1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * standalone DLM module
8 * Copyright (C) 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
28 #include <linux/module.h>
30 #include <linux/types.h>
31 #include <linux/slab.h>
32 #include <linux/highmem.h>
33 #include <linux/init.h>
34 #include <linux/sysctl.h>
35 #include <linux/random.h>
36 #include <linux/blkdev.h>
37 #include <linux/socket.h>
38 #include <linux/inet.h>
39 #include <linux/spinlock.h>
40 #include <linux/delay.h>
43 #include "cluster/heartbeat.h"
44 #include "cluster/nodemanager.h"
45 #include "cluster/tcp.h"
48 #include "dlmcommon.h"
49 #include "dlmdomain.h"
52 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
53 #include "cluster/masklog.h"
55 static void dlm_mle_node_down(struct dlm_ctxt
*dlm
,
56 struct dlm_master_list_entry
*mle
,
57 struct o2nm_node
*node
,
59 static void dlm_mle_node_up(struct dlm_ctxt
*dlm
,
60 struct dlm_master_list_entry
*mle
,
61 struct o2nm_node
*node
,
64 static void dlm_assert_master_worker(struct dlm_work_item
*item
, void *data
);
65 static int dlm_do_assert_master(struct dlm_ctxt
*dlm
,
66 struct dlm_lock_resource
*res
,
67 void *nodemap
, u32 flags
);
68 static void dlm_deref_lockres_worker(struct dlm_work_item
*item
, void *data
);
70 static inline int dlm_mle_equal(struct dlm_ctxt
*dlm
,
71 struct dlm_master_list_entry
*mle
,
78 if (namelen
!= mle
->mnamelen
||
79 memcmp(name
, mle
->mname
, namelen
) != 0)
85 static struct kmem_cache
*dlm_lockres_cache
;
86 static struct kmem_cache
*dlm_lockname_cache
;
87 static struct kmem_cache
*dlm_mle_cache
;
89 static void dlm_mle_release(struct kref
*kref
);
90 static void dlm_init_mle(struct dlm_master_list_entry
*mle
,
91 enum dlm_mle_type type
,
93 struct dlm_lock_resource
*res
,
95 unsigned int namelen
);
96 static void dlm_put_mle(struct dlm_master_list_entry
*mle
);
97 static void __dlm_put_mle(struct dlm_master_list_entry
*mle
);
98 static int dlm_find_mle(struct dlm_ctxt
*dlm
,
99 struct dlm_master_list_entry
**mle
,
100 char *name
, unsigned int namelen
);
102 static int dlm_do_master_request(struct dlm_lock_resource
*res
,
103 struct dlm_master_list_entry
*mle
, int to
);
106 static int dlm_wait_for_lock_mastery(struct dlm_ctxt
*dlm
,
107 struct dlm_lock_resource
*res
,
108 struct dlm_master_list_entry
*mle
,
110 static int dlm_restart_lock_mastery(struct dlm_ctxt
*dlm
,
111 struct dlm_lock_resource
*res
,
112 struct dlm_master_list_entry
*mle
,
114 static int dlm_add_migration_mle(struct dlm_ctxt
*dlm
,
115 struct dlm_lock_resource
*res
,
116 struct dlm_master_list_entry
*mle
,
117 struct dlm_master_list_entry
**oldmle
,
118 const char *name
, unsigned int namelen
,
119 u8 new_master
, u8 master
);
121 static u8
dlm_pick_migration_target(struct dlm_ctxt
*dlm
,
122 struct dlm_lock_resource
*res
);
123 static void dlm_remove_nonlocal_locks(struct dlm_ctxt
*dlm
,
124 struct dlm_lock_resource
*res
);
125 static int dlm_mark_lockres_migrating(struct dlm_ctxt
*dlm
,
126 struct dlm_lock_resource
*res
,
128 static int dlm_pre_master_reco_lockres(struct dlm_ctxt
*dlm
,
129 struct dlm_lock_resource
*res
);
132 int dlm_is_host_down(int errno
)
149 case -EINVAL
: /* if returned from our tcp code,
150 this means there is no socket */
158 * MASTER LIST FUNCTIONS
163 * regarding master list entries and heartbeat callbacks:
165 * in order to avoid sleeping and allocation that occurs in
166 * heartbeat, master list entries are simply attached to the
167 * dlm's established heartbeat callbacks. the mle is attached
168 * when it is created, and since the dlm->spinlock is held at
169 * that time, any heartbeat event will be properly discovered
170 * by the mle. the mle needs to be detached from the
171 * dlm->mle_hb_events list as soon as heartbeat events are no
172 * longer useful to the mle, and before the mle is freed.
174 * as a general rule, heartbeat events are no longer needed by
175 * the mle once an "answer" regarding the lock master has been
178 static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt
*dlm
,
179 struct dlm_master_list_entry
*mle
)
181 assert_spin_locked(&dlm
->spinlock
);
183 list_add_tail(&mle
->hb_events
, &dlm
->mle_hb_events
);
187 static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt
*dlm
,
188 struct dlm_master_list_entry
*mle
)
190 if (!list_empty(&mle
->hb_events
))
191 list_del_init(&mle
->hb_events
);
195 static inline void dlm_mle_detach_hb_events(struct dlm_ctxt
*dlm
,
196 struct dlm_master_list_entry
*mle
)
198 spin_lock(&dlm
->spinlock
);
199 __dlm_mle_detach_hb_events(dlm
, mle
);
200 spin_unlock(&dlm
->spinlock
);
203 static void dlm_get_mle_inuse(struct dlm_master_list_entry
*mle
)
205 struct dlm_ctxt
*dlm
;
208 assert_spin_locked(&dlm
->spinlock
);
209 assert_spin_locked(&dlm
->master_lock
);
211 kref_get(&mle
->mle_refs
);
214 static void dlm_put_mle_inuse(struct dlm_master_list_entry
*mle
)
216 struct dlm_ctxt
*dlm
;
219 spin_lock(&dlm
->spinlock
);
220 spin_lock(&dlm
->master_lock
);
223 spin_unlock(&dlm
->master_lock
);
224 spin_unlock(&dlm
->spinlock
);
228 /* remove from list and free */
229 static void __dlm_put_mle(struct dlm_master_list_entry
*mle
)
231 struct dlm_ctxt
*dlm
;
234 assert_spin_locked(&dlm
->spinlock
);
235 assert_spin_locked(&dlm
->master_lock
);
236 if (!kref_read(&mle
->mle_refs
)) {
237 /* this may or may not crash, but who cares.
239 mlog(ML_ERROR
, "bad mle: %p\n", mle
);
240 dlm_print_one_mle(mle
);
243 kref_put(&mle
->mle_refs
, dlm_mle_release
);
247 /* must not have any spinlocks coming in */
248 static void dlm_put_mle(struct dlm_master_list_entry
*mle
)
250 struct dlm_ctxt
*dlm
;
253 spin_lock(&dlm
->spinlock
);
254 spin_lock(&dlm
->master_lock
);
256 spin_unlock(&dlm
->master_lock
);
257 spin_unlock(&dlm
->spinlock
);
260 static inline void dlm_get_mle(struct dlm_master_list_entry
*mle
)
262 kref_get(&mle
->mle_refs
);
265 static void dlm_init_mle(struct dlm_master_list_entry
*mle
,
266 enum dlm_mle_type type
,
267 struct dlm_ctxt
*dlm
,
268 struct dlm_lock_resource
*res
,
270 unsigned int namelen
)
272 assert_spin_locked(&dlm
->spinlock
);
276 INIT_HLIST_NODE(&mle
->master_hash_node
);
277 INIT_LIST_HEAD(&mle
->hb_events
);
278 memset(mle
->maybe_map
, 0, sizeof(mle
->maybe_map
));
279 spin_lock_init(&mle
->spinlock
);
280 init_waitqueue_head(&mle
->wq
);
281 atomic_set(&mle
->woken
, 0);
282 kref_init(&mle
->mle_refs
);
283 memset(mle
->response_map
, 0, sizeof(mle
->response_map
));
284 mle
->master
= O2NM_MAX_NODES
;
285 mle
->new_master
= O2NM_MAX_NODES
;
288 BUG_ON(mle
->type
!= DLM_MLE_BLOCK
&&
289 mle
->type
!= DLM_MLE_MASTER
&&
290 mle
->type
!= DLM_MLE_MIGRATION
);
292 if (mle
->type
== DLM_MLE_MASTER
) {
295 memcpy(mle
->mname
, res
->lockname
.name
, res
->lockname
.len
);
296 mle
->mnamelen
= res
->lockname
.len
;
297 mle
->mnamehash
= res
->lockname
.hash
;
301 memcpy(mle
->mname
, name
, namelen
);
302 mle
->mnamelen
= namelen
;
303 mle
->mnamehash
= dlm_lockid_hash(name
, namelen
);
306 atomic_inc(&dlm
->mle_tot_count
[mle
->type
]);
307 atomic_inc(&dlm
->mle_cur_count
[mle
->type
]);
309 /* copy off the node_map and register hb callbacks on our copy */
310 memcpy(mle
->node_map
, dlm
->domain_map
, sizeof(mle
->node_map
));
311 memcpy(mle
->vote_map
, dlm
->domain_map
, sizeof(mle
->vote_map
));
312 clear_bit(dlm
->node_num
, mle
->vote_map
);
313 clear_bit(dlm
->node_num
, mle
->node_map
);
315 /* attach the mle to the domain node up/down events */
316 __dlm_mle_attach_hb_events(dlm
, mle
);
319 void __dlm_unlink_mle(struct dlm_ctxt
*dlm
, struct dlm_master_list_entry
*mle
)
321 assert_spin_locked(&dlm
->spinlock
);
322 assert_spin_locked(&dlm
->master_lock
);
324 if (!hlist_unhashed(&mle
->master_hash_node
))
325 hlist_del_init(&mle
->master_hash_node
);
328 void __dlm_insert_mle(struct dlm_ctxt
*dlm
, struct dlm_master_list_entry
*mle
)
330 struct hlist_head
*bucket
;
332 assert_spin_locked(&dlm
->master_lock
);
334 bucket
= dlm_master_hash(dlm
, mle
->mnamehash
);
335 hlist_add_head(&mle
->master_hash_node
, bucket
);
338 /* returns 1 if found, 0 if not */
339 static int dlm_find_mle(struct dlm_ctxt
*dlm
,
340 struct dlm_master_list_entry
**mle
,
341 char *name
, unsigned int namelen
)
343 struct dlm_master_list_entry
*tmpmle
;
344 struct hlist_head
*bucket
;
347 assert_spin_locked(&dlm
->master_lock
);
349 hash
= dlm_lockid_hash(name
, namelen
);
350 bucket
= dlm_master_hash(dlm
, hash
);
351 hlist_for_each_entry(tmpmle
, bucket
, master_hash_node
) {
352 if (!dlm_mle_equal(dlm
, tmpmle
, name
, namelen
))
361 void dlm_hb_event_notify_attached(struct dlm_ctxt
*dlm
, int idx
, int node_up
)
363 struct dlm_master_list_entry
*mle
;
365 assert_spin_locked(&dlm
->spinlock
);
367 list_for_each_entry(mle
, &dlm
->mle_hb_events
, hb_events
) {
369 dlm_mle_node_up(dlm
, mle
, NULL
, idx
);
371 dlm_mle_node_down(dlm
, mle
, NULL
, idx
);
375 static void dlm_mle_node_down(struct dlm_ctxt
*dlm
,
376 struct dlm_master_list_entry
*mle
,
377 struct o2nm_node
*node
, int idx
)
379 spin_lock(&mle
->spinlock
);
381 if (!test_bit(idx
, mle
->node_map
))
382 mlog(0, "node %u already removed from nodemap!\n", idx
);
384 clear_bit(idx
, mle
->node_map
);
386 spin_unlock(&mle
->spinlock
);
389 static void dlm_mle_node_up(struct dlm_ctxt
*dlm
,
390 struct dlm_master_list_entry
*mle
,
391 struct o2nm_node
*node
, int idx
)
393 spin_lock(&mle
->spinlock
);
395 if (test_bit(idx
, mle
->node_map
))
396 mlog(0, "node %u already in node map!\n", idx
);
398 set_bit(idx
, mle
->node_map
);
400 spin_unlock(&mle
->spinlock
);
404 int dlm_init_mle_cache(void)
406 dlm_mle_cache
= kmem_cache_create("o2dlm_mle",
407 sizeof(struct dlm_master_list_entry
),
408 0, SLAB_HWCACHE_ALIGN
,
410 if (dlm_mle_cache
== NULL
)
415 void dlm_destroy_mle_cache(void)
418 kmem_cache_destroy(dlm_mle_cache
);
421 static void dlm_mle_release(struct kref
*kref
)
423 struct dlm_master_list_entry
*mle
;
424 struct dlm_ctxt
*dlm
;
426 mle
= container_of(kref
, struct dlm_master_list_entry
, mle_refs
);
429 assert_spin_locked(&dlm
->spinlock
);
430 assert_spin_locked(&dlm
->master_lock
);
432 mlog(0, "Releasing mle for %.*s, type %d\n", mle
->mnamelen
, mle
->mname
,
435 /* remove from list if not already */
436 __dlm_unlink_mle(dlm
, mle
);
438 /* detach the mle from the domain node up/down events */
439 __dlm_mle_detach_hb_events(dlm
, mle
);
441 atomic_dec(&dlm
->mle_cur_count
[mle
->type
]);
443 /* NOTE: kfree under spinlock here.
444 * if this is bad, we can move this to a freelist. */
445 kmem_cache_free(dlm_mle_cache
, mle
);
450 * LOCK RESOURCE FUNCTIONS
453 int dlm_init_master_caches(void)
455 dlm_lockres_cache
= kmem_cache_create("o2dlm_lockres",
456 sizeof(struct dlm_lock_resource
),
457 0, SLAB_HWCACHE_ALIGN
, NULL
);
458 if (!dlm_lockres_cache
)
461 dlm_lockname_cache
= kmem_cache_create("o2dlm_lockname",
462 DLM_LOCKID_NAME_MAX
, 0,
463 SLAB_HWCACHE_ALIGN
, NULL
);
464 if (!dlm_lockname_cache
)
469 dlm_destroy_master_caches();
473 void dlm_destroy_master_caches(void)
475 if (dlm_lockname_cache
) {
476 kmem_cache_destroy(dlm_lockname_cache
);
477 dlm_lockname_cache
= NULL
;
480 if (dlm_lockres_cache
) {
481 kmem_cache_destroy(dlm_lockres_cache
);
482 dlm_lockres_cache
= NULL
;
486 static void dlm_lockres_release(struct kref
*kref
)
488 struct dlm_lock_resource
*res
;
489 struct dlm_ctxt
*dlm
;
491 res
= container_of(kref
, struct dlm_lock_resource
, refs
);
494 /* This should not happen -- all lockres' have a name
495 * associated with them at init time. */
496 BUG_ON(!res
->lockname
.name
);
498 mlog(0, "destroying lockres %.*s\n", res
->lockname
.len
,
501 atomic_dec(&dlm
->res_cur_count
);
503 if (!hlist_unhashed(&res
->hash_node
) ||
504 !list_empty(&res
->granted
) ||
505 !list_empty(&res
->converting
) ||
506 !list_empty(&res
->blocked
) ||
507 !list_empty(&res
->dirty
) ||
508 !list_empty(&res
->recovering
) ||
509 !list_empty(&res
->purge
)) {
511 "Going to BUG for resource %.*s."
