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;
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);
478 if (dlm_lockres_cache)
479 kmem_cache_destroy(dlm_lockres_cache);
482 static void dlm_lockres_release(struct kref *kref)
484 struct dlm_lock_resource *res;
485 struct dlm_ctxt *dlm;
487 res = container_of(kref, struct dlm_lock_resource, refs);
490 /* This should not happen -- all lockres' have a name
491 * associated with them at init time. */
492 BUG_ON(!res->lockname.name);
494 mlog(0, "destroying lockres %.*s\n", res->lockname.len,
497 spin_lock(&dlm->track_lock);
498 if (!list_empty(&res->tracking))
499 list_del_init(&res->tracking);
501 mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n",
502 res->lockname.len, res->lockname.name);
503 dlm_print_one_lock_resource(res);
505 spin_unlock(&dlm->track_lock);
507 atomic_dec(&dlm->res_cur_count);
509 if (!hlist_unhashed(&res->hash_node) ||
510 !list_empty(&res->granted) ||
511 !list_empty(&res->converting) ||
512 !list_empty(&res->blocked) ||
513 !list_empty(&res->dirty) ||
514 !list_empty(&res->recovering) ||
515 !list_empty(&res->purge)) {
517 "Going to BUG for resource %.*s."
518 " We're on a list! [%c%c%c%c%c%c%c]\n",
519 res->lockname.len, res->lockname.name,
520 !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
521 !list_empty(&res->granted) ? 'G' : ' ',
522 !list_empty(&res->converting) ? 'C' : ' ',
523 !list_empty(&res->blocked) ? 'B' : ' ',
524 !list_empty(&res->dirty) ? 'D' : ' ',
525 !list_empty(&res->recovering) ? 'R' : ' ',
526 !list_empty(&res->purge) ? 'P' : ' ');
528 dlm_print_one_lock_resource(res);
531 /* By the time we're ready to blow this guy away, we shouldn't
532 * be on any lists. */
533 BUG_ON(!hlist_unhashed(&res->hash_node));
534 BUG_ON(!list_empty(&res->granted));
535 BUG_ON(!list_empty(&res->converting));
536 BUG_ON(!list_empty(&res->blocked));
537 BUG_ON(!list_empty(&res->dirty));
538 BUG_ON(!list_empty(&res->recovering));
539 BUG_ON(!list_empty(&res->purge));
541 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
543 kmem_cache_free(dlm_lockres_cache, res);
546 void dlm_lockres_put(struct dlm_lock_resource *res)
548 kref_put(&res->refs, dlm_lockres_release);
551 static void dlm_init_lockres(struct dlm_ctxt *dlm,
552 struct dlm_lock_resource *res,
553 const char *name, unsigned int namelen)
557 /* If we memset here, we lose our reference to the kmalloc'd
558 * res->lockname.name, so be sure to init every field
561 qname = (char *) res->lockname.name;
562 memcpy(qname, name, namelen);
564 res->lockname.len = namelen;
565 res->lockname.hash = dlm_lockid_hash(name, namelen);
567 init_waitqueue_head(&res->wq);
568 spin_lock_init(&res->spinlock);
569 INIT_HLIST_NODE(&res->hash_node);
570 INIT_LIST_HEAD(&res->granted);
571 INIT_LIST_HEAD(&res->converting);
572 INIT_LIST_HEAD(&res->blocked);
573 INIT_LIST_HEAD(&res->dirty);
574 INIT_LIST_HEAD(&res->recovering);
575 INIT_LIST_HEAD(&res->purge);
576 INIT_LIST_HEAD(&res->tracking);
577 atomic_set(&res->asts_reserved, 0);
578 res->migration_pending = 0;
579 res->inflight_locks = 0;
583 kref_init(&res->refs);
585 atomic_inc(&dlm->res_tot_count);
586 atomic_inc(&dlm->res_cur_count);
588 /* just for consistency */
589 spin_lock(&res->spinlock);
590 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
591 spin_unlock(&res->spinlock);
593 res->state = DLM_LOCK_RES_IN_PROGRESS;
597 spin_lock(&dlm->spinlock);
598 list_add_tail(&res->tracking, &dlm->tracking_list);
599 spin_unlock(&dlm->spinlock);
601 memset(res->lvb, 0, DLM_LVB_LEN);
602 memset(res->refmap, 0, sizeof(res->refmap));
605 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
607 unsigned int namelen)
609 struct dlm_lock_resource *res = NULL;
611 res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
615 res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
616 if (!res->lockname.name)
619 dlm_init_lockres(dlm, res, name, namelen);
623 if (res && res->lockname.name)
624 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
627 kmem_cache_free(dlm_lockres_cache, res);
631 void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm,
632 struct dlm_lock_resource *res, int bit)
634 assert_spin_locked(&res->spinlock);
636 mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len,
637 res->lockname.name, bit, __builtin_return_address(0));
639 set_bit(bit, res->refmap);
642 void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm,
643 struct dlm_lock_resource *res, int bit)
645 assert_spin_locked(&res->spinlock);
647 mlog(0, "res %.*s, clr node %u, %ps()\n", res->lockname.len,
648 res->lockname.name, bit, __builtin_return_address(0));
650 clear_bit(bit, res->refmap);
654 void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
655 struct dlm_lock_resource *res)
657 assert_spin_locked(&res->spinlock);
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_drop_inflight_ref(struct dlm_ctxt *dlm,
667 struct dlm_lock_resource *res)
669 assert_spin_locked(&res->spinlock);
671 BUG_ON(res->inflight_locks == 0);
673 res->inflight_locks--;
675 mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm->name,
676 res->lockname.len, res->lockname.name, res->inflight_locks,
677 __builtin_return_address(0));
683 * lookup a lock resource by name.
684 * may already exist in the hashtable.
685 * lockid is null terminated
687 * if not, allocate enough for the lockres and for
688 * the temporary structure used in doing the mastering.
690 * also, do a lookup in the dlm->master_list to see
691 * if another node has begun mastering the same lock.
692 * if so, there should be a block entry in there
693 * for this name, and we should *not* attempt to master
694 * the lock here. need to wait around for that node
695 * to assert_master (or die).
698 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
703 struct dlm_lock_resource *tmpres=NULL, *res=NULL;
704 struct dlm_master_list_entry *mle = NULL;
705 struct dlm_master_list_entry *alloc_mle = NULL;
708 struct dlm_node_iter iter;
711 int bit, wait_on_recovery = 0;
715 hash = dlm_lockid_hash(lockid, namelen);
717 mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
720 spin_lock(&dlm->spinlock);
721 tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
723 spin_unlock(&dlm->spinlock);
724 spin_lock(&tmpres->spinlock);
725 /* Wait on the thread that is mastering the resource */
726 if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
727 __dlm_wait_on_lockres(tmpres);
728 BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
729 spin_unlock(&tmpres->spinlock);
730 dlm_lockres_put(tmpres);
735 /* Wait on the resource purge to complete before continuing */
736 if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) {
737 BUG_ON(tmpres->owner == dlm->node_num);
738 __dlm_wait_on_lockres_flags(tmpres,
739 DLM_LOCK_RES_DROPPING_REF);
740 spin_unlock(&tmpres->spinlock);
741 dlm_lockres_put(tmpres);
746 /* Grab inflight ref to pin the resource */
747 dlm_lockres_grab_inflight_ref(dlm, tmpres);
749 spin_unlock(&tmpres->spinlock);
751 dlm_lockres_put(res);
757 spin_unlock(&dlm->spinlock);
758 mlog(0, "allocating a new resource\n");
759 /* nothing found and we need to allocate one. */
760 alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
763 res = dlm_new_lockres(dlm, lockid, namelen);
769 mlog(0, "no lockres found, allocated our own: %p\n", res);
771 if (flags & LKM_LOCAL) {
772 /* caller knows it's safe to assume it's not mastered elsewhere
773 * DONE! return right away */
774 spin_lock(&res->spinlock);
775 dlm_change_lockres_owner(dlm, res, dlm->node_num);
776 __dlm_insert_lockres(dlm, res);
777 dlm_lockres_grab_inflight_ref(dlm, res);
778 spin_unlock(&res->spinlock);
779 spin_unlock(&dlm->spinlock);
780 /* lockres still marked IN_PROGRESS */
784 /* check master list to see if another node has started mastering it */
785 spin_lock(&dlm->master_lock);
787 /* if we found a block, wait for lock to be mastered by another node */
788 blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
791 if (mle->type == DLM_MLE_MASTER) {
792 mlog(ML_ERROR, "master entry for nonexistent lock!\n");
795 mig = (mle->type == DLM_MLE_MIGRATION);
796 /* if there is a migration in progress, let the migration
797 * finish before continuing. we can wait for the absence
798 * of the MIGRATION mle: either the migrate finished or
799 * one of the nodes died and the mle was cleaned up.
800 * if there is a BLOCK here, but it already has a master
801 * set, we are too late. the master does not have a ref
802 * for us in the refmap. detach the mle and drop it.
