc(&dlm->res_tot_count);
589 atomic_inc(&dlm->res_cur_count);
591 /* just for consistency */
592 spin_lock(&res->spinlock);
593 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
594 spin_unlock(&res->spinlock);
596 res->state = DLM_LOCK_RES_IN_PROGRESS;
600 spin_lock(&dlm->spinlock);
601 list_add_tail(&res->tracking, &dlm->tracking_list);
602 spin_unlock(&dlm->spinlock);
604 memset(res->lvb, 0, DLM_LVB_LEN);
605 memset(res->refmap, 0, sizeof(res->refmap));
608 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
610 unsigned int namelen)
612 struct dlm_lock_resource *res = NULL;
614 res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
618 res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
619 if (!res->lockname.name)
622 dlm_init_lockres(dlm, res, name, namelen);
626 if (res && res->lockname.name)
627 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
630 kmem_cache_free(dlm_lockres_cache, res);
634 void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm,
635 struct dlm_lock_resource *res, int bit)
637 assert_spin_locked(&res->spinlock);
639 mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len,
640 res->lockname.name, bit, __builtin_return_address(0));
642 set_bit(bit, res->refmap);
645 void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm,
646 struct dlm_lock_resource *res, int bit)
648 assert_spin_locked(&res->spinlock);
650 mlog(0, "res %.*s, clr node %u, %ps()\n", res->lockname.len,
651 res->lockname.name, bit, __builtin_return_address(0));
653 clear_bit(bit, res->refmap);
656 static void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
657 struct dlm_lock_resource *res)
659 res->inflight_locks++;
661 mlog(0, "%s: res %.*s, inflight++: now %u, %ps()\n", dlm->name,
662 res->lockname.len, res->lockname.name, res->inflight_locks,
663 __builtin_return_address(0));
666 void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
667 struct dlm_lock_resource *res)
669 assert_spin_locked(&res->spinlock);
670 __dlm_lockres_grab_inflight_ref(dlm, res);
673 void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
674 struct dlm_lock_resource *res)
676 assert_spin_locked(&res->spinlock);
678 BUG_ON(res->inflight_locks == 0);
680 res->inflight_locks--;
682 mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm->name,
683 res->lockname.len, res->lockname.name, res->inflight_locks,
684 __builtin_return_address(0));
690 * lookup a lock resource by name.
691 * may already exist in the hashtable.
692 * lockid is null terminated
694 * if not, allocate enough for the lockres and for
695 * the temporary structure used in doing the mastering.
697 * also, do a lookup in the dlm->master_list to see
698 * if another node has begun mastering the same lock.
699 * if so, there should be a block entry in there
700 * for this name, and we should *not* attempt to master
701 * the lock here. need to wait around for that node
702 * to assert_master (or die).
705 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
710 struct dlm_lock_resource *tmpres=NULL, *res=NULL;
711 struct dlm_master_list_entry *mle = NULL;
712 struct dlm_master_list_entry *alloc_mle = NULL;
715 struct dlm_node_iter iter;
718 int bit, wait_on_recovery = 0;
722 hash = dlm_lockid_hash(lockid, namelen);
724 mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
727 spin_lock(&dlm->spinlock);
728 tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
730 spin_unlock(&dlm->spinlock);
731 spin_lock(&tmpres->spinlock);
732 /* Wait on the thread that is mastering the resource */
733 if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
734 __dlm_wait_on_lockres(tmpres);
735 BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
736 spin_unlock(&tmpres->spinlock);
737 dlm_lockres_put(tmpres);
742 /* Wait on the resource purge to complete before continuing */
743 if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) {
744 BUG_ON(tmpres->owner == dlm->node_num);
745 __dlm_wait_on_lockres_flags(tmpres,
746 DLM_LOCK_RES_DROPPING_REF);
747 spin_unlock(&tmpres->spinlock);
748 dlm_lockres_put(tmpres);
753 /* Grab inflight ref to pin the resource */
754 dlm_lockres_grab_inflight_ref(dlm, tmpres);
756 spin_unlock(&tmpres->spinlock);
758 dlm_lockres_put(res);
764 spin_unlock(&dlm->spinlock);
765 mlog(0, "allocating a new resource\n");
766 /* nothing found and we need to allocate one. */
767 alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
770 res = dlm_new_lockres(dlm, lockid, namelen);
776 mlog(0, "no lockres found, allocated our own: %p\n", res);
778 if (flags & LKM_LOCAL) {
779 /* caller knows it's safe to assume it's not mastered elsewhere
780 * DONE! return right away */
781 spin_lock(&res->spinlock);
782 dlm_change_lockres_owner(dlm, res, dlm->node_num);
783 __dlm_insert_lockres(dlm, res);
784 dlm_lockres_grab_inflight_ref(dlm, res);
785 spin_unlock(&res->spinlock);
786 spin_unlock(&dlm->spinlock);
787 /* lockres still marked IN_PROGRESS */
791 /* check master list to see if another node has started mastering it */
792 spin_lock(&dlm->master_lock);
794 /* if we found a block, wait for lock to be mastered by another node */
795 blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
798 if (mle->type == DLM_MLE_MASTER) {
799 mlog(ML_ERROR, "master entry for nonexistent lock!\n");
802 mig = (mle->type == DLM_MLE_MIGRATION);
803 /* if there is a migration in progress, let the migration
804 * finish before continuing. we can wait for the absence
805 * of the MIGRATION mle: either the migrate finished or
806 * one of the nodes died and the mle was cleaned up.
807 * if there is a BLOCK here, but it already has a master
808 * set, we are too late. the master does not have a ref
809 * for us in the refmap. detach the mle and drop it.
810 * either way, go back to the top and start over. */
811 if (mig || mle->master != O2NM_MAX_NODES) {
812 BUG_ON(mig && mle->master == dlm->node_num);
813 /* we arrived too late. the master does not
814 * have a ref for us. retry. */
815 mlog(0, "%s:%.*s: late on %s\n",
816 dlm->name, namelen, lockid,
817 mig ? "MIGRATION" : "BLOCK");
818 spin_unlock(&dlm->master_lock);
819 spin_unlock(&dlm->spinlock);
821 /* master is known, detach */
823 dlm_mle_detach_hb_events(dlm, mle);
826 /* this is lame, but we can't wait on either
827 * the mle or lockres waitqueue here */
833 /* go ahead and try to master lock on this node */
835 /* make sure this does not get freed below */
837 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
838 set_bit(dlm->node_num, mle->maybe_map);
839 __dlm_insert_mle(dlm, mle);
841 /* still holding the dlm spinlock, check the recovery map
842 * to see if there are any nodes that still need to be
843 * considered. these will not appear in the mle nodemap
844 * but they might own this lockres. wait on them. */
845 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
846 if (bit < O2NM_MAX_NODES) {
847 mlog(0, "%s: res %.*s, At least one node (%d) "
848 "to recover before lock mastery can begin\n",
849 dlm->name, namelen, (char *)lockid, bit);
850 wait_on_recovery = 1;
854 /* at this point there is either a DLM_MLE_BLOCK or a
855 * DLM_MLE_MASTER on the master list, so it's safe to add the
856 * lockres to the hashtable. anyone who finds the lock will
857 * still have to wait on the IN_PROGRESS. */
859 /* finally add the lockres to its hash bucket */
860 __dlm_insert_lockres(dlm, res);
862 /* since this lockres is new it doesn't not require the spinlock */
863 __dlm_lockres_grab_inflight_ref(dlm, res);
865 /* get an extra ref on the mle in case this is a BLOCK
866 * if so, the creator of the BLOCK may try to put the last
867 * ref at this time in the assert master handler, so we
868 * need an extra one to keep from a bad ptr deref. */
869 dlm_get_mle_inuse(mle);
870 spin_unlock(&dlm->master_lock);
871 spin_unlock(&dlm->spinlock);
874 while (wait_on_recovery) {
875 /* any cluster changes that occurred after dropping the
876 * dlm spinlock would be detectable be a change on the mle,
877 * so we only need to clear out the recovery map once. */
878 if (dlm_is_recovery_lock(lockid, namelen)) {
879 mlog(0, "%s: Recovery map is not empty, but must "
880 "master $RECOVERY lock now\n", dlm->name);
881 if (!dlm_pre_master_reco_lockres(dlm, res))
882 wait_on_recovery = 0;
884 mlog(0, "%s: waiting 500ms for heartbeat state "
885 "change\n", dlm->name);
891 dlm_kick_recovery_thread(dlm);
893 dlm_wait_for_recovery(dlm);
895 spin_lock(&dlm->spinlock);
896 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
897 if (bit < O2NM_MAX_NODES) {
898 mlog(0, "%s: res %.*s, At least one node (%d) "
899 "to recover before lock mastery can begin\n",
900 dlm->name, namelen, (char *)lockid, bit);
901 wait_on_recovery = 1;
903 wait_on_recovery = 0;
904 spin_unlock(&dlm->spinlock);
906 if (wait_on_recovery)
907 dlm_wait_for_node_recovery(dlm, bit, 10000);
910 /* must wait for lock to be mastered elsewhere */
915 dlm_node_iter_init(mle->vote_map, &iter);
916 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
917 ret = dlm_do_master_request(res, mle, nodenum);
920 if (mle->master != O2NM_MAX_NODES) {
921 /* found a master ! */
922 if (mle->master <= nodenum)
924 /* if our master request has not reached the master
925 * yet, keep going until it does. this is how the
926 * master will know that asserts are needed back to
927 * the lower nodes. */
928 mlog(0, "%s: res %.*s, Requests only up to %u but "
929 "master is %u, keep going\n", dlm->name, namelen,
930 lockid, nodenum, mle->master);
935 /* keep going until the response map includes all nodes */
936 ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
938 wait_on_recovery = 1;
939 mlog(0, "%s: res %.*s, Node map changed, redo the master "
940 "request now, blocked=%d\n", dlm->name, res->lockname.len,
941 res->lockname.name, blocked);
943 mlog(ML_ERROR, "%s: res %.*s, Spinning on "
