aboutsummaryrefslogtreecommitdiffstats
path: root/fs/ocfs2/dlm/dlmmaster.c
diff options
context:
space:
mode:
Diffstat (limited to 'fs/ocfs2/dlm/dlmmaster.c')
-rw-r--r--fs/ocfs2/dlm/dlmmaster.c2664
1 files changed, 2664 insertions, 0 deletions
diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
new file mode 100644
index 000000000000..27e984f7e4cd
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmmaster.c
@@ -0,0 +1,2664 @@
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"
50#include "dlmdebug.h"
51#include "dlmdomain.h"
52
53#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
54#include "cluster/masklog.h"
55
56enum dlm_mle_type {
57 DLM_MLE_BLOCK,
58 DLM_MLE_MASTER,
59 DLM_MLE_MIGRATION
60};
61
62struct dlm_lock_name
63{
64 u8 len;
65 u8 name[DLM_LOCKID_NAME_MAX];
66};
67
68struct dlm_master_list_entry
69{
70 struct list_head list;
71 struct list_head hb_events;
72 struct dlm_ctxt *dlm;
73 spinlock_t spinlock;
74 wait_queue_head_t wq;
75 atomic_t woken;
76 struct kref mle_refs;
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);
102static int dlm_do_assert_master(struct dlm_ctxt *dlm, const char *lockname,
103 unsigned int namelen, void *nodemap,
104 u32 flags);
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
130#if 0
131/* Code here is included but defined out as it aids debugging */
132
133void dlm_print_one_mle(struct dlm_master_list_entry *mle)
134{
135 int i = 0, refs;
136 char *type;
137 char attached;
138 u8 master;
139 unsigned int namelen;
140 const char *name;
141 struct kref *k;
142
143 k = &mle->mle_refs;
144 if (mle->type == DLM_MLE_BLOCK)
145 type = "BLK";
146 else if (mle->type == DLM_MLE_MASTER)
147 type = "MAS";
148 else
149 type = "MIG";
150 refs = atomic_read(&k->refcount);
151 master = mle->master;
152 attached = (list_empty(&mle->hb_events) ? 'N' : 'Y');
153
154 if (mle->type != DLM_MLE_MASTER) {
155 namelen = mle->u.name.len;
156 name = mle->u.name.name;
157 } else {
158 namelen = mle->u.res->lockname.len;
159 name = mle->u.res->lockname.name;
160 }
161
162 mlog(ML_NOTICE, " #%3d: %3s %3d %3u %3u %c (%d)%.*s\n",
163 i, type, refs, master, mle->new_master, attached,
164 namelen, namelen, name);
165}
166
167static void dlm_dump_mles(struct dlm_ctxt *dlm)
168{
169 struct dlm_master_list_entry *mle;
170 struct list_head *iter;
171
172 mlog(ML_NOTICE, "dumping all mles for domain %s:\n", dlm->name);
173 mlog(ML_NOTICE, " ####: type refs owner new events? lockname nodemap votemap respmap maybemap\n");
174 spin_lock(&dlm->master_lock);
175 list_for_each(iter, &dlm->master_list) {
176 mle = list_entry(iter, struct dlm_master_list_entry, list);
177 dlm_print_one_mle(mle);
178 }
179 spin_unlock(&dlm->master_lock);
180}
181
182int dlm_dump_all_mles(const char __user *data, unsigned int len)
183{
184 struct list_head *iter;
185 struct dlm_ctxt *dlm;
186
187 spin_lock(&dlm_domain_lock);
188 list_for_each(iter, &dlm_domains) {
189 dlm = list_entry (iter, struct dlm_ctxt, list);
190 mlog(ML_NOTICE, "found dlm: %p, name=%s\n", dlm, dlm->name);
191 dlm_dump_mles(dlm);
192 }
193 spin_unlock(&dlm_domain_lock);
194 return len;
195}
196EXPORT_SYMBOL_GPL(dlm_dump_all_mles);
197
198#endif /* 0 */
199
200
201static kmem_cache_t *dlm_mle_cache = NULL;
202
203
204static void dlm_mle_release(struct kref *kref);
205static void dlm_init_mle(struct dlm_master_list_entry *mle,
206 enum dlm_mle_type type,
207 struct dlm_ctxt *dlm,
208 struct dlm_lock_resource *res,
209 const char *name,
210 unsigned int namelen);
211static void dlm_put_mle(struct dlm_master_list_entry *mle);
212static void __dlm_put_mle(struct dlm_master_list_entry *mle);
213static int dlm_find_mle(struct dlm_ctxt *dlm,
214 struct dlm_master_list_entry **mle,
215 char *name, unsigned int namelen);
216
217static int dlm_do_master_request(struct dlm_master_list_entry *mle, int to);
218
219
220static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
221 struct dlm_lock_resource *res,
222 struct dlm_master_list_entry *mle,
223 int *blocked);
224static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
225 struct dlm_lock_resource *res,
226 struct dlm_master_list_entry *mle,
227 int blocked);
228static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
229 struct dlm_lock_resource *res,
230 struct dlm_master_list_entry *mle,
231 struct dlm_master_list_entry **oldmle,
232 const char *name, unsigned int namelen,
233 u8 new_master, u8 master);
234
235static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
236 struct dlm_lock_resource *res);
237static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
238 struct dlm_lock_resource *res);
239static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
240 struct dlm_lock_resource *res,
241 u8 target);
242
243
244int dlm_is_host_down(int errno)
245{
246 switch (errno) {
247 case -EBADF:
248 case -ECONNREFUSED:
249 case -ENOTCONN:
250 case -ECONNRESET:
251 case -EPIPE:
252 case -EHOSTDOWN:
253 case -EHOSTUNREACH:
254 case -ETIMEDOUT:
255 case -ECONNABORTED:
256 case -ENETDOWN:
257 case -ENETUNREACH:
258 case -ENETRESET:
259 case -ESHUTDOWN:
260 case -ENOPROTOOPT:
261 case -EINVAL: /* if returned from our tcp code,
262 this means there is no socket */
263 return 1;
264 }
265 return 0;
266}
267
268
269/*
270 * MASTER LIST FUNCTIONS
271 */
272
273
274/*
275 * regarding master list entries and heartbeat callbacks:
276 *
277 * in order to avoid sleeping and allocation that occurs in
278 * heartbeat, master list entries are simply attached to the
279 * dlm's established heartbeat callbacks. the mle is attached
280 * when it is created, and since the dlm->spinlock is held at
281 * that time, any heartbeat event will be properly discovered
282 * by the mle. the mle needs to be detached from the
283 * dlm->mle_hb_events list as soon as heartbeat events are no
284 * longer useful to the mle, and before the mle is freed.
285 *
286 * as a general rule, heartbeat events are no longer needed by
287 * the mle once an "answer" regarding the lock master has been
288 * received.