803 * either way, go back to the top and start over. */
804 if (mig || mle->master != O2NM_MAX_NODES) {
805 BUG_ON(mig && mle->master == dlm->node_num);
806 /* we arrived too late. the master does not
807 * have a ref for us. retry. */
808 mlog(0, "%s:%.*s: late on %s\n",
809 dlm->name, namelen, lockid,
810 mig ? "MIGRATION" : "BLOCK");
811 spin_unlock(&dlm->master_lock);
812 spin_unlock(&dlm->spinlock);
814 /* master is known, detach */
816 dlm_mle_detach_hb_events(dlm, mle);
819 /* this is lame, but we can't wait on either
820 * the mle or lockres waitqueue here */
826 /* go ahead and try to master lock on this node */
828 /* make sure this does not get freed below */
830 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
831 set_bit(dlm->node_num, mle->maybe_map);
832 __dlm_insert_mle(dlm, mle);
834 /* still holding the dlm spinlock, check the recovery map
835 * to see if there are any nodes that still need to be
836 * considered. these will not appear in the mle nodemap
837 * but they might own this lockres. wait on them. */
838 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
839 if (bit < O2NM_MAX_NODES) {
840 mlog(0, "%s: res %.*s, At least one node (%d) "
841 "to recover before lock mastery can begin\n",
842 dlm->name, namelen, (char *)lockid, bit);
843 wait_on_recovery = 1;
847 /* at this point there is either a DLM_MLE_BLOCK or a
848 * DLM_MLE_MASTER on the master list, so it's safe to add the
849 * lockres to the hashtable. anyone who finds the lock will
850 * still have to wait on the IN_PROGRESS. */
852 /* finally add the lockres to its hash bucket */
853 __dlm_insert_lockres(dlm, res);
855 /* Grab inflight ref to pin the resource */
856 spin_lock(&res->spinlock);
857 dlm_lockres_grab_inflight_ref(dlm, res);
858 spin_unlock(&res->spinlock);
860 /* get an extra ref on the mle in case this is a BLOCK
861 * if so, the creator of the BLOCK may try to put the last
862 * ref at this time in the assert master handler, so we
863 * need an extra one to keep from a bad ptr deref. */
864 dlm_get_mle_inuse(mle);
865 spin_unlock(&dlm->master_lock);
866 spin_unlock(&dlm->spinlock);
869 while (wait_on_recovery) {
870 /* any cluster changes that occurred after dropping the
871 * dlm spinlock would be detectable be a change on the mle,
872 * so we only need to clear out the recovery map once. */
873 if (dlm_is_recovery_lock(lockid, namelen)) {
874 mlog(0, "%s: Recovery map is not empty, but must "
875 "master $RECOVERY lock now\n", dlm->name);
876 if (!dlm_pre_master_reco_lockres(dlm, res))
877 wait_on_recovery = 0;
879 mlog(0, "%s: waiting 500ms for heartbeat state "
880 "change\n", dlm->name);
886 dlm_kick_recovery_thread(dlm);
888 dlm_wait_for_recovery(dlm);
890 spin_lock(&dlm->spinlock);
891 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
892 if (bit < O2NM_MAX_NODES) {
893 mlog(0, "%s: res %.*s, At least one node (%d) "
894 "to recover before lock mastery can begin\n",
895 dlm->name, namelen, (char *)lockid, bit);
896 wait_on_recovery = 1;
898 wait_on_recovery = 0;
899 spin_unlock(&dlm->spinlock);
901 if (wait_on_recovery)
902 dlm_wait_for_node_recovery(dlm, bit, 10000);
905 /* must wait for lock to be mastered elsewhere */
910 dlm_node_iter_init(mle->vote_map, &iter);
911 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
912 ret = dlm_do_master_request(res, mle, nodenum);
915 if (mle->master != O2NM_MAX_NODES) {
916 /* found a master ! */
917 if (mle->master <= nodenum)
919 /* if our master request has not reached the master
920 * yet, keep going until it does. this is how the
921 * master will know that asserts are needed back to
922 * the lower nodes. */
923 mlog(0, "%s: res %.*s, Requests only up to %u but "
924 "master is %u, keep going\n", dlm->name, namelen,
925 lockid, nodenum, mle->master);
930 /* keep going until the response map includes all nodes */
931 ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
933 wait_on_recovery = 1;
934 mlog(0, "%s: res %.*s, Node map changed, redo the master "
935 "request now, blocked=%d\n", dlm->name, res->lockname.len,
936 res->lockname.name, blocked);
938 mlog(ML_ERROR, "%s: res %.*s, Spinning on "
939 "dlm_wait_for_lock_mastery, blocked = %d\n",
940 dlm->name, res->lockname.len,
941 res->lockname.name, blocked);
942 dlm_print_one_lock_resource(res);
943 dlm_print_one_mle(mle);
949 mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len,
950 res->lockname.name, res->owner);
951 /* make sure we never continue without this */
952 BUG_ON(res->owner == O2NM_MAX_NODES);
954 /* master is known, detach if not already detached */
955 dlm_mle_detach_hb_events(dlm, mle);
957 /* put the extra ref */
958 dlm_put_mle_inuse(mle);
961 spin_lock(&res->spinlock);
962 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
963 spin_unlock(&res->spinlock);
967 /* need to free the unused mle */
969 kmem_cache_free(dlm_mle_cache, alloc_mle);
975 #define DLM_MASTERY_TIMEOUT_MS 5000
977 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
978 struct dlm_lock_resource *res,
979 struct dlm_master_list_entry *mle,
984 int map_changed, voting_done;
991 /* check if another node has already become the owner */
992 spin_lock(&res->spinlock);
993 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
994 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
995 res->lockname.len, res->lockname.name, res->owner);
996 spin_unlock(&res->spinlock);
997 /* this will cause the master to re-assert across
998 * the whole cluster, freeing up mles */
999 if (res->owner != dlm->node_num) {
1000 ret = dlm_do_master_request(res, mle, res->owner);
1002 /* give recovery a chance to run */
1003 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1011 spin_unlock(&res->spinlock);
1013 spin_lock(&mle->spinlock);
1015 map_changed = (memcmp(mle->vote_map, mle->node_map,
1016 sizeof(mle->vote_map)) != 0);
1017 voting_done = (memcmp(mle->vote_map, mle->response_map,
1018 sizeof(mle->vote_map)) == 0);
1020 /* restart if we hit any errors */
1023 mlog(0, "%s: %.*s: node map changed, restarting\n",
1024 dlm->name, res->lockname.len, res->lockname.name);
1025 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1026 b = (mle->type == DLM_MLE_BLOCK);
1027 if ((*blocked && !b) || (!*blocked && b)) {
1028 mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1029 dlm->name, res->lockname.len, res->lockname.name,
1033 spin_unlock(&mle->spinlock);
1038 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1039 "rechecking now\n", dlm->name, res->lockname.len,
1040 res->lockname.name);
1044 mlog(0, "map not changed and voting not done "
1045 "for %s:%.*s\n", dlm->name, res->lockname.len,
1046 res->lockname.name);
1050 if (m != O2NM_MAX_NODES) {
1051 /* another node has done an assert!
1056 /* have all nodes responded? */
1057 if (voting_done && !*blocked) {
1058 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1059 if (dlm->node_num <= bit) {
1060 /* my node number is lowest.
1061 * now tell other nodes that I am
1062 * mastering this. */
1063 mle->master = dlm->node_num;
1064 /* ref was grabbed in get_lock_resource
1065 * will be dropped in dlmlock_master */
1069 /* if voting is done, but we have not received
1070 * an assert master yet, we must sleep */
1074 spin_unlock(&mle->spinlock);
1076 /* sleep if we haven't finished voting yet */
1078 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1081 if (atomic_read(&mle->mle_refs.refcount) < 2)
1082 mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1083 atomic_read(&mle->mle_refs.refcount),
1084 res->lockname.len, res->lockname.name);
1086 atomic_set(&mle->woken, 0);
1087 (void)wait_event_timeout(mle->wq,
1088 (atomic_read(&mle->woken) == 1),
1090 if (res->owner == O2NM_MAX_NODES) {
1091 mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1092 res->lockname.len, res->lockname.name);
1095 mlog(0, "done waiting, master is %u\n", res->owner);
1103 mlog(0, "about to master %.*s here, this=%u\n",
1104 res->lockname.len, res->lockname.name, m);
1105 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1107 /* This is a failure in the network path,
1108 * not in the response to the assert_master
1109 * (any nonzero response is a BUG on this node).
1110 * Most likely a socket just got disconnected
1111 * due to node death. */
1114 /* no longer need to restart lock mastery.
1115 * all living nodes have been contacted. */
1119 /* set the lockres owner */
1120 spin_lock(&res->spinlock);
1121 /* mastery reference obtained either during
1122 * assert_master_handler or in get_lock_resource */
1123 dlm_change_lockres_owner(dlm, res, m);
1124 spin_unlock(&res->spinlock);
1130 struct dlm_bitmap_diff_iter
1133 unsigned long *orig_bm;
1134 unsigned long *cur_bm;
1135 unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1138 enum dlm_node_state_change
1145 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1146 unsigned long *orig_bm,
1147 unsigned long *cur_bm)
1149 unsigned long p1, p2;
1153 iter->orig_bm = orig_bm;
1154 iter->cur_bm = cur_bm;
1156 for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1157 p1 = *(iter->orig_bm + i);
1158 p2 = *(iter->cur_bm + i);
1159 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1163 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1164 enum dlm_node_state_change *state)
1168 if (iter->curnode >= O2NM_MAX_NODES)
1171 bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1173 if (bit >= O2NM_MAX_NODES) {
1174 iter->curnode = O2NM_MAX_NODES;
1178 /* if it was there in the original then this node died */
1179 if (test_bit(bit, iter->orig_bm))
1184 iter->curnode = bit;
1189 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1190 struct dlm_lock_resource *res,
1191 struct dlm_master_list_entry *mle,
1194 struct dlm_bitmap_diff_iter bdi;
1195 enum dlm_node_state_change sc;
1199 mlog(0, "something happened such that the "
1200 "master process may need to be restarted!\n");
1202 assert_spin_locked(&mle->spinlock);
1204 dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1205 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1207 if (sc == NODE_UP) {
1208 /* a node came up. clear any old vote from
1209 * the response map and set it in the vote map
1210 * then restart the mastery. */
1211 mlog(ML_NOTICE, "node %d up while restarting\n", node);
1213 /* redo the master request, but only for the new node */
1214 mlog(0, "sending request to new node\n");
1215 clear_bit(node, mle->response_map);
1216 set_bit(node, mle->vote_map);
1218 mlog(ML_ERROR, "node down! %d\n", node);
1220 int lowest = find_next_bit(mle->maybe_map,
1223 /* act like it was never there */
1224 clear_bit(node, mle->maybe_map);
1226 if (node == lowest) {
1227 mlog(0, "expected master %u died"
1228 " while this node was blocked "
1229 "waiting on it!\n", node);
1230 lowest = find_next_bit(mle->maybe_map,
1233 if (lowest < O2NM_MAX_NODES) {
1234 mlog(0, "%s:%.*s:still "
1235 "blocked. waiting on %u "
1241 /* mle is an MLE_BLOCK, but
1242 * there is now nothing left to
1243 * block on. we need to return
1244 * all the way back out and try
1245 * again with an MLE_MASTER.