512 " We're on a list! [%c%c%c%c%c%c%c]\n",
513 res
->lockname
.len
, res
->lockname
.name
,
514 !hlist_unhashed(&res
->hash_node
) ? 'H' : ' ',
515 !list_empty(&res
->granted
) ? 'G' : ' ',
516 !list_empty(&res
->converting
) ? 'C' : ' ',
517 !list_empty(&res
->blocked
) ? 'B' : ' ',
518 !list_empty(&res
->dirty
) ? 'D' : ' ',
519 !list_empty(&res
->recovering
) ? 'R' : ' ',
520 !list_empty(&res
->purge
) ? 'P' : ' ');
522 dlm_print_one_lock_resource(res
);
525 /* By the time we're ready to blow this guy away, we shouldn't
526 * be on any lists. */
527 BUG_ON(!hlist_unhashed(&res
->hash_node
));
528 BUG_ON(!list_empty(&res
->granted
));
529 BUG_ON(!list_empty(&res
->converting
));
530 BUG_ON(!list_empty(&res
->blocked
));
531 BUG_ON(!list_empty(&res
->dirty
));
532 BUG_ON(!list_empty(&res
->recovering
));
533 BUG_ON(!list_empty(&res
->purge
));
535 kmem_cache_free(dlm_lockname_cache
, (void *)res
->lockname
.name
);
537 kmem_cache_free(dlm_lockres_cache
, res
);
540 void dlm_lockres_put(struct dlm_lock_resource
*res
)
542 kref_put(&res
->refs
, dlm_lockres_release
);
545 static void dlm_init_lockres(struct dlm_ctxt
*dlm
,
546 struct dlm_lock_resource
*res
,
547 const char *name
, unsigned int namelen
)
551 /* If we memset here, we lose our reference to the kmalloc'd
552 * res->lockname.name, so be sure to init every field
555 qname
= (char *) res
->lockname
.name
;
556 memcpy(qname
, name
, namelen
);
558 res
->lockname
.len
= namelen
;
559 res
->lockname
.hash
= dlm_lockid_hash(name
, namelen
);
561 init_waitqueue_head(&res
->wq
);
562 spin_lock_init(&res
->spinlock
);
563 INIT_HLIST_NODE(&res
->hash_node
);
564 INIT_LIST_HEAD(&res
->granted
);
565 INIT_LIST_HEAD(&res
->converting
);
566 INIT_LIST_HEAD(&res
->blocked
);
567 INIT_LIST_HEAD(&res
->dirty
);
568 INIT_LIST_HEAD(&res
->recovering
);
569 INIT_LIST_HEAD(&res
->purge
);
570 INIT_LIST_HEAD(&res
->tracking
);
571 atomic_set(&res
->asts_reserved
, 0);
572 res
->migration_pending
= 0;
573 res
->inflight_locks
= 0;
574 res
->inflight_assert_workers
= 0;
578 kref_init(&res
->refs
);
580 atomic_inc(&dlm
->res_tot_count
);
581 atomic_inc(&dlm
->res_cur_count
);
583 /* just for consistency */
584 spin_lock(&res
->spinlock
);
585 dlm_set_lockres_owner(dlm
, res
, DLM_LOCK_RES_OWNER_UNKNOWN
);
586 spin_unlock(&res
->spinlock
);
588 res
->state
= DLM_LOCK_RES_IN_PROGRESS
;
592 spin_lock(&dlm
->spinlock
);
593 list_add_tail(&res
->tracking
, &dlm
->tracking_list
);
594 spin_unlock(&dlm
->spinlock
);
596 memset(res
->lvb
, 0, DLM_LVB_LEN
);
597 memset(res
->refmap
, 0, sizeof(res
->refmap
));
600 struct dlm_lock_resource
*dlm_new_lockres(struct dlm_ctxt
*dlm
,
602 unsigned int namelen
)
604 struct dlm_lock_resource
*res
= NULL
;
606 res
= kmem_cache_zalloc(dlm_lockres_cache
, GFP_NOFS
);
610 res
->lockname
.name
= kmem_cache_zalloc(dlm_lockname_cache
, GFP_NOFS
);
611 if (!res
->lockname
.name
)
614 dlm_init_lockres(dlm
, res
, name
, namelen
);
619 kmem_cache_free(dlm_lockres_cache
, res
);
623 void dlm_lockres_set_refmap_bit(struct dlm_ctxt
*dlm
,
624 struct dlm_lock_resource
*res
, int bit
)
626 assert_spin_locked(&res
->spinlock
);
628 mlog(0, "res %.*s, set node %u, %ps()\n", res
->lockname
.len
,
629 res
->lockname
.name
, bit
, __builtin_return_address(0));
631 set_bit(bit
, res
->refmap
);
634 void dlm_lockres_clear_refmap_bit(struct dlm_ctxt
*dlm
,
635 struct dlm_lock_resource
*res
, int bit
)
637 assert_spin_locked(&res
->spinlock
);
639 mlog(0, "res %.*s, clr node %u, %ps()\n", res
->lockname
.len
,
640 res
->lockname
.name
, bit
, __builtin_return_address(0));
642 clear_bit(bit
, res
->refmap
);
645 static void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt
*dlm
,
646 struct dlm_lock_resource
*res
)
648 res
->inflight_locks
++;
650 mlog(0, "%s: res %.*s, inflight++: now %u, %ps()\n", dlm
->name
,
651 res
->lockname
.len
, res
->lockname
.name
, res
->inflight_locks
,
652 __builtin_return_address(0));
655 void dlm_lockres_grab_inflight_ref(struct dlm_ctxt
*dlm
,
656 struct dlm_lock_resource
*res
)
658 assert_spin_locked(&res
->spinlock
);
659 __dlm_lockres_grab_inflight_ref(dlm
, res
);
662 void dlm_lockres_drop_inflight_ref(struct dlm_ctxt
*dlm
,
663 struct dlm_lock_resource
*res
)
665 assert_spin_locked(&res
->spinlock
);
667 BUG_ON(res
->inflight_locks
== 0);
669 res
->inflight_locks
--;
671 mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm
->name
,
672 res
->lockname
.len
, res
->lockname
.name
, res
->inflight_locks
,
673 __builtin_return_address(0));
678 void __dlm_lockres_grab_inflight_worker(struct dlm_ctxt
*dlm
,
679 struct dlm_lock_resource
*res
)
681 assert_spin_locked(&res
->spinlock
);
682 res
->inflight_assert_workers
++;
683 mlog(0, "%s:%.*s: inflight assert worker++: now %u\n",
684 dlm
->name
, res
->lockname
.len
, res
->lockname
.name
,
685 res
->inflight_assert_workers
);
688 static void __dlm_lockres_drop_inflight_worker(struct dlm_ctxt
*dlm
,
689 struct dlm_lock_resource
*res
)
691 assert_spin_locked(&res
->spinlock
);
692 BUG_ON(res
->inflight_assert_workers
== 0);
693 res
->inflight_assert_workers
--;
694 mlog(0, "%s:%.*s: inflight assert worker--: now %u\n",
695 dlm
->name
, res
->lockname
.len
, res
->lockname
.name
,
696 res
->inflight_assert_workers
);
699 static void dlm_lockres_drop_inflight_worker(struct dlm_ctxt
*dlm
,
700 struct dlm_lock_resource
*res
)
702 spin_lock(&res
->spinlock
);
703 __dlm_lockres_drop_inflight_worker(dlm
, res
);
704 spin_unlock(&res
->spinlock
);
708 * lookup a lock resource by name.
709 * may already exist in the hashtable.
710 * lockid is null terminated
712 * if not, allocate enough for the lockres and for
713 * the temporary structure used in doing the mastering.
715 * also, do a lookup in the dlm->master_list to see
716 * if another node has begun mastering the same lock.
717 * if so, there should be a block entry in there
718 * for this name, and we should *not* attempt to master
719 * the lock here. need to wait around for that node
720 * to assert_master (or die).
723 struct dlm_lock_resource
* dlm_get_lock_resource(struct dlm_ctxt
*dlm
,
728 struct dlm_lock_resource
*tmpres
=NULL
, *res
=NULL
;
729 struct dlm_master_list_entry
*mle
= NULL
;
730 struct dlm_master_list_entry
*alloc_mle
= NULL
;
733 struct dlm_node_iter iter
;
736 int bit
, wait_on_recovery
= 0;
740 hash
= dlm_lockid_hash(lockid
, namelen
);
742 mlog(0, "get lockres %s (len %d)\n", lockid
, namelen
);
745 spin_lock(&dlm
->spinlock
);
746 tmpres
= __dlm_lookup_lockres_full(dlm
, lockid
, namelen
, hash
);
748 spin_unlock(&dlm
->spinlock
);
749 spin_lock(&tmpres
->spinlock
);
752 * Right after dlm spinlock was released, dlm_thread could have
753 * purged the lockres. Check if lockres got unhashed. If so
756 if (hlist_unhashed(&tmpres
->hash_node
)) {
757 spin_unlock(&tmpres
->spinlock
);
758 dlm_lockres_put(tmpres
);
763 /* Wait on the thread that is mastering the resource */
764 if (tmpres
->owner
== DLM_LOCK_RES_OWNER_UNKNOWN
) {
765 __dlm_wait_on_lockres(tmpres
);
766 BUG_ON(tmpres
->owner
== DLM_LOCK_RES_OWNER_UNKNOWN
);
767 spin_unlock(&tmpres
->spinlock
);
768 dlm_lockres_put(tmpres
);
773 /* Wait on the resource purge to complete before continuing */
774 if (tmpres
->state
& DLM_LOCK_RES_DROPPING_REF
) {
775 BUG_ON(tmpres
->owner
== dlm
->node_num
);
776 __dlm_wait_on_lockres_flags(tmpres
,
777 DLM_LOCK_RES_DROPPING_REF
);
778 spin_unlock(&tmpres
->spinlock
);
779 dlm_lockres_put(tmpres
);
784 /* Grab inflight ref to pin the resource */
785 dlm_lockres_grab_inflight_ref(dlm
, tmpres
);
787 spin_unlock(&tmpres
->spinlock
);
789 spin_lock(&dlm
->track_lock
);
790 if (!list_empty(&res
->tracking
))
791 list_del_init(&res
->tracking
);
793 mlog(ML_ERROR
, "Resource %.*s not "
794 "on the Tracking list\n",
797 spin_unlock(&dlm
->track_lock
);
798 dlm_lockres_put(res
);
805 spin_unlock(&dlm
->spinlock
);
806 mlog(0, "allocating a new resource\n");
807 /* nothing found and we need to allocate one. */
808 alloc_mle
= kmem_cache_alloc(dlm_mle_cache
, GFP_NOFS
);
811 res
= dlm_new_lockres(dlm
, lockid
, namelen
);
817 mlog(0, "no lockres found, allocated our own: %p\n", res
);
819 if (flags
& LKM_LOCAL
) {
820 /* caller knows it's safe to assume it's not mastered elsewhere
821 * DONE! return right away */
822 spin_lock(&res
->spinlock
);
823 dlm_change_lockres_owner(dlm
, res
, dlm
->node_num
);
824 __dlm_insert_lockres(dlm
, res
);
825 dlm_lockres_grab_inflight_ref(dlm
, res
);
826 spin_unlock(&res
->spinlock
);
827 spin_unlock(&dlm
->spinlock
);
828 /* lockres still marked IN_PROGRESS */
832 /* check master list to see if another node has started mastering it */
833 spin_lock(&dlm
->master_lock
);
835 /* if we found a block, wait for lock to be mastered by another node */
836 blocked
= dlm_find_mle(dlm
, &mle
, (char *)lockid
, namelen
);
839 if (mle
->type
== DLM_MLE_MASTER
) {
840 mlog(ML_ERROR
, "master entry for nonexistent lock!\n");
843 mig
= (mle
->type
== DLM_MLE_MIGRATION
);
844 /* if there is a migration in progress, let the migration
845 * finish before continuing. we can wait for the absence
846 * of the MIGRATION mle: either the migrate finished or
847 * one of the nodes died and the mle was cleaned up.
848 * if there is a BLOCK here, but it already has a master
849 * set, we are too late. the master does not have a ref
850 * for us in the refmap. detach the mle and drop it.
