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/utsname.h>
34 #include <linux/init.h>
35 #include <linux/sysctl.h>
36 #include <linux/random.h>
37 #include <linux/blkdev.h>
38 #include <linux/socket.h>
39 #include <linux/inet.h>
40 #include <linux/spinlock.h>
41 #include <linux/delay.h>
44 #include "cluster/heartbeat.h"
45 #include "cluster/nodemanager.h"
46 #include "cluster/tcp.h"
49 #include "dlmcommon.h"
50 #include "dlmdomain.h"
52 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
53 #include "cluster/masklog.h"
64 u8 name[DLM_LOCKID_NAME_MAX];
67 struct dlm_master_list_entry
69 struct list_head list;
70 struct list_head hb_events;
77 unsigned long maybe_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
78 unsigned long vote_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
79 unsigned long response_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
80 unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
83 enum dlm_mle_type type;
84 struct o2hb_callback_func mle_hb_up;
85 struct o2hb_callback_func mle_hb_down;
87 struct dlm_lock_resource *res;
88 struct dlm_lock_name name;
92 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
93 struct dlm_master_list_entry *mle,
94 struct o2nm_node *node,
96 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
97 struct dlm_master_list_entry *mle,
98 struct o2nm_node *node,
101 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
102 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
103 struct dlm_lock_resource *res,
104 void *nodemap, u32 flags);
106 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
107 struct dlm_master_list_entry *mle,
109 unsigned int namelen)
111 struct dlm_lock_resource *res;
116 if (mle->type == DLM_MLE_BLOCK ||
117 mle->type == DLM_MLE_MIGRATION) {
118 if (namelen != mle->u.name.len ||
119 memcmp(name, mle->u.name.name, namelen)!=0)
123 if (namelen != res->lockname.len ||
124 memcmp(res->lockname.name, name, namelen) != 0)
130 #define dlm_print_nodemap(m) _dlm_print_nodemap(m,#m)
131 static void _dlm_print_nodemap(unsigned long *map, const char *mapname)
134 printk("%s=[ ", mapname);
135 for (i=0; i<O2NM_MAX_NODES; i++)
136 if (test_bit(i, map))
141 static void dlm_print_one_mle(struct dlm_master_list_entry *mle)
147 unsigned int namelen;
150 unsigned long *maybe = mle->maybe_map,
151 *vote = mle->vote_map,
152 *resp = mle->response_map,
153 *node = mle->node_map;
156 if (mle->type == DLM_MLE_BLOCK)
158 else if (mle->type == DLM_MLE_MASTER)
162 refs = atomic_read(&k->refcount);
163 master = mle->master;
164 attached = (list_empty(&mle->hb_events) ? 'N' : 'Y');
166 if (mle->type != DLM_MLE_MASTER) {
167 namelen = mle->u.name.len;
168 name = mle->u.name.name;
170 namelen = mle->u.res->lockname.len;
171 name = mle->u.res->lockname.name;
174 mlog(ML_NOTICE, "%.*s: %3s refs=%3d mas=%3u new=%3u evt=%c inuse=%d ",
175 namelen, name, type, refs, master, mle->new_master, attached,
177 dlm_print_nodemap(maybe);
179 dlm_print_nodemap(vote);
181 dlm_print_nodemap(resp);
183 dlm_print_nodemap(node);
189 /* Code here is included but defined out as it aids debugging */
191 static void dlm_dump_mles(struct dlm_ctxt *dlm)
193 struct dlm_master_list_entry *mle;
194 struct list_head *iter;
196 mlog(ML_NOTICE, "dumping all mles for domain %s:\n", dlm->name);
197 spin_lock(&dlm->master_lock);
198 list_for_each(iter, &dlm->master_list) {
199 mle = list_entry(iter, struct dlm_master_list_entry, list);
200 dlm_print_one_mle(mle);
202 spin_unlock(&dlm->master_lock);
205 int dlm_dump_all_mles(const char __user *data, unsigned int len)
207 struct list_head *iter;
208 struct dlm_ctxt *dlm;
210 spin_lock(&dlm_domain_lock);
211 list_for_each(iter, &dlm_domains) {
212 dlm = list_entry (iter, struct dlm_ctxt, list);
213 mlog(ML_NOTICE, "found dlm: %p, name=%s\n", dlm, dlm->name);
216 spin_unlock(&dlm_domain_lock);
219 EXPORT_SYMBOL_GPL(dlm_dump_all_mles);
224 static struct kmem_cache *dlm_mle_cache = NULL;
227 static void dlm_mle_release(struct kref *kref);
228 static void dlm_init_mle(struct dlm_master_list_entry *mle,
229 enum dlm_mle_type type,
230 struct dlm_ctxt *dlm,
231 struct dlm_lock_resource *res,
233 unsigned int namelen);
234 static void dlm_put_mle(struct dlm_master_list_entry *mle);
235 static void __dlm_put_mle(struct dlm_master_list_entry *mle);
236 static int dlm_find_mle(struct dlm_ctxt *dlm,
237 struct dlm_master_list_entry **mle,
238 char *name, unsigned int namelen);
240 static int dlm_do_master_request(struct dlm_lock_resource *res,
241 struct dlm_master_list_entry *mle, int to);
244 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
245 struct dlm_lock_resource *res,
246 struct dlm_master_list_entry *mle,
248 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
249 struct dlm_lock_resource *res,
250 struct dlm_master_list_entry *mle,
252 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
253 struct dlm_lock_resource *res,
254 struct dlm_master_list_entry *mle,
255 struct dlm_master_list_entry **oldmle,
256 const char *name, unsigned int namelen,
257 u8 new_master, u8 master);
259 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
260 struct dlm_lock_resource *res);
261 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
262 struct dlm_lock_resource *res);
263 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
264 struct dlm_lock_resource *res,
266 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
267 struct dlm_lock_resource *res);
270 int dlm_is_host_down(int errno)
287 case -EINVAL: /* if returned from our tcp code,
288 this means there is no socket */
296 * MASTER LIST FUNCTIONS
301 * regarding master list entries and heartbeat callbacks:
303 * in order to avoid sleeping and allocation that occurs in
304 * heartbeat, master list entries are simply attached to the
305 * dlm's established heartbeat callbacks. the mle is attached
306 * when it is created, and since the dlm->spinlock is held at
307 * that time, any heartbeat event will be properly discovered
308 * by the mle. the mle needs to be detached from the
309 * dlm->mle_hb_events list as soon as heartbeat events are no
310 * longer useful to the mle, and before the mle is freed.
312 * as a general rule, heartbeat events are no longer needed by
313 * the mle once an "answer" regarding the lock master has been
316 static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
317 struct dlm_master_list_entry *mle)
319 assert_spin_locked(&dlm->spinlock);
321 list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
325 static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
326 struct dlm_master_list_entry *mle)
328 if (!list_empty(&mle->hb_events))
329 list_del_init(&mle->hb_events);
333 static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
334 struct dlm_master_list_entry *mle)
336 spin_lock(&dlm->spinlock);
337 __dlm_mle_detach_hb_events(dlm, mle);
338 spin_unlock(&dlm->spinlock);
341 static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
343 struct dlm_ctxt *dlm;
346 assert_spin_locked(&dlm->spinlock);
347 assert_spin_locked(&dlm->master_lock);
349 kref_get(&mle->mle_refs);
352 static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
354 struct dlm_ctxt *dlm;
357 spin_lock(&dlm->spinlock);
358 spin_lock(&dlm->master_lock);
361 spin_unlock(&dlm->master_lock);
362 spin_unlock(&dlm->spinlock);
366 /* remove from list and free */
367 static void __dlm_put_mle(struct dlm_master_list_entry *mle)
369 struct dlm_ctxt *dlm;
372 assert_spin_locked(&dlm->spinlock);
373 assert_spin_locked(&dlm->master_lock);
374 if (!atomic_read(&mle->mle_refs.refcount)) {
375 /* this may or may not crash, but who cares.
377 mlog(ML_ERROR, "bad mle: %p\n", mle);
378 dlm_print_one_mle(mle);
381 kref_put(&mle->mle_refs, dlm_mle_release);
385 /* must not have any spinlocks coming in */
386 static void dlm_put_mle(struct dlm_master_list_entry *mle)
388 struct dlm_ctxt *dlm;
391 spin_lock(&dlm->spinlock);
392 spin_lock(&dlm->master_lock);
394 spin_unlock(&dlm->master_lock);
395 spin_unlock(&dlm->spinlock);
398 static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
400 kref_get(&mle->mle_refs);
403 static void dlm_init_mle(struct dlm_master_list_entry *mle,
404 enum dlm_mle_type type,
405 struct dlm_ctxt *dlm,
406 struct dlm_lock_resource *res,
408 unsigned int namelen)
410 assert_spin_locked(&dlm->spinlock);
414 INIT_LIST_HEAD(&mle->list);
415 INIT_LIST_HEAD(&mle->hb_events);
416 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
417 spin_lock_init(&mle->spinlock);
418 init_waitqueue_head(&mle->wq);
419 atomic_set(&mle->woken, 0);
420 kref_init(&mle->mle_refs);
421 memset(mle->response_map, 0, sizeof(mle->response_map));
422 mle->master = O2NM_MAX_NODES;
423 mle->new_master = O2NM_MAX_NODES;
426 if (mle->type == DLM_MLE_MASTER) {
429 } else if (mle->type == DLM_MLE_BLOCK) {
431 memcpy(mle->u.name.name, name, namelen);
432 mle->u.name.len = namelen;
433 } else /* DLM_MLE_MIGRATION */ {
435 memcpy(mle->u.name.name, name, namelen);
436 mle->u.name.len = namelen;
439 /* copy off the node_map and register hb callbacks on our copy */
440 memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
441 memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
442 clear_bit(dlm->node_num, mle->vote_map);
443 clear_bit(dlm->node_num, mle->node_map);
445 /* attach the mle to the domain node up/down events */
446 __dlm_mle_attach_hb_events(dlm, mle);
450 /* returns 1 if found, 0 if not */
451 static int dlm_find_mle(struct dlm_ctxt *dlm,
452 struct dlm_master_list_entry **mle,
453 char *name, unsigned int namelen)
455 struct dlm_master_list_entry *tmpmle;
456 struct list_head *iter;
458 assert_spin_locked(&dlm->master_lock);
460 list_for_each(iter, &dlm->master_list) {
461 tmpmle = list_entry(iter, struct dlm_master_list_entry, list);
462 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
471 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
473 struct dlm_master_list_entry *mle;
474 struct list_head *iter;
476 assert_spin_locked(&dlm->spinlock);
478 list_for_each(iter, &dlm->mle_hb_events) {
479 mle = list_entry(iter, struct dlm_master_list_entry,
482 dlm_mle_node_up(dlm, mle, NULL, idx);
484 dlm_mle_node_down(dlm, mle, NULL, idx);
488 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
489 struct dlm_master_list_entry *mle,
490 struct o2nm_node *node, int idx)
492 spin_lock(&mle->spinlock);
494 if (!test_bit(idx, mle->node_map))
495 mlog(0, "node %u already removed from nodemap!\n", idx);
497 clear_bit(idx, mle->node_map);
499 spin_unlock(&mle->spinlock);
502 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
503 struct dlm_master_list_entry *mle,
504 struct o2nm_node *node, int idx)
506 spin_lock(&mle->spinlock);
508 if (test_bit(idx, mle->node_map))
509 mlog(0, "node %u already in node map!\n", idx);
511 set_bit(idx, mle->node_map);
513 spin_unlock(&mle->spinlock);
517 int dlm_init_mle_cache(void)
519 dlm_mle_cache = kmem_cache_create("dlm_mle_cache",
520 sizeof(struct dlm_master_list_entry),
521 0, SLAB_HWCACHE_ALIGN,
523 if (dlm_mle_cache == NULL)
528 void dlm_destroy_mle_cache(void)
531 kmem_cache_destroy(dlm_mle_cache);
534 static void dlm_mle_release(struct kref *kref)
536 struct dlm_master_list_entry *mle;
537 struct dlm_ctxt *dlm;
541 mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
544 if (mle->type != DLM_MLE_MASTER) {
545 mlog(0, "calling mle_release for %.*s, type %d\n",
546 mle->u.name.len, mle->u.name.name, mle->type);
548 mlog(0, "calling mle_release for %.*s, type %d\n",
549 mle->u.res->lockname.len,
550 mle->u.res->lockname.name, mle->type);
552 assert_spin_locked(&dlm->spinlock);
553 assert_spin_locked(&dlm->master_lock);
555 /* remove from list if not already */
556 if (!list_empty(&mle->list))
557 list_del_init(&mle->list);
559 /* detach the mle from the domain node up/down events */
560 __dlm_mle_detach_hb_events(dlm, mle);
562 /* NOTE: kfree under spinlock here.