1246 * dlm_do_local_recovery_cleanup
1247 * has already run, so the mle
1249 mlog(0, "%s:%.*s: no "
1250 "longer blocking. try to "
1251 "master this here\n",
1254 res->lockname.name);
1255 mle->type = DLM_MLE_MASTER;
1261 /* now blank out everything, as if we had never
1262 * contacted anyone */
1263 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1264 memset(mle->response_map, 0, sizeof(mle->response_map));
1265 /* reset the vote_map to the current node_map */
1266 memcpy(mle->vote_map, mle->node_map,
1267 sizeof(mle->node_map));
1268 /* put myself into the maybe map */
1269 if (mle->type != DLM_MLE_BLOCK)
1270 set_bit(dlm->node_num, mle->maybe_map);
1273 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1280 * DLM_MASTER_REQUEST_MSG
1282 * returns: 0 on success,
1283 * -errno on a network error
1285 * on error, the caller should assume the target node is "dead"
1289 static int dlm_do_master_request(struct dlm_lock_resource *res,
1290 struct dlm_master_list_entry *mle, int to)
1292 struct dlm_ctxt *dlm = mle->dlm;
1293 struct dlm_master_request request;
1294 int ret, response=0, resend;
1296 memset(&request, 0, sizeof(request));
1297 request.node_idx = dlm->node_num;
1299 BUG_ON(mle->type == DLM_MLE_MIGRATION);
1301 request.namelen = (u8)mle->mnamelen;
1302 memcpy(request.name, mle->mname, request.namelen);
1305 ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1306 sizeof(request), to, &response);
1308 if (ret == -ESRCH) {
1309 /* should never happen */
1310 mlog(ML_ERROR, "TCP stack not ready!\n");
1312 } else if (ret == -EINVAL) {
1313 mlog(ML_ERROR, "bad args passed to o2net!\n");
1315 } else if (ret == -ENOMEM) {
1316 mlog(ML_ERROR, "out of memory while trying to send "
1317 "network message! retrying\n");
1318 /* this is totally crude */
1321 } else if (!dlm_is_host_down(ret)) {
1322 /* not a network error. bad. */
1324 mlog(ML_ERROR, "unhandled error!");
1327 /* all other errors should be network errors,
1328 * and likely indicate node death */
1329 mlog(ML_ERROR, "link to %d went down!\n", to);
1335 spin_lock(&mle->spinlock);
1337 case DLM_MASTER_RESP_YES:
1338 set_bit(to, mle->response_map);
1339 mlog(0, "node %u is the master, response=YES\n", to);
1340 mlog(0, "%s:%.*s: master node %u now knows I have a "
1341 "reference\n", dlm->name, res->lockname.len,
1342 res->lockname.name, to);
1345 case DLM_MASTER_RESP_NO:
1346 mlog(0, "node %u not master, response=NO\n", to);
1347 set_bit(to, mle->response_map);
1349 case DLM_MASTER_RESP_MAYBE:
1350 mlog(0, "node %u not master, response=MAYBE\n", to);
1351 set_bit(to, mle->response_map);
1352 set_bit(to, mle->maybe_map);
1354 case DLM_MASTER_RESP_ERROR:
1355 mlog(0, "node %u hit an error, resending\n", to);
1360 mlog(ML_ERROR, "bad response! %u\n", response);
1363 spin_unlock(&mle->spinlock);
1365 /* this is also totally crude */
1375 * locks that can be taken here:
1381 * if possible, TRIM THIS DOWN!!!
1383 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1386 u8 response = DLM_MASTER_RESP_MAYBE;
1387 struct dlm_ctxt *dlm = data;
1388 struct dlm_lock_resource *res = NULL;
1389 struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1390 struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1392 unsigned int namelen, hash;
1395 int dispatch_assert = 0;
1398 return DLM_MASTER_RESP_NO;
1400 if (!dlm_domain_fully_joined(dlm)) {
1401 response = DLM_MASTER_RESP_NO;
1405 name = request->name;
1406 namelen = request->namelen;
1407 hash = dlm_lockid_hash(name, namelen);
1409 if (namelen > DLM_LOCKID_NAME_MAX) {
1410 response = DLM_IVBUFLEN;
1415 spin_lock(&dlm->spinlock);
1416 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1418 spin_unlock(&dlm->spinlock);
1420 /* take care of the easy cases up front */
1421 spin_lock(&res->spinlock);
1422 if (res->state & (DLM_LOCK_RES_RECOVERING|
1423 DLM_LOCK_RES_MIGRATING)) {
1424 spin_unlock(&res->spinlock);
1425 mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1426 "being recovered/migrated\n");
1427 response = DLM_MASTER_RESP_ERROR;
1429 kmem_cache_free(dlm_mle_cache, mle);
1433 if (res->owner == dlm->node_num) {
1434 dlm_lockres_set_refmap_bit(dlm, res, request->node_idx);
1435 spin_unlock(&res->spinlock);
1436 response = DLM_MASTER_RESP_YES;
1438 kmem_cache_free(dlm_mle_cache, mle);
1440 /* this node is the owner.
1441 * there is some extra work that needs to
1442 * happen now. the requesting node has
1443 * caused all nodes up to this one to
1444 * create mles. this node now needs to
1445 * go back and clean those up. */
1446 dispatch_assert = 1;
1448 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1449 spin_unlock(&res->spinlock);
1450 // mlog(0, "node %u is the master\n", res->owner);
1451 response = DLM_MASTER_RESP_NO;
1453 kmem_cache_free(dlm_mle_cache, mle);
1457 /* ok, there is no owner. either this node is
1458 * being blocked, or it is actively trying to
1459 * master this lock. */
1460 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1461 mlog(ML_ERROR, "lock with no owner should be "
1466 // mlog(0, "lockres is in progress...\n");
1467 spin_lock(&dlm->master_lock);
1468 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1470 mlog(ML_ERROR, "no mle found for this lock!\n");
1474 spin_lock(&tmpmle->spinlock);
1475 if (tmpmle->type == DLM_MLE_BLOCK) {
1476 // mlog(0, "this node is waiting for "
1477 // "lockres to be mastered\n");
1478 response = DLM_MASTER_RESP_NO;
1479 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1480 mlog(0, "node %u is master, but trying to migrate to "
1481 "node %u.\n", tmpmle->master, tmpmle->new_master);
1482 if (tmpmle->master == dlm->node_num) {
1483 mlog(ML_ERROR, "no owner on lockres, but this "
1484 "node is trying to migrate it to %u?!\n",
1485 tmpmle->new_master);
1488 /* the real master can respond on its own */
1489 response = DLM_MASTER_RESP_NO;
1491 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1493 if (tmpmle->master == dlm->node_num) {
1494 response = DLM_MASTER_RESP_YES;
1495 /* this node will be the owner.
1496 * go back and clean the mles on any
1498 dispatch_assert = 1;
1499 dlm_lockres_set_refmap_bit(dlm, res,
1502 response = DLM_MASTER_RESP_NO;
1504 // mlog(0, "this node is attempting to "
1505 // "master lockres\n");
1506 response = DLM_MASTER_RESP_MAYBE;
1509 set_bit(request->node_idx, tmpmle->maybe_map);
1510 spin_unlock(&tmpmle->spinlock);
1512 spin_unlock(&dlm->master_lock);
1513 spin_unlock(&res->spinlock);
1515 /* keep the mle attached to heartbeat events */
1516 dlm_put_mle(tmpmle);
1518 kmem_cache_free(dlm_mle_cache, mle);
1523 * lockres doesn't exist on this node
1524 * if there is an MLE_BLOCK, return NO
1525 * if there is an MLE_MASTER, return MAYBE
1526 * otherwise, add an MLE_BLOCK, return NO
1528 spin_lock(&dlm->master_lock);
1529 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1531 /* this lockid has never been seen on this node yet */
1532 // mlog(0, "no mle found\n");
1534 spin_unlock(&dlm->master_lock);
1535 spin_unlock(&dlm->spinlock);
1537 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1539 response = DLM_MASTER_RESP_ERROR;
1540 mlog_errno(-ENOMEM);
1546 // mlog(0, "this is second time thru, already allocated, "
1547 // "add the block.\n");
1548 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1549 set_bit(request->node_idx, mle->maybe_map);
1550 __dlm_insert_mle(dlm, mle);
1551 response = DLM_MASTER_RESP_NO;
1553 // mlog(0, "mle was found\n");
1555 spin_lock(&tmpmle->spinlock);
1556 if (tmpmle->master == dlm->node_num) {
1557 mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1560 if (tmpmle->type == DLM_MLE_BLOCK)
1561 response = DLM_MASTER_RESP_NO;
1562 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1563 mlog(0, "migration mle was found (%u->%u)\n",
1564 tmpmle->master, tmpmle->new_master);
1565 /* real master can respond on its own */
1566 response = DLM_MASTER_RESP_NO;
1568 response = DLM_MASTER_RESP_MAYBE;
1570 set_bit(request->node_idx, tmpmle->maybe_map);
1571 spin_unlock(&tmpmle->spinlock);
1573 spin_unlock(&dlm->master_lock);
1574 spin_unlock(&dlm->spinlock);
1577 /* keep the mle attached to heartbeat events */
1578 dlm_put_mle(tmpmle);
1582 * __dlm_lookup_lockres() grabbed a reference to this lockres.