289 */
290static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
291 struct dlm_master_list_entry *mle)
292{
293 assert_spin_locked(&dlm->spinlock);
294
295 list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
296}
297
298
299static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
300 struct dlm_master_list_entry *mle)
301{
302 if (!list_empty(&mle->hb_events))
303 list_del_init(&mle->hb_events);
304}
305
306
307static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
308 struct dlm_master_list_entry *mle)
309{
310 spin_lock(&dlm->spinlock);
311 __dlm_mle_detach_hb_events(dlm, mle);
312 spin_unlock(&dlm->spinlock);
313}
314
315/* remove from list and free */
316static void __dlm_put_mle(struct dlm_master_list_entry *mle)
317{
318 struct dlm_ctxt *dlm;
319 dlm = mle->dlm;
320
321 assert_spin_locked(&dlm->spinlock);
322 assert_spin_locked(&dlm->master_lock);
323 BUG_ON(!atomic_read(&mle->mle_refs.refcount));
324
325 kref_put(&mle->mle_refs, dlm_mle_release);
326}
327
328
329/* must not have any spinlocks coming in */
330static void dlm_put_mle(struct dlm_master_list_entry *mle)
331{
332 struct dlm_ctxt *dlm;
333 dlm = mle->dlm;
334
335 spin_lock(&dlm->spinlock);
336 spin_lock(&dlm->master_lock);
337 __dlm_put_mle(mle);
338 spin_unlock(&dlm->master_lock);
339 spin_unlock(&dlm->spinlock);
340}
341
342static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
343{
344 kref_get(&mle->mle_refs);
345}
346
347static void dlm_init_mle(struct dlm_master_list_entry *mle,
348 enum dlm_mle_type type,
349 struct dlm_ctxt *dlm,
350 struct dlm_lock_resource *res,
351 const char *name,
352 unsigned int namelen)
353{
354 assert_spin_locked(&dlm->spinlock);
355
356 mle->dlm = dlm;
357 mle->type = type;
358 INIT_LIST_HEAD(&mle->list);
359 INIT_LIST_HEAD(&mle->hb_events);
360 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
361 spin_lock_init(&mle->spinlock);
362 init_waitqueue_head(&mle->wq);
363 atomic_set(&mle->woken, 0);
364 kref_init(&mle->mle_refs);
365 memset(mle->response_map, 0, sizeof(mle->response_map));
366 mle->master = O2NM_MAX_NODES;
367 mle->new_master = O2NM_MAX_NODES;
368
369 if (mle->type == DLM_MLE_MASTER) {
370 BUG_ON(!res);
371 mle->u.res = res;
372 } else if (mle->type == DLM_MLE_BLOCK) {
373 BUG_ON(!name);
374 memcpy(mle->u.name.name, name, namelen);
375 mle->u.name.len = namelen;
376 } else /* DLM_MLE_MIGRATION */ {
377 BUG_ON(!name);
378 memcpy(mle->u.name.name, name, namelen);
379 mle->u.name.len = namelen;
380 }
381
382 /* copy off the node_map and register hb callbacks on our copy */
383 memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
384 memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
385 clear_bit(dlm->node_num, mle->vote_map);
386 clear_bit(dlm->node_num, mle->node_map);
387
388 /* attach the mle to the domain node up/down events */
389 __dlm_mle_attach_hb_events(dlm, mle);
390}
391
392
393/* returns 1 if found, 0 if not */
394static int dlm_find_mle(struct dlm_ctxt *dlm,
395 struct dlm_master_list_entry **mle,
396 char *name, unsigned int namelen)
397{
398 struct dlm_master_list_entry *tmpmle;
399 struct list_head *iter;
400
401 assert_spin_locked(&dlm->master_lock);
402
403 list_for_each(iter, &dlm->master_list) {
404 tmpmle = list_entry(iter, struct dlm_master_list_entry, list);
405 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
406 continue;
407 dlm_get_mle(tmpmle);
408 *mle = tmpmle;
409 return 1;
410 }
411 return 0;
412}
413
414void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
415{
416 struct dlm_master_list_entry *mle;
417 struct list_head *iter;
418
419 assert_spin_locked(&dlm->spinlock);
420
421 list_for_each(iter, &dlm->mle_hb_events) {
422 mle = list_entry(iter, struct dlm_master_list_entry,
423 hb_events);
424 if (node_up)
425 dlm_mle_node_up(dlm, mle, NULL, idx);
426 else
427 dlm_mle_node_down(dlm, mle, NULL, idx);
428 }
429}
430
431static void dlm_mle_node_down(struct dlm_ctxt *dlm,
432 struct dlm_master_list_entry *mle,
433 struct o2nm_node *node, int idx)
434{
435 spin_lock(&mle->spinlock);
436
437 if (!test_bit(idx, mle->node_map))
438 mlog(0, "node %u already removed from nodemap!\n", idx);
439 else
440 clear_bit(idx, mle->node_map);
441
442 spin_unlock(&mle->spinlock);
443}
444
445static void dlm_mle_node_up(struct dlm_ctxt *dlm,
446 struct dlm_master_list_entry *mle,
447 struct o2nm_node *node, int idx)
448{
449 spin_lock(&mle->spinlock);
450
451 if (test_bit(idx, mle->node_map))
452 mlog(0, "node %u already in node map!\n", idx);
453 else
454 set_bit(idx, mle->node_map);
455
456 spin_unlock(&mle->spinlock);
457}
458
459
460int dlm_init_mle_cache(void)
461{
462 dlm_mle_cache = kmem_cache_create("dlm_mle_cache",
463 sizeof(struct dlm_master_list_entry),
464 0, SLAB_HWCACHE_ALIGN,
465 NULL, NULL);
466 if (dlm_mle_cache == NULL)
467 return -ENOMEM;
468 return 0;
469}
470
471void dlm_destroy_mle_cache(void)
472{
473 if (dlm_mle_cache)
474 kmem_cache_destroy(dlm_mle_cache);
475}
476
477static void dlm_mle_release(struct kref *kref)
478{
479 struct dlm_master_list_entry *mle;
480 struct dlm_ctxt *dlm;
481
482 mlog_entry_void();
483
484 mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
485 dlm = mle->dlm;
486
487 if (mle->type != DLM_MLE_MASTER) {
488 mlog(0, "calling mle_release for %.*s, type %d\n",
489 mle->u.name.len, mle->u.name.name, mle->type);
490 } else {
491 mlog(0, "calling mle_release for %.*s, type %d\n",
492 mle->u.res->lockname.len,
493 mle->u.res->lockname.name, mle->type);
494 }
495 assert_spin_locked(&dlm->spinlock);
496 assert_spin_locked(&dlm->master_lock);
497
498 /* remove from list if not already */
499 if (!list_empty(&mle->list))
500 list_del_init(&mle->list);
501
502 /* detach the mle from the domain node up/down events */
503 __dlm_mle_detach_hb_events(dlm, mle);
504
505 /* NOTE: kfree under spinlock here.