851 * either way, go back to the top and start over. */
852 if (mig
|| mle
->master
!= O2NM_MAX_NODES
) {
853 BUG_ON(mig
&& mle
->master
== dlm
->node_num
);
854 /* we arrived too late. the master does not
855 * have a ref for us. retry. */
856 mlog(0, "%s:%.*s: late on %s\n",
857 dlm
->name
, namelen
, lockid
,
858 mig
? "MIGRATION" : "BLOCK");
859 spin_unlock(&dlm
->master_lock
);
860 spin_unlock(&dlm
->spinlock
);
862 /* master is known, detach */
864 dlm_mle_detach_hb_events(dlm
, mle
);
867 /* this is lame, but we can't wait on either
868 * the mle or lockres waitqueue here */
874 /* go ahead and try to master lock on this node */
876 /* make sure this does not get freed below */
878 dlm_init_mle(mle
, DLM_MLE_MASTER
, dlm
, res
, NULL
, 0);
879 set_bit(dlm
->node_num
, mle
->maybe_map
);
880 __dlm_insert_mle(dlm
, mle
);
882 /* still holding the dlm spinlock, check the recovery map
883 * to see if there are any nodes that still need to be
884 * considered. these will not appear in the mle nodemap
885 * but they might own this lockres. wait on them. */
886 bit
= find_next_bit(dlm
->recovery_map
, O2NM_MAX_NODES
, 0);
887 if (bit
< O2NM_MAX_NODES
) {
888 mlog(0, "%s: res %.*s, At least one node (%d) "
889 "to recover before lock mastery can begin\n",
890 dlm
->name
, namelen
, (char *)lockid
, bit
);
891 wait_on_recovery
= 1;
895 /* at this point there is either a DLM_MLE_BLOCK or a
896 * DLM_MLE_MASTER on the master list, so it's safe to add the
897 * lockres to the hashtable. anyone who finds the lock will
898 * still have to wait on the IN_PROGRESS. */
900 /* finally add the lockres to its hash bucket */
901 __dlm_insert_lockres(dlm
, res
);
903 /* since this lockres is new it doesn't not require the spinlock */
904 __dlm_lockres_grab_inflight_ref(dlm
, res
);
906 /* get an extra ref on the mle in case this is a BLOCK
907 * if so, the creator of the BLOCK may try to put the last
908 * ref at this time in the assert master handler, so we
909 * need an extra one to keep from a bad ptr deref. */
910 dlm_get_mle_inuse(mle
);
911 spin_unlock(&dlm
->master_lock
);
912 spin_unlock(&dlm
->spinlock
);
915 while (wait_on_recovery
) {
916 /* any cluster changes that occurred after dropping the
917 * dlm spinlock would be detectable be a change on the mle,
918 * so we only need to clear out the recovery map once. */
919 if (dlm_is_recovery_lock(lockid
, namelen
)) {
920 mlog(0, "%s: Recovery map is not empty, but must "
921 "master $RECOVERY lock now\n", dlm
->name
);
922 if (!dlm_pre_master_reco_lockres(dlm
, res
))
923 wait_on_recovery
= 0;
925 mlog(0, "%s: waiting 500ms for heartbeat state "
926 "change\n", dlm
->name
);
932 dlm_kick_recovery_thread(dlm
);
934 dlm_wait_for_recovery(dlm
);
936 spin_lock(&dlm
->spinlock
);
937 bit
= find_next_bit(dlm
->recovery_map
, O2NM_MAX_NODES
, 0);
938 if (bit
< O2NM_MAX_NODES
) {
939 mlog(0, "%s: res %.*s, At least one node (%d) "
940 "to recover before lock mastery can begin\n",
941 dlm
->name
, namelen
, (char *)lockid
, bit
);
942 wait_on_recovery
= 1;
944 wait_on_recovery
= 0;
945 spin_unlock(&dlm
->spinlock
);
947 if (wait_on_recovery
)
948 dlm_wait_for_node_recovery(dlm
, bit
, 10000);
951 /* must wait for lock to be mastered elsewhere */
956 dlm_node_iter_init(mle
->vote_map
, &iter
);
957 while ((nodenum
= dlm_node_iter_next(&iter
)) >= 0) {
958 ret
= dlm_do_master_request(res
, mle
, nodenum
);
961 if (mle
->master
!= O2NM_MAX_NODES
) {
962 /* found a master ! */
963 if (mle
->master
<= nodenum
)
965 /* if our master request has not reached the master
966 * yet, keep going until it does. this is how the
967 * master will know that asserts are needed back to
968 * the lower nodes. */
969 mlog(0, "%s: res %.*s, Requests only up to %u but "
970 "master is %u, keep going\n", dlm
->name
, namelen
,
971 lockid
, nodenum
, mle
->master
);
976 /* keep going until the response map includes all nodes */
977 ret
= dlm_wait_for_lock_mastery(dlm
, res
, mle
, &blocked
);
979 wait_on_recovery
= 1;
980 mlog(0, "%s: res %.*s, Node map changed, redo the master "
981 "request now, blocked=%d\n", dlm
->name
, res
->lockname
.len
,
982 res
->lockname
.name
, blocked
);
984 mlog(ML_ERROR
, "%s: res %.*s, Spinning on "
985 "dlm_wait_for_lock_mastery, blocked = %d\n",
986 dlm
->name
, res
->lockname
.len
,
987 res
->lockname
.name
, blocked
);
988 dlm_print_one_lock_resource(res
);
989 dlm_print_one_mle(mle
);
995 mlog(0, "%s: res %.*s, Mastered by %u\n", dlm
->name
, res
->lockname
.len
,
996 res
->lockname
.name
, res
->owner
);
997 /* make sure we never continue without this */
998 BUG_ON(res
->owner
== O2NM_MAX_NODES
);
1000 /* master is known, detach if not already detached */
1001 dlm_mle_detach_hb_events(dlm
, mle
);
1003 /* put the extra ref */
1004 dlm_put_mle_inuse(mle
);
1007 spin_lock(&res
->spinlock
);
1008 res
->state
&= ~DLM_LOCK_RES_IN_PROGRESS
;
1009 spin_unlock(&res
->spinlock
);
1013 /* need to free the unused mle */
1015 kmem_cache_free(dlm_mle_cache
, alloc_mle
);
1021 #define DLM_MASTERY_TIMEOUT_MS 5000
1023 static int dlm_wait_for_lock_mastery(struct dlm_ctxt
*dlm
,
1024 struct dlm_lock_resource
*res
,
1025 struct dlm_master_list_entry
*mle
,
1030 int map_changed
, voting_done
;
1037 /* check if another node has already become the owner */
1038 spin_lock(&res
->spinlock
);
1039 if (res
->owner
!= DLM_LOCK_RES_OWNER_UNKNOWN
) {
1040 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm
->name
,
1041 res
->lockname
.len
, res
->lockname
.name
, res
->owner
);
1042 spin_unlock(&res
->spinlock
);
1043 /* this will cause the master to re-assert across
1044 * the whole cluster, freeing up mles */
1045 if (res
->owner
!= dlm
->node_num
) {
1046 ret
= dlm_do_master_request(res
, mle
, res
->owner
);
1048 /* give recovery a chance to run */
1049 mlog(ML_ERROR
, "link to %u went down?: %d\n", res
->owner
, ret
);
1057 spin_unlock(&res
->spinlock
);
1059 spin_lock(&mle
->spinlock
);
1061 map_changed
= (memcmp(mle
->vote_map
, mle
->node_map
,
1062 sizeof(mle
->vote_map
)) != 0);
1063 voting_done
= (memcmp(mle
->vote_map
, mle
->response_map
,
1064 sizeof(mle
->vote_map
)) == 0);
1066 /* restart if we hit any errors */
1069 mlog(0, "%s: %.*s: node map changed, restarting\n",
1070 dlm
->name
, res
->lockname
.len
, res
->lockname
.name
);
1071 ret
= dlm_restart_lock_mastery(dlm
, res
, mle
, *blocked
);
1072 b
= (mle
->type
== DLM_MLE_BLOCK
);
1073 if ((*blocked
&& !b
) || (!*blocked
&& b
)) {
1074 mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1075 dlm
->name
, res
->lockname
.len
, res
->lockname
.name
,
1079 spin_unlock(&mle
->spinlock
);
1084 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1085 "rechecking now\n", dlm
->name
, res
->lockname
.len
,
1086 res
->lockname
.name
);
1090 mlog(0, "map not changed and voting not done "
1091 "for %s:%.*s\n", dlm
->name
, res
->lockname
.len
,
1092 res
->lockname
.name
);
1096 if (m
!= O2NM_MAX_NODES
) {
1097 /* another node has done an assert!
1102 /* have all nodes responded? */
1103 if (voting_done
&& !*blocked
) {
1104 bit
= find_next_bit(mle
->maybe_map
, O2NM_MAX_NODES
, 0);
1105 if (dlm
->node_num
<= bit
) {
1106 /* my node number is lowest.
1107 * now tell other nodes that I am
1108 * mastering this. */
1109 mle
->master
= dlm
->node_num
;
1110 /* ref was grabbed in get_lock_resource
1111 * will be dropped in dlmlock_master */
1115 /* if voting is done, but we have not received
1116 * an assert master yet, we must sleep */
1120 spin_unlock(&mle
->spinlock
);
1122 /* sleep if we haven't finished voting yet */
1124 unsigned long timeo
= msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS
);
1125 atomic_set(&mle
->woken
, 0);
1126 (void)wait_event_timeout(mle
->wq
,
1127 (atomic_read(&mle
->woken
) == 1),
1129 if (res
->owner
== O2NM_MAX_NODES
) {
1130 mlog(0, "%s:%.*s: waiting again\n", dlm
->name
,
1131 res
->lockname
.len
, res
->lockname
.name
);
1134 mlog(0, "done waiting, master is %u\n", res
->owner
);
1142 mlog(0, "about to master %.*s here, this=%u\n",
1143 res
->lockname
.len
, res
->lockname
.name
, m
);
1144 ret
= dlm_do_assert_master(dlm
, res
, mle
->vote_map
, 0);
1146 /* This is a failure in the network path,
1147 * not in the response to the assert_master
1148 * (any nonzero response is a BUG on this node).
1149 * Most likely a socket just got disconnected
1150 * due to node death. */
1153 /* no longer need to restart lock mastery.
1154 * all living nodes have been contacted. */
1158 /* set the lockres owner */
1159 spin_lock(&res
->spinlock
);
1160 /* mastery reference obtained either during
1161 * assert_master_handler or in get_lock_resource */
1162 dlm_change_lockres_owner(dlm
, res
, m
);
1163 spin_unlock(&res
->spinlock
);
1169 struct dlm_bitmap_diff_iter
1172 unsigned long *orig_bm
;
1173 unsigned long *cur_bm
;
1174 unsigned long diff_bm
[BITS_TO_LONGS(O2NM_MAX_NODES
)];
1177 enum dlm_node_state_change
1184 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter
*iter
,
1185 unsigned long *orig_bm
,
1186 unsigned long *cur_bm
)
1188 unsigned long p1
, p2
;
1192 iter
->orig_bm
= orig_bm
;
1193 iter
->cur_bm
= cur_bm
;
1195 for (i
= 0; i
< BITS_TO_LONGS(O2NM_MAX_NODES
); i
++) {
1196 p1
= *(iter
->orig_bm
+ i
);
1197 p2
= *(iter
->cur_bm
+ i
);
1198 iter
->diff_bm
[i
] = (p1
& ~p2
) | (p2
& ~p1
);
1202 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter
*iter
,
1203 enum dlm_node_state_change
*state
)
1207 if (iter
->curnode
>= O2NM_MAX_NODES
)
1210 bit
= find_next_bit(iter
->diff_bm
, O2NM_MAX_NODES
,
1212 if (bit
>= O2NM_MAX_NODES
) {
1213 iter
->curnode
= O2NM_MAX_NODES
;
1217 /* if it was there in the original then this node died */
1218 if (test_bit(bit
, iter
->orig_bm
))
1223 iter
->curnode
= bit
;
1228 static int dlm_restart_lock_mastery(struct dlm_ctxt
*dlm
,
1229 struct dlm_lock_resource
*res
,
1230 struct dlm_master_list_entry
*mle
,
1233 struct dlm_bitmap_diff_iter bdi
;
1234 enum dlm_node_state_change sc
;
1238 mlog(0, "something happened such that the "
1239 "master process may need to be restarted!\n");
1241 assert_spin_locked(&mle
->spinlock
);
1243 dlm_bitmap_diff_iter_init(&bdi
, mle
->vote_map
, mle
->node_map
);
1244 node
= dlm_bitmap_diff_iter_next(&bdi
, &sc
);
1246 if (sc
== NODE_UP
) {
1247 /* a node came up. clear any old vote from
1248 * the response map and set it in the vote map
1249 * then restart the mastery. */
1250 mlog(ML_NOTICE
, "node %d up while restarting\n", node
);
1252 /* redo the master request, but only for the new node */
1253 mlog(0, "sending request to new node\n");
1254 clear_bit(node
, mle
->response_map
);
1255 set_bit(node
, mle
->vote_map
);
1257 mlog(ML_ERROR
, "node down! %d\n", node
);
1259 int lowest
= find_next_bit(mle
->maybe_map
,
1262 /* act like it was never there */
1263 clear_bit(node
, mle
->maybe_map
);
1265 if (node
== lowest
) {
1266 mlog(0, "expected master %u died"
1267 " while this node was blocked "
1268 "waiting on it!\n", node
);
1269 lowest
= find_next_bit(mle
->maybe_map
,
1272 if (lowest
< O2NM_MAX_NODES
) {
1273 mlog(0, "%s:%.*s:still "
1274 "blocked. waiting on %u "
1280 /* mle is an MLE_BLOCK, but
1281 * there is now nothing left to
1282 * block on. we need to return
1283 * all the way back out and try
1284 * again with an MLE_MASTER.
1285 * dlm_do_local_recovery_cleanup
1286 * has already run, so the mle
1288 mlog(0, "%s:%.*s: no "
1289 "longer blocking. try to "
1290 "master this here\n",
1293 res
->lockname
.name
);
1294 mle
->type
= DLM_MLE_MASTER
;
1300 /* now blank out everything, as if we had never
1301 * contacted anyone */
1302 memset(mle
->maybe_map
, 0, sizeof(mle
->maybe_map
));
1303 memset(mle
->response_map
, 0, sizeof(mle
->response_map
));
1304 /* reset the vote_map to the current node_map */
1305 memcpy(mle
->vote_map
, mle
->node_map
,
1306 sizeof(mle
->node_map
));
1307 /* put myself into the maybe map */
1308 if (mle
->type
!= DLM_MLE_BLOCK
)
1309 set_bit(dlm
->node_num
, mle
->maybe_map
);
1312 node
= dlm_bitmap_diff_iter_next(&bdi
, &sc
);
1319 * DLM_MASTER_REQUEST_MSG
1321 * returns: 0 on success,
1322 * -errno on a network error
1324 * on error, the caller should assume the target node is "dead"
1328 static int dlm_do_master_request(struct dlm_lock_resource
*res
,
1329 struct dlm_master_list_entry
*mle
, int to
)
1331 struct dlm_ctxt
*dlm
= mle
->dlm
;
1332 struct dlm_master_request request
;
1333 int ret
, response
=0, resend
;
1335 memset(&request
, 0, sizeof(request
));
1336 request
.node_idx
= dlm
->node_num
;
1338 BUG_ON(mle
->type
== DLM_MLE_MIGRATION
);
1340 request
.namelen
= (u8
)mle
->mnamelen
;
1341 memcpy(request
.name
, mle
->mname
, request
.namelen
);
1344 ret
= o2net_send_message(DLM_MASTER_REQUEST_MSG
, dlm
->key
, &request
,
1345 sizeof(request
), to
, &response
);
1347 if (ret
== -ESRCH
) {
1348 /* should never happen */
1349 mlog(ML_ERROR
, "TCP stack not ready!\n");
1351 } else if (ret
== -EINVAL
) {
1352 mlog(ML_ERROR
, "bad args passed to o2net!\n");
1354 } else if (ret
== -ENOMEM
) {
1355 mlog(ML_ERROR
, "out of memory while trying to send "
1356 "network message! retrying\n");
1357 /* this is totally crude */
1360 } else if (!dlm_is_host_down(ret
)) {
1361 /* not a network error. bad. */
1363 mlog(ML_ERROR
, "unhandled error!");
1366 /* all other errors should be network errors,
1367 * and likely indicate node death */
1368 mlog(ML_ERROR
, "link to %d went down!\n", to
);
1374 spin_lock(&mle
->spinlock
);
1376 case DLM_MASTER_RESP_YES
:
1377 set_bit(to
, mle
->response_map
);
1378 mlog(0, "node %u is the master, response=YES\n", to
);
1379 mlog(0, "%s:%.*s: master node %u now knows I have a "
1380 "reference\n", dlm
->name
, res
->lockname
.len
,
1381 res
->lockname
.name
, to
);
1384 case DLM_MASTER_RESP_NO
:
1385 mlog(0, "node %u not master, response=NO\n", to
);
1386 set_bit(to
, mle
->response_map
);
1388 case DLM_MASTER_RESP_MAYBE
:
1389 mlog(0, "node %u not master, response=MAYBE\n", to
);
1390 set_bit(to
, mle
->response_map
);
1391 set_bit(to
, mle
->maybe_map
);
1393 case DLM_MASTER_RESP_ERROR
:
1394 mlog(0, "node %u hit an error, resending\n", to
);
1399 mlog(ML_ERROR
, "bad response! %u\n", response
);
1402 spin_unlock(&mle
->spinlock
);
1404 /* this is also totally crude */
1414 * locks that can be taken here:
1420 * if possible, TRIM THIS DOWN!!!
