projects / deliverable / linux.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
ocfs2: reflink: fix slow unlink for refcounted file
[deliverable/linux.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
1a5c4e2a
FF		 85static struct kmem_cache *dlm_lockres_cache;
86static struct kmem_cache *dlm_lockname_cache;
87static struct kmem_cache *dlm_mle_cache;
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 344 struct hlist_head *bucket;
2ed6c750 345 unsigned int hash;
6714d8e8
KH		 346
347 assert_spin_locked(&dlm->master_lock);
348
2ed6c750
SM		 349 hash = dlm_lockid_hash(name, namelen);
350 bucket = dlm_master_hash(dlm, hash);
df53cd3b 351 hlist_for_each_entry(tmpmle, bucket, master_hash_node) {
6714d8e8
KH		 352 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
353 continue;
354 dlm_get_mle(tmpmle);
355 *mle = tmpmle;
356 return 1;
357 }
358 return 0;
359}
360
361void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
362{
363 struct dlm_master_list_entry *mle;
6714d8e8
KH		 364
365 assert_spin_locked(&dlm->spinlock);
2bd63216 366
800deef3 367 list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
6714d8e8
KH		 368 if (node_up)
369 dlm_mle_node_up(dlm, mle, NULL, idx);
370 else
371 dlm_mle_node_down(dlm, mle, NULL, idx);
372 }
373}
374
375static void dlm_mle_node_down(struct dlm_ctxt *dlm,
376 struct dlm_master_list_entry *mle,
377 struct o2nm_node *node, int idx)
378{
379 spin_lock(&mle->spinlock);
380
381 if (!test_bit(idx, mle->node_map))
382 mlog(0, "node %u already removed from nodemap!\n", idx);
383 else
384 clear_bit(idx, mle->node_map);
385
386 spin_unlock(&mle->spinlock);
387}
388
389static void dlm_mle_node_up(struct dlm_ctxt *dlm,
390 struct dlm_master_list_entry *mle,
391 struct o2nm_node *node, int idx)
392{
393 spin_lock(&mle->spinlock);
394
395 if (test_bit(idx, mle->node_map))
396 mlog(0, "node %u already in node map!\n", idx);
397 else
398 set_bit(idx, mle->node_map);
399
400 spin_unlock(&mle->spinlock);
401}
402
403
404int dlm_init_mle_cache(void)
405{
12eb0035 406 dlm_mle_cache = kmem_cache_create("o2dlm_mle",
6714d8e8
KH		 407 sizeof(struct dlm_master_list_entry),
408 0, SLAB_HWCACHE_ALIGN,
20c2df83 409 NULL);
6714d8e8
KH		 410 if (dlm_mle_cache == NULL)
411 return -ENOMEM;
412 return 0;
413}
414
415void dlm_destroy_mle_cache(void)
416{
417 if (dlm_mle_cache)
418 kmem_cache_destroy(dlm_mle_cache);
419}
420
421static void dlm_mle_release(struct kref *kref)
422{
423 struct dlm_master_list_entry *mle;
424 struct dlm_ctxt *dlm;
425
6714d8e8
KH		 426 mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
427 dlm = mle->dlm;
428
6714d8e8
KH		 429 assert_spin_locked(&dlm->spinlock);
430 assert_spin_locked(&dlm->master_lock);
431
7141514b
SM		 432 mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
433 mle->type);
2ed6c750 434
6714d8e8 435 /* remove from list if not already */
1c084577 436 __dlm_unlink_mle(dlm, mle);
6714d8e8
KH		 437
438 /* detach the mle from the domain node up/down events */
439 __dlm_mle_detach_hb_events(dlm, mle);
440
2041d8fd
SM		 441 atomic_dec(&dlm->mle_cur_count[mle->type]);
442
6714d8e8
KH		 443 /* NOTE: kfree under spinlock here.
444 * if this is bad, we can move this to a freelist. */
445 kmem_cache_free(dlm_mle_cache, mle);
446}
447
448
449/*
450 * LOCK RESOURCE FUNCTIONS
451 */
452
724bdca9
SM		 453int dlm_init_master_caches(void)
454{
455 dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
456 sizeof(struct dlm_lock_resource),
457 0, SLAB_HWCACHE_ALIGN, NULL);
458 if (!dlm_lockres_cache)
459 goto bail;
460
461 dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
462 DLM_LOCKID_NAME_MAX, 0,
463 SLAB_HWCACHE_ALIGN, NULL);
464 if (!dlm_lockname_cache)
465 goto bail;
466
467 return 0;
468bail:
469 dlm_destroy_master_caches();
470 return -ENOMEM;
471}
472
473void dlm_destroy_master_caches(void)
474{
66db6cfd 475 if (dlm_lockname_cache) {
724bdca9 476 kmem_cache_destroy(dlm_lockname_cache);
66db6cfd
JQ		 477 dlm_lockname_cache = NULL;
478 }
724bdca9 479
66db6cfd 480 if (dlm_lockres_cache) {
724bdca9 481 kmem_cache_destroy(dlm_lockres_cache);
66db6cfd
JQ		 482 dlm_lockres_cache = NULL;
483 }
724bdca9
SM		 484}
485
6714d8e8
KH		 486static void dlm_lockres_release(struct kref *kref)
487{
488 struct dlm_lock_resource *res;
b0d4f817 489 struct dlm_ctxt *dlm;
6714d8e8
KH		 490
491 res = container_of(kref, struct dlm_lock_resource, refs);
b0d4f817 492 dlm = res->dlm;
6714d8e8
KH		 493
494 /* This should not happen -- all lockres' have a name
495 * associated with them at init time. */
496 BUG_ON(!res->lockname.name);
497
498 mlog(0, "destroying lockres %.*s\n", res->lockname.len,
499 res->lockname.name);
500
b0d4f817 501 spin_lock(&dlm->track_lock);
29576f8b
SM		 502 if (!list_empty(&res->tracking))
503 list_del_init(&res->tracking);
504 else {
505 mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n",
506 res->lockname.len, res->lockname.name);
507 dlm_print_one_lock_resource(res);
508 }
b0d4f817
SM		 509 spin_unlock(&dlm->track_lock);
510
6800791a
SM		 511 atomic_dec(&dlm->res_cur_count);
512
a7f90d83
KH		 513 if (!hlist_unhashed(&res->hash_node) ||
514 !list_empty(&res->granted) ||
515 !list_empty(&res->converting) ||
516 !list_empty(&res->blocked) ||
517 !list_empty(&res->dirty) ||
518 !list_empty(&res->recovering) ||
519 !list_empty(&res->purge)) {
520 mlog(ML_ERROR,
521 "Going to BUG for resource %.*s."
522 " We're on a list! [%c%c%c%c%c%c%c]\n",
523 res->lockname.len, res->lockname.name,
524 !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
525 !list_empty(&res->granted) ? 'G' : ' ',
526 !list_empty(&res->converting) ? 'C' : ' ',
527 !list_empty(&res->blocked) ? 'B' : ' ',
528 !list_empty(&res->dirty) ? 'D' : ' ',
529 !list_empty(&res->recovering) ? 'R' : ' ',
530 !list_empty(&res->purge) ? 'P' : ' ');
531
532 dlm_print_one_lock_resource(res);
533 }
534
6714d8e8
KH		 535 /* By the time we're ready to blow this guy away, we shouldn't
536 * be on any lists. */
81f2094a 537 BUG_ON(!hlist_unhashed(&res->hash_node));
6714d8e8
KH		 538 BUG_ON(!list_empty(&res->granted));
539 BUG_ON(!list_empty(&res->converting));
540 BUG_ON(!list_empty(&res->blocked));
541 BUG_ON(!list_empty(&res->dirty));
542 BUG_ON(!list_empty(&res->recovering));
543 BUG_ON(!list_empty(&res->purge));
544
724bdca9 545 kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
6714d8e8 546
724bdca9 547 kmem_cache_free(dlm_lockres_cache, res);
6714d8e8
KH		 548}
549
6714d8e8
KH		 550void dlm_lockres_put(struct dlm_lock_resource *res)
551{
552 kref_put(&res->refs, dlm_lockres_release);
553}
554
555static void dlm_init_lockres(struct dlm_ctxt *dlm,
556 struct dlm_lock_resource *res,
557 const char *name, unsigned int namelen)
558{
559 char *qname;
560
561 /* If we memset here, we lose our reference to the kmalloc'd
562 * res->lockname.name, so be sure to init every field
563 * correctly! */
564
565 qname = (char *) res->lockname.name;
566 memcpy(qname, name, namelen);
567
568 res->lockname.len = namelen;
a3d33291 569 res->lockname.hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 570
571 init_waitqueue_head(&res->wq);
572 spin_lock_init(&res->spinlock);
81f2094a 573 INIT_HLIST_NODE(&res->hash_node);
6714d8e8
KH		 574 INIT_LIST_HEAD(&res->granted);
575 INIT_LIST_HEAD(&res->converting);
576 INIT_LIST_HEAD(&res->blocked);
577 INIT_LIST_HEAD(&res->dirty);
578 INIT_LIST_HEAD(&res->recovering);
579 INIT_LIST_HEAD(&res->purge);
29576f8b 580 INIT_LIST_HEAD(&res->tracking);
6714d8e8
KH		 581 atomic_set(&res->asts_reserved, 0);
582 res->migration_pending = 0;
ba2bf218 583 res->inflight_locks = 0;
ac4fef4d 584 res->inflight_assert_workers = 0;
6714d8e8 585
b0d4f817
SM		 586 res->dlm = dlm;
587
6714d8e8
KH		 588 kref_init(&res->refs);
589
6800791a
SM		 590 atomic_inc(&dlm->res_tot_count);
591 atomic_inc(&dlm->res_cur_count);
592
6714d8e8
KH		 593 /* just for consistency */
594 spin_lock(&res->spinlock);
595 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
596 spin_unlock(&res->spinlock);
597
598 res->state = DLM_LOCK_RES_IN_PROGRESS;
599
600 res->last_used = 0;
601
18c6ac38 602 spin_lock(&dlm->spinlock);
29576f8b 603 list_add_tail(&res->tracking, &dlm->tracking_list);
18c6ac38 604 spin_unlock(&dlm->spinlock);
29576f8b 605
6714d8e8 606 memset(res->lvb, 0, DLM_LVB_LEN);
ba2bf218 607 memset(res->refmap, 0, sizeof(res->refmap));
6714d8e8
KH		 608}
609
610struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
611 const char *name,
612 unsigned int namelen)
613{
724bdca9 614 struct dlm_lock_resource *res = NULL;
6714d8e8 615
3914ed0c 616 res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
6714d8e8 617 if (!res)
724bdca9 618 goto error;
6714d8e8 619
3914ed0c 620 res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
724bdca9
SM		 621 if (!res->lockname.name)
622 goto error;
6714d8e8
KH		 623
624 dlm_init_lockres(dlm, res, name, namelen);
625 return res;
724bdca9
SM		 626
627error:
724bdca9
SM		 628 if (res)
629 kmem_cache_free(dlm_lockres_cache, res);
630 return NULL;
6714d8e8
KH		 631}
632
8d400b81
SM		 633void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm,
634 struct dlm_lock_resource *res, int bit)
ba2bf218 635{
8d400b81
SM		 636 assert_spin_locked(&res->spinlock);
637
638 mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len,
639 res->lockname.name, bit, __builtin_return_address(0));
640
641 set_bit(bit, res->refmap);
642}
643
644void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm,
645 struct dlm_lock_resource *res, int bit)
646{
647 assert_spin_locked(&res->spinlock);
648
649 mlog(0, "res %.*s, clr node %u, %ps()\n", res->lockname.len,
650 res->lockname.name, bit, __builtin_return_address(0));
651
652 clear_bit(bit, res->refmap);
653}
654
5760a97c 655static void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
8d400b81
SM		 656 struct dlm_lock_resource *res)
657{
ba2bf218 658 res->inflight_locks++;
ff0a522e 659
8d400b81
SM		 660 mlog(0, "%s: res %.*s, inflight++: now %u, %ps()\n", dlm->name,
661 res->lockname.len, res->lockname.name, res->inflight_locks,
662 __builtin_return_address(0));
ba2bf218
KH		 663}
664
5760a97c
JQ		 665void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
666 struct dlm_lock_resource *res)
667{
668 assert_spin_locked(&res->spinlock);
669 __dlm_lockres_grab_inflight_ref(dlm, res);
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--;
ff0a522e 680
8d400b81
SM		 681 mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm->name,
682 res->lockname.len, res->lockname.name, res->inflight_locks,
683 __builtin_return_address(0));
684
ba2bf218
KH		 685 wake_up(&res->wq);
686}
687
ac4fef4d
X		 688void __dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm,
689 struct dlm_lock_resource *res)
690{
691 assert_spin_locked(&res->spinlock);
692 res->inflight_assert_workers++;
693 mlog(0, "%s:%.*s: inflight assert worker++: now %u\n",
694 dlm->name, res->lockname.len, res->lockname.name,
695 res->inflight_assert_workers);
696}
697
698static void dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm,
699 struct dlm_lock_resource *res)
700{
701 spin_lock(&res->spinlock);
702 __dlm_lockres_grab_inflight_worker(dlm, res);
703 spin_unlock(&res->spinlock);
704}
705
706static void __dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
707 struct dlm_lock_resource *res)
708{
709 assert_spin_locked(&res->spinlock);
710 BUG_ON(res->inflight_assert_workers == 0);
711 res->inflight_assert_workers--;
712 mlog(0, "%s:%.*s: inflight assert worker--: now %u\n",
713 dlm->name, res->lockname.len, res->lockname.name,
714 res->inflight_assert_workers);
715}
716
717static void dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
718 struct dlm_lock_resource *res)
719{
720 spin_lock(&res->spinlock);
721 __dlm_lockres_drop_inflight_worker(dlm, res);
722 spin_unlock(&res->spinlock);
723}
724
6714d8e8
KH		 725/*
726 * lookup a lock resource by name.
727 * may already exist in the hashtable.
728 * lockid is null terminated
729 *
730 * if not, allocate enough for the lockres and for
731 * the temporary structure used in doing the mastering.
732 *
733 * also, do a lookup in the dlm->master_list to see
734 * if another node has begun mastering the same lock.
735 * if so, there should be a block entry in there
736 * for this name, and we should *not* attempt to master
737 * the lock here. need to wait around for that node
738 * to assert_master (or die).