506 * if this is bad, we can move this to a freelist. */
507 kmem_cache_free(dlm_mle_cache, mle);
508}
509
510
511/*
512 * LOCK RESOURCE FUNCTIONS
513 */
514
515static void dlm_set_lockres_owner(struct dlm_ctxt *dlm,
516 struct dlm_lock_resource *res,
517 u8 owner)
518{
519 assert_spin_locked(&res->spinlock);
520
521 mlog_entry("%.*s, %u\n", res->lockname.len, res->lockname.name, owner);
522
523 if (owner == dlm->node_num)
524 atomic_inc(&dlm->local_resources);
525 else if (owner == DLM_LOCK_RES_OWNER_UNKNOWN)
526 atomic_inc(&dlm->unknown_resources);
527 else
528 atomic_inc(&dlm->remote_resources);
529
530 res->owner = owner;
531}
532
533void dlm_change_lockres_owner(struct dlm_ctxt *dlm,
534 struct dlm_lock_resource *res, u8 owner)
535{
536 assert_spin_locked(&res->spinlock);
537
538 if (owner == res->owner)
539 return;
540
541 if (res->owner == dlm->node_num)
542 atomic_dec(&dlm->local_resources);
543 else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN)
544 atomic_dec(&dlm->unknown_resources);
545 else
546 atomic_dec(&dlm->remote_resources);
547
548 dlm_set_lockres_owner(dlm, res, owner);
549}
550
551
552static void dlm_lockres_release(struct kref *kref)
553{
554 struct dlm_lock_resource *res;
555
556 res = container_of(kref, struct dlm_lock_resource, refs);
557
558 /* This should not happen -- all lockres' have a name
559 * associated with them at init time. */
560 BUG_ON(!res->lockname.name);
561
562 mlog(0, "destroying lockres %.*s\n", res->lockname.len,
563 res->lockname.name);
564
565 /* By the time we're ready to blow this guy away, we shouldn't
566 * be on any lists. */
567 BUG_ON(!list_empty(&res->list));
568 BUG_ON(!list_empty(&res->granted));
569 BUG_ON(!list_empty(&res->converting));
570 BUG_ON(!list_empty(&res->blocked));
571 BUG_ON(!list_empty(&res->dirty));
572 BUG_ON(!list_empty(&res->recovering));
573 BUG_ON(!list_empty(&res->purge));
574
575 kfree(res->lockname.name);
576
577 kfree(res);
578}
579
580void dlm_lockres_get(struct dlm_lock_resource *res)
581{
582 kref_get(&res->refs);
583}
584
585void dlm_lockres_put(struct dlm_lock_resource *res)
586{
587 kref_put(&res->refs, dlm_lockres_release);
588}
589
590static void dlm_init_lockres(struct dlm_ctxt *dlm,
591 struct dlm_lock_resource *res,
592 const char *name, unsigned int namelen)
593{
594 char *qname;
595
596 /* If we memset here, we lose our reference to the kmalloc'd
597 * res->lockname.name, so be sure to init every field
598 * correctly! */
599
600 qname = (char *) res->lockname.name;
601 memcpy(qname, name, namelen);
602
603 res->lockname.len = namelen;
604 res->lockname.hash = full_name_hash(name, namelen);
605
606 init_waitqueue_head(&res->wq);
607 spin_lock_init(&res->spinlock);
608 INIT_LIST_HEAD(&res->list);
609 INIT_LIST_HEAD(&res->granted);
610 INIT_LIST_HEAD(&res->converting);
611 INIT_LIST_HEAD(&res->blocked);
612 INIT_LIST_HEAD(&res->dirty);
613 INIT_LIST_HEAD(&res->recovering);
614 INIT_LIST_HEAD(&res->purge);
615 atomic_set(&res->asts_reserved, 0);
616 res->migration_pending = 0;
617
618 kref_init(&res->refs);
619
620 /* just for consistency */
621 spin_lock(&res->spinlock);
622 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
623 spin_unlock(&res->spinlock);
624
625 res->state = DLM_LOCK_RES_IN_PROGRESS;
626
627 res->last_used = 0;
628
629 memset(res->lvb, 0, DLM_LVB_LEN);
630}
631
632struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
633 const char *name,
634 unsigned int namelen)
635{
636 struct dlm_lock_resource *res;
637
638 res = kmalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
639 if (!res)
640 return NULL;
641
642 res->lockname.name = kmalloc(namelen, GFP_KERNEL);
643 if (!res->lockname.name) {
644 kfree(res);
645 return NULL;
646 }
647
648 dlm_init_lockres(dlm, res, name, namelen);
649 return res;
650}
651
652/*
653 * lookup a lock resource by name.
654 * may already exist in the hashtable.
655 * lockid is null terminated
656 *
657 * if not, allocate enough for the lockres and for
658 * the temporary structure used in doing the mastering.
659 *
660 * also, do a lookup in the dlm->master_list to see
661 * if another node has begun mastering the same lock.
662 * if so, there should be a block entry in there
663 * for this name, and we should *not* attempt to master
664 * the lock here. need to wait around for that node
665 * to assert_master (or die).
666 *
667 */
668struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
669 const char *lockid,
670 int flags)
671{
672 struct dlm_lock_resource *tmpres=NULL, *res=NULL;
673 struct dlm_master_list_entry *mle = NULL;
674 struct dlm_master_list_entry *alloc_mle = NULL;
675 int blocked = 0;
676 int ret, nodenum;
677 struct dlm_node_iter iter;
678 unsigned int namelen;
679 int tries = 0;
680
681 BUG_ON(!lockid);
682
683 namelen = strlen(lockid);
684
685 mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
686
687lookup:
688 spin_lock(&dlm->spinlock);
689 tmpres = __dlm_lookup_lockres(dlm, lockid, namelen);
690 if (tmpres) {
691 spin_unlock(&dlm->spinlock);
692 mlog(0, "found in hash!\n");
693 if (res)
694 dlm_lockres_put(res);
695 res = tmpres;
696 goto leave;
697 }
698
699 if (!res) {
700 spin_unlock(&dlm->spinlock);
701 mlog(0, "allocating a new resource\n");
702 /* nothing found and we need to allocate one. */
703 alloc_mle = (struct dlm_master_list_entry *)
704 kmem_cache_alloc(dlm_mle_cache, GFP_KERNEL);
705 if (!alloc_mle)
706 goto leave;
707 res = dlm_new_lockres(dlm, lockid, namelen);
708 if (!res)
709 goto leave;
710 goto lookup;
711 }
712
713 mlog(0, "no lockres found, allocated our own: %p\n", res);
714
715 if (flags & LKM_LOCAL) {
716 /* caller knows it's safe to assume it's not mastered elsewhere
717 * DONE! return right away */
718 spin_lock(&res->spinlock);
719 dlm_change_lockres_owner(dlm, res, dlm->node_num);
720 __dlm_insert_lockres(dlm, res);
721 spin_unlock(&res->spinlock);
722 spin_unlock(&dlm->spinlock);
723 /* lockres still marked IN_PROGRESS */
724 goto wake_waiters;
725 }
726
727 /* check master list to see if another node has started mastering it */
728 spin_lock(&dlm->master_lock);
729
730 /* if we found a block, wait for lock to be mastered by another node */
731 blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
732 if (blocked) {
733 if (mle->type == DLM_MLE_MASTER) {
734 mlog(ML_ERROR, "master entry for nonexistent lock!\n");
735 BUG();
736 } else if (mle->type == DLM_MLE_MIGRATION) {
737 /* migration is in progress! */
738 /* the good news is that we now know the
