projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
ocfs2/dlm: Activate dlm->master_hash for master list entries
[linux-2.6-block.git] / fs / ocfs2 / dlm / dlmmaster.c
CommitLineData
6714d8e8
KH		 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/utsname.h>
34#include <linux/init.h>
35#include <linux/sysctl.h>
36#include <linux/random.h>
37#include <linux/blkdev.h>
38#include <linux/socket.h>
39#include <linux/inet.h>
40#include <linux/spinlock.h>
41#include <linux/delay.h>
42
43
44#include "cluster/heartbeat.h"
45#include "cluster/nodemanager.h"
46#include "cluster/tcp.h"
47
48#include "dlmapi.h"
49#include "dlmcommon.h"
82353b59 50#include "dlmdomain.h"
e5a0334c 51#include "dlmdebug.h"
6714d8e8
KH		 52
53#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
54#include "cluster/masklog.h"
55
6714d8e8
KH		 56static void dlm_mle_node_down(struct dlm_ctxt *dlm,
57 struct dlm_master_list_entry *mle,
58 struct o2nm_node *node,
59 int idx);
60static void dlm_mle_node_up(struct dlm_ctxt *dlm,
61 struct dlm_master_list_entry *mle,
62 struct o2nm_node *node,
63 int idx);
64
65static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
ba2bf218
KH		 66static int dlm_do_assert_master(struct dlm_ctxt *dlm,
67 struct dlm_lock_resource *res,
68 void *nodemap, u32 flags);
f3f85464 69static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
6714d8e8 70
f77a9a78 71static inline void __dlm_mle_name(struct dlm_master_list_entry *mle,
2ed6c750
SM		 72 unsigned char **name, unsigned int *namelen,
73 unsigned int *namehash)
f77a9a78
SM		 74{
75 BUG_ON(mle->type != DLM_MLE_BLOCK &&
76 mle->type != DLM_MLE_MASTER &&
77 mle->type != DLM_MLE_MIGRATION);
78
79 if (mle->type != DLM_MLE_MASTER) {
80 *name = mle->u.mlename.name;
81 *namelen = mle->u.mlename.len;
2ed6c750
SM		 82 if (namehash)
83 *namehash = mle->u.mlename.hash;
f77a9a78
SM		 84 } else {
85 *name = (unsigned char *)mle->u.mleres->lockname.name;
86 *namelen = mle->u.mleres->lockname.len;
2ed6c750
SM		 87 if (namehash)
88 *namehash = mle->u.mleres->lockname.hash;
f77a9a78
SM		 89 }
90}
91
6714d8e8
KH		 92static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
93 struct dlm_master_list_entry *mle,
94 const char *name,
95 unsigned int namelen)
96{
f77a9a78
SM		 97 unsigned char *mlename;
98 unsigned int mlelen;
6714d8e8
KH		 99
100 if (dlm != mle->dlm)
101 return 0;
102
2ed6c750 103 __dlm_mle_name(mle, &mlename, &mlelen, NULL);
f77a9a78
SM		 104
105 if (namelen != mlelen || memcmp(name, mlename, namelen) != 0)
106 return 0;
107
6714d8e8
KH		 108 return 1;
109}
110
724bdca9
SM		 111static struct kmem_cache *dlm_lockres_cache = NULL;
112static struct kmem_cache *dlm_lockname_cache = NULL;
e18b890b 113static struct kmem_cache *dlm_mle_cache = NULL;
6714d8e8 114
6714d8e8
KH		 115static void dlm_mle_release(struct kref *kref);
116static void dlm_init_mle(struct dlm_master_list_entry *mle,
117 enum dlm_mle_type type,
118 struct dlm_ctxt *dlm,
119 struct dlm_lock_resource *res,
120 const char *name,
121 unsigned int namelen);
122static void dlm_put_mle(struct dlm_master_list_entry *mle);
123static void __dlm_put_mle(struct dlm_master_list_entry *mle);
124static int dlm_find_mle(struct dlm_ctxt *dlm,
125 struct dlm_master_list_entry **mle,
126 char *name, unsigned int namelen);
127
ba2bf218
KH		 128static int dlm_do_master_request(struct dlm_lock_resource *res,
129 struct dlm_master_list_entry *mle, int to);
6714d8e8
KH		 130
131
132static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
133 struct dlm_lock_resource *res,
134 struct dlm_master_list_entry *mle,
135 int *blocked);
136static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
137 struct dlm_lock_resource *res,
138 struct dlm_master_list_entry *mle,
139 int blocked);
140static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
141 struct dlm_lock_resource *res,
142 struct dlm_master_list_entry *mle,
143 struct dlm_master_list_entry **oldmle,
144 const char *name, unsigned int namelen,
145 u8 new_master, u8 master);
146
147static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
148 struct dlm_lock_resource *res);
149static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
150 struct dlm_lock_resource *res);
151static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
152 struct dlm_lock_resource *res,
153 u8 target);
c03872f5
KH		 154static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
155 struct dlm_lock_resource *res);
6714d8e8
KH		 156
157
158int dlm_is_host_down(int errno)
159{
160 switch (errno) {
161 case -EBADF:
162 case -ECONNREFUSED:
163 case -ENOTCONN:
164 case -ECONNRESET:
165 case -EPIPE:
166 case -EHOSTDOWN:
167 case -EHOSTUNREACH:
168 case -ETIMEDOUT:
169 case -ECONNABORTED:
170 case -ENETDOWN:
171 case -ENETUNREACH:
172 case -ENETRESET:
173 case -ESHUTDOWN:
174 case -ENOPROTOOPT:
175 case -EINVAL: /* if returned from our tcp code,
176 this means there is no socket */
177 return 1;
178 }
179 return 0;
180}
181
182
183/*
184 * MASTER LIST FUNCTIONS
185 */
186
187
188/*
189 * regarding master list entries and heartbeat callbacks:
190 *
191 * in order to avoid sleeping and allocation that occurs in
192 * heartbeat, master list entries are simply attached to the
193 * dlm's established heartbeat callbacks. the mle is attached
194 * when it is created, and since the dlm->spinlock is held at
195 * that time, any heartbeat event will be properly discovered
196 * by the mle. the mle needs to be detached from the
197 * dlm->mle_hb_events list as soon as heartbeat events are no
198 * longer useful to the mle, and before the mle is freed.
199 *
200 * as a general rule, heartbeat events are no longer needed by
201 * the mle once an "answer" regarding the lock master has been
202 * received.
203 */
204static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
205 struct dlm_master_list_entry *mle)
206{
207 assert_spin_locked(&dlm->spinlock);
208
209 list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
210}
211
212
213static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
214 struct dlm_master_list_entry *mle)
215{
216 if (!list_empty(&mle->hb_events))
217 list_del_init(&mle->hb_events);
218}
219
220
221static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
222 struct dlm_master_list_entry *mle)
223{
224 spin_lock(&dlm->spinlock);
225 __dlm_mle_detach_hb_events(dlm, mle);
226 spin_unlock(&dlm->spinlock);
227}
228
a2bf0477
KH		 229static void dlm_get_mle_inuse(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 mle->inuse++;
237 kref_get(&mle->mle_refs);
238}
239
240static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
241{
242 struct dlm_ctxt *dlm;
243 dlm = mle->dlm;
244
245 spin_lock(&dlm->spinlock);
246 spin_lock(&dlm->master_lock);
247 mle->inuse--;
248 __dlm_put_mle(mle);
249 spin_unlock(&dlm->master_lock);
250 spin_unlock(&dlm->spinlock);
251
252}
253
6714d8e8
KH		 254/* remove from list and free */
255static void __dlm_put_mle(struct dlm_master_list_entry *mle)
256{
257 struct dlm_ctxt *dlm;
258 dlm = mle->dlm;
259
260 assert_spin_locked(&dlm->spinlock);
261 assert_spin_locked(&dlm->master_lock);
aa852354
KH		 262 if (!atomic_read(&mle->mle_refs.refcount)) {
263 /* this may or may not crash, but who cares.
264 * it's a BUG. */
265 mlog(ML_ERROR, "bad mle: %p\n", mle);
266 dlm_print_one_mle(mle);
267 BUG();
268 } else
269 kref_put(&mle->mle_refs, dlm_mle_release);
6714d8e8
KH		 270}
271
272
273/* must not have any spinlocks coming in */
274static void dlm_put_mle(struct dlm_master_list_entry *mle)
275{
276 struct dlm_ctxt *dlm;
277 dlm = mle->dlm;
278
279 spin_lock(&dlm->spinlock);
280 spin_lock(&dlm->master_lock);
281 __dlm_put_mle(mle);
282 spin_unlock(&dlm->master_lock);
283 spin_unlock(&dlm->spinlock);
284}
285
286static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
287{
288 kref_get(&mle->mle_refs);
289}
290
291static void dlm_init_mle(struct dlm_master_list_entry *mle,
292 enum dlm_mle_type type,
293 struct dlm_ctxt *dlm,
294 struct dlm_lock_resource *res,
295 const char *name,
296 unsigned int namelen)
297{
298 assert_spin_locked(&dlm->spinlock);
299
300 mle->dlm = dlm;
301 mle->type = type;
2ed6c750 302 INIT_HLIST_NODE(&mle->master_hash_node);
6714d8e8
KH		 303 INIT_LIST_HEAD(&mle->hb_events);
304 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
305 spin_lock_init(&mle->spinlock);
306 init_waitqueue_head(&mle->wq);
307 atomic_set(&mle->woken, 0);
308 kref_init(&mle->mle_refs);
309 memset(mle->response_map, 0, sizeof(mle->response_map));
310 mle->master = O2NM_MAX_NODES;
311 mle->new_master = O2NM_MAX_NODES;
a2bf0477 312 mle->inuse = 0;
6714d8e8 313
f77a9a78
SM		 314 BUG_ON(mle->type != DLM_MLE_BLOCK &&
315 mle->type != DLM_MLE_MASTER &&
316 mle->type != DLM_MLE_MIGRATION);
317
6714d8e8
KH		 318 if (mle->type == DLM_MLE_MASTER) {
319 BUG_ON(!res);
f77a9a78
SM		 320 mle->u.mleres = res;
321 } else {
6714d8e8 322 BUG_ON(!name);
f77a9a78
SM		 323 memcpy(mle->u.mlename.name, name, namelen);
324 mle->u.mlename.len = namelen;
2ed6c750 325 mle->u.mlename.hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 326 }
327
328 /* copy off the node_map and register hb callbacks on our copy */
329 memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
330 memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
331 clear_bit(dlm->node_num, mle->vote_map);
332 clear_bit(dlm->node_num, mle->node_map);
333
334 /* attach the mle to the domain node up/down events */
335 __dlm_mle_attach_hb_events(dlm, mle);
336}
337
1c084577
SM		 338void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
339{
340 assert_spin_locked(&dlm->spinlock);
341 assert_spin_locked(&dlm->master_lock);
342
2ed6c750
SM		 343 if (!hlist_unhashed(&mle->master_hash_node))
344 hlist_del_init(&mle->master_hash_node);
1c084577
SM		 345}
346
347void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
348{
2ed6c750
SM		 349 struct hlist_head *bucket;
350 unsigned char *mname;
351 unsigned int mlen, hash;
352
1c084577
SM		 353 assert_spin_locked(&dlm->master_lock);
354
2ed6c750
SM		 355 __dlm_mle_name(mle, &mname, &mlen, &hash);
356 bucket = dlm_master_hash(dlm, hash);
357 hlist_add_head(&mle->master_hash_node, bucket);
1c084577 358}
6714d8e8
KH		 359
360/* returns 1 if found, 0 if not */
361static int dlm_find_mle(struct dlm_ctxt *dlm,
362 struct dlm_master_list_entry **mle,
363 char *name, unsigned int namelen)
364{
365 struct dlm_master_list_entry *tmpmle;
2ed6c750
SM		 366 struct hlist_head *bucket;
367 struct hlist_node *list;
368 unsigned int hash;
6714d8e8
KH		 369
370 assert_spin_locked(&dlm->master_lock);
371
2ed6c750
SM		 372 hash = dlm_lockid_hash(name, namelen);
373 bucket = dlm_master_hash(dlm, hash);
374 hlist_for_each(list, bucket) {
375 tmpmle = hlist_entry(list, struct dlm_master_list_entry,
376 master_hash_node);
6714d8e8
KH		 377 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
378 continue;
379 dlm_get_mle(tmpmle);
380 *mle = tmpmle;
381 return 1;
382 }
383 return 0;
384}
385
386void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
387{
388 struct dlm_master_list_entry *mle;
6714d8e8
KH		 389
390 assert_spin_locked(&dlm->spinlock);
391
800deef3 392 list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
6714d8e8
KH		 393 if (node_up)
394 dlm_mle_node_up(dlm, mle, NULL, idx);
395 else
396 dlm_mle_node_down(dlm, mle, NULL, idx);
397 }
398}
399
400static void dlm_mle_node_down(struct dlm_ctxt *dlm,
401 struct dlm_master_list_entry *mle,
402 struct o2nm_node *node, int idx)
403{
404 spin_lock(&mle->spinlock);
405
406 if (!test_bit(idx, mle->node_map))
407 mlog(0, "node %u already removed from nodemap!\n", idx);
408 else
409 clear_bit(idx, mle->node_map);
410
411 spin_unlock(&mle->spinlock);
412}
413
414static void dlm_mle_node_up(struct dlm_ctxt *dlm,
415 struct dlm_master_list_entry *mle,
416 struct o2nm_node *node, int idx)
417{
418 spin_lock(&mle->spinlock);
419
420 if (test_bit(idx, mle->node_map))
421 mlog(0, "node %u already in node map!\n", idx);
422 else
423 set_bit(idx, mle->node_map);
424
425 spin_unlock(&mle->spinlock);
426}
427
428
429int dlm_init_mle_cache(void)
430{
12eb0035 431 dlm_mle_cache = kmem_cache_create("o2dlm_mle",
6714d8e8
KH		 432 sizeof(struct dlm_master_list_entry),
433 0, SLAB_HWCACHE_ALIGN,
20c2df83 434 NULL);
6714d8e8
KH		 435 if (dlm_mle_cache == NULL)
436 return -ENOMEM;
437 return 0;
438}
439
440void dlm_destroy_mle_cache(void)
441{
442 if (dlm_mle_cache)
443 kmem_cache_destroy(dlm_mle_cache);
444}
445
446static void dlm_mle_release(struct kref *kref)
447{
448 struct dlm_master_list_entry *mle;
449 struct dlm_ctxt *dlm;
2ed6c750
SM		 450 unsigned char *mname;
451 unsigned int mlen;
6714d8e8
KH		 452
453 mlog_entry_void();
454
455 mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
456 dlm = mle->dlm;
457
6714d8e8
KH		 458 assert_spin_locked(&dlm->spinlock);
459 assert_spin_locked(&dlm->master_lock);
460
2ed6c750
SM		 461 __dlm_mle_name(mle, &mname, &mlen, NULL);
462 mlog(0, "Releasing mle for %.*s, type %d\n", mlen, mname, mle->type);
463
6714d8e8 464 /* remove from list if not already */
1c084577 465 __dlm_unlink_mle(dlm, mle);
6714d8e8
KH		 466
467 /* detach the mle from the domain node up/down events */
468 __dlm_mle_detach_hb_events(dlm, mle);
469
470 /* NOTE: kfree under spinlock here.
