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"
51 #include "dlmdomain.h"
53 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
54 #include "cluster/masklog.h"
65 u8 name[DLM_LOCKID_NAME_MAX];
68 struct dlm_master_list_entry
70 struct list_head list;
71 struct list_head hb_events;
78 unsigned long maybe_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
79 unsigned long vote_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
80 unsigned long response_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
81 unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
84 enum dlm_mle_type type;
85 struct o2hb_callback_func mle_hb_up;
86 struct o2hb_callback_func mle_hb_down;
88 struct dlm_lock_resource *res;
89 struct dlm_lock_name name;
93 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
94 struct dlm_master_list_entry *mle,
95 struct o2nm_node *node,
97 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
98 struct dlm_master_list_entry *mle,
99 struct o2nm_node *node,
102 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
103 static int dlm_do_assert_master(struct dlm_ctxt *dlm, const char *lockname,
104 unsigned int namelen, void *nodemap,
107 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
108 struct dlm_master_list_entry *mle,
110 unsigned int namelen)
112 struct dlm_lock_resource *res;
117 if (mle->type == DLM_MLE_BLOCK ||
118 mle->type == DLM_MLE_MIGRATION) {
119 if (namelen != mle->u.name.len ||
120 memcmp(name, mle->u.name.name, namelen)!=0)
124 if (namelen != res->lockname.len ||
125 memcmp(res->lockname.name, name, namelen) != 0)
132 /* Code here is included but defined out as it aids debugging */
134 #define dlm_print_nodemap(m) _dlm_print_nodemap(m,#m)
135 void _dlm_print_nodemap(unsigned long *map, const char *mapname)
138 printk("%s=[ ", mapname);
139 for (i=0; i<O2NM_MAX_NODES; i++)
140 if (test_bit(i, map))
145 void dlm_print_one_mle(struct dlm_master_list_entry *mle)
151 unsigned int namelen;
154 unsigned long *maybe = mle->maybe_map,
155 *vote = mle->vote_map,
156 *resp = mle->response_map,
157 *node = mle->node_map;
160 if (mle->type == DLM_MLE_BLOCK)
162 else if (mle->type == DLM_MLE_MASTER)
166 refs = atomic_read(&k->refcount);
167 master = mle->master;
168 attached = (list_empty(&mle->hb_events) ? 'N' : 'Y');
170 if (mle->type != DLM_MLE_MASTER) {
171 namelen = mle->u.name.len;
172 name = mle->u.name.name;
174 namelen = mle->u.res->lockname.len;
175 name = mle->u.res->lockname.name;
178 mlog(ML_NOTICE, "%.*s: %3s refs=%3d mas=%3u new=%3u evt=%c inuse=%d ",
179 namelen, name, type, refs, master, mle->new_master, attached,
181 dlm_print_nodemap(maybe);
183 dlm_print_nodemap(vote);
185 dlm_print_nodemap(resp);
187 dlm_print_nodemap(node);
192 static void dlm_dump_mles(struct dlm_ctxt *dlm)
194 struct dlm_master_list_entry *mle;
195 struct list_head *iter;
197 mlog(ML_NOTICE, "dumping all mles for domain %s:\n", dlm->name);
198 spin_lock(&dlm->master_lock);
199 list_for_each(iter, &dlm->master_list) {
200 mle = list_entry(iter, struct dlm_master_list_entry, list);
201 dlm_print_one_mle(mle);
203 spin_unlock(&dlm->master_lock);
206 int dlm_dump_all_mles(const char __user *data, unsigned int len)
208 struct list_head *iter;
209 struct dlm_ctxt *dlm;
211 spin_lock(&dlm_domain_lock);
212 list_for_each(iter, &dlm_domains) {
213 dlm = list_entry (iter, struct dlm_ctxt, list);
214 mlog(ML_NOTICE, "found dlm: %p, name=%s\n", dlm, dlm->name);
217 spin_unlock(&dlm_domain_lock);
220 EXPORT_SYMBOL_GPL(dlm_dump_all_mles);
225 static kmem_cache_t *dlm_mle_cache = NULL;
228 static void dlm_mle_release(struct kref *kref);
229 static void dlm_init_mle(struct dlm_master_list_entry *mle,
230 enum dlm_mle_type type,
231 struct dlm_ctxt *dlm,
232 struct dlm_lock_resource *res,
234 unsigned int namelen);
235 static void dlm_put_mle(struct dlm_master_list_entry *mle);
236 static void __dlm_put_mle(struct dlm_master_list_entry *mle);
237 static int dlm_find_mle(struct dlm_ctxt *dlm,
238 struct dlm_master_list_entry **mle,
239 char *name, unsigned int namelen);
241 static int dlm_do_master_request(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;
692 kref_init(&res->refs);
694 /* just for consistency */
695 spin_lock(&res->spinlock);
696 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
697 spin_unlock(&res->spinlock);
699 res->state = DLM_LOCK_RES_IN_PROGRESS;
703 memset(res->lvb, 0, DLM_LVB_LEN);
706 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
708 unsigned int namelen)
710 struct dlm_lock_resource *res;
712 res = kmalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
716 res->lockname.name = kmalloc(namelen, GFP_KERNEL);
717 if (!res->lockname.name) {
722 dlm_init_lockres(dlm, res, name, namelen);
727 * lookup a lock resource by name.
728 * may already exist in the hashtable.
729 * lockid is null terminated
731 * if not, allocate enough for the lockres and for
732 * the temporary structure used in doing the mastering.
734 * also, do a lookup in the dlm->master_list to see
735 * if another node has begun mastering the same lock.
736 * if so, there should be a block entry in there
737 * for this name, and we should *not* attempt to master
738 * the lock here. need to wait around for that node
739 * to assert_master (or die).
742 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
746 struct dlm_lock_resource *tmpres=NULL, *res=NULL;
747 struct dlm_master_list_entry *mle = NULL;
748 struct dlm_master_list_entry *alloc_mle = NULL;
751 struct dlm_node_iter iter;
752 unsigned int namelen, hash;
754 int bit, wait_on_recovery = 0;
758 namelen = strlen(lockid);
759 hash = dlm_lockid_hash(lockid, namelen);
761 mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
764 spin_lock(&dlm->spinlock);
765 tmpres = __dlm_lookup_lockres(dlm, lockid, namelen, hash);
767 spin_unlock(&dlm->spinlock);
768 mlog(0, "found in hash!\n");
770 dlm_lockres_put(res);
776 spin_unlock(&dlm->spinlock);
777 mlog(0, "allocating a new resource\n");
778 /* nothing found and we need to allocate one. */
779 alloc_mle = (struct dlm_master_list_entry *)
780 kmem_cache_alloc(dlm_mle_cache, GFP_KERNEL);
783 res = dlm_new_lockres(dlm, lockid, namelen);
789 mlog(0, "no lockres found, allocated our own: %p\n", res);
791 if (flags & LKM_LOCAL) {
792 /* caller knows it's safe to assume it's not mastered elsewhere
793 * DONE! return right away */
794 spin_lock(&res->spinlock);
795 dlm_change_lockres_owner(dlm, res, dlm->node_num);
796 __dlm_insert_lockres(dlm, res);
797 spin_unlock(&res->spinlock);
798 spin_unlock(&dlm->spinlock);
799 /* lockres still marked IN_PROGRESS */
803 /* check master list to see if another node has started mastering it */
804 spin_lock(&dlm->master_lock);
806 /* if we found a block, wait for lock to be mastered by another node */
807 blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
809 if (mle->type == DLM_MLE_MASTER) {
810 mlog(ML_ERROR, "master entry for nonexistent lock!\n");
812 } else if (mle->type == DLM_MLE_MIGRATION) {
813 /* migration is in progress! */
814 /* the good news is that we now know the
815 * "current" master (mle->master). */
817 spin_unlock(&dlm->master_lock);
818 assert_spin_locked(&dlm->spinlock);
820 /* set the lockres owner and hash it */
821 spin_lock(&res->spinlock);
822 dlm_set_lockres_owner(dlm, res, mle->master);
823 __dlm_insert_lockres(dlm, res);
824 spin_unlock(&res->spinlock);
825 spin_unlock(&dlm->spinlock);
827 /* master is known, detach */
828 dlm_mle_detach_hb_events(dlm, mle);
834 /* go ahead and try to master lock on this node */
836 /* make sure this does not get freed below */
838 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
839 set_bit(dlm->node_num, mle->maybe_map);
840 list_add(&mle->list, &dlm->master_list);
842 /* still holding the dlm spinlock, check the recovery map
843 * to see if there are any nodes that still need to be
844 * considered. these will not appear in the mle nodemap
845 * but they might own this lockres. wait on them. */
846 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
847 if (bit < O2NM_MAX_NODES) {
848 mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to"
849 "recover before lock mastery can begin\n",
850 dlm->name, namelen, (char *)lockid, bit);
851 wait_on_recovery = 1;
855 /* at this point there is either a DLM_MLE_BLOCK or a
