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;
86 static struct kmem_cache *dlm_lockname_cache;
87 static struct kmem_cache *dlm_mle_cache;
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;
347 assert_spin_locked(&dlm->master_lock);
349 hash = dlm_lockid_hash(name, namelen);
350 bucket = dlm_master_hash(dlm, hash);
351 hlist_for_each_entry(tmpmle, bucket, master_hash_node) {
352 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
361 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
363 struct dlm_master_list_entry *mle;
365 assert_spin_locked(&dlm->spinlock);
367 list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
369 dlm_mle_node_up(dlm, mle, NULL, idx);
371 dlm_mle_node_down(dlm, mle, NULL, idx);
375 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
376 struct dlm_master_list_entry *mle,
377 struct o2nm_node *node, int idx)
379 spin_lock(&mle->spinlock);
381 if (!test_bit(idx, mle->node_map))
382 mlog(0, "node %u already removed from nodemap!\n", idx);
384 clear_bit(idx, mle->node_map);
386 spin_unlock(&mle->spinlock);
389 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
390 struct dlm_master_list_entry *mle,
391 struct o2nm_node *node, int idx)
393 spin_lock(&mle->spinlock);
395 if (test_bit(idx, mle->node_map))
396 mlog(0, "node %u already in node map!\n", idx);
398 set_bit(idx, mle->node_map);
400 spin_unlock(&mle->spinlock);
404 int dlm_init_mle_cache(void)
406 dlm_mle_cache = kmem_cache_create("o2dlm_mle",
407 sizeof(struct dlm_master_list_entry),
408 0, SLAB_HWCACHE_ALIGN,
410 if (dlm_mle_cache == NULL)
415 void dlm_destroy_mle_cache(void)
418 kmem_cache_destroy(dlm_mle_cache);
421 static void dlm_mle_release(struct kref *kref)
423 struct dlm_master_list_entry *mle;
424 struct dlm_ctxt *dlm;
426 mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
429 assert_spin_locked(&dlm->spinlock);
430 assert_spin_locked(&dlm->master_lock);
432 mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
435 /* remove from list if not already */
436 __dlm_unlink_mle(dlm, mle);
438 /* detach the mle from the domain node up/down events */
439 __dlm_mle_detach_hb_events(dlm, mle);
441 atomic_dec(&dlm->mle_cur_count[mle->type]);
443 /* NOTE: kfree under spinlock here.
444 * if this is bad, we can move this to a freelist. */
445 kmem_cache_free(dlm_mle_cache, mle);
450 * LOCK RESOURCE FUNCTIONS
453 int dlm_init_master_caches(void)
455 dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
456 sizeof(struct dlm_lock_resource),
457 0, SLAB_HWCACHE_ALIGN, NULL);
458 if (!dlm_lockres_cache)
461 dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
462 DLM_LOCKID_NAME_MAX, 0,
463 SLAB_HWCACHE_ALIGN, NULL);
464 if (!dlm_lockname_cache)
469 dlm_destroy_master_caches();
473 void dlm_destroy_master_caches(void)
475 if (dlm_lockname_cache) {
476 kmem_cache_destroy(dlm_lockname_cache);
477 dlm_lockname_cache = NULL;
480 if (dlm_lockres_cache) {
481 kmem_cache_destroy(dlm_lockres_cache);
482 dlm_lockres_cache = NULL;
486 static void dlm_lockres_release(struct kref *kref)
488 struct dlm_lock_resource *res;
489 struct dlm_ctxt *dlm;
491 res = container_of(kref, struct dlm_lock_resource, refs);
494 /* This should not happen -- all lockres' have a name
495 * associated with them at init time. */
496 BUG_ON(!res->lockname.name);
498 mlog(0, "destroying lockres %.*s\n", res->lockname.len,
501 atomic_dec(&dlm->res_cur_count);
503 if (!hlist_unhashed(&res->hash_node) ||
504 !list_empty(&res->granted) ||
505 !list_empty(&res->converting) ||
506 !list_empty(&res->blocked) ||
507 !list_empty(&res->dirty) ||
508 !list_empty(&res->recovering) ||
509 !list_empty(&res->purge)) {
511 "Going to BUG for resource %.*s."
512 " We're on a list! [%c%c%c%c%c%c%c]\n",
513 res->lockname.len, res->lockname.name,
514 !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
515 !list_empty(&res->granted) ? 'G' : ' ',
516 !list_empty(&res->converting) ? 'C' : ' ',
517 !list_empty(&res->blocked) ? 'B' : ' ',
518 !list_empty(&res->dirty) ? 'D' : ' ',
519 !list_empty(&res->recovering) ? 'R' : ' ',
520 !list_empty(&res->purge) ? 'P' : ' ');
522 dlm_print_one_lock_resource(res);
525 /* By the time we're ready to blow this guy away, we shouldn't
526 * be on any lists. */
527 BUG_ON(!hlist_unhashed(&res->hash_node));
528 BUG_ON(!list_empty(&res->granted));
529 BUG_ON(!list_empty(&res->converting));
530 BUG_ON(!list_empty(&res->blocked));
531 BUG_ON(!list_empty(&res->dirty));
532 BUG_ON(!list_empty(&res->recovering));
533 BUG_ON(!list_empty(&res->purge));
535 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
537 kmem_cache_free(dlm_lockres_cache, res);
540 void dlm_lockres_put(struct dlm_lock_resource *res)
542 kref_put(&res->refs, dlm_lockres_release);
545 static void dlm_init_lockres(struct dlm_ctxt *dlm,
546 struct dlm_lock_resource *res,
547 const char *name, unsigned int namelen)
551 /* If we memset here, we lose our reference to the kmalloc'd
552 * res->lockname.name, so be sure to init every field
555 qname = (char *) res->lockname.name;
556 memcpy(qname, name, namelen);
558 res->lockname.len = namelen;
559 res->lockname.hash = dlm_lockid_hash(name, namelen);
561 init_waitqueue_head(&res->wq);
562 spin_lock_init(&res->spinlock);
563 INIT_HLIST_NODE(&res->hash_node);
564 INIT_LIST_HEAD(&res->granted);
565 INIT_LIST_HEAD(&res->converting);
566 INIT_LIST_HEAD(&res->blocked);
567 INIT_LIST_HEAD(&res->dirty);
568 INIT_LIST_HEAD(&res->recovering);
569 INIT_LIST_HEAD(&res->purge);
570 INIT_LIST_HEAD(&res->tracking);
571 atomic_set(&res->asts_reserved, 0);
572 res->migration_pending = 0;
573 res->inflight_locks = 0;
574 res->inflight_assert_workers = 0;
578 kref_init(&res->refs);
580 atomic_inc(&dlm->res_tot_count);
581 atomic_inc(&dlm->res_cur_count);
583 /* just for consistency */
584 spin_lock(&res->spinlock);
585 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
586 spin_unlock(&res->spinlock);
588 res->state = DLM_LOCK_RES_IN_PROGRESS;
592 spin_lock(&dlm->spinlock);
593 list_add_tail(&res->tracking, &dlm->tracking_list);
594 spin_unlock(&dlm->spinlock);
596 memset(res->lvb, 0, DLM_LVB_LEN);
597 memset(res->refmap, 0, sizeof(res->refmap));
600 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
602 unsigned int namelen)
604 struct dlm_lock_resource *res = NULL;
606 res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
610 res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
611 if (!res->lockname.name)
614 dlm_init_lockres(dlm, res, name, namelen);
619 kmem_cache_free(dlm_lockres_cache, res);
623 void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm,
624 struct dlm_lock_resource *res, int bit)
626 assert_spin_locked(&res->spinlock);
628 mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len,
629 res->lockname.name, bit, __builtin_return_address(0));
631 set_bit(bit, res->refmap);
634 void dlm_lockres_clear_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, clr node %u, %ps()\n", res->lockname.len,
640 res->lockname.name, bit, __builtin_return_address(0));
642 clear_bit(bit, res->refmap);
645 static void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
646 struct dlm_lock_resource *res)
648 res->inflight_locks++;
650 mlog(0, "%s: res %.*s, inflight++: now %u, %ps()\n", dlm->name,
651 res->lockname.len, res->lockname.name, res->inflight_locks,
652 __builtin_return_address(0));
655 void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
656 struct dlm_lock_resource *res)
658 assert_spin_locked(&res->spinlock);
659 __dlm_lockres_grab_inflight_ref(dlm, res);
662 void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
663 struct dlm_lock_resource *res)
665 assert_spin_locked(&res->spinlock);
667 BUG_ON(res->inflight_locks == 0);
669 res->inflight_locks--;
671 mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm->name,
672 res->lockname.len, res->lockname.name, res->inflight_locks,
673 __builtin_return_address(0));
678 void __dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm,
679 struct dlm_lock_resource *res)
681 assert_spin_locked(&res->spinlock);
682 res->inflight_assert_workers++;
683 mlog(0, "%s:%.*s: inflight assert worker++: now %u\n",
684 dlm->name, res->lockname.len, res->lockname.name,
685 res->inflight_assert_workers);
688 static void __dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
689 struct dlm_lock_resource *res)
691 assert_spin_locked(&res->spinlock);
692 BUG_ON(res->inflight_assert_workers == 0);
693 res->inflight_assert_workers--;
694 mlog(0, "%s:%.*s: inflight assert worker--: now %u\n",
695 dlm->name, res->lockname.len, res->lockname.name,
696 res->inflight_assert_workers);
699 static void dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
700 struct dlm_lock_resource *res)
702 spin_lock(&res->spinlock);
703 __dlm_lockres_drop_inflight_worker(dlm, res);
704 spin_unlock(&res->spinlock);
708 * lookup a lock resource by name.
709 * may already exist in the hashtable.
710 * lockid is null terminated
712 * if not, allocate enough for the lockres and for
713 * the temporary structure used in doing the mastering.
715 * also, do a lookup in the dlm->master_list to see
716 * if another node has begun mastering the same lock.
717 * if so, there should be a block entry in there
718 * for this name, and we should *not* attempt to master
719 * the lock here. need to wait around for that node
720 * to assert_master (or die).