563 * if this is bad, we can move this to a freelist. */
564 kmem_cache_free(dlm_mle_cache, mle);
569 * LOCK RESOURCE FUNCTIONS
572 static void dlm_set_lockres_owner(struct dlm_ctxt *dlm,
573 struct dlm_lock_resource *res,
576 assert_spin_locked(&res->spinlock);
578 mlog_entry("%.*s, %u\n", res->lockname.len, res->lockname.name, owner);
580 if (owner == dlm->node_num)
581 atomic_inc(&dlm->local_resources);
582 else if (owner == DLM_LOCK_RES_OWNER_UNKNOWN)
583 atomic_inc(&dlm->unknown_resources);
585 atomic_inc(&dlm->remote_resources);
590 void dlm_change_lockres_owner(struct dlm_ctxt *dlm,
591 struct dlm_lock_resource *res, u8 owner)
593 assert_spin_locked(&res->spinlock);
595 if (owner == res->owner)
598 if (res->owner == dlm->node_num)
599 atomic_dec(&dlm->local_resources);
600 else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN)
601 atomic_dec(&dlm->unknown_resources);
603 atomic_dec(&dlm->remote_resources);
605 dlm_set_lockres_owner(dlm, res, owner);
609 static void dlm_lockres_release(struct kref *kref)
611 struct dlm_lock_resource *res;
613 res = container_of(kref, struct dlm_lock_resource, refs);
615 /* This should not happen -- all lockres' have a name
616 * associated with them at init time. */
617 BUG_ON(!res->lockname.name);
619 mlog(0, "destroying lockres %.*s\n", res->lockname.len,
622 if (!hlist_unhashed(&res->hash_node) ||
623 !list_empty(&res->granted) ||
624 !list_empty(&res->converting) ||
625 !list_empty(&res->blocked) ||
626 !list_empty(&res->dirty) ||
627 !list_empty(&res->recovering) ||
628 !list_empty(&res->purge)) {
630 "Going to BUG for resource %.*s."
631 " We're on a list! [%c%c%c%c%c%c%c]\n",
632 res->lockname.len, res->lockname.name,
633 !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
634 !list_empty(&res->granted) ? 'G' : ' ',
635 !list_empty(&res->converting) ? 'C' : ' ',
636 !list_empty(&res->blocked) ? 'B' : ' ',
637 !list_empty(&res->dirty) ? 'D' : ' ',
638 !list_empty(&res->recovering) ? 'R' : ' ',
639 !list_empty(&res->purge) ? 'P' : ' ');
641 dlm_print_one_lock_resource(res);
644 /* By the time we're ready to blow this guy away, we shouldn't
645 * be on any lists. */
646 BUG_ON(!hlist_unhashed(&res->hash_node));
647 BUG_ON(!list_empty(&res->granted));
648 BUG_ON(!list_empty(&res->converting));
649 BUG_ON(!list_empty(&res->blocked));
650 BUG_ON(!list_empty(&res->dirty));
651 BUG_ON(!list_empty(&res->recovering));
652 BUG_ON(!list_empty(&res->purge));
654 kfree(res->lockname.name);
659 void dlm_lockres_put(struct dlm_lock_resource *res)
661 kref_put(&res->refs, dlm_lockres_release);
664 static void dlm_init_lockres(struct dlm_ctxt *dlm,
665 struct dlm_lock_resource *res,
666 const char *name, unsigned int namelen)
670 /* If we memset here, we lose our reference to the kmalloc'd
671 * res->lockname.name, so be sure to init every field
674 qname = (char *) res->lockname.name;
675 memcpy(qname, name, namelen);
677 res->lockname.len = namelen;
678 res->lockname.hash = dlm_lockid_hash(name, namelen);
680 init_waitqueue_head(&res->wq);
681 spin_lock_init(&res->spinlock);
682 INIT_HLIST_NODE(&res->hash_node);
683 INIT_LIST_HEAD(&res->granted);
684 INIT_LIST_HEAD(&res->converting);
685 INIT_LIST_HEAD(&res->blocked);
686 INIT_LIST_HEAD(&res->dirty);
687 INIT_LIST_HEAD(&res->recovering);
688 INIT_LIST_HEAD(&res->purge);
689 atomic_set(&res->asts_reserved, 0);
690 res->migration_pending = 0;
691 res->inflight_locks = 0;
693 kref_init(&res->refs);
695 /* just for consistency */
696 spin_lock(&res->spinlock);
697 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
698 spin_unlock(&res->spinlock);
700 res->state = DLM_LOCK_RES_IN_PROGRESS;
704 memset(res->lvb, 0, DLM_LVB_LEN);
705 memset(res->refmap, 0, sizeof(res->refmap));
708 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
710 unsigned int namelen)
712 struct dlm_lock_resource *res;
714 res = kmalloc(sizeof(struct dlm_lock_resource), GFP_NOFS);
718 res->lockname.name = kmalloc(namelen, GFP_NOFS);
719 if (!res->lockname.name) {
724 dlm_init_lockres(dlm, res, name, namelen);
728 void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
729 struct dlm_lock_resource *res,
735 assert_spin_locked(&res->spinlock);
737 if (!test_bit(dlm->node_num, res->refmap)) {
738 BUG_ON(res->inflight_locks != 0);
739 dlm_lockres_set_refmap_bit(dlm->node_num, res);
741 res->inflight_locks++;
742 mlog(0, "%s:%.*s: inflight++: now %u\n",
743 dlm->name, res->lockname.len, res->lockname.name,
744 res->inflight_locks);
747 void __dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
748 struct dlm_lock_resource *res,
752 assert_spin_locked(&res->spinlock);
754 BUG_ON(res->inflight_locks == 0);
755 res->inflight_locks--;
756 mlog(0, "%s:%.*s: inflight--: now %u\n",
757 dlm->name, res->lockname.len, res->lockname.name,
758 res->inflight_locks);
759 if (res->inflight_locks == 0)
760 dlm_lockres_clear_refmap_bit(dlm->node_num, res);
765 * lookup a lock resource by name.
766 * may already exist in the hashtable.
767 * lockid is null terminated
769 * if not, allocate enough for the lockres and for
770 * the temporary structure used in doing the mastering.
772 * also, do a lookup in the dlm->master_list to see
773 * if another node has begun mastering the same lock.
774 * if so, there should be a block entry in there
775 * for this name, and we should *not* attempt to master
776 * the lock here. need to wait around for that node
777 * to assert_master (or die).
780 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
785 struct dlm_lock_resource *tmpres=NULL, *res=NULL;
786 struct dlm_master_list_entry *mle = NULL;
787 struct dlm_master_list_entry *alloc_mle = NULL;
790 struct dlm_node_iter iter;
793 int bit, wait_on_recovery = 0;
794 int drop_inflight_if_nonlocal = 0;
798 hash = dlm_lockid_hash(lockid, namelen);
800 mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
803 spin_lock(&dlm->spinlock);
804 tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
806 int dropping_ref = 0;
808 spin_lock(&tmpres->spinlock);
809 if (tmpres->owner == dlm->node_num) {
810 BUG_ON(tmpres->state & DLM_LOCK_RES_DROPPING_REF);
811 dlm_lockres_grab_inflight_ref(dlm, tmpres);
812 } else if (tmpres->state & DLM_LOCK_RES_DROPPING_REF)
814 spin_unlock(&tmpres->spinlock);
815 spin_unlock(&dlm->spinlock);
817 /* wait until done messaging the master, drop our ref to allow
818 * the lockres to be purged, start over. */
820 spin_lock(&tmpres->spinlock);
821 __dlm_wait_on_lockres_flags(tmpres, DLM_LOCK_RES_DROPPING_REF);
822 spin_unlock(&tmpres->spinlock);
823 dlm_lockres_put(tmpres);
828 mlog(0, "found in hash!\n");
830 dlm_lockres_put(res);
836 spin_unlock(&dlm->spinlock);
837 mlog(0, "allocating a new resource\n");
838 /* nothing found and we need to allocate one. */
839 alloc_mle = (struct dlm_master_list_entry *)
840 kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
843 res = dlm_new_lockres(dlm, lockid, namelen);
849 mlog(0, "no lockres found, allocated our own: %p\n", res);
851 if (flags & LKM_LOCAL) {
852 /* caller knows it's safe to assume it's not mastered elsewhere
853 * DONE! return right away */
854 spin_lock(&res->spinlock);
855 dlm_change_lockres_owner(dlm, res, dlm->node_num);
856 __dlm_insert_lockres(dlm, res);
857 dlm_lockres_grab_inflight_ref(dlm, res);
858 spin_unlock(&res->spinlock);
859 spin_unlock(&dlm->spinlock);
860 /* lockres still marked IN_PROGRESS */
864 /* check master list to see if another node has started mastering it */
865 spin_lock(&dlm->master_lock);
867 /* if we found a block, wait for lock to be mastered by another node */
868 blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
871 if (mle->type == DLM_MLE_MASTER) {
872 mlog(ML_ERROR, "master entry for nonexistent lock!\n");
875 mig = (mle->type == DLM_MLE_MIGRATION);
876 /* if there is a migration in progress, let the migration
877 * finish before continuing. we can wait for the absence
878 * of the MIGRATION mle: either the migrate finished or
879 * one of the nodes died and the mle was cleaned up.