739 * "current" master (mle->master). */
740
741 spin_unlock(&dlm->master_lock);
742 assert_spin_locked(&dlm->spinlock);
743
744 /* set the lockres owner and hash it */
745 spin_lock(&res->spinlock);
746 dlm_set_lockres_owner(dlm, res, mle->master);
747 __dlm_insert_lockres(dlm, res);
748 spin_unlock(&res->spinlock);
749 spin_unlock(&dlm->spinlock);
750
751 /* master is known, detach */
752 dlm_mle_detach_hb_events(dlm, mle);
753 dlm_put_mle(mle);
754 mle = NULL;
755 goto wake_waiters;
756 }
757 } else {
758 /* go ahead and try to master lock on this node */
759 mle = alloc_mle;
760 /* make sure this does not get freed below */
761 alloc_mle = NULL;
762 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
763 set_bit(dlm->node_num, mle->maybe_map);
764 list_add(&mle->list, &dlm->master_list);
765 }
766
767 /* at this point there is either a DLM_MLE_BLOCK or a
768 * DLM_MLE_MASTER on the master list, so it's safe to add the
769 * lockres to the hashtable. anyone who finds the lock will
770 * still have to wait on the IN_PROGRESS. */
771
772 /* finally add the lockres to its hash bucket */
773 __dlm_insert_lockres(dlm, res);
774 /* get an extra ref on the mle in case this is a BLOCK
775 * if so, the creator of the BLOCK may try to put the last
776 * ref at this time in the assert master handler, so we
777 * need an extra one to keep from a bad ptr deref. */
778 dlm_get_mle(mle);
779 spin_unlock(&dlm->master_lock);
780 spin_unlock(&dlm->spinlock);
781
782 /* must wait for lock to be mastered elsewhere */
783 if (blocked)
784 goto wait;
785
786redo_request:
787 ret = -EINVAL;
788 dlm_node_iter_init(mle->vote_map, &iter);
789 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
790 ret = dlm_do_master_request(mle, nodenum);
791 if (ret < 0)
792 mlog_errno(ret);
793 if (mle->master != O2NM_MAX_NODES) {
794 /* found a master ! */
795 break;
796 }
797 }
798
799wait:
800 /* keep going until the response map includes all nodes */
801 ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
802 if (ret < 0) {
803 mlog(0, "%s:%.*s: node map changed, redo the "
804 "master request now, blocked=%d\n",
805 dlm->name, res->lockname.len,
806 res->lockname.name, blocked);
807 if (++tries > 20) {
808 mlog(ML_ERROR, "%s:%.*s: spinning on "
809 "dlm_wait_for_lock_mastery, blocked=%d\n",
810 dlm->name, res->lockname.len,
811 res->lockname.name, blocked);
812 dlm_print_one_lock_resource(res);
813 /* dlm_print_one_mle(mle); */
814 tries = 0;
815 }
816 goto redo_request;
817 }
818
819 mlog(0, "lockres mastered by %u\n", res->owner);
820 /* make sure we never continue without this */
821 BUG_ON(res->owner == O2NM_MAX_NODES);
822
823 /* master is known, detach if not already detached */
824 dlm_mle_detach_hb_events(dlm, mle);
825 dlm_put_mle(mle);
826 /* put the extra ref */
827 dlm_put_mle(mle);
828
829wake_waiters:
830 spin_lock(&res->spinlock);
831 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
832 spin_unlock(&res->spinlock);
833 wake_up(&res->wq);
834
835leave:
836 /* need to free the unused mle */
837 if (alloc_mle)
838 kmem_cache_free(dlm_mle_cache, alloc_mle);
839
840 return res;
841}
842
843
844#define DLM_MASTERY_TIMEOUT_MS 5000
845
846static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
847 struct dlm_lock_resource *res,
848 struct dlm_master_list_entry *mle,
849 int *blocked)
850{
851 u8 m;
852 int ret, bit;
853 int map_changed, voting_done;
854 int assert, sleep;
855
856recheck:
857 ret = 0;
858 assert = 0;
859
860 /* check if another node has already become the owner */
861 spin_lock(&res->spinlock);
862 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
863 spin_unlock(&res->spinlock);
864 goto leave;
865 }
866 spin_unlock(&res->spinlock);
867
868 spin_lock(&mle->spinlock);
869 m = mle->master;
870 map_changed = (memcmp(mle->vote_map, mle->node_map,
871 sizeof(mle->vote_map)) != 0);
872 voting_done = (memcmp(mle->vote_map, mle->response_map,
873 sizeof(mle->vote_map)) == 0);
874
875 /* restart if we hit any errors */
876 if (map_changed) {
877 int b;
878 mlog(0, "%s: %.*s: node map changed, restarting\n",
879 dlm->name, res->lockname.len, res->lockname.name);
880 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
881 b = (mle->type == DLM_MLE_BLOCK);
882 if ((*blocked && !b) || (!*blocked && b)) {
883 mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
884 dlm->name, res->lockname.len, res->lockname.name,
885 *blocked, b);
886 *blocked = b;
887 }
888 spin_unlock(&mle->spinlock);
889 if (ret < 0) {
890 mlog_errno(ret);
891 goto leave;
892 }
893 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
894 "rechecking now\n", dlm->name, res->lockname.len,
895 res->lockname.name);
896 goto recheck;
897 }
898
899 if (m != O2NM_MAX_NODES) {
900 /* another node has done an assert!
901 * all done! */
902 sleep = 0;
903 } else {
904 sleep = 1;
905 /* have all nodes responded? */
906 if (voting_done && !*blocked) {
907 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
908 if (dlm->node_num <= bit) {
909 /* my node number is lowest.
910 * now tell other nodes that I am
911 * mastering this. */
912 mle->master = dlm->node_num;
913 assert = 1;
914 sleep = 0;
915 }
916 /* if voting is done, but we have not received
917 * an assert master yet, we must sleep */
918 }
919 }
920
921 spin_unlock(&mle->spinlock);
922
923 /* sleep if we haven't finished voting yet */
924 if (sleep) {
925 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
926
927 /*
928 if (atomic_read(&mle->mle_refs.refcount) < 2)
929 mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
930 atomic_read(&mle->mle_refs.refcount),
931 res->lockname.len, res->lockname.name);
932 */
933 atomic_set(&mle->woken, 0);
934 (void)wait_event_timeout(mle->wq,
935 (atomic_read(&mle->woken) == 1),
936 timeo);
937 if (res->owner == O2NM_MAX_NODES) {
938 mlog(0, "waiting again\n");
939 goto recheck;
940 }
941 mlog(0, "done waiting, master is %u\n", res->owner);
942 ret = 0;
943 goto leave;
944 }
945
946 ret = 0; /* done */
947 if (assert) {
948 m = dlm->node_num;
949 mlog(0, "about to master %.*s here, this=%u\n",
950 res->lockname.len, res->lockname.name, m);
951 ret = dlm_do_assert_master(dlm, res->lockname.name,
952 res->lockname.len, mle->vote_map, 0);
953 if (ret) {
954 /* This is a failure in the network path,
955 * not in the response to the assert_master
956 * (any nonzero response is a BUG on this node).
957 * Most likely a socket just got disconnected
958 * due to node death. */
959 mlog_errno(ret);
960 }
961 /* no longer need to restart lock mastery.
