projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
ocfs2/dlm: Cleanup dlm_wait_for_node_death() and dlm_wait_for_node_recovery()
[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>
6714d8e8
KH		 33#include <linux/init.h>
34#include <linux/sysctl.h>
35#include <linux/random.h>
36#include <linux/blkdev.h>
37#include <linux/socket.h>
38#include <linux/inet.h>
39#include <linux/spinlock.h>
40#include <linux/delay.h>
41
42
43#include "cluster/heartbeat.h"
44#include "cluster/nodemanager.h"
45#include "cluster/tcp.h"
46
47#include "dlmapi.h"
48#include "dlmcommon.h"
82353b59 49#include "dlmdomain.h"
e5a0334c 50#include "dlmdebug.h"
6714d8e8
KH		 51
52#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
53#include "cluster/masklog.h"
54
6714d8e8
KH		 55static void dlm_mle_node_down(struct dlm_ctxt *dlm,
56 struct dlm_master_list_entry *mle,
57 struct o2nm_node *node,
58 int idx);
59static void dlm_mle_node_up(struct dlm_ctxt *dlm,
60 struct dlm_master_list_entry *mle,
61 struct o2nm_node *node,
62 int idx);
63
64static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
ba2bf218
KH		 65static int dlm_do_assert_master(struct dlm_ctxt *dlm,
66 struct dlm_lock_resource *res,
67 void *nodemap, u32 flags);
f3f85464 68static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
6714d8e8
KH		 69
70static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
71 struct dlm_master_list_entry *mle,
72 const char *name,
73 unsigned int namelen)
74{
6714d8e8
KH		 75 if (dlm != mle->dlm)
76 return 0;
77
7141514b
SM		 78 if (namelen != mle->mnamelen ||
79 memcmp(name, mle->mname, namelen) != 0)
f77a9a78
SM		 80 return 0;
81
6714d8e8
KH		 82 return 1;
83}
84
724bdca9
SM		 85static struct kmem_cache *dlm_lockres_cache = NULL;
86static struct kmem_cache *dlm_lockname_cache = NULL;
e18b890b 87static struct kmem_cache *dlm_mle_cache = NULL;
6714d8e8 88
6714d8e8
KH		 89static void dlm_mle_release(struct kref *kref);
90static void dlm_init_mle(struct dlm_master_list_entry *mle,
91 enum dlm_mle_type type,
92 struct dlm_ctxt *dlm,
93 struct dlm_lock_resource *res,
94 const char *name,
95 unsigned int namelen);
96static void dlm_put_mle(struct dlm_master_list_entry *mle);
97static void __dlm_put_mle(struct dlm_master_list_entry *mle);
98static int dlm_find_mle(struct dlm_ctxt *dlm,
99 struct dlm_master_list_entry **mle,
100 char *name, unsigned int namelen);
101
ba2bf218
KH		 102static int dlm_do_master_request(struct dlm_lock_resource *res,
103 struct dlm_master_list_entry *mle, int to);
6714d8e8
KH		 104
105
106static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
107 struct dlm_lock_resource *res,
108 struct dlm_master_list_entry *mle,
109 int *blocked);
110static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
111 struct dlm_lock_resource *res,
112 struct dlm_master_list_entry *mle,
113 int blocked);
114static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
115 struct dlm_lock_resource *res,
116 struct dlm_master_list_entry *mle,
117 struct dlm_master_list_entry **oldmle,
118 const char *name, unsigned int namelen,
119 u8 new_master, u8 master);
120
121static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
122 struct dlm_lock_resource *res);
123static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
124 struct dlm_lock_resource *res);
125static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
126 struct dlm_lock_resource *res,
127 u8 target);
c03872f5
KH		 128static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
129 struct dlm_lock_resource *res);
6714d8e8
KH		 130
131
132int dlm_is_host_down(int errno)
133{
134 switch (errno) {
135 case -EBADF:
136 case -ECONNREFUSED:
137 case -ENOTCONN:
138 case -ECONNRESET:
139 case -EPIPE:
140 case -EHOSTDOWN:
141 case -EHOSTUNREACH:
142 case -ETIMEDOUT:
143 case -ECONNABORTED:
144 case -ENETDOWN:
145 case -ENETUNREACH:
146 case -ENETRESET:
147 case -ESHUTDOWN:
148 case -ENOPROTOOPT:
149 case -EINVAL: /* if returned from our tcp code,
150 this means there is no socket */
151 return 1;
152 }
153 return 0;
154}
155
156
157/*
158 * MASTER LIST FUNCTIONS
159 */
160
161
162/*
163 * regarding master list entries and heartbeat callbacks:
164 *
165 * in order to avoid sleeping and allocation that occurs in
166 * heartbeat, master list entries are simply attached to the
167 * dlm's established heartbeat callbacks. the mle is attached
168 * when it is created, and since the dlm->spinlock is held at
169 * that time, any heartbeat event will be properly discovered
170 * by the mle. the mle needs to be detached from the
171 * dlm->mle_hb_events list as soon as heartbeat events are no
172 * longer useful to the mle, and before the mle is freed.
173 *
174 * as a general rule, heartbeat events are no longer needed by
175 * the mle once an "answer" regarding the lock master has been
176 * received.
177 */
178static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
179 struct dlm_master_list_entry *mle)
180{
181 assert_spin_locked(&dlm->spinlock);
182
183 list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
184}
185
186
187static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
188 struct dlm_master_list_entry *mle)
189{
190 if (!list_empty(&mle->hb_events))
191 list_del_init(&mle->hb_events);
192}
193
194
195static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
196 struct dlm_master_list_entry *mle)
197{
198 spin_lock(&dlm->spinlock);
199 __dlm_mle_detach_hb_events(dlm, mle);
200 spin_unlock(&dlm->spinlock);
201}
202
a2bf0477
KH		 203static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
204{
205 struct dlm_ctxt *dlm;
206 dlm = mle->dlm;
207
208 assert_spin_locked(&dlm->spinlock);
209 assert_spin_locked(&dlm->master_lock);
210 mle->inuse++;
211 kref_get(&mle->mle_refs);
212}
213
214static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
215{
216 struct dlm_ctxt *dlm;
217 dlm = mle->dlm;
218
219 spin_lock(&dlm->spinlock);
220 spin_lock(&dlm->master_lock);
221 mle->inuse--;
222 __dlm_put_mle(mle);
223 spin_unlock(&dlm->master_lock);
224 spin_unlock(&dlm->spinlock);
225
226}
227
6714d8e8
KH		 228/* remove from list and free */
229static void __dlm_put_mle(struct dlm_master_list_entry *mle)
230{
231 struct dlm_ctxt *dlm;
232 dlm = mle->dlm;
233
234 assert_spin_locked(&dlm->spinlock);
235 assert_spin_locked(&dlm->master_lock);
aa852354
KH		 236 if (!atomic_read(&mle->mle_refs.refcount)) {
237 /* this may or may not crash, but who cares.
238 * it's a BUG. */
239 mlog(ML_ERROR, "bad mle: %p\n", mle);
240 dlm_print_one_mle(mle);
241 BUG();
242 } else
243 kref_put(&mle->mle_refs, dlm_mle_release);
6714d8e8
KH		 244}
245
246
247/* must not have any spinlocks coming in */
248static void dlm_put_mle(struct dlm_master_list_entry *mle)
249{
250 struct dlm_ctxt *dlm;
251 dlm = mle->dlm;
252
253 spin_lock(&dlm->spinlock);
254 spin_lock(&dlm->master_lock);
255 __dlm_put_mle(mle);
256 spin_unlock(&dlm->master_lock);
257 spin_unlock(&dlm->spinlock);
258}
259
260static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
261{
262 kref_get(&mle->mle_refs);
263}
264
265static void dlm_init_mle(struct dlm_master_list_entry *mle,
266 enum dlm_mle_type type,
267 struct dlm_ctxt *dlm,
268 struct dlm_lock_resource *res,
269 const char *name,
270 unsigned int namelen)
271{
272 assert_spin_locked(&dlm->spinlock);
273
274 mle->dlm = dlm;
275 mle->type = type;
2ed6c750 276 INIT_HLIST_NODE(&mle->master_hash_node);
6714d8e8
KH		 277 INIT_LIST_HEAD(&mle->hb_events);
278 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
279 spin_lock_init(&mle->spinlock);
280 init_waitqueue_head(&mle->wq);
281 atomic_set(&mle->woken, 0);
282 kref_init(&mle->mle_refs);
283 memset(mle->response_map, 0, sizeof(mle->response_map));
284 mle->master = O2NM_MAX_NODES;
285 mle->new_master = O2NM_MAX_NODES;
a2bf0477 286 mle->inuse = 0;
6714d8e8 287
f77a9a78
SM		 288 BUG_ON(mle->type != DLM_MLE_BLOCK &&
289 mle->type != DLM_MLE_MASTER &&
290 mle->type != DLM_MLE_MIGRATION);
291
6714d8e8
KH		 292 if (mle->type == DLM_MLE_MASTER) {
293 BUG_ON(!res);
7141514b
SM		 294 mle->mleres = res;
295 memcpy(mle->mname, res->lockname.name, res->lockname.len);
296 mle->mnamelen = res->lockname.len;
297 mle->mnamehash = res->lockname.hash;
f77a9a78 298 } else {
6714d8e8 299 BUG_ON(!name);
7141514b
SM		 300 mle->mleres = NULL;
301 memcpy(mle->mname, name, namelen);
302 mle->mnamelen = namelen;
303 mle->mnamehash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 304 }
305
2041d8fd
SM		 306 atomic_inc(&dlm->mle_tot_count[mle->type]);
307 atomic_inc(&dlm->mle_cur_count[mle->type]);
308
6714d8e8
KH		 309 /* copy off the node_map and register hb callbacks on our copy */
310 memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
311 memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
312 clear_bit(dlm->node_num, mle->vote_map);
313 clear_bit(dlm->node_num, mle->node_map);
314
315 /* attach the mle to the domain node up/down events */
316 __dlm_mle_attach_hb_events(dlm, mle);
317}
318
1c084577
SM		 319void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
320{
321 assert_spin_locked(&dlm->spinlock);
322 assert_spin_locked(&dlm->master_lock);
323
2ed6c750
SM		 324 if (!hlist_unhashed(&mle->master_hash_node))
325 hlist_del_init(&mle->master_hash_node);
1c084577
SM		 326}
327
328void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
329{
2ed6c750 330 struct hlist_head *bucket;
2ed6c750 331
1c084577
SM		 332 assert_spin_locked(&dlm->master_lock);
333
7141514b 334 bucket = dlm_master_hash(dlm, mle->mnamehash);
2ed6c750 335 hlist_add_head(&mle->master_hash_node, bucket);
1c084577 336}
6714d8e8
KH		 337
338/* returns 1 if found, 0 if not */
339static int dlm_find_mle(struct dlm_ctxt *dlm,
340 struct dlm_master_list_entry **mle,
341 char *name, unsigned int namelen)
342{
343 struct dlm_master_list_entry *tmpmle;
2ed6c750
SM		 344 struct hlist_head *bucket;
345 struct hlist_node *list;
346 unsigned int hash;
6714d8e8
KH		 347
348 assert_spin_locked(&dlm->master_lock);
349
2ed6c750
SM		 350 hash = dlm_lockid_hash(name, namelen);
351 bucket = dlm_master_hash(dlm, hash);
352 hlist_for_each(list, bucket) {
353 tmpmle = hlist_entry(list, struct dlm_master_list_entry,
354 master_hash_node);
6714d8e8
KH		 355 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
356 continue;
357 dlm_get_mle(tmpmle);
358 *mle = tmpmle;
359 return 1;
360 }
361 return 0;
362}
363
364void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
365{
366 struct dlm_master_list_entry *mle;
6714d8e8
KH		 367
368 assert_spin_locked(&dlm->spinlock);
2bd63216 369
800deef3 370 list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
6714d8e8
KH		 371 if (node_up)
372 dlm_mle_node_up(dlm, mle, NULL, idx);
373 else
374 dlm_mle_node_down(dlm, mle, NULL, idx);
375 }
376}
377
378static void dlm_mle_node_down(struct dlm_ctxt *dlm,
379 struct dlm_master_list_entry *mle,
380 struct o2nm_node *node, int idx)
381{
382 spin_lock(&mle->spinlock);
383
384 if (!test_bit(idx, mle->node_map))
385 mlog(0, "node %u already removed from nodemap!\n", idx);
386 else
387 clear_bit(idx, mle->node_map);
388
389 spin_unlock(&mle->spinlock);
390}
391
392static void dlm_mle_node_up(struct dlm_ctxt *dlm,
393 struct dlm_master_list_entry *mle,
394 struct o2nm_node *node, int idx)
395{
396 spin_lock(&mle->spinlock);
397
398 if (test_bit(idx, mle->node_map))
399 mlog(0, "node %u already in node map!\n", idx);
400 else
401 set_bit(idx, mle->node_map);
402
403 spin_unlock(&mle->spinlock);
404}
405
406
407int dlm_init_mle_cache(void)
408{
12eb0035 409 dlm_mle_cache = kmem_cache_create("o2dlm_mle",
6714d8e8
KH		 410 sizeof(struct dlm_master_list_entry),
411 0, SLAB_HWCACHE_ALIGN,
20c2df83 412 NULL);
6714d8e8
KH		 413 if (dlm_mle_cache == NULL)
414 return -ENOMEM;
415 return 0;
416}
417
418void dlm_destroy_mle_cache(void)
419{
420 if (dlm_mle_cache)
421 kmem_cache_destroy(dlm_mle_cache);
422}
423
424static void dlm_mle_release(struct kref *kref)
425{
426 struct dlm_master_list_entry *mle;
427 struct dlm_ctxt *dlm;
428
6714d8e8
KH		 429 mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
430 dlm = mle->dlm;
431
6714d8e8
KH		 432 assert_spin_locked(&dlm->spinlock);
433 assert_spin_locked(&dlm->master_lock);
434
7141514b
SM		 435 mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
436 mle->type);
2ed6c750 437
6714d8e8 438 /* remove from list if not already */
1c084577 439 __dlm_unlink_mle(dlm, mle);
6714d8e8
KH		 440
441 /* detach the mle from the domain node up/down events */
442 __dlm_mle_detach_hb_events(dlm, mle);
443
2041d8fd
SM		 444 atomic_dec(&dlm->mle_cur_count[mle->type]);
445
6714d8e8
KH		 446 /* NOTE: kfree under spinlock here.
447 * if this is bad, we can move this to a freelist. */
448 kmem_cache_free(dlm_mle_cache, mle);
449}
450
451
452/*
453 * LOCK RESOURCE FUNCTIONS
454 */
455
724bdca9
SM		 456int dlm_init_master_caches(void)
457{
458 dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
459 sizeof(struct dlm_lock_resource),
460 0, SLAB_HWCACHE_ALIGN, NULL);
461 if (!dlm_lockres_cache)
462 goto bail;
463
464 dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
465 DLM_LOCKID_NAME_MAX, 0,
466 SLAB_HWCACHE_ALIGN, NULL);
467 if (!dlm_lockname_cache)
468 goto bail;
469
470 return 0;
471bail:
472 dlm_destroy_master_caches();
473 return -ENOMEM;
474}
475
476void dlm_destroy_master_caches(void)
477{
478 if (dlm_lockname_cache)
479 kmem_cache_destroy(dlm_lockname_cache);
480
481 if (dlm_lockres_cache)
482 kmem_cache_destroy(dlm_lockres_cache);
483}
484
6714d8e8
KH		 485static void dlm_lockres_release(struct kref *kref)
486{
487 struct dlm_lock_resource *res;
b0d4f817 488 struct dlm_ctxt *dlm;
6714d8e8
KH		 489
490 res = container_of(kref, struct dlm_lock_resource, refs);
b0d4f817 491 dlm = res->dlm;
6714d8e8
KH		 492
493 /* This should not happen -- all lockres' have a name
494 * associated with them at init time. */
495 BUG_ON(!res->lockname.name);
496
497 mlog(0, "destroying lockres %.*s\n", res->lockname.len,
498 res->lockname.name);
499
b0d4f817 500 spin_lock(&dlm->track_lock);
29576f8b
SM		 501 if (!list_empty(&res->tracking))
502 list_del_init(&res->tracking);
503 else {
504 mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n",
505 res->lockname.len, res->lockname.name);
506 dlm_print_one_lock_resource(res);
507 }
b0d4f817
SM		 508 spin_unlock(&dlm->track_lock);
509
6800791a
SM		 510 atomic_dec(&dlm->res_cur_count);
511
a7f90d83
KH		 512 if (!hlist_unhashed(&res->hash_node) ||
513 !list_empty(&res->granted) ||
514 !list_empty(&res->converting) ||
515 !list_empty(&res->blocked) ||
516 !list_empty(&res->dirty) ||
517 !list_empty(&res->recovering) ||
518 !list_empty(&res->purge)) {
519 mlog(ML_ERROR,
520 "Going to BUG for resource %.*s."