856 * DLM_MLE_MASTER on the master list, so it's safe to add the
857 * lockres to the hashtable. anyone who finds the lock will
858 * still have to wait on the IN_PROGRESS. */
860 /* finally add the lockres to its hash bucket */
861 __dlm_insert_lockres(dlm, res);
862 /* get an extra ref on the mle in case this is a BLOCK
863 * if so, the creator of the BLOCK may try to put the last
864 * ref at this time in the assert master handler, so we
865 * need an extra one to keep from a bad ptr deref. */
866 dlm_get_mle_inuse(mle);
867 spin_unlock(&dlm->master_lock);
868 spin_unlock(&dlm->spinlock);
871 while (wait_on_recovery) {
872 /* any cluster changes that occurred after dropping the
873 * dlm spinlock would be detectable be a change on the mle,
874 * so we only need to clear out the recovery map once. */
875 if (dlm_is_recovery_lock(lockid, namelen)) {
876 mlog(ML_NOTICE, "%s: recovery map is not empty, but "
877 "must master $RECOVERY lock now\n", dlm->name);
878 if (!dlm_pre_master_reco_lockres(dlm, res))
879 wait_on_recovery = 0;
881 mlog(0, "%s: waiting 500ms for heartbeat state "
882 "change\n", dlm->name);
888 dlm_kick_recovery_thread(dlm);
890 dlm_wait_for_recovery(dlm);
892 spin_lock(&dlm->spinlock);
893 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
894 if (bit < O2NM_MAX_NODES) {
895 mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to"
896 "recover before lock mastery can begin\n",
897 dlm->name, namelen, (char *)lockid, bit);
898 wait_on_recovery = 1;
900 wait_on_recovery = 0;
901 spin_unlock(&dlm->spinlock);
903 if (wait_on_recovery)
904 dlm_wait_for_node_recovery(dlm, bit, 10000);
907 /* must wait for lock to be mastered elsewhere */
912 dlm_node_iter_init(mle->vote_map, &iter);
913 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
914 ret = dlm_do_master_request(mle, nodenum);
917 if (mle->master != O2NM_MAX_NODES) {
918 /* found a master ! */
919 if (mle->master <= nodenum)
921 /* if our master request has not reached the master
922 * yet, keep going until it does. this is how the
923 * master will know that asserts are needed back to
924 * the lower nodes. */
925 mlog(0, "%s:%.*s: requests only up to %u but master "
926 "is %u, keep going\n", dlm->name, namelen,
927 lockid, nodenum, mle->master);
932 /* keep going until the response map includes all nodes */
933 ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
935 wait_on_recovery = 1;
936 mlog(0, "%s:%.*s: node map changed, redo the "
937 "master request now, blocked=%d\n",
938 dlm->name, res->lockname.len,
939 res->lockname.name, blocked);
941 mlog(ML_ERROR, "%s:%.*s: spinning on "
942 "dlm_wait_for_lock_mastery, blocked=%d\n",
943 dlm->name, res->lockname.len,
944 res->lockname.name, blocked);
945 dlm_print_one_lock_resource(res);
946 /* dlm_print_one_mle(mle); */
952 mlog(0, "lockres mastered by %u\n", res->owner);
953 /* make sure we never continue without this */
954 BUG_ON(res->owner == O2NM_MAX_NODES);
956 /* master is known, detach if not already detached */
957 dlm_mle_detach_hb_events(dlm, mle);
959 /* put the extra ref */
960 dlm_put_mle_inuse(mle);
963 spin_lock(&res->spinlock);
964 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
965 spin_unlock(&res->spinlock);
969 /* need to free the unused mle */
971 kmem_cache_free(dlm_mle_cache, alloc_mle);
977 #define DLM_MASTERY_TIMEOUT_MS 5000
979 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
980 struct dlm_lock_resource *res,
981 struct dlm_master_list_entry *mle,
986 int map_changed, voting_done;
993 /* check if another node has already become the owner */
994 spin_lock(&res->spinlock);
995 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
996 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
997 res->lockname.len, res->lockname.name, res->owner);
998 spin_unlock(&res->spinlock);
999 /* this will cause the master to re-assert across
1000 * the whole cluster, freeing up mles */
1001 if (res->owner != dlm->node_num) {
1002 ret = dlm_do_master_request(mle, res->owner);
1004 /* give recovery a chance to run */
1005 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1013 spin_unlock(&res->spinlock);
1015 spin_lock(&mle->spinlock);
1017 map_changed = (memcmp(mle->vote_map, mle->node_map,
1018 sizeof(mle->vote_map)) != 0);
1019 voting_done = (memcmp(mle->vote_map, mle->response_map,
1020 sizeof(mle->vote_map)) == 0);
1022 /* restart if we hit any errors */
1025 mlog(0, "%s: %.*s: node map changed, restarting\n",
1026 dlm->name, res->lockname.len, res->lockname.name);
1027 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1028 b = (mle->type == DLM_MLE_BLOCK);
1029 if ((*blocked && !b) || (!*blocked && b)) {
1030 mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1031 dlm->name, res->lockname.len, res->lockname.name,
1035 spin_unlock(&mle->spinlock);
1040 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1041 "rechecking now\n", dlm->name, res->lockname.len,
1042 res->lockname.name);
1046 mlog(0, "map not changed and voting not done "
1047 "for %s:%.*s\n", dlm->name, res->lockname.len,
1048 res->lockname.name);
1052 if (m != O2NM_MAX_NODES) {
1053 /* another node has done an assert!
1058 /* have all nodes responded? */
1059 if (voting_done && !*blocked) {
1060 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1061 if (dlm->node_num <= bit) {
1062 /* my node number is lowest.
1063 * now tell other nodes that I am
1064 * mastering this. */
1065 mle->master = dlm->node_num;
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, "waiting again\n");
1094 mlog(0, "done waiting, master is %u\n", res->owner);
1102 mlog(0, "about to master %.*s here, this=%u\n",
1103 res->lockname.len, res->lockname.name, m);
1104 ret = dlm_do_assert_master(dlm, res->lockname.name,
1105 res->lockname.len, 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 dlm_change_lockres_owner(dlm, res, m);
1122 spin_unlock(&res->spinlock);
1128 struct dlm_bitmap_diff_iter
1131 unsigned long *orig_bm;
1132 unsigned long *cur_bm;
1133 unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1136 enum dlm_node_state_change
1143 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1144 unsigned long *orig_bm,
1145 unsigned long *cur_bm)
1147 unsigned long p1, p2;
1151 iter->orig_bm = orig_bm;
1152 iter->cur_bm = cur_bm;
1154 for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1155 p1 = *(iter->orig_bm + i);
1156 p2 = *(iter->cur_bm + i);
1157 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1161 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1162 enum dlm_node_state_change *state)
1166 if (iter->curnode >= O2NM_MAX_NODES)
1169 bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1171 if (bit >= O2NM_MAX_NODES) {
1172 iter->curnode = O2NM_MAX_NODES;
1176 /* if it was there in the original then this node died */
1177 if (test_bit(bit, iter->orig_bm))
1182 iter->curnode = bit;
1187 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1188 struct dlm_lock_resource *res,
1189 struct dlm_master_list_entry *mle,
1192 struct dlm_bitmap_diff_iter bdi;
1193 enum dlm_node_state_change sc;
1197 mlog(0, "something happened such that the "
1198 "master process may need to be restarted!\n");
1200 assert_spin_locked(&mle->spinlock);
1202 dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1203 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1205 if (sc == NODE_UP) {
1206 /* a node came up. clear any old vote from
1207 * the response map and set it in the vote map
1208 * then restart the mastery. */
1209 mlog(ML_NOTICE, "node %d up while restarting\n", node);
1211 /* redo the master request, but only for the new node */
1212 mlog(0, "sending request to new node\n");
1213 clear_bit(node, mle->response_map);
1214 set_bit(node, mle->vote_map);
1216 mlog(ML_ERROR, "node down! %d\n", node);
1218 int lowest = find_next_bit(mle->maybe_map,
1221 /* act like it was never there */
1222 clear_bit(node, mle->maybe_map);
1224 if (node == lowest) {
1225 mlog(0, "expected master %u died"
1226 " while this node was blocked "
1227 "waiting on it!\n", node);
1228 lowest = find_next_bit(mle->maybe_map,
1231 if (lowest < O2NM_MAX_NODES) {
1232 mlog(0, "%s:%.*s:still "
1233 "blocked. waiting on %u "
1239 /* mle is an MLE_BLOCK, but
1240 * there is now nothing left to
1241 * block on. we need to return
1242 * all the way back out and try
1243 * again with an MLE_MASTER.