471 * if this is bad, we can move this to a freelist. */
472 kmem_cache_free(dlm_mle_cache, mle);
473}
474
475
476/*
477 * LOCK RESOURCE FUNCTIONS
478 */
479
724bdca9
SM		 480int dlm_init_master_caches(void)
481{
482 dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
483 sizeof(struct dlm_lock_resource),
484 0, SLAB_HWCACHE_ALIGN, NULL);
485 if (!dlm_lockres_cache)
486 goto bail;
487
488 dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
489 DLM_LOCKID_NAME_MAX, 0,
490 SLAB_HWCACHE_ALIGN, NULL);
491 if (!dlm_lockname_cache)
492 goto bail;
493
494 return 0;
495bail:
496 dlm_destroy_master_caches();
497 return -ENOMEM;
498}
499
500void dlm_destroy_master_caches(void)
501{
502 if (dlm_lockname_cache)
503 kmem_cache_destroy(dlm_lockname_cache);
504
505 if (dlm_lockres_cache)
506 kmem_cache_destroy(dlm_lockres_cache);
507}
508
6714d8e8
KH		 509static void dlm_set_lockres_owner(struct dlm_ctxt *dlm,
510 struct dlm_lock_resource *res,
511 u8 owner)
512{
513 assert_spin_locked(&res->spinlock);
514
515 mlog_entry("%.*s, %u\n", res->lockname.len, res->lockname.name, owner);
516
517 if (owner == dlm->node_num)
518 atomic_inc(&dlm->local_resources);
519 else if (owner == DLM_LOCK_RES_OWNER_UNKNOWN)
520 atomic_inc(&dlm->unknown_resources);
521 else
522 atomic_inc(&dlm->remote_resources);
523
524 res->owner = owner;
525}
526
527void dlm_change_lockres_owner(struct dlm_ctxt *dlm,
528 struct dlm_lock_resource *res, u8 owner)
529{
530 assert_spin_locked(&res->spinlock);
531
532 if (owner == res->owner)
533 return;
534
535 if (res->owner == dlm->node_num)
536 atomic_dec(&dlm->local_resources);
537 else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN)
538 atomic_dec(&dlm->unknown_resources);
539 else
540 atomic_dec(&dlm->remote_resources);
541
542 dlm_set_lockres_owner(dlm, res, owner);
543}
544
545
546static void dlm_lockres_release(struct kref *kref)
547{
548 struct dlm_lock_resource *res;
b0d4f817 549 struct dlm_ctxt *dlm;
6714d8e8
KH		 550
551 res = container_of(kref, struct dlm_lock_resource, refs);
b0d4f817 552 dlm = res->dlm;
6714d8e8
KH		 553
554 /* This should not happen -- all lockres' have a name
555 * associated with them at init time. */
556 BUG_ON(!res->lockname.name);
557
558 mlog(0, "destroying lockres %.*s\n", res->lockname.len,
559 res->lockname.name);
560
b0d4f817 561 spin_lock(&dlm->track_lock);
29576f8b
SM		 562 if (!list_empty(&res->tracking))
563 list_del_init(&res->tracking);
564 else {
565 mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n",
566 res->lockname.len, res->lockname.name);
567 dlm_print_one_lock_resource(res);
568 }
b0d4f817
SM		 569 spin_unlock(&dlm->track_lock);
570
571 dlm_put(dlm);
29576f8b 572
a7f90d83
KH		 573 if (!hlist_unhashed(&res->hash_node) ||
574 !list_empty(&res->granted) ||
575 !list_empty(&res->converting) ||
576 !list_empty(&res->blocked) ||
577 !list_empty(&res->dirty) ||
578 !list_empty(&res->recovering) ||
579 !list_empty(&res->purge)) {
580 mlog(ML_ERROR,
581 "Going to BUG for resource %.*s."
582 " We're on a list! [%c%c%c%c%c%c%c]\n",
583 res->lockname.len, res->lockname.name,
584 !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
585 !list_empty(&res->granted) ? 'G' : ' ',
586 !list_empty(&res->converting) ? 'C' : ' ',
587 !list_empty(&res->blocked) ? 'B' : ' ',
588 !list_empty(&res->dirty) ? 'D' : ' ',
589 !list_empty(&res->recovering) ? 'R' : ' ',
590 !list_empty(&res->purge) ? 'P' : ' ');
591
592 dlm_print_one_lock_resource(res);
593 }
594
6714d8e8
KH		 595 /* By the time we're ready to blow this guy away, we shouldn't
596 * be on any lists. */
81f2094a 597 BUG_ON(!hlist_unhashed(&res->hash_node));
6714d8e8
KH		 598 BUG_ON(!list_empty(&res->granted));
599 BUG_ON(!list_empty(&res->converting));
600 BUG_ON(!list_empty(&res->blocked));
601 BUG_ON(!list_empty(&res->dirty));
602 BUG_ON(!list_empty(&res->recovering));
603 BUG_ON(!list_empty(&res->purge));
604
724bdca9 605 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
6714d8e8 606
724bdca9 607 kmem_cache_free(dlm_lockres_cache, res);
6714d8e8
KH		 608}
609
6714d8e8
KH		 610void dlm_lockres_put(struct dlm_lock_resource *res)
611{
612 kref_put(&res->refs, dlm_lockres_release);
613}
614
615static void dlm_init_lockres(struct dlm_ctxt *dlm,
616 struct dlm_lock_resource *res,
617 const char *name, unsigned int namelen)
618{
619 char *qname;
620
621 /* If we memset here, we lose our reference to the kmalloc'd
622 * res->lockname.name, so be sure to init every field
623 * correctly! */
624
625 qname = (char *) res->lockname.name;
626 memcpy(qname, name, namelen);
627
628 res->lockname.len = namelen;
a3d33291 629 res->lockname.hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 630
631 init_waitqueue_head(&res->wq);
632 spin_lock_init(&res->spinlock);
81f2094a 633 INIT_HLIST_NODE(&res->hash_node);
6714d8e8
KH		 634 INIT_LIST_HEAD(&res->granted);
635 INIT_LIST_HEAD(&res->converting);
636 INIT_LIST_HEAD(&res->blocked);
637 INIT_LIST_HEAD(&res->dirty);
638 INIT_LIST_HEAD(&res->recovering);
639 INIT_LIST_HEAD(&res->purge);
29576f8b 640 INIT_LIST_HEAD(&res->tracking);
6714d8e8
KH		 641 atomic_set(&res->asts_reserved, 0);
642 res->migration_pending = 0;
ba2bf218 643 res->inflight_locks = 0;
6714d8e8 644
b0d4f817
SM		 645 /* put in dlm_lockres_release */
646 dlm_grab(dlm);
647 res->dlm = dlm;
648
6714d8e8
KH		 649 kref_init(&res->refs);
650
651 /* just for consistency */
652 spin_lock(&res->spinlock);
653 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
654 spin_unlock(&res->spinlock);
655
656 res->state = DLM_LOCK_RES_IN_PROGRESS;
657
658 res->last_used = 0;
659
18c6ac38 660 spin_lock(&dlm->spinlock);
29576f8b 661 list_add_tail(&res->tracking, &dlm->tracking_list);
18c6ac38 662 spin_unlock(&dlm->spinlock);
29576f8b 663
6714d8e8 664 memset(res->lvb, 0, DLM_LVB_LEN);
ba2bf218 665 memset(res->refmap, 0, sizeof(res->refmap));
6714d8e8
KH		 666}
667
668struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
669 const char *name,
670 unsigned int namelen)
671{
724bdca9 672 struct dlm_lock_resource *res = NULL;
6714d8e8 673
724bdca9
SM		 674 res = (struct dlm_lock_resource *)
675 kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
6714d8e8 676 if (!res)
724bdca9 677 goto error;
6714d8e8 678
724bdca9
SM		 679 res->lockname.name = (char *)
680 kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
681 if (!res->lockname.name)
682 goto error;
6714d8e8
KH		 683
684 dlm_init_lockres(dlm, res, name, namelen);
685 return res;
724bdca9
SM		 686
687error:
688 if (res && res->lockname.name)
689 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
690
691 if (res)
692 kmem_cache_free(dlm_lockres_cache, res);
693 return NULL;
6714d8e8
KH		 694}
695
ba2bf218
KH		 696void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
697 struct dlm_lock_resource *res,
698 int new_lockres,
699 const char *file,
700 int line)
701{
702 if (!new_lockres)
703 assert_spin_locked(&res->spinlock);
704
705 if (!test_bit(dlm->node_num, res->refmap)) {
706 BUG_ON(res->inflight_locks != 0);
707 dlm_lockres_set_refmap_bit(dlm->node_num, res);
708 }
709 res->inflight_locks++;
710 mlog(0, "%s:%.*s: inflight++: now %u\n",
711 dlm->name, res->lockname.len, res->lockname.name,
712 res->inflight_locks);
713}
714
715void __dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
716 struct dlm_lock_resource *res,
717 const char *file,
718 int line)
719{
720 assert_spin_locked(&res->spinlock);
721
722 BUG_ON(res->inflight_locks == 0);
723 res->inflight_locks--;
724 mlog(0, "%s:%.*s: inflight--: now %u\n",
725 dlm->name, res->lockname.len, res->lockname.name,
726 res->inflight_locks);
727 if (res->inflight_locks == 0)
728 dlm_lockres_clear_refmap_bit(dlm->node_num, res);
729 wake_up(&res->wq);
730}
731
6714d8e8
KH		 732/*
733 * lookup a lock resource by name.
734 * may already exist in the hashtable.
735 * lockid is null terminated
736 *
737 * if not, allocate enough for the lockres and for
738 * the temporary structure used in doing the mastering.
739 *
740 * also, do a lookup in the dlm->master_list to see
741 * if another node has begun mastering the same lock.
742 * if so, there should be a block entry in there
743 * for this name, and we should *not* attempt to master
744 * the lock here. need to wait around for that node
745 * to assert_master (or die).
746 *
747 */
748struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
749 const char *lockid,
3384f3df 750 int namelen,
6714d8e8
KH		 751 int flags)
752{
753 struct dlm_lock_resource *tmpres=NULL, *res=NULL;
754 struct dlm_master_list_entry *mle = NULL;
755 struct dlm_master_list_entry *alloc_mle = NULL;
756 int blocked = 0;
757 int ret, nodenum;
758 struct dlm_node_iter iter;
3384f3df 759 unsigned int hash;
6714d8e8 760 int tries = 0;
c03872f5 761 int bit, wait_on_recovery = 0;
ba2bf218 762 int drop_inflight_if_nonlocal = 0;
6714d8e8
KH		 763
764 BUG_ON(!lockid);
765
a3d33291 766 hash = dlm_lockid_hash(lockid, namelen);
6714d8e8
KH		 767
768 mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
769
770lookup:
771 spin_lock(&dlm->spinlock);
ba2bf218 772 tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
6714d8e8 773 if (tmpres) {
ba2bf218
KH		 774 int dropping_ref = 0;
775
7b791d68
SM		 776 spin_unlock(&dlm->spinlock);
777
ba2bf218 778 spin_lock(&tmpres->spinlock);
7b791d68
SM		 779 /* We wait for the other thread that is mastering the resource */
780 if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
781 __dlm_wait_on_lockres(tmpres);
782 BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
783 }
784
ba2bf218
KH		 785 if (tmpres->owner == dlm->node_num) {
786 BUG_ON(tmpres->state & DLM_LOCK_RES_DROPPING_REF);
787 dlm_lockres_grab_inflight_ref(dlm, tmpres);
788 } else if (tmpres->state & DLM_LOCK_RES_DROPPING_REF)
789 dropping_ref = 1;
790 spin_unlock(&tmpres->spinlock);
ba2bf218
KH		 791
792 /* wait until done messaging the master, drop our ref to allow
793 * the lockres to be purged, start over. */
794 if (dropping_ref) {
795 spin_lock(&tmpres->spinlock);
796 __dlm_wait_on_lockres_flags(tmpres, DLM_LOCK_RES_DROPPING_REF);
797 spin_unlock(&tmpres->spinlock);
798 dlm_lockres_put(tmpres);
799 tmpres = NULL;
800 goto lookup;
801 }
802
6714d8e8
KH		 803 mlog(0, "found in hash!\n");
804 if (res)
805 dlm_lockres_put(res);
806 res = tmpres;
807 goto leave;
808 }
809
810 if (!res) {
811 spin_unlock(&dlm->spinlock);
812 mlog(0, "allocating a new resource\n");
813 /* nothing found and we need to allocate one. */
814 alloc_mle = (struct dlm_master_list_entry *)
ad8100e0 815 kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
6714d8e8
KH		 816 if (!alloc_mle)
817 goto leave;
818 res = dlm_new_lockres(dlm, lockid, namelen);
819 if (!res)
820 goto leave;
821 goto lookup;
822 }
823
824 mlog(0, "no lockres found, allocated our own: %p\n", res);
825
826 if (flags & LKM_LOCAL) {
827 /* caller knows it's safe to assume it's not mastered elsewhere
828 * DONE! return right away */
829 spin_lock(&res->spinlock);
830 dlm_change_lockres_owner(dlm, res, dlm->node_num);
831 __dlm_insert_lockres(dlm, res);
ba2bf218 832 dlm_lockres_grab_inflight_ref(dlm, res);
6714d8e8
KH		 833 spin_unlock(&res->spinlock);
834 spin_unlock(&dlm->spinlock);
835 /* lockres still marked IN_PROGRESS */
836 goto wake_waiters;
837 }
838
839 /* check master list to see if another node has started mastering it */
840 spin_lock(&dlm->master_lock);
841
842 /* if we found a block, wait for lock to be mastered by another node */
843 blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
844 if (blocked) {
ba2bf218 845 int mig;
6714d8e8
KH		 846 if (mle->type == DLM_MLE_MASTER) {
847 mlog(ML_ERROR, "master entry for nonexistent lock!\n");
848 BUG();
ba2bf218
KH		 849 }
850 mig = (mle->type == DLM_MLE_MIGRATION);
851 /* if there is a migration in progress, let the migration
852 * finish before continuing. we can wait for the absence
853 * of the MIGRATION mle: either the migrate finished or
854 * one of the nodes died and the mle was cleaned up.
855 * if there is a BLOCK here, but it already has a master
856 * set, we are too late. the master does not have a ref
857 * for us in the refmap. detach the mle and drop it.