521 " We're on a list! [%c%c%c%c%c%c%c]\n",
522 res->lockname.len, res->lockname.name,
523 !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
524 !list_empty(&res->granted) ? 'G' : ' ',
525 !list_empty(&res->converting) ? 'C' : ' ',
526 !list_empty(&res->blocked) ? 'B' : ' ',
527 !list_empty(&res->dirty) ? 'D' : ' ',
528 !list_empty(&res->recovering) ? 'R' : ' ',
529 !list_empty(&res->purge) ? 'P' : ' ');
530
531 dlm_print_one_lock_resource(res);
532 }
533
6714d8e8
KH		 534 /* By the time we're ready to blow this guy away, we shouldn't
535 * be on any lists. */
81f2094a 536 BUG_ON(!hlist_unhashed(&res->hash_node));
6714d8e8
KH		 537 BUG_ON(!list_empty(&res->granted));
538 BUG_ON(!list_empty(&res->converting));
539 BUG_ON(!list_empty(&res->blocked));
540 BUG_ON(!list_empty(&res->dirty));
541 BUG_ON(!list_empty(&res->recovering));
542 BUG_ON(!list_empty(&res->purge));
543
724bdca9 544 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
6714d8e8 545
724bdca9 546 kmem_cache_free(dlm_lockres_cache, res);
6714d8e8
KH		 547}
548
6714d8e8
KH		 549void dlm_lockres_put(struct dlm_lock_resource *res)
550{
551 kref_put(&res->refs, dlm_lockres_release);
552}
553
554static void dlm_init_lockres(struct dlm_ctxt *dlm,
555 struct dlm_lock_resource *res,
556 const char *name, unsigned int namelen)
557{
558 char *qname;
559
560 /* If we memset here, we lose our reference to the kmalloc'd
561 * res->lockname.name, so be sure to init every field
562 * correctly! */
563
564 qname = (char *) res->lockname.name;
565 memcpy(qname, name, namelen);
566
567 res->lockname.len = namelen;
a3d33291 568 res->lockname.hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 569
570 init_waitqueue_head(&res->wq);
571 spin_lock_init(&res->spinlock);
81f2094a 572 INIT_HLIST_NODE(&res->hash_node);
6714d8e8
KH		 573 INIT_LIST_HEAD(&res->granted);
574 INIT_LIST_HEAD(&res->converting);
575 INIT_LIST_HEAD(&res->blocked);
576 INIT_LIST_HEAD(&res->dirty);
577 INIT_LIST_HEAD(&res->recovering);
578 INIT_LIST_HEAD(&res->purge);
29576f8b 579 INIT_LIST_HEAD(&res->tracking);
6714d8e8
KH		 580 atomic_set(&res->asts_reserved, 0);
581 res->migration_pending = 0;
ba2bf218 582 res->inflight_locks = 0;
6714d8e8 583
b0d4f817
SM		 584 res->dlm = dlm;
585
6714d8e8
KH		 586 kref_init(&res->refs);
587
6800791a
SM		 588 atomic_inc(&dlm->res_tot_count);
589 atomic_inc(&dlm->res_cur_count);
590
6714d8e8
KH		 591 /* just for consistency */
592 spin_lock(&res->spinlock);
593 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
594 spin_unlock(&res->spinlock);
595
596 res->state = DLM_LOCK_RES_IN_PROGRESS;
597
598 res->last_used = 0;
599
18c6ac38 600 spin_lock(&dlm->spinlock);
29576f8b 601 list_add_tail(&res->tracking, &dlm->tracking_list);
18c6ac38 602 spin_unlock(&dlm->spinlock);
29576f8b 603
6714d8e8 604 memset(res->lvb, 0, DLM_LVB_LEN);
ba2bf218 605 memset(res->refmap, 0, sizeof(res->refmap));
6714d8e8
KH		 606}
607
608struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
609 const char *name,
610 unsigned int namelen)
611{
724bdca9 612 struct dlm_lock_resource *res = NULL;
6714d8e8 613
3914ed0c 614 res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
6714d8e8 615 if (!res)
724bdca9 616 goto error;
6714d8e8 617
3914ed0c 618 res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
724bdca9
SM		 619 if (!res->lockname.name)
620 goto error;
6714d8e8
KH		 621
622 dlm_init_lockres(dlm, res, name, namelen);
623 return res;
724bdca9
SM		 624
625error:
626 if (res && res->lockname.name)
627 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
628
629 if (res)
630 kmem_cache_free(dlm_lockres_cache, res);
631 return NULL;
6714d8e8
KH		 632}
633
8d400b81
SM		 634void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm,
635 struct dlm_lock_resource *res, int bit)
ba2bf218 636{
8d400b81
SM		 637 assert_spin_locked(&res->spinlock);
638
639 mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len,
640 res->lockname.name, bit, __builtin_return_address(0));
641
642 set_bit(bit, res->refmap);
643}
644
645void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm,
646 struct dlm_lock_resource *res, int bit)
647{
648 assert_spin_locked(&res->spinlock);
649
650 mlog(0, "res %.*s, clr node %u, %ps()\n", res->lockname.len,
651 res->lockname.name, bit, __builtin_return_address(0));
652
653 clear_bit(bit, res->refmap);
654}
655
656
657void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
658 struct dlm_lock_resource *res)
659{
660 assert_spin_locked(&res->spinlock);
ba2bf218
KH		 661
662 if (!test_bit(dlm->node_num, res->refmap)) {
663 BUG_ON(res->inflight_locks != 0);
8d400b81 664 dlm_lockres_set_refmap_bit(dlm, res, dlm->node_num);
ba2bf218
KH		 665 }
666 res->inflight_locks++;
8d400b81
SM		 667 mlog(0, "%s: res %.*s, inflight++: now %u, %ps()\n", dlm->name,
668 res->lockname.len, res->lockname.name, res->inflight_locks,
669 __builtin_return_address(0));
ba2bf218
KH		 670}
671
8d400b81
SM		 672void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
673 struct dlm_lock_resource *res)
ba2bf218
KH		 674{
675 assert_spin_locked(&res->spinlock);
676
677 BUG_ON(res->inflight_locks == 0);
8d400b81 678
ba2bf218 679 res->inflight_locks--;
8d400b81
SM		 680 mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm->name,
681 res->lockname.len, res->lockname.name, res->inflight_locks,
682 __builtin_return_address(0));
683
ba2bf218 684 if (res->inflight_locks == 0)
8d400b81 685 dlm_lockres_clear_refmap_bit(dlm, res, dlm->node_num);
ba2bf218
KH		 686 wake_up(&res->wq);
687}
688
6714d8e8
KH		 689/*
690 * lookup a lock resource by name.
691 * may already exist in the hashtable.
692 * lockid is null terminated
693 *
694 * if not, allocate enough for the lockres and for
695 * the temporary structure used in doing the mastering.
696 *
697 * also, do a lookup in the dlm->master_list to see
698 * if another node has begun mastering the same lock.
699 * if so, there should be a block entry in there
700 * for this name, and we should *not* attempt to master
701 * the lock here. need to wait around for that node
702 * to assert_master (or die).
703 *
704 */
705struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
706 const char *lockid,
3384f3df 707 int namelen,
6714d8e8
KH		 708 int flags)
709{
710 struct dlm_lock_resource *tmpres=NULL, *res=NULL;
711 struct dlm_master_list_entry *mle = NULL;
712 struct dlm_master_list_entry *alloc_mle = NULL;
713 int blocked = 0;
714 int ret, nodenum;
715 struct dlm_node_iter iter;
3384f3df 716 unsigned int hash;
6714d8e8 717 int tries = 0;
c03872f5 718 int bit, wait_on_recovery = 0;
ba2bf218 719 int drop_inflight_if_nonlocal = 0;
6714d8e8
KH		 720
721 BUG_ON(!lockid);
722
a3d33291 723 hash = dlm_lockid_hash(lockid, namelen);
6714d8e8
KH		 724
725 mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
726
727lookup:
728 spin_lock(&dlm->spinlock);
ba2bf218 729 tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
6714d8e8 730 if (tmpres) {
ba2bf218
KH		 731 int dropping_ref = 0;
732
7b791d68
SM		 733 spin_unlock(&dlm->spinlock);
734
ba2bf218 735 spin_lock(&tmpres->spinlock);
7b791d68
SM		 736 /* We wait for the other thread that is mastering the resource */
737 if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
738 __dlm_wait_on_lockres(tmpres);
739 BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
740 }
741
ba2bf218
KH		 742 if (tmpres->owner == dlm->node_num) {
743 BUG_ON(tmpres->state & DLM_LOCK_RES_DROPPING_REF);
744 dlm_lockres_grab_inflight_ref(dlm, tmpres);
745 } else if (tmpres->state & DLM_LOCK_RES_DROPPING_REF)
746 dropping_ref = 1;
747 spin_unlock(&tmpres->spinlock);
ba2bf218
KH		 748
749 /* wait until done messaging the master, drop our ref to allow
750 * the lockres to be purged, start over. */
751 if (dropping_ref) {
752 spin_lock(&tmpres->spinlock);
753 __dlm_wait_on_lockres_flags(tmpres, DLM_LOCK_RES_DROPPING_REF);
754 spin_unlock(&tmpres->spinlock);
755 dlm_lockres_put(tmpres);
756 tmpres = NULL;
757 goto lookup;
758 }
759
6714d8e8
KH		 760 mlog(0, "found in hash!\n");
761 if (res)
762 dlm_lockres_put(res);
763 res = tmpres;
764 goto leave;
765 }
766
767 if (!res) {
768 spin_unlock(&dlm->spinlock);
769 mlog(0, "allocating a new resource\n");
770 /* nothing found and we need to allocate one. */
3914ed0c 771 alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
6714d8e8
KH		 772 if (!alloc_mle)
773 goto leave;
774 res = dlm_new_lockres(dlm, lockid, namelen);
775 if (!res)
776 goto leave;
777 goto lookup;
778 }
779
780 mlog(0, "no lockres found, allocated our own: %p\n", res);
781
782 if (flags & LKM_LOCAL) {
783 /* caller knows it's safe to assume it's not mastered elsewhere
784 * DONE! return right away */
785 spin_lock(&res->spinlock);
786 dlm_change_lockres_owner(dlm, res, dlm->node_num);
787 __dlm_insert_lockres(dlm, res);
ba2bf218 788 dlm_lockres_grab_inflight_ref(dlm, res);
6714d8e8
KH		 789 spin_unlock(&res->spinlock);
790 spin_unlock(&dlm->spinlock);
791 /* lockres still marked IN_PROGRESS */
792 goto wake_waiters;
793 }
794
795 /* check master list to see if another node has started mastering it */
796 spin_lock(&dlm->master_lock);
797
798 /* if we found a block, wait for lock to be mastered by another node */
799 blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
800 if (blocked) {
ba2bf218 801 int mig;
6714d8e8
KH		 802 if (mle->type == DLM_MLE_MASTER) {
803 mlog(ML_ERROR, "master entry for nonexistent lock!\n");
804 BUG();
ba2bf218
KH		 805 }
806 mig = (mle->type == DLM_MLE_MIGRATION);
807 /* if there is a migration in progress, let the migration
808 * finish before continuing. we can wait for the absence
809 * of the MIGRATION mle: either the migrate finished or
810 * one of the nodes died and the mle was cleaned up.
811 * if there is a BLOCK here, but it already has a master
812 * set, we are too late. the master does not have a ref
813 * for us in the refmap. detach the mle and drop it.