944 "dlm_wait_for_lock_mastery, blocked = %d\n",
945 dlm->name, res->lockname.len,
946 res->lockname.name, blocked);
947 dlm_print_one_lock_resource(res);
948 dlm_print_one_mle(mle);
954 mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len,
955 res->lockname.name, res->owner);
956 /* make sure we never continue without this */
957 BUG_ON(res->owner == O2NM_MAX_NODES);
959 /* master is known, detach if not already detached */
960 dlm_mle_detach_hb_events(dlm, mle);
962 /* put the extra ref */
963 dlm_put_mle_inuse(mle);
966 spin_lock(&res->spinlock);
967 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
968 spin_unlock(&res->spinlock);
972 /* need to free the unused mle */
974 kmem_cache_free(dlm_mle_cache, alloc_mle);
980 #define DLM_MASTERY_TIMEOUT_MS 5000
982 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
983 struct dlm_lock_resource *res,
984 struct dlm_master_list_entry *mle,
989 int map_changed, voting_done;
996 /* check if another node has already become the owner */
997 spin_lock(&res->spinlock);
998 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
999 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1000 res->lockname.len, res->lockname.name, res->owner);
1001 spin_unlock(&res->spinlock);
1002 /* this will cause the master to re-assert across
1003 * the whole cluster, freeing up mles */
1004 if (res->owner != dlm->node_num) {
1005 ret = dlm_do_master_request(res, mle, res->owner);
1007 /* give recovery a chance to run */
1008 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1016 spin_unlock(&res->spinlock);
1018 spin_lock(&mle->spinlock);
1020 map_changed = (memcmp(mle->vote_map, mle->node_map,
1021 sizeof(mle->vote_map)) != 0);
1022 voting_done = (memcmp(mle->vote_map, mle->response_map,
1023 sizeof(mle->vote_map)) == 0);
1025 /* restart if we hit any errors */
1028 mlog(0, "%s: %.*s: node map changed, restarting\n",
1029 dlm->name, res->lockname.len, res->lockname.name);
1030 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1031 b = (mle->type == DLM_MLE_BLOCK);
1032 if ((*blocked && !b) || (!*blocked && b)) {
1033 mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1034 dlm->name, res->lockname.len, res->lockname.name,
1038 spin_unlock(&mle->spinlock);
1043 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1044 "rechecking now\n", dlm->name, res->lockname.len,
1045 res->lockname.name);
1049 mlog(0, "map not changed and voting not done "
1050 "for %s:%.*s\n", dlm->name, res->lockname.len,
1051 res->lockname.name);
1055 if (m != O2NM_MAX_NODES) {
1056 /* another node has done an assert!
1061 /* have all nodes responded? */
1062 if (voting_done && !*blocked) {
1063 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1064 if (dlm->node_num <= bit) {
1065 /* my node number is lowest.
1066 * now tell other nodes that I am
1067 * mastering this. */
1068 mle->master = dlm->node_num;
1069 /* ref was grabbed in get_lock_resource
1070 * will be dropped in dlmlock_master */
1074 /* if voting is done, but we have not received
1075 * an assert master yet, we must sleep */
1079 spin_unlock(&mle->spinlock);
1081 /* sleep if we haven't finished voting yet */
1083 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1086 if (atomic_read(&mle->mle_refs.refcount) < 2)
1087 mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1088 atomic_read(&mle->mle_refs.refcount),
1089 res->lockname.len, res->lockname.name);
1091 atomic_set(&mle->woken, 0);
1092 (void)wait_event_timeout(mle->wq,
1093 (atomic_read(&mle->woken) == 1),
1095 if (res->owner == O2NM_MAX_NODES) {
1096 mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1097 res->lockname.len, res->lockname.name);
1100 mlog(0, "done waiting, master is %u\n", res->owner);
1108 mlog(0, "about to master %.*s here, this=%u\n",
1109 res->lockname.len, res->lockname.name, m);
1110 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1112 /* This is a failure in the network path,
1113 * not in the response to the assert_master
1114 * (any nonzero response is a BUG on this node).
1115 * Most likely a socket just got disconnected
1116 * due to node death. */
1119 /* no longer need to restart lock mastery.
1120 * all living nodes have been contacted. */
1124 /* set the lockres owner */
1125 spin_lock(&res->spinlock);
1126 /* mastery reference obtained either during
1127 * assert_master_handler or in get_lock_resource */
1128 dlm_change_lockres_owner(dlm, res, m);
1129 spin_unlock(&res->spinlock);
1135 struct dlm_bitmap_diff_iter
1138 unsigned long *orig_bm;
1139 unsigned long *cur_bm;
1140 unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1143 enum dlm_node_state_change
1150 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1151 unsigned long *orig_bm,
1152 unsigned long *cur_bm)
1154 unsigned long p1, p2;
1158 iter->orig_bm = orig_bm;
1159 iter->cur_bm = cur_bm;
1161 for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1162 p1 = *(iter->orig_bm + i);
1163 p2 = *(iter->cur_bm + i);
1164 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1168 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1169 enum dlm_node_state_change *state)
1173 if (iter->curnode >= O2NM_MAX_NODES)
1176 bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1178 if (bit >= O2NM_MAX_NODES) {
1179 iter->curnode = O2NM_MAX_NODES;
1183 /* if it was there in the original then this node died */
1184 if (test_bit(bit, iter->orig_bm))
1189 iter->curnode = bit;
1194 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1195 struct dlm_lock_resource *res,
1196 struct dlm_master_list_entry *mle,
1199 struct dlm_bitmap_diff_iter bdi;
1200 enum dlm_node_state_change sc;
1204 mlog(0, "something happened such that the "
1205 "master process may need to be restarted!\n");
1207 assert_spin_locked(&mle->spinlock);
1209 dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1210 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1212 if (sc == NODE_UP) {
1213 /* a node came up. clear any old vote from
1214 * the response map and set it in the vote map
1215 * then restart the mastery. */
1216 mlog(ML_NOTICE, "node %d up while restarting\n", node);
1218 /* redo the master request, but only for the new node */
1219 mlog(0, "sending request to new node\n");
1220 clear_bit(node, mle->response_map);
1221 set_bit(node, mle->vote_map);
1223 mlog(ML_ERROR, "node down! %d\n", node);
1225 int lowest = find_next_bit(mle->maybe_map,
1228 /* act like it was never there */
1229 clear_bit(node, mle->maybe_map);
1231 if (node == lowest) {
1232 mlog(0, "expected master %u died"
1233 " while this node was blocked "
1234 "waiting on it!\n", node);
1235 lowest = find_next_bit(mle->maybe_map,
1238 if (lowest < O2NM_MAX_NODES) {
1239 mlog(0, "%s:%.*s:still "
1240 "blocked. waiting on %u "
1246 /* mle is an MLE_BLOCK, but
1247 * there is now nothing left to
1248 * block on. we need to return
1249 * all the way back out and try
1250 * again with an MLE_MASTER.
1251 * dlm_do_local_recovery_cleanup
1252 * has already run, so the mle
1254 mlog(0, "%s:%.*s: no "
1255 "longer blocking. try to "
1256 "master this here\n",
1259 res->lockname.name);
1260 mle->type = DLM_MLE_MASTER;
1266 /* now blank out everything, as if we had never
1267 * contacted anyone */
1268 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1269 memset(mle->response_map, 0, sizeof(mle->response_map));
1270 /* reset the vote_map to the current node_map */
1271 memcpy(mle->vote_map, mle->node_map,
1272 sizeof(mle->node_map));
1273 /* put myself into the maybe map */
1274 if (mle->type != DLM_MLE_BLOCK)
1275 set_bit(dlm->node_num, mle->maybe_map);
1278 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1285 * DLM_MASTER_REQUEST_MSG
1287 * returns: 0 on success,
1288 * -errno on a network error
1290 * on error, the caller should assume the target node is "dead"
1294 static int dlm_do_master_request(struct dlm_lock_resource *res,
1295 struct dlm_master_list_entry *mle, int to)
1297 struct dlm_ctxt *dlm = mle->dlm;
1298 struct dlm_master_request request;
1299 int ret, response=0, resend;
1301 memset(&request, 0, sizeof(request));
1302 request.node_idx = dlm->node_num;
1304 BUG_ON(mle->type == DLM_MLE_MIGRATION);
1306 request.namelen = (u8)mle->mnamelen;
1307 memcpy(request.name, mle->mname, request.namelen);
1310 ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1311 sizeof(request), to, &response);
1313 if (ret == -ESRCH) {
1314 /* should never happen */
1315 mlog(ML_ERROR, "TCP stack not ready!\n");
1317 } else if (ret == -EINVAL) {
1318 mlog(ML_ERROR, "bad args passed to o2net!\n");
1320 } else if (ret == -ENOMEM) {
1321 mlog(ML_ERROR, "out of memory while trying to send "
1322 "network message! retrying\n");
1323 /* this is totally crude */
1326 } else if (!dlm_is_host_down(ret)) {
1327 /* not a network error. bad. */
1329 mlog(ML_ERROR, "unhandled error!");
1332 /* all other errors should be network errors,
1333 * and likely indicate node death */
1334 mlog(ML_ERROR, "link to %d went down!\n", to);
1340 spin_lock(&mle->spinlock);
1342 case DLM_MASTER_RESP_YES:
1343 set_bit(to, mle->response_map);
1344 mlog(0, "node %u is the master, response=YES\n", to);
1345 mlog(0, "%s:%.*s: master node %u now knows I have a "
1346 "reference\n", dlm->name, res->lockname.len,
1347 res->lockname.name, to);
1350 case DLM_MASTER_RESP_NO:
1351 mlog(0, "node %u not master, response=NO\n", to);
1352 set_bit(to, mle->response_map);
1354 case DLM_MASTER_RESP_MAYBE:
1355 mlog(0, "node %u not master, response=MAYBE\n", to);
1356 set_bit(to, mle->response_map);
1357 set_bit(to, mle->maybe_map);
1359 case DLM_MASTER_RESP_ERROR:
1360 mlog(0, "node %u hit an error, resending\n", to);
1365 mlog(ML_ERROR, "bad response! %u\n", response);
1368 spin_unlock(&mle->spinlock);
1370 /* this is also totally crude */
1380 * locks that can be taken here:
1386 * if possible, TRIM THIS DOWN!!!