962 * all living nodes have been contacted. */
963 ret = 0;
964 }
965
966 /* set the lockres owner */
967 spin_lock(&res->spinlock);
968 dlm_change_lockres_owner(dlm, res, m);
969 spin_unlock(&res->spinlock);
970
971leave:
972 return ret;
973}
974
975struct dlm_bitmap_diff_iter
976{
977 int curnode;
978 unsigned long *orig_bm;
979 unsigned long *cur_bm;
980 unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
981};
982
983enum dlm_node_state_change
984{
985 NODE_DOWN = -1,
986 NODE_NO_CHANGE = 0,
987 NODE_UP
988};
989
990static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
991 unsigned long *orig_bm,
992 unsigned long *cur_bm)
993{
994 unsigned long p1, p2;
995 int i;
996
997 iter->curnode = -1;
998 iter->orig_bm = orig_bm;
999 iter->cur_bm = cur_bm;
1000
1001 for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1002 p1 = *(iter->orig_bm + i);
1003 p2 = *(iter->cur_bm + i);
1004 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1005 }
1006}
1007
1008static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1009 enum dlm_node_state_change *state)
1010{
1011 int bit;
1012
1013 if (iter->curnode >= O2NM_MAX_NODES)
1014 return -ENOENT;
1015
1016 bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1017 iter->curnode+1);
1018 if (bit >= O2NM_MAX_NODES) {
1019 iter->curnode = O2NM_MAX_NODES;
1020 return -ENOENT;
1021 }
1022
1023 /* if it was there in the original then this node died */
1024 if (test_bit(bit, iter->orig_bm))
1025 *state = NODE_DOWN;
1026 else
1027 *state = NODE_UP;
1028
1029 iter->curnode = bit;
1030 return bit;
1031}
1032
1033
1034static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1035 struct dlm_lock_resource *res,
1036 struct dlm_master_list_entry *mle,
1037 int blocked)
1038{
1039 struct dlm_bitmap_diff_iter bdi;
1040 enum dlm_node_state_change sc;
1041 int node;
1042 int ret = 0;
1043
1044 mlog(0, "something happened such that the "
1045 "master process may need to be restarted!\n");
1046
1047 assert_spin_locked(&mle->spinlock);
1048
1049 dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1050 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1051 while (node >= 0) {
1052 if (sc == NODE_UP) {
1053 /* a node came up. easy. might not even need
1054 * to talk to it if its node number is higher
1055 * or if we are already blocked. */
1056 mlog(0, "node up! %d\n", node);
1057 if (blocked)
1058 goto next;
1059
1060 if (node > dlm->node_num) {
1061 mlog(0, "node > this node. skipping.\n");
1062 goto next;
1063 }
1064
1065 /* redo the master request, but only for the new node */
1066 mlog(0, "sending request to new node\n");
1067 clear_bit(node, mle->response_map);
1068 set_bit(node, mle->vote_map);
1069 } else {
1070 mlog(ML_ERROR, "node down! %d\n", node);
1071
1072 /* if the node wasn't involved in mastery skip it,
1073 * but clear it out from the maps so that it will
1074 * not affect mastery of this lockres */
1075 clear_bit(node, mle->response_map);
1076 clear_bit(node, mle->vote_map);
1077 if (!test_bit(node, mle->maybe_map))
1078 goto next;
1079
1080 /* if we're already blocked on lock mastery, and the
1081 * dead node wasn't the expected master, or there is
1082 * another node in the maybe_map, keep waiting */
1083 if (blocked) {
1084 int lowest = find_next_bit(mle->maybe_map,
1085 O2NM_MAX_NODES, 0);
1086
1087 /* act like it was never there */
1088 clear_bit(node, mle->maybe_map);
1089
1090 if (node != lowest)
1091 goto next;
1092
1093 mlog(ML_ERROR, "expected master %u died while "
1094 "this node was blocked waiting on it!\n",
1095 node);
1096 lowest = find_next_bit(mle->maybe_map,
1097 O2NM_MAX_NODES,
1098 lowest+1);
1099 if (lowest < O2NM_MAX_NODES) {
1100 mlog(0, "still blocked. waiting "
1101 "on %u now\n", lowest);
1102 goto next;
1103 }
1104
1105 /* mle is an MLE_BLOCK, but there is now
1106 * nothing left to block on. we need to return
1107 * all the way back out and try again with
1108 * an MLE_MASTER. dlm_do_local_recovery_cleanup
1109 * has already run, so the mle refcount is ok */
1110 mlog(0, "no longer blocking. we can "
1111 "try to master this here\n");
1112 mle->type = DLM_MLE_MASTER;
1113 memset(mle->maybe_map, 0,
1114 sizeof(mle->maybe_map));
1115 memset(mle->response_map, 0,
1116 sizeof(mle->maybe_map));
1117 memcpy(mle->vote_map, mle->node_map,
1118 sizeof(mle->node_map));
1119 mle->u.res = res;
1120 set_bit(dlm->node_num, mle->maybe_map);
1121
1122 ret = -EAGAIN;
1123 goto next;
1124 }
1125
1126 clear_bit(node, mle->maybe_map);
1127 if (node > dlm->node_num)
1128 goto next;
1129
1130 mlog(0, "dead node in map!\n");
1131 /* yuck. go back and re-contact all nodes
1132 * in the vote_map, removing this node. */
1133 memset(mle->response_map, 0,
1134 sizeof(mle->response_map));
1135 }
1136 ret = -EAGAIN;
1137next:
1138 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1139 }
1140 return ret;
1141}
1142
1143
1144/*
1145 * DLM_MASTER_REQUEST_MSG
1146 *
1147 * returns: 0 on success,
1148 * -errno on a network error
1149 *
1150 * on error, the caller should assume the target node is "dead"
1151 *
1152 */
1153
1154static int dlm_do_master_request(struct dlm_master_list_entry *mle, int to)
1155{
1156 struct dlm_ctxt *dlm = mle->dlm;
1157 struct dlm_master_request request;
1158 int ret, response=0, resend;
1159
1160 memset(&request, 0, sizeof(request));
1161 request.node_idx = dlm->node_num;
1162
1163 BUG_ON(mle->type == DLM_MLE_MIGRATION);
1164
1165 if (mle->type != DLM_MLE_MASTER) {
1166 request.namelen = mle->u.name.len;
1167 memcpy(request.name, mle->u.name.name, request.namelen);
1168 } else {
1169 request.namelen = mle->u.res->lockname.len;
1170 memcpy(request.name, mle->u.res->lockname.name,
1171 request.namelen);
1172 }
1173
1174again:
1175 ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1176 sizeof(request), to, &response);
1177 if (ret < 0) {
1178 if (ret == -ESRCH) {
1179 /* should never happen */
1180 mlog(ML_ERROR, "TCP stack not ready!\n");
1181 BUG();
1182 } else if (ret == -EINVAL) {
1183 mlog(ML_ERROR, "bad args passed to o2net!\n");
1184 BUG();
1185 } else if (ret == -ENOMEM) {
1186 mlog(ML_ERROR, "out of memory while trying to send "
1187 "network message! retrying\n");
1188 /* this is totally crude */
1189 msleep(50);
1190 goto again;
1191 } else if (!dlm_is_host_down(ret)) {
1192 /* not a network error. bad. */
1193 mlog_errno(ret);
1194 mlog(ML_ERROR, "unhandled error!");
1195 BUG();
1196 }
1197 /* all other errors should be network errors,
1198 * and likely indicate node death */
1199 mlog(ML_ERROR, "link to %d went down!\n", to);
1200 goto out;
1201 }
1202
1203 ret = 0;
1204 resend = 0;
1205 spin_lock(&mle->spinlock);
1206 switch (response) {
1207 case DLM_MASTER_RESP_YES:
1208 set_bit(to, mle->response_map);
1209 mlog(0, "node %u is the master, response=YES\n", to);
1210 mle->master = to;
1211 break;
1212 case DLM_MASTER_RESP_NO:
1213 mlog(0, "node %u not master, response=NO\n", to);
1214 set_bit(to, mle->response_map);
1215 break;
1216 case DLM_MASTER_RESP_MAYBE:
1217 mlog(0, "node %u not master, response=MAYBE\n", to);
1218 set_bit(to, mle->response_map);
1219 set_bit(to, mle->maybe_map);
1220 break;
1221 case DLM_MASTER_RESP_ERROR:
1222 mlog(0, "node %u hit an error, resending\n", to);
1223 resend = 1;
1224 response = 0;
1225 break;
1226 default:
1227 mlog(ML_ERROR, "bad response! %u\n", response);
1228 BUG();
1229 }
1230 spin_unlock(&mle->spinlock);
1231 if (resend) {
1232 /* this is also totally crude */
1233 msleep(50);
1234 goto again;
1235 }
1236
1237out:
1238 return ret;
1239}
1240
1241/*
1242 * locks that can be taken here:
1243 * dlm->spinlock
1244 * res->spinlock
1245 * mle->spinlock
1246 * dlm->master_list
1247 *
1248 * if possible, TRIM THIS DOWN!!!