814 * either way, go back to the top and start over. */
815 if (mig || mle->master != O2NM_MAX_NODES) {
816 BUG_ON(mig && mle->master == dlm->node_num);
817 /* we arrived too late. the master does not
818 * have a ref for us. retry. */
819 mlog(0, "%s:%.*s: late on %s\n",
820 dlm->name, namelen, lockid,
821 mig ? "MIGRATION" : "BLOCK");
6714d8e8 822 spin_unlock(&dlm->master_lock);
6714d8e8
KH		 823 spin_unlock(&dlm->spinlock);
824
825 /* master is known, detach */
ba2bf218
KH		 826 if (!mig)
827 dlm_mle_detach_hb_events(dlm, mle);
6714d8e8
KH		 828 dlm_put_mle(mle);
829 mle = NULL;
25985edc 830 /* this is lame, but we can't wait on either
ba2bf218
KH		 831 * the mle or lockres waitqueue here */
832 if (mig)
833 msleep(100);
834 goto lookup;
6714d8e8
KH		 835 }
836 } else {
837 /* go ahead and try to master lock on this node */
838 mle = alloc_mle;
839 /* make sure this does not get freed below */
840 alloc_mle = NULL;
841 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
842 set_bit(dlm->node_num, mle->maybe_map);
1c084577 843 __dlm_insert_mle(dlm, mle);
c03872f5
KH		 844
845 /* still holding the dlm spinlock, check the recovery map
2bd63216 846 * to see if there are any nodes that still need to be
c03872f5
KH		 847 * considered. these will not appear in the mle nodemap
848 * but they might own this lockres. wait on them. */
849 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
850 if (bit < O2NM_MAX_NODES) {
8decab3c
SM		 851 mlog(0, "%s: res %.*s, At least one node (%d) "
852 "to recover before lock mastery can begin\n",
c03872f5
KH		 853 dlm->name, namelen, (char *)lockid, bit);
854 wait_on_recovery = 1;
855 }
6714d8e8
KH		 856 }
857
858 /* at this point there is either a DLM_MLE_BLOCK or a
859 * DLM_MLE_MASTER on the master list, so it's safe to add the
860 * lockres to the hashtable. anyone who finds the lock will
861 * still have to wait on the IN_PROGRESS. */
862
863 /* finally add the lockres to its hash bucket */
864 __dlm_insert_lockres(dlm, res);
8d400b81
SM		 865
866 spin_lock(&res->spinlock);
867 dlm_lockres_grab_inflight_ref(dlm, res);
868 spin_unlock(&res->spinlock);
ba2bf218
KH		 869
870 /* if this node does not become the master make sure to drop
871 * this inflight reference below */
872 drop_inflight_if_nonlocal = 1;
873
6714d8e8
KH		 874 /* get an extra ref on the mle in case this is a BLOCK
875 * if so, the creator of the BLOCK may try to put the last
876 * ref at this time in the assert master handler, so we
877 * need an extra one to keep from a bad ptr deref. */
a2bf0477 878 dlm_get_mle_inuse(mle);
6714d8e8
KH		 879 spin_unlock(&dlm->master_lock);
880 spin_unlock(&dlm->spinlock);
881
e7e69eb3 882redo_request:
c03872f5
KH		 883 while (wait_on_recovery) {
884 /* any cluster changes that occurred after dropping the
885 * dlm spinlock would be detectable be a change on the mle,
886 * so we only need to clear out the recovery map once. */
887 if (dlm_is_recovery_lock(lockid, namelen)) {
8decab3c
SM		 888 mlog(0, "%s: Recovery map is not empty, but must "
889 "master $RECOVERY lock now\n", dlm->name);
c03872f5
KH		 890 if (!dlm_pre_master_reco_lockres(dlm, res))
891 wait_on_recovery = 0;
892 else {
893 mlog(0, "%s: waiting 500ms for heartbeat state "
894 "change\n", dlm->name);
895 msleep(500);
896 }
897 continue;
2bd63216 898 }
c03872f5
KH		 899
900 dlm_kick_recovery_thread(dlm);
aa087b84 901 msleep(1000);
c03872f5
KH		 902 dlm_wait_for_recovery(dlm);
903
904 spin_lock(&dlm->spinlock);
905 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
906 if (bit < O2NM_MAX_NODES) {
8decab3c
SM		 907 mlog(0, "%s: res %.*s, At least one node (%d) "
908 "to recover before lock mastery can begin\n",
c03872f5
KH		 909 dlm->name, namelen, (char *)lockid, bit);
910 wait_on_recovery = 1;
911 } else
912 wait_on_recovery = 0;
913 spin_unlock(&dlm->spinlock);
b7084ab5
KH		 914
915 if (wait_on_recovery)
916 dlm_wait_for_node_recovery(dlm, bit, 10000);
c03872f5
KH		 917 }
918
6714d8e8
KH		 919 /* must wait for lock to be mastered elsewhere */
920 if (blocked)
921 goto wait;
922
6714d8e8
KH		 923 ret = -EINVAL;
924 dlm_node_iter_init(mle->vote_map, &iter);
925 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
ba2bf218 926 ret = dlm_do_master_request(res, mle, nodenum);
6714d8e8
KH		 927 if (ret < 0)
928 mlog_errno(ret);
929 if (mle->master != O2NM_MAX_NODES) {
930 /* found a master ! */
9c6510a5
KH		 931 if (mle->master <= nodenum)
932 break;
933 /* if our master request has not reached the master
934 * yet, keep going until it does. this is how the
935 * master will know that asserts are needed back to
936 * the lower nodes. */
8decab3c
SM		 937 mlog(0, "%s: res %.*s, Requests only up to %u but "
938 "master is %u, keep going\n", dlm->name, namelen,
9c6510a5 939 lockid, nodenum, mle->master);
6714d8e8
KH		 940 }
941 }
942
943wait:
944 /* keep going until the response map includes all nodes */
945 ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
946 if (ret < 0) {
e7e69eb3 947 wait_on_recovery = 1;
8decab3c
SM		 948 mlog(0, "%s: res %.*s, Node map changed, redo the master "
949 "request now, blocked=%d\n", dlm->name, res->lockname.len,
6714d8e8
KH		 950 res->lockname.name, blocked);
951 if (++tries > 20) {
8decab3c
SM		 952 mlog(ML_ERROR, "%s: res %.*s, Spinning on "
953 "dlm_wait_for_lock_mastery, blocked = %d\n",
2bd63216 954 dlm->name, res->lockname.len,
6714d8e8
KH		 955 res->lockname.name, blocked);
956 dlm_print_one_lock_resource(res);
8a9343fa 957 dlm_print_one_mle(mle);
6714d8e8
KH		 958 tries = 0;
959 }
960 goto redo_request;
961 }
962
8decab3c
SM		 963 mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len,
964 res->lockname.name, res->owner);
6714d8e8
KH		 965 /* make sure we never continue without this */
966 BUG_ON(res->owner == O2NM_MAX_NODES);
967
968 /* master is known, detach if not already detached */
969 dlm_mle_detach_hb_events(dlm, mle);
970 dlm_put_mle(mle);
971 /* put the extra ref */
a2bf0477 972 dlm_put_mle_inuse(mle);
6714d8e8
KH		 973
974wake_waiters:
975 spin_lock(&res->spinlock);
ba2bf218
KH		 976 if (res->owner != dlm->node_num && drop_inflight_if_nonlocal)
977 dlm_lockres_drop_inflight_ref(dlm, res);
6714d8e8
KH		 978 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
979 spin_unlock(&res->spinlock);
980 wake_up(&res->wq);
981
982leave:
983 /* need to free the unused mle */
984 if (alloc_mle)
985 kmem_cache_free(dlm_mle_cache, alloc_mle);
986
987 return res;
988}
989
990
991#define DLM_MASTERY_TIMEOUT_MS 5000
992
993static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
994 struct dlm_lock_resource *res,
995 struct dlm_master_list_entry *mle,
996 int *blocked)
997{
998 u8 m;
999 int ret, bit;
1000 int map_changed, voting_done;
1001 int assert, sleep;
1002
1003recheck:
1004 ret = 0;
1005 assert = 0;
1006
1007 /* check if another node has already become the owner */
1008 spin_lock(&res->spinlock);
1009 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
9c6510a5
KH		 1010 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1011 res->lockname.len, res->lockname.name, res->owner);
6714d8e8 1012 spin_unlock(&res->spinlock);
9c6510a5
KH		 1013 /* this will cause the master to re-assert across
1014 * the whole cluster, freeing up mles */
588e0090 1015 if (res->owner != dlm->node_num) {
ba2bf218 1016 ret = dlm_do_master_request(res, mle, res->owner);
588e0090
KH		 1017 if (ret < 0) {
1018 /* give recovery a chance to run */
1019 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1020 msleep(500);
1021 goto recheck;
1022 }
9c6510a5
KH		 1023 }
1024 ret = 0;
6714d8e8
KH		 1025 goto leave;
1026 }
1027 spin_unlock(&res->spinlock);
1028
1029 spin_lock(&mle->spinlock);
1030 m = mle->master;
1031 map_changed = (memcmp(mle->vote_map, mle->node_map,
1032 sizeof(mle->vote_map)) != 0);
1033 voting_done = (memcmp(mle->vote_map, mle->response_map,
1034 sizeof(mle->vote_map)) == 0);
1035
1036 /* restart if we hit any errors */
1037 if (map_changed) {
1038 int b;
1039 mlog(0, "%s: %.*s: node map changed, restarting\n",
1040 dlm->name, res->lockname.len, res->lockname.name);
1041 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1042 b = (mle->type == DLM_MLE_BLOCK);
1043 if ((*blocked && !b) || (!*blocked && b)) {
2bd63216 1044 mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
6714d8e8
KH		 1045 dlm->name, res->lockname.len, res->lockname.name,
1046 *blocked, b);
1047 *blocked = b;
1048 }
1049 spin_unlock(&mle->spinlock);
1050 if (ret < 0) {
1051 mlog_errno(ret);
1052 goto leave;
1053 }
1054 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1055 "rechecking now\n", dlm->name, res->lockname.len,
1056 res->lockname.name);
1057 goto recheck;
aa852354
KH		 1058 } else {
1059 if (!voting_done) {
1060 mlog(0, "map not changed and voting not done "
1061 "for %s:%.*s\n", dlm->name, res->lockname.len,
1062 res->lockname.name);
1063 }
6714d8e8
KH		 1064 }
1065
1066 if (m != O2NM_MAX_NODES) {
1067 /* another node has done an assert!
1068 * all done! */
1069 sleep = 0;
1070 } else {
1071 sleep = 1;
1072 /* have all nodes responded? */
1073 if (voting_done && !*blocked) {
1074 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1075 if (dlm->node_num <= bit) {
1076 /* my node number is lowest.
1077 * now tell other nodes that I am
1078 * mastering this. */
1079 mle->master = dlm->node_num;
ba2bf218
KH		 1080 /* ref was grabbed in get_lock_resource
1081 * will be dropped in dlmlock_master */
6714d8e8
KH		 1082 assert = 1;
1083 sleep = 0;
1084 }
1085 /* if voting is done, but we have not received
1086 * an assert master yet, we must sleep */
1087 }
1088 }
1089
1090 spin_unlock(&mle->spinlock);
1091
1092 /* sleep if we haven't finished voting yet */
1093 if (sleep) {
1094 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1095
1096 /*
1097 if (atomic_read(&mle->mle_refs.refcount) < 2)
1098 mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1099 atomic_read(&mle->mle_refs.refcount),
1100 res->lockname.len, res->lockname.name);
1101 */
1102 atomic_set(&mle->woken, 0);
1103 (void)wait_event_timeout(mle->wq,
1104 (atomic_read(&mle->woken) == 1),
1105 timeo);
1106 if (res->owner == O2NM_MAX_NODES) {
ba2bf218
KH		 1107 mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1108 res->lockname.len, res->lockname.name);
6714d8e8
KH		 1109 goto recheck;
1110 }
1111 mlog(0, "done waiting, master is %u\n", res->owner);
1112 ret = 0;
1113 goto leave;
1114 }
1115
1116 ret = 0; /* done */
1117 if (assert) {
1118 m = dlm->node_num;
1119 mlog(0, "about to master %.*s here, this=%u\n",
1120 res->lockname.len, res->lockname.name, m);
ba2bf218 1121 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
6714d8e8
KH		 1122 if (ret) {
1123 /* This is a failure in the network path,
1124 * not in the response to the assert_master
1125 * (any nonzero response is a BUG on this node).
1126 * Most likely a socket just got disconnected
1127 * due to node death. */
1128 mlog_errno(ret);
1129 }
1130 /* no longer need to restart lock mastery.
1131 * all living nodes have been contacted. */
1132 ret = 0;
1133 }
1134
1135 /* set the lockres owner */
1136 spin_lock(&res->spinlock);
ba2bf218
KH		 1137 /* mastery reference obtained either during
1138 * assert_master_handler or in get_lock_resource */
6714d8e8
KH		 1139 dlm_change_lockres_owner(dlm, res, m);
1140 spin_unlock(&res->spinlock);
1141
1142leave:
1143 return ret;
1144}
1145
1146struct dlm_bitmap_diff_iter
1147{
1148 int curnode;
1149 unsigned long *orig_bm;
1150 unsigned long *cur_bm;
1151 unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1152};
1153
1154enum dlm_node_state_change
1155{
1156 NODE_DOWN = -1,
1157 NODE_NO_CHANGE = 0,
1158 NODE_UP
1159};
1160
1161static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1162 unsigned long *orig_bm,
1163 unsigned long *cur_bm)
1164{
1165 unsigned long p1, p2;
1166 int i;
1167
1168 iter->curnode = -1;
1169 iter->orig_bm = orig_bm;
1170 iter->cur_bm = cur_bm;
1171
1172 for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1173 p1 = *(iter->orig_bm + i);
1174 p2 = *(iter->cur_bm + i);
1175 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1176 }
1177}
1178
1179static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1180 enum dlm_node_state_change *state)
1181{
1182 int bit;
1183
1184 if (iter->curnode >= O2NM_MAX_NODES)
1185 return -ENOENT;
1186
1187 bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1188 iter->curnode+1);
1189 if (bit >= O2NM_MAX_NODES) {
1190 iter->curnode = O2NM_MAX_NODES;
1191 return -ENOENT;
1192 }
1193
1194 /* if it was there in the original then this node died */
1195 if (test_bit(bit, iter->orig_bm))
1196 *state = NODE_DOWN;
1197 else
1198 *state = NODE_UP;
1199
1200 iter->curnode = bit;
1201 return bit;
1202}
1203
1204
1205static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1206 struct dlm_lock_resource *res,
1207 struct dlm_master_list_entry *mle,
1208 int blocked)
1209{
1210 struct dlm_bitmap_diff_iter bdi;
1211 enum dlm_node_state_change sc;
1212 int node;
1213 int ret = 0;
1214
1215 mlog(0, "something happened such that the "
1216 "master process may need to be restarted!\n");
1217
1218 assert_spin_locked(&mle->spinlock);
1219
1220 dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1221 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1222 while (node >= 0) {
1223 if (sc == NODE_UP) {
e2faea4c
KH		 1224 /* a node came up. clear any old vote from
1225 * the response map and set it in the vote map
1226 * then restart the mastery. */
1227 mlog(ML_NOTICE, "node %d up while restarting\n", node);
6714d8e8
KH		 1228
1229 /* redo the master request, but only for the new node */
1230 mlog(0, "sending request to new node\n");
1231 clear_bit(node, mle->response_map);
1232 set_bit(node, mle->vote_map);
1233 } else {
1234 mlog(ML_ERROR, "node down! %d\n", node);
6714d8e8
KH		 1235 if (blocked) {
1236 int lowest = find_next_bit(mle->maybe_map,
1237 O2NM_MAX_NODES, 0);
1238
1239 /* act like it was never there */
1240 clear_bit(node, mle->maybe_map);
1241
e7e69eb3
KH		 1242 if (node == lowest) {
1243 mlog(0, "expected master %u died"
1244 " while this node was blocked "
1245 "waiting on it!\n", node);
1246 lowest = find_next_bit(mle->maybe_map,
1247 O2NM_MAX_NODES,
1248 lowest+1);
1249 if (lowest < O2NM_MAX_NODES) {
1250 mlog(0, "%s:%.*s:still "
1251 "blocked. waiting on %u "
1252 "now\n", dlm->name,
1253 res->lockname.len,
1254 res->lockname.name,
1255 lowest);
1256 } else {
1257 /* mle is an MLE_BLOCK, but
1258 * there is now nothing left to
1259 * block on. we need to return
1260 * all the way back out and try
1261 * again with an MLE_MASTER.
1262 * dlm_do_local_recovery_cleanup
1263 * has already run, so the mle
1264 * refcount is ok */
1265 mlog(0, "%s:%.*s: no "
1266 "longer blocking. try to "
1267 "master this here\n",
1268 dlm->name,
1269 res->lockname.len,
1270 res->lockname.name);
1271 mle->type = DLM_MLE_MASTER;
7141514b 1272 mle->mleres = res;
e7e69eb3 1273 }
6714d8e8 1274 }
6714d8e8
KH		 1275 }
1276
e7e69eb3
KH		 1277 /* now blank out everything, as if we had never
1278 * contacted anyone */
1279 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1280 memset(mle->response_map, 0, sizeof(mle->response_map));
1281 /* reset the vote_map to the current node_map */
1282 memcpy(mle->vote_map, mle->node_map,
1283 sizeof(mle->node_map));
1284 /* put myself into the maybe map */
1285 if (mle->type != DLM_MLE_BLOCK)
1286 set_bit(dlm->node_num, mle->maybe_map);
6714d8e8
KH		 1287 }
1288 ret = -EAGAIN;
6714d8e8
KH		 1289 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1290 }
1291 return ret;
1292}
1293
1294
1295/*
1296 * DLM_MASTER_REQUEST_MSG
1297 *
1298 * returns: 0 on success,
1299 * -errno on a network error
1300 *
1301 * on error, the caller should assume the target node is "dead"
1302 *
1303 */
1304
ba2bf218
KH		 1305static int dlm_do_master_request(struct dlm_lock_resource *res,
1306 struct dlm_master_list_entry *mle, int to)
6714d8e8
KH		 1307{
1308 struct dlm_ctxt *dlm = mle->dlm;
1309 struct dlm_master_request request;
1310 int ret, response=0, resend;
1311
1312 memset(&request, 0, sizeof(request));
1313 request.node_idx = dlm->node_num;
1314
1315 BUG_ON(mle->type == DLM_MLE_MIGRATION);
1316
7141514b
SM		 1317 request.namelen = (u8)mle->mnamelen;
1318 memcpy(request.name, mle->mname, request.namelen);
6714d8e8
KH		 1319
1320again:
1321 ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1322 sizeof(request), to, &response);
1323 if (ret < 0) {
1324 if (ret == -ESRCH) {
1325 /* should never happen */
1326 mlog(ML_ERROR, "TCP stack not ready!\n");
1327 BUG();
1328 } else if (ret == -EINVAL) {
1329 mlog(ML_ERROR, "bad args passed to o2net!\n");
1330 BUG();
1331 } else if (ret == -ENOMEM) {
1332 mlog(ML_ERROR, "out of memory while trying to send "
1333 "network message! retrying\n");
1334 /* this is totally crude */
1335 msleep(50);
1336 goto again;
1337 } else if (!dlm_is_host_down(ret)) {
1338 /* not a network error. bad. */
1339 mlog_errno(ret);
1340 mlog(ML_ERROR, "unhandled error!");
1341 BUG();
1342 }
1343 /* all other errors should be network errors,
1344 * and likely indicate node death */
1345 mlog(ML_ERROR, "link to %d went down!\n", to);
1346 goto out;
1347 }
1348
1349 ret = 0;
1350 resend = 0;
1351 spin_lock(&mle->spinlock);
1352 switch (response) {
1353 case DLM_MASTER_RESP_YES:
1354 set_bit(to, mle->response_map);
1355 mlog(0, "node %u is the master, response=YES\n", to);
ba2bf218
KH		 1356 mlog(0, "%s:%.*s: master node %u now knows I have a "
1357 "reference\n", dlm->name, res->lockname.len,
1358 res->lockname.name, to);
6714d8e8
KH		 1359 mle->master = to;
1360 break;
1361 case DLM_MASTER_RESP_NO:
1362 mlog(0, "node %u not master, response=NO\n", to);
1363 set_bit(to, mle->response_map);
1364 break;
1365 case DLM_MASTER_RESP_MAYBE:
1366 mlog(0, "node %u not master, response=MAYBE\n", to);
1367 set_bit(to, mle->response_map);
1368 set_bit(to, mle->maybe_map);
1369 break;
1370 case DLM_MASTER_RESP_ERROR:
1371 mlog(0, "node %u hit an error, resending\n", to);
1372 resend = 1;
1373 response = 0;
1374 break;
1375 default:
1376 mlog(ML_ERROR, "bad response! %u\n", response);
1377 BUG();
1378 }
1379 spin_unlock(&mle->spinlock);
1380 if (resend) {
1381 /* this is also totally crude */
1382 msleep(50);
1383 goto again;
1384 }
1385
1386out:
1387 return ret;
1388}
1389
1390/*
1391 * locks that can be taken here:
1392 * dlm->spinlock
1393 * res->spinlock
1394 * mle->spinlock
1395 * dlm->master_list
1396 *
1397 * if possible, TRIM THIS DOWN!!!