1244 * dlm_do_local_recovery_cleanup
1245 * has already run, so the mle
1247 mlog(0, "%s:%.*s: no "
1248 "longer blocking. try to "
1249 "master this here\n",
1252 res->lockname.name);
1253 mle->type = DLM_MLE_MASTER;
1259 /* now blank out everything, as if we had never
1260 * contacted anyone */
1261 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1262 memset(mle->response_map, 0, sizeof(mle->response_map));
1263 /* reset the vote_map to the current node_map */
1264 memcpy(mle->vote_map, mle->node_map,
1265 sizeof(mle->node_map));
1266 /* put myself into the maybe map */
1267 if (mle->type != DLM_MLE_BLOCK)
1268 set_bit(dlm->node_num, mle->maybe_map);
1271 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1278 * DLM_MASTER_REQUEST_MSG
1280 * returns: 0 on success,
1281 * -errno on a network error
1283 * on error, the caller should assume the target node is "dead"
1287 static int dlm_do_master_request(struct dlm_master_list_entry *mle, int to)
1289 struct dlm_ctxt *dlm = mle->dlm;
1290 struct dlm_master_request request;
1291 int ret, response=0, resend;
1293 memset(&request, 0, sizeof(request));
1294 request.node_idx = dlm->node_num;
1296 BUG_ON(mle->type == DLM_MLE_MIGRATION);
1298 if (mle->type != DLM_MLE_MASTER) {
1299 request.namelen = mle->u.name.len;
1300 memcpy(request.name, mle->u.name.name, request.namelen);
1302 request.namelen = mle->u.res->lockname.len;
1303 memcpy(request.name, mle->u.res->lockname.name,
1308 ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1309 sizeof(request), to, &response);
1311 if (ret == -ESRCH) {
1312 /* should never happen */
1313 mlog(ML_ERROR, "TCP stack not ready!\n");
1315 } else if (ret == -EINVAL) {
1316 mlog(ML_ERROR, "bad args passed to o2net!\n");
1318 } else if (ret == -ENOMEM) {
1319 mlog(ML_ERROR, "out of memory while trying to send "
1320 "network message! retrying\n");
1321 /* this is totally crude */
1324 } else if (!dlm_is_host_down(ret)) {
1325 /* not a network error. bad. */
1327 mlog(ML_ERROR, "unhandled error!");
1330 /* all other errors should be network errors,
1331 * and likely indicate node death */
1332 mlog(ML_ERROR, "link to %d went down!\n", to);
1338 spin_lock(&mle->spinlock);
1340 case DLM_MASTER_RESP_YES:
1341 set_bit(to, mle->response_map);
1342 mlog(0, "node %u is the master, response=YES\n", 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)
1385 u8 response = DLM_MASTER_RESP_MAYBE;
1386 struct dlm_ctxt *dlm = data;
1387 struct dlm_lock_resource *res = NULL;
1388 struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1389 struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1391 unsigned int namelen, hash;
1394 int dispatch_assert = 0;
1397 return DLM_MASTER_RESP_NO;
1399 if (!dlm_domain_fully_joined(dlm)) {
1400 response = DLM_MASTER_RESP_NO;
1404 name = request->name;
1405 namelen = request->namelen;
1406 hash = dlm_lockid_hash(name, namelen);
1408 if (namelen > DLM_LOCKID_NAME_MAX) {
1409 response = DLM_IVBUFLEN;
1414 spin_lock(&dlm->spinlock);
1415 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1417 spin_unlock(&dlm->spinlock);
1419 /* take care of the easy cases up front */
1420 spin_lock(&res->spinlock);
1421 if (res->state & DLM_LOCK_RES_RECOVERING) {
1422 spin_unlock(&res->spinlock);
1423 mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1424 "being recovered\n");
1425 response = DLM_MASTER_RESP_ERROR;
1427 kmem_cache_free(dlm_mle_cache, mle);
1431 if (res->owner == dlm->node_num) {
1432 spin_unlock(&res->spinlock);
1433 // mlog(0, "this node is the master\n");
1434 response = DLM_MASTER_RESP_YES;
1436 kmem_cache_free(dlm_mle_cache, mle);
1438 /* this node is the owner.
1439 * there is some extra work that needs to
1440 * happen now. the requesting node has
1441 * caused all nodes up to this one to
1442 * create mles. this node now needs to
1443 * go back and clean those up. */
1444 dispatch_assert = 1;
1446 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1447 spin_unlock(&res->spinlock);
1448 // mlog(0, "node %u is the master\n", res->owner);
1449 response = DLM_MASTER_RESP_NO;
1451 kmem_cache_free(dlm_mle_cache, mle);
1455 /* ok, there is no owner. either this node is
1456 * being blocked, or it is actively trying to
1457 * master this lock. */
1458 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1459 mlog(ML_ERROR, "lock with no owner should be "
1464 // mlog(0, "lockres is in progress...\n");
1465 spin_lock(&dlm->master_lock);
1466 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1468 mlog(ML_ERROR, "no mle found for this lock!\n");
1472 spin_lock(&tmpmle->spinlock);
1473 if (tmpmle->type == DLM_MLE_BLOCK) {
1474 // mlog(0, "this node is waiting for "
1475 // "lockres to be mastered\n");
1476 response = DLM_MASTER_RESP_NO;
1477 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1478 mlog(0, "node %u is master, but trying to migrate to "
1479 "node %u.\n", tmpmle->master, tmpmle->new_master);
1480 if (tmpmle->master == dlm->node_num) {
1481 response = DLM_MASTER_RESP_YES;
1482 mlog(ML_ERROR, "no owner on lockres, but this "
1483 "node is trying to migrate it to %u?!\n",
1484 tmpmle->new_master);
1487 /* the real master can respond on its own */
1488 response = DLM_MASTER_RESP_NO;
1490 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1492 if (tmpmle->master == dlm->node_num) {
1493 response = DLM_MASTER_RESP_YES;
1494 /* this node will be the owner.
