projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
ocfs2_dlm: Dlm dispatch was stopping too early
[linux-2.6-block.git] / fs / ocfs2 / dlm / dlmmaster.c
CommitLineData
6714d8e8
KH		 1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmmod.c
5 *
6 * standalone DLM module
7 *
8 * Copyright (C) 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 *
25 */
26
27
28#include <linux/module.h>
29#include <linux/fs.h>
30#include <linux/types.h>
31#include <linux/slab.h>
32#include <linux/highmem.h>
33#include <linux/utsname.h>
34#include <linux/init.h>
35#include <linux/sysctl.h>
36#include <linux/random.h>
37#include <linux/blkdev.h>
38#include <linux/socket.h>
39#include <linux/inet.h>
40#include <linux/spinlock.h>
41#include <linux/delay.h>
42
43
44#include "cluster/heartbeat.h"
45#include "cluster/nodemanager.h"
46#include "cluster/tcp.h"
47
48#include "dlmapi.h"
49#include "dlmcommon.h"
82353b59 50#include "dlmdomain.h"
6714d8e8
KH		 51
52#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
53#include "cluster/masklog.h"
54
55enum dlm_mle_type {
56 DLM_MLE_BLOCK,
57 DLM_MLE_MASTER,
58 DLM_MLE_MIGRATION
59};
60
61struct dlm_lock_name
62{
63 u8 len;
64 u8 name[DLM_LOCKID_NAME_MAX];
65};
66
67struct dlm_master_list_entry
68{
69 struct list_head list;
70 struct list_head hb_events;
71 struct dlm_ctxt *dlm;
72 spinlock_t spinlock;
73 wait_queue_head_t wq;
74 atomic_t woken;
75 struct kref mle_refs;
a2bf0477 76 int inuse;
6714d8e8
KH		 77 unsigned long maybe_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
78 unsigned long vote_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
79 unsigned long response_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
80 unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
81 u8 master;
82 u8 new_master;
83 enum dlm_mle_type type;
84 struct o2hb_callback_func mle_hb_up;
85 struct o2hb_callback_func mle_hb_down;
86 union {
87 struct dlm_lock_resource *res;
88 struct dlm_lock_name name;
89 } u;
90};
91
92static void dlm_mle_node_down(struct dlm_ctxt *dlm,
93 struct dlm_master_list_entry *mle,
94 struct o2nm_node *node,
95 int idx);
96static void dlm_mle_node_up(struct dlm_ctxt *dlm,
97 struct dlm_master_list_entry *mle,
98 struct o2nm_node *node,
99 int idx);
100
101static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
ba2bf218
KH		 102static int dlm_do_assert_master(struct dlm_ctxt *dlm,
103 struct dlm_lock_resource *res,
104 void *nodemap, u32 flags);
6714d8e8
KH		 105
106static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
107 struct dlm_master_list_entry *mle,
108 const char *name,
109 unsigned int namelen)
110{
111 struct dlm_lock_resource *res;
112
113 if (dlm != mle->dlm)
114 return 0;
115
116 if (mle->type == DLM_MLE_BLOCK ||
117 mle->type == DLM_MLE_MIGRATION) {
118 if (namelen != mle->u.name.len ||
119 memcmp(name, mle->u.name.name, namelen)!=0)
120 return 0;
121 } else {
122 res = mle->u.res;
123 if (namelen != res->lockname.len ||
124 memcmp(res->lockname.name, name, namelen) != 0)
125 return 0;
126 }
127 return 1;
128}
129
95883719 130#define dlm_print_nodemap(m) _dlm_print_nodemap(m,#m)
8a9343fa 131static void _dlm_print_nodemap(unsigned long *map, const char *mapname)
95883719
KH		 132{
133 int i;
134 printk("%s=[ ", mapname);
135 for (i=0; i<O2NM_MAX_NODES; i++)
136 if (test_bit(i, map))
137 printk("%d ", i);
138 printk("]");
139}
140
8a9343fa 141static void dlm_print_one_mle(struct dlm_master_list_entry *mle)
6714d8e8 142{
95883719 143 int refs;
6714d8e8
KH		 144 char *type;
145 char attached;
146 u8 master;
147 unsigned int namelen;
148 const char *name;
149 struct kref *k;
95883719
KH		 150 unsigned long *maybe = mle->maybe_map,
151 *vote = mle->vote_map,
152 *resp = mle->response_map,
153 *node = mle->node_map;
6714d8e8
KH		 154
155 k = &mle->mle_refs;
156 if (mle->type == DLM_MLE_BLOCK)
157 type = "BLK";
158 else if (mle->type == DLM_MLE_MASTER)
159 type = "MAS";
160 else
161 type = "MIG";
162 refs = atomic_read(&k->refcount);
163 master = mle->master;
164 attached = (list_empty(&mle->hb_events) ? 'N' : 'Y');
165
166 if (mle->type != DLM_MLE_MASTER) {
167 namelen = mle->u.name.len;
168 name = mle->u.name.name;
169 } else {
170 namelen = mle->u.res->lockname.len;
171 name = mle->u.res->lockname.name;
172 }
173
95883719
KH		 174 mlog(ML_NOTICE, "%.*s: %3s refs=%3d mas=%3u new=%3u evt=%c inuse=%d ",
175 namelen, name, type, refs, master, mle->new_master, attached,
176 mle->inuse);
177 dlm_print_nodemap(maybe);
178 printk(", ");
179 dlm_print_nodemap(vote);
180 printk(", ");
181 dlm_print_nodemap(resp);
182 printk(", ");
183 dlm_print_nodemap(node);
184 printk(", ");
185 printk("\n");
6714d8e8
KH		 186}
187
8a9343fa
MF		 188#if 0
189/* Code here is included but defined out as it aids debugging */
190
6714d8e8
KH		 191static void dlm_dump_mles(struct dlm_ctxt *dlm)
192{
193 struct dlm_master_list_entry *mle;
194 struct list_head *iter;
195
196 mlog(ML_NOTICE, "dumping all mles for domain %s:\n", dlm->name);
6714d8e8
KH		 197 spin_lock(&dlm->master_lock);
198 list_for_each(iter, &dlm->master_list) {
199 mle = list_entry(iter, struct dlm_master_list_entry, list);
200 dlm_print_one_mle(mle);
201 }
202 spin_unlock(&dlm->master_lock);
203}
204
6714d8e8
KH		 205int dlm_dump_all_mles(const char __user *data, unsigned int len)
206{
207 struct list_head *iter;
208 struct dlm_ctxt *dlm;
209
210 spin_lock(&dlm_domain_lock);
211 list_for_each(iter, &dlm_domains) {
212 dlm = list_entry (iter, struct dlm_ctxt, list);
213 mlog(ML_NOTICE, "found dlm: %p, name=%s\n", dlm, dlm->name);
214 dlm_dump_mles(dlm);
215 }
216 spin_unlock(&dlm_domain_lock);
217 return len;
218}
219EXPORT_SYMBOL_GPL(dlm_dump_all_mles);
220
221#endif /* 0 */
222
223
e18b890b 224static struct kmem_cache *dlm_mle_cache = NULL;
6714d8e8
KH		 225
226
227static void dlm_mle_release(struct kref *kref);
228static void dlm_init_mle(struct dlm_master_list_entry *mle,
229 enum dlm_mle_type type,
230 struct dlm_ctxt *dlm,
231 struct dlm_lock_resource *res,
232 const char *name,
233 unsigned int namelen);
234static void dlm_put_mle(struct dlm_master_list_entry *mle);
235static void __dlm_put_mle(struct dlm_master_list_entry *mle);
236static int dlm_find_mle(struct dlm_ctxt *dlm,
237 struct dlm_master_list_entry **mle,
238 char *name, unsigned int namelen);
239
ba2bf218
KH		 240static int dlm_do_master_request(struct dlm_lock_resource *res,
241 struct dlm_master_list_entry *mle, int to);
6714d8e8
KH		 242
243
244static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
245 struct dlm_lock_resource *res,
246 struct dlm_master_list_entry *mle,
247 int *blocked);
248static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
249 struct dlm_lock_resource *res,
250 struct dlm_master_list_entry *mle,
251 int blocked);
252static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
253 struct dlm_lock_resource *res,
254 struct dlm_master_list_entry *mle,
255 struct dlm_master_list_entry **oldmle,
256 const char *name, unsigned int namelen,
257 u8 new_master, u8 master);
258
259static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
260 struct dlm_lock_resource *res);
261static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
262 struct dlm_lock_resource *res);
263static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
264 struct dlm_lock_resource *res,
265 u8 target);
c03872f5
KH		 266static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
267 struct dlm_lock_resource *res);
6714d8e8
KH		 268
269
270int dlm_is_host_down(int errno)
271{
272 switch (errno) {
273 case -EBADF:
274 case -ECONNREFUSED:
275 case -ENOTCONN:
276 case -ECONNRESET:
277 case -EPIPE:
278 case -EHOSTDOWN:
279 case -EHOSTUNREACH:
280 case -ETIMEDOUT:
281 case -ECONNABORTED:
282 case -ENETDOWN:
283 case -ENETUNREACH:
284 case -ENETRESET:
285 case -ESHUTDOWN:
286 case -ENOPROTOOPT:
287 case -EINVAL: /* if returned from our tcp code,
288 this means there is no socket */
289 return 1;
290 }
291 return 0;
292}
293
294
295/*
296 * MASTER LIST FUNCTIONS
297 */
298
299
300/*
301 * regarding master list entries and heartbeat callbacks:
302 *
303 * in order to avoid sleeping and allocation that occurs in
304 * heartbeat, master list entries are simply attached to the
305 * dlm's established heartbeat callbacks. the mle is attached
306 * when it is created, and since the dlm->spinlock is held at
307 * that time, any heartbeat event will be properly discovered
308 * by the mle. the mle needs to be detached from the
309 * dlm->mle_hb_events list as soon as heartbeat events are no
310 * longer useful to the mle, and before the mle is freed.
311 *
312 * as a general rule, heartbeat events are no longer needed by
313 * the mle once an "answer" regarding the lock master has been
314 * received.
315 */
316static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
317 struct dlm_master_list_entry *mle)
318{
319 assert_spin_locked(&dlm->spinlock);
320
321 list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
322}
323
324
325static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
326 struct dlm_master_list_entry *mle)
327{
328 if (!list_empty(&mle->hb_events))
329 list_del_init(&mle->hb_events);
330}
331
332
333static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
334 struct dlm_master_list_entry *mle)
335{
336 spin_lock(&dlm->spinlock);
337 __dlm_mle_detach_hb_events(dlm, mle);
338 spin_unlock(&dlm->spinlock);
339}
340
a2bf0477
KH		 341static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
342{
343 struct dlm_ctxt *dlm;
344 dlm = mle->dlm;
345
346 assert_spin_locked(&dlm->spinlock);
347 assert_spin_locked(&dlm->master_lock);
348 mle->inuse++;
349 kref_get(&mle->mle_refs);
350}
351
352static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
353{
354 struct dlm_ctxt *dlm;
355 dlm = mle->dlm;
356
357 spin_lock(&dlm->spinlock);
358 spin_lock(&dlm->master_lock);
359 mle->inuse--;
360 __dlm_put_mle(mle);
361 spin_unlock(&dlm->master_lock);
362 spin_unlock(&dlm->spinlock);
363
364}
365
6714d8e8
KH		 366/* remove from list and free */
367static void __dlm_put_mle(struct dlm_master_list_entry *mle)
368{
369 struct dlm_ctxt *dlm;
370 dlm = mle->dlm;
371
372 assert_spin_locked(&dlm->spinlock);
373 assert_spin_locked(&dlm->master_lock);
aa852354
KH		 374 if (!atomic_read(&mle->mle_refs.refcount)) {
375 /* this may or may not crash, but who cares.
376 * it's a BUG. */
377 mlog(ML_ERROR, "bad mle: %p\n", mle);
378 dlm_print_one_mle(mle);
379 BUG();
380 } else
381 kref_put(&mle->mle_refs, dlm_mle_release);
6714d8e8
KH		 382}
383
384
385/* must not have any spinlocks coming in */
386static void dlm_put_mle(struct dlm_master_list_entry *mle)
387{
388 struct dlm_ctxt *dlm;
389 dlm = mle->dlm;
390
391 spin_lock(&dlm->spinlock);
392 spin_lock(&dlm->master_lock);
393 __dlm_put_mle(mle);
394 spin_unlock(&dlm->master_lock);
395 spin_unlock(&dlm->spinlock);
396}
397
398static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
399{
400 kref_get(&mle->mle_refs);
401}
402
403static void dlm_init_mle(struct dlm_master_list_entry *mle,
404 enum dlm_mle_type type,
405 struct dlm_ctxt *dlm,
406 struct dlm_lock_resource *res,
407 const char *name,
408 unsigned int namelen)
409{
410 assert_spin_locked(&dlm->spinlock);
411
412 mle->dlm = dlm;
413 mle->type = type;
414 INIT_LIST_HEAD(&mle->list);
415 INIT_LIST_HEAD(&mle->hb_events);
416 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
417 spin_lock_init(&mle->spinlock);
418 init_waitqueue_head(&mle->wq);
419 atomic_set(&mle->woken, 0);
420 kref_init(&mle->mle_refs);
421 memset(mle->response_map, 0, sizeof(mle->response_map));
422 mle->master = O2NM_MAX_NODES;
423 mle->new_master = O2NM_MAX_NODES;
a2bf0477 424 mle->inuse = 0;
6714d8e8
KH		 425
426 if (mle->type == DLM_MLE_MASTER) {
427 BUG_ON(!res);
428 mle->u.res = res;
429 } else if (mle->type == DLM_MLE_BLOCK) {
430 BUG_ON(!name);
431 memcpy(mle->u.name.name, name, namelen);
432 mle->u.name.len = namelen;
433 } else /* DLM_MLE_MIGRATION */ {
434 BUG_ON(!name);
435 memcpy(mle->u.name.name, name, namelen);
436 mle->u.name.len = namelen;
437 }
438
439 /* copy off the node_map and register hb callbacks on our copy */
440 memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
441 memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
442 clear_bit(dlm->node_num, mle->vote_map);
443 clear_bit(dlm->node_num, mle->node_map);
444
445 /* attach the mle to the domain node up/down events */
446 __dlm_mle_attach_hb_events(dlm, mle);
447}
448
449
450/* returns 1 if found, 0 if not */
451static int dlm_find_mle(struct dlm_ctxt *dlm,
452 struct dlm_master_list_entry **mle,
453 char *name, unsigned int namelen)
454{
455 struct dlm_master_list_entry *tmpmle;
456 struct list_head *iter;
457
458 assert_spin_locked(&dlm->master_lock);
459
460 list_for_each(iter, &dlm->master_list) {
461 tmpmle = list_entry(iter, struct dlm_master_list_entry, list);
462 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
463 continue;
464 dlm_get_mle(tmpmle);
465 *mle = tmpmle;
466 return 1;
467 }
468 return 0;
469}
470
471void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
472{
473 struct dlm_master_list_entry *mle;
474 struct list_head *iter;
475
476 assert_spin_locked(&dlm->spinlock);
477
478 list_for_each(iter, &dlm->mle_hb_events) {
479 mle = list_entry(iter, struct dlm_master_list_entry,
480 hb_events);
481 if (node_up)
482 dlm_mle_node_up(dlm, mle, NULL, idx);
483 else
484 dlm_mle_node_down(dlm, mle, NULL, idx);
485 }
486}
487
488static void dlm_mle_node_down(struct dlm_ctxt *dlm,
489 struct dlm_master_list_entry *mle,
490 struct o2nm_node *node, int idx)
491{
492 spin_lock(&mle->spinlock);
493
494 if (!test_bit(idx, mle->node_map))
495 mlog(0, "node %u already removed from nodemap!\n", idx);
496 else
497 clear_bit(idx, mle->node_map);
498
499 spin_unlock(&mle->spinlock);
500}
501
502static void dlm_mle_node_up(struct dlm_ctxt *dlm,
503 struct dlm_master_list_entry *mle,
504 struct o2nm_node *node, int idx)
505{
506 spin_lock(&mle->spinlock);
507
508 if (test_bit(idx, mle->node_map))
509 mlog(0, "node %u already in node map!\n", idx);
510 else
511 set_bit(idx, mle->node_map);
512
513 spin_unlock(&mle->spinlock);
514}
515
516
517int dlm_init_mle_cache(void)
518{
519 dlm_mle_cache = kmem_cache_create("dlm_mle_cache",
520 sizeof(struct dlm_master_list_entry),
521 0, SLAB_HWCACHE_ALIGN,
522 NULL, NULL);
523 if (dlm_mle_cache == NULL)
524 return -ENOMEM;
525 return 0;
526}
527
528void dlm_destroy_mle_cache(void)
529{
530 if (dlm_mle_cache)
531 kmem_cache_destroy(dlm_mle_cache);
532}
533
534static void dlm_mle_release(struct kref *kref)
535{
536 struct dlm_master_list_entry *mle;
537 struct dlm_ctxt *dlm;
538
539 mlog_entry_void();
540
541 mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
542 dlm = mle->dlm;
543
544 if (mle->type != DLM_MLE_MASTER) {
545 mlog(0, "calling mle_release for %.*s, type %d\n",
546 mle->u.name.len, mle->u.name.name, mle->type);
547 } else {
548 mlog(0, "calling mle_release for %.*s, type %d\n",
549 mle->u.res->lockname.len,
550 mle->u.res->lockname.name, mle->type);
551 }
552 assert_spin_locked(&dlm->spinlock);
553 assert_spin_locked(&dlm->master_lock);
554
555 /* remove from list if not already */
556 if (!list_empty(&mle->list))
557 list_del_init(&mle->list);
558
559 /* detach the mle from the domain node up/down events */
560 __dlm_mle_detach_hb_events(dlm, mle);
561
562 /* NOTE: kfree under spinlock here.