1388 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1391 u8 response = DLM_MASTER_RESP_MAYBE;
1392 struct dlm_ctxt *dlm = data;
1393 struct dlm_lock_resource *res = NULL;
1394 struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1395 struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1397 unsigned int namelen, hash;
1400 int dispatch_assert = 0;
1403 return DLM_MASTER_RESP_NO;
1405 if (!dlm_domain_fully_joined(dlm)) {
1406 response = DLM_MASTER_RESP_NO;
1410 name = request->name;
1411 namelen = request->namelen;
1412 hash = dlm_lockid_hash(name, namelen);
1414 if (namelen > DLM_LOCKID_NAME_MAX) {
1415 response = DLM_IVBUFLEN;
1420 spin_lock(&dlm->spinlock);
1421 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1423 spin_unlock(&dlm->spinlock);
1425 /* take care of the easy cases up front */
1426 spin_lock(&res->spinlock);
1427 if (res->state & (DLM_LOCK_RES_RECOVERING|
1428 DLM_LOCK_RES_MIGRATING)) {
1429 spin_unlock(&res->spinlock);
1430 mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1431 "being recovered/migrated\n");
1432 response = DLM_MASTER_RESP_ERROR;
1434 kmem_cache_free(dlm_mle_cache, mle);
1438 if (res->owner == dlm->node_num) {
1439 dlm_lockres_set_refmap_bit(dlm, res, request->node_idx);
1440 spin_unlock(&res->spinlock);
1441 response = DLM_MASTER_RESP_YES;
1443 kmem_cache_free(dlm_mle_cache, mle);
1445 /* this node is the owner.
1446 * there is some extra work that needs to
1447 * happen now. the requesting node has
1448 * caused all nodes up to this one to
1449 * create mles. this node now needs to
1450 * go back and clean those up. */
1451 dispatch_assert = 1;
1453 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1454 spin_unlock(&res->spinlock);
1455 // mlog(0, "node %u is the master\n", res->owner);
1456 response = DLM_MASTER_RESP_NO;
1458 kmem_cache_free(dlm_mle_cache, mle);
1462 /* ok, there is no owner. either this node is
1463 * being blocked, or it is actively trying to
1464 * master this lock. */
1465 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1466 mlog(ML_ERROR, "lock with no owner should be "
1471 // mlog(0, "lockres is in progress...\n");
1472 spin_lock(&dlm->master_lock);
1473 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1475 mlog(ML_ERROR, "no mle found for this lock!\n");
1479 spin_lock(&tmpmle->spinlock);
1480 if (tmpmle->type == DLM_MLE_BLOCK) {
1481 // mlog(0, "this node is waiting for "
1482 // "lockres to be mastered\n");
1483 response = DLM_MASTER_RESP_NO;
1484 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1485 mlog(0, "node %u is master, but trying to migrate to "
1486 "node %u.\n", tmpmle->master, tmpmle->new_master);
1487 if (tmpmle->master == dlm->node_num) {
1488 mlog(ML_ERROR, "no owner on lockres, but this "
1489 "node is trying to migrate it to %u?!\n",
1490 tmpmle->new_master);
1493 /* the real master can respond on its own */
1494 response = DLM_MASTER_RESP_NO;
1496 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1498 if (tmpmle->master == dlm->node_num) {
1499 response = DLM_MASTER_RESP_YES;
1500 /* this node will be the owner.
1501 * go back and clean the mles on any
1503 dispatch_assert = 1;
1504 dlm_lockres_set_refmap_bit(dlm, res,
1507 response = DLM_MASTER_RESP_NO;
1509 // mlog(0, "this node is attempting to "
1510 // "master lockres\n");
1511 response = DLM_MASTER_RESP_MAYBE;
1514 set_bit(request->node_idx, tmpmle->maybe_map);
1515 spin_unlock(&tmpmle->spinlock);
1517 spin_unlock(&dlm->master_lock);
1518 spin_unlock(&res->spinlock);
1520 /* keep the mle attached to heartbeat events */
1521 dlm_put_mle(tmpmle);
1523 kmem_cache_free(dlm_mle_cache, mle);
1528 * lockres doesn't exist on this node
1529 * if there is an MLE_BLOCK, return NO
1530 * if there is an MLE_MASTER, return MAYBE
1531 * otherwise, add an MLE_BLOCK, return NO
1533 spin_lock(&dlm->master_lock);
1534 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1536 /* this lockid has never been seen on this node yet */
1537 // mlog(0, "no mle found\n");
1539 spin_unlock(&dlm->master_lock);
1540 spin_unlock(&dlm->spinlock);
1542 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1544 response = DLM_MASTER_RESP_ERROR;
1545 mlog_errno(-ENOMEM);
1551 // mlog(0, "this is second time thru, already allocated, "
1552 // "add the block.\n");
1553 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1554 set_bit(request->node_idx, mle->maybe_map);
1555 __dlm_insert_mle(dlm, mle);
1556 response = DLM_MASTER_RESP_NO;
1558 // mlog(0, "mle was found\n");
1560 spin_lock(&tmpmle->spinlock);
1561 if (tmpmle->master == dlm->node_num) {
1562 mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1565 if (tmpmle->type == DLM_MLE_BLOCK)
1566 response = DLM_MASTER_RESP_NO;
1567 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1568 mlog(0, "migration mle was found (%u->%u)\n",
1569 tmpmle->master, tmpmle->new_master);
1570 /* real master can respond on its own */
1571 response = DLM_MASTER_RESP_NO;
1573 response = DLM_MASTER_RESP_MAYBE;
1575 set_bit(request->node_idx, tmpmle->maybe_map);
1576 spin_unlock(&tmpmle->spinlock);
1578 spin_unlock(&dlm->master_lock);
1579 spin_unlock(&dlm->spinlock);
1582 /* keep the mle attached to heartbeat events */
1583 dlm_put_mle(tmpmle);
1587 * __dlm_lookup_lockres() grabbed a reference to this lockres.
1588 * The reference is released by dlm_assert_master_worker() under
1589 * the call to dlm_dispatch_assert_master(). If
1590 * dlm_assert_master_worker() isn't called, we drop it here.
1592 if (dispatch_assert) {
1593 if (response != DLM_MASTER_RESP_YES)
1594 mlog(ML_ERROR, "invalid response %d\n", response);
1596 mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1599 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1600 dlm->node_num, res->lockname.len, res->lockname.name);
1601 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1602 DLM_ASSERT_MASTER_MLE_CLEANUP);
1604 mlog(ML_ERROR, "failed to dispatch assert master work\n");
1605 response = DLM_MASTER_RESP_ERROR;
1606 dlm_lockres_put(res);
1610 dlm_lockres_put(res);
1618 * DLM_ASSERT_MASTER_MSG
1623 * NOTE: this can be used for debugging
1624 * can periodically run all locks owned by this node
1625 * and re-assert across the cluster...