1398 */
d74c9803
KH		 1399int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1400 void **ret_data)
6714d8e8
KH		 1401{
1402 u8 response = DLM_MASTER_RESP_MAYBE;
1403 struct dlm_ctxt *dlm = data;
9c6510a5 1404 struct dlm_lock_resource *res = NULL;
6714d8e8
KH		 1405 struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1406 struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1407 char *name;
a3d33291 1408 unsigned int namelen, hash;
6714d8e8
KH		 1409 int found, ret;
1410 int set_maybe;
9c6510a5 1411 int dispatch_assert = 0;
6714d8e8
KH		 1412
1413 if (!dlm_grab(dlm))
1414 return DLM_MASTER_RESP_NO;
1415
1416 if (!dlm_domain_fully_joined(dlm)) {
1417 response = DLM_MASTER_RESP_NO;
1418 goto send_response;
1419 }
1420
1421 name = request->name;
1422 namelen = request->namelen;
a3d33291 1423 hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 1424
1425 if (namelen > DLM_LOCKID_NAME_MAX) {
1426 response = DLM_IVBUFLEN;
1427 goto send_response;
1428 }
1429
1430way_up_top:
1431 spin_lock(&dlm->spinlock);
a3d33291 1432 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
6714d8e8
KH		 1433 if (res) {
1434 spin_unlock(&dlm->spinlock);
1435
1436 /* take care of the easy cases up front */
1437 spin_lock(&res->spinlock);
1cd04dbe
KH		 1438 if (res->state & (DLM_LOCK_RES_RECOVERING|
1439 DLM_LOCK_RES_MIGRATING)) {
6714d8e8
KH		 1440 spin_unlock(&res->spinlock);
1441 mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1cd04dbe 1442 "being recovered/migrated\n");
6714d8e8
KH		 1443 response = DLM_MASTER_RESP_ERROR;
1444 if (mle)
1445 kmem_cache_free(dlm_mle_cache, mle);
1446 goto send_response;
1447 }
1448
1449 if (res->owner == dlm->node_num) {
8d400b81 1450 dlm_lockres_set_refmap_bit(dlm, res, request->node_idx);
6714d8e8 1451 spin_unlock(&res->spinlock);
6714d8e8
KH		 1452 response = DLM_MASTER_RESP_YES;
1453 if (mle)
1454 kmem_cache_free(dlm_mle_cache, mle);
1455
1456 /* this node is the owner.
1457 * there is some extra work that needs to
1458 * happen now. the requesting node has
1459 * caused all nodes up to this one to
1460 * create mles. this node now needs to
1461 * go back and clean those up. */
9c6510a5 1462 dispatch_assert = 1;
6714d8e8
KH		 1463 goto send_response;
1464 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1465 spin_unlock(&res->spinlock);
1466 // mlog(0, "node %u is the master\n", res->owner);
1467 response = DLM_MASTER_RESP_NO;
1468 if (mle)
1469 kmem_cache_free(dlm_mle_cache, mle);
1470 goto send_response;
1471 }
1472
1473 /* ok, there is no owner. either this node is
1474 * being blocked, or it is actively trying to
1475 * master this lock. */
1476 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1477 mlog(ML_ERROR, "lock with no owner should be "
1478 "in-progress!\n");
1479 BUG();
1480 }
1481
1482 // mlog(0, "lockres is in progress...\n");
1483 spin_lock(&dlm->master_lock);
1484 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1485 if (!found) {
1486 mlog(ML_ERROR, "no mle found for this lock!\n");
1487 BUG();
1488 }
1489 set_maybe = 1;
1490 spin_lock(&tmpmle->spinlock);
1491 if (tmpmle->type == DLM_MLE_BLOCK) {
1492 // mlog(0, "this node is waiting for "
1493 // "lockres to be mastered\n");
1494 response = DLM_MASTER_RESP_NO;
1495 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1496 mlog(0, "node %u is master, but trying to migrate to "
1497 "node %u.\n", tmpmle->master, tmpmle->new_master);
1498 if (tmpmle->master == dlm->node_num) {
6714d8e8
KH		 1499 mlog(ML_ERROR, "no owner on lockres, but this "
1500 "node is trying to migrate it to %u?!\n",
1501 tmpmle->new_master);
1502 BUG();
1503 } else {
1504 /* the real master can respond on its own */
1505 response = DLM_MASTER_RESP_NO;
1506 }
1507 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1508 set_maybe = 0;
9c6510a5 1509 if (tmpmle->master == dlm->node_num) {
6714d8e8 1510 response = DLM_MASTER_RESP_YES;
9c6510a5
KH		 1511 /* this node will be the owner.
1512 * go back and clean the mles on any
1513 * other nodes */
1514 dispatch_assert = 1;
8d400b81
SM		 1515 dlm_lockres_set_refmap_bit(dlm, res,
1516 request->node_idx);
9c6510a5 1517 } else
6714d8e8
KH		 1518 response = DLM_MASTER_RESP_NO;
1519 } else {
1520 // mlog(0, "this node is attempting to "
1521 // "master lockres\n");
1522 response = DLM_MASTER_RESP_MAYBE;
1523 }
1524 if (set_maybe)
1525 set_bit(request->node_idx, tmpmle->maybe_map);
1526 spin_unlock(&tmpmle->spinlock);
1527
1528 spin_unlock(&dlm->master_lock);
1529 spin_unlock(&res->spinlock);
1530
1531 /* keep the mle attached to heartbeat events */
1532 dlm_put_mle(tmpmle);
1533 if (mle)
1534 kmem_cache_free(dlm_mle_cache, mle);
1535 goto send_response;
1536 }
1537
1538 /*
1539 * lockres doesn't exist on this node
1540 * if there is an MLE_BLOCK, return NO
1541 * if there is an MLE_MASTER, return MAYBE
1542 * otherwise, add an MLE_BLOCK, return NO
1543 */
1544 spin_lock(&dlm->master_lock);
1545 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1546 if (!found) {
1547 /* this lockid has never been seen on this node yet */
1548 // mlog(0, "no mle found\n");
1549 if (!mle) {
1550 spin_unlock(&dlm->master_lock);
1551 spin_unlock(&dlm->spinlock);
1552
3914ed0c 1553 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
6714d8e8 1554 if (!mle) {
6714d8e8 1555 response = DLM_MASTER_RESP_ERROR;
9c6510a5 1556 mlog_errno(-ENOMEM);
6714d8e8
KH		 1557 goto send_response;
1558 }
6714d8e8
KH		 1559 goto way_up_top;
1560 }
1561
1562 // mlog(0, "this is second time thru, already allocated, "
1563 // "add the block.\n");
41b8c8a1 1564 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
6714d8e8 1565 set_bit(request->node_idx, mle->maybe_map);
1c084577 1566 __dlm_insert_mle(dlm, mle);
6714d8e8
KH		 1567 response = DLM_MASTER_RESP_NO;
1568 } else {
1569 // mlog(0, "mle was found\n");
1570 set_maybe = 1;
1571 spin_lock(&tmpmle->spinlock);
9c6510a5
KH		 1572 if (tmpmle->master == dlm->node_num) {
1573 mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1574 BUG();
1575 }
6714d8e8
KH		 1576 if (tmpmle->type == DLM_MLE_BLOCK)
1577 response = DLM_MASTER_RESP_NO;
1578 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1579 mlog(0, "migration mle was found (%u->%u)\n",
1580 tmpmle->master, tmpmle->new_master);
6714d8e8
KH		 1581 /* real master can respond on its own */
1582 response = DLM_MASTER_RESP_NO;
9c6510a5
KH		 1583 } else
1584 response = DLM_MASTER_RESP_MAYBE;
6714d8e8
KH		 1585 if (set_maybe)
1586 set_bit(request->node_idx, tmpmle->maybe_map);
1587 spin_unlock(&tmpmle->spinlock);
1588 }
1589 spin_unlock(&dlm->master_lock);
1590 spin_unlock(&dlm->spinlock);
1591
1592 if (found) {
1593 /* keep the mle attached to heartbeat events */
1594 dlm_put_mle(tmpmle);
1595 }
1596send_response:
b31cfc02
SM		 1597 /*
1598 * __dlm_lookup_lockres() grabbed a reference to this lockres.
1599 * The reference is released by dlm_assert_master_worker() under
1600 * the call to dlm_dispatch_assert_master(). If
1601 * dlm_assert_master_worker() isn't called, we drop it here.
1602 */
9c6510a5
KH		 1603 if (dispatch_assert) {
1604 if (response != DLM_MASTER_RESP_YES)
1605 mlog(ML_ERROR, "invalid response %d\n", response);
1606 if (!res) {
1607 mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1608 BUG();
1609 }
1610 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1611 dlm->node_num, res->lockname.len, res->lockname.name);
2bd63216 1612 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
9c6510a5
KH		 1613 DLM_ASSERT_MASTER_MLE_CLEANUP);
1614 if (ret < 0) {
1615 mlog(ML_ERROR, "failed to dispatch assert master work\n");
1616 response = DLM_MASTER_RESP_ERROR;
b31cfc02 1617 dlm_lockres_put(res);
9c6510a5 1618 }
b31cfc02
SM		 1619 } else {
1620 if (res)
1621 dlm_lockres_put(res);
9c6510a5
KH		 1622 }
1623
6714d8e8
KH		 1624 dlm_put(dlm);
1625 return response;
1626}
1627
1628/*
1629 * DLM_ASSERT_MASTER_MSG
1630 */
1631
1632
1633/*
1634 * NOTE: this can be used for debugging
1635 * can periodically run all locks owned by this node
1636 * and re-assert across the cluster...
1637 */
05488bbe
AB		 1638static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1639 struct dlm_lock_resource *res,
1640 void *nodemap, u32 flags)
6714d8e8
KH		 1641{
1642 struct dlm_assert_master assert;
1643 int to, tmpret;
1644 struct dlm_node_iter iter;
1645 int ret = 0;
9c6510a5 1646 int reassert;
ba2bf218
KH		 1647 const char *lockname = res->lockname.name;
1648 unsigned int namelen = res->lockname.len;
6714d8e8
KH		 1649
1650 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
f3f85464
SM		 1651
1652 spin_lock(&res->spinlock);
1653 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1654 spin_unlock(&res->spinlock);
1655
9c6510a5
KH		 1656again:
1657 reassert = 0;
6714d8e8
KH		 1658
1659 /* note that if this nodemap is empty, it returns 0 */
1660 dlm_node_iter_init(nodemap, &iter);
1661 while ((to = dlm_node_iter_next(&iter)) >= 0) {
1662 int r = 0;
a9ee4c8a
KH		 1663 struct dlm_master_list_entry *mle = NULL;
1664
6714d8e8
KH		 1665 mlog(0, "sending assert master to %d (%.*s)\n", to,
1666 namelen, lockname);
1667 memset(&assert, 0, sizeof(assert));
1668 assert.node_idx = dlm->node_num;
1669 assert.namelen = namelen;
1670 memcpy(assert.name, lockname, namelen);
1671 assert.flags = cpu_to_be32(flags);
1672
1673 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1674 &assert, sizeof(assert), to, &r);
1675 if (tmpret < 0) {
a5196ec5
WW		 1676 mlog(ML_ERROR, "Error %d when sending message %u (key "
1677 "0x%x) to node %u\n", tmpret,
1678 DLM_ASSERT_MASTER_MSG, dlm->key, to);
6714d8e8 1679 if (!dlm_is_host_down(tmpret)) {
3b3b84a8 1680 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
6714d8e8
KH		 1681 BUG();
1682 }
1683 /* a node died. finish out the rest of the nodes. */
3b3b84a8 1684 mlog(0, "link to %d went down!\n", to);
6714d8e8
KH		 1685 /* any nonzero status return will do */
1686 ret = tmpret;
ba2bf218 1687 r = 0;
6714d8e8
KH		 1688 } else if (r < 0) {
1689 /* ok, something horribly messed. kill thyself. */
1690 mlog(ML_ERROR,"during assert master of %.*s to %u, "
1691 "got %d.\n", namelen, lockname, to, r);
a9ee4c8a
KH		 1692 spin_lock(&dlm->spinlock);
1693 spin_lock(&dlm->master_lock);
1694 if (dlm_find_mle(dlm, &mle, (char *)lockname,
1695 namelen)) {
1696 dlm_print_one_mle(mle);
1697 __dlm_put_mle(mle);
1698 }
1699 spin_unlock(&dlm->master_lock);
1700 spin_unlock(&dlm->spinlock);
6714d8e8 1701 BUG();
ba2bf218
KH		 1702 }
1703
1704 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1705 !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1706 mlog(ML_ERROR, "%.*s: very strange, "
1707 "master MLE but no lockres on %u\n",
1708 namelen, lockname, to);
1709 }
1710
1711 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
9c6510a5 1712 mlog(0, "%.*s: node %u create mles on other "
2bd63216 1713 "nodes and requests a re-assert\n",
9c6510a5
KH		 1714 namelen, lockname, to);
1715 reassert = 1;
6714d8e8 1716 }
ba2bf218
KH		 1717 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1718 mlog(0, "%.*s: node %u has a reference to this "
1719 "lockres, set the bit in the refmap\n",
1720 namelen, lockname, to);
1721 spin_lock(&res->spinlock);
8d400b81 1722 dlm_lockres_set_refmap_bit(dlm, res, to);
ba2bf218
KH		 1723 spin_unlock(&res->spinlock);
1724 }
6714d8e8
KH		 1725 }
1726
9c6510a5
KH		 1727 if (reassert)
1728 goto again;
1729
f3f85464
SM		 1730 spin_lock(&res->spinlock);
1731 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1732 spin_unlock(&res->spinlock);
1733 wake_up(&res->wq);
1734
6714d8e8
KH		 1735 return ret;
1736}
1737
1738/*
1739 * locks that can be taken here:
1740 * dlm->spinlock
1741 * res->spinlock
1742 * mle->spinlock
1743 * dlm->master_list
1744 *
1745 * if possible, TRIM THIS DOWN!!!
