Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/* -*- mode: c; c-basic-offset: 8; -*-
3 * vim: noexpandtab sw=8 ts=8 sts=0:
4 *
5 * dlmmod.c
6 *
7 * standalone DLM module
8 *
9 * Copyright (C) 2004 Oracle. All rights reserved.
10 */
11
12
13#include <linux/module.h>
14#include <linux/fs.h>
15#include <linux/types.h>
16#include <linux/slab.h>
17#include <linux/highmem.h>
18#include <linux/init.h>
19#include <linux/sysctl.h>
20#include <linux/random.h>
21#include <linux/blkdev.h>
22#include <linux/socket.h>
23#include <linux/inet.h>
24#include <linux/spinlock.h>
25#include <linux/delay.h>
26
27
28#include "cluster/heartbeat.h"
29#include "cluster/nodemanager.h"
30#include "cluster/tcp.h"
31
32#include "dlmapi.h"
33#include "dlmcommon.h"
34#include "dlmdomain.h"
35#include "dlmdebug.h"
36
37#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
38#include "cluster/masklog.h"
39
40static void dlm_mle_node_down(struct dlm_ctxt *dlm,
41 struct dlm_master_list_entry *mle,
42 struct o2nm_node *node,
43 int idx);
44static void dlm_mle_node_up(struct dlm_ctxt *dlm,
45 struct dlm_master_list_entry *mle,
46 struct o2nm_node *node,
47 int idx);
48
49static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
50static int dlm_do_assert_master(struct dlm_ctxt *dlm,
51 struct dlm_lock_resource *res,
52 void *nodemap, u32 flags);
53static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
54
55static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
56 struct dlm_master_list_entry *mle,
57 const char *name,
58 unsigned int namelen)
59{
60 if (dlm != mle->dlm)
61 return 0;
62
63 if (namelen != mle->mnamelen ||
64 memcmp(name, mle->mname, namelen) != 0)
65 return 0;
66
67 return 1;
68}
69
70static struct kmem_cache *dlm_lockres_cache;
71static struct kmem_cache *dlm_lockname_cache;
72static struct kmem_cache *dlm_mle_cache;
73
74static void dlm_mle_release(struct kref *kref);
75static void dlm_init_mle(struct dlm_master_list_entry *mle,
76 enum dlm_mle_type type,
77 struct dlm_ctxt *dlm,
78 struct dlm_lock_resource *res,
79 const char *name,
80 unsigned int namelen);
81static void dlm_put_mle(struct dlm_master_list_entry *mle);
82static void __dlm_put_mle(struct dlm_master_list_entry *mle);
83static int dlm_find_mle(struct dlm_ctxt *dlm,
84 struct dlm_master_list_entry **mle,
85 char *name, unsigned int namelen);
86
87static int dlm_do_master_request(struct dlm_lock_resource *res,
88 struct dlm_master_list_entry *mle, int to);
89
90
91static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
92 struct dlm_lock_resource *res,
93 struct dlm_master_list_entry *mle,
94 int *blocked);
95static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
96 struct dlm_lock_resource *res,
97 struct dlm_master_list_entry *mle,
98 int blocked);
99static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
100 struct dlm_lock_resource *res,
101 struct dlm_master_list_entry *mle,
102 struct dlm_master_list_entry **oldmle,
103 const char *name, unsigned int namelen,
104 u8 new_master, u8 master);
105
106static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
107 struct dlm_lock_resource *res);
108static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
109 struct dlm_lock_resource *res);
110static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
111 struct dlm_lock_resource *res,
112 u8 target);
113static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
114 struct dlm_lock_resource *res);
115
116
117int dlm_is_host_down(int errno)
118{
119 switch (errno) {
120 case -EBADF:
121 case -ECONNREFUSED:
122 case -ENOTCONN:
123 case -ECONNRESET:
124 case -EPIPE:
125 case -EHOSTDOWN:
126 case -EHOSTUNREACH:
127 case -ETIMEDOUT:
128 case -ECONNABORTED:
129 case -ENETDOWN:
130 case -ENETUNREACH:
131 case -ENETRESET:
132 case -ESHUTDOWN:
133 case -ENOPROTOOPT:
134 case -EINVAL: /* if returned from our tcp code,
135 this means there is no socket */
136 return 1;
137 }
138 return 0;
139}
140
141
142/*
143 * MASTER LIST FUNCTIONS
144 */
145
146
147/*
148 * regarding master list entries and heartbeat callbacks:
149 *
150 * in order to avoid sleeping and allocation that occurs in
151 * heartbeat, master list entries are simply attached to the
152 * dlm's established heartbeat callbacks. the mle is attached
153 * when it is created, and since the dlm->spinlock is held at
154 * that time, any heartbeat event will be properly discovered
155 * by the mle. the mle needs to be detached from the
156 * dlm->mle_hb_events list as soon as heartbeat events are no
157 * longer useful to the mle, and before the mle is freed.
158 *
159 * as a general rule, heartbeat events are no longer needed by
160 * the mle once an "answer" regarding the lock master has been
161 * received.
162 */
163static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
164 struct dlm_master_list_entry *mle)
165{
166 assert_spin_locked(&dlm->spinlock);
167
168 list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
169}
170
171
172static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
173 struct dlm_master_list_entry *mle)
174{
175 if (!list_empty(&mle->hb_events))
176 list_del_init(&mle->hb_events);
177}
178
179
180static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
181 struct dlm_master_list_entry *mle)
182{
183 spin_lock(&dlm->spinlock);
184 __dlm_mle_detach_hb_events(dlm, mle);
185 spin_unlock(&dlm->spinlock);
186}
187
188static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
189{
190 struct dlm_ctxt *dlm;
191 dlm = mle->dlm;
192
193 assert_spin_locked(&dlm->spinlock);
194 assert_spin_locked(&dlm->master_lock);
195 mle->inuse++;
196 kref_get(&mle->mle_refs);
197}
198
199static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
200{
201 struct dlm_ctxt *dlm;
202 dlm = mle->dlm;
203
204 spin_lock(&dlm->spinlock);
205 spin_lock(&dlm->master_lock);
206 mle->inuse--;
207 __dlm_put_mle(mle);
208 spin_unlock(&dlm->master_lock);
209 spin_unlock(&dlm->spinlock);
210
211}
212
213/* remove from list and free */
214static void __dlm_put_mle(struct dlm_master_list_entry *mle)
215{
216 struct dlm_ctxt *dlm;
217 dlm = mle->dlm;
218
219 assert_spin_locked(&dlm->spinlock);
220 assert_spin_locked(&dlm->master_lock);
221 if (!kref_read(&mle->mle_refs)) {
222 /* this may or may not crash, but who cares.
223 * it's a BUG. */
224 mlog(ML_ERROR, "bad mle: %p\n", mle);
225 dlm_print_one_mle(mle);
226 BUG();
227 } else
228 kref_put(&mle->mle_refs, dlm_mle_release);
229}
230
231
232/* must not have any spinlocks coming in */
233static void dlm_put_mle(struct dlm_master_list_entry *mle)
234{
235 struct dlm_ctxt *dlm;
236 dlm = mle->dlm;
237
238 spin_lock(&dlm->spinlock);
239 spin_lock(&dlm->master_lock);
240 __dlm_put_mle(mle);
241 spin_unlock(&dlm->master_lock);
242 spin_unlock(&dlm->spinlock);
243}
244
245static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
246{
247 kref_get(&mle->mle_refs);
248}
249
250static void dlm_init_mle(struct dlm_master_list_entry *mle,
251 enum dlm_mle_type type,
252 struct dlm_ctxt *dlm,
253 struct dlm_lock_resource *res,
254 const char *name,
255 unsigned int namelen)
256{
257 assert_spin_locked(&dlm->spinlock);
258
259 mle->dlm = dlm;
260 mle->type = type;
261 INIT_HLIST_NODE(&mle->master_hash_node);
262 INIT_LIST_HEAD(&mle->hb_events);
263 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
264 spin_lock_init(&mle->spinlock);
265 init_waitqueue_head(&mle->wq);
266 atomic_set(&mle->woken, 0);
267 kref_init(&mle->mle_refs);
268 memset(mle->response_map, 0, sizeof(mle->response_map));
269 mle->master = O2NM_MAX_NODES;
270 mle->new_master = O2NM_MAX_NODES;
271 mle->inuse = 0;
272
273 BUG_ON(mle->type != DLM_MLE_BLOCK &&
274 mle->type != DLM_MLE_MASTER &&
275 mle->type != DLM_MLE_MIGRATION);
276
277 if (mle->type == DLM_MLE_MASTER) {
278 BUG_ON(!res);
279 mle->mleres = res;
280 memcpy(mle->mname, res->lockname.name, res->lockname.len);
281 mle->mnamelen = res->lockname.len;
282 mle->mnamehash = res->lockname.hash;
283 } else {
284 BUG_ON(!name);
285 mle->mleres = NULL;
286 memcpy(mle->mname, name, namelen);
287 mle->mnamelen = namelen;
288 mle->mnamehash = dlm_lockid_hash(name, namelen);
289 }
290
291 atomic_inc(&dlm->mle_tot_count[mle->type]);
292 atomic_inc(&dlm->mle_cur_count[mle->type]);
293
294 /* copy off the node_map and register hb callbacks on our copy */
295 memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
296 memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
297 clear_bit(dlm->node_num, mle->vote_map);
298 clear_bit(dlm->node_num, mle->node_map);
299
300 /* attach the mle to the domain node up/down events */
301 __dlm_mle_attach_hb_events(dlm, mle);
302}
303
304void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
305{
306 assert_spin_locked(&dlm->spinlock);
307 assert_spin_locked(&dlm->master_lock);
308
309 if (!hlist_unhashed(&mle->master_hash_node))
310 hlist_del_init(&mle->master_hash_node);
311}
312
313void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
314{
315 struct hlist_head *bucket;
316
317 assert_spin_locked(&dlm->master_lock);
318
319 bucket = dlm_master_hash(dlm, mle->mnamehash);
320 hlist_add_head(&mle->master_hash_node, bucket);
321}
322
323/* returns 1 if found, 0 if not */
324static int dlm_find_mle(struct dlm_ctxt *dlm,
325 struct dlm_master_list_entry **mle,
326 char *name, unsigned int namelen)
327{
328 struct dlm_master_list_entry *tmpmle;
329 struct hlist_head *bucket;
330 unsigned int hash;
331
332 assert_spin_locked(&dlm->master_lock);
333
334 hash = dlm_lockid_hash(name, namelen);
335 bucket = dlm_master_hash(dlm, hash);
336 hlist_for_each_entry(tmpmle, bucket, master_hash_node) {
337 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
338 continue;
339 dlm_get_mle(tmpmle);
340 *mle = tmpmle;
341 return 1;
342 }
343 return 0;
344}
345
346void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
347{
348 struct dlm_master_list_entry *mle;
349
350 assert_spin_locked(&dlm->spinlock);
351
352 list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
353 if (node_up)
354 dlm_mle_node_up(dlm, mle, NULL, idx);
355 else
356 dlm_mle_node_down(dlm, mle, NULL, idx);
357 }
358}
359
360static void dlm_mle_node_down(struct dlm_ctxt *dlm,
361 struct dlm_master_list_entry *mle,
362 struct o2nm_node *node, int idx)
363{
364 spin_lock(&mle->spinlock);
365
366 if (!test_bit(idx, mle->node_map))
367 mlog(0, "node %u already removed from nodemap!\n", idx);
368 else
369 clear_bit(idx, mle->node_map);
370
371 spin_unlock(&mle->spinlock);
372}
373
374static void dlm_mle_node_up(struct dlm_ctxt *dlm,
375 struct dlm_master_list_entry *mle,
376 struct o2nm_node *node, int idx)
377{
378 spin_lock(&mle->spinlock);
379
380 if (test_bit(idx, mle->node_map))
381 mlog(0, "node %u already in node map!\n", idx);
382 else
383 set_bit(idx, mle->node_map);
384
385 spin_unlock(&mle->spinlock);
386}
387
388
389int dlm_init_mle_cache(void)
390{
391 dlm_mle_cache = kmem_cache_create("o2dlm_mle",
392 sizeof(struct dlm_master_list_entry),
393 0, SLAB_HWCACHE_ALIGN,
394 NULL);
395 if (dlm_mle_cache == NULL)
396 return -ENOMEM;
397 return 0;
398}
399
400void dlm_destroy_mle_cache(void)
401{
402 kmem_cache_destroy(dlm_mle_cache);
403}
404
405static void dlm_mle_release(struct kref *kref)
406{
407 struct dlm_master_list_entry *mle;
408 struct dlm_ctxt *dlm;
409
410 mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
411 dlm = mle->dlm;
412
413 assert_spin_locked(&dlm->spinlock);
414 assert_spin_locked(&dlm->master_lock);
415
416 mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
417 mle->type);
418
419 /* remove from list if not already */
420 __dlm_unlink_mle(dlm, mle);
421
422 /* detach the mle from the domain node up/down events */
423 __dlm_mle_detach_hb_events(dlm, mle);
424
425 atomic_dec(&dlm->mle_cur_count[mle->type]);
426
427 /* NOTE: kfree under spinlock here.
428 * if this is bad, we can move this to a freelist. */
429 kmem_cache_free(dlm_mle_cache, mle);
430}
431
432
433/*
434 * LOCK RESOURCE FUNCTIONS
435 */
436
437int dlm_init_master_caches(void)
438{
439 dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
440 sizeof(struct dlm_lock_resource),
441 0, SLAB_HWCACHE_ALIGN, NULL);
442 if (!dlm_lockres_cache)
443 goto bail;
444
445 dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
446 DLM_LOCKID_NAME_MAX, 0,
447 SLAB_HWCACHE_ALIGN, NULL);
448 if (!dlm_lockname_cache)
449 goto bail;
450
451 return 0;
452bail:
453 dlm_destroy_master_caches();
454 return -ENOMEM;
455}
456
457void dlm_destroy_master_caches(void)
458{
459 kmem_cache_destroy(dlm_lockname_cache);
460 dlm_lockname_cache = NULL;
461
462 kmem_cache_destroy(dlm_lockres_cache);
463 dlm_lockres_cache = NULL;
464}
465
466static void dlm_lockres_release(struct kref *kref)
467{
468 struct dlm_lock_resource *res;
469 struct dlm_ctxt *dlm;
470
471 res = container_of(kref, struct dlm_lock_resource, refs);
472 dlm = res->dlm;
473
474 /* This should not happen -- all lockres' have a name
475 * associated with them at init time. */
476 BUG_ON(!res->lockname.name);
477
478 mlog(0, "destroying lockres %.*s\n", res->lockname.len,
479 res->lockname.name);
480
481 atomic_dec(&dlm->res_cur_count);
482
483 if (!hlist_unhashed(&res->hash_node) ||
484 !list_empty(&res->granted) ||
485 !list_empty(&res->converting) ||
486 !list_empty(&res->blocked) ||
487 !list_empty(&res->dirty) ||
488 !list_empty(&res->recovering) ||
489 !list_empty(&res->purge)) {
490 mlog(ML_ERROR,
491 "Going to BUG for resource %.*s."
492 " We're on a list! [%c%c%c%c%c%c%c]\n",
493 res->lockname.len, res->lockname.name,
494 !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
495 !list_empty(&res->granted) ? 'G' : ' ',
496 !list_empty(&res->converting) ? 'C' : ' ',
497 !list_empty(&res->blocked) ? 'B' : ' ',
498 !list_empty(&res->dirty) ? 'D' : ' ',
499 !list_empty(&res->recovering) ? 'R' : ' ',
500 !list_empty(&res->purge) ? 'P' : ' ');
501
502 dlm_print_one_lock_resource(res);
503 }
504
505 /* By the time we're ready to blow this guy away, we shouldn't
506 * be on any lists. */
507 BUG_ON(!hlist_unhashed(&res->hash_node));
508 BUG_ON(!list_empty(&res->granted));
509 BUG_ON(!list_empty(&res->converting));
510 BUG_ON(!list_empty(&res->blocked));
511 BUG_ON(!list_empty(&res->dirty));
512 BUG_ON(!list_empty(&res->recovering));
513 BUG_ON(!list_empty(&res->purge));
514
515 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
516
517 kmem_cache_free(dlm_lockres_cache, res);
518}
519
520void dlm_lockres_put(struct dlm_lock_resource *res)
521{
522 kref_put(&res->refs, dlm_lockres_release);
523}
524
525static void dlm_init_lockres(struct dlm_ctxt *dlm,
526 struct dlm_lock_resource *res,
527 const char *name, unsigned int namelen)
528{
529 char *qname;
530
531 /* If we memset here, we lose our reference to the kmalloc'd
532 * res->lockname.name, so be sure to init every field
533 * correctly! */
534
535 qname = (char *) res->lockname.name;
536 memcpy(qname, name, namelen);
537
538 res->lockname.len = namelen;
539 res->lockname.hash = dlm_lockid_hash(name, namelen);
540
541 init_waitqueue_head(&res->wq);
542 spin_lock_init(&res->spinlock);
543 INIT_HLIST_NODE(&res->hash_node);
544 INIT_LIST_HEAD(&res->granted);
545 INIT_LIST_HEAD(&res->converting);
546 INIT_LIST_HEAD(&res->blocked);
547 INIT_LIST_HEAD(&res->dirty);
548 INIT_LIST_HEAD(&res->recovering);
549 INIT_LIST_HEAD(&res->purge);
550 INIT_LIST_HEAD(&res->tracking);
551 atomic_set(&res->asts_reserved, 0);
552 res->migration_pending = 0;
553 res->inflight_locks = 0;
554 res->inflight_assert_workers = 0;
555
556 res->dlm = dlm;
557
558 kref_init(&res->refs);
559
560 atomic_inc(&dlm->res_tot_count);
561 atomic_inc(&dlm->res_cur_count);
562
563 /* just for consistency */
564 spin_lock(&res->spinlock);
565 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
566 spin_unlock(&res->spinlock);
567
568 res->state = DLM_LOCK_RES_IN_PROGRESS;
569
570 res->last_used = 0;
571
572 spin_lock(&dlm->track_lock);
573 list_add_tail(&res->tracking, &dlm->tracking_list);
574 spin_unlock(&dlm->track_lock);
575
576 memset(res->lvb, 0, DLM_LVB_LEN);
577 memset(res->refmap, 0, sizeof(res->refmap));
578}
579
580struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
581 const char *name,
582 unsigned int namelen)
583{
584 struct dlm_lock_resource *res = NULL;
585
586 res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
587 if (!res)
588 goto error;
589
590 res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
591 if (!res->lockname.name)
592 goto error;
593
594 dlm_init_lockres(dlm, res, name, namelen);
595 return res;
596
597error:
598 if (res)
599 kmem_cache_free(dlm_lockres_cache, res);
600 return NULL;
601}
602
603void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm,
604 struct dlm_lock_resource *res, int bit)
605{
606 assert_spin_locked(&res->spinlock);
607
608 mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len,
609 res->lockname.name, bit, __builtin_return_address(0));
610
611 set_bit(bit, res->refmap);
612}
613
614void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm,
615 struct dlm_lock_resource *res, int bit)
616{
617 assert_spin_locked(&res->spinlock);
618
619 mlog(0, "res %.*s, clr node %u, %ps()\n", res->lockname.len,
620 res->lockname.name, bit, __builtin_return_address(0));
621
622 clear_bit(bit, res->refmap);
623}
624
625static void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
626 struct dlm_lock_resource *res)
627{
628 res->inflight_locks++;
629
630 mlog(0, "%s: res %.*s, inflight++: now %u, %ps()\n", dlm->name,
631 res->lockname.len, res->lockname.name, res->inflight_locks,
632 __builtin_return_address(0));
633}
634
635void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
636 struct dlm_lock_resource *res)
637{
638 assert_spin_locked(&res->spinlock);
639 __dlm_lockres_grab_inflight_ref(dlm, res);
640}
641
642void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
643 struct dlm_lock_resource *res)
644{
645 assert_spin_locked(&res->spinlock);
646
647 BUG_ON(res->inflight_locks == 0);
648
649 res->inflight_locks--;
650
651 mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm->name,
652 res->lockname.len, res->lockname.name, res->inflight_locks,
653 __builtin_return_address(0));
654
655 wake_up(&res->wq);
656}
657
658void __dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm,
659 struct dlm_lock_resource *res)
660{
661 assert_spin_locked(&res->spinlock);
662 res->inflight_assert_workers++;
663 mlog(0, "%s:%.*s: inflight assert worker++: now %u\n",
664 dlm->name, res->lockname.len, res->lockname.name,
665 res->inflight_assert_workers);
666}
667
668static void __dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
669 struct dlm_lock_resource *res)
670{
671 assert_spin_locked(&res->spinlock);
672 BUG_ON(res->inflight_assert_workers == 0);
673 res->inflight_assert_workers--;
674 mlog(0, "%s:%.*s: inflight assert worker--: now %u\n",
675 dlm->name, res->lockname.len, res->lockname.name,
676 res->inflight_assert_workers);
677}
678
679static void dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
680 struct dlm_lock_resource *res)
681{
682 spin_lock(&res->spinlock);
683 __dlm_lockres_drop_inflight_worker(dlm, res);
684 spin_unlock(&res->spinlock);
685}
686
687/*
688 * lookup a lock resource by name.
689 * may already exist in the hashtable.
690 * lockid is null terminated
691 *
692 * if not, allocate enough for the lockres and for
693 * the temporary structure used in doing the mastering.
694 *
695 * also, do a lookup in the dlm->master_list to see
696 * if another node has begun mastering the same lock.
697 * if so, there should be a block entry in there
698 * for this name, and we should *not* attempt to master
699 * the lock here. need to wait around for that node
700 * to assert_master (or die).
701 *
702 */
703struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
704 const char *lockid,
705 int namelen,
706 int flags)
707{
708 struct dlm_lock_resource *tmpres=NULL, *res=NULL;
709 struct dlm_master_list_entry *mle = NULL;
710 struct dlm_master_list_entry *alloc_mle = NULL;
711 int blocked = 0;
712 int ret, nodenum;
713 struct dlm_node_iter iter;
714 unsigned int hash;
715 int tries = 0;
716 int bit, wait_on_recovery = 0;
717
718 BUG_ON(!lockid);
719
720 hash = dlm_lockid_hash(lockid, namelen);
721
722 mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
723
724lookup:
725 spin_lock(&dlm->spinlock);
726 tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
727 if (tmpres) {
728 spin_unlock(&dlm->spinlock);
729 spin_lock(&tmpres->spinlock);
730
731 /*
732 * Right after dlm spinlock was released, dlm_thread could have
733 * purged the lockres. Check if lockres got unhashed. If so
734 * start over.
735 */
736 if (hlist_unhashed(&tmpres->hash_node)) {
737 spin_unlock(&tmpres->spinlock);
738 dlm_lockres_put(tmpres);
739 tmpres = NULL;
740 goto lookup;
741 }
742
743 /* Wait on the thread that is mastering the resource */
744 if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
745 __dlm_wait_on_lockres(tmpres);
746 BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
747 spin_unlock(&tmpres->spinlock);
748 dlm_lockres_put(tmpres);
749 tmpres = NULL;
750 goto lookup;
751 }
752
753 /* Wait on the resource purge to complete before continuing */
754 if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) {
755 BUG_ON(tmpres->owner == dlm->node_num);
756 __dlm_wait_on_lockres_flags(tmpres,
757 DLM_LOCK_RES_DROPPING_REF);
758 spin_unlock(&tmpres->spinlock);
759 dlm_lockres_put(tmpres);
760 tmpres = NULL;
761 goto lookup;
762 }
763
764 /* Grab inflight ref to pin the resource */
765 dlm_lockres_grab_inflight_ref(dlm, tmpres);
766
767 spin_unlock(&tmpres->spinlock);
768 if (res) {
769 spin_lock(&dlm->track_lock);
770 if (!list_empty(&res->tracking))
771 list_del_init(&res->tracking);
772 else
773 mlog(ML_ERROR, "Resource %.*s not "
774 "on the Tracking list\n",
775 res->lockname.len,
776 res->lockname.name);
777 spin_unlock(&dlm->track_lock);
778 dlm_lockres_put(res);
779 }
780 res = tmpres;
781 goto leave;
782 }
783
784 if (!res) {
785 spin_unlock(&dlm->spinlock);
786 mlog(0, "allocating a new resource\n");
787 /* nothing found and we need to allocate one. */
788 alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
789 if (!alloc_mle)
790 goto leave;
791 res = dlm_new_lockres(dlm, lockid, namelen);
792 if (!res)
793 goto leave;
794 goto lookup;
795 }
796
797 mlog(0, "no lockres found, allocated our own: %p\n", res);
798
799 if (flags & LKM_LOCAL) {
800 /* caller knows it's safe to assume it's not mastered elsewhere
801 * DONE! return right away */
802 spin_lock(&res->spinlock);
803 dlm_change_lockres_owner(dlm, res, dlm->node_num);
804 __dlm_insert_lockres(dlm, res);
805 dlm_lockres_grab_inflight_ref(dlm, res);
806 spin_unlock(&res->spinlock);
807 spin_unlock(&dlm->spinlock);
808 /* lockres still marked IN_PROGRESS */
809 goto wake_waiters;
810 }
811
812 /* check master list to see if another node has started mastering it */
813 spin_lock(&dlm->master_lock);
814
815 /* if we found a block, wait for lock to be mastered by another node */
816 blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
817 if (blocked) {
818 int mig;
819 if (mle->type == DLM_MLE_MASTER) {
820 mlog(ML_ERROR, "master entry for nonexistent lock!\n");
821 BUG();
822 }
823 mig = (mle->type == DLM_MLE_MIGRATION);
824 /* if there is a migration in progress, let the migration
825 * finish before continuing. we can wait for the absence
826 * of the MIGRATION mle: either the migrate finished or
827 * one of the nodes died and the mle was cleaned up.
828 * if there is a BLOCK here, but it already has a master
829 * set, we are too late. the master does not have a ref
830 * for us in the refmap. detach the mle and drop it.