739 *
740 */
741struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
742 const char *lockid,
3384f3df 743 int namelen,
6714d8e8
KH		 744 int flags)
745{
746 struct dlm_lock_resource *tmpres=NULL, *res=NULL;
747 struct dlm_master_list_entry *mle = NULL;
748 struct dlm_master_list_entry *alloc_mle = NULL;
749 int blocked = 0;
750 int ret, nodenum;
751 struct dlm_node_iter iter;
3384f3df 752 unsigned int hash;
6714d8e8 753 int tries = 0;
c03872f5 754 int bit, wait_on_recovery = 0;
6714d8e8
KH		 755
756 BUG_ON(!lockid);
757
a3d33291 758 hash = dlm_lockid_hash(lockid, namelen);
6714d8e8
KH		 759
760 mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
761
762lookup:
763 spin_lock(&dlm->spinlock);
ba2bf218 764 tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
6714d8e8 765 if (tmpres) {
7b791d68 766 spin_unlock(&dlm->spinlock);
ba2bf218 767 spin_lock(&tmpres->spinlock);
ff0a522e 768 /* Wait on the thread that is mastering the resource */
7b791d68
SM		 769 if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
770 __dlm_wait_on_lockres(tmpres);
771 BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
ff0a522e
SM		 772 spin_unlock(&tmpres->spinlock);
773 dlm_lockres_put(tmpres);
774 tmpres = NULL;
775 goto lookup;
7b791d68
SM		 776 }
777
ff0a522e
SM		 778 /* Wait on the resource purge to complete before continuing */
779 if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) {
780 BUG_ON(tmpres->owner == dlm->node_num);
781 __dlm_wait_on_lockres_flags(tmpres,
782 DLM_LOCK_RES_DROPPING_REF);
ba2bf218
KH		 783 spin_unlock(&tmpres->spinlock);
784 dlm_lockres_put(tmpres);
785 tmpres = NULL;
786 goto lookup;
787 }
788
ff0a522e
SM		 789 /* Grab inflight ref to pin the resource */
790 dlm_lockres_grab_inflight_ref(dlm, tmpres);
791
792 spin_unlock(&tmpres->spinlock);
6714d8e8
KH		 793 if (res)
794 dlm_lockres_put(res);
795 res = tmpres;
796 goto leave;
797 }
798
799 if (!res) {
800 spin_unlock(&dlm->spinlock);
801 mlog(0, "allocating a new resource\n");
802 /* nothing found and we need to allocate one. */
3914ed0c 803 alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
6714d8e8
KH		 804 if (!alloc_mle)
805 goto leave;
806 res = dlm_new_lockres(dlm, lockid, namelen);
807 if (!res)
808 goto leave;
809 goto lookup;
810 }
811
812 mlog(0, "no lockres found, allocated our own: %p\n", res);
813
814 if (flags & LKM_LOCAL) {
815 /* caller knows it's safe to assume it's not mastered elsewhere
816 * DONE! return right away */
817 spin_lock(&res->spinlock);
818 dlm_change_lockres_owner(dlm, res, dlm->node_num);
819 __dlm_insert_lockres(dlm, res);
ba2bf218 820 dlm_lockres_grab_inflight_ref(dlm, res);
6714d8e8
KH		 821 spin_unlock(&res->spinlock);
822 spin_unlock(&dlm->spinlock);
823 /* lockres still marked IN_PROGRESS */
824 goto wake_waiters;
825 }
826
827 /* check master list to see if another node has started mastering it */
828 spin_lock(&dlm->master_lock);
829
830 /* if we found a block, wait for lock to be mastered by another node */
831 blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
832 if (blocked) {
ba2bf218 833 int mig;
6714d8e8
KH		 834 if (mle->type == DLM_MLE_MASTER) {
835 mlog(ML_ERROR, "master entry for nonexistent lock!\n");
836 BUG();
ba2bf218
KH		 837 }
838 mig = (mle->type == DLM_MLE_MIGRATION);
839 /* if there is a migration in progress, let the migration
840 * finish before continuing. we can wait for the absence
841 * of the MIGRATION mle: either the migrate finished or
842 * one of the nodes died and the mle was cleaned up.
843 * if there is a BLOCK here, but it already has a master
844 * set, we are too late. the master does not have a ref
845 * for us in the refmap. detach the mle and drop it.
846 * either way, go back to the top and start over. */
847 if (mig || mle->master != O2NM_MAX_NODES) {
848 BUG_ON(mig && mle->master == dlm->node_num);
849 /* we arrived too late. the master does not
850 * have a ref for us. retry. */
851 mlog(0, "%s:%.*s: late on %s\n",
852 dlm->name, namelen, lockid,
853 mig ? "MIGRATION" : "BLOCK");
6714d8e8 854 spin_unlock(&dlm->master_lock);
6714d8e8
KH		 855 spin_unlock(&dlm->spinlock);
856
857 /* master is known, detach */
ba2bf218
KH		 858 if (!mig)
859 dlm_mle_detach_hb_events(dlm, mle);
6714d8e8
KH		 860 dlm_put_mle(mle);
861 mle = NULL;
25985edc 862 /* this is lame, but we can't wait on either
ba2bf218
KH		 863 * the mle or lockres waitqueue here */
864 if (mig)
865 msleep(100);
866 goto lookup;
6714d8e8
KH		 867 }
868 } else {
869 /* go ahead and try to master lock on this node */
870 mle = alloc_mle;
871 /* make sure this does not get freed below */
872 alloc_mle = NULL;
873 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
874 set_bit(dlm->node_num, mle->maybe_map);
1c084577 875 __dlm_insert_mle(dlm, mle);
c03872f5
KH		 876
877 /* still holding the dlm spinlock, check the recovery map
2bd63216 878 * to see if there are any nodes that still need to be
c03872f5
KH		 879 * considered. these will not appear in the mle nodemap
880 * but they might own this lockres. wait on them. */
881 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
882 if (bit < O2NM_MAX_NODES) {
8decab3c
SM		 883 mlog(0, "%s: res %.*s, At least one node (%d) "
884 "to recover before lock mastery can begin\n",
c03872f5
KH		 885 dlm->name, namelen, (char *)lockid, bit);
886 wait_on_recovery = 1;
887 }
6714d8e8
KH		 888 }
889
890 /* at this point there is either a DLM_MLE_BLOCK or a
891 * DLM_MLE_MASTER on the master list, so it's safe to add the
892 * lockres to the hashtable. anyone who finds the lock will
893 * still have to wait on the IN_PROGRESS. */
894
895 /* finally add the lockres to its hash bucket */
896 __dlm_insert_lockres(dlm, res);
8d400b81 897
5760a97c
JQ		 898 /* since this lockres is new it doesn't not require the spinlock */
899 __dlm_lockres_grab_inflight_ref(dlm, res);
ba2bf218 900
6714d8e8
KH		 901 /* get an extra ref on the mle in case this is a BLOCK
902 * if so, the creator of the BLOCK may try to put the last
903 * ref at this time in the assert master handler, so we
904 * need an extra one to keep from a bad ptr deref. */
a2bf0477 905 dlm_get_mle_inuse(mle);
6714d8e8
KH		 906 spin_unlock(&dlm->master_lock);
907 spin_unlock(&dlm->spinlock);
908
e7e69eb3 909redo_request:
c03872f5
KH		 910 while (wait_on_recovery) {
911 /* any cluster changes that occurred after dropping the
912 * dlm spinlock would be detectable be a change on the mle,
913 * so we only need to clear out the recovery map once. */
914 if (dlm_is_recovery_lock(lockid, namelen)) {
8decab3c
SM		 915 mlog(0, "%s: Recovery map is not empty, but must "
916 "master $RECOVERY lock now\n", dlm->name);
c03872f5
KH		 917 if (!dlm_pre_master_reco_lockres(dlm, res))
918 wait_on_recovery = 0;
919 else {
920 mlog(0, "%s: waiting 500ms for heartbeat state "
921 "change\n", dlm->name);
922 msleep(500);
923 }
924 continue;
2bd63216 925 }
c03872f5
KH		 926
927 dlm_kick_recovery_thread(dlm);
aa087b84 928 msleep(1000);
c03872f5
KH		 929 dlm_wait_for_recovery(dlm);
930
931 spin_lock(&dlm->spinlock);
932 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
933 if (bit < O2NM_MAX_NODES) {
8decab3c
SM		 934 mlog(0, "%s: res %.*s, At least one node (%d) "
935 "to recover before lock mastery can begin\n",
c03872f5
KH		 936 dlm->name, namelen, (char *)lockid, bit);
937 wait_on_recovery = 1;
938 } else
939 wait_on_recovery = 0;
940 spin_unlock(&dlm->spinlock);
b7084ab5
KH		 941
942 if (wait_on_recovery)
943 dlm_wait_for_node_recovery(dlm, bit, 10000);
c03872f5
KH		 944 }
945
6714d8e8
KH		 946 /* must wait for lock to be mastered elsewhere */
947 if (blocked)
948 goto wait;
949
6714d8e8
KH		 950 ret = -EINVAL;
951 dlm_node_iter_init(mle->vote_map, &iter);
952 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
ba2bf218 953 ret = dlm_do_master_request(res, mle, nodenum);
6714d8e8
KH		 954 if (ret < 0)
955 mlog_errno(ret);
956 if (mle->master != O2NM_MAX_NODES) {
957 /* found a master ! */
9c6510a5
KH		 958 if (mle->master <= nodenum)
959 break;
960 /* if our master request has not reached the master
961 * yet, keep going until it does. this is how the
962 * master will know that asserts are needed back to
963 * the lower nodes. */
8decab3c
SM		 964 mlog(0, "%s: res %.*s, Requests only up to %u but "
965 "master is %u, keep going\n", dlm->name, namelen,
9c6510a5 966 lockid, nodenum, mle->master);
6714d8e8
KH		 967 }
968 }
969
970wait:
971 /* keep going until the response map includes all nodes */
972 ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
973 if (ret < 0) {
e7e69eb3 974 wait_on_recovery = 1;
8decab3c
SM		 975 mlog(0, "%s: res %.*s, Node map changed, redo the master "
976 "request now, blocked=%d\n", dlm->name, res->lockname.len,
6714d8e8
KH		 977 res->lockname.name, blocked);
978 if (++tries > 20) {
8decab3c
SM		 979 mlog(ML_ERROR, "%s: res %.*s, Spinning on "
980 "dlm_wait_for_lock_mastery, blocked = %d\n",
2bd63216 981 dlm->name, res->lockname.len,
6714d8e8
KH		 982 res->lockname.name, blocked);
983 dlm_print_one_lock_resource(res);
8a9343fa 984 dlm_print_one_mle(mle);
6714d8e8
KH		 985 tries = 0;
986 }
987 goto redo_request;
988 }
989
8decab3c
SM		 990 mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len,
991 res->lockname.name, res->owner);
6714d8e8
KH		 992 /* make sure we never continue without this */
993 BUG_ON(res->owner == O2NM_MAX_NODES);
994
995 /* master is known, detach if not already detached */
996 dlm_mle_detach_hb_events(dlm, mle);
997 dlm_put_mle(mle);
998 /* put the extra ref */
a2bf0477 999 dlm_put_mle_inuse(mle);
6714d8e8
KH		 1000
1001wake_waiters:
1002 spin_lock(&res->spinlock);
1003 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1004 spin_unlock(&res->spinlock);
1005 wake_up(&res->wq);
1006
1007leave:
1008 /* need to free the unused mle */
1009 if (alloc_mle)
1010 kmem_cache_free(dlm_mle_cache, alloc_mle);
1011
1012 return res;
1013}
1014
1015
1016#define DLM_MASTERY_TIMEOUT_MS 5000
1017
1018static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
1019 struct dlm_lock_resource *res,
1020 struct dlm_master_list_entry *mle,
1021 int *blocked)
1022{
1023 u8 m;
1024 int ret, bit;
1025 int map_changed, voting_done;
1026 int assert, sleep;
1027
1028recheck:
1029 ret = 0;
1030 assert = 0;
1031
1032 /* check if another node has already become the owner */
1033 spin_lock(&res->spinlock);
1034 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
9c6510a5
KH		 1035 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1036 res->lockname.len, res->lockname.name, res->owner);
6714d8e8 1037 spin_unlock(&res->spinlock);
9c6510a5
KH		 1038 /* this will cause the master to re-assert across
1039 * the whole cluster, freeing up mles */
588e0090 1040 if (res->owner != dlm->node_num) {
ba2bf218 1041 ret = dlm_do_master_request(res, mle, res->owner);
588e0090
KH		 1042 if (ret < 0) {
1043 /* give recovery a chance to run */
1044 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1045 msleep(500);
1046 goto recheck;
1047 }
9c6510a5
KH		 1048 }
1049 ret = 0;
6714d8e8
KH		 1050 goto leave;
1051 }
1052 spin_unlock(&res->spinlock);
1053
1054 spin_lock(&mle->spinlock);
1055 m = mle->master;
1056 map_changed = (memcmp(mle->vote_map, mle->node_map,
1057 sizeof(mle->vote_map)) != 0);
1058 voting_done = (memcmp(mle->vote_map, mle->response_map,
1059 sizeof(mle->vote_map)) == 0);
1060
1061 /* restart if we hit any errors */
1062 if (map_changed) {
1063 int b;
1064 mlog(0, "%s: %.*s: node map changed, restarting\n",
1065 dlm->name, res->lockname.len, res->lockname.name);
1066 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1067 b = (mle->type == DLM_MLE_BLOCK);
1068 if ((*blocked && !b) || (!*blocked && b)) {
2bd63216 1069 mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
6714d8e8
KH		 1070 dlm->name, res->lockname.len, res->lockname.name,
1071 *blocked, b);
1072 *blocked = b;
1073 }
1074 spin_unlock(&mle->spinlock);
1075 if (ret < 0) {
1076 mlog_errno(ret);
1077 goto leave;
1078 }
1079 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1080 "rechecking now\n", dlm->name, res->lockname.len,
1081 res->lockname.name);
1082 goto recheck;
aa852354
KH		 1083 } else {
1084 if (!voting_done) {
1085 mlog(0, "map not changed and voting not done "
1086 "for %s:%.*s\n", dlm->name, res->lockname.len,
1087 res->lockname.name);
1088 }
6714d8e8
KH		 1089 }
1090
1091 if (m != O2NM_MAX_NODES) {
1092 /* another node has done an assert!
1093 * all done! */
1094 sleep = 0;
1095 } else {
1096 sleep = 1;
1097 /* have all nodes responded? */
1098 if (voting_done && !*blocked) {
1099 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1100 if (dlm->node_num <= bit) {
1101 /* my node number is lowest.