723 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
728 struct dlm_lock_resource *tmpres=NULL, *res=NULL;
729 struct dlm_master_list_entry *mle = NULL;
730 struct dlm_master_list_entry *alloc_mle = NULL;
733 struct dlm_node_iter iter;
736 int bit, wait_on_recovery = 0;
740 hash = dlm_lockid_hash(lockid, namelen);
742 mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
745 spin_lock(&dlm->spinlock);
746 tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
748 spin_unlock(&dlm->spinlock);
749 spin_lock(&tmpres->spinlock);
752 * Right after dlm spinlock was released, dlm_thread could have
753 * purged the lockres. Check if lockres got unhashed. If so
756 if (hlist_unhashed(&tmpres->hash_node)) {
757 spin_unlock(&tmpres->spinlock);
758 dlm_lockres_put(tmpres);
763 /* Wait on the thread that is mastering the resource */
764 if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
765 __dlm_wait_on_lockres(tmpres);
766 BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
767 spin_unlock(&tmpres->spinlock);
768 dlm_lockres_put(tmpres);
773 /* Wait on the resource purge to complete before continuing */
774 if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) {
775 BUG_ON(tmpres->owner == dlm->node_num);
776 __dlm_wait_on_lockres_flags(tmpres,
777 DLM_LOCK_RES_DROPPING_REF);
778 spin_unlock(&tmpres->spinlock);
779 dlm_lockres_put(tmpres);
784 /* Grab inflight ref to pin the resource */
785 dlm_lockres_grab_inflight_ref(dlm, tmpres);
787 spin_unlock(&tmpres->spinlock);
789 spin_lock(&dlm->track_lock);
790 if (!list_empty(&res->tracking))
791 list_del_init(&res->tracking);
793 mlog(ML_ERROR, "Resource %.*s not "
794 "on the Tracking list\n",
797 spin_unlock(&dlm->track_lock);
798 dlm_lockres_put(res);
805 spin_unlock(&dlm->spinlock);
806 mlog(0, "allocating a new resource\n");
807 /* nothing found and we need to allocate one. */
808 alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
811 res = dlm_new_lockres(dlm, lockid, namelen);
817 mlog(0, "no lockres found, allocated our own: %p\n", res);
819 if (flags & LKM_LOCAL) {
820 /* caller knows it's safe to assume it's not mastered elsewhere
821 * DONE! return right away */
822 spin_lock(&res->spinlock);
823 dlm_change_lockres_owner(dlm, res, dlm->node_num);
824 __dlm_insert_lockres(dlm, res);
825 dlm_lockres_grab_inflight_ref(dlm, res);
826 spin_unlock(&res->spinlock);
827 spin_unlock(&dlm->spinlock);
828 /* lockres still marked IN_PROGRESS */
832 /* check master list to see if another node has started mastering it */
833 spin_lock(&dlm->master_lock);
835 /* if we found a block, wait for lock to be mastered by another node */
836 blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
839 if (mle->type == DLM_MLE_MASTER) {
840 mlog(ML_ERROR, "master entry for nonexistent lock!\n");
843 mig = (mle->type == DLM_MLE_MIGRATION);
844 /* if there is a migration in progress, let the migration
845 * finish before continuing. we can wait for the absence
846 * of the MIGRATION mle: either the migrate finished or
847 * one of the nodes died and the mle was cleaned up.
848 * if there is a BLOCK here, but it already has a master
849 * set, we are too late. the master does not have a ref
850 * for us in the refmap. detach the mle and drop it.
851 * either way, go back to the top and start over. */
852 if (mig || mle->master != O2NM_MAX_NODES) {
853 BUG_ON(mig && mle->master == dlm->node_num);
854 /* we arrived too late. the master does not
855 * have a ref for us. retry. */
856 mlog(0, "%s:%.*s: late on %s\n",
857 dlm->name, namelen, lockid,
858 mig ? "MIGRATION" : "BLOCK");
859 spin_unlock(&dlm->master_lock);
860 spin_unlock(&dlm->spinlock);
862 /* master is known, detach */
864 dlm_mle_detach_hb_events(dlm, mle);
867 /* this is lame, but we can't wait on either
868 * the mle or lockres waitqueue here */
874 /* go ahead and try to master lock on this node */
876 /* make sure this does not get freed below */
878 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
879 set_bit(dlm->node_num, mle->maybe_map);
880 __dlm_insert_mle(dlm, mle);
882 /* still holding the dlm spinlock, check the recovery map
883 * to see if there are any nodes that still need to be
884 * considered. these will not appear in the mle nodemap
885 * but they might own this lockres. wait on them. */
886 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
887 if (bit < O2NM_MAX_NODES) {
888 mlog(0, "%s: res %.*s, At least one node (%d) "
889 "to recover before lock mastery can begin\n",
890 dlm->name, namelen, (char *)lockid, bit);
891 wait_on_recovery = 1;
895 /* at this point there is either a DLM_MLE_BLOCK or a
896 * DLM_MLE_MASTER on the master list, so it's safe to add the
897 * lockres to the hashtable. anyone who finds the lock will
898 * still have to wait on the IN_PROGRESS. */
900 /* finally add the lockres to its hash bucket */
901 __dlm_insert_lockres(dlm, res);
903 /* since this lockres is new it doesn't not require the spinlock */
904 __dlm_lockres_grab_inflight_ref(dlm, res);
906 /* get an extra ref on the mle in case this is a BLOCK
907 * if so, the creator of the BLOCK may try to put the last
908 * ref at this time in the assert master handler, so we
909 * need an extra one to keep from a bad ptr deref. */
910 dlm_get_mle_inuse(mle);
911 spin_unlock(&dlm->master_lock);
912 spin_unlock(&dlm->spinlock);
915 while (wait_on_recovery) {
916 /* any cluster changes that occurred after dropping the
917 * dlm spinlock would be detectable be a change on the mle,
918 * so we only need to clear out the recovery map once. */
919 if (dlm_is_recovery_lock(lockid, namelen)) {
920 mlog(0, "%s: Recovery map is not empty, but must "
921 "master $RECOVERY lock now\n", dlm->name);
922 if (!dlm_pre_master_reco_lockres(dlm, res))
923 wait_on_recovery = 0;
925 mlog(0, "%s: waiting 500ms for heartbeat state "
926 "change\n", dlm->name);
932 dlm_kick_recovery_thread(dlm);
934 dlm_wait_for_recovery(dlm);
936 spin_lock(&dlm->spinlock);
937 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
938 if (bit < O2NM_MAX_NODES) {
939 mlog(0, "%s: res %.*s, At least one node (%d) "
940 "to recover before lock mastery can begin\n",
941 dlm->name, namelen, (char *)lockid, bit);
942 wait_on_recovery = 1;
944 wait_on_recovery = 0;
945 spin_unlock(&dlm->spinlock);
947 if (wait_on_recovery)
948 dlm_wait_for_node_recovery(dlm, bit, 10000);
951 /* must wait for lock to be mastered elsewhere */
956 dlm_node_iter_init(mle->vote_map, &iter);
957 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
958 ret = dlm_do_master_request(res, mle, nodenum);
961 if (mle->master != O2NM_MAX_NODES) {
962 /* found a master ! */
963 if (mle->master <= nodenum)
965 /* if our master request has not reached the master
966 * yet, keep going until it does. this is how the
967 * master will know that asserts are needed back to
968 * the lower nodes. */
969 mlog(0, "%s: res %.*s, Requests only up to %u but "
970 "master is %u, keep going\n", dlm->name, namelen,
971 lockid, nodenum, mle->master);
976 /* keep going until the response map includes all nodes */
977 ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
979 wait_on_recovery = 1;
980 mlog(0, "%s: res %.*s, Node map changed, redo the master "
981 "request now, blocked=%d\n", dlm->name, res->lockname.len,
982 res->lockname.name, blocked);
984 mlog(ML_ERROR, "%s: res %.*s, Spinning on "
985 "dlm_wait_for_lock_mastery, blocked = %d\n",
986 dlm->name, res->lockname.len,
987 res->lockname.name, blocked);
988 dlm_print_one_lock_resource(res);
989 dlm_print_one_mle(mle);
995 mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len,
996 res->lockname.name, res->owner);
997 /* make sure we never continue without this */
998 BUG_ON(res->owner == O2NM_MAX_NODES);
1000 /* master is known, detach if not already detached */
1001 dlm_mle_detach_hb_events(dlm, mle);
1003 /* put the extra ref */
1004 dlm_put_mle_inuse(mle);
1007 spin_lock(&res->spinlock);
1008 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1009 spin_unlock(&res->spinlock);
1013 /* need to free the unused mle */
1015 kmem_cache_free(dlm_mle_cache, alloc_mle);
1021 #define DLM_MASTERY_TIMEOUT_MS 5000
1023 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
1024 struct dlm_lock_resource *res,
1025 struct dlm_master_list_entry *mle,
1030 int map_changed, voting_done;
1037 /* check if another node has already become the owner */
1038 spin_lock(&res->spinlock);
1039 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1040 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1041 res->lockname.len, res->lockname.name, res->owner);
1042 spin_unlock(&res->spinlock);
1043 /* this will cause the master to re-assert across
1044 * the whole cluster, freeing up mles */
1045 if (res->owner != dlm->node_num) {
1046 ret = dlm_do_master_request(res, mle, res->owner);
1048 /* give recovery a chance to run */
1049 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1057 spin_unlock(&res->spinlock);
1059 spin_lock(&mle->spinlock);
1061 map_changed = (memcmp(mle->vote_map, mle->node_map,
1062 sizeof(mle->vote_map)) != 0);
1063 voting_done = (memcmp(mle->vote_map, mle->response_map,
1064 sizeof(mle->vote_map)) == 0);
1066 /* restart if we hit any errors */
1069 mlog(0, "%s: %.*s: node map changed, restarting\n",
1070 dlm->name, res->lockname.len, res->lockname.name);
1071 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1072 b = (mle->type == DLM_MLE_BLOCK);
1073 if ((*blocked && !b) || (!*blocked && b)) {
1074 mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1075 dlm->name, res->lockname.len, res->lockname.name,
1079 spin_unlock(&mle->spinlock);
1084 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1085 "rechecking now\n", dlm->name, res->lockname.len,
1086 res->lockname.name);
1090 mlog(0, "map not changed and voting not done "
1091 "for %s:%.*s\n", dlm->name, res->lockname.len,
1092 res->lockname.name);
1096 if (m != O2NM_MAX_NODES) {
1097 /* another node has done an assert!
1102 /* have all nodes responded? */
1103 if (voting_done && !*blocked) {
1104 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1105 if (dlm->node_num <= bit) {
1106 /* my node number is lowest.