831 * either way, go back to the top and start over. */
832 if (mig || mle->master != O2NM_MAX_NODES) {
833 BUG_ON(mig && mle->master == dlm->node_num);
834 /* we arrived too late. the master does not
835 * have a ref for us. retry. */
836 mlog(0, "%s:%.*s: late on %s\n",
837 dlm->name, namelen, lockid,
838 mig ? "MIGRATION" : "BLOCK");
839 spin_unlock(&dlm->master_lock);
840 spin_unlock(&dlm->spinlock);
841
842 /* master is known, detach */
843 if (!mig)
844 dlm_mle_detach_hb_events(dlm, mle);
845 dlm_put_mle(mle);
846 mle = NULL;
847 /* this is lame, but we can't wait on either
848 * the mle or lockres waitqueue here */
849 if (mig)
850 msleep(100);
851 goto lookup;
852 }
853 } else {
854 /* go ahead and try to master lock on this node */
855 mle = alloc_mle;
856 /* make sure this does not get freed below */
857 alloc_mle = NULL;
858 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
859 set_bit(dlm->node_num, mle->maybe_map);
860 __dlm_insert_mle(dlm, mle);
861
862 /* still holding the dlm spinlock, check the recovery map
863 * to see if there are any nodes that still need to be
864 * considered. these will not appear in the mle nodemap
865 * but they might own this lockres. wait on them. */
866 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
867 if (bit < O2NM_MAX_NODES) {
868 mlog(0, "%s: res %.*s, At least one node (%d) "
869 "to recover before lock mastery can begin\n",
870 dlm->name, namelen, (char *)lockid, bit);
871 wait_on_recovery = 1;
872 }
873 }
874
875 /* at this point there is either a DLM_MLE_BLOCK or a
876 * DLM_MLE_MASTER on the master list, so it's safe to add the
877 * lockres to the hashtable. anyone who finds the lock will
878 * still have to wait on the IN_PROGRESS. */
879
880 /* finally add the lockres to its hash bucket */
881 __dlm_insert_lockres(dlm, res);
882
883 /* since this lockres is new it doesn't not require the spinlock */
884 __dlm_lockres_grab_inflight_ref(dlm, res);
885
886 /* get an extra ref on the mle in case this is a BLOCK
887 * if so, the creator of the BLOCK may try to put the last
888 * ref at this time in the assert master handler, so we
889 * need an extra one to keep from a bad ptr deref. */
890 dlm_get_mle_inuse(mle);
891 spin_unlock(&dlm->master_lock);
892 spin_unlock(&dlm->spinlock);
893
894redo_request:
895 while (wait_on_recovery) {
896 /* any cluster changes that occurred after dropping the
897 * dlm spinlock would be detectable be a change on the mle,
898 * so we only need to clear out the recovery map once. */
899 if (dlm_is_recovery_lock(lockid, namelen)) {
900 mlog(0, "%s: Recovery map is not empty, but must "
901 "master $RECOVERY lock now\n", dlm->name);
902 if (!dlm_pre_master_reco_lockres(dlm, res))
903 wait_on_recovery = 0;
904 else {
905 mlog(0, "%s: waiting 500ms for heartbeat state "
906 "change\n", dlm->name);
907 msleep(500);
908 }
909 continue;
910 }
911
912 dlm_kick_recovery_thread(dlm);
913 msleep(1000);
914 dlm_wait_for_recovery(dlm);
915
916 spin_lock(&dlm->spinlock);
917 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
918 if (bit < O2NM_MAX_NODES) {
919 mlog(0, "%s: res %.*s, At least one node (%d) "
920 "to recover before lock mastery can begin\n",
921 dlm->name, namelen, (char *)lockid, bit);
922 wait_on_recovery = 1;
923 } else
924 wait_on_recovery = 0;
925 spin_unlock(&dlm->spinlock);
926
927 if (wait_on_recovery)
928 dlm_wait_for_node_recovery(dlm, bit, 10000);
929 }
930
931 /* must wait for lock to be mastered elsewhere */
932 if (blocked)
933 goto wait;
934
935 ret = -EINVAL;
936 dlm_node_iter_init(mle->vote_map, &iter);
937 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
938 ret = dlm_do_master_request(res, mle, nodenum);
939 if (ret < 0)
940 mlog_errno(ret);
941 if (mle->master != O2NM_MAX_NODES) {
942 /* found a master ! */
943 if (mle->master <= nodenum)
944 break;
945 /* if our master request has not reached the master
946 * yet, keep going until it does. this is how the
947 * master will know that asserts are needed back to
948 * the lower nodes. */
949 mlog(0, "%s: res %.*s, Requests only up to %u but "
950 "master is %u, keep going\n", dlm->name, namelen,
951 lockid, nodenum, mle->master);
952 }
953 }
954
955wait:
956 /* keep going until the response map includes all nodes */
957 ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
958 if (ret < 0) {
959 wait_on_recovery = 1;
960 mlog(0, "%s: res %.*s, Node map changed, redo the master "
961 "request now, blocked=%d\n", dlm->name, res->lockname.len,
962 res->lockname.name, blocked);
963 if (++tries > 20) {
964 mlog(ML_ERROR, "%s: res %.*s, Spinning on "
965 "dlm_wait_for_lock_mastery, blocked = %d\n",
966 dlm->name, res->lockname.len,
967 res->lockname.name, blocked);
968 dlm_print_one_lock_resource(res);
969 dlm_print_one_mle(mle);
970 tries = 0;
971 }
972 goto redo_request;
973 }
974
975 mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len,
976 res->lockname.name, res->owner);
977 /* make sure we never continue without this */
978 BUG_ON(res->owner == O2NM_MAX_NODES);
979
980 /* master is known, detach if not already detached */
981 dlm_mle_detach_hb_events(dlm, mle);
982 dlm_put_mle(mle);
983 /* put the extra ref */
984 dlm_put_mle_inuse(mle);
985
986wake_waiters:
987 spin_lock(&res->spinlock);
988 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
989 spin_unlock(&res->spinlock);
990 wake_up(&res->wq);
991
992leave:
993 /* need to free the unused mle */
994 if (alloc_mle)
995 kmem_cache_free(dlm_mle_cache, alloc_mle);
996
997 return res;
998}
999
1000
1001#define DLM_MASTERY_TIMEOUT_MS 5000
1002
1003static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
1004 struct dlm_lock_resource *res,
1005 struct dlm_master_list_entry *mle,
1006 int *blocked)
1007{
1008 u8 m;
1009 int ret, bit;
1010 int map_changed, voting_done;
1011 int assert, sleep;
1012
1013recheck:
1014 ret = 0;
1015 assert = 0;
1016
1017 /* check if another node has already become the owner */
1018 spin_lock(&res->spinlock);
1019 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1020 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1021 res->lockname.len, res->lockname.name, res->owner);
1022 spin_unlock(&res->spinlock);
1023 /* this will cause the master to re-assert across
1024 * the whole cluster, freeing up mles */
1025 if (res->owner != dlm->node_num) {
1026 ret = dlm_do_master_request(res, mle, res->owner);
1027 if (ret < 0) {
1028 /* give recovery a chance to run */
1029 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1030 msleep(500);
1031 goto recheck;
1032 }
1033 }
1034 ret = 0;
1035 goto leave;
1036 }
1037 spin_unlock(&res->spinlock);
1038
1039 spin_lock(&mle->spinlock);
1040 m = mle->master;
1041 map_changed = (memcmp(mle->vote_map, mle->node_map,
1042 sizeof(mle->vote_map)) != 0);
1043 voting_done = (memcmp(mle->vote_map, mle->response_map,
1044 sizeof(mle->vote_map)) == 0);
1045
1046 /* restart if we hit any errors */
1047 if (map_changed) {
1048 int b;
1049 mlog(0, "%s: %.*s: node map changed, restarting\n",
1050 dlm->name, res->lockname.len, res->lockname.name);
1051 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1052 b = (mle->type == DLM_MLE_BLOCK);
1053 if ((*blocked && !b) || (!*blocked && b)) {
1054 mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1055 dlm->name, res->lockname.len, res->lockname.name,
1056 *blocked, b);
1057 *blocked = b;
1058 }
1059 spin_unlock(&mle->spinlock);
1060 if (ret < 0) {
1061 mlog_errno(ret);
1062 goto leave;
1063 }
1064 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1065 "rechecking now\n", dlm->name, res->lockname.len,
1066 res->lockname.name);
1067 goto recheck;
1068 } else {
1069 if (!voting_done) {
1070 mlog(0, "map not changed and voting not done "
1071 "for %s:%.*s\n", dlm->name, res->lockname.len,
1072 res->lockname.name);
1073 }
1074 }
1075
1076 if (m != O2NM_MAX_NODES) {
1077 /* another node has done an assert!
1078 * all done! */
1079 sleep = 0;
1080 } else {
1081 sleep = 1;
1082 /* have all nodes responded? */
1083 if (voting_done && !*blocked) {
1084 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1085 if (dlm->node_num <= bit) {
1086 /* my node number is lowest.
1087 * now tell other nodes that I am
1088 * mastering this. */
1089 mle->master = dlm->node_num;
1090 /* ref was grabbed in get_lock_resource
1091 * will be dropped in dlmlock_master */
1092 assert = 1;
1093 sleep = 0;
1094 }
1095 /* if voting is done, but we have not received
1096 * an assert master yet, we must sleep */
1097 }
1098 }
1099
1100 spin_unlock(&mle->spinlock);
1101
1102 /* sleep if we haven't finished voting yet */
1103 if (sleep) {
1104 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1105 atomic_set(&mle->woken, 0);
1106 (void)wait_event_timeout(mle->wq,
1107 (atomic_read(&mle->woken) == 1),
1108 timeo);
1109 if (res->owner == O2NM_MAX_NODES) {
1110 mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1111 res->lockname.len, res->lockname.name);
1112 goto recheck;
1113 }
1114 mlog(0, "done waiting, master is %u\n", res->owner);
1115 ret = 0;
1116 goto leave;
1117 }
1118
1119 ret = 0; /* done */
1120 if (assert) {
1121 m = dlm->node_num;
1122 mlog(0, "about to master %.*s here, this=%u\n",
1123 res->lockname.len, res->lockname.name, m);
1124 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1125 if (ret) {
1126 /* This is a failure in the network path,
1127 * not in the response to the assert_master
1128 * (any nonzero response is a BUG on this node).
1129 * Most likely a socket just got disconnected
1130 * due to node death. */
1131 mlog_errno(ret);
1132 }
1133 /* no longer need to restart lock mastery.
1134 * all living nodes have been contacted. */
1135 ret = 0;
1136 }
1137
1138 /* set the lockres owner */
1139 spin_lock(&res->spinlock);
1140 /* mastery reference obtained either during
1141 * assert_master_handler or in get_lock_resource */
1142 dlm_change_lockres_owner(dlm, res, m);
1143 spin_unlock(&res->spinlock);
1144
1145leave:
1146 return ret;
1147}
1148
1149struct dlm_bitmap_diff_iter
1150{
1151 int curnode;
1152 unsigned long *orig_bm;
1153 unsigned long *cur_bm;
1154 unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1155};
1156
1157enum dlm_node_state_change
1158{
1159 NODE_DOWN = -1,
1160 NODE_NO_CHANGE = 0,
1161 NODE_UP
1162};
1163
1164static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1165 unsigned long *orig_bm,
1166 unsigned long *cur_bm)
1167{
1168 unsigned long p1, p2;
1169 int i;
1170
1171 iter->curnode = -1;
1172 iter->orig_bm = orig_bm;
1173 iter->cur_bm = cur_bm;
1174
1175 for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1176 p1 = *(iter->orig_bm + i);
1177 p2 = *(iter->cur_bm + i);
1178 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1179 }
1180}
1181
1182static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1183 enum dlm_node_state_change *state)
1184{
1185 int bit;
1186
1187 if (iter->curnode >= O2NM_MAX_NODES)
1188 return -ENOENT;
1189
1190 bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1191 iter->curnode+1);
1192 if (bit >= O2NM_MAX_NODES) {
1193 iter->curnode = O2NM_MAX_NODES;
1194 return -ENOENT;
1195 }
1196
1197 /* if it was there in the original then this node died */
1198 if (test_bit(bit, iter->orig_bm))
1199 *state = NODE_DOWN;
1200 else
1201 *state = NODE_UP;
1202
1203 iter->curnode = bit;
1204 return bit;
1205}
1206
1207
1208static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1209 struct dlm_lock_resource *res,
1210 struct dlm_master_list_entry *mle,
1211 int blocked)
1212{
1213 struct dlm_bitmap_diff_iter bdi;
1214 enum dlm_node_state_change sc;
1215 int node;
1216 int ret = 0;
1217
1218 mlog(0, "something happened such that the "
1219 "master process may need to be restarted!\n");
1220
1221 assert_spin_locked(&mle->spinlock);
1222
1223 dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1224 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1225 while (node >= 0) {
1226 if (sc == NODE_UP) {
1227 /* a node came up. clear any old vote from
1228 * the response map and set it in the vote map
1229 * then restart the mastery. */
1230 mlog(ML_NOTICE, "node %d up while restarting\n", node);
1231
1232 /* redo the master request, but only for the new node */
1233 mlog(0, "sending request to new node\n");
1234 clear_bit(node, mle->response_map);
1235 set_bit(node, mle->vote_map);
1236 } else {
1237 mlog(ML_ERROR, "node down! %d\n", node);
1238 if (blocked) {
1239 int lowest = find_next_bit(mle->maybe_map,
1240 O2NM_MAX_NODES, 0);
1241
1242 /* act like it was never there */
1243 clear_bit(node, mle->maybe_map);
1244
1245 if (node == lowest) {
1246 mlog(0, "expected master %u died"
1247 " while this node was blocked "
1248 "waiting on it!\n", node);
1249 lowest = find_next_bit(mle->maybe_map,
1250 O2NM_MAX_NODES,
1251 lowest+1);
1252 if (lowest < O2NM_MAX_NODES) {
1253 mlog(0, "%s:%.*s:still "
1254 "blocked. waiting on %u "
1255 "now\n", dlm->name,
1256 res->lockname.len,
1257 res->lockname.name,
1258 lowest);
1259 } else {
1260 /* mle is an MLE_BLOCK, but
1261 * there is now nothing left to
1262 * block on. we need to return
1263 * all the way back out and try
1264 * again with an MLE_MASTER.
1265 * dlm_do_local_recovery_cleanup
1266 * has already run, so the mle
1267 * refcount is ok */
1268 mlog(0, "%s:%.*s: no "
1269 "longer blocking. try to "
1270 "master this here\n",
1271 dlm->name,
1272 res->lockname.len,
1273 res->lockname.name);
1274 mle->type = DLM_MLE_MASTER;
1275 mle->mleres = res;
1276 }
1277 }
1278 }
1279
1280 /* now blank out everything, as if we had never
1281 * contacted anyone */
1282 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1283 memset(mle->response_map, 0, sizeof(mle->response_map));
1284 /* reset the vote_map to the current node_map */
1285 memcpy(mle->vote_map, mle->node_map,
1286 sizeof(mle->node_map));
1287 /* put myself into the maybe map */
1288 if (mle->type != DLM_MLE_BLOCK)
1289 set_bit(dlm->node_num, mle->maybe_map);
1290 }
1291 ret = -EAGAIN;
1292 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1293 }
1294 return ret;
1295}
1296
1297
1298/*
1299 * DLM_MASTER_REQUEST_MSG
1300 *
1301 * returns: 0 on success,
1302 * -errno on a network error
1303 *
1304 * on error, the caller should assume the target node is "dead"
1305 *
1306 */
1307
1308static int dlm_do_master_request(struct dlm_lock_resource *res,
1309 struct dlm_master_list_entry *mle, int to)
1310{
1311 struct dlm_ctxt *dlm = mle->dlm;
1312 struct dlm_master_request request;
1313 int ret, response=0, resend;
1314
1315 memset(&request, 0, sizeof(request));
1316 request.node_idx = dlm->node_num;
1317
1318 BUG_ON(mle->type == DLM_MLE_MIGRATION);
1319
1320 request.namelen = (u8)mle->mnamelen;
1321 memcpy(request.name, mle->mname, request.namelen);
1322
1323again:
1324 ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1325 sizeof(request), to, &response);
1326 if (ret < 0) {
1327 if (ret == -ESRCH) {
1328 /* should never happen */
1329 mlog(ML_ERROR, "TCP stack not ready!\n");
1330 BUG();
1331 } else if (ret == -EINVAL) {
1332 mlog(ML_ERROR, "bad args passed to o2net!\n");
1333 BUG();
1334 } else if (ret == -ENOMEM) {
1335 mlog(ML_ERROR, "out of memory while trying to send "
1336 "network message! retrying\n");
1337 /* this is totally crude */
1338 msleep(50);
1339 goto again;
1340 } else if (!dlm_is_host_down(ret)) {
1341 /* not a network error. bad. */
1342 mlog_errno(ret);
1343 mlog(ML_ERROR, "unhandled error!");
1344 BUG();
1345 }
1346 /* all other errors should be network errors,
1347 * and likely indicate node death */
1348 mlog(ML_ERROR, "link to %d went down!\n", to);
1349 goto out;
1350 }
1351
1352 ret = 0;
1353 resend = 0;
1354 spin_lock(&mle->spinlock);
1355 switch (response) {
1356 case DLM_MASTER_RESP_YES:
1357 set_bit(to, mle->response_map);
1358 mlog(0, "node %u is the master, response=YES\n", to);
1359 mlog(0, "%s:%.*s: master node %u now knows I have a "
1360 "reference\n", dlm->name, res->lockname.len,
1361 res->lockname.name, to);
1362 mle->master = to;
1363 break;
1364 case DLM_MASTER_RESP_NO:
1365 mlog(0, "node %u not master, response=NO\n", to);
1366 set_bit(to, mle->response_map);
1367 break;
1368 case DLM_MASTER_RESP_MAYBE:
1369 mlog(0, "node %u not master, response=MAYBE\n", to);
1370 set_bit(to, mle->response_map);
1371 set_bit(to, mle->maybe_map);
1372 break;
1373 case DLM_MASTER_RESP_ERROR:
1374 mlog(0, "node %u hit an error, resending\n", to);
1375 resend = 1;
1376 response = 0;
1377 break;
1378 default:
1379 mlog(ML_ERROR, "bad response! %u\n", response);
1380 BUG();
1381 }
1382 spin_unlock(&mle->spinlock);
1383 if (resend) {
1384 /* this is also totally crude */
1385 msleep(50);
1386 goto again;
1387 }
1388
1389out:
1390 return ret;
1391}
1392
1393/*
1394 * locks that can be taken here:
1395 * dlm->spinlock
1396 * res->spinlock
1397 * mle->spinlock
1398 * dlm->master_list
1399 *
1400 * if possible, TRIM THIS DOWN!!!
1401 */
1402int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1403 void **ret_data)
1404{
1405 u8 response = DLM_MASTER_RESP_MAYBE;
1406 struct dlm_ctxt *dlm = data;
1407 struct dlm_lock_resource *res = NULL;
1408 struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1409 struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1410 char *name;
1411 unsigned int namelen, hash;
1412 int found, ret;
1413 int set_maybe;
1414 int dispatch_assert = 0;
1415 int dispatched = 0;
1416
1417 if (!dlm_grab(dlm))
1418 return DLM_MASTER_RESP_NO;
1419
1420 if (!dlm_domain_fully_joined(dlm)) {
1421 response = DLM_MASTER_RESP_NO;
1422 goto send_response;
1423 }
1424
1425 name = request->name;
1426 namelen = request->namelen;
1427 hash = dlm_lockid_hash(name, namelen);
1428
1429 if (namelen > DLM_LOCKID_NAME_MAX) {
1430 response = DLM_IVBUFLEN;
1431 goto send_response;
1432 }
1433
1434way_up_top:
1435 spin_lock(&dlm->spinlock);
1436 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1437 if (res) {
1438 spin_unlock(&dlm->spinlock);
1439
1440 /* take care of the easy cases up front */
1441 spin_lock(&res->spinlock);
1442
1443 /*
1444 * Right after dlm spinlock was released, dlm_thread could have
1445 * purged the lockres. Check if lockres got unhashed. If so
1446 * start over.
1447 */
1448 if (hlist_unhashed(&res->hash_node)) {
1449 spin_unlock(&res->spinlock);
1450 dlm_lockres_put(res);
1451 goto way_up_top;
1452 }
1453
1454 if (res->state & (DLM_LOCK_RES_RECOVERING|
1455 DLM_LOCK_RES_MIGRATING)) {
1456 spin_unlock(&res->spinlock);
1457 mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1458 "being recovered/migrated\n");
1459 response = DLM_MASTER_RESP_ERROR;
1460 if (mle)
1461 kmem_cache_free(dlm_mle_cache, mle);
1462 goto send_response;
1463 }
1464
1465 if (res->owner == dlm->node_num) {
1466 dlm_lockres_set_refmap_bit(dlm, res, request->node_idx);
1467 spin_unlock(&res->spinlock);
1468 response = DLM_MASTER_RESP_YES;
1469 if (mle)
1470 kmem_cache_free(dlm_mle_cache, mle);
1471
1472 /* this node is the owner.
1473 * there is some extra work that needs to
1474 * happen now. the requesting node has
1475 * caused all nodes up to this one to
1476 * create mles. this node now needs to
1477 * go back and clean those up. */
1478 dispatch_assert = 1;
1479 goto send_response;
1480 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1481 spin_unlock(&res->spinlock);
1482 // mlog(0, "node %u is the master\n", res->owner);
1483 response = DLM_MASTER_RESP_NO;
1484 if (mle)
1485 kmem_cache_free(dlm_mle_cache, mle);
1486 goto send_response;
1487 }
1488
1489 /* ok, there is no owner. either this node is
1490 * being blocked, or it is actively trying to
1491 * master this lock. */
1492 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1493 mlog(ML_ERROR, "lock with no owner should be "
1494 "in-progress!\n");
1495 BUG();
1496 }
1497
1498 // mlog(0, "lockres is in progress...\n");
1499 spin_lock(&dlm->master_lock);
1500 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1501 if (!found) {
1502 mlog(ML_ERROR, "no mle found for this lock!\n");
1503 BUG();
1504 }
1505 set_maybe = 1;
1506 spin_lock(&tmpmle->spinlock);
1507 if (tmpmle->type == DLM_MLE_BLOCK) {
1508 // mlog(0, "this node is waiting for "
1509 // "lockres to be mastered\n");
1510 response = DLM_MASTER_RESP_NO;
1511 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1512 mlog(0, "node %u is master, but trying to migrate to "
1513 "node %u.\n", tmpmle->master, tmpmle->new_master);
1514 if (tmpmle->master == dlm->node_num) {
1515 mlog(ML_ERROR, "no owner on lockres, but this "
1516 "node is trying to migrate it to %u?!\n",
1517 tmpmle->new_master);
1518 BUG();
1519 } else {
1520 /* the real master can respond on its own */
1521 response = DLM_MASTER_RESP_NO;
1522 }
1523 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1524 set_maybe = 0;
1525 if (tmpmle->master == dlm->node_num) {
1526 response = DLM_MASTER_RESP_YES;
1527 /* this node will be the owner.
1528 * go back and clean the mles on any
1529 * other nodes */
1530 dispatch_assert = 1;
1531 dlm_lockres_set_refmap_bit(dlm, res,
1532 request->node_idx);
1533 } else
1534 response = DLM_MASTER_RESP_NO;
1535 } else {
1536 // mlog(0, "this node is attempting to "
1537 // "master lockres\n");
1538 response = DLM_MASTER_RESP_MAYBE;
1539 }
1540 if (set_maybe)
1541 set_bit(request->node_idx, tmpmle->maybe_map);
1542 spin_unlock(&tmpmle->spinlock);
1543
1544 spin_unlock(&dlm->master_lock);
1545 spin_unlock(&res->spinlock);
1546
1547 /* keep the mle attached to heartbeat events */
1548 dlm_put_mle(tmpmle);
1549 if (mle)
1550 kmem_cache_free(dlm_mle_cache, mle);
1551 goto send_response;
1552 }
1553
1554 /*
1555 * lockres doesn't exist on this node
1556 * if there is an MLE_BLOCK, return NO
1557 * if there is an MLE_MASTER, return MAYBE
1558 * otherwise, add an MLE_BLOCK, return NO
1559 */
1560 spin_lock(&dlm->master_lock);
1561 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1562 if (!found) {
1563 /* this lockid has never been seen on this node yet */
1564 // mlog(0, "no mle found\n");
1565 if (!mle) {
1566 spin_unlock(&dlm->master_lock);
1567 spin_unlock(&dlm->spinlock);
1568
1569 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1570 if (!mle) {
1571 response = DLM_MASTER_RESP_ERROR;
1572 mlog_errno(-ENOMEM);
1573 goto send_response;
1574 }
1575 goto way_up_top;
1576 }
1577
1578 // mlog(0, "this is second time thru, already allocated, "
1579 // "add the block.\n");
1580 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1581 set_bit(request->node_idx, mle->maybe_map);
1582 __dlm_insert_mle(dlm, mle);
1583 response = DLM_MASTER_RESP_NO;
1584 } else {
1585 spin_lock(&tmpmle->spinlock);
1586 if (tmpmle->master == dlm->node_num) {
1587 mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1588 BUG();
1589 }
1590 if (tmpmle->type == DLM_MLE_BLOCK)
1591 response = DLM_MASTER_RESP_NO;
1592 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1593 mlog(0, "migration mle was found (%u->%u)\n",
1594 tmpmle->master, tmpmle->new_master);
1595 /* real master can respond on its own */
1596 response = DLM_MASTER_RESP_NO;
1597 } else
1598 response = DLM_MASTER_RESP_MAYBE;
1599 set_bit(request->node_idx, tmpmle->maybe_map);
1600 spin_unlock(&tmpmle->spinlock);
1601 }
1602 spin_unlock(&dlm->master_lock);
1603 spin_unlock(&dlm->spinlock);
1604
1605 if (found) {
1606 /* keep the mle attached to heartbeat events */
1607 dlm_put_mle(tmpmle);
1608 }
1609send_response:
1610 /*
1611 * __dlm_lookup_lockres() grabbed a reference to this lockres.
1612 * The reference is released by dlm_assert_master_worker() under
1613 * the call to dlm_dispatch_assert_master(). If
1614 * dlm_assert_master_worker() isn't called, we drop it here.
1615 */
1616 if (dispatch_assert) {
1617 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1618 dlm->node_num, res->lockname.len, res->lockname.name);
1619 spin_lock(&res->spinlock);
1620 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1621 DLM_ASSERT_MASTER_MLE_CLEANUP);
1622 if (ret < 0) {
1623 mlog(ML_ERROR, "failed to dispatch assert master work\n");
1624 response = DLM_MASTER_RESP_ERROR;
1625 spin_unlock(&res->spinlock);
1626 dlm_lockres_put(res);
1627 } else {
1628 dispatched = 1;
1629 __dlm_lockres_grab_inflight_worker(dlm, res);
1630 spin_unlock(&res->spinlock);
1631 }
1632 } else {
1633 if (res)
1634 dlm_lockres_put(res);
1635 }
1636
1637 if (!dispatched)
1638 dlm_put(dlm);
1639 return response;
1640}
1641
1642/*
1643 * DLM_ASSERT_MASTER_MSG
1644 */
1645
1646
1647/*
1648 * NOTE: this can be used for debugging
1649 * can periodically run all locks owned by this node
1650 * and re-assert across the cluster...
1651 */
1652static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1653 struct dlm_lock_resource *res,
1654 void *nodemap, u32 flags)
1655{
1656 struct dlm_assert_master assert;
1657 int to, tmpret;
1658 struct dlm_node_iter iter;
1659 int ret = 0;
1660 int reassert;
1661 const char *lockname = res->lockname.name;
1662 unsigned int namelen = res->lockname.len;
1663
1664 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1665
1666 spin_lock(&res->spinlock);
1667 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1668 spin_unlock(&res->spinlock);
1669
1670again:
1671 reassert = 0;
1672
1673 /* note that if this nodemap is empty, it returns 0 */
1674 dlm_node_iter_init(nodemap, &iter);
1675 while ((to = dlm_node_iter_next(&iter)) >= 0) {
1676 int r = 0;
1677 struct dlm_master_list_entry *mle = NULL;
1678
1679 mlog(0, "sending assert master to %d (%.*s)\n", to,
1680 namelen, lockname);
1681 memset(&assert, 0, sizeof(assert));
1682 assert.node_idx = dlm->node_num;
1683 assert.namelen = namelen;
1684 memcpy(assert.name, lockname, namelen);
1685 assert.flags = cpu_to_be32(flags);
1686
1687 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1688 &assert, sizeof(assert), to, &r);
1689 if (tmpret < 0) {
1690 mlog(ML_ERROR, "Error %d when sending message %u (key "
1691 "0x%x) to node %u\n", tmpret,
1692 DLM_ASSERT_MASTER_MSG, dlm->key, to);
1693 if (!dlm_is_host_down(tmpret)) {
1694 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1695 BUG();
1696 }
1697 /* a node died. finish out the rest of the nodes. */
1698 mlog(0, "link to %d went down!\n", to);
1699 /* any nonzero status return will do */
1700 ret = tmpret;
1701 r = 0;
1702 } else if (r < 0) {
1703 /* ok, something horribly messed. kill thyself. */
1704 mlog(ML_ERROR,"during assert master of %.*s to %u, "
1705 "got %d.\n", namelen, lockname, to, r);
1706 spin_lock(&dlm->spinlock);
1707 spin_lock(&dlm->master_lock);
1708 if (dlm_find_mle(dlm, &mle, (char *)lockname,
1709 namelen)) {
1710 dlm_print_one_mle(mle);
1711 __dlm_put_mle(mle);
1712 }
1713 spin_unlock(&dlm->master_lock);
1714 spin_unlock(&dlm->spinlock);
1715 BUG();
1716 }
1717
1718 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1719 !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1720 mlog(ML_ERROR, "%.*s: very strange, "
1721 "master MLE but no lockres on %u\n",
1722 namelen, lockname, to);
1723 }
1724
1725 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1726 mlog(0, "%.*s: node %u create mles on other "
1727 "nodes and requests a re-assert\n",
1728 namelen, lockname, to);
1729 reassert = 1;
1730 }
1731 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1732 mlog(0, "%.*s: node %u has a reference to this "
1733 "lockres, set the bit in the refmap\n",
1734 namelen, lockname, to);
1735 spin_lock(&res->spinlock);
1736 dlm_lockres_set_refmap_bit(dlm, res, to);
1737 spin_unlock(&res->spinlock);
1738 }
1739 }
1740
1741 if (reassert)
1742 goto again;
1743
1744 spin_lock(&res->spinlock);
1745 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1746 spin_unlock(&res->spinlock);
1747 wake_up(&res->wq);
1748
1749 return ret;
1750}
1751
1752/*
1753 * locks that can be taken here:
1754 * dlm->spinlock
1755 * res->spinlock
1756 * mle->spinlock
1757 * dlm->master_list
1758 *
1759 * if possible, TRIM THIS DOWN!!!