1102 * now tell other nodes that I am
1103 * mastering this. */
1104 mle->master = dlm->node_num;
ba2bf218
KH		 1105 /* ref was grabbed in get_lock_resource
1106 * will be dropped in dlmlock_master */
6714d8e8
KH		 1107 assert = 1;
1108 sleep = 0;
1109 }
1110 /* if voting is done, but we have not received
1111 * an assert master yet, we must sleep */
1112 }
1113 }
1114
1115 spin_unlock(&mle->spinlock);
1116
1117 /* sleep if we haven't finished voting yet */
1118 if (sleep) {
1119 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1120
1121 /*
1122 if (atomic_read(&mle->mle_refs.refcount) < 2)
1123 mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1124 atomic_read(&mle->mle_refs.refcount),
1125 res->lockname.len, res->lockname.name);
1126 */
1127 atomic_set(&mle->woken, 0);
1128 (void)wait_event_timeout(mle->wq,
1129 (atomic_read(&mle->woken) == 1),
1130 timeo);
1131 if (res->owner == O2NM_MAX_NODES) {
ba2bf218
KH		 1132 mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1133 res->lockname.len, res->lockname.name);
6714d8e8
KH		 1134 goto recheck;
1135 }
1136 mlog(0, "done waiting, master is %u\n", res->owner);
1137 ret = 0;
1138 goto leave;
1139 }
1140
1141 ret = 0; /* done */
1142 if (assert) {
1143 m = dlm->node_num;
1144 mlog(0, "about to master %.*s here, this=%u\n",
1145 res->lockname.len, res->lockname.name, m);
ba2bf218 1146 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
6714d8e8
KH		 1147 if (ret) {
1148 /* This is a failure in the network path,
1149 * not in the response to the assert_master
1150 * (any nonzero response is a BUG on this node).
1151 * Most likely a socket just got disconnected
1152 * due to node death. */
1153 mlog_errno(ret);
1154 }
1155 /* no longer need to restart lock mastery.
1156 * all living nodes have been contacted. */
1157 ret = 0;
1158 }
1159
1160 /* set the lockres owner */
1161 spin_lock(&res->spinlock);
ba2bf218
KH		 1162 /* mastery reference obtained either during
1163 * assert_master_handler or in get_lock_resource */
6714d8e8
KH		 1164 dlm_change_lockres_owner(dlm, res, m);
1165 spin_unlock(&res->spinlock);
1166
1167leave:
1168 return ret;
1169}
1170
1171struct dlm_bitmap_diff_iter
1172{
1173 int curnode;
1174 unsigned long *orig_bm;
1175 unsigned long *cur_bm;
1176 unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1177};
1178
1179enum dlm_node_state_change
1180{
1181 NODE_DOWN = -1,
1182 NODE_NO_CHANGE = 0,
1183 NODE_UP
1184};
1185
1186static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1187 unsigned long *orig_bm,
1188 unsigned long *cur_bm)
1189{
1190 unsigned long p1, p2;
1191 int i;
1192
1193 iter->curnode = -1;
1194 iter->orig_bm = orig_bm;
1195 iter->cur_bm = cur_bm;
1196
1197 for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1198 p1 = *(iter->orig_bm + i);
1199 p2 = *(iter->cur_bm + i);
1200 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1201 }
1202}
1203
1204static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1205 enum dlm_node_state_change *state)
1206{
1207 int bit;
1208
1209 if (iter->curnode >= O2NM_MAX_NODES)
1210 return -ENOENT;
1211
1212 bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1213 iter->curnode+1);
1214 if (bit >= O2NM_MAX_NODES) {
1215 iter->curnode = O2NM_MAX_NODES;
1216 return -ENOENT;
1217 }
1218
1219 /* if it was there in the original then this node died */
1220 if (test_bit(bit, iter->orig_bm))
1221 *state = NODE_DOWN;
1222 else
1223 *state = NODE_UP;
1224
1225 iter->curnode = bit;
1226 return bit;
1227}
1228
1229
1230static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1231 struct dlm_lock_resource *res,
1232 struct dlm_master_list_entry *mle,
1233 int blocked)
1234{
1235 struct dlm_bitmap_diff_iter bdi;
1236 enum dlm_node_state_change sc;
1237 int node;
1238 int ret = 0;
1239
1240 mlog(0, "something happened such that the "
1241 "master process may need to be restarted!\n");
1242
1243 assert_spin_locked(&mle->spinlock);
1244
1245 dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1246 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1247 while (node >= 0) {
1248 if (sc == NODE_UP) {
e2faea4c
KH		 1249 /* a node came up. clear any old vote from
1250 * the response map and set it in the vote map
1251 * then restart the mastery. */
1252 mlog(ML_NOTICE, "node %d up while restarting\n", node);
6714d8e8
KH		 1253
1254 /* redo the master request, but only for the new node */
1255 mlog(0, "sending request to new node\n");
1256 clear_bit(node, mle->response_map);
1257 set_bit(node, mle->vote_map);
1258 } else {
1259 mlog(ML_ERROR, "node down! %d\n", node);
6714d8e8
KH		 1260 if (blocked) {
1261 int lowest = find_next_bit(mle->maybe_map,
1262 O2NM_MAX_NODES, 0);
1263
1264 /* act like it was never there */
1265 clear_bit(node, mle->maybe_map);
1266
e7e69eb3
KH		 1267 if (node == lowest) {
1268 mlog(0, "expected master %u died"
1269 " while this node was blocked "
1270 "waiting on it!\n", node);
1271 lowest = find_next_bit(mle->maybe_map,
1272 O2NM_MAX_NODES,
1273 lowest+1);
1274 if (lowest < O2NM_MAX_NODES) {
1275 mlog(0, "%s:%.*s:still "
1276 "blocked. waiting on %u "
1277 "now\n", dlm->name,
1278 res->lockname.len,
1279 res->lockname.name,
1280 lowest);
1281 } else {
1282 /* mle is an MLE_BLOCK, but
1283 * there is now nothing left to
1284 * block on. we need to return
1285 * all the way back out and try
1286 * again with an MLE_MASTER.
1287 * dlm_do_local_recovery_cleanup
1288 * has already run, so the mle
1289 * refcount is ok */
1290 mlog(0, "%s:%.*s: no "
1291 "longer blocking. try to "
1292 "master this here\n",
1293 dlm->name,
1294 res->lockname.len,
1295 res->lockname.name);
1296 mle->type = DLM_MLE_MASTER;
7141514b 1297 mle->mleres = res;
e7e69eb3 1298 }
6714d8e8 1299 }
6714d8e8
KH		 1300 }
1301
e7e69eb3
KH		 1302 /* now blank out everything, as if we had never
1303 * contacted anyone */
1304 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1305 memset(mle->response_map, 0, sizeof(mle->response_map));
1306 /* reset the vote_map to the current node_map */
1307 memcpy(mle->vote_map, mle->node_map,
1308 sizeof(mle->node_map));
1309 /* put myself into the maybe map */
1310 if (mle->type != DLM_MLE_BLOCK)
1311 set_bit(dlm->node_num, mle->maybe_map);
6714d8e8
KH		 1312 }
1313 ret = -EAGAIN;
6714d8e8
KH		 1314 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1315 }
1316 return ret;
1317}
1318
1319
1320/*
1321 * DLM_MASTER_REQUEST_MSG
1322 *
1323 * returns: 0 on success,
1324 * -errno on a network error
1325 *
1326 * on error, the caller should assume the target node is "dead"
1327 *
1328 */
1329
ba2bf218
KH		 1330static int dlm_do_master_request(struct dlm_lock_resource *res,
1331 struct dlm_master_list_entry *mle, int to)
6714d8e8
KH		 1332{
1333 struct dlm_ctxt *dlm = mle->dlm;
1334 struct dlm_master_request request;
1335 int ret, response=0, resend;
1336
1337 memset(&request, 0, sizeof(request));
1338 request.node_idx = dlm->node_num;
1339
1340 BUG_ON(mle->type == DLM_MLE_MIGRATION);
1341
7141514b
SM		 1342 request.namelen = (u8)mle->mnamelen;
1343 memcpy(request.name, mle->mname, request.namelen);
6714d8e8
KH		 1344
1345again:
1346 ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1347 sizeof(request), to, &response);
1348 if (ret < 0) {
1349 if (ret == -ESRCH) {
1350 /* should never happen */
1351 mlog(ML_ERROR, "TCP stack not ready!\n");
1352 BUG();
1353 } else if (ret == -EINVAL) {
1354 mlog(ML_ERROR, "bad args passed to o2net!\n");
1355 BUG();
1356 } else if (ret == -ENOMEM) {
1357 mlog(ML_ERROR, "out of memory while trying to send "
1358 "network message! retrying\n");
1359 /* this is totally crude */
1360 msleep(50);
1361 goto again;
1362 } else if (!dlm_is_host_down(ret)) {
1363 /* not a network error. bad. */
1364 mlog_errno(ret);
1365 mlog(ML_ERROR, "unhandled error!");
1366 BUG();
1367 }
1368 /* all other errors should be network errors,
1369 * and likely indicate node death */
1370 mlog(ML_ERROR, "link to %d went down!\n", to);
1371 goto out;
1372 }
1373
1374 ret = 0;
1375 resend = 0;
1376 spin_lock(&mle->spinlock);
1377 switch (response) {
1378 case DLM_MASTER_RESP_YES:
1379 set_bit(to, mle->response_map);
1380 mlog(0, "node %u is the master, response=YES\n", to);
ba2bf218
KH		 1381 mlog(0, "%s:%.*s: master node %u now knows I have a "
1382 "reference\n", dlm->name, res->lockname.len,
1383 res->lockname.name, to);
6714d8e8
KH		 1384 mle->master = to;
1385 break;
1386 case DLM_MASTER_RESP_NO:
1387 mlog(0, "node %u not master, response=NO\n", to);
1388 set_bit(to, mle->response_map);
1389 break;
1390 case DLM_MASTER_RESP_MAYBE:
1391 mlog(0, "node %u not master, response=MAYBE\n", to);
1392 set_bit(to, mle->response_map);
1393 set_bit(to, mle->maybe_map);
1394 break;
1395 case DLM_MASTER_RESP_ERROR:
1396 mlog(0, "node %u hit an error, resending\n", to);
1397 resend = 1;
1398 response = 0;
1399 break;
1400 default:
1401 mlog(ML_ERROR, "bad response! %u\n", response);
1402 BUG();
1403 }
1404 spin_unlock(&mle->spinlock);
1405 if (resend) {
1406 /* this is also totally crude */
1407 msleep(50);
1408 goto again;
1409 }
1410
1411out:
1412 return ret;
1413}
1414
1415/*
1416 * locks that can be taken here:
1417 * dlm->spinlock
1418 * res->spinlock
1419 * mle->spinlock
1420 * dlm->master_list
1421 *
1422 * if possible, TRIM THIS DOWN!!!
1423 */
d74c9803
KH		 1424int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1425 void **ret_data)
6714d8e8
KH		 1426{
1427 u8 response = DLM_MASTER_RESP_MAYBE;
1428 struct dlm_ctxt *dlm = data;
9c6510a5 1429 struct dlm_lock_resource *res = NULL;
6714d8e8
KH		 1430 struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1431 struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1432 char *name;
a3d33291 1433 unsigned int namelen, hash;
6714d8e8
KH		 1434 int found, ret;
1435 int set_maybe;
9c6510a5 1436 int dispatch_assert = 0;
6714d8e8
KH		 1437
1438 if (!dlm_grab(dlm))
1439 return DLM_MASTER_RESP_NO;
1440
1441 if (!dlm_domain_fully_joined(dlm)) {
1442 response = DLM_MASTER_RESP_NO;
1443 goto send_response;
1444 }
1445
1446 name = request->name;
1447 namelen = request->namelen;
a3d33291 1448 hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 1449
1450 if (namelen > DLM_LOCKID_NAME_MAX) {
1451 response = DLM_IVBUFLEN;
1452 goto send_response;
1453 }
1454
1455way_up_top:
1456 spin_lock(&dlm->spinlock);
a3d33291 1457 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
6714d8e8
KH		 1458 if (res) {
1459 spin_unlock(&dlm->spinlock);
1460
1461 /* take care of the easy cases up front */
1462 spin_lock(&res->spinlock);
cb79662b
SE		 1463
1464 /*
1465 * Right after dlm spinlock was released, dlm_thread could have
1466 * purged the lockres. Check if lockres got unhashed. If so
1467 * start over.
1468 */
1469 if (hlist_unhashed(&res->hash_node)) {
1470 spin_unlock(&res->spinlock);
1471 dlm_lockres_put(res);
1472 goto way_up_top;
1473 }
1474
1cd04dbe
KH		 1475 if (res->state & (DLM_LOCK_RES_RECOVERING|
1476 DLM_LOCK_RES_MIGRATING)) {
6714d8e8
KH		 1477 spin_unlock(&res->spinlock);
1478 mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1cd04dbe 1479 "being recovered/migrated\n");
6714d8e8
KH		 1480 response = DLM_MASTER_RESP_ERROR;
1481 if (mle)
1482 kmem_cache_free(dlm_mle_cache, mle);
1483 goto send_response;
1484 }
1485
1486 if (res->owner == dlm->node_num) {
8d400b81 1487 dlm_lockres_set_refmap_bit(dlm, res, request->node_idx);
6714d8e8 1488 spin_unlock(&res->spinlock);
6714d8e8
KH		 1489 response = DLM_MASTER_RESP_YES;
1490 if (mle)
1491 kmem_cache_free(dlm_mle_cache, mle);
1492
1493 /* this node is the owner.
1494 * there is some extra work that needs to
1495 * happen now. the requesting node has
1496 * caused all nodes up to this one to
1497 * create mles. this node now needs to
1498 * go back and clean those up. */
9c6510a5 1499 dispatch_assert = 1;
6714d8e8
KH		 1500 goto send_response;
1501 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1502 spin_unlock(&res->spinlock);
1503 // mlog(0, "node %u is the master\n", res->owner);
1504 response = DLM_MASTER_RESP_NO;
1505 if (mle)
1506 kmem_cache_free(dlm_mle_cache, mle);
1507 goto send_response;
1508 }
1509
1510 /* ok, there is no owner. either this node is
1511 * being blocked, or it is actively trying to
1512 * master this lock. */
1513 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1514 mlog(ML_ERROR, "lock with no owner should be "
1515 "in-progress!\n");
1516 BUG();
1517 }
1518
1519 // mlog(0, "lockres is in progress...\n");
1520 spin_lock(&dlm->master_lock);
1521 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1522 if (!found) {
1523 mlog(ML_ERROR, "no mle found for this lock!\n");
1524 BUG();
1525 }
1526 set_maybe = 1;
1527 spin_lock(&tmpmle->spinlock);
1528 if (tmpmle->type == DLM_MLE_BLOCK) {
1529 // mlog(0, "this node is waiting for "
1530 // "lockres to be mastered\n");
1531 response = DLM_MASTER_RESP_NO;
1532 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1533 mlog(0, "node %u is master, but trying to migrate to "
1534 "node %u.\n", tmpmle->master, tmpmle->new_master);
1535 if (tmpmle->master == dlm->node_num) {
6714d8e8
KH		 1536 mlog(ML_ERROR, "no owner on lockres, but this "
1537 "node is trying to migrate it to %u?!\n",
1538 tmpmle->new_master);
1539 BUG();
1540 } else {
1541 /* the real master can respond on its own */
1542 response = DLM_MASTER_RESP_NO;
1543 }
1544 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1545 set_maybe = 0;
9c6510a5 1546 if (tmpmle->master == dlm->node_num) {
6714d8e8 1547 response = DLM_MASTER_RESP_YES;
9c6510a5
KH		 1548 /* this node will be the owner.