1627 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1628 struct dlm_lock_resource *res,
1629 void *nodemap, u32 flags)
1631 struct dlm_assert_master assert;
1633 struct dlm_node_iter iter;
1636 const char *lockname = res->lockname.name;
1637 unsigned int namelen = res->lockname.len;
1639 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1641 spin_lock(&res->spinlock);
1642 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1643 spin_unlock(&res->spinlock);
1648 /* note that if this nodemap is empty, it returns 0 */
1649 dlm_node_iter_init(nodemap, &iter);
1650 while ((to = dlm_node_iter_next(&iter)) >= 0) {
1652 struct dlm_master_list_entry *mle = NULL;
1654 mlog(0, "sending assert master to %d (%.*s)\n", to,
1656 memset(&assert, 0, sizeof(assert));
1657 assert.node_idx = dlm->node_num;
1658 assert.namelen = namelen;
1659 memcpy(assert.name, lockname, namelen);
1660 assert.flags = cpu_to_be32(flags);
1662 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1663 &assert, sizeof(assert), to, &r);
1665 mlog(ML_ERROR, "Error %d when sending message %u (key "
1666 "0x%x) to node %u\n", tmpret,
1667 DLM_ASSERT_MASTER_MSG, dlm->key, to);
1668 if (!dlm_is_host_down(tmpret)) {
1669 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1672 /* a node died. finish out the rest of the nodes. */
1673 mlog(0, "link to %d went down!\n", to);
1674 /* any nonzero status return will do */
1678 /* ok, something horribly messed. kill thyself. */
1679 mlog(ML_ERROR,"during assert master of %.*s to %u, "
1680 "got %d.\n", namelen, lockname, to, r);
1681 spin_lock(&dlm->spinlock);
1682 spin_lock(&dlm->master_lock);
1683 if (dlm_find_mle(dlm, &mle, (char *)lockname,
1685 dlm_print_one_mle(mle);
1688 spin_unlock(&dlm->master_lock);
1689 spin_unlock(&dlm->spinlock);
1693 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1694 !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1695 mlog(ML_ERROR, "%.*s: very strange, "
1696 "master MLE but no lockres on %u\n",
1697 namelen, lockname, to);
1700 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1701 mlog(0, "%.*s: node %u create mles on other "
1702 "nodes and requests a re-assert\n",
1703 namelen, lockname, to);
1706 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1707 mlog(0, "%.*s: node %u has a reference to this "
1708 "lockres, set the bit in the refmap\n",
1709 namelen, lockname, to);
1710 spin_lock(&res->spinlock);
1711 dlm_lockres_set_refmap_bit(dlm, res, to);
1712 spin_unlock(&res->spinlock);
1719 spin_lock(&res->spinlock);
1720 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1721 spin_unlock(&res->spinlock);
1728 * locks that can be taken here:
1734 * if possible, TRIM THIS DOWN!!!
1736 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1739 struct dlm_ctxt *dlm = data;
1740 struct dlm_master_list_entry *mle = NULL;
1741 struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1742 struct dlm_lock_resource *res = NULL;
1744 unsigned int namelen, hash;
1746 int master_request = 0, have_lockres_ref = 0;
1752 name = assert->name;
1753 namelen = assert->namelen;
1754 hash = dlm_lockid_hash(name, namelen);
1755 flags = be32_to_cpu(assert->flags);
1757 if (namelen > DLM_LOCKID_NAME_MAX) {
1758 mlog(ML_ERROR, "Invalid name length!");
1762 spin_lock(&dlm->spinlock);
1765 mlog(0, "assert_master with flags: %u\n", flags);
1768 spin_lock(&dlm->master_lock);
1769 if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1770 /* not an error, could be master just re-asserting */
1771 mlog(0, "just got an assert_master from %u, but no "
1772 "MLE for it! (%.*s)\n", assert->node_idx,
1775 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1776 if (bit >= O2NM_MAX_NODES) {
1777 /* not necessarily an error, though less likely.
1778 * could be master just re-asserting. */
1779 mlog(0, "no bits set in the maybe_map, but %u "
1780 "is asserting! (%.*s)\n", assert->node_idx,
1782 } else if (bit != assert->node_idx) {
1783 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1784 mlog(0, "master %u was found, %u should "
1785 "back off\n", assert->node_idx, bit);
1787 /* with the fix for bug 569, a higher node
1788 * number winning the mastery will respond
1789 * YES to mastery requests, but this node
1790 * had no way of knowing. let it pass. */
1791 mlog(0, "%u is the lowest node, "
1792 "%u is asserting. (%.*s) %u must "
1793 "have begun after %u won.\n", bit,
1794 assert->node_idx, namelen, name, bit,
1798 if (mle->type == DLM_MLE_MIGRATION) {
1799 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1800 mlog(0, "%s:%.*s: got cleanup assert"
1801 " from %u for migration\n",
1802 dlm->name, namelen, name,
1804 } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1805 mlog(0, "%s:%.*s: got unrelated assert"
1806 " from %u for migration, ignoring\n",
1807 dlm->name, namelen, name,
1810 spin_unlock(&dlm->master_lock);
1811 spin_unlock(&dlm->spinlock);
1816 spin_unlock(&dlm->master_lock);
1818 /* ok everything checks out with the MLE
1819 * now check to see if there is a lockres */
1820 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1822 spin_lock(&res->spinlock);
1823 if (res->state & DLM_LOCK_RES_RECOVERING) {
1824 mlog(ML_ERROR, "%u asserting but %.*s is "
1825 "RECOVERING!\n", assert->node_idx, namelen, name);
1829 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1830 res->owner != assert->node_idx) {
1831 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1832 "but current owner is %u! (%.*s)\n",
1833 assert->node_idx, res->owner, namelen,
1835 __dlm_print_one_lock_resource(res);
1838 } else if (mle->type != DLM_MLE_MIGRATION) {
1839 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1840 /* owner is just re-asserting */
1841 if (res->owner == assert->node_idx) {
1842 mlog(0, "owner %u re-asserting on "
1843 "lock %.*s\n", assert->node_idx,
1847 mlog(ML_ERROR, "got assert_master from "
1848 "node %u, but %u is the owner! "
1849 "(%.*s)\n", assert->node_idx,
1850 res->owner, namelen, name);
1853 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1854 mlog(ML_ERROR, "got assert from %u, but lock "
1855 "with no owner should be "
1856 "in-progress! (%.*s)\n",
1861 } else /* mle->type == DLM_MLE_MIGRATION */ {
1862 /* should only be getting an assert from new master */
1863 if (assert->node_idx != mle->new_master) {
1864 mlog(ML_ERROR, "got assert from %u, but "
1865 "new master is %u, and old master "
1867 assert->node_idx, mle->new_master,
1868 mle->master, namelen, name);
1874 spin_unlock(&res->spinlock);
1877 // mlog(0, "woo! got an assert_master from node %u!\n",
1878 // assert->node_idx);
1884 spin_lock(&mle->spinlock);
1885 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1888 /* MASTER mle: if any bits set in the response map
1889 * then the calling node needs to re-assert to clear
1890 * up nodes that this node contacted */
1891 while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1892 nn+1)) < O2NM_MAX_NODES) {
1893 if (nn != dlm->node_num && nn != assert->node_idx)
1897 mle->master = assert->node_idx;
1898 atomic_set(&mle->woken, 1);
1900 spin_unlock(&mle->spinlock);
1904 spin_lock(&res->spinlock);
1905 if (mle->type == DLM_MLE_MIGRATION) {
1906 mlog(0, "finishing off migration of lockres %.*s, "
1908 res->lockname.len, res->lockname.name,
1909 dlm->node_num, mle->new_master);
1910 res->state &= ~DLM_LOCK_RES_MIGRATING;
1912 dlm_change_lockres_owner(dlm, res, mle->new_master);
1913 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1915 dlm_change_lockres_owner(dlm, res, mle->master);
1917 spin_unlock(&res->spinlock);
1918 have_lockres_ref = 1;
1923 /* master is known, detach if not already detached.
1924 * ensures that only one assert_master call will happen
1926 spin_lock(&dlm->master_lock);
1928 rr = atomic_read(&mle->mle_refs.refcount);
1929 if (mle->inuse > 0) {
1930 if (extra_ref && rr < 3)
1932 else if (!extra_ref && rr < 2)
1935 if (extra_ref && rr < 2)
1937 else if (!extra_ref && rr < 1)
1941 mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
1942 "that will mess up this node, refs=%d, extra=%d, "
1943 "inuse=%d\n", dlm->name, namelen, name,
1944 assert->node_idx, rr, extra_ref, mle->inuse);
1945 dlm_print_one_mle(mle);
1947 __dlm_unlink_mle(dlm, mle);
1948 __dlm_mle_detach_hb_events(dlm, mle);
1951 /* the assert master message now balances the extra
1952 * ref given by the master / migration request message.