1107 * now tell other nodes that I am
1108 * mastering this. */
1109 mle->master = dlm->node_num;
1110 /* ref was grabbed in get_lock_resource
1111 * will be dropped in dlmlock_master */
1115 /* if voting is done, but we have not received
1116 * an assert master yet, we must sleep */
1120 spin_unlock(&mle->spinlock);
1122 /* sleep if we haven't finished voting yet */
1124 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1127 if (atomic_read(&mle->mle_refs.refcount) < 2)
1128 mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1129 atomic_read(&mle->mle_refs.refcount),
1130 res->lockname.len, res->lockname.name);
1132 atomic_set(&mle->woken, 0);
1133 (void)wait_event_timeout(mle->wq,
1134 (atomic_read(&mle->woken) == 1),
1136 if (res->owner == O2NM_MAX_NODES) {
1137 mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1138 res->lockname.len, res->lockname.name);
1141 mlog(0, "done waiting, master is %u\n", res->owner);
1149 mlog(0, "about to master %.*s here, this=%u\n",
1150 res->lockname.len, res->lockname.name, m);
1151 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1153 /* This is a failure in the network path,
1154 * not in the response to the assert_master
1155 * (any nonzero response is a BUG on this node).
1156 * Most likely a socket just got disconnected
1157 * due to node death. */
1160 /* no longer need to restart lock mastery.
1161 * all living nodes have been contacted. */
1165 /* set the lockres owner */
1166 spin_lock(&res->spinlock);
1167 /* mastery reference obtained either during
1168 * assert_master_handler or in get_lock_resource */
1169 dlm_change_lockres_owner(dlm, res, m);
1170 spin_unlock(&res->spinlock);
1176 struct dlm_bitmap_diff_iter
1179 unsigned long *orig_bm;
1180 unsigned long *cur_bm;
1181 unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1184 enum dlm_node_state_change
1191 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1192 unsigned long *orig_bm,
1193 unsigned long *cur_bm)
1195 unsigned long p1, p2;
1199 iter->orig_bm = orig_bm;
1200 iter->cur_bm = cur_bm;
1202 for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1203 p1 = *(iter->orig_bm + i);
1204 p2 = *(iter->cur_bm + i);
1205 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1209 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1210 enum dlm_node_state_change *state)
1214 if (iter->curnode >= O2NM_MAX_NODES)
1217 bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1219 if (bit >= O2NM_MAX_NODES) {
1220 iter->curnode = O2NM_MAX_NODES;
1224 /* if it was there in the original then this node died */
1225 if (test_bit(bit, iter->orig_bm))
1230 iter->curnode = bit;
1235 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1236 struct dlm_lock_resource *res,
1237 struct dlm_master_list_entry *mle,
1240 struct dlm_bitmap_diff_iter bdi;
1241 enum dlm_node_state_change sc;
1245 mlog(0, "something happened such that the "
1246 "master process may need to be restarted!\n");
1248 assert_spin_locked(&mle->spinlock);
1250 dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1251 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1253 if (sc == NODE_UP) {
1254 /* a node came up. clear any old vote from
1255 * the response map and set it in the vote map
1256 * then restart the mastery. */
1257 mlog(ML_NOTICE, "node %d up while restarting\n", node);
1259 /* redo the master request, but only for the new node */
1260 mlog(0, "sending request to new node\n");
1261 clear_bit(node, mle->response_map);
1262 set_bit(node, mle->vote_map);
1264 mlog(ML_ERROR, "node down! %d\n", node);
1266 int lowest = find_next_bit(mle->maybe_map,
1269 /* act like it was never there */
1270 clear_bit(node, mle->maybe_map);
1272 if (node == lowest) {
1273 mlog(0, "expected master %u died"
1274 " while this node was blocked "
1275 "waiting on it!\n", node);
1276 lowest = find_next_bit(mle->maybe_map,
1279 if (lowest < O2NM_MAX_NODES) {
1280 mlog(0, "%s:%.*s:still "
1281 "blocked. waiting on %u "
1287 /* mle is an MLE_BLOCK, but
1288 * there is now nothing left to
1289 * block on. we need to return
1290 * all the way back out and try
1291 * again with an MLE_MASTER.
1292 * dlm_do_local_recovery_cleanup
1293 * has already run, so the mle
1295 mlog(0, "%s:%.*s: no "
1296 "longer blocking. try to "
1297 "master this here\n",
1300 res->lockname.name);
1301 mle->type = DLM_MLE_MASTER;
1307 /* now blank out everything, as if we had never
1308 * contacted anyone */
1309 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1310 memset(mle->response_map, 0, sizeof(mle->response_map));
1311 /* reset the vote_map to the current node_map */
1312 memcpy(mle->vote_map, mle->node_map,
1313 sizeof(mle->node_map));
1314 /* put myself into the maybe map */
1315 if (mle->type != DLM_MLE_BLOCK)
1316 set_bit(dlm->node_num, mle->maybe_map);
1319 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1326 * DLM_MASTER_REQUEST_MSG
1328 * returns: 0 on success,
1329 * -errno on a network error
1331 * on error, the caller should assume the target node is "dead"
1335 static int dlm_do_master_request(struct dlm_lock_resource *res,
1336 struct dlm_master_list_entry *mle, int to)
1338 struct dlm_ctxt *dlm = mle->dlm;
1339 struct dlm_master_request request;
1340 int ret, response=0, resend;
1342 memset(&request, 0, sizeof(request));
1343 request.node_idx = dlm->node_num;
1345 BUG_ON(mle->type == DLM_MLE_MIGRATION);
1347 request.namelen = (u8)mle->mnamelen;
1348 memcpy(request.name, mle->mname, request.namelen);
1351 ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1352 sizeof(request), to, &response);
1354 if (ret == -ESRCH) {
1355 /* should never happen */
1356 mlog(ML_ERROR, "TCP stack not ready!\n");
1358 } else if (ret == -EINVAL) {
1359 mlog(ML_ERROR, "bad args passed to o2net!\n");
1361 } else if (ret == -ENOMEM) {
1362 mlog(ML_ERROR, "out of memory while trying to send "
1363 "network message! retrying\n");
1364 /* this is totally crude */
1367 } else if (!dlm_is_host_down(ret)) {
1368 /* not a network error. bad. */
1370 mlog(ML_ERROR, "unhandled error!");
1373 /* all other errors should be network errors,
1374 * and likely indicate node death */
1375 mlog(ML_ERROR, "link to %d went down!\n", to);
1381 spin_lock(&mle->spinlock);
1383 case DLM_MASTER_RESP_YES:
1384 set_bit(to, mle->response_map);
1385 mlog(0, "node %u is the master, response=YES\n", to);
1386 mlog(0, "%s:%.*s: master node %u now knows I have a "
1387 "reference\n", dlm->name, res->lockname.len,
1388 res->lockname.name, to);
1391 case DLM_MASTER_RESP_NO:
1392 mlog(0, "node %u not master, response=NO\n", to);
1393 set_bit(to, mle->response_map);
1395 case DLM_MASTER_RESP_MAYBE:
1396 mlog(0, "node %u not master, response=MAYBE\n", to);
1397 set_bit(to, mle->response_map);
1398 set_bit(to, mle->maybe_map);
1400 case DLM_MASTER_RESP_ERROR:
1401 mlog(0, "node %u hit an error, resending\n", to);
1406 mlog(ML_ERROR, "bad response! %u\n", response);
1409 spin_unlock(&mle->spinlock);
1411 /* this is also totally crude */
1421 * locks that can be taken here:
1427 * if possible, TRIM THIS DOWN!!!
1429 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1432 u8 response = DLM_MASTER_RESP_MAYBE;
1433 struct dlm_ctxt *dlm = data;
1434 struct dlm_lock_resource *res = NULL;
1435 struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1436 struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1438 unsigned int namelen, hash;
1441 int dispatch_assert = 0;
1445 return DLM_MASTER_RESP_NO;
1447 if (!dlm_domain_fully_joined(dlm)) {
1448 response = DLM_MASTER_RESP_NO;
1452 name = request->name;
1453 namelen = request->namelen;
1454 hash = dlm_lockid_hash(name, namelen);
1456 if (namelen > DLM_LOCKID_NAME_MAX) {
1457 response = DLM_IVBUFLEN;
1462 spin_lock(&dlm->spinlock);
1463 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1465 spin_unlock(&dlm->spinlock);
1467 /* take care of the easy cases up front */
1468 spin_lock(&res->spinlock);
1471 * Right after dlm spinlock was released, dlm_thread could have
1472 * purged the lockres. Check if lockres got unhashed. If so
1475 if (hlist_unhashed(&res->hash_node)) {
1476 spin_unlock(&res->spinlock);
1477 dlm_lockres_put(res);
1481 if (res->state & (DLM_LOCK_RES_RECOVERING|
1482 DLM_LOCK_RES_MIGRATING)) {
1483 spin_unlock(&res->spinlock);
1484 mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1485 "being recovered/migrated\n");
1486 response = DLM_MASTER_RESP_ERROR;
1488 kmem_cache_free(dlm_mle_cache, mle);
1492 if (res->owner == dlm->node_num) {
1493 dlm_lockres_set_refmap_bit(dlm, res, request->node_idx);
1494 spin_unlock(&res->spinlock);
1495 response = DLM_MASTER_RESP_YES;
1497 kmem_cache_free(dlm_mle_cache, mle);
1499 /* this node is the owner.
1500 * there is some extra work that needs to
1501 * happen now. the requesting node has
1502 * caused all nodes up to this one to
1503 * create mles. this node now needs to
1504 * go back and clean those up. */
1505 dispatch_assert = 1;
1507 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1508 spin_unlock(&res->spinlock);
1509 // mlog(0, "node %u is the master\n", res->owner);
1510 response = DLM_MASTER_RESP_NO;
1512 kmem_cache_free(dlm_mle_cache, mle);
1516 /* ok, there is no owner. either this node is
1517 * being blocked, or it is actively trying to
1518 * master this lock. */
1519 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1520 mlog(ML_ERROR, "lock with no owner should be "
1525 // mlog(0, "lockres is in progress...\n");
1526 spin_lock(&dlm->master_lock);
1527 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1529 mlog(ML_ERROR, "no mle found for this lock!\n");
1533 spin_lock(&tmpmle->spinlock);
1534 if (tmpmle->type == DLM_MLE_BLOCK) {
1535 // mlog(0, "this node is waiting for "
1536 // "lockres to be mastered\n");
1537 response = DLM_MASTER_RESP_NO;
1538 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1539 mlog(0, "node %u is master, but trying to migrate to "
1540 "node %u.\n", tmpmle->master, tmpmle->new_master);
1541 if (tmpmle->master == dlm->node_num) {
1542 mlog(ML_ERROR, "no owner on lockres, but this "
1543 "node is trying to migrate it to %u?!\n",
1544 tmpmle->new_master);
1547 /* the real master can respond on its own */
1548 response = DLM_MASTER_RESP_NO;
1550 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1552 if (tmpmle->master == dlm->node_num) {
1553 response = DLM_MASTER_RESP_YES;
1554 /* this node will be the owner.