1549 * go back and clean the mles on any
1550 * other nodes */
1551 dispatch_assert = 1;
8d400b81
SM		 1552 dlm_lockres_set_refmap_bit(dlm, res,
1553 request->node_idx);
9c6510a5 1554 } else
6714d8e8
KH		 1555 response = DLM_MASTER_RESP_NO;
1556 } else {
1557 // mlog(0, "this node is attempting to "
1558 // "master lockres\n");
1559 response = DLM_MASTER_RESP_MAYBE;
1560 }
1561 if (set_maybe)
1562 set_bit(request->node_idx, tmpmle->maybe_map);
1563 spin_unlock(&tmpmle->spinlock);
1564
1565 spin_unlock(&dlm->master_lock);
1566 spin_unlock(&res->spinlock);
1567
1568 /* keep the mle attached to heartbeat events */
1569 dlm_put_mle(tmpmle);
1570 if (mle)
1571 kmem_cache_free(dlm_mle_cache, mle);
1572 goto send_response;
1573 }
1574
1575 /*
1576 * lockres doesn't exist on this node
1577 * if there is an MLE_BLOCK, return NO
1578 * if there is an MLE_MASTER, return MAYBE
1579 * otherwise, add an MLE_BLOCK, return NO
1580 */
1581 spin_lock(&dlm->master_lock);
1582 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1583 if (!found) {
1584 /* this lockid has never been seen on this node yet */
1585 // mlog(0, "no mle found\n");
1586 if (!mle) {
1587 spin_unlock(&dlm->master_lock);
1588 spin_unlock(&dlm->spinlock);
1589
3914ed0c 1590 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
6714d8e8 1591 if (!mle) {
6714d8e8 1592 response = DLM_MASTER_RESP_ERROR;
9c6510a5 1593 mlog_errno(-ENOMEM);
6714d8e8
KH		 1594 goto send_response;
1595 }
6714d8e8
KH		 1596 goto way_up_top;
1597 }
1598
1599 // mlog(0, "this is second time thru, already allocated, "
1600 // "add the block.\n");
41b8c8a1 1601 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
6714d8e8 1602 set_bit(request->node_idx, mle->maybe_map);
1c084577 1603 __dlm_insert_mle(dlm, mle);
6714d8e8
KH		 1604 response = DLM_MASTER_RESP_NO;
1605 } else {
1606 // mlog(0, "mle was found\n");
1607 set_maybe = 1;
1608 spin_lock(&tmpmle->spinlock);
9c6510a5
KH		 1609 if (tmpmle->master == dlm->node_num) {
1610 mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1611 BUG();
1612 }
6714d8e8
KH		 1613 if (tmpmle->type == DLM_MLE_BLOCK)
1614 response = DLM_MASTER_RESP_NO;
1615 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1616 mlog(0, "migration mle was found (%u->%u)\n",
1617 tmpmle->master, tmpmle->new_master);
6714d8e8
KH		 1618 /* real master can respond on its own */
1619 response = DLM_MASTER_RESP_NO;
9c6510a5
KH		 1620 } else
1621 response = DLM_MASTER_RESP_MAYBE;
6714d8e8
KH		 1622 if (set_maybe)
1623 set_bit(request->node_idx, tmpmle->maybe_map);
1624 spin_unlock(&tmpmle->spinlock);
1625 }
1626 spin_unlock(&dlm->master_lock);
1627 spin_unlock(&dlm->spinlock);
1628
1629 if (found) {
1630 /* keep the mle attached to heartbeat events */
1631 dlm_put_mle(tmpmle);
1632 }
1633send_response:
b31cfc02
SM		 1634 /*
1635 * __dlm_lookup_lockres() grabbed a reference to this lockres.
1636 * The reference is released by dlm_assert_master_worker() under
1637 * the call to dlm_dispatch_assert_master(). If
1638 * dlm_assert_master_worker() isn't called, we drop it here.
1639 */
9c6510a5
KH		 1640 if (dispatch_assert) {
1641 if (response != DLM_MASTER_RESP_YES)
1642 mlog(ML_ERROR, "invalid response %d\n", response);
1643 if (!res) {
1644 mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1645 BUG();
1646 }
1647 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1648 dlm->node_num, res->lockname.len, res->lockname.name);
2bd63216 1649 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
9c6510a5
KH		 1650 DLM_ASSERT_MASTER_MLE_CLEANUP);
1651 if (ret < 0) {
1652 mlog(ML_ERROR, "failed to dispatch assert master work\n");
1653 response = DLM_MASTER_RESP_ERROR;
b31cfc02 1654 dlm_lockres_put(res);
ac4fef4d
X		 1655 } else
1656 dlm_lockres_grab_inflight_worker(dlm, res);
b31cfc02
SM		 1657 } else {
1658 if (res)
1659 dlm_lockres_put(res);
9c6510a5
KH		 1660 }
1661
6714d8e8
KH		 1662 dlm_put(dlm);
1663 return response;
1664}
1665
1666/*
1667 * DLM_ASSERT_MASTER_MSG
1668 */
1669
1670
1671/*
1672 * NOTE: this can be used for debugging
1673 * can periodically run all locks owned by this node
1674 * and re-assert across the cluster...
1675 */
05488bbe
AB		 1676static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1677 struct dlm_lock_resource *res,
1678 void *nodemap, u32 flags)
6714d8e8
KH		 1679{
1680 struct dlm_assert_master assert;
1681 int to, tmpret;
1682 struct dlm_node_iter iter;
1683 int ret = 0;
9c6510a5 1684 int reassert;
ba2bf218
KH		 1685 const char *lockname = res->lockname.name;
1686 unsigned int namelen = res->lockname.len;
6714d8e8
KH		 1687
1688 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
f3f85464
SM		 1689
1690 spin_lock(&res->spinlock);
1691 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1692 spin_unlock(&res->spinlock);
1693
9c6510a5
KH		 1694again:
1695 reassert = 0;
6714d8e8
KH		 1696
1697 /* note that if this nodemap is empty, it returns 0 */
1698 dlm_node_iter_init(nodemap, &iter);
1699 while ((to = dlm_node_iter_next(&iter)) >= 0) {
1700 int r = 0;
a9ee4c8a
KH		 1701 struct dlm_master_list_entry *mle = NULL;
1702
6714d8e8
KH		 1703 mlog(0, "sending assert master to %d (%.*s)\n", to,
1704 namelen, lockname);
1705 memset(&assert, 0, sizeof(assert));
1706 assert.node_idx = dlm->node_num;
1707 assert.namelen = namelen;
1708 memcpy(assert.name, lockname, namelen);
1709 assert.flags = cpu_to_be32(flags);
1710
1711 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1712 &assert, sizeof(assert), to, &r);
1713 if (tmpret < 0) {
a5196ec5
WW		 1714 mlog(ML_ERROR, "Error %d when sending message %u (key "
1715 "0x%x) to node %u\n", tmpret,
1716 DLM_ASSERT_MASTER_MSG, dlm->key, to);
6714d8e8 1717 if (!dlm_is_host_down(tmpret)) {
3b3b84a8 1718 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
6714d8e8
KH		 1719 BUG();
1720 }
1721 /* a node died. finish out the rest of the nodes. */
3b3b84a8 1722 mlog(0, "link to %d went down!\n", to);
6714d8e8
KH		 1723 /* any nonzero status return will do */
1724 ret = tmpret;
ba2bf218 1725 r = 0;
6714d8e8
KH		 1726 } else if (r < 0) {
1727 /* ok, something horribly messed. kill thyself. */
1728 mlog(ML_ERROR,"during assert master of %.*s to %u, "
1729 "got %d.\n", namelen, lockname, to, r);
a9ee4c8a
KH		 1730 spin_lock(&dlm->spinlock);
1731 spin_lock(&dlm->master_lock);
1732 if (dlm_find_mle(dlm, &mle, (char *)lockname,
1733 namelen)) {
1734 dlm_print_one_mle(mle);
1735 __dlm_put_mle(mle);
1736 }
1737 spin_unlock(&dlm->master_lock);
1738 spin_unlock(&dlm->spinlock);
6714d8e8 1739 BUG();
ba2bf218
KH		 1740 }
1741
1742 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1743 !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1744 mlog(ML_ERROR, "%.*s: very strange, "
1745 "master MLE but no lockres on %u\n",
1746 namelen, lockname, to);
1747 }
1748
1749 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
9c6510a5 1750 mlog(0, "%.*s: node %u create mles on other "
2bd63216 1751 "nodes and requests a re-assert\n",
9c6510a5
KH		 1752 namelen, lockname, to);
1753 reassert = 1;
6714d8e8 1754 }
ba2bf218
KH		 1755 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1756 mlog(0, "%.*s: node %u has a reference to this "
1757 "lockres, set the bit in the refmap\n",
1758 namelen, lockname, to);
1759 spin_lock(&res->spinlock);
8d400b81 1760 dlm_lockres_set_refmap_bit(dlm, res, to);
ba2bf218
KH		 1761 spin_unlock(&res->spinlock);
1762 }
6714d8e8
KH		 1763 }
1764
9c6510a5
KH		 1765 if (reassert)
1766 goto again;
1767
f3f85464
SM		 1768 spin_lock(&res->spinlock);
1769 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1770 spin_unlock(&res->spinlock);
1771 wake_up(&res->wq);
1772
6714d8e8
KH		 1773 return ret;
1774}
1775
1776/*
1777 * locks that can be taken here:
1778 * dlm->spinlock
1779 * res->spinlock
1780 * mle->spinlock
1781 * dlm->master_list
1782 *
1783 * if possible, TRIM THIS DOWN!!!
1784 */
d74c9803
KH		 1785int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1786 void **ret_data)
6714d8e8
KH		 1787{
1788 struct dlm_ctxt *dlm = data;
1789 struct dlm_master_list_entry *mle = NULL;
1790 struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1791 struct dlm_lock_resource *res = NULL;
1792 char *name;
a3d33291 1793 unsigned int namelen, hash;
6714d8e8 1794 u32 flags;
ba2bf218 1795 int master_request = 0, have_lockres_ref = 0;
9c6510a5 1796 int ret = 0;
6714d8e8
KH		 1797
1798 if (!dlm_grab(dlm))
1799 return 0;
1800
1801 name = assert->name;
1802 namelen = assert->namelen;
a3d33291 1803 hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 1804 flags = be32_to_cpu(assert->flags);
1805
1806 if (namelen > DLM_LOCKID_NAME_MAX) {
1807 mlog(ML_ERROR, "Invalid name length!");
1808 goto done;
1809 }
1810
1811 spin_lock(&dlm->spinlock);
1812
1813 if (flags)
1814 mlog(0, "assert_master with flags: %u\n", flags);
1815
1816 /* find the MLE */
1817 spin_lock(&dlm->master_lock);
1818 if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1819 /* not an error, could be master just re-asserting */
1820 mlog(0, "just got an assert_master from %u, but no "
1821 "MLE for it! (%.*s)\n", assert->node_idx,
1822 namelen, name);
1823 } else {
1824 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1825 if (bit >= O2NM_MAX_NODES) {
1826 /* not necessarily an error, though less likely.
1827 * could be master just re-asserting. */
aa852354 1828 mlog(0, "no bits set in the maybe_map, but %u "
6714d8e8
KH		 1829 "is asserting! (%.*s)\n", assert->node_idx,
1830 namelen, name);
1831 } else if (bit != assert->node_idx) {
1832 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1833 mlog(0, "master %u was found, %u should "
1834 "back off\n", assert->node_idx, bit);
1835 } else {
1836 /* with the fix for bug 569, a higher node
1837 * number winning the mastery will respond
1838 * YES to mastery requests, but this node
1839 * had no way of knowing. let it pass. */
aa852354 1840 mlog(0, "%u is the lowest node, "
6714d8e8
KH		 1841 "%u is asserting. (%.*s) %u must "
1842 "have begun after %u won.\n", bit,
1843 assert->node_idx, namelen, name, bit,
1844 assert->node_idx);
1845 }
1846 }
2d1a868c
KH		 1847 if (mle->type == DLM_MLE_MIGRATION) {
1848 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1849 mlog(0, "%s:%.*s: got cleanup assert"
1850 " from %u for migration\n",
1851 dlm->name, namelen, name,
1852 assert->node_idx);
1853 } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1854 mlog(0, "%s:%.*s: got unrelated assert"
1855 " from %u for migration, ignoring\n",
1856 dlm->name, namelen, name,
1857 assert->node_idx);
1858 __dlm_put_mle(mle);
1859 spin_unlock(&dlm->master_lock);
1860 spin_unlock(&dlm->spinlock);
1861 goto done;
2bd63216 1862 }
2d1a868c 1863 }
6714d8e8
KH		 1864 }
1865 spin_unlock(&dlm->master_lock);
1866
1867 /* ok everything checks out with the MLE
1868 * now check to see if there is a lockres */
a3d33291 1869 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
6714d8e8
KH		 1870 if (res) {
1871 spin_lock(&res->spinlock);
1872 if (res->state & DLM_LOCK_RES_RECOVERING) {
1873 mlog(ML_ERROR, "%u asserting but %.*s is "
1874 "RECOVERING!\n", assert->node_idx, namelen, name);
1875 goto kill;
1876 }
1877 if (!mle) {
dc2ed195
KH		 1878 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1879 res->owner != assert->node_idx) {
53ecd25e
SM		 1880 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1881 "but current owner is %u! (%.*s)\n",
1882 assert->node_idx, res->owner, namelen,
1883 name);
1884 __dlm_print_one_lock_resource(res);
1885 BUG();
6714d8e8
KH		 1886 }
1887 } else if (mle->type != DLM_MLE_MIGRATION) {
1888 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1889 /* owner is just re-asserting */
1890 if (res->owner == assert->node_idx) {
1891 mlog(0, "owner %u re-asserting on "
1892 "lock %.*s\n", assert->node_idx,
1893 namelen, name);
1894 goto ok;
1895 }
1896 mlog(ML_ERROR, "got assert_master from "
1897 "node %u, but %u is the owner! "
1898 "(%.*s)\n", assert->node_idx,
1899 res->owner, namelen, name);
1900 goto kill;
1901 }
1902 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1903 mlog(ML_ERROR, "got assert from %u, but lock "
1904 "with no owner should be "
1905 "in-progress! (%.*s)\n",
1906 assert->node_idx,
1907 namelen, name);
1908 goto kill;
1909 }
1910 } else /* mle->type == DLM_MLE_MIGRATION */ {
1911 /* should only be getting an assert from new master */
1912 if (assert->node_idx != mle->new_master) {
1913 mlog(ML_ERROR, "got assert from %u, but "
1914 "new master is %u, and old master "
1915 "was %u (%.*s)\n",
1916 assert->node_idx, mle->new_master,
1917 mle->master, namelen, name);
1918 goto kill;
1919 }
1920
1921 }
1922ok:
1923 spin_unlock(&res->spinlock);
1924 }
6714d8e8
KH		 1925
1926 // mlog(0, "woo! got an assert_master from node %u!\n",
1927 // assert->node_idx);
1928 if (mle) {
9c6510a5
KH		 1929 int extra_ref = 0;
1930 int nn = -1;
a2bf0477 1931 int rr, err = 0;
2bd63216 1932
6714d8e8 1933 spin_lock(&mle->spinlock);
9c6510a5
KH		 1934 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1935 extra_ref = 1;
1936 else {
1937 /* MASTER mle: if any bits set in the response map
1938 * then the calling node needs to re-assert to clear
1939 * up nodes that this node contacted */
2bd63216 1940 while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
9c6510a5 1941 nn+1)) < O2NM_MAX_NODES) {
728b9805 1942 if (nn != dlm->node_num && nn != assert->node_idx) {
9c6510a5 1943 master_request = 1;
728b9805
JB		 1944 break;
1945 }
9c6510a5
KH		 1946 }
1947 }
6714d8e8
KH		 1948 mle->master = assert->node_idx;
1949 atomic_set(&mle->woken, 1);
1950 wake_up(&mle->wq);
1951 spin_unlock(&mle->spinlock);
1952
a2bf0477 1953 if (res) {
a6fa3640 1954 int wake = 0;
6714d8e8 1955 spin_lock(&res->spinlock);
a2bf0477
KH		 1956 if (mle->type == DLM_MLE_MIGRATION) {
1957 mlog(0, "finishing off migration of lockres %.*s, "
1958 "from %u to %u\n",
1959 res->lockname.len, res->lockname.name,
1960 dlm->node_num, mle->new_master);
1961 res->state &= ~DLM_LOCK_RES_MIGRATING;
a6fa3640 1962 wake = 1;
a2bf0477
KH		 1963 dlm_change_lockres_owner(dlm, res, mle->new_master);
1964 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1965 } else {
1966 dlm_change_lockres_owner(dlm, res, mle->master);
1967 }
6714d8e8 1968 spin_unlock(&res->spinlock);
ba2bf218 1969 have_lockres_ref = 1;
a6fa3640
KH		 1970 if (wake)
1971 wake_up(&res->wq);
6714d8e8 1972 }
a2bf0477
KH		 1973
1974 /* master is known, detach if not already detached.