1495 * go back and clean the mles on any
1497 dispatch_assert = 1;
1499 response = DLM_MASTER_RESP_NO;
1501 // mlog(0, "this node is attempting to "
1502 // "master lockres\n");
1503 response = DLM_MASTER_RESP_MAYBE;
1506 set_bit(request->node_idx, tmpmle->maybe_map);
1507 spin_unlock(&tmpmle->spinlock);
1509 spin_unlock(&dlm->master_lock);
1510 spin_unlock(&res->spinlock);
1512 /* keep the mle attached to heartbeat events */
1513 dlm_put_mle(tmpmle);
1515 kmem_cache_free(dlm_mle_cache, mle);
1520 * lockres doesn't exist on this node
1521 * if there is an MLE_BLOCK, return NO
1522 * if there is an MLE_MASTER, return MAYBE
1523 * otherwise, add an MLE_BLOCK, return NO
1525 spin_lock(&dlm->master_lock);
1526 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1528 /* this lockid has never been seen on this node yet */
1529 // mlog(0, "no mle found\n");
1531 spin_unlock(&dlm->master_lock);
1532 spin_unlock(&dlm->spinlock);
1534 mle = (struct dlm_master_list_entry *)
1535 kmem_cache_alloc(dlm_mle_cache, GFP_KERNEL);
1537 response = DLM_MASTER_RESP_ERROR;
1538 mlog_errno(-ENOMEM);
1544 // mlog(0, "this is second time thru, already allocated, "
1545 // "add the block.\n");
1546 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1547 set_bit(request->node_idx, mle->maybe_map);
1548 list_add(&mle->list, &dlm->master_list);
1549 response = DLM_MASTER_RESP_NO;
1551 // mlog(0, "mle was found\n");
1553 spin_lock(&tmpmle->spinlock);
1554 if (tmpmle->master == dlm->node_num) {
1555 mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1558 if (tmpmle->type == DLM_MLE_BLOCK)
1559 response = DLM_MASTER_RESP_NO;
1560 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1561 mlog(0, "migration mle was found (%u->%u)\n",
1562 tmpmle->master, tmpmle->new_master);
1563 /* real master can respond on its own */
1564 response = DLM_MASTER_RESP_NO;
1566 response = DLM_MASTER_RESP_MAYBE;
1568 set_bit(request->node_idx, tmpmle->maybe_map);
1569 spin_unlock(&tmpmle->spinlock);
1571 spin_unlock(&dlm->master_lock);
1572 spin_unlock(&dlm->spinlock);
1575 /* keep the mle attached to heartbeat events */
1576 dlm_put_mle(tmpmle);
1580 if (dispatch_assert) {
1581 if (response != DLM_MASTER_RESP_YES)
1582 mlog(ML_ERROR, "invalid response %d\n", response);
1584 mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1587 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1588 dlm->node_num, res->lockname.len, res->lockname.name);
1589 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1590 DLM_ASSERT_MASTER_MLE_CLEANUP);
1592 mlog(ML_ERROR, "failed to dispatch assert master work\n");
1593 response = DLM_MASTER_RESP_ERROR;
1602 * DLM_ASSERT_MASTER_MSG
1607 * NOTE: this can be used for debugging
1608 * can periodically run all locks owned by this node
1609 * and re-assert across the cluster...
1611 static int dlm_do_assert_master(struct dlm_ctxt *dlm, const char *lockname,
1612 unsigned int namelen, void *nodemap,
1615 struct dlm_assert_master assert;
1617 struct dlm_node_iter iter;
1621 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1625 /* note that if this nodemap is empty, it returns 0 */
1626 dlm_node_iter_init(nodemap, &iter);
1627 while ((to = dlm_node_iter_next(&iter)) >= 0) {
1629 struct dlm_master_list_entry *mle = NULL;
1631 mlog(0, "sending assert master to %d (%.*s)\n", to,
1633 memset(&assert, 0, sizeof(assert));
1634 assert.node_idx = dlm->node_num;
1635 assert.namelen = namelen;
1636 memcpy(assert.name, lockname, namelen);
1637 assert.flags = cpu_to_be32(flags);
1639 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1640 &assert, sizeof(assert), to, &r);
1642 mlog(ML_ERROR, "assert_master returned %d!\n", tmpret);
1643 if (!dlm_is_host_down(tmpret)) {
1644 mlog(ML_ERROR, "unhandled error!\n");
1647 /* a node died. finish out the rest of the nodes. */
1648 mlog(ML_ERROR, "link to %d went down!\n", to);
1649 /* any nonzero status return will do */
1652 /* ok, something horribly messed. kill thyself. */
1653 mlog(ML_ERROR,"during assert master of %.*s to %u, "
1654 "got %d.\n", namelen, lockname, to, r);
1655 spin_lock(&dlm->spinlock);
1656 spin_lock(&dlm->master_lock);
1657 if (dlm_find_mle(dlm, &mle, (char *)lockname,
1659 dlm_print_one_mle(mle);
1662 spin_unlock(&dlm->master_lock);
1663 spin_unlock(&dlm->spinlock);
1665 } else if (r == EAGAIN) {
1666 mlog(0, "%.*s: node %u create mles on other "
1667 "nodes and requests a re-assert\n",
1668 namelen, lockname, to);
1680 * locks that can be taken here:
1686 * if possible, TRIM THIS DOWN!!!
1688 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data)
1690 struct dlm_ctxt *dlm = data;
1691 struct dlm_master_list_entry *mle = NULL;
1692 struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1693 struct dlm_lock_resource *res = NULL;
1695 unsigned int namelen, hash;
1697 int master_request = 0;
1703 name = assert->name;
1704 namelen = assert->namelen;
1705 hash = dlm_lockid_hash(name, namelen);
1706 flags = be32_to_cpu(assert->flags);
1708 if (namelen > DLM_LOCKID_NAME_MAX) {
1709 mlog(ML_ERROR, "Invalid name length!");
1713 spin_lock(&dlm->spinlock);
1716 mlog(0, "assert_master with flags: %u\n", flags);
1719 spin_lock(&dlm->master_lock);
1720 if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1721 /* not an error, could be master just re-asserting */
1722 mlog(0, "just got an assert_master from %u, but no "
1723 "MLE for it! (%.*s)\n", assert->node_idx,
1726 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1727 if (bit >= O2NM_MAX_NODES) {
1728 /* not necessarily an error, though less likely.
1729 * could be master just re-asserting. */
1730 mlog(0, "no bits set in the maybe_map, but %u "
1731 "is asserting! (%.*s)\n", assert->node_idx,
1733 } else if (bit != assert->node_idx) {
1734 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1735 mlog(0, "master %u was found, %u should "
1736 "back off\n", assert->node_idx, bit);
1738 /* with the fix for bug 569, a higher node
1739 * number winning the mastery will respond
1740 * YES to mastery requests, but this node
1741 * had no way of knowing. let it pass. */
1742 mlog(0, "%u is the lowest node, "
1743 "%u is asserting. (%.*s) %u must "
1744 "have begun after %u won.\n", bit,
1745 assert->node_idx, namelen, name, bit,
1749 if (mle->type == DLM_MLE_MIGRATION) {
1750 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1751 mlog(0, "%s:%.*s: got cleanup assert"
1752 " from %u for migration\n",
1753 dlm->name, namelen, name,
1755 } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1756 mlog(0, "%s:%.*s: got unrelated assert"
1757 " from %u for migration, ignoring\n",
1758 dlm->name, namelen, name,
1761 spin_unlock(&dlm->master_lock);
1762 spin_unlock(&dlm->spinlock);
1767 spin_unlock(&dlm->master_lock);
1769 /* ok everything checks out with the MLE
1770 * now check to see if there is a lockres */
1771 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1773 spin_lock(&res->spinlock);
1774 if (res->state & DLM_LOCK_RES_RECOVERING) {
1775 mlog(ML_ERROR, "%u asserting but %.*s is "
1776 "RECOVERING!\n", assert->node_idx, namelen, name);
1780 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1781 res->owner != assert->node_idx) {
1782 mlog(ML_ERROR, "assert_master from "
1783 "%u, but current owner is "
1785 assert->node_idx, res->owner,
1789 } else if (mle->type != DLM_MLE_MIGRATION) {
1790 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1791 /* owner is just re-asserting */
1792 if (res->owner == assert->node_idx) {
1793 mlog(0, "owner %u re-asserting on "
1794 "lock %.*s\n", assert->node_idx,
1798 mlog(ML_ERROR, "got assert_master from "
1799 "node %u, but %u is the owner! "
1800 "(%.*s)\n", assert->node_idx,
1801 res->owner, namelen, name);
1804 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1805 mlog(ML_ERROR, "got assert from %u, but lock "
1806 "with no owner should be "
1807 "in-progress! (%.*s)\n",
1812 } else /* mle->type == DLM_MLE_MIGRATION */ {
1813 /* should only be getting an assert from new master */
1814 if (assert->node_idx != mle->new_master) {
1815 mlog(ML_ERROR, "got assert from %u, but "
1816 "new master is %u, and old master "
1818 assert->node_idx, mle->new_master,
1819 mle->master, namelen, name);
1825 spin_unlock(&res->spinlock);
1827 spin_unlock(&dlm->spinlock);
1829 // mlog(0, "woo! got an assert_master from node %u!\n",
1830 // assert->node_idx);
1836 spin_lock(&mle->spinlock);
1837 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1840 /* MASTER mle: if any bits set in the response map
1841 * then the calling node needs to re-assert to clear
1842 * up nodes that this node contacted */
1843 while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1844 nn+1)) < O2NM_MAX_NODES) {
1845 if (nn != dlm->node_num && nn != assert->node_idx)
1849 mle->master = assert->node_idx;
1850 atomic_set(&mle->woken, 1);
1852 spin_unlock(&mle->spinlock);
1855 spin_lock(&res->spinlock);
1856 if (mle->type == DLM_MLE_MIGRATION) {
1857 mlog(0, "finishing off migration of lockres %.*s, "
1859 res->lockname.len, res->lockname.name,
1860 dlm->node_num, mle->new_master);
1861 res->state &= ~DLM_LOCK_RES_MIGRATING;
1862 dlm_change_lockres_owner(dlm, res, mle->new_master);
1863 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1865 dlm_change_lockres_owner(dlm, res, mle->master);
1867 spin_unlock(&res->spinlock);
1870 /* master is known, detach if not already detached.