880 * if there is a BLOCK here, but it already has a master
881 * set, we are too late. the master does not have a ref
882 * for us in the refmap. detach the mle and drop it.
883 * either way, go back to the top and start over. */
884 if (mig || mle->master != O2NM_MAX_NODES) {
885 BUG_ON(mig && mle->master == dlm->node_num);
886 /* we arrived too late. the master does not
887 * have a ref for us. retry. */
888 mlog(0, "%s:%.*s: late on %s\n",
889 dlm->name, namelen, lockid,
890 mig ? "MIGRATION" : "BLOCK");
891 spin_unlock(&dlm->master_lock);
892 spin_unlock(&dlm->spinlock);
894 /* master is known, detach */
896 dlm_mle_detach_hb_events(dlm, mle);
899 /* this is lame, but we cant wait on either
900 * the mle or lockres waitqueue here */
906 /* go ahead and try to master lock on this node */
908 /* make sure this does not get freed below */
910 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
911 set_bit(dlm->node_num, mle->maybe_map);
912 list_add(&mle->list, &dlm->master_list);
914 /* still holding the dlm spinlock, check the recovery map
915 * to see if there are any nodes that still need to be
916 * considered. these will not appear in the mle nodemap
917 * but they might own this lockres. wait on them. */
918 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
919 if (bit < O2NM_MAX_NODES) {
920 mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to"
921 "recover before lock mastery can begin\n",
922 dlm->name, namelen, (char *)lockid, bit);
923 wait_on_recovery = 1;
927 /* at this point there is either a DLM_MLE_BLOCK or a
928 * DLM_MLE_MASTER on the master list, so it's safe to add the
929 * lockres to the hashtable. anyone who finds the lock will
930 * still have to wait on the IN_PROGRESS. */
932 /* finally add the lockres to its hash bucket */
933 __dlm_insert_lockres(dlm, res);
934 /* since this lockres is new it doesnt not require the spinlock */
935 dlm_lockres_grab_inflight_ref_new(dlm, res);
937 /* if this node does not become the master make sure to drop
938 * this inflight reference below */
939 drop_inflight_if_nonlocal = 1;
941 /* get an extra ref on the mle in case this is a BLOCK
942 * if so, the creator of the BLOCK may try to put the last
943 * ref at this time in the assert master handler, so we
944 * need an extra one to keep from a bad ptr deref. */
945 dlm_get_mle_inuse(mle);
946 spin_unlock(&dlm->master_lock);
947 spin_unlock(&dlm->spinlock);
950 while (wait_on_recovery) {
951 /* any cluster changes that occurred after dropping the
952 * dlm spinlock would be detectable be a change on the mle,
953 * so we only need to clear out the recovery map once. */
954 if (dlm_is_recovery_lock(lockid, namelen)) {
955 mlog(ML_NOTICE, "%s: recovery map is not empty, but "
956 "must master $RECOVERY lock now\n", dlm->name);
957 if (!dlm_pre_master_reco_lockres(dlm, res))
958 wait_on_recovery = 0;
960 mlog(0, "%s: waiting 500ms for heartbeat state "
961 "change\n", dlm->name);
967 dlm_kick_recovery_thread(dlm);
969 dlm_wait_for_recovery(dlm);
971 spin_lock(&dlm->spinlock);
972 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
973 if (bit < O2NM_MAX_NODES) {
974 mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to"
975 "recover before lock mastery can begin\n",
976 dlm->name, namelen, (char *)lockid, bit);
977 wait_on_recovery = 1;
979 wait_on_recovery = 0;
980 spin_unlock(&dlm->spinlock);
982 if (wait_on_recovery)
983 dlm_wait_for_node_recovery(dlm, bit, 10000);
986 /* must wait for lock to be mastered elsewhere */
991 dlm_node_iter_init(mle->vote_map, &iter);
992 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
993 ret = dlm_do_master_request(res, mle, nodenum);
996 if (mle->master != O2NM_MAX_NODES) {
997 /* found a master ! */
998 if (mle->master <= nodenum)
1000 /* if our master request has not reached the master
1001 * yet, keep going until it does. this is how the
1002 * master will know that asserts are needed back to
1003 * the lower nodes. */
1004 mlog(0, "%s:%.*s: requests only up to %u but master "
1005 "is %u, keep going\n", dlm->name, namelen,
1006 lockid, nodenum, mle->master);
1011 /* keep going until the response map includes all nodes */
1012 ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
1014 wait_on_recovery = 1;
1015 mlog(0, "%s:%.*s: node map changed, redo the "
1016 "master request now, blocked=%d\n",
1017 dlm->name, res->lockname.len,
1018 res->lockname.name, blocked);
1020 mlog(ML_ERROR, "%s:%.*s: spinning on "
1021 "dlm_wait_for_lock_mastery, blocked=%d\n",
1022 dlm->name, res->lockname.len,
1023 res->lockname.name, blocked);
1024 dlm_print_one_lock_resource(res);
1025 dlm_print_one_mle(mle);
1031 mlog(0, "lockres mastered by %u\n", res->owner);
1032 /* make sure we never continue without this */
1033 BUG_ON(res->owner == O2NM_MAX_NODES);
1035 /* master is known, detach if not already detached */
1036 dlm_mle_detach_hb_events(dlm, mle);
1038 /* put the extra ref */
1039 dlm_put_mle_inuse(mle);
1042 spin_lock(&res->spinlock);
1043 if (res->owner != dlm->node_num && drop_inflight_if_nonlocal)
1044 dlm_lockres_drop_inflight_ref(dlm, res);
1045 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1046 spin_unlock(&res->spinlock);
1050 /* need to free the unused mle */
1052 kmem_cache_free(dlm_mle_cache, alloc_mle);
1058 #define DLM_MASTERY_TIMEOUT_MS 5000
1060 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
1061 struct dlm_lock_resource *res,
1062 struct dlm_master_list_entry *mle,
1067 int map_changed, voting_done;
1074 /* check if another node has already become the owner */
1075 spin_lock(&res->spinlock);
1076 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1077 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1078 res->lockname.len, res->lockname.name, res->owner);
1079 spin_unlock(&res->spinlock);
1080 /* this will cause the master to re-assert across
1081 * the whole cluster, freeing up mles */
1082 if (res->owner != dlm->node_num) {
1083 ret = dlm_do_master_request(res, mle, res->owner);
1085 /* give recovery a chance to run */
1086 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1094 spin_unlock(&res->spinlock);
1096 spin_lock(&mle->spinlock);
1098 map_changed = (memcmp(mle->vote_map, mle->node_map,
1099 sizeof(mle->vote_map)) != 0);
1100 voting_done = (memcmp(mle->vote_map, mle->response_map,
1101 sizeof(mle->vote_map)) == 0);
1103 /* restart if we hit any errors */
1106 mlog(0, "%s: %.*s: node map changed, restarting\n",
1107 dlm->name, res->lockname.len, res->lockname.name);
1108 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1109 b = (mle->type == DLM_MLE_BLOCK);
1110 if ((*blocked && !b) || (!*blocked && b)) {
1111 mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1112 dlm->name, res->lockname.len, res->lockname.name,
1116 spin_unlock(&mle->spinlock);
1121 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1122 "rechecking now\n", dlm->name, res->lockname.len,
1123 res->lockname.name);
1127 mlog(0, "map not changed and voting not done "
1128 "for %s:%.*s\n", dlm->name, res->lockname.len,
1129 res->lockname.name);
1133 if (m != O2NM_MAX_NODES) {
1134 /* another node has done an assert!
1139 /* have all nodes responded? */
1140 if (voting_done && !*blocked) {
1141 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1142 if (dlm->node_num <= bit) {
1143 /* my node number is lowest.
1144 * now tell other nodes that I am
1145 * mastering this. */
1146 mle->master = dlm->node_num;
1147 /* ref was grabbed in get_lock_resource
1148 * will be dropped in dlmlock_master */
1152 /* if voting is done, but we have not received
1153 * an assert master yet, we must sleep */
1157 spin_unlock(&mle->spinlock);
1159 /* sleep if we haven't finished voting yet */
1161 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1164 if (atomic_read(&mle->mle_refs.refcount) < 2)
1165 mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1166 atomic_read(&mle->mle_refs.refcount),
1167 res->lockname.len, res->lockname.name);
1169 atomic_set(&mle->woken, 0);
1170 (void)wait_event_timeout(mle->wq,
1171 (atomic_read(&mle->woken) == 1),
1173 if (res->owner == O2NM_MAX_NODES) {
1174 mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1175 res->lockname.len, res->lockname.name);
1178 mlog(0, "done waiting, master is %u\n", res->owner);
1186 mlog(0, "about to master %.*s here, this=%u\n",
1187 res->lockname.len, res->lockname.name, m);
1188 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1190 /* This is a failure in the network path,
1191 * not in the response to the assert_master
1192 * (any nonzero response is a BUG on this node).
1193 * Most likely a socket just got disconnected
1194 * due to node death. */
1197 /* no longer need to restart lock mastery.
1198 * all living nodes have been contacted. */
1202 /* set the lockres owner */
1203 spin_lock(&res->spinlock);
1204 /* mastery reference obtained either during
1205 * assert_master_handler or in get_lock_resource */
1206 dlm_change_lockres_owner(dlm, res, m);
1207 spin_unlock(&res->spinlock);
1213 struct dlm_bitmap_diff_iter
1216 unsigned long *orig_bm;
1217 unsigned long *cur_bm;
1218 unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1221 enum dlm_node_state_change
1228 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1229 unsigned long *orig_bm,
1230 unsigned long *cur_bm)
1232 unsigned long p1, p2;
1236 iter->orig_bm = orig_bm;
1237 iter->cur_bm = cur_bm;
1239 for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1240 p1 = *(iter->orig_bm + i);
1241 p2 = *(iter->cur_bm + i);
1242 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1246 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1247 enum dlm_node_state_change *state)
1251 if (iter->curnode >= O2NM_MAX_NODES)
1254 bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1256 if (bit >= O2NM_MAX_NODES) {
1257 iter->curnode = O2NM_MAX_NODES;
1261 /* if it was there in the original then this node died */
1262 if (test_bit(bit, iter->orig_bm))
1267 iter->curnode = bit;
1272 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1273 struct dlm_lock_resource *res,
1274 struct dlm_master_list_entry *mle,
1277 struct dlm_bitmap_diff_iter bdi;
1278 enum dlm_node_state_change sc;
1282 mlog(0, "something happened such that the "
1283 "master process may need to be restarted!\n");
1285 assert_spin_locked(&mle->spinlock);
1287 dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1288 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1290 if (sc == NODE_UP) {
1291 /* a node came up. clear any old vote from
1292 * the response map and set it in the vote map
1293 * then restart the mastery. */
1294 mlog(ML_NOTICE, "node %d up while restarting\n", node);
1296 /* redo the master request, but only for the new node */
1297 mlog(0, "sending request to new node\n");
1298 clear_bit(node, mle->response_map);
1299 set_bit(node, mle->vote_map);
1301 mlog(ML_ERROR, "node down! %d\n", node);
1303 int lowest = find_next_bit(mle->maybe_map,
1306 /* act like it was never there */
1307 clear_bit(node, mle->maybe_map);
1309 if (node == lowest) {
1310 mlog(0, "expected master %u died"
1311 " while this node was blocked "
1312 "waiting on it!\n", node);
1313 lowest = find_next_bit(mle->maybe_map,
1316 if (lowest < O2NM_MAX_NODES) {
1317 mlog(0, "%s:%.*s:still "
1318 "blocked. waiting on %u "
1324 /* mle is an MLE_BLOCK, but
1325 * there is now nothing left to
1326 * block on. we need to return
1327 * all the way back out and try
1328 * again with an MLE_MASTER.