1746 */
d74c9803
KH		 1747int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1748 void **ret_data)
6714d8e8
KH		 1749{
1750 struct dlm_ctxt *dlm = data;
1751 struct dlm_master_list_entry *mle = NULL;
1752 struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1753 struct dlm_lock_resource *res = NULL;
1754 char *name;
a3d33291 1755 unsigned int namelen, hash;
6714d8e8 1756 u32 flags;
ba2bf218 1757 int master_request = 0, have_lockres_ref = 0;
9c6510a5 1758 int ret = 0;
6714d8e8
KH		 1759
1760 if (!dlm_grab(dlm))
1761 return 0;
1762
1763 name = assert->name;
1764 namelen = assert->namelen;
a3d33291 1765 hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 1766 flags = be32_to_cpu(assert->flags);
1767
1768 if (namelen > DLM_LOCKID_NAME_MAX) {
1769 mlog(ML_ERROR, "Invalid name length!");
1770 goto done;
1771 }
1772
1773 spin_lock(&dlm->spinlock);
1774
1775 if (flags)
1776 mlog(0, "assert_master with flags: %u\n", flags);
1777
1778 /* find the MLE */
1779 spin_lock(&dlm->master_lock);
1780 if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1781 /* not an error, could be master just re-asserting */
1782 mlog(0, "just got an assert_master from %u, but no "
1783 "MLE for it! (%.*s)\n", assert->node_idx,
1784 namelen, name);
1785 } else {
1786 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1787 if (bit >= O2NM_MAX_NODES) {
1788 /* not necessarily an error, though less likely.
1789 * could be master just re-asserting. */
aa852354 1790 mlog(0, "no bits set in the maybe_map, but %u "
6714d8e8
KH		 1791 "is asserting! (%.*s)\n", assert->node_idx,
1792 namelen, name);
1793 } else if (bit != assert->node_idx) {
1794 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1795 mlog(0, "master %u was found, %u should "
1796 "back off\n", assert->node_idx, bit);
1797 } else {
1798 /* with the fix for bug 569, a higher node
1799 * number winning the mastery will respond
1800 * YES to mastery requests, but this node
1801 * had no way of knowing. let it pass. */
aa852354 1802 mlog(0, "%u is the lowest node, "
6714d8e8
KH		 1803 "%u is asserting. (%.*s) %u must "
1804 "have begun after %u won.\n", bit,
1805 assert->node_idx, namelen, name, bit,
1806 assert->node_idx);
1807 }
1808 }
2d1a868c
KH		 1809 if (mle->type == DLM_MLE_MIGRATION) {
1810 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1811 mlog(0, "%s:%.*s: got cleanup assert"
1812 " from %u for migration\n",
1813 dlm->name, namelen, name,
1814 assert->node_idx);
1815 } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1816 mlog(0, "%s:%.*s: got unrelated assert"
1817 " from %u for migration, ignoring\n",
1818 dlm->name, namelen, name,
1819 assert->node_idx);
1820 __dlm_put_mle(mle);
1821 spin_unlock(&dlm->master_lock);
1822 spin_unlock(&dlm->spinlock);
1823 goto done;
2bd63216 1824 }
2d1a868c 1825 }
6714d8e8
KH		 1826 }
1827 spin_unlock(&dlm->master_lock);
1828
1829 /* ok everything checks out with the MLE
1830 * now check to see if there is a lockres */
a3d33291 1831 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
6714d8e8
KH		 1832 if (res) {
1833 spin_lock(&res->spinlock);
1834 if (res->state & DLM_LOCK_RES_RECOVERING) {
1835 mlog(ML_ERROR, "%u asserting but %.*s is "
1836 "RECOVERING!\n", assert->node_idx, namelen, name);
1837 goto kill;
1838 }
1839 if (!mle) {
dc2ed195
KH		 1840 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1841 res->owner != assert->node_idx) {
53ecd25e
SM		 1842 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1843 "but current owner is %u! (%.*s)\n",
1844 assert->node_idx, res->owner, namelen,
1845 name);
1846 __dlm_print_one_lock_resource(res);
1847 BUG();
6714d8e8
KH		 1848 }
1849 } else if (mle->type != DLM_MLE_MIGRATION) {
1850 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1851 /* owner is just re-asserting */
1852 if (res->owner == assert->node_idx) {
1853 mlog(0, "owner %u re-asserting on "
1854 "lock %.*s\n", assert->node_idx,
1855 namelen, name);
1856 goto ok;
1857 }
1858 mlog(ML_ERROR, "got assert_master from "
1859 "node %u, but %u is the owner! "
1860 "(%.*s)\n", assert->node_idx,
1861 res->owner, namelen, name);
1862 goto kill;
1863 }
1864 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1865 mlog(ML_ERROR, "got assert from %u, but lock "
1866 "with no owner should be "
1867 "in-progress! (%.*s)\n",
1868 assert->node_idx,
1869 namelen, name);
1870 goto kill;
1871 }
1872 } else /* mle->type == DLM_MLE_MIGRATION */ {
1873 /* should only be getting an assert from new master */
1874 if (assert->node_idx != mle->new_master) {
1875 mlog(ML_ERROR, "got assert from %u, but "
1876 "new master is %u, and old master "
1877 "was %u (%.*s)\n",
1878 assert->node_idx, mle->new_master,
1879 mle->master, namelen, name);
1880 goto kill;
1881 }
1882
1883 }
1884ok:
1885 spin_unlock(&res->spinlock);
1886 }
6714d8e8
KH		 1887
1888 // mlog(0, "woo! got an assert_master from node %u!\n",
1889 // assert->node_idx);
1890 if (mle) {
9c6510a5
KH		 1891 int extra_ref = 0;
1892 int nn = -1;
a2bf0477 1893 int rr, err = 0;
2bd63216 1894
6714d8e8 1895 spin_lock(&mle->spinlock);
9c6510a5
KH		 1896 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1897 extra_ref = 1;
1898 else {
1899 /* MASTER mle: if any bits set in the response map
1900 * then the calling node needs to re-assert to clear
1901 * up nodes that this node contacted */
2bd63216 1902 while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
9c6510a5
KH		 1903 nn+1)) < O2NM_MAX_NODES) {
1904 if (nn != dlm->node_num && nn != assert->node_idx)
1905 master_request = 1;
1906 }
1907 }
6714d8e8
KH		 1908 mle->master = assert->node_idx;
1909 atomic_set(&mle->woken, 1);
1910 wake_up(&mle->wq);
1911 spin_unlock(&mle->spinlock);
1912
a2bf0477 1913 if (res) {
a6fa3640 1914 int wake = 0;
6714d8e8 1915 spin_lock(&res->spinlock);
a2bf0477
KH		 1916 if (mle->type == DLM_MLE_MIGRATION) {
1917 mlog(0, "finishing off migration of lockres %.*s, "
1918 "from %u to %u\n",
1919 res->lockname.len, res->lockname.name,
1920 dlm->node_num, mle->new_master);
1921 res->state &= ~DLM_LOCK_RES_MIGRATING;
a6fa3640 1922 wake = 1;
a2bf0477
KH		 1923 dlm_change_lockres_owner(dlm, res, mle->new_master);
1924 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1925 } else {
1926 dlm_change_lockres_owner(dlm, res, mle->master);
1927 }
6714d8e8 1928 spin_unlock(&res->spinlock);
ba2bf218 1929 have_lockres_ref = 1;
a6fa3640
KH		 1930 if (wake)
1931 wake_up(&res->wq);
6714d8e8 1932 }
a2bf0477
KH		 1933
1934 /* master is known, detach if not already detached.
1935 * ensures that only one assert_master call will happen
1936 * on this mle. */
a2bf0477
KH		 1937 spin_lock(&dlm->master_lock);
1938
1939 rr = atomic_read(&mle->mle_refs.refcount);
1940 if (mle->inuse > 0) {
1941 if (extra_ref && rr < 3)
1942 err = 1;
1943 else if (!extra_ref && rr < 2)
1944 err = 1;
1945 } else {
1946 if (extra_ref && rr < 2)
1947 err = 1;
1948 else if (!extra_ref && rr < 1)
1949 err = 1;
1950 }
1951 if (err) {
1952 mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
1953 "that will mess up this node, refs=%d, extra=%d, "
1954 "inuse=%d\n", dlm->name, namelen, name,
1955 assert->node_idx, rr, extra_ref, mle->inuse);
1956 dlm_print_one_mle(mle);
1957 }
1c084577 1958 __dlm_unlink_mle(dlm, mle);
a2bf0477
KH		 1959 __dlm_mle_detach_hb_events(dlm, mle);
1960 __dlm_put_mle(mle);
6714d8e8
KH		 1961 if (extra_ref) {
1962 /* the assert master message now balances the extra
1963 * ref given by the master / migration request message.
1964 * if this is the last put, it will be removed
1965 * from the list. */
a2bf0477
KH		 1966 __dlm_put_mle(mle);
1967 }
1968 spin_unlock(&dlm->master_lock);
a2bf0477
KH		 1969 } else if (res) {
1970 if (res->owner != assert->node_idx) {
1971 mlog(0, "assert_master from %u, but current "
1972 "owner is %u (%.*s), no mle\n", assert->node_idx,
1973 res->owner, namelen, name);
6714d8e8
KH		 1974 }
1975 }
14741472 1976 spin_unlock(&dlm->spinlock);
6714d8e8
KH		 1977
1978done:
9c6510a5 1979 ret = 0;
3b8118cf
KH		 1980 if (res) {
1981 spin_lock(&res->spinlock);
1982 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1983 spin_unlock(&res->spinlock);
1984 *ret_data = (void *)res;
1985 }
6714d8e8 1986 dlm_put(dlm);
9c6510a5
KH		 1987 if (master_request) {
1988 mlog(0, "need to tell master to reassert\n");
ba2bf218
KH		 1989 /* positive. negative would shoot down the node. */
1990 ret |= DLM_ASSERT_RESPONSE_REASSERT;
1991 if (!have_lockres_ref) {
1992 mlog(ML_ERROR, "strange, got assert from %u, MASTER "
1993 "mle present here for %s:%.*s, but no lockres!\n",
1994 assert->node_idx, dlm->name, namelen, name);
1995 }
1996 }
1997 if (have_lockres_ref) {
1998 /* let the master know we have a reference to the lockres */
1999 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2000 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2001 dlm->name, namelen, name, assert->node_idx);
9c6510a5
KH		 2002 }
2003 return ret;
6714d8e8
KH		 2004
2005kill:
2006 /* kill the caller! */
a9ee4c8a
KH		 2007 mlog(ML_ERROR, "Bad message received from another node. Dumping state "
2008 "and killing the other node now! This node is OK and can continue.\n");
2009 __dlm_print_one_lock_resource(res);
6714d8e8
KH		 2010 spin_unlock(&res->spinlock);
2011 spin_unlock(&dlm->spinlock);
2bd63216 2012 *ret_data = (void *)res;
6714d8e8
KH		 2013 dlm_put(dlm);
2014 return -EINVAL;
2015}
2016
3b8118cf
KH		 2017void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2018{
2019 struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2020
2021 if (ret_data) {
2022 spin_lock(&res->spinlock);
2023 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2024 spin_unlock(&res->spinlock);
2025 wake_up(&res->wq);
2026 dlm_lockres_put(res);
2027 }
2028 return;
2029}
2030
6714d8e8
KH		 2031int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2032 struct dlm_lock_resource *res,
2033 int ignore_higher, u8 request_from, u32 flags)
2034{
2035 struct dlm_work_item *item;
cd861280 2036 item = kzalloc(sizeof(*item), GFP_NOFS);
6714d8e8
KH		 2037 if (!item)
2038 return -ENOMEM;
2039
2040
2041 /* queue up work for dlm_assert_master_worker */
2042 dlm_grab(dlm); /* get an extra ref for the work item */
2043 dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2044 item->u.am.lockres = res; /* already have a ref */
2045 /* can optionally ignore node numbers higher than this node */
2046 item->u.am.ignore_higher = ignore_higher;
2047 item->u.am.request_from = request_from;
2048 item->u.am.flags = flags;
2049
2bd63216
SM		 2050 if (ignore_higher)
2051 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
9c6510a5 2052 res->lockname.name);
2bd63216 2053
6714d8e8
KH		 2054 spin_lock(&dlm->work_lock);
2055 list_add_tail(&item->list, &dlm->work_list);
2056 spin_unlock(&dlm->work_lock);
2057
3156d267 2058 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
6714d8e8
KH		 2059 return 0;
2060}
2061
2062static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2063{
2064 struct dlm_ctxt *dlm = data;
2065 int ret = 0;
2066 struct dlm_lock_resource *res;
2067 unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2068 int ignore_higher;
2069 int bit;
2070 u8 request_from;
2071 u32 flags;
2072
2073 dlm = item->dlm;
2074 res = item->u.am.lockres;
2075 ignore_higher = item->u.am.ignore_higher;
2076 request_from = item->u.am.request_from;
2077 flags = item->u.am.flags;
2078
2079 spin_lock(&dlm->spinlock);
2080 memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2081 spin_unlock(&dlm->spinlock);
2082
2083 clear_bit(dlm->node_num, nodemap);
2084 if (ignore_higher) {
2085 /* if is this just to clear up mles for nodes below
2086 * this node, do not send the message to the original
2087 * caller or any node number higher than this */
2088 clear_bit(request_from, nodemap);
2089 bit = dlm->node_num;
2090 while (1) {
2091 bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2092 bit+1);
2093 if (bit >= O2NM_MAX_NODES)
2094 break;
2095 clear_bit(bit, nodemap);
2096 }
2097 }
2098
36407488
KH		 2099 /*
2100 * If we're migrating this lock to someone else, we are no
2101 * longer allowed to assert out own mastery. OTOH, we need to
2102 * prevent migration from starting while we're still asserting
2103 * our dominance. The reserved ast delays migration.