1953 * if this is the last put, it will be removed
1957 spin_unlock(&dlm->master_lock);
1959 if (res->owner != assert->node_idx) {
1960 mlog(0, "assert_master from %u, but current "
1961 "owner is %u (%.*s), no mle\n", assert->node_idx,
1962 res->owner, namelen, name);
1965 spin_unlock(&dlm->spinlock);
1970 spin_lock(&res->spinlock);
1971 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1972 spin_unlock(&res->spinlock);
1973 *ret_data = (void *)res;
1976 if (master_request) {
1977 mlog(0, "need to tell master to reassert\n");
1978 /* positive. negative would shoot down the node. */
1979 ret |= DLM_ASSERT_RESPONSE_REASSERT;
1980 if (!have_lockres_ref) {
1981 mlog(ML_ERROR, "strange, got assert from %u, MASTER "
1982 "mle present here for %s:%.*s, but no lockres!\n",
1983 assert->node_idx, dlm->name, namelen, name);
1986 if (have_lockres_ref) {
1987 /* let the master know we have a reference to the lockres */
1988 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
1989 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
1990 dlm->name, namelen, name, assert->node_idx);
1995 /* kill the caller! */
1996 mlog(ML_ERROR, "Bad message received from another node. Dumping state "
1997 "and killing the other node now! This node is OK and can continue.\n");
1998 __dlm_print_one_lock_resource(res);
1999 spin_unlock(&res->spinlock);
2000 spin_unlock(&dlm->spinlock);
2001 *ret_data = (void *)res;
2006 void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2008 struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2011 spin_lock(&res->spinlock);
2012 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2013 spin_unlock(&res->spinlock);
2015 dlm_lockres_put(res);
2020 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2021 struct dlm_lock_resource *res,
2022 int ignore_higher, u8 request_from, u32 flags)
2024 struct dlm_work_item *item;
2025 item = kzalloc(sizeof(*item), GFP_ATOMIC);
2030 /* queue up work for dlm_assert_master_worker */
2031 dlm_grab(dlm); /* get an extra ref for the work item */
2032 dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2033 item->u.am.lockres = res; /* already have a ref */
2034 /* can optionally ignore node numbers higher than this node */
2035 item->u.am.ignore_higher = ignore_higher;
2036 item->u.am.request_from = request_from;
2037 item->u.am.flags = flags;
2040 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2041 res->lockname.name);
2043 spin_lock(&dlm->work_lock);
2044 list_add_tail(&item->list, &dlm->work_list);
2045 spin_unlock(&dlm->work_lock);
2047 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2051 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2053 struct dlm_ctxt *dlm = data;
2055 struct dlm_lock_resource *res;
2056 unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2063 res = item->u.am.lockres;
2064 ignore_higher = item->u.am.ignore_higher;
2065 request_from = item->u.am.request_from;
2066 flags = item->u.am.flags;
2068 spin_lock(&dlm->spinlock);
2069 memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2070 spin_unlock(&dlm->spinlock);
2072 clear_bit(dlm->node_num, nodemap);
2073 if (ignore_higher) {
2074 /* if is this just to clear up mles for nodes below
2075 * this node, do not send the message to the original
2076 * caller or any node number higher than this */
2077 clear_bit(request_from, nodemap);
2078 bit = dlm->node_num;
2080 bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2082 if (bit >= O2NM_MAX_NODES)
2084 clear_bit(bit, nodemap);
2089 * If we're migrating this lock to someone else, we are no
2090 * longer allowed to assert out own mastery. OTOH, we need to
2091 * prevent migration from starting while we're still asserting
2092 * our dominance. The reserved ast delays migration.
2094 spin_lock(&res->spinlock);
2095 if (res->state & DLM_LOCK_RES_MIGRATING) {
2096 mlog(0, "Someone asked us to assert mastery, but we're "
2097 "in the middle of migration. Skipping assert, "
2098 "the new master will handle that.\n");
2099 spin_unlock(&res->spinlock);
2102 __dlm_lockres_reserve_ast(res);
2103 spin_unlock(&res->spinlock);
2105 /* this call now finishes out the nodemap
2106 * even if one or more nodes die */
2107 mlog(0, "worker about to master %.*s here, this=%u\n",
2108 res->lockname.len, res->lockname.name, dlm->node_num);
2109 ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2111 /* no need to restart, we are done */
2112 if (!dlm_is_host_down(ret))
2116 /* Ok, we've asserted ourselves. Let's let migration start. */
2117 dlm_lockres_release_ast(dlm, res);
2120 dlm_lockres_put(res);
2122 mlog(0, "finished with dlm_assert_master_worker\n");
2125 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2126 * We cannot wait for node recovery to complete to begin mastering this
2127 * lockres because this lockres is used to kick off recovery! ;-)
2128 * So, do a pre-check on all living nodes to see if any of those nodes
2129 * think that $RECOVERY is currently mastered by a dead node. If so,
2130 * we wait a short time to allow that node to get notified by its own
2131 * heartbeat stack, then check again. All $RECOVERY lock resources
2132 * mastered by dead nodes are purged when the hearbeat callback is
2133 * fired, so we can know for sure that it is safe to continue once
2134 * the node returns a live node or no node. */
2135 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2136 struct dlm_lock_resource *res)
2138 struct dlm_node_iter iter;
2141 u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2143 spin_lock(&dlm->spinlock);
2144 dlm_node_iter_init(dlm->domain_map, &iter);
2145 spin_unlock(&dlm->spinlock);
2147 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2148 /* do not send to self */
2149 if (nodenum == dlm->node_num)
2151 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2154 if (!dlm_is_host_down(ret))
2156 /* host is down, so answer for that node would be
2157 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
2161 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2162 /* check to see if this master is in the recovery map */
2163 spin_lock(&dlm->spinlock);
2164 if (test_bit(master, dlm->recovery_map)) {
2165 mlog(ML_NOTICE, "%s: node %u has not seen "
2166 "node %u go down yet, and thinks the "
2167 "dead node is mastering the recovery "
2168 "lock. must wait.\n", dlm->name,
2172 spin_unlock(&dlm->spinlock);
2173 mlog(0, "%s: reco lock master is %u\n", dlm->name,
2182 * DLM_DEREF_LOCKRES_MSG
2185 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2187 struct dlm_deref_lockres deref;
2189 const char *lockname;
2190 unsigned int namelen;
2192 lockname = res->lockname.name;
2193 namelen = res->lockname.len;
2194 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2196 memset(&deref, 0, sizeof(deref));
2197 deref.node_idx = dlm->node_num;
2198 deref.namelen = namelen;
2199 memcpy(deref.name, lockname, namelen);
2201 ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2202 &deref, sizeof(deref), res->owner, &r);
2204 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n",
2205 dlm->name, namelen, lockname, ret, res->owner);
2207 /* BAD. other node says I did not have a ref. */
2208 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2209 dlm->name, namelen, lockname, res->owner, r);
2210 dlm_print_one_lock_resource(res);
2216 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2219 struct dlm_ctxt *dlm = data;
2220 struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2221 struct dlm_lock_resource *res = NULL;
2223 unsigned int namelen;
2227 struct dlm_work_item *item;
2235 namelen = deref->namelen;
2236 node = deref->node_idx;
2238 if (namelen > DLM_LOCKID_NAME_MAX) {
2239 mlog(ML_ERROR, "Invalid name length!");
2242 if (deref->node_idx >= O2NM_MAX_NODES) {
2243 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2247 hash = dlm_lockid_hash(name, namelen);
2249 spin_lock(&dlm->spinlock);
2250 res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2252 spin_unlock(&dlm->spinlock);
2253 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2254 dlm->name, namelen, name);
2257 spin_unlock(&dlm->spinlock);
2259 spin_lock(&res->spinlock);
2260 if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2263 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2264 if (test_bit(node, res->refmap)) {
2265 dlm_lockres_clear_refmap_bit(dlm, res, node);
2269 spin_unlock(&res->spinlock);
2273 dlm_lockres_calc_usage(dlm, res);
2275 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2276 "but it is already dropped!\n", dlm->name,
2277 res->lockname.len, res->lockname.name, node);
2278 dlm_print_one_lock_resource(res);
2284 item = kzalloc(sizeof(*item), GFP_NOFS);
2291 dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2292 item->u.dl.deref_res = res;
2293 item->u.dl.deref_node = node;
2295 spin_lock(&dlm->work_lock);
2296 list_add_tail(&item->list, &dlm->work_list);
2297 spin_unlock(&dlm->work_lock);
2299 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2304 dlm_lockres_put(res);
2310 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2312 struct dlm_ctxt *dlm;
2313 struct dlm_lock_resource *res;
2318 res = item->u.dl.deref_res;
2319 node = item->u.dl.deref_node;
2321 spin_lock(&res->spinlock);
2322 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2323 if (test_bit(node, res->refmap)) {
2324 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2325 dlm_lockres_clear_refmap_bit(dlm, res, node);
2328 spin_unlock(&res->spinlock);
2331 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2332 dlm->name, res->lockname.len, res->lockname.name, node);
2333 dlm_lockres_calc_usage(dlm, res);
2335 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2336 "but it is already dropped!\n", dlm->name,
2337 res->lockname.len, res->lockname.name, node);
2338 dlm_print_one_lock_resource(res);
2341 dlm_lockres_put(res);
2345 * A migrateable resource is one that is :
2346 * 1. locally mastered, and,
2347 * 2. zero local locks, and,
2348 * 3. one or more non-local locks, or, one or more references
2349 * Returns 1 if yes, 0 if not.