1329 * dlm_do_local_recovery_cleanup
1330 * has already run, so the mle
1332 mlog(0, "%s:%.*s: no "
1333 "longer blocking. try to "
1334 "master this here\n",
1337 res->lockname.name);
1338 mle->type = DLM_MLE_MASTER;
1344 /* now blank out everything, as if we had never
1345 * contacted anyone */
1346 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1347 memset(mle->response_map, 0, sizeof(mle->response_map));
1348 /* reset the vote_map to the current node_map */
1349 memcpy(mle->vote_map, mle->node_map,
1350 sizeof(mle->node_map));
1351 /* put myself into the maybe map */
1352 if (mle->type != DLM_MLE_BLOCK)
1353 set_bit(dlm->node_num, mle->maybe_map);
1356 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1363 * DLM_MASTER_REQUEST_MSG
1365 * returns: 0 on success,
1366 * -errno on a network error
1368 * on error, the caller should assume the target node is "dead"
1372 static int dlm_do_master_request(struct dlm_lock_resource *res,
1373 struct dlm_master_list_entry *mle, int to)
1375 struct dlm_ctxt *dlm = mle->dlm;
1376 struct dlm_master_request request;
1377 int ret, response=0, resend;
1379 memset(&request, 0, sizeof(request));
1380 request.node_idx = dlm->node_num;
1382 BUG_ON(mle->type == DLM_MLE_MIGRATION);
1384 if (mle->type != DLM_MLE_MASTER) {
1385 request.namelen = mle->u.name.len;
1386 memcpy(request.name, mle->u.name.name, request.namelen);
1388 request.namelen = mle->u.res->lockname.len;
1389 memcpy(request.name, mle->u.res->lockname.name,
1394 ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1395 sizeof(request), to, &response);
1397 if (ret == -ESRCH) {
1398 /* should never happen */
1399 mlog(ML_ERROR, "TCP stack not ready!\n");
1401 } else if (ret == -EINVAL) {
1402 mlog(ML_ERROR, "bad args passed to o2net!\n");
1404 } else if (ret == -ENOMEM) {
1405 mlog(ML_ERROR, "out of memory while trying to send "
1406 "network message! retrying\n");
1407 /* this is totally crude */
1410 } else if (!dlm_is_host_down(ret)) {
1411 /* not a network error. bad. */
1413 mlog(ML_ERROR, "unhandled error!");
1416 /* all other errors should be network errors,
1417 * and likely indicate node death */
1418 mlog(ML_ERROR, "link to %d went down!\n", to);
1424 spin_lock(&mle->spinlock);
1426 case DLM_MASTER_RESP_YES:
1427 set_bit(to, mle->response_map);
1428 mlog(0, "node %u is the master, response=YES\n", to);
1429 mlog(0, "%s:%.*s: master node %u now knows I have a "
1430 "reference\n", dlm->name, res->lockname.len,
1431 res->lockname.name, to);
1434 case DLM_MASTER_RESP_NO:
1435 mlog(0, "node %u not master, response=NO\n", to);
1436 set_bit(to, mle->response_map);
1438 case DLM_MASTER_RESP_MAYBE:
1439 mlog(0, "node %u not master, response=MAYBE\n", to);
1440 set_bit(to, mle->response_map);
1441 set_bit(to, mle->maybe_map);
1443 case DLM_MASTER_RESP_ERROR:
1444 mlog(0, "node %u hit an error, resending\n", to);
1449 mlog(ML_ERROR, "bad response! %u\n", response);
1452 spin_unlock(&mle->spinlock);
1454 /* this is also totally crude */
1464 * locks that can be taken here:
1470 * if possible, TRIM THIS DOWN!!!
1472 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data)
1474 u8 response = DLM_MASTER_RESP_MAYBE;
1475 struct dlm_ctxt *dlm = data;
1476 struct dlm_lock_resource *res = NULL;
1477 struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1478 struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1480 unsigned int namelen, hash;
1483 int dispatch_assert = 0;
1486 return DLM_MASTER_RESP_NO;
1488 if (!dlm_domain_fully_joined(dlm)) {
1489 response = DLM_MASTER_RESP_NO;
1493 name = request->name;
1494 namelen = request->namelen;
1495 hash = dlm_lockid_hash(name, namelen);
1497 if (namelen > DLM_LOCKID_NAME_MAX) {
1498 response = DLM_IVBUFLEN;
1503 spin_lock(&dlm->spinlock);
1504 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1506 spin_unlock(&dlm->spinlock);
1508 /* take care of the easy cases up front */
1509 spin_lock(&res->spinlock);
1510 if (res->state & (DLM_LOCK_RES_RECOVERING|
1511 DLM_LOCK_RES_MIGRATING)) {
1512 spin_unlock(&res->spinlock);
1513 mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1514 "being recovered/migrated\n");
1515 response = DLM_MASTER_RESP_ERROR;
1517 kmem_cache_free(dlm_mle_cache, mle);
1521 if (res->owner == dlm->node_num) {
1522 mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1523 dlm->name, namelen, name, request->node_idx);
1524 dlm_lockres_set_refmap_bit(request->node_idx, res);
1525 spin_unlock(&res->spinlock);
1526 response = DLM_MASTER_RESP_YES;
1528 kmem_cache_free(dlm_mle_cache, mle);
1530 /* this node is the owner.
1531 * there is some extra work that needs to
1532 * happen now. the requesting node has
1533 * caused all nodes up to this one to
1534 * create mles. this node now needs to
1535 * go back and clean those up. */
1536 dispatch_assert = 1;
1538 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1539 spin_unlock(&res->spinlock);
1540 // mlog(0, "node %u is the master\n", res->owner);
1541 response = DLM_MASTER_RESP_NO;
1543 kmem_cache_free(dlm_mle_cache, mle);
1547 /* ok, there is no owner. either this node is
1548 * being blocked, or it is actively trying to
1549 * master this lock. */
1550 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1551 mlog(ML_ERROR, "lock with no owner should be "
1556 // mlog(0, "lockres is in progress...\n");
1557 spin_lock(&dlm->master_lock);
1558 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1560 mlog(ML_ERROR, "no mle found for this lock!\n");
1564 spin_lock(&tmpmle->spinlock);
1565 if (tmpmle->type == DLM_MLE_BLOCK) {
1566 // mlog(0, "this node is waiting for "
1567 // "lockres to be mastered\n");
1568 response = DLM_MASTER_RESP_NO;
1569 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1570 mlog(0, "node %u is master, but trying to migrate to "
1571 "node %u.\n", tmpmle->master, tmpmle->new_master);
1572 if (tmpmle->master == dlm->node_num) {
1573 mlog(ML_ERROR, "no owner on lockres, but this "
1574 "node is trying to migrate it to %u?!\n",
1575 tmpmle->new_master);
1578 /* the real master can respond on its own */
1579 response = DLM_MASTER_RESP_NO;
1581 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1583 if (tmpmle->master == dlm->node_num) {
1584 response = DLM_MASTER_RESP_YES;
1585 /* this node will be the owner.
1586 * go back and clean the mles on any
1588 dispatch_assert = 1;
1589 dlm_lockres_set_refmap_bit(request->node_idx, res);
1590 mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1591 dlm->name, namelen, name,
1594 response = DLM_MASTER_RESP_NO;
1596 // mlog(0, "this node is attempting to "
1597 // "master lockres\n");
1598 response = DLM_MASTER_RESP_MAYBE;
1601 set_bit(request->node_idx, tmpmle->maybe_map);
1602 spin_unlock(&tmpmle->spinlock);
1604 spin_unlock(&dlm->master_lock);
1605 spin_unlock(&res->spinlock);
1607 /* keep the mle attached to heartbeat events */
1608 dlm_put_mle(tmpmle);
1610 kmem_cache_free(dlm_mle_cache, mle);
1615 * lockres doesn't exist on this node
1616 * if there is an MLE_BLOCK, return NO
1617 * if there is an MLE_MASTER, return MAYBE
1618 * otherwise, add an MLE_BLOCK, return NO
1620 spin_lock(&dlm->master_lock);
1621 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1623 /* this lockid has never been seen on this node yet */
1624 // mlog(0, "no mle found\n");
1626 spin_unlock(&dlm->master_lock);
1627 spin_unlock(&dlm->spinlock);
1629 mle = (struct dlm_master_list_entry *)
1630 kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1632 response = DLM_MASTER_RESP_ERROR;
1633 mlog_errno(-ENOMEM);
1639 // mlog(0, "this is second time thru, already allocated, "
1640 // "add the block.\n");
1641 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1642 set_bit(request->node_idx, mle->maybe_map);
1643 list_add(&mle->list, &dlm->master_list);
1644 response = DLM_MASTER_RESP_NO;
1646 // mlog(0, "mle was found\n");
1648 spin_lock(&tmpmle->spinlock);
1649 if (tmpmle->master == dlm->node_num) {
1650 mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1653 if (tmpmle->type == DLM_MLE_BLOCK)
1654 response = DLM_MASTER_RESP_NO;
1655 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1656 mlog(0, "migration mle was found (%u->%u)\n",
1657 tmpmle->master, tmpmle->new_master);
1658 /* real master can respond on its own */
1659 response = DLM_MASTER_RESP_NO;
1661 response = DLM_MASTER_RESP_MAYBE;
1663 set_bit(request->node_idx, tmpmle->maybe_map);
1664 spin_unlock(&tmpmle->spinlock);
1666 spin_unlock(&dlm->master_lock);
1667 spin_unlock(&dlm->spinlock);
1670 /* keep the mle attached to heartbeat events */
1671 dlm_put_mle(tmpmle);
1675 if (dispatch_assert) {
1676 if (response != DLM_MASTER_RESP_YES)
1677 mlog(ML_ERROR, "invalid response %d\n", response);
1679 mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1682 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1683 dlm->node_num, res->lockname.len, res->lockname.name);
1684 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1685 DLM_ASSERT_MASTER_MLE_CLEANUP);
1687 mlog(ML_ERROR, "failed to dispatch assert master work\n");
1688 response = DLM_MASTER_RESP_ERROR;
1697 * DLM_ASSERT_MASTER_MSG
1702 * NOTE: this can be used for debugging
1703 * can periodically run all locks owned by this node
1704 * and re-assert across the cluster...