858 * either way, go back to the top and start over. */
859 if (mig || mle->master != O2NM_MAX_NODES) {
860 BUG_ON(mig && mle->master == dlm->node_num);
861 /* we arrived too late. the master does not
862 * have a ref for us. retry. */
863 mlog(0, "%s:%.*s: late on %s\n",
864 dlm->name, namelen, lockid,
865 mig ? "MIGRATION" : "BLOCK");
6714d8e8 866 spin_unlock(&dlm->master_lock);
6714d8e8
KH		 867 spin_unlock(&dlm->spinlock);
868
869 /* master is known, detach */
ba2bf218
KH		 870 if (!mig)
871 dlm_mle_detach_hb_events(dlm, mle);
6714d8e8
KH		 872 dlm_put_mle(mle);
873 mle = NULL;
ba2bf218
KH		 874 /* this is lame, but we cant wait on either
875 * the mle or lockres waitqueue here */
876 if (mig)
877 msleep(100);
878 goto lookup;
6714d8e8
KH		 879 }
880 } else {
881 /* go ahead and try to master lock on this node */
882 mle = alloc_mle;
883 /* make sure this does not get freed below */
884 alloc_mle = NULL;
885 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
886 set_bit(dlm->node_num, mle->maybe_map);
1c084577 887 __dlm_insert_mle(dlm, mle);
c03872f5
KH		 888
889 /* still holding the dlm spinlock, check the recovery map
890 * to see if there are any nodes that still need to be
891 * considered. these will not appear in the mle nodemap
892 * but they might own this lockres. wait on them. */
893 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
894 if (bit < O2NM_MAX_NODES) {
2759236f 895 mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to "
c03872f5
KH		 896 "recover before lock mastery can begin\n",
897 dlm->name, namelen, (char *)lockid, bit);
898 wait_on_recovery = 1;
899 }
6714d8e8
KH		 900 }
901
902 /* at this point there is either a DLM_MLE_BLOCK or a
903 * DLM_MLE_MASTER on the master list, so it's safe to add the
904 * lockres to the hashtable. anyone who finds the lock will
905 * still have to wait on the IN_PROGRESS. */
906
907 /* finally add the lockres to its hash bucket */
908 __dlm_insert_lockres(dlm, res);
ba2bf218
KH		 909 /* since this lockres is new it doesnt not require the spinlock */
910 dlm_lockres_grab_inflight_ref_new(dlm, res);
911
912 /* if this node does not become the master make sure to drop
913 * this inflight reference below */
914 drop_inflight_if_nonlocal = 1;
915
6714d8e8
KH		 916 /* get an extra ref on the mle in case this is a BLOCK
917 * if so, the creator of the BLOCK may try to put the last
918 * ref at this time in the assert master handler, so we
919 * need an extra one to keep from a bad ptr deref. */
a2bf0477 920 dlm_get_mle_inuse(mle);
6714d8e8
KH		 921 spin_unlock(&dlm->master_lock);
922 spin_unlock(&dlm->spinlock);
923
e7e69eb3 924redo_request:
c03872f5
KH		 925 while (wait_on_recovery) {
926 /* any cluster changes that occurred after dropping the
927 * dlm spinlock would be detectable be a change on the mle,
928 * so we only need to clear out the recovery map once. */
929 if (dlm_is_recovery_lock(lockid, namelen)) {
930 mlog(ML_NOTICE, "%s: recovery map is not empty, but "
931 "must master $RECOVERY lock now\n", dlm->name);
932 if (!dlm_pre_master_reco_lockres(dlm, res))
933 wait_on_recovery = 0;
934 else {
935 mlog(0, "%s: waiting 500ms for heartbeat state "
936 "change\n", dlm->name);
937 msleep(500);
938 }
939 continue;
940 }
941
942 dlm_kick_recovery_thread(dlm);
aa087b84 943 msleep(1000);
c03872f5
KH		 944 dlm_wait_for_recovery(dlm);
945
946 spin_lock(&dlm->spinlock);
947 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
948 if (bit < O2NM_MAX_NODES) {
2759236f 949 mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to "
c03872f5
KH		 950 "recover before lock mastery can begin\n",
951 dlm->name, namelen, (char *)lockid, bit);
952 wait_on_recovery = 1;
953 } else
954 wait_on_recovery = 0;
955 spin_unlock(&dlm->spinlock);
b7084ab5
KH		 956
957 if (wait_on_recovery)
958 dlm_wait_for_node_recovery(dlm, bit, 10000);
c03872f5
KH		 959 }
960
6714d8e8
KH		 961 /* must wait for lock to be mastered elsewhere */
962 if (blocked)
963 goto wait;
964
6714d8e8
KH		 965 ret = -EINVAL;
966 dlm_node_iter_init(mle->vote_map, &iter);
967 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
ba2bf218 968 ret = dlm_do_master_request(res, mle, nodenum);
6714d8e8
KH		 969 if (ret < 0)
970 mlog_errno(ret);
971 if (mle->master != O2NM_MAX_NODES) {
972 /* found a master ! */
9c6510a5
KH		 973 if (mle->master <= nodenum)
974 break;
975 /* if our master request has not reached the master
976 * yet, keep going until it does. this is how the
977 * master will know that asserts are needed back to
978 * the lower nodes. */
979 mlog(0, "%s:%.*s: requests only up to %u but master "
980 "is %u, keep going\n", dlm->name, namelen,
981 lockid, nodenum, mle->master);
6714d8e8
KH		 982 }
983 }
984
985wait:
986 /* keep going until the response map includes all nodes */
987 ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
988 if (ret < 0) {
e7e69eb3 989 wait_on_recovery = 1;
6714d8e8
KH		 990 mlog(0, "%s:%.*s: node map changed, redo the "
991 "master request now, blocked=%d\n",
992 dlm->name, res->lockname.len,
993 res->lockname.name, blocked);
994 if (++tries > 20) {
995 mlog(ML_ERROR, "%s:%.*s: spinning on "
996 "dlm_wait_for_lock_mastery, blocked=%d\n",
997 dlm->name, res->lockname.len,
998 res->lockname.name, blocked);
999 dlm_print_one_lock_resource(res);
8a9343fa 1000 dlm_print_one_mle(mle);
6714d8e8
KH		 1001 tries = 0;
1002 }
1003 goto redo_request;
1004 }
1005
1006 mlog(0, "lockres mastered by %u\n", res->owner);
1007 /* make sure we never continue without this */
1008 BUG_ON(res->owner == O2NM_MAX_NODES);
1009
1010 /* master is known, detach if not already detached */
1011 dlm_mle_detach_hb_events(dlm, mle);
1012 dlm_put_mle(mle);
1013 /* put the extra ref */
a2bf0477 1014 dlm_put_mle_inuse(mle);
6714d8e8
KH		 1015
1016wake_waiters:
1017 spin_lock(&res->spinlock);
ba2bf218
KH		 1018 if (res->owner != dlm->node_num && drop_inflight_if_nonlocal)
1019 dlm_lockres_drop_inflight_ref(dlm, res);
6714d8e8
KH		 1020 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1021 spin_unlock(&res->spinlock);
1022 wake_up(&res->wq);
1023
1024leave:
1025 /* need to free the unused mle */
1026 if (alloc_mle)
1027 kmem_cache_free(dlm_mle_cache, alloc_mle);
1028
1029 return res;
1030}
1031
1032
1033#define DLM_MASTERY_TIMEOUT_MS 5000
1034
1035static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
1036 struct dlm_lock_resource *res,
1037 struct dlm_master_list_entry *mle,
1038 int *blocked)
1039{
1040 u8 m;
1041 int ret, bit;
1042 int map_changed, voting_done;
1043 int assert, sleep;
1044
1045recheck:
1046 ret = 0;
1047 assert = 0;
1048
1049 /* check if another node has already become the owner */
1050 spin_lock(&res->spinlock);
1051 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
9c6510a5
KH		 1052 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1053 res->lockname.len, res->lockname.name, res->owner);
6714d8e8 1054 spin_unlock(&res->spinlock);
9c6510a5
KH		 1055 /* this will cause the master to re-assert across
1056 * the whole cluster, freeing up mles */
588e0090 1057 if (res->owner != dlm->node_num) {
ba2bf218 1058 ret = dlm_do_master_request(res, mle, res->owner);
588e0090
KH		 1059 if (ret < 0) {
1060 /* give recovery a chance to run */
1061 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1062 msleep(500);
1063 goto recheck;
1064 }
9c6510a5
KH		 1065 }
1066 ret = 0;
6714d8e8
KH		 1067 goto leave;
1068 }
1069 spin_unlock(&res->spinlock);
1070
1071 spin_lock(&mle->spinlock);
1072 m = mle->master;
1073 map_changed = (memcmp(mle->vote_map, mle->node_map,
1074 sizeof(mle->vote_map)) != 0);
1075 voting_done = (memcmp(mle->vote_map, mle->response_map,
1076 sizeof(mle->vote_map)) == 0);
1077
1078 /* restart if we hit any errors */
1079 if (map_changed) {
1080 int b;
1081 mlog(0, "%s: %.*s: node map changed, restarting\n",
1082 dlm->name, res->lockname.len, res->lockname.name);
1083 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1084 b = (mle->type == DLM_MLE_BLOCK);
1085 if ((*blocked && !b) || (!*blocked && b)) {
1086 mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1087 dlm->name, res->lockname.len, res->lockname.name,
1088 *blocked, b);
1089 *blocked = b;
1090 }
1091 spin_unlock(&mle->spinlock);
1092 if (ret < 0) {
1093 mlog_errno(ret);
1094 goto leave;
1095 }
1096 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1097 "rechecking now\n", dlm->name, res->lockname.len,
1098 res->lockname.name);
1099 goto recheck;
aa852354
KH		 1100 } else {
1101 if (!voting_done) {
1102 mlog(0, "map not changed and voting not done "
1103 "for %s:%.*s\n", dlm->name, res->lockname.len,
1104 res->lockname.name);
1105 }
6714d8e8
KH		 1106 }
1107
1108 if (m != O2NM_MAX_NODES) {
1109 /* another node has done an assert!
1110 * all done! */
1111 sleep = 0;
1112 } else {
1113 sleep = 1;
1114 /* have all nodes responded? */
1115 if (voting_done && !*blocked) {
1116 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1117 if (dlm->node_num <= bit) {
1118 /* my node number is lowest.
1119 * now tell other nodes that I am
1120 * mastering this. */
1121 mle->master = dlm->node_num;
ba2bf218
KH		 1122 /* ref was grabbed in get_lock_resource
1123 * will be dropped in dlmlock_master */
6714d8e8
KH		 1124 assert = 1;
1125 sleep = 0;
1126 }
1127 /* if voting is done, but we have not received
1128 * an assert master yet, we must sleep */
1129 }
1130 }
1131
1132 spin_unlock(&mle->spinlock);
1133
1134 /* sleep if we haven't finished voting yet */
1135 if (sleep) {
1136 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1137
1138 /*
1139 if (atomic_read(&mle->mle_refs.refcount) < 2)
1140 mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1141 atomic_read(&mle->mle_refs.refcount),
1142 res->lockname.len, res->lockname.name);
1143 */
1144 atomic_set(&mle->woken, 0);
1145 (void)wait_event_timeout(mle->wq,
1146 (atomic_read(&mle->woken) == 1),
1147 timeo);
1148 if (res->owner == O2NM_MAX_NODES) {
ba2bf218
KH		 1149 mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1150 res->lockname.len, res->lockname.name);
6714d8e8
KH		 1151 goto recheck;
1152 }
1153 mlog(0, "done waiting, master is %u\n", res->owner);
1154 ret = 0;
1155 goto leave;
1156 }
1157
1158 ret = 0; /* done */
1159 if (assert) {
1160 m = dlm->node_num;
1161 mlog(0, "about to master %.*s here, this=%u\n",
1162 res->lockname.len, res->lockname.name, m);
ba2bf218 1163 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
6714d8e8
KH		 1164 if (ret) {
1165 /* This is a failure in the network path,
1166 * not in the response to the assert_master
1167 * (any nonzero response is a BUG on this node).
1168 * Most likely a socket just got disconnected
1169 * due to node death. */
1170 mlog_errno(ret);
1171 }
1172 /* no longer need to restart lock mastery.
1173 * all living nodes have been contacted. */
1174 ret = 0;
1175 }
1176
1177 /* set the lockres owner */
1178 spin_lock(&res->spinlock);
ba2bf218
KH		 1179 /* mastery reference obtained either during
1180 * assert_master_handler or in get_lock_resource */
6714d8e8
KH		 1181 dlm_change_lockres_owner(dlm, res, m);
1182 spin_unlock(&res->spinlock);
1183
1184leave:
1185 return ret;
1186}
1187
1188struct dlm_bitmap_diff_iter
1189{
1190 int curnode;
1191 unsigned long *orig_bm;
1192 unsigned long *cur_bm;
1193 unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1194};
1195
1196enum dlm_node_state_change
1197{
1198 NODE_DOWN = -1,
1199 NODE_NO_CHANGE = 0,
1200 NODE_UP
1201};
1202
1203static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1204 unsigned long *orig_bm,
1205 unsigned long *cur_bm)
1206{
1207 unsigned long p1, p2;
1208 int i;
1209
1210 iter->curnode = -1;
1211 iter->orig_bm = orig_bm;
1212 iter->cur_bm = cur_bm;
1213
1214 for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1215 p1 = *(iter->orig_bm + i);
1216 p2 = *(iter->cur_bm + i);
1217 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1218 }
1219}
1220
1221static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1222 enum dlm_node_state_change *state)
1223{
1224 int bit;
1225
1226 if (iter->curnode >= O2NM_MAX_NODES)
1227 return -ENOENT;
1228
1229 bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1230 iter->curnode+1);
1231 if (bit >= O2NM_MAX_NODES) {
1232 iter->curnode = O2NM_MAX_NODES;
1233 return -ENOENT;
1234 }
1235
1236 /* if it was there in the original then this node died */
1237 if (test_bit(bit, iter->orig_bm))
1238 *state = NODE_DOWN;
1239 else
1240 *state = NODE_UP;
1241
1242 iter->curnode = bit;
1243 return bit;
1244}
1245
1246
1247static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1248 struct dlm_lock_resource *res,
1249 struct dlm_master_list_entry *mle,
1250 int blocked)
1251{
1252 struct dlm_bitmap_diff_iter bdi;
1253 enum dlm_node_state_change sc;
1254 int node;
1255 int ret = 0;
1256
1257 mlog(0, "something happened such that the "
1258 "master process may need to be restarted!\n");
1259
1260 assert_spin_locked(&mle->spinlock);
1261
1262 dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1263 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1264 while (node >= 0) {
1265 if (sc == NODE_UP) {
e2faea4c
KH		 1266 /* a node came up. clear any old vote from
1267 * the response map and set it in the vote map
1268 * then restart the mastery. */
1269 mlog(ML_NOTICE, "node %d up while restarting\n", node);
6714d8e8
KH		 1270
1271 /* redo the master request, but only for the new node */
1272 mlog(0, "sending request to new node\n");
1273 clear_bit(node, mle->response_map);
1274 set_bit(node, mle->vote_map);
1275 } else {
1276 mlog(ML_ERROR, "node down! %d\n", node);
6714d8e8
KH		 1277 if (blocked) {
1278 int lowest = find_next_bit(mle->maybe_map,
1279 O2NM_MAX_NODES, 0);
1280
1281 /* act like it was never there */
1282 clear_bit(node, mle->maybe_map);
1283
e7e69eb3
KH		 1284 if (node == lowest) {
1285 mlog(0, "expected master %u died"
1286 " while this node was blocked "
1287 "waiting on it!\n", node);
1288 lowest = find_next_bit(mle->maybe_map,
1289 O2NM_MAX_NODES,
1290 lowest+1);
1291 if (lowest < O2NM_MAX_NODES) {
1292 mlog(0, "%s:%.*s:still "
1293 "blocked. waiting on %u "
1294 "now\n", dlm->name,
1295 res->lockname.len,
1296 res->lockname.name,
1297 lowest);
1298 } else {
1299 /* mle is an MLE_BLOCK, but
1300 * there is now nothing left to
1301 * block on. we need to return
1302 * all the way back out and try
1303 * again with an MLE_MASTER.