1249 */
1250int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data)
1251{
1252 u8 response = DLM_MASTER_RESP_MAYBE;
1253 struct dlm_ctxt *dlm = data;
1254 struct dlm_lock_resource *res;
1255 struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1256 struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1257 char *name;
1258 unsigned int namelen;
1259 int found, ret;
1260 int set_maybe;
1261
1262 if (!dlm_grab(dlm))
1263 return DLM_MASTER_RESP_NO;
1264
1265 if (!dlm_domain_fully_joined(dlm)) {
1266 response = DLM_MASTER_RESP_NO;
1267 goto send_response;
1268 }
1269
1270 name = request->name;
1271 namelen = request->namelen;
1272
1273 if (namelen > DLM_LOCKID_NAME_MAX) {
1274 response = DLM_IVBUFLEN;
1275 goto send_response;
1276 }
1277
1278way_up_top:
1279 spin_lock(&dlm->spinlock);
1280 res = __dlm_lookup_lockres(dlm, name, namelen);
1281 if (res) {
1282 spin_unlock(&dlm->spinlock);
1283
1284 /* take care of the easy cases up front */
1285 spin_lock(&res->spinlock);
1286 if (res->state & DLM_LOCK_RES_RECOVERING) {
1287 spin_unlock(&res->spinlock);
1288 mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1289 "being recovered\n");
1290 response = DLM_MASTER_RESP_ERROR;
1291 if (mle)
1292 kmem_cache_free(dlm_mle_cache, mle);
1293 goto send_response;
1294 }
1295
1296 if (res->owner == dlm->node_num) {
1297 u32 flags = DLM_ASSERT_MASTER_MLE_CLEANUP;
1298 spin_unlock(&res->spinlock);
1299 // mlog(0, "this node is the master\n");
1300 response = DLM_MASTER_RESP_YES;
1301 if (mle)
1302 kmem_cache_free(dlm_mle_cache, mle);
1303
1304 /* this node is the owner.
1305 * there is some extra work that needs to
1306 * happen now. the requesting node has
1307 * caused all nodes up to this one to
1308 * create mles. this node now needs to
1309 * go back and clean those up. */
1310 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1311 dlm->node_num, res->lockname.len, res->lockname.name);
1312 ret = dlm_dispatch_assert_master(dlm, res, 1,
1313 request->node_idx,
1314 flags);
1315 if (ret < 0) {
1316 mlog(ML_ERROR, "failed to dispatch assert "
1317 "master work\n");
1318 response = DLM_MASTER_RESP_ERROR;
1319 }
1320 goto send_response;
1321 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1322 spin_unlock(&res->spinlock);
1323 // mlog(0, "node %u is the master\n", res->owner);
1324 response = DLM_MASTER_RESP_NO;
1325 if (mle)
1326 kmem_cache_free(dlm_mle_cache, mle);
1327 goto send_response;
1328 }
1329
1330 /* ok, there is no owner. either this node is
1331 * being blocked, or it is actively trying to
1332 * master this lock. */
1333 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1334 mlog(ML_ERROR, "lock with no owner should be "
1335 "in-progress!\n");
1336 BUG();
1337 }
1338
1339 // mlog(0, "lockres is in progress...\n");
1340 spin_lock(&dlm->master_lock);
1341 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1342 if (!found) {
1343 mlog(ML_ERROR, "no mle found for this lock!\n");
1344 BUG();
1345 }
1346 set_maybe = 1;
1347 spin_lock(&tmpmle->spinlock);
1348 if (tmpmle->type == DLM_MLE_BLOCK) {
1349 // mlog(0, "this node is waiting for "
1350 // "lockres to be mastered\n");
1351 response = DLM_MASTER_RESP_NO;
1352 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1353 mlog(0, "node %u is master, but trying to migrate to "
1354 "node %u.\n", tmpmle->master, tmpmle->new_master);
1355 if (tmpmle->master == dlm->node_num) {
1356 response = DLM_MASTER_RESP_YES;
1357 mlog(ML_ERROR, "no owner on lockres, but this "
1358 "node is trying to migrate it to %u?!\n",
1359 tmpmle->new_master);
1360 BUG();
1361 } else {
1362 /* the real master can respond on its own */
1363 response = DLM_MASTER_RESP_NO;
1364 }
1365 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1366 set_maybe = 0;
1367 if (tmpmle->master == dlm->node_num)
1368 response = DLM_MASTER_RESP_YES;
1369 else
1370 response = DLM_MASTER_RESP_NO;
1371 } else {
1372 // mlog(0, "this node is attempting to "
1373 // "master lockres\n");
1374 response = DLM_MASTER_RESP_MAYBE;
1375 }
1376 if (set_maybe)
1377 set_bit(request->node_idx, tmpmle->maybe_map);
1378 spin_unlock(&tmpmle->spinlock);
1379
1380 spin_unlock(&dlm->master_lock);
1381 spin_unlock(&res->spinlock);
1382
1383 /* keep the mle attached to heartbeat events */
1384 dlm_put_mle(tmpmle);
1385 if (mle)
1386 kmem_cache_free(dlm_mle_cache, mle);
1387 goto send_response;
1388 }
1389
1390 /*
1391 * lockres doesn't exist on this node
1392 * if there is an MLE_BLOCK, return NO
1393 * if there is an MLE_MASTER, return MAYBE
1394 * otherwise, add an MLE_BLOCK, return NO
1395 */
1396 spin_lock(&dlm->master_lock);
1397 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1398 if (!found) {
1399 /* this lockid has never been seen on this node yet */
1400 // mlog(0, "no mle found\n");
1401 if (!mle) {
1402 spin_unlock(&dlm->master_lock);
1403 spin_unlock(&dlm->spinlock);
1404
1405 mle = (struct dlm_master_list_entry *)
1406 kmem_cache_alloc(dlm_mle_cache, GFP_KERNEL);
1407 if (!mle) {
1408 // bad bad bad... this sucks.