563 * if this is bad, we can move this to a freelist. */
564 kmem_cache_free(dlm_mle_cache, mle);
565}
566
567
568/*
569 * LOCK RESOURCE FUNCTIONS
570 */
571
572static void dlm_set_lockres_owner(struct dlm_ctxt *dlm,
573 struct dlm_lock_resource *res,
574 u8 owner)
575{
576 assert_spin_locked(&res->spinlock);
577
578 mlog_entry("%.*s, %u\n", res->lockname.len, res->lockname.name, owner);
579
580 if (owner == dlm->node_num)
581 atomic_inc(&dlm->local_resources);
582 else if (owner == DLM_LOCK_RES_OWNER_UNKNOWN)
583 atomic_inc(&dlm->unknown_resources);
584 else
585 atomic_inc(&dlm->remote_resources);
586
587 res->owner = owner;
588}
589
590void dlm_change_lockres_owner(struct dlm_ctxt *dlm,
591 struct dlm_lock_resource *res, u8 owner)
592{
593 assert_spin_locked(&res->spinlock);
594
595 if (owner == res->owner)
596 return;
597
598 if (res->owner == dlm->node_num)
599 atomic_dec(&dlm->local_resources);
600 else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN)
601 atomic_dec(&dlm->unknown_resources);
602 else
603 atomic_dec(&dlm->remote_resources);
604
605 dlm_set_lockres_owner(dlm, res, owner);
606}
607
608
609static void dlm_lockres_release(struct kref *kref)
610{
611 struct dlm_lock_resource *res;
612
613 res = container_of(kref, struct dlm_lock_resource, refs);
614
615 /* This should not happen -- all lockres' have a name
616 * associated with them at init time. */
617 BUG_ON(!res->lockname.name);
618
619 mlog(0, "destroying lockres %.*s\n", res->lockname.len,
620 res->lockname.name);
621
a7f90d83
KH		 622 if (!hlist_unhashed(&res->hash_node) ||
623 !list_empty(&res->granted) ||
624 !list_empty(&res->converting) ||
625 !list_empty(&res->blocked) ||
626 !list_empty(&res->dirty) ||
627 !list_empty(&res->recovering) ||
628 !list_empty(&res->purge)) {
629 mlog(ML_ERROR,
630 "Going to BUG for resource %.*s."
631 " We're on a list! [%c%c%c%c%c%c%c]\n",
632 res->lockname.len, res->lockname.name,
633 !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
634 !list_empty(&res->granted) ? 'G' : ' ',
635 !list_empty(&res->converting) ? 'C' : ' ',
636 !list_empty(&res->blocked) ? 'B' : ' ',
637 !list_empty(&res->dirty) ? 'D' : ' ',
638 !list_empty(&res->recovering) ? 'R' : ' ',
639 !list_empty(&res->purge) ? 'P' : ' ');
640
641 dlm_print_one_lock_resource(res);
642 }
643
6714d8e8
KH		 644 /* By the time we're ready to blow this guy away, we shouldn't
645 * be on any lists. */
81f2094a 646 BUG_ON(!hlist_unhashed(&res->hash_node));
6714d8e8
KH		 647 BUG_ON(!list_empty(&res->granted));
648 BUG_ON(!list_empty(&res->converting));
649 BUG_ON(!list_empty(&res->blocked));
650 BUG_ON(!list_empty(&res->dirty));
651 BUG_ON(!list_empty(&res->recovering));
652 BUG_ON(!list_empty(&res->purge));
653
654 kfree(res->lockname.name);
655
656 kfree(res);
657}
658
6714d8e8
KH		 659void dlm_lockres_put(struct dlm_lock_resource *res)
660{
661 kref_put(&res->refs, dlm_lockres_release);
662}
663
664static void dlm_init_lockres(struct dlm_ctxt *dlm,
665 struct dlm_lock_resource *res,
666 const char *name, unsigned int namelen)
667{
668 char *qname;
669
670 /* If we memset here, we lose our reference to the kmalloc'd
671 * res->lockname.name, so be sure to init every field
672 * correctly! */
673
674 qname = (char *) res->lockname.name;
675 memcpy(qname, name, namelen);
676
677 res->lockname.len = namelen;
a3d33291 678 res->lockname.hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 679
680 init_waitqueue_head(&res->wq);
681 spin_lock_init(&res->spinlock);
81f2094a 682 INIT_HLIST_NODE(&res->hash_node);
6714d8e8
KH		 683 INIT_LIST_HEAD(&res->granted);
684 INIT_LIST_HEAD(&res->converting);
685 INIT_LIST_HEAD(&res->blocked);
686 INIT_LIST_HEAD(&res->dirty);
687 INIT_LIST_HEAD(&res->recovering);
688 INIT_LIST_HEAD(&res->purge);
689 atomic_set(&res->asts_reserved, 0);
690 res->migration_pending = 0;
ba2bf218 691 res->inflight_locks = 0;
6714d8e8
KH		 692
693 kref_init(&res->refs);
694
695 /* just for consistency */
696 spin_lock(&res->spinlock);
697 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
698 spin_unlock(&res->spinlock);
699
700 res->state = DLM_LOCK_RES_IN_PROGRESS;
701
702 res->last_used = 0;
703
704 memset(res->lvb, 0, DLM_LVB_LEN);
ba2bf218 705 memset(res->refmap, 0, sizeof(res->refmap));
6714d8e8
KH		 706}
707
708struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
709 const char *name,
710 unsigned int namelen)
711{
712 struct dlm_lock_resource *res;
713
ad8100e0 714 res = kmalloc(sizeof(struct dlm_lock_resource), GFP_NOFS);
6714d8e8
KH		 715 if (!res)
716 return NULL;
717
ad8100e0 718 res->lockname.name = kmalloc(namelen, GFP_NOFS);
6714d8e8
KH		 719 if (!res->lockname.name) {
720 kfree(res);
721 return NULL;
722 }
723
724 dlm_init_lockres(dlm, res, name, namelen);
725 return res;
726}
727
ba2bf218
KH		 728void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
729 struct dlm_lock_resource *res,
730 int new_lockres,
731 const char *file,
732 int line)
733{
734 if (!new_lockres)
735 assert_spin_locked(&res->spinlock);
736
737 if (!test_bit(dlm->node_num, res->refmap)) {
738 BUG_ON(res->inflight_locks != 0);
739 dlm_lockres_set_refmap_bit(dlm->node_num, res);
740 }
741 res->inflight_locks++;
742 mlog(0, "%s:%.*s: inflight++: now %u\n",
743 dlm->name, res->lockname.len, res->lockname.name,
744 res->inflight_locks);
745}
746
747void __dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
748 struct dlm_lock_resource *res,
749 const char *file,
750 int line)
751{
752 assert_spin_locked(&res->spinlock);
753
754 BUG_ON(res->inflight_locks == 0);
755 res->inflight_locks--;
756 mlog(0, "%s:%.*s: inflight--: now %u\n",
757 dlm->name, res->lockname.len, res->lockname.name,
758 res->inflight_locks);
759 if (res->inflight_locks == 0)
760 dlm_lockres_clear_refmap_bit(dlm->node_num, res);
761 wake_up(&res->wq);
762}
763
6714d8e8
KH		 764/*
765 * lookup a lock resource by name.
766 * may already exist in the hashtable.
767 * lockid is null terminated
768 *
769 * if not, allocate enough for the lockres and for
770 * the temporary structure used in doing the mastering.
771 *
772 * also, do a lookup in the dlm->master_list to see
773 * if another node has begun mastering the same lock.
774 * if so, there should be a block entry in there
775 * for this name, and we should *not* attempt to master
776 * the lock here. need to wait around for that node
777 * to assert_master (or die).
778 *
779 */
780struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
781 const char *lockid,
3384f3df 782 int namelen,
6714d8e8
KH		 783 int flags)
784{
785 struct dlm_lock_resource *tmpres=NULL, *res=NULL;
786 struct dlm_master_list_entry *mle = NULL;
787 struct dlm_master_list_entry *alloc_mle = NULL;
788 int blocked = 0;
789 int ret, nodenum;
790 struct dlm_node_iter iter;
3384f3df 791 unsigned int hash;
6714d8e8 792 int tries = 0;
c03872f5 793 int bit, wait_on_recovery = 0;
ba2bf218 794 int drop_inflight_if_nonlocal = 0;
6714d8e8
KH		 795
796 BUG_ON(!lockid);
797
a3d33291 798 hash = dlm_lockid_hash(lockid, namelen);
6714d8e8
KH		 799
800 mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
801
802lookup:
803 spin_lock(&dlm->spinlock);
ba2bf218 804 tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
6714d8e8 805 if (tmpres) {
ba2bf218
KH		 806 int dropping_ref = 0;
807
808 spin_lock(&tmpres->spinlock);
809 if (tmpres->owner == dlm->node_num) {
810 BUG_ON(tmpres->state & DLM_LOCK_RES_DROPPING_REF);
811 dlm_lockres_grab_inflight_ref(dlm, tmpres);
812 } else if (tmpres->state & DLM_LOCK_RES_DROPPING_REF)
813 dropping_ref = 1;
814 spin_unlock(&tmpres->spinlock);
6714d8e8 815 spin_unlock(&dlm->spinlock);
ba2bf218
KH		 816
817 /* wait until done messaging the master, drop our ref to allow
818 * the lockres to be purged, start over. */
819 if (dropping_ref) {
820 spin_lock(&tmpres->spinlock);
821 __dlm_wait_on_lockres_flags(tmpres, DLM_LOCK_RES_DROPPING_REF);
822 spin_unlock(&tmpres->spinlock);
823 dlm_lockres_put(tmpres);
824 tmpres = NULL;
825 goto lookup;
826 }
827
6714d8e8
KH		 828 mlog(0, "found in hash!\n");
829 if (res)
830 dlm_lockres_put(res);
831 res = tmpres;
832 goto leave;
833 }
834
835 if (!res) {
836 spin_unlock(&dlm->spinlock);
837 mlog(0, "allocating a new resource\n");
838 /* nothing found and we need to allocate one. */
839 alloc_mle = (struct dlm_master_list_entry *)
ad8100e0 840 kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
6714d8e8
KH		 841 if (!alloc_mle)
842 goto leave;
843 res = dlm_new_lockres(dlm, lockid, namelen);
844 if (!res)
845 goto leave;
846 goto lookup;
847 }
848
849 mlog(0, "no lockres found, allocated our own: %p\n", res);
850
851 if (flags & LKM_LOCAL) {
852 /* caller knows it's safe to assume it's not mastered elsewhere
853 * DONE! return right away */
854 spin_lock(&res->spinlock);
855 dlm_change_lockres_owner(dlm, res, dlm->node_num);
856 __dlm_insert_lockres(dlm, res);
ba2bf218 857 dlm_lockres_grab_inflight_ref(dlm, res);
6714d8e8
KH		 858 spin_unlock(&res->spinlock);
859 spin_unlock(&dlm->spinlock);
860 /* lockres still marked IN_PROGRESS */
861 goto wake_waiters;
862 }
863
864 /* check master list to see if another node has started mastering it */
865 spin_lock(&dlm->master_lock);
866
867 /* if we found a block, wait for lock to be mastered by another node */
868 blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
869 if (blocked) {
ba2bf218 870 int mig;
6714d8e8
KH		 871 if (mle->type == DLM_MLE_MASTER) {
872 mlog(ML_ERROR, "master entry for nonexistent lock!\n");
873 BUG();
ba2bf218
KH		 874 }
875 mig = (mle->type == DLM_MLE_MIGRATION);
876 /* if there is a migration in progress, let the migration
877 * finish before continuing. we can wait for the absence
878 * of the MIGRATION mle: either the migrate finished or
879 * one of the nodes died and the mle was cleaned up.
880 * if there is a BLOCK here, but it already has a master
881 * set, we are too late. the master does not have a ref
882 * for us in the refmap. detach the mle and drop it.