1555 * go back and clean the mles on any
1557 dispatch_assert = 1;
1558 dlm_lockres_set_refmap_bit(dlm, res,
1561 response = DLM_MASTER_RESP_NO;
1563 // mlog(0, "this node is attempting to "
1564 // "master lockres\n");
1565 response = DLM_MASTER_RESP_MAYBE;
1568 set_bit(request->node_idx, tmpmle->maybe_map);
1569 spin_unlock(&tmpmle->spinlock);
1571 spin_unlock(&dlm->master_lock);
1572 spin_unlock(&res->spinlock);
1574 /* keep the mle attached to heartbeat events */
1575 dlm_put_mle(tmpmle);
1577 kmem_cache_free(dlm_mle_cache, mle);
1582 * lockres doesn't exist on this node
1583 * if there is an MLE_BLOCK, return NO
1584 * if there is an MLE_MASTER, return MAYBE
1585 * otherwise, add an MLE_BLOCK, return NO
1587 spin_lock(&dlm->master_lock);
1588 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1590 /* this lockid has never been seen on this node yet */
1591 // mlog(0, "no mle found\n");
1593 spin_unlock(&dlm->master_lock);
1594 spin_unlock(&dlm->spinlock);
1596 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1598 response = DLM_MASTER_RESP_ERROR;
1599 mlog_errno(-ENOMEM);
1605 // mlog(0, "this is second time thru, already allocated, "
1606 // "add the block.\n");
1607 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1608 set_bit(request->node_idx, mle->maybe_map);
1609 __dlm_insert_mle(dlm, mle);
1610 response = DLM_MASTER_RESP_NO;
1612 // mlog(0, "mle was found\n");
1614 spin_lock(&tmpmle->spinlock);
1615 if (tmpmle->master == dlm->node_num) {
1616 mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1619 if (tmpmle->type == DLM_MLE_BLOCK)
1620 response = DLM_MASTER_RESP_NO;
1621 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1622 mlog(0, "migration mle was found (%u->%u)\n",
1623 tmpmle->master, tmpmle->new_master);
1624 /* real master can respond on its own */
1625 response = DLM_MASTER_RESP_NO;
1627 response = DLM_MASTER_RESP_MAYBE;
1629 set_bit(request->node_idx, tmpmle->maybe_map);
1630 spin_unlock(&tmpmle->spinlock);
1632 spin_unlock(&dlm->master_lock);
1633 spin_unlock(&dlm->spinlock);
1636 /* keep the mle attached to heartbeat events */
1637 dlm_put_mle(tmpmle);
1641 * __dlm_lookup_lockres() grabbed a reference to this lockres.
1642 * The reference is released by dlm_assert_master_worker() under
1643 * the call to dlm_dispatch_assert_master(). If
1644 * dlm_assert_master_worker() isn't called, we drop it here.
1646 if (dispatch_assert) {
1647 if (response != DLM_MASTER_RESP_YES)
1648 mlog(ML_ERROR, "invalid response %d\n", response);
1650 mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1653 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1654 dlm->node_num, res->lockname.len, res->lockname.name);
1655 spin_lock(&res->spinlock);
1656 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1657 DLM_ASSERT_MASTER_MLE_CLEANUP);
1659 mlog(ML_ERROR, "failed to dispatch assert master work\n");
1660 response = DLM_MASTER_RESP_ERROR;
1661 dlm_lockres_put(res);
1664 __dlm_lockres_grab_inflight_worker(dlm, res);
1666 spin_unlock(&res->spinlock);
1669 dlm_lockres_put(res);
1678 * DLM_ASSERT_MASTER_MSG
1683 * NOTE: this can be used for debugging
1684 * can periodically run all locks owned by this node
1685 * and re-assert across the cluster...
1687 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1688 struct dlm_lock_resource *res,
1689 void *nodemap, u32 flags)
1691 struct dlm_assert_master assert;
1693 struct dlm_node_iter iter;
1696 const char *lockname = res->lockname.name;
1697 unsigned int namelen = res->lockname.len;
1699 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1701 spin_lock(&res->spinlock);
1702 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1703 spin_unlock(&res->spinlock);
1708 /* note that if this nodemap is empty, it returns 0 */
1709 dlm_node_iter_init(nodemap, &iter);
1710 while ((to = dlm_node_iter_next(&iter)) >= 0) {
1712 struct dlm_master_list_entry *mle = NULL;
1714 mlog(0, "sending assert master to %d (%.*s)\n", to,
1716 memset(&assert, 0, sizeof(assert));
1717 assert.node_idx = dlm->node_num;
1718 assert.namelen = namelen;
1719 memcpy(assert.name, lockname, namelen);
1720 assert.flags = cpu_to_be32(flags);
1722 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1723 &assert, sizeof(assert), to, &r);
1725 mlog(ML_ERROR, "Error %d when sending message %u (key "
1726 "0x%x) to node %u\n", tmpret,
1727 DLM_ASSERT_MASTER_MSG, dlm->key, to);
1728 if (!dlm_is_host_down(tmpret)) {
1729 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1732 /* a node died. finish out the rest of the nodes. */
1733 mlog(0, "link to %d went down!\n", to);
1734 /* any nonzero status return will do */
1738 /* ok, something horribly messed. kill thyself. */
1739 mlog(ML_ERROR,"during assert master of %.*s to %u, "
1740 "got %d.\n", namelen, lockname, to, r);
1741 spin_lock(&dlm->spinlock);
1742 spin_lock(&dlm->master_lock);
1743 if (dlm_find_mle(dlm, &mle, (char *)lockname,
1745 dlm_print_one_mle(mle);
1748 spin_unlock(&dlm->master_lock);
1749 spin_unlock(&dlm->spinlock);
1753 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1754 !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1755 mlog(ML_ERROR, "%.*s: very strange, "
1756 "master MLE but no lockres on %u\n",
1757 namelen, lockname, to);
1760 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1761 mlog(0, "%.*s: node %u create mles on other "
1762 "nodes and requests a re-assert\n",
1763 namelen, lockname, to);
1766 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1767 mlog(0, "%.*s: node %u has a reference to this "
1768 "lockres, set the bit in the refmap\n",
1769 namelen, lockname, to);
1770 spin_lock(&res->spinlock);
1771 dlm_lockres_set_refmap_bit(dlm, res, to);
1772 spin_unlock(&res->spinlock);
1779 spin_lock(&res->spinlock);
1780 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1781 spin_unlock(&res->spinlock);
1788 * locks that can be taken here:
1794 * if possible, TRIM THIS DOWN!!!
1796 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1799 struct dlm_ctxt *dlm = data;
1800 struct dlm_master_list_entry *mle = NULL;
1801 struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1802 struct dlm_lock_resource *res = NULL;
1804 unsigned int namelen, hash;
1806 int master_request = 0, have_lockres_ref = 0;
1812 name = assert->name;
1813 namelen = assert->namelen;
1814 hash = dlm_lockid_hash(name, namelen);
1815 flags = be32_to_cpu(assert->flags);
1817 if (namelen > DLM_LOCKID_NAME_MAX) {
1818 mlog(ML_ERROR, "Invalid name length!");
1822 spin_lock(&dlm->spinlock);
1825 mlog(0, "assert_master with flags: %u\n", flags);
1828 spin_lock(&dlm->master_lock);
1829 if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1830 /* not an error, could be master just re-asserting */
1831 mlog(0, "just got an assert_master from %u, but no "
1832 "MLE for it! (%.*s)\n", assert->node_idx,
1835 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1836 if (bit >= O2NM_MAX_NODES) {
1837 /* not necessarily an error, though less likely.
1838 * could be master just re-asserting. */
1839 mlog(0, "no bits set in the maybe_map, but %u "
1840 "is asserting! (%.*s)\n", assert->node_idx,
1842 } else if (bit != assert->node_idx) {
1843 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1844 mlog(0, "master %u was found, %u should "
1845 "back off\n", assert->node_idx, bit);
1847 /* with the fix for bug 569, a higher node
1848 * number winning the mastery will respond
1849 * YES to mastery requests, but this node
1850 * had no way of knowing. let it pass. */
1851 mlog(0, "%u is the lowest node, "
1852 "%u is asserting. (%.*s) %u must "
1853 "have begun after %u won.\n", bit,
1854 assert->node_idx, namelen, name, bit,
1858 if (mle->type == DLM_MLE_MIGRATION) {
1859 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1860 mlog(0, "%s:%.*s: got cleanup assert"
1861 " from %u for migration\n",
1862 dlm->name, namelen, name,
1864 } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1865 mlog(0, "%s:%.*s: got unrelated assert"
1866 " from %u for migration, ignoring\n",
1867 dlm->name, namelen, name,
1870 spin_unlock(&dlm->master_lock);
1871 spin_unlock(&dlm->spinlock);
1876 spin_unlock(&dlm->master_lock);
1878 /* ok everything checks out with the MLE
1879 * now check to see if there is a lockres */
1880 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1882 spin_lock(&res->spinlock);
1883 if (res->state & DLM_LOCK_RES_RECOVERING) {
1884 mlog(ML_ERROR, "%u asserting but %.*s is "
1885 "RECOVERING!\n", assert->node_idx, namelen, name);
1889 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1890 res->owner != assert->node_idx) {
1891 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1892 "but current owner is %u! (%.*s)\n",
1893 assert->node_idx, res->owner, namelen,
1895 __dlm_print_one_lock_resource(res);
1898 } else if (mle->type != DLM_MLE_MIGRATION) {
1899 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1900 /* owner is just re-asserting */
1901 if (res->owner == assert->node_idx) {
1902 mlog(0, "owner %u re-asserting on "
1903 "lock %.*s\n", assert->node_idx,
1907 mlog(ML_ERROR, "got assert_master from "
1908 "node %u, but %u is the owner! "
1909 "(%.*s)\n", assert->node_idx,
1910 res->owner, namelen, name);
1913 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1914 mlog(ML_ERROR, "got assert from %u, but lock "
1915 "with no owner should be "
1916 "in-progress! (%.*s)\n",
1921 } else /* mle->type == DLM_MLE_MIGRATION */ {
1922 /* should only be getting an assert from new master */
1923 if (assert->node_idx != mle->new_master) {
1924 mlog(ML_ERROR, "got assert from %u, but "
1925 "new master is %u, and old master "
1927 assert->node_idx, mle->new_master,
1928 mle->master, namelen, name);
1934 spin_unlock(&res->spinlock);
1937 // mlog(0, "woo! got an assert_master from node %u!\n",
1938 // assert->node_idx);
1944 spin_lock(&mle->spinlock);
1945 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1948 /* MASTER mle: if any bits set in the response map
1949 * then the calling node needs to re-assert to clear
1950 * up nodes that this node contacted */
1951 while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1952 nn+1)) < O2NM_MAX_NODES) {
1953 if (nn != dlm->node_num && nn != assert->node_idx) {
1959 mle->master = assert->node_idx;
1960 atomic_set(&mle->woken, 1);
1962 spin_unlock(&mle->spinlock);
1966 spin_lock(&res->spinlock);
1967 if (mle->type == DLM_MLE_MIGRATION) {
1968 mlog(0, "finishing off migration of lockres %.*s, "
1970 res->lockname.len, res->lockname.name,
1971 dlm->node_num, mle->new_master);
1972 res->state &= ~DLM_LOCK_RES_MIGRATING;
1974 dlm_change_lockres_owner(dlm, res, mle->new_master);
1975 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1977 dlm_change_lockres_owner(dlm, res, mle->master);
1979 spin_unlock(&res->spinlock);
1980 have_lockres_ref = 1;
1985 /* master is known, detach if not already detached.