1304 * dlm_do_local_recovery_cleanup
1305 * has already run, so the mle
1306 * refcount is ok */
1307 mlog(0, "%s:%.*s: no "
1308 "longer blocking. try to "
1309 "master this here\n",
1310 dlm->name,
1311 res->lockname.len,
1312 res->lockname.name);
1313 mle->type = DLM_MLE_MASTER;
f77a9a78 1314 mle->u.mleres = res;
e7e69eb3 1315 }
6714d8e8 1316 }
6714d8e8
KH		 1317 }
1318
e7e69eb3
KH		 1319 /* now blank out everything, as if we had never
1320 * contacted anyone */
1321 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1322 memset(mle->response_map, 0, sizeof(mle->response_map));
1323 /* reset the vote_map to the current node_map */
1324 memcpy(mle->vote_map, mle->node_map,
1325 sizeof(mle->node_map));
1326 /* put myself into the maybe map */
1327 if (mle->type != DLM_MLE_BLOCK)
1328 set_bit(dlm->node_num, mle->maybe_map);
6714d8e8
KH		 1329 }
1330 ret = -EAGAIN;
6714d8e8
KH		 1331 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1332 }
1333 return ret;
1334}
1335
1336
1337/*
1338 * DLM_MASTER_REQUEST_MSG
1339 *
1340 * returns: 0 on success,
1341 * -errno on a network error
1342 *
1343 * on error, the caller should assume the target node is "dead"
1344 *
1345 */
1346
ba2bf218
KH		 1347static int dlm_do_master_request(struct dlm_lock_resource *res,
1348 struct dlm_master_list_entry *mle, int to)
6714d8e8
KH		 1349{
1350 struct dlm_ctxt *dlm = mle->dlm;
1351 struct dlm_master_request request;
1352 int ret, response=0, resend;
f77a9a78
SM		 1353 unsigned char *mlename;
1354 unsigned int mlenamelen;
6714d8e8
KH		 1355
1356 memset(&request, 0, sizeof(request));
1357 request.node_idx = dlm->node_num;
1358
1359 BUG_ON(mle->type == DLM_MLE_MIGRATION);
1360
2ed6c750 1361 __dlm_mle_name(mle, &mlename, &mlenamelen, NULL);
f77a9a78
SM		 1362
1363 request.namelen = (u8)mlenamelen;
1364 memcpy(request.name, mlename, request.namelen);
6714d8e8
KH		 1365
1366again:
1367 ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1368 sizeof(request), to, &response);
1369 if (ret < 0) {
1370 if (ret == -ESRCH) {
1371 /* should never happen */
1372 mlog(ML_ERROR, "TCP stack not ready!\n");
1373 BUG();
1374 } else if (ret == -EINVAL) {
1375 mlog(ML_ERROR, "bad args passed to o2net!\n");
1376 BUG();
1377 } else if (ret == -ENOMEM) {
1378 mlog(ML_ERROR, "out of memory while trying to send "
1379 "network message! retrying\n");
1380 /* this is totally crude */
1381 msleep(50);
1382 goto again;
1383 } else if (!dlm_is_host_down(ret)) {
1384 /* not a network error. bad. */
1385 mlog_errno(ret);
1386 mlog(ML_ERROR, "unhandled error!");
1387 BUG();
1388 }
1389 /* all other errors should be network errors,
1390 * and likely indicate node death */
1391 mlog(ML_ERROR, "link to %d went down!\n", to);
1392 goto out;
1393 }
1394
1395 ret = 0;
1396 resend = 0;
1397 spin_lock(&mle->spinlock);
1398 switch (response) {
1399 case DLM_MASTER_RESP_YES:
1400 set_bit(to, mle->response_map);
1401 mlog(0, "node %u is the master, response=YES\n", to);
ba2bf218
KH		 1402 mlog(0, "%s:%.*s: master node %u now knows I have a "
1403 "reference\n", dlm->name, res->lockname.len,
1404 res->lockname.name, to);
6714d8e8
KH		 1405 mle->master = to;
1406 break;
1407 case DLM_MASTER_RESP_NO:
1408 mlog(0, "node %u not master, response=NO\n", to);
1409 set_bit(to, mle->response_map);
1410 break;
1411 case DLM_MASTER_RESP_MAYBE:
1412 mlog(0, "node %u not master, response=MAYBE\n", to);
1413 set_bit(to, mle->response_map);
1414 set_bit(to, mle->maybe_map);
1415 break;
1416 case DLM_MASTER_RESP_ERROR:
1417 mlog(0, "node %u hit an error, resending\n", to);
1418 resend = 1;
1419 response = 0;
1420 break;
1421 default:
1422 mlog(ML_ERROR, "bad response! %u\n", response);
1423 BUG();
1424 }
1425 spin_unlock(&mle->spinlock);
1426 if (resend) {
1427 /* this is also totally crude */
1428 msleep(50);
1429 goto again;
1430 }
1431
1432out:
1433 return ret;
1434}
1435
1436/*
1437 * locks that can be taken here:
1438 * dlm->spinlock
1439 * res->spinlock
1440 * mle->spinlock
1441 * dlm->master_list
1442 *
1443 * if possible, TRIM THIS DOWN!!!
1444 */
d74c9803
KH		 1445int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1446 void **ret_data)
6714d8e8
KH		 1447{
1448 u8 response = DLM_MASTER_RESP_MAYBE;
1449 struct dlm_ctxt *dlm = data;
9c6510a5 1450 struct dlm_lock_resource *res = NULL;
6714d8e8
KH		 1451 struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1452 struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1453 char *name;
a3d33291 1454 unsigned int namelen, hash;
6714d8e8
KH		 1455 int found, ret;
1456 int set_maybe;
9c6510a5 1457 int dispatch_assert = 0;
6714d8e8
KH		 1458
1459 if (!dlm_grab(dlm))
1460 return DLM_MASTER_RESP_NO;
1461
1462 if (!dlm_domain_fully_joined(dlm)) {
1463 response = DLM_MASTER_RESP_NO;
1464 goto send_response;
1465 }
1466
1467 name = request->name;
1468 namelen = request->namelen;
a3d33291 1469 hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 1470
1471 if (namelen > DLM_LOCKID_NAME_MAX) {
1472 response = DLM_IVBUFLEN;
1473 goto send_response;
1474 }
1475
1476way_up_top:
1477 spin_lock(&dlm->spinlock);
a3d33291 1478 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
6714d8e8
KH		 1479 if (res) {
1480 spin_unlock(&dlm->spinlock);
1481
1482 /* take care of the easy cases up front */
1483 spin_lock(&res->spinlock);
1cd04dbe
KH		 1484 if (res->state & (DLM_LOCK_RES_RECOVERING|
1485 DLM_LOCK_RES_MIGRATING)) {
6714d8e8
KH		 1486 spin_unlock(&res->spinlock);
1487 mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1cd04dbe 1488 "being recovered/migrated\n");
6714d8e8
KH		 1489 response = DLM_MASTER_RESP_ERROR;
1490 if (mle)
1491 kmem_cache_free(dlm_mle_cache, mle);
1492 goto send_response;
1493 }
1494
1495 if (res->owner == dlm->node_num) {
ba2bf218
KH		 1496 mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1497 dlm->name, namelen, name, request->node_idx);
1498 dlm_lockres_set_refmap_bit(request->node_idx, res);
6714d8e8 1499 spin_unlock(&res->spinlock);
6714d8e8
KH		 1500 response = DLM_MASTER_RESP_YES;
1501 if (mle)
1502 kmem_cache_free(dlm_mle_cache, mle);
1503
1504 /* this node is the owner.
1505 * there is some extra work that needs to
1506 * happen now. the requesting node has
1507 * caused all nodes up to this one to
1508 * create mles. this node now needs to
1509 * go back and clean those up. */
9c6510a5 1510 dispatch_assert = 1;
6714d8e8
KH		 1511 goto send_response;
1512 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1513 spin_unlock(&res->spinlock);
1514 // mlog(0, "node %u is the master\n", res->owner);
1515 response = DLM_MASTER_RESP_NO;
1516 if (mle)
1517 kmem_cache_free(dlm_mle_cache, mle);
1518 goto send_response;
1519 }
1520
1521 /* ok, there is no owner. either this node is
1522 * being blocked, or it is actively trying to
1523 * master this lock. */
1524 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1525 mlog(ML_ERROR, "lock with no owner should be "
1526 "in-progress!\n");
1527 BUG();
1528 }
1529
1530 // mlog(0, "lockres is in progress...\n");
1531 spin_lock(&dlm->master_lock);
1532 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1533 if (!found) {
1534 mlog(ML_ERROR, "no mle found for this lock!\n");
1535 BUG();
1536 }
1537 set_maybe = 1;
1538 spin_lock(&tmpmle->spinlock);
1539 if (tmpmle->type == DLM_MLE_BLOCK) {
1540 // mlog(0, "this node is waiting for "
1541 // "lockres to be mastered\n");
1542 response = DLM_MASTER_RESP_NO;
1543 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1544 mlog(0, "node %u is master, but trying to migrate to "
1545 "node %u.\n", tmpmle->master, tmpmle->new_master);
1546 if (tmpmle->master == dlm->node_num) {
6714d8e8
KH		 1547 mlog(ML_ERROR, "no owner on lockres, but this "
1548 "node is trying to migrate it to %u?!\n",
1549 tmpmle->new_master);
1550 BUG();
1551 } else {
1552 /* the real master can respond on its own */
1553 response = DLM_MASTER_RESP_NO;
1554 }
1555 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1556 set_maybe = 0;
9c6510a5 1557 if (tmpmle->master == dlm->node_num) {
6714d8e8 1558 response = DLM_MASTER_RESP_YES;
9c6510a5
KH		 1559 /* this node will be the owner.
1560 * go back and clean the mles on any
1561 * other nodes */
1562 dispatch_assert = 1;
ba2bf218
KH		 1563 dlm_lockres_set_refmap_bit(request->node_idx, res);
1564 mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1565 dlm->name, namelen, name,
1566 request->node_idx);
9c6510a5 1567 } else
6714d8e8
KH		 1568 response = DLM_MASTER_RESP_NO;
1569 } else {
1570 // mlog(0, "this node is attempting to "
1571 // "master lockres\n");
1572 response = DLM_MASTER_RESP_MAYBE;
1573 }
1574 if (set_maybe)
1575 set_bit(request->node_idx, tmpmle->maybe_map);
1576 spin_unlock(&tmpmle->spinlock);
1577
1578 spin_unlock(&dlm->master_lock);
1579 spin_unlock(&res->spinlock);
1580
1581 /* keep the mle attached to heartbeat events */
1582 dlm_put_mle(tmpmle);
1583 if (mle)
1584 kmem_cache_free(dlm_mle_cache, mle);
1585 goto send_response;
1586 }
1587
1588 /*
1589 * lockres doesn't exist on this node
1590 * if there is an MLE_BLOCK, return NO
1591 * if there is an MLE_MASTER, return MAYBE
1592 * otherwise, add an MLE_BLOCK, return NO
1593 */
1594 spin_lock(&dlm->master_lock);
1595 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1596 if (!found) {
1597 /* this lockid has never been seen on this node yet */
1598 // mlog(0, "no mle found\n");
1599 if (!mle) {
1600 spin_unlock(&dlm->master_lock);
1601 spin_unlock(&dlm->spinlock);
1602
1603 mle = (struct dlm_master_list_entry *)
ad8100e0 1604 kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
6714d8e8 1605 if (!mle) {
6714d8e8 1606 response = DLM_MASTER_RESP_ERROR;
9c6510a5 1607 mlog_errno(-ENOMEM);
6714d8e8
KH		 1608 goto send_response;
1609 }
6714d8e8
KH		 1610 goto way_up_top;
1611 }
1612
1613 // mlog(0, "this is second time thru, already allocated, "
1614 // "add the block.\n");
41b8c8a1 1615 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
6714d8e8 1616 set_bit(request->node_idx, mle->maybe_map);
1c084577 1617 __dlm_insert_mle(dlm, mle);
6714d8e8
KH		 1618 response = DLM_MASTER_RESP_NO;
1619 } else {
1620 // mlog(0, "mle was found\n");
1621 set_maybe = 1;
1622 spin_lock(&tmpmle->spinlock);
9c6510a5
KH		 1623 if (tmpmle->master == dlm->node_num) {
1624 mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1625 BUG();
1626 }
6714d8e8
KH		 1627 if (tmpmle->type == DLM_MLE_BLOCK)
1628 response = DLM_MASTER_RESP_NO;
1629 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1630 mlog(0, "migration mle was found (%u->%u)\n",
1631 tmpmle->master, tmpmle->new_master);
6714d8e8
KH		 1632 /* real master can respond on its own */
1633 response = DLM_MASTER_RESP_NO;
9c6510a5
KH		 1634 } else
1635 response = DLM_MASTER_RESP_MAYBE;
6714d8e8
KH		 1636 if (set_maybe)
1637 set_bit(request->node_idx, tmpmle->maybe_map);
1638 spin_unlock(&tmpmle->spinlock);
1639 }
1640 spin_unlock(&dlm->master_lock);
1641 spin_unlock(&dlm->spinlock);
1642
1643 if (found) {
1644 /* keep the mle attached to heartbeat events */
1645 dlm_put_mle(tmpmle);
1646 }
1647send_response:
b31cfc02
SM		 1648 /*
1649 * __dlm_lookup_lockres() grabbed a reference to this lockres.
1650 * The reference is released by dlm_assert_master_worker() under
1651 * the call to dlm_dispatch_assert_master(). If
1652 * dlm_assert_master_worker() isn't called, we drop it here.
1653 */
9c6510a5
KH		 1654 if (dispatch_assert) {
1655 if (response != DLM_MASTER_RESP_YES)
1656 mlog(ML_ERROR, "invalid response %d\n", response);
1657 if (!res) {
1658 mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1659 BUG();
1660 }
1661 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1662 dlm->node_num, res->lockname.len, res->lockname.name);
1663 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1664 DLM_ASSERT_MASTER_MLE_CLEANUP);
1665 if (ret < 0) {
1666 mlog(ML_ERROR, "failed to dispatch assert master work\n");
1667 response = DLM_MASTER_RESP_ERROR;
b31cfc02 1668 dlm_lockres_put(res);
9c6510a5 1669 }
b31cfc02
SM		 1670 } else {
1671 if (res)
1672 dlm_lockres_put(res);
9c6510a5
KH		 1673 }
1674
6714d8e8
KH		 1675 dlm_put(dlm);
1676 return response;
1677}
1678
1679/*
1680 * DLM_ASSERT_MASTER_MSG
1681 */
1682
1683
1684/*
1685 * NOTE: this can be used for debugging
1686 * can periodically run all locks owned by this node
1687 * and re-assert across the cluster...
1688 */
05488bbe
AB		 1689static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1690 struct dlm_lock_resource *res,
1691 void *nodemap, u32 flags)
6714d8e8
KH		 1692{
1693 struct dlm_assert_master assert;
1694 int to, tmpret;
1695 struct dlm_node_iter iter;
1696 int ret = 0;
9c6510a5 1697 int reassert;
ba2bf218
KH		 1698 const char *lockname = res->lockname.name;
1699 unsigned int namelen = res->lockname.len;
6714d8e8
KH		 1700
1701 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
f3f85464
SM		 1702
1703 spin_lock(&res->spinlock);
1704 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1705 spin_unlock(&res->spinlock);
1706
9c6510a5
KH		 1707again:
1708 reassert = 0;
6714d8e8
KH		 1709
1710 /* note that if this nodemap is empty, it returns 0 */
1711 dlm_node_iter_init(nodemap, &iter);
1712 while ((to = dlm_node_iter_next(&iter)) >= 0) {
1713 int r = 0;
a9ee4c8a
KH		 1714 struct dlm_master_list_entry *mle = NULL;
1715
6714d8e8
KH		 1716 mlog(0, "sending assert master to %d (%.*s)\n", to,
1717 namelen, lockname);
1718 memset(&assert, 0, sizeof(assert));
1719 assert.node_idx = dlm->node_num;
1720 assert.namelen = namelen;
1721 memcpy(assert.name, lockname, namelen);
1722 assert.flags = cpu_to_be32(flags);
1723
1724 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1725 &assert, sizeof(assert), to, &r);
1726 if (tmpret < 0) {
3b3b84a8 1727 mlog(0, "assert_master returned %d!\n", tmpret);
6714d8e8 1728 if (!dlm_is_host_down(tmpret)) {
3b3b84a8 1729 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
6714d8e8
KH		 1730 BUG();
1731 }
1732 /* a node died. finish out the rest of the nodes. */
3b3b84a8 1733 mlog(0, "link to %d went down!\n", to);
6714d8e8
KH		 1734 /* any nonzero status return will do */
1735 ret = tmpret;
ba2bf218 1736 r = 0;
6714d8e8
KH		 1737 } else if (r < 0) {
1738 /* ok, something horribly messed. kill thyself. */
1739 mlog(ML_ERROR,"during assert master of %.*s to %u, "
1740 "got %d.\n", namelen, lockname, to, r);
a9ee4c8a
KH		 1741 spin_lock(&dlm->spinlock);
1742 spin_lock(&dlm->master_lock);
1743 if (dlm_find_mle(dlm, &mle, (char *)lockname,
1744 namelen)) {
1745 dlm_print_one_mle(mle);
1746 __dlm_put_mle(mle);
1747 }
1748 spin_unlock(&dlm->master_lock);
1749 spin_unlock(&dlm->spinlock);
6714d8e8 1750 BUG();
ba2bf218
KH		 1751 }
1752
1753 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1754 !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1755 mlog(ML_ERROR, "%.*s: very strange, "
1756 "master MLE but no lockres on %u\n",
1757 namelen, lockname, to);
1758 }
1759
1760 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
9c6510a5
KH		 1761 mlog(0, "%.*s: node %u create mles on other "
1762 "nodes and requests a re-assert\n",
1763 namelen, lockname, to);
1764 reassert = 1;
6714d8e8 1765 }
ba2bf218
KH		 1766 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1767 mlog(0, "%.*s: node %u has a reference to this "
1768 "lockres, set the bit in the refmap\n",
1769 namelen, lockname, to);
1770 spin_lock(&res->spinlock);
1771 dlm_lockres_set_refmap_bit(to, res);
1772 spin_unlock(&res->spinlock);
1773 }
6714d8e8
KH		 1774 }
1775
9c6510a5
KH		 1776 if (reassert)
1777 goto again;
1778
f3f85464
SM		 1779 spin_lock(&res->spinlock);
1780 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1781 spin_unlock(&res->spinlock);
1782 wake_up(&res->wq);
1783
6714d8e8
KH		 1784 return ret;
1785}
1786
1787/*
1788 * locks that can be taken here:
1789 * dlm->spinlock
1790 * res->spinlock
1791 * mle->spinlock
1792 * dlm->master_list
1793 *
1794 * if possible, TRIM THIS DOWN!!!