1422 int dlm_master_request_handler(struct o2net_msg
*msg
, u32 len
, void *data
,
1425 u8 response
= DLM_MASTER_RESP_MAYBE
;
1426 struct dlm_ctxt
*dlm
= data
;
1427 struct dlm_lock_resource
*res
= NULL
;
1428 struct dlm_master_request
*request
= (struct dlm_master_request
*) msg
->buf
;
1429 struct dlm_master_list_entry
*mle
= NULL
, *tmpmle
= NULL
;
1431 unsigned int namelen
, hash
;
1434 int dispatch_assert
= 0;
1438 return DLM_MASTER_RESP_NO
;
1440 if (!dlm_domain_fully_joined(dlm
)) {
1441 response
= DLM_MASTER_RESP_NO
;
1445 name
= request
->name
;
1446 namelen
= request
->namelen
;
1447 hash
= dlm_lockid_hash(name
, namelen
);
1449 if (namelen
> DLM_LOCKID_NAME_MAX
) {
1450 response
= DLM_IVBUFLEN
;
1455 spin_lock(&dlm
->spinlock
);
1456 res
= __dlm_lookup_lockres(dlm
, name
, namelen
, hash
);
1458 spin_unlock(&dlm
->spinlock
);
1460 /* take care of the easy cases up front */
1461 spin_lock(&res
->spinlock
);
1464 * Right after dlm spinlock was released, dlm_thread could have
1465 * purged the lockres. Check if lockres got unhashed. If so
1468 if (hlist_unhashed(&res
->hash_node
)) {
1469 spin_unlock(&res
->spinlock
);
1470 dlm_lockres_put(res
);
1474 if (res
->state
& (DLM_LOCK_RES_RECOVERING
|
1475 DLM_LOCK_RES_MIGRATING
)) {
1476 spin_unlock(&res
->spinlock
);
1477 mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1478 "being recovered/migrated\n");
1479 response
= DLM_MASTER_RESP_ERROR
;
1481 kmem_cache_free(dlm_mle_cache
, mle
);
1485 if (res
->owner
== dlm
->node_num
) {
1486 dlm_lockres_set_refmap_bit(dlm
, res
, request
->node_idx
);
1487 spin_unlock(&res
->spinlock
);
1488 response
= DLM_MASTER_RESP_YES
;
1490 kmem_cache_free(dlm_mle_cache
, mle
);
1492 /* this node is the owner.
1493 * there is some extra work that needs to
1494 * happen now. the requesting node has
1495 * caused all nodes up to this one to
1496 * create mles. this node now needs to
1497 * go back and clean those up. */
1498 dispatch_assert
= 1;
1500 } else if (res
->owner
!= DLM_LOCK_RES_OWNER_UNKNOWN
) {
1501 spin_unlock(&res
->spinlock
);
1502 // mlog(0, "node %u is the master\n", res->owner);
1503 response
= DLM_MASTER_RESP_NO
;
1505 kmem_cache_free(dlm_mle_cache
, mle
);
1509 /* ok, there is no owner. either this node is
1510 * being blocked, or it is actively trying to
1511 * master this lock. */
1512 if (!(res
->state
& DLM_LOCK_RES_IN_PROGRESS
)) {
1513 mlog(ML_ERROR
, "lock with no owner should be "
1518 // mlog(0, "lockres is in progress...\n");
1519 spin_lock(&dlm
->master_lock
);
1520 found
= dlm_find_mle(dlm
, &tmpmle
, name
, namelen
);
1522 mlog(ML_ERROR
, "no mle found for this lock!\n");
1526 spin_lock(&tmpmle
->spinlock
);
1527 if (tmpmle
->type
== DLM_MLE_BLOCK
) {
1528 // mlog(0, "this node is waiting for "
1529 // "lockres to be mastered\n");
1530 response
= DLM_MASTER_RESP_NO
;
1531 } else if (tmpmle
->type
== DLM_MLE_MIGRATION
) {
1532 mlog(0, "node %u is master, but trying to migrate to "
1533 "node %u.\n", tmpmle
->master
, tmpmle
->new_master
);
1534 if (tmpmle
->master
== dlm
->node_num
) {
1535 mlog(ML_ERROR
, "no owner on lockres, but this "
1536 "node is trying to migrate it to %u?!\n",
1537 tmpmle
->new_master
);
1540 /* the real master can respond on its own */
1541 response
= DLM_MASTER_RESP_NO
;
1543 } else if (tmpmle
->master
!= DLM_LOCK_RES_OWNER_UNKNOWN
) {
1545 if (tmpmle
->master
== dlm
->node_num
) {
1546 response
= DLM_MASTER_RESP_YES
;
1547 /* this node will be the owner.
1548 * go back and clean the mles on any
1550 dispatch_assert
= 1;
1551 dlm_lockres_set_refmap_bit(dlm
, res
,
1554 response
= DLM_MASTER_RESP_NO
;
1556 // mlog(0, "this node is attempting to "
1557 // "master lockres\n");
1558 response
= DLM_MASTER_RESP_MAYBE
;
1561 set_bit(request
->node_idx
, tmpmle
->maybe_map
);
1562 spin_unlock(&tmpmle
->spinlock
);
1564 spin_unlock(&dlm
->master_lock
);
1565 spin_unlock(&res
->spinlock
);
1567 /* keep the mle attached to heartbeat events */
1568 dlm_put_mle(tmpmle
);
1570 kmem_cache_free(dlm_mle_cache
, mle
);
1575 * lockres doesn't exist on this node
1576 * if there is an MLE_BLOCK, return NO
1577 * if there is an MLE_MASTER, return MAYBE
1578 * otherwise, add an MLE_BLOCK, return NO
1580 spin_lock(&dlm
->master_lock
);
1581 found
= dlm_find_mle(dlm
, &tmpmle
, name
, namelen
);
1583 /* this lockid has never been seen on this node yet */
1584 // mlog(0, "no mle found\n");
1586 spin_unlock(&dlm
->master_lock
);
1587 spin_unlock(&dlm
->spinlock
);
1589 mle
= kmem_cache_alloc(dlm_mle_cache
, GFP_NOFS
);
1591 response
= DLM_MASTER_RESP_ERROR
;
1592 mlog_errno(-ENOMEM
);
1598 // mlog(0, "this is second time thru, already allocated, "
1599 // "add the block.\n");
1600 dlm_init_mle(mle
, DLM_MLE_BLOCK
, dlm
, NULL
, name
, namelen
);
1601 set_bit(request
->node_idx
, mle
->maybe_map
);
1602 __dlm_insert_mle(dlm
, mle
);
1603 response
= DLM_MASTER_RESP_NO
;
1605 spin_lock(&tmpmle
->spinlock
);
1606 if (tmpmle
->master
== dlm
->node_num
) {
1607 mlog(ML_ERROR
, "no lockres, but an mle with this node as master!\n");
1610 if (tmpmle
->type
== DLM_MLE_BLOCK
)
1611 response
= DLM_MASTER_RESP_NO
;
1612 else if (tmpmle
->type
== DLM_MLE_MIGRATION
) {
1613 mlog(0, "migration mle was found (%u->%u)\n",
1614 tmpmle
->master
, tmpmle
->new_master
);
1615 /* real master can respond on its own */
1616 response
= DLM_MASTER_RESP_NO
;
1618 response
= DLM_MASTER_RESP_MAYBE
;
1619 set_bit(request
->node_idx
, tmpmle
->maybe_map
);
1620 spin_unlock(&tmpmle
->spinlock
);
1622 spin_unlock(&dlm
->master_lock
);
1623 spin_unlock(&dlm
->spinlock
);
1626 /* keep the mle attached to heartbeat events */
1627 dlm_put_mle(tmpmle
);
1631 * __dlm_lookup_lockres() grabbed a reference to this lockres.
1632 * The reference is released by dlm_assert_master_worker() under
1633 * the call to dlm_dispatch_assert_master(). If
1634 * dlm_assert_master_worker() isn't called, we drop it here.
1636 if (dispatch_assert
) {
1637 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1638 dlm
->node_num
, res
->lockname
.len
, res
->lockname
.name
);
1639 spin_lock(&res
->spinlock
);
1640 ret
= dlm_dispatch_assert_master(dlm
, res
, 0, request
->node_idx
,
1641 DLM_ASSERT_MASTER_MLE_CLEANUP
);
1643 mlog(ML_ERROR
, "failed to dispatch assert master work\n");
1644 response
= DLM_MASTER_RESP_ERROR
;
1645 spin_unlock(&res
->spinlock
);
1646 dlm_lockres_put(res
);
1649 __dlm_lockres_grab_inflight_worker(dlm
, res
);
1650 spin_unlock(&res
->spinlock
);
1654 dlm_lockres_put(res
);
1663 * DLM_ASSERT_MASTER_MSG
1668 * NOTE: this can be used for debugging
1669 * can periodically run all locks owned by this node
1670 * and re-assert across the cluster...
1672 static int dlm_do_assert_master(struct dlm_ctxt
*dlm
,
1673 struct dlm_lock_resource
*res
,
1674 void *nodemap
, u32 flags
)
1676 struct dlm_assert_master
assert;
1678 struct dlm_node_iter iter
;
1681 const char *lockname
= res
->lockname
.name
;
1682 unsigned int namelen
= res
->lockname
.len
;
1684 BUG_ON(namelen
> O2NM_MAX_NAME_LEN
);
1686 spin_lock(&res
->spinlock
);
1687 res
->state
|= DLM_LOCK_RES_SETREF_INPROG
;
1688 spin_unlock(&res
->spinlock
);
1693 /* note that if this nodemap is empty, it returns 0 */
1694 dlm_node_iter_init(nodemap
, &iter
);
1695 while ((to
= dlm_node_iter_next(&iter
)) >= 0) {
1697 struct dlm_master_list_entry
*mle
= NULL
;
1699 mlog(0, "sending assert master to %d (%.*s)\n", to
,
1701 memset(&assert, 0, sizeof(assert));
1702 assert.node_idx
= dlm
->node_num
;
1703 assert.namelen
= namelen
;
1704 memcpy(assert.name
, lockname
, namelen
);
1705 assert.flags
= cpu_to_be32(flags
);
1707 tmpret
= o2net_send_message(DLM_ASSERT_MASTER_MSG
, dlm
->key
,
1708 &assert, sizeof(assert), to
, &r
);
1710 mlog(ML_ERROR
, "Error %d when sending message %u (key "
1711 "0x%x) to node %u\n", tmpret
,
1712 DLM_ASSERT_MASTER_MSG
, dlm
->key
, to
);
1713 if (!dlm_is_host_down(tmpret
)) {
1714 mlog(ML_ERROR
, "unhandled error=%d!\n", tmpret
);
1717 /* a node died. finish out the rest of the nodes. */
1718 mlog(0, "link to %d went down!\n", to
);
1719 /* any nonzero status return will do */
1723 /* ok, something horribly messed. kill thyself. */
1724 mlog(ML_ERROR
,"during assert master of %.*s to %u, "
1725 "got %d.\n", namelen
, lockname
, to
, r
);
1726 spin_lock(&dlm
->spinlock
);
1727 spin_lock(&dlm
->master_lock
);
1728 if (dlm_find_mle(dlm
, &mle
, (char *)lockname
,
1730 dlm_print_one_mle(mle
);
1733 spin_unlock(&dlm
->master_lock
);
1734 spin_unlock(&dlm
->spinlock
);
1738 if (r
& DLM_ASSERT_RESPONSE_REASSERT
&&
1739 !(r
& DLM_ASSERT_RESPONSE_MASTERY_REF
)) {
1740 mlog(ML_ERROR
, "%.*s: very strange, "
1741 "master MLE but no lockres on %u\n",
1742 namelen
, lockname
, to
);
1745 if (r
& DLM_ASSERT_RESPONSE_REASSERT
) {
1746 mlog(0, "%.*s: node %u create mles on other "
1747 "nodes and requests a re-assert\n",
1748 namelen
, lockname
, to
);
1751 if (r
& DLM_ASSERT_RESPONSE_MASTERY_REF
) {
1752 mlog(0, "%.*s: node %u has a reference to this "
1753 "lockres, set the bit in the refmap\n",
1754 namelen
, lockname
, to
);
1755 spin_lock(&res
->spinlock
);
1756 dlm_lockres_set_refmap_bit(dlm
, res
, to
);
1757 spin_unlock(&res
->spinlock
);
1764 spin_lock(&res
->spinlock
);
1765 res
->state
&= ~DLM_LOCK_RES_SETREF_INPROG
;
1766 spin_unlock(&res
->spinlock
);
1773 * locks that can be taken here:
1779 * if possible, TRIM THIS DOWN!!!
1781 int dlm_assert_master_handler(struct o2net_msg
*msg
, u32 len
, void *data
,
1784 struct dlm_ctxt
*dlm
= data
;
1785 struct dlm_master_list_entry
*mle
= NULL
;
1786 struct dlm_assert_master
*assert = (struct dlm_assert_master
*)msg
->buf
;
1787 struct dlm_lock_resource
*res
= NULL
;
1789 unsigned int namelen
, hash
;
1791 int master_request
= 0, have_lockres_ref
= 0;
1797 name
= assert->name
;
1798 namelen
= assert->namelen
;
1799 hash
= dlm_lockid_hash(name
, namelen
);
1800 flags
= be32_to_cpu(assert->flags
);
1802 if (namelen
> DLM_LOCKID_NAME_MAX
) {
1803 mlog(ML_ERROR
, "Invalid name length!");
1807 spin_lock(&dlm
->spinlock
);
1810 mlog(0, "assert_master with flags: %u\n", flags
);
1813 spin_lock(&dlm
->master_lock
);
1814 if (!dlm_find_mle(dlm
, &mle
, name
, namelen
)) {
1815 /* not an error, could be master just re-asserting */
1816 mlog(0, "just got an assert_master from %u, but no "
1817 "MLE for it! (%.*s)\n", assert->node_idx
,
1820 int bit
= find_next_bit (mle
->maybe_map
, O2NM_MAX_NODES
, 0);
1821 if (bit
>= O2NM_MAX_NODES
) {
1822 /* not necessarily an error, though less likely.