1871 * ensures that only one assert_master call will happen
1873 spin_lock(&dlm->spinlock);
1874 spin_lock(&dlm->master_lock);
1876 rr = atomic_read(&mle->mle_refs.refcount);
1877 if (mle->inuse > 0) {
1878 if (extra_ref && rr < 3)
1880 else if (!extra_ref && rr < 2)
1883 if (extra_ref && rr < 2)
1885 else if (!extra_ref && rr < 1)
1889 mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
1890 "that will mess up this node, refs=%d, extra=%d, "
1891 "inuse=%d\n", dlm->name, namelen, name,
1892 assert->node_idx, rr, extra_ref, mle->inuse);
1893 dlm_print_one_mle(mle);
1895 list_del_init(&mle->list);
1896 __dlm_mle_detach_hb_events(dlm, mle);
1899 /* the assert master message now balances the extra
1900 * ref given by the master / migration request message.
1901 * if this is the last put, it will be removed
1905 spin_unlock(&dlm->master_lock);
1906 spin_unlock(&dlm->spinlock);
1908 if (res->owner != assert->node_idx) {
1909 mlog(0, "assert_master from %u, but current "
1910 "owner is %u (%.*s), no mle\n", assert->node_idx,
1911 res->owner, namelen, name);
1918 dlm_lockres_put(res);
1920 if (master_request) {
1921 mlog(0, "need to tell master to reassert\n");
1922 ret = EAGAIN; // positive. negative would shoot down the node.
1927 /* kill the caller! */
1928 mlog(ML_ERROR, "Bad message received from another node. Dumping state "
1929 "and killing the other node now! This node is OK and can continue.\n");
1930 __dlm_print_one_lock_resource(res);
1931 spin_unlock(&res->spinlock);
1932 spin_unlock(&dlm->spinlock);
1933 dlm_lockres_put(res);
1938 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
1939 struct dlm_lock_resource *res,
1940 int ignore_higher, u8 request_from, u32 flags)
1942 struct dlm_work_item *item;
1943 item = kcalloc(1, sizeof(*item), GFP_KERNEL);
1948 /* queue up work for dlm_assert_master_worker */
1949 dlm_grab(dlm); /* get an extra ref for the work item */
1950 dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
1951 item->u.am.lockres = res; /* already have a ref */
1952 /* can optionally ignore node numbers higher than this node */
1953 item->u.am.ignore_higher = ignore_higher;
1954 item->u.am.request_from = request_from;
1955 item->u.am.flags = flags;
1958 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
1959 res->lockname.name);
1961 spin_lock(&dlm->work_lock);
1962 list_add_tail(&item->list, &dlm->work_list);
1963 spin_unlock(&dlm->work_lock);
1965 schedule_work(&dlm->dispatched_work);
1969 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
1971 struct dlm_ctxt *dlm = data;
1973 struct dlm_lock_resource *res;
1974 unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
1981 res = item->u.am.lockres;
1982 ignore_higher = item->u.am.ignore_higher;
1983 request_from = item->u.am.request_from;
1984 flags = item->u.am.flags;
1986 spin_lock(&dlm->spinlock);
1987 memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
1988 spin_unlock(&dlm->spinlock);
1990 clear_bit(dlm->node_num, nodemap);
1991 if (ignore_higher) {
1992 /* if is this just to clear up mles for nodes below
1993 * this node, do not send the message to the original
1994 * caller or any node number higher than this */
1995 clear_bit(request_from, nodemap);
1996 bit = dlm->node_num;
1998 bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2000 if (bit >= O2NM_MAX_NODES)
2002 clear_bit(bit, nodemap);
2007 * If we're migrating this lock to someone else, we are no
2008 * longer allowed to assert out own mastery. OTOH, we need to
2009 * prevent migration from starting while we're still asserting
2010 * our dominance. The reserved ast delays migration.
2012 spin_lock(&res->spinlock);
2013 if (res->state & DLM_LOCK_RES_MIGRATING) {
2014 mlog(0, "Someone asked us to assert mastery, but we're "
2015 "in the middle of migration. Skipping assert, "
2016 "the new master will handle that.\n");
2017 spin_unlock(&res->spinlock);
2020 __dlm_lockres_reserve_ast(res);
2021 spin_unlock(&res->spinlock);
2023 /* this call now finishes out the nodemap
2024 * even if one or more nodes die */
2025 mlog(0, "worker about to master %.*s here, this=%u\n",
2026 res->lockname.len, res->lockname.name, dlm->node_num);
2027 ret = dlm_do_assert_master(dlm, res->lockname.name,
2031 /* no need to restart, we are done */
2035 /* Ok, we've asserted ourselves. Let's let migration start. */
2036 dlm_lockres_release_ast(dlm, res);
2039 dlm_lockres_put(res);
2041 mlog(0, "finished with dlm_assert_master_worker\n");
2044 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2045 * We cannot wait for node recovery to complete to begin mastering this
2046 * lockres because this lockres is used to kick off recovery! ;-)
2047 * So, do a pre-check on all living nodes to see if any of those nodes
2048 * think that $RECOVERY is currently mastered by a dead node. If so,
2049 * we wait a short time to allow that node to get notified by its own
2050 * heartbeat stack, then check again. All $RECOVERY lock resources
2051 * mastered by dead nodes are purged when the hearbeat callback is
2052 * fired, so we can know for sure that it is safe to continue once
2053 * the node returns a live node or no node. */
2054 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2055 struct dlm_lock_resource *res)
2057 struct dlm_node_iter iter;
2060 u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2062 spin_lock(&dlm->spinlock);
2063 dlm_node_iter_init(dlm->domain_map, &iter);
2064 spin_unlock(&dlm->spinlock);
2066 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2067 /* do not send to self */
2068 if (nodenum == dlm->node_num)
2070 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2073 if (!dlm_is_host_down(ret))
2075 /* host is down, so answer for that node would be
2076 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
2080 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2081 /* check to see if this master is in the recovery map */
2082 spin_lock(&dlm->spinlock);
2083 if (test_bit(master, dlm->recovery_map)) {
2084 mlog(ML_NOTICE, "%s: node %u has not seen "
2085 "node %u go down yet, and thinks the "
2086 "dead node is mastering the recovery "
2087 "lock. must wait.\n", dlm->name,
2091 spin_unlock(&dlm->spinlock);
2092 mlog(0, "%s: reco lock master is %u\n", dlm->name,
2102 * DLM_MIGRATE_LOCKRES
2106 int dlm_migrate_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
2109 struct dlm_master_list_entry *mle = NULL;
2110 struct dlm_master_list_entry *oldmle = NULL;
2111 struct dlm_migratable_lockres *mres = NULL;
2114 unsigned int namelen;
2116 struct list_head *queue, *iter;
2118 struct dlm_lock *lock;
2124 name = res->lockname.name;
2125 namelen = res->lockname.len;
2127 mlog(0, "migrating %.*s to %u\n", namelen, name, target);
2130 * ensure this lockres is a proper candidate for migration
2132 spin_lock(&res->spinlock);
2133 if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
2134 mlog(0, "cannot migrate lockres with unknown owner!\n");
2135 spin_unlock(&res->spinlock);
2138 if (res->owner != dlm->node_num) {
2139 mlog(0, "cannot migrate lockres this node doesn't own!\n");
2140 spin_unlock(&res->spinlock);
2143 mlog(0, "checking queues...\n");
2144 queue = &res->granted;
2145 for (i=0; i<3; i++) {
2146 list_for_each(iter, queue) {
2147 lock = list_entry (iter, struct dlm_lock, list);
2149 if (lock->ml.node == dlm->node_num) {
2150 mlog(0, "found a lock owned by this node "
2151 "still on the %s queue! will not "
2152 "migrate this lockres\n",
2154 (i==1 ? "converting" : "blocked"));
2155 spin_unlock(&res->spinlock);
2162 mlog(0, "all locks on this lockres are nonlocal. continuing\n");
2163 spin_unlock(&res->spinlock);
2167 mlog(0, "no locks were found on this lockres! done!\n");
2173 * preallocate up front
2174 * if this fails, abort
2178 mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_KERNEL);