1795 */
d74c9803
KH		 1796int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1797 void **ret_data)
6714d8e8
KH		 1798{
1799 struct dlm_ctxt *dlm = data;
1800 struct dlm_master_list_entry *mle = NULL;
1801 struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1802 struct dlm_lock_resource *res = NULL;
1803 char *name;
a3d33291 1804 unsigned int namelen, hash;
6714d8e8 1805 u32 flags;
ba2bf218 1806 int master_request = 0, have_lockres_ref = 0;
9c6510a5 1807 int ret = 0;
6714d8e8
KH		 1808
1809 if (!dlm_grab(dlm))
1810 return 0;
1811
1812 name = assert->name;
1813 namelen = assert->namelen;
a3d33291 1814 hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 1815 flags = be32_to_cpu(assert->flags);
1816
1817 if (namelen > DLM_LOCKID_NAME_MAX) {
1818 mlog(ML_ERROR, "Invalid name length!");
1819 goto done;
1820 }
1821
1822 spin_lock(&dlm->spinlock);
1823
1824 if (flags)
1825 mlog(0, "assert_master with flags: %u\n", flags);
1826
1827 /* find the MLE */
1828 spin_lock(&dlm->master_lock);
1829 if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1830 /* not an error, could be master just re-asserting */
1831 mlog(0, "just got an assert_master from %u, but no "
1832 "MLE for it! (%.*s)\n", assert->node_idx,
1833 namelen, name);
1834 } else {
1835 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1836 if (bit >= O2NM_MAX_NODES) {
1837 /* not necessarily an error, though less likely.
1838 * could be master just re-asserting. */
aa852354 1839 mlog(0, "no bits set in the maybe_map, but %u "
6714d8e8
KH		 1840 "is asserting! (%.*s)\n", assert->node_idx,
1841 namelen, name);
1842 } else if (bit != assert->node_idx) {
1843 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1844 mlog(0, "master %u was found, %u should "
1845 "back off\n", assert->node_idx, bit);
1846 } else {
1847 /* with the fix for bug 569, a higher node
1848 * number winning the mastery will respond
1849 * YES to mastery requests, but this node
1850 * had no way of knowing. let it pass. */
aa852354 1851 mlog(0, "%u is the lowest node, "
6714d8e8
KH		 1852 "%u is asserting. (%.*s) %u must "
1853 "have begun after %u won.\n", bit,
1854 assert->node_idx, namelen, name, bit,
1855 assert->node_idx);
1856 }
1857 }
2d1a868c
KH		 1858 if (mle->type == DLM_MLE_MIGRATION) {
1859 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1860 mlog(0, "%s:%.*s: got cleanup assert"
1861 " from %u for migration\n",
1862 dlm->name, namelen, name,
1863 assert->node_idx);
1864 } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1865 mlog(0, "%s:%.*s: got unrelated assert"
1866 " from %u for migration, ignoring\n",
1867 dlm->name, namelen, name,
1868 assert->node_idx);
1869 __dlm_put_mle(mle);
1870 spin_unlock(&dlm->master_lock);
1871 spin_unlock(&dlm->spinlock);
1872 goto done;
1873 }
1874 }
6714d8e8
KH		 1875 }
1876 spin_unlock(&dlm->master_lock);
1877
1878 /* ok everything checks out with the MLE
1879 * now check to see if there is a lockres */
a3d33291 1880 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
6714d8e8
KH		 1881 if (res) {
1882 spin_lock(&res->spinlock);
1883 if (res->state & DLM_LOCK_RES_RECOVERING) {
1884 mlog(ML_ERROR, "%u asserting but %.*s is "
1885 "RECOVERING!\n", assert->node_idx, namelen, name);
1886 goto kill;
1887 }
1888 if (!mle) {
dc2ed195
KH		 1889 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1890 res->owner != assert->node_idx) {
53ecd25e
SM		 1891 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1892 "but current owner is %u! (%.*s)\n",
1893 assert->node_idx, res->owner, namelen,
1894 name);
1895 __dlm_print_one_lock_resource(res);
1896 BUG();
6714d8e8
KH		 1897 }
1898 } else if (mle->type != DLM_MLE_MIGRATION) {
1899 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1900 /* owner is just re-asserting */
1901 if (res->owner == assert->node_idx) {
1902 mlog(0, "owner %u re-asserting on "
1903 "lock %.*s\n", assert->node_idx,
1904 namelen, name);
1905 goto ok;
1906 }
1907 mlog(ML_ERROR, "got assert_master from "
1908 "node %u, but %u is the owner! "
1909 "(%.*s)\n", assert->node_idx,
1910 res->owner, namelen, name);
1911 goto kill;
1912 }
1913 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1914 mlog(ML_ERROR, "got assert from %u, but lock "
1915 "with no owner should be "
1916 "in-progress! (%.*s)\n",
1917 assert->node_idx,
1918 namelen, name);
1919 goto kill;
1920 }
1921 } else /* mle->type == DLM_MLE_MIGRATION */ {
1922 /* should only be getting an assert from new master */
1923 if (assert->node_idx != mle->new_master) {
1924 mlog(ML_ERROR, "got assert from %u, but "
1925 "new master is %u, and old master "
1926 "was %u (%.*s)\n",
1927 assert->node_idx, mle->new_master,
1928 mle->master, namelen, name);
1929 goto kill;
1930 }
1931
1932 }
1933ok:
1934 spin_unlock(&res->spinlock);
1935 }
1936 spin_unlock(&dlm->spinlock);
1937
1938 // mlog(0, "woo! got an assert_master from node %u!\n",
1939 // assert->node_idx);
1940 if (mle) {
9c6510a5
KH		 1941 int extra_ref = 0;
1942 int nn = -1;
a2bf0477 1943 int rr, err = 0;
6714d8e8
KH		 1944
1945 spin_lock(&mle->spinlock);
9c6510a5
KH		 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 }
1957 }
6714d8e8
KH		 1958 mle->master = assert->node_idx;
1959 atomic_set(&mle->woken, 1);
1960 wake_up(&mle->wq);
1961 spin_unlock(&mle->spinlock);
1962
a2bf0477 1963 if (res) {
a6fa3640 1964 int wake = 0;
6714d8e8 1965 spin_lock(&res->spinlock);
a2bf0477
KH		 1966 if (mle->type == DLM_MLE_MIGRATION) {
1967 mlog(0, "finishing off migration of lockres %.*s, "
1968 "from %u to %u\n",
1969 res->lockname.len, res->lockname.name,
1970 dlm->node_num, mle->new_master);
1971 res->state &= ~DLM_LOCK_RES_MIGRATING;
a6fa3640 1972 wake = 1;
a2bf0477
KH		 1973 dlm_change_lockres_owner(dlm, res, mle->new_master);
1974 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1975 } else {
1976 dlm_change_lockres_owner(dlm, res, mle->master);
1977 }
6714d8e8 1978 spin_unlock(&res->spinlock);
ba2bf218 1979 have_lockres_ref = 1;
a6fa3640
KH		 1980 if (wake)
1981 wake_up(&res->wq);
6714d8e8 1982 }
a2bf0477
KH		 1983
1984 /* master is known, detach if not already detached.
1985 * ensures that only one assert_master call will happen
1986 * on this mle. */
1987 spin_lock(&dlm->spinlock);
1988 spin_lock(&dlm->master_lock);
1989
1990 rr = atomic_read(&mle->mle_refs.refcount);
1991 if (mle->inuse > 0) {
1992 if (extra_ref && rr < 3)
1993 err = 1;
1994 else if (!extra_ref && rr < 2)
1995 err = 1;
1996 } else {
1997 if (extra_ref && rr < 2)
1998 err = 1;
1999 else if (!extra_ref && rr < 1)
2000 err = 1;
2001 }
2002 if (err) {
2003 mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
2004 "that will mess up this node, refs=%d, extra=%d, "
2005 "inuse=%d\n", dlm->name, namelen, name,
2006 assert->node_idx, rr, extra_ref, mle->inuse);
2007 dlm_print_one_mle(mle);
2008 }
1c084577 2009 __dlm_unlink_mle(dlm, mle);
a2bf0477
KH		 2010 __dlm_mle_detach_hb_events(dlm, mle);
2011 __dlm_put_mle(mle);
6714d8e8
KH		 2012 if (extra_ref) {
2013 /* the assert master message now balances the extra
2014 * ref given by the master / migration request message.
2015 * if this is the last put, it will be removed
2016 * from the list. */
a2bf0477
KH		 2017 __dlm_put_mle(mle);
2018 }
2019 spin_unlock(&dlm->master_lock);
2020 spin_unlock(&dlm->spinlock);
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);
6714d8e8
KH		 2026 }
2027 }
2028
2029done:
9c6510a5 2030 ret = 0;
3b8118cf
KH		 2031 if (res) {
2032 spin_lock(&res->spinlock);
2033 res->state |= DLM_LOCK_RES_SETREF_INPROG;
2034 spin_unlock(&res->spinlock);
2035 *ret_data = (void *)res;
2036 }
6714d8e8 2037 dlm_put(dlm);
9c6510a5
KH		 2038 if (master_request) {
2039 mlog(0, "need to tell master to reassert\n");
ba2bf218
KH		 2040 /* positive. negative would shoot down the node. */
2041 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2042 if (!have_lockres_ref) {
2043 mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2044 "mle present here for %s:%.*s, but no lockres!\n",
2045 assert->node_idx, dlm->name, namelen, name);
2046 }
2047 }
2048 if (have_lockres_ref) {
2049 /* let the master know we have a reference to the lockres */
2050 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2051 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2052 dlm->name, namelen, name, assert->node_idx);
9c6510a5
KH		 2053 }
2054 return ret;
6714d8e8
KH		 2055
2056kill:
2057 /* kill the caller! */
a9ee4c8a
KH		 2058 mlog(ML_ERROR, "Bad message received from another node. Dumping state "
2059 "and killing the other node now! This node is OK and can continue.\n");
2060 __dlm_print_one_lock_resource(res);
6714d8e8
KH		 2061 spin_unlock(&res->spinlock);
2062 spin_unlock(&dlm->spinlock);
3b8118cf 2063 *ret_data = (void *)res;
6714d8e8
KH		 2064 dlm_put(dlm);
2065 return -EINVAL;
2066}
2067
3b8118cf
KH		 2068void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2069{
2070 struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2071
2072 if (ret_data) {
2073 spin_lock(&res->spinlock);
2074 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2075 spin_unlock(&res->spinlock);
2076 wake_up(&res->wq);
2077 dlm_lockres_put(res);
2078 }
2079 return;
2080}
2081
6714d8e8
KH		 2082int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2083 struct dlm_lock_resource *res,
2084 int ignore_higher, u8 request_from, u32 flags)
2085{
2086 struct dlm_work_item *item;
cd861280 2087 item = kzalloc(sizeof(*item), GFP_NOFS);
6714d8e8
KH		 2088 if (!item)
2089 return -ENOMEM;
2090
2091
2092 /* queue up work for dlm_assert_master_worker */
2093 dlm_grab(dlm); /* get an extra ref for the work item */
2094 dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2095 item->u.am.lockres = res; /* already have a ref */
2096 /* can optionally ignore node numbers higher than this node */
2097 item->u.am.ignore_higher = ignore_higher;
2098 item->u.am.request_from = request_from;
2099 item->u.am.flags = flags;
2100
9c6510a5
KH		 2101 if (ignore_higher)
2102 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2103 res->lockname.name);
2104
6714d8e8
KH		 2105 spin_lock(&dlm->work_lock);
2106 list_add_tail(&item->list, &dlm->work_list);
2107 spin_unlock(&dlm->work_lock);
2108
3156d267 2109 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
6714d8e8
KH		 2110 return 0;
2111}
2112
2113static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2114{
2115 struct dlm_ctxt *dlm = data;
2116 int ret = 0;
2117 struct dlm_lock_resource *res;
2118 unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2119 int ignore_higher;
2120 int bit;
2121 u8 request_from;
2122 u32 flags;
2123
2124 dlm = item->dlm;
2125 res = item->u.am.lockres;
2126 ignore_higher = item->u.am.ignore_higher;
2127 request_from = item->u.am.request_from;
2128 flags = item->u.am.flags;
2129
2130 spin_lock(&dlm->spinlock);
2131 memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2132 spin_unlock(&dlm->spinlock);
2133
2134 clear_bit(dlm->node_num, nodemap);
2135 if (ignore_higher) {
2136 /* if is this just to clear up mles for nodes below
2137 * this node, do not send the message to the original
2138 * caller or any node number higher than this */
2139 clear_bit(request_from, nodemap);
2140 bit = dlm->node_num;
2141 while (1) {
2142 bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2143 bit+1);
2144 if (bit >= O2NM_MAX_NODES)
2145 break;
2146 clear_bit(bit, nodemap);
2147 }
2148 }
2149
36407488
KH		 2150 /*
2151 * If we're migrating this lock to someone else, we are no
2152 * longer allowed to assert out own mastery. OTOH, we need to
2153 * prevent migration from starting while we're still asserting
2154 * our dominance. The reserved ast delays migration.