2351 static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
2352 struct dlm_lock_resource *res)
2354 enum dlm_lockres_list idx;
2355 int nonlocal = 0, node_ref;
2356 struct list_head *queue;
2357 struct dlm_lock *lock;
2360 assert_spin_locked(&res->spinlock);
2362 if (res->owner != dlm->node_num)
2365 for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2366 queue = dlm_list_idx_to_ptr(res, idx);
2367 list_for_each_entry(lock, queue, list) {
2368 if (lock->ml.node != dlm->node_num) {
2372 cookie = be64_to_cpu(lock->ml.cookie);
2373 mlog(0, "%s: Not migrateable res %.*s, lock %u:%llu on "
2374 "%s list\n", dlm->name, res->lockname.len,
2376 dlm_get_lock_cookie_node(cookie),
2377 dlm_get_lock_cookie_seq(cookie),
2378 dlm_list_in_text(idx));
2384 node_ref = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
2385 if (node_ref >= O2NM_MAX_NODES)
2389 mlog(0, "%s: res %.*s, Migrateable\n", dlm->name, res->lockname.len,
2390 res->lockname.name);
2396 * DLM_MIGRATE_LOCKRES
2400 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2401 struct dlm_lock_resource *res, u8 target)
2403 struct dlm_master_list_entry *mle = NULL;
2404 struct dlm_master_list_entry *oldmle = NULL;
2405 struct dlm_migratable_lockres *mres = NULL;
2408 unsigned int namelen;
2415 BUG_ON(target == O2NM_MAX_NODES);
2417 name = res->lockname.name;
2418 namelen = res->lockname.len;
2420 mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name,
2423 /* preallocate up front. if this fails, abort */
2425 mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2431 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2439 * clear any existing master requests and
2440 * add the migration mle to the list
2442 spin_lock(&dlm->spinlock);
2443 spin_lock(&dlm->master_lock);
2444 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2445 namelen, target, dlm->node_num);
2446 spin_unlock(&dlm->master_lock);
2447 spin_unlock(&dlm->spinlock);
2449 if (ret == -EEXIST) {
2450 mlog(0, "another process is already migrating it\n");
2456 * set the MIGRATING flag and flush asts
2457 * if we fail after this we need to re-dirty the lockres
2459 if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2460 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2461 "the target went down.\n", res->lockname.len,
2462 res->lockname.name, target);
2463 spin_lock(&res->spinlock);
2464 res->state &= ~DLM_LOCK_RES_MIGRATING;
2466 spin_unlock(&res->spinlock);
2472 /* master is known, detach if not already detached */
2473 dlm_mle_detach_hb_events(dlm, oldmle);
2474 dlm_put_mle(oldmle);
2479 dlm_mle_detach_hb_events(dlm, mle);
2482 kmem_cache_free(dlm_mle_cache, mle);
2489 * at this point, we have a migration target, an mle
2490 * in the master list, and the MIGRATING flag set on
2494 /* now that remote nodes are spinning on the MIGRATING flag,
2495 * ensure that all assert_master work is flushed. */
2496 flush_workqueue(dlm->dlm_worker);
2498 /* get an extra reference on the mle.
2499 * otherwise the assert_master from the new
2500 * master will destroy this.
2501 * also, make sure that all callers of dlm_get_mle
2502 * take both dlm->spinlock and dlm->master_lock */
2503 spin_lock(&dlm->spinlock);
2504 spin_lock(&dlm->master_lock);
2505 dlm_get_mle_inuse(mle);
2506 spin_unlock(&dlm->master_lock);
2507 spin_unlock(&dlm->spinlock);
2509 /* notify new node and send all lock state */
2510 /* call send_one_lockres with migration flag.
2511 * this serves as notice to the target node that a
2512 * migration is starting. */
2513 ret = dlm_send_one_lockres(dlm, res, mres, target,
2514 DLM_MRES_MIGRATION);
2517 mlog(0, "migration to node %u failed with %d\n",
2519 /* migration failed, detach and clean up mle */
2520 dlm_mle_detach_hb_events(dlm, mle);
2522 dlm_put_mle_inuse(mle);
2523 spin_lock(&res->spinlock);
2524 res->state &= ~DLM_LOCK_RES_MIGRATING;
2526 spin_unlock(&res->spinlock);
2527 if (dlm_is_host_down(ret))
2528 dlm_wait_for_node_death(dlm, target,
2529 DLM_NODE_DEATH_WAIT_MAX);
2533 /* at this point, the target sends a message to all nodes,
2534 * (using dlm_do_migrate_request). this node is skipped since
2535 * we had to put an mle in the list to begin the process. this
2536 * node now waits for target to do an assert master. this node
2537 * will be the last one notified, ensuring that the migration
2538 * is complete everywhere. if the target dies while this is
2539 * going on, some nodes could potentially see the target as the
2540 * master, so it is important that my recovery finds the migration
2541 * mle and sets the master to UNKNOWN. */
2544 /* wait for new node to assert master */
2546 ret = wait_event_interruptible_timeout(mle->wq,
2547 (atomic_read(&mle->woken) == 1),
2548 msecs_to_jiffies(5000));
2551 if (atomic_read(&mle->woken) == 1 ||
2552 res->owner == target)
2555 mlog(0, "%s:%.*s: timed out during migration\n",
2556 dlm->name, res->lockname.len, res->lockname.name);
2557 /* avoid hang during shutdown when migrating lockres
2558 * to a node which also goes down */
2559 if (dlm_is_node_dead(dlm, target)) {
2560 mlog(0, "%s:%.*s: expected migration "
2561 "target %u is no longer up, restarting\n",
2562 dlm->name, res->lockname.len,
2563 res->lockname.name, target);
2565 /* migration failed, detach and clean up mle */
2566 dlm_mle_detach_hb_events(dlm, mle);
2568 dlm_put_mle_inuse(mle);
2569 spin_lock(&res->spinlock);
2570 res->state &= ~DLM_LOCK_RES_MIGRATING;
2572 spin_unlock(&res->spinlock);
2576 mlog(0, "%s:%.*s: caught signal during migration\n",
2577 dlm->name, res->lockname.len, res->lockname.name);
2580 /* all done, set the owner, clear the flag */
2581 spin_lock(&res->spinlock);
2582 dlm_set_lockres_owner(dlm, res, target);
2583 res->state &= ~DLM_LOCK_RES_MIGRATING;
2584 dlm_remove_nonlocal_locks(dlm, res);
2585 spin_unlock(&res->spinlock);
2588 /* master is known, detach if not already detached */
2589 dlm_mle_detach_hb_events(dlm, mle);
2590 dlm_put_mle_inuse(mle);
2593 dlm_lockres_calc_usage(dlm, res);
2596 /* re-dirty the lockres if we failed */
2598 dlm_kick_thread(dlm, res);
2600 /* wake up waiters if the MIGRATING flag got set
2601 * but migration failed */
2606 free_page((unsigned long)mres);
2610 mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen,
2615 #define DLM_MIGRATION_RETRY_MS 100
2618 * Should be called only after beginning the domain leave process.
2619 * There should not be any remaining locks on nonlocal lock resources,
2620 * and there should be no local locks left on locally mastered resources.
2622 * Called with the dlm spinlock held, may drop it to do migration, but
2623 * will re-acquire before exit.
2625 * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2627 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2630 int lock_dropped = 0;
2631 u8 target = O2NM_MAX_NODES;
2633 assert_spin_locked(&dlm->spinlock);
2635 spin_lock(&res->spinlock);
2636 if (dlm_is_lockres_migrateable(dlm, res))
2637 target = dlm_pick_migration_target(dlm, res);
2638 spin_unlock(&res->spinlock);
2640 if (target == O2NM_MAX_NODES)
2643 /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2644 spin_unlock(&dlm->spinlock);
2646 ret = dlm_migrate_lockres(dlm, res, target);
2648 mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2649 dlm->name, res->lockname.len, res->lockname.name,
2651 spin_lock(&dlm->spinlock);
2653 return lock_dropped;
2656 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2659 spin_lock(&dlm->ast_lock);
2660 spin_lock(&lock->spinlock);
2661 ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2662 spin_unlock(&lock->spinlock);
2663 spin_unlock(&dlm->ast_lock);
2667 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2668 struct dlm_lock_resource *res,
2672 spin_lock(&res->spinlock);
2673 can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2674 spin_unlock(&res->spinlock);
2676 /* target has died, so make the caller break out of the
2677 * wait_event, but caller must recheck the domain_map */
2678 spin_lock(&dlm->spinlock);
2679 if (!test_bit(mig_target, dlm->domain_map))
2681 spin_unlock(&dlm->spinlock);
2685 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2686 struct dlm_lock_resource *res)
2689 spin_lock(&res->spinlock);
2690 ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2691 spin_unlock(&res->spinlock);
2696 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2697 struct dlm_lock_resource *res,
2702 mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2703 res->lockname.len, res->lockname.name, dlm->node_num,
2705 /* need to set MIGRATING flag on lockres. this is done by
2706 * ensuring that all asts have been flushed for this lockres. */
2707 spin_lock(&res->spinlock);
2708 BUG_ON(res->migration_pending);
2709 res->migration_pending = 1;
2710 /* strategy is to reserve an extra ast then release
2711 * it below, letting the release do all of the work */
2712 __dlm_lockres_reserve_ast(res);
2713 spin_unlock(&res->spinlock);
2715 /* now flush all the pending asts */
2716 dlm_kick_thread(dlm, res);
2717 /* before waiting on DIRTY, block processes which may
2718 * try to dirty the lockres before MIGRATING is set */
2719 spin_lock(&res->spinlock);
2720 BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2721 res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2722 spin_unlock(&res->spinlock);
2723 /* now wait on any pending asts and the DIRTY state */
2724 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2725 dlm_lockres_release_ast(dlm, res);
2727 mlog(0, "about to wait on migration_wq, dirty=%s\n",
2728 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2729 /* if the extra ref we just put was the final one, this
2730 * will pass thru immediately. otherwise, we need to wait
2731 * for the last ast to finish. */
2733 ret = wait_event_interruptible_timeout(dlm->migration_wq,
2734 dlm_migration_can_proceed(dlm, res, target),
2735 msecs_to_jiffies(1000));
2737 mlog(0, "woken again: migrating? %s, dead? %s\n",
2738 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2739 test_bit(target, dlm->domain_map) ? "no":"yes");
2741 mlog(0, "all is well: migrating? %s, dead? %s\n",
2742 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2743 test_bit(target, dlm->domain_map) ? "no":"yes");
2745 if (!dlm_migration_can_proceed(dlm, res, target)) {
2746 mlog(0, "trying again...\n");
2751 /* did the target go down or die? */
2752 spin_lock(&dlm->spinlock);
2753 if (!test_bit(target, dlm->domain_map)) {
2754 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2758 spin_unlock(&dlm->spinlock);
2761 * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2762 * another try; otherwise, we are sure the MIGRATING state is there,
2763 * drop the unneded state which blocked threads trying to DIRTY
2765 spin_lock(&res->spinlock);
2766 BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2767 res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2769 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2770 spin_unlock(&res->spinlock);
2775 * o the DLM_LOCK_RES_MIGRATING flag is set if target not down
2776 * o there are no pending asts on this lockres
2777 * o all processes trying to reserve an ast on this
2778 * lockres must wait for the MIGRATING flag to clear
2783 /* last step in the migration process.