1706 int dlm_do_assert_master(struct dlm_ctxt *dlm,
1707 struct dlm_lock_resource *res,
1708 void *nodemap, u32 flags)
1710 struct dlm_assert_master assert;
1712 struct dlm_node_iter iter;
1715 const char *lockname = res->lockname.name;
1716 unsigned int namelen = res->lockname.len;
1718 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1722 /* note that if this nodemap is empty, it returns 0 */
1723 dlm_node_iter_init(nodemap, &iter);
1724 while ((to = dlm_node_iter_next(&iter)) >= 0) {
1726 struct dlm_master_list_entry *mle = NULL;
1728 mlog(0, "sending assert master to %d (%.*s)\n", to,
1730 memset(&assert, 0, sizeof(assert));
1731 assert.node_idx = dlm->node_num;
1732 assert.namelen = namelen;
1733 memcpy(assert.name, lockname, namelen);
1734 assert.flags = cpu_to_be32(flags);
1736 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1737 &assert, sizeof(assert), to, &r);
1739 mlog(0, "assert_master returned %d!\n", tmpret);
1740 if (!dlm_is_host_down(tmpret)) {
1741 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1744 /* a node died. finish out the rest of the nodes. */
1745 mlog(0, "link to %d went down!\n", to);
1746 /* any nonzero status return will do */
1750 /* ok, something horribly messed. kill thyself. */
1751 mlog(ML_ERROR,"during assert master of %.*s to %u, "
1752 "got %d.\n", namelen, lockname, to, r);
1753 spin_lock(&dlm->spinlock);
1754 spin_lock(&dlm->master_lock);
1755 if (dlm_find_mle(dlm, &mle, (char *)lockname,
1757 dlm_print_one_mle(mle);
1760 spin_unlock(&dlm->master_lock);
1761 spin_unlock(&dlm->spinlock);
1765 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1766 !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1767 mlog(ML_ERROR, "%.*s: very strange, "
1768 "master MLE but no lockres on %u\n",
1769 namelen, lockname, to);
1772 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1773 mlog(0, "%.*s: node %u create mles on other "
1774 "nodes and requests a re-assert\n",
1775 namelen, lockname, to);
1778 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1779 mlog(0, "%.*s: node %u has a reference to this "
1780 "lockres, set the bit in the refmap\n",
1781 namelen, lockname, to);
1782 spin_lock(&res->spinlock);
1783 dlm_lockres_set_refmap_bit(to, res);
1784 spin_unlock(&res->spinlock);
1795 * locks that can be taken here:
1801 * if possible, TRIM THIS DOWN!!!
1803 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data)
1805 struct dlm_ctxt *dlm = data;
1806 struct dlm_master_list_entry *mle = NULL;
1807 struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1808 struct dlm_lock_resource *res = NULL;
1810 unsigned int namelen, hash;
1812 int master_request = 0, have_lockres_ref = 0;
1818 name = assert->name;
1819 namelen = assert->namelen;
1820 hash = dlm_lockid_hash(name, namelen);
1821 flags = be32_to_cpu(assert->flags);
1823 if (namelen > DLM_LOCKID_NAME_MAX) {
1824 mlog(ML_ERROR, "Invalid name length!");
1828 spin_lock(&dlm->spinlock);
1831 mlog(0, "assert_master with flags: %u\n", flags);
1834 spin_lock(&dlm->master_lock);
1835 if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1836 /* not an error, could be master just re-asserting */
1837 mlog(0, "just got an assert_master from %u, but no "
1838 "MLE for it! (%.*s)\n", assert->node_idx,
1841 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1842 if (bit >= O2NM_MAX_NODES) {
1843 /* not necessarily an error, though less likely.
1844 * could be master just re-asserting. */
1845 mlog(0, "no bits set in the maybe_map, but %u "
1846 "is asserting! (%.*s)\n", assert->node_idx,
1848 } else if (bit != assert->node_idx) {
1849 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1850 mlog(0, "master %u was found, %u should "
1851 "back off\n", assert->node_idx, bit);
1853 /* with the fix for bug 569, a higher node
1854 * number winning the mastery will respond
1855 * YES to mastery requests, but this node
1856 * had no way of knowing. let it pass. */
1857 mlog(0, "%u is the lowest node, "
1858 "%u is asserting. (%.*s) %u must "
1859 "have begun after %u won.\n", bit,
1860 assert->node_idx, namelen, name, bit,
1864 if (mle->type == DLM_MLE_MIGRATION) {
1865 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1866 mlog(0, "%s:%.*s: got cleanup assert"
1867 " from %u for migration\n",
1868 dlm->name, namelen, name,
1870 } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1871 mlog(0, "%s:%.*s: got unrelated assert"
1872 " from %u for migration, ignoring\n",
1873 dlm->name, namelen, name,
1876 spin_unlock(&dlm->master_lock);
1877 spin_unlock(&dlm->spinlock);
1882 spin_unlock(&dlm->master_lock);
1884 /* ok everything checks out with the MLE
1885 * now check to see if there is a lockres */
1886 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1888 spin_lock(&res->spinlock);
1889 if (res->state & DLM_LOCK_RES_RECOVERING) {
1890 mlog(ML_ERROR, "%u asserting but %.*s is "
1891 "RECOVERING!\n", assert->node_idx, namelen, name);
1895 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1896 res->owner != assert->node_idx) {
1897 mlog(ML_ERROR, "assert_master from "
1898 "%u, but current owner is "
1900 assert->node_idx, res->owner,
1904 } else if (mle->type != DLM_MLE_MIGRATION) {
1905 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1906 /* owner is just re-asserting */
1907 if (res->owner == assert->node_idx) {
1908 mlog(0, "owner %u re-asserting on "
1909 "lock %.*s\n", assert->node_idx,
1913 mlog(ML_ERROR, "got assert_master from "
1914 "node %u, but %u is the owner! "
1915 "(%.*s)\n", assert->node_idx,
1916 res->owner, namelen, name);
1919 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1920 mlog(ML_ERROR, "got assert from %u, but lock "
1921 "with no owner should be "
1922 "in-progress! (%.*s)\n",
1927 } else /* mle->type == DLM_MLE_MIGRATION */ {
1928 /* should only be getting an assert from new master */
1929 if (assert->node_idx != mle->new_master) {
1930 mlog(ML_ERROR, "got assert from %u, but "
1931 "new master is %u, and old master "
1933 assert->node_idx, mle->new_master,
1934 mle->master, namelen, name);
1940 spin_unlock(&res->spinlock);
1942 spin_unlock(&dlm->spinlock);
1944 // mlog(0, "woo! got an assert_master from node %u!\n",
1945 // assert->node_idx);
1951 spin_lock(&mle->spinlock);
1952 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1955 /* MASTER mle: if any bits set in the response map
1956 * then the calling node needs to re-assert to clear
1957 * up nodes that this node contacted */
1958 while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1959 nn+1)) < O2NM_MAX_NODES) {
1960 if (nn != dlm->node_num && nn != assert->node_idx)
1964 mle->master = assert->node_idx;
1965 atomic_set(&mle->woken, 1);
1967 spin_unlock(&mle->spinlock);
1970 spin_lock(&res->spinlock);
1971 if (mle->type == DLM_MLE_MIGRATION) {
1972 mlog(0, "finishing off migration of lockres %.*s, "
1974 res->lockname.len, res->lockname.name,
1975 dlm->node_num, mle->new_master);
1976 res->state &= ~DLM_LOCK_RES_MIGRATING;
1977 dlm_change_lockres_owner(dlm, res, mle->new_master);
1978 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1980 dlm_change_lockres_owner(dlm, res, mle->master);
1982 spin_unlock(&res->spinlock);
1983 have_lockres_ref = 1;
1986 /* master is known, detach if not already detached.
1987 * ensures that only one assert_master call will happen
1989 spin_lock(&dlm->spinlock);
1990 spin_lock(&dlm->master_lock);
1992 rr = atomic_read(&mle->mle_refs.refcount);
1993 if (mle->inuse > 0) {
1994 if (extra_ref && rr < 3)
1996 else if (!extra_ref && rr < 2)
1999 if (extra_ref && rr < 2)
2001 else if (!extra_ref && rr < 1)
2005 mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
2006 "that will mess up this node, refs=%d, extra=%d, "
2007 "inuse=%d\n", dlm->name, namelen, name,
2008 assert->node_idx, rr, extra_ref, mle->inuse);
2009 dlm_print_one_mle(mle);
2011 list_del_init(&mle->list);
2012 __dlm_mle_detach_hb_events(dlm, mle);
2015 /* the assert master message now balances the extra
2016 * ref given by the master / migration request message.
2017 * if this is the last put, it will be removed
2021 spin_unlock(&dlm->master_lock);
2022 spin_unlock(&dlm->spinlock);
2024 if (res->owner != assert->node_idx) {
2025 mlog(0, "assert_master from %u, but current "
2026 "owner is %u (%.*s), no mle\n", assert->node_idx,
2027 res->owner, namelen, name);
2034 dlm_lockres_put(res);
2036 if (master_request) {
2037 mlog(0, "need to tell master to reassert\n");
2038 /* positive. negative would shoot down the node. */
2039 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2040 if (!have_lockres_ref) {
2041 mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2042 "mle present here for %s:%.*s, but no lockres!\n",
2043 assert->node_idx, dlm->name, namelen, name);
2046 if (have_lockres_ref) {
2047 /* let the master know we have a reference to the lockres */
2048 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2049 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2050 dlm->name, namelen, name, assert->node_idx);
2055 /* kill the caller! */
2056 mlog(ML_ERROR, "Bad message received from another node. Dumping state "
2057 "and killing the other node now! This node is OK and can continue.\n");
2058 __dlm_print_one_lock_resource(res);
2059 spin_unlock(&res->spinlock);
2060 spin_unlock(&dlm->spinlock);
2061 dlm_lockres_put(res);
2066 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2067 struct dlm_lock_resource *res,
2068 int ignore_higher, u8 request_from, u32 flags)
2070 struct dlm_work_item *item;
2071 item = kzalloc(sizeof(*item), GFP_NOFS);
2076 /* queue up work for dlm_assert_master_worker */
2077 dlm_grab(dlm); /* get an extra ref for the work item */
2078 dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2079 item->u.am.lockres = res; /* already have a ref */
2080 /* can optionally ignore node numbers higher than this node */
2081 item->u.am.ignore_higher = ignore_higher;
2082 item->u.am.request_from = request_from;
2083 item->u.am.flags = flags;
2086 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2087 res->lockname.name);
2089 spin_lock(&dlm->work_lock);
2090 list_add_tail(&item->list, &dlm->work_list);
2091 spin_unlock(&dlm->work_lock);
2093 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2097 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2099 struct dlm_ctxt *dlm = data;
2101 struct dlm_lock_resource *res;
2102 unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2109 res = item->u.am.lockres;
2110 ignore_higher = item->u.am.ignore_higher;
2111 request_from = item->u.am.request_from;
2112 flags = item->u.am.flags;
2114 spin_lock(&dlm->spinlock);
2115 memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2116 spin_unlock(&dlm->spinlock);
2118 clear_bit(dlm->node_num, nodemap);
2119 if (ignore_higher) {
2120 /* if is this just to clear up mles for nodes below
2121 * this node, do not send the message to the original
2122 * caller or any node number higher than this */
2123 clear_bit(request_from, nodemap);
2124 bit = dlm->node_num;
2126 bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2128 if (bit >= O2NM_MAX_NODES)
2130 clear_bit(bit, nodemap);
2135 * If we're migrating this lock to someone else, we are no
2136 * longer allowed to assert out own mastery. OTOH, we need to
2137 * prevent migration from starting while we're still asserting
2138 * our dominance. The reserved ast delays migration.