1760 */
1761int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1762 void **ret_data)
1763{
1764 struct dlm_ctxt *dlm = data;
1765 struct dlm_master_list_entry *mle = NULL;
1766 struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1767 struct dlm_lock_resource *res = NULL;
1768 char *name;
1769 unsigned int namelen, hash;
1770 u32 flags;
1771 int master_request = 0, have_lockres_ref = 0;
1772 int ret = 0;
1773
1774 if (!dlm_grab(dlm))
1775 return 0;
1776
1777 name = assert->name;
1778 namelen = assert->namelen;
1779 hash = dlm_lockid_hash(name, namelen);
1780 flags = be32_to_cpu(assert->flags);
1781
1782 if (namelen > DLM_LOCKID_NAME_MAX) {
1783 mlog(ML_ERROR, "Invalid name length!");
1784 goto done;
1785 }
1786
1787 spin_lock(&dlm->spinlock);
1788
1789 if (flags)
1790 mlog(0, "assert_master with flags: %u\n", flags);
1791
1792 /* find the MLE */
1793 spin_lock(&dlm->master_lock);
1794 if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1795 /* not an error, could be master just re-asserting */
1796 mlog(0, "just got an assert_master from %u, but no "
1797 "MLE for it! (%.*s)\n", assert->node_idx,
1798 namelen, name);
1799 } else {
1800 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1801 if (bit >= O2NM_MAX_NODES) {
1802 /* not necessarily an error, though less likely.
1803 * could be master just re-asserting. */
1804 mlog(0, "no bits set in the maybe_map, but %u "
1805 "is asserting! (%.*s)\n", assert->node_idx,
1806 namelen, name);
1807 } else if (bit != assert->node_idx) {
1808 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1809 mlog(0, "master %u was found, %u should "
1810 "back off\n", assert->node_idx, bit);
1811 } else {
1812 /* with the fix for bug 569, a higher node
1813 * number winning the mastery will respond
1814 * YES to mastery requests, but this node
1815 * had no way of knowing. let it pass. */
1816 mlog(0, "%u is the lowest node, "
1817 "%u is asserting. (%.*s) %u must "
1818 "have begun after %u won.\n", bit,
1819 assert->node_idx, namelen, name, bit,
1820 assert->node_idx);
1821 }
1822 }
1823 if (mle->type == DLM_MLE_MIGRATION) {
1824 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1825 mlog(0, "%s:%.*s: got cleanup assert"
1826 " from %u for migration\n",
1827 dlm->name, namelen, name,
1828 assert->node_idx);
1829 } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1830 mlog(0, "%s:%.*s: got unrelated assert"
1831 " from %u for migration, ignoring\n",
1832 dlm->name, namelen, name,
1833 assert->node_idx);
1834 __dlm_put_mle(mle);
1835 spin_unlock(&dlm->master_lock);
1836 spin_unlock(&dlm->spinlock);
1837 goto done;
1838 }
1839 }
1840 }
1841 spin_unlock(&dlm->master_lock);
1842
1843 /* ok everything checks out with the MLE
1844 * now check to see if there is a lockres */
1845 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1846 if (res) {
1847 spin_lock(&res->spinlock);
1848 if (res->state & DLM_LOCK_RES_RECOVERING) {
1849 mlog(ML_ERROR, "%u asserting but %.*s is "
1850 "RECOVERING!\n", assert->node_idx, namelen, name);
1851 goto kill;
1852 }
1853 if (!mle) {
1854 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1855 res->owner != assert->node_idx) {
1856 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1857 "but current owner is %u! (%.*s)\n",
1858 assert->node_idx, res->owner, namelen,
1859 name);
1860 __dlm_print_one_lock_resource(res);
1861 BUG();
1862 }
1863 } else if (mle->type != DLM_MLE_MIGRATION) {
1864 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1865 /* owner is just re-asserting */
1866 if (res->owner == assert->node_idx) {
1867 mlog(0, "owner %u re-asserting on "
1868 "lock %.*s\n", assert->node_idx,
1869 namelen, name);
1870 goto ok;
1871 }
1872 mlog(ML_ERROR, "got assert_master from "
1873 "node %u, but %u is the owner! "
1874 "(%.*s)\n", assert->node_idx,
1875 res->owner, namelen, name);
1876 goto kill;
1877 }
1878 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1879 mlog(ML_ERROR, "got assert from %u, but lock "
1880 "with no owner should be "
1881 "in-progress! (%.*s)\n",
1882 assert->node_idx,
1883 namelen, name);
1884 goto kill;
1885 }
1886 } else /* mle->type == DLM_MLE_MIGRATION */ {
1887 /* should only be getting an assert from new master */
1888 if (assert->node_idx != mle->new_master) {
1889 mlog(ML_ERROR, "got assert from %u, but "
1890 "new master is %u, and old master "
1891 "was %u (%.*s)\n",
1892 assert->node_idx, mle->new_master,
1893 mle->master, namelen, name);
1894 goto kill;
1895 }
1896
1897 }
1898ok:
1899 spin_unlock(&res->spinlock);
1900 }
1901
1902 // mlog(0, "woo! got an assert_master from node %u!\n",
1903 // assert->node_idx);
1904 if (mle) {
1905 int extra_ref = 0;
1906 int nn = -1;
1907 int rr, err = 0;
1908
1909 spin_lock(&mle->spinlock);
1910 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1911 extra_ref = 1;
1912 else {
1913 /* MASTER mle: if any bits set in the response map
1914 * then the calling node needs to re-assert to clear
1915 * up nodes that this node contacted */
1916 while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1917 nn+1)) < O2NM_MAX_NODES) {
1918 if (nn != dlm->node_num && nn != assert->node_idx) {
1919 master_request = 1;
1920 break;
1921 }
1922 }
1923 }
1924 mle->master = assert->node_idx;
1925 atomic_set(&mle->woken, 1);
1926 wake_up(&mle->wq);
1927 spin_unlock(&mle->spinlock);
1928
1929 if (res) {
1930 int wake = 0;
1931 spin_lock(&res->spinlock);
1932 if (mle->type == DLM_MLE_MIGRATION) {
1933 mlog(0, "finishing off migration of lockres %.*s, "
1934 "from %u to %u\n",
1935 res->lockname.len, res->lockname.name,
1936 dlm->node_num, mle->new_master);
1937 res->state &= ~DLM_LOCK_RES_MIGRATING;
1938 wake = 1;
1939 dlm_change_lockres_owner(dlm, res, mle->new_master);
1940 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1941 } else {
1942 dlm_change_lockres_owner(dlm, res, mle->master);
1943 }
1944 spin_unlock(&res->spinlock);
1945 have_lockres_ref = 1;
1946 if (wake)
1947 wake_up(&res->wq);
1948 }
1949
1950 /* master is known, detach if not already detached.
1951 * ensures that only one assert_master call will happen
1952 * on this mle. */
1953 spin_lock(&dlm->master_lock);
1954
1955 rr = kref_read(&mle->mle_refs);
1956 if (mle->inuse > 0) {
1957 if (extra_ref && rr < 3)
1958 err = 1;
1959 else if (!extra_ref && rr < 2)
1960 err = 1;
1961 } else {
1962 if (extra_ref && rr < 2)
1963 err = 1;
1964 else if (!extra_ref && rr < 1)
1965 err = 1;
1966 }
1967 if (err) {
1968 mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
1969 "that will mess up this node, refs=%d, extra=%d, "
1970 "inuse=%d\n", dlm->name, namelen, name,
1971 assert->node_idx, rr, extra_ref, mle->inuse);
1972 dlm_print_one_mle(mle);
1973 }
1974 __dlm_unlink_mle(dlm, mle);
1975 __dlm_mle_detach_hb_events(dlm, mle);
1976 __dlm_put_mle(mle);
1977 if (extra_ref) {
1978 /* the assert master message now balances the extra
1979 * ref given by the master / migration request message.
1980 * if this is the last put, it will be removed
1981 * from the list. */
1982 __dlm_put_mle(mle);
1983 }
1984 spin_unlock(&dlm->master_lock);
1985 } else if (res) {
1986 if (res->owner != assert->node_idx) {
1987 mlog(0, "assert_master from %u, but current "
1988 "owner is %u (%.*s), no mle\n", assert->node_idx,
1989 res->owner, namelen, name);
1990 }
1991 }
1992 spin_unlock(&dlm->spinlock);
1993
1994done:
1995 ret = 0;
1996 if (res) {
1997 spin_lock(&res->spinlock);
1998 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1999 spin_unlock(&res->spinlock);
2000 *ret_data = (void *)res;
2001 }
2002 dlm_put(dlm);
2003 if (master_request) {
2004 mlog(0, "need to tell master to reassert\n");
2005 /* positive. negative would shoot down the node. */
2006 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2007 if (!have_lockres_ref) {
2008 mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2009 "mle present here for %s:%.*s, but no lockres!\n",
2010 assert->node_idx, dlm->name, namelen, name);
2011 }
2012 }
2013 if (have_lockres_ref) {
2014 /* let the master know we have a reference to the lockres */
2015 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2016 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2017 dlm->name, namelen, name, assert->node_idx);
2018 }
2019 return ret;
2020
2021kill:
2022 /* kill the caller! */
2023 mlog(ML_ERROR, "Bad message received from another node. Dumping state "
2024 "and killing the other node now! This node is OK and can continue.\n");
2025 __dlm_print_one_lock_resource(res);
2026 spin_unlock(&res->spinlock);
2027 spin_lock(&dlm->master_lock);
2028 if (mle)
2029 __dlm_put_mle(mle);
2030 spin_unlock(&dlm->master_lock);
2031 spin_unlock(&dlm->spinlock);
2032 *ret_data = (void *)res;
2033 dlm_put(dlm);
2034 return -EINVAL;
2035}
2036
2037void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2038{
2039 struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2040
2041 if (ret_data) {
2042 spin_lock(&res->spinlock);
2043 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2044 spin_unlock(&res->spinlock);
2045 wake_up(&res->wq);
2046 dlm_lockres_put(res);
2047 }
2048 return;
2049}
2050
2051int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2052 struct dlm_lock_resource *res,
2053 int ignore_higher, u8 request_from, u32 flags)
2054{
2055 struct dlm_work_item *item;
2056 item = kzalloc(sizeof(*item), GFP_ATOMIC);
2057 if (!item)
2058 return -ENOMEM;
2059
2060
2061 /* queue up work for dlm_assert_master_worker */
2062 dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2063 item->u.am.lockres = res; /* already have a ref */
2064 /* can optionally ignore node numbers higher than this node */
2065 item->u.am.ignore_higher = ignore_higher;
2066 item->u.am.request_from = request_from;
2067 item->u.am.flags = flags;
2068
2069 if (ignore_higher)
2070 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2071 res->lockname.name);
2072
2073 spin_lock(&dlm->work_lock);
2074 list_add_tail(&item->list, &dlm->work_list);
2075 spin_unlock(&dlm->work_lock);
2076
2077 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2078 return 0;
2079}
2080
2081static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2082{
2083 struct dlm_ctxt *dlm = data;
2084 int ret = 0;
2085 struct dlm_lock_resource *res;
2086 unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2087 int ignore_higher;
2088 int bit;
2089 u8 request_from;
2090 u32 flags;
2091
2092 dlm = item->dlm;
2093 res = item->u.am.lockres;
2094 ignore_higher = item->u.am.ignore_higher;
2095 request_from = item->u.am.request_from;
2096 flags = item->u.am.flags;
2097
2098 spin_lock(&dlm->spinlock);
2099 memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2100 spin_unlock(&dlm->spinlock);
2101
2102 clear_bit(dlm->node_num, nodemap);
2103 if (ignore_higher) {
2104 /* if is this just to clear up mles for nodes below
2105 * this node, do not send the message to the original
2106 * caller or any node number higher than this */
2107 clear_bit(request_from, nodemap);
2108 bit = dlm->node_num;
2109 while (1) {
2110 bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2111 bit+1);
2112 if (bit >= O2NM_MAX_NODES)
2113 break;
2114 clear_bit(bit, nodemap);
2115 }
2116 }
2117
2118 /*
2119 * If we're migrating this lock to someone else, we are no
2120 * longer allowed to assert out own mastery. OTOH, we need to
2121 * prevent migration from starting while we're still asserting
2122 * our dominance. The reserved ast delays migration.
2123 */
2124 spin_lock(&res->spinlock);
2125 if (res->state & DLM_LOCK_RES_MIGRATING) {
2126 mlog(0, "Someone asked us to assert mastery, but we're "
2127 "in the middle of migration. Skipping assert, "
2128 "the new master will handle that.\n");
2129 spin_unlock(&res->spinlock);
2130 goto put;
2131 } else
2132 __dlm_lockres_reserve_ast(res);
2133 spin_unlock(&res->spinlock);
2134
2135 /* this call now finishes out the nodemap
2136 * even if one or more nodes die */
2137 mlog(0, "worker about to master %.*s here, this=%u\n",
2138 res->lockname.len, res->lockname.name, dlm->node_num);
2139 ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2140 if (ret < 0) {
2141 /* no need to restart, we are done */
2142 if (!dlm_is_host_down(ret))
2143 mlog_errno(ret);
2144 }
2145
2146 /* Ok, we've asserted ourselves. Let's let migration start. */
2147 dlm_lockres_release_ast(dlm, res);
2148
2149put:
2150 dlm_lockres_drop_inflight_worker(dlm, res);
2151
2152 dlm_lockres_put(res);
2153
2154 mlog(0, "finished with dlm_assert_master_worker\n");
2155}
2156
2157/* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2158 * We cannot wait for node recovery to complete to begin mastering this
2159 * lockres because this lockres is used to kick off recovery! ;-)
2160 * So, do a pre-check on all living nodes to see if any of those nodes
2161 * think that $RECOVERY is currently mastered by a dead node. If so,
2162 * we wait a short time to allow that node to get notified by its own
2163 * heartbeat stack, then check again. All $RECOVERY lock resources
2164 * mastered by dead nodes are purged when the heartbeat callback is
2165 * fired, so we can know for sure that it is safe to continue once
2166 * the node returns a live node or no node. */
2167static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2168 struct dlm_lock_resource *res)
2169{
2170 struct dlm_node_iter iter;
2171 int nodenum;
2172 int ret = 0;
2173 u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2174
2175 spin_lock(&dlm->spinlock);
2176 dlm_node_iter_init(dlm->domain_map, &iter);
2177 spin_unlock(&dlm->spinlock);
2178
2179 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2180 /* do not send to self */
2181 if (nodenum == dlm->node_num)
2182 continue;
2183 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2184 if (ret < 0) {
2185 mlog_errno(ret);
2186 if (!dlm_is_host_down(ret))
2187 BUG();
2188 /* host is down, so answer for that node would be
2189 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
2190 ret = 0;
2191 }
2192
2193 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2194 /* check to see if this master is in the recovery map */
2195 spin_lock(&dlm->spinlock);
2196 if (test_bit(master, dlm->recovery_map)) {
2197 mlog(ML_NOTICE, "%s: node %u has not seen "
2198 "node %u go down yet, and thinks the "
2199 "dead node is mastering the recovery "
2200 "lock. must wait.\n", dlm->name,
2201 nodenum, master);
2202 ret = -EAGAIN;
2203 }
2204 spin_unlock(&dlm->spinlock);
2205 mlog(0, "%s: reco lock master is %u\n", dlm->name,
2206 master);
2207 break;
2208 }
2209 }
2210 return ret;
2211}
2212
2213/*
2214 * DLM_DEREF_LOCKRES_MSG
2215 */
2216
2217int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2218{
2219 struct dlm_deref_lockres deref;
2220 int ret = 0, r;
2221 const char *lockname;
2222 unsigned int namelen;
2223
2224 lockname = res->lockname.name;
2225 namelen = res->lockname.len;
2226 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2227
2228 memset(&deref, 0, sizeof(deref));
2229 deref.node_idx = dlm->node_num;
2230 deref.namelen = namelen;
2231 memcpy(deref.name, lockname, namelen);
2232
2233 ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2234 &deref, sizeof(deref), res->owner, &r);
2235 if (ret < 0)
2236 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n",
2237 dlm->name, namelen, lockname, ret, res->owner);
2238 else if (r < 0) {
2239 /* BAD. other node says I did not have a ref. */
2240 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2241 dlm->name, namelen, lockname, res->owner, r);
2242 dlm_print_one_lock_resource(res);
2243 if (r == -ENOMEM)
2244 BUG();
2245 } else
2246 ret = r;
2247
2248 return ret;
2249}
2250
2251int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2252 void **ret_data)
2253{
2254 struct dlm_ctxt *dlm = data;
2255 struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2256 struct dlm_lock_resource *res = NULL;
2257 char *name;
2258 unsigned int namelen;
2259 int ret = -EINVAL;
2260 u8 node;
2261 unsigned int hash;
2262 struct dlm_work_item *item;
2263 int cleared = 0;
2264 int dispatch = 0;
2265
2266 if (!dlm_grab(dlm))
2267 return 0;
2268
2269 name = deref->name;
2270 namelen = deref->namelen;
2271 node = deref->node_idx;
2272
2273 if (namelen > DLM_LOCKID_NAME_MAX) {
2274 mlog(ML_ERROR, "Invalid name length!");
2275 goto done;
2276 }
2277 if (deref->node_idx >= O2NM_MAX_NODES) {
2278 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2279 goto done;
2280 }
2281
2282 hash = dlm_lockid_hash(name, namelen);
2283
2284 spin_lock(&dlm->spinlock);
2285 res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2286 if (!res) {
2287 spin_unlock(&dlm->spinlock);
2288 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2289 dlm->name, namelen, name);
2290 goto done;
2291 }
2292 spin_unlock(&dlm->spinlock);
2293
2294 spin_lock(&res->spinlock);
2295 if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2296 dispatch = 1;
2297 else {
2298 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2299 if (test_bit(node, res->refmap)) {
2300 dlm_lockres_clear_refmap_bit(dlm, res, node);
2301 cleared = 1;
2302 }
2303 }
2304 spin_unlock(&res->spinlock);
2305
2306 if (!dispatch) {
2307 if (cleared)
2308 dlm_lockres_calc_usage(dlm, res);
2309 else {
2310 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2311 "but it is already dropped!\n", dlm->name,
2312 res->lockname.len, res->lockname.name, node);
2313 dlm_print_one_lock_resource(res);
2314 }
2315 ret = DLM_DEREF_RESPONSE_DONE;
2316 goto done;
2317 }
2318
2319 item = kzalloc(sizeof(*item), GFP_NOFS);
2320 if (!item) {
2321 ret = -ENOMEM;
2322 mlog_errno(ret);
2323 goto done;
2324 }
2325
2326 dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2327 item->u.dl.deref_res = res;
2328 item->u.dl.deref_node = node;
2329
2330 spin_lock(&dlm->work_lock);
2331 list_add_tail(&item->list, &dlm->work_list);
2332 spin_unlock(&dlm->work_lock);
2333
2334 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2335 return DLM_DEREF_RESPONSE_INPROG;
2336
2337done:
2338 if (res)
2339 dlm_lockres_put(res);
2340 dlm_put(dlm);
2341
2342 return ret;
2343}
2344
2345int dlm_deref_lockres_done_handler(struct o2net_msg *msg, u32 len, void *data,
2346 void **ret_data)
2347{
2348 struct dlm_ctxt *dlm = data;
2349 struct dlm_deref_lockres_done *deref
2350 = (struct dlm_deref_lockres_done *)msg->buf;
2351 struct dlm_lock_resource *res = NULL;
2352 char *name;
2353 unsigned int namelen;
2354 int ret = -EINVAL;
2355 u8 node;
2356 unsigned int hash;
2357
2358 if (!dlm_grab(dlm))
2359 return 0;
2360
2361 name = deref->name;
2362 namelen = deref->namelen;
2363 node = deref->node_idx;
2364
2365 if (namelen > DLM_LOCKID_NAME_MAX) {
2366 mlog(ML_ERROR, "Invalid name length!");
2367 goto done;
2368 }
2369 if (deref->node_idx >= O2NM_MAX_NODES) {
2370 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2371 goto done;
2372 }
2373
2374 hash = dlm_lockid_hash(name, namelen);
2375
2376 spin_lock(&dlm->spinlock);
2377 res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2378 if (!res) {
2379 spin_unlock(&dlm->spinlock);
2380 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2381 dlm->name, namelen, name);
2382 goto done;
2383 }
2384
2385 spin_lock(&res->spinlock);
2386 if (!(res->state & DLM_LOCK_RES_DROPPING_REF)) {
2387 spin_unlock(&res->spinlock);
2388 spin_unlock(&dlm->spinlock);
2389 mlog(ML_NOTICE, "%s:%.*s: node %u sends deref done "
2390 "but it is already derefed!\n", dlm->name,
2391 res->lockname.len, res->lockname.name, node);
2392 ret = 0;
2393 goto done;
2394 }
2395
2396 __dlm_do_purge_lockres(dlm, res);
2397 spin_unlock(&res->spinlock);
2398 wake_up(&res->wq);
2399
2400 spin_unlock(&dlm->spinlock);
2401
2402 ret = 0;
2403done:
2404 if (res)
2405 dlm_lockres_put(res);
2406 dlm_put(dlm);
2407 return ret;
2408}
2409
2410static void dlm_drop_lockres_ref_done(struct dlm_ctxt *dlm,
2411 struct dlm_lock_resource *res, u8 node)
2412{
2413 struct dlm_deref_lockres_done deref;
2414 int ret = 0, r;
2415 const char *lockname;
2416 unsigned int namelen;
2417
2418 lockname = res->lockname.name;
2419 namelen = res->lockname.len;
2420 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2421
2422 memset(&deref, 0, sizeof(deref));
2423 deref.node_idx = dlm->node_num;
2424 deref.namelen = namelen;
2425 memcpy(deref.name, lockname, namelen);
2426
2427 ret = o2net_send_message(DLM_DEREF_LOCKRES_DONE, dlm->key,
2428 &deref, sizeof(deref), node, &r);
2429 if (ret < 0) {
2430 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF DONE "
2431 " to node %u\n", dlm->name, namelen,
2432 lockname, ret, node);
2433 } else if (r < 0) {
2434 /* ignore the error */
2435 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2436 dlm->name, namelen, lockname, node, r);
2437 dlm_print_one_lock_resource(res);
2438 }
2439}
2440
2441static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2442{
2443 struct dlm_ctxt *dlm;
2444 struct dlm_lock_resource *res;
2445 u8 node;
2446 u8 cleared = 0;
2447
2448 dlm = item->dlm;
2449 res = item->u.dl.deref_res;
2450 node = item->u.dl.deref_node;
2451
2452 spin_lock(&res->spinlock);
2453 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2454 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2455 if (test_bit(node, res->refmap)) {
2456 dlm_lockres_clear_refmap_bit(dlm, res, node);
2457 cleared = 1;
2458 }
2459 spin_unlock(&res->spinlock);
2460
2461 dlm_drop_lockres_ref_done(dlm, res, node);
2462
2463 if (cleared) {
2464 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2465 dlm->name, res->lockname.len, res->lockname.name, node);
2466 dlm_lockres_calc_usage(dlm, res);
2467 } else {
2468 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2469 "but it is already dropped!\n", dlm->name,
2470 res->lockname.len, res->lockname.name, node);
2471 dlm_print_one_lock_resource(res);
2472 }
2473
2474 dlm_lockres_put(res);
2475}
2476
2477/*
2478 * A migratable resource is one that is :
2479 * 1. locally mastered, and,
2480 * 2. zero local locks, and,
2481 * 3. one or more non-local locks, or, one or more references
2482 * Returns 1 if yes, 0 if not.
2483 */
2484static int dlm_is_lockres_migratable(struct dlm_ctxt *dlm,
2485 struct dlm_lock_resource *res)
2486{
2487 enum dlm_lockres_list idx;
2488 int nonlocal = 0, node_ref;
2489 struct list_head *queue;
2490 struct dlm_lock *lock;
2491 u64 cookie;
2492
2493 assert_spin_locked(&res->spinlock);
2494
2495 /* delay migration when the lockres is in MIGRATING state */
2496 if (res->state & DLM_LOCK_RES_MIGRATING)
2497 return 0;
2498
2499 /* delay migration when the lockres is in RECOCERING state */
2500 if (res->state & (DLM_LOCK_RES_RECOVERING|
2501 DLM_LOCK_RES_RECOVERY_WAITING))
2502 return 0;
2503
2504 if (res->owner != dlm->node_num)
2505 return 0;
2506
2507 for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2508 queue = dlm_list_idx_to_ptr(res, idx);
2509 list_for_each_entry(lock, queue, list) {
2510 if (lock->ml.node != dlm->node_num) {
2511 nonlocal++;
2512 continue;
2513 }
2514 cookie = be64_to_cpu(lock->ml.cookie);
2515 mlog(0, "%s: Not migratable res %.*s, lock %u:%llu on "
2516 "%s list\n", dlm->name, res->lockname.len,
2517 res->lockname.name,
2518 dlm_get_lock_cookie_node(cookie),
2519 dlm_get_lock_cookie_seq(cookie),
2520 dlm_list_in_text(idx));
2521 return 0;
2522 }
2523 }
2524
2525 if (!nonlocal) {
2526 node_ref = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
2527 if (node_ref >= O2NM_MAX_NODES)
2528 return 0;
2529 }
2530
2531 mlog(0, "%s: res %.*s, Migratable\n", dlm->name, res->lockname.len,
2532 res->lockname.name);
2533
2534 return 1;
2535}
2536
2537/*
2538 * DLM_MIGRATE_LOCKRES
2539 */
2540
2541
2542static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2543 struct dlm_lock_resource *res, u8 target)
2544{
2545 struct dlm_master_list_entry *mle = NULL;
2546 struct dlm_master_list_entry *oldmle = NULL;
2547 struct dlm_migratable_lockres *mres = NULL;
2548 int ret = 0;
2549 const char *name;
2550 unsigned int namelen;
2551 int mle_added = 0;
2552 int wake = 0;
2553
2554 if (!dlm_grab(dlm))
2555 return -EINVAL;
2556
2557 BUG_ON(target == O2NM_MAX_NODES);
2558
2559 name = res->lockname.name;
2560 namelen = res->lockname.len;
2561
2562 mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name,
2563 target);
2564
2565 /* preallocate up front. if this fails, abort */
2566 ret = -ENOMEM;
2567 mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2568 if (!mres) {
2569 mlog_errno(ret);
2570 goto leave;
2571 }
2572
2573 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2574 if (!mle) {
2575 mlog_errno(ret);
2576 goto leave;
2577 }
2578 ret = 0;
2579
2580 /*
2581 * clear any existing master requests and
2582 * add the migration mle to the list
2583 */
2584 spin_lock(&dlm->spinlock);
2585 spin_lock(&dlm->master_lock);
2586 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2587 namelen, target, dlm->node_num);
2588 /* get an extra reference on the mle.