883 * either way, go back to the top and start over. */
884 if (mig || mle->master != O2NM_MAX_NODES) {
885 BUG_ON(mig && mle->master == dlm->node_num);
886 /* we arrived too late. the master does not
887 * have a ref for us. retry. */
888 mlog(0, "%s:%.*s: late on %s\n",
889 dlm->name, namelen, lockid,
890 mig ? "MIGRATION" : "BLOCK");
6714d8e8 891 spin_unlock(&dlm->master_lock);
6714d8e8
KH		 892 spin_unlock(&dlm->spinlock);
893
894 /* master is known, detach */
ba2bf218
KH		 895 if (!mig)
896 dlm_mle_detach_hb_events(dlm, mle);
6714d8e8
KH		 897 dlm_put_mle(mle);
898 mle = NULL;
ba2bf218
KH		 899 /* this is lame, but we cant wait on either
900 * the mle or lockres waitqueue here */
901 if (mig)
902 msleep(100);
903 goto lookup;
6714d8e8
KH		 904 }
905 } else {
906 /* go ahead and try to master lock on this node */
907 mle = alloc_mle;
908 /* make sure this does not get freed below */
909 alloc_mle = NULL;
910 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
911 set_bit(dlm->node_num, mle->maybe_map);
912 list_add(&mle->list, &dlm->master_list);
c03872f5
KH		 913
914 /* still holding the dlm spinlock, check the recovery map
915 * to see if there are any nodes that still need to be
916 * considered. these will not appear in the mle nodemap
917 * but they might own this lockres. wait on them. */
918 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
919 if (bit < O2NM_MAX_NODES) {
920 mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to"
921 "recover before lock mastery can begin\n",
922 dlm->name, namelen, (char *)lockid, bit);
923 wait_on_recovery = 1;
924 }
6714d8e8
KH		 925 }
926
927 /* at this point there is either a DLM_MLE_BLOCK or a
928 * DLM_MLE_MASTER on the master list, so it's safe to add the
929 * lockres to the hashtable. anyone who finds the lock will
930 * still have to wait on the IN_PROGRESS. */
931
932 /* finally add the lockres to its hash bucket */
933 __dlm_insert_lockres(dlm, res);
ba2bf218
KH		 934 /* since this lockres is new it doesnt not require the spinlock */
935 dlm_lockres_grab_inflight_ref_new(dlm, res);
936
937 /* if this node does not become the master make sure to drop
938 * this inflight reference below */
939 drop_inflight_if_nonlocal = 1;
940
6714d8e8
KH		 941 /* get an extra ref on the mle in case this is a BLOCK
942 * if so, the creator of the BLOCK may try to put the last
943 * ref at this time in the assert master handler, so we
944 * need an extra one to keep from a bad ptr deref. */
a2bf0477 945 dlm_get_mle_inuse(mle);
6714d8e8
KH		 946 spin_unlock(&dlm->master_lock);
947 spin_unlock(&dlm->spinlock);
948
e7e69eb3 949redo_request:
c03872f5
KH		 950 while (wait_on_recovery) {
951 /* any cluster changes that occurred after dropping the
952 * dlm spinlock would be detectable be a change on the mle,
953 * so we only need to clear out the recovery map once. */
954 if (dlm_is_recovery_lock(lockid, namelen)) {
955 mlog(ML_NOTICE, "%s: recovery map is not empty, but "
956 "must master $RECOVERY lock now\n", dlm->name);
957 if (!dlm_pre_master_reco_lockres(dlm, res))
958 wait_on_recovery = 0;
959 else {
960 mlog(0, "%s: waiting 500ms for heartbeat state "
961 "change\n", dlm->name);
962 msleep(500);
963 }
964 continue;
965 }
966
967 dlm_kick_recovery_thread(dlm);
aa087b84 968 msleep(1000);
c03872f5
KH		 969 dlm_wait_for_recovery(dlm);
970
971 spin_lock(&dlm->spinlock);
972 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
973 if (bit < O2NM_MAX_NODES) {
974 mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to"
975 "recover before lock mastery can begin\n",
976 dlm->name, namelen, (char *)lockid, bit);
977 wait_on_recovery = 1;
978 } else
979 wait_on_recovery = 0;
980 spin_unlock(&dlm->spinlock);
b7084ab5
KH		 981
982 if (wait_on_recovery)
983 dlm_wait_for_node_recovery(dlm, bit, 10000);
c03872f5
KH		 984 }
985
6714d8e8
KH		 986 /* must wait for lock to be mastered elsewhere */
987 if (blocked)
988 goto wait;
989
6714d8e8
KH		 990 ret = -EINVAL;
991 dlm_node_iter_init(mle->vote_map, &iter);
992 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
ba2bf218 993 ret = dlm_do_master_request(res, mle, nodenum);
6714d8e8
KH		 994 if (ret < 0)
995 mlog_errno(ret);
996 if (mle->master != O2NM_MAX_NODES) {
997 /* found a master ! */
9c6510a5
KH		 998 if (mle->master <= nodenum)
999 break;
1000 /* if our master request has not reached the master
1001 * yet, keep going until it does. this is how the
1002 * master will know that asserts are needed back to
1003 * the lower nodes. */
1004 mlog(0, "%s:%.*s: requests only up to %u but master "
1005 "is %u, keep going\n", dlm->name, namelen,
1006 lockid, nodenum, mle->master);
6714d8e8
KH		 1007 }
1008 }
1009
1010wait:
1011 /* keep going until the response map includes all nodes */
1012 ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
1013 if (ret < 0) {
e7e69eb3 1014 wait_on_recovery = 1;
6714d8e8
KH		 1015 mlog(0, "%s:%.*s: node map changed, redo the "
1016 "master request now, blocked=%d\n",
1017 dlm->name, res->lockname.len,
1018 res->lockname.name, blocked);
1019 if (++tries > 20) {
1020 mlog(ML_ERROR, "%s:%.*s: spinning on "
1021 "dlm_wait_for_lock_mastery, blocked=%d\n",
1022 dlm->name, res->lockname.len,
1023 res->lockname.name, blocked);
1024 dlm_print_one_lock_resource(res);
8a9343fa 1025 dlm_print_one_mle(mle);
6714d8e8
KH		 1026 tries = 0;
1027 }
1028 goto redo_request;
1029 }
1030
1031 mlog(0, "lockres mastered by %u\n", res->owner);
1032 /* make sure we never continue without this */
1033 BUG_ON(res->owner == O2NM_MAX_NODES);
1034
1035 /* master is known, detach if not already detached */
1036 dlm_mle_detach_hb_events(dlm, mle);
1037 dlm_put_mle(mle);
1038 /* put the extra ref */
a2bf0477 1039 dlm_put_mle_inuse(mle);
6714d8e8
KH		 1040
1041wake_waiters:
1042 spin_lock(&res->spinlock);
ba2bf218
KH		 1043 if (res->owner != dlm->node_num && drop_inflight_if_nonlocal)
1044 dlm_lockres_drop_inflight_ref(dlm, res);
6714d8e8
KH		 1045 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1046 spin_unlock(&res->spinlock);
1047 wake_up(&res->wq);
1048
1049leave:
1050 /* need to free the unused mle */
1051 if (alloc_mle)
1052 kmem_cache_free(dlm_mle_cache, alloc_mle);
1053
1054 return res;
1055}
1056
1057
1058#define DLM_MASTERY_TIMEOUT_MS 5000
1059
1060static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
1061 struct dlm_lock_resource *res,
1062 struct dlm_master_list_entry *mle,
1063 int *blocked)
1064{
1065 u8 m;
1066 int ret, bit;
1067 int map_changed, voting_done;
1068 int assert, sleep;
1069
1070recheck:
1071 ret = 0;
1072 assert = 0;
1073
1074 /* check if another node has already become the owner */
1075 spin_lock(&res->spinlock);
1076 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
9c6510a5
KH		 1077 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1078 res->lockname.len, res->lockname.name, res->owner);
6714d8e8 1079 spin_unlock(&res->spinlock);
9c6510a5
KH		 1080 /* this will cause the master to re-assert across
1081 * the whole cluster, freeing up mles */
588e0090 1082 if (res->owner != dlm->node_num) {
ba2bf218 1083 ret = dlm_do_master_request(res, mle, res->owner);
588e0090
KH		 1084 if (ret < 0) {
1085 /* give recovery a chance to run */
1086 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1087 msleep(500);
1088 goto recheck;
1089 }
9c6510a5
KH		 1090 }
1091 ret = 0;
6714d8e8
KH		 1092 goto leave;
1093 }
1094 spin_unlock(&res->spinlock);
1095
1096 spin_lock(&mle->spinlock);
1097 m = mle->master;
1098 map_changed = (memcmp(mle->vote_map, mle->node_map,
1099 sizeof(mle->vote_map)) != 0);
1100 voting_done = (memcmp(mle->vote_map, mle->response_map,
1101 sizeof(mle->vote_map)) == 0);
1102
1103 /* restart if we hit any errors */
1104 if (map_changed) {
1105 int b;
1106 mlog(0, "%s: %.*s: node map changed, restarting\n",
1107 dlm->name, res->lockname.len, res->lockname.name);
1108 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1109 b = (mle->type == DLM_MLE_BLOCK);
1110 if ((*blocked && !b) || (!*blocked && b)) {
1111 mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1112 dlm->name, res->lockname.len, res->lockname.name,
1113 *blocked, b);
1114 *blocked = b;
1115 }
1116 spin_unlock(&mle->spinlock);
1117 if (ret < 0) {
1118 mlog_errno(ret);
1119 goto leave;
1120 }
1121 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1122 "rechecking now\n", dlm->name, res->lockname.len,
1123 res->lockname.name);
1124 goto recheck;
aa852354
KH		 1125 } else {
1126 if (!voting_done) {
1127 mlog(0, "map not changed and voting not done "
1128 "for %s:%.*s\n", dlm->name, res->lockname.len,
1129 res->lockname.name);
1130 }
6714d8e8
KH		 1131 }
1132
1133 if (m != O2NM_MAX_NODES) {
1134 /* another node has done an assert!
1135 * all done! */
1136 sleep = 0;
1137 } else {
1138 sleep = 1;
1139 /* have all nodes responded? */
1140 if (voting_done && !*blocked) {
1141 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1142 if (dlm->node_num <= bit) {
1143 /* my node number is lowest.
1144 * now tell other nodes that I am
1145 * mastering this. */
1146 mle->master = dlm->node_num;
ba2bf218
KH		 1147 /* ref was grabbed in get_lock_resource
1148 * will be dropped in dlmlock_master */
6714d8e8
KH		 1149 assert = 1;
1150 sleep = 0;
1151 }
1152 /* if voting is done, but we have not received
1153 * an assert master yet, we must sleep */
1154 }
1155 }
1156
1157 spin_unlock(&mle->spinlock);
1158
1159 /* sleep if we haven't finished voting yet */
1160 if (sleep) {
1161 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1162
1163 /*
1164 if (atomic_read(&mle->mle_refs.refcount) < 2)
1165 mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1166 atomic_read(&mle->mle_refs.refcount),
1167 res->lockname.len, res->lockname.name);
1168 */
1169 atomic_set(&mle->woken, 0);
1170 (void)wait_event_timeout(mle->wq,
1171 (atomic_read(&mle->woken) == 1),
1172 timeo);
1173 if (res->owner == O2NM_MAX_NODES) {
ba2bf218
KH		 1174 mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1175 res->lockname.len, res->lockname.name);
6714d8e8
KH		 1176 goto recheck;
1177 }
1178 mlog(0, "done waiting, master is %u\n", res->owner);
1179 ret = 0;
1180 goto leave;
1181 }
1182
1183 ret = 0; /* done */
1184 if (assert) {
1185 m = dlm->node_num;
1186 mlog(0, "about to master %.*s here, this=%u\n",
1187 res->lockname.len, res->lockname.name, m);
ba2bf218 1188 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
6714d8e8
KH		 1189 if (ret) {
1190 /* This is a failure in the network path,
1191 * not in the response to the assert_master
1192 * (any nonzero response is a BUG on this node).
1193 * Most likely a socket just got disconnected
1194 * due to node death. */
1195 mlog_errno(ret);
1196 }
1197 /* no longer need to restart lock mastery.
1198 * all living nodes have been contacted. */
1199 ret = 0;
1200 }
1201
1202 /* set the lockres owner */
1203 spin_lock(&res->spinlock);
ba2bf218
KH		 1204 /* mastery reference obtained either during
1205 * assert_master_handler or in get_lock_resource */
6714d8e8
KH		 1206 dlm_change_lockres_owner(dlm, res, m);
1207 spin_unlock(&res->spinlock);
1208
1209leave:
1210 return ret;
1211}
1212
1213struct dlm_bitmap_diff_iter
1214{
1215 int curnode;
1216 unsigned long *orig_bm;
1217 unsigned long *cur_bm;
1218 unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1219};
1220
1221enum dlm_node_state_change
1222{
1223 NODE_DOWN = -1,
1224 NODE_NO_CHANGE = 0,
1225 NODE_UP
1226};
1227
1228static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1229 unsigned long *orig_bm,
1230 unsigned long *cur_bm)
1231{
1232 unsigned long p1, p2;
1233 int i;
1234
1235 iter->curnode = -1;
1236 iter->orig_bm = orig_bm;
1237 iter->cur_bm = cur_bm;
1238
1239 for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1240 p1 = *(iter->orig_bm + i);
1241 p2 = *(iter->cur_bm + i);
1242 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1243 }
1244}
1245
1246static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1247 enum dlm_node_state_change *state)
1248{
1249 int bit;
1250
1251 if (iter->curnode >= O2NM_MAX_NODES)
1252 return -ENOENT;
1253
1254 bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1255 iter->curnode+1);
1256 if (bit >= O2NM_MAX_NODES) {
1257 iter->curnode = O2NM_MAX_NODES;
1258 return -ENOENT;
1259 }
1260
1261 /* if it was there in the original then this node died */
1262 if (test_bit(bit, iter->orig_bm))
1263 *state = NODE_DOWN;
1264 else
1265 *state = NODE_UP;
1266
1267 iter->curnode = bit;
1268 return bit;
1269}
1270
1271
1272static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1273 struct dlm_lock_resource *res,
1274 struct dlm_master_list_entry *mle,
1275 int blocked)
1276{
1277 struct dlm_bitmap_diff_iter bdi;
1278 enum dlm_node_state_change sc;
1279 int node;
1280 int ret = 0;
1281
1282 mlog(0, "something happened such that the "
1283 "master process may need to be restarted!\n");
1284
1285 assert_spin_locked(&mle->spinlock);
1286
1287 dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1288 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1289 while (node >= 0) {
1290 if (sc == NODE_UP) {
e2faea4c
KH		 1291 /* a node came up. clear any old vote from
1292 * the response map and set it in the vote map
1293 * then restart the mastery. */
1294 mlog(ML_NOTICE, "node %d up while restarting\n", node);
6714d8e8
KH		 1295
1296 /* redo the master request, but only for the new node */
1297 mlog(0, "sending request to new node\n");
1298 clear_bit(node, mle->response_map);
1299 set_bit(node, mle->vote_map);
1300 } else {
1301 mlog(ML_ERROR, "node down! %d\n", node);
6714d8e8
KH		 1302 if (blocked) {
1303 int lowest = find_next_bit(mle->maybe_map,
1304 O2NM_MAX_NODES, 0);
1305
1306 /* act like it was never there */
1307 clear_bit(node, mle->maybe_map);
1308
e7e69eb3
KH		 1309 if (node == lowest) {
1310 mlog(0, "expected master %u died"
1311 " while this node was blocked "
1312 "waiting on it!\n", node);
1313 lowest = find_next_bit(mle->maybe_map,
1314 O2NM_MAX_NODES,
1315 lowest+1);
1316 if (lowest < O2NM_MAX_NODES) {
1317 mlog(0, "%s:%.*s:still "
1318 "blocked. waiting on %u "
1319 "now\n", dlm->name,
1320 res->lockname.len,
1321 res->lockname.name,
1322 lowest);
1323 } else {
1324 /* mle is an MLE_BLOCK, but
1325 * there is now nothing left to
1326 * block on. we need to return
1327 * all the way back out and try
1328 * again with an MLE_MASTER.