1823 * could be master just re-asserting. */
1824 mlog(0, "no bits set in the maybe_map, but %u "
1825 "is asserting! (%.*s)\n", assert->node_idx
,
1827 } else if (bit
!= assert->node_idx
) {
1828 if (flags
& DLM_ASSERT_MASTER_MLE_CLEANUP
) {
1829 mlog(0, "master %u was found, %u should "
1830 "back off\n", assert->node_idx
, bit
);
1832 /* with the fix for bug 569, a higher node
1833 * number winning the mastery will respond
1834 * YES to mastery requests, but this node
1835 * had no way of knowing. let it pass. */
1836 mlog(0, "%u is the lowest node, "
1837 "%u is asserting. (%.*s) %u must "
1838 "have begun after %u won.\n", bit
,
1839 assert->node_idx
, namelen
, name
, bit
,
1843 if (mle
->type
== DLM_MLE_MIGRATION
) {
1844 if (flags
& DLM_ASSERT_MASTER_MLE_CLEANUP
) {
1845 mlog(0, "%s:%.*s: got cleanup assert"
1846 " from %u for migration\n",
1847 dlm
->name
, namelen
, name
,
1849 } else if (!(flags
& DLM_ASSERT_MASTER_FINISH_MIGRATION
)) {
1850 mlog(0, "%s:%.*s: got unrelated assert"
1851 " from %u for migration, ignoring\n",
1852 dlm
->name
, namelen
, name
,
1855 spin_unlock(&dlm
->master_lock
);
1856 spin_unlock(&dlm
->spinlock
);
1861 spin_unlock(&dlm
->master_lock
);
1863 /* ok everything checks out with the MLE
1864 * now check to see if there is a lockres */
1865 res
= __dlm_lookup_lockres(dlm
, name
, namelen
, hash
);
1867 spin_lock(&res
->spinlock
);
1868 if (res
->state
& DLM_LOCK_RES_RECOVERING
) {
1869 mlog(ML_ERROR
, "%u asserting but %.*s is "
1870 "RECOVERING!\n", assert->node_idx
, namelen
, name
);
1874 if (res
->owner
!= DLM_LOCK_RES_OWNER_UNKNOWN
&&
1875 res
->owner
!= assert->node_idx
) {
1876 mlog(ML_ERROR
, "DIE! Mastery assert from %u, "
1877 "but current owner is %u! (%.*s)\n",
1878 assert->node_idx
, res
->owner
, namelen
,
1880 __dlm_print_one_lock_resource(res
);
1883 } else if (mle
->type
!= DLM_MLE_MIGRATION
) {
1884 if (res
->owner
!= DLM_LOCK_RES_OWNER_UNKNOWN
) {
1885 /* owner is just re-asserting */
1886 if (res
->owner
== assert->node_idx
) {
1887 mlog(0, "owner %u re-asserting on "
1888 "lock %.*s\n", assert->node_idx
,
1892 mlog(ML_ERROR
, "got assert_master from "
1893 "node %u, but %u is the owner! "
1894 "(%.*s)\n", assert->node_idx
,
1895 res
->owner
, namelen
, name
);
1898 if (!(res
->state
& DLM_LOCK_RES_IN_PROGRESS
)) {
1899 mlog(ML_ERROR
, "got assert from %u, but lock "
1900 "with no owner should be "
1901 "in-progress! (%.*s)\n",
1906 } else /* mle->type == DLM_MLE_MIGRATION */ {
1907 /* should only be getting an assert from new master */
1908 if (assert->node_idx
!= mle
->new_master
) {
1909 mlog(ML_ERROR
, "got assert from %u, but "
1910 "new master is %u, and old master "
1912 assert->node_idx
, mle
->new_master
,
1913 mle
->master
, namelen
, name
);
1919 spin_unlock(&res
->spinlock
);
1922 // mlog(0, "woo! got an assert_master from node %u!\n",
1923 // assert->node_idx);
1929 spin_lock(&mle
->spinlock
);
1930 if (mle
->type
== DLM_MLE_BLOCK
|| mle
->type
== DLM_MLE_MIGRATION
)
1933 /* MASTER mle: if any bits set in the response map
1934 * then the calling node needs to re-assert to clear
1935 * up nodes that this node contacted */
1936 while ((nn
= find_next_bit (mle
->response_map
, O2NM_MAX_NODES
,
1937 nn
+1)) < O2NM_MAX_NODES
) {
1938 if (nn
!= dlm
->node_num
&& nn
!= assert->node_idx
) {
1944 mle
->master
= assert->node_idx
;
1945 atomic_set(&mle
->woken
, 1);
1947 spin_unlock(&mle
->spinlock
);
1951 spin_lock(&res
->spinlock
);
1952 if (mle
->type
== DLM_MLE_MIGRATION
) {
1953 mlog(0, "finishing off migration of lockres %.*s, "
1955 res
->lockname
.len
, res
->lockname
.name
,
1956 dlm
->node_num
, mle
->new_master
);
1957 res
->state
&= ~DLM_LOCK_RES_MIGRATING
;
1959 dlm_change_lockres_owner(dlm
, res
, mle
->new_master
);
1960 BUG_ON(res
->state
& DLM_LOCK_RES_DIRTY
);
1962 dlm_change_lockres_owner(dlm
, res
, mle
->master
);
1964 spin_unlock(&res
->spinlock
);
1965 have_lockres_ref
= 1;
1970 /* master is known, detach if not already detached.
1971 * ensures that only one assert_master call will happen
1973 spin_lock(&dlm
->master_lock
);
1975 rr
= kref_read(&mle
->mle_refs
);
1976 if (mle
->inuse
> 0) {
1977 if (extra_ref
&& rr
< 3)
1979 else if (!extra_ref
&& rr
< 2)
1982 if (extra_ref
&& rr
< 2)
1984 else if (!extra_ref
&& rr
< 1)
1988 mlog(ML_ERROR
, "%s:%.*s: got assert master from %u "
1989 "that will mess up this node, refs=%d, extra=%d, "
1990 "inuse=%d\n", dlm
->name
, namelen
, name
,
1991 assert->node_idx
, rr
, extra_ref
, mle
->inuse
);
1992 dlm_print_one_mle(mle
);
1994 __dlm_unlink_mle(dlm
, mle
);
1995 __dlm_mle_detach_hb_events(dlm
, mle
);
1998 /* the assert master message now balances the extra
1999 * ref given by the master / migration request message.
2000 * if this is the last put, it will be removed
2004 spin_unlock(&dlm
->master_lock
);
2006 if (res
->owner
!= assert->node_idx
) {
2007 mlog(0, "assert_master from %u, but current "
2008 "owner is %u (%.*s), no mle\n", assert->node_idx
,
2009 res
->owner
, namelen
, name
);
2012 spin_unlock(&dlm
->spinlock
);
2017 spin_lock(&res
->spinlock
);
2018 res
->state
|= DLM_LOCK_RES_SETREF_INPROG
;
2019 spin_unlock(&res
->spinlock
);
2020 *ret_data
= (void *)res
;
2023 if (master_request
) {
2024 mlog(0, "need to tell master to reassert\n");
2025 /* positive. negative would shoot down the node. */
2026 ret
|= DLM_ASSERT_RESPONSE_REASSERT
;
2027 if (!have_lockres_ref
) {
2028 mlog(ML_ERROR
, "strange, got assert from %u, MASTER "
2029 "mle present here for %s:%.*s, but no lockres!\n",
2030 assert->node_idx
, dlm
->name
, namelen
, name
);
2033 if (have_lockres_ref
) {
2034 /* let the master know we have a reference to the lockres */
2035 ret
|= DLM_ASSERT_RESPONSE_MASTERY_REF
;
2036 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2037 dlm
->name
, namelen
, name
, assert->node_idx
);
2042 /* kill the caller! */
2043 mlog(ML_ERROR
, "Bad message received from another node. Dumping state "
2044 "and killing the other node now! This node is OK and can continue.\n");
2045 __dlm_print_one_lock_resource(res
);
2046 spin_unlock(&res
->spinlock
);
2047 spin_lock(&dlm
->master_lock
);
2050 spin_unlock(&dlm
->master_lock
);
2051 spin_unlock(&dlm
->spinlock
);
2052 *ret_data
= (void *)res
;
2057 void dlm_assert_master_post_handler(int status
, void *data
, void *ret_data
)
2059 struct dlm_lock_resource
*res
= (struct dlm_lock_resource
*)ret_data
;
2062 spin_lock(&res
->spinlock
);
2063 res
->state
&= ~DLM_LOCK_RES_SETREF_INPROG
;
2064 spin_unlock(&res
->spinlock
);
2066 dlm_lockres_put(res
);
2071 int dlm_dispatch_assert_master(struct dlm_ctxt
*dlm
,
2072 struct dlm_lock_resource
*res
,
2073 int ignore_higher
, u8 request_from
, u32 flags
)
2075 struct dlm_work_item
*item
;
2076 item
= kzalloc(sizeof(*item
), GFP_ATOMIC
);
2081 /* queue up work for dlm_assert_master_worker */
2082 dlm_init_work_item(dlm
, item
, dlm_assert_master_worker
, NULL
);
2083 item
->u
.am
.lockres
= res
; /* already have a ref */
2084 /* can optionally ignore node numbers higher than this node */
2085 item
->u
.am
.ignore_higher
= ignore_higher
;
2086 item
->u
.am
.request_from
= request_from
;
2087 item
->u
.am
.flags
= flags
;
2090 mlog(0, "IGNORE HIGHER: %.*s\n", res
->lockname
.len
,
2091 res
->lockname
.name
);
2093 spin_lock(&dlm
->work_lock
);
2094 list_add_tail(&item
->list
, &dlm
->work_list
);
2095 spin_unlock(&dlm
->work_lock
);
2097 queue_work(dlm
->dlm_worker
, &dlm
->dispatched_work
);
2101 static void dlm_assert_master_worker(struct dlm_work_item
*item
, void *data
)
2103 struct dlm_ctxt
*dlm
= data
;
2105 struct dlm_lock_resource
*res
;
2106 unsigned long nodemap
[BITS_TO_LONGS(O2NM_MAX_NODES
)];
2113 res
= item
->u
.am
.lockres
;
2114 ignore_higher
= item
->u
.am
.ignore_higher
;
2115 request_from
= item
->u
.am
.request_from
;
2116 flags
= item
->u
.am
.flags
;
2118 spin_lock(&dlm
->spinlock
);
2119 memcpy(nodemap
, dlm
->domain_map
, sizeof(nodemap
));
2120 spin_unlock(&dlm
->spinlock
);
2122 clear_bit(dlm
->node_num
, nodemap
);
2123 if (ignore_higher
) {
2124 /* if is this just to clear up mles for nodes below
2125 * this node, do not send the message to the original
2126 * caller or any node number higher than this */
2127 clear_bit(request_from
, nodemap
);
2128 bit
= dlm
->node_num
;
2130 bit
= find_next_bit(nodemap
, O2NM_MAX_NODES
,
2132 if (bit
>= O2NM_MAX_NODES
)
2134 clear_bit(bit
, nodemap
);
2139 * If we're migrating this lock to someone else, we are no
2140 * longer allowed to assert out own mastery. OTOH, we need to
2141 * prevent migration from starting while we're still asserting
2142 * our dominance. The reserved ast delays migration.