1975 * ensures that only one assert_master call will happen
1976 * on this mle. */
a2bf0477
KH		 1977 spin_lock(&dlm->master_lock);
1978
1979 rr = atomic_read(&mle->mle_refs.refcount);
1980 if (mle->inuse > 0) {
1981 if (extra_ref && rr < 3)
1982 err = 1;
1983 else if (!extra_ref && rr < 2)
1984 err = 1;
1985 } else {
1986 if (extra_ref && rr < 2)
1987 err = 1;
1988 else if (!extra_ref && rr < 1)
1989 err = 1;
1990 }
1991 if (err) {
1992 mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
1993 "that will mess up this node, refs=%d, extra=%d, "
1994 "inuse=%d\n", dlm->name, namelen, name,
1995 assert->node_idx, rr, extra_ref, mle->inuse);
1996 dlm_print_one_mle(mle);
1997 }
1c084577 1998 __dlm_unlink_mle(dlm, mle);
a2bf0477
KH		 1999 __dlm_mle_detach_hb_events(dlm, mle);
2000 __dlm_put_mle(mle);
6714d8e8
KH		 2001 if (extra_ref) {
2002 /* the assert master message now balances the extra
2003 * ref given by the master / migration request message.
2004 * if this is the last put, it will be removed
2005 * from the list. */
a2bf0477
KH		 2006 __dlm_put_mle(mle);
2007 }
2008 spin_unlock(&dlm->master_lock);
a2bf0477
KH		 2009 } else if (res) {
2010 if (res->owner != assert->node_idx) {
2011 mlog(0, "assert_master from %u, but current "
2012 "owner is %u (%.*s), no mle\n", assert->node_idx,
2013 res->owner, namelen, name);
6714d8e8
KH		 2014 }
2015 }
14741472 2016 spin_unlock(&dlm->spinlock);
6714d8e8
KH		 2017
2018done:
9c6510a5 2019 ret = 0;
3b8118cf
KH		 2020 if (res) {
2021 spin_lock(&res->spinlock);
2022 res->state |= DLM_LOCK_RES_SETREF_INPROG;
2023 spin_unlock(&res->spinlock);
2024 *ret_data = (void *)res;
2025 }
6714d8e8 2026 dlm_put(dlm);
9c6510a5
KH		 2027 if (master_request) {
2028 mlog(0, "need to tell master to reassert\n");
ba2bf218
KH		 2029 /* positive. negative would shoot down the node. */
2030 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2031 if (!have_lockres_ref) {
2032 mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2033 "mle present here for %s:%.*s, but no lockres!\n",
2034 assert->node_idx, dlm->name, namelen, name);
2035 }
2036 }
2037 if (have_lockres_ref) {
2038 /* let the master know we have a reference to the lockres */
2039 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2040 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2041 dlm->name, namelen, name, assert->node_idx);
9c6510a5
KH		 2042 }
2043 return ret;
6714d8e8
KH		 2044
2045kill:
2046 /* kill the caller! */
a9ee4c8a
KH		 2047 mlog(ML_ERROR, "Bad message received from another node. Dumping state "
2048 "and killing the other node now! This node is OK and can continue.\n");
2049 __dlm_print_one_lock_resource(res);
6714d8e8 2050 spin_unlock(&res->spinlock);
55dacd22 2051 spin_lock(&dlm->master_lock);
2052 if (mle)
2053 __dlm_put_mle(mle);
2054 spin_unlock(&dlm->master_lock);
6714d8e8 2055 spin_unlock(&dlm->spinlock);
2bd63216 2056 *ret_data = (void *)res;
6714d8e8
KH		 2057 dlm_put(dlm);
2058 return -EINVAL;
2059}
2060
3b8118cf
KH		 2061void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2062{
2063 struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2064
2065 if (ret_data) {
2066 spin_lock(&res->spinlock);
2067 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2068 spin_unlock(&res->spinlock);
2069 wake_up(&res->wq);
2070 dlm_lockres_put(res);
2071 }
2072 return;
2073}
2074
6714d8e8
KH		 2075int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2076 struct dlm_lock_resource *res,
2077 int ignore_higher, u8 request_from, u32 flags)
2078{
2079 struct dlm_work_item *item;
b24ae0b5 2080 item = kzalloc(sizeof(*item), GFP_ATOMIC);
6714d8e8
KH		 2081 if (!item)
2082 return -ENOMEM;
2083
2084
2085 /* queue up work for dlm_assert_master_worker */
2086 dlm_grab(dlm); /* get an extra ref for the work item */
2087 dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2088 item->u.am.lockres = res; /* already have a ref */
2089 /* can optionally ignore node numbers higher than this node */
2090 item->u.am.ignore_higher = ignore_higher;
2091 item->u.am.request_from = request_from;
2092 item->u.am.flags = flags;
2093
2bd63216
SM		 2094 if (ignore_higher)
2095 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
9c6510a5 2096 res->lockname.name);
2bd63216 2097
6714d8e8
KH		 2098 spin_lock(&dlm->work_lock);
2099 list_add_tail(&item->list, &dlm->work_list);
2100 spin_unlock(&dlm->work_lock);
2101
3156d267 2102 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
6714d8e8
KH		 2103 return 0;
2104}
2105
2106static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2107{
2108 struct dlm_ctxt *dlm = data;
2109 int ret = 0;
2110 struct dlm_lock_resource *res;
2111 unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2112 int ignore_higher;
2113 int bit;
2114 u8 request_from;
2115 u32 flags;
2116
2117 dlm = item->dlm;
2118 res = item->u.am.lockres;
2119 ignore_higher = item->u.am.ignore_higher;
2120 request_from = item->u.am.request_from;
2121 flags = item->u.am.flags;
2122
2123 spin_lock(&dlm->spinlock);
2124 memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2125 spin_unlock(&dlm->spinlock);
2126
2127 clear_bit(dlm->node_num, nodemap);
2128 if (ignore_higher) {
2129 /* if is this just to clear up mles for nodes below
2130 * this node, do not send the message to the original
2131 * caller or any node number higher than this */
2132 clear_bit(request_from, nodemap);
2133 bit = dlm->node_num;
2134 while (1) {
2135 bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2136 bit+1);
2137 if (bit >= O2NM_MAX_NODES)
2138 break;
2139 clear_bit(bit, nodemap);
2140 }
2141 }
2142
36407488
KH		 2143 /*
2144 * If we're migrating this lock to someone else, we are no
2145 * longer allowed to assert out own mastery. OTOH, we need to
2146 * prevent migration from starting while we're still asserting
2147 * our dominance. The reserved ast delays migration.
2148 */
2149 spin_lock(&res->spinlock);
2150 if (res->state & DLM_LOCK_RES_MIGRATING) {
2151 mlog(0, "Someone asked us to assert mastery, but we're "
2152 "in the middle of migration. Skipping assert, "
2153 "the new master will handle that.\n");
2154 spin_unlock(&res->spinlock);
2155 goto put;
2156 } else
2157 __dlm_lockres_reserve_ast(res);
2158 spin_unlock(&res->spinlock);
2159
6714d8e8
KH		 2160 /* this call now finishes out the nodemap
2161 * even if one or more nodes die */
2162 mlog(0, "worker about to master %.*s here, this=%u\n",
2163 res->lockname.len, res->lockname.name, dlm->node_num);
ba2bf218 2164 ret = dlm_do_assert_master(dlm, res, nodemap, flags);
6714d8e8
KH		 2165 if (ret < 0) {
2166 /* no need to restart, we are done */
3b3b84a8
KH		 2167 if (!dlm_is_host_down(ret))
2168 mlog_errno(ret);
6714d8e8
KH		 2169 }
2170
36407488
KH		 2171 /* Ok, we've asserted ourselves. Let's let migration start. */
2172 dlm_lockres_release_ast(dlm, res);
2173
2174put:
ac4fef4d
X		 2175 dlm_lockres_drop_inflight_worker(dlm, res);
2176
6714d8e8
KH		 2177 dlm_lockres_put(res);
2178
2179 mlog(0, "finished with dlm_assert_master_worker\n");
2180}
2181
c03872f5
KH		 2182/* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2183 * We cannot wait for node recovery to complete to begin mastering this
2184 * lockres because this lockres is used to kick off recovery! ;-)
2185 * So, do a pre-check on all living nodes to see if any of those nodes
2186 * think that $RECOVERY is currently mastered by a dead node. If so,
2187 * we wait a short time to allow that node to get notified by its own
2188 * heartbeat stack, then check again. All $RECOVERY lock resources
2bd63216 2189 * mastered by dead nodes are purged when the hearbeat callback is
c03872f5
KH		 2190 * fired, so we can know for sure that it is safe to continue once
2191 * the node returns a live node or no node. */
2192static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2193 struct dlm_lock_resource *res)
2194{
2195 struct dlm_node_iter iter;
2196 int nodenum;
2197 int ret = 0;
2198 u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2199
2200 spin_lock(&dlm->spinlock);
2201 dlm_node_iter_init(dlm->domain_map, &iter);
2202 spin_unlock(&dlm->spinlock);
2203
2204 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2205 /* do not send to self */
2206 if (nodenum == dlm->node_num)
2207 continue;
2208 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2209 if (ret < 0) {
2210 mlog_errno(ret);
2211 if (!dlm_is_host_down(ret))
2212 BUG();
2213 /* host is down, so answer for that node would be
2214 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
f42a100b 2215 ret = 0;
c03872f5
KH		 2216 }
2217
2218 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2219 /* check to see if this master is in the recovery map */
2220 spin_lock(&dlm->spinlock);
2221 if (test_bit(master, dlm->recovery_map)) {
2222 mlog(ML_NOTICE, "%s: node %u has not seen "
2223 "node %u go down yet, and thinks the "
2224 "dead node is mastering the recovery "
2225 "lock. must wait.\n", dlm->name,
2226 nodenum, master);
2227 ret = -EAGAIN;
2228 }
2229 spin_unlock(&dlm->spinlock);
2bd63216 2230 mlog(0, "%s: reco lock master is %u\n", dlm->name,
c03872f5
KH		 2231 master);
2232 break;
2233 }
2234 }
2235 return ret;
2236}
2237
ba2bf218
KH		 2238/*
2239 * DLM_DEREF_LOCKRES_MSG
2240 */
2241
2242int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2243{
2244 struct dlm_deref_lockres deref;
2245 int ret = 0, r;
2246 const char *lockname;
2247 unsigned int namelen;
2248
2249 lockname = res->lockname.name;
2250 namelen = res->lockname.len;
2251 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2252
ba2bf218
KH		 2253 memset(&deref, 0, sizeof(deref));
2254 deref.node_idx = dlm->node_num;
2255 deref.namelen = namelen;
2256 memcpy(deref.name, lockname, namelen);
2257
2258 ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2259 &deref, sizeof(deref), res->owner, &r);
2260 if (ret < 0)
8decab3c
SM		 2261 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n",
2262 dlm->name, namelen, lockname, ret, res->owner);
ba2bf218
KH		 2263 else if (r < 0) {
2264 /* BAD. other node says I did not have a ref. */
8decab3c
SM		 2265 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2266 dlm->name, namelen, lockname, res->owner, r);
ba2bf218
KH		 2267 dlm_print_one_lock_resource(res);
2268 BUG();
2269 }
2270 return ret;
2271}
2272
d74c9803
KH		 2273int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2274 void **ret_data)
ba2bf218
KH		 2275{
2276 struct dlm_ctxt *dlm = data;
2277 struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2278 struct dlm_lock_resource *res = NULL;
2279 char *name;
2280 unsigned int namelen;
2281 int ret = -EINVAL;
2282 u8 node;
2283 unsigned int hash;
f3f85464
SM		 2284 struct dlm_work_item *item;
2285 int cleared = 0;
2286 int dispatch = 0;
ba2bf218
KH		 2287
2288 if (!dlm_grab(dlm))
2289 return 0;
2290
2291 name = deref->name;
2292 namelen = deref->namelen;
2293 node = deref->node_idx;
2294
2295 if (namelen > DLM_LOCKID_NAME_MAX) {
2296 mlog(ML_ERROR, "Invalid name length!");
2297 goto done;
2298 }
2299 if (deref->node_idx >= O2NM_MAX_NODES) {
2300 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2301 goto done;
2302 }
2303
2304 hash = dlm_lockid_hash(name, namelen);
2305
2306 spin_lock(&dlm->spinlock);
2307 res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2308 if (!res) {
2309 spin_unlock(&dlm->spinlock);
2310 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2311 dlm->name, namelen, name);
2312 goto done;
2313 }
2314 spin_unlock(&dlm->spinlock);
2315
2316 spin_lock(&res->spinlock);
f3f85464
SM		 2317 if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2318 dispatch = 1;
2319 else {
2320 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2321 if (test_bit(node, res->refmap)) {
8d400b81 2322 dlm_lockres_clear_refmap_bit(dlm, res, node);
f3f85464
SM		 2323 cleared = 1;
2324 }
ba2bf218
KH		 2325 }
2326 spin_unlock(&res->spinlock);
2327
f3f85464
SM		 2328 if (!dispatch) {
2329 if (cleared)
2330 dlm_lockres_calc_usage(dlm, res);
2331 else {
2332 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2333 "but it is already dropped!\n", dlm->name,
2334 res->lockname.len, res->lockname.name, node);
2af37ce8 2335 dlm_print_one_lock_resource(res);
f3f85464
SM		 2336 }
2337 ret = 0;
2338 goto done;
2339 }
2340
2341 item = kzalloc(sizeof(*item), GFP_NOFS);
2342 if (!item) {
2343 ret = -ENOMEM;
2344 mlog_errno(ret);
2345 goto done;
2346 }
2347
2348 dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2349 item->u.dl.deref_res = res;
2350 item->u.dl.deref_node = node;
2351
2352 spin_lock(&dlm->work_lock);
2353 list_add_tail(&item->list, &dlm->work_list);
2354 spin_unlock(&dlm->work_lock);
2355
2356 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2357 return 0;
2358
ba2bf218
KH		 2359done:
2360 if (res)
2361 dlm_lockres_put(res);
2362 dlm_put(dlm);
f3f85464 2363
ba2bf218
KH		 2364 return ret;
2365}
2366
f3f85464
SM		 2367static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2368{
2369 struct dlm_ctxt *dlm;
2370 struct dlm_lock_resource *res;
2371 u8 node;
2372 u8 cleared = 0;
2373
2374 dlm = item->dlm;
2375 res = item->u.dl.deref_res;
2376 node = item->u.dl.deref_node;
2377
2378 spin_lock(&res->spinlock);
2379 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2380 if (test_bit(node, res->refmap)) {
2381 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
8d400b81 2382 dlm_lockres_clear_refmap_bit(dlm, res, node);
f3f85464
SM		 2383 cleared = 1;
2384 }
2385 spin_unlock(&res->spinlock);
2386
2387 if (cleared) {
2388 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2389 dlm->name, res->lockname.len, res->lockname.name, node);
2390 dlm_lockres_calc_usage(dlm, res);
2391 } else {
2392 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2393 "but it is already dropped!\n", dlm->name,
2394 res->lockname.len, res->lockname.name, node);
2af37ce8 2395 dlm_print_one_lock_resource(res);
f3f85464
SM		 2396 }
2397
2398 dlm_lockres_put(res);
2399}
2400
9f62e960
SM		 2401/*
2402 * A migrateable resource is one that is :
2403 * 1. locally mastered, and,
2404 * 2. zero local locks, and,
2405 * 3. one or more non-local locks, or, one or more references
2406 * Returns 1 if yes, 0 if not.