1409 response = DLM_MASTER_RESP_ERROR;
1410 goto send_response;
1411 }
1412 spin_lock(&dlm->spinlock);
1413 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL,
1414 name, namelen);
1415 spin_unlock(&dlm->spinlock);
1416 goto way_up_top;
1417 }
1418
1419 // mlog(0, "this is second time thru, already allocated, "
1420 // "add the block.\n");
1421 set_bit(request->node_idx, mle->maybe_map);
1422 list_add(&mle->list, &dlm->master_list);
1423 response = DLM_MASTER_RESP_NO;
1424 } else {
1425 // mlog(0, "mle was found\n");
1426 set_maybe = 1;
1427 spin_lock(&tmpmle->spinlock);
1428 if (tmpmle->type == DLM_MLE_BLOCK)
1429 response = DLM_MASTER_RESP_NO;
1430 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1431 mlog(0, "migration mle was found (%u->%u)\n",
1432 tmpmle->master, tmpmle->new_master);
1433 if (tmpmle->master == dlm->node_num) {
1434 mlog(ML_ERROR, "no lockres, but migration mle "
1435 "says that this node is master!\n");
1436 BUG();
1437 }
1438 /* real master can respond on its own */
1439 response = DLM_MASTER_RESP_NO;
1440 } else {
1441 if (tmpmle->master == dlm->node_num) {
1442 response = DLM_MASTER_RESP_YES;
1443 set_maybe = 0;
1444 } else
1445 response = DLM_MASTER_RESP_MAYBE;
1446 }
1447 if (set_maybe)
1448 set_bit(request->node_idx, tmpmle->maybe_map);
1449 spin_unlock(&tmpmle->spinlock);
1450 }
1451 spin_unlock(&dlm->master_lock);
1452 spin_unlock(&dlm->spinlock);
1453
1454 if (found) {
1455 /* keep the mle attached to heartbeat events */
1456 dlm_put_mle(tmpmle);
1457 }
1458send_response:
1459 dlm_put(dlm);
1460 return response;
1461}
1462
1463/*
1464 * DLM_ASSERT_MASTER_MSG
1465 */
1466
1467
1468/*
1469 * NOTE: this can be used for debugging
1470 * can periodically run all locks owned by this node
1471 * and re-assert across the cluster...
1472 */
1473static int dlm_do_assert_master(struct dlm_ctxt *dlm, const char *lockname,
1474 unsigned int namelen, void *nodemap,
1475 u32 flags)
1476{
1477 struct dlm_assert_master assert;
1478 int to, tmpret;
1479 struct dlm_node_iter iter;
1480 int ret = 0;
1481
1482 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1483
1484 /* note that if this nodemap is empty, it returns 0 */
1485 dlm_node_iter_init(nodemap, &iter);
1486 while ((to = dlm_node_iter_next(&iter)) >= 0) {
1487 int r = 0;
1488 mlog(0, "sending assert master to %d (%.*s)\n", to,
1489 namelen, lockname);
1490 memset(&assert, 0, sizeof(assert));
1491 assert.node_idx = dlm->node_num;
1492 assert.namelen = namelen;
1493 memcpy(assert.name, lockname, namelen);
1494 assert.flags = cpu_to_be32(flags);
1495
1496 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1497 &assert, sizeof(assert), to, &r);
1498 if (tmpret < 0) {
1499 mlog(ML_ERROR, "assert_master returned %d!\n", tmpret);
1500 if (!dlm_is_host_down(tmpret)) {
1501 mlog(ML_ERROR, "unhandled error!\n");
1502 BUG();
1503 }
1504 /* a node died. finish out the rest of the nodes. */
1505 mlog(ML_ERROR, "link to %d went down!\n", to);
1506 /* any nonzero status return will do */
1507 ret = tmpret;
1508 } else if (r < 0) {
1509 /* ok, something horribly messed. kill thyself. */
1510 mlog(ML_ERROR,"during assert master of %.*s to %u, "
1511 "got %d.\n", namelen, lockname, to, r);
1512 dlm_dump_lock_resources(dlm);
1513 BUG();
1514 }
1515 }
1516
1517 return ret;
1518}
1519
1520/*
1521 * locks that can be taken here:
1522 * dlm->spinlock
1523 * res->spinlock
1524 * mle->spinlock
1525 * dlm->master_list
1526 *
1527 * if possible, TRIM THIS DOWN!!!
1528 */
1529int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data)
1530{
1531 struct dlm_ctxt *dlm = data;
1532 struct dlm_master_list_entry *mle = NULL;
1533 struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1534 struct dlm_lock_resource *res = NULL;
1535 char *name;
1536 unsigned int namelen;
1537 u32 flags;
1538
1539 if (!dlm_grab(dlm))
1540 return 0;
1541
1542 name = assert->name;
1543 namelen = assert->namelen;
1544 flags = be32_to_cpu(assert->flags);
1545
1546 if (namelen > DLM_LOCKID_NAME_MAX) {
1547 mlog(ML_ERROR, "Invalid name length!");
1548 goto done;
1549 }
1550
1551 spin_lock(&dlm->spinlock);
1552
1553 if (flags)
1554 mlog(0, "assert_master with flags: %u\n", flags);
1555
1556 /* find the MLE */
1557 spin_lock(&dlm->master_lock);
1558 if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1559 /* not an error, could be master just re-asserting */
1560 mlog(0, "just got an assert_master from %u, but no "
1561 "MLE for it! (%.*s)\n", assert->node_idx,
1562 namelen, name);
1563 } else {
1564 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1565 if (bit >= O2NM_MAX_NODES) {
1566 /* not necessarily an error, though less likely.