2144 spin_lock(&res
->spinlock
);
2145 if (res
->state
& DLM_LOCK_RES_MIGRATING
) {
2146 mlog(0, "Someone asked us to assert mastery, but we're "
2147 "in the middle of migration. Skipping assert, "
2148 "the new master will handle that.\n");
2149 spin_unlock(&res
->spinlock
);
2152 __dlm_lockres_reserve_ast(res
);
2153 spin_unlock(&res
->spinlock
);
2155 /* this call now finishes out the nodemap
2156 * even if one or more nodes die */
2157 mlog(0, "worker about to master %.*s here, this=%u\n",
2158 res
->lockname
.len
, res
->lockname
.name
, dlm
->node_num
);
2159 ret
= dlm_do_assert_master(dlm
, res
, nodemap
, flags
);
2161 /* no need to restart, we are done */
2162 if (!dlm_is_host_down(ret
))
2166 /* Ok, we've asserted ourselves. Let's let migration start. */
2167 dlm_lockres_release_ast(dlm
, res
);
2170 dlm_lockres_drop_inflight_worker(dlm
, res
);
2172 dlm_lockres_put(res
);
2174 mlog(0, "finished with dlm_assert_master_worker\n");
2177 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2178 * We cannot wait for node recovery to complete to begin mastering this
2179 * lockres because this lockres is used to kick off recovery! ;-)
2180 * So, do a pre-check on all living nodes to see if any of those nodes
2181 * think that $RECOVERY is currently mastered by a dead node. If so,
2182 * we wait a short time to allow that node to get notified by its own
2183 * heartbeat stack, then check again. All $RECOVERY lock resources
2184 * mastered by dead nodes are purged when the hearbeat callback is
2185 * fired, so we can know for sure that it is safe to continue once
2186 * the node returns a live node or no node. */
2187 static int dlm_pre_master_reco_lockres(struct dlm_ctxt
*dlm
,
2188 struct dlm_lock_resource
*res
)
2190 struct dlm_node_iter iter
;
2193 u8 master
= DLM_LOCK_RES_OWNER_UNKNOWN
;
2195 spin_lock(&dlm
->spinlock
);
2196 dlm_node_iter_init(dlm
->domain_map
, &iter
);
2197 spin_unlock(&dlm
->spinlock
);
2199 while ((nodenum
= dlm_node_iter_next(&iter
)) >= 0) {
2200 /* do not send to self */
2201 if (nodenum
== dlm
->node_num
)
2203 ret
= dlm_do_master_requery(dlm
, res
, nodenum
, &master
);
2206 if (!dlm_is_host_down(ret
))
2208 /* host is down, so answer for that node would be
2209 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
2213 if (master
!= DLM_LOCK_RES_OWNER_UNKNOWN
) {
2214 /* check to see if this master is in the recovery map */
2215 spin_lock(&dlm
->spinlock
);
2216 if (test_bit(master
, dlm
->recovery_map
)) {
2217 mlog(ML_NOTICE
, "%s: node %u has not seen "
2218 "node %u go down yet, and thinks the "
2219 "dead node is mastering the recovery "
2220 "lock. must wait.\n", dlm
->name
,
2224 spin_unlock(&dlm
->spinlock
);
2225 mlog(0, "%s: reco lock master is %u\n", dlm
->name
,
2234 * DLM_DEREF_LOCKRES_MSG
2237 int dlm_drop_lockres_ref(struct dlm_ctxt
*dlm
, struct dlm_lock_resource
*res
)
2239 struct dlm_deref_lockres deref
;
2241 const char *lockname
;
2242 unsigned int namelen
;
2244 lockname
= res
->lockname
.name
;
2245 namelen
= res
->lockname
.len
;
2246 BUG_ON(namelen
> O2NM_MAX_NAME_LEN
);
2248 memset(&deref
, 0, sizeof(deref
));
2249 deref
.node_idx
= dlm
->node_num
;
2250 deref
.namelen
= namelen
;
2251 memcpy(deref
.name
, lockname
, namelen
);
2253 ret
= o2net_send_message(DLM_DEREF_LOCKRES_MSG
, dlm
->key
,
2254 &deref
, sizeof(deref
), res
->owner
, &r
);
2256 mlog(ML_ERROR
, "%s: res %.*s, error %d send DEREF to node %u\n",
2257 dlm
->name
, namelen
, lockname
, ret
, res
->owner
);
2259 /* BAD. other node says I did not have a ref. */
2260 mlog(ML_ERROR
, "%s: res %.*s, DEREF to node %u got %d\n",
2261 dlm
->name
, namelen
, lockname
, res
->owner
, r
);
2262 dlm_print_one_lock_resource(res
);
2271 int dlm_deref_lockres_handler(struct o2net_msg
*msg
, u32 len
, void *data
,
2274 struct dlm_ctxt
*dlm
= data
;
2275 struct dlm_deref_lockres
*deref
= (struct dlm_deref_lockres
*)msg
->buf
;
2276 struct dlm_lock_resource
*res
= NULL
;
2278 unsigned int namelen
;
2282 struct dlm_work_item
*item
;
2290 namelen
= deref
->namelen
;
2291 node
= deref
->node_idx
;
2293 if (namelen
> DLM_LOCKID_NAME_MAX
) {
2294 mlog(ML_ERROR
, "Invalid name length!");
2297 if (deref
->node_idx
>= O2NM_MAX_NODES
) {
2298 mlog(ML_ERROR
, "Invalid node number: %u\n", node
);
2302 hash
= dlm_lockid_hash(name
, namelen
);
2304 spin_lock(&dlm
->spinlock
);
2305 res
= __dlm_lookup_lockres_full(dlm
, name
, namelen
, hash
);
2307 spin_unlock(&dlm
->spinlock
);
2308 mlog(ML_ERROR
, "%s:%.*s: bad lockres name\n",
2309 dlm
->name
, namelen
, name
);
2312 spin_unlock(&dlm
->spinlock
);
2314 spin_lock(&res
->spinlock
);
2315 if (res
->state
& DLM_LOCK_RES_SETREF_INPROG
)
2318 BUG_ON(res
->state
& DLM_LOCK_RES_DROPPING_REF
);
2319 if (test_bit(node
, res
->refmap
)) {
2320 dlm_lockres_clear_refmap_bit(dlm
, res
, node
);
2324 spin_unlock(&res
->spinlock
);
2328 dlm_lockres_calc_usage(dlm
, res
);
2330 mlog(ML_ERROR
, "%s:%.*s: node %u trying to drop ref "
2331 "but it is already dropped!\n", dlm
->name
,
2332 res
->lockname
.len
, res
->lockname
.name
, node
);
2333 dlm_print_one_lock_resource(res
);
2335 ret
= DLM_DEREF_RESPONSE_DONE
;
2339 item
= kzalloc(sizeof(*item
), GFP_NOFS
);
2346 dlm_init_work_item(dlm
, item
, dlm_deref_lockres_worker
, NULL
);
2347 item
->u
.dl
.deref_res
= res
;
2348 item
->u
.dl
.deref_node
= node
;
2350 spin_lock(&dlm
->work_lock
);
2351 list_add_tail(&item
->list
, &dlm
->work_list
);
2352 spin_unlock(&dlm
->work_lock
);
2354 queue_work(dlm
->dlm_worker
, &dlm
->dispatched_work
);
2355 return DLM_DEREF_RESPONSE_INPROG
;
2359 dlm_lockres_put(res
);
2365 int dlm_deref_lockres_done_handler(struct o2net_msg
*msg
, u32 len
, void *data
,
2368 struct dlm_ctxt
*dlm
= data
;
2369 struct dlm_deref_lockres_done
*deref
2370 = (struct dlm_deref_lockres_done
*)msg
->buf
;
2371 struct dlm_lock_resource
*res
= NULL
;
2373 unsigned int namelen
;
2382 namelen
= deref
->namelen
;
2383 node
= deref
->node_idx
;
2385 if (namelen
> DLM_LOCKID_NAME_MAX
) {
2386 mlog(ML_ERROR
, "Invalid name length!");
2389 if (deref
->node_idx
>= O2NM_MAX_NODES
) {
2390 mlog(ML_ERROR
, "Invalid node number: %u\n", node
);
2394 hash
= dlm_lockid_hash(name
, namelen
);
2396 spin_lock(&dlm
->spinlock
);
2397 res
= __dlm_lookup_lockres_full(dlm
, name
, namelen
, hash
);
2399 spin_unlock(&dlm
->spinlock
);
2400 mlog(ML_ERROR
, "%s:%.*s: bad lockres name\n",
2401 dlm
->name
, namelen
, name
);
2405 spin_lock(&res
->spinlock
);
2406 if (!(res
->state
& DLM_LOCK_RES_DROPPING_REF
)) {
2407 spin_unlock(&res
->spinlock
);
2408 spin_unlock(&dlm
->spinlock
);
2409 mlog(ML_NOTICE
, "%s:%.*s: node %u sends deref done "
2410 "but it is already derefed!\n", dlm
->name
,
2411 res
->lockname
.len
, res
->lockname
.name
, node
);
2416 __dlm_do_purge_lockres(dlm
, res
);
2417 spin_unlock(&res
->spinlock
);
2420 spin_unlock(&dlm
->spinlock
);
2425 dlm_lockres_put(res
);
2430 static void dlm_drop_lockres_ref_done(struct dlm_ctxt
*dlm
,
2431 struct dlm_lock_resource
*res
, u8 node
)
2433 struct dlm_deref_lockres_done deref
;
2435 const char *lockname
;
2436 unsigned int namelen
;
2438 lockname
= res
->lockname
.name
;
2439 namelen
= res
->lockname
.len
;
2440 BUG_ON(namelen
> O2NM_MAX_NAME_LEN
);
2442 memset(&deref
, 0, sizeof(deref
));
2443 deref
.node_idx
= dlm
->node_num
;
2444 deref
.namelen
= namelen
;
2445 memcpy(deref
.name
, lockname
, namelen
);
2447 ret
= o2net_send_message(DLM_DEREF_LOCKRES_DONE
, dlm
->key
,
2448 &deref
, sizeof(deref
), node
, &r
);
2450 mlog(ML_ERROR
, "%s: res %.*s, error %d send DEREF DONE "
2451 " to node %u\n", dlm
->name
, namelen
,
2452 lockname
, ret
, node
);
2454 /* ignore the error */
2455 mlog(ML_ERROR
, "%s: res %.*s, DEREF to node %u got %d\n",
2456 dlm
->name
, namelen
, lockname
, node
, r
);
2457 dlm_print_one_lock_resource(res
);
2461 static void dlm_deref_lockres_worker(struct dlm_work_item
*item
, void *data
)
2463 struct dlm_ctxt
*dlm
;
2464 struct dlm_lock_resource
*res
;
2469 res
= item
->u
.dl
.deref_res
;
2470 node
= item
->u
.dl
.deref_node
;
2472 spin_lock(&res
->spinlock
);
2473 BUG_ON(res
->state
& DLM_LOCK_RES_DROPPING_REF
);
2474 __dlm_wait_on_lockres_flags(res
, DLM_LOCK_RES_SETREF_INPROG
);
2475 if (test_bit(node
, res
->refmap
)) {
2476 dlm_lockres_clear_refmap_bit(dlm
, res
, node
);
2479 spin_unlock(&res
->spinlock
);
2481 dlm_drop_lockres_ref_done(dlm
, res
, node
);
2484 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2485 dlm
->name
, res
->lockname
.len
, res
->lockname
.name
, node
);
2486 dlm_lockres_calc_usage(dlm
, res
);
2488 mlog(ML_ERROR
, "%s:%.*s: node %u trying to drop ref "
2489 "but it is already dropped!\n", dlm
->name
,
2490 res
->lockname
.len
, res
->lockname
.name
, node
);
2491 dlm_print_one_lock_resource(res
);
2494 dlm_lockres_put(res
);
2498 * A migrateable resource is one that is :
2499 * 1. locally mastered, and,
2500 * 2. zero local locks, and,
2501 * 3. one or more non-local locks, or, one or more references
2502 * Returns 1 if yes, 0 if not.
2504 static int dlm_is_lockres_migrateable(struct dlm_ctxt
*dlm
,
2505 struct dlm_lock_resource
*res
)
2507 enum dlm_lockres_list idx
;
2508 int nonlocal
= 0, node_ref
;
2509 struct list_head
*queue
;
2510 struct dlm_lock
*lock
;
2513 assert_spin_locked(&res
->spinlock
);
2515 /* delay migration when the lockres is in MIGRATING state */
2516 if (res
->state
& DLM_LOCK_RES_MIGRATING
)
2519 /* delay migration when the lockres is in RECOCERING state */
2520 if (res
->state
& (DLM_LOCK_RES_RECOVERING
|
2521 DLM_LOCK_RES_RECOVERY_WAITING
))
2524 if (res
->owner
!= dlm
->node_num
)
2527 for (idx
= DLM_GRANTED_LIST
; idx
<= DLM_BLOCKED_LIST
; idx
++) {
2528 queue
= dlm_list_idx_to_ptr(res
, idx
);
2529 list_for_each_entry(lock
, queue
, list
) {
2530 if (lock
->ml
.node
!= dlm
->node_num
) {
2534 cookie
= be64_to_cpu(lock
->ml
.cookie
);
2535 mlog(0, "%s: Not migrateable res %.*s, lock %u:%llu on "
2536 "%s list\n", dlm
->name
, res
->lockname
.len
,
2538 dlm_get_lock_cookie_node(cookie
),
2539 dlm_get_lock_cookie_seq(cookie
),
2540 dlm_list_in_text(idx
));
2546 node_ref
= find_next_bit(res
->refmap
, O2NM_MAX_NODES
, 0);
2547 if (node_ref
>= O2NM_MAX_NODES
)
2551 mlog(0, "%s: res %.*s, Migrateable\n", dlm
->name
, res
->lockname
.len
,
2552 res
->lockname
.name
);
2558 * DLM_MIGRATE_LOCKRES
2562 static int dlm_migrate_lockres(struct dlm_ctxt
*dlm
,
2563 struct dlm_lock_resource
*res
, u8 target
)
2565 struct dlm_master_list_entry
*mle
= NULL
;
2566 struct dlm_master_list_entry
*oldmle
= NULL
;
2567 struct dlm_migratable_lockres
*mres
= NULL
;
2570 unsigned int namelen
;
2577 BUG_ON(target
== O2NM_MAX_NODES
);
2579 name
= res
->lockname
.name
;
2580 namelen
= res
->lockname
.len
;
2582 mlog(0, "%s: Migrating %.*s to node %u\n", dlm
->name
, namelen
, name
,
2585 /* preallocate up front. if this fails, abort */
2587 mres
= (struct dlm_migratable_lockres
*) __get_free_page(GFP_NOFS
);
2593 mle
= kmem_cache_alloc(dlm_mle_cache
, GFP_NOFS
);
2601 * clear any existing master requests and
2602 * add the migration mle to the list
2604 spin_lock(&dlm
->spinlock
);
2605 spin_lock(&dlm
->master_lock
);
2606 ret
= dlm_add_migration_mle(dlm
, res
, mle
, &oldmle
, name
,
2607 namelen
, target
, dlm
->node_num
);
2608 /* get an extra reference on the mle.
2609 * otherwise the assert_master from the new
2610 * master will destroy this.
2613 dlm_get_mle_inuse(mle
);
2615 spin_unlock(&dlm
->master_lock
);
2616 spin_unlock(&dlm
->spinlock
);
2618 if (ret
== -EEXIST
) {
2619 mlog(0, "another process is already migrating it\n");
2625 * set the MIGRATING flag and flush asts
2626 * if we fail after this we need to re-dirty the lockres
2628 if (dlm_mark_lockres_migrating(dlm
, res
, target
) < 0) {
2629 mlog(ML_ERROR
, "tried to migrate %.*s to %u, but "
2630 "the target went down.\n", res
->lockname
.len
,
2631 res
->lockname
.name
, target
);
2632 spin_lock(&res
->spinlock
);
2633 res
->state
&= ~DLM_LOCK_RES_MIGRATING
;
2635 spin_unlock(&res
->spinlock
);
2640 if (ret
!= -EEXIST
&& oldmle
) {
2641 /* master is known, detach if not already detached */
2642 dlm_mle_detach_hb_events(dlm
, oldmle
);
2643 dlm_put_mle(oldmle
);
2648 dlm_mle_detach_hb_events(dlm
, mle
);
2650 dlm_put_mle_inuse(mle
);
2652 kmem_cache_free(dlm_mle_cache
, mle
);
2659 * at this point, we have a migration target, an mle
2660 * in the master list, and the MIGRATING flag set on
2664 /* now that remote nodes are spinning on the MIGRATING flag,
2665 * ensure that all assert_master work is flushed. */
2666 flush_workqueue(dlm
->dlm_worker
);
2668 /* notify new node and send all lock state */
2669 /* call send_one_lockres with migration flag.
2670 * this serves as notice to the target node that a
2671 * migration is starting. */
2672 ret
= dlm_send_one_lockres(dlm
, res
, mres
, target
,
2673 DLM_MRES_MIGRATION
);
2676 mlog(0, "migration to node %u failed with %d\n",
2678 /* migration failed, detach and clean up mle */
2679 dlm_mle_detach_hb_events(dlm
, mle
);
2681 dlm_put_mle_inuse(mle
);
2682 spin_lock(&res
->spinlock
);
2683 res
->state
&= ~DLM_LOCK_RES_MIGRATING
;
2685 spin_unlock(&res
->spinlock
);
2686 if (dlm_is_host_down(ret
))
2687 dlm_wait_for_node_death(dlm
, target
,
2688 DLM_NODE_DEATH_WAIT_MAX
);
2692 /* at this point, the target sends a message to all nodes,
2693 * (using dlm_do_migrate_request). this node is skipped since
2694 * we had to put an mle in the list to begin the process. this
2695 * node now waits for target to do an assert master. this node
2696 * will be the last one notified, ensuring that the migration
2697 * is complete everywhere. if the target dies while this is
2698 * going on, some nodes could potentially see the target as the
2699 * master, so it is important that my recovery finds the migration
2700 * mle and sets the master to UNKNOWN. */
2703 /* wait for new node to assert master */
2705 ret
= wait_event_interruptible_timeout(mle
->wq
,
2706 (atomic_read(&mle
->woken
) == 1),
2707 msecs_to_jiffies(5000));
2710 if (atomic_read(&mle
->woken
) == 1 ||
2711 res
->owner
== target
)
2714 mlog(0, "%s:%.*s: timed out during migration\n",
2715 dlm
->name
, res
->lockname
.len
, res
->lockname
.name
);
2716 /* avoid hang during shutdown when migrating lockres
2717 * to a node which also goes down */
2718 if (dlm_is_node_dead(dlm
, target
)) {
2719 mlog(0, "%s:%.*s: expected migration "
2720 "target %u is no longer up, restarting\n",
2721 dlm
->name
, res
->lockname
.len
,
2722 res
->lockname
.name
, target
);
2724 /* migration failed, detach and clean up mle */
2725 dlm_mle_detach_hb_events(dlm
, mle
);
2727 dlm_put_mle_inuse(mle
);
2728 spin_lock(&res
->spinlock
);
2729 res
->state
&= ~DLM_LOCK_RES_MIGRATING
;
2731 spin_unlock(&res
->spinlock
);
2735 mlog(0, "%s:%.*s: caught signal during migration\n",
2736 dlm
->name
, res
->lockname
.len
, res
->lockname
.name
);
2739 /* all done, set the owner, clear the flag */
2740 spin_lock(&res
->spinlock
);
2741 dlm_set_lockres_owner(dlm
, res
, target
);
2742 res
->state
&= ~DLM_LOCK_RES_MIGRATING
;
2743 dlm_remove_nonlocal_locks(dlm
, res
);
2744 spin_unlock(&res
->spinlock
);
2747 /* master is known, detach if not already detached */
2748 dlm_mle_detach_hb_events(dlm
, mle
);
2749 dlm_put_mle_inuse(mle
);
2752 dlm_lockres_calc_usage(dlm
, res
);
2755 /* re-dirty the lockres if we failed */
2757 dlm_kick_thread(dlm
, res
);
2759 /* wake up waiters if the MIGRATING flag got set
2760 * but migration failed */
2765 free_page((unsigned long)mres
);
2769 mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm
->name
, namelen
,
2774 #define DLM_MIGRATION_RETRY_MS 100
2777 * Should be called only after beginning the domain leave process.