2184 mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
2193 * find a node to migrate the lockres to
2196 mlog(0, "picking a migration node\n");
2197 spin_lock(&dlm->spinlock);
2198 /* pick a new node */
2199 if (!test_bit(target, dlm->domain_map) ||
2200 target >= O2NM_MAX_NODES) {
2201 target = dlm_pick_migration_target(dlm, res);
2203 mlog(0, "node %u chosen for migration\n", target);
2205 if (target >= O2NM_MAX_NODES ||
2206 !test_bit(target, dlm->domain_map)) {
2207 /* target chosen is not alive */
2212 spin_unlock(&dlm->spinlock);
2216 mlog(0, "continuing with target = %u\n", target);
2219 * clear any existing master requests and
2220 * add the migration mle to the list
2222 spin_lock(&dlm->master_lock);
2223 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2224 namelen, target, dlm->node_num);
2225 spin_unlock(&dlm->master_lock);
2226 spin_unlock(&dlm->spinlock);
2228 if (ret == -EEXIST) {
2229 mlog(0, "another process is already migrating it\n");
2235 * set the MIGRATING flag and flush asts
2236 * if we fail after this we need to re-dirty the lockres
2238 if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2239 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2240 "the target went down.\n", res->lockname.len,
2241 res->lockname.name, target);
2242 spin_lock(&res->spinlock);
2243 res->state &= ~DLM_LOCK_RES_MIGRATING;
2244 spin_unlock(&res->spinlock);
2250 /* master is known, detach if not already detached */
2251 dlm_mle_detach_hb_events(dlm, oldmle);
2252 dlm_put_mle(oldmle);
2257 dlm_mle_detach_hb_events(dlm, mle);
2260 kmem_cache_free(dlm_mle_cache, mle);
2266 * at this point, we have a migration target, an mle
2267 * in the master list, and the MIGRATING flag set on
2272 /* get an extra reference on the mle.
2273 * otherwise the assert_master from the new
2274 * master will destroy this.
2275 * also, make sure that all callers of dlm_get_mle
2276 * take both dlm->spinlock and dlm->master_lock */
2277 spin_lock(&dlm->spinlock);
2278 spin_lock(&dlm->master_lock);
2279 dlm_get_mle_inuse(mle);
2280 spin_unlock(&dlm->master_lock);
2281 spin_unlock(&dlm->spinlock);
2283 /* notify new node and send all lock state */
2284 /* call send_one_lockres with migration flag.
2285 * this serves as notice to the target node that a
2286 * migration is starting. */
2287 ret = dlm_send_one_lockres(dlm, res, mres, target,
2288 DLM_MRES_MIGRATION);
2291 mlog(0, "migration to node %u failed with %d\n",
2293 /* migration failed, detach and clean up mle */
2294 dlm_mle_detach_hb_events(dlm, mle);
2296 dlm_put_mle_inuse(mle);
2297 spin_lock(&res->spinlock);
2298 res->state &= ~DLM_LOCK_RES_MIGRATING;
2299 spin_unlock(&res->spinlock);
2303 /* at this point, the target sends a message to all nodes,
2304 * (using dlm_do_migrate_request). this node is skipped since
2305 * we had to put an mle in the list to begin the process. this
2306 * node now waits for target to do an assert master. this node
2307 * will be the last one notified, ensuring that the migration
2308 * is complete everywhere. if the target dies while this is
2309 * going on, some nodes could potentially see the target as the
2310 * master, so it is important that my recovery finds the migration
2311 * mle and sets the master to UNKNONWN. */
2314 /* wait for new node to assert master */
2316 ret = wait_event_interruptible_timeout(mle->wq,
2317 (atomic_read(&mle->woken) == 1),
2318 msecs_to_jiffies(5000));
2321 if (atomic_read(&mle->woken) == 1 ||
2322 res->owner == target)
2325 mlog(0, "timed out during migration\n");
2326 /* avoid hang during shutdown when migrating lockres
2327 * to a node which also goes down */
2328 if (dlm_is_node_dead(dlm, target)) {
2329 mlog(0, "%s:%.*s: expected migration "
2330 "target %u is no longer up, restarting\n",
2331 dlm->name, res->lockname.len,
2332 res->lockname.name, target);
2336 if (ret == -ERESTARTSYS) {
2337 /* migration failed, detach and clean up mle */
2338 dlm_mle_detach_hb_events(dlm, mle);
2340 dlm_put_mle_inuse(mle);
2341 spin_lock(&res->spinlock);
2342 res->state &= ~DLM_LOCK_RES_MIGRATING;
2343 spin_unlock(&res->spinlock);
2346 /* TODO: if node died: stop, clean up, return error */
2349 /* all done, set the owner, clear the flag */
2350 spin_lock(&res->spinlock);
2351 dlm_set_lockres_owner(dlm, res, target);
2352 res->state &= ~DLM_LOCK_RES_MIGRATING;
2353 dlm_remove_nonlocal_locks(dlm, res);
2354 spin_unlock(&res->spinlock);
2357 /* master is known, detach if not already detached */
2358 dlm_mle_detach_hb_events(dlm, mle);
2359 dlm_put_mle_inuse(mle);
2362 dlm_lockres_calc_usage(dlm, res);
2365 /* re-dirty the lockres if we failed */
2367 dlm_kick_thread(dlm, res);
2371 free_page((unsigned long)mres);
2375 mlog(0, "returning %d\n", ret);
2378 EXPORT_SYMBOL_GPL(dlm_migrate_lockres);
2380 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2383 spin_lock(&dlm->ast_lock);
2384 spin_lock(&lock->spinlock);
2385 ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2386 spin_unlock(&lock->spinlock);
2387 spin_unlock(&dlm->ast_lock);
2391 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2392 struct dlm_lock_resource *res,
2396 spin_lock(&res->spinlock);
2397 can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2398 spin_unlock(&res->spinlock);
2400 /* target has died, so make the caller break out of the
2401 * wait_event, but caller must recheck the domain_map */
2402 spin_lock(&dlm->spinlock);
2403 if (!test_bit(mig_target, dlm->domain_map))
2405 spin_unlock(&dlm->spinlock);
2409 int dlm_lockres_is_dirty(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2412 spin_lock(&res->spinlock);
2413 ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2414 spin_unlock(&res->spinlock);
2419 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2420 struct dlm_lock_resource *res,
2425 mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2426 res->lockname.len, res->lockname.name, dlm->node_num,
2428 /* need to set MIGRATING flag on lockres. this is done by
2429 * ensuring that all asts have been flushed for this lockres. */
2430 spin_lock(&res->spinlock);
2431 BUG_ON(res->migration_pending);
2432 res->migration_pending = 1;
2433 /* strategy is to reserve an extra ast then release
2434 * it below, letting the release do all of the work */
2435 __dlm_lockres_reserve_ast(res);
2436 spin_unlock(&res->spinlock);
2438 /* now flush all the pending asts.. hang out for a bit */
2439 dlm_kick_thread(dlm, res);
2440 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2441 dlm_lockres_release_ast(dlm, res);
2443 mlog(0, "about to wait on migration_wq, dirty=%s\n",
2444 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2445 /* if the extra ref we just put was the final one, this
2446 * will pass thru immediately. otherwise, we need to wait
2447 * for the last ast to finish. */
2449 ret = wait_event_interruptible_timeout(dlm->migration_wq,
2450 dlm_migration_can_proceed(dlm, res, target),
2451 msecs_to_jiffies(1000));
2453 mlog(0, "woken again: migrating? %s, dead? %s\n",
2454 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2455 test_bit(target, dlm->domain_map) ? "no":"yes");
2457 mlog(0, "all is well: migrating? %s, dead? %s\n",
2458 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2459 test_bit(target, dlm->domain_map) ? "no":"yes");
2461 if (!dlm_migration_can_proceed(dlm, res, target)) {
2462 mlog(0, "trying again...\n");
2466 /* did the target go down or die? */
2467 spin_lock(&dlm->spinlock);
2468 if (!test_bit(target, dlm->domain_map)) {
2469 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2473 spin_unlock(&dlm->spinlock);
2478 * o the DLM_LOCK_RES_MIGRATING flag is set
2479 * o there are no pending asts on this lockres
2480 * o all processes trying to reserve an ast on this
2481 * lockres must wait for the MIGRATING flag to clear
2486 /* last step in the migration process.