2784 * original master calls this to free all of the dlm_lock
2785 * structures that used to be for other nodes. */
2786 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2787 struct dlm_lock_resource *res)
2789 struct list_head *queue = &res->granted;
2791 struct dlm_lock *lock, *next;
2793 assert_spin_locked(&res->spinlock);
2795 BUG_ON(res->owner == dlm->node_num);
2797 for (i=0; i<3; i++) {
2798 list_for_each_entry_safe(lock, next, queue, list) {
2799 if (lock->ml.node != dlm->node_num) {
2800 mlog(0, "putting lock for node %u\n",
2802 /* be extra careful */
2803 BUG_ON(!list_empty(&lock->ast_list));
2804 BUG_ON(!list_empty(&lock->bast_list));
2805 BUG_ON(lock->ast_pending);
2806 BUG_ON(lock->bast_pending);
2807 dlm_lockres_clear_refmap_bit(dlm, res,
2809 list_del_init(&lock->list);
2811 /* In a normal unlock, we would have added a
2812 * DLM_UNLOCK_FREE_LOCK action. Force it. */
2820 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2821 if (bit >= O2NM_MAX_NODES)
2823 /* do not clear the local node reference, if there is a
2824 * process holding this, let it drop the ref itself */
2825 if (bit != dlm->node_num) {
2826 mlog(0, "%s:%.*s: node %u had a ref to this "
2827 "migrating lockres, clearing\n", dlm->name,
2828 res->lockname.len, res->lockname.name, bit);
2829 dlm_lockres_clear_refmap_bit(dlm, res, bit);
2836 * Pick a node to migrate the lock resource to. This function selects a
2837 * potential target based first on the locks and then on refmap. It skips
2838 * nodes that are in the process of exiting the domain.
2840 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2841 struct dlm_lock_resource *res)
2843 enum dlm_lockres_list idx;
2844 struct list_head *queue = &res->granted;
2845 struct dlm_lock *lock;
2847 u8 nodenum = O2NM_MAX_NODES;
2849 assert_spin_locked(&dlm->spinlock);
2850 assert_spin_locked(&res->spinlock);
2852 /* Go through all the locks */
2853 for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2854 queue = dlm_list_idx_to_ptr(res, idx);
2855 list_for_each_entry(lock, queue, list) {
2856 if (lock->ml.node == dlm->node_num)
2858 if (test_bit(lock->ml.node, dlm->exit_domain_map))
2860 nodenum = lock->ml.node;
2865 /* Go thru the refmap */
2868 noderef = find_next_bit(res->refmap, O2NM_MAX_NODES,
2870 if (noderef >= O2NM_MAX_NODES)
2872 if (noderef == dlm->node_num)
2874 if (test_bit(noderef, dlm->exit_domain_map))
2884 /* this is called by the new master once all lockres
2885 * data has been received */
2886 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2887 struct dlm_lock_resource *res,
2888 u8 master, u8 new_master,
2889 struct dlm_node_iter *iter)
2891 struct dlm_migrate_request migrate;
2892 int ret, skip, status = 0;
2895 memset(&migrate, 0, sizeof(migrate));
2896 migrate.namelen = res->lockname.len;
2897 memcpy(migrate.name, res->lockname.name, migrate.namelen);
2898 migrate.new_master = new_master;
2899 migrate.master = master;
2903 /* send message to all nodes, except the master and myself */
2904 while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2905 if (nodenum == master ||
2906 nodenum == new_master)
2909 /* We could race exit domain. If exited, skip. */
2910 spin_lock(&dlm->spinlock);
2911 skip = (!test_bit(nodenum, dlm->domain_map));
2912 spin_unlock(&dlm->spinlock);
2914 clear_bit(nodenum, iter->node_map);
2918 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2919 &migrate, sizeof(migrate), nodenum,
2922 mlog(ML_ERROR, "%s: res %.*s, Error %d send "
2923 "MIGRATE_REQUEST to node %u\n", dlm->name,
2924 migrate.namelen, migrate.name, ret, nodenum);
2925 if (!dlm_is_host_down(ret)) {
2926 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
2929 clear_bit(nodenum, iter->node_map);
2931 } else if (status < 0) {
2932 mlog(0, "migrate request (node %u) returned %d!\n",
2935 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
2936 /* during the migration request we short-circuited
2937 * the mastery of the lockres. make sure we have
2938 * a mastery ref for nodenum */
2939 mlog(0, "%s:%.*s: need ref for node %u\n",
2940 dlm->name, res->lockname.len, res->lockname.name,
2942 spin_lock(&res->spinlock);
2943 dlm_lockres_set_refmap_bit(dlm, res, nodenum);
2944 spin_unlock(&res->spinlock);
2951 mlog(0, "returning ret=%d\n", ret);
2956 /* if there is an existing mle for this lockres, we now know who the master is.
2957 * (the one who sent us *this* message) we can clear it up right away.
2958 * since the process that put the mle on the list still has a reference to it,
2959 * we can unhash it now, set the master and wake the process. as a result,
2960 * we will have no mle in the list to start with. now we can add an mle for
2961 * the migration and this should be the only one found for those scanning the
2963 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
2966 struct dlm_ctxt *dlm = data;
2967 struct dlm_lock_resource *res = NULL;
2968 struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
2969 struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
2971 unsigned int namelen, hash;
2977 name = migrate->name;
2978 namelen = migrate->namelen;
2979 hash = dlm_lockid_hash(name, namelen);
2981 /* preallocate.. if this fails, abort */
2982 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2989 /* check for pre-existing lock */
2990 spin_lock(&dlm->spinlock);
2991 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
2993 spin_lock(&res->spinlock);
2994 if (res->state & DLM_LOCK_RES_RECOVERING) {
2995 /* if all is working ok, this can only mean that we got
2996 * a migrate request from a node that we now see as
2997 * dead. what can we do here? drop it to the floor? */
2998 spin_unlock(&res->spinlock);
2999 mlog(ML_ERROR, "Got a migrate request, but the "
3000 "lockres is marked as recovering!");
3001 kmem_cache_free(dlm_mle_cache, mle);
3002 ret = -EINVAL; /* need a better solution */
3005 res->state |= DLM_LOCK_RES_MIGRATING;
3006 spin_unlock(&res->spinlock);
3009 spin_lock(&dlm->master_lock);
3010 /* ignore status. only nonzero status would BUG. */
3011 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3013 migrate->new_master,
3016 spin_unlock(&dlm->master_lock);
3018 spin_unlock(&dlm->spinlock);
3021 /* master is known, detach if not already detached */
3022 dlm_mle_detach_hb_events(dlm, oldmle);
3023 dlm_put_mle(oldmle);
3027 dlm_lockres_put(res);
3033 /* must be holding dlm->spinlock and dlm->master_lock
3034 * when adding a migration mle, we can clear any other mles
3035 * in the master list because we know with certainty that
3036 * the master is "master". so we remove any old mle from
3037 * the list after setting it's master field, and then add
3038 * the new migration mle. this way we can hold with the rule
3039 * of having only one mle for a given lock name at all times. */
3040 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3041 struct dlm_lock_resource *res,
3042 struct dlm_master_list_entry *mle,
3043 struct dlm_master_list_entry **oldmle,
3044 const char *name, unsigned int namelen,
3045 u8 new_master, u8 master)
3052 assert_spin_locked(&dlm->spinlock);
3053 assert_spin_locked(&dlm->master_lock);
3055 /* caller is responsible for any ref taken here on oldmle */
3056 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3058 struct dlm_master_list_entry *tmp = *oldmle;
3059 spin_lock(&tmp->spinlock);
3060 if (tmp->type == DLM_MLE_MIGRATION) {
3061 if (master == dlm->node_num) {
3062 /* ah another process raced me to it */
3063 mlog(0, "tried to migrate %.*s, but some "
3064 "process beat me to it\n",
3068 /* bad. 2 NODES are trying to migrate! */
3069 mlog(ML_ERROR, "migration error mle: "
3070 "master=%u new_master=%u // request: "
3071 "master=%u new_master=%u // "
3073 tmp->master, tmp->new_master,
3079 /* this is essentially what assert_master does */
3080 tmp->master = master;
3081 atomic_set(&tmp->woken, 1);
3083 /* remove it so that only one mle will be found */
3084 __dlm_unlink_mle(dlm, tmp);
3085 __dlm_mle_detach_hb_events(dlm, tmp);
3086 ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3087 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3088 "telling master to get ref for cleared out mle "
3089 "during migration\n", dlm->name, namelen, name,
3090 master, new_master);
3092 spin_unlock(&tmp->spinlock);
3095 /* now add a migration mle to the tail of the list */
3096 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3097 mle->new_master = new_master;
3098 /* the new master will be sending an assert master for this.