2778 * There should not be any remaining locks on nonlocal lock resources,
2779 * and there should be no local locks left on locally mastered resources.
2781 * Called with the dlm spinlock held, may drop it to do migration, but
2782 * will re-acquire before exit.
2784 * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2786 int dlm_empty_lockres(struct dlm_ctxt
*dlm
, struct dlm_lock_resource
*res
)
2789 int lock_dropped
= 0;
2790 u8 target
= O2NM_MAX_NODES
;
2792 assert_spin_locked(&dlm
->spinlock
);
2794 spin_lock(&res
->spinlock
);
2795 if (dlm_is_lockres_migrateable(dlm
, res
))
2796 target
= dlm_pick_migration_target(dlm
, res
);
2797 spin_unlock(&res
->spinlock
);
2799 if (target
== O2NM_MAX_NODES
)
2802 /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2803 spin_unlock(&dlm
->spinlock
);
2805 ret
= dlm_migrate_lockres(dlm
, res
, target
);
2807 mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2808 dlm
->name
, res
->lockname
.len
, res
->lockname
.name
,
2810 spin_lock(&dlm
->spinlock
);
2812 return lock_dropped
;
2815 int dlm_lock_basts_flushed(struct dlm_ctxt
*dlm
, struct dlm_lock
*lock
)
2818 spin_lock(&dlm
->ast_lock
);
2819 spin_lock(&lock
->spinlock
);
2820 ret
= (list_empty(&lock
->bast_list
) && !lock
->bast_pending
);
2821 spin_unlock(&lock
->spinlock
);
2822 spin_unlock(&dlm
->ast_lock
);
2826 static int dlm_migration_can_proceed(struct dlm_ctxt
*dlm
,
2827 struct dlm_lock_resource
*res
,
2831 spin_lock(&res
->spinlock
);
2832 can_proceed
= !!(res
->state
& DLM_LOCK_RES_MIGRATING
);
2833 spin_unlock(&res
->spinlock
);
2835 /* target has died, so make the caller break out of the
2836 * wait_event, but caller must recheck the domain_map */
2837 spin_lock(&dlm
->spinlock
);
2838 if (!test_bit(mig_target
, dlm
->domain_map
))
2840 spin_unlock(&dlm
->spinlock
);
2844 static int dlm_lockres_is_dirty(struct dlm_ctxt
*dlm
,
2845 struct dlm_lock_resource
*res
)
2848 spin_lock(&res
->spinlock
);
2849 ret
= !!(res
->state
& DLM_LOCK_RES_DIRTY
);
2850 spin_unlock(&res
->spinlock
);
2855 static int dlm_mark_lockres_migrating(struct dlm_ctxt
*dlm
,
2856 struct dlm_lock_resource
*res
,
2861 mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2862 res
->lockname
.len
, res
->lockname
.name
, dlm
->node_num
,
2864 /* need to set MIGRATING flag on lockres. this is done by
2865 * ensuring that all asts have been flushed for this lockres. */
2866 spin_lock(&res
->spinlock
);
2867 BUG_ON(res
->migration_pending
);
2868 res
->migration_pending
= 1;
2869 /* strategy is to reserve an extra ast then release
2870 * it below, letting the release do all of the work */
2871 __dlm_lockres_reserve_ast(res
);
2872 spin_unlock(&res
->spinlock
);
2874 /* now flush all the pending asts */
2875 dlm_kick_thread(dlm
, res
);
2876 /* before waiting on DIRTY, block processes which may
2877 * try to dirty the lockres before MIGRATING is set */
2878 spin_lock(&res
->spinlock
);
2879 BUG_ON(res
->state
& DLM_LOCK_RES_BLOCK_DIRTY
);
2880 res
->state
|= DLM_LOCK_RES_BLOCK_DIRTY
;
2881 spin_unlock(&res
->spinlock
);
2882 /* now wait on any pending asts and the DIRTY state */
2883 wait_event(dlm
->ast_wq
, !dlm_lockres_is_dirty(dlm
, res
));
2884 dlm_lockres_release_ast(dlm
, res
);
2886 mlog(0, "about to wait on migration_wq, dirty=%s\n",
2887 res
->state
& DLM_LOCK_RES_DIRTY
? "yes" : "no");
2888 /* if the extra ref we just put was the final one, this
2889 * will pass thru immediately. otherwise, we need to wait
2890 * for the last ast to finish. */
2892 ret
= wait_event_interruptible_timeout(dlm
->migration_wq
,
2893 dlm_migration_can_proceed(dlm
, res
, target
),
2894 msecs_to_jiffies(1000));
2896 mlog(0, "woken again: migrating? %s, dead? %s\n",
2897 res
->state
& DLM_LOCK_RES_MIGRATING
? "yes":"no",
2898 test_bit(target
, dlm
->domain_map
) ? "no":"yes");
2900 mlog(0, "all is well: migrating? %s, dead? %s\n",
2901 res
->state
& DLM_LOCK_RES_MIGRATING
? "yes":"no",
2902 test_bit(target
, dlm
->domain_map
) ? "no":"yes");
2904 if (!dlm_migration_can_proceed(dlm
, res
, target
)) {
2905 mlog(0, "trying again...\n");
2910 /* did the target go down or die? */
2911 spin_lock(&dlm
->spinlock
);
2912 if (!test_bit(target
, dlm
->domain_map
)) {
2913 mlog(ML_ERROR
, "aha. migration target %u just went down\n",
2917 spin_unlock(&dlm
->spinlock
);
2920 * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2921 * another try; otherwise, we are sure the MIGRATING state is there,
2922 * drop the unneeded state which blocked threads trying to DIRTY
2924 spin_lock(&res
->spinlock
);
2925 BUG_ON(!(res
->state
& DLM_LOCK_RES_BLOCK_DIRTY
));
2926 res
->state
&= ~DLM_LOCK_RES_BLOCK_DIRTY
;
2928 BUG_ON(!(res
->state
& DLM_LOCK_RES_MIGRATING
));
2930 res
->migration_pending
= 0;
2931 spin_unlock(&res
->spinlock
);
2936 * o the DLM_LOCK_RES_MIGRATING flag is set if target not down
2937 * o there are no pending asts on this lockres
2938 * o all processes trying to reserve an ast on this
2939 * lockres must wait for the MIGRATING flag to clear
2944 /* last step in the migration process.
2945 * original master calls this to free all of the dlm_lock
2946 * structures that used to be for other nodes. */
2947 static void dlm_remove_nonlocal_locks(struct dlm_ctxt
*dlm
,
2948 struct dlm_lock_resource
*res
)
2950 struct list_head
*queue
= &res
->granted
;
2952 struct dlm_lock
*lock
, *next
;
2954 assert_spin_locked(&res
->spinlock
);
2956 BUG_ON(res
->owner
== dlm
->node_num
);
2958 for (i
=0; i
<3; i
++) {
2959 list_for_each_entry_safe(lock
, next
, queue
, list
) {
2960 if (lock
->ml
.node
!= dlm
->node_num
) {
2961 mlog(0, "putting lock for node %u\n",
2963 /* be extra careful */
2964 BUG_ON(!list_empty(&lock
->ast_list
));
2965 BUG_ON(!list_empty(&lock
->bast_list
));
2966 BUG_ON(lock
->ast_pending
);
2967 BUG_ON(lock
->bast_pending
);
2968 dlm_lockres_clear_refmap_bit(dlm
, res
,
2970 list_del_init(&lock
->list
);
2972 /* In a normal unlock, we would have added a
2973 * DLM_UNLOCK_FREE_LOCK action. Force it. */
2981 bit
= find_next_bit(res
->refmap
, O2NM_MAX_NODES
, bit
);
2982 if (bit
>= O2NM_MAX_NODES
)
2984 /* do not clear the local node reference, if there is a
2985 * process holding this, let it drop the ref itself */
2986 if (bit
!= dlm
->node_num
) {
2987 mlog(0, "%s:%.*s: node %u had a ref to this "
2988 "migrating lockres, clearing\n", dlm
->name
,
2989 res
->lockname
.len
, res
->lockname
.name
, bit
);
2990 dlm_lockres_clear_refmap_bit(dlm
, res
, bit
);
2997 * Pick a node to migrate the lock resource to. This function selects a
2998 * potential target based first on the locks and then on refmap. It skips
2999 * nodes that are in the process of exiting the domain.
3001 static u8
dlm_pick_migration_target(struct dlm_ctxt
*dlm
,
3002 struct dlm_lock_resource
*res
)
3004 enum dlm_lockres_list idx
;
3005 struct list_head
*queue
= &res
->granted
;
3006 struct dlm_lock
*lock
;
3008 u8 nodenum
= O2NM_MAX_NODES
;
3010 assert_spin_locked(&dlm
->spinlock
);
3011 assert_spin_locked(&res
->spinlock
);
3013 /* Go through all the locks */
3014 for (idx
= DLM_GRANTED_LIST
; idx
<= DLM_BLOCKED_LIST
; idx
++) {
3015 queue
= dlm_list_idx_to_ptr(res
, idx
);
3016 list_for_each_entry(lock
, queue
, list
) {
3017 if (lock
->ml
.node
== dlm
->node_num
)
3019 if (test_bit(lock
->ml
.node
, dlm
->exit_domain_map
))
3021 nodenum
= lock
->ml
.node
;
3026 /* Go thru the refmap */
3029 noderef
= find_next_bit(res
->refmap
, O2NM_MAX_NODES
,
3031 if (noderef
>= O2NM_MAX_NODES
)
3033 if (noderef
== dlm
->node_num
)
3035 if (test_bit(noderef
, dlm
->exit_domain_map
))
3045 /* this is called by the new master once all lockres
3046 * data has been received */
3047 static int dlm_do_migrate_request(struct dlm_ctxt
*dlm
,
3048 struct dlm_lock_resource
*res
,
3049 u8 master
, u8 new_master
,
3050 struct dlm_node_iter
*iter
)
3052 struct dlm_migrate_request migrate
;
3053 int ret
, skip
, status
= 0;
3056 memset(&migrate
, 0, sizeof(migrate
));
3057 migrate
.namelen
= res
->lockname
.len
;
3058 memcpy(migrate
.name
, res
->lockname
.name
, migrate
.namelen
);
3059 migrate
.new_master
= new_master
;
3060 migrate
.master
= master
;
3064 /* send message to all nodes, except the master and myself */
3065 while ((nodenum
= dlm_node_iter_next(iter
)) >= 0) {
3066 if (nodenum
== master
||
3067 nodenum
== new_master
)
3070 /* We could race exit domain. If exited, skip. */
3071 spin_lock(&dlm
->spinlock
);
3072 skip
= (!test_bit(nodenum
, dlm
->domain_map
));
3073 spin_unlock(&dlm
->spinlock
);
3075 clear_bit(nodenum
, iter
->node_map
);
3079 ret
= o2net_send_message(DLM_MIGRATE_REQUEST_MSG
, dlm
->key
,
3080 &migrate
, sizeof(migrate
), nodenum
,
3083 mlog(ML_ERROR
, "%s: res %.*s, Error %d send "
3084 "MIGRATE_REQUEST to node %u\n", dlm
->name
,
3085 migrate
.namelen
, migrate
.name
, ret
, nodenum
);
3086 if (!dlm_is_host_down(ret
)) {
3087 mlog(ML_ERROR
, "unhandled error=%d!\n", ret
);
3090 clear_bit(nodenum
, iter
->node_map
);
3092 } else if (status
< 0) {
3093 mlog(0, "migrate request (node %u) returned %d!\n",
3096 } else if (status
== DLM_MIGRATE_RESPONSE_MASTERY_REF
) {
3097 /* during the migration request we short-circuited
3098 * the mastery of the lockres. make sure we have
3099 * a mastery ref for nodenum */
3100 mlog(0, "%s:%.*s: need ref for node %u\n",
3101 dlm
->name
, res
->lockname
.len
, res
->lockname
.name
,
3103 spin_lock(&res
->spinlock
);
3104 dlm_lockres_set_refmap_bit(dlm
, res
, nodenum
);
3105 spin_unlock(&res
->spinlock
);
3112 mlog(0, "returning ret=%d\n", ret
);
3117 /* if there is an existing mle for this lockres, we now know who the master is.
3118 * (the one who sent us *this* message) we can clear it up right away.