1583 * The reference is released by dlm_assert_master_worker() under
1584 * the call to dlm_dispatch_assert_master(). If
1585 * dlm_assert_master_worker() isn't called, we drop it here.
1587 if (dispatch_assert) {
1588 if (response != DLM_MASTER_RESP_YES)
1589 mlog(ML_ERROR, "invalid response %d\n", response);
1591 mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1594 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1595 dlm->node_num, res->lockname.len, res->lockname.name);
1596 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1597 DLM_ASSERT_MASTER_MLE_CLEANUP);
1599 mlog(ML_ERROR, "failed to dispatch assert master work\n");
1600 response = DLM_MASTER_RESP_ERROR;
1601 dlm_lockres_put(res);
1605 dlm_lockres_put(res);
1613 * DLM_ASSERT_MASTER_MSG
1618 * NOTE: this can be used for debugging
1619 * can periodically run all locks owned by this node
1620 * and re-assert across the cluster...
1622 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1623 struct dlm_lock_resource *res,
1624 void *nodemap, u32 flags)
1626 struct dlm_assert_master assert;
1628 struct dlm_node_iter iter;
1631 const char *lockname = res->lockname.name;
1632 unsigned int namelen = res->lockname.len;
1634 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1636 spin_lock(&res->spinlock);
1637 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1638 spin_unlock(&res->spinlock);
1643 /* note that if this nodemap is empty, it returns 0 */
1644 dlm_node_iter_init(nodemap, &iter);
1645 while ((to = dlm_node_iter_next(&iter)) >= 0) {
1647 struct dlm_master_list_entry *mle = NULL;
1649 mlog(0, "sending assert master to %d (%.*s)\n", to,
1651 memset(&assert, 0, sizeof(assert));
1652 assert.node_idx = dlm->node_num;
1653 assert.namelen = namelen;
1654 memcpy(assert.name, lockname, namelen);
1655 assert.flags = cpu_to_be32(flags);
1657 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1658 &assert, sizeof(assert), to, &r);
1660 mlog(ML_ERROR, "Error %d when sending message %u (key "
1661 "0x%x) to node %u\n", tmpret,
1662 DLM_ASSERT_MASTER_MSG, dlm->key, to);
1663 if (!dlm_is_host_down(tmpret)) {
1664 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1667 /* a node died. finish out the rest of the nodes. */
1668 mlog(0, "link to %d went down!\n", to);
1669 /* any nonzero status return will do */
1673 /* ok, something horribly messed. kill thyself. */
1674 mlog(ML_ERROR,"during assert master of %.*s to %u, "
1675 "got %d.\n", namelen, lockname, to, r);
1676 spin_lock(&dlm->spinlock);
1677 spin_lock(&dlm->master_lock);
1678 if (dlm_find_mle(dlm, &mle, (char *)lockname,
1680 dlm_print_one_mle(mle);
1683 spin_unlock(&dlm->master_lock);
1684 spin_unlock(&dlm->spinlock);
1688 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1689 !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1690 mlog(ML_ERROR, "%.*s: very strange, "
1691 "master MLE but no lockres on %u\n",
1692 namelen, lockname, to);
1695 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1696 mlog(0, "%.*s: node %u create mles on other "
1697 "nodes and requests a re-assert\n",
1698 namelen, lockname, to);
1701 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1702 mlog(0, "%.*s: node %u has a reference to this "
1703 "lockres, set the bit in the refmap\n",
1704 namelen, lockname, to);
1705 spin_lock(&res->spinlock);
1706 dlm_lockres_set_refmap_bit(dlm, res, to);
1707 spin_unlock(&res->spinlock);
1714 spin_lock(&res->spinlock);
1715 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1716 spin_unlock(&res->spinlock);
1723 * locks that can be taken here:
1729 * if possible, TRIM THIS DOWN!!!
1731 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1734 struct dlm_ctxt *dlm = data;
1735 struct dlm_master_list_entry *mle = NULL;
1736 struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1737 struct dlm_lock_resource *res = NULL;
1739 unsigned int namelen, hash;
1741 int master_request = 0, have_lockres_ref = 0;
1747 name = assert->name;
1748 namelen = assert->namelen;
1749 hash = dlm_lockid_hash(name, namelen);
1750 flags = be32_to_cpu(assert->flags);
1752 if (namelen > DLM_LOCKID_NAME_MAX) {
1753 mlog(ML_ERROR, "Invalid name length!");
1757 spin_lock(&dlm->spinlock);
1760 mlog(0, "assert_master with flags: %u\n", flags);
1763 spin_lock(&dlm->master_lock);
1764 if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1765 /* not an error, could be master just re-asserting */
1766 mlog(0, "just got an assert_master from %u, but no "
1767 "MLE for it! (%.*s)\n", assert->node_idx,
1770 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1771 if (bit >= O2NM_MAX_NODES) {
1772 /* not necessarily an error, though less likely.
1773 * could be master just re-asserting. */
1774 mlog(0, "no bits set in the maybe_map, but %u "
1775 "is asserting! (%.*s)\n", assert->node_idx,
1777 } else if (bit != assert->node_idx) {
1778 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1779 mlog(0, "master %u was found, %u should "
1780 "back off\n", assert->node_idx, bit);
1782 /* with the fix for bug 569, a higher node
1783 * number winning the mastery will respond
1784 * YES to mastery requests, but this node
1785 * had no way of knowing. let it pass. */
1786 mlog(0, "%u is the lowest node, "
1787 "%u is asserting. (%.*s) %u must "
1788 "have begun after %u won.\n", bit,
1789 assert->node_idx, namelen, name, bit,
1793 if (mle->type == DLM_MLE_MIGRATION) {
1794 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1795 mlog(0, "%s:%.*s: got cleanup assert"
1796 " from %u for migration\n",
1797 dlm->name, namelen, name,
1799 } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1800 mlog(0, "%s:%.*s: got unrelated assert"
1801 " from %u for migration, ignoring\n",
1802 dlm->name, namelen, name,
1805 spin_unlock(&dlm->master_lock);
1806 spin_unlock(&dlm->spinlock);
1811 spin_unlock(&dlm->master_lock);
1813 /* ok everything checks out with the MLE
1814 * now check to see if there is a lockres */
1815 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1817 spin_lock(&res->spinlock);
1818 if (res->state & DLM_LOCK_RES_RECOVERING) {
1819 mlog(ML_ERROR, "%u asserting but %.*s is "
1820 "RECOVERING!\n", assert->node_idx, namelen, name);
1824 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1825 res->owner != assert->node_idx) {
1826 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1827 "but current owner is %u! (%.*s)\n",
1828 assert->node_idx, res->owner, namelen,
1830 __dlm_print_one_lock_resource(res);
1833 } else if (mle->type != DLM_MLE_MIGRATION) {
1834 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1835 /* owner is just re-asserting */
1836 if (res->owner == assert->node_idx) {
1837 mlog(0, "owner %u re-asserting on "
1838 "lock %.*s\n", assert->node_idx,
1842 mlog(ML_ERROR, "got assert_master from "
1843 "node %u, but %u is the owner! "
1844 "(%.*s)\n", assert->node_idx,
1845 res->owner, namelen, name);
1848 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1849 mlog(ML_ERROR, "got assert from %u, but lock "
1850 "with no owner should be "
1851 "in-progress! (%.*s)\n",
1856 } else /* mle->type == DLM_MLE_MIGRATION */ {
1857 /* should only be getting an assert from new master */
1858 if (assert->node_idx != mle->new_master) {
1859 mlog(ML_ERROR, "got assert from %u, but "
1860 "new master is %u, and old master "
1862 assert->node_idx, mle->new_master,
1863 mle->master, namelen, name);
1869 spin_unlock(&res->spinlock);
1872 // mlog(0, "woo! got an assert_master from node %u!\n",
1873 // assert->node_idx);
1879 spin_lock(&mle->spinlock);
1880 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1883 /* MASTER mle: if any bits set in the response map
1884 * then the calling node needs to re-assert to clear
1885 * up nodes that this node contacted */
1886 while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1887 nn+1)) < O2NM_MAX_NODES) {
1888 if (nn != dlm->node_num && nn != assert->node_idx)
1892 mle->master = assert->node_idx;
1893 atomic_set(&mle->woken, 1);
1895 spin_unlock(&mle->spinlock);
1899 spin_lock(&res->spinlock);
1900 if (mle->type == DLM_MLE_MIGRATION) {
1901 mlog(0, "finishing off migration of lockres %.*s, "
1903 res->lockname.len, res->lockname.name,
1904 dlm->node_num, mle->new_master);
1905 res->state &= ~DLM_LOCK_RES_MIGRATING;
1907 dlm_change_lockres_owner(dlm, res, mle->new_master);
1908 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1910 dlm_change_lockres_owner(dlm, res, mle->master);
1912 spin_unlock(&res->spinlock);
1913 have_lockres_ref = 1;
1918 /* master is known, detach if not already detached.
1919 * ensures that only one assert_master call will happen
1921 spin_lock(&dlm->master_lock);
1923 rr = atomic_read(&mle->mle_refs.refcount);
1924 if (mle->inuse > 0) {
1925 if (extra_ref && rr < 3)
1927 else if (!extra_ref && rr < 2)
1930 if (extra_ref && rr < 2)
1932 else if (!extra_ref && rr < 1)
1936 mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
1937 "that will mess up this node, refs=%d, extra=%d, "
1938 "inuse=%d\n", dlm->name, namelen, name,
1939 assert->node_idx, rr, extra_ref, mle->inuse);
1940 dlm_print_one_mle(mle);
1942 __dlm_unlink_mle(dlm, mle);
1943 __dlm_mle_detach_hb_events(dlm, mle);
1946 /* the assert master message now balances the extra
1947 * ref given by the master / migration request message.