2155 */
2156 spin_lock(&res->spinlock);
2157 if (res->state & DLM_LOCK_RES_MIGRATING) {
2158 mlog(0, "Someone asked us to assert mastery, but we're "
2159 "in the middle of migration. Skipping assert, "
2160 "the new master will handle that.\n");
2161 spin_unlock(&res->spinlock);
2162 goto put;
2163 } else
2164 __dlm_lockres_reserve_ast(res);
2165 spin_unlock(&res->spinlock);
2166
6714d8e8
KH		 2167 /* this call now finishes out the nodemap
2168 * even if one or more nodes die */
2169 mlog(0, "worker about to master %.*s here, this=%u\n",
2170 res->lockname.len, res->lockname.name, dlm->node_num);
ba2bf218 2171 ret = dlm_do_assert_master(dlm, res, nodemap, flags);
6714d8e8
KH		 2172 if (ret < 0) {
2173 /* no need to restart, we are done */
3b3b84a8
KH		 2174 if (!dlm_is_host_down(ret))
2175 mlog_errno(ret);
6714d8e8
KH		 2176 }
2177
36407488
KH		 2178 /* Ok, we've asserted ourselves. Let's let migration start. */
2179 dlm_lockres_release_ast(dlm, res);
2180
2181put:
6714d8e8
KH		 2182 dlm_lockres_put(res);
2183
2184 mlog(0, "finished with dlm_assert_master_worker\n");
2185}
2186
c03872f5
KH		 2187/* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2188 * We cannot wait for node recovery to complete to begin mastering this
2189 * lockres because this lockres is used to kick off recovery! ;-)
2190 * So, do a pre-check on all living nodes to see if any of those nodes
2191 * think that $RECOVERY is currently mastered by a dead node. If so,
2192 * we wait a short time to allow that node to get notified by its own
2193 * heartbeat stack, then check again. All $RECOVERY lock resources
2194 * mastered by dead nodes are purged when the hearbeat callback is
2195 * fired, so we can know for sure that it is safe to continue once
2196 * the node returns a live node or no node. */
2197static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2198 struct dlm_lock_resource *res)
2199{
2200 struct dlm_node_iter iter;
2201 int nodenum;
2202 int ret = 0;
2203 u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2204
2205 spin_lock(&dlm->spinlock);
2206 dlm_node_iter_init(dlm->domain_map, &iter);
2207 spin_unlock(&dlm->spinlock);
2208
2209 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2210 /* do not send to self */
2211 if (nodenum == dlm->node_num)
2212 continue;
2213 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2214 if (ret < 0) {
2215 mlog_errno(ret);
2216 if (!dlm_is_host_down(ret))
2217 BUG();
2218 /* host is down, so answer for that node would be
2219 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
f42a100b 2220 ret = 0;
c03872f5
KH		 2221 }
2222
2223 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2224 /* check to see if this master is in the recovery map */
2225 spin_lock(&dlm->spinlock);
2226 if (test_bit(master, dlm->recovery_map)) {
2227 mlog(ML_NOTICE, "%s: node %u has not seen "
2228 "node %u go down yet, and thinks the "
2229 "dead node is mastering the recovery "
2230 "lock. must wait.\n", dlm->name,
2231 nodenum, master);
2232 ret = -EAGAIN;
2233 }
2234 spin_unlock(&dlm->spinlock);
2235 mlog(0, "%s: reco lock master is %u\n", dlm->name,
2236 master);
2237 break;
2238 }
2239 }
2240 return ret;
2241}
2242
ba2bf218
KH		 2243/*
2244 * DLM_DEREF_LOCKRES_MSG
2245 */
2246
2247int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2248{
2249 struct dlm_deref_lockres deref;
2250 int ret = 0, r;
2251 const char *lockname;
2252 unsigned int namelen;
2253
2254 lockname = res->lockname.name;
2255 namelen = res->lockname.len;
2256 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2257
2258 mlog(0, "%s:%.*s: sending deref to %d\n",
2259 dlm->name, namelen, lockname, res->owner);
2260 memset(&deref, 0, sizeof(deref));
2261 deref.node_idx = dlm->node_num;
2262 deref.namelen = namelen;
2263 memcpy(deref.name, lockname, namelen);
2264
2265 ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2266 &deref, sizeof(deref), res->owner, &r);
2267 if (ret < 0)
2268 mlog_errno(ret);
2269 else if (r < 0) {
2270 /* BAD. other node says I did not have a ref. */
2271 mlog(ML_ERROR,"while dropping ref on %s:%.*s "
2272 "(master=%u) got %d.\n", dlm->name, namelen,
2273 lockname, res->owner, r);
2274 dlm_print_one_lock_resource(res);
2275 BUG();
2276 }
2277 return ret;
2278}
2279
d74c9803
KH		 2280int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2281 void **ret_data)
ba2bf218
KH		 2282{
2283 struct dlm_ctxt *dlm = data;
2284 struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2285 struct dlm_lock_resource *res = NULL;
2286 char *name;
2287 unsigned int namelen;
2288 int ret = -EINVAL;
2289 u8 node;
2290 unsigned int hash;
f3f85464
SM		 2291 struct dlm_work_item *item;
2292 int cleared = 0;
2293 int dispatch = 0;
ba2bf218
KH		 2294
2295 if (!dlm_grab(dlm))
2296 return 0;
2297
2298 name = deref->name;
2299 namelen = deref->namelen;
2300 node = deref->node_idx;
2301
2302 if (namelen > DLM_LOCKID_NAME_MAX) {
2303 mlog(ML_ERROR, "Invalid name length!");
2304 goto done;
2305 }
2306 if (deref->node_idx >= O2NM_MAX_NODES) {
2307 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2308 goto done;
2309 }
2310
2311 hash = dlm_lockid_hash(name, namelen);
2312
2313 spin_lock(&dlm->spinlock);
2314 res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2315 if (!res) {
2316 spin_unlock(&dlm->spinlock);
2317 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2318 dlm->name, namelen, name);
2319 goto done;
2320 }
2321 spin_unlock(&dlm->spinlock);
2322
2323 spin_lock(&res->spinlock);
f3f85464
SM		 2324 if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2325 dispatch = 1;
2326 else {
2327 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2328 if (test_bit(node, res->refmap)) {
2329 dlm_lockres_clear_refmap_bit(node, res);
2330 cleared = 1;
2331 }
ba2bf218
KH		 2332 }
2333 spin_unlock(&res->spinlock);
2334
f3f85464
SM		 2335 if (!dispatch) {
2336 if (cleared)
2337 dlm_lockres_calc_usage(dlm, res);
2338 else {
2339 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2340 "but it is already dropped!\n", dlm->name,
2341 res->lockname.len, res->lockname.name, node);
2af37ce8 2342 dlm_print_one_lock_resource(res);
f3f85464
SM		 2343 }
2344 ret = 0;
2345 goto done;
2346 }
2347
2348 item = kzalloc(sizeof(*item), GFP_NOFS);
2349 if (!item) {
2350 ret = -ENOMEM;
2351 mlog_errno(ret);
2352 goto done;
2353 }
2354
2355 dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2356 item->u.dl.deref_res = res;
2357 item->u.dl.deref_node = node;
2358
2359 spin_lock(&dlm->work_lock);
2360 list_add_tail(&item->list, &dlm->work_list);
2361 spin_unlock(&dlm->work_lock);
2362
2363 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2364 return 0;
2365
ba2bf218
KH		 2366done:
2367 if (res)
2368 dlm_lockres_put(res);
2369 dlm_put(dlm);
f3f85464 2370
ba2bf218
KH		 2371 return ret;
2372}
2373
f3f85464
SM		 2374static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2375{
2376 struct dlm_ctxt *dlm;
2377 struct dlm_lock_resource *res;
2378 u8 node;
2379 u8 cleared = 0;
2380
2381 dlm = item->dlm;
2382 res = item->u.dl.deref_res;
2383 node = item->u.dl.deref_node;
2384
2385 spin_lock(&res->spinlock);
2386 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2387 if (test_bit(node, res->refmap)) {
2388 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2389 dlm_lockres_clear_refmap_bit(node, res);
2390 cleared = 1;
2391 }
2392 spin_unlock(&res->spinlock);
2393
2394 if (cleared) {
2395 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2396 dlm->name, res->lockname.len, res->lockname.name, node);
2397 dlm_lockres_calc_usage(dlm, res);
2398 } else {
2399 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2400 "but it is already dropped!\n", dlm->name,
2401 res->lockname.len, res->lockname.name, node);
2af37ce8 2402 dlm_print_one_lock_resource(res);
f3f85464
SM		 2403 }
2404
2405 dlm_lockres_put(res);
2406}
2407
2f5bf1f2
SM		 2408/* Checks whether the lockres can be migrated. Returns 0 if yes, < 0
2409 * if not. If 0, numlocks is set to the number of locks in the lockres.
2410 */
2411static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
2412 struct dlm_lock_resource *res,
2413 int *numlocks)
2414{
2415 int ret;
2416 int i;
2417 int count = 0;
800deef3 2418 struct list_head *queue;
2f5bf1f2
SM		 2419 struct dlm_lock *lock;
2420
2421 assert_spin_locked(&res->spinlock);
2422
2423 ret = -EINVAL;
2424 if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
2425 mlog(0, "cannot migrate lockres with unknown owner!\n");
2426 goto leave;
2427 }
2428
2429 if (res->owner != dlm->node_num) {
2430 mlog(0, "cannot migrate lockres this node doesn't own!\n");
2431 goto leave;
2432 }
2433
2434 ret = 0;
2435 queue = &res->granted;
2436 for (i = 0; i < 3; i++) {
800deef3 2437 list_for_each_entry(lock, queue, list) {
2f5bf1f2
SM		 2438 ++count;
2439 if (lock->ml.node == dlm->node_num) {
2440 mlog(0, "found a lock owned by this node still "
2441 "on the %s queue! will not migrate this "
2442 "lockres\n", (i == 0 ? "granted" :
2443 (i == 1 ? "converting" :
2444 "blocked")));
2445 ret = -ENOTEMPTY;
2446 goto leave;
2447 }
2448 }
2449 queue++;
2450 }
2451
2452 *numlocks = count;
2453 mlog(0, "migrateable lockres having %d locks\n", *numlocks);
2454
2455leave:
2456 return ret;
2457}
6714d8e8
KH		 2458
2459/*
2460 * DLM_MIGRATE_LOCKRES
2461 */
2462
2463
faf0ec9f
AB		 2464static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2465 struct dlm_lock_resource *res,
2466 u8 target)
6714d8e8
KH		 2467{
2468 struct dlm_master_list_entry *mle = NULL;
2469 struct dlm_master_list_entry *oldmle = NULL;
2470 struct dlm_migratable_lockres *mres = NULL;
2f5bf1f2 2471 int ret = 0;
6714d8e8
KH		 2472 const char *name;
2473 unsigned int namelen;
2474 int mle_added = 0;
2f5bf1f2
SM		 2475 int numlocks;
2476 int wake = 0;
6714d8e8
KH		 2477
2478 if (!dlm_grab(dlm))
2479 return -EINVAL;
2480
2481 name = res->lockname.name;
2482 namelen = res->lockname.len;
2483
2484 mlog(0, "migrating %.*s to %u\n", namelen, name, target);
2485
2486 /*
2487 * ensure this lockres is a proper candidate for migration
2488 */
2489 spin_lock(&res->spinlock);
2f5bf1f2
SM		 2490 ret = dlm_is_lockres_migrateable(dlm, res, &numlocks);
2491 if (ret < 0) {
6714d8e8
KH		 2492 spin_unlock(&res->spinlock);
2493 goto leave;
2494 }
6714d8e8
KH		 2495 spin_unlock(&res->spinlock);
2496
2497 /* no work to do */
2f5bf1f2 2498 if (numlocks == 0) {
6714d8e8 2499 mlog(0, "no locks were found on this lockres! done!\n");
6714d8e8
KH		 2500 goto leave;
2501 }
2502
2503 /*
2504 * preallocate up front
2505 * if this fails, abort
2506 */
2507
2508 ret = -ENOMEM;
ad8100e0 2509 mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
6714d8e8
KH		 2510 if (!mres) {
2511 mlog_errno(ret);
2512 goto leave;
2513 }
2514
2515 mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
ad8100e0 2516 GFP_NOFS);
6714d8e8
KH		 2517 if (!mle) {
2518 mlog_errno(ret);
2519 goto leave;
2520 }
2521 ret = 0;
2522
2523 /*
2524 * find a node to migrate the lockres to
2525 */
2526
2527 mlog(0, "picking a migration node\n");
2528 spin_lock(&dlm->spinlock);
2529 /* pick a new node */
2530 if (!test_bit(target, dlm->domain_map) ||
2531 target >= O2NM_MAX_NODES) {
2532 target = dlm_pick_migration_target(dlm, res);
2533 }
2534 mlog(0, "node %u chosen for migration\n", target);
2535
2536 if (target >= O2NM_MAX_NODES ||
2537 !test_bit(target, dlm->domain_map)) {
2538 /* target chosen is not alive */
2539 ret = -EINVAL;
2540 }
2541
2542 if (ret) {
2543 spin_unlock(&dlm->spinlock);
2544 goto fail;
2545 }
2546
2547 mlog(0, "continuing with target = %u\n", target);
2548
2549 /*
2550 * clear any existing master requests and
2551 * add the migration mle to the list
2552 */
2553 spin_lock(&dlm->master_lock);
2554 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2555 namelen, target, dlm->node_num);
2556 spin_unlock(&dlm->master_lock);
2557 spin_unlock(&dlm->spinlock);
2558
2559 if (ret == -EEXIST) {
2560 mlog(0, "another process is already migrating it\n");
2561 goto fail;
2562 }
2563 mle_added = 1;
2564
2565 /*
2566 * set the MIGRATING flag and flush asts
2567 * if we fail after this we need to re-dirty the lockres
2568 */
2569 if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2570 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2571 "the target went down.\n", res->lockname.len,
2572 res->lockname.name, target);
2573 spin_lock(&res->spinlock);
2574 res->state &= ~DLM_LOCK_RES_MIGRATING;
a6fa3640 2575 wake = 1;
6714d8e8
KH		 2576 spin_unlock(&res->spinlock);
2577 ret = -EINVAL;
2578 }
2579
2580fail:
2581 if (oldmle) {
2582 /* master is known, detach if not already detached */
2583 dlm_mle_detach_hb_events(dlm, oldmle);
2584 dlm_put_mle(oldmle);
2585 }
2586
2587 if (ret < 0) {
2588 if (mle_added) {
2589 dlm_mle_detach_hb_events(dlm, mle);
2590 dlm_put_mle(mle);
2591 } else if (mle) {
2592 kmem_cache_free(dlm_mle_cache, mle);
2593 }
2594 goto leave;
2595 }
2596
2597 /*
2598 * at this point, we have a migration target, an mle
2599 * in the master list, and the MIGRATING flag set on
2600 * the lockres
2601 */
2602
1cd04dbe
KH		 2603 /* now that remote nodes are spinning on the MIGRATING flag,
2604 * ensure that all assert_master work is flushed. */
2605 flush_workqueue(dlm->dlm_worker);
6714d8e8
KH		 2606
2607 /* get an extra reference on the mle.
2608 * otherwise the assert_master from the new
2609 * master will destroy this.
2610 * also, make sure that all callers of dlm_get_mle
2611 * take both dlm->spinlock and dlm->master_lock */
2612 spin_lock(&dlm->spinlock);
2613 spin_lock(&dlm->master_lock);
a2bf0477 2614 dlm_get_mle_inuse(mle);
6714d8e8
KH		 2615 spin_unlock(&dlm->master_lock);
2616 spin_unlock(&dlm->spinlock);
2617
2618 /* notify new node and send all lock state */
2619 /* call send_one_lockres with migration flag.