1986 * ensures that only one assert_master call will happen
1988 spin_lock(&dlm->master_lock);
1990 rr = atomic_read(&mle->mle_refs.refcount);
1991 if (mle->inuse > 0) {
1992 if (extra_ref && rr < 3)
1994 else if (!extra_ref && rr < 2)
1997 if (extra_ref && rr < 2)
1999 else if (!extra_ref && rr < 1)
2003 mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
2004 "that will mess up this node, refs=%d, extra=%d, "
2005 "inuse=%d\n", dlm->name, namelen, name,
2006 assert->node_idx, rr, extra_ref, mle->inuse);
2007 dlm_print_one_mle(mle);
2009 __dlm_unlink_mle(dlm, mle);
2010 __dlm_mle_detach_hb_events(dlm, mle);
2013 /* the assert master message now balances the extra
2014 * ref given by the master / migration request message.
2015 * if this is the last put, it will be removed
2019 spin_unlock(&dlm->master_lock);
2021 if (res->owner != assert->node_idx) {
2022 mlog(0, "assert_master from %u, but current "
2023 "owner is %u (%.*s), no mle\n", assert->node_idx,
2024 res->owner, namelen, name);
2027 spin_unlock(&dlm->spinlock);
2032 spin_lock(&res->spinlock);
2033 res->state |= DLM_LOCK_RES_SETREF_INPROG;
2034 spin_unlock(&res->spinlock);
2035 *ret_data = (void *)res;
2038 if (master_request) {
2039 mlog(0, "need to tell master to reassert\n");
2040 /* positive. negative would shoot down the node. */
2041 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2042 if (!have_lockres_ref) {
2043 mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2044 "mle present here for %s:%.*s, but no lockres!\n",
2045 assert->node_idx, dlm->name, namelen, name);
2048 if (have_lockres_ref) {
2049 /* let the master know we have a reference to the lockres */
2050 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2051 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2052 dlm->name, namelen, name, assert->node_idx);
2057 /* kill the caller! */
2058 mlog(ML_ERROR, "Bad message received from another node. Dumping state "
2059 "and killing the other node now! This node is OK and can continue.\n");
2060 __dlm_print_one_lock_resource(res);
2061 spin_unlock(&res->spinlock);
2062 spin_lock(&dlm->master_lock);
2065 spin_unlock(&dlm->master_lock);
2066 spin_unlock(&dlm->spinlock);
2067 *ret_data = (void *)res;
2072 void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2074 struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2077 spin_lock(&res->spinlock);
2078 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2079 spin_unlock(&res->spinlock);
2081 dlm_lockres_put(res);
2086 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2087 struct dlm_lock_resource *res,
2088 int ignore_higher, u8 request_from, u32 flags)
2090 struct dlm_work_item *item;
2091 item = kzalloc(sizeof(*item), GFP_ATOMIC);
2096 /* queue up work for dlm_assert_master_worker */
2097 dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2098 item->u.am.lockres = res; /* already have a ref */
2099 /* can optionally ignore node numbers higher than this node */
2100 item->u.am.ignore_higher = ignore_higher;
2101 item->u.am.request_from = request_from;
2102 item->u.am.flags = flags;
2105 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2106 res->lockname.name);
2108 spin_lock(&dlm->work_lock);
2109 list_add_tail(&item->list, &dlm->work_list);
2110 spin_unlock(&dlm->work_lock);
2112 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2116 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2118 struct dlm_ctxt *dlm = data;
2120 struct dlm_lock_resource *res;
2121 unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2128 res = item->u.am.lockres;
2129 ignore_higher = item->u.am.ignore_higher;
2130 request_from = item->u.am.request_from;
2131 flags = item->u.am.flags;
2133 spin_lock(&dlm->spinlock);
2134 memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2135 spin_unlock(&dlm->spinlock);
2137 clear_bit(dlm->node_num, nodemap);
2138 if (ignore_higher) {
2139 /* if is this just to clear up mles for nodes below
2140 * this node, do not send the message to the original
2141 * caller or any node number higher than this */
2142 clear_bit(request_from, nodemap);
2143 bit = dlm->node_num;
2145 bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2147 if (bit >= O2NM_MAX_NODES)
2149 clear_bit(bit, nodemap);
2154 * If we're migrating this lock to someone else, we are no
2155 * longer allowed to assert out own mastery. OTOH, we need to
2156 * prevent migration from starting while we're still asserting
2157 * our dominance. The reserved ast delays migration.
2159 spin_lock(&res->spinlock);
2160 if (res->state & DLM_LOCK_RES_MIGRATING) {
2161 mlog(0, "Someone asked us to assert mastery, but we're "
2162 "in the middle of migration. Skipping assert, "
2163 "the new master will handle that.\n");
2164 spin_unlock(&res->spinlock);
2167 __dlm_lockres_reserve_ast(res);
2168 spin_unlock(&res->spinlock);
2170 /* this call now finishes out the nodemap
2171 * even if one or more nodes die */
2172 mlog(0, "worker about to master %.*s here, this=%u\n",
2173 res->lockname.len, res->lockname.name, dlm->node_num);
2174 ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2176 /* no need to restart, we are done */
2177 if (!dlm_is_host_down(ret))
2181 /* Ok, we've asserted ourselves. Let's let migration start. */
2182 dlm_lockres_release_ast(dlm, res);
2185 dlm_lockres_drop_inflight_worker(dlm, res);
2187 dlm_lockres_put(res);
2189 mlog(0, "finished with dlm_assert_master_worker\n");
2192 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2193 * We cannot wait for node recovery to complete to begin mastering this
2194 * lockres because this lockres is used to kick off recovery! ;-)
2195 * So, do a pre-check on all living nodes to see if any of those nodes
2196 * think that $RECOVERY is currently mastered by a dead node. If so,
2197 * we wait a short time to allow that node to get notified by its own
2198 * heartbeat stack, then check again. All $RECOVERY lock resources
2199 * mastered by dead nodes are purged when the hearbeat callback is
2200 * fired, so we can know for sure that it is safe to continue once
2201 * the node returns a live node or no node. */
2202 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2203 struct dlm_lock_resource *res)
2205 struct dlm_node_iter iter;
2208 u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2210 spin_lock(&dlm->spinlock);
2211 dlm_node_iter_init(dlm->domain_map, &iter);
2212 spin_unlock(&dlm->spinlock);
2214 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2215 /* do not send to self */
2216 if (nodenum == dlm->node_num)
2218 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2221 if (!dlm_is_host_down(ret))
2223 /* host is down, so answer for that node would be
2224 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
2228 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2229 /* check to see if this master is in the recovery map */
2230 spin_lock(&dlm->spinlock);
2231 if (test_bit(master, dlm->recovery_map)) {
2232 mlog(ML_NOTICE, "%s: node %u has not seen "
2233 "node %u go down yet, and thinks the "
2234 "dead node is mastering the recovery "
2235 "lock. must wait.\n", dlm->name,
2239 spin_unlock(&dlm->spinlock);
2240 mlog(0, "%s: reco lock master is %u\n", dlm->name,
2249 * DLM_DEREF_LOCKRES_MSG
2252 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2254 struct dlm_deref_lockres deref;
2256 const char *lockname;
2257 unsigned int namelen;
2259 lockname = res->lockname.name;
2260 namelen = res->lockname.len;
2261 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2263 memset(&deref, 0, sizeof(deref));
2264 deref.node_idx = dlm->node_num;
2265 deref.namelen = namelen;
2266 memcpy(deref.name, lockname, namelen);
2268 ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2269 &deref, sizeof(deref), res->owner, &r);
2271 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n",
2272 dlm->name, namelen, lockname, ret, res->owner);
2274 /* BAD. other node says I did not have a ref. */
2275 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2276 dlm->name, namelen, lockname, res->owner, r);
2277 dlm_print_one_lock_resource(res);
2283 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2286 struct dlm_ctxt *dlm = data;
2287 struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2288 struct dlm_lock_resource *res = NULL;
2290 unsigned int namelen;
2294 struct dlm_work_item *item;
2302 namelen = deref->namelen;
2303 node = deref->node_idx;
2305 if (namelen > DLM_LOCKID_NAME_MAX) {
2306 mlog(ML_ERROR, "Invalid name length!");
2309 if (deref->node_idx >= O2NM_MAX_NODES) {
2310 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2314 hash = dlm_lockid_hash(name, namelen);
2316 spin_lock(&dlm->spinlock);
2317 res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2319 spin_unlock(&dlm->spinlock);
2320 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2321 dlm->name, namelen, name);
2324 spin_unlock(&dlm->spinlock);
2326 spin_lock(&res->spinlock);
2327 if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2330 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2331 if (test_bit(node, res->refmap)) {
2332 dlm_lockres_clear_refmap_bit(dlm, res, node);
2336 spin_unlock(&res->spinlock);
2340 dlm_lockres_calc_usage(dlm, res);
2342 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2343 "but it is already dropped!\n", dlm->name,
2344 res->lockname.len, res->lockname.name, node);
2345 dlm_print_one_lock_resource(res);
2351 item = kzalloc(sizeof(*item), GFP_NOFS);
2358 dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2359 item->u.dl.deref_res = res;
2360 item->u.dl.deref_node = node;
2362 spin_lock(&dlm->work_lock);
2363 list_add_tail(&item->list, &dlm->work_list);
2364 spin_unlock(&dlm->work_lock);
2366 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2371 dlm_lockres_put(res);
2377 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2379 struct dlm_ctxt *dlm;
2380 struct dlm_lock_resource *res;
2385 res = item->u.dl.deref_res;
2386 node = item->u.dl.deref_node;
2388 spin_lock(&res->spinlock);
2389 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2390 if (test_bit(node, res->refmap)) {
2391 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2392 dlm_lockres_clear_refmap_bit(dlm, res, node);
2395 spin_unlock(&res->spinlock);
2398 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2399 dlm->name, res->lockname.len, res->lockname.name, node);
2400 dlm_lockres_calc_usage(dlm, res);
2402 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2403 "but it is already dropped!\n", dlm->name,
2404 res->lockname.len, res->lockname.name, node);
2405 dlm_print_one_lock_resource(res);
2408 dlm_lockres_put(res);
2412 * A migrateable resource is one that is :
2413 * 1. locally mastered, and,
2414 * 2. zero local locks, and,
2415 * 3. one or more non-local locks, or, one or more references
2416 * Returns 1 if yes, 0 if not.