2140 spin_lock(&res->spinlock);
2141 if (res->state & DLM_LOCK_RES_MIGRATING) {
2142 mlog(0, "Someone asked us to assert mastery, but we're "
2143 "in the middle of migration. Skipping assert, "
2144 "the new master will handle that.\n");
2145 spin_unlock(&res->spinlock);
2148 __dlm_lockres_reserve_ast(res);
2149 spin_unlock(&res->spinlock);
2151 /* this call now finishes out the nodemap
2152 * even if one or more nodes die */
2153 mlog(0, "worker about to master %.*s here, this=%u\n",
2154 res->lockname.len, res->lockname.name, dlm->node_num);
2155 ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2157 /* no need to restart, we are done */
2158 if (!dlm_is_host_down(ret))
2162 /* Ok, we've asserted ourselves. Let's let migration start. */
2163 dlm_lockres_release_ast(dlm, res);
2166 dlm_lockres_put(res);
2168 mlog(0, "finished with dlm_assert_master_worker\n");
2171 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2172 * We cannot wait for node recovery to complete to begin mastering this
2173 * lockres because this lockres is used to kick off recovery! ;-)
2174 * So, do a pre-check on all living nodes to see if any of those nodes
2175 * think that $RECOVERY is currently mastered by a dead node. If so,
2176 * we wait a short time to allow that node to get notified by its own
2177 * heartbeat stack, then check again. All $RECOVERY lock resources
2178 * mastered by dead nodes are purged when the hearbeat callback is
2179 * fired, so we can know for sure that it is safe to continue once
2180 * the node returns a live node or no node. */
2181 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2182 struct dlm_lock_resource *res)
2184 struct dlm_node_iter iter;
2187 u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2189 spin_lock(&dlm->spinlock);
2190 dlm_node_iter_init(dlm->domain_map, &iter);
2191 spin_unlock(&dlm->spinlock);
2193 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2194 /* do not send to self */
2195 if (nodenum == dlm->node_num)
2197 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2200 if (!dlm_is_host_down(ret))
2202 /* host is down, so answer for that node would be
2203 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
2207 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2208 /* check to see if this master is in the recovery map */
2209 spin_lock(&dlm->spinlock);
2210 if (test_bit(master, dlm->recovery_map)) {
2211 mlog(ML_NOTICE, "%s: node %u has not seen "
2212 "node %u go down yet, and thinks the "
2213 "dead node is mastering the recovery "
2214 "lock. must wait.\n", dlm->name,
2218 spin_unlock(&dlm->spinlock);
2219 mlog(0, "%s: reco lock master is %u\n", dlm->name,
2228 * DLM_DEREF_LOCKRES_MSG
2231 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2233 struct dlm_deref_lockres deref;
2235 const char *lockname;
2236 unsigned int namelen;
2238 lockname = res->lockname.name;
2239 namelen = res->lockname.len;
2240 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2242 mlog(0, "%s:%.*s: sending deref to %d\n",
2243 dlm->name, namelen, lockname, res->owner);
2244 memset(&deref, 0, sizeof(deref));
2245 deref.node_idx = dlm->node_num;
2246 deref.namelen = namelen;
2247 memcpy(deref.name, lockname, namelen);
2249 ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2250 &deref, sizeof(deref), res->owner, &r);
2254 /* BAD. other node says I did not have a ref. */
2255 mlog(ML_ERROR,"while dropping ref on %s:%.*s "
2256 "(master=%u) got %d.\n", dlm->name, namelen,
2257 lockname, res->owner, r);
2258 dlm_print_one_lock_resource(res);
2264 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data)
2266 struct dlm_ctxt *dlm = data;
2267 struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2268 struct dlm_lock_resource *res = NULL;
2270 unsigned int namelen;
2279 namelen = deref->namelen;
2280 node = deref->node_idx;
2282 if (namelen > DLM_LOCKID_NAME_MAX) {
2283 mlog(ML_ERROR, "Invalid name length!");
2286 if (deref->node_idx >= O2NM_MAX_NODES) {
2287 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2291 hash = dlm_lockid_hash(name, namelen);
2293 spin_lock(&dlm->spinlock);
2294 res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2296 spin_unlock(&dlm->spinlock);
2297 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2298 dlm->name, namelen, name);
2301 spin_unlock(&dlm->spinlock);
2303 spin_lock(&res->spinlock);
2304 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2305 if (test_bit(node, res->refmap)) {
2307 dlm_lockres_clear_refmap_bit(node, res);
2309 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2310 "but it is already dropped!\n", dlm->name, namelen,
2312 __dlm_print_one_lock_resource(res);
2314 spin_unlock(&res->spinlock);
2317 dlm_lockres_calc_usage(dlm, res);
2320 dlm_lockres_put(res);
2327 * DLM_MIGRATE_LOCKRES
2331 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2332 struct dlm_lock_resource *res,
2335 struct dlm_master_list_entry *mle = NULL;
2336 struct dlm_master_list_entry *oldmle = NULL;
2337 struct dlm_migratable_lockres *mres = NULL;
2340 unsigned int namelen;
2342 struct list_head *queue, *iter;
2344 struct dlm_lock *lock;
2350 name = res->lockname.name;
2351 namelen = res->lockname.len;
2353 mlog(0, "migrating %.*s to %u\n", namelen, name, target);
2356 * ensure this lockres is a proper candidate for migration
2358 spin_lock(&res->spinlock);
2359 if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
2360 mlog(0, "cannot migrate lockres with unknown owner!\n");
2361 spin_unlock(&res->spinlock);
2364 if (res->owner != dlm->node_num) {
2365 mlog(0, "cannot migrate lockres this node doesn't own!\n");
2366 spin_unlock(&res->spinlock);
2369 mlog(0, "checking queues...\n");
2370 queue = &res->granted;
2371 for (i=0; i<3; i++) {
2372 list_for_each(iter, queue) {
2373 lock = list_entry (iter, struct dlm_lock, list);
2375 if (lock->ml.node == dlm->node_num) {
2376 mlog(0, "found a lock owned by this node "
2377 "still on the %s queue! will not "
2378 "migrate this lockres\n",
2380 (i==1 ? "converting" : "blocked"));
2381 spin_unlock(&res->spinlock);
2388 mlog(0, "all locks on this lockres are nonlocal. continuing\n");
2389 spin_unlock(&res->spinlock);
2393 mlog(0, "no locks were found on this lockres! done!\n");
2399 * preallocate up front
2400 * if this fails, abort
2404 mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2410 mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
2419 * find a node to migrate the lockres to
2422 mlog(0, "picking a migration node\n");
2423 spin_lock(&dlm->spinlock);
2424 /* pick a new node */
2425 if (!test_bit(target, dlm->domain_map) ||
2426 target >= O2NM_MAX_NODES) {
2427 target = dlm_pick_migration_target(dlm, res);
2429 mlog(0, "node %u chosen for migration\n", target);
2431 if (target >= O2NM_MAX_NODES ||
2432 !test_bit(target, dlm->domain_map)) {
2433 /* target chosen is not alive */
2438 spin_unlock(&dlm->spinlock);
2442 mlog(0, "continuing with target = %u\n", target);
2445 * clear any existing master requests and
2446 * add the migration mle to the list
2448 spin_lock(&dlm->master_lock);
2449 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2450 namelen, target, dlm->node_num);
2451 spin_unlock(&dlm->master_lock);
2452 spin_unlock(&dlm->spinlock);
2454 if (ret == -EEXIST) {
2455 mlog(0, "another process is already migrating it\n");
2461 * set the MIGRATING flag and flush asts
2462 * if we fail after this we need to re-dirty the lockres
2464 if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2465 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2466 "the target went down.\n", res->lockname.len,
2467 res->lockname.name, target);
2468 spin_lock(&res->spinlock);
2469 res->state &= ~DLM_LOCK_RES_MIGRATING;
2470 spin_unlock(&res->spinlock);
2476 /* master is known, detach if not already detached */
2477 dlm_mle_detach_hb_events(dlm, oldmle);
2478 dlm_put_mle(oldmle);
2483 dlm_mle_detach_hb_events(dlm, mle);
2486 kmem_cache_free(dlm_mle_cache, mle);
2492 * at this point, we have a migration target, an mle
2493 * in the master list, and the MIGRATING flag set on
2497 /* now that remote nodes are spinning on the MIGRATING flag,
2498 * ensure that all assert_master work is flushed. */
2499 flush_workqueue(dlm->dlm_worker);
2501 /* get an extra reference on the mle.
2502 * otherwise the assert_master from the new
2503 * master will destroy this.
2504 * also, make sure that all callers of dlm_get_mle
2505 * take both dlm->spinlock and dlm->master_lock */
2506 spin_lock(&dlm->spinlock);
2507 spin_lock(&dlm->master_lock);
2508 dlm_get_mle_inuse(mle);
2509 spin_unlock(&dlm->master_lock);
2510 spin_unlock(&dlm->spinlock);
2512 /* notify new node and send all lock state */
2513 /* call send_one_lockres with migration flag.