2f5bf1f2
SM		 2407 */
2408static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
9f62e960 2409 struct dlm_lock_resource *res)
2f5bf1f2 2410{
9f62e960
SM		 2411 enum dlm_lockres_list idx;
2412 int nonlocal = 0, node_ref;
800deef3 2413 struct list_head *queue;
2f5bf1f2 2414 struct dlm_lock *lock;
9f62e960 2415 u64 cookie;
2f5bf1f2
SM		 2416
2417 assert_spin_locked(&res->spinlock);
2418
fae477b6
X		 2419 /* delay migration when the lockres is in MIGRATING state */
2420 if (res->state & DLM_LOCK_RES_MIGRATING)
2421 return 0;
2422
bba1cb17
TS		 2423 /* delay migration when the lockres is in RECOCERING state */
2424 if (res->state & DLM_LOCK_RES_RECOVERING)
2425 return 0;
2426
9f62e960
SM		 2427 if (res->owner != dlm->node_num)
2428 return 0;
2f5bf1f2 2429
9f62e960
SM		 2430 for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2431 queue = dlm_list_idx_to_ptr(res, idx);
800deef3 2432 list_for_each_entry(lock, queue, list) {
9f62e960
SM		 2433 if (lock->ml.node != dlm->node_num) {
2434 nonlocal++;
2435 continue;
2f5bf1f2 2436 }
9f62e960
SM		 2437 cookie = be64_to_cpu(lock->ml.cookie);
2438 mlog(0, "%s: Not migrateable res %.*s, lock %u:%llu on "
2439 "%s list\n", dlm->name, res->lockname.len,
2440 res->lockname.name,
2441 dlm_get_lock_cookie_node(cookie),
2442 dlm_get_lock_cookie_seq(cookie),
2443 dlm_list_in_text(idx));
2444 return 0;
2f5bf1f2 2445 }
2f5bf1f2
SM		 2446 }
2447
9f62e960
SM		 2448 if (!nonlocal) {
2449 node_ref = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
2450 if (node_ref >= O2NM_MAX_NODES)
2451 return 0;
2452 }
388c4bcb 2453
9f62e960
SM		 2454 mlog(0, "%s: res %.*s, Migrateable\n", dlm->name, res->lockname.len,
2455 res->lockname.name);
2f5bf1f2 2456
9f62e960 2457 return 1;
2f5bf1f2 2458}
6714d8e8
KH		 2459
2460/*
2461 * DLM_MIGRATE_LOCKRES
2462 */
2463
2464
faf0ec9f 2465static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
66effd3c 2466 struct dlm_lock_resource *res, u8 target)
6714d8e8
KH		 2467{
2468 struct dlm_master_list_entry *mle = NULL;
2469 struct dlm_master_list_entry *oldmle = NULL;
2470 struct dlm_migratable_lockres *mres = NULL;
2f5bf1f2 2471 int ret = 0;
6714d8e8
KH		 2472 const char *name;
2473 unsigned int namelen;
2474 int mle_added = 0;
2f5bf1f2 2475 int wake = 0;
6714d8e8
KH		 2476
2477 if (!dlm_grab(dlm))
2478 return -EINVAL;
2479
66effd3c
SM		 2480 BUG_ON(target == O2NM_MAX_NODES);
2481
6714d8e8
KH		 2482 name = res->lockname.name;
2483 namelen = res->lockname.len;
2484
66effd3c
SM		 2485 mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name,
2486 target);
6714d8e8 2487
66effd3c 2488 /* preallocate up front. if this fails, abort */
6714d8e8 2489 ret = -ENOMEM;
ad8100e0 2490 mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
6714d8e8
KH		 2491 if (!mres) {
2492 mlog_errno(ret);
2493 goto leave;
2494 }
2495
3914ed0c 2496 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
6714d8e8
KH		 2497 if (!mle) {
2498 mlog_errno(ret);
2499 goto leave;
2500 }
2501 ret = 0;
2502
6714d8e8
KH		 2503 /*
2504 * clear any existing master requests and
2505 * add the migration mle to the list
2506 */
66effd3c 2507 spin_lock(&dlm->spinlock);
6714d8e8
KH		 2508 spin_lock(&dlm->master_lock);
2509 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2510 namelen, target, dlm->node_num);
2511 spin_unlock(&dlm->master_lock);
2512 spin_unlock(&dlm->spinlock);
2513
2514 if (ret == -EEXIST) {
2515 mlog(0, "another process is already migrating it\n");
2516 goto fail;
2517 }
2518 mle_added = 1;
2519
2520 /*
2521 * set the MIGRATING flag and flush asts
2522 * if we fail after this we need to re-dirty the lockres
2523 */
2524 if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2525 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2526 "the target went down.\n", res->lockname.len,
2527 res->lockname.name, target);
2528 spin_lock(&res->spinlock);
2529 res->state &= ~DLM_LOCK_RES_MIGRATING;
a6fa3640 2530 wake = 1;
6714d8e8
KH		 2531 spin_unlock(&res->spinlock);
2532 ret = -EINVAL;
2533 }
2534
2535fail:
2536 if (oldmle) {
2537 /* master is known, detach if not already detached */
2538 dlm_mle_detach_hb_events(dlm, oldmle);
2539 dlm_put_mle(oldmle);
2540 }
2541
2542 if (ret < 0) {
2543 if (mle_added) {
2544 dlm_mle_detach_hb_events(dlm, mle);
2545 dlm_put_mle(mle);
2546 } else if (mle) {
2547 kmem_cache_free(dlm_mle_cache, mle);
66effd3c 2548 mle = NULL;
6714d8e8
KH		 2549 }
2550 goto leave;
2551 }
2552
2553 /*
2554 * at this point, we have a migration target, an mle
2555 * in the master list, and the MIGRATING flag set on
2556 * the lockres
2557 */
2558
1cd04dbe
KH		 2559 /* now that remote nodes are spinning on the MIGRATING flag,
2560 * ensure that all assert_master work is flushed. */
2561 flush_workqueue(dlm->dlm_worker);
6714d8e8
KH		 2562
2563 /* get an extra reference on the mle.
2564 * otherwise the assert_master from the new
2565 * master will destroy this.
2566 * also, make sure that all callers of dlm_get_mle
2567 * take both dlm->spinlock and dlm->master_lock */
2568 spin_lock(&dlm->spinlock);
2569 spin_lock(&dlm->master_lock);
a2bf0477 2570 dlm_get_mle_inuse(mle);
6714d8e8
KH		 2571 spin_unlock(&dlm->master_lock);
2572 spin_unlock(&dlm->spinlock);
2573
2574 /* notify new node and send all lock state */
2575 /* call send_one_lockres with migration flag.
2576 * this serves as notice to the target node that a
2577 * migration is starting. */
2578 ret = dlm_send_one_lockres(dlm, res, mres, target,
2579 DLM_MRES_MIGRATION);
2580
2581 if (ret < 0) {
2582 mlog(0, "migration to node %u failed with %d\n",
2583 target, ret);
2584 /* migration failed, detach and clean up mle */
2585 dlm_mle_detach_hb_events(dlm, mle);
2586 dlm_put_mle(mle);
a2bf0477
KH		 2587 dlm_put_mle_inuse(mle);
2588 spin_lock(&res->spinlock);
2589 res->state &= ~DLM_LOCK_RES_MIGRATING;
a6fa3640 2590 wake = 1;
a2bf0477 2591 spin_unlock(&res->spinlock);
df016c66
SM		 2592 if (dlm_is_host_down(ret))
2593 dlm_wait_for_node_death(dlm, target,
2594 DLM_NODE_DEATH_WAIT_MAX);
6714d8e8
KH		 2595 goto leave;
2596 }
2597
2598 /* at this point, the target sends a message to all nodes,
2599 * (using dlm_do_migrate_request). this node is skipped since
2600 * we had to put an mle in the list to begin the process. this
2601 * node now waits for target to do an assert master. this node
2602 * will be the last one notified, ensuring that the migration
2603 * is complete everywhere. if the target dies while this is
2604 * going on, some nodes could potentially see the target as the
2605 * master, so it is important that my recovery finds the migration
af901ca1 2606 * mle and sets the master to UNKNOWN. */
6714d8e8
KH		 2607
2608
2609 /* wait for new node to assert master */
2610 while (1) {
2611 ret = wait_event_interruptible_timeout(mle->wq,
2612 (atomic_read(&mle->woken) == 1),
2613 msecs_to_jiffies(5000));
2614
2615 if (ret >= 0) {
2616 if (atomic_read(&mle->woken) == 1 ||
2617 res->owner == target)
2618 break;
2619
1cd04dbe
KH		 2620 mlog(0, "%s:%.*s: timed out during migration\n",
2621 dlm->name, res->lockname.len, res->lockname.name);
2bd63216 2622 /* avoid hang during shutdown when migrating lockres
e2faea4c
KH		 2623 * to a node which also goes down */
2624 if (dlm_is_node_dead(dlm, target)) {
aa852354
KH		 2625 mlog(0, "%s:%.*s: expected migration "
2626 "target %u is no longer up, restarting\n",
e2faea4c
KH		 2627 dlm->name, res->lockname.len,
2628 res->lockname.name, target);
1cd04dbe
KH		 2629 ret = -EINVAL;
2630 /* migration failed, detach and clean up mle */
2631 dlm_mle_detach_hb_events(dlm, mle);
2632 dlm_put_mle(mle);
2633 dlm_put_mle_inuse(mle);
2634 spin_lock(&res->spinlock);
2635 res->state &= ~DLM_LOCK_RES_MIGRATING;
a6fa3640 2636 wake = 1;
1cd04dbe
KH		 2637 spin_unlock(&res->spinlock);
2638 goto leave;
e2faea4c 2639 }
1cd04dbe
KH		 2640 } else
2641 mlog(0, "%s:%.*s: caught signal during migration\n",
2642 dlm->name, res->lockname.len, res->lockname.name);
6714d8e8
KH		 2643 }
2644
2645 /* all done, set the owner, clear the flag */
2646 spin_lock(&res->spinlock);
2647 dlm_set_lockres_owner(dlm, res, target);
2648 res->state &= ~DLM_LOCK_RES_MIGRATING;
2649 dlm_remove_nonlocal_locks(dlm, res);
2650 spin_unlock(&res->spinlock);
2651 wake_up(&res->wq);
2652
2653 /* master is known, detach if not already detached */
2654 dlm_mle_detach_hb_events(dlm, mle);
a2bf0477 2655 dlm_put_mle_inuse(mle);
6714d8e8
KH		 2656 ret = 0;
2657
2658 dlm_lockres_calc_usage(dlm, res);
2659
2660leave:
2661 /* re-dirty the lockres if we failed */
2662 if (ret < 0)
2663 dlm_kick_thread(dlm, res);
2664
a6fa3640
KH		 2665 /* wake up waiters if the MIGRATING flag got set
2666 * but migration failed */
2667 if (wake)
2668 wake_up(&res->wq);
2669
6714d8e8
KH		 2670 if (mres)
2671 free_page((unsigned long)mres);
2672
2673 dlm_put(dlm);
2674
9f62e960
SM		 2675 mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen,
2676 name, target, ret);
6714d8e8
KH		 2677 return ret;
2678}
6714d8e8 2679
ba2bf218
KH		 2680#define DLM_MIGRATION_RETRY_MS 100
2681
9f62e960
SM		 2682/*
2683 * Should be called only after beginning the domain leave process.
ba2bf218
KH		 2684 * There should not be any remaining locks on nonlocal lock resources,
2685 * and there should be no local locks left on locally mastered resources.
2686 *
2687 * Called with the dlm spinlock held, may drop it to do migration, but
2688 * will re-acquire before exit.