1329 * dlm_do_local_recovery_cleanup
1330 * has already run, so the mle
1331 * refcount is ok */
1332 mlog(0, "%s:%.*s: no "
1333 "longer blocking. try to "
1334 "master this here\n",
1335 dlm->name,
1336 res->lockname.len,
1337 res->lockname.name);
1338 mle->type = DLM_MLE_MASTER;
1339 mle->u.res = res;
1340 }
6714d8e8 1341 }
6714d8e8
KH		 1342 }
1343
e7e69eb3
KH		 1344 /* now blank out everything, as if we had never
1345 * contacted anyone */
1346 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1347 memset(mle->response_map, 0, sizeof(mle->response_map));
1348 /* reset the vote_map to the current node_map */
1349 memcpy(mle->vote_map, mle->node_map,
1350 sizeof(mle->node_map));
1351 /* put myself into the maybe map */
1352 if (mle->type != DLM_MLE_BLOCK)
1353 set_bit(dlm->node_num, mle->maybe_map);
6714d8e8
KH		 1354 }
1355 ret = -EAGAIN;
6714d8e8
KH		 1356 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1357 }
1358 return ret;
1359}
1360
1361
1362/*
1363 * DLM_MASTER_REQUEST_MSG
1364 *
1365 * returns: 0 on success,
1366 * -errno on a network error
1367 *
1368 * on error, the caller should assume the target node is "dead"
1369 *
1370 */
1371
ba2bf218
KH		 1372static int dlm_do_master_request(struct dlm_lock_resource *res,
1373 struct dlm_master_list_entry *mle, int to)
6714d8e8
KH		 1374{
1375 struct dlm_ctxt *dlm = mle->dlm;
1376 struct dlm_master_request request;
1377 int ret, response=0, resend;
1378
1379 memset(&request, 0, sizeof(request));
1380 request.node_idx = dlm->node_num;
1381
1382 BUG_ON(mle->type == DLM_MLE_MIGRATION);
1383
1384 if (mle->type != DLM_MLE_MASTER) {
1385 request.namelen = mle->u.name.len;
1386 memcpy(request.name, mle->u.name.name, request.namelen);
1387 } else {
1388 request.namelen = mle->u.res->lockname.len;
1389 memcpy(request.name, mle->u.res->lockname.name,
1390 request.namelen);
1391 }
1392
1393again:
1394 ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1395 sizeof(request), to, &response);
1396 if (ret < 0) {
1397 if (ret == -ESRCH) {
1398 /* should never happen */
1399 mlog(ML_ERROR, "TCP stack not ready!\n");
1400 BUG();
1401 } else if (ret == -EINVAL) {
1402 mlog(ML_ERROR, "bad args passed to o2net!\n");
1403 BUG();
1404 } else if (ret == -ENOMEM) {
1405 mlog(ML_ERROR, "out of memory while trying to send "
1406 "network message! retrying\n");
1407 /* this is totally crude */
1408 msleep(50);
1409 goto again;
1410 } else if (!dlm_is_host_down(ret)) {
1411 /* not a network error. bad. */
1412 mlog_errno(ret);
1413 mlog(ML_ERROR, "unhandled error!");
1414 BUG();
1415 }
1416 /* all other errors should be network errors,
1417 * and likely indicate node death */
1418 mlog(ML_ERROR, "link to %d went down!\n", to);
1419 goto out;
1420 }
1421
1422 ret = 0;
1423 resend = 0;
1424 spin_lock(&mle->spinlock);
1425 switch (response) {
1426 case DLM_MASTER_RESP_YES:
1427 set_bit(to, mle->response_map);
1428 mlog(0, "node %u is the master, response=YES\n", to);
ba2bf218
KH		 1429 mlog(0, "%s:%.*s: master node %u now knows I have a "
1430 "reference\n", dlm->name, res->lockname.len,
1431 res->lockname.name, to);
6714d8e8
KH		 1432 mle->master = to;
1433 break;
1434 case DLM_MASTER_RESP_NO:
1435 mlog(0, "node %u not master, response=NO\n", to);
1436 set_bit(to, mle->response_map);
1437 break;
1438 case DLM_MASTER_RESP_MAYBE:
1439 mlog(0, "node %u not master, response=MAYBE\n", to);
1440 set_bit(to, mle->response_map);
1441 set_bit(to, mle->maybe_map);
1442 break;
1443 case DLM_MASTER_RESP_ERROR:
1444 mlog(0, "node %u hit an error, resending\n", to);
1445 resend = 1;
1446 response = 0;
1447 break;
1448 default:
1449 mlog(ML_ERROR, "bad response! %u\n", response);
1450 BUG();
1451 }
1452 spin_unlock(&mle->spinlock);
1453 if (resend) {
1454 /* this is also totally crude */
1455 msleep(50);
1456 goto again;
1457 }
1458
1459out:
1460 return ret;
1461}
1462
1463/*
1464 * locks that can be taken here:
1465 * dlm->spinlock
1466 * res->spinlock
1467 * mle->spinlock
1468 * dlm->master_list
1469 *
1470 * if possible, TRIM THIS DOWN!!!
1471 */
1472int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data)
1473{
1474 u8 response = DLM_MASTER_RESP_MAYBE;
1475 struct dlm_ctxt *dlm = data;
9c6510a5 1476 struct dlm_lock_resource *res = NULL;
6714d8e8
KH		 1477 struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1478 struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1479 char *name;
a3d33291 1480 unsigned int namelen, hash;
6714d8e8
KH		 1481 int found, ret;
1482 int set_maybe;
9c6510a5 1483 int dispatch_assert = 0;
6714d8e8
KH		 1484
1485 if (!dlm_grab(dlm))
1486 return DLM_MASTER_RESP_NO;
1487
1488 if (!dlm_domain_fully_joined(dlm)) {
1489 response = DLM_MASTER_RESP_NO;
1490 goto send_response;
1491 }
1492
1493 name = request->name;
1494 namelen = request->namelen;
a3d33291 1495 hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 1496
1497 if (namelen > DLM_LOCKID_NAME_MAX) {
1498 response = DLM_IVBUFLEN;
1499 goto send_response;
1500 }
1501
1502way_up_top:
1503 spin_lock(&dlm->spinlock);
a3d33291 1504 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
6714d8e8
KH		 1505 if (res) {
1506 spin_unlock(&dlm->spinlock);
1507
1508 /* take care of the easy cases up front */
1509 spin_lock(&res->spinlock);
1cd04dbe
KH		 1510 if (res->state & (DLM_LOCK_RES_RECOVERING|
1511 DLM_LOCK_RES_MIGRATING)) {
6714d8e8
KH		 1512 spin_unlock(&res->spinlock);
1513 mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1cd04dbe 1514 "being recovered/migrated\n");
6714d8e8
KH		 1515 response = DLM_MASTER_RESP_ERROR;
1516 if (mle)
1517 kmem_cache_free(dlm_mle_cache, mle);
1518 goto send_response;
1519 }
1520
1521 if (res->owner == dlm->node_num) {
ba2bf218
KH		 1522 mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1523 dlm->name, namelen, name, request->node_idx);
1524 dlm_lockres_set_refmap_bit(request->node_idx, res);
6714d8e8 1525 spin_unlock(&res->spinlock);
6714d8e8
KH		 1526 response = DLM_MASTER_RESP_YES;
1527 if (mle)
1528 kmem_cache_free(dlm_mle_cache, mle);
1529
1530 /* this node is the owner.
1531 * there is some extra work that needs to
1532 * happen now. the requesting node has
1533 * caused all nodes up to this one to
1534 * create mles. this node now needs to
1535 * go back and clean those up. */
9c6510a5 1536 dispatch_assert = 1;
6714d8e8
KH		 1537 goto send_response;
1538 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1539 spin_unlock(&res->spinlock);
1540 // mlog(0, "node %u is the master\n", res->owner);
1541 response = DLM_MASTER_RESP_NO;
1542 if (mle)
1543 kmem_cache_free(dlm_mle_cache, mle);
1544 goto send_response;
1545 }
1546
1547 /* ok, there is no owner. either this node is
1548 * being blocked, or it is actively trying to
1549 * master this lock. */
1550 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1551 mlog(ML_ERROR, "lock with no owner should be "
1552 "in-progress!\n");
1553 BUG();
1554 }
1555
1556 // mlog(0, "lockres is in progress...\n");
1557 spin_lock(&dlm->master_lock);
1558 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1559 if (!found) {
1560 mlog(ML_ERROR, "no mle found for this lock!\n");
1561 BUG();
1562 }
1563 set_maybe = 1;
1564 spin_lock(&tmpmle->spinlock);
1565 if (tmpmle->type == DLM_MLE_BLOCK) {
1566 // mlog(0, "this node is waiting for "
1567 // "lockres to be mastered\n");
1568 response = DLM_MASTER_RESP_NO;
1569 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1570 mlog(0, "node %u is master, but trying to migrate to "
1571 "node %u.\n", tmpmle->master, tmpmle->new_master);
1572 if (tmpmle->master == dlm->node_num) {
6714d8e8
KH		 1573 mlog(ML_ERROR, "no owner on lockres, but this "
1574 "node is trying to migrate it to %u?!\n",
1575 tmpmle->new_master);
1576 BUG();
1577 } else {
1578 /* the real master can respond on its own */
1579 response = DLM_MASTER_RESP_NO;
1580 }
1581 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1582 set_maybe = 0;
9c6510a5 1583 if (tmpmle->master == dlm->node_num) {
6714d8e8 1584 response = DLM_MASTER_RESP_YES;
9c6510a5
KH		 1585 /* this node will be the owner.
1586 * go back and clean the mles on any
1587 * other nodes */
1588 dispatch_assert = 1;
ba2bf218
KH		 1589 dlm_lockres_set_refmap_bit(request->node_idx, res);
1590 mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1591 dlm->name, namelen, name,
1592 request->node_idx);
9c6510a5 1593 } else
6714d8e8
KH		 1594 response = DLM_MASTER_RESP_NO;
1595 } else {
1596 // mlog(0, "this node is attempting to "
1597 // "master lockres\n");
1598 response = DLM_MASTER_RESP_MAYBE;
1599 }
1600 if (set_maybe)
1601 set_bit(request->node_idx, tmpmle->maybe_map);
1602 spin_unlock(&tmpmle->spinlock);
1603
1604 spin_unlock(&dlm->master_lock);
1605 spin_unlock(&res->spinlock);
1606
1607 /* keep the mle attached to heartbeat events */
1608 dlm_put_mle(tmpmle);
1609 if (mle)
1610 kmem_cache_free(dlm_mle_cache, mle);
1611 goto send_response;
1612 }
1613
1614 /*
1615 * lockres doesn't exist on this node
1616 * if there is an MLE_BLOCK, return NO
1617 * if there is an MLE_MASTER, return MAYBE
1618 * otherwise, add an MLE_BLOCK, return NO
1619 */
1620 spin_lock(&dlm->master_lock);
1621 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1622 if (!found) {
1623 /* this lockid has never been seen on this node yet */
1624 // mlog(0, "no mle found\n");
1625 if (!mle) {
1626 spin_unlock(&dlm->master_lock);
1627 spin_unlock(&dlm->spinlock);
1628
1629 mle = (struct dlm_master_list_entry *)
ad8100e0 1630 kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
6714d8e8 1631 if (!mle) {
6714d8e8 1632 response = DLM_MASTER_RESP_ERROR;
9c6510a5 1633 mlog_errno(-ENOMEM);
6714d8e8
KH		 1634 goto send_response;
1635 }
6714d8e8
KH		 1636 goto way_up_top;
1637 }
1638
1639 // mlog(0, "this is second time thru, already allocated, "
1640 // "add the block.\n");
41b8c8a1 1641 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
6714d8e8
KH		 1642 set_bit(request->node_idx, mle->maybe_map);
1643 list_add(&mle->list, &dlm->master_list);
1644 response = DLM_MASTER_RESP_NO;
1645 } else {
1646 // mlog(0, "mle was found\n");
1647 set_maybe = 1;
1648 spin_lock(&tmpmle->spinlock);
9c6510a5
KH		 1649 if (tmpmle->master == dlm->node_num) {
1650 mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1651 BUG();
1652 }
6714d8e8
KH		 1653 if (tmpmle->type == DLM_MLE_BLOCK)
1654 response = DLM_MASTER_RESP_NO;
1655 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1656 mlog(0, "migration mle was found (%u->%u)\n",
1657 tmpmle->master, tmpmle->new_master);
6714d8e8
KH		 1658 /* real master can respond on its own */
1659 response = DLM_MASTER_RESP_NO;
9c6510a5
KH		 1660 } else
1661 response = DLM_MASTER_RESP_MAYBE;
6714d8e8
KH		 1662 if (set_maybe)
1663 set_bit(request->node_idx, tmpmle->maybe_map);
1664 spin_unlock(&tmpmle->spinlock);
1665 }
1666 spin_unlock(&dlm->master_lock);
1667 spin_unlock(&dlm->spinlock);
1668
1669 if (found) {
1670 /* keep the mle attached to heartbeat events */
1671 dlm_put_mle(tmpmle);
1672 }
1673send_response:
9c6510a5
KH		 1674
1675 if (dispatch_assert) {
1676 if (response != DLM_MASTER_RESP_YES)
1677 mlog(ML_ERROR, "invalid response %d\n", response);
1678 if (!res) {
1679 mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1680 BUG();
1681 }
1682 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1683 dlm->node_num, res->lockname.len, res->lockname.name);
1684 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1685 DLM_ASSERT_MASTER_MLE_CLEANUP);
1686 if (ret < 0) {
1687 mlog(ML_ERROR, "failed to dispatch assert master work\n");
1688 response = DLM_MASTER_RESP_ERROR;
1689 }
1690 }
1691
6714d8e8
KH		 1692 dlm_put(dlm);
1693 return response;
1694}
1695
1696/*
1697 * DLM_ASSERT_MASTER_MSG
1698 */
1699
1700
1701/*
1702 * NOTE: this can be used for debugging
1703 * can periodically run all locks owned by this node
1704 * and re-assert across the cluster...