1567 * could be master just re-asserting. */
1568 mlog(ML_ERROR, "no bits set in the maybe_map, but %u "
1569 "is asserting! (%.*s)\n", assert->node_idx,
1570 namelen, name);
1571 } else if (bit != assert->node_idx) {
1572 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1573 mlog(0, "master %u was found, %u should "
1574 "back off\n", assert->node_idx, bit);
1575 } else {
1576 /* with the fix for bug 569, a higher node
1577 * number winning the mastery will respond
1578 * YES to mastery requests, but this node
1579 * had no way of knowing. let it pass. */
1580 mlog(ML_ERROR, "%u is the lowest node, "
1581 "%u is asserting. (%.*s) %u must "
1582 "have begun after %u won.\n", bit,
1583 assert->node_idx, namelen, name, bit,
1584 assert->node_idx);
1585 }
1586 }
1587 }
1588 spin_unlock(&dlm->master_lock);
1589
1590 /* ok everything checks out with the MLE
1591 * now check to see if there is a lockres */
1592 res = __dlm_lookup_lockres(dlm, name, namelen);
1593 if (res) {
1594 spin_lock(&res->spinlock);
1595 if (res->state & DLM_LOCK_RES_RECOVERING) {
1596 mlog(ML_ERROR, "%u asserting but %.*s is "
1597 "RECOVERING!\n", assert->node_idx, namelen, name);
1598 goto kill;
1599 }
1600 if (!mle) {
1601 if (res->owner != assert->node_idx) {
1602 mlog(ML_ERROR, "assert_master from "
1603 "%u, but current owner is "
1604 "%u! (%.*s)\n",
1605 assert->node_idx, res->owner,
1606 namelen, name);
1607 goto kill;
1608 }
1609 } else if (mle->type != DLM_MLE_MIGRATION) {
1610 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1611 /* owner is just re-asserting */
1612 if (res->owner == assert->node_idx) {
1613 mlog(0, "owner %u re-asserting on "
1614 "lock %.*s\n", assert->node_idx,
1615 namelen, name);
1616 goto ok;
1617 }
1618 mlog(ML_ERROR, "got assert_master from "
1619 "node %u, but %u is the owner! "
1620 "(%.*s)\n", assert->node_idx,
1621 res->owner, namelen, name);
1622 goto kill;
1623 }
1624 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1625 mlog(ML_ERROR, "got assert from %u, but lock "
1626 "with no owner should be "
1627 "in-progress! (%.*s)\n",
1628 assert->node_idx,
1629 namelen, name);
1630 goto kill;
1631 }
1632 } else /* mle->type == DLM_MLE_MIGRATION */ {
1633 /* should only be getting an assert from new master */
1634 if (assert->node_idx != mle->new_master) {
1635 mlog(ML_ERROR, "got assert from %u, but "
1636 "new master is %u, and old master "
1637 "was %u (%.*s)\n",
1638 assert->node_idx, mle->new_master,
1639 mle->master, namelen, name);
1640 goto kill;
1641 }
1642
1643 }
1644ok:
1645 spin_unlock(&res->spinlock);
1646 }
1647 spin_unlock(&dlm->spinlock);
1648
1649 // mlog(0, "woo! got an assert_master from node %u!\n",
1650 // assert->node_idx);
1651 if (mle) {
1652 int extra_ref;
1653
1654 spin_lock(&mle->spinlock);
1655 extra_ref = !!(mle->type == DLM_MLE_BLOCK
1656 || mle->type == DLM_MLE_MIGRATION);
1657 mle->master = assert->node_idx;
1658 atomic_set(&mle->woken, 1);
1659 wake_up(&mle->wq);
1660 spin_unlock(&mle->spinlock);
1661
1662 if (mle->type == DLM_MLE_MIGRATION && res) {
1663 mlog(0, "finishing off migration of lockres %.*s, "
1664 "from %u to %u\n",
1665 res->lockname.len, res->lockname.name,
1666 dlm->node_num, mle->new_master);
1667 spin_lock(&res->spinlock);
1668 res->state &= ~DLM_LOCK_RES_MIGRATING;
1669 dlm_change_lockres_owner(dlm, res, mle->new_master);
1670 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1671 spin_unlock(&res->spinlock);
1672 }
1673 /* master is known, detach if not already detached */
1674 dlm_mle_detach_hb_events(dlm, mle);
1675 dlm_put_mle(mle);
1676
1677 if (extra_ref) {
1678 /* the assert master message now balances the extra
1679 * ref given by the master / migration request message.
1680 * if this is the last put, it will be removed
1681 * from the list. */
1682 dlm_put_mle(mle);
1683 }
1684 }
1685
1686done:
1687 if (res)
1688 dlm_lockres_put(res);
1689 dlm_put(dlm);
1690 return 0;
1691
1692kill:
1693 /* kill the caller! */
1694 spin_unlock(&res->spinlock);
1695 spin_unlock(&dlm->spinlock);
1696 dlm_lockres_put(res);
1697 mlog(ML_ERROR, "Bad message received from another node. Dumping state "
1698 "and killing the other node now! This node is OK and can continue.\n");
1699 dlm_dump_lock_resources(dlm);
1700 dlm_put(dlm);
1701 return -EINVAL;
1702}
1703
1704int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
1705 struct dlm_lock_resource *res,
1706 int ignore_higher, u8 request_from, u32 flags)
1707{
1708 struct dlm_work_item *item;
1709 item = kcalloc(1, sizeof(*item), GFP_KERNEL);
1710 if (!item)
1711 return -ENOMEM;
1712
1713
1714 /* queue up work for dlm_assert_master_worker */
1715 dlm_grab(dlm); /* get an extra ref for the work item */
1716 dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
1717 item->u.am.lockres = res; /* already have a ref */
1718 /* can optionally ignore node numbers higher than this node */
1719 item->u.am.ignore_higher = ignore_higher;
1720 item->u.am.request_from = request_from;
1721 item->u.am.flags = flags;
1722
1723 spin_lock(&dlm->work_lock);
1724 list_add_tail(&item->list, &dlm->work_list);
1725 spin_unlock(&dlm->work_lock);
1726
1727 schedule_work(&dlm->dispatched_work);
1728 return 0;
1729}
1730
1731static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
1732{
1733 struct dlm_ctxt *dlm = data;
1734 int ret = 0;
1735 struct dlm_lock_resource *res;
1736 unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
1737 int ignore_higher;
1738 int bit;
1739 u8 request_from;
1740 u32 flags;
1741
1742 dlm = item->dlm;
1743 res = item->u.am.lockres;
1744 ignore_higher = item->u.am.ignore_higher;
1745 request_from = item->u.am.request_from;
1746 flags = item->u.am.flags;
1747
1748 spin_lock(&dlm->spinlock);
1749 memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
1750 spin_unlock(&dlm->spinlock);
1751
1752 clear_bit(dlm->node_num, nodemap);
1753 if (ignore_higher) {
1754 /* if is this just to clear up mles for nodes below
1755 * this node, do not send the message to the original
1756 * caller or any node number higher than this */
1757 clear_bit(request_from, nodemap);
1758 bit = dlm->node_num;
1759 while (1) {
1760 bit = find_next_bit(nodemap, O2NM_MAX_NODES,
1761 bit+1);
1762 if (bit >= O2NM_MAX_NODES)
1763 break;
1764 clear_bit(bit, nodemap);
1765 }
1766 }
1767
1768 /* this call now finishes out the nodemap
1769 * even if one or more nodes die */
1770 mlog(0, "worker about to master %.*s here, this=%u\n",
1771 res->lockname.len, res->lockname.name, dlm->node_num);
1772 ret = dlm_do_assert_master(dlm, res->lockname.name,
1773 res->lockname.len,
1774 nodemap, flags);
1775 if (ret < 0) {
1776 /* no need to restart, we are done */
1777 mlog_errno(ret);
1778 }
1779
1780 dlm_lockres_put(res);
1781
1782 mlog(0, "finished with dlm_assert_master_worker\n");
1783}
1784
1785
1786/*
1787 * DLM_MIGRATE_LOCKRES
1788 */
1789
1790
1791int dlm_migrate_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1792 u8 target)
1793{
1794 struct dlm_master_list_entry *mle = NULL;
1795 struct dlm_master_list_entry *oldmle = NULL;
1796 struct dlm_migratable_lockres *mres = NULL;
1797 int ret = -EINVAL;
1798 const char *name;
1799 unsigned int namelen;
1800 int mle_added = 0;
1801 struct list_head *queue, *iter;
1802 int i;
1803 struct dlm_lock *lock;
1804 int empty = 1;
1805
1806 if (!dlm_grab(dlm))
1