2589 * otherwise the assert_master from the new
2590 * master will destroy this.
2591 */
2592 if (ret != -EEXIST)
2593 dlm_get_mle_inuse(mle);
2594
2595 spin_unlock(&dlm->master_lock);
2596 spin_unlock(&dlm->spinlock);
2597
2598 if (ret == -EEXIST) {
2599 mlog(0, "another process is already migrating it\n");
2600 goto fail;
2601 }
2602 mle_added = 1;
2603
2604 /*
2605 * set the MIGRATING flag and flush asts
2606 * if we fail after this we need to re-dirty the lockres
2607 */
2608 if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2609 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2610 "the target went down.\n", res->lockname.len,
2611 res->lockname.name, target);
2612 spin_lock(&res->spinlock);
2613 res->state &= ~DLM_LOCK_RES_MIGRATING;
2614 wake = 1;
2615 spin_unlock(&res->spinlock);
2616 ret = -EINVAL;
2617 }
2618
2619fail:
2620 if (ret != -EEXIST && oldmle) {
2621 /* master is known, detach if not already detached */
2622 dlm_mle_detach_hb_events(dlm, oldmle);
2623 dlm_put_mle(oldmle);
2624 }
2625
2626 if (ret < 0) {
2627 if (mle_added) {
2628 dlm_mle_detach_hb_events(dlm, mle);
2629 dlm_put_mle(mle);
2630 dlm_put_mle_inuse(mle);
2631 } else if (mle) {
2632 kmem_cache_free(dlm_mle_cache, mle);
2633 mle = NULL;
2634 }
2635 goto leave;
2636 }
2637
2638 /*
2639 * at this point, we have a migration target, an mle
2640 * in the master list, and the MIGRATING flag set on
2641 * the lockres
2642 */
2643
2644 /* now that remote nodes are spinning on the MIGRATING flag,
2645 * ensure that all assert_master work is flushed. */
2646 flush_workqueue(dlm->dlm_worker);
2647
2648 /* notify new node and send all lock state */
2649 /* call send_one_lockres with migration flag.
2650 * this serves as notice to the target node that a
2651 * migration is starting. */
2652 ret = dlm_send_one_lockres(dlm, res, mres, target,
2653 DLM_MRES_MIGRATION);
2654
2655 if (ret < 0) {
2656 mlog(0, "migration to node %u failed with %d\n",
2657 target, ret);
2658 /* migration failed, detach and clean up mle */
2659 dlm_mle_detach_hb_events(dlm, mle);
2660 dlm_put_mle(mle);
2661 dlm_put_mle_inuse(mle);
2662 spin_lock(&res->spinlock);
2663 res->state &= ~DLM_LOCK_RES_MIGRATING;
2664 wake = 1;
2665 spin_unlock(&res->spinlock);
2666 if (dlm_is_host_down(ret))
2667 dlm_wait_for_node_death(dlm, target,
2668 DLM_NODE_DEATH_WAIT_MAX);
2669 goto leave;
2670 }
2671
2672 /* at this point, the target sends a message to all nodes,
2673 * (using dlm_do_migrate_request). this node is skipped since
2674 * we had to put an mle in the list to begin the process. this
2675 * node now waits for target to do an assert master. this node
2676 * will be the last one notified, ensuring that the migration
2677 * is complete everywhere. if the target dies while this is
2678 * going on, some nodes could potentially see the target as the
2679 * master, so it is important that my recovery finds the migration
2680 * mle and sets the master to UNKNOWN. */
2681
2682
2683 /* wait for new node to assert master */
2684 while (1) {
2685 ret = wait_event_interruptible_timeout(mle->wq,
2686 (atomic_read(&mle->woken) == 1),
2687 msecs_to_jiffies(5000));
2688
2689 if (ret >= 0) {
2690 if (atomic_read(&mle->woken) == 1 ||
2691 res->owner == target)
2692 break;
2693
2694 mlog(0, "%s:%.*s: timed out during migration\n",
2695 dlm->name, res->lockname.len, res->lockname.name);
2696 /* avoid hang during shutdown when migrating lockres
2697 * to a node which also goes down */
2698 if (dlm_is_node_dead(dlm, target)) {
2699 mlog(0, "%s:%.*s: expected migration "
2700 "target %u is no longer up, restarting\n",
2701 dlm->name, res->lockname.len,
2702 res->lockname.name, target);
2703 ret = -EINVAL;
2704 /* migration failed, detach and clean up mle */
2705 dlm_mle_detach_hb_events(dlm, mle);
2706 dlm_put_mle(mle);
2707 dlm_put_mle_inuse(mle);
2708 spin_lock(&res->spinlock);
2709 res->state &= ~DLM_LOCK_RES_MIGRATING;
2710 wake = 1;
2711 spin_unlock(&res->spinlock);
2712 goto leave;
2713 }
2714 } else
2715 mlog(0, "%s:%.*s: caught signal during migration\n",
2716 dlm->name, res->lockname.len, res->lockname.name);
2717 }
2718
2719 /* all done, set the owner, clear the flag */
2720 spin_lock(&res->spinlock);
2721 dlm_set_lockres_owner(dlm, res, target);
2722 res->state &= ~DLM_LOCK_RES_MIGRATING;
2723 dlm_remove_nonlocal_locks(dlm, res);
2724 spin_unlock(&res->spinlock);
2725 wake_up(&res->wq);
2726
2727 /* master is known, detach if not already detached */
2728 dlm_mle_detach_hb_events(dlm, mle);
2729 dlm_put_mle_inuse(mle);
2730 ret = 0;
2731
2732 dlm_lockres_calc_usage(dlm, res);
2733
2734leave:
2735 /* re-dirty the lockres if we failed */
2736 if (ret < 0)
2737 dlm_kick_thread(dlm, res);
2738
2739 /* wake up waiters if the MIGRATING flag got set
2740 * but migration failed */
2741 if (wake)
2742 wake_up(&res->wq);
2743
2744 if (mres)
2745 free_page((unsigned long)mres);
2746
2747 dlm_put(dlm);
2748
2749 mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen,
2750 name, target, ret);
2751 return ret;
2752}
2753
2754#define DLM_MIGRATION_RETRY_MS 100
2755
2756/*
2757 * Should be called only after beginning the domain leave process.
2758 * There should not be any remaining locks on nonlocal lock resources,
2759 * and there should be no local locks left on locally mastered resources.
2760 *
2761 * Called with the dlm spinlock held, may drop it to do migration, but
2762 * will re-acquire before exit.
2763 *
2764 * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2765 */
2766int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2767{
2768 int ret;
2769 int lock_dropped = 0;
2770 u8 target = O2NM_MAX_NODES;
2771
2772 assert_spin_locked(&dlm->spinlock);
2773
2774 spin_lock(&res->spinlock);
2775 if (dlm_is_lockres_migratable(dlm, res))
2776 target = dlm_pick_migration_target(dlm, res);
2777 spin_unlock(&res->spinlock);
2778
2779 if (target == O2NM_MAX_NODES)
2780 goto leave;
2781
2782 /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2783 spin_unlock(&dlm->spinlock);
2784 lock_dropped = 1;
2785 ret = dlm_migrate_lockres(dlm, res, target);
2786 if (ret)
2787 mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2788 dlm->name, res->lockname.len, res->lockname.name,
2789 target, ret);
2790 spin_lock(&dlm->spinlock);
2791leave:
2792 return lock_dropped;
2793}
2794
2795int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2796{
2797 int ret;
2798 spin_lock(&dlm->ast_lock);
2799 spin_lock(&lock->spinlock);
2800 ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2801 spin_unlock(&lock->spinlock);
2802 spin_unlock(&dlm->ast_lock);
2803 return ret;
2804}
2805
2806static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2807 struct dlm_lock_resource *res,
2808 u8 mig_target)
2809{
2810 int can_proceed;
2811 spin_lock(&res->spinlock);
2812 can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2813 spin_unlock(&res->spinlock);
2814
2815 /* target has died, so make the caller break out of the
2816 * wait_event, but caller must recheck the domain_map */
2817 spin_lock(&dlm->spinlock);
2818 if (!test_bit(mig_target, dlm->domain_map))
2819 can_proceed = 1;
2820 spin_unlock(&dlm->spinlock);
2821 return can_proceed;
2822}
2823
2824static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2825 struct dlm_lock_resource *res)
2826{
2827 int ret;
2828 spin_lock(&res->spinlock);
2829 ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2830 spin_unlock(&res->spinlock);
2831 return ret;
2832}
2833
2834
2835static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2836 struct dlm_lock_resource *res,
2837 u8 target)
2838{
2839 int ret = 0;
2840
2841 mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2842 res->lockname.len, res->lockname.name, dlm->node_num,
2843 target);
2844 /* need to set MIGRATING flag on lockres. this is done by
2845 * ensuring that all asts have been flushed for this lockres. */
2846 spin_lock(&res->spinlock);
2847 BUG_ON(res->migration_pending);
2848 res->migration_pending = 1;
2849 /* strategy is to reserve an extra ast then release
2850 * it below, letting the release do all of the work */
2851 __dlm_lockres_reserve_ast(res);
2852 spin_unlock(&res->spinlock);
2853
2854 /* now flush all the pending asts */
2855 dlm_kick_thread(dlm, res);
2856 /* before waiting on DIRTY, block processes which may
2857 * try to dirty the lockres before MIGRATING is set */
2858 spin_lock(&res->spinlock);
2859 BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2860 res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2861 spin_unlock(&res->spinlock);
2862 /* now wait on any pending asts and the DIRTY state */
2863 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2864 dlm_lockres_release_ast(dlm, res);
2865
2866 mlog(0, "about to wait on migration_wq, dirty=%s\n",
2867 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2868 /* if the extra ref we just put was the final one, this
2869 * will pass thru immediately. otherwise, we need to wait
2870 * for the last ast to finish. */
2871again:
2872 ret = wait_event_interruptible_timeout(dlm->migration_wq,
2873 dlm_migration_can_proceed(dlm, res, target),
2874 msecs_to_jiffies(1000));
2875 if (ret < 0) {
2876 mlog(0, "woken again: migrating? %s, dead? %s\n",
2877 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2878 test_bit(target, dlm->domain_map) ? "no":"yes");
2879 } else {
2880 mlog(0, "all is well: migrating? %s, dead? %s\n",
2881 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2882 test_bit(target, dlm->domain_map) ? "no":"yes");
2883 }
2884 if (!dlm_migration_can_proceed(dlm, res, target)) {
2885 mlog(0, "trying again...\n");
2886 goto again;
2887 }
2888
2889 ret = 0;
2890 /* did the target go down or die? */
2891 spin_lock(&dlm->spinlock);
2892 if (!test_bit(target, dlm->domain_map)) {
2893 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2894 target);
2895 ret = -EHOSTDOWN;
2896 }
2897 spin_unlock(&dlm->spinlock);
2898
2899 /*
2900 * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2901 * another try; otherwise, we are sure the MIGRATING state is there,
2902 * drop the unneeded state which blocked threads trying to DIRTY
2903 */
2904 spin_lock(&res->spinlock);
2905 BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2906 res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2907 if (!ret)
2908 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2909 else
2910 res->migration_pending = 0;
2911 spin_unlock(&res->spinlock);
2912
2913 /*
2914 * at this point:
2915 *
2916 * o the DLM_LOCK_RES_MIGRATING flag is set if target not down
2917 * o there are no pending asts on this lockres
2918 * o all processes trying to reserve an ast on this
2919 * lockres must wait for the MIGRATING flag to clear
2920 */
2921 return ret;
2922}
2923
2924/* last step in the migration process.
2925 * original master calls this to free all of the dlm_lock
2926 * structures that used to be for other nodes. */
2927static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2928 struct dlm_lock_resource *res)
2929{
2930 struct list_head *queue = &res->granted;
2931 int i, bit;
2932 struct dlm_lock *lock, *next;
2933
2934 assert_spin_locked(&res->spinlock);
2935
2936 BUG_ON(res->owner == dlm->node_num);
2937
2938 for (i=0; i<3; i++) {
2939 list_for_each_entry_safe(lock, next, queue, list) {
2940 if (lock->ml.node != dlm->node_num) {
2941 mlog(0, "putting lock for node %u\n",
2942 lock->ml.node);
2943 /* be extra careful */
2944 BUG_ON(!list_empty(&lock->ast_list));
2945 BUG_ON(!list_empty(&lock->bast_list));
2946 BUG_ON(lock->ast_pending);
2947 BUG_ON(lock->bast_pending);
2948 dlm_lockres_clear_refmap_bit(dlm, res,
2949 lock->ml.node);
2950 list_del_init(&lock->list);
2951 dlm_lock_put(lock);
2952 /* In a normal unlock, we would have added a
2953 * DLM_UNLOCK_FREE_LOCK action. Force it. */
2954 dlm_lock_put(lock);
2955 }
2956 }
2957 queue++;
2958 }
2959 bit = 0;
2960 while (1) {
2961 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2962 if (bit >= O2NM_MAX_NODES)
2963 break;
2964 /* do not clear the local node reference, if there is a
2965 * process holding this, let it drop the ref itself */
2966 if (bit != dlm->node_num) {
2967 mlog(0, "%s:%.*s: node %u had a ref to this "
2968 "migrating lockres, clearing\n", dlm->name,
2969 res->lockname.len, res->lockname.name, bit);
2970 dlm_lockres_clear_refmap_bit(dlm, res, bit);
2971 }
2972 bit++;
2973 }
2974}
2975
2976/*
2977 * Pick a node to migrate the lock resource to. This function selects a
2978 * potential target based first on the locks and then on refmap. It skips
2979 * nodes that are in the process of exiting the domain.
2980 */
2981static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2982 struct dlm_lock_resource *res)
2983{
2984 enum dlm_lockres_list idx;
2985 struct list_head *queue = &res->granted;
2986 struct dlm_lock *lock;
2987 int noderef;
2988 u8 nodenum = O2NM_MAX_NODES;
2989
2990 assert_spin_locked(&dlm->spinlock);
2991 assert_spin_locked(&res->spinlock);
2992
2993 /* Go through all the locks */
2994 for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2995 queue = dlm_list_idx_to_ptr(res, idx);
2996 list_for_each_entry(lock, queue, list) {
2997 if (lock->ml.node == dlm->node_num)
2998 continue;
2999 if (test_bit(lock->ml.node, dlm->exit_domain_map))
3000 continue;
3001 nodenum = lock->ml.node;
3002 goto bail;
3003 }
3004 }
3005
3006 /* Go thru the refmap */
3007 noderef = -1;
3008 while (1) {
3009 noderef = find_next_bit(res->refmap, O2NM_MAX_NODES,
3010 noderef + 1);
3011 if (noderef >= O2NM_MAX_NODES)
3012 break;
3013 if (noderef == dlm->node_num)
3014 continue;
3015 if (test_bit(noderef, dlm->exit_domain_map))
3016 continue;
3017 nodenum = noderef;
3018 goto bail;
3019 }
3020
3021bail:
3022 return nodenum;
3023}
3024
3025/* this is called by the new master once all lockres
3026 * data has been received */
3027static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
3028 struct dlm_lock_resource *res,
3029 u8 master, u8 new_master,
3030 struct dlm_node_iter *iter)
3031{
3032 struct dlm_migrate_request migrate;
3033 int ret, skip, status = 0;
3034 int nodenum;
3035
3036 memset(&migrate, 0, sizeof(migrate));
3037 migrate.namelen = res->lockname.len;
3038 memcpy(migrate.name, res->lockname.name, migrate.namelen);
3039 migrate.new_master = new_master;
3040 migrate.master = master;
3041
3042 ret = 0;
3043
3044 /* send message to all nodes, except the master and myself */
3045 while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
3046 if (nodenum == master ||
3047 nodenum == new_master)
3048 continue;
3049
3050 /* We could race exit domain. If exited, skip. */
3051 spin_lock(&dlm->spinlock);
3052 skip = (!test_bit(nodenum, dlm->domain_map));
3053 spin_unlock(&dlm->spinlock);
3054 if (skip) {
3055 clear_bit(nodenum, iter->node_map);
3056 continue;
3057 }
3058
3059 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
3060 &migrate, sizeof(migrate), nodenum,
3061 &status);
3062 if (ret < 0) {
3063 mlog(ML_ERROR, "%s: res %.*s, Error %d send "
3064 "MIGRATE_REQUEST to node %u\n", dlm->name,
3065 migrate.namelen, migrate.name, ret, nodenum);
3066 if (!dlm_is_host_down(ret)) {
3067 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
3068 BUG();
3069 }
3070 clear_bit(nodenum, iter->node_map);
3071 ret = 0;
3072 } else if (status < 0) {
3073 mlog(0, "migrate request (node %u) returned %d!\n",
3074 nodenum, status);
3075 ret = status;
3076 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
3077 /* during the migration request we short-circuited
3078 * the mastery of the lockres. make sure we have
3079 * a mastery ref for nodenum */
3080 mlog(0, "%s:%.*s: need ref for node %u\n",
3081 dlm->name, res->lockname.len, res->lockname.name,
3082 nodenum);
3083 spin_lock(&res->spinlock);
3084 dlm_lockres_set_refmap_bit(dlm, res, nodenum);
3085 spin_unlock(&res->spinlock);
3086 }
3087 }
3088
3089 if (ret < 0)
3090 mlog_errno(ret);
3091
3092 mlog(0, "returning ret=%d\n", ret);
3093 return ret;
3094}
3095
3096
3097/* if there is an existing mle for this lockres, we now know who the master is.
3098 * (the one who sent us *this* message) we can clear it up right away.
3099 * since the process that put the mle on the list still has a reference to it,
3100 * we can unhash it now, set the master and wake the process. as a result,
3101 * we will have no mle in the list to start with. now we can add an mle for
3102 * the migration and this should be the only one found for those scanning the
3103 * list. */
3104int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
3105 void **ret_data)
3106{
3107 struct dlm_ctxt *dlm = data;
3108 struct dlm_lock_resource *res = NULL;
3109 struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
3110 struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
3111 const char *name;
3112 unsigned int namelen, hash;
3113 int ret = 0;
3114
3115 if (!dlm_grab(dlm))
3116 return 0;
3117
3118 name = migrate->name;
3119 namelen = migrate->namelen;
3120 hash = dlm_lockid_hash(name, namelen);
3121
3122 /* preallocate.. if this fails, abort */
3123 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
3124
3125 if (!mle) {
3126 ret = -ENOMEM;
3127 goto leave;
3128 }
3129
3130 /* check for pre-existing lock */
3131 spin_lock(&dlm->spinlock);
3132 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
3133 if (res) {
3134 spin_lock(&res->spinlock);
3135 if (res->state & DLM_LOCK_RES_RECOVERING) {
3136 /* if all is working ok, this can only mean that we got
3137 * a migrate request from a node that we now see as
3138 * dead. what can we do here? drop it to the floor? */
3139 spin_unlock(&res->spinlock);
3140 mlog(ML_ERROR, "Got a migrate request, but the "
3141 "lockres is marked as recovering!");
3142 kmem_cache_free(dlm_mle_cache, mle);
3143 ret = -EINVAL; /* need a better solution */
3144 goto unlock;
3145 }
3146 res->state |= DLM_LOCK_RES_MIGRATING;
3147 spin_unlock(&res->spinlock);
3148 }
3149
3150 spin_lock(&dlm->master_lock);
3151 /* ignore status. only nonzero status would BUG. */
3152 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3153 name, namelen,
3154 migrate->new_master,
3155 migrate->master);
3156
3157 if (ret < 0)
3158 kmem_cache_free(dlm_mle_cache, mle);
3159
3160 spin_unlock(&dlm->master_lock);
3161unlock:
3162 spin_unlock(&dlm->spinlock);
3163
3164 if (oldmle) {
3165 /* master is known, detach if not already detached */
3166 dlm_mle_detach_hb_events(dlm, oldmle);
3167 dlm_put_mle(oldmle);
3168 }
3169
3170 if (res)
3171 dlm_lockres_put(res);
3172leave:
3173 dlm_put(dlm);
3174 return ret;
3175}
3176
3177/* must be holding dlm->spinlock and dlm->master_lock
3178 * when adding a migration mle, we can clear any other mles
3179 * in the master list because we know with certainty that
3180 * the master is "master". so we remove any old mle from
3181 * the list after setting it's master field, and then add
3182 * the new migration mle. this way we can hold with the rule
3183 * of having only one mle for a given lock name at all times. */
3184static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3185 struct dlm_lock_resource *res,
3186 struct dlm_master_list_entry *mle,
3187 struct dlm_master_list_entry **oldmle,
3188 const char *name, unsigned int namelen,
3189 u8 new_master, u8 master)
3190{
3191 int found;
3192 int ret = 0;
3193
3194 *oldmle = NULL;
3195
3196 assert_spin_locked(&dlm->spinlock);
3197 assert_spin_locked(&dlm->master_lock);
3198
3199 /* caller is responsible for any ref taken here on oldmle */
3200 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3201 if (found) {
3202 struct dlm_master_list_entry *tmp = *oldmle;
3203 spin_lock(&tmp->spinlock);
3204 if (tmp->type == DLM_MLE_MIGRATION) {
3205 if (master == dlm->node_num) {
3206 /* ah another process raced me to it */
3207 mlog(0, "tried to migrate %.*s, but some "
3208 "process beat me to it\n",
3209 namelen, name);
3210 spin_unlock(&tmp->spinlock);
3211 return -EEXIST;
3212 } else {
3213 /* bad. 2 NODES are trying to migrate! */
3214 mlog(ML_ERROR, "migration error mle: "
3215 "master=%u new_master=%u // request: "
3216 "master=%u new_master=%u // "
3217 "lockres=%.*s\n",
3218 tmp->master, tmp->new_master,
3219 master, new_master,
3220 namelen, name);
3221 BUG();
3222 }
3223 } else {
3224 /* this is essentially what assert_master does */
3225 tmp->master = master;
3226 atomic_set(&tmp->woken, 1);
3227 wake_up(&tmp->wq);
3228 /* remove it so that only one mle will be found */
3229 __dlm_unlink_mle(dlm, tmp);
3230 __dlm_mle_detach_hb_events(dlm, tmp);
3231 if (tmp->type == DLM_MLE_MASTER) {
3232 ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3233 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3234 "telling master to get ref "
3235 "for cleared out mle during "
3236 "migration\n", dlm->name,
3237 namelen, name, master,
3238 new_master);
3239 }
3240 }
3241 spin_unlock(&tmp->spinlock);
3242 }
3243
3244 /* now add a migration mle to the tail of the list */
3245 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3246 mle->new_master = new_master;
3247 /* the new master will be sending an assert master for this.
3248 * at that point we will get the refmap reference */
3249 mle->master = master;
3250 /* do this for consistency with other mle types */
3251 set_bit(new_master, mle->maybe_map);
3252 __dlm_insert_mle(dlm, mle);
3253
3254 return ret;
3255}
3256
3257/*
3258 * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3259 */
3260static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3261 struct dlm_master_list_entry *mle)
3262{
3263 struct dlm_lock_resource *res;
3264
3265 /* Find the lockres associated to the mle and set its owner to UNK */
3266 res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3267 mle->mnamehash);
3268 if (res) {
3269 spin_unlock(&dlm->master_lock);
3270
3271 /* move lockres onto recovery list */
3272 spin_lock(&res->spinlock);
3273 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3274 dlm_move_lockres_to_recovery_list(dlm, res);
3275 spin_unlock(&res->spinlock);
3276 dlm_lockres_put(res);
3277
3278 /* about to get rid of mle, detach from heartbeat */
3279 __dlm_mle_detach_hb_events(dlm, mle);
3280
3281 /* dump the mle */
3282 spin_lock(&dlm->master_lock);
3283 __dlm_put_mle(mle);
3284 spin_unlock(&dlm->master_lock);
3285 }
3286
3287 return res;
3288}
3289
3290static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3291 struct dlm_master_list_entry *mle)
3292{
3293 __dlm_mle_detach_hb_events(dlm, mle);
3294
3295 spin_lock(&mle->spinlock);
3296 __dlm_unlink_mle(dlm, mle);
3297 atomic_set(&mle->woken, 1);
3298 spin_unlock(&mle->spinlock);
3299
3300 wake_up(&mle->wq);
3301}
3302
3303static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3304 struct dlm_master_list_entry *mle, u8 dead_node)
3305{
3306 int bit;
3307
3308 BUG_ON(mle->type != DLM_MLE_BLOCK);
3309
3310 spin_lock(&mle->spinlock);
3311 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3312 if (bit != dead_node) {
3313 mlog(0, "mle found, but dead node %u would not have been "
3314 "master\n", dead_node);
3315 spin_unlock(&mle->spinlock);
3316 } else {
3317 /* Must drop the refcount by one since the assert_master will
3318 * never arrive. This may result in the mle being unlinked and
3319 * freed, but there may still be a process waiting in the
3320 * dlmlock path which is fine. */
3321 mlog(0, "node %u was expected master\n", dead_node);
3322 atomic_set(&mle->woken, 1);
3323 spin_unlock(&mle->spinlock);
3324 wake_up(&mle->wq);
3325
3326 /* Do not need events any longer, so detach from heartbeat */
3327 __dlm_mle_detach_hb_events(dlm, mle);
3328 __dlm_put_mle(mle);
3329 }
3330}
3331
3332void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3333{
3334 struct dlm_master_list_entry *mle;
3335 struct dlm_lock_resource *res;
3336 struct hlist_head *bucket;
3337 struct hlist_node *tmp;
3338 unsigned int i;
3339
3340 mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
3341top:
3342 assert_spin_locked(&dlm->spinlock);
3343
3344 /* clean the master list */
3345 spin_lock(&dlm->master_lock);
3346 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3347 bucket = dlm_master_hash(dlm, i);
3348 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3349 BUG_ON(mle->type != DLM_MLE_BLOCK &&
3350 mle->type != DLM_MLE_MASTER &&
3351 mle->type != DLM_MLE_MIGRATION);
3352
3353 /* MASTER mles are initiated locally. The waiting
3354 * process will notice the node map change shortly.
3355 * Let that happen as normal. */
3356 if (mle->type == DLM_MLE_MASTER)
3357 continue;
3358
3359 /* BLOCK mles are initiated by other nodes. Need to
3360 * clean up if the dead node would have been the
3361 * master. */
3362 if (mle->type == DLM_MLE_BLOCK) {
3363 dlm_clean_block_mle(dlm, mle, dead_node);
3364 continue;
3365 }
3366
3367 /* Everything else is a MIGRATION mle */
3368
3369 /* The rule for MIGRATION mles is that the master
3370 * becomes UNKNOWN if *either* the original or the new
3371 * master dies. All UNKNOWN lockres' are sent to
3372 * whichever node becomes the recovery master. The new
3373 * master is responsible for determining if there is
3374 * still a master for this lockres, or if he needs to
3375 * take over mastery. Either way, this node should
3376 * expect another message to resolve this. */
3377
3378 if (mle->master != dead_node &&
3379 mle->new_master != dead_node)
3380 continue;
3381
3382 if (mle->new_master == dead_node && mle->inuse) {
3383 mlog(ML_NOTICE, "%s: target %u died during "
3384 "migration from %u, the MLE is "
3385 "still keep used, ignore it!\n",
3386 dlm->name, dead_node,
3387 mle->master);
3388 continue;
3389 }
3390
3391 /* If we have reached this point, this mle needs to be
3392 * removed from the list and freed. */
3393 dlm_clean_migration_mle(dlm, mle);
3394
3395 mlog(0, "%s: node %u died during migration from "
3396 "%u to %u!\n", dlm->name, dead_node, mle->master,
3397 mle->new_master);
3398
3399 /* If we find a lockres associated with the mle, we've
3400 * hit this rare case that messes up our lock ordering.