2620 * this serves as notice to the target node that a
2621 * migration is starting. */
2622 ret = dlm_send_one_lockres(dlm, res, mres, target,
2623 DLM_MRES_MIGRATION);
2624
2625 if (ret < 0) {
2626 mlog(0, "migration to node %u failed with %d\n",
2627 target, ret);
2628 /* migration failed, detach and clean up mle */
2629 dlm_mle_detach_hb_events(dlm, mle);
2630 dlm_put_mle(mle);
a2bf0477
KH		 2631 dlm_put_mle_inuse(mle);
2632 spin_lock(&res->spinlock);
2633 res->state &= ~DLM_LOCK_RES_MIGRATING;
a6fa3640 2634 wake = 1;
a2bf0477 2635 spin_unlock(&res->spinlock);
6714d8e8
KH		 2636 goto leave;
2637 }
2638
2639 /* at this point, the target sends a message to all nodes,
2640 * (using dlm_do_migrate_request). this node is skipped since
2641 * we had to put an mle in the list to begin the process. this
2642 * node now waits for target to do an assert master. this node
2643 * will be the last one notified, ensuring that the migration
2644 * is complete everywhere. if the target dies while this is
2645 * going on, some nodes could potentially see the target as the
2646 * master, so it is important that my recovery finds the migration
2647 * mle and sets the master to UNKNONWN. */
2648
2649
2650 /* wait for new node to assert master */
2651 while (1) {
2652 ret = wait_event_interruptible_timeout(mle->wq,
2653 (atomic_read(&mle->woken) == 1),
2654 msecs_to_jiffies(5000));
2655
2656 if (ret >= 0) {
2657 if (atomic_read(&mle->woken) == 1 ||
2658 res->owner == target)
2659 break;
2660
1cd04dbe
KH		 2661 mlog(0, "%s:%.*s: timed out during migration\n",
2662 dlm->name, res->lockname.len, res->lockname.name);
e2faea4c
KH		 2663 /* avoid hang during shutdown when migrating lockres
2664 * to a node which also goes down */
2665 if (dlm_is_node_dead(dlm, target)) {
aa852354
KH		 2666 mlog(0, "%s:%.*s: expected migration "
2667 "target %u is no longer up, restarting\n",
e2faea4c
KH		 2668 dlm->name, res->lockname.len,
2669 res->lockname.name, target);
1cd04dbe
KH		 2670 ret = -EINVAL;
2671 /* migration failed, detach and clean up mle */
2672 dlm_mle_detach_hb_events(dlm, mle);
2673 dlm_put_mle(mle);
2674 dlm_put_mle_inuse(mle);
2675 spin_lock(&res->spinlock);
2676 res->state &= ~DLM_LOCK_RES_MIGRATING;
a6fa3640 2677 wake = 1;
1cd04dbe
KH		 2678 spin_unlock(&res->spinlock);
2679 goto leave;
e2faea4c 2680 }
1cd04dbe
KH		 2681 } else
2682 mlog(0, "%s:%.*s: caught signal during migration\n",
2683 dlm->name, res->lockname.len, res->lockname.name);
6714d8e8
KH		 2684 }
2685
2686 /* all done, set the owner, clear the flag */
2687 spin_lock(&res->spinlock);
2688 dlm_set_lockres_owner(dlm, res, target);
2689 res->state &= ~DLM_LOCK_RES_MIGRATING;
2690 dlm_remove_nonlocal_locks(dlm, res);
2691 spin_unlock(&res->spinlock);
2692 wake_up(&res->wq);
2693
2694 /* master is known, detach if not already detached */
2695 dlm_mle_detach_hb_events(dlm, mle);
a2bf0477 2696 dlm_put_mle_inuse(mle);
6714d8e8
KH		 2697 ret = 0;
2698
2699 dlm_lockres_calc_usage(dlm, res);
2700
2701leave:
2702 /* re-dirty the lockres if we failed */
2703 if (ret < 0)
2704 dlm_kick_thread(dlm, res);
2705
a6fa3640
KH		 2706 /* wake up waiters if the MIGRATING flag got set
2707 * but migration failed */
2708 if (wake)
2709 wake_up(&res->wq);
2710
6714d8e8
KH		 2711 /* TODO: cleanup */
2712 if (mres)
2713 free_page((unsigned long)mres);
2714
2715 dlm_put(dlm);
2716
2717 mlog(0, "returning %d\n", ret);
2718 return ret;
2719}
6714d8e8 2720
ba2bf218
KH		 2721#define DLM_MIGRATION_RETRY_MS 100
2722
2723/* Should be called only after beginning the domain leave process.
2724 * There should not be any remaining locks on nonlocal lock resources,
2725 * and there should be no local locks left on locally mastered resources.
2726 *
2727 * Called with the dlm spinlock held, may drop it to do migration, but
2728 * will re-acquire before exit.
2729 *
2730 * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped */
2731int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2732{
2733 int ret;
2734 int lock_dropped = 0;
2f5bf1f2 2735 int numlocks;
ba2bf218 2736
b36c3f84 2737 spin_lock(&res->spinlock);
ba2bf218
KH		 2738 if (res->owner != dlm->node_num) {
2739 if (!__dlm_lockres_unused(res)) {
2740 mlog(ML_ERROR, "%s:%.*s: this node is not master, "
2741 "trying to free this but locks remain\n",
2742 dlm->name, res->lockname.len, res->lockname.name);
2743 }
b36c3f84 2744 spin_unlock(&res->spinlock);
ba2bf218
KH		 2745 goto leave;
2746 }
2f5bf1f2
SM		 2747
2748 /* No need to migrate a lockres having no locks */
2749 ret = dlm_is_lockres_migrateable(dlm, res, &numlocks);
2750 if (ret >= 0 && numlocks == 0) {
2751 spin_unlock(&res->spinlock);
2752 goto leave;
2753 }
b36c3f84 2754 spin_unlock(&res->spinlock);
ba2bf218
KH		 2755
2756 /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2757 spin_unlock(&dlm->spinlock);
2758 lock_dropped = 1;
2759 while (1) {
2760 ret = dlm_migrate_lockres(dlm, res, O2NM_MAX_NODES);
2761 if (ret >= 0)
2762 break;
2763 if (ret == -ENOTEMPTY) {
2764 mlog(ML_ERROR, "lockres %.*s still has local locks!\n",
2765 res->lockname.len, res->lockname.name);
2766 BUG();
2767 }
2768
2769 mlog(0, "lockres %.*s: migrate failed, "
2770 "retrying\n", res->lockname.len,
2771 res->lockname.name);
2772 msleep(DLM_MIGRATION_RETRY_MS);
2773 }
2774 spin_lock(&dlm->spinlock);
2775leave:
2776 return lock_dropped;
2777}
2778
6714d8e8
KH		 2779int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2780{
2781 int ret;
2782 spin_lock(&dlm->ast_lock);
2783 spin_lock(&lock->spinlock);
2784 ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2785 spin_unlock(&lock->spinlock);
2786 spin_unlock(&dlm->ast_lock);
2787 return ret;
2788}
2789
2790static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2791 struct dlm_lock_resource *res,
2792 u8 mig_target)
2793{
2794 int can_proceed;
2795 spin_lock(&res->spinlock);
2796 can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2797 spin_unlock(&res->spinlock);
2798
2799 /* target has died, so make the caller break out of the
2800 * wait_event, but caller must recheck the domain_map */
2801 spin_lock(&dlm->spinlock);
2802 if (!test_bit(mig_target, dlm->domain_map))
2803 can_proceed = 1;
2804 spin_unlock(&dlm->spinlock);
2805 return can_proceed;
2806}
2807
faf0ec9f
AB		 2808static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2809 struct dlm_lock_resource *res)
6714d8e8
KH		 2810{
2811 int ret;
2812 spin_lock(&res->spinlock);
2813 ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2814 spin_unlock(&res->spinlock);
2815 return ret;
2816}
2817
2818
2819static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2820 struct dlm_lock_resource *res,
2821 u8 target)
2822{
2823 int ret = 0;
2824
2825 mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2826 res->lockname.len, res->lockname.name, dlm->node_num,
2827 target);
2828 /* need to set MIGRATING flag on lockres. this is done by
2829 * ensuring that all asts have been flushed for this lockres. */
2830 spin_lock(&res->spinlock);
2831 BUG_ON(res->migration_pending);
2832 res->migration_pending = 1;
2833 /* strategy is to reserve an extra ast then release
2834 * it below, letting the release do all of the work */
2835 __dlm_lockres_reserve_ast(res);
2836 spin_unlock(&res->spinlock);
2837
ddc09c8d 2838 /* now flush all the pending asts */
6714d8e8 2839 dlm_kick_thread(dlm, res);
ddc09c8d
KH		 2840 /* before waiting on DIRTY, block processes which may
2841 * try to dirty the lockres before MIGRATING is set */
2842 spin_lock(&res->spinlock);
2843 BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2844 res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2845 spin_unlock(&res->spinlock);
2846 /* now wait on any pending asts and the DIRTY state */
6714d8e8
KH		 2847 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2848 dlm_lockres_release_ast(dlm, res);
2849
2850 mlog(0, "about to wait on migration_wq, dirty=%s\n",
2851 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2852 /* if the extra ref we just put was the final one, this
2853 * will pass thru immediately. otherwise, we need to wait
2854 * for the last ast to finish. */
2855again:
2856 ret = wait_event_interruptible_timeout(dlm->migration_wq,
2857 dlm_migration_can_proceed(dlm, res, target),
2858 msecs_to_jiffies(1000));
2859 if (ret < 0) {
2860 mlog(0, "woken again: migrating? %s, dead? %s\n",
2861 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2862 test_bit(target, dlm->domain_map) ? "no":"yes");
2863 } else {
2864 mlog(0, "all is well: migrating? %s, dead? %s\n",
2865 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2866 test_bit(target, dlm->domain_map) ? "no":"yes");
2867 }
2868 if (!dlm_migration_can_proceed(dlm, res, target)) {
2869 mlog(0, "trying again...\n");
2870 goto again;
2871 }
ddc09c8d
KH		 2872 /* now that we are sure the MIGRATING state is there, drop
2873 * the unneded state which blocked threads trying to DIRTY */
2874 spin_lock(&res->spinlock);
2875 BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2876 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2877 res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2878 spin_unlock(&res->spinlock);
6714d8e8
KH		 2879
2880 /* did the target go down or die? */
2881 spin_lock(&dlm->spinlock);
2882 if (!test_bit(target, dlm->domain_map)) {
2883 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2884 target);
2885 ret = -EHOSTDOWN;
2886 }
2887 spin_unlock(&dlm->spinlock);
2888
2889 /*
2890 * at this point:
2891 *
2892 * o the DLM_LOCK_RES_MIGRATING flag is set
2893 * o there are no pending asts on this lockres
2894 * o all processes trying to reserve an ast on this
2895 * lockres must wait for the MIGRATING flag to clear
2896 */
2897 return ret;
2898}
2899
2900/* last step in the migration process.
2901 * original master calls this to free all of the dlm_lock
2902 * structures that used to be for other nodes. */
2903static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2904 struct dlm_lock_resource *res)
2905{
6714d8e8 2906 struct list_head *queue = &res->granted;
ba2bf218 2907 int i, bit;
800deef3 2908 struct dlm_lock *lock, *next;
6714d8e8
KH		 2909
2910 assert_spin_locked(&res->spinlock);
2911
2912 BUG_ON(res->owner == dlm->node_num);
2913
2914 for (i=0; i<3; i++) {
800deef3 2915 list_for_each_entry_safe(lock, next, queue, list) {
6714d8e8
KH		 2916 if (lock->ml.node != dlm->node_num) {
2917 mlog(0, "putting lock for node %u\n",
2918 lock->ml.node);
2919 /* be extra careful */
2920 BUG_ON(!list_empty(&lock->ast_list));
2921 BUG_ON(!list_empty(&lock->bast_list));
2922 BUG_ON(lock->ast_pending);
2923 BUG_ON(lock->bast_pending);
ba2bf218 2924 dlm_lockres_clear_refmap_bit(lock->ml.node, res);
6714d8e8
KH		 2925 list_del_init(&lock->list);
2926 dlm_lock_put(lock);
2c5c54ac
SM		 2927 /* In a normal unlock, we would have added a
2928 * DLM_UNLOCK_FREE_LOCK action. Force it. */
2929 dlm_lock_put(lock);
6714d8e8
KH		 2930 }
2931 }
2932 queue++;
2933 }
ba2bf218
KH		 2934 bit = 0;
2935 while (1) {
2936 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2937 if (bit >= O2NM_MAX_NODES)
2938 break;
2939 /* do not clear the local node reference, if there is a
2940 * process holding this, let it drop the ref itself */
2941 if (bit != dlm->node_num) {
2942 mlog(0, "%s:%.*s: node %u had a ref to this "
2943 "migrating lockres, clearing\n", dlm->name,
2944 res->lockname.len, res->lockname.name, bit);
2945 dlm_lockres_clear_refmap_bit(bit, res);
2946 }
2947 bit++;
2948 }
6714d8e8
KH		 2949}
2950
2951/* for now this is not too intelligent. we will
2952 * need stats to make this do the right thing.
2953 * this just finds the first lock on one of the
2954 * queues and uses that node as the target. */
2955static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2956 struct dlm_lock_resource *res)
2957{
2958 int i;
2959 struct list_head *queue = &res->granted;
6714d8e8
KH		 2960 struct dlm_lock *lock;
2961 int nodenum;
2962
2963 assert_spin_locked(&dlm->spinlock);
2964
2965 spin_lock(&res->spinlock);
2966 for (i=0; i<3; i++) {
800deef3 2967 list_for_each_entry(lock, queue, list) {
6714d8e8
KH		 2968 /* up to the caller to make sure this node
2969 * is alive */
6714d8e8
KH		 2970 if (lock->ml.node != dlm->node_num) {
2971 spin_unlock(&res->spinlock);
2972 return lock->ml.node;
2973 }
2974 }
2975 queue++;
2976 }
2977 spin_unlock(&res->spinlock);
2978 mlog(0, "have not found a suitable target yet! checking domain map\n");
2979
2980 /* ok now we're getting desperate. pick anyone alive. */
2981 nodenum = -1;
2982 while (1) {
2983 nodenum = find_next_bit(dlm->domain_map,
2984 O2NM_MAX_NODES, nodenum+1);
2985 mlog(0, "found %d in domain map\n", nodenum);
2986 if (nodenum >= O2NM_MAX_NODES)
2987 break;
2988 if (nodenum != dlm->node_num) {
2989 mlog(0, "picking %d\n", nodenum);
2990 return nodenum;
2991 }
2992 }
2993
2994 mlog(0, "giving up. no master to migrate to\n");
2995 return DLM_LOCK_RES_OWNER_UNKNOWN;
2996}
2997
2998
2999
3000/* this is called by the new master once all lockres
3001 * data has been received */
3002static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
3003 struct dlm_lock_resource *res,
3004 u8 master, u8 new_master,
3005 struct dlm_node_iter *iter)
3006{
3007 struct dlm_migrate_request migrate;
2b832564 3008 int ret, skip, status = 0;
6714d8e8
KH		 3009 int nodenum;
3010
3011 memset(&migrate, 0, sizeof(migrate));
3012 migrate.namelen = res->lockname.len;
3013 memcpy(migrate.name, res->lockname.name, migrate.namelen);
3014 migrate.new_master = new_master;
3015 migrate.master = master;
3016
3017 ret = 0;
3018
3019 /* send message to all nodes, except the master and myself */
3020 while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
3021 if (nodenum == master ||
3022 nodenum == new_master)
3023 continue;
3024
2b832564
SM		 3025 /* We could race exit domain. If exited, skip. */
3026 spin_lock(&dlm->spinlock);
3027 skip = (!test_bit(nodenum, dlm->domain_map));
3028 spin_unlock(&dlm->spinlock);
3029 if (skip) {
3030 clear_bit(nodenum, iter->node_map);
3031 continue;
3032 }
3033
6714d8e8
KH		 3034 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
3035 &migrate, sizeof(migrate), nodenum,
3036 &status);
2b832564
SM		 3037 if (ret < 0) {
3038 mlog(0, "migrate_request returned %d!\n", ret);
3039 if (!dlm_is_host_down(ret)) {
3040 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
3041 BUG();
3042 }
3043 clear_bit(nodenum, iter->node_map);
3044 ret = 0;
3045 } else if (status < 0) {
6714d8e8
KH		 3046 mlog(0, "migrate request (node %u) returned %d!\n",
3047 nodenum, status);
3048 ret = status;
ba2bf218
KH		 3049 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
3050 /* during the migration request we short-circuited
3051 * the mastery of the lockres. make sure we have
3052 * a mastery ref for nodenum */
3053 mlog(0, "%s:%.*s: need ref for node %u\n",
3054 dlm->name, res->lockname.len, res->lockname.name,
3055 nodenum);
3056 spin_lock(&res->spinlock);
3057 dlm_lockres_set_refmap_bit(nodenum, res);
3058 spin_unlock(&res->spinlock);
6714d8e8
KH		 3059 }
3060 }
3061
3062 if (ret < 0)
3063 mlog_errno(ret);
3064
3065 mlog(0, "returning ret=%d\n", ret);
3066 return ret;
3067}
3068
3069
3070/* if there is an existing mle for this lockres, we now know who the master is.