1948 * if this is the last put, it will be removed
1952 spin_unlock(&dlm->master_lock);
1954 if (res->owner != assert->node_idx) {
1955 mlog(0, "assert_master from %u, but current "
1956 "owner is %u (%.*s), no mle\n", assert->node_idx,
1957 res->owner, namelen, name);
1960 spin_unlock(&dlm->spinlock);
1965 spin_lock(&res->spinlock);
1966 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1967 spin_unlock(&res->spinlock);
1968 *ret_data = (void *)res;
1971 if (master_request) {
1972 mlog(0, "need to tell master to reassert\n");
1973 /* positive. negative would shoot down the node. */
1974 ret |= DLM_ASSERT_RESPONSE_REASSERT;
1975 if (!have_lockres_ref) {
1976 mlog(ML_ERROR, "strange, got assert from %u, MASTER "
1977 "mle present here for %s:%.*s, but no lockres!\n",
1978 assert->node_idx, dlm->name, namelen, name);
1981 if (have_lockres_ref) {
1982 /* let the master know we have a reference to the lockres */
1983 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
1984 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
1985 dlm->name, namelen, name, assert->node_idx);
1990 /* kill the caller! */
1991 mlog(ML_ERROR, "Bad message received from another node. Dumping state "
1992 "and killing the other node now! This node is OK and can continue.\n");
1993 __dlm_print_one_lock_resource(res);
1994 spin_unlock(&res->spinlock);
1995 spin_unlock(&dlm->spinlock);
1996 *ret_data = (void *)res;
2001 void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2003 struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2006 spin_lock(&res->spinlock);
2007 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2008 spin_unlock(&res->spinlock);
2010 dlm_lockres_put(res);
2015 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2016 struct dlm_lock_resource *res,
2017 int ignore_higher, u8 request_from, u32 flags)
2019 struct dlm_work_item *item;
2020 item = kzalloc(sizeof(*item), GFP_ATOMIC);
2025 /* queue up work for dlm_assert_master_worker */
2026 dlm_grab(dlm); /* get an extra ref for the work item */
2027 dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2028 item->u.am.lockres = res; /* already have a ref */
2029 /* can optionally ignore node numbers higher than this node */
2030 item->u.am.ignore_higher = ignore_higher;
2031 item->u.am.request_from = request_from;
2032 item->u.am.flags = flags;
2035 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2036 res->lockname.name);
2038 spin_lock(&dlm->work_lock);
2039 list_add_tail(&item->list, &dlm->work_list);
2040 spin_unlock(&dlm->work_lock);
2042 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2046 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2048 struct dlm_ctxt *dlm = data;
2050 struct dlm_lock_resource *res;
2051 unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2058 res = item->u.am.lockres;
2059 ignore_higher = item->u.am.ignore_higher;
2060 request_from = item->u.am.request_from;
2061 flags = item->u.am.flags;
2063 spin_lock(&dlm->spinlock);
2064 memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2065 spin_unlock(&dlm->spinlock);
2067 clear_bit(dlm->node_num, nodemap);
2068 if (ignore_higher) {
2069 /* if is this just to clear up mles for nodes below
2070 * this node, do not send the message to the original
2071 * caller or any node number higher than this */
2072 clear_bit(request_from, nodemap);
2073 bit = dlm->node_num;
2075 bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2077 if (bit >= O2NM_MAX_NODES)
2079 clear_bit(bit, nodemap);
2084 * If we're migrating this lock to someone else, we are no
2085 * longer allowed to assert out own mastery. OTOH, we need to
2086 * prevent migration from starting while we're still asserting
2087 * our dominance. The reserved ast delays migration.
2089 spin_lock(&res->spinlock);
2090 if (res->state & DLM_LOCK_RES_MIGRATING) {
2091 mlog(0, "Someone asked us to assert mastery, but we're "
2092 "in the middle of migration. Skipping assert, "
2093 "the new master will handle that.\n");
2094 spin_unlock(&res->spinlock);
2097 __dlm_lockres_reserve_ast(res);
2098 spin_unlock(&res->spinlock);
2100 /* this call now finishes out the nodemap
2101 * even if one or more nodes die */
2102 mlog(0, "worker about to master %.*s here, this=%u\n",
2103 res->lockname.len, res->lockname.name, dlm->node_num);
2104 ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2106 /* no need to restart, we are done */
2107 if (!dlm_is_host_down(ret))
2111 /* Ok, we've asserted ourselves. Let's let migration start. */
2112 dlm_lockres_release_ast(dlm, res);
2115 dlm_lockres_put(res);
2117 mlog(0, "finished with dlm_assert_master_worker\n");
2120 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2121 * We cannot wait for node recovery to complete to begin mastering this
2122 * lockres because this lockres is used to kick off recovery! ;-)
2123 * So, do a pre-check on all living nodes to see if any of those nodes
2124 * think that $RECOVERY is currently mastered by a dead node. If so,
2125 * we wait a short time to allow that node to get notified by its own
2126 * heartbeat stack, then check again. All $RECOVERY lock resources
2127 * mastered by dead nodes are purged when the hearbeat callback is
2128 * fired, so we can know for sure that it is safe to continue once
2129 * the node returns a live node or no node. */
2130 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2131 struct dlm_lock_resource *res)
2133 struct dlm_node_iter iter;
2136 u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2138 spin_lock(&dlm->spinlock);
2139 dlm_node_iter_init(dlm->domain_map, &iter);
2140 spin_unlock(&dlm->spinlock);
2142 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2143 /* do not send to self */
2144 if (nodenum == dlm->node_num)
2146 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2149 if (!dlm_is_host_down(ret))
2151 /* host is down, so answer for that node would be
2152 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
2156 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2157 /* check to see if this master is in the recovery map */
2158 spin_lock(&dlm->spinlock);
2159 if (test_bit(master, dlm->recovery_map)) {
2160 mlog(ML_NOTICE, "%s: node %u has not seen "
2161 "node %u go down yet, and thinks the "
2162 "dead node is mastering the recovery "
2163 "lock. must wait.\n", dlm->name,
2167 spin_unlock(&dlm->spinlock);
2168 mlog(0, "%s: reco lock master is %u\n", dlm->name,
2177 * DLM_DEREF_LOCKRES_MSG
2180 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2182 struct dlm_deref_lockres deref;
2184 const char *lockname;
2185 unsigned int namelen;
2187 lockname = res->lockname.name;
2188 namelen = res->lockname.len;
2189 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2191 memset(&deref, 0, sizeof(deref));
2192 deref.node_idx = dlm->node_num;
2193 deref.namelen = namelen;
2194 memcpy(deref.name, lockname, namelen);
2196 ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2197 &deref, sizeof(deref), res->owner, &r);
2199 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n",
2200 dlm->name, namelen, lockname, ret, res->owner);
2202 /* BAD. other node says I did not have a ref. */
2203 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2204 dlm->name, namelen, lockname, res->owner, r);
2205 dlm_print_one_lock_resource(res);
2211 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2214 struct dlm_ctxt *dlm = data;
2215 struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2216 struct dlm_lock_resource *res = NULL;
2218 unsigned int namelen;
2222 struct dlm_work_item *item;
2230 namelen = deref->namelen;
2231 node = deref->node_idx;
2233 if (namelen > DLM_LOCKID_NAME_MAX) {
2234 mlog(ML_ERROR, "Invalid name length!");
2237 if (deref->node_idx >= O2NM_MAX_NODES) {
2238 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2242 hash = dlm_lockid_hash(name, namelen);
2244 spin_lock(&dlm->spinlock);
2245 res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2247 spin_unlock(&dlm->spinlock);
2248 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2249 dlm->name, namelen, name);
2252 spin_unlock(&dlm->spinlock);
2254 spin_lock(&res->spinlock);
2255 if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2258 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2259 if (test_bit(node, res->refmap)) {
2260 dlm_lockres_clear_refmap_bit(dlm, res, node);
2264 spin_unlock(&res->spinlock);
2268 dlm_lockres_calc_usage(dlm, res);
2270 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2271 "but it is already dropped!\n", dlm->name,
2272 res->lockname.len, res->lockname.name, node);
2273 dlm_print_one_lock_resource(res);
2279 item = kzalloc(sizeof(*item), GFP_NOFS);
2286 dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2287 item->u.dl.deref_res = res;
2288 item->u.dl.deref_node = node;
2290 spin_lock(&dlm->work_lock);
2291 list_add_tail(&item->list, &dlm->work_list);
2292 spin_unlock(&dlm->work_lock);
2294 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2299 dlm_lockres_put(res);
2305 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2307 struct dlm_ctxt *dlm;
2308 struct dlm_lock_resource *res;
2313 res = item->u.dl.deref_res;
2314 node = item->u.dl.deref_node;
2316 spin_lock(&res->spinlock);
2317 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2318 if (test_bit(node, res->refmap)) {
2319 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2320 dlm_lockres_clear_refmap_bit(dlm, res, node);
2323 spin_unlock(&res->spinlock);
2326 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2327 dlm->name, res->lockname.len, res->lockname.name, node);
2328 dlm_lockres_calc_usage(dlm, res);
2330 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2331 "but it is already dropped!\n", dlm->name,
2332 res->lockname.len, res->lockname.name, node);
2333 dlm_print_one_lock_resource(res);
2336 dlm_lockres_put(res);
2340 * A migrateable resource is one that is :
2341 * 1. locally mastered, and,
2342 * 2. zero local locks, and,
2343 * 3. one or more non-local locks, or, one or more references
2344 * Returns 1 if yes, 0 if not.