3401 * If so, we need to drop the master lock so that we can
3402 * take the lockres lock, meaning that we will have to
3403 * restart from the head of list. */
3404 res = dlm_reset_mleres_owner(dlm, mle);
3405 if (res)
3406 /* restart */
3407 goto top;
3408
3409 /* This may be the last reference */
3410 __dlm_put_mle(mle);
3411 }
3412 }
3413 spin_unlock(&dlm->master_lock);
3414}
3415
3416int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3417 u8 old_master)
3418{
3419 struct dlm_node_iter iter;
3420 int ret = 0;
3421
3422 spin_lock(&dlm->spinlock);
3423 dlm_node_iter_init(dlm->domain_map, &iter);
3424 clear_bit(old_master, iter.node_map);
3425 clear_bit(dlm->node_num, iter.node_map);
3426 spin_unlock(&dlm->spinlock);
3427
3428 /* ownership of the lockres is changing. account for the
3429 * mastery reference here since old_master will briefly have
3430 * a reference after the migration completes */
3431 spin_lock(&res->spinlock);
3432 dlm_lockres_set_refmap_bit(dlm, res, old_master);
3433 spin_unlock(&res->spinlock);
3434
3435 mlog(0, "now time to do a migrate request to other nodes\n");
3436 ret = dlm_do_migrate_request(dlm, res, old_master,
3437 dlm->node_num, &iter);
3438 if (ret < 0) {
3439 mlog_errno(ret);
3440 goto leave;
3441 }
3442
3443 mlog(0, "doing assert master of %.*s to all except the original node\n",
3444 res->lockname.len, res->lockname.name);
3445 /* this call now finishes out the nodemap
3446 * even if one or more nodes die */
3447 ret = dlm_do_assert_master(dlm, res, iter.node_map,
3448 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3449 if (ret < 0) {
3450 /* no longer need to retry. all living nodes contacted. */
3451 mlog_errno(ret);
3452 ret = 0;
3453 }
3454
3455 memset(iter.node_map, 0, sizeof(iter.node_map));
3456 set_bit(old_master, iter.node_map);
3457 mlog(0, "doing assert master of %.*s back to %u\n",
3458 res->lockname.len, res->lockname.name, old_master);
3459 ret = dlm_do_assert_master(dlm, res, iter.node_map,
3460 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3461 if (ret < 0) {
3462 mlog(0, "assert master to original master failed "
3463 "with %d.\n", ret);
3464 /* the only nonzero status here would be because of
3465 * a dead original node. we're done. */
3466 ret = 0;
3467 }
3468
3469 /* all done, set the owner, clear the flag */
3470 spin_lock(&res->spinlock);
3471 dlm_set_lockres_owner(dlm, res, dlm->node_num);
3472 res->state &= ~DLM_LOCK_RES_MIGRATING;
3473 spin_unlock(&res->spinlock);
3474 /* re-dirty it on the new master */
3475 dlm_kick_thread(dlm, res);
3476 wake_up(&res->wq);
3477leave:
3478 return ret;
3479}
3480
3481/*
3482 * LOCKRES AST REFCOUNT
3483 * this is integral to migration
3484 */
3485
3486/* for future intent to call an ast, reserve one ahead of time.
3487 * this should be called only after waiting on the lockres
3488 * with dlm_wait_on_lockres, and while still holding the
3489 * spinlock after the call. */
3490void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3491{
3492 assert_spin_locked(&res->spinlock);
3493 if (res->state & DLM_LOCK_RES_MIGRATING) {
3494 __dlm_print_one_lock_resource(res);
3495 }
3496 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3497
3498 atomic_inc(&res->asts_reserved);
3499}
3500
3501/*
3502 * used to drop the reserved ast, either because it went unused,
3503 * or because the ast/bast was actually called.
3504 *
3505 * also, if there is a pending migration on this lockres,
3506 * and this was the last pending ast on the lockres,
3507 * atomically set the MIGRATING flag before we drop the lock.
3508 * this is how we ensure that migration can proceed with no
3509 * asts in progress. note that it is ok if the state of the
3510 * queues is such that a lock should be granted in the future
3511 * or that a bast should be fired, because the new master will
3512 * shuffle the lists on this lockres as soon as it is migrated.
3513 */
3514void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3515 struct dlm_lock_resource *res)
3516{
3517 if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3518 return;
3519
3520 if (!res->migration_pending) {
3521 spin_unlock(&res->spinlock);
3522 return;
3523 }
3524
3525 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3526 res->migration_pending = 0;
3527 res->state |= DLM_LOCK_RES_MIGRATING;
3528 spin_unlock(&res->spinlock);
3529 wake_up(&res->wq);
3530 wake_up(&dlm->migration_wq);
3531}
3532
3533void dlm_force_free_mles(struct dlm_ctxt *dlm)
3534{
3535 int i;
3536 struct hlist_head *bucket;
3537 struct dlm_master_list_entry *mle;
3538 struct hlist_node *tmp;
3539
3540 /*
3541 * We notified all other nodes that we are exiting the domain and
3542 * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3543 * around we force free them and wake any processes that are waiting
3544 * on the mles
3545 */
3546 spin_lock(&dlm->spinlock);
3547 spin_lock(&dlm->master_lock);
3548
3549 BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3550 BUG_ON((find_next_bit(dlm->domain_map, O2NM_MAX_NODES, 0) < O2NM_MAX_NODES));
3551
3552 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3553 bucket = dlm_master_hash(dlm, i);
3554 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3555 if (mle->type != DLM_MLE_BLOCK) {
3556 mlog(ML_ERROR, "bad mle: %p\n", mle);
3557 dlm_print_one_mle(mle);
3558 }
3559 atomic_set(&mle->woken, 1);
3560 wake_up(&mle->wq);
3561
3562 __dlm_unlink_mle(dlm, mle);
3563 __dlm_mle_detach_hb_events(dlm, mle);
3564 __dlm_put_mle(mle);
3565 }
3566 }
3567 spin_unlock(&dlm->master_lock);
3568 spin_unlock(&dlm->spinlock);
3569}
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmmod.c
5 *
6 * standalone DLM module
7 *
8 * Copyright (C) 2004 Oracle. All rights reserved.
9 *
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.
14 *
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.
19 *
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.
24 *
25 */
26
27
28#include <linux/module.h>
29#include <linux/fs.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>
41
42
43#include "cluster/heartbeat.h"
44#include "cluster/nodemanager.h"
45#include "cluster/tcp.h"
46
47#include "dlmapi.h"
48#include "dlmcommon.h"
49#include "dlmdomain.h"
50#include "dlmdebug.h"
51
52#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
53#include "cluster/masklog.h"
54
55static void dlm_mle_node_down(struct dlm_ctxt *dlm,
56 struct dlm_master_list_entry *mle,
57 struct o2nm_node *node,
58 int idx);
59static void dlm_mle_node_up(struct dlm_ctxt *dlm,
60 struct dlm_master_list_entry *mle,
61 struct o2nm_node *node,
62 int idx);
63
64static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
65static int dlm_do_assert_master(struct dlm_ctxt *dlm,
66 struct dlm_lock_resource *res,
67 void *nodemap, u32 flags);
68static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
69
70static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
71 struct dlm_master_list_entry *mle,
72 const char *name,
73 unsigned int namelen)
74{
75 if (dlm != mle->dlm)
76 return 0;
77
78 if (namelen != mle->mnamelen ||
79 memcmp(name, mle->mname, namelen) != 0)
80 return 0;
81
82 return 1;
83}
84
85static struct kmem_cache *dlm_lockres_cache;
86static struct kmem_cache *dlm_lockname_cache;
87static struct kmem_cache *dlm_mle_cache;
88
89static void dlm_mle_release(struct kref *kref);
90static void dlm_init_mle(struct dlm_master_list_entry *mle,
91 enum dlm_mle_type type,
92 struct dlm_ctxt *dlm,
93 struct dlm_lock_resource *res,
94 const char *name,
95 unsigned int namelen);
96static void dlm_put_mle(struct dlm_master_list_entry *mle);
97static void __dlm_put_mle(struct dlm_master_list_entry *mle);
98static int dlm_find_mle(struct dlm_ctxt *dlm,
99 struct dlm_master_list_entry **mle,
100 char *name, unsigned int namelen);
101
102static int dlm_do_master_request(struct dlm_lock_resource *res,
103 struct dlm_master_list_entry *mle, int to);
104
105
106static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
107 struct dlm_lock_resource *res,
108 struct dlm_master_list_entry *mle,
109 int *blocked);
110static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
111 struct dlm_lock_resource *res,
112 struct dlm_master_list_entry *mle,
113 int blocked);
114static 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);
120
121static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
122 struct dlm_lock_resource *res);
123static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
124 struct dlm_lock_resource *res);
125static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
126 struct dlm_lock_resource *res,
127 u8 target);
128static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
129 struct dlm_lock_resource *res);
130
131
132int dlm_is_host_down(int errno)
133{
134 switch (errno) {
135 case -EBADF:
136 case -ECONNREFUSED:
137 case -ENOTCONN:
138 case -ECONNRESET:
139 case -EPIPE:
140 case -EHOSTDOWN:
141 case -EHOSTUNREACH:
142 case -ETIMEDOUT:
143 case -ECONNABORTED:
144 case -ENETDOWN:
145 case -ENETUNREACH:
146 case -ENETRESET:
147 case -ESHUTDOWN:
148 case -ENOPROTOOPT:
149 case -EINVAL: /* if returned from our tcp code,
150 this means there is no socket */
151 return 1;
152 }
153 return 0;
154}
155
156
157/*
158 * MASTER LIST FUNCTIONS
159 */
160
161
162/*
163 * regarding master list entries and heartbeat callbacks:
164 *
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.
173 *
174 * as a general rule, heartbeat events are no longer needed by
175 * the mle once an "answer" regarding the lock master has been
176 * received.
177 */
178static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
179 struct dlm_master_list_entry *mle)
180{
181 assert_spin_locked(&dlm->spinlock);
182
183 list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
184}
185
186
187static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
188 struct dlm_master_list_entry *mle)
189{
190 if (!list_empty(&mle->hb_events))
191 list_del_init(&mle->hb_events);
192}
193
194
195static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
196 struct dlm_master_list_entry *mle)
197{
198 spin_lock(&dlm->spinlock);
199 __dlm_mle_detach_hb_events(dlm, mle);
200 spin_unlock(&dlm->spinlock);
201}
202
203static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
204{
205 struct dlm_ctxt *dlm;
206 dlm = mle->dlm;
207
208 assert_spin_locked(&dlm->spinlock);
209 assert_spin_locked(&dlm->master_lock);
210 mle->inuse++;
211 kref_get(&mle->mle_refs);
212}
213
214static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
215{
216 struct dlm_ctxt *dlm;
217 dlm = mle->dlm;
218
219 spin_lock(&dlm->spinlock);
220 spin_lock(&dlm->master_lock);
221 mle->inuse--;
222 __dlm_put_mle(mle);
223 spin_unlock(&dlm->master_lock);
224 spin_unlock(&dlm->spinlock);
225
226}
227
228/* remove from list and free */
229static void __dlm_put_mle(struct dlm_master_list_entry *mle)
230{
231 struct dlm_ctxt *dlm;
232 dlm = mle->dlm;
233
234 assert_spin_locked(&dlm->spinlock);
235 assert_spin_locked(&dlm->master_lock);
236 if (!atomic_read(&mle->mle_refs.refcount)) {
237 /* this may or may not crash, but who cares.
238 * it's a BUG. */
239 mlog(ML_ERROR, "bad mle: %p\n", mle);
240 dlm_print_one_mle(mle);
241 BUG();
242 } else
243 kref_put(&mle->mle_refs, dlm_mle_release);
244}
245
246
247/* must not have any spinlocks coming in */
248static void dlm_put_mle(struct dlm_master_list_entry *mle)
249{
250 struct dlm_ctxt *dlm;
251 dlm = mle->dlm;
252
253 spin_lock(&dlm->spinlock);
254 spin_lock(&dlm->master_lock);
255 __dlm_put_mle(mle);
256 spin_unlock(&dlm->master_lock);
257 spin_unlock(&dlm->spinlock);
258}
259
260static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
261{
262 kref_get(&mle->mle_refs);
263}
264
265static 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,
269 const char *name,
270 unsigned int namelen)
271{
272 assert_spin_locked(&dlm->spinlock);
273
274 mle->dlm = dlm;
275 mle->type = type;
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;
286 mle->inuse = 0;
287
288 BUG_ON(mle->type != DLM_MLE_BLOCK &&
289 mle->type != DLM_MLE_MASTER &&
290 mle->type != DLM_MLE_MIGRATION);
291
292 if (mle->type == DLM_MLE_MASTER) {
293 BUG_ON(!res);
294 mle->mleres = res;
295 memcpy(mle->mname, res->lockname.name, res->lockname.len);
296 mle->mnamelen = res->lockname.len;
297 mle->mnamehash = res->lockname.hash;
298 } else {
299 BUG_ON(!name);
300 mle->mleres = NULL;
301 memcpy(mle->mname, name, namelen);
302 mle->mnamelen = namelen;
303 mle->mnamehash = dlm_lockid_hash(name, namelen);
304 }
305
306 atomic_inc(&dlm->mle_tot_count[mle->type]);
307 atomic_inc(&dlm->mle_cur_count[mle->type]);
308
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);
314
315 /* attach the mle to the domain node up/down events */
316 __dlm_mle_attach_hb_events(dlm, mle);
317}
318
319void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
320{
321 assert_spin_locked(&dlm->spinlock);
322 assert_spin_locked(&dlm->master_lock);
323
324 if (!hlist_unhashed(&mle->master_hash_node))
325 hlist_del_init(&mle->master_hash_node);
326}
327
328void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
329{
330 struct hlist_head *bucket;
331
332 assert_spin_locked(&dlm->master_lock);
333
334 bucket = dlm_master_hash(dlm, mle->mnamehash);
335 hlist_add_head(&mle->master_hash_node, bucket);
336}
337
338/* returns 1 if found, 0 if not */
339static int dlm_find_mle(struct dlm_ctxt *dlm,
340 struct dlm_master_list_entry **mle,
341 char *name, unsigned int namelen)
342{
343 struct dlm_master_list_entry *tmpmle;
344 struct hlist_head *bucket;
345 unsigned int hash;
346
347 assert_spin_locked(&dlm->master_lock);
348
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))
353 continue;
354 dlm_get_mle(tmpmle);
355 *mle = tmpmle;
356 return 1;
357 }
358 return 0;
359}
360
361void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
362{
363 struct dlm_master_list_entry *mle;
364
365 assert_spin_locked(&dlm->spinlock);
366
367 list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
368 if (node_up)
369 dlm_mle_node_up(dlm, mle, NULL, idx);
370 else
371 dlm_mle_node_down(dlm, mle, NULL, idx);
372 }
373}
374
375static void dlm_mle_node_down(struct dlm_ctxt *dlm,
376 struct dlm_master_list_entry *mle,
377 struct o2nm_node *node, int idx)
378{
379 spin_lock(&mle->spinlock);
380
381 if (!test_bit(idx, mle->node_map))
382 mlog(0, "node %u already removed from nodemap!\n", idx);
383 else
384 clear_bit(idx, mle->node_map);
385
386 spin_unlock(&mle->spinlock);
387}
388
389static void dlm_mle_node_up(struct dlm_ctxt *dlm,
390 struct dlm_master_list_entry *mle,
391 struct o2nm_node *node, int idx)
392{
393 spin_lock(&mle->spinlock);
394
395 if (test_bit(idx, mle->node_map))
396 mlog(0, "node %u already in node map!\n", idx);
397 else
398 set_bit(idx, mle->node_map);
399
400 spin_unlock(&mle->spinlock);
401}
402
403
404int dlm_init_mle_cache(void)
405{
406 dlm_mle_cache = kmem_cache_create("o2dlm_mle",
407 sizeof(struct dlm_master_list_entry),
408 0, SLAB_HWCACHE_ALIGN,
409 NULL);
410 if (dlm_mle_cache == NULL)
411 return -ENOMEM;
412 return 0;
413}
414
415void dlm_destroy_mle_cache(void)
416{
417 if (dlm_mle_cache)
418 kmem_cache_destroy(dlm_mle_cache);
419}
420
421static void dlm_mle_release(struct kref *kref)
422{
423 struct dlm_master_list_entry *mle;
424 struct dlm_ctxt *dlm;
425
426 mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
427 dlm = mle->dlm;
428
429 assert_spin_locked(&dlm->spinlock);
430 assert_spin_locked(&dlm->master_lock);
431
432 mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
433 mle->type);
434
435 /* remove from list if not already */
436 __dlm_unlink_mle(dlm, mle);
437
438 /* detach the mle from the domain node up/down events */
439 __dlm_mle_detach_hb_events(dlm, mle);
440
441 atomic_dec(&dlm->mle_cur_count[mle->type]);
442
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);
446}
447
448
449/*
450 * LOCK RESOURCE FUNCTIONS
451 */
452
453int dlm_init_master_caches(void)
454{
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)
459 goto bail;
460
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)
465 goto bail;
466
467 return 0;
468bail:
469 dlm_destroy_master_caches();
470 return -ENOMEM;
471}
472
473void dlm_destroy_master_caches(void)
474{
475 if (dlm_lockname_cache) {
476 kmem_cache_destroy(dlm_lockname_cache);
477 dlm_lockname_cache = NULL;
478 }
479
480 if (dlm_lockres_cache) {
481 kmem_cache_destroy(dlm_lockres_cache);
482 dlm_lockres_cache = NULL;
483 }
484}
485
486static void dlm_lockres_release(struct kref *kref)
487{
488 struct dlm_lock_resource *res;
489 struct dlm_ctxt *dlm;
490
491 res = container_of(kref, struct dlm_lock_resource, refs);
492 dlm = res->dlm;
493
494 /* This should not happen -- all lockres' have a name
495 * associated with them at init time. */
496 BUG_ON(!res->lockname.name);
497
498 mlog(0, "destroying lockres %.*s\n", res->lockname.len,
499 res->lockname.name);
500
501 atomic_dec(&dlm->res_cur_count);
502
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)) {
510 mlog(ML_ERROR,
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' : ' ');
521
522 dlm_print_one_lock_resource(res);
523 }
524
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));
534
535 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
536
537 kmem_cache_free(dlm_lockres_cache, res);
538}
539
540void dlm_lockres_put(struct dlm_lock_resource *res)
541{
542 kref_put(&res->refs, dlm_lockres_release);
543}
544
545static void dlm_init_lockres(struct dlm_ctxt *dlm,
546 struct dlm_lock_resource *res,
547 const char *name, unsigned int namelen)
548{
549 char *qname;
550
551 /* If we memset here, we lose our reference to the kmalloc'd
552 * res->lockname.name, so be sure to init every field
553 * correctly! */
554
555 qname = (char *) res->lockname.name;
556 memcpy(qname, name, namelen);
557
558 res->lockname.len = namelen;
559 res->lockname.hash = dlm_lockid_hash(name, namelen);
560
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;
575
576 res->dlm = dlm;
577
578 kref_init(&res->refs);
579
580 atomic_inc(&dlm->res_tot_count);
581 atomic_inc(&dlm->res_cur_count);
582
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);
587
588 res->state = DLM_LOCK_RES_IN_PROGRESS;
589
590 res->last_used = 0;
591
592 spin_lock(&dlm->spinlock);
593 list_add_tail(&res->tracking, &dlm->tracking_list);
594 spin_unlock(&dlm->spinlock);
595
596 memset(res->lvb, 0, DLM_LVB_LEN);
597 memset(res->refmap, 0, sizeof(res->refmap));
598}
599
600struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
601 const char *name,
602 unsigned int namelen)
603{
604 struct dlm_lock_resource *res = NULL;
605
606 res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
607 if (!res)
608 goto error;
609
610 res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
611 if (!res->lockname.name)
612 goto error;
613
614 dlm_init_lockres(dlm, res, name, namelen);
615 return res;
616
617error:
618 if (res)
619 kmem_cache_free(dlm_lockres_cache, res);
620 return NULL;
621}
622
623void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm,
624 struct dlm_lock_resource *res, int bit)
625{
626 assert_spin_locked(&res->spinlock);
627
628 mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len,
629 res->lockname.name, bit, __builtin_return_address(0));
630
631 set_bit(bit, res->refmap);
632}
633
634void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm,
635 struct dlm_lock_resource *res, int bit)
636{
637 assert_spin_locked(&res->spinlock);
638
639 mlog(0, "res %.*s, clr node %u, %ps()\n", res->lockname.len,
640 res->lockname.name, bit, __builtin_return_address(0));
641
642 clear_bit(bit, res->refmap);
643}
644
645static void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
646 struct dlm_lock_resource *res)
647{
648 res->inflight_locks++;
649
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));
653}
654
655void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
656 struct dlm_lock_resource *res)
657{
658 assert_spin_locked(&res->spinlock);
659 __dlm_lockres_grab_inflight_ref(dlm, res);
660}
661
662void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
663 struct dlm_lock_resource *res)
664{
665 assert_spin_locked(&res->spinlock);
666
667 BUG_ON(res->inflight_locks == 0);
668
669 res->inflight_locks--;
670
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));
674
675 wake_up(&res->wq);
676}
677
678void __dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm,
679 struct dlm_lock_resource *res)
680{
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);
686}
687
688static void __dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
689 struct dlm_lock_resource *res)
690{
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);
697}
698
699static void dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
700 struct dlm_lock_resource *res)
701{
702 spin_lock(&res->spinlock);
703 __dlm_lockres_drop_inflight_worker(dlm, res);
704 spin_unlock(&res->spinlock);
705}
706
707/*
708 * lookup a lock resource by name.
709 * may already exist in the hashtable.
710 * lockid is null terminated
711 *
712 * if not, allocate enough for the lockres and for
713 * the temporary structure used in doing the mastering.
714 *
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).
721 *
722 */
723struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
724 const char *lockid,
725 int namelen,
726 int flags)
727{
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;
731 int blocked = 0;
732 int ret, nodenum;
733 struct dlm_node_iter iter;
734 unsigned int hash;
735 int tries = 0;
736 int bit, wait_on_recovery = 0;
737
738 BUG_ON(!lockid);
739
740 hash = dlm_lockid_hash(lockid, namelen);
741
742 mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
743
744lookup:
745 spin_lock(&dlm->spinlock);
746 tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
747 if (tmpres) {
748 spin_unlock(&dlm->spinlock);
749 spin_lock(&tmpres->spinlock);
750
751 /*
752 * Right after dlm spinlock was released, dlm_thread could have
753 * purged the lockres. Check if lockres got unhashed. If so
754 * start over.
755 */
756 if (hlist_unhashed(&tmpres->hash_node)) {
757 spin_unlock(&tmpres->spinlock);
758 dlm_lockres_put(tmpres);
759 tmpres = NULL;
760 goto lookup;
761 }
762
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);
769 tmpres = NULL;
770 goto lookup;
771 }
772
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);
780 tmpres = NULL;
781 goto lookup;
782 }
783
784 /* Grab inflight ref to pin the resource */
785 dlm_lockres_grab_inflight_ref(dlm, tmpres);
786
787 spin_unlock(&tmpres->spinlock);
788 if (res) {
789 spin_lock(&dlm->track_lock);
790 if (!list_empty(&res->tracking))
791 list_del_init(&res->tracking);
792 else
793 mlog(ML_ERROR, "Resource %.*s not "
794 "on the Tracking list\n",
795 res->lockname.len,
796 res->lockname.name);
797 spin_unlock(&dlm->track_lock);
798 dlm_lockres_put(res);
799 }
800 res = tmpres;
801 goto leave;
802 }
803
804 if (!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);
809 if (!alloc_mle)
810 goto leave;
811 res = dlm_new_lockres(dlm, lockid, namelen);
812 if (!res)
813 goto leave;
814 goto lookup;
815 }
816
817 mlog(0, "no lockres found, allocated our own: %p\n", res);
818
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 */
829 goto wake_waiters;
830 }
831
832 /* check master list to see if another node has started mastering it */
833 spin_lock(&dlm->master_lock);
834
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);
837 if (blocked) {
838 int mig;
839 if (mle->type == DLM_MLE_MASTER) {
840 mlog(ML_ERROR, "master entry for nonexistent lock!\n");
841 BUG();
842 }
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);
861
862 /* master is known, detach */
863 if (!mig)
864 dlm_mle_detach_hb_events(dlm, mle);
865 dlm_put_mle(mle);
866 mle = NULL;
867 /* this is lame, but we can't wait on either
868 * the mle or lockres waitqueue here */
869 if (mig)
870 msleep(100);
871 goto lookup;
872 }
873 } else {
874 /* go ahead and try to master lock on this node */
875 mle = alloc_mle;
876 /* make sure this does not get freed below */
877 alloc_mle = NULL;
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);
881
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;
892 }
893 }
894
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. */
899
900 /* finally add the lockres to its hash bucket */
901 __dlm_insert_lockres(dlm, res);
902
903 /* since this lockres is new it doesn't not require the spinlock */
904 __dlm_lockres_grab_inflight_ref(dlm, res);
905
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);
913
914redo_request:
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;
924 else {
925 mlog(0, "%s: waiting 500ms for heartbeat state "
926 "change\n", dlm->name);
927 msleep(500);
928 }
929 continue;
930 }
931
932 dlm_kick_recovery_thread(dlm);
933 msleep(1000);
934 dlm_wait_for_recovery(dlm);
935
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;
943 } else
944 wait_on_recovery = 0;
945 spin_unlock(&dlm->spinlock);
946
947 if (wait_on_recovery)
948 dlm_wait_for_node_recovery(dlm, bit, 10000);
949 }
950
951 /* must wait for lock to be mastered elsewhere */
952 if (blocked)
953 goto wait;
954
955 ret = -EINVAL;
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);
959 if (ret < 0)
960 mlog_errno(ret);
961 if (mle->master != O2NM_MAX_NODES) {
962 /* found a master ! */
963 if (mle->master <= nodenum)
964 break;
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);
972 }
973 }
974
975wait:
976 /* keep going until the response map includes all nodes */
977 ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
978 if (ret < 0) {
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);
983 if (++tries > 20) {
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);
990 tries = 0;
991 }
992 goto redo_request;
993 }
994
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);
999
1000 /* master is known, detach if not already detached */
1001 dlm_mle_detach_hb_events(dlm, mle);
1002 dlm_put_mle(mle);
1003 /* put the extra ref */
1004 dlm_put_mle_inuse(mle);
1005
1006wake_waiters:
1007 spin_lock(&res->spinlock);
1008 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1009 spin_unlock(&res->spinlock);
1010 wake_up(&res->wq);
1011
1012leave:
1013 /* need to free the unused mle */
1014 if (alloc_mle)
1015 kmem_cache_free(dlm_mle_cache, alloc_mle);
1016
1017 return res;
1018}
1019
1020
1021#define DLM_MASTERY_TIMEOUT_MS 5000
1022
1023static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
1024 struct dlm_lock_resource *res,
1025 struct dlm_master_list_entry *mle,
1026 int *blocked)
1027{
1028 u8 m;
1029 int ret, bit;
1030 int map_changed, voting_done;
1031 int assert, sleep;
1032
1033recheck:
1034 ret = 0;
1035 assert = 0;
1036
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);
1047 if (ret < 0) {
1048 /* give recovery a chance to run */
1049 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1050 msleep(500);
1051 goto recheck;
1052 }
1053 }
1054 ret = 0;
1055 goto leave;
1056 }
1057 spin_unlock(&res->spinlock);
1058
1059 spin_lock(&mle->spinlock);
1060 m = mle->master;
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);
1065
1066 /* restart if we hit any errors */
1067 if (map_changed) {
1068 int b;
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,
1076 *blocked, b);
1077 *blocked = b;
1078 }
1079 spin_unlock(&mle->spinlock);
1080 if (ret < 0) {
1081 mlog_errno(ret);
1082 goto leave;
1083 }
1084 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1085 "rechecking now\n", dlm->name, res->lockname.len,
1086 res->lockname.name);
1087 goto recheck;
1088 } else {
1089 if (!voting_done) {
1090 mlog(0, "map not changed and voting not done "
1091 "for %s:%.*s\n", dlm->name, res->lockname.len,
1092 res->lockname.name);
1093 }
1094 }
1095
1096 if (m != O2NM_MAX_NODES) {
1097 /* another node has done an assert!