2104 */
2105 spin_lock(&res->spinlock);
2106 if (res->state & DLM_LOCK_RES_MIGRATING) {
2107 mlog(0, "Someone asked us to assert mastery, but we're "
2108 "in the middle of migration. Skipping assert, "
2109 "the new master will handle that.\n");
2110 spin_unlock(&res->spinlock);
2111 goto put;
2112 } else
2113 __dlm_lockres_reserve_ast(res);
2114 spin_unlock(&res->spinlock);
2115
6714d8e8
KH		 2116 /* this call now finishes out the nodemap
2117 * even if one or more nodes die */
2118 mlog(0, "worker about to master %.*s here, this=%u\n",
2119 res->lockname.len, res->lockname.name, dlm->node_num);
ba2bf218 2120 ret = dlm_do_assert_master(dlm, res, nodemap, flags);
6714d8e8
KH		 2121 if (ret < 0) {
2122 /* no need to restart, we are done */
3b3b84a8
KH		 2123 if (!dlm_is_host_down(ret))
2124 mlog_errno(ret);
6714d8e8
KH		 2125 }
2126
36407488
KH		 2127 /* Ok, we've asserted ourselves. Let's let migration start. */
2128 dlm_lockres_release_ast(dlm, res);
2129
2130put:
6714d8e8
KH		 2131 dlm_lockres_put(res);
2132
2133 mlog(0, "finished with dlm_assert_master_worker\n");
2134}
2135
c03872f5
KH		 2136/* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2137 * We cannot wait for node recovery to complete to begin mastering this
2138 * lockres because this lockres is used to kick off recovery! ;-)
2139 * So, do a pre-check on all living nodes to see if any of those nodes
2140 * think that $RECOVERY is currently mastered by a dead node. If so,
2141 * we wait a short time to allow that node to get notified by its own
2142 * heartbeat stack, then check again. All $RECOVERY lock resources
2bd63216 2143 * mastered by dead nodes are purged when the hearbeat callback is
c03872f5
KH		 2144 * fired, so we can know for sure that it is safe to continue once
2145 * the node returns a live node or no node. */
2146static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2147 struct dlm_lock_resource *res)
2148{
2149 struct dlm_node_iter iter;
2150 int nodenum;
2151 int ret = 0;
2152 u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2153
2154 spin_lock(&dlm->spinlock);
2155 dlm_node_iter_init(dlm->domain_map, &iter);
2156 spin_unlock(&dlm->spinlock);
2157
2158 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2159 /* do not send to self */
2160 if (nodenum == dlm->node_num)
2161 continue;
2162 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2163 if (ret < 0) {
2164 mlog_errno(ret);
2165 if (!dlm_is_host_down(ret))
2166 BUG();
2167 /* host is down, so answer for that node would be
2168 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
f42a100b 2169 ret = 0;
c03872f5
KH		 2170 }
2171
2172 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2173 /* check to see if this master is in the recovery map */
2174 spin_lock(&dlm->spinlock);
2175 if (test_bit(master, dlm->recovery_map)) {
2176 mlog(ML_NOTICE, "%s: node %u has not seen "
2177 "node %u go down yet, and thinks the "
2178 "dead node is mastering the recovery "
2179 "lock. must wait.\n", dlm->name,
2180 nodenum, master);
2181 ret = -EAGAIN;
2182 }
2183 spin_unlock(&dlm->spinlock);
2bd63216 2184 mlog(0, "%s: reco lock master is %u\n", dlm->name,
c03872f5
KH		 2185 master);
2186 break;
2187 }
2188 }
2189 return ret;
2190}
2191
ba2bf218
KH		 2192/*
2193 * DLM_DEREF_LOCKRES_MSG
2194 */
2195
2196int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2197{
2198 struct dlm_deref_lockres deref;
2199 int ret = 0, r;
2200 const char *lockname;
2201 unsigned int namelen;
2202
2203 lockname = res->lockname.name;
2204 namelen = res->lockname.len;
2205 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2206
ba2bf218
KH		 2207 memset(&deref, 0, sizeof(deref));
2208 deref.node_idx = dlm->node_num;
2209 deref.namelen = namelen;
2210 memcpy(deref.name, lockname, namelen);
2211
2212 ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2213 &deref, sizeof(deref), res->owner, &r);
2214 if (ret < 0)
8decab3c
SM		 2215 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n",
2216 dlm->name, namelen, lockname, ret, res->owner);
ba2bf218
KH		 2217 else if (r < 0) {
2218 /* BAD. other node says I did not have a ref. */
8decab3c
SM		 2219 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2220 dlm->name, namelen, lockname, res->owner, r);
ba2bf218
KH		 2221 dlm_print_one_lock_resource(res);
2222 BUG();
2223 }
2224 return ret;
2225}
2226
d74c9803
KH		 2227int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2228 void **ret_data)
ba2bf218
KH		 2229{
2230 struct dlm_ctxt *dlm = data;
2231 struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2232 struct dlm_lock_resource *res = NULL;
2233 char *name;
2234 unsigned int namelen;
2235 int ret = -EINVAL;
2236 u8 node;
2237 unsigned int hash;
f3f85464
SM		 2238 struct dlm_work_item *item;
2239 int cleared = 0;
2240 int dispatch = 0;
ba2bf218
KH		 2241
2242 if (!dlm_grab(dlm))
2243 return 0;
2244
2245 name = deref->name;
2246 namelen = deref->namelen;
2247 node = deref->node_idx;
2248
2249 if (namelen > DLM_LOCKID_NAME_MAX) {
2250 mlog(ML_ERROR, "Invalid name length!");
2251 goto done;
2252 }
2253 if (deref->node_idx >= O2NM_MAX_NODES) {
2254 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2255 goto done;
2256 }
2257
2258 hash = dlm_lockid_hash(name, namelen);
2259
2260 spin_lock(&dlm->spinlock);
2261 res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2262 if (!res) {
2263 spin_unlock(&dlm->spinlock);
2264 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2265 dlm->name, namelen, name);
2266 goto done;
2267 }
2268 spin_unlock(&dlm->spinlock);
2269
2270 spin_lock(&res->spinlock);
f3f85464
SM		 2271 if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2272 dispatch = 1;
2273 else {
2274 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2275 if (test_bit(node, res->refmap)) {
8d400b81 2276 dlm_lockres_clear_refmap_bit(dlm, res, node);
f3f85464
SM		 2277 cleared = 1;
2278 }
ba2bf218
KH		 2279 }
2280 spin_unlock(&res->spinlock);
2281
f3f85464
SM		 2282 if (!dispatch) {
2283 if (cleared)
2284 dlm_lockres_calc_usage(dlm, res);
2285 else {
2286 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2287 "but it is already dropped!\n", dlm->name,
2288 res->lockname.len, res->lockname.name, node);
2af37ce8 2289 dlm_print_one_lock_resource(res);
f3f85464
SM		 2290 }
2291 ret = 0;
2292 goto done;
2293 }
2294
2295 item = kzalloc(sizeof(*item), GFP_NOFS);
2296 if (!item) {
2297 ret = -ENOMEM;
2298 mlog_errno(ret);
2299 goto done;
2300 }
2301
2302 dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2303 item->u.dl.deref_res = res;
2304 item->u.dl.deref_node = node;
2305
2306 spin_lock(&dlm->work_lock);
2307 list_add_tail(&item->list, &dlm->work_list);
2308 spin_unlock(&dlm->work_lock);
2309
2310 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2311 return 0;
2312
ba2bf218
KH		 2313done:
2314 if (res)
2315 dlm_lockres_put(res);
2316 dlm_put(dlm);
f3f85464 2317
ba2bf218
KH		 2318 return ret;
2319}
2320
f3f85464
SM		 2321static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2322{
2323 struct dlm_ctxt *dlm;
2324 struct dlm_lock_resource *res;
2325 u8 node;
2326 u8 cleared = 0;
2327
2328 dlm = item->dlm;
2329 res = item->u.dl.deref_res;
2330 node = item->u.dl.deref_node;
2331
2332 spin_lock(&res->spinlock);
2333 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2334 if (test_bit(node, res->refmap)) {
2335 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
8d400b81 2336 dlm_lockres_clear_refmap_bit(dlm, res, node);
f3f85464
SM		 2337 cleared = 1;
2338 }
2339 spin_unlock(&res->spinlock);
2340
2341 if (cleared) {
2342 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2343 dlm->name, res->lockname.len, res->lockname.name, node);
2344 dlm_lockres_calc_usage(dlm, res);
2345 } else {
2346 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2347 "but it is already dropped!\n", dlm->name,
2348 res->lockname.len, res->lockname.name, node);
2af37ce8 2349 dlm_print_one_lock_resource(res);
f3f85464
SM		 2350 }
2351
2352 dlm_lockres_put(res);
2353}
2354
9f62e960
SM		 2355/*
2356 * A migrateable resource is one that is :
2357 * 1. locally mastered, and,
2358 * 2. zero local locks, and,
2359 * 3. one or more non-local locks, or, one or more references
2360 * Returns 1 if yes, 0 if not.
2f5bf1f2
SM		 2361 */
2362static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
9f62e960 2363 struct dlm_lock_resource *res)
2f5bf1f2 2364{
9f62e960
SM		 2365 enum dlm_lockres_list idx;
2366 int nonlocal = 0, node_ref;
800deef3 2367 struct list_head *queue;
2f5bf1f2 2368 struct dlm_lock *lock;
9f62e960 2369 u64 cookie;
2f5bf1f2
SM		 2370
2371 assert_spin_locked(&res->spinlock);
2372
9f62e960
SM		 2373 if (res->owner != dlm->node_num)
2374 return 0;
2f5bf1f2 2375
9f62e960
SM		 2376 for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2377 queue = dlm_list_idx_to_ptr(res, idx);
800deef3 2378 list_for_each_entry(lock, queue, list) {
9f62e960
SM		 2379 if (lock->ml.node != dlm->node_num) {
2380 nonlocal++;
2381 continue;
2f5bf1f2 2382 }
9f62e960
SM		 2383 cookie = be64_to_cpu(lock->ml.cookie);
2384 mlog(0, "%s: Not migrateable res %.*s, lock %u:%llu on "
2385 "%s list\n", dlm->name, res->lockname.len,
2386 res->lockname.name,
2387 dlm_get_lock_cookie_node(cookie),
2388 dlm_get_lock_cookie_seq(cookie),
2389 dlm_list_in_text(idx));
2390 return 0;
2f5bf1f2 2391 }
2f5bf1f2
SM		 2392 }
2393
9f62e960
SM		 2394 if (!nonlocal) {
2395 node_ref = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
2396 if (node_ref >= O2NM_MAX_NODES)
2397 return 0;
2398 }
388c4bcb 2399
9f62e960
SM		 2400 mlog(0, "%s: res %.*s, Migrateable\n", dlm->name, res->lockname.len,
2401 res->lockname.name);
2f5bf1f2 2402
9f62e960 2403 return 1;
2f5bf1f2 2404}
6714d8e8
KH		 2405
2406/*
2407 * DLM_MIGRATE_LOCKRES
2408 */
2409
2410
faf0ec9f 2411static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
66effd3c 2412 struct dlm_lock_resource *res, u8 target)
6714d8e8
KH		 2413{
2414 struct dlm_master_list_entry *mle = NULL;
2415 struct dlm_master_list_entry *oldmle = NULL;
2416 struct dlm_migratable_lockres *mres = NULL;
2f5bf1f2 2417 int ret = 0;
6714d8e8
KH		 2418 const char *name;
2419 unsigned int namelen;
2420 int mle_added = 0;
2f5bf1f2 2421 int wake = 0;
6714d8e8
KH		 2422
2423 if (!dlm_grab(dlm))
2424 return -EINVAL;
2425
66effd3c
SM		 2426 BUG_ON(target == O2NM_MAX_NODES);
2427
6714d8e8
KH		 2428 name = res->lockname.name;
2429 namelen = res->lockname.len;
2430
66effd3c
SM		 2431 mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name,
2432 target);
6714d8e8 2433
66effd3c 2434 /* preallocate up front. if this fails, abort */
6714d8e8 2435 ret = -ENOMEM;
ad8100e0 2436 mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
6714d8e8
KH		 2437 if (!mres) {
2438 mlog_errno(ret);
2439 goto leave;
2440 }
2441
3914ed0c 2442 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
6714d8e8
KH		 2443 if (!mle) {
2444 mlog_errno(ret);
2445 goto leave;
2446 }
2447 ret = 0;
2448
6714d8e8
KH		 2449 /*
2450 * clear any existing master requests and
2451 * add the migration mle to the list
2452 */
66effd3c 2453 spin_lock(&dlm->spinlock);
6714d8e8
KH		 2454 spin_lock(&dlm->master_lock);
2455 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2456 namelen, target, dlm->node_num);
2457 spin_unlock(&dlm->master_lock);
2458 spin_unlock(&dlm->spinlock);
2459
2460 if (ret == -EEXIST) {
2461 mlog(0, "another process is already migrating it\n");
2462 goto fail;
2463 }
2464 mle_added = 1;
2465
2466 /*
2467 * set the MIGRATING flag and flush asts
2468 * if we fail after this we need to re-dirty the lockres
2469 */
2470 if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2471 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2472 "the target went down.\n", res->lockname.len,
2473 res->lockname.name, target);
2474 spin_lock(&res->spinlock);
2475 res->state &= ~DLM_LOCK_RES_MIGRATING;
a6fa3640 2476 wake = 1;
6714d8e8
KH		 2477 spin_unlock(&res->spinlock);
2478 ret = -EINVAL;
2479 }
2480
2481fail:
2482 if (oldmle) {
2483 /* master is known, detach if not already detached */
2484 dlm_mle_detach_hb_events(dlm, oldmle);
2485 dlm_put_mle(oldmle);
2486 }
2487
2488 if (ret < 0) {
2489 if (mle_added) {
2490 dlm_mle_detach_hb_events(dlm, mle);
2491 dlm_put_mle(mle);
2492 } else if (mle) {
2493 kmem_cache_free(dlm_mle_cache, mle);
66effd3c 2494 mle = NULL;
6714d8e8
KH		 2495 }
2496 goto leave;
2497 }
2498
2499 /*
2500 * at this point, we have a migration target, an mle
2501 * in the master list, and the MIGRATING flag set on
2502 * the lockres
2503 */
2504
1cd04dbe
KH		 2505 /* now that remote nodes are spinning on the MIGRATING flag,
2506 * ensure that all assert_master work is flushed. */
2507 flush_workqueue(dlm->dlm_worker);
6714d8e8
KH		 2508
2509 /* get an extra reference on the mle.
2510 * otherwise the assert_master from the new
2511 * master will destroy this.
2512 * also, make sure that all callers of dlm_get_mle
2513 * take both dlm->spinlock and dlm->master_lock */
2514 spin_lock(&dlm->spinlock);
2515 spin_lock(&dlm->master_lock);
a2bf0477 2516 dlm_get_mle_inuse(mle);
6714d8e8
KH		 2517 spin_unlock(&dlm->master_lock);
2518 spin_unlock(&dlm->spinlock);
2519
2520 /* notify new node and send all lock state */
2521 /* call send_one_lockres with migration flag.
2522 * this serves as notice to the target node that a
2523 * migration is starting. */
2524 ret = dlm_send_one_lockres(dlm, res, mres, target,
2525 DLM_MRES_MIGRATION);
2526
2527 if (ret < 0) {
2528 mlog(0, "migration to node %u failed with %d\n",
2529 target, ret);
2530 /* migration failed, detach and clean up mle */
2531 dlm_mle_detach_hb_events(dlm, mle);
2532 dlm_put_mle(mle);
a2bf0477
KH		 2533 dlm_put_mle_inuse(mle);
2534 spin_lock(&res->spinlock);
2535 res->state &= ~DLM_LOCK_RES_MIGRATING;
a6fa3640 2536 wake = 1;
a2bf0477 2537 spin_unlock(&res->spinlock);
df016c66
SM		 2538 if (dlm_is_host_down(ret))
2539 dlm_wait_for_node_death(dlm, target,
2540 DLM_NODE_DEATH_WAIT_MAX);
6714d8e8
KH		 2541 goto leave;
2542 }
2543
2544 /* at this point, the target sends a message to all nodes,
2545 * (using dlm_do_migrate_request). this node is skipped since
2546 * we had to put an mle in the list to begin the process. this
2547 * node now waits for target to do an assert master. this node
2548 * will be the last one notified, ensuring that the migration
2549 * is complete everywhere. if the target dies while this is
2550 * going on, some nodes could potentially see the target as the
2551 * master, so it is important that my recovery finds the migration
af901ca1 2552 * mle and sets the master to UNKNOWN. */
6714d8e8
KH		 2553
2554
2555 /* wait for new node to assert master */
2556 while (1) {
2557 ret = wait_event_interruptible_timeout(mle->wq,
2558 (atomic_read(&mle->woken) == 1),
2559 msecs_to_jiffies(5000));
2560
2561 if (ret >= 0) {
2562 if (atomic_read(&mle->woken) == 1 ||
2563 res->owner == target)
2564 break;
2565
1cd04dbe
KH		 2566 mlog(0, "%s:%.*s: timed out during migration\n",
2567 dlm->name, res->lockname.len, res->lockname.name);
2bd63216 2568 /* avoid hang during shutdown when migrating lockres
e2faea4c
KH		 2569 * to a node which also goes down */
2570 if (dlm_is_node_dead(dlm, target)) {
aa852354
KH		 2571 mlog(0, "%s:%.*s: expected migration "
2572 "target %u is no longer up, restarting\n",
e2faea4c
KH		 2573 dlm->name, res->lockname.len,
2574 res->lockname.name, target);
1cd04dbe
KH		 2575 ret = -EINVAL;
2576 /* migration failed, detach and clean up mle */
2577 dlm_mle_detach_hb_events(dlm, mle);
2578 dlm_put_mle(mle);
2579 dlm_put_mle_inuse(mle);
2580 spin_lock(&res->spinlock);
2581 res->state &= ~DLM_LOCK_RES_MIGRATING;
a6fa3640 2582 wake = 1;
1cd04dbe
KH		 2583 spin_unlock(&res->spinlock);
2584 goto leave;
e2faea4c 2585 }
1cd04dbe
KH		 2586 } else
2587 mlog(0, "%s:%.*s: caught signal during migration\n",
2588 dlm->name, res->lockname.len, res->lockname.name);
6714d8e8
KH		 2589 }
2590
2591 /* all done, set the owner, clear the flag */
2592 spin_lock(&res->spinlock);
2593 dlm_set_lockres_owner(dlm, res, target);
2594 res->state &= ~DLM_LOCK_RES_MIGRATING;
2595 dlm_remove_nonlocal_locks(dlm, res);
2596 spin_unlock(&res->spinlock);
2597 wake_up(&res->wq);
2598
2599 /* master is known, detach if not already detached */
2600 dlm_mle_detach_hb_events(dlm, mle);
a2bf0477 2601 dlm_put_mle_inuse(mle);
6714d8e8
KH		 2602 ret = 0;
2603
2604 dlm_lockres_calc_usage(dlm, res);
2605
2606leave:
2607 /* re-dirty the lockres if we failed */
2608 if (ret < 0)
2609 dlm_kick_thread(dlm, res);
2610
a6fa3640
KH		 2611 /* wake up waiters if the MIGRATING flag got set
2612 * but migration failed */
2613 if (wake)
2614 wake_up(&res->wq);
2615
6714d8e8
KH		 2616 if (mres)
2617 free_page((unsigned long)mres);
2618
2619 dlm_put(dlm);
2620
9f62e960
SM		 2621 mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen,
2622 name, target, ret);
6714d8e8
KH		 2623 return ret;
2624}
6714d8e8 2625
ba2bf218
KH		 2626#define DLM_MIGRATION_RETRY_MS 100
2627
9f62e960
SM		 2628/*
2629 * Should be called only after beginning the domain leave process.
ba2bf218
KH		 2630 * There should not be any remaining locks on nonlocal lock resources,
2631 * and there should be no local locks left on locally mastered resources.