2514 * this serves as notice to the target node that a
2515 * migration is starting. */
2516 ret = dlm_send_one_lockres(dlm, res, mres, target,
2517 DLM_MRES_MIGRATION);
2520 mlog(0, "migration to node %u failed with %d\n",
2522 /* migration failed, detach and clean up mle */
2523 dlm_mle_detach_hb_events(dlm, mle);
2525 dlm_put_mle_inuse(mle);
2526 spin_lock(&res->spinlock);
2527 res->state &= ~DLM_LOCK_RES_MIGRATING;
2528 spin_unlock(&res->spinlock);
2532 /* at this point, the target sends a message to all nodes,
2533 * (using dlm_do_migrate_request). this node is skipped since
2534 * we had to put an mle in the list to begin the process. this
2535 * node now waits for target to do an assert master. this node
2536 * will be the last one notified, ensuring that the migration
2537 * is complete everywhere. if the target dies while this is
2538 * going on, some nodes could potentially see the target as the
2539 * master, so it is important that my recovery finds the migration
2540 * mle and sets the master to UNKNONWN. */
2543 /* wait for new node to assert master */
2545 ret = wait_event_interruptible_timeout(mle->wq,
2546 (atomic_read(&mle->woken) == 1),
2547 msecs_to_jiffies(5000));
2550 if (atomic_read(&mle->woken) == 1 ||
2551 res->owner == target)
2554 mlog(0, "%s:%.*s: timed out during migration\n",
2555 dlm->name, res->lockname.len, res->lockname.name);
2556 /* avoid hang during shutdown when migrating lockres
2557 * to a node which also goes down */
2558 if (dlm_is_node_dead(dlm, target)) {
2559 mlog(0, "%s:%.*s: expected migration "
2560 "target %u is no longer up, restarting\n",
2561 dlm->name, res->lockname.len,
2562 res->lockname.name, target);
2564 /* migration failed, detach and clean up mle */
2565 dlm_mle_detach_hb_events(dlm, mle);
2567 dlm_put_mle_inuse(mle);
2568 spin_lock(&res->spinlock);
2569 res->state &= ~DLM_LOCK_RES_MIGRATING;
2570 spin_unlock(&res->spinlock);
2574 mlog(0, "%s:%.*s: caught signal during migration\n",
2575 dlm->name, res->lockname.len, res->lockname.name);
2578 /* all done, set the owner, clear the flag */
2579 spin_lock(&res->spinlock);
2580 dlm_set_lockres_owner(dlm, res, target);
2581 res->state &= ~DLM_LOCK_RES_MIGRATING;
2582 dlm_remove_nonlocal_locks(dlm, res);
2583 spin_unlock(&res->spinlock);
2586 /* master is known, detach if not already detached */
2587 dlm_mle_detach_hb_events(dlm, mle);
2588 dlm_put_mle_inuse(mle);
2591 dlm_lockres_calc_usage(dlm, res);
2594 /* re-dirty the lockres if we failed */
2596 dlm_kick_thread(dlm, res);
2600 free_page((unsigned long)mres);
2604 mlog(0, "returning %d\n", ret);
2608 #define DLM_MIGRATION_RETRY_MS 100
2610 /* Should be called only after beginning the domain leave process.
2611 * There should not be any remaining locks on nonlocal lock resources,
2612 * and there should be no local locks left on locally mastered resources.
2614 * Called with the dlm spinlock held, may drop it to do migration, but
2615 * will re-acquire before exit.
2617 * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped */
2618 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2621 int lock_dropped = 0;
2623 if (res->owner != dlm->node_num) {
2624 if (!__dlm_lockres_unused(res)) {
2625 mlog(ML_ERROR, "%s:%.*s: this node is not master, "
2626 "trying to free this but locks remain\n",
2627 dlm->name, res->lockname.len, res->lockname.name);
2632 /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2633 spin_unlock(&dlm->spinlock);
2636 ret = dlm_migrate_lockres(dlm, res, O2NM_MAX_NODES);
2639 if (ret == -ENOTEMPTY) {
2640 mlog(ML_ERROR, "lockres %.*s still has local locks!\n",
2641 res->lockname.len, res->lockname.name);
2645 mlog(0, "lockres %.*s: migrate failed, "
2646 "retrying\n", res->lockname.len,
2647 res->lockname.name);
2648 msleep(DLM_MIGRATION_RETRY_MS);
2650 spin_lock(&dlm->spinlock);
2652 return lock_dropped;
2655 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2658 spin_lock(&dlm->ast_lock);
2659 spin_lock(&lock->spinlock);
2660 ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2661 spin_unlock(&lock->spinlock);
2662 spin_unlock(&dlm->ast_lock);
2666 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2667 struct dlm_lock_resource *res,
2671 spin_lock(&res->spinlock);
2672 can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2673 spin_unlock(&res->spinlock);
2675 /* target has died, so make the caller break out of the
2676 * wait_event, but caller must recheck the domain_map */
2677 spin_lock(&dlm->spinlock);
2678 if (!test_bit(mig_target, dlm->domain_map))
2680 spin_unlock(&dlm->spinlock);
2684 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2685 struct dlm_lock_resource *res)
2688 spin_lock(&res->spinlock);
2689 ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2690 spin_unlock(&res->spinlock);
2695 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2696 struct dlm_lock_resource *res,
2701 mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2702 res->lockname.len, res->lockname.name, dlm->node_num,
2704 /* need to set MIGRATING flag on lockres. this is done by
2705 * ensuring that all asts have been flushed for this lockres. */
2706 spin_lock(&res->spinlock);
2707 BUG_ON(res->migration_pending);
2708 res->migration_pending = 1;
2709 /* strategy is to reserve an extra ast then release
2710 * it below, letting the release do all of the work */
2711 __dlm_lockres_reserve_ast(res);
2712 spin_unlock(&res->spinlock);
2714 /* now flush all the pending asts */
2715 dlm_kick_thread(dlm, res);
2716 /* before waiting on DIRTY, block processes which may
2717 * try to dirty the lockres before MIGRATING is set */
2718 spin_lock(&res->spinlock);
2719 BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2720 res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2721 spin_unlock(&res->spinlock);
2722 /* now wait on any pending asts and the DIRTY state */
2723 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2724 dlm_lockres_release_ast(dlm, res);
2726 mlog(0, "about to wait on migration_wq, dirty=%s\n",
2727 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2728 /* if the extra ref we just put was the final one, this
2729 * will pass thru immediately. otherwise, we need to wait
2730 * for the last ast to finish. */
2732 ret = wait_event_interruptible_timeout(dlm->migration_wq,
2733 dlm_migration_can_proceed(dlm, res, target),
2734 msecs_to_jiffies(1000));
2736 mlog(0, "woken again: migrating? %s, dead? %s\n",
2737 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2738 test_bit(target, dlm->domain_map) ? "no":"yes");
2740 mlog(0, "all is well: migrating? %s, dead? %s\n",
2741 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2742 test_bit(target, dlm->domain_map) ? "no":"yes");
2744 if (!dlm_migration_can_proceed(dlm, res, target)) {
2745 mlog(0, "trying again...\n");
2748 /* now that we are sure the MIGRATING state is there, drop
2749 * the unneded state which blocked threads trying to DIRTY */
2750 spin_lock(&res->spinlock);
2751 BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2752 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2753 res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2754 spin_unlock(&res->spinlock);
2756 /* did the target go down or die? */
2757 spin_lock(&dlm->spinlock);
2758 if (!test_bit(target, dlm->domain_map)) {
2759 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2763 spin_unlock(&dlm->spinlock);
2768 * o the DLM_LOCK_RES_MIGRATING flag is set
2769 * o there are no pending asts on this lockres
2770 * o all processes trying to reserve an ast on this
2771 * lockres must wait for the MIGRATING flag to clear
2776 /* last step in the migration process.
2777 * original master calls this to free all of the dlm_lock
2778 * structures that used to be for other nodes. */
2779 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2780 struct dlm_lock_resource *res)
2782 struct list_head *iter, *iter2;
2783 struct list_head *queue = &res->granted;
2785 struct dlm_lock *lock;
2787 assert_spin_locked(&res->spinlock);
2789 BUG_ON(res->owner == dlm->node_num);
2791 for (i=0; i<3; i++) {
2792 list_for_each_safe(iter, iter2, queue) {
2793 lock = list_entry (iter, struct dlm_lock, 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(lock->ml.node, res);
2803 list_del_init(&lock->list);
2811 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2812 if (bit >= O2NM_MAX_NODES)
2814 /* do not clear the local node reference, if there is a
2815 * process holding this, let it drop the ref itself */
2816 if (bit != dlm->node_num) {
2817 mlog(0, "%s:%.*s: node %u had a ref to this "
2818 "migrating lockres, clearing\n", dlm->name,
2819 res->lockname.len, res->lockname.name, bit);
2820 dlm_lockres_clear_refmap_bit(bit, res);
2826 /* for now this is not too intelligent. we will
2827 * need stats to make this do the right thing.
2828 * this just finds the first lock on one of the
2829 * queues and uses that node as the target. */
2830 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2831 struct dlm_lock_resource *res)
2834 struct list_head *queue = &res->granted;
2835 struct list_head *iter;
2836 struct dlm_lock *lock;
2839 assert_spin_locked(&dlm->spinlock);
2841 spin_lock(&res->spinlock);
2842 for (i=0; i<3; i++) {
2843 list_for_each(iter, queue) {
2844 /* up to the caller to make sure this node
2846 lock = list_entry (iter, struct dlm_lock, list);
2847 if (lock->ml.node != dlm->node_num) {
2848 spin_unlock(&res->spinlock);
2849 return lock->ml.node;
2854 spin_unlock(&res->spinlock);
2855 mlog(0, "have not found a suitable target yet! checking domain map\n");
2857 /* ok now we're getting desperate. pick anyone alive. */
2860 nodenum = find_next_bit(dlm->domain_map,
2861 O2NM_MAX_NODES, nodenum+1);
2862 mlog(0, "found %d in domain map\n", nodenum);
2863 if (nodenum >= O2NM_MAX_NODES)
2865 if (nodenum != dlm->node_num) {
2866 mlog(0, "picking %d\n", nodenum);
2871 mlog(0, "giving up. no master to migrate to\n");
2872 return DLM_LOCK_RES_OWNER_UNKNOWN;
2877 /* this is called by the new master once all lockres
2878 * data has been received */
2879 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2880 struct dlm_lock_resource *res,
2881 u8 master, u8 new_master,
2882 struct dlm_node_iter *iter)
2884 struct dlm_migrate_request migrate;
2885 int ret, status = 0;
2888 memset(&migrate, 0, sizeof(migrate));
2889 migrate.namelen = res->lockname.len;
2890 memcpy(migrate.name, res->lockname.name, migrate.namelen);
2891 migrate.new_master = new_master;
2892 migrate.master = master;
2896 /* send message to all nodes, except the master and myself */
2897 while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2898 if (nodenum == master ||
2899 nodenum == new_master)
2902 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2903 &migrate, sizeof(migrate), nodenum,
2907 else if (status < 0) {
2908 mlog(0, "migrate request (node %u) returned %d!\n",
2911 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
2912 /* during the migration request we short-circuited
2913 * the mastery of the lockres. make sure we have
2914 * a mastery ref for nodenum */
2915 mlog(0, "%s:%.*s: need ref for node %u\n",
2916 dlm->name, res->lockname.len, res->lockname.name,
2918 spin_lock(&res->spinlock);
2919 dlm_lockres_set_refmap_bit(nodenum, res);
2920 spin_unlock(&res->spinlock);
2927 mlog(0, "returning ret=%d\n", ret);
2932 /* if there is an existing mle for this lockres, we now know who the master is.
2933 * (the one who sent us *this* message) we can clear it up right away.