1098 * all done! */
1099 sleep = 0;
1100 } else {
1101 sleep = 1;
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 */
1112 assert = 1;
1113 sleep = 0;
1114 }
1115 /* if voting is done, but we have not received
1116 * an assert master yet, we must sleep */
1117 }
1118 }
1119
1120 spin_unlock(&mle->spinlock);
1121
1122 /* sleep if we haven't finished voting yet */
1123 if (sleep) {
1124 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1125
1126 /*
1127 if (atomic_read(&mle->mle_refs.refcount) < 2)
1128 mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1129 atomic_read(&mle->mle_refs.refcount),
1130 res->lockname.len, res->lockname.name);
1131 */
1132 atomic_set(&mle->woken, 0);
1133 (void)wait_event_timeout(mle->wq,
1134 (atomic_read(&mle->woken) == 1),
1135 timeo);
1136 if (res->owner == O2NM_MAX_NODES) {
1137 mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1138 res->lockname.len, res->lockname.name);
1139 goto recheck;
1140 }
1141 mlog(0, "done waiting, master is %u\n", res->owner);
1142 ret = 0;
1143 goto leave;
1144 }
1145
1146 ret = 0; /* done */
1147 if (assert) {
1148 m = dlm->node_num;
1149 mlog(0, "about to master %.*s here, this=%u\n",
1150 res->lockname.len, res->lockname.name, m);
1151 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1152 if (ret) {
1153 /* This is a failure in the network path,
1154 * not in the response to the assert_master
1155 * (any nonzero response is a BUG on this node).
1156 * Most likely a socket just got disconnected
1157 * due to node death. */
1158 mlog_errno(ret);
1159 }
1160 /* no longer need to restart lock mastery.
1161 * all living nodes have been contacted. */
1162 ret = 0;
1163 }
1164
1165 /* set the lockres owner */
1166 spin_lock(&res->spinlock);
1167 /* mastery reference obtained either during
1168 * assert_master_handler or in get_lock_resource */
1169 dlm_change_lockres_owner(dlm, res, m);
1170 spin_unlock(&res->spinlock);
1171
1172leave:
1173 return ret;
1174}
1175
1176struct dlm_bitmap_diff_iter
1177{
1178 int curnode;
1179 unsigned long *orig_bm;
1180 unsigned long *cur_bm;
1181 unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1182};
1183
1184enum dlm_node_state_change
1185{
1186 NODE_DOWN = -1,
1187 NODE_NO_CHANGE = 0,
1188 NODE_UP
1189};
1190
1191static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1192 unsigned long *orig_bm,
1193 unsigned long *cur_bm)
1194{
1195 unsigned long p1, p2;
1196 int i;
1197
1198 iter->curnode = -1;
1199 iter->orig_bm = orig_bm;
1200 iter->cur_bm = cur_bm;
1201
1202 for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1203 p1 = *(iter->orig_bm + i);
1204 p2 = *(iter->cur_bm + i);
1205 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1206 }
1207}
1208
1209static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1210 enum dlm_node_state_change *state)
1211{
1212 int bit;
1213
1214 if (iter->curnode >= O2NM_MAX_NODES)
1215 return -ENOENT;
1216
1217 bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1218 iter->curnode+1);
1219 if (bit >= O2NM_MAX_NODES) {
1220 iter->curnode = O2NM_MAX_NODES;
1221 return -ENOENT;
1222 }
1223
1224 /* if it was there in the original then this node died */
1225 if (test_bit(bit, iter->orig_bm))
1226 *state = NODE_DOWN;
1227 else
1228 *state = NODE_UP;
1229
1230 iter->curnode = bit;
1231 return bit;
1232}
1233
1234
1235static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1236 struct dlm_lock_resource *res,
1237 struct dlm_master_list_entry *mle,
1238 int blocked)
1239{
1240 struct dlm_bitmap_diff_iter bdi;
1241 enum dlm_node_state_change sc;
1242 int node;
1243 int ret = 0;
1244
1245 mlog(0, "something happened such that the "
1246 "master process may need to be restarted!\n");
1247
1248 assert_spin_locked(&mle->spinlock);
1249
1250 dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1251 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1252 while (node >= 0) {
1253 if (sc == NODE_UP) {
1254 /* a node came up. clear any old vote from
1255 * the response map and set it in the vote map
1256 * then restart the mastery. */
1257 mlog(ML_NOTICE, "node %d up while restarting\n", node);
1258
1259 /* redo the master request, but only for the new node */
1260 mlog(0, "sending request to new node\n");
1261 clear_bit(node, mle->response_map);
1262 set_bit(node, mle->vote_map);
1263 } else {
1264 mlog(ML_ERROR, "node down! %d\n", node);
1265 if (blocked) {
1266 int lowest = find_next_bit(mle->maybe_map,
1267 O2NM_MAX_NODES, 0);
1268
1269 /* act like it was never there */
1270 clear_bit(node, mle->maybe_map);
1271
1272 if (node == lowest) {
1273 mlog(0, "expected master %u died"
1274 " while this node was blocked "
1275 "waiting on it!\n", node);
1276 lowest = find_next_bit(mle->maybe_map,
1277 O2NM_MAX_NODES,
1278 lowest+1);
1279 if (lowest < O2NM_MAX_NODES) {
1280 mlog(0, "%s:%.*s:still "
1281 "blocked. waiting on %u "
1282 "now\n", dlm->name,
1283 res->lockname.len,
1284 res->lockname.name,
1285 lowest);
1286 } else {
1287 /* mle is an MLE_BLOCK, but
1288 * there is now nothing left to
1289 * block on. we need to return
1290 * all the way back out and try
1291 * again with an MLE_MASTER.
1292 * dlm_do_local_recovery_cleanup
1293 * has already run, so the mle
1294 * refcount is ok */
1295 mlog(0, "%s:%.*s: no "
1296 "longer blocking. try to "
1297 "master this here\n",
1298 dlm->name,
1299 res->lockname.len,
1300 res->lockname.name);
1301 mle->type = DLM_MLE_MASTER;
1302 mle->mleres = res;
1303 }
1304 }
1305 }
1306
1307 /* now blank out everything, as if we had never
1308 * contacted anyone */
1309 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1310 memset(mle->response_map, 0, sizeof(mle->response_map));
1311 /* reset the vote_map to the current node_map */
1312 memcpy(mle->vote_map, mle->node_map,
1313 sizeof(mle->node_map));
1314 /* put myself into the maybe map */
1315 if (mle->type != DLM_MLE_BLOCK)
1316 set_bit(dlm->node_num, mle->maybe_map);
1317 }
1318 ret = -EAGAIN;
1319 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1320 }
1321 return ret;
1322}
1323
1324
1325/*
1326 * DLM_MASTER_REQUEST_MSG
1327 *
1328 * returns: 0 on success,
1329 * -errno on a network error
1330 *
1331 * on error, the caller should assume the target node is "dead"
1332 *
1333 */
1334
1335static int dlm_do_master_request(struct dlm_lock_resource *res,
1336 struct dlm_master_list_entry *mle, int to)
1337{
1338 struct dlm_ctxt *dlm = mle->dlm;
1339 struct dlm_master_request request;
1340 int ret, response=0, resend;
1341
1342 memset(&request, 0, sizeof(request));
1343 request.node_idx = dlm->node_num;
1344
1345 BUG_ON(mle->type == DLM_MLE_MIGRATION);
1346
1347 request.namelen = (u8)mle->mnamelen;
1348 memcpy(request.name, mle->mname, request.namelen);
1349
1350again:
1351 ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1352 sizeof(request), to, &response);
1353 if (ret < 0) {
1354 if (ret == -ESRCH) {
1355 /* should never happen */
1356 mlog(ML_ERROR, "TCP stack not ready!\n");
1357 BUG();
1358 } else if (ret == -EINVAL) {
1359 mlog(ML_ERROR, "bad args passed to o2net!\n");
1360 BUG();
1361 } else if (ret == -ENOMEM) {
1362 mlog(ML_ERROR, "out of memory while trying to send "
1363 "network message! retrying\n");
1364 /* this is totally crude */
1365 msleep(50);
1366 goto again;
1367 } else if (!dlm_is_host_down(ret)) {
1368 /* not a network error. bad. */
1369 mlog_errno(ret);
1370 mlog(ML_ERROR, "unhandled error!");
1371 BUG();
1372 }
1373 /* all other errors should be network errors,
1374 * and likely indicate node death */
1375 mlog(ML_ERROR, "link to %d went down!\n", to);
1376 goto out;
1377 }
1378
1379 ret = 0;
1380 resend = 0;
1381 spin_lock(&mle->spinlock);
1382 switch (response) {
1383 case DLM_MASTER_RESP_YES:
1384 set_bit(to, mle->response_map);
1385 mlog(0, "node %u is the master, response=YES\n", to);
1386 mlog(0, "%s:%.*s: master node %u now knows I have a "
1387 "reference\n", dlm->name, res->lockname.len,
1388 res->lockname.name, to);
1389 mle->master = to;
1390 break;
1391 case DLM_MASTER_RESP_NO:
1392 mlog(0, "node %u not master, response=NO\n", to);
1393 set_bit(to, mle->response_map);
1394 break;
1395 case DLM_MASTER_RESP_MAYBE:
1396 mlog(0, "node %u not master, response=MAYBE\n", to);
1397 set_bit(to, mle->response_map);
1398 set_bit(to, mle->maybe_map);
1399 break;
1400 case DLM_MASTER_RESP_ERROR:
1401 mlog(0, "node %u hit an error, resending\n", to);
1402 resend = 1;
1403 response = 0;
1404 break;
1405 default:
1406 mlog(ML_ERROR, "bad response! %u\n", response);
1407 BUG();
1408 }
1409 spin_unlock(&mle->spinlock);
1410 if (resend) {
1411 /* this is also totally crude */
1412 msleep(50);
1413 goto again;
1414 }
1415
1416out:
1417 return ret;
1418}
1419
1420/*
1421 * locks that can be taken here:
1422 * dlm->spinlock
1423 * res->spinlock
1424 * mle->spinlock
1425 * dlm->master_list
1426 *
1427 * if possible, TRIM THIS DOWN!!!
1428 */
1429int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1430 void **ret_data)
1431{
1432 u8 response = DLM_MASTER_RESP_MAYBE;
1433 struct dlm_ctxt *dlm = data;
1434 struct dlm_lock_resource *res = NULL;
1435 struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1436 struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1437 char *name;
1438 unsigned int namelen, hash;
1439 int found, ret;
1440 int set_maybe;
1441 int dispatch_assert = 0;
1442 int dispatched = 0;
1443
1444 if (!dlm_grab(dlm))
1445 return DLM_MASTER_RESP_NO;
1446
1447 if (!dlm_domain_fully_joined(dlm)) {
1448 response = DLM_MASTER_RESP_NO;
1449 goto send_response;
1450 }
1451
1452 name = request->name;
1453 namelen = request->namelen;
1454 hash = dlm_lockid_hash(name, namelen);
1455
1456 if (namelen > DLM_LOCKID_NAME_MAX) {
1457 response = DLM_IVBUFLEN;
1458 goto send_response;
1459 }
1460
1461way_up_top:
1462 spin_lock(&dlm->spinlock);
1463 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1464 if (res) {
1465 spin_unlock(&dlm->spinlock);
1466
1467 /* take care of the easy cases up front */
1468 spin_lock(&res->spinlock);
1469
1470 /*
1471 * Right after dlm spinlock was released, dlm_thread could have
1472 * purged the lockres. Check if lockres got unhashed. If so
1473 * start over.
1474 */
1475 if (hlist_unhashed(&res->hash_node)) {
1476 spin_unlock(&res->spinlock);
1477 dlm_lockres_put(res);
1478 goto way_up_top;
1479 }
1480
1481 if (res->state & (DLM_LOCK_RES_RECOVERING|
1482 DLM_LOCK_RES_MIGRATING)) {
1483 spin_unlock(&res->spinlock);
1484 mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1485 "being recovered/migrated\n");
1486 response = DLM_MASTER_RESP_ERROR;
1487 if (mle)
1488 kmem_cache_free(dlm_mle_cache, mle);
1489 goto send_response;
1490 }
1491
1492 if (res->owner == dlm->node_num) {
1493 dlm_lockres_set_refmap_bit(dlm, res, request->node_idx);
1494 spin_unlock(&res->spinlock);
1495 response = DLM_MASTER_RESP_YES;
1496 if (mle)
1497 kmem_cache_free(dlm_mle_cache, mle);
1498
1499 /* this node is the owner.
1500 * there is some extra work that needs to
1501 * happen now. the requesting node has
1502 * caused all nodes up to this one to
1503 * create mles. this node now needs to
1504 * go back and clean those up. */
1505 dispatch_assert = 1;
1506 goto send_response;
1507 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1508 spin_unlock(&res->spinlock);
1509 // mlog(0, "node %u is the master\n", res->owner);
1510 response = DLM_MASTER_RESP_NO;
1511 if (mle)
1512 kmem_cache_free(dlm_mle_cache, mle);
1513 goto send_response;
1514 }
1515
1516 /* ok, there is no owner. either this node is
1517 * being blocked, or it is actively trying to
1518 * master this lock. */
1519 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1520 mlog(ML_ERROR, "lock with no owner should be "
1521 "in-progress!\n");
1522 BUG();
1523 }
1524
1525 // mlog(0, "lockres is in progress...\n");
1526 spin_lock(&dlm->master_lock);
1527 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1528 if (!found) {
1529 mlog(ML_ERROR, "no mle found for this lock!\n");
1530 BUG();
1531 }
1532 set_maybe = 1;
1533 spin_lock(&tmpmle->spinlock);
1534 if (tmpmle->type == DLM_MLE_BLOCK) {
1535 // mlog(0, "this node is waiting for "
1536 // "lockres to be mastered\n");
1537 response = DLM_MASTER_RESP_NO;
1538 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1539 mlog(0, "node %u is master, but trying to migrate to "
1540 "node %u.\n", tmpmle->master, tmpmle->new_master);
1541 if (tmpmle->master == dlm->node_num) {
1542 mlog(ML_ERROR, "no owner on lockres, but this "
1543 "node is trying to migrate it to %u?!\n",
1544 tmpmle->new_master);
1545 BUG();
1546 } else {
1547 /* the real master can respond on its own */
1548 response = DLM_MASTER_RESP_NO;
1549 }
1550 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1551 set_maybe = 0;
1552 if (tmpmle->master == dlm->node_num) {
1553 response = DLM_MASTER_RESP_YES;
1554 /* this node will be the owner.
1555 * go back and clean the mles on any
1556 * other nodes */
1557 dispatch_assert = 1;
1558 dlm_lockres_set_refmap_bit(dlm, res,
1559 request->node_idx);
1560 } else
1561 response = DLM_MASTER_RESP_NO;
1562 } else {
1563 // mlog(0, "this node is attempting to "
1564 // "master lockres\n");
1565 response = DLM_MASTER_RESP_MAYBE;
1566 }
1567 if (set_maybe)
1568 set_bit(request->node_idx, tmpmle->maybe_map);
1569 spin_unlock(&tmpmle->spinlock);
1570
1571 spin_unlock(&dlm->master_lock);
1572 spin_unlock(&res->spinlock);
1573
1574 /* keep the mle attached to heartbeat events */
1575 dlm_put_mle(tmpmle);
1576 if (mle)
1577 kmem_cache_free(dlm_mle_cache, mle);
1578 goto send_response;
1579 }
1580
1581 /*
1582 * lockres doesn't exist on this node
1583 * if there is an MLE_BLOCK, return NO
1584 * if there is an MLE_MASTER, return MAYBE
1585 * otherwise, add an MLE_BLOCK, return NO
1586 */
1587 spin_lock(&dlm->master_lock);
1588 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1589 if (!found) {
1590 /* this lockid has never been seen on this node yet */
1591 // mlog(0, "no mle found\n");
1592 if (!mle) {
1593 spin_unlock(&dlm->master_lock);
1594 spin_unlock(&dlm->spinlock);
1595
1596 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1597 if (!mle) {
1598 response = DLM_MASTER_RESP_ERROR;
1599 mlog_errno(-ENOMEM);
1600 goto send_response;
1601 }
1602 goto way_up_top;
1603 }
1604
1605 // mlog(0, "this is second time thru, already allocated, "
1606 // "add the block.\n");
1607 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1608 set_bit(request->node_idx, mle->maybe_map);
1609 __dlm_insert_mle(dlm, mle);
1610 response = DLM_MASTER_RESP_NO;
1611 } else {
1612 // mlog(0, "mle was found\n");
1613 set_maybe = 1;
1614 spin_lock(&tmpmle->spinlock);
1615 if (tmpmle->master == dlm->node_num) {
1616 mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1617 BUG();
1618 }
1619 if (tmpmle->type == DLM_MLE_BLOCK)
1620 response = DLM_MASTER_RESP_NO;
1621 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1622 mlog(0, "migration mle was found (%u->%u)\n",
1623 tmpmle->master, tmpmle->new_master);
1624 /* real master can respond on its own */
1625 response = DLM_MASTER_RESP_NO;
1626 } else
1627 response = DLM_MASTER_RESP_MAYBE;
1628 if (set_maybe)
1629 set_bit(request->node_idx, tmpmle->maybe_map);
1630 spin_unlock(&tmpmle->spinlock);
1631 }
1632 spin_unlock(&dlm->master_lock);
1633 spin_unlock(&dlm->spinlock);
1634
1635 if (found) {
1636 /* keep the mle attached to heartbeat events */
1637 dlm_put_mle(tmpmle);
1638 }
1639send_response:
1640 /*
1641 * __dlm_lookup_lockres() grabbed a reference to this lockres.
1642 * The reference is released by dlm_assert_master_worker() under
1643 * the call to dlm_dispatch_assert_master(). If
1644 * dlm_assert_master_worker() isn't called, we drop it here.
1645 */
1646 if (dispatch_assert) {
1647 if (response != DLM_MASTER_RESP_YES)
1648 mlog(ML_ERROR, "invalid response %d\n", response);
1649 if (!res) {
1650 mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1651 BUG();
1652 }
1653 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1654 dlm->node_num, res->lockname.len, res->lockname.name);
1655 spin_lock(&res->spinlock);
1656 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1657 DLM_ASSERT_MASTER_MLE_CLEANUP);
1658 if (ret < 0) {
1659 mlog(ML_ERROR, "failed to dispatch assert master work\n");
1660 response = DLM_MASTER_RESP_ERROR;
1661 spin_unlock(&res->spinlock);
1662 dlm_lockres_put(res);
1663 } else {
1664 dispatched = 1;
1665 __dlm_lockres_grab_inflight_worker(dlm, res);
1666 spin_unlock(&res->spinlock);
1667 }
1668 } else {
1669 if (res)
1670 dlm_lockres_put(res);
1671 }
1672
1673 if (!dispatched)
1674 dlm_put(dlm);
1675 return response;
1676}
1677
1678/*
1679 * DLM_ASSERT_MASTER_MSG
1680 */
1681
1682
1683/*
1684 * NOTE: this can be used for debugging
1685 * can periodically run all locks owned by this node
1686 * and re-assert across the cluster...
1687 */
1688static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1689 struct dlm_lock_resource *res,
1690 void *nodemap, u32 flags)
1691{
1692 struct dlm_assert_master assert;
1693 int to, tmpret;
1694 struct dlm_node_iter iter;
1695 int ret = 0;
1696 int reassert;
1697 const char *lockname = res->lockname.name;
1698 unsigned int namelen = res->lockname.len;
1699
1700 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1701
1702 spin_lock(&res->spinlock);
1703 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1704 spin_unlock(&res->spinlock);
1705
1706again:
1707 reassert = 0;
1708
1709 /* note that if this nodemap is empty, it returns 0 */
1710 dlm_node_iter_init(nodemap, &iter);
1711 while ((to = dlm_node_iter_next(&iter)) >= 0) {
1712 int r = 0;
1713 struct dlm_master_list_entry *mle = NULL;
1714
1715 mlog(0, "sending assert master to %d (%.*s)\n", to,
1716 namelen, lockname);
1717 memset(&assert, 0, sizeof(assert));
1718 assert.node_idx = dlm->node_num;
1719 assert.namelen = namelen;
1720 memcpy(assert.name, lockname, namelen);
1721 assert.flags = cpu_to_be32(flags);
1722
1723 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1724 &assert, sizeof(assert), to, &r);
1725 if (tmpret < 0) {
1726 mlog(ML_ERROR, "Error %d when sending message %u (key "
1727 "0x%x) to node %u\n", tmpret,
1728 DLM_ASSERT_MASTER_MSG, dlm->key, to);
1729 if (!dlm_is_host_down(tmpret)) {
1730 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1731 BUG();
1732 }
1733 /* a node died. finish out the rest of the nodes. */
1734 mlog(0, "link to %d went down!\n", to);
1735 /* any nonzero status return will do */
1736 ret = tmpret;
1737 r = 0;
1738 } else if (r < 0) {
1739 /* ok, something horribly messed. kill thyself. */
1740 mlog(ML_ERROR,"during assert master of %.*s to %u, "
1741 "got %d.\n", namelen, lockname, to, r);
1742 spin_lock(&dlm->spinlock);
1743 spin_lock(&dlm->master_lock);
1744 if (dlm_find_mle(dlm, &mle, (char *)lockname,
1745 namelen)) {
1746 dlm_print_one_mle(mle);
1747 __dlm_put_mle(mle);
1748 }
1749 spin_unlock(&dlm->master_lock);
1750 spin_unlock(&dlm->spinlock);
1751 BUG();
1752 }
1753
1754 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1755 !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1756 mlog(ML_ERROR, "%.*s: very strange, "
1757 "master MLE but no lockres on %u\n",
1758 namelen, lockname, to);
1759 }
1760
1761 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1762 mlog(0, "%.*s: node %u create mles on other "
1763 "nodes and requests a re-assert\n",
1764 namelen, lockname, to);
1765 reassert = 1;
1766 }
1767 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1768 mlog(0, "%.*s: node %u has a reference to this "
1769 "lockres, set the bit in the refmap\n",
1770 namelen, lockname, to);
1771 spin_lock(&res->spinlock);
1772 dlm_lockres_set_refmap_bit(dlm, res, to);
1773 spin_unlock(&res->spinlock);
1774 }
1775 }
1776
1777 if (reassert)
1778 goto again;
1779
1780 spin_lock(&res->spinlock);
1781 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1782 spin_unlock(&res->spinlock);
1783 wake_up(&res->wq);
1784
1785 return ret;
1786}
1787
1788/*
1789 * locks that can be taken here:
1790 * dlm->spinlock
1791 * res->spinlock
1792 * mle->spinlock
1793 * dlm->master_list
1794 *
1795 * if possible, TRIM THIS DOWN!!!
1796 */
1797int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1798 void **ret_data)
1799{
1800 struct dlm_ctxt *dlm = data;
1801 struct dlm_master_list_entry *mle = NULL;
1802 struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1803 struct dlm_lock_resource *res = NULL;
1804 char *name;
1805 unsigned int namelen, hash;
1806 u32 flags;
1807 int master_request = 0, have_lockres_ref = 0;
1808 int ret = 0;
1809
1810 if (!dlm_grab(dlm))
1811 return 0;
1812
1813 name = assert->name;
1814 namelen = assert->namelen;
1815 hash = dlm_lockid_hash(name, namelen);
1816 flags = be32_to_cpu(assert->flags);
1817
1818 if (namelen > DLM_LOCKID_NAME_MAX) {
1819 mlog(ML_ERROR, "Invalid name length!");
1820 goto done;
1821 }
1822
1823 spin_lock(&dlm->spinlock);
1824
1825 if (flags)
1826 mlog(0, "assert_master with flags: %u\n", flags);
1827
1828 /* find the MLE */
1829 spin_lock(&dlm->master_lock);
1830 if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1831 /* not an error, could be master just re-asserting */
1832 mlog(0, "just got an assert_master from %u, but no "
1833 "MLE for it! (%.*s)\n", assert->node_idx,
1834 namelen, name);
1835 } else {
1836 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1837 if (bit >= O2NM_MAX_NODES) {
1838 /* not necessarily an error, though less likely.
1839 * could be master just re-asserting. */
1840 mlog(0, "no bits set in the maybe_map, but %u "
1841 "is asserting! (%.*s)\n", assert->node_idx,
1842 namelen, name);
1843 } else if (bit != assert->node_idx) {
1844 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1845 mlog(0, "master %u was found, %u should "
1846 "back off\n", assert->node_idx, bit);
1847 } else {
1848 /* with the fix for bug 569, a higher node
1849 * number winning the mastery will respond
1850 * YES to mastery requests, but this node
1851 * had no way of knowing. let it pass. */
1852 mlog(0, "%u is the lowest node, "
1853 "%u is asserting. (%.*s) %u must "
1854 "have begun after %u won.\n", bit,
1855 assert->node_idx, namelen, name, bit,
1856 assert->node_idx);
1857 }
1858 }
1859 if (mle->type == DLM_MLE_MIGRATION) {
1860 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1861 mlog(0, "%s:%.*s: got cleanup assert"
1862 " from %u for migration\n",
1863 dlm->name, namelen, name,
1864 assert->node_idx);
1865 } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1866 mlog(0, "%s:%.*s: got unrelated assert"
1867 " from %u for migration, ignoring\n",
1868 dlm->name, namelen, name,
1869 assert->node_idx);
1870 __dlm_put_mle(mle);
1871 spin_unlock(&dlm->master_lock);
1872 spin_unlock(&dlm->spinlock);
1873 goto done;
1874 }
1875 }
1876 }
1877 spin_unlock(&dlm->master_lock);
1878
1879 /* ok everything checks out with the MLE
1880 * now check to see if there is a lockres */
1881 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1882 if (res) {
1883 spin_lock(&res->spinlock);
1884 if (res->state & DLM_LOCK_RES_RECOVERING) {
1885 mlog(ML_ERROR, "%u asserting but %.*s is "
1886 "RECOVERING!\n", assert->node_idx, namelen, name);
1887 goto kill;
1888 }
1889 if (!mle) {
1890 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1891 res->owner != assert->node_idx) {
1892 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1893 "but current owner is %u! (%.*s)\n",
1894 assert->node_idx, res->owner, namelen,
1895 name);
1896 __dlm_print_one_lock_resource(res);
1897 BUG();
1898 }
1899 } else if (mle->type != DLM_MLE_MIGRATION) {
1900 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1901 /* owner is just re-asserting */
1902 if (res->owner == assert->node_idx) {
1903 mlog(0, "owner %u re-asserting on "
1904 "lock %.*s\n", assert->node_idx,
1905 namelen, name);
1906 goto ok;
1907 }
1908 mlog(ML_ERROR, "got assert_master from "
1909 "node %u, but %u is the owner! "
1910 "(%.*s)\n", assert->node_idx,
1911 res->owner, namelen, name);
1912 goto kill;
1913 }
1914 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1915 mlog(ML_ERROR, "got assert from %u, but lock "
1916 "with no owner should be "
1917 "in-progress! (%.*s)\n",
1918 assert->node_idx,
1919 namelen, name);
1920 goto kill;
1921 }
1922 } else /* mle->type == DLM_MLE_MIGRATION */ {
1923 /* should only be getting an assert from new master */
1924 if (assert->node_idx != mle->new_master) {
1925 mlog(ML_ERROR, "got assert from %u, but "
1926 "new master is %u, and old master "
1927 "was %u (%.*s)\n",
1928 assert->node_idx, mle->new_master,
1929 mle->master, namelen, name);
1930 goto kill;
1931 }
1932
1933 }
1934ok:
1935 spin_unlock(&res->spinlock);
1936 }
1937
1938 // mlog(0, "woo! got an assert_master from node %u!\n",
1939 // assert->node_idx);
1940 if (mle) {
1941 int extra_ref = 0;
1942 int nn = -1;
1943 int rr, err = 0;
1944
1945 spin_lock(&mle->spinlock);
1946 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1947 extra_ref = 1;
1948 else {
1949 /* MASTER mle: if any bits set in the response map
1950 * then the calling node needs to re-assert to clear
1951 * up nodes that this node contacted */
1952 while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1953 nn+1)) < O2NM_MAX_NODES) {
1954 if (nn != dlm->node_num && nn != assert->node_idx) {
1955 master_request = 1;
1956 break;
1957 }
1958 }
1959 }
1960 mle->master = assert->node_idx;
1961 atomic_set(&mle->woken, 1);
1962 wake_up(&mle->wq);
1963 spin_unlock(&mle->spinlock);
1964
1965 if (res) {
1966 int wake = 0;
1967 spin_lock(&res->spinlock);
1968 if (mle->type == DLM_MLE_MIGRATION) {
1969 mlog(0, "finishing off migration of lockres %.*s, "
1970 "from %u to %u\n",
1971 res->lockname.len, res->lockname.name,
1972 dlm->node_num, mle->new_master);
1973 res->state &= ~DLM_LOCK_RES_MIGRATING;
1974 wake = 1;
1975 dlm_change_lockres_owner(dlm, res, mle->new_master);
1976 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1977 } else {
1978 dlm_change_lockres_owner(dlm, res, mle->master);
1979 }
1980 spin_unlock(&res->spinlock);
1981 have_lockres_ref = 1;
1982 if (wake)
1983 wake_up(&res->wq);
1984 }
1985
1986 /* master is known, detach if not already detached.