2418 static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
2419 struct dlm_lock_resource *res)
2421 enum dlm_lockres_list idx;
2422 int nonlocal = 0, node_ref;
2423 struct list_head *queue;
2424 struct dlm_lock *lock;
2427 assert_spin_locked(&res->spinlock);
2429 /* delay migration when the lockres is in MIGRATING state */
2430 if (res->state & DLM_LOCK_RES_MIGRATING)
2433 /* delay migration when the lockres is in RECOCERING state */
2434 if (res->state & DLM_LOCK_RES_RECOVERING)
2437 if (res->owner != dlm->node_num)
2440 for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2441 queue = dlm_list_idx_to_ptr(res, idx);
2442 list_for_each_entry(lock, queue, list) {
2443 if (lock->ml.node != dlm->node_num) {
2447 cookie = be64_to_cpu(lock->ml.cookie);
2448 mlog(0, "%s: Not migrateable res %.*s, lock %u:%llu on "
2449 "%s list\n", dlm->name, res->lockname.len,
2451 dlm_get_lock_cookie_node(cookie),
2452 dlm_get_lock_cookie_seq(cookie),
2453 dlm_list_in_text(idx));
2459 node_ref = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
2460 if (node_ref >= O2NM_MAX_NODES)
2464 mlog(0, "%s: res %.*s, Migrateable\n", dlm->name, res->lockname.len,
2465 res->lockname.name);
2471 * DLM_MIGRATE_LOCKRES
2475 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2476 struct dlm_lock_resource *res, u8 target)
2478 struct dlm_master_list_entry *mle = NULL;
2479 struct dlm_master_list_entry *oldmle = NULL;
2480 struct dlm_migratable_lockres *mres = NULL;
2483 unsigned int namelen;
2490 BUG_ON(target == O2NM_MAX_NODES);
2492 name = res->lockname.name;
2493 namelen = res->lockname.len;
2495 mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name,
2498 /* preallocate up front. if this fails, abort */
2500 mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2506 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2514 * clear any existing master requests and
2515 * add the migration mle to the list
2517 spin_lock(&dlm->spinlock);
2518 spin_lock(&dlm->master_lock);
2519 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2520 namelen, target, dlm->node_num);
2521 spin_unlock(&dlm->master_lock);
2522 spin_unlock(&dlm->spinlock);
2524 if (ret == -EEXIST) {
2525 mlog(0, "another process is already migrating it\n");
2531 * set the MIGRATING flag and flush asts
2532 * if we fail after this we need to re-dirty the lockres
2534 if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2535 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2536 "the target went down.\n", res->lockname.len,
2537 res->lockname.name, target);
2538 spin_lock(&res->spinlock);
2539 res->state &= ~DLM_LOCK_RES_MIGRATING;
2541 spin_unlock(&res->spinlock);
2547 /* master is known, detach if not already detached */
2548 dlm_mle_detach_hb_events(dlm, oldmle);
2549 dlm_put_mle(oldmle);
2554 dlm_mle_detach_hb_events(dlm, mle);
2557 kmem_cache_free(dlm_mle_cache, mle);
2564 * at this point, we have a migration target, an mle
2565 * in the master list, and the MIGRATING flag set on
2569 /* now that remote nodes are spinning on the MIGRATING flag,
2570 * ensure that all assert_master work is flushed. */
2571 flush_workqueue(dlm->dlm_worker);
2573 /* get an extra reference on the mle.
2574 * otherwise the assert_master from the new
2575 * master will destroy this.
2576 * also, make sure that all callers of dlm_get_mle
2577 * take both dlm->spinlock and dlm->master_lock */
2578 spin_lock(&dlm->spinlock);
2579 spin_lock(&dlm->master_lock);
2580 dlm_get_mle_inuse(mle);
2581 spin_unlock(&dlm->master_lock);
2582 spin_unlock(&dlm->spinlock);
2584 /* notify new node and send all lock state */
2585 /* call send_one_lockres with migration flag.
2586 * this serves as notice to the target node that a
2587 * migration is starting. */
2588 ret = dlm_send_one_lockres(dlm, res, mres, target,
2589 DLM_MRES_MIGRATION);
2592 mlog(0, "migration to node %u failed with %d\n",
2594 /* migration failed, detach and clean up mle */
2595 dlm_mle_detach_hb_events(dlm, mle);
2597 dlm_put_mle_inuse(mle);
2598 spin_lock(&res->spinlock);
2599 res->state &= ~DLM_LOCK_RES_MIGRATING;
2601 spin_unlock(&res->spinlock);
2602 if (dlm_is_host_down(ret))
2603 dlm_wait_for_node_death(dlm, target,
2604 DLM_NODE_DEATH_WAIT_MAX);
2608 /* at this point, the target sends a message to all nodes,
2609 * (using dlm_do_migrate_request). this node is skipped since
2610 * we had to put an mle in the list to begin the process. this
2611 * node now waits for target to do an assert master. this node
2612 * will be the last one notified, ensuring that the migration
2613 * is complete everywhere. if the target dies while this is
2614 * going on, some nodes could potentially see the target as the
2615 * master, so it is important that my recovery finds the migration
2616 * mle and sets the master to UNKNOWN. */
2619 /* wait for new node to assert master */
2621 ret = wait_event_interruptible_timeout(mle->wq,
2622 (atomic_read(&mle->woken) == 1),
2623 msecs_to_jiffies(5000));
2626 if (atomic_read(&mle->woken) == 1 ||
2627 res->owner == target)
2630 mlog(0, "%s:%.*s: timed out during migration\n",
2631 dlm->name, res->lockname.len, res->lockname.name);
2632 /* avoid hang during shutdown when migrating lockres
2633 * to a node which also goes down */
2634 if (dlm_is_node_dead(dlm, target)) {
2635 mlog(0, "%s:%.*s: expected migration "
2636 "target %u is no longer up, restarting\n",
2637 dlm->name, res->lockname.len,
2638 res->lockname.name, target);
2640 /* migration failed, detach and clean up mle */
2641 dlm_mle_detach_hb_events(dlm, mle);
2643 dlm_put_mle_inuse(mle);
2644 spin_lock(&res->spinlock);
2645 res->state &= ~DLM_LOCK_RES_MIGRATING;
2647 spin_unlock(&res->spinlock);
2651 mlog(0, "%s:%.*s: caught signal during migration\n",
2652 dlm->name, res->lockname.len, res->lockname.name);
2655 /* all done, set the owner, clear the flag */
2656 spin_lock(&res->spinlock);
2657 dlm_set_lockres_owner(dlm, res, target);
2658 res->state &= ~DLM_LOCK_RES_MIGRATING;
2659 dlm_remove_nonlocal_locks(dlm, res);
2660 spin_unlock(&res->spinlock);
2663 /* master is known, detach if not already detached */
2664 dlm_mle_detach_hb_events(dlm, mle);
2665 dlm_put_mle_inuse(mle);
2668 dlm_lockres_calc_usage(dlm, res);
2671 /* re-dirty the lockres if we failed */
2673 dlm_kick_thread(dlm, res);
2675 /* wake up waiters if the MIGRATING flag got set
2676 * but migration failed */
2681 free_page((unsigned long)mres);
2685 mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen,
2690 #define DLM_MIGRATION_RETRY_MS 100
2693 * Should be called only after beginning the domain leave process.
2694 * There should not be any remaining locks on nonlocal lock resources,
2695 * and there should be no local locks left on locally mastered resources.
2697 * Called with the dlm spinlock held, may drop it to do migration, but
2698 * will re-acquire before exit.