1705 */
ba2bf218
KH		 1706int dlm_do_assert_master(struct dlm_ctxt *dlm,
1707 struct dlm_lock_resource *res,
1708 void *nodemap, u32 flags)
6714d8e8
KH		 1709{
1710 struct dlm_assert_master assert;
1711 int to, tmpret;
1712 struct dlm_node_iter iter;
1713 int ret = 0;
9c6510a5 1714 int reassert;
ba2bf218
KH		 1715 const char *lockname = res->lockname.name;
1716 unsigned int namelen = res->lockname.len;
6714d8e8
KH		 1717
1718 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
9c6510a5
KH		 1719again:
1720 reassert = 0;
6714d8e8
KH		 1721
1722 /* note that if this nodemap is empty, it returns 0 */
1723 dlm_node_iter_init(nodemap, &iter);
1724 while ((to = dlm_node_iter_next(&iter)) >= 0) {
1725 int r = 0;
a9ee4c8a
KH		 1726 struct dlm_master_list_entry *mle = NULL;
1727
6714d8e8
KH		 1728 mlog(0, "sending assert master to %d (%.*s)\n", to,
1729 namelen, lockname);
1730 memset(&assert, 0, sizeof(assert));
1731 assert.node_idx = dlm->node_num;
1732 assert.namelen = namelen;
1733 memcpy(assert.name, lockname, namelen);
1734 assert.flags = cpu_to_be32(flags);
1735
1736 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1737 &assert, sizeof(assert), to, &r);
1738 if (tmpret < 0) {
3b3b84a8 1739 mlog(0, "assert_master returned %d!\n", tmpret);
6714d8e8 1740 if (!dlm_is_host_down(tmpret)) {
3b3b84a8 1741 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
6714d8e8
KH		 1742 BUG();
1743 }
1744 /* a node died. finish out the rest of the nodes. */
3b3b84a8 1745 mlog(0, "link to %d went down!\n", to);
6714d8e8
KH		 1746 /* any nonzero status return will do */
1747 ret = tmpret;
ba2bf218 1748 r = 0;
6714d8e8
KH		 1749 } else if (r < 0) {
1750 /* ok, something horribly messed. kill thyself. */
1751 mlog(ML_ERROR,"during assert master of %.*s to %u, "
1752 "got %d.\n", namelen, lockname, to, r);
a9ee4c8a
KH		 1753 spin_lock(&dlm->spinlock);
1754 spin_lock(&dlm->master_lock);
1755 if (dlm_find_mle(dlm, &mle, (char *)lockname,
1756 namelen)) {
1757 dlm_print_one_mle(mle);
1758 __dlm_put_mle(mle);
1759 }
1760 spin_unlock(&dlm->master_lock);
1761 spin_unlock(&dlm->spinlock);
6714d8e8 1762 BUG();
ba2bf218
KH		 1763 }
1764
1765 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1766 !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1767 mlog(ML_ERROR, "%.*s: very strange, "
1768 "master MLE but no lockres on %u\n",
1769 namelen, lockname, to);
1770 }
1771
1772 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
9c6510a5
KH		 1773 mlog(0, "%.*s: node %u create mles on other "
1774 "nodes and requests a re-assert\n",
1775 namelen, lockname, to);
1776 reassert = 1;
6714d8e8 1777 }
ba2bf218
KH		 1778 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1779 mlog(0, "%.*s: node %u has a reference to this "
1780 "lockres, set the bit in the refmap\n",
1781 namelen, lockname, to);
1782 spin_lock(&res->spinlock);
1783 dlm_lockres_set_refmap_bit(to, res);
1784 spin_unlock(&res->spinlock);
1785 }
6714d8e8
KH		 1786 }
1787
9c6510a5
KH		 1788 if (reassert)
1789 goto again;
1790
6714d8e8
KH		 1791 return ret;
1792}
1793
1794/*
1795 * locks that can be taken here:
1796 * dlm->spinlock
1797 * res->spinlock
1798 * mle->spinlock
1799 * dlm->master_list
1800 *
1801 * if possible, TRIM THIS DOWN!!!
1802 */
1803int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data)
1804{
1805 struct dlm_ctxt *dlm = data;
1806 struct dlm_master_list_entry *mle = NULL;
1807 struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1808 struct dlm_lock_resource *res = NULL;
1809 char *name;
a3d33291 1810 unsigned int namelen, hash;
6714d8e8 1811 u32 flags;
ba2bf218 1812 int master_request = 0, have_lockres_ref = 0;
9c6510a5 1813 int ret = 0;
6714d8e8
KH		 1814
1815 if (!dlm_grab(dlm))
1816 return 0;
1817
1818 name = assert->name;
1819 namelen = assert->namelen;
a3d33291 1820 hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 1821 flags = be32_to_cpu(assert->flags);
1822
1823 if (namelen > DLM_LOCKID_NAME_MAX) {
1824 mlog(ML_ERROR, "Invalid name length!");
1825 goto done;
1826 }
1827
1828 spin_lock(&dlm->spinlock);
1829
1830 if (flags)
1831 mlog(0, "assert_master with flags: %u\n", flags);
1832
1833 /* find the MLE */
1834 spin_lock(&dlm->master_lock);
1835 if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1836 /* not an error, could be master just re-asserting */
1837 mlog(0, "just got an assert_master from %u, but no "
1838 "MLE for it! (%.*s)\n", assert->node_idx,
1839 namelen, name);
1840 } else {
1841 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1842 if (bit >= O2NM_MAX_NODES) {
1843 /* not necessarily an error, though less likely.
1844 * could be master just re-asserting. */
aa852354 1845 mlog(0, "no bits set in the maybe_map, but %u "
6714d8e8
KH		 1846 "is asserting! (%.*s)\n", assert->node_idx,
1847 namelen, name);
1848 } else if (bit != assert->node_idx) {
1849 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1850 mlog(0, "master %u was found, %u should "
1851 "back off\n", assert->node_idx, bit);
1852 } else {
1853 /* with the fix for bug 569, a higher node
1854 * number winning the mastery will respond
1855 * YES to mastery requests, but this node
1856 * had no way of knowing. let it pass. */
aa852354 1857 mlog(0, "%u is the lowest node, "
6714d8e8
KH		 1858 "%u is asserting. (%.*s) %u must "
1859 "have begun after %u won.\n", bit,
1860 assert->node_idx, namelen, name, bit,
1861 assert->node_idx);
1862 }
1863 }
2d1a868c
KH		 1864 if (mle->type == DLM_MLE_MIGRATION) {
1865 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1866 mlog(0, "%s:%.*s: got cleanup assert"
1867 " from %u for migration\n",
1868 dlm->name, namelen, name,
1869 assert->node_idx);
1870 } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1871 mlog(0, "%s:%.*s: got unrelated assert"
1872 " from %u for migration, ignoring\n",
1873 dlm->name, namelen, name,
1874 assert->node_idx);
1875 __dlm_put_mle(mle);
1876 spin_unlock(&dlm->master_lock);
1877 spin_unlock(&dlm->spinlock);
1878 goto done;
1879 }
1880 }
6714d8e8
KH		 1881 }
1882 spin_unlock(&dlm->master_lock);
1883
1884 /* ok everything checks out with the MLE
1885 * now check to see if there is a lockres */
a3d33291 1886 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
6714d8e8
KH		 1887 if (res) {
1888 spin_lock(&res->spinlock);
1889 if (res->state & DLM_LOCK_RES_RECOVERING) {
1890 mlog(ML_ERROR, "%u asserting but %.*s is "
1891 "RECOVERING!\n", assert->node_idx, namelen, name);
1892 goto kill;
1893 }
1894 if (!mle) {
dc2ed195
KH		 1895 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1896 res->owner != assert->node_idx) {
6714d8e8
KH		 1897 mlog(ML_ERROR, "assert_master from "
1898 "%u, but current owner is "
1899 "%u! (%.*s)\n",
1900 assert->node_idx, res->owner,
1901 namelen, name);
1902 goto kill;
1903 }
1904 } else if (mle->type != DLM_MLE_MIGRATION) {
1905 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1906 /* owner is just re-asserting */
1907 if (res->owner == assert->node_idx) {
1908 mlog(0, "owner %u re-asserting on "
1909 "lock %.*s\n", assert->node_idx,
1910 namelen, name);
1911 goto ok;
1912 }
1913 mlog(ML_ERROR, "got assert_master from "
1914 "node %u, but %u is the owner! "
1915 "(%.*s)\n", assert->node_idx,
1916 res->owner, namelen, name);
1917 goto kill;
1918 }
1919 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1920 mlog(ML_ERROR, "got assert from %u, but lock "
1921 "with no owner should be "
1922 "in-progress! (%.*s)\n",
1923 assert->node_idx,
1924 namelen, name);
1925 goto kill;
1926 }
1927 } else /* mle->type == DLM_MLE_MIGRATION */ {
1928 /* should only be getting an assert from new master */
1929 if (assert->node_idx != mle->new_master) {
1930 mlog(ML_ERROR, "got assert from %u, but "
1931 "new master is %u, and old master "
1932 "was %u (%.*s)\n",
1933 assert->node_idx, mle->new_master,
1934 mle->master, namelen, name);
1935 goto kill;
1936 }
1937
1938 }
1939ok:
1940 spin_unlock(&res->spinlock);
1941 }
1942 spin_unlock(&dlm->spinlock);
1943
1944 // mlog(0, "woo! got an assert_master from node %u!\n",
1945 // assert->node_idx);
1946 if (mle) {
9c6510a5
KH		 1947 int extra_ref = 0;
1948 int nn = -1;
a2bf0477 1949 int rr, err = 0;
6714d8e8
KH		 1950
1951 spin_lock(&mle->spinlock);
9c6510a5
KH		 1952 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1953 extra_ref = 1;
1954 else {
1955 /* MASTER mle: if any bits set in the response map
1956 * then the calling node needs to re-assert to clear
1957 * up nodes that this node contacted */
1958 while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1959 nn+1)) < O2NM_MAX_NODES) {
1960 if (nn != dlm->node_num && nn != assert->node_idx)
1961 master_request = 1;
1962 }
1963 }
6714d8e8
KH		 1964 mle->master = assert->node_idx;
1965 atomic_set(&mle->woken, 1);
1966 wake_up(&mle->wq);
1967 spin_unlock(&mle->spinlock);
1968
a2bf0477 1969 if (res) {
6714d8e8 1970 spin_lock(&res->spinlock);
a2bf0477
KH		 1971 if (mle->type == DLM_MLE_MIGRATION) {
1972 mlog(0, "finishing off migration of lockres %.*s, "
1973 "from %u to %u\n",
1974 res->lockname.len, res->lockname.name,
1975 dlm->node_num, mle->new_master);
1976 res->state &= ~DLM_LOCK_RES_MIGRATING;
1977 dlm_change_lockres_owner(dlm, res, mle->new_master);
1978 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1979 } else {
1980 dlm_change_lockres_owner(dlm, res, mle->master);
1981 }
6714d8e8 1982 spin_unlock(&res->spinlock);
ba2bf218 1983 have_lockres_ref = 1;
6714d8e8 1984 }
a2bf0477
KH		 1985
1986 /* master is known, detach if not already detached.
1987 * ensures that only one assert_master call will happen
1988 * on this mle. */
1989 spin_lock(&dlm->spinlock);
1990 spin_lock(&dlm->master_lock);
1991
1992 rr = atomic_read(&mle->mle_refs.refcount);
1993 if (mle->inuse > 0) {
1994 if (extra_ref && rr < 3)
1995 err = 1;
1996 else if (!extra_ref && rr < 2)
1997 err = 1;
1998 } else {
1999 if (extra_ref && rr < 2)
2000 err = 1;
2001 else if (!extra_ref && rr < 1)
2002 err = 1;
2003 }
2004 if (err) {
2005 mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
2006 "that will mess up this node, refs=%d, extra=%d, "
2007 "inuse=%d\n", dlm->name, namelen, name,
2008 assert->node_idx, rr, extra_ref, mle->inuse);
2009 dlm_print_one_mle(mle);
2010 }
2011 list_del_init(&mle->list);
2012 __dlm_mle_detach_hb_events(dlm, mle);
2013 __dlm_put_mle(mle);
6714d8e8
KH		 2014 if (extra_ref) {
2015 /* the assert master message now balances the extra
2016 * ref given by the master / migration request message.
2017 * if this is the last put, it will be removed
2018 * from the list. */
a2bf0477
KH		 2019 __dlm_put_mle(mle);
2020 }
2021 spin_unlock(&dlm->master_lock);
2022 spin_unlock(&dlm->spinlock);
2023 } else if (res) {
2024 if (res->owner != assert->node_idx) {
2025 mlog(0, "assert_master from %u, but current "
2026 "owner is %u (%.*s), no mle\n", assert->node_idx,
2027 res->owner, namelen, name);
6714d8e8
KH		 2028 }
2029 }
2030
2031done:
9c6510a5 2032 ret = 0;
6714d8e8
KH		 2033 if (res)
2034 dlm_lockres_put(res);
2035 dlm_put(dlm);
9c6510a5
KH		 2036 if (master_request) {
2037 mlog(0, "need to tell master to reassert\n");
ba2bf218
KH		 2038 /* positive. negative would shoot down the node. */
2039 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2040 if (!have_lockres_ref) {
2041 mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2042 "mle present here for %s:%.*s, but no lockres!\n",
2043 assert->node_idx, dlm->name, namelen, name);
2044 }
2045 }
2046 if (have_lockres_ref) {
2047 /* let the master know we have a reference to the lockres */
2048 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2049 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2050 dlm->name, namelen, name, assert->node_idx);
9c6510a5
KH		 2051 }
2052 return ret;
6714d8e8
KH		 2053
2054kill:
2055 /* kill the caller! */
a9ee4c8a
KH		 2056 mlog(ML_ERROR, "Bad message received from another node. Dumping state "
2057 "and killing the other node now! This node is OK and can continue.\n");
2058 __dlm_print_one_lock_resource(res);
6714d8e8
KH		 2059 spin_unlock(&res->spinlock);
2060 spin_unlock(&dlm->spinlock);
2061 dlm_lockres_put(res);
6714d8e8
KH		 2062 dlm_put(dlm);
2063 return -EINVAL;
2064}
2065
2066int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2067 struct dlm_lock_resource *res,
2068 int ignore_higher, u8 request_from, u32 flags)
2069{
2070 struct dlm_work_item *item;
cd861280 2071 item = kzalloc(sizeof(*item), GFP_NOFS);
6714d8e8
KH		 2072 if (!item)
2073 return -ENOMEM;
2074
2075
2076 /* queue up work for dlm_assert_master_worker */
2077 dlm_grab(dlm); /* get an extra ref for the work item */
2078 dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2079 item->u.am.lockres = res; /* already have a ref */
2080 /* can optionally ignore node numbers higher than this node */
2081 item->u.am.ignore_higher = ignore_higher;
2082 item->u.am.request_from = request_from;
2083 item->u.am.flags = flags;
2084
9c6510a5
KH		 2085 if (ignore_higher)
2086 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2087 res->lockname.name);
2088
6714d8e8
KH		 2089 spin_lock(&dlm->work_lock);
2090 list_add_tail(&item->list, &dlm->work_list);
2091 spin_unlock(&dlm->work_lock);
2092
3156d267 2093 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
6714d8e8
KH		 2094 return 0;
2095}
2096
2097static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2098{
2099 struct dlm_ctxt *dlm = data;
2100 int ret = 0;
2101 struct dlm_lock_resource *res;
2102 unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2103 int ignore_higher;
2104 int bit;
2105 u8 request_from;
2106 u32 flags;
2107
2108 dlm = item->dlm;
2109 res = item->u.am.lockres;
2110 ignore_higher = item->u.am.ignore_higher;
2111 request_from = item->u.am.request_from;
2112 flags = item->u.am.flags;
2113
2114 spin_lock(&dlm->spinlock);
2115 memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2116 spin_unlock(&dlm->spinlock);
2117
2118 clear_bit(dlm->node_num, nodemap);
2119 if (ignore_higher) {
2120 /* if is this just to clear up mles for nodes below
2121 * this node, do not send the message to the original
2122 * caller or any node number higher than this */
2123 clear_bit(request_from, nodemap);
2124 bit = dlm->node_num;
2125 while (1) {
2126 bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2127 bit+1);
2128 if (bit >= O2NM_MAX_NODES)
2129 break;
2130 clear_bit(bit, nodemap);
2131 }
2132 }
2133
36407488
KH		 2134 /*
2135 * If we're migrating this lock to someone else, we are no
2136 * longer allowed to assert out own mastery. OTOH, we need to
2137 * prevent migration from starting while we're still asserting
2138 * our dominance. The reserved ast delays migration.