2346 static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
2347 struct dlm_lock_resource *res)
2349 enum dlm_lockres_list idx;
2350 int nonlocal = 0, node_ref;
2351 struct list_head *queue;
2352 struct dlm_lock *lock;
2355 assert_spin_locked(&res->spinlock);
2357 if (res->owner != dlm->node_num)
2360 for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2361 queue = dlm_list_idx_to_ptr(res, idx);
2362 list_for_each_entry(lock, queue, list) {
2363 if (lock->ml.node != dlm->node_num) {
2367 cookie = be64_to_cpu(lock->ml.cookie);
2368 mlog(0, "%s: Not migrateable res %.*s, lock %u:%llu on "
2369 "%s list\n", dlm->name, res->lockname.len,
2371 dlm_get_lock_cookie_node(cookie),
2372 dlm_get_lock_cookie_seq(cookie),
2373 dlm_list_in_text(idx));
2379 node_ref = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
2380 if (node_ref >= O2NM_MAX_NODES)
2384 mlog(0, "%s: res %.*s, Migrateable\n", dlm->name, res->lockname.len,
2385 res->lockname.name);
2391 * DLM_MIGRATE_LOCKRES
2395 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2396 struct dlm_lock_resource *res, u8 target)
2398 struct dlm_master_list_entry *mle = NULL;
2399 struct dlm_master_list_entry *oldmle = NULL;
2400 struct dlm_migratable_lockres *mres = NULL;
2403 unsigned int namelen;
2410 BUG_ON(target == O2NM_MAX_NODES);
2412 name = res->lockname.name;
2413 namelen = res->lockname.len;
2415 mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name,
2418 /* preallocate up front. if this fails, abort */
2420 mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2426 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2434 * clear any existing master requests and
2435 * add the migration mle to the list
2437 spin_lock(&dlm->spinlock);
2438 spin_lock(&dlm->master_lock);
2439 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2440 namelen, target, dlm->node_num);
2441 spin_unlock(&dlm->master_lock);
2442 spin_unlock(&dlm->spinlock);
2444 if (ret == -EEXIST) {
2445 mlog(0, "another process is already migrating it\n");
2451 * set the MIGRATING flag and flush asts
2452 * if we fail after this we need to re-dirty the lockres
2454 if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2455 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2456 "the target went down.\n", res->lockname.len,
2457 res->lockname.name, target);
2458 spin_lock(&res->spinlock);
2459 res->state &= ~DLM_LOCK_RES_MIGRATING;
2461 spin_unlock(&res->spinlock);
2467 /* master is known, detach if not already detached */
2468 dlm_mle_detach_hb_events(dlm, oldmle);
2469 dlm_put_mle(oldmle);
2474 dlm_mle_detach_hb_events(dlm, mle);
2477 kmem_cache_free(dlm_mle_cache, mle);
2484 * at this point, we have a migration target, an mle
2485 * in the master list, and the MIGRATING flag set on
2489 /* now that remote nodes are spinning on the MIGRATING flag,
2490 * ensure that all assert_master work is flushed. */
2491 flush_workqueue(dlm->dlm_worker);
2493 /* get an extra reference on the mle.
2494 * otherwise the assert_master from the new
2495 * master will destroy this.
2496 * also, make sure that all callers of dlm_get_mle
2497 * take both dlm->spinlock and dlm->master_lock */
2498 spin_lock(&dlm->spinlock);
2499 spin_lock(&dlm->master_lock);
2500 dlm_get_mle_inuse(mle);
2501 spin_unlock(&dlm->master_lock);
2502 spin_unlock(&dlm->spinlock);
2504 /* notify new node and send all lock state */
2505 /* call send_one_lockres with migration flag.
2506 * this serves as notice to the target node that a
2507 * migration is starting. */
2508 ret = dlm_send_one_lockres(dlm, res, mres, target,
2509 DLM_MRES_MIGRATION);
2512 mlog(0, "migration to node %u failed with %d\n",
2514 /* migration failed, detach and clean up mle */
2515 dlm_mle_detach_hb_events(dlm, mle);
2517 dlm_put_mle_inuse(mle);
2518 spin_lock(&res->spinlock);
2519 res->state &= ~DLM_LOCK_RES_MIGRATING;
2521 spin_unlock(&res->spinlock);
2522 if (dlm_is_host_down(ret))
2523 dlm_wait_for_node_death(dlm, target,
2524 DLM_NODE_DEATH_WAIT_MAX);
2528 /* at this point, the target sends a message to all nodes,
2529 * (using dlm_do_migrate_request). this node is skipped since
2530 * we had to put an mle in the list to begin the process. this
2531 * node now waits for target to do an assert master. this node
2532 * will be the last one notified, ensuring that the migration
2533 * is complete everywhere. if the target dies while this is
2534 * going on, some nodes could potentially see the target as the
2535 * master, so it is important that my recovery finds the migration
2536 * mle and sets the master to UNKNOWN. */
2539 /* wait for new node to assert master */
2541 ret = wait_event_interruptible_timeout(mle->wq,
2542 (atomic_read(&mle->woken) == 1),
2543 msecs_to_jiffies(5000));
2546 if (atomic_read(&mle->woken) == 1 ||
2547 res->owner == target)
2550 mlog(0, "%s:%.*s: timed out during migration\n",
2551 dlm->name, res->lockname.len, res->lockname.name);
2552 /* avoid hang during shutdown when migrating lockres
2553 * to a node which also goes down */
2554 if (dlm_is_node_dead(dlm, target)) {
2555 mlog(0, "%s:%.*s: expected migration "
2556 "target %u is no longer up, restarting\n",
2557 dlm->name, res->lockname.len,
2558 res->lockname.name, target);
2560 /* migration failed, detach and clean up mle */
2561 dlm_mle_detach_hb_events(dlm, mle);
2563 dlm_put_mle_inuse(mle);
2564 spin_lock(&res->spinlock);
2565 res->state &= ~DLM_LOCK_RES_MIGRATING;
2567 spin_unlock(&res->spinlock);
2571 mlog(0, "%s:%.*s: caught signal during migration\n",
2572 dlm->name, res->lockname.len, res->lockname.name);
2575 /* all done, set the owner, clear the flag */
2576 spin_lock(&res->spinlock);
2577 dlm_set_lockres_owner(dlm, res, target);
2578 res->state &= ~DLM_LOCK_RES_MIGRATING;
2579 dlm_remove_nonlocal_locks(dlm, res);
2580 spin_unlock(&res->spinlock);
2583 /* master is known, detach if not already detached */
2584 dlm_mle_detach_hb_events(dlm, mle);
2585 dlm_put_mle_inuse(mle);
2588 dlm_lockres_calc_usage(dlm, res);
2591 /* re-dirty the lockres if we failed */
2593 dlm_kick_thread(dlm, res);
2595 /* wake up waiters if the MIGRATING flag got set
2596 * but migration failed */
2601 free_page((unsigned long)mres);
2605 mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen,
2610 #define DLM_MIGRATION_RETRY_MS 100
2613 * Should be called only after beginning the domain leave process.
2614 * There should not be any remaining locks on nonlocal lock resources,
2615 * and there should be no local locks left on locally mastered resources.
2617 * Called with the dlm spinlock held, may drop it to do migration, but
2618 * will re-acquire before exit.
2620 * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2622 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2625 int lock_dropped = 0;
2626 u8 target = O2NM_MAX_NODES;
2628 assert_spin_locked(&dlm->spinlock);
2630 spin_lock(&res->spinlock);
2631 if (dlm_is_lockres_migrateable(dlm, res))
2632 target = dlm_pick_migration_target(dlm, res);
2633 spin_unlock(&res->spinlock);
2635 if (target == O2NM_MAX_NODES)
2638 /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2639 spin_unlock(&dlm->spinlock);
2641 ret = dlm_migrate_lockres(dlm, res, target);
2643 mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2644 dlm->name, res->lockname.len, res->lockname.name,
2646 spin_lock(&dlm->spinlock);
2648 return lock_dropped;
2651 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2654 spin_lock(&dlm->ast_lock);
2655 spin_lock(&lock->spinlock);
2656 ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2657 spin_unlock(&lock->spinlock);
2658 spin_unlock(&dlm->ast_lock);
2662 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2663 struct dlm_lock_resource *res,
2667 spin_lock(&res->spinlock);
2668 can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2669 spin_unlock(&res->spinlock);
2671 /* target has died, so make the caller break out of the
2672 * wait_event, but caller must recheck the domain_map */
2673 spin_lock(&dlm->spinlock);
2674 if (!test_bit(mig_target, dlm->domain_map))
2676 spin_unlock(&dlm->spinlock);
2680 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2681 struct dlm_lock_resource *res)
2684 spin_lock(&res->spinlock);
2685 ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2686 spin_unlock(&res->spinlock);
2691 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2692 struct dlm_lock_resource *res,
2697 mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2698 res->lockname.len, res->lockname.name, dlm->node_num,
2700 /* need to set MIGRATING flag on lockres. this is done by
2701 * ensuring that all asts have been flushed for this lockres. */
2702 spin_lock(&res->spinlock);
2703 BUG_ON(res->migration_pending);
2704 res->migration_pending = 1;
2705 /* strategy is to reserve an extra ast then release
2706 * it below, letting the release do all of the work */
2707 __dlm_lockres_reserve_ast(res);
2708 spin_unlock(&res->spinlock);
2710 /* now flush all the pending asts */
2711 dlm_kick_thread(dlm, res);
2712 /* before waiting on DIRTY, block processes which may
2713 * try to dirty the lockres before MIGRATING is set */
2714 spin_lock(&res->spinlock);
2715 BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2716 res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2717 spin_unlock(&res->spinlock);
2718 /* now wait on any pending asts and the DIRTY state */
2719 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2720 dlm_lockres_release_ast(dlm, res);
2722 mlog(0, "about to wait on migration_wq, dirty=%s\n",
2723 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2724 /* if the extra ref we just put was the final one, this
2725 * will pass thru immediately. otherwise, we need to wait
2726 * for the last ast to finish. */
2728 ret = wait_event_interruptible_timeout(dlm->migration_wq,
2729 dlm_migration_can_proceed(dlm, res, target),
2730 msecs_to_jiffies(1000));
2732 mlog(0, "woken again: migrating? %s, dead? %s\n",
2733 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2734 test_bit(target, dlm->domain_map) ? "no":"yes");
2736 mlog(0, "all is well: migrating? %s, dead? %s\n",
2737 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2738 test_bit(target, dlm->domain_map) ? "no":"yes");
2740 if (!dlm_migration_can_proceed(dlm, res, target)) {
2741 mlog(0, "trying again...\n");
2746 /* did the target go down or die? */
2747 spin_lock(&dlm->spinlock);
2748 if (!test_bit(target, dlm->domain_map)) {
2749 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2753 spin_unlock(&dlm->spinlock);
2756 * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2757 * another try; otherwise, we are sure the MIGRATING state is there,
2758 * drop the unneded state which blocked threads trying to DIRTY
2760 spin_lock(&res->spinlock);
2761 BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2762 res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2764 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2765 spin_unlock(&res->spinlock);
2770 * o the DLM_LOCK_RES_MIGRATING flag is set if target not down
2771 * o there are no pending asts on this lockres
2772 * o all processes trying to reserve an ast on this
2773 * lockres must wait for the MIGRATING flag to clear
2778 /* last step in the migration process.