3119 * since the process that put the mle on the list still has a reference to it,
3120 * we can unhash it now, set the master and wake the process. as a result,
3121 * we will have no mle in the list to start with. now we can add an mle for
3122 * the migration and this should be the only one found for those scanning the
3124 int dlm_migrate_request_handler(struct o2net_msg
*msg
, u32 len
, void *data
,
3127 struct dlm_ctxt
*dlm
= data
;
3128 struct dlm_lock_resource
*res
= NULL
;
3129 struct dlm_migrate_request
*migrate
= (struct dlm_migrate_request
*) msg
->buf
;
3130 struct dlm_master_list_entry
*mle
= NULL
, *oldmle
= NULL
;
3132 unsigned int namelen
, hash
;
3138 name
= migrate
->name
;
3139 namelen
= migrate
->namelen
;
3140 hash
= dlm_lockid_hash(name
, namelen
);
3142 /* preallocate.. if this fails, abort */
3143 mle
= kmem_cache_alloc(dlm_mle_cache
, GFP_NOFS
);
3150 /* check for pre-existing lock */
3151 spin_lock(&dlm
->spinlock
);
3152 res
= __dlm_lookup_lockres(dlm
, name
, namelen
, hash
);
3154 spin_lock(&res
->spinlock
);
3155 if (res
->state
& DLM_LOCK_RES_RECOVERING
) {
3156 /* if all is working ok, this can only mean that we got
3157 * a migrate request from a node that we now see as
3158 * dead. what can we do here? drop it to the floor? */
3159 spin_unlock(&res
->spinlock
);
3160 mlog(ML_ERROR
, "Got a migrate request, but the "
3161 "lockres is marked as recovering!");
3162 kmem_cache_free(dlm_mle_cache
, mle
);
3163 ret
= -EINVAL
; /* need a better solution */
3166 res
->state
|= DLM_LOCK_RES_MIGRATING
;
3167 spin_unlock(&res
->spinlock
);
3170 spin_lock(&dlm
->master_lock
);
3171 /* ignore status. only nonzero status would BUG. */
3172 ret
= dlm_add_migration_mle(dlm
, res
, mle
, &oldmle
,
3174 migrate
->new_master
,
3178 kmem_cache_free(dlm_mle_cache
, mle
);
3180 spin_unlock(&dlm
->master_lock
);
3182 spin_unlock(&dlm
->spinlock
);
3185 /* master is known, detach if not already detached */
3186 dlm_mle_detach_hb_events(dlm
, oldmle
);
3187 dlm_put_mle(oldmle
);
3191 dlm_lockres_put(res
);
3197 /* must be holding dlm->spinlock and dlm->master_lock
3198 * when adding a migration mle, we can clear any other mles
3199 * in the master list because we know with certainty that
3200 * the master is "master". so we remove any old mle from
3201 * the list after setting it's master field, and then add
3202 * the new migration mle. this way we can hold with the rule
3203 * of having only one mle for a given lock name at all times. */
3204 static int dlm_add_migration_mle(struct dlm_ctxt
*dlm
,
3205 struct dlm_lock_resource
*res
,
3206 struct dlm_master_list_entry
*mle
,
3207 struct dlm_master_list_entry
**oldmle
,
3208 const char *name
, unsigned int namelen
,
3209 u8 new_master
, u8 master
)
3216 assert_spin_locked(&dlm
->spinlock
);
3217 assert_spin_locked(&dlm
->master_lock
);
3219 /* caller is responsible for any ref taken here on oldmle */
3220 found
= dlm_find_mle(dlm
, oldmle
, (char *)name
, namelen
);
3222 struct dlm_master_list_entry
*tmp
= *oldmle
;
3223 spin_lock(&tmp
->spinlock
);
3224 if (tmp
->type
== DLM_MLE_MIGRATION
) {
3225 if (master
== dlm
->node_num
) {
3226 /* ah another process raced me to it */
3227 mlog(0, "tried to migrate %.*s, but some "
3228 "process beat me to it\n",
3230 spin_unlock(&tmp
->spinlock
);
3233 /* bad. 2 NODES are trying to migrate! */
3234 mlog(ML_ERROR
, "migration error mle: "
3235 "master=%u new_master=%u // request: "
3236 "master=%u new_master=%u // "
3238 tmp
->master
, tmp
->new_master
,
3244 /* this is essentially what assert_master does */
3245 tmp
->master
= master
;
3246 atomic_set(&tmp
->woken
, 1);
3248 /* remove it so that only one mle will be found */
3249 __dlm_unlink_mle(dlm
, tmp
);
3250 __dlm_mle_detach_hb_events(dlm
, tmp
);
3251 if (tmp
->type
== DLM_MLE_MASTER
) {
3252 ret
= DLM_MIGRATE_RESPONSE_MASTERY_REF
;
3253 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3254 "telling master to get ref "
3255 "for cleared out mle during "
3256 "migration\n", dlm
->name
,
3257 namelen
, name
, master
,
3261 spin_unlock(&tmp
->spinlock
);
3264 /* now add a migration mle to the tail of the list */
3265 dlm_init_mle(mle
, DLM_MLE_MIGRATION
, dlm
, res
, name
, namelen
);
3266 mle
->new_master
= new_master
;
3267 /* the new master will be sending an assert master for this.
3268 * at that point we will get the refmap reference */
3269 mle
->master
= master
;
3270 /* do this for consistency with other mle types */
3271 set_bit(new_master
, mle
->maybe_map
);
3272 __dlm_insert_mle(dlm
, mle
);
3278 * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3280 static struct dlm_lock_resource
*dlm_reset_mleres_owner(struct dlm_ctxt
*dlm
,
3281 struct dlm_master_list_entry
*mle
)
3283 struct dlm_lock_resource
*res
;
3285 /* Find the lockres associated to the mle and set its owner to UNK */
3286 res
= __dlm_lookup_lockres(dlm
, mle
->mname
, mle
->mnamelen
,
3289 spin_unlock(&dlm
->master_lock
);
3291 /* move lockres onto recovery list */
3292 spin_lock(&res
->spinlock
);
3293 dlm_set_lockres_owner(dlm
, res
, DLM_LOCK_RES_OWNER_UNKNOWN
);
3294 dlm_move_lockres_to_recovery_list(dlm
, res
);
3295 spin_unlock(&res
->spinlock
);
3296 dlm_lockres_put(res
);
3298 /* about to get rid of mle, detach from heartbeat */
3299 __dlm_mle_detach_hb_events(dlm
, mle
);
3302 spin_lock(&dlm
->master_lock
);
3304 spin_unlock(&dlm
->master_lock
);
3310 static void dlm_clean_migration_mle(struct dlm_ctxt
*dlm
,
3311 struct dlm_master_list_entry
*mle
)
3313 __dlm_mle_detach_hb_events(dlm
, mle
);
3315 spin_lock(&mle
->spinlock
);
3316 __dlm_unlink_mle(dlm
, mle
);
3317 atomic_set(&mle
->woken
, 1);
3318 spin_unlock(&mle
->spinlock
);
3323 static void dlm_clean_block_mle(struct dlm_ctxt
*dlm
,
3324 struct dlm_master_list_entry
*mle
, u8 dead_node
)
3328 BUG_ON(mle
->type
!= DLM_MLE_BLOCK
);
3330 spin_lock(&mle
->spinlock
);
3331 bit
= find_next_bit(mle
->maybe_map
, O2NM_MAX_NODES
, 0);
3332 if (bit
!= dead_node
) {
3333 mlog(0, "mle found, but dead node %u would not have been "
3334 "master\n", dead_node
);
3335 spin_unlock(&mle
->spinlock
);
3337 /* Must drop the refcount by one since the assert_master will
3338 * never arrive. This may result in the mle being unlinked and
3339 * freed, but there may still be a process waiting in the
3340 * dlmlock path which is fine. */
3341 mlog(0, "node %u was expected master\n", dead_node
);
3342 atomic_set(&mle
->woken
, 1);
3343 spin_unlock(&mle
->spinlock
);
3346 /* Do not need events any longer, so detach from heartbeat */
3347 __dlm_mle_detach_hb_events(dlm
, mle
);
3352 void dlm_clean_master_list(struct dlm_ctxt
*dlm
, u8 dead_node
)
3354 struct dlm_master_list_entry
*mle
;
3355 struct dlm_lock_resource
*res
;
3356 struct hlist_head
*bucket
;
3357 struct hlist_node
*tmp
;
3360 mlog(0, "dlm=%s, dead node=%u\n", dlm
->name
, dead_node
);
3362 assert_spin_locked(&dlm
->spinlock
);
3364 /* clean the master list */
3365 spin_lock(&dlm
->master_lock
);
3366 for (i
= 0; i
< DLM_HASH_BUCKETS
; i
++) {
3367 bucket
= dlm_master_hash(dlm
, i
);
3368 hlist_for_each_entry_safe(mle
, tmp
, bucket
, master_hash_node
) {
3369 BUG_ON(mle
->type
!= DLM_MLE_BLOCK
&&
3370 mle
->type
!= DLM_MLE_MASTER
&&
3371 mle
->type
!= DLM_MLE_MIGRATION
);
3373 /* MASTER mles are initiated locally. The waiting
3374 * process will notice the node map change shortly.
3375 * Let that happen as normal. */
3376 if (mle
->type
== DLM_MLE_MASTER
)
3379 /* BLOCK mles are initiated by other nodes. Need to
3380 * clean up if the dead node would have been the
3382 if (mle
->type
== DLM_MLE_BLOCK
) {
3383 dlm_clean_block_mle(dlm
, mle
, dead_node
);
3387 /* Everything else is a MIGRATION mle */
3389 /* The rule for MIGRATION mles is that the master
3390 * becomes UNKNOWN if *either* the original or the new
3391 * master dies. All UNKNOWN lockres' are sent to
3392 * whichever node becomes the recovery master. The new
3393 * master is responsible for determining if there is
3394 * still a master for this lockres, or if he needs to
3395 * take over mastery. Either way, this node should
3396 * expect another message to resolve this. */
3398 if (mle
->master
!= dead_node
&&
3399 mle
->new_master
!= dead_node
)
3402 if (mle
->new_master
== dead_node
&& mle
->inuse
) {
3403 mlog(ML_NOTICE
, "%s: target %u died during "
3404 "migration from %u, the MLE is "
3405 "still keep used, ignore it!\n",
3406 dlm
->name
, dead_node
,
3411 /* If we have reached this point, this mle needs to be
3412 * removed from the list and freed. */
3413 dlm_clean_migration_mle(dlm
, mle
);
3415 mlog(0, "%s: node %u died during migration from "
3416 "%u to %u!\n", dlm
->name
, dead_node
, mle
->master
,
3419 /* If we find a lockres associated with the mle, we've
3420 * hit this rare case that messes up our lock ordering.
3421 * If so, we need to drop the master lock so that we can
3422 * take the lockres lock, meaning that we will have to
3423 * restart from the head of list. */
3424 res
= dlm_reset_mleres_owner(dlm
, mle
);
3429 /* This may be the last reference */
3433 spin_unlock(&dlm
->master_lock
);
3436 int dlm_finish_migration(struct dlm_ctxt
*dlm
, struct dlm_lock_resource
*res
,
3439 struct dlm_node_iter iter
;
3442 spin_lock(&dlm
->spinlock
);
3443 dlm_node_iter_init(dlm
->domain_map
, &iter
);
3444 clear_bit(old_master
, iter
.node_map
);
3445 clear_bit(dlm
->node_num
, iter
.node_map
);
3446 spin_unlock(&dlm
->spinlock
);
3448 /* ownership of the lockres is changing. account for the
3449 * mastery reference here since old_master will briefly have
3450 * a reference after the migration completes */
3451 spin_lock(&res
->spinlock
);
3452 dlm_lockres_set_refmap_bit(dlm
, res
, old_master
);
3453 spin_unlock(&res
->spinlock
);
3455 mlog(0, "now time to do a migrate request to other nodes\n");
3456 ret
= dlm_do_migrate_request(dlm
, res
, old_master
,
3457 dlm
->node_num
, &iter
);
3463 mlog(0, "doing assert master of %.*s to all except the original node\n",
3464 res
->lockname
.len
, res
->lockname
.name
);
3465 /* this call now finishes out the nodemap
3466 * even if one or more nodes die */
3467 ret
= dlm_do_assert_master(dlm
, res
, iter
.node_map
,
3468 DLM_ASSERT_MASTER_FINISH_MIGRATION
);
3470 /* no longer need to retry. all living nodes contacted. */
3475 memset(iter
.node_map
, 0, sizeof(iter
.node_map
));
3476 set_bit(old_master
, iter
.node_map
);
3477 mlog(0, "doing assert master of %.*s back to %u\n",
3478 res
->lockname
.len
, res
->lockname
.name
, old_master
);
3479 ret
= dlm_do_assert_master(dlm
, res
, iter
.node_map
,
3480 DLM_ASSERT_MASTER_FINISH_MIGRATION
);
3482 mlog(0, "assert master to original master failed "
3484 /* the only nonzero status here would be because of
3485 * a dead original node. we're done. */
3489 /* all done, set the owner, clear the flag */
3490 spin_lock(&res
->spinlock
);
3491 dlm_set_lockres_owner(dlm
, res
, dlm
->node_num
);
3492 res
->state
&= ~DLM_LOCK_RES_MIGRATING
;
3493 spin_unlock(&res
->spinlock
);
3494 /* re-dirty it on the new master */
3495 dlm_kick_thread(dlm
, res
);
3502 * LOCKRES AST REFCOUNT
3503 * this is integral to migration
3506 /* for future intent to call an ast, reserve one ahead of time.
3507 * this should be called only after waiting on the lockres
3508 * with dlm_wait_on_lockres, and while still holding the
3509 * spinlock after the call. */
3510 void __dlm_lockres_reserve_ast(struct dlm_lock_resource
*res
)
3512 assert_spin_locked(&res
->spinlock
);
3513 if (res
->state
& DLM_LOCK_RES_MIGRATING
) {
3514 __dlm_print_one_lock_resource(res
);
3516 BUG_ON(res
->state
& DLM_LOCK_RES_MIGRATING
);
3518 atomic_inc(&res
->asts_reserved
);
3522 * used to drop the reserved ast, either because it went unused,
3523 * or because the ast/bast was actually called.
3525 * also, if there is a pending migration on this lockres,
3526 * and this was the last pending ast on the lockres,
3527 * atomically set the MIGRATING flag before we drop the lock.
3528 * this is how we ensure that migration can proceed with no
3529 * asts in progress. note that it is ok if the state of the
3530 * queues is such that a lock should be granted in the future
3531 * or that a bast should be fired, because the new master will
3532 * shuffle the lists on this lockres as soon as it is migrated.
3534 void dlm_lockres_release_ast(struct dlm_ctxt
*dlm
,
3535 struct dlm_lock_resource
*res
)
3537 if (!atomic_dec_and_lock(&res
->asts_reserved
, &res
->spinlock
))
3540 if (!res
->migration_pending
) {
3541 spin_unlock(&res
->spinlock
);
3545 BUG_ON(res
->state
& DLM_LOCK_RES_MIGRATING
);
3546 res
->migration_pending
= 0;
3547 res
->state
|= DLM_LOCK_RES_MIGRATING
;
3548 spin_unlock(&res
->spinlock
);
3550 wake_up(&dlm
->migration_wq
);
3553 void dlm_force_free_mles(struct dlm_ctxt
*dlm
)
3556 struct hlist_head
*bucket
;
3557 struct dlm_master_list_entry
*mle
;
3558 struct hlist_node
*tmp
;
3561 * We notified all other nodes that we are exiting the domain and
3562 * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3563 * around we force free them and wake any processes that are waiting
3566 spin_lock(&dlm
->spinlock
);
3567 spin_lock(&dlm
->master_lock
);
3569 BUG_ON(dlm
->dlm_state
!= DLM_CTXT_LEAVING
);
3570 BUG_ON((find_next_bit(dlm
->domain_map
, O2NM_MAX_NODES
, 0) < O2NM_MAX_NODES
));
3572 for (i
= 0; i
< DLM_HASH_BUCKETS
; i
++) {
3573 bucket
= dlm_master_hash(dlm
, i
);
3574 hlist_for_each_entry_safe(mle
, tmp
, bucket
, master_hash_node
) {
3575 if (mle
->type
!= DLM_MLE_BLOCK
) {
3576 mlog(ML_ERROR
, "bad mle: %p\n", mle
);
3577 dlm_print_one_mle(mle
);
3579 atomic_set(&mle
->woken
, 1);
3582 __dlm_unlink_mle(dlm
, mle
);
3583 __dlm_mle_detach_hb_events(dlm
, mle
);
3587 spin_unlock(&dlm
->master_lock
);
3588 spin_unlock(&dlm
->spinlock
);