2139 */
2140 spin_lock(&res->spinlock);
2141 if (res->state & DLM_LOCK_RES_MIGRATING) {
2142 mlog(0, "Someone asked us to assert mastery, but we're "
2143 "in the middle of migration. Skipping assert, "
2144 "the new master will handle that.\n");
2145 spin_unlock(&res->spinlock);
2146 goto put;
2147 } else
2148 __dlm_lockres_reserve_ast(res);
2149 spin_unlock(&res->spinlock);
2150
6714d8e8
KH		 2151 /* this call now finishes out the nodemap
2152 * even if one or more nodes die */
2153 mlog(0, "worker about to master %.*s here, this=%u\n",
2154 res->lockname.len, res->lockname.name, dlm->node_num);
ba2bf218 2155 ret = dlm_do_assert_master(dlm, res, nodemap, flags);
6714d8e8
KH		 2156 if (ret < 0) {
2157 /* no need to restart, we are done */
3b3b84a8
KH		 2158 if (!dlm_is_host_down(ret))
2159 mlog_errno(ret);
6714d8e8
KH		 2160 }
2161
36407488
KH		 2162 /* Ok, we've asserted ourselves. Let's let migration start. */
2163 dlm_lockres_release_ast(dlm, res);
2164
2165put:
6714d8e8
KH		 2166 dlm_lockres_put(res);
2167
2168 mlog(0, "finished with dlm_assert_master_worker\n");
2169}
2170
c03872f5
KH		 2171/* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2172 * We cannot wait for node recovery to complete to begin mastering this
2173 * lockres because this lockres is used to kick off recovery! ;-)
2174 * So, do a pre-check on all living nodes to see if any of those nodes
2175 * think that $RECOVERY is currently mastered by a dead node. If so,
2176 * we wait a short time to allow that node to get notified by its own
2177 * heartbeat stack, then check again. All $RECOVERY lock resources
2178 * mastered by dead nodes are purged when the hearbeat callback is
2179 * fired, so we can know for sure that it is safe to continue once
2180 * the node returns a live node or no node. */
2181static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2182 struct dlm_lock_resource *res)
2183{
2184 struct dlm_node_iter iter;
2185 int nodenum;
2186 int ret = 0;
2187 u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2188
2189 spin_lock(&dlm->spinlock);
2190 dlm_node_iter_init(dlm->domain_map, &iter);
2191 spin_unlock(&dlm->spinlock);
2192
2193 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2194 /* do not send to self */
2195 if (nodenum == dlm->node_num)
2196 continue;
2197 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2198 if (ret < 0) {
2199 mlog_errno(ret);
2200 if (!dlm_is_host_down(ret))
2201 BUG();
2202 /* host is down, so answer for that node would be
2203 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
f42a100b 2204 ret = 0;
c03872f5
KH		 2205 }
2206
2207 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2208 /* check to see if this master is in the recovery map */
2209 spin_lock(&dlm->spinlock);
2210 if (test_bit(master, dlm->recovery_map)) {
2211 mlog(ML_NOTICE, "%s: node %u has not seen "
2212 "node %u go down yet, and thinks the "
2213 "dead node is mastering the recovery "
2214 "lock. must wait.\n", dlm->name,
2215 nodenum, master);
2216 ret = -EAGAIN;
2217 }
2218 spin_unlock(&dlm->spinlock);
2219 mlog(0, "%s: reco lock master is %u\n", dlm->name,
2220 master);
2221 break;
2222 }
2223 }
2224 return ret;
2225}
2226
ba2bf218
KH		 2227/*
2228 * DLM_DEREF_LOCKRES_MSG
2229 */
2230
2231int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2232{
2233 struct dlm_deref_lockres deref;
2234 int ret = 0, r;
2235 const char *lockname;
2236 unsigned int namelen;
2237
2238 lockname = res->lockname.name;
2239 namelen = res->lockname.len;
2240 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2241
2242 mlog(0, "%s:%.*s: sending deref to %d\n",
2243 dlm->name, namelen, lockname, res->owner);
2244 memset(&deref, 0, sizeof(deref));
2245 deref.node_idx = dlm->node_num;
2246 deref.namelen = namelen;
2247 memcpy(deref.name, lockname, namelen);
2248
2249 ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2250 &deref, sizeof(deref), res->owner, &r);
2251 if (ret < 0)
2252 mlog_errno(ret);
2253 else if (r < 0) {
2254 /* BAD. other node says I did not have a ref. */
2255 mlog(ML_ERROR,"while dropping ref on %s:%.*s "
2256 "(master=%u) got %d.\n", dlm->name, namelen,
2257 lockname, res->owner, r);
2258 dlm_print_one_lock_resource(res);
2259 BUG();
2260 }
2261 return ret;
2262}
2263
2264int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data)
2265{
2266 struct dlm_ctxt *dlm = data;
2267 struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2268 struct dlm_lock_resource *res = NULL;
2269 char *name;
2270 unsigned int namelen;
2271 int ret = -EINVAL;
2272 u8 node;
2273 unsigned int hash;
2274
2275 if (!dlm_grab(dlm))
2276 return 0;
2277
2278 name = deref->name;
2279 namelen = deref->namelen;
2280 node = deref->node_idx;
2281
2282 if (namelen > DLM_LOCKID_NAME_MAX) {
2283 mlog(ML_ERROR, "Invalid name length!");
2284 goto done;
2285 }
2286 if (deref->node_idx >= O2NM_MAX_NODES) {
2287 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2288 goto done;
2289 }
2290
2291 hash = dlm_lockid_hash(name, namelen);
2292
2293 spin_lock(&dlm->spinlock);
2294 res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2295 if (!res) {
2296 spin_unlock(&dlm->spinlock);
2297 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2298 dlm->name, namelen, name);
2299 goto done;
2300 }
2301 spin_unlock(&dlm->spinlock);
2302
2303 spin_lock(&res->spinlock);
2304 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2305 if (test_bit(node, res->refmap)) {
2306 ret = 0;
2307 dlm_lockres_clear_refmap_bit(node, res);
2308 } else {
2309 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2310 "but it is already dropped!\n", dlm->name, namelen,
2311 name, node);
2312 __dlm_print_one_lock_resource(res);
2313 }
2314 spin_unlock(&res->spinlock);
2315
2316 if (!ret)
2317 dlm_lockres_calc_usage(dlm, res);
2318done:
2319 if (res)
2320 dlm_lockres_put(res);
2321 dlm_put(dlm);
2322 return ret;
2323}
2324
6714d8e8
KH		 2325
2326/*
2327 * DLM_MIGRATE_LOCKRES
2328 */
2329
2330
faf0ec9f
AB		 2331static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2332 struct dlm_lock_resource *res,
2333 u8 target)
6714d8e8
KH		 2334{
2335 struct dlm_master_list_entry *mle = NULL;
2336 struct dlm_master_list_entry *oldmle = NULL;
2337 struct dlm_migratable_lockres *mres = NULL;
2338 int ret = -EINVAL;
2339 const char *name;
2340 unsigned int namelen;
2341 int mle_added = 0;
2342 struct list_head *queue, *iter;
2343 int i;
2344 struct dlm_lock *lock;
2345 int empty = 1;
2346
2347 if (!dlm_grab(dlm))
2348 return -EINVAL;
2349
2350 name = res->lockname.name;
2351 namelen = res->lockname.len;
2352
2353 mlog(0, "migrating %.*s to %u\n", namelen, name, target);
2354
2355 /*
2356 * ensure this lockres is a proper candidate for migration
2357 */
2358 spin_lock(&res->spinlock);
2359 if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
2360 mlog(0, "cannot migrate lockres with unknown owner!\n");
2361 spin_unlock(&res->spinlock);
2362 goto leave;
2363 }
2364 if (res->owner != dlm->node_num) {
2365 mlog(0, "cannot migrate lockres this node doesn't own!\n");
2366 spin_unlock(&res->spinlock);
2367 goto leave;
2368 }
2369 mlog(0, "checking queues...\n");
2370 queue = &res->granted;
2371 for (i=0; i<3; i++) {
2372 list_for_each(iter, queue) {
2373 lock = list_entry (iter, struct dlm_lock, list);
2374 empty = 0;
2375 if (lock->ml.node == dlm->node_num) {
2376 mlog(0, "found a lock owned by this node "
2377 "still on the %s queue! will not "
2378 "migrate this lockres\n",
2379 i==0 ? "granted" :
2380 (i==1 ? "converting" : "blocked"));
2381 spin_unlock(&res->spinlock);
2382 ret = -ENOTEMPTY;
2383 goto leave;
2384 }
2385 }
2386 queue++;
2387 }
2388 mlog(0, "all locks on this lockres are nonlocal. continuing\n");
2389 spin_unlock(&res->spinlock);
2390
2391 /* no work to do */
2392 if (empty) {
2393 mlog(0, "no locks were found on this lockres! done!\n");
2394 ret = 0;
2395 goto leave;
2396 }
2397
2398 /*
2399 * preallocate up front
2400 * if this fails, abort
2401 */
2402
2403 ret = -ENOMEM;
ad8100e0 2404 mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
6714d8e8
KH		 2405 if (!mres) {
2406 mlog_errno(ret);
2407 goto leave;
2408 }
2409
2410 mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
ad8100e0 2411 GFP_NOFS);
6714d8e8
KH		 2412 if (!mle) {
2413 mlog_errno(ret);
2414 goto leave;
2415 }
2416 ret = 0;
2417
2418 /*
2419 * find a node to migrate the lockres to
2420 */
2421
2422 mlog(0, "picking a migration node\n");
2423 spin_lock(&dlm->spinlock);
2424 /* pick a new node */
2425 if (!test_bit(target, dlm->domain_map) ||
2426 target >= O2NM_MAX_NODES) {
2427 target = dlm_pick_migration_target(dlm, res);
2428 }
2429 mlog(0, "node %u chosen for migration\n", target);
2430
2431 if (target >= O2NM_MAX_NODES ||
2432 !test_bit(target, dlm->domain_map)) {
2433 /* target chosen is not alive */
2434 ret = -EINVAL;
2435 }
2436
2437 if (ret) {
2438 spin_unlock(&dlm->spinlock);
2439 goto fail;
2440 }
2441
2442 mlog(0, "continuing with target = %u\n", target);
2443
2444 /*
2445 * clear any existing master requests and
2446 * add the migration mle to the list
2447 */
2448 spin_lock(&dlm->master_lock);
2449 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2450 namelen, target, dlm->node_num);
2451 spin_unlock(&dlm->master_lock);
2452 spin_unlock(&dlm->spinlock);
2453
2454 if (ret == -EEXIST) {
2455 mlog(0, "another process is already migrating it\n");
2456 goto fail;
2457 }
2458 mle_added = 1;
2459
2460 /*
2461 * set the MIGRATING flag and flush asts
2462 * if we fail after this we need to re-dirty the lockres
2463 */
2464 if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2465 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2466 "the target went down.\n", res->lockname.len,
2467 res->lockname.name, target);
2468 spin_lock(&res->spinlock);
2469 res->state &= ~DLM_LOCK_RES_MIGRATING;
2470 spin_unlock(&res->spinlock);
2471 ret = -EINVAL;
2472 }
2473
2474fail:
2475 if (oldmle) {
2476 /* master is known, detach if not already detached */
2477 dlm_mle_detach_hb_events(dlm, oldmle);
2478 dlm_put_mle(oldmle);
2479 }
2480
2481 if (ret < 0) {
2482 if (mle_added) {
2483 dlm_mle_detach_hb_events(dlm, mle);
2484 dlm_put_mle(mle);
2485 } else if (mle) {
2486 kmem_cache_free(dlm_mle_cache, mle);
2487 }
2488 goto leave;
2489 }
2490
2491 /*
2492 * at this point, we have a migration target, an mle
2493 * in the master list, and the MIGRATING flag set on
2494 * the lockres
2495 */
2496
1cd04dbe
KH		 2497 /* now that remote nodes are spinning on the MIGRATING flag,
2498 * ensure that all assert_master work is flushed. */
2499 flush_workqueue(dlm->dlm_worker);
6714d8e8
KH		 2500
2501 /* get an extra reference on the mle.
2502 * otherwise the assert_master from the new
2503 * master will destroy this.
2504 * also, make sure that all callers of dlm_get_mle
2505 * take both dlm->spinlock and dlm->master_lock */
2506 spin_lock(&dlm->spinlock);
2507 spin_lock(&dlm->master_lock);
a2bf0477 2508 dlm_get_mle_inuse(mle);
6714d8e8
KH		 2509 spin_unlock(&dlm->master_lock);
2510 spin_unlock(&dlm->spinlock);
2511
2512 /* notify new node and send all lock state */
2513 /* call send_one_lockres with migration flag.