1987 * ensures that only one assert_master call will happen
1988 * on this mle. */
1989 spin_lock(&dlm->master_lock);
1990
1991 rr = atomic_read(&mle->mle_refs.refcount);
1992 if (mle->inuse > 0) {
1993 if (extra_ref && rr < 3)
1994 err = 1;
1995 else if (!extra_ref && rr < 2)
1996 err = 1;
1997 } else {
1998 if (extra_ref && rr < 2)
1999 err = 1;
2000 else if (!extra_ref && rr < 1)
2001 err = 1;
2002 }
2003 if (err) {
2004 mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
2005 "that will mess up this node, refs=%d, extra=%d, "
2006 "inuse=%d\n", dlm->name, namelen, name,
2007 assert->node_idx, rr, extra_ref, mle->inuse);
2008 dlm_print_one_mle(mle);
2009 }
2010 __dlm_unlink_mle(dlm, mle);
2011 __dlm_mle_detach_hb_events(dlm, mle);
2012 __dlm_put_mle(mle);
2013 if (extra_ref) {
2014 /* the assert master message now balances the extra
2015 * ref given by the master / migration request message.
2016 * if this is the last put, it will be removed
2017 * from the list. */
2018 __dlm_put_mle(mle);
2019 }
2020 spin_unlock(&dlm->master_lock);
2021 } else if (res) {
2022 if (res->owner != assert->node_idx) {
2023 mlog(0, "assert_master from %u, but current "
2024 "owner is %u (%.*s), no mle\n", assert->node_idx,
2025 res->owner, namelen, name);
2026 }
2027 }
2028 spin_unlock(&dlm->spinlock);
2029
2030done:
2031 ret = 0;
2032 if (res) {
2033 spin_lock(&res->spinlock);
2034 res->state |= DLM_LOCK_RES_SETREF_INPROG;
2035 spin_unlock(&res->spinlock);
2036 *ret_data = (void *)res;
2037 }
2038 dlm_put(dlm);
2039 if (master_request) {
2040 mlog(0, "need to tell master to reassert\n");
2041 /* positive. negative would shoot down the node. */
2042 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2043 if (!have_lockres_ref) {
2044 mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2045 "mle present here for %s:%.*s, but no lockres!\n",
2046 assert->node_idx, dlm->name, namelen, name);
2047 }
2048 }
2049 if (have_lockres_ref) {
2050 /* let the master know we have a reference to the lockres */
2051 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2052 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2053 dlm->name, namelen, name, assert->node_idx);
2054 }
2055 return ret;
2056
2057kill:
2058 /* kill the caller! */
2059 mlog(ML_ERROR, "Bad message received from another node. Dumping state "
2060 "and killing the other node now! This node is OK and can continue.\n");
2061 __dlm_print_one_lock_resource(res);
2062 spin_unlock(&res->spinlock);
2063 spin_lock(&dlm->master_lock);
2064 if (mle)
2065 __dlm_put_mle(mle);
2066 spin_unlock(&dlm->master_lock);
2067 spin_unlock(&dlm->spinlock);
2068 *ret_data = (void *)res;
2069 dlm_put(dlm);
2070 return -EINVAL;
2071}
2072
2073void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2074{
2075 struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2076
2077 if (ret_data) {
2078 spin_lock(&res->spinlock);
2079 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2080 spin_unlock(&res->spinlock);
2081 wake_up(&res->wq);
2082 dlm_lockres_put(res);
2083 }
2084 return;
2085}
2086
2087int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2088 struct dlm_lock_resource *res,
2089 int ignore_higher, u8 request_from, u32 flags)
2090{
2091 struct dlm_work_item *item;
2092 item = kzalloc(sizeof(*item), GFP_ATOMIC);
2093 if (!item)
2094 return -ENOMEM;
2095
2096
2097 /* queue up work for dlm_assert_master_worker */
2098 dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2099 item->u.am.lockres = res; /* already have a ref */
2100 /* can optionally ignore node numbers higher than this node */
2101 item->u.am.ignore_higher = ignore_higher;
2102 item->u.am.request_from = request_from;
2103 item->u.am.flags = flags;
2104
2105 if (ignore_higher)
2106 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2107 res->lockname.name);
2108
2109 spin_lock(&dlm->work_lock);
2110 list_add_tail(&item->list, &dlm->work_list);
2111 spin_unlock(&dlm->work_lock);
2112
2113 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2114 return 0;
2115}
2116
2117static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2118{
2119 struct dlm_ctxt *dlm = data;
2120 int ret = 0;
2121 struct dlm_lock_resource *res;
2122 unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2123 int ignore_higher;
2124 int bit;
2125 u8 request_from;
2126 u32 flags;
2127
2128 dlm = item->dlm;
2129 res = item->u.am.lockres;
2130 ignore_higher = item->u.am.ignore_higher;
2131 request_from = item->u.am.request_from;
2132 flags = item->u.am.flags;
2133
2134 spin_lock(&dlm->spinlock);
2135 memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2136 spin_unlock(&dlm->spinlock);
2137
2138 clear_bit(dlm->node_num, nodemap);
2139 if (ignore_higher) {
2140 /* if is this just to clear up mles for nodes below
2141 * this node, do not send the message to the original
2142 * caller or any node number higher than this */
2143 clear_bit(request_from, nodemap);
2144 bit = dlm->node_num;
2145 while (1) {
2146 bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2147 bit+1);
2148 if (bit >= O2NM_MAX_NODES)
2149 break;
2150 clear_bit(bit, nodemap);
2151 }
2152 }
2153
2154 /*
2155 * If we're migrating this lock to someone else, we are no
2156 * longer allowed to assert out own mastery. OTOH, we need to
2157 * prevent migration from starting while we're still asserting
2158 * our dominance. The reserved ast delays migration.
2159 */
2160 spin_lock(&res->spinlock);
2161 if (res->state & DLM_LOCK_RES_MIGRATING) {
2162 mlog(0, "Someone asked us to assert mastery, but we're "
2163 "in the middle of migration. Skipping assert, "
2164 "the new master will handle that.\n");
2165 spin_unlock(&res->spinlock);
2166 goto put;
2167 } else
2168 __dlm_lockres_reserve_ast(res);
2169 spin_unlock(&res->spinlock);
2170
2171 /* this call now finishes out the nodemap
2172 * even if one or more nodes die */
2173 mlog(0, "worker about to master %.*s here, this=%u\n",
2174 res->lockname.len, res->lockname.name, dlm->node_num);
2175 ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2176 if (ret < 0) {
2177 /* no need to restart, we are done */
2178 if (!dlm_is_host_down(ret))
2179 mlog_errno(ret);
2180 }
2181
2182 /* Ok, we've asserted ourselves. Let's let migration start. */
2183 dlm_lockres_release_ast(dlm, res);
2184
2185put:
2186 dlm_lockres_drop_inflight_worker(dlm, res);
2187
2188 dlm_lockres_put(res);
2189
2190 mlog(0, "finished with dlm_assert_master_worker\n");
2191}
2192
2193/* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2194 * We cannot wait for node recovery to complete to begin mastering this
2195 * lockres because this lockres is used to kick off recovery! ;-)
2196 * So, do a pre-check on all living nodes to see if any of those nodes
2197 * think that $RECOVERY is currently mastered by a dead node. If so,
2198 * we wait a short time to allow that node to get notified by its own
2199 * heartbeat stack, then check again. All $RECOVERY lock resources
2200 * mastered by dead nodes are purged when the hearbeat callback is
2201 * fired, so we can know for sure that it is safe to continue once
2202 * the node returns a live node or no node. */
2203static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2204 struct dlm_lock_resource *res)
2205{
2206 struct dlm_node_iter iter;
2207 int nodenum;
2208 int ret = 0;
2209 u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2210
2211 spin_lock(&dlm->spinlock);
2212 dlm_node_iter_init(dlm->domain_map, &iter);
2213 spin_unlock(&dlm->spinlock);
2214
2215 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2216 /* do not send to self */
2217 if (nodenum == dlm->node_num)
2218 continue;
2219 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2220 if (ret < 0) {
2221 mlog_errno(ret);
2222 if (!dlm_is_host_down(ret))
2223 BUG();
2224 /* host is down, so answer for that node would be
2225 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
2226 ret = 0;
2227 }
2228
2229 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2230 /* check to see if this master is in the recovery map */
2231 spin_lock(&dlm->spinlock);
2232 if (test_bit(master, dlm->recovery_map)) {
2233 mlog(ML_NOTICE, "%s: node %u has not seen "
2234 "node %u go down yet, and thinks the "
2235 "dead node is mastering the recovery "
2236 "lock. must wait.\n", dlm->name,
2237 nodenum, master);
2238 ret = -EAGAIN;
2239 }
2240 spin_unlock(&dlm->spinlock);
2241 mlog(0, "%s: reco lock master is %u\n", dlm->name,
2242 master);
2243 break;
2244 }
2245 }
2246 return ret;
2247}
2248
2249/*
2250 * DLM_DEREF_LOCKRES_MSG
2251 */
2252
2253int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2254{
2255 struct dlm_deref_lockres deref;
2256 int ret = 0, r;
2257 const char *lockname;
2258 unsigned int namelen;
2259
2260 lockname = res->lockname.name;
2261 namelen = res->lockname.len;
2262 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2263
2264 memset(&deref, 0, sizeof(deref));
2265 deref.node_idx = dlm->node_num;
2266 deref.namelen = namelen;
2267 memcpy(deref.name, lockname, namelen);
2268
2269 ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2270 &deref, sizeof(deref), res->owner, &r);
2271 if (ret < 0)
2272 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n",
2273 dlm->name, namelen, lockname, ret, res->owner);
2274 else if (r < 0) {
2275 /* BAD. other node says I did not have a ref. */
2276 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2277 dlm->name, namelen, lockname, res->owner, r);
2278 dlm_print_one_lock_resource(res);
2279 BUG();
2280 }
2281 return ret ? ret : r;
2282}
2283
2284int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2285 void **ret_data)
2286{
2287 struct dlm_ctxt *dlm = data;
2288 struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2289 struct dlm_lock_resource *res = NULL;
2290 char *name;
2291 unsigned int namelen;
2292 int ret = -EINVAL;
2293 u8 node;
2294 unsigned int hash;
2295 struct dlm_work_item *item;
2296 int cleared = 0;
2297 int dispatch = 0;
2298
2299 if (!dlm_grab(dlm))
2300 return 0;
2301
2302 name = deref->name;
2303 namelen = deref->namelen;
2304 node = deref->node_idx;
2305
2306 if (namelen > DLM_LOCKID_NAME_MAX) {
2307 mlog(ML_ERROR, "Invalid name length!");
2308 goto done;
2309 }
2310 if (deref->node_idx >= O2NM_MAX_NODES) {
2311 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2312 goto done;
2313 }
2314
2315 hash = dlm_lockid_hash(name, namelen);
2316
2317 spin_lock(&dlm->spinlock);
2318 res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2319 if (!res) {
2320 spin_unlock(&dlm->spinlock);
2321 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2322 dlm->name, namelen, name);
2323 goto done;
2324 }
2325 spin_unlock(&dlm->spinlock);
2326
2327 spin_lock(&res->spinlock);
2328 if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2329 dispatch = 1;
2330 else {
2331 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2332 if (test_bit(node, res->refmap)) {
2333 dlm_lockres_clear_refmap_bit(dlm, res, node);
2334 cleared = 1;
2335 }
2336 }
2337 spin_unlock(&res->spinlock);
2338
2339 if (!dispatch) {
2340 if (cleared)
2341 dlm_lockres_calc_usage(dlm, res);
2342 else {
2343 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2344 "but it is already dropped!\n", dlm->name,
2345 res->lockname.len, res->lockname.name, node);
2346 dlm_print_one_lock_resource(res);
2347 }
2348 ret = DLM_DEREF_RESPONSE_DONE;
2349 goto done;
2350 }
2351
2352 item = kzalloc(sizeof(*item), GFP_NOFS);
2353 if (!item) {
2354 ret = -ENOMEM;
2355 mlog_errno(ret);
2356 goto done;
2357 }
2358
2359 dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2360 item->u.dl.deref_res = res;
2361 item->u.dl.deref_node = node;
2362
2363 spin_lock(&dlm->work_lock);
2364 list_add_tail(&item->list, &dlm->work_list);
2365 spin_unlock(&dlm->work_lock);
2366
2367 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2368 return DLM_DEREF_RESPONSE_INPROG;
2369
2370done:
2371 if (res)
2372 dlm_lockres_put(res);
2373 dlm_put(dlm);
2374
2375 return ret;
2376}
2377
2378int dlm_deref_lockres_done_handler(struct o2net_msg *msg, u32 len, void *data,
2379 void **ret_data)
2380{
2381 struct dlm_ctxt *dlm = data;
2382 struct dlm_deref_lockres_done *deref
2383 = (struct dlm_deref_lockres_done *)msg->buf;
2384 struct dlm_lock_resource *res = NULL;
2385 char *name;
2386 unsigned int namelen;
2387 int ret = -EINVAL;
2388 u8 node;
2389 unsigned int hash;
2390
2391 if (!dlm_grab(dlm))
2392 return 0;
2393
2394 name = deref->name;
2395 namelen = deref->namelen;
2396 node = deref->node_idx;
2397
2398 if (namelen > DLM_LOCKID_NAME_MAX) {
2399 mlog(ML_ERROR, "Invalid name length!");
2400 goto done;
2401 }
2402 if (deref->node_idx >= O2NM_MAX_NODES) {
2403 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2404 goto done;
2405 }
2406
2407 hash = dlm_lockid_hash(name, namelen);
2408
2409 spin_lock(&dlm->spinlock);
2410 res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2411 if (!res) {
2412 spin_unlock(&dlm->spinlock);
2413 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2414 dlm->name, namelen, name);
2415 goto done;
2416 }
2417
2418 spin_lock(&res->spinlock);
2419 BUG_ON(!(res->state & DLM_LOCK_RES_DROPPING_REF));
2420 if (!list_empty(&res->purge)) {
2421 mlog(0, "%s: Removing res %.*s from purgelist\n",
2422 dlm->name, res->lockname.len, res->lockname.name);
2423 list_del_init(&res->purge);
2424 dlm_lockres_put(res);
2425 dlm->purge_count--;
2426 }
2427
2428 if (!__dlm_lockres_unused(res)) {
2429 mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
2430 dlm->name, res->lockname.len, res->lockname.name);
2431 __dlm_print_one_lock_resource(res);
2432 BUG();
2433 }
2434
2435 __dlm_unhash_lockres(dlm, res);
2436
2437 spin_lock(&dlm->track_lock);
2438 if (!list_empty(&res->tracking))
2439 list_del_init(&res->tracking);
2440 else {
2441 mlog(ML_ERROR, "%s: Resource %.*s not on the Tracking list\n",
2442 dlm->name, res->lockname.len, res->lockname.name);
2443 __dlm_print_one_lock_resource(res);
2444 }
2445 spin_unlock(&dlm->track_lock);
2446
2447 /* lockres is not in the hash now. drop the flag and wake up
2448 * any processes waiting in dlm_get_lock_resource.
2449 */
2450 res->state &= ~DLM_LOCK_RES_DROPPING_REF;
2451 spin_unlock(&res->spinlock);
2452 wake_up(&res->wq);
2453
2454 dlm_lockres_put(res);
2455
2456 spin_unlock(&dlm->spinlock);
2457
2458 ret = 0;
2459
2460done:
2461 dlm_put(dlm);
2462 return ret;
2463}
2464
2465static void dlm_drop_lockres_ref_done(struct dlm_ctxt *dlm,
2466 struct dlm_lock_resource *res, u8 node)
2467{
2468 struct dlm_deref_lockres_done deref;
2469 int ret = 0, r;
2470 const char *lockname;
2471 unsigned int namelen;
2472
2473 lockname = res->lockname.name;
2474 namelen = res->lockname.len;
2475 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2476
2477 memset(&deref, 0, sizeof(deref));
2478 deref.node_idx = dlm->node_num;
2479 deref.namelen = namelen;
2480 memcpy(deref.name, lockname, namelen);
2481
2482 ret = o2net_send_message(DLM_DEREF_LOCKRES_DONE, dlm->key,
2483 &deref, sizeof(deref), node, &r);
2484 if (ret < 0) {
2485 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF DONE "
2486 " to node %u\n", dlm->name, namelen,
2487 lockname, ret, node);
2488 } else if (r < 0) {
2489 /* ignore the error */
2490 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2491 dlm->name, namelen, lockname, node, r);
2492 dlm_print_one_lock_resource(res);
2493 }
2494}
2495
2496static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2497{
2498 struct dlm_ctxt *dlm;
2499 struct dlm_lock_resource *res;
2500 u8 node;
2501 u8 cleared = 0;
2502
2503 dlm = item->dlm;
2504 res = item->u.dl.deref_res;
2505 node = item->u.dl.deref_node;
2506
2507 spin_lock(&res->spinlock);
2508 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2509 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2510 if (test_bit(node, res->refmap)) {
2511 dlm_lockres_clear_refmap_bit(dlm, res, node);
2512 cleared = 1;
2513 }
2514 spin_unlock(&res->spinlock);
2515
2516 dlm_drop_lockres_ref_done(dlm, res, node);
2517
2518 if (cleared) {
2519 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2520 dlm->name, res->lockname.len, res->lockname.name, node);
2521 dlm_lockres_calc_usage(dlm, res);
2522 } else {
2523 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2524 "but it is already dropped!\n", dlm->name,
2525 res->lockname.len, res->lockname.name, node);
2526 dlm_print_one_lock_resource(res);
2527 }
2528
2529 dlm_lockres_put(res);
2530}
2531
2532/*
2533 * A migrateable resource is one that is :
2534 * 1. locally mastered, and,
2535 * 2. zero local locks, and,
2536 * 3. one or more non-local locks, or, one or more references
2537 * Returns 1 if yes, 0 if not.
2538 */
2539static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
2540 struct dlm_lock_resource *res)
2541{
2542 enum dlm_lockres_list idx;
2543 int nonlocal = 0, node_ref;
2544 struct list_head *queue;
2545 struct dlm_lock *lock;
2546 u64 cookie;
2547
2548 assert_spin_locked(&res->spinlock);
2549
2550 /* delay migration when the lockres is in MIGRATING state */
2551 if (res->state & DLM_LOCK_RES_MIGRATING)
2552 return 0;
2553
2554 /* delay migration when the lockres is in RECOCERING state */
2555 if (res->state & (DLM_LOCK_RES_RECOVERING|
2556 DLM_LOCK_RES_RECOVERY_WAITING))
2557 return 0;
2558
2559 if (res->owner != dlm->node_num)
2560 return 0;
2561
2562 for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2563 queue = dlm_list_idx_to_ptr(res, idx);
2564 list_for_each_entry(lock, queue, list) {
2565 if (lock->ml.node != dlm->node_num) {
2566 nonlocal++;
2567 continue;
2568 }
2569 cookie = be64_to_cpu(lock->ml.cookie);
2570 mlog(0, "%s: Not migrateable res %.*s, lock %u:%llu on "
2571 "%s list\n", dlm->name, res->lockname.len,
2572 res->lockname.name,
2573 dlm_get_lock_cookie_node(cookie),
2574 dlm_get_lock_cookie_seq(cookie),
2575 dlm_list_in_text(idx));
2576 return 0;
2577 }
2578 }
2579
2580 if (!nonlocal) {
2581 node_ref = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
2582 if (node_ref >= O2NM_MAX_NODES)
2583 return 0;
2584 }
2585
2586 mlog(0, "%s: res %.*s, Migrateable\n", dlm->name, res->lockname.len,
2587 res->lockname.name);
2588
2589 return 1;
2590}
2591
2592/*
2593 * DLM_MIGRATE_LOCKRES
2594 */
2595
2596
2597static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2598 struct dlm_lock_resource *res, u8 target)
2599{
2600 struct dlm_master_list_entry *mle = NULL;
2601 struct dlm_master_list_entry *oldmle = NULL;
2602 struct dlm_migratable_lockres *mres = NULL;
2603 int ret = 0;
2604 const char *name;
2605 unsigned int namelen;
2606 int mle_added = 0;
2607 int wake = 0;
2608
2609 if (!dlm_grab(dlm))
2610 return -EINVAL;
2611
2612 BUG_ON(target == O2NM_MAX_NODES);
2613
2614 name = res->lockname.name;
2615 namelen = res->lockname.len;
2616
2617 mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name,
2618 target);
2619
2620 /* preallocate up front. if this fails, abort */
2621 ret = -ENOMEM;
2622 mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2623 if (!mres) {
2624 mlog_errno(ret);
2625 goto leave;
2626 }
2627
2628 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2629 if (!mle) {
2630 mlog_errno(ret);
2631 goto leave;
2632 }
2633 ret = 0;
2634
2635 /*
2636 * clear any existing master requests and
2637 * add the migration mle to the list
2638 */
2639 spin_lock(&dlm->spinlock);
2640 spin_lock(&dlm->master_lock);
2641 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2642 namelen, target, dlm->node_num);
2643 /* get an extra reference on the mle.
2644 * otherwise the assert_master from the new
2645 * master will destroy this.
2646 */
2647 dlm_get_mle_inuse(mle);
2648 spin_unlock(&dlm->master_lock);
2649 spin_unlock(&dlm->spinlock);
2650
2651 if (ret == -EEXIST) {
2652 mlog(0, "another process is already migrating it\n");
2653 goto fail;
2654 }
2655 mle_added = 1;
2656
2657 /*
2658 * set the MIGRATING flag and flush asts
2659 * if we fail after this we need to re-dirty the lockres
2660 */
2661 if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2662 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2663 "the target went down.\n", res->lockname.len,
2664 res->lockname.name, target);
2665 spin_lock(&res->spinlock);
2666 res->state &= ~DLM_LOCK_RES_MIGRATING;
2667 wake = 1;
2668 spin_unlock(&res->spinlock);
2669 ret = -EINVAL;
2670 }
2671
2672fail:
2673 if (ret != -EEXIST && oldmle) {
2674 /* master is known, detach if not already detached */
2675 dlm_mle_detach_hb_events(dlm, oldmle);
2676 dlm_put_mle(oldmle);
2677 }
2678
2679 if (ret < 0) {
2680 if (mle_added) {
2681 dlm_mle_detach_hb_events(dlm, mle);
2682 dlm_put_mle(mle);
2683 dlm_put_mle_inuse(mle);
2684 } else if (mle) {
2685 kmem_cache_free(dlm_mle_cache, mle);
2686 mle = NULL;
2687 }
2688 goto leave;
2689 }
2690
2691 /*
2692 * at this point, we have a migration target, an mle
2693 * in the master list, and the MIGRATING flag set on
2694 * the lockres
2695 */
2696
2697 /* now that remote nodes are spinning on the MIGRATING flag,
2698 * ensure that all assert_master work is flushed. */
2699 flush_workqueue(dlm->dlm_worker);
2700
2701 /* notify new node and send all lock state */
2702 /* call send_one_lockres with migration flag.
2703 * this serves as notice to the target node that a
2704 * migration is starting. */
2705 ret = dlm_send_one_lockres(dlm, res, mres, target,
2706 DLM_MRES_MIGRATION);
2707
2708 if (ret < 0) {
2709 mlog(0, "migration to node %u failed with %d\n",
2710 target, ret);
2711 /* migration failed, detach and clean up mle */
2712 dlm_mle_detach_hb_events(dlm, mle);
2713 dlm_put_mle(mle);
2714 dlm_put_mle_inuse(mle);
2715 spin_lock(&res->spinlock);
2716 res->state &= ~DLM_LOCK_RES_MIGRATING;
2717 wake = 1;
2718 spin_unlock(&res->spinlock);
2719 if (dlm_is_host_down(ret))
2720 dlm_wait_for_node_death(dlm, target,
2721 DLM_NODE_DEATH_WAIT_MAX);
2722 goto leave;
2723 }
2724
2725 /* at this point, the target sends a message to all nodes,
2726 * (using dlm_do_migrate_request). this node is skipped since
2727 * we had to put an mle in the list to begin the process. this
2728 * node now waits for target to do an assert master. this node
2729 * will be the last one notified, ensuring that the migration
2730 * is complete everywhere. if the target dies while this is
2731 * going on, some nodes could potentially see the target as the
2732 * master, so it is important that my recovery finds the migration
2733 * mle and sets the master to UNKNOWN. */
2734
2735
2736 /* wait for new node to assert master */
2737 while (1) {
2738 ret = wait_event_interruptible_timeout(mle->wq,
2739 (atomic_read(&mle->woken) == 1),
2740 msecs_to_jiffies(5000));
2741
2742 if (ret >= 0) {
2743 if (atomic_read(&mle->woken) == 1 ||
2744 res->owner == target)
2745 break;
2746
2747 mlog(0, "%s:%.*s: timed out during migration\n",
2748 dlm->name, res->lockname.len, res->lockname.name);
2749 /* avoid hang during shutdown when migrating lockres
2750 * to a node which also goes down */
2751 if (dlm_is_node_dead(dlm, target)) {
2752 mlog(0, "%s:%.*s: expected migration "
2753 "target %u is no longer up, restarting\n",
2754 dlm->name, res->lockname.len,
2755 res->lockname.name, target);
2756 ret = -EINVAL;
2757 /* migration failed, detach and clean up mle */
2758 dlm_mle_detach_hb_events(dlm, mle);
2759 dlm_put_mle(mle);
2760 dlm_put_mle_inuse(mle);
2761 spin_lock(&res->spinlock);
2762 res->state &= ~DLM_LOCK_RES_MIGRATING;
2763 wake = 1;
2764 spin_unlock(&res->spinlock);
2765 goto leave;
2766 }
2767 } else
2768 mlog(0, "%s:%.*s: caught signal during migration\n",
2769 dlm->name, res->lockname.len, res->lockname.name);
2770 }
2771
2772 /* all done, set the owner, clear the flag */
2773 spin_lock(&res->spinlock);
2774 dlm_set_lockres_owner(dlm, res, target);
2775 res->state &= ~DLM_LOCK_RES_MIGRATING;
2776 dlm_remove_nonlocal_locks(dlm, res);
2777 spin_unlock(&res->spinlock);
2778 wake_up(&res->wq);
2779
2780 /* master is known, detach if not already detached */
2781 dlm_mle_detach_hb_events(dlm, mle);
2782 dlm_put_mle_inuse(mle);
2783 ret = 0;
2784
2785 dlm_lockres_calc_usage(dlm, res);
2786
2787leave:
2788 /* re-dirty the lockres if we failed */
2789 if (ret < 0)
2790 dlm_kick_thread(dlm, res);
2791
2792 /* wake up waiters if the MIGRATING flag got set
2793 * but migration failed */
2794 if (wake)
2795 wake_up(&res->wq);
2796
2797 if (mres)
2798 free_page((unsigned long)mres);
2799
2800 dlm_put(dlm);
2801
2802 mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen,
2803 name, target, ret);
2804 return ret;
2805}
2806
2807#define DLM_MIGRATION_RETRY_MS 100
2808
2809/*
2810 * Should be called only after beginning the domain leave process.