2487 * original master calls this to free all of the dlm_lock
2488 * structures that used to be for other nodes. */
2489 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2490 struct dlm_lock_resource *res)
2492 struct list_head *iter, *iter2;
2493 struct list_head *queue = &res->granted;
2495 struct dlm_lock *lock;
2497 assert_spin_locked(&res->spinlock);
2499 BUG_ON(res->owner == dlm->node_num);
2501 for (i=0; i<3; i++) {
2502 list_for_each_safe(iter, iter2, queue) {
2503 lock = list_entry (iter, struct dlm_lock, list);
2504 if (lock->ml.node != dlm->node_num) {
2505 mlog(0, "putting lock for node %u\n",
2507 /* be extra careful */
2508 BUG_ON(!list_empty(&lock->ast_list));
2509 BUG_ON(!list_empty(&lock->bast_list));
2510 BUG_ON(lock->ast_pending);
2511 BUG_ON(lock->bast_pending);
2512 list_del_init(&lock->list);
2520 /* for now this is not too intelligent. we will
2521 * need stats to make this do the right thing.
2522 * this just finds the first lock on one of the
2523 * queues and uses that node as the target. */
2524 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2525 struct dlm_lock_resource *res)
2528 struct list_head *queue = &res->granted;
2529 struct list_head *iter;
2530 struct dlm_lock *lock;
2533 assert_spin_locked(&dlm->spinlock);
2535 spin_lock(&res->spinlock);
2536 for (i=0; i<3; i++) {
2537 list_for_each(iter, queue) {
2538 /* up to the caller to make sure this node
2540 lock = list_entry (iter, struct dlm_lock, list);
2541 if (lock->ml.node != dlm->node_num) {
2542 spin_unlock(&res->spinlock);
2543 return lock->ml.node;
2548 spin_unlock(&res->spinlock);
2549 mlog(0, "have not found a suitable target yet! checking domain map\n");
2551 /* ok now we're getting desperate. pick anyone alive. */
2554 nodenum = find_next_bit(dlm->domain_map,
2555 O2NM_MAX_NODES, nodenum+1);
2556 mlog(0, "found %d in domain map\n", nodenum);
2557 if (nodenum >= O2NM_MAX_NODES)
2559 if (nodenum != dlm->node_num) {
2560 mlog(0, "picking %d\n", nodenum);
2565 mlog(0, "giving up. no master to migrate to\n");
2566 return DLM_LOCK_RES_OWNER_UNKNOWN;
2571 /* this is called by the new master once all lockres
2572 * data has been received */
2573 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2574 struct dlm_lock_resource *res,
2575 u8 master, u8 new_master,
2576 struct dlm_node_iter *iter)
2578 struct dlm_migrate_request migrate;
2579 int ret, status = 0;
2582 memset(&migrate, 0, sizeof(migrate));
2583 migrate.namelen = res->lockname.len;
2584 memcpy(migrate.name, res->lockname.name, migrate.namelen);
2585 migrate.new_master = new_master;
2586 migrate.master = master;
2590 /* send message to all nodes, except the master and myself */
2591 while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2592 if (nodenum == master ||
2593 nodenum == new_master)
2596 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2597 &migrate, sizeof(migrate), nodenum,
2601 else if (status < 0) {
2602 mlog(0, "migrate request (node %u) returned %d!\n",
2611 mlog(0, "returning ret=%d\n", ret);
2616 /* if there is an existing mle for this lockres, we now know who the master is.
2617 * (the one who sent us *this* message) we can clear it up right away.
2618 * since the process that put the mle on the list still has a reference to it,
2619 * we can unhash it now, set the master and wake the process. as a result,
2620 * we will have no mle in the list to start with. now we can add an mle for
2621 * the migration and this should be the only one found for those scanning the
2623 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data)
2625 struct dlm_ctxt *dlm = data;
2626 struct dlm_lock_resource *res = NULL;
2627 struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
2628 struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
2630 unsigned int namelen, hash;
2636 name = migrate->name;
2637 namelen = migrate->namelen;
2638 hash = dlm_lockid_hash(name, namelen);
2640 /* preallocate.. if this fails, abort */
2641 mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
2649 /* check for pre-existing lock */
2650 spin_lock(&dlm->spinlock);
2651 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
2652 spin_lock(&dlm->master_lock);
2655 spin_lock(&res->spinlock);
2656 if (res->state & DLM_LOCK_RES_RECOVERING) {
2657 /* if all is working ok, this can only mean that we got
2658 * a migrate request from a node that we now see as
2659 * dead. what can we do here? drop it to the floor? */
2660 spin_unlock(&res->spinlock);
2661 mlog(ML_ERROR, "Got a migrate request, but the "
2662 "lockres is marked as recovering!");
2663 kmem_cache_free(dlm_mle_cache, mle);
2664 ret = -EINVAL; /* need a better solution */
2667 res->state |= DLM_LOCK_RES_MIGRATING;
2668 spin_unlock(&res->spinlock);
2671 /* ignore status. only nonzero status would BUG. */
2672 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
2674 migrate->new_master,
2678 spin_unlock(&dlm->master_lock);
2679 spin_unlock(&dlm->spinlock);
2682 /* master is known, detach if not already detached */
2683 dlm_mle_detach_hb_events(dlm, oldmle);
2684 dlm_put_mle(oldmle);
2688 dlm_lockres_put(res);
2694 /* must be holding dlm->spinlock and dlm->master_lock
2695 * when adding a migration mle, we can clear any other mles
2696 * in the master list because we know with certainty that
2697 * the master is "master". so we remove any old mle from
2698 * the list after setting it's master field, and then add
2699 * the new migration mle. this way we can hold with the rule
2700 * of having only one mle for a given lock name at all times. */
2701 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
2702 struct dlm_lock_resource *res,
2703 struct dlm_master_list_entry *mle,
2704 struct dlm_master_list_entry **oldmle,
2705 const char *name, unsigned int namelen,
2706 u8 new_master, u8 master)
2715 assert_spin_locked(&dlm->spinlock);
2716 assert_spin_locked(&dlm->master_lock);
2718 /* caller is responsible for any ref taken here on oldmle */
2719 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
2721 struct dlm_master_list_entry *tmp = *oldmle;
2722 spin_lock(&tmp->spinlock);
2723 if (tmp->type == DLM_MLE_MIGRATION) {
2724 if (master == dlm->node_num) {
2725 /* ah another process raced me to it */
2726 mlog(0, "tried to migrate %.*s, but some "
2727 "process beat me to it\n",
2731 /* bad. 2 NODES are trying to migrate! */
2732 mlog(ML_ERROR, "migration error mle: "
2733 "master=%u new_master=%u // request: "
2734 "master=%u new_master=%u // "
2736 tmp->master, tmp->new_master,
2742 /* this is essentially what assert_master does */
2743 tmp->master = master;
2744 atomic_set(&tmp->woken, 1);
2746 /* remove it from the list so that only one
2747 * mle will be found */
2748 list_del_init(&tmp->list);
2749 __dlm_mle_detach_hb_events(dlm, mle);
2751 spin_unlock(&tmp->spinlock);
2754 /* now add a migration mle to the tail of the list */
2755 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
2756 mle->new_master = new_master;
2757 mle->master = master;
2758 /* do this for consistency with other mle types */
2759 set_bit(new_master, mle->maybe_map);
2760 list_add(&mle->list, &dlm->master_list);
2766 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
2768 struct list_head *iter, *iter2;
2769 struct dlm_master_list_entry *mle;
2770 struct dlm_lock_resource *res;
2773 mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node);
2775 assert_spin_locked(&dlm->spinlock);
2777 /* clean the master list */
2778 spin_lock(&dlm->master_lock);
2779 list_for_each_safe(iter, iter2, &dlm->master_list) {
2780 mle = list_entry(iter, struct dlm_master_list_entry, list);
2782 BUG_ON(mle->type != DLM_MLE_BLOCK &&
2783 mle->type != DLM_MLE_MASTER &&
2784 mle->type != DLM_MLE_MIGRATION);
2786 /* MASTER mles are initiated locally. the waiting
2787 * process will notice the node map change
2788 * shortly. let that happen as normal. */
2789 if (mle->type == DLM_MLE_MASTER)
2793 /* BLOCK mles are initiated by other nodes.