2632 *
2633 * Called with the dlm spinlock held, may drop it to do migration, but
2634 * will re-acquire before exit.
2635 *
9f62e960
SM		 2636 * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2637 */
ba2bf218
KH		 2638int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2639{
66effd3c 2640 int ret;
ba2bf218 2641 int lock_dropped = 0;
66effd3c 2642 u8 target = O2NM_MAX_NODES;
ba2bf218 2643
9f62e960 2644 assert_spin_locked(&dlm->spinlock);
2f5bf1f2 2645
9f62e960 2646 spin_lock(&res->spinlock);
66effd3c
SM		 2647 if (dlm_is_lockres_migrateable(dlm, res))
2648 target = dlm_pick_migration_target(dlm, res);
b36c3f84 2649 spin_unlock(&res->spinlock);
66effd3c
SM		 2650
2651 if (target == O2NM_MAX_NODES)
9f62e960 2652 goto leave;
ba2bf218
KH		 2653
2654 /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2655 spin_unlock(&dlm->spinlock);
2656 lock_dropped = 1;
66effd3c
SM		 2657 ret = dlm_migrate_lockres(dlm, res, target);
2658 if (ret)
2659 mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2660 dlm->name, res->lockname.len, res->lockname.name,
2661 target, ret);
ba2bf218
KH		 2662 spin_lock(&dlm->spinlock);
2663leave:
2664 return lock_dropped;
2665}
2666
6714d8e8
KH		 2667int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2668{
2669 int ret;
2670 spin_lock(&dlm->ast_lock);
2671 spin_lock(&lock->spinlock);
2672 ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2673 spin_unlock(&lock->spinlock);
2674 spin_unlock(&dlm->ast_lock);
2675 return ret;
2676}
2677
2678static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2679 struct dlm_lock_resource *res,
2680 u8 mig_target)
2681{
2682 int can_proceed;
2683 spin_lock(&res->spinlock);
2684 can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2685 spin_unlock(&res->spinlock);
2686
2bd63216 2687 /* target has died, so make the caller break out of the
6714d8e8
KH		 2688 * wait_event, but caller must recheck the domain_map */
2689 spin_lock(&dlm->spinlock);
2690 if (!test_bit(mig_target, dlm->domain_map))
2691 can_proceed = 1;
2692 spin_unlock(&dlm->spinlock);
2693 return can_proceed;
2694}
2695
faf0ec9f
AB		 2696static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2697 struct dlm_lock_resource *res)
6714d8e8
KH		 2698{
2699 int ret;
2700 spin_lock(&res->spinlock);
2701 ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2702 spin_unlock(&res->spinlock);
2703 return ret;
2704}
2705
2706
2707static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2708 struct dlm_lock_resource *res,
2709 u8 target)
2710{
2711 int ret = 0;
2712
2713 mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2714 res->lockname.len, res->lockname.name, dlm->node_num,
2715 target);
2716 /* need to set MIGRATING flag on lockres. this is done by
2717 * ensuring that all asts have been flushed for this lockres. */
2718 spin_lock(&res->spinlock);
2719 BUG_ON(res->migration_pending);
2720 res->migration_pending = 1;
2721 /* strategy is to reserve an extra ast then release
2722 * it below, letting the release do all of the work */
2723 __dlm_lockres_reserve_ast(res);
2724 spin_unlock(&res->spinlock);
2725
ddc09c8d 2726 /* now flush all the pending asts */
6714d8e8 2727 dlm_kick_thread(dlm, res);
ddc09c8d
KH		 2728 /* before waiting on DIRTY, block processes which may
2729 * try to dirty the lockres before MIGRATING is set */
2730 spin_lock(&res->spinlock);
2731 BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2732 res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2733 spin_unlock(&res->spinlock);
2734 /* now wait on any pending asts and the DIRTY state */
6714d8e8
KH		 2735 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2736 dlm_lockres_release_ast(dlm, res);
2737
2738 mlog(0, "about to wait on migration_wq, dirty=%s\n",
2739 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2740 /* if the extra ref we just put was the final one, this
2741 * will pass thru immediately. otherwise, we need to wait
2742 * for the last ast to finish. */
2743again:
2744 ret = wait_event_interruptible_timeout(dlm->migration_wq,
2745 dlm_migration_can_proceed(dlm, res, target),
2746 msecs_to_jiffies(1000));
2747 if (ret < 0) {
2748 mlog(0, "woken again: migrating? %s, dead? %s\n",
2749 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2750 test_bit(target, dlm->domain_map) ? "no":"yes");
2751 } else {
2752 mlog(0, "all is well: migrating? %s, dead? %s\n",
2753 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2754 test_bit(target, dlm->domain_map) ? "no":"yes");
2755 }
2756 if (!dlm_migration_can_proceed(dlm, res, target)) {
2757 mlog(0, "trying again...\n");
2758 goto again;
2759 }
2760
a39953dd 2761 ret = 0;
6714d8e8
KH		 2762 /* did the target go down or die? */
2763 spin_lock(&dlm->spinlock);
2764 if (!test_bit(target, dlm->domain_map)) {
2765 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2766 target);
2767 ret = -EHOSTDOWN;
2768 }
2769 spin_unlock(&dlm->spinlock);
2770
a39953dd
WW		 2771 /*
2772 * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2773 * another try; otherwise, we are sure the MIGRATING state is there,
2774 * drop the unneded state which blocked threads trying to DIRTY
2775 */
2776 spin_lock(&res->spinlock);
2777 BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2778 res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2779 if (!ret)
2780 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2781 spin_unlock(&res->spinlock);
2782
6714d8e8
KH		 2783 /*
2784 * at this point:
2785 *
a39953dd 2786 * o the DLM_LOCK_RES_MIGRATING flag is set if target not down
6714d8e8
KH		 2787 * o there are no pending asts on this lockres
2788 * o all processes trying to reserve an ast on this
2789 * lockres must wait for the MIGRATING flag to clear
2790 */
2791 return ret;
2792}
2793
2794/* last step in the migration process.
2795 * original master calls this to free all of the dlm_lock
2796 * structures that used to be for other nodes. */
2797static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2798 struct dlm_lock_resource *res)
2799{
6714d8e8 2800 struct list_head *queue = &res->granted;
ba2bf218 2801 int i, bit;
800deef3 2802 struct dlm_lock *lock, *next;
6714d8e8
KH		 2803
2804 assert_spin_locked(&res->spinlock);
2805
2806 BUG_ON(res->owner == dlm->node_num);
2807
2808 for (i=0; i<3; i++) {
800deef3 2809 list_for_each_entry_safe(lock, next, queue, list) {
6714d8e8
KH		 2810 if (lock->ml.node != dlm->node_num) {
2811 mlog(0, "putting lock for node %u\n",
2812 lock->ml.node);
2813 /* be extra careful */
2814 BUG_ON(!list_empty(&lock->ast_list));
2815 BUG_ON(!list_empty(&lock->bast_list));
2816 BUG_ON(lock->ast_pending);
2817 BUG_ON(lock->bast_pending);
8d400b81
SM		 2818 dlm_lockres_clear_refmap_bit(dlm, res,
2819 lock->ml.node);
6714d8e8
KH		 2820 list_del_init(&lock->list);
2821 dlm_lock_put(lock);
2c5c54ac
SM		 2822 /* In a normal unlock, we would have added a
2823 * DLM_UNLOCK_FREE_LOCK action. Force it. */
2824 dlm_lock_put(lock);
6714d8e8
KH		 2825 }
2826 }
2827 queue++;
2828 }
ba2bf218
KH		 2829 bit = 0;
2830 while (1) {
2831 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2832 if (bit >= O2NM_MAX_NODES)
2833 break;
2834 /* do not clear the local node reference, if there is a
2835 * process holding this, let it drop the ref itself */
2836 if (bit != dlm->node_num) {
2837 mlog(0, "%s:%.*s: node %u had a ref to this "
2838 "migrating lockres, clearing\n", dlm->name,
2839 res->lockname.len, res->lockname.name, bit);
8d400b81 2840 dlm_lockres_clear_refmap_bit(dlm, res, bit);
ba2bf218
KH		 2841 }
2842 bit++;
2843 }
6714d8e8
KH		 2844}
2845
66effd3c
SM		 2846/*
2847 * Pick a node to migrate the lock resource to. This function selects a
2848 * potential target based first on the locks and then on refmap. It skips
2849 * nodes that are in the process of exiting the domain.
2850 */
6714d8e8
KH		 2851static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2852 struct dlm_lock_resource *res)
2853{
66effd3c 2854 enum dlm_lockres_list idx;
6714d8e8 2855 struct list_head *queue = &res->granted;
6714d8e8 2856 struct dlm_lock *lock;
66effd3c
SM		 2857 int noderef;
2858 u8 nodenum = O2NM_MAX_NODES;
6714d8e8
KH		 2859
2860 assert_spin_locked(&dlm->spinlock);
66effd3c 2861 assert_spin_locked(&res->spinlock);
6714d8e8 2862
66effd3c
SM		 2863 /* Go through all the locks */
2864 for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2865 queue = dlm_list_idx_to_ptr(res, idx);
800deef3 2866 list_for_each_entry(lock, queue, list) {
66effd3c
SM		 2867 if (lock->ml.node == dlm->node_num)
2868 continue;
2869 if (test_bit(lock->ml.node, dlm->exit_domain_map))
2870 continue;
2871 nodenum = lock->ml.node;
2872 goto bail;
6714d8e8 2873 }
6714d8e8 2874 }
388c4bcb 2875
66effd3c
SM		 2876 /* Go thru the refmap */
2877 noderef = -1;
6714d8e8 2878 while (1) {
66effd3c
SM		 2879 noderef = find_next_bit(res->refmap, O2NM_MAX_NODES,
2880 noderef + 1);
2881 if (noderef >= O2NM_MAX_NODES)
6714d8e8 2882 break;
66effd3c
SM		 2883 if (noderef == dlm->node_num)
2884 continue;
2885 if (test_bit(noderef, dlm->exit_domain_map))
2886 continue;
2887 nodenum = noderef;
2888 goto bail;
6714d8e8
KH		 2889 }
2890
66effd3c
SM		 2891bail:
2892 return nodenum;
6714d8e8
KH		 2893}
2894
6714d8e8
KH		 2895/* this is called by the new master once all lockres
2896 * data has been received */
2897static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2898 struct dlm_lock_resource *res,
2899 u8 master, u8 new_master,
2900 struct dlm_node_iter *iter)
2901{
2902 struct dlm_migrate_request migrate;
2b832564 2903 int ret, skip, status = 0;
6714d8e8
KH		 2904 int nodenum;
2905
2906 memset(&migrate, 0, sizeof(migrate));
2907 migrate.namelen = res->lockname.len;
2908 memcpy(migrate.name, res->lockname.name, migrate.namelen);
2909 migrate.new_master = new_master;
2910 migrate.master = master;
2911
2912 ret = 0;
2913
2914 /* send message to all nodes, except the master and myself */
2915 while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2916 if (nodenum == master ||
2917 nodenum == new_master)
2918 continue;
2919
2b832564
SM		 2920 /* We could race exit domain. If exited, skip. */
2921 spin_lock(&dlm->spinlock);
2922 skip = (!test_bit(nodenum, dlm->domain_map));
2923 spin_unlock(&dlm->spinlock);
2924 if (skip) {
2925 clear_bit(nodenum, iter->node_map);
2926 continue;
2927 }
2928
6714d8e8
KH		 2929 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2930 &migrate, sizeof(migrate), nodenum,
2931 &status);
2b832564 2932 if (ret < 0) {
8decab3c
SM		 2933 mlog(ML_ERROR, "%s: res %.*s, Error %d send "
2934 "MIGRATE_REQUEST to node %u\n", dlm->name,
2935 migrate.namelen, migrate.name, ret, nodenum);
2b832564
SM		 2936 if (!dlm_is_host_down(ret)) {
2937 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
2938 BUG();
2939 }
2940 clear_bit(nodenum, iter->node_map);
2941 ret = 0;
2942 } else if (status < 0) {
6714d8e8
KH		 2943 mlog(0, "migrate request (node %u) returned %d!\n",
2944 nodenum, status);
2945 ret = status;
ba2bf218
KH		 2946 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
2947 /* during the migration request we short-circuited
2948 * the mastery of the lockres. make sure we have
2949 * a mastery ref for nodenum */
2950 mlog(0, "%s:%.*s: need ref for node %u\n",
2951 dlm->name, res->lockname.len, res->lockname.name,
2952 nodenum);
2953 spin_lock(&res->spinlock);
8d400b81 2954 dlm_lockres_set_refmap_bit(dlm, res, nodenum);
ba2bf218 2955 spin_unlock(&res->spinlock);
6714d8e8
KH		 2956 }
2957 }
2958
2959 if (ret < 0)
2960 mlog_errno(ret);
2961
2962 mlog(0, "returning ret=%d\n", ret);
2963 return ret;
2964}
2965
2966
2967/* if there is an existing mle for this lockres, we now know who the master is.
2968 * (the one who sent us *this* message) we can clear it up right away.