2689 *
9f62e960
SM		 2690 * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2691 */
ba2bf218
KH		 2692int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2693{
66effd3c 2694 int ret;
ba2bf218 2695 int lock_dropped = 0;
66effd3c 2696 u8 target = O2NM_MAX_NODES;
ba2bf218 2697
9f62e960 2698 assert_spin_locked(&dlm->spinlock);
2f5bf1f2 2699
9f62e960 2700 spin_lock(&res->spinlock);
66effd3c
SM		 2701 if (dlm_is_lockres_migrateable(dlm, res))
2702 target = dlm_pick_migration_target(dlm, res);
b36c3f84 2703 spin_unlock(&res->spinlock);
66effd3c
SM		 2704
2705 if (target == O2NM_MAX_NODES)
9f62e960 2706 goto leave;
ba2bf218
KH		 2707
2708 /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2709 spin_unlock(&dlm->spinlock);
2710 lock_dropped = 1;
66effd3c
SM		 2711 ret = dlm_migrate_lockres(dlm, res, target);
2712 if (ret)
2713 mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2714 dlm->name, res->lockname.len, res->lockname.name,
2715 target, ret);
ba2bf218
KH		 2716 spin_lock(&dlm->spinlock);
2717leave:
2718 return lock_dropped;
2719}
2720
6714d8e8
KH		 2721int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2722{
2723 int ret;
2724 spin_lock(&dlm->ast_lock);
2725 spin_lock(&lock->spinlock);
2726 ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2727 spin_unlock(&lock->spinlock);
2728 spin_unlock(&dlm->ast_lock);
2729 return ret;
2730}
2731
2732static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2733 struct dlm_lock_resource *res,
2734 u8 mig_target)
2735{
2736 int can_proceed;
2737 spin_lock(&res->spinlock);
2738 can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2739 spin_unlock(&res->spinlock);
2740
2bd63216 2741 /* target has died, so make the caller break out of the
6714d8e8
KH		 2742 * wait_event, but caller must recheck the domain_map */
2743 spin_lock(&dlm->spinlock);
2744 if (!test_bit(mig_target, dlm->domain_map))
2745 can_proceed = 1;
2746 spin_unlock(&dlm->spinlock);
2747 return can_proceed;
2748}
2749
faf0ec9f
AB		 2750static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2751 struct dlm_lock_resource *res)
6714d8e8
KH		 2752{
2753 int ret;
2754 spin_lock(&res->spinlock);
2755 ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2756 spin_unlock(&res->spinlock);
2757 return ret;
2758}
2759
2760
2761static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2762 struct dlm_lock_resource *res,
2763 u8 target)
2764{
2765 int ret = 0;
2766
2767 mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2768 res->lockname.len, res->lockname.name, dlm->node_num,
2769 target);
2770 /* need to set MIGRATING flag on lockres. this is done by
2771 * ensuring that all asts have been flushed for this lockres. */
2772 spin_lock(&res->spinlock);
2773 BUG_ON(res->migration_pending);
2774 res->migration_pending = 1;
2775 /* strategy is to reserve an extra ast then release
2776 * it below, letting the release do all of the work */
2777 __dlm_lockres_reserve_ast(res);
2778 spin_unlock(&res->spinlock);
2779
ddc09c8d 2780 /* now flush all the pending asts */
6714d8e8 2781 dlm_kick_thread(dlm, res);
ddc09c8d
KH		 2782 /* before waiting on DIRTY, block processes which may
2783 * try to dirty the lockres before MIGRATING is set */
2784 spin_lock(&res->spinlock);
2785 BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2786 res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2787 spin_unlock(&res->spinlock);
2788 /* now wait on any pending asts and the DIRTY state */
6714d8e8
KH		 2789 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2790 dlm_lockres_release_ast(dlm, res);
2791
2792 mlog(0, "about to wait on migration_wq, dirty=%s\n",
2793 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2794 /* if the extra ref we just put was the final one, this
2795 * will pass thru immediately. otherwise, we need to wait
2796 * for the last ast to finish. */
2797again:
2798 ret = wait_event_interruptible_timeout(dlm->migration_wq,
2799 dlm_migration_can_proceed(dlm, res, target),
2800 msecs_to_jiffies(1000));
2801 if (ret < 0) {
2802 mlog(0, "woken again: migrating? %s, dead? %s\n",
2803 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2804 test_bit(target, dlm->domain_map) ? "no":"yes");
2805 } else {
2806 mlog(0, "all is well: migrating? %s, dead? %s\n",
2807 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2808 test_bit(target, dlm->domain_map) ? "no":"yes");
2809 }
2810 if (!dlm_migration_can_proceed(dlm, res, target)) {
2811 mlog(0, "trying again...\n");
2812 goto again;
2813 }
2814
a39953dd 2815 ret = 0;
6714d8e8
KH		 2816 /* did the target go down or die? */
2817 spin_lock(&dlm->spinlock);
2818 if (!test_bit(target, dlm->domain_map)) {
2819 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2820 target);
2821 ret = -EHOSTDOWN;
2822 }
2823 spin_unlock(&dlm->spinlock);
2824
a39953dd
WW		 2825 /*
2826 * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2827 * another try; otherwise, we are sure the MIGRATING state is there,
2828 * drop the unneded state which blocked threads trying to DIRTY
2829 */
2830 spin_lock(&res->spinlock);
2831 BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2832 res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2833 if (!ret)
2834 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2835 spin_unlock(&res->spinlock);
2836
6714d8e8
KH		 2837 /*
2838 * at this point:
2839 *
a39953dd 2840 * o the DLM_LOCK_RES_MIGRATING flag is set if target not down
6714d8e8
KH		 2841 * o there are no pending asts on this lockres
2842 * o all processes trying to reserve an ast on this
2843 * lockres must wait for the MIGRATING flag to clear
2844 */
2845 return ret;
2846}
2847
2848/* last step in the migration process.
2849 * original master calls this to free all of the dlm_lock
2850 * structures that used to be for other nodes. */
2851static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2852 struct dlm_lock_resource *res)
2853{
6714d8e8 2854 struct list_head *queue = &res->granted;
ba2bf218 2855 int i, bit;
800deef3 2856 struct dlm_lock *lock, *next;
6714d8e8
KH		 2857
2858 assert_spin_locked(&res->spinlock);
2859
2860 BUG_ON(res->owner == dlm->node_num);
2861
2862 for (i=0; i<3; i++) {
800deef3 2863 list_for_each_entry_safe(lock, next, queue, list) {
6714d8e8
KH		 2864 if (lock->ml.node != dlm->node_num) {
2865 mlog(0, "putting lock for node %u\n",
2866 lock->ml.node);
2867 /* be extra careful */
2868 BUG_ON(!list_empty(&lock->ast_list));
2869 BUG_ON(!list_empty(&lock->bast_list));
2870 BUG_ON(lock->ast_pending);
2871 BUG_ON(lock->bast_pending);
8d400b81
SM		 2872 dlm_lockres_clear_refmap_bit(dlm, res,
2873 lock->ml.node);
6714d8e8
KH		 2874 list_del_init(&lock->list);
2875 dlm_lock_put(lock);
2c5c54ac
SM		 2876 /* In a normal unlock, we would have added a
2877 * DLM_UNLOCK_FREE_LOCK action. Force it. */
2878 dlm_lock_put(lock);
6714d8e8
KH		 2879 }
2880 }
2881 queue++;
2882 }
ba2bf218
KH		 2883 bit = 0;
2884 while (1) {
2885 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2886 if (bit >= O2NM_MAX_NODES)
2887 break;
2888 /* do not clear the local node reference, if there is a
2889 * process holding this, let it drop the ref itself */
2890 if (bit != dlm->node_num) {
2891 mlog(0, "%s:%.*s: node %u had a ref to this "
2892 "migrating lockres, clearing\n", dlm->name,
2893 res->lockname.len, res->lockname.name, bit);
8d400b81 2894 dlm_lockres_clear_refmap_bit(dlm, res, bit);
ba2bf218
KH		 2895 }
2896 bit++;
2897 }
6714d8e8
KH		 2898}
2899
66effd3c
SM		 2900/*
2901 * Pick a node to migrate the lock resource to. This function selects a
2902 * potential target based first on the locks and then on refmap. It skips
2903 * nodes that are in the process of exiting the domain.
2904 */
6714d8e8
KH		 2905static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2906 struct dlm_lock_resource *res)
2907{
66effd3c 2908 enum dlm_lockres_list idx;
6714d8e8 2909 struct list_head *queue = &res->granted;
6714d8e8 2910 struct dlm_lock *lock;
66effd3c
SM		 2911 int noderef;
2912 u8 nodenum = O2NM_MAX_NODES;
6714d8e8
KH		 2913
2914 assert_spin_locked(&dlm->spinlock);
66effd3c 2915 assert_spin_locked(&res->spinlock);
6714d8e8 2916
66effd3c
SM		 2917 /* Go through all the locks */
2918 for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2919 queue = dlm_list_idx_to_ptr(res, idx);
800deef3 2920 list_for_each_entry(lock, queue, list) {
66effd3c
SM		 2921 if (lock->ml.node == dlm->node_num)
2922 continue;
2923 if (test_bit(lock->ml.node, dlm->exit_domain_map))
2924 continue;
2925 nodenum = lock->ml.node;
2926 goto bail;
6714d8e8 2927 }
6714d8e8 2928 }
388c4bcb 2929
66effd3c
SM		 2930 /* Go thru the refmap */
2931 noderef = -1;
6714d8e8 2932 while (1) {
66effd3c
SM		 2933 noderef = find_next_bit(res->refmap, O2NM_MAX_NODES,
2934 noderef + 1);
2935 if (noderef >= O2NM_MAX_NODES)
6714d8e8 2936 break;
66effd3c
SM		 2937 if (noderef == dlm->node_num)
2938 continue;
2939 if (test_bit(noderef, dlm->exit_domain_map))
2940 continue;
2941 nodenum = noderef;
2942 goto bail;
6714d8e8
KH		 2943 }
2944
66effd3c
SM		 2945bail:
2946 return nodenum;
6714d8e8
KH		 2947}
2948
6714d8e8
KH		 2949/* this is called by the new master once all lockres
2950 * data has been received */
2951static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2952 struct dlm_lock_resource *res,
2953 u8 master, u8 new_master,
2954 struct dlm_node_iter *iter)
2955{
2956 struct dlm_migrate_request migrate;
2b832564 2957 int ret, skip, status = 0;
6714d8e8
KH		 2958 int nodenum;
2959
2960 memset(&migrate, 0, sizeof(migrate));
2961 migrate.namelen = res->lockname.len;
2962 memcpy(migrate.name, res->lockname.name, migrate.namelen);
2963 migrate.new_master = new_master;
2964 migrate.master = master;
2965
2966 ret = 0;
2967
2968 /* send message to all nodes, except the master and myself */
2969 while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2970 if (nodenum == master ||
2971 nodenum == new_master)
2972 continue;
2973
2b832564
SM		 2974 /* We could race exit domain. If exited, skip. */
2975 spin_lock(&dlm->spinlock);
2976 skip = (!test_bit(nodenum, dlm->domain_map));
2977 spin_unlock(&dlm->spinlock);
2978 if (skip) {
2979 clear_bit(nodenum, iter->node_map);
2980 continue;
2981 }
2982
6714d8e8
KH		 2983 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2984 &migrate, sizeof(migrate), nodenum,
2985 &status);
2b832564 2986 if (ret < 0) {
8decab3c
SM		 2987 mlog(ML_ERROR, "%s: res %.*s, Error %d send "
2988 "MIGRATE_REQUEST to node %u\n", dlm->name,
2989 migrate.namelen, migrate.name, ret, nodenum);
2b832564
SM		 2990 if (!dlm_is_host_down(ret)) {
2991 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
2992 BUG();
2993 }
2994 clear_bit(nodenum, iter->node_map);
2995 ret = 0;
2996 } else if (status < 0) {
6714d8e8
KH		 2997 mlog(0, "migrate request (node %u) returned %d!\n",
2998 nodenum, status);
2999 ret = status;
ba2bf218
KH		 3000 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
3001 /* during the migration request we short-circuited
3002 * the mastery of the lockres. make sure we have
3003 * a mastery ref for nodenum */
3004 mlog(0, "%s:%.*s: need ref for node %u\n",
3005 dlm->name, res->lockname.len, res->lockname.name,
3006 nodenum);
3007 spin_lock(&res->spinlock);
8d400b81 3008 dlm_lockres_set_refmap_bit(dlm, res, nodenum);
ba2bf218 3009 spin_unlock(&res->spinlock);
6714d8e8
KH		 3010 }
3011 }
3012
3013 if (ret < 0)
3014 mlog_errno(ret);
3015
3016 mlog(0, "returning ret=%d\n", ret);
3017 return ret;
3018}
3019
3020
3021/* if there is an existing mle for this lockres, we now know who the master is.
3022 * (the one who sent us *this* message) we can clear it up right away.