2779 * original master calls this to free all of the dlm_lock
2780 * structures that used to be for other nodes. */
2781 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2782 struct dlm_lock_resource *res)
2784 struct list_head *queue = &res->granted;
2786 struct dlm_lock *lock, *next;
2788 assert_spin_locked(&res->spinlock);
2790 BUG_ON(res->owner == dlm->node_num);
2792 for (i=0; i<3; i++) {
2793 list_for_each_entry_safe(lock, next, queue, list) {
2794 if (lock->ml.node != dlm->node_num) {
2795 mlog(0, "putting lock for node %u\n",
2797 /* be extra careful */
2798 BUG_ON(!list_empty(&lock->ast_list));
2799 BUG_ON(!list_empty(&lock->bast_list));
2800 BUG_ON(lock->ast_pending);
2801 BUG_ON(lock->bast_pending);
2802 dlm_lockres_clear_refmap_bit(dlm, res,
2804 list_del_init(&lock->list);
2806 /* In a normal unlock, we would have added a
2807 * DLM_UNLOCK_FREE_LOCK action. Force it. */
2815 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2816 if (bit >= O2NM_MAX_NODES)
2818 /* do not clear the local node reference, if there is a
2819 * process holding this, let it drop the ref itself */
2820 if (bit != dlm->node_num) {
2821 mlog(0, "%s:%.*s: node %u had a ref to this "
2822 "migrating lockres, clearing\n", dlm->name,
2823 res->lockname.len, res->lockname.name, bit);
2824 dlm_lockres_clear_refmap_bit(dlm, res, bit);
2831 * Pick a node to migrate the lock resource to. This function selects a
2832 * potential target based first on the locks and then on refmap. It skips
2833 * nodes that are in the process of exiting the domain.
2835 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2836 struct dlm_lock_resource *res)
2838 enum dlm_lockres_list idx;
2839 struct list_head *queue = &res->granted;
2840 struct dlm_lock *lock;
2842 u8 nodenum = O2NM_MAX_NODES;
2844 assert_spin_locked(&dlm->spinlock);
2845 assert_spin_locked(&res->spinlock);
2847 /* Go through all the locks */
2848 for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2849 queue = dlm_list_idx_to_ptr(res, idx);
2850 list_for_each_entry(lock, queue, list) {
2851 if (lock->ml.node == dlm->node_num)
2853 if (test_bit(lock->ml.node, dlm->exit_domain_map))
2855 nodenum = lock->ml.node;
2860 /* Go thru the refmap */
2863 noderef = find_next_bit(res->refmap, O2NM_MAX_NODES,
2865 if (noderef >= O2NM_MAX_NODES)
2867 if (noderef == dlm->node_num)
2869 if (test_bit(noderef, dlm->exit_domain_map))
2879 /* this is called by the new master once all lockres
2880 * data has been received */
2881 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2882 struct dlm_lock_resource *res,
2883 u8 master, u8 new_master,
2884 struct dlm_node_iter *iter)
2886 struct dlm_migrate_request migrate;
2887 int ret, skip, status = 0;
2890 memset(&migrate, 0, sizeof(migrate));
2891 migrate.namelen = res->lockname.len;
2892 memcpy(migrate.name, res->lockname.name, migrate.namelen);
2893 migrate.new_master = new_master;
2894 migrate.master = master;
2898 /* send message to all nodes, except the master and myself */
2899 while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2900 if (nodenum == master ||
2901 nodenum == new_master)
2904 /* We could race exit domain. If exited, skip. */
2905 spin_lock(&dlm->spinlock);
2906 skip = (!test_bit(nodenum, dlm->domain_map));
2907 spin_unlock(&dlm->spinlock);
2909 clear_bit(nodenum, iter->node_map);
2913 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2914 &migrate, sizeof(migrate), nodenum,
2917 mlog(ML_ERROR, "%s: res %.*s, Error %d send "
2918 "MIGRATE_REQUEST to node %u\n", dlm->name,
2919 migrate.namelen, migrate.name, ret, nodenum);
2920 if (!dlm_is_host_down(ret)) {
2921 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
2924 clear_bit(nodenum, iter->node_map);
2926 } else if (status < 0) {
2927 mlog(0, "migrate request (node %u) returned %d!\n",
2930 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
2931 /* during the migration request we short-circuited
2932 * the mastery of the lockres. make sure we have
2933 * a mastery ref for nodenum */
2934 mlog(0, "%s:%.*s: need ref for node %u\n",
2935 dlm->name, res->lockname.len, res->lockname.name,
2937 spin_lock(&res->spinlock);
2938 dlm_lockres_set_refmap_bit(dlm, res, nodenum);
2939 spin_unlock(&res->spinlock);
2946 mlog(0, "returning ret=%d\n", ret);
2951 /* if there is an existing mle for this lockres, we now know who the master is.
2952 * (the one who sent us *this* message) we can clear it up right away.
2953 * since the process that put the mle on the list still has a reference to it,
2954 * we can unhash it now, set the master and wake the process. as a result,
2955 * we will have no mle in the list to start with. now we can add an mle for
2956 * the migration and this should be the only one found for those scanning the
2958 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
2961 struct dlm_ctxt *dlm = data;
2962 struct dlm_lock_resource *res = NULL;
2963 struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
2964 struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
2966 unsigned int namelen, hash;
2972 name = migrate->name;
2973 namelen = migrate->namelen;
2974 hash = dlm_lockid_hash(name, namelen);
2976 /* preallocate.. if this fails, abort */
2977 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2984 /* check for pre-existing lock */
2985 spin_lock(&dlm->spinlock);
2986 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
2988 spin_lock(&res->spinlock);
2989 if (res->state & DLM_LOCK_RES_RECOVERING) {
2990 /* if all is working ok, this can only mean that we got
2991 * a migrate request from a node that we now see as
2992 * dead. what can we do here? drop it to the floor? */
2993 spin_unlock(&res->spinlock);
2994 mlog(ML_ERROR, "Got a migrate request, but the "
2995 "lockres is marked as recovering!");
2996 kmem_cache_free(dlm_mle_cache, mle);
2997 ret = -EINVAL; /* need a better solution */
3000 res->state |= DLM_LOCK_RES_MIGRATING;
3001 spin_unlock(&res->spinlock);
3004 spin_lock(&dlm->master_lock);
3005 /* ignore status. only nonzero status would BUG. */
3006 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3008 migrate->new_master,
3011 spin_unlock(&dlm->master_lock);
3013 spin_unlock(&dlm->spinlock);
3016 /* master is known, detach if not already detached */
3017 dlm_mle_detach_hb_events(dlm, oldmle);
3018 dlm_put_mle(oldmle);
3022 dlm_lockres_put(res);
3028 /* must be holding dlm->spinlock and dlm->master_lock
3029 * when adding a migration mle, we can clear any other mles
3030 * in the master list because we know with certainty that
3031 * the master is "master". so we remove any old mle from
3032 * the list after setting it's master field, and then add
3033 * the new migration mle. this way we can hold with the rule
3034 * of having only one mle for a given lock name at all times. */
3035 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3036 struct dlm_lock_resource *res,
3037 struct dlm_master_list_entry *mle,
3038 struct dlm_master_list_entry **oldmle,
3039 const char *name, unsigned int namelen,
3040 u8 new_master, u8 master)
3047 assert_spin_locked(&dlm->spinlock);
3048 assert_spin_locked(&dlm->master_lock);
3050 /* caller is responsible for any ref taken here on oldmle */
3051 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3053 struct dlm_master_list_entry *tmp = *oldmle;
3054 spin_lock(&tmp->spinlock);
3055 if (tmp->type == DLM_MLE_MIGRATION) {
3056 if (master == dlm->node_num) {
3057 /* ah another process raced me to it */
3058 mlog(0, "tried to migrate %.*s, but some "
3059 "process beat me to it\n",
3063 /* bad. 2 NODES are trying to migrate! */
3064 mlog(ML_ERROR, "migration error mle: "
3065 "master=%u new_master=%u // request: "
3066 "master=%u new_master=%u // "
3068 tmp->master, tmp->new_master,
3074 /* this is essentially what assert_master does */
3075 tmp->master = master;
3076 atomic_set(&tmp->woken, 1);
3078 /* remove it so that only one mle will be found */
3079 __dlm_unlink_mle(dlm, tmp);
3080 __dlm_mle_detach_hb_events(dlm, tmp);
3081 ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3082 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3083 "telling master to get ref for cleared out mle "
3084 "during migration\n", dlm->name, namelen, name,
3085 master, new_master);
3087 spin_unlock(&tmp->spinlock);
3090 /* now add a migration mle to the tail of the list */
3091 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3092 mle->new_master = new_master;
3093 /* the new master will be sending an assert master for this.