3099 * at that point we will get the refmap reference */
3100 mle->master = master;
3101 /* do this for consistency with other mle types */
3102 set_bit(new_master, mle->maybe_map);
3103 __dlm_insert_mle(dlm, mle);
3109 * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3111 static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3112 struct dlm_master_list_entry *mle)
3114 struct dlm_lock_resource *res;
3116 /* Find the lockres associated to the mle and set its owner to UNK */
3117 res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3120 spin_unlock(&dlm->master_lock);
3122 /* move lockres onto recovery list */
3123 spin_lock(&res->spinlock);
3124 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3125 dlm_move_lockres_to_recovery_list(dlm, res);
3126 spin_unlock(&res->spinlock);
3127 dlm_lockres_put(res);
3129 /* about to get rid of mle, detach from heartbeat */
3130 __dlm_mle_detach_hb_events(dlm, mle);
3133 spin_lock(&dlm->master_lock);
3135 spin_unlock(&dlm->master_lock);
3141 static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3142 struct dlm_master_list_entry *mle)
3144 __dlm_mle_detach_hb_events(dlm, mle);
3146 spin_lock(&mle->spinlock);
3147 __dlm_unlink_mle(dlm, mle);
3148 atomic_set(&mle->woken, 1);
3149 spin_unlock(&mle->spinlock);
3154 static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3155 struct dlm_master_list_entry *mle, u8 dead_node)
3159 BUG_ON(mle->type != DLM_MLE_BLOCK);
3161 spin_lock(&mle->spinlock);
3162 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3163 if (bit != dead_node) {
3164 mlog(0, "mle found, but dead node %u would not have been "
3165 "master\n", dead_node);
3166 spin_unlock(&mle->spinlock);
3168 /* Must drop the refcount by one since the assert_master will
3169 * never arrive. This may result in the mle being unlinked and
3170 * freed, but there may still be a process waiting in the
3171 * dlmlock path which is fine. */
3172 mlog(0, "node %u was expected master\n", dead_node);
3173 atomic_set(&mle->woken, 1);
3174 spin_unlock(&mle->spinlock);
3177 /* Do not need events any longer, so detach from heartbeat */
3178 __dlm_mle_detach_hb_events(dlm, mle);
3183 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3185 struct dlm_master_list_entry *mle;
3186 struct dlm_lock_resource *res;
3187 struct hlist_head *bucket;
3188 struct hlist_node *list;
3191 mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
3193 assert_spin_locked(&dlm->spinlock);
3195 /* clean the master list */
3196 spin_lock(&dlm->master_lock);
3197 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3198 bucket = dlm_master_hash(dlm, i);
3199 hlist_for_each(list, bucket) {
3200 mle = hlist_entry(list, struct dlm_master_list_entry,
3203 BUG_ON(mle->type != DLM_MLE_BLOCK &&
3204 mle->type != DLM_MLE_MASTER &&
3205 mle->type != DLM_MLE_MIGRATION);
3207 /* MASTER mles are initiated locally. The waiting
3208 * process will notice the node map change shortly.
3209 * Let that happen as normal. */
3210 if (mle->type == DLM_MLE_MASTER)
3213 /* BLOCK mles are initiated by other nodes. Need to
3214 * clean up if the dead node would have been the
3216 if (mle->type == DLM_MLE_BLOCK) {
3217 dlm_clean_block_mle(dlm, mle, dead_node);
3221 /* Everything else is a MIGRATION mle */
3223 /* The rule for MIGRATION mles is that the master
3224 * becomes UNKNOWN if *either* the original or the new
3225 * master dies. All UNKNOWN lockres' are sent to
3226 * whichever node becomes the recovery master. The new
3227 * master is responsible for determining if there is
3228 * still a master for this lockres, or if he needs to
3229 * take over mastery. Either way, this node should
3230 * expect another message to resolve this. */
3232 if (mle->master != dead_node &&
3233 mle->new_master != dead_node)
3236 /* If we have reached this point, this mle needs to be
3237 * removed from the list and freed. */
3238 dlm_clean_migration_mle(dlm, mle);
3240 mlog(0, "%s: node %u died during migration from "
3241 "%u to %u!\n", dlm->name, dead_node, mle->master,
3244 /* If we find a lockres associated with the mle, we've
3245 * hit this rare case that messes up our lock ordering.
3246 * If so, we need to drop the master lock so that we can
3247 * take the lockres lock, meaning that we will have to
3248 * restart from the head of list. */
3249 res = dlm_reset_mleres_owner(dlm, mle);
3254 /* This may be the last reference */
3258 spin_unlock(&dlm->master_lock);
3261 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3264 struct dlm_node_iter iter;
3267 spin_lock(&dlm->spinlock);
3268 dlm_node_iter_init(dlm->domain_map, &iter);
3269 clear_bit(old_master, iter.node_map);
3270 clear_bit(dlm->node_num, iter.node_map);
3271 spin_unlock(&dlm->spinlock);
3273 /* ownership of the lockres is changing. account for the
3274 * mastery reference here since old_master will briefly have
3275 * a reference after the migration completes */
3276 spin_lock(&res->spinlock);
3277 dlm_lockres_set_refmap_bit(dlm, res, old_master);
3278 spin_unlock(&res->spinlock);
3280 mlog(0, "now time to do a migrate request to other nodes\n");
3281 ret = dlm_do_migrate_request(dlm, res, old_master,
3282 dlm->node_num, &iter);
3288 mlog(0, "doing assert master of %.*s to all except the original node\n",
3289 res->lockname.len, res->lockname.name);
3290 /* this call now finishes out the nodemap
3291 * even if one or more nodes die */
3292 ret = dlm_do_assert_master(dlm, res, iter.node_map,
3293 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3295 /* no longer need to retry. all living nodes contacted. */
3300 memset(iter.node_map, 0, sizeof(iter.node_map));
3301 set_bit(old_master, iter.node_map);
3302 mlog(0, "doing assert master of %.*s back to %u\n",
3303 res->lockname.len, res->lockname.name, old_master);
3304 ret = dlm_do_assert_master(dlm, res, iter.node_map,
3305 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3307 mlog(0, "assert master to original master failed "
3309 /* the only nonzero status here would be because of
3310 * a dead original node. we're done. */
3314 /* all done, set the owner, clear the flag */
3315 spin_lock(&res->spinlock);
3316 dlm_set_lockres_owner(dlm, res, dlm->node_num);
3317 res->state &= ~DLM_LOCK_RES_MIGRATING;
3318 spin_unlock(&res->spinlock);
3319 /* re-dirty it on the new master */
3320 dlm_kick_thread(dlm, res);
3327 * LOCKRES AST REFCOUNT
3328 * this is integral to migration
3331 /* for future intent to call an ast, reserve one ahead of time.
3332 * this should be called only after waiting on the lockres
3333 * with dlm_wait_on_lockres, and while still holding the
3334 * spinlock after the call. */
3335 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3337 assert_spin_locked(&res->spinlock);
3338 if (res->state & DLM_LOCK_RES_MIGRATING) {
3339 __dlm_print_one_lock_resource(res);
3341 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3343 atomic_inc(&res->asts_reserved);
3347 * used to drop the reserved ast, either because it went unused,
3348 * or because the ast/bast was actually called.
3350 * also, if there is a pending migration on this lockres,
3351 * and this was the last pending ast on the lockres,
3352 * atomically set the MIGRATING flag before we drop the lock.
3353 * this is how we ensure that migration can proceed with no
3354 * asts in progress. note that it is ok if the state of the
3355 * queues is such that a lock should be granted in the future
3356 * or that a bast should be fired, because the new master will
3357 * shuffle the lists on this lockres as soon as it is migrated.
3359 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3360 struct dlm_lock_resource *res)
3362 if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3365 if (!res->migration_pending) {
3366 spin_unlock(&res->spinlock);
3370 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3371 res->migration_pending = 0;
3372 res->state |= DLM_LOCK_RES_MIGRATING;
3373 spin_unlock(&res->spinlock);
3375 wake_up(&dlm->migration_wq);
3378 void dlm_force_free_mles(struct dlm_ctxt *dlm)
3381 struct hlist_head *bucket;
3382 struct dlm_master_list_entry *mle;
3383 struct hlist_node *tmp, *list;
3386 * We notified all other nodes that we are exiting the domain and
3387 * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3388 * around we force free them and wake any processes that are waiting
3391 spin_lock(&dlm->spinlock);
3392 spin_lock(&dlm->master_lock);
3394 BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3395 BUG_ON((find_next_bit(dlm->domain_map, O2NM_MAX_NODES, 0) < O2NM_MAX_NODES));
3397 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3398 bucket = dlm_master_hash(dlm, i);
3399 hlist_for_each_safe(list, tmp, bucket) {
3400 mle = hlist_entry(list, struct dlm_master_list_entry,
3402 if (mle->type != DLM_MLE_BLOCK) {
3403 mlog(ML_ERROR, "bad mle: %p\n", mle);
3404 dlm_print_one_mle(mle);
3406 atomic_set(&mle->woken, 1);
3409 __dlm_unlink_mle(dlm, mle);
3410 __dlm_mle_detach_hb_events(dlm, mle);
3414 spin_unlock(&dlm->master_lock);
3415 spin_unlock(&dlm->spinlock);