3071 * (the one who sent us *this* message) we can clear it up right away.
3072 * since the process that put the mle on the list still has a reference to it,
3073 * we can unhash it now, set the master and wake the process. as a result,
3074 * we will have no mle in the list to start with. now we can add an mle for
3075 * the migration and this should be the only one found for those scanning the
3076 * list. */
d74c9803
KH		 3077int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
3078 void **ret_data)
6714d8e8
KH		 3079{
3080 struct dlm_ctxt *dlm = data;
3081 struct dlm_lock_resource *res = NULL;
3082 struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
3083 struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
3084 const char *name;
a3d33291 3085 unsigned int namelen, hash;
6714d8e8
KH		 3086 int ret = 0;
3087
3088 if (!dlm_grab(dlm))
3089 return -EINVAL;
3090
3091 name = migrate->name;
3092 namelen = migrate->namelen;
a3d33291 3093 hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 3094
3095 /* preallocate.. if this fails, abort */
3096 mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
ad8100e0 3097 GFP_NOFS);
6714d8e8
KH		 3098
3099 if (!mle) {
3100 ret = -ENOMEM;
3101 goto leave;
3102 }
3103
3104 /* check for pre-existing lock */
3105 spin_lock(&dlm->spinlock);
a3d33291 3106 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
6714d8e8
KH		 3107 spin_lock(&dlm->master_lock);
3108
3109 if (res) {
3110 spin_lock(&res->spinlock);
3111 if (res->state & DLM_LOCK_RES_RECOVERING) {
3112 /* if all is working ok, this can only mean that we got
3113 * a migrate request from a node that we now see as
3114 * dead. what can we do here? drop it to the floor? */
3115 spin_unlock(&res->spinlock);
3116 mlog(ML_ERROR, "Got a migrate request, but the "
3117 "lockres is marked as recovering!");
3118 kmem_cache_free(dlm_mle_cache, mle);
3119 ret = -EINVAL; /* need a better solution */
3120 goto unlock;
3121 }
3122 res->state |= DLM_LOCK_RES_MIGRATING;
3123 spin_unlock(&res->spinlock);
3124 }
3125
3126 /* ignore status. only nonzero status would BUG. */
3127 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3128 name, namelen,
3129 migrate->new_master,
3130 migrate->master);
3131
3132unlock:
3133 spin_unlock(&dlm->master_lock);
3134 spin_unlock(&dlm->spinlock);
3135
3136 if (oldmle) {
3137 /* master is known, detach if not already detached */
3138 dlm_mle_detach_hb_events(dlm, oldmle);
3139 dlm_put_mle(oldmle);
3140 }
3141
3142 if (res)
3143 dlm_lockres_put(res);
3144leave:
3145 dlm_put(dlm);
3146 return ret;
3147}
3148
3149/* must be holding dlm->spinlock and dlm->master_lock
3150 * when adding a migration mle, we can clear any other mles
3151 * in the master list because we know with certainty that
3152 * the master is "master". so we remove any old mle from
3153 * the list after setting it's master field, and then add
3154 * the new migration mle. this way we can hold with the rule
3155 * of having only one mle for a given lock name at all times. */
3156static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3157 struct dlm_lock_resource *res,
3158 struct dlm_master_list_entry *mle,
3159 struct dlm_master_list_entry **oldmle,
3160 const char *name, unsigned int namelen,
3161 u8 new_master, u8 master)
3162{
3163 int found;
3164 int ret = 0;
3165
3166 *oldmle = NULL;
3167
3168 mlog_entry_void();
3169
3170 assert_spin_locked(&dlm->spinlock);
3171 assert_spin_locked(&dlm->master_lock);
3172
3173 /* caller is responsible for any ref taken here on oldmle */
3174 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3175 if (found) {
3176 struct dlm_master_list_entry *tmp = *oldmle;
3177 spin_lock(&tmp->spinlock);
3178 if (tmp->type == DLM_MLE_MIGRATION) {
3179 if (master == dlm->node_num) {
3180 /* ah another process raced me to it */
3181 mlog(0, "tried to migrate %.*s, but some "
3182 "process beat me to it\n",
3183 namelen, name);
3184 ret = -EEXIST;
3185 } else {
3186 /* bad. 2 NODES are trying to migrate! */
3187 mlog(ML_ERROR, "migration error mle: "
3188 "master=%u new_master=%u // request: "
3189 "master=%u new_master=%u // "
3190 "lockres=%.*s\n",
3191 tmp->master, tmp->new_master,
3192 master, new_master,
3193 namelen, name);
3194 BUG();
3195 }
3196 } else {
3197 /* this is essentially what assert_master does */
3198 tmp->master = master;
3199 atomic_set(&tmp->woken, 1);
3200 wake_up(&tmp->wq);
1c084577
SM		 3201 /* remove it so that only one mle will be found */
3202 __dlm_unlink_mle(dlm, tmp);
ba2bf218
KH		 3203 __dlm_mle_detach_hb_events(dlm, tmp);
3204 ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3205 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3206 "telling master to get ref for cleared out mle "
3207 "during migration\n", dlm->name, namelen, name,
3208 master, new_master);
6714d8e8
KH		 3209 }
3210 spin_unlock(&tmp->spinlock);
3211 }
3212
3213 /* now add a migration mle to the tail of the list */
3214 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3215 mle->new_master = new_master;
ba2bf218
KH		 3216 /* the new master will be sending an assert master for this.
3217 * at that point we will get the refmap reference */
6714d8e8
KH		 3218 mle->master = master;
3219 /* do this for consistency with other mle types */
3220 set_bit(new_master, mle->maybe_map);
1c084577 3221 __dlm_insert_mle(dlm, mle);
6714d8e8
KH		 3222
3223 return ret;
3224}
3225
c2cd4a44
SM		 3226/*
3227 * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3228 */
3229static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3230 struct dlm_master_list_entry *mle)
3231{
3232 struct dlm_lock_resource *res;
3233 unsigned int hash;
3234
3235 /* Find the lockres associated to the mle and set its owner to UNK */
3236 hash = dlm_lockid_hash(mle->u.mlename.name, mle->u.mlename.len);
3237 res = __dlm_lookup_lockres(dlm, mle->u.mlename.name, mle->u.mlename.len,
3238 hash);
3239 if (res) {
3240 spin_unlock(&dlm->master_lock);
3241
3242 /* move lockres onto recovery list */
3243 spin_lock(&res->spinlock);
3244 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3245 dlm_move_lockres_to_recovery_list(dlm, res);
3246 spin_unlock(&res->spinlock);
3247 dlm_lockres_put(res);
3248
3249 /* about to get rid of mle, detach from heartbeat */
3250 __dlm_mle_detach_hb_events(dlm, mle);
3251
3252 /* dump the mle */
3253 spin_lock(&dlm->master_lock);
3254 __dlm_put_mle(mle);
3255 spin_unlock(&dlm->master_lock);
3256 }
3257
3258 return res;
3259}
3260
3261static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3262 struct dlm_master_list_entry *mle)
3263{
3264 __dlm_mle_detach_hb_events(dlm, mle);
3265
3266 spin_lock(&mle->spinlock);
3267 __dlm_unlink_mle(dlm, mle);
3268 atomic_set(&mle->woken, 1);
3269 spin_unlock(&mle->spinlock);
3270
3271 wake_up(&mle->wq);
3272}
3273
3274static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3275 struct dlm_master_list_entry *mle, u8 dead_node)
3276{
3277 int bit;
3278
3279 BUG_ON(mle->type != DLM_MLE_BLOCK);
3280
3281 spin_lock(&mle->spinlock);
3282 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3283 if (bit != dead_node) {
3284 mlog(0, "mle found, but dead node %u would not have been "
3285 "master\n", dead_node);
3286 spin_unlock(&mle->spinlock);
3287 } else {
3288 /* Must drop the refcount by one since the assert_master will
3289 * never arrive. This may result in the mle being unlinked and
3290 * freed, but there may still be a process waiting in the
3291 * dlmlock path which is fine. */
3292 mlog(0, "node %u was expected master\n", dead_node);
3293 atomic_set(&mle->woken, 1);
3294 spin_unlock(&mle->spinlock);
3295 wake_up(&mle->wq);
3296
3297 /* Do not need events any longer, so detach from heartbeat */
3298 __dlm_mle_detach_hb_events(dlm, mle);
3299 __dlm_put_mle(mle);
3300 }
3301}
6714d8e8
KH		 3302
3303void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3304{
2ed6c750 3305 struct dlm_master_list_entry *mle;
6714d8e8 3306 struct dlm_lock_resource *res;
2ed6c750
SM		 3307 struct hlist_head *bucket;
3308 struct hlist_node *list;
3309 unsigned int i;
6714d8e8
KH		 3310
3311 mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node);
3312top:
3313 assert_spin_locked(&dlm->spinlock);
3314
3315 /* clean the master list */
3316 spin_lock(&dlm->master_lock);
2ed6c750
SM		 3317 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3318 bucket = dlm_master_hash(dlm, i);
3319 hlist_for_each(list, bucket) {
3320 mle = hlist_entry(list, struct dlm_master_list_entry,
3321 master_hash_node);
3322
6714d8e8
KH		 3323 BUG_ON(mle->type != DLM_MLE_BLOCK &&
3324 mle->type != DLM_MLE_MASTER &&
3325 mle->type != DLM_MLE_MIGRATION);
3326
3327 /* MASTER mles are initiated locally. the waiting
3328 * process will notice the node map change
3329 * shortly. let that happen as normal. */
3330 if (mle->type == DLM_MLE_MASTER)
3331 continue;
3332
3333
3334 /* BLOCK mles are initiated by other nodes.
3335 * need to clean up if the dead node would have
3336 * been the master. */
3337 if (mle->type == DLM_MLE_BLOCK) {
c2cd4a44 3338 dlm_clean_block_mle(dlm, mle, dead_node);
6714d8e8
KH		 3339 continue;
3340 }
3341
3342 /* everything else is a MIGRATION mle */
3343
3344 /* the rule for MIGRATION mles is that the master
3345 * becomes UNKNOWN if *either* the original or
3346 * the new master dies. all UNKNOWN lockreses
3347 * are sent to whichever node becomes the recovery
3348 * master. the new master is responsible for
3349 * determining if there is still a master for
3350 * this lockres, or if he needs to take over
3351 * mastery. either way, this node should expect
3352 * another message to resolve this. */
3353 if (mle->master != dead_node &&
3354 mle->new_master != dead_node)
3355 continue;
3356
3357 /* if we have reached this point, this mle needs to
3358 * be removed from the list and freed. */
c2cd4a44 3359 dlm_clean_migration_mle(dlm, mle);
6714d8e8 3360
aa852354
KH		 3361 mlog(0, "%s: node %u died during migration from "
3362 "%u to %u!\n", dlm->name, dead_node,
6714d8e8 3363 mle->master, mle->new_master);
6714d8e8 3364
c2cd4a44
SM		 3365 /* If we find a lockres associated with the mle, we've
3366 * hit this rare case that messes up our lock ordering.
3367 * If so, we need to drop the master lock so that we can
3368 * take the lockres lock, meaning that we will have to
3369 * restart from the head of list. */
3370 res = dlm_reset_mleres_owner(dlm, mle);
3371 if (res)
6714d8e8
KH		 3372 /* restart */
3373 goto top;
6714d8e8
KH		 3374
3375 /* this may be the last reference */
3376 __dlm_put_mle(mle);
3377 }
2ed6c750 3378 }
6714d8e8
KH		 3379 spin_unlock(&dlm->master_lock);
3380}
3381
3382
3383int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3384 u8 old_master)
3385{
3386 struct dlm_node_iter iter;
3387 int ret = 0;
3388
3389 spin_lock(&dlm->spinlock);
3390 dlm_node_iter_init(dlm->domain_map, &iter);
3391 clear_bit(old_master, iter.node_map);
3392 clear_bit(dlm->node_num, iter.node_map);
3393 spin_unlock(&dlm->spinlock);
3394
ba2bf218
KH		 3395 /* ownership of the lockres is changing. account for the
3396 * mastery reference here since old_master will briefly have
3397 * a reference after the migration completes */
3398 spin_lock(&res->spinlock);
3399 dlm_lockres_set_refmap_bit(old_master, res);
3400 spin_unlock(&res->spinlock);
3401
6714d8e8
KH		 3402 mlog(0, "now time to do a migrate request to other nodes\n");
3403 ret = dlm_do_migrate_request(dlm, res, old_master,
3404 dlm->node_num, &iter);
3405 if (ret < 0) {
3406 mlog_errno(ret);
3407 goto leave;
3408 }
3409
3410 mlog(0, "doing assert master of %.*s to all except the original node\n",
3411 res->lockname.len, res->lockname.name);
3412 /* this call now finishes out the nodemap
3413 * even if one or more nodes die */
ba2bf218 3414 ret = dlm_do_assert_master(dlm, res, iter.node_map,
6714d8e8
KH		 3415 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3416 if (ret < 0) {
3417 /* no longer need to retry. all living nodes contacted. */
3418 mlog_errno(ret);
3419 ret = 0;
3420 }
3421
3422 memset(iter.node_map, 0, sizeof(iter.node_map));
3423 set_bit(old_master, iter.node_map);
3424 mlog(0, "doing assert master of %.*s back to %u\n",
3425 res->lockname.len, res->lockname.name, old_master);
ba2bf218 3426 ret = dlm_do_assert_master(dlm, res, iter.node_map,
6714d8e8
KH		 3427 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3428 if (ret < 0) {
3429 mlog(0, "assert master to original master failed "
3430 "with %d.\n", ret);
3431 /* the only nonzero status here would be because of
3432 * a dead original node. we're done. */
3433 ret = 0;
3434 }
3435
3436 /* all done, set the owner, clear the flag */
3437 spin_lock(&res->spinlock);
3438 dlm_set_lockres_owner(dlm, res, dlm->node_num);
3439 res->state &= ~DLM_LOCK_RES_MIGRATING;
3440 spin_unlock(&res->spinlock);
3441 /* re-dirty it on the new master */
3442 dlm_kick_thread(dlm, res);
3443 wake_up(&res->wq);
3444leave:
3445 return ret;
3446}
3447
3448/*
3449 * LOCKRES AST REFCOUNT
3450 * this is integral to migration
3451 */
3452
3453/* for future intent to call an ast, reserve one ahead of time.
3454 * this should be called only after waiting on the lockres
3455 * with dlm_wait_on_lockres, and while still holding the
3456 * spinlock after the call. */
3457void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3458{
3459 assert_spin_locked(&res->spinlock);
3460 if (res->state & DLM_LOCK_RES_MIGRATING) {
3461 __dlm_print_one_lock_resource(res);
3462 }
3463 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3464
3465 atomic_inc(&res->asts_reserved);
3466}
3467
3468/*
3469 * used to drop the reserved ast, either because it went unused,
3470 * or because the ast/bast was actually called.
3471 *
3472 * also, if there is a pending migration on this lockres,
3473 * and this was the last pending ast on the lockres,
3474 * atomically set the MIGRATING flag before we drop the lock.
3475 * this is how we ensure that migration can proceed with no
3476 * asts in progress. note that it is ok if the state of the
3477 * queues is such that a lock should be granted in the future
3478 * or that a bast should be fired, because the new master will
3479 * shuffle the lists on this lockres as soon as it is migrated.
3480 */
3481void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3482 struct dlm_lock_resource *res)
3483{
3484 if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3485 return;
3486
3487 if (!res->migration_pending) {
3488 spin_unlock(&res->spinlock);
3489 return;
3490 }
3491
3492 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3493 res->migration_pending = 0;
3494 res->state |= DLM_LOCK_RES_MIGRATING;
3495 spin_unlock(&res->spinlock);
3496 wake_up(&res->wq);
3497 wake_up(&dlm->migration_wq);
3498}
Linux 6.x block layer and io_uring tree(s)