3023 * since the process that put the mle on the list still has a reference to it,
3024 * we can unhash it now, set the master and wake the process. as a result,
3025 * we will have no mle in the list to start with. now we can add an mle for
3026 * the migration and this should be the only one found for those scanning the
3027 * list. */
d74c9803
KH		 3028int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
3029 void **ret_data)
6714d8e8
KH		 3030{
3031 struct dlm_ctxt *dlm = data;
3032 struct dlm_lock_resource *res = NULL;
3033 struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
3034 struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
3035 const char *name;
a3d33291 3036 unsigned int namelen, hash;
6714d8e8
KH		 3037 int ret = 0;
3038
3039 if (!dlm_grab(dlm))
3040 return -EINVAL;
3041
3042 name = migrate->name;
3043 namelen = migrate->namelen;
a3d33291 3044 hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 3045
3046 /* preallocate.. if this fails, abort */
3914ed0c 3047 mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
6714d8e8
KH		 3048
3049 if (!mle) {
3050 ret = -ENOMEM;
3051 goto leave;
3052 }
3053
3054 /* check for pre-existing lock */
3055 spin_lock(&dlm->spinlock);
a3d33291 3056 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
6714d8e8
KH		 3057 if (res) {
3058 spin_lock(&res->spinlock);
3059 if (res->state & DLM_LOCK_RES_RECOVERING) {
3060 /* if all is working ok, this can only mean that we got
3061 * a migrate request from a node that we now see as
3062 * dead. what can we do here? drop it to the floor? */
3063 spin_unlock(&res->spinlock);
3064 mlog(ML_ERROR, "Got a migrate request, but the "
3065 "lockres is marked as recovering!");
3066 kmem_cache_free(dlm_mle_cache, mle);
3067 ret = -EINVAL; /* need a better solution */
3068 goto unlock;
3069 }
3070 res->state |= DLM_LOCK_RES_MIGRATING;
3071 spin_unlock(&res->spinlock);
3072 }
3073
6d98c3cc 3074 spin_lock(&dlm->master_lock);
6714d8e8
KH		 3075 /* ignore status. only nonzero status would BUG. */
3076 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3077 name, namelen,
3078 migrate->new_master,
3079 migrate->master);
3080
6714d8e8 3081 spin_unlock(&dlm->master_lock);
6d98c3cc 3082unlock:
6714d8e8
KH		 3083 spin_unlock(&dlm->spinlock);
3084
3085 if (oldmle) {
3086 /* master is known, detach if not already detached */
3087 dlm_mle_detach_hb_events(dlm, oldmle);
3088 dlm_put_mle(oldmle);
3089 }
3090
3091 if (res)
3092 dlm_lockres_put(res);
3093leave:
3094 dlm_put(dlm);
3095 return ret;
3096}
3097
3098/* must be holding dlm->spinlock and dlm->master_lock
3099 * when adding a migration mle, we can clear any other mles
3100 * in the master list because we know with certainty that
3101 * the master is "master". so we remove any old mle from
3102 * the list after setting it's master field, and then add
3103 * the new migration mle. this way we can hold with the rule
3104 * of having only one mle for a given lock name at all times. */
3105static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3106 struct dlm_lock_resource *res,
3107 struct dlm_master_list_entry *mle,
3108 struct dlm_master_list_entry **oldmle,
3109 const char *name, unsigned int namelen,
3110 u8 new_master, u8 master)
3111{
3112 int found;
3113 int ret = 0;
3114
3115 *oldmle = NULL;
3116
6714d8e8
KH		 3117 assert_spin_locked(&dlm->spinlock);
3118 assert_spin_locked(&dlm->master_lock);
3119
3120 /* caller is responsible for any ref taken here on oldmle */
3121 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3122 if (found) {
3123 struct dlm_master_list_entry *tmp = *oldmle;
3124 spin_lock(&tmp->spinlock);
3125 if (tmp->type == DLM_MLE_MIGRATION) {
3126 if (master == dlm->node_num) {
3127 /* ah another process raced me to it */
3128 mlog(0, "tried to migrate %.*s, but some "
3129 "process beat me to it\n",
3130 namelen, name);
3131 ret = -EEXIST;
3132 } else {
3133 /* bad. 2 NODES are trying to migrate! */
3134 mlog(ML_ERROR, "migration error mle: "
3135 "master=%u new_master=%u // request: "
3136 "master=%u new_master=%u // "
3137 "lockres=%.*s\n",
3138 tmp->master, tmp->new_master,
3139 master, new_master,
3140 namelen, name);
3141 BUG();
3142 }
3143 } else {
3144 /* this is essentially what assert_master does */
3145 tmp->master = master;
3146 atomic_set(&tmp->woken, 1);
3147 wake_up(&tmp->wq);
1c084577
SM		 3148 /* remove it so that only one mle will be found */
3149 __dlm_unlink_mle(dlm, tmp);
ba2bf218 3150 __dlm_mle_detach_hb_events(dlm, tmp);
b9aaac5a 3151 if (tmp->type == DLM_MLE_MASTER) {
3152 ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3153 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3154 "telling master to get ref "
3155 "for cleared out mle during "
3156 "migration\n", dlm->name,
3157 namelen, name, master,
3158 new_master);
3159 }
6714d8e8
KH		 3160 }
3161 spin_unlock(&tmp->spinlock);
3162 }
3163
3164 /* now add a migration mle to the tail of the list */
3165 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3166 mle->new_master = new_master;
ba2bf218
KH		 3167 /* the new master will be sending an assert master for this.
3168 * at that point we will get the refmap reference */
6714d8e8
KH		 3169 mle->master = master;
3170 /* do this for consistency with other mle types */
3171 set_bit(new_master, mle->maybe_map);
1c084577 3172 __dlm_insert_mle(dlm, mle);
6714d8e8
KH		 3173
3174 return ret;
3175}
3176
c2cd4a44
SM		 3177/*
3178 * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3179 */
3180static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3181 struct dlm_master_list_entry *mle)
3182{
3183 struct dlm_lock_resource *res;
c2cd4a44
SM		 3184
3185 /* Find the lockres associated to the mle and set its owner to UNK */
7141514b
SM		 3186 res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3187 mle->mnamehash);
c2cd4a44
SM		 3188 if (res) {
3189 spin_unlock(&dlm->master_lock);
3190
3191 /* move lockres onto recovery list */
3192 spin_lock(&res->spinlock);
3193 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3194 dlm_move_lockres_to_recovery_list(dlm, res);
3195 spin_unlock(&res->spinlock);
3196 dlm_lockres_put(res);
3197
3198 /* about to get rid of mle, detach from heartbeat */
3199 __dlm_mle_detach_hb_events(dlm, mle);
3200
3201 /* dump the mle */
3202 spin_lock(&dlm->master_lock);
3203 __dlm_put_mle(mle);
3204 spin_unlock(&dlm->master_lock);
3205 }
3206
3207 return res;
3208}
3209
3210static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3211 struct dlm_master_list_entry *mle)
3212{
3213 __dlm_mle_detach_hb_events(dlm, mle);
3214
3215 spin_lock(&mle->spinlock);
3216 __dlm_unlink_mle(dlm, mle);
3217 atomic_set(&mle->woken, 1);
3218 spin_unlock(&mle->spinlock);
3219
3220 wake_up(&mle->wq);
3221}
3222
3223static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3224 struct dlm_master_list_entry *mle, u8 dead_node)
3225{
3226 int bit;
3227
3228 BUG_ON(mle->type != DLM_MLE_BLOCK);
3229
3230 spin_lock(&mle->spinlock);
3231 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3232 if (bit != dead_node) {
3233 mlog(0, "mle found, but dead node %u would not have been "
3234 "master\n", dead_node);
3235 spin_unlock(&mle->spinlock);
3236 } else {
3237 /* Must drop the refcount by one since the assert_master will
3238 * never arrive. This may result in the mle being unlinked and
3239 * freed, but there may still be a process waiting in the
3240 * dlmlock path which is fine. */
3241 mlog(0, "node %u was expected master\n", dead_node);
3242 atomic_set(&mle->woken, 1);
3243 spin_unlock(&mle->spinlock);
3244 wake_up(&mle->wq);
3245
3246 /* Do not need events any longer, so detach from heartbeat */
3247 __dlm_mle_detach_hb_events(dlm, mle);
3248 __dlm_put_mle(mle);
3249 }
3250}
6714d8e8
KH		 3251
3252void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3253{
2ed6c750 3254 struct dlm_master_list_entry *mle;
6714d8e8 3255 struct dlm_lock_resource *res;
2ed6c750 3256 struct hlist_head *bucket;
df53cd3b 3257 struct hlist_node *tmp;
2ed6c750 3258 unsigned int i;
6714d8e8 3259
ef6b689b 3260 mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
6714d8e8
KH		 3261top:
3262 assert_spin_locked(&dlm->spinlock);
3263
3264 /* clean the master list */
3265 spin_lock(&dlm->master_lock);
2ed6c750
SM		 3266 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3267 bucket = dlm_master_hash(dlm, i);
df53cd3b 3268 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
67ae1f06
SM		 3269 BUG_ON(mle->type != DLM_MLE_BLOCK &&
3270 mle->type != DLM_MLE_MASTER &&
3271 mle->type != DLM_MLE_MIGRATION);
3272
3273 /* MASTER mles are initiated locally. The waiting
3274 * process will notice the node map change shortly.
3275 * Let that happen as normal. */
3276 if (mle->type == DLM_MLE_MASTER)
3277 continue;
3278
3279 /* BLOCK mles are initiated by other nodes. Need to
3280 * clean up if the dead node would have been the
3281 * master. */
3282 if (mle->type == DLM_MLE_BLOCK) {
3283 dlm_clean_block_mle(dlm, mle, dead_node);
3284 continue;
3285 }
6714d8e8 3286
67ae1f06
SM		 3287 /* Everything else is a MIGRATION mle */
3288
3289 /* The rule for MIGRATION mles is that the master
3290 * becomes UNKNOWN if *either* the original or the new
3291 * master dies. All UNKNOWN lockres' are sent to
3292 * whichever node becomes the recovery master. The new
3293 * master is responsible for determining if there is
3294 * still a master for this lockres, or if he needs to
3295 * take over mastery. Either way, this node should
3296 * expect another message to resolve this. */
3297
3298 if (mle->master != dead_node &&
3299 mle->new_master != dead_node)
3300 continue;
3301
3302 /* If we have reached this point, this mle needs to be
3303 * removed from the list and freed. */
3304 dlm_clean_migration_mle(dlm, mle);
3305
3306 mlog(0, "%s: node %u died during migration from "
3307 "%u to %u!\n", dlm->name, dead_node, mle->master,
3308 mle->new_master);
3309
3310 /* If we find a lockres associated with the mle, we've
3311 * hit this rare case that messes up our lock ordering.
3312 * If so, we need to drop the master lock so that we can
3313 * take the lockres lock, meaning that we will have to
3314 * restart from the head of list. */
3315 res = dlm_reset_mleres_owner(dlm, mle);
3316 if (res)
3317 /* restart */
3318 goto top;
3319
3320 /* This may be the last reference */
3321 __dlm_put_mle(mle);
6714d8e8 3322 }
2ed6c750 3323 }
6714d8e8
KH		 3324 spin_unlock(&dlm->master_lock);
3325}
3326
6714d8e8
KH		 3327int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3328 u8 old_master)
3329{
3330 struct dlm_node_iter iter;
3331 int ret = 0;
3332
3333 spin_lock(&dlm->spinlock);
3334 dlm_node_iter_init(dlm->domain_map, &iter);
3335 clear_bit(old_master, iter.node_map);
3336 clear_bit(dlm->node_num, iter.node_map);
3337 spin_unlock(&dlm->spinlock);
3338
ba2bf218
KH		 3339 /* ownership of the lockres is changing. account for the
3340 * mastery reference here since old_master will briefly have
3341 * a reference after the migration completes */
3342 spin_lock(&res->spinlock);
8d400b81 3343 dlm_lockres_set_refmap_bit(dlm, res, old_master);
ba2bf218
KH		 3344 spin_unlock(&res->spinlock);
3345
6714d8e8
KH		 3346 mlog(0, "now time to do a migrate request to other nodes\n");
3347 ret = dlm_do_migrate_request(dlm, res, old_master,
3348 dlm->node_num, &iter);
3349 if (ret < 0) {
3350 mlog_errno(ret);
3351 goto leave;
3352 }
3353
3354 mlog(0, "doing assert master of %.*s to all except the original node\n",
3355 res->lockname.len, res->lockname.name);
3356 /* this call now finishes out the nodemap
3357 * even if one or more nodes die */
ba2bf218 3358 ret = dlm_do_assert_master(dlm, res, iter.node_map,
6714d8e8
KH		 3359 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3360 if (ret < 0) {
3361 /* no longer need to retry. all living nodes contacted. */
3362 mlog_errno(ret);
3363 ret = 0;
3364 }
3365
3366 memset(iter.node_map, 0, sizeof(iter.node_map));
3367 set_bit(old_master, iter.node_map);
3368 mlog(0, "doing assert master of %.*s back to %u\n",
3369 res->lockname.len, res->lockname.name, old_master);
ba2bf218 3370 ret = dlm_do_assert_master(dlm, res, iter.node_map,
6714d8e8
KH		 3371 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3372 if (ret < 0) {
3373 mlog(0, "assert master to original master failed "
3374 "with %d.\n", ret);
3375 /* the only nonzero status here would be because of
3376 * a dead original node. we're done. */
3377 ret = 0;
3378 }
3379
3380 /* all done, set the owner, clear the flag */
3381 spin_lock(&res->spinlock);
3382 dlm_set_lockres_owner(dlm, res, dlm->node_num);
3383 res->state &= ~DLM_LOCK_RES_MIGRATING;
3384 spin_unlock(&res->spinlock);
3385 /* re-dirty it on the new master */
3386 dlm_kick_thread(dlm, res);
3387 wake_up(&res->wq);
3388leave:
3389 return ret;
3390}
3391
3392/*
3393 * LOCKRES AST REFCOUNT
3394 * this is integral to migration
3395 */
3396
3397/* for future intent to call an ast, reserve one ahead of time.
3398 * this should be called only after waiting on the lockres
3399 * with dlm_wait_on_lockres, and while still holding the
3400 * spinlock after the call. */
3401void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3402{
3403 assert_spin_locked(&res->spinlock);
3404 if (res->state & DLM_LOCK_RES_MIGRATING) {
3405 __dlm_print_one_lock_resource(res);
3406 }
3407 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3408
3409 atomic_inc(&res->asts_reserved);
3410}
3411
3412/*
3413 * used to drop the reserved ast, either because it went unused,
3414 * or because the ast/bast was actually called.
3415 *
3416 * also, if there is a pending migration on this lockres,
3417 * and this was the last pending ast on the lockres,
3418 * atomically set the MIGRATING flag before we drop the lock.
3419 * this is how we ensure that migration can proceed with no
3420 * asts in progress. note that it is ok if the state of the
3421 * queues is such that a lock should be granted in the future
3422 * or that a bast should be fired, because the new master will
3423 * shuffle the lists on this lockres as soon as it is migrated.
3424 */
3425void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3426 struct dlm_lock_resource *res)
3427{
3428 if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3429 return;
3430
3431 if (!res->migration_pending) {
3432 spin_unlock(&res->spinlock);
3433 return;
3434 }
3435
3436 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3437 res->migration_pending = 0;
3438 res->state |= DLM_LOCK_RES_MIGRATING;
3439 spin_unlock(&res->spinlock);
3440 wake_up(&res->wq);
3441 wake_up(&dlm->migration_wq);
3442}
5dad6c39
SE		 3443
3444void dlm_force_free_mles(struct dlm_ctxt *dlm)
3445{
3446 int i;
3447 struct hlist_head *bucket;
3448 struct dlm_master_list_entry *mle;
df53cd3b 3449 struct hlist_node *tmp;
5dad6c39
SE		 3450
3451 /*
3452 * We notified all other nodes that we are exiting the domain and
3453 * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3454 * around we force free them and wake any processes that are waiting
3455 * on the mles
3456 */
3457 spin_lock(&dlm->spinlock);
3458 spin_lock(&dlm->master_lock);
3459
3460 BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3461 BUG_ON((find_next_bit(dlm->domain_map, O2NM_MAX_NODES, 0) < O2NM_MAX_NODES));
3462
3463 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3464 bucket = dlm_master_hash(dlm, i);
df53cd3b 3465 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
5dad6c39
SE		 3466 if (mle->type != DLM_MLE_BLOCK) {
3467 mlog(ML_ERROR, "bad mle: %p\n", mle);
3468 dlm_print_one_mle(mle);
3469 }
3470 atomic_set(&mle->woken, 1);
3471 wake_up(&mle->wq);
3472
3473 __dlm_unlink_mle(dlm, mle);
3474 __dlm_mle_detach_hb_events(dlm, mle);
3475 __dlm_put_mle(mle);
3476 }
3477 }
3478 spin_unlock(&dlm->master_lock);
3479 spin_unlock(&dlm->spinlock);
3480}
Linux kernel - Deliverables
This page took 0.845396 seconds and 5 git commands to generate.