2934 * since the process that put the mle on the list still has a reference to it,
2935 * we can unhash it now, set the master and wake the process. as a result,
2936 * we will have no mle in the list to start with. now we can add an mle for
2937 * the migration and this should be the only one found for those scanning the
2939 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data)
2941 struct dlm_ctxt *dlm = data;
2942 struct dlm_lock_resource *res = NULL;
2943 struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
2944 struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
2946 unsigned int namelen, hash;
2952 name = migrate->name;
2953 namelen = migrate->namelen;
2954 hash = dlm_lockid_hash(name, namelen);
2956 /* preallocate.. if this fails, abort */
2957 mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
2965 /* check for pre-existing lock */
2966 spin_lock(&dlm->spinlock);
2967 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
2968 spin_lock(&dlm->master_lock);
2971 spin_lock(&res->spinlock);
2972 if (res->state & DLM_LOCK_RES_RECOVERING) {
2973 /* if all is working ok, this can only mean that we got
2974 * a migrate request from a node that we now see as
2975 * dead. what can we do here? drop it to the floor? */
2976 spin_unlock(&res->spinlock);
2977 mlog(ML_ERROR, "Got a migrate request, but the "
2978 "lockres is marked as recovering!");
2979 kmem_cache_free(dlm_mle_cache, mle);
2980 ret = -EINVAL; /* need a better solution */
2983 res->state |= DLM_LOCK_RES_MIGRATING;
2984 spin_unlock(&res->spinlock);
2987 /* ignore status. only nonzero status would BUG. */
2988 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
2990 migrate->new_master,
2994 spin_unlock(&dlm->master_lock);
2995 spin_unlock(&dlm->spinlock);
2998 /* master is known, detach if not already detached */
2999 dlm_mle_detach_hb_events(dlm, oldmle);
3000 dlm_put_mle(oldmle);
3004 dlm_lockres_put(res);
3010 /* must be holding dlm->spinlock and dlm->master_lock
3011 * when adding a migration mle, we can clear any other mles
3012 * in the master list because we know with certainty that
3013 * the master is "master". so we remove any old mle from
3014 * the list after setting it's master field, and then add
3015 * the new migration mle. this way we can hold with the rule
3016 * of having only one mle for a given lock name at all times. */
3017 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3018 struct dlm_lock_resource *res,
3019 struct dlm_master_list_entry *mle,
3020 struct dlm_master_list_entry **oldmle,
3021 const char *name, unsigned int namelen,
3022 u8 new_master, u8 master)
3031 assert_spin_locked(&dlm->spinlock);
3032 assert_spin_locked(&dlm->master_lock);
3034 /* caller is responsible for any ref taken here on oldmle */
3035 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3037 struct dlm_master_list_entry *tmp = *oldmle;
3038 spin_lock(&tmp->spinlock);
3039 if (tmp->type == DLM_MLE_MIGRATION) {
3040 if (master == dlm->node_num) {
3041 /* ah another process raced me to it */
3042 mlog(0, "tried to migrate %.*s, but some "
3043 "process beat me to it\n",
3047 /* bad. 2 NODES are trying to migrate! */
3048 mlog(ML_ERROR, "migration error mle: "
3049 "master=%u new_master=%u // request: "
3050 "master=%u new_master=%u // "
3052 tmp->master, tmp->new_master,
3058 /* this is essentially what assert_master does */
3059 tmp->master = master;
3060 atomic_set(&tmp->woken, 1);
3062 /* remove it from the list so that only one
3063 * mle will be found */
3064 list_del_init(&tmp->list);
3065 /* this was obviously WRONG. mle is uninited here. should be tmp. */
3066 __dlm_mle_detach_hb_events(dlm, tmp);
3067 ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3068 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3069 "telling master to get ref for cleared out mle "
3070 "during migration\n", dlm->name, namelen, name,
3071 master, new_master);
3073 spin_unlock(&tmp->spinlock);
3076 /* now add a migration mle to the tail of the list */
3077 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3078 mle->new_master = new_master;
3079 /* the new master will be sending an assert master for this.
3080 * at that point we will get the refmap reference */
3081 mle->master = master;
3082 /* do this for consistency with other mle types */
3083 set_bit(new_master, mle->maybe_map);
3084 list_add(&mle->list, &dlm->master_list);
3090 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3092 struct list_head *iter, *iter2;
3093 struct dlm_master_list_entry *mle;
3094 struct dlm_lock_resource *res;
3097 mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node);
3099 assert_spin_locked(&dlm->spinlock);
3101 /* clean the master list */
3102 spin_lock(&dlm->master_lock);
3103 list_for_each_safe(iter, iter2, &dlm->master_list) {
3104 mle = list_entry(iter, struct dlm_master_list_entry, list);
3106 BUG_ON(mle->type != DLM_MLE_BLOCK &&
3107 mle->type != DLM_MLE_MASTER &&
3108 mle->type != DLM_MLE_MIGRATION);
3110 /* MASTER mles are initiated locally. the waiting
3111 * process will notice the node map change
3112 * shortly. let that happen as normal. */
3113 if (mle->type == DLM_MLE_MASTER)
3117 /* BLOCK mles are initiated by other nodes.
3118 * need to clean up if the dead node would have
3119 * been the master. */
3120 if (mle->type == DLM_MLE_BLOCK) {
3123 spin_lock(&mle->spinlock);
3124 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3125 if (bit != dead_node) {
3126 mlog(0, "mle found, but dead node %u would "
3127 "not have been master\n", dead_node);
3128 spin_unlock(&mle->spinlock);
3130 /* must drop the refcount by one since the
3131 * assert_master will never arrive. this
3132 * may result in the mle being unlinked and
3133 * freed, but there may still be a process
3134 * waiting in the dlmlock path which is fine. */
3135 mlog(0, "node %u was expected master\n",
3137 atomic_set(&mle->woken, 1);
3138 spin_unlock(&mle->spinlock);
3140 /* do not need events any longer, so detach
3142 __dlm_mle_detach_hb_events(dlm, mle);
3148 /* everything else is a MIGRATION mle */
3150 /* the rule for MIGRATION mles is that the master
3151 * becomes UNKNOWN if *either* the original or
3152 * the new master dies. all UNKNOWN lockreses
3153 * are sent to whichever node becomes the recovery
3154 * master. the new master is responsible for
3155 * determining if there is still a master for
3156 * this lockres, or if he needs to take over
3157 * mastery. either way, this node should expect
3158 * another message to resolve this. */
3159 if (mle->master != dead_node &&
3160 mle->new_master != dead_node)
3163 /* if we have reached this point, this mle needs to
3164 * be removed from the list and freed. */
3166 /* remove from the list early. NOTE: unlinking
3167 * list_head while in list_for_each_safe */
3168 __dlm_mle_detach_hb_events(dlm, mle);
3169 spin_lock(&mle->spinlock);
3170 list_del_init(&mle->list);
3171 atomic_set(&mle->woken, 1);
3172 spin_unlock(&mle->spinlock);
3175 mlog(0, "%s: node %u died during migration from "
3176 "%u to %u!\n", dlm->name, dead_node,
3177 mle->master, mle->new_master);
3178 /* if there is a lockres associated with this
3179 * mle, find it and set its owner to UNKNOWN */
3180 hash = dlm_lockid_hash(mle->u.name.name, mle->u.name.len);
3181 res = __dlm_lookup_lockres(dlm, mle->u.name.name,
3182 mle->u.name.len, hash);
3184 /* unfortunately if we hit this rare case, our
3185 * lock ordering is messed. we need to drop
3186 * the master lock so that we can take the
3187 * lockres lock, meaning that we will have to
3188 * restart from the head of list. */
3189 spin_unlock(&dlm->master_lock);
3191 /* move lockres onto recovery list */
3192 spin_lock(&res->spinlock);
3193 dlm_set_lockres_owner(dlm, res,
3194 DLM_LOCK_RES_OWNER_UNKNOWN);
3195 dlm_move_lockres_to_recovery_list(dlm, res);
3196 spin_unlock(&res->spinlock);
3197 dlm_lockres_put(res);
3199 /* about to get rid of mle, detach from heartbeat */
3200 __dlm_mle_detach_hb_events(dlm, mle);
3203 spin_lock(&dlm->master_lock);
3205 spin_unlock(&dlm->master_lock);
3211 /* this may be the last reference */
3214 spin_unlock(&dlm->master_lock);
3218 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3221 struct dlm_node_iter iter;
3224 spin_lock(&dlm->spinlock);
3225 dlm_node_iter_init(dlm->domain_map, &iter);
3226 clear_bit(old_master, iter.node_map);
3227 clear_bit(dlm->node_num, iter.node_map);
3228 spin_unlock(&dlm->spinlock);
3230 /* ownership of the lockres is changing. account for the
3231 * mastery reference here since old_master will briefly have
3232 * a reference after the migration completes */
3233 spin_lock(&res->spinlock);
3234 dlm_lockres_set_refmap_bit(old_master, res);
3235 spin_unlock(&res->spinlock);
3237 mlog(0, "now time to do a migrate request to other nodes\n");
3238 ret = dlm_do_migrate_request(dlm, res, old_master,
3239 dlm->node_num, &iter);
3245 mlog(0, "doing assert master of %.*s to all except the original node\n",
3246 res->lockname.len, res->lockname.name);
3247 /* this call now finishes out the nodemap
3248 * even if one or more nodes die */
3249 ret = dlm_do_assert_master(dlm, res, iter.node_map,
3250 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3252 /* no longer need to retry. all living nodes contacted. */
3257 memset(iter.node_map, 0, sizeof(iter.node_map));
3258 set_bit(old_master, iter.node_map);
3259 mlog(0, "doing assert master of %.*s back to %u\n",
3260 res->lockname.len, res->lockname.name, old_master);
3261 ret = dlm_do_assert_master(dlm, res, iter.node_map,
3262 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3264 mlog(0, "assert master to original master failed "
3266 /* the only nonzero status here would be because of
3267 * a dead original node. we're done. */
3271 /* all done, set the owner, clear the flag */
3272 spin_lock(&res->spinlock);
3273 dlm_set_lockres_owner(dlm, res, dlm->node_num);
3274 res->state &= ~DLM_LOCK_RES_MIGRATING;
3275 spin_unlock(&res->spinlock);
3276 /* re-dirty it on the new master */
3277 dlm_kick_thread(dlm, res);
3284 * LOCKRES AST REFCOUNT
3285 * this is integral to migration
3288 /* for future intent to call an ast, reserve one ahead of time.
3289 * this should be called only after waiting on the lockres
3290 * with dlm_wait_on_lockres, and while still holding the
3291 * spinlock after the call. */
3292 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3294 assert_spin_locked(&res->spinlock);
3295 if (res->state & DLM_LOCK_RES_MIGRATING) {
3296 __dlm_print_one_lock_resource(res);
3298 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3300 atomic_inc(&res->asts_reserved);
3304 * used to drop the reserved ast, either because it went unused,
3305 * or because the ast/bast was actually called.
3307 * also, if there is a pending migration on this lockres,
3308 * and this was the last pending ast on the lockres,
3309 * atomically set the MIGRATING flag before we drop the lock.
3310 * this is how we ensure that migration can proceed with no
3311 * asts in progress. note that it is ok if the state of the
3312 * queues is such that a lock should be granted in the future
3313 * or that a bast should be fired, because the new master will
3314 * shuffle the lists on this lockres as soon as it is migrated.
3316 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3317 struct dlm_lock_resource *res)
3319 if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3322 if (!res->migration_pending) {
3323 spin_unlock(&res->spinlock);
3327 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3328 res->migration_pending = 0;
3329 res->state |= DLM_LOCK_RES_MIGRATING;
3330 spin_unlock(&res->spinlock);
3332 wake_up(&dlm->migration_wq);