2811 * There should not be any remaining locks on nonlocal lock resources,
2812 * and there should be no local locks left on locally mastered resources.
2813 *
2814 * Called with the dlm spinlock held, may drop it to do migration, but
2815 * will re-acquire before exit.
2816 *
2817 * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2818 */
2819int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2820{
2821 int ret;
2822 int lock_dropped = 0;
2823 u8 target = O2NM_MAX_NODES;
2824
2825 assert_spin_locked(&dlm->spinlock);
2826
2827 spin_lock(&res->spinlock);
2828 if (dlm_is_lockres_migrateable(dlm, res))
2829 target = dlm_pick_migration_target(dlm, res);
2830 spin_unlock(&res->spinlock);
2831
2832 if (target == O2NM_MAX_NODES)
2833 goto leave;
2834
2835 /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2836 spin_unlock(&dlm->spinlock);
2837 lock_dropped = 1;
2838 ret = dlm_migrate_lockres(dlm, res, target);
2839 if (ret)
2840 mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2841 dlm->name, res->lockname.len, res->lockname.name,
2842 target, ret);
2843 spin_lock(&dlm->spinlock);
2844leave:
2845 return lock_dropped;
2846}
2847
2848int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2849{
2850 int ret;
2851 spin_lock(&dlm->ast_lock);
2852 spin_lock(&lock->spinlock);
2853 ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2854 spin_unlock(&lock->spinlock);
2855 spin_unlock(&dlm->ast_lock);
2856 return ret;
2857}
2858
2859static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2860 struct dlm_lock_resource *res,
2861 u8 mig_target)
2862{
2863 int can_proceed;
2864 spin_lock(&res->spinlock);
2865 can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2866 spin_unlock(&res->spinlock);
2867
2868 /* target has died, so make the caller break out of the
2869 * wait_event, but caller must recheck the domain_map */
2870 spin_lock(&dlm->spinlock);
2871 if (!test_bit(mig_target, dlm->domain_map))
2872 can_proceed = 1;
2873 spin_unlock(&dlm->spinlock);
2874 return can_proceed;
2875}
2876
2877static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2878 struct dlm_lock_resource *res)
2879{
2880 int ret;
2881 spin_lock(&res->spinlock);
2882 ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2883 spin_unlock(&res->spinlock);
2884 return ret;
2885}
2886
2887
2888static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2889 struct dlm_lock_resource *res,
2890 u8 target)
2891{
2892 int ret = 0;
2893
2894 mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2895 res->lockname.len, res->lockname.name, dlm->node_num,
2896 target);
2897 /* need to set MIGRATING flag on lockres. this is done by
2898 * ensuring that all asts have been flushed for this lockres. */
2899 spin_lock(&res->spinlock);
2900 BUG_ON(res->migration_pending);
2901 res->migration_pending = 1;
2902 /* strategy is to reserve an extra ast then release
2903 * it below, letting the release do all of the work */
2904 __dlm_lockres_reserve_ast(res);
2905 spin_unlock(&res->spinlock);
2906
2907 /* now flush all the pending asts */
2908 dlm_kick_thread(dlm, res);
2909 /* before waiting on DIRTY, block processes which may
2910 * try to dirty the lockres before MIGRATING is set */
2911 spin_lock(&res->spinlock);
2912 BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2913 res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2914 spin_unlock(&res->spinlock);
2915 /* now wait on any pending asts and the DIRTY state */
2916 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2917 dlm_lockres_release_ast(dlm, res);
2918
2919 mlog(0, "about to wait on migration_wq, dirty=%s\n",
2920 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2921 /* if the extra ref we just put was the final one, this
2922 * will pass thru immediately. otherwise, we need to wait
2923 * for the last ast to finish. */
2924again:
2925 ret = wait_event_interruptible_timeout(dlm->migration_wq,
2926 dlm_migration_can_proceed(dlm, res, target),
2927 msecs_to_jiffies(1000));
2928 if (ret < 0) {
2929 mlog(0, "woken again: migrating? %s, dead? %s\n",
2930 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2931 test_bit(target, dlm->domain_map) ? "no":"yes");
2932 } else {
2933 mlog(0, "all is well: migrating? %s, dead? %s\n",
2934 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2935 test_bit(target, dlm->domain_map) ? "no":"yes");
2936 }
2937 if (!dlm_migration_can_proceed(dlm, res, target)) {
2938 mlog(0, "trying again...\n");
2939 goto again;
2940 }
2941
2942 ret = 0;
2943 /* did the target go down or die? */
2944 spin_lock(&dlm->spinlock);
2945 if (!test_bit(target, dlm->domain_map)) {
2946 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2947 target);
2948 ret = -EHOSTDOWN;
2949 }
2950 spin_unlock(&dlm->spinlock);
2951
2952 /*
2953 * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2954 * another try; otherwise, we are sure the MIGRATING state is there,
2955 * drop the unneded state which blocked threads trying to DIRTY
2956 */
2957 spin_lock(&res->spinlock);
2958 BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2959 res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2960 if (!ret)
2961 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2962 else
2963 res->migration_pending = 0;
2964 spin_unlock(&res->spinlock);
2965
2966 /*
2967 * at this point:
2968 *
2969 * o the DLM_LOCK_RES_MIGRATING flag is set if target not down
2970 * o there are no pending asts on this lockres
2971 * o all processes trying to reserve an ast on this
2972 * lockres must wait for the MIGRATING flag to clear
2973 */
2974 return ret;
2975}
2976
2977/* last step in the migration process.
2978 * original master calls this to free all of the dlm_lock
2979 * structures that used to be for other nodes. */
2980static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2981 struct dlm_lock_resource *res)
2982{
2983 struct list_head *queue = &res->granted;
2984 int i, bit;
2985 struct dlm_lock *lock, *next;
2986
2987 assert_spin_locked(&res->spinlock);
2988
2989 BUG_ON(res->owner == dlm->node_num);
2990
2991 for (i=0; i<3; i++) {
2992 list_for_each_entry_safe(lock, next, queue, list) {
2993 if (lock->ml.node != dlm->node_num) {
2994 mlog(0, "putting lock for node %u\n",
2995 lock->ml.node);
2996 /* be extra careful */
2997 BUG_ON(!list_empty(&lock->ast_list));
2998 BUG_ON(!list_empty(&lock->bast_list));
2999 BUG_ON(lock->ast_pending);
3000 BUG_ON(lock->bast_pending);
3001 dlm_lockres_clear_refmap_bit(dlm, res,
3002 lock->ml.node);
3003 list_del_init(&lock->list);
3004 dlm_lock_put(lock);
3005 /* In a normal unlock, we would have added a
3006 * DLM_UNLOCK_FREE_LOCK action. Force it. */
3007 dlm_lock_put(lock);
3008 }
3009 }
3010 queue++;
3011 }
3012 bit = 0;
3013 while (1) {
3014 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
3015 if (bit >= O2NM_MAX_NODES)
3016 break;
3017 /* do not clear the local node reference, if there is a
3018 * process holding this, let it drop the ref itself */
3019 if (bit != dlm->node_num) {
3020 mlog(0, "%s:%.*s: node %u had a ref to this "
3021 "migrating lockres, clearing\n", dlm->name,
3022 res->lockname.len, res->lockname.name, bit);
3023 dlm_lockres_clear_refmap_bit(dlm, res, bit);
3024 }
3025 bit++;
3026 }
3027}
3028
3029/*
3030 * Pick a node to migrate the lock resource to. This function selects a
3031 * potential target based first on the locks and then on refmap. It skips
3032 * nodes that are in the process of exiting the domain.
3033 */
3034static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
3035 struct dlm_lock_resource *res)
3036{
3037 enum dlm_lockres_list idx;
3038 struct list_head *queue = &res->granted;
3039 struct dlm_lock *lock;
3040 int noderef;
3041 u8 nodenum = O2NM_MAX_NODES;
3042
3043 assert_spin_locked(&dlm->spinlock);
3044 assert_spin_locked(&res->spinlock);
3045
3046 /* Go through all the locks */
3047 for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
3048 queue = dlm_list_idx_to_ptr(res, idx);
3049 list_for_each_entry(lock, queue, list) {
3050 if (lock->ml.node == dlm->node_num)
3051 continue;
3052 if (test_bit(lock->ml.node, dlm->exit_domain_map))
3053 continue;
3054 nodenum = lock->ml.node;
3055 goto bail;
3056 }
3057 }
3058
3059 /* Go thru the refmap */
3060 noderef = -1;
3061 while (1) {
3062 noderef = find_next_bit(res->refmap, O2NM_MAX_NODES,
3063 noderef + 1);
3064 if (noderef >= O2NM_MAX_NODES)
3065 break;
3066 if (noderef == dlm->node_num)
3067 continue;
3068 if (test_bit(noderef, dlm->exit_domain_map))
3069 continue;
3070 nodenum = noderef;
3071 goto bail;
3072 }
3073
3074bail:
3075 return nodenum;
3076}
3077
3078/* this is called by the new master once all lockres
3079 * data has been received */
3080static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
3081 struct dlm_lock_resource *res,
3082 u8 master, u8 new_master,
3083 struct dlm_node_iter *iter)
3084{
3085 struct dlm_migrate_request migrate;
3086 int ret, skip, status = 0;
3087 int nodenum;
3088
3089 memset(&migrate, 0, sizeof(migrate));
3090 migrate.namelen = res->lockname.len;
3091 memcpy(migrate.name, res->lockname.name, migrate.namelen);
3092 migrate.new_master = new_master;
3093 migrate.master = master;
3094
3095 ret = 0;
3096
3097 /* send message to all nodes, except the master and myself */
3098 while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
3099 if (nodenum == master ||
3100 nodenum == new_master)
3101 continue;
3102
3103 /* We could race exit domain. If exited, skip. */
3104 spin_lock(&dlm->spinlock);
3105 skip = (!test_bit(nodenum, dlm->domain_map));
3106 spin_unlock(&dlm->spinlock);
3107 if (skip) {
3108 clear_bit(nodenum, iter->node_map);
3109 continue;
3110 }
3111
3112 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
3113 &migrate, sizeof(migrate), nodenum,
3114 &status);
3115 if (ret < 0) {
3116 mlog(ML_ERROR, "%s: res %.*s, Error %d send "
3117 "MIGRATE_REQUEST to node %u\n", dlm->name,
3118 migrate.namelen, migrate.name, ret, nodenum);
3119 if (!dlm_is_host_down(ret)) {
3120 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
3121 BUG();
3122 }
3123 clear_bit(nodenum, iter->node_map);
3124 ret = 0;
3125 } else if (status < 0) {
3126 mlog(0, "migrate request (node %u) returned %d!\n",
3127 nodenum, status);
3128 ret = status;
3129 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
3130 /* during the migration request we short-circuited
3131 * the mastery of the lockres. make sure we have
3132 * a mastery ref for nodenum */
3133 mlog(0, "%s:%.*s: need ref for node %u\n",
3134 dlm->name, res->lockname.len, res->lockname.name,
3135 nodenum);
3136 spin_lock(&res->spinlock);
3137 dlm_lockres_set_refmap_bit(dlm, res, nodenum);
3138 spin_unlock(&res->spinlock);
3139 }
3140 }
3141
3142 if (ret < 0)
3143 mlog_errno(ret);
3144
3145 mlog(0, "returning ret=%d\n", ret);
3146 return ret;
3147}
3148
3149
3150/* if there is an existing mle for this lockres, we now know who the master is.
3151 * (the one who sent us *this* message) we can clear it up right away.
3152 * since the process that put the mle on the list still has a reference to it,
3153 * we can unhash it now, set the master and wake the process. as a result,
3154 * we will have no mle in the list to start with. now we can add an mle for
3155 * the migration and this should be the only one found for those scanning the
3156 * list. */
3157int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
3158 void **ret_data)
3159{
3160 struct dlm_ctxt *dlm = data;
3161 struct dlm_lock_resource *res = NULL;
3162 struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
3163 struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
3164 const char *name;
3165 unsigned int namelen, hash;
3166 int ret = 0;
3167
3168 if (!dlm_grab(dlm))
3169 return 0;
3170
3171 name = migrate->name;
3172 namelen = migrate->namelen;
3173 hash = dlm_lockid_hash(name, namelen);
3174
3175 /* preallocate.. if this fails, abort */
3176 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
3177
3178 if (!mle) {
3179 ret = -ENOMEM;
3180 goto leave;
3181 }
3182
3183 /* check for pre-existing lock */
3184 spin_lock(&dlm->spinlock);
3185 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
3186 if (res) {
3187 spin_lock(&res->spinlock);
3188 if (res->state & DLM_LOCK_RES_RECOVERING) {
3189 /* if all is working ok, this can only mean that we got
3190 * a migrate request from a node that we now see as
3191 * dead. what can we do here? drop it to the floor? */
3192 spin_unlock(&res->spinlock);
3193 mlog(ML_ERROR, "Got a migrate request, but the "
3194 "lockres is marked as recovering!");
3195 kmem_cache_free(dlm_mle_cache, mle);
3196 ret = -EINVAL; /* need a better solution */
3197 goto unlock;
3198 }
3199 res->state |= DLM_LOCK_RES_MIGRATING;
3200 spin_unlock(&res->spinlock);
3201 }
3202
3203 spin_lock(&dlm->master_lock);
3204 /* ignore status. only nonzero status would BUG. */
3205 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3206 name, namelen,
3207 migrate->new_master,
3208 migrate->master);
3209
3210 spin_unlock(&dlm->master_lock);
3211unlock:
3212 spin_unlock(&dlm->spinlock);
3213
3214 if (oldmle) {
3215 /* master is known, detach if not already detached */
3216 dlm_mle_detach_hb_events(dlm, oldmle);
3217 dlm_put_mle(oldmle);
3218 }
3219
3220 if (res)
3221 dlm_lockres_put(res);
3222leave:
3223 dlm_put(dlm);
3224 return ret;
3225}
3226
3227/* must be holding dlm->spinlock and dlm->master_lock
3228 * when adding a migration mle, we can clear any other mles
3229 * in the master list because we know with certainty that
3230 * the master is "master". so we remove any old mle from
3231 * the list after setting it's master field, and then add
3232 * the new migration mle. this way we can hold with the rule
3233 * of having only one mle for a given lock name at all times. */
3234static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3235 struct dlm_lock_resource *res,
3236 struct dlm_master_list_entry *mle,
3237 struct dlm_master_list_entry **oldmle,
3238 const char *name, unsigned int namelen,
3239 u8 new_master, u8 master)
3240{
3241 int found;
3242 int ret = 0;
3243
3244 *oldmle = NULL;
3245
3246 assert_spin_locked(&dlm->spinlock);
3247 assert_spin_locked(&dlm->master_lock);
3248
3249 /* caller is responsible for any ref taken here on oldmle */
3250 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3251 if (found) {
3252 struct dlm_master_list_entry *tmp = *oldmle;
3253 spin_lock(&tmp->spinlock);
3254 if (tmp->type == DLM_MLE_MIGRATION) {
3255 if (master == dlm->node_num) {
3256 /* ah another process raced me to it */
3257 mlog(0, "tried to migrate %.*s, but some "
3258 "process beat me to it\n",
3259 namelen, name);
3260 spin_unlock(&tmp->spinlock);
3261 return -EEXIST;
3262 } else {
3263 /* bad. 2 NODES are trying to migrate! */
3264 mlog(ML_ERROR, "migration error mle: "
3265 "master=%u new_master=%u // request: "
3266 "master=%u new_master=%u // "
3267 "lockres=%.*s\n",
3268 tmp->master, tmp->new_master,
3269 master, new_master,
3270 namelen, name);
3271 BUG();
3272 }
3273 } else {
3274 /* this is essentially what assert_master does */
3275 tmp->master = master;
3276 atomic_set(&tmp->woken, 1);
3277 wake_up(&tmp->wq);
3278 /* remove it so that only one mle will be found */
3279 __dlm_unlink_mle(dlm, tmp);
3280 __dlm_mle_detach_hb_events(dlm, tmp);
3281 if (tmp->type == DLM_MLE_MASTER) {
3282 ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3283 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3284 "telling master to get ref "
3285 "for cleared out mle during "
3286 "migration\n", dlm->name,
3287 namelen, name, master,
3288 new_master);
3289 }
3290 }
3291 spin_unlock(&tmp->spinlock);
3292 }
3293
3294 /* now add a migration mle to the tail of the list */
3295 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3296 mle->new_master = new_master;
3297 /* the new master will be sending an assert master for this.
3298 * at that point we will get the refmap reference */
3299 mle->master = master;
3300 /* do this for consistency with other mle types */
3301 set_bit(new_master, mle->maybe_map);
3302 __dlm_insert_mle(dlm, mle);
3303
3304 return ret;
3305}
3306
3307/*
3308 * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3309 */
3310static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3311 struct dlm_master_list_entry *mle)
3312{
3313 struct dlm_lock_resource *res;
3314
3315 /* Find the lockres associated to the mle and set its owner to UNK */
3316 res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3317 mle->mnamehash);
3318 if (res) {
3319 spin_unlock(&dlm->master_lock);
3320
3321 /* move lockres onto recovery list */
3322 spin_lock(&res->spinlock);
3323 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3324 dlm_move_lockres_to_recovery_list(dlm, res);
3325 spin_unlock(&res->spinlock);
3326 dlm_lockres_put(res);
3327
3328 /* about to get rid of mle, detach from heartbeat */
3329 __dlm_mle_detach_hb_events(dlm, mle);
3330
3331 /* dump the mle */
3332 spin_lock(&dlm->master_lock);
3333 __dlm_put_mle(mle);
3334 spin_unlock(&dlm->master_lock);
3335 }
3336
3337 return res;
3338}
3339
3340static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3341 struct dlm_master_list_entry *mle)
3342{
3343 __dlm_mle_detach_hb_events(dlm, mle);
3344
3345 spin_lock(&mle->spinlock);
3346 __dlm_unlink_mle(dlm, mle);
3347 atomic_set(&mle->woken, 1);
3348 spin_unlock(&mle->spinlock);
3349
3350 wake_up(&mle->wq);
3351}
3352
3353static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3354 struct dlm_master_list_entry *mle, u8 dead_node)
3355{
3356 int bit;
3357
3358 BUG_ON(mle->type != DLM_MLE_BLOCK);
3359
3360 spin_lock(&mle->spinlock);
3361 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3362 if (bit != dead_node) {
3363 mlog(0, "mle found, but dead node %u would not have been "
3364 "master\n", dead_node);
3365 spin_unlock(&mle->spinlock);
3366 } else {
3367 /* Must drop the refcount by one since the assert_master will
3368 * never arrive. This may result in the mle being unlinked and
3369 * freed, but there may still be a process waiting in the
3370 * dlmlock path which is fine. */
3371 mlog(0, "node %u was expected master\n", dead_node);
3372 atomic_set(&mle->woken, 1);
3373 spin_unlock(&mle->spinlock);
3374 wake_up(&mle->wq);
3375
3376 /* Do not need events any longer, so detach from heartbeat */
3377 __dlm_mle_detach_hb_events(dlm, mle);
3378 __dlm_put_mle(mle);
3379 }
3380}
3381
3382void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3383{
3384 struct dlm_master_list_entry *mle;
3385 struct dlm_lock_resource *res;
3386 struct hlist_head *bucket;
3387 struct hlist_node *tmp;
3388 unsigned int i;
3389
3390 mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
3391top:
3392 assert_spin_locked(&dlm->spinlock);
3393
3394 /* clean the master list */
3395 spin_lock(&dlm->master_lock);
3396 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3397 bucket = dlm_master_hash(dlm, i);
3398 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3399 BUG_ON(mle->type != DLM_MLE_BLOCK &&
3400 mle->type != DLM_MLE_MASTER &&
3401 mle->type != DLM_MLE_MIGRATION);
3402
3403 /* MASTER mles are initiated locally. The waiting
3404 * process will notice the node map change shortly.
3405 * Let that happen as normal. */
3406 if (mle->type == DLM_MLE_MASTER)
3407 continue;
3408
3409 /* BLOCK mles are initiated by other nodes. Need to
3410 * clean up if the dead node would have been the
3411 * master. */
3412 if (mle->type == DLM_MLE_BLOCK) {
3413 dlm_clean_block_mle(dlm, mle, dead_node);
3414 continue;
3415 }
3416
3417 /* Everything else is a MIGRATION mle */
3418
3419 /* The rule for MIGRATION mles is that the master
3420 * becomes UNKNOWN if *either* the original or the new
3421 * master dies. All UNKNOWN lockres' are sent to
3422 * whichever node becomes the recovery master. The new
3423 * master is responsible for determining if there is
3424 * still a master for this lockres, or if he needs to
3425 * take over mastery. Either way, this node should
3426 * expect another message to resolve this. */
3427
3428 if (mle->master != dead_node &&
3429 mle->new_master != dead_node)
3430 continue;
3431
3432 if (mle->new_master == dead_node && mle->inuse) {
3433 mlog(ML_NOTICE, "%s: target %u died during "
3434 "migration from %u, the MLE is "
3435 "still keep used, ignore it!\n",
3436 dlm->name, dead_node,
3437 mle->master);
3438 continue;
3439 }
3440
3441 /* If we have reached this point, this mle needs to be
3442 * removed from the list and freed. */
3443 dlm_clean_migration_mle(dlm, mle);
3444
3445 mlog(0, "%s: node %u died during migration from "
3446 "%u to %u!\n", dlm->name, dead_node, mle->master,
3447 mle->new_master);
3448
3449 /* If we find a lockres associated with the mle, we've
3450 * hit this rare case that messes up our lock ordering.
3451 * If so, we need to drop the master lock so that we can
3452 * take the lockres lock, meaning that we will have to
3453 * restart from the head of list. */
3454 res = dlm_reset_mleres_owner(dlm, mle);
3455 if (res)
3456 /* restart */
3457 goto top;
3458
3459 /* This may be the last reference */
3460 __dlm_put_mle(mle);
3461 }
3462 }
3463 spin_unlock(&dlm->master_lock);
3464}
3465
3466int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3467 u8 old_master)
3468{
3469 struct dlm_node_iter iter;
3470 int ret = 0;
3471
3472 spin_lock(&dlm->spinlock);
3473 dlm_node_iter_init(dlm->domain_map, &iter);
3474 clear_bit(old_master, iter.node_map);
3475 clear_bit(dlm->node_num, iter.node_map);
3476 spin_unlock(&dlm->spinlock);
3477
3478 /* ownership of the lockres is changing. account for the
3479 * mastery reference here since old_master will briefly have
3480 * a reference after the migration completes */
3481 spin_lock(&res->spinlock);
3482 dlm_lockres_set_refmap_bit(dlm, res, old_master);
3483 spin_unlock(&res->spinlock);
3484
3485 mlog(0, "now time to do a migrate request to other nodes\n");
3486 ret = dlm_do_migrate_request(dlm, res, old_master,
3487 dlm->node_num, &iter);
3488 if (ret < 0) {
3489 mlog_errno(ret);
3490 goto leave;
3491 }
3492
3493 mlog(0, "doing assert master of %.*s to all except the original node\n",
3494 res->lockname.len, res->lockname.name);
3495 /* this call now finishes out the nodemap
3496 * even if one or more nodes die */
3497 ret = dlm_do_assert_master(dlm, res, iter.node_map,
3498 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3499 if (ret < 0) {
3500 /* no longer need to retry. all living nodes contacted. */
3501 mlog_errno(ret);
3502 ret = 0;
3503 }
3504
3505 memset(iter.node_map, 0, sizeof(iter.node_map));
3506 set_bit(old_master, iter.node_map);
3507 mlog(0, "doing assert master of %.*s back to %u\n",
3508 res->lockname.len, res->lockname.name, old_master);
3509 ret = dlm_do_assert_master(dlm, res, iter.node_map,
3510 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3511 if (ret < 0) {
3512 mlog(0, "assert master to original master failed "
3513 "with %d.\n", ret);
3514 /* the only nonzero status here would be because of
3515 * a dead original node. we're done. */
3516 ret = 0;
3517 }
3518
3519 /* all done, set the owner, clear the flag */
3520 spin_lock(&res->spinlock);
3521 dlm_set_lockres_owner(dlm, res, dlm->node_num);
3522 res->state &= ~DLM_LOCK_RES_MIGRATING;
3523 spin_unlock(&res->spinlock);
3524 /* re-dirty it on the new master */
3525 dlm_kick_thread(dlm, res);
3526 wake_up(&res->wq);
3527leave:
3528 return ret;
3529}
3530
3531/*
3532 * LOCKRES AST REFCOUNT
3533 * this is integral to migration
3534 */
3535
3536/* for future intent to call an ast, reserve one ahead of time.
3537 * this should be called only after waiting on the lockres
3538 * with dlm_wait_on_lockres, and while still holding the
3539 * spinlock after the call. */
3540void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3541{
3542 assert_spin_locked(&res->spinlock);
3543 if (res->state & DLM_LOCK_RES_MIGRATING) {
3544 __dlm_print_one_lock_resource(res);
3545 }
3546 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3547
3548 atomic_inc(&res->asts_reserved);
3549}
3550
3551/*
3552 * used to drop the reserved ast, either because it went unused,
3553 * or because the ast/bast was actually called.
3554 *
3555 * also, if there is a pending migration on this lockres,
3556 * and this was the last pending ast on the lockres,
3557 * atomically set the MIGRATING flag before we drop the lock.
3558 * this is how we ensure that migration can proceed with no
3559 * asts in progress. note that it is ok if the state of the
3560 * queues is such that a lock should be granted in the future
3561 * or that a bast should be fired, because the new master will
3562 * shuffle the lists on this lockres as soon as it is migrated.
3563 */
3564void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3565 struct dlm_lock_resource *res)
3566{
3567 if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3568 return;
3569
3570 if (!res->migration_pending) {
3571 spin_unlock(&res->spinlock);
3572 return;
3573 }
3574
3575 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3576 res->migration_pending = 0;
3577 res->state |= DLM_LOCK_RES_MIGRATING;
3578 spin_unlock(&res->spinlock);
3579 wake_up(&res->wq);
3580 wake_up(&dlm->migration_wq);
3581}
3582
3583void dlm_force_free_mles(struct dlm_ctxt *dlm)
3584{
3585 int i;
3586 struct hlist_head *bucket;
3587 struct dlm_master_list_entry *mle;
3588 struct hlist_node *tmp;
3589
3590 /*
3591 * We notified all other nodes that we are exiting the domain and
3592 * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3593 * around we force free them and wake any processes that are waiting
3594 * on the mles
3595 */
3596 spin_lock(&dlm->spinlock);
3597 spin_lock(&dlm->master_lock);
3598
3599 BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3600 BUG_ON((find_next_bit(dlm->domain_map, O2NM_MAX_NODES, 0) < O2NM_MAX_NODES));
3601
3602 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3603 bucket = dlm_master_hash(dlm, i);
3604 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3605 if (mle->type != DLM_MLE_BLOCK) {
3606 mlog(ML_ERROR, "bad mle: %p\n", mle);
3607 dlm_print_one_mle(mle);
3608 }
3609 atomic_set(&mle->woken, 1);
3610 wake_up(&mle->wq);
3611
3612 __dlm_unlink_mle(dlm, mle);
3613 __dlm_mle_detach_hb_events(dlm, mle);
3614 __dlm_put_mle(mle);
3615 }
3616 }
3617 spin_unlock(&dlm->master_lock);
3618 spin_unlock(&dlm->spinlock);
3619}