2514 * this serves as notice to the target node that a
2515 * migration is starting. */
2516 ret = dlm_send_one_lockres(dlm, res, mres, target,
2517 DLM_MRES_MIGRATION);
2518
2519 if (ret < 0) {
2520 mlog(0, "migration to node %u failed with %d\n",
2521 target, ret);
2522 /* migration failed, detach and clean up mle */
2523 dlm_mle_detach_hb_events(dlm, mle);
2524 dlm_put_mle(mle);
a2bf0477
KH		 2525 dlm_put_mle_inuse(mle);
2526 spin_lock(&res->spinlock);
2527 res->state &= ~DLM_LOCK_RES_MIGRATING;
2528 spin_unlock(&res->spinlock);
6714d8e8
KH		 2529 goto leave;
2530 }
2531
2532 /* at this point, the target sends a message to all nodes,
2533 * (using dlm_do_migrate_request). this node is skipped since
2534 * we had to put an mle in the list to begin the process. this
2535 * node now waits for target to do an assert master. this node
2536 * will be the last one notified, ensuring that the migration
2537 * is complete everywhere. if the target dies while this is
2538 * going on, some nodes could potentially see the target as the
2539 * master, so it is important that my recovery finds the migration
2540 * mle and sets the master to UNKNONWN. */
2541
2542
2543 /* wait for new node to assert master */
2544 while (1) {
2545 ret = wait_event_interruptible_timeout(mle->wq,
2546 (atomic_read(&mle->woken) == 1),
2547 msecs_to_jiffies(5000));
2548
2549 if (ret >= 0) {
2550 if (atomic_read(&mle->woken) == 1 ||
2551 res->owner == target)
2552 break;
2553
1cd04dbe
KH		 2554 mlog(0, "%s:%.*s: timed out during migration\n",
2555 dlm->name, res->lockname.len, res->lockname.name);
e2faea4c
KH		 2556 /* avoid hang during shutdown when migrating lockres
2557 * to a node which also goes down */
2558 if (dlm_is_node_dead(dlm, target)) {
aa852354
KH		 2559 mlog(0, "%s:%.*s: expected migration "
2560 "target %u is no longer up, restarting\n",
e2faea4c
KH		 2561 dlm->name, res->lockname.len,
2562 res->lockname.name, target);
1cd04dbe
KH		 2563 ret = -EINVAL;
2564 /* migration failed, detach and clean up mle */
2565 dlm_mle_detach_hb_events(dlm, mle);
2566 dlm_put_mle(mle);
2567 dlm_put_mle_inuse(mle);
2568 spin_lock(&res->spinlock);
2569 res->state &= ~DLM_LOCK_RES_MIGRATING;
2570 spin_unlock(&res->spinlock);
2571 goto leave;
e2faea4c 2572 }
1cd04dbe
KH		 2573 } else
2574 mlog(0, "%s:%.*s: caught signal during migration\n",
2575 dlm->name, res->lockname.len, res->lockname.name);
6714d8e8
KH		 2576 }
2577
2578 /* all done, set the owner, clear the flag */
2579 spin_lock(&res->spinlock);
2580 dlm_set_lockres_owner(dlm, res, target);
2581 res->state &= ~DLM_LOCK_RES_MIGRATING;
2582 dlm_remove_nonlocal_locks(dlm, res);
2583 spin_unlock(&res->spinlock);
2584 wake_up(&res->wq);
2585
2586 /* master is known, detach if not already detached */
2587 dlm_mle_detach_hb_events(dlm, mle);
a2bf0477 2588 dlm_put_mle_inuse(mle);
6714d8e8
KH		 2589 ret = 0;
2590
2591 dlm_lockres_calc_usage(dlm, res);
2592
2593leave:
2594 /* re-dirty the lockres if we failed */
2595 if (ret < 0)
2596 dlm_kick_thread(dlm, res);
2597
2598 /* TODO: cleanup */
2599 if (mres)
2600 free_page((unsigned long)mres);
2601
2602 dlm_put(dlm);
2603
2604 mlog(0, "returning %d\n", ret);
2605 return ret;
2606}
6714d8e8 2607
ba2bf218
KH		 2608#define DLM_MIGRATION_RETRY_MS 100
2609
2610/* Should be called only after beginning the domain leave process.
2611 * There should not be any remaining locks on nonlocal lock resources,
2612 * and there should be no local locks left on locally mastered resources.
2613 *
2614 * Called with the dlm spinlock held, may drop it to do migration, but
2615 * will re-acquire before exit.
2616 *
2617 * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped */
2618int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2619{
2620 int ret;
2621 int lock_dropped = 0;
2622
2623 if (res->owner != dlm->node_num) {
2624 if (!__dlm_lockres_unused(res)) {
2625 mlog(ML_ERROR, "%s:%.*s: this node is not master, "
2626 "trying to free this but locks remain\n",
2627 dlm->name, res->lockname.len, res->lockname.name);
2628 }
2629 goto leave;
2630 }
2631
2632 /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2633 spin_unlock(&dlm->spinlock);
2634 lock_dropped = 1;
2635 while (1) {
2636 ret = dlm_migrate_lockres(dlm, res, O2NM_MAX_NODES);
2637 if (ret >= 0)
2638 break;
2639 if (ret == -ENOTEMPTY) {
2640 mlog(ML_ERROR, "lockres %.*s still has local locks!\n",
2641 res->lockname.len, res->lockname.name);
2642 BUG();
2643 }
2644
2645 mlog(0, "lockres %.*s: migrate failed, "
2646 "retrying\n", res->lockname.len,
2647 res->lockname.name);
2648 msleep(DLM_MIGRATION_RETRY_MS);
2649 }
2650 spin_lock(&dlm->spinlock);
2651leave:
2652 return lock_dropped;
2653}
2654
6714d8e8
KH		 2655int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2656{
2657 int ret;
2658 spin_lock(&dlm->ast_lock);
2659 spin_lock(&lock->spinlock);
2660 ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2661 spin_unlock(&lock->spinlock);
2662 spin_unlock(&dlm->ast_lock);
2663 return ret;
2664}
2665
2666static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2667 struct dlm_lock_resource *res,
2668 u8 mig_target)
2669{
2670 int can_proceed;
2671 spin_lock(&res->spinlock);
2672 can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2673 spin_unlock(&res->spinlock);
2674
2675 /* target has died, so make the caller break out of the
2676 * wait_event, but caller must recheck the domain_map */
2677 spin_lock(&dlm->spinlock);
2678 if (!test_bit(mig_target, dlm->domain_map))
2679 can_proceed = 1;
2680 spin_unlock(&dlm->spinlock);
2681 return can_proceed;
2682}
2683
faf0ec9f
AB		 2684static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2685 struct dlm_lock_resource *res)
6714d8e8
KH		 2686{
2687 int ret;
2688 spin_lock(&res->spinlock);
2689 ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2690 spin_unlock(&res->spinlock);
2691 return ret;
2692}
2693
2694
2695static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2696 struct dlm_lock_resource *res,
2697 u8 target)
2698{
2699 int ret = 0;
2700
2701 mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2702 res->lockname.len, res->lockname.name, dlm->node_num,
2703 target);
2704 /* need to set MIGRATING flag on lockres. this is done by
2705 * ensuring that all asts have been flushed for this lockres. */
2706 spin_lock(&res->spinlock);
2707 BUG_ON(res->migration_pending);
2708 res->migration_pending = 1;
2709 /* strategy is to reserve an extra ast then release
2710 * it below, letting the release do all of the work */
2711 __dlm_lockres_reserve_ast(res);
2712 spin_unlock(&res->spinlock);
2713
ddc09c8d 2714 /* now flush all the pending asts */
6714d8e8 2715 dlm_kick_thread(dlm, res);
ddc09c8d
KH		 2716 /* before waiting on DIRTY, block processes which may
2717 * try to dirty the lockres before MIGRATING is set */
2718 spin_lock(&res->spinlock);
2719 BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2720 res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2721 spin_unlock(&res->spinlock);
2722 /* now wait on any pending asts and the DIRTY state */
6714d8e8
KH		 2723 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2724 dlm_lockres_release_ast(dlm, res);
2725
2726 mlog(0, "about to wait on migration_wq, dirty=%s\n",
2727 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2728 /* if the extra ref we just put was the final one, this
2729 * will pass thru immediately. otherwise, we need to wait
2730 * for the last ast to finish. */
2731again:
2732 ret = wait_event_interruptible_timeout(dlm->migration_wq,
2733 dlm_migration_can_proceed(dlm, res, target),
2734 msecs_to_jiffies(1000));
2735 if (ret < 0) {
2736 mlog(0, "woken again: migrating? %s, dead? %s\n",
2737 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2738 test_bit(target, dlm->domain_map) ? "no":"yes");
2739 } else {
2740 mlog(0, "all is well: migrating? %s, dead? %s\n",
2741 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2742 test_bit(target, dlm->domain_map) ? "no":"yes");
2743 }
2744 if (!dlm_migration_can_proceed(dlm, res, target)) {
2745 mlog(0, "trying again...\n");
2746 goto again;
2747 }
ddc09c8d
KH		 2748 /* now that we are sure the MIGRATING state is there, drop
2749 * the unneded state which blocked threads trying to DIRTY */
2750 spin_lock(&res->spinlock);
2751 BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2752 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2753 res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2754 spin_unlock(&res->spinlock);
6714d8e8
KH		 2755
2756 /* did the target go down or die? */
2757 spin_lock(&dlm->spinlock);
2758 if (!test_bit(target, dlm->domain_map)) {
2759 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2760 target);
2761 ret = -EHOSTDOWN;
2762 }
2763 spin_unlock(&dlm->spinlock);
2764
2765 /*
2766 * at this point:
2767 *
2768 * o the DLM_LOCK_RES_MIGRATING flag is set
2769 * o there are no pending asts on this lockres
2770 * o all processes trying to reserve an ast on this
2771 * lockres must wait for the MIGRATING flag to clear
2772 */
2773 return ret;
2774}
2775
2776/* last step in the migration process.
2777 * original master calls this to free all of the dlm_lock
2778 * structures that used to be for other nodes. */
2779static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2780 struct dlm_lock_resource *res)
2781{
2782 struct list_head *iter, *iter2;
2783 struct list_head *queue = &res->granted;
ba2bf218 2784 int i, bit;
6714d8e8
KH		 2785 struct dlm_lock *lock;
2786
2787 assert_spin_locked(&res->spinlock);
2788
2789 BUG_ON(res->owner == dlm->node_num);
2790
2791 for (i=0; i<3; i++) {
2792 list_for_each_safe(iter, iter2, queue) {
2793 lock = list_entry (iter, struct dlm_lock, list);
2794 if (lock->ml.node != dlm->node_num) {
2795 mlog(0, "putting lock for node %u\n",
2796 lock->ml.node);
2797 /* be extra careful */
2798 BUG_ON(!list_empty(&lock->ast_list));
2799 BUG_ON(!list_empty(&lock->bast_list));
2800 BUG_ON(lock->ast_pending);
2801 BUG_ON(lock->bast_pending);
ba2bf218 2802 dlm_lockres_clear_refmap_bit(lock->ml.node, res);
6714d8e8
KH		 2803 list_del_init(&lock->list);
2804 dlm_lock_put(lock);
2805 }
2806 }
2807 queue++;
2808 }
ba2bf218
KH		 2809 bit = 0;
2810 while (1) {
2811 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2812 if (bit >= O2NM_MAX_NODES)
2813 break;
2814 /* do not clear the local node reference, if there is a
2815 * process holding this, let it drop the ref itself */
2816 if (bit != dlm->node_num) {
2817 mlog(0, "%s:%.*s: node %u had a ref to this "
2818 "migrating lockres, clearing\n", dlm->name,
2819 res->lockname.len, res->lockname.name, bit);
2820 dlm_lockres_clear_refmap_bit(bit, res);
2821 }
2822 bit++;
2823 }
6714d8e8
KH		 2824}
2825
2826/* for now this is not too intelligent. we will
2827 * need stats to make this do the right thing.
2828 * this just finds the first lock on one of the
2829 * queues and uses that node as the target. */
2830static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2831 struct dlm_lock_resource *res)
2832{
2833 int i;
2834 struct list_head *queue = &res->granted;
2835 struct list_head *iter;
2836 struct dlm_lock *lock;
2837 int nodenum;
2838
2839 assert_spin_locked(&dlm->spinlock);
2840
2841 spin_lock(&res->spinlock);
2842 for (i=0; i<3; i++) {
2843 list_for_each(iter, queue) {
2844 /* up to the caller to make sure this node
2845 * is alive */
2846 lock = list_entry (iter, struct dlm_lock, list);
2847 if (lock->ml.node != dlm->node_num) {
2848 spin_unlock(&res->spinlock);
2849 return lock->ml.node;
2850 }
2851 }
2852 queue++;
2853 }
2854 spin_unlock(&res->spinlock);
2855 mlog(0, "have not found a suitable target yet! checking domain map\n");
2856
2857 /* ok now we're getting desperate. pick anyone alive. */
2858 nodenum = -1;
2859 while (1) {
2860 nodenum = find_next_bit(dlm->domain_map,
2861 O2NM_MAX_NODES, nodenum+1);
2862 mlog(0, "found %d in domain map\n", nodenum);
2863 if (nodenum >= O2NM_MAX_NODES)
2864 break;
2865 if (nodenum != dlm->node_num) {
2866 mlog(0, "picking %d\n", nodenum);
2867 return nodenum;
2868 }
2869 }
2870
2871 mlog(0, "giving up. no master to migrate to\n");
2872 return DLM_LOCK_RES_OWNER_UNKNOWN;
2873}
2874
2875
2876
2877/* this is called by the new master once all lockres
2878 * data has been received */
2879static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2880 struct dlm_lock_resource *res,
2881 u8 master, u8 new_master,
2882 struct dlm_node_iter *iter)
2883{
2884 struct dlm_migrate_request migrate;
2885 int ret, status = 0;
2886 int nodenum;
2887
2888 memset(&migrate, 0, sizeof(migrate));
2889 migrate.namelen = res->lockname.len;
2890 memcpy(migrate.name, res->lockname.name, migrate.namelen);
2891 migrate.new_master = new_master;
2892 migrate.master = master;
2893
2894 ret = 0;
2895
2896 /* send message to all nodes, except the master and myself */
2897 while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2898 if (nodenum == master ||
2899 nodenum == new_master)
2900 continue;
2901
2902 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2903 &migrate, sizeof(migrate), nodenum,
2904 &status);
2905 if (ret < 0)
2906 mlog_errno(ret);
2907 else if (status < 0) {
2908 mlog(0, "migrate request (node %u) returned %d!\n",
2909 nodenum, status);
2910 ret = status;
ba2bf218
KH		 2911 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
2912 /* during the migration request we short-circuited
2913 * the mastery of the lockres. make sure we have
2914 * a mastery ref for nodenum */
2915 mlog(0, "%s:%.*s: need ref for node %u\n",
2916 dlm->name, res->lockname.len, res->lockname.name,
2917 nodenum);
2918 spin_lock(&res->spinlock);
2919 dlm_lockres_set_refmap_bit(nodenum, res);
2920 spin_unlock(&res->spinlock);
6714d8e8
KH		 2921 }
2922 }
2923
2924 if (ret < 0)
2925 mlog_errno(ret);
2926
2927 mlog(0, "returning ret=%d\n", ret);
2928 return ret;
2929}
2930
2931
2932/* if there is an existing mle for this lockres, we now know who the master is.
2933 * (the one who sent us *this* message) we can clear it up right away.