2969 * since the process that put the mle on the list still has a reference to it,
2970 * we can unhash it now, set the master and wake the process. as a result,
2971 * we will have no mle in the list to start with. now we can add an mle for
2972 * the migration and this should be the only one found for those scanning the
2973 * list. */
d74c9803
KH		 2974int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
2975 void **ret_data)
6714d8e8
KH		 2976{
2977 struct dlm_ctxt *dlm = data;
2978 struct dlm_lock_resource *res = NULL;
2979 struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
2980 struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
2981 const char *name;
a3d33291 2982 unsigned int namelen, hash;
6714d8e8
KH		 2983 int ret = 0;
2984
2985 if (!dlm_grab(dlm))
2986 return -EINVAL;
2987
2988 name = migrate->name;
2989 namelen = migrate->namelen;
a3d33291 2990 hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 2991
2992 /* preallocate.. if this fails, abort */
3914ed0c 2993 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
6714d8e8
KH		 2994
2995 if (!mle) {
2996 ret = -ENOMEM;
2997 goto leave;
2998 }
2999
3000 /* check for pre-existing lock */
3001 spin_lock(&dlm->spinlock);
a3d33291 3002 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
6714d8e8
KH		 3003 if (res) {
3004 spin_lock(&res->spinlock);
3005 if (res->state & DLM_LOCK_RES_RECOVERING) {
3006 /* if all is working ok, this can only mean that we got
3007 * a migrate request from a node that we now see as
3008 * dead. what can we do here? drop it to the floor? */
3009 spin_unlock(&res->spinlock);
3010 mlog(ML_ERROR, "Got a migrate request, but the "
3011 "lockres is marked as recovering!");
3012 kmem_cache_free(dlm_mle_cache, mle);
3013 ret = -EINVAL; /* need a better solution */
3014 goto unlock;
3015 }
3016 res->state |= DLM_LOCK_RES_MIGRATING;
3017 spin_unlock(&res->spinlock);
3018 }
3019
6d98c3cc 3020 spin_lock(&dlm->master_lock);
6714d8e8
KH		 3021 /* ignore status. only nonzero status would BUG. */
3022 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3023 name, namelen,
3024 migrate->new_master,
3025 migrate->master);
3026
6714d8e8 3027 spin_unlock(&dlm->master_lock);
6d98c3cc 3028unlock:
6714d8e8
KH		 3029 spin_unlock(&dlm->spinlock);
3030
3031 if (oldmle) {
3032 /* master is known, detach if not already detached */
3033 dlm_mle_detach_hb_events(dlm, oldmle);
3034 dlm_put_mle(oldmle);
3035 }
3036
3037 if (res)
3038 dlm_lockres_put(res);
3039leave:
3040 dlm_put(dlm);
3041 return ret;
3042}
3043
3044/* must be holding dlm->spinlock and dlm->master_lock
3045 * when adding a migration mle, we can clear any other mles
3046 * in the master list because we know with certainty that
3047 * the master is "master". so we remove any old mle from
3048 * the list after setting it's master field, and then add
3049 * the new migration mle. this way we can hold with the rule
3050 * of having only one mle for a given lock name at all times. */
3051static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3052 struct dlm_lock_resource *res,
3053 struct dlm_master_list_entry *mle,
3054 struct dlm_master_list_entry **oldmle,
3055 const char *name, unsigned int namelen,
3056 u8 new_master, u8 master)
3057{
3058 int found;
3059 int ret = 0;
3060
3061 *oldmle = NULL;
3062
6714d8e8
KH		 3063 assert_spin_locked(&dlm->spinlock);
3064 assert_spin_locked(&dlm->master_lock);
3065
3066 /* caller is responsible for any ref taken here on oldmle */
3067 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3068 if (found) {
3069 struct dlm_master_list_entry *tmp = *oldmle;
3070 spin_lock(&tmp->spinlock);
3071 if (tmp->type == DLM_MLE_MIGRATION) {
3072 if (master == dlm->node_num) {
3073 /* ah another process raced me to it */
3074 mlog(0, "tried to migrate %.*s, but some "
3075 "process beat me to it\n",
3076 namelen, name);
3077 ret = -EEXIST;
3078 } else {
3079 /* bad. 2 NODES are trying to migrate! */
3080 mlog(ML_ERROR, "migration error mle: "
3081 "master=%u new_master=%u // request: "
3082 "master=%u new_master=%u // "
3083 "lockres=%.*s\n",
3084 tmp->master, tmp->new_master,
3085 master, new_master,
3086 namelen, name);
3087 BUG();
3088 }
3089 } else {
3090 /* this is essentially what assert_master does */
3091 tmp->master = master;
3092 atomic_set(&tmp->woken, 1);
3093 wake_up(&tmp->wq);
1c084577
SM		 3094 /* remove it so that only one mle will be found */
3095 __dlm_unlink_mle(dlm, tmp);
ba2bf218
KH		 3096 __dlm_mle_detach_hb_events(dlm, tmp);
3097 ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3098 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3099 "telling master to get ref for cleared out mle "
3100 "during migration\n", dlm->name, namelen, name,
3101 master, new_master);
6714d8e8
KH		 3102 }
3103 spin_unlock(&tmp->spinlock);
3104 }
3105
3106 /* now add a migration mle to the tail of the list */
3107 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3108 mle->new_master = new_master;
ba2bf218
KH		 3109 /* the new master will be sending an assert master for this.
3110 * at that point we will get the refmap reference */
6714d8e8
KH		 3111 mle->master = master;
3112 /* do this for consistency with other mle types */
3113 set_bit(new_master, mle->maybe_map);
1c084577 3114 __dlm_insert_mle(dlm, mle);
6714d8e8
KH		 3115
3116 return ret;
3117}
3118
c2cd4a44
SM		 3119/*
3120 * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3121 */
3122static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3123 struct dlm_master_list_entry *mle)
3124{
3125 struct dlm_lock_resource *res;
c2cd4a44
SM		 3126
3127 /* Find the lockres associated to the mle and set its owner to UNK */
7141514b
SM		 3128 res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3129 mle->mnamehash);
c2cd4a44
SM		 3130 if (res) {
3131 spin_unlock(&dlm->master_lock);
3132
3133 /* move lockres onto recovery list */
3134 spin_lock(&res->spinlock);
3135 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3136 dlm_move_lockres_to_recovery_list(dlm, res);
3137 spin_unlock(&res->spinlock);
3138 dlm_lockres_put(res);
3139
3140 /* about to get rid of mle, detach from heartbeat */
3141 __dlm_mle_detach_hb_events(dlm, mle);
3142
3143 /* dump the mle */
3144 spin_lock(&dlm->master_lock);
3145 __dlm_put_mle(mle);
3146 spin_unlock(&dlm->master_lock);
3147 }
3148
3149 return res;
3150}
3151
3152static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3153 struct dlm_master_list_entry *mle)
3154{
3155 __dlm_mle_detach_hb_events(dlm, mle);
3156
3157 spin_lock(&mle->spinlock);
3158 __dlm_unlink_mle(dlm, mle);
3159 atomic_set(&mle->woken, 1);
3160 spin_unlock(&mle->spinlock);
3161
3162 wake_up(&mle->wq);
3163}
3164
3165static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3166 struct dlm_master_list_entry *mle, u8 dead_node)
3167{
3168 int bit;
3169
3170 BUG_ON(mle->type != DLM_MLE_BLOCK);
3171
3172 spin_lock(&mle->spinlock);
3173 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3174 if (bit != dead_node) {
3175 mlog(0, "mle found, but dead node %u would not have been "
3176 "master\n", dead_node);
3177 spin_unlock(&mle->spinlock);
3178 } else {
3179 /* Must drop the refcount by one since the assert_master will
3180 * never arrive. This may result in the mle being unlinked and
3181 * freed, but there may still be a process waiting in the
3182 * dlmlock path which is fine. */
3183 mlog(0, "node %u was expected master\n", dead_node);
3184 atomic_set(&mle->woken, 1);
3185 spin_unlock(&mle->spinlock);
3186 wake_up(&mle->wq);
3187
3188 /* Do not need events any longer, so detach from heartbeat */
3189 __dlm_mle_detach_hb_events(dlm, mle);
3190 __dlm_put_mle(mle);
3191 }
3192}
6714d8e8
KH		 3193
3194void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3195{
2ed6c750 3196 struct dlm_master_list_entry *mle;
6714d8e8 3197 struct dlm_lock_resource *res;
2ed6c750
SM		 3198 struct hlist_head *bucket;
3199 struct hlist_node *list;
3200 unsigned int i;
6714d8e8 3201
ef6b689b 3202 mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
6714d8e8
KH		 3203top:
3204 assert_spin_locked(&dlm->spinlock);
3205
3206 /* clean the master list */
3207 spin_lock(&dlm->master_lock);
2ed6c750
SM		 3208 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3209 bucket = dlm_master_hash(dlm, i);
3210 hlist_for_each(list, bucket) {
3211 mle = hlist_entry(list, struct dlm_master_list_entry,
3212 master_hash_node);
3213
67ae1f06
SM		 3214 BUG_ON(mle->type != DLM_MLE_BLOCK &&
3215 mle->type != DLM_MLE_MASTER &&
3216 mle->type != DLM_MLE_MIGRATION);
3217
3218 /* MASTER mles are initiated locally. The waiting
3219 * process will notice the node map change shortly.
3220 * Let that happen as normal. */
3221 if (mle->type == DLM_MLE_MASTER)
3222 continue;
3223
3224 /* BLOCK mles are initiated by other nodes. Need to
3225 * clean up if the dead node would have been the
3226 * master. */
3227 if (mle->type == DLM_MLE_BLOCK) {
3228 dlm_clean_block_mle(dlm, mle, dead_node);
3229 continue;
3230 }
6714d8e8 3231
67ae1f06
SM		 3232 /* Everything else is a MIGRATION mle */
3233
3234 /* The rule for MIGRATION mles is that the master
3235 * becomes UNKNOWN if *either* the original or the new
3236 * master dies. All UNKNOWN lockres' are sent to
3237 * whichever node becomes the recovery master. The new
3238 * master is responsible for determining if there is
3239 * still a master for this lockres, or if he needs to
3240 * take over mastery. Either way, this node should
3241 * expect another message to resolve this. */
3242
3243 if (mle->master != dead_node &&
3244 mle->new_master != dead_node)
3245 continue;
3246
3247 /* If we have reached this point, this mle needs to be
3248 * removed from the list and freed. */
3249 dlm_clean_migration_mle(dlm, mle);
3250
3251 mlog(0, "%s: node %u died during migration from "
3252 "%u to %u!\n", dlm->name, dead_node, mle->master,
3253 mle->new_master);
3254
3255 /* If we find a lockres associated with the mle, we've
3256 * hit this rare case that messes up our lock ordering.
3257 * If so, we need to drop the master lock so that we can
3258 * take the lockres lock, meaning that we will have to
3259 * restart from the head of list. */
3260 res = dlm_reset_mleres_owner(dlm, mle);
3261 if (res)
3262 /* restart */
3263 goto top;
3264
3265 /* This may be the last reference */
3266 __dlm_put_mle(mle);
6714d8e8 3267 }
2ed6c750 3268 }
6714d8e8
KH		 3269 spin_unlock(&dlm->master_lock);
3270}
3271
6714d8e8
KH		 3272int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3273 u8 old_master)
3274{
3275 struct dlm_node_iter iter;
3276 int ret = 0;
3277
3278 spin_lock(&dlm->spinlock);
3279 dlm_node_iter_init(dlm->domain_map, &iter);
3280 clear_bit(old_master, iter.node_map);
3281 clear_bit(dlm->node_num, iter.node_map);
3282 spin_unlock(&dlm->spinlock);
3283
ba2bf218
KH		 3284 /* ownership of the lockres is changing. account for the
3285 * mastery reference here since old_master will briefly have
3286 * a reference after the migration completes */
3287 spin_lock(&res->spinlock);
8d400b81 3288 dlm_lockres_set_refmap_bit(dlm, res, old_master);
ba2bf218
KH		 3289 spin_unlock(&res->spinlock);
3290
6714d8e8
KH		 3291 mlog(0, "now time to do a migrate request to other nodes\n");
3292 ret = dlm_do_migrate_request(dlm, res, old_master,
3293 dlm->node_num, &iter);
3294 if (ret < 0) {
3295 mlog_errno(ret);
3296 goto leave;
3297 }
3298
3299 mlog(0, "doing assert master of %.*s to all except the original node\n",
3300 res->lockname.len, res->lockname.name);
3301 /* this call now finishes out the nodemap
3302 * even if one or more nodes die */
ba2bf218 3303 ret = dlm_do_assert_master(dlm, res, iter.node_map,
6714d8e8
KH		 3304 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3305 if (ret < 0) {
3306 /* no longer need to retry. all living nodes contacted. */
3307 mlog_errno(ret);
3308 ret = 0;
3309 }
3310
3311 memset(iter.node_map, 0, sizeof(iter.node_map));
3312 set_bit(old_master, iter.node_map);
3313 mlog(0, "doing assert master of %.*s back to %u\n",
3314 res->lockname.len, res->lockname.name, old_master);
ba2bf218 3315 ret = dlm_do_assert_master(dlm, res, iter.node_map,
6714d8e8
KH		 3316 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3317 if (ret < 0) {
3318 mlog(0, "assert master to original master failed "
3319 "with %d.\n", ret);
3320 /* the only nonzero status here would be because of
3321 * a dead original node. we're done. */
3322 ret = 0;
3323 }
3324
3325 /* all done, set the owner, clear the flag */
3326 spin_lock(&res->spinlock);
3327 dlm_set_lockres_owner(dlm, res, dlm->node_num);
3328 res->state &= ~DLM_LOCK_RES_MIGRATING;
3329 spin_unlock(&res->spinlock);
3330 /* re-dirty it on the new master */
3331 dlm_kick_thread(dlm, res);
3332 wake_up(&res->wq);
3333leave:
3334 return ret;
3335}
3336
3337/*
3338 * LOCKRES AST REFCOUNT
3339 * this is integral to migration
3340 */
3341
3342/* for future intent to call an ast, reserve one ahead of time.
3343 * this should be called only after waiting on the lockres
3344 * with dlm_wait_on_lockres, and while still holding the
3345 * spinlock after the call. */
3346void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3347{
3348 assert_spin_locked(&res->spinlock);
3349 if (res->state & DLM_LOCK_RES_MIGRATING) {
3350 __dlm_print_one_lock_resource(res);
3351 }
3352 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3353
3354 atomic_inc(&res->asts_reserved);
3355}
3356
3357/*
3358 * used to drop the reserved ast, either because it went unused,
3359 * or because the ast/bast was actually called.
3360 *
3361 * also, if there is a pending migration on this lockres,
3362 * and this was the last pending ast on the lockres,
3363 * atomically set the MIGRATING flag before we drop the lock.
3364 * this is how we ensure that migration can proceed with no
3365 * asts in progress. note that it is ok if the state of the
3366 * queues is such that a lock should be granted in the future
3367 * or that a bast should be fired, because the new master will
3368 * shuffle the lists on this lockres as soon as it is migrated.
3369 */
3370void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3371 struct dlm_lock_resource *res)
3372{
3373 if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3374 return;
3375
3376 if (!res->migration_pending) {
3377 spin_unlock(&res->spinlock);
3378 return;
3379 }
3380
3381 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3382 res->migration_pending = 0;
3383 res->state |= DLM_LOCK_RES_MIGRATING;
3384 spin_unlock(&res->spinlock);
3385 wake_up(&res->wq);
3386 wake_up(&dlm->migration_wq);
3387}
5dad6c39
SE		 3388
3389void dlm_force_free_mles(struct dlm_ctxt *dlm)
3390{
3391 int i;
3392 struct hlist_head *bucket;
3393 struct dlm_master_list_entry *mle;
3394 struct hlist_node *tmp, *list;
3395
3396 /*
3397 * We notified all other nodes that we are exiting the domain and
3398 * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3399 * around we force free them and wake any processes that are waiting
3400 * on the mles
3401 */
3402 spin_lock(&dlm->spinlock);
3403 spin_lock(&dlm->master_lock);
3404
3405 BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3406 BUG_ON((find_next_bit(dlm->domain_map, O2NM_MAX_NODES, 0) < O2NM_MAX_NODES));
3407
3408 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3409 bucket = dlm_master_hash(dlm, i);
3410 hlist_for_each_safe(list, tmp, bucket) {
3411 mle = hlist_entry(list, struct dlm_master_list_entry,
3412 master_hash_node);
3413 if (mle->type != DLM_MLE_BLOCK) {
3414 mlog(ML_ERROR, "bad mle: %p\n", mle);
3415 dlm_print_one_mle(mle);
3416 }
3417 atomic_set(&mle->woken, 1);
3418 wake_up(&mle->wq);
3419
3420 __dlm_unlink_mle(dlm, mle);
3421 __dlm_mle_detach_hb_events(dlm, mle);
3422 __dlm_put_mle(mle);
3423 }
3424 }
3425 spin_unlock(&dlm->master_lock);
3426 spin_unlock(&dlm->spinlock);
3427}
Linux 6.x block layer and io_uring tree(s)