3094 * at that point we will get the refmap reference */
3095 mle->master = master;
3096 /* do this for consistency with other mle types */
3097 set_bit(new_master, mle->maybe_map);
3098 __dlm_insert_mle(dlm, mle);
3104 * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3106 static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3107 struct dlm_master_list_entry *mle)
3109 struct dlm_lock_resource *res;
3111 /* Find the lockres associated to the mle and set its owner to UNK */
3112 res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3115 spin_unlock(&dlm->master_lock);
3117 /* move lockres onto recovery list */
3118 spin_lock(&res->spinlock);
3119 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3120 dlm_move_lockres_to_recovery_list(dlm, res);
3121 spin_unlock(&res->spinlock);
3122 dlm_lockres_put(res);
3124 /* about to get rid of mle, detach from heartbeat */
3125 __dlm_mle_detach_hb_events(dlm, mle);
3128 spin_lock(&dlm->master_lock);
3130 spin_unlock(&dlm->master_lock);
3136 static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3137 struct dlm_master_list_entry *mle)
3139 __dlm_mle_detach_hb_events(dlm, mle);
3141 spin_lock(&mle->spinlock);
3142 __dlm_unlink_mle(dlm, mle);
3143 atomic_set(&mle->woken, 1);
3144 spin_unlock(&mle->spinlock);
3149 static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3150 struct dlm_master_list_entry *mle, u8 dead_node)
3154 BUG_ON(mle->type != DLM_MLE_BLOCK);
3156 spin_lock(&mle->spinlock);
3157 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3158 if (bit != dead_node) {
3159 mlog(0, "mle found, but dead node %u would not have been "
3160 "master\n", dead_node);
3161 spin_unlock(&mle->spinlock);
3163 /* Must drop the refcount by one since the assert_master will
3164 * never arrive. This may result in the mle being unlinked and
3165 * freed, but there may still be a process waiting in the
3166 * dlmlock path which is fine. */
3167 mlog(0, "node %u was expected master\n", dead_node);
3168 atomic_set(&mle->woken, 1);
3169 spin_unlock(&mle->spinlock);
3172 /* Do not need events any longer, so detach from heartbeat */
3173 __dlm_mle_detach_hb_events(dlm, mle);
3178 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3180 struct dlm_master_list_entry *mle;
3181 struct dlm_lock_resource *res;
3182 struct hlist_head *bucket;
3183 struct hlist_node *tmp;
3186 mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
3188 assert_spin_locked(&dlm->spinlock);
3190 /* clean the master list */
3191 spin_lock(&dlm->master_lock);
3192 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3193 bucket = dlm_master_hash(dlm, i);
3194 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3195 BUG_ON(mle->type != DLM_MLE_BLOCK &&
3196 mle->type != DLM_MLE_MASTER &&
3197 mle->type != DLM_MLE_MIGRATION);
3199 /* MASTER mles are initiated locally. The waiting
3200 * process will notice the node map change shortly.
3201 * Let that happen as normal. */
3202 if (mle->type == DLM_MLE_MASTER)
3205 /* BLOCK mles are initiated by other nodes. Need to
3206 * clean up if the dead node would have been the
3208 if (mle->type == DLM_MLE_BLOCK) {
3209 dlm_clean_block_mle(dlm, mle, dead_node);
3213 /* Everything else is a MIGRATION mle */
3215 /* The rule for MIGRATION mles is that the master
3216 * becomes UNKNOWN if *either* the original or the new
3217 * master dies. All UNKNOWN lockres' are sent to
3218 * whichever node becomes the recovery master. The new
3219 * master is responsible for determining if there is
3220 * still a master for this lockres, or if he needs to
3221 * take over mastery. Either way, this node should
3222 * expect another message to resolve this. */
3224 if (mle->master != dead_node &&
3225 mle->new_master != dead_node)
3228 /* If we have reached this point, this mle needs to be
3229 * removed from the list and freed. */
3230 dlm_clean_migration_mle(dlm, mle);
3232 mlog(0, "%s: node %u died during migration from "
3233 "%u to %u!\n", dlm->name, dead_node, mle->master,
3236 /* If we find a lockres associated with the mle, we've
3237 * hit this rare case that messes up our lock ordering.
3238 * If so, we need to drop the master lock so that we can
3239 * take the lockres lock, meaning that we will have to
3240 * restart from the head of list. */
3241 res = dlm_reset_mleres_owner(dlm, mle);
3246 /* This may be the last reference */
3250 spin_unlock(&dlm->master_lock);
3253 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3256 struct dlm_node_iter iter;
3259 spin_lock(&dlm->spinlock);
3260 dlm_node_iter_init(dlm->domain_map, &iter);
3261 clear_bit(old_master, iter.node_map);
3262 clear_bit(dlm->node_num, iter.node_map);
3263 spin_unlock(&dlm->spinlock);
3265 /* ownership of the lockres is changing. account for the
3266 * mastery reference here since old_master will briefly have
3267 * a reference after the migration completes */
3268 spin_lock(&res->spinlock);
3269 dlm_lockres_set_refmap_bit(dlm, res, old_master);
3270 spin_unlock(&res->spinlock);
3272 mlog(0, "now time to do a migrate request to other nodes\n");
3273 ret = dlm_do_migrate_request(dlm, res, old_master,
3274 dlm->node_num, &iter);
3280 mlog(0, "doing assert master of %.*s to all except the original node\n",
3281 res->lockname.len, res->lockname.name);
3282 /* this call now finishes out the nodemap
3283 * even if one or more nodes die */
3284 ret = dlm_do_assert_master(dlm, res, iter.node_map,
3285 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3287 /* no longer need to retry. all living nodes contacted. */
3292 memset(iter.node_map, 0, sizeof(iter.node_map));
3293 set_bit(old_master, iter.node_map);
3294 mlog(0, "doing assert master of %.*s back to %u\n",
3295 res->lockname.len, res->lockname.name, old_master);
3296 ret = dlm_do_assert_master(dlm, res, iter.node_map,
3297 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3299 mlog(0, "assert master to original master failed "
3301 /* the only nonzero status here would be because of
3302 * a dead original node. we're done. */
3306 /* all done, set the owner, clear the flag */
3307 spin_lock(&res->spinlock);
3308 dlm_set_lockres_owner(dlm, res, dlm->node_num);
3309 res->state &= ~DLM_LOCK_RES_MIGRATING;
3310 spin_unlock(&res->spinlock);
3311 /* re-dirty it on the new master */
3312 dlm_kick_thread(dlm, res);
3319 * LOCKRES AST REFCOUNT
3320 * this is integral to migration
3323 /* for future intent to call an ast, reserve one ahead of time.
3324 * this should be called only after waiting on the lockres
3325 * with dlm_wait_on_lockres, and while still holding the
3326 * spinlock after the call. */
3327 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3329 assert_spin_locked(&res->spinlock);
3330 if (res->state & DLM_LOCK_RES_MIGRATING) {
3331 __dlm_print_one_lock_resource(res);
3333 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3335 atomic_inc(&res->asts_reserved);
3339 * used to drop the reserved ast, either because it went unused,
3340 * or because the ast/bast was actually called.
3342 * also, if there is a pending migration on this lockres,
3343 * and this was the last pending ast on the lockres,
3344 * atomically set the MIGRATING flag before we drop the lock.
3345 * this is how we ensure that migration can proceed with no
3346 * asts in progress. note that it is ok if the state of the
3347 * queues is such that a lock should be granted in the future
3348 * or that a bast should be fired, because the new master will
3349 * shuffle the lists on this lockres as soon as it is migrated.
3351 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3352 struct dlm_lock_resource *res)
3354 if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3357 if (!res->migration_pending) {
3358 spin_unlock(&res->spinlock);
3362 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3363 res->migration_pending = 0;
3364 res->state |= DLM_LOCK_RES_MIGRATING;
3365 spin_unlock(&res->spinlock);
3367 wake_up(&dlm->migration_wq);
3370 void dlm_force_free_mles(struct dlm_ctxt *dlm)
3373 struct hlist_head *bucket;
3374 struct dlm_master_list_entry *mle;
3375 struct hlist_node *tmp;
3378 * We notified all other nodes that we are exiting the domain and
3379 * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3380 * around we force free them and wake any processes that are waiting
3383 spin_lock(&dlm->spinlock);
3384 spin_lock(&dlm->master_lock);
3386 BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3387 BUG_ON((find_next_bit(dlm->domain_map, O2NM_MAX_NODES, 0) < O2NM_MAX_NODES));
3389 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3390 bucket = dlm_master_hash(dlm, i);
3391 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3392 if (mle->type != DLM_MLE_BLOCK) {
3393 mlog(ML_ERROR, "bad mle: %p\n", mle);
3394 dlm_print_one_mle(mle);
3396 atomic_set(&mle->woken, 1);
3399 __dlm_unlink_mle(dlm, mle);
3400 __dlm_mle_detach_hb_events(dlm, mle);
3404 spin_unlock(&dlm->master_lock);
3405 spin_unlock(&dlm->spinlock);