2700 * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2702 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2705 int lock_dropped = 0;
2706 u8 target = O2NM_MAX_NODES;
2708 assert_spin_locked(&dlm->spinlock);
2710 spin_lock(&res->spinlock);
2711 if (dlm_is_lockres_migrateable(dlm, res))
2712 target = dlm_pick_migration_target(dlm, res);
2713 spin_unlock(&res->spinlock);
2715 if (target == O2NM_MAX_NODES)
2718 /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2719 spin_unlock(&dlm->spinlock);
2721 ret = dlm_migrate_lockres(dlm, res, target);
2723 mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2724 dlm->name, res->lockname.len, res->lockname.name,
2726 spin_lock(&dlm->spinlock);
2728 return lock_dropped;
2731 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2734 spin_lock(&dlm->ast_lock);
2735 spin_lock(&lock->spinlock);
2736 ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2737 spin_unlock(&lock->spinlock);
2738 spin_unlock(&dlm->ast_lock);
2742 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2743 struct dlm_lock_resource *res,
2747 spin_lock(&res->spinlock);
2748 can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2749 spin_unlock(&res->spinlock);
2751 /* target has died, so make the caller break out of the
2752 * wait_event, but caller must recheck the domain_map */
2753 spin_lock(&dlm->spinlock);
2754 if (!test_bit(mig_target, dlm->domain_map))
2756 spin_unlock(&dlm->spinlock);
2760 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2761 struct dlm_lock_resource *res)
2764 spin_lock(&res->spinlock);
2765 ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2766 spin_unlock(&res->spinlock);
2771 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2772 struct dlm_lock_resource *res,
2777 mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2778 res->lockname.len, res->lockname.name, dlm->node_num,
2780 /* need to set MIGRATING flag on lockres. this is done by
2781 * ensuring that all asts have been flushed for this lockres. */
2782 spin_lock(&res->spinlock);
2783 BUG_ON(res->migration_pending);
2784 res->migration_pending = 1;
2785 /* strategy is to reserve an extra ast then release
2786 * it below, letting the release do all of the work */
2787 __dlm_lockres_reserve_ast(res);
2788 spin_unlock(&res->spinlock);
2790 /* now flush all the pending asts */
2791 dlm_kick_thread(dlm, res);
2792 /* before waiting on DIRTY, block processes which may
2793 * try to dirty the lockres before MIGRATING is set */
2794 spin_lock(&res->spinlock);
2795 BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2796 res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2797 spin_unlock(&res->spinlock);
2798 /* now wait on any pending asts and the DIRTY state */
2799 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2800 dlm_lockres_release_ast(dlm, res);
2802 mlog(0, "about to wait on migration_wq, dirty=%s\n",
2803 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2804 /* if the extra ref we just put was the final one, this
2805 * will pass thru immediately. otherwise, we need to wait
2806 * for the last ast to finish. */
2808 ret = wait_event_interruptible_timeout(dlm->migration_wq,
2809 dlm_migration_can_proceed(dlm, res, target),
2810 msecs_to_jiffies(1000));
2812 mlog(0, "woken again: migrating? %s, dead? %s\n",
2813 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2814 test_bit(target, dlm->domain_map) ? "no":"yes");
2816 mlog(0, "all is well: migrating? %s, dead? %s\n",
2817 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2818 test_bit(target, dlm->domain_map) ? "no":"yes");
2820 if (!dlm_migration_can_proceed(dlm, res, target)) {
2821 mlog(0, "trying again...\n");
2826 /* did the target go down or die? */
2827 spin_lock(&dlm->spinlock);
2828 if (!test_bit(target, dlm->domain_map)) {
2829 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2833 spin_unlock(&dlm->spinlock);
2836 * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2837 * another try; otherwise, we are sure the MIGRATING state is there,
2838 * drop the unneded state which blocked threads trying to DIRTY
2840 spin_lock(&res->spinlock);
2841 BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2842 res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2844 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2845 spin_unlock(&res->spinlock);
2850 * o the DLM_LOCK_RES_MIGRATING flag is set if target not down
2851 * o there are no pending asts on this lockres
2852 * o all processes trying to reserve an ast on this
2853 * lockres must wait for the MIGRATING flag to clear
2858 /* last step in the migration process.
2859 * original master calls this to free all of the dlm_lock
2860 * structures that used to be for other nodes. */
2861 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2862 struct dlm_lock_resource *res)
2864 struct list_head *queue = &res->granted;
2866 struct dlm_lock *lock, *next;
2868 assert_spin_locked(&res->spinlock);
2870 BUG_ON(res->owner == dlm->node_num);
2872 for (i=0; i<3; i++) {
2873 list_for_each_entry_safe(lock, next, queue, list) {
2874 if (lock->ml.node != dlm->node_num) {
2875 mlog(0, "putting lock for node %u\n",
2877 /* be extra careful */
2878 BUG_ON(!list_empty(&lock->ast_list));
2879 BUG_ON(!list_empty(&lock->bast_list));
2880 BUG_ON(lock->ast_pending);
2881 BUG_ON(lock->bast_pending);
2882 dlm_lockres_clear_refmap_bit(dlm, res,
2884 list_del_init(&lock->list);
2886 /* In a normal unlock, we would have added a
2887 * DLM_UNLOCK_FREE_LOCK action. Force it. */
2895 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2896 if (bit >= O2NM_MAX_NODES)
2898 /* do not clear the local node reference, if there is a
2899 * process holding this, let it drop the ref itself */
2900 if (bit != dlm->node_num) {
2901 mlog(0, "%s:%.*s: node %u had a ref to this "
2902 "migrating lockres, clearing\n", dlm->name,
2903 res->lockname.len, res->lockname.name, bit);
2904 dlm_lockres_clear_refmap_bit(dlm, res, bit);
2911 * Pick a node to migrate the lock resource to. This function selects a
2912 * potential target based first on the locks and then on refmap. It skips
2913 * nodes that are in the process of exiting the domain.
2915 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2916 struct dlm_lock_resource *res)
2918 enum dlm_lockres_list idx;
2919 struct list_head *queue = &res->granted;
2920 struct dlm_lock *lock;
2922 u8 nodenum = O2NM_MAX_NODES;
2924 assert_spin_locked(&dlm->spinlock);
2925 assert_spin_locked(&res->spinlock);
2927 /* Go through all the locks */
2928 for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2929 queue = dlm_list_idx_to_ptr(res, idx);
2930 list_for_each_entry(lock, queue, list) {
2931 if (lock->ml.node == dlm->node_num)
2933 if (test_bit(lock->ml.node, dlm->exit_domain_map))
2935 nodenum = lock->ml.node;
2940 /* Go thru the refmap */
2943 noderef = find_next_bit(res->refmap, O2NM_MAX_NODES,
2945 if (noderef >= O2NM_MAX_NODES)
2947 if (noderef == dlm->node_num)
2949 if (test_bit(noderef, dlm->exit_domain_map))
2959 /* this is called by the new master once all lockres
2960 * data has been received */
2961 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2962 struct dlm_lock_resource *res,
2963 u8 master, u8 new_master,
2964 struct dlm_node_iter *iter)
2966 struct dlm_migrate_request migrate;
2967 int ret, skip, status = 0;
2970 memset(&migrate, 0, sizeof(migrate));
2971 migrate.namelen = res->lockname.len;
2972 memcpy(migrate.name, res->lockname.name, migrate.namelen);
2973 migrate.new_master = new_master;
2974 migrate.master = master;
2978 /* send message to all nodes, except the master and myself */
2979 while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2980 if (nodenum == master ||
2981 nodenum == new_master)
2984 /* We could race exit domain. If exited, skip. */
2985 spin_lock(&dlm->spinlock);
2986 skip = (!test_bit(nodenum, dlm->domain_map));
2987 spin_unlock(&dlm->spinlock);
2989 clear_bit(nodenum, iter->node_map);
2993 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2994 &migrate, sizeof(migrate), nodenum,
2997 mlog(ML_ERROR, "%s: res %.*s, Error %d send "
2998 "MIGRATE_REQUEST to node %u\n", dlm->name,
2999 migrate.namelen, migrate.name, ret, nodenum);
3000 if (!dlm_is_host_down(ret)) {
3001 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
3004 clear_bit(nodenum, iter->node_map);
3006 } else if (status < 0) {
3007 mlog(0, "migrate request (node %u) returned %d!\n",
3010 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
3011 /* during the migration request we short-circuited
3012 * the mastery of the lockres. make sure we have
3013 * a mastery ref for nodenum */
3014 mlog(0, "%s:%.*s: need ref for node %u\n",
3015 dlm->name, res->lockname.len, res->lockname.name,
3017 spin_lock(&res->spinlock);
3018 dlm_lockres_set_refmap_bit(dlm, res, nodenum);
3019 spin_unlock(&res->spinlock);
3026 mlog(0, "returning ret=%d\n", ret);
3031 /* if there is an existing mle for this lockres, we now know who the master is.
3032 * (the one who sent us *this* message) we can clear it up right away.
3033 * since the process that put the mle on the list still has a reference to it,
3034 * we can unhash it now, set the master and wake the process. as a result,
3035 * we will have no mle in the list to start with. now we can add an mle for
3036 * the migration and this should be the only one found for those scanning the
3038 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
3041 struct dlm_ctxt *dlm = data;
3042 struct dlm_lock_resource *res = NULL;
3043 struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
3044 struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
3046 unsigned int namelen, hash;
3052 name = migrate->name;
3053 namelen = migrate->namelen;
3054 hash = dlm_lockid_hash(name, namelen);
3056 /* preallocate.. if this fails, abort */
3057 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
3064 /* check for pre-existing lock */
3065 spin_lock(&dlm->spinlock);
3066 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
3068 spin_lock(&res->spinlock);
3069 if (res->state & DLM_LOCK_RES_RECOVERING) {
3070 /* if all is working ok, this can only mean that we got
3071 * a migrate request from a node that we now see as
3072 * dead. what can we do here? drop it to the floor? */
3073 spin_unlock(&res->spinlock);
3074 mlog(ML_ERROR, "Got a migrate request, but the "
3075 "lockres is marked as recovering!");
3076 kmem_cache_free(dlm_mle_cache, mle);
3077 ret = -EINVAL; /* need a better solution */
3080 res->state |= DLM_LOCK_RES_MIGRATING;
3081 spin_unlock(&res->spinlock);
3084 spin_lock(&dlm->master_lock);
3085 /* ignore status. only nonzero status would BUG. */
3086 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3088 migrate->new_master,
3091 spin_unlock(&dlm->master_lock);
3093 spin_unlock(&dlm->spinlock);
3096 /* master is known, detach if not already detached */
3097 dlm_mle_detach_hb_events(dlm, oldmle);
3098 dlm_put_mle(oldmle);
3102 dlm_lockres_put(res);
3108 /* must be holding dlm->spinlock and dlm->master_lock
3109 * when adding a migration mle, we can clear any other mles
3110 * in the master list because we know with certainty that
3111 * the master is "master". so we remove any old mle from
3112 * the list after setting it's master field, and then add
3113 * the new migration mle. this way we can hold with the rule
3114 * of having only one mle for a given lock name at all times. */
3115 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3116 struct dlm_lock_resource *res,
3117 struct dlm_master_list_entry *mle,
3118 struct dlm_master_list_entry **oldmle,
3119 const char *name, unsigned int namelen,
3120 u8 new_master, u8 master)
3127 assert_spin_locked(&dlm->spinlock);
3128 assert_spin_locked(&dlm->master_lock);
3130 /* caller is responsible for any ref taken here on oldmle */
3131 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3133 struct dlm_master_list_entry *tmp = *oldmle;
3134 spin_lock(&tmp->spinlock);
3135 if (tmp->type == DLM_MLE_MIGRATION) {
3136 if (master == dlm->node_num) {
3137 /* ah another process raced me to it */
3138 mlog(0, "tried to migrate %.*s, but some "
3139 "process beat me to it\n",
3143 /* bad. 2 NODES are trying to migrate! */
3144 mlog(ML_ERROR, "migration error mle: "
3145 "master=%u new_master=%u // request: "
3146 "master=%u new_master=%u // "
3148 tmp->master, tmp->new_master,
3154 /* this is essentially what assert_master does */
3155 tmp->master = master;
3156 atomic_set(&tmp->woken, 1);
3158 /* remove it so that only one mle will be found */
3159 __dlm_unlink_mle(dlm, tmp);
3160 __dlm_mle_detach_hb_events(dlm, tmp);
3161 if (tmp->type == DLM_MLE_MASTER) {
3162 ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3163 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3164 "telling master to get ref "
3165 "for cleared out mle during "
3166 "migration\n", dlm->name,
3167 namelen, name, master,
3171 spin_unlock(&tmp->spinlock);
3174 /* now add a migration mle to the tail of the list */
3175 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3176 mle->new_master = new_master;
3177 /* the new master will be sending an assert master for this.