2794 * need to clean up if the dead node would have
2795 * been the master. */
2796 if (mle->type == DLM_MLE_BLOCK) {
2799 spin_lock(&mle->spinlock);
2800 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
2801 if (bit != dead_node) {
2802 mlog(0, "mle found, but dead node %u would "
2803 "not have been master\n", dead_node);
2804 spin_unlock(&mle->spinlock);
2806 /* must drop the refcount by one since the
2807 * assert_master will never arrive. this
2808 * may result in the mle being unlinked and
2809 * freed, but there may still be a process
2810 * waiting in the dlmlock path which is fine. */
2811 mlog(ML_ERROR, "node %u was expected master\n",
2813 atomic_set(&mle->woken, 1);
2814 spin_unlock(&mle->spinlock);
2816 /* do not need events any longer, so detach
2818 __dlm_mle_detach_hb_events(dlm, mle);
2824 /* everything else is a MIGRATION mle */
2826 /* the rule for MIGRATION mles is that the master
2827 * becomes UNKNOWN if *either* the original or
2828 * the new master dies. all UNKNOWN lockreses
2829 * are sent to whichever node becomes the recovery
2830 * master. the new master is responsible for
2831 * determining if there is still a master for
2832 * this lockres, or if he needs to take over
2833 * mastery. either way, this node should expect
2834 * another message to resolve this. */
2835 if (mle->master != dead_node &&
2836 mle->new_master != dead_node)
2839 /* if we have reached this point, this mle needs to
2840 * be removed from the list and freed. */
2842 /* remove from the list early. NOTE: unlinking
2843 * list_head while in list_for_each_safe */
2844 __dlm_mle_detach_hb_events(dlm, mle);
2845 spin_lock(&mle->spinlock);
2846 list_del_init(&mle->list);
2847 atomic_set(&mle->woken, 1);
2848 spin_unlock(&mle->spinlock);
2851 mlog(0, "%s: node %u died during migration from "
2852 "%u to %u!\n", dlm->name, dead_node,
2853 mle->master, mle->new_master);
2854 /* if there is a lockres associated with this
2855 * mle, find it and set its owner to UNKNOWN */
2856 hash = dlm_lockid_hash(mle->u.name.name, mle->u.name.len);
2857 res = __dlm_lookup_lockres(dlm, mle->u.name.name,
2858 mle->u.name.len, hash);
2860 /* unfortunately if we hit this rare case, our
2861 * lock ordering is messed. we need to drop
2862 * the master lock so that we can take the
2863 * lockres lock, meaning that we will have to
2864 * restart from the head of list. */
2865 spin_unlock(&dlm->master_lock);
2867 /* move lockres onto recovery list */
2868 spin_lock(&res->spinlock);
2869 dlm_set_lockres_owner(dlm, res,
2870 DLM_LOCK_RES_OWNER_UNKNOWN);
2871 dlm_move_lockres_to_recovery_list(dlm, res);
2872 spin_unlock(&res->spinlock);
2873 dlm_lockres_put(res);
2875 /* about to get rid of mle, detach from heartbeat */
2876 __dlm_mle_detach_hb_events(dlm, mle);
2879 spin_lock(&dlm->master_lock);
2881 spin_unlock(&dlm->master_lock);
2887 /* this may be the last reference */
2890 spin_unlock(&dlm->master_lock);
2894 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
2897 struct dlm_node_iter iter;
2900 spin_lock(&dlm->spinlock);
2901 dlm_node_iter_init(dlm->domain_map, &iter);
2902 clear_bit(old_master, iter.node_map);
2903 clear_bit(dlm->node_num, iter.node_map);
2904 spin_unlock(&dlm->spinlock);
2906 mlog(0, "now time to do a migrate request to other nodes\n");
2907 ret = dlm_do_migrate_request(dlm, res, old_master,
2908 dlm->node_num, &iter);
2914 mlog(0, "doing assert master of %.*s to all except the original node\n",
2915 res->lockname.len, res->lockname.name);
2916 /* this call now finishes out the nodemap
2917 * even if one or more nodes die */
2918 ret = dlm_do_assert_master(dlm, res->lockname.name,
2919 res->lockname.len, iter.node_map,
2920 DLM_ASSERT_MASTER_FINISH_MIGRATION);
2922 /* no longer need to retry. all living nodes contacted. */
2927 memset(iter.node_map, 0, sizeof(iter.node_map));
2928 set_bit(old_master, iter.node_map);
2929 mlog(0, "doing assert master of %.*s back to %u\n",
2930 res->lockname.len, res->lockname.name, old_master);
2931 ret = dlm_do_assert_master(dlm, res->lockname.name,
2932 res->lockname.len, iter.node_map,
2933 DLM_ASSERT_MASTER_FINISH_MIGRATION);
2935 mlog(0, "assert master to original master failed "
2937 /* the only nonzero status here would be because of
2938 * a dead original node. we're done. */
2942 /* all done, set the owner, clear the flag */
2943 spin_lock(&res->spinlock);
2944 dlm_set_lockres_owner(dlm, res, dlm->node_num);
2945 res->state &= ~DLM_LOCK_RES_MIGRATING;
2946 spin_unlock(&res->spinlock);
2947 /* re-dirty it on the new master */
2948 dlm_kick_thread(dlm, res);
2955 * LOCKRES AST REFCOUNT
2956 * this is integral to migration
2959 /* for future intent to call an ast, reserve one ahead of time.
2960 * this should be called only after waiting on the lockres
2961 * with dlm_wait_on_lockres, and while still holding the
2962 * spinlock after the call. */
2963 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
2965 assert_spin_locked(&res->spinlock);
2966 if (res->state & DLM_LOCK_RES_MIGRATING) {
2967 __dlm_print_one_lock_resource(res);
2969 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
2971 atomic_inc(&res->asts_reserved);
2975 * used to drop the reserved ast, either because it went unused,
2976 * or because the ast/bast was actually called.
2978 * also, if there is a pending migration on this lockres,
2979 * and this was the last pending ast on the lockres,
2980 * atomically set the MIGRATING flag before we drop the lock.
2981 * this is how we ensure that migration can proceed with no
2982 * asts in progress. note that it is ok if the state of the
2983 * queues is such that a lock should be granted in the future
2984 * or that a bast should be fired, because the new master will
2985 * shuffle the lists on this lockres as soon as it is migrated.
2987 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
2988 struct dlm_lock_resource *res)
2990 if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
2993 if (!res->migration_pending) {
2994 spin_unlock(&res->spinlock);
2998 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
2999 res->migration_pending = 0;
3000 res->state |= DLM_LOCK_RES_MIGRATING;
3001 spin_unlock(&res->spinlock);
3003 wake_up(&dlm->migration_wq);