2934 * since the process that put the mle on the list still has a reference to it,
2935 * we can unhash it now, set the master and wake the process. as a result,
2936 * we will have no mle in the list to start with. now we can add an mle for
2937 * the migration and this should be the only one found for those scanning the
2938 * list. */
2939int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data)
2940{
2941 struct dlm_ctxt *dlm = data;
2942 struct dlm_lock_resource *res = NULL;
2943 struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
2944 struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
2945 const char *name;
a3d33291 2946 unsigned int namelen, hash;
6714d8e8
KH		 2947 int ret = 0;
2948
2949 if (!dlm_grab(dlm))
2950 return -EINVAL;
2951
2952 name = migrate->name;
2953 namelen = migrate->namelen;
a3d33291 2954 hash = dlm_lockid_hash(name, namelen);
6714d8e8
KH		 2955
2956 /* preallocate.. if this fails, abort */
2957 mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
ad8100e0 2958 GFP_NOFS);
6714d8e8
KH		 2959
2960 if (!mle) {
2961 ret = -ENOMEM;
2962 goto leave;
2963 }
2964
2965 /* check for pre-existing lock */
2966 spin_lock(&dlm->spinlock);
a3d33291 2967 res = __dlm_lookup_lockres(dlm, name, namelen, hash);
6714d8e8
KH		 2968 spin_lock(&dlm->master_lock);
2969
2970 if (res) {
2971 spin_lock(&res->spinlock);
2972 if (res->state & DLM_LOCK_RES_RECOVERING) {
2973 /* if all is working ok, this can only mean that we got
2974 * a migrate request from a node that we now see as
2975 * dead. what can we do here? drop it to the floor? */
2976 spin_unlock(&res->spinlock);
2977 mlog(ML_ERROR, "Got a migrate request, but the "
2978 "lockres is marked as recovering!");
2979 kmem_cache_free(dlm_mle_cache, mle);
2980 ret = -EINVAL; /* need a better solution */
2981 goto unlock;
2982 }
2983 res->state |= DLM_LOCK_RES_MIGRATING;
2984 spin_unlock(&res->spinlock);
2985 }
2986
2987 /* ignore status. only nonzero status would BUG. */
2988 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
2989 name, namelen,
2990 migrate->new_master,
2991 migrate->master);
2992
2993unlock:
2994 spin_unlock(&dlm->master_lock);
2995 spin_unlock(&dlm->spinlock);
2996
2997 if (oldmle) {
2998 /* master is known, detach if not already detached */
2999 dlm_mle_detach_hb_events(dlm, oldmle);
3000 dlm_put_mle(oldmle);
3001 }
3002
3003 if (res)
3004 dlm_lockres_put(res);
3005leave:
3006 dlm_put(dlm);
3007 return ret;
3008}
3009
3010/* must be holding dlm->spinlock and dlm->master_lock
3011 * when adding a migration mle, we can clear any other mles
3012 * in the master list because we know with certainty that
3013 * the master is "master". so we remove any old mle from
3014 * the list after setting it's master field, and then add
3015 * the new migration mle. this way we can hold with the rule
3016 * of having only one mle for a given lock name at all times. */
3017static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3018 struct dlm_lock_resource *res,
3019 struct dlm_master_list_entry *mle,
3020 struct dlm_master_list_entry **oldmle,
3021 const char *name, unsigned int namelen,
3022 u8 new_master, u8 master)
3023{
3024 int found;
3025 int ret = 0;
3026
3027 *oldmle = NULL;
3028
3029 mlog_entry_void();
3030
3031 assert_spin_locked(&dlm->spinlock);
3032 assert_spin_locked(&dlm->master_lock);
3033
3034 /* caller is responsible for any ref taken here on oldmle */
3035 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3036 if (found) {
3037 struct dlm_master_list_entry *tmp = *oldmle;
3038 spin_lock(&tmp->spinlock);
3039 if (tmp->type == DLM_MLE_MIGRATION) {
3040 if (master == dlm->node_num) {
3041 /* ah another process raced me to it */
3042 mlog(0, "tried to migrate %.*s, but some "
3043 "process beat me to it\n",
3044 namelen, name);
3045 ret = -EEXIST;
3046 } else {
3047 /* bad. 2 NODES are trying to migrate! */
3048 mlog(ML_ERROR, "migration error mle: "
3049 "master=%u new_master=%u // request: "
3050 "master=%u new_master=%u // "
3051 "lockres=%.*s\n",
3052 tmp->master, tmp->new_master,
3053 master, new_master,
3054 namelen, name);
3055 BUG();
3056 }
3057 } else {
3058 /* this is essentially what assert_master does */
3059 tmp->master = master;
3060 atomic_set(&tmp->woken, 1);
3061 wake_up(&tmp->wq);
3062 /* remove it from the list so that only one
3063 * mle will be found */
3064 list_del_init(&tmp->list);
ba2bf218
KH		 3065 /* this was obviously WRONG. mle is uninited here. should be tmp. */
3066 __dlm_mle_detach_hb_events(dlm, tmp);
3067 ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3068 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3069 "telling master to get ref for cleared out mle "
3070 "during migration\n", dlm->name, namelen, name,
3071 master, new_master);
6714d8e8
KH		 3072 }
3073 spin_unlock(&tmp->spinlock);
3074 }
3075
3076 /* now add a migration mle to the tail of the list */
3077 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3078 mle->new_master = new_master;
ba2bf218
KH		 3079 /* the new master will be sending an assert master for this.
3080 * at that point we will get the refmap reference */
6714d8e8
KH		 3081 mle->master = master;
3082 /* do this for consistency with other mle types */
3083 set_bit(new_master, mle->maybe_map);
3084 list_add(&mle->list, &dlm->master_list);
3085
3086 return ret;
3087}
3088
3089
3090void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3091{
3092 struct list_head *iter, *iter2;
3093 struct dlm_master_list_entry *mle;
3094 struct dlm_lock_resource *res;
a3d33291 3095 unsigned int hash;
6714d8e8
KH		 3096
3097 mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node);
3098top:
3099 assert_spin_locked(&dlm->spinlock);
3100
3101 /* clean the master list */
3102 spin_lock(&dlm->master_lock);
3103 list_for_each_safe(iter, iter2, &dlm->master_list) {
3104 mle = list_entry(iter, struct dlm_master_list_entry, list);
3105
3106 BUG_ON(mle->type != DLM_MLE_BLOCK &&
3107 mle->type != DLM_MLE_MASTER &&
3108 mle->type != DLM_MLE_MIGRATION);
3109
3110 /* MASTER mles are initiated locally. the waiting
3111 * process will notice the node map change
3112 * shortly. let that happen as normal. */
3113 if (mle->type == DLM_MLE_MASTER)
3114 continue;
3115
3116
3117 /* BLOCK mles are initiated by other nodes.
3118 * need to clean up if the dead node would have
3119 * been the master. */
3120 if (mle->type == DLM_MLE_BLOCK) {
3121 int bit;
3122
3123 spin_lock(&mle->spinlock);
3124 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3125 if (bit != dead_node) {
3126 mlog(0, "mle found, but dead node %u would "
3127 "not have been master\n", dead_node);
3128 spin_unlock(&mle->spinlock);
3129 } else {
3130 /* must drop the refcount by one since the
3131 * assert_master will never arrive. this
3132 * may result in the mle being unlinked and
3133 * freed, but there may still be a process
3134 * waiting in the dlmlock path which is fine. */
3b3b84a8 3135 mlog(0, "node %u was expected master\n",
6714d8e8
KH		 3136 dead_node);
3137 atomic_set(&mle->woken, 1);
3138 spin_unlock(&mle->spinlock);
3139 wake_up(&mle->wq);
f671c09b
KH		 3140 /* do not need events any longer, so detach
3141 * from heartbeat */
3142 __dlm_mle_detach_hb_events(dlm, mle);
6714d8e8
KH		 3143 __dlm_put_mle(mle);
3144 }
3145 continue;
3146 }
3147
3148 /* everything else is a MIGRATION mle */
3149
3150 /* the rule for MIGRATION mles is that the master
3151 * becomes UNKNOWN if *either* the original or
3152 * the new master dies. all UNKNOWN lockreses
3153 * are sent to whichever node becomes the recovery
3154 * master. the new master is responsible for
3155 * determining if there is still a master for
3156 * this lockres, or if he needs to take over
3157 * mastery. either way, this node should expect
3158 * another message to resolve this. */
3159 if (mle->master != dead_node &&
3160 mle->new_master != dead_node)
3161 continue;
3162
3163 /* if we have reached this point, this mle needs to
3164 * be removed from the list and freed. */
3165
3166 /* remove from the list early. NOTE: unlinking
3167 * list_head while in list_for_each_safe */
da01ad05 3168 __dlm_mle_detach_hb_events(dlm, mle);
6714d8e8
KH		 3169 spin_lock(&mle->spinlock);
3170 list_del_init(&mle->list);
3171 atomic_set(&mle->woken, 1);
3172 spin_unlock(&mle->spinlock);
3173 wake_up(&mle->wq);
3174
aa852354
KH		 3175 mlog(0, "%s: node %u died during migration from "
3176 "%u to %u!\n", dlm->name, dead_node,
6714d8e8
KH		 3177 mle->master, mle->new_master);
3178 /* if there is a lockres associated with this
3179 * mle, find it and set its owner to UNKNOWN */
a3d33291 3180 hash = dlm_lockid_hash(mle->u.name.name, mle->u.name.len);
6714d8e8 3181 res = __dlm_lookup_lockres(dlm, mle->u.name.name,
a3d33291 3182 mle->u.name.len, hash);
6714d8e8
KH		 3183 if (res) {
3184 /* unfortunately if we hit this rare case, our
3185 * lock ordering is messed. we need to drop
3186 * the master lock so that we can take the
3187 * lockres lock, meaning that we will have to
3188 * restart from the head of list. */
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,
3194 DLM_LOCK_RES_OWNER_UNKNOWN);
3195 dlm_move_lockres_to_recovery_list(dlm, res);
3196 spin_unlock(&res->spinlock);
3197 dlm_lockres_put(res);
3198
f671c09b
KH		 3199 /* about to get rid of mle, detach from heartbeat */
3200 __dlm_mle_detach_hb_events(dlm, mle);
3201
6714d8e8
KH		 3202 /* dump the mle */
3203 spin_lock(&dlm->master_lock);
3204 __dlm_put_mle(mle);
3205 spin_unlock(&dlm->master_lock);
3206
3207 /* restart */
3208 goto top;
3209 }
3210
3211 /* this may be the last reference */
3212 __dlm_put_mle(mle);
3213 }
3214 spin_unlock(&dlm->master_lock);
3215}
3216
3217
3218int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3219 u8 old_master)
3220{
3221 struct dlm_node_iter iter;
3222 int ret = 0;
3223
3224 spin_lock(&dlm->spinlock);
3225 dlm_node_iter_init(dlm->domain_map, &iter);
3226 clear_bit(old_master, iter.node_map);
3227 clear_bit(dlm->node_num, iter.node_map);
3228 spin_unlock(&dlm->spinlock);
3229
ba2bf218
KH		 3230 /* ownership of the lockres is changing. account for the
3231 * mastery reference here since old_master will briefly have
3232 * a reference after the migration completes */
3233 spin_lock(&res->spinlock);
3234 dlm_lockres_set_refmap_bit(old_master, res);
3235 spin_unlock(&res->spinlock);
3236
6714d8e8
KH		 3237 mlog(0, "now time to do a migrate request to other nodes\n");
3238 ret = dlm_do_migrate_request(dlm, res, old_master,
3239 dlm->node_num, &iter);
3240 if (ret < 0) {
3241 mlog_errno(ret);
3242 goto leave;
3243 }
3244
3245 mlog(0, "doing assert master of %.*s to all except the original node\n",
3246 res->lockname.len, res->lockname.name);
3247 /* this call now finishes out the nodemap
3248 * even if one or more nodes die */
ba2bf218 3249 ret = dlm_do_assert_master(dlm, res, iter.node_map,
6714d8e8
KH		 3250 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3251 if (ret < 0) {
3252 /* no longer need to retry. all living nodes contacted. */
3253 mlog_errno(ret);
3254 ret = 0;
3255 }
3256
3257 memset(iter.node_map, 0, sizeof(iter.node_map));
3258 set_bit(old_master, iter.node_map);
3259 mlog(0, "doing assert master of %.*s back to %u\n",
3260 res->lockname.len, res->lockname.name, old_master);
ba2bf218 3261 ret = dlm_do_assert_master(dlm, res, iter.node_map,
6714d8e8
KH		 3262 DLM_ASSERT_MASTER_FINISH_MIGRATION);
3263 if (ret < 0) {
3264 mlog(0, "assert master to original master failed "
3265 "with %d.\n", ret);
3266 /* the only nonzero status here would be because of
3267 * a dead original node. we're done. */
3268 ret = 0;
3269 }
3270
3271 /* all done, set the owner, clear the flag */
3272 spin_lock(&res->spinlock);
3273 dlm_set_lockres_owner(dlm, res, dlm->node_num);
3274 res->state &= ~DLM_LOCK_RES_MIGRATING;
3275 spin_unlock(&res->spinlock);
3276 /* re-dirty it on the new master */
3277 dlm_kick_thread(dlm, res);
3278 wake_up(&res->wq);
3279leave:
3280 return ret;
3281}
3282
3283/*
3284 * LOCKRES AST REFCOUNT
3285 * this is integral to migration
3286 */
3287
3288/* for future intent to call an ast, reserve one ahead of time.
3289 * this should be called only after waiting on the lockres
3290 * with dlm_wait_on_lockres, and while still holding the
3291 * spinlock after the call. */
3292void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3293{
3294 assert_spin_locked(&res->spinlock);
3295 if (res->state & DLM_LOCK_RES_MIGRATING) {
3296 __dlm_print_one_lock_resource(res);
3297 }
3298 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3299
3300 atomic_inc(&res->asts_reserved);
3301}
3302
3303/*
3304 * used to drop the reserved ast, either because it went unused,
3305 * or because the ast/bast was actually called.
3306 *
3307 * also, if there is a pending migration on this lockres,
3308 * and this was the last pending ast on the lockres,
3309 * atomically set the MIGRATING flag before we drop the lock.
3310 * this is how we ensure that migration can proceed with no
3311 * asts in progress. note that it is ok if the state of the
3312 * queues is such that a lock should be granted in the future
3313 * or that a bast should be fired, because the new master will
3314 * shuffle the lists on this lockres as soon as it is migrated.
3315 */
3316void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3317 struct dlm_lock_resource *res)
3318{
3319 if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3320 return;
3321
3322 if (!res->migration_pending) {
3323 spin_unlock(&res->spinlock);
3324 return;
3325 }
3326
3327 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3328 res->migration_pending = 0;
3329 res->state |= DLM_LOCK_RES_MIGRATING;
3330 spin_unlock(&res->spinlock);
3331 wake_up(&res->wq);
3332 wake_up(&dlm->migration_wq);
3333}
Linux 6.x block layer and io_uring tree(s)