3178 * at that point we will get the refmap reference */
3179 mle->master = master;
3180 /* do this for consistency with other mle types */
3181 set_bit(new_master, mle->maybe_map);
3182 __dlm_insert_mle(dlm, mle);
3188 * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3190 static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3191 struct dlm_master_list_entry *mle)
3193 struct dlm_lock_resource *res;
3195 /* Find the lockres associated to the mle and set its owner to UNK */
3196 res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3199 spin_unlock(&dlm->master_lock);
3201 /* move lockres onto recovery list */
3202 spin_lock(&res->spinlock);
3203 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3204 dlm_move_lockres_to_recovery_list(dlm, res);
3205 spin_unlock(&res->spinlock);
3206 dlm_lockres_put(res);
3208 /* about to get rid of mle, detach from heartbeat */
3209 __dlm_mle_detach_hb_events(dlm, mle);
3212 spin_lock(&dlm->master_lock);
3214 spin_unlock(&dlm->master_lock);
3220 static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3221 struct dlm_master_list_entry *mle)
3223 __dlm_mle_detach_hb_events(dlm, mle);
3225 spin_lock(&mle->spinlock);
3226 __dlm_unlink_mle(dlm, mle);
3227 atomic_set(&mle->woken, 1);
3228 spin_unlock(&mle->spinlock);
3233 static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3234 struct dlm_master_list_entry *mle, u8 dead_node)
3238 BUG_ON(mle->type != DLM_MLE_BLOCK);
3240 spin_lock(&mle->spinlock);
3241 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3242 if (bit != dead_node) {
3243 mlog(0, "mle found, but dead node %u would not have been "
3244 "master\n", dead_node);
3245 spin_unlock(&mle->spinlock);
3247 /* Must drop the refcount by one since the assert_master will
3248 * never arrive. This may result in the mle being unlinked and
3249 * freed, but there may still be a process waiting in the
3250 * dlmlock path which is fine. */
3251 mlog(0, "node %u was expected master\n", dead_node);
3252 atomic_set(&mle->woken, 1);
3253 spin_unlock(&mle->spinlock);
3256 /* Do not need events any longer, so detach from heartbeat */
3257 __dlm_mle_detach_hb_events(dlm, mle);
3262 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3264 struct dlm_master_list_entry *mle;
3265 struct dlm_lock_resource *res;
3266 struct hlist_head *bucket;
3267 struct hlist_node *tmp;
3270 mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
3272 assert_spin_locked(&dlm->spinlock);
3274 /* clean the master list */
3275 spin_lock(&dlm->master_lock);
3276 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3277 bucket = dlm_master_hash(dlm, i);
3278 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3279 BUG_ON(mle->type != DLM_MLE_BLOCK &&
3280 mle->type != DLM_MLE_MASTER &&
3281 mle->type != DLM_MLE_MIGRATION);
3283 /* MASTER mles are initiated locally. The waiting
3284 * process will notice the node map change shortly.
3285 * Let that happen as normal. */
3286 if (mle->type == DLM_MLE_MASTER)
3289 /* BLOCK mles are initiated by other nodes. Need to
3290 * clean up if the dead node would have been the
3292 if (mle->type == DLM_MLE_BLOCK) {
3293 dlm_clean_block_mle(dlm, mle, dead_node);
3297 /* Everything else is a MIGRATION mle */
3299 /* The rule for MIGRATION mles is that the master
3300 * becomes UNKNOWN if *either* the original or the new
3301 * master dies. All UNKNOWN lockres' are sent to
3302 * whichever node becomes the recovery master. The new
3303 * master is responsible for determining if there is
3304 * still a master for this lockres, or if he needs to
3305 * take over mastery. Either way, this node should
3306 * expect another message to resolve this. */
3308 if (mle->master != dead_node &&
3309 mle->new_master != dead_node)
3312 /* If we have reached this point, this mle needs to be
3313 * removed from the list and freed. */
3314 dlm_clean_migration_mle(dlm, mle);
3316 mlog(0, "%s: node %u died during migration from "
3317 "%u to %u!\n", dlm->name, dead_node, mle->master,
3320 /* If we find a lockres associated with the mle, we've
3321 * hit this rare case that messes up our lock ordering.
3322 * If so, we need to drop the master lock so that we can
3323 * take the lockres lock, meaning that we will have to
3324 * restart from the head of list. */
3325 res = dlm_reset_mleres_owner(dlm, mle);
3330 /* This may be the last reference */
3334 spin_unlock(&dlm->master_lock);
3337 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3340 struct dlm_node_iter iter;
3343 spin_lock(&dlm->spinlock);
3344 dlm_node_iter_init(dlm->domain_map, &iter);
3345 clear_bit(old_master, iter.node_map);
3346 clear_bit(dlm->node_num, iter.node_map);
3347 spin_unlock(&dlm->spinlock);
3349 /* ownership of the lockres is changing. account for the
3350 * mastery reference here since old_master will briefly have
3351 * a reference after the migration completes */
3352 spin_lock(&res->spinlock);
3353 dlm_lockres_set_refmap_bit(dlm, res, old_master);
3354 spin_unlock(&res->spinlock);
3356 mlog(0, "now time to do a migrate request to other nodes\n");
3357 ret = dlm_do_migrate_request(dlm, res, old_master,
3358 dlm->node_num, &iter);
3364 mlog(0, "doing assert master of %.*s to all except the original node\n",
3365 res->lockname.len, res->lockname.name);
3366 /* this call now finishes out the nodemap
3367 * even if one or more nodes die */
3368 ret = dlm_do_assert_master(dlm, res, iter.node_map,
3369 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3371 /* no longer need to retry. all living nodes contacted. */
3376 memset(iter.node_map, 0, sizeof(iter.node_map));
3377 set_bit(old_master, iter.node_map);
3378 mlog(0, "doing assert master of %.*s back to %u\n",
3379 res->lockname.len, res->lockname.name, old_master);
3380 ret = dlm_do_assert_master(dlm, res, iter.node_map,
3381 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3383 mlog(0, "assert master to original master failed "
3385 /* the only nonzero status here would be because of
3386 * a dead original node. we're done. */
3390 /* all done, set the owner, clear the flag */
3391 spin_lock(&res->spinlock);
3392 dlm_set_lockres_owner(dlm, res, dlm->node_num);
3393 res->state &= ~DLM_LOCK_RES_MIGRATING;
3394 spin_unlock(&res->spinlock);
3395 /* re-dirty it on the new master */
3396 dlm_kick_thread(dlm, res);
3403 * LOCKRES AST REFCOUNT
3404 * this is integral to migration
3407 /* for future intent to call an ast, reserve one ahead of time.
3408 * this should be called only after waiting on the lockres
3409 * with dlm_wait_on_lockres, and while still holding the
3410 * spinlock after the call. */
3411 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3413 assert_spin_locked(&res->spinlock);
3414 if (res->state & DLM_LOCK_RES_MIGRATING) {
3415 __dlm_print_one_lock_resource(res);
3417 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3419 atomic_inc(&res->asts_reserved);
3423 * used to drop the reserved ast, either because it went unused,
3424 * or because the ast/bast was actually called.
3426 * also, if there is a pending migration on this lockres,
3427 * and this was the last pending ast on the lockres,
3428 * atomically set the MIGRATING flag before we drop the lock.
3429 * this is how we ensure that migration can proceed with no
3430 * asts in progress. note that it is ok if the state of the
3431 * queues is such that a lock should be granted in the future
3432 * or that a bast should be fired, because the new master will
3433 * shuffle the lists on this lockres as soon as it is migrated.
3435 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3436 struct dlm_lock_resource *res)
3438 if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3441 if (!res->migration_pending) {
3442 spin_unlock(&res->spinlock);
3446 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3447 res->migration_pending = 0;
3448 res->state |= DLM_LOCK_RES_MIGRATING;
3449 spin_unlock(&res->spinlock);
3451 wake_up(&dlm->migration_wq);
3454 void dlm_force_free_mles(struct dlm_ctxt *dlm)
3457 struct hlist_head *bucket;
3458 struct dlm_master_list_entry *mle;
3459 struct hlist_node *tmp;
3462 * We notified all other nodes that we are exiting the domain and
3463 * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3464 * around we force free them and wake any processes that are waiting
3467 spin_lock(&dlm->spinlock);
3468 spin_lock(&dlm->master_lock);
3470 BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3471 BUG_ON((find_next_bit(dlm->domain_map, O2NM_MAX_NODES, 0) < O2NM_MAX_NODES));
3473 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3474 bucket = dlm_master_hash(dlm, i);
3475 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3476 if (mle->type != DLM_MLE_BLOCK) {
3477 mlog(ML_ERROR, "bad mle: %p\n", mle);
3478 dlm_print_one_mle(mle);
3480 atomic_set(&mle->woken, 1);
3483 __dlm_unlink_mle(dlm, mle);
3484 __dlm_mle_detach_hb_events(dlm, mle);
3488 spin_unlock(&dlm->master_lock);
3489 spin_unlock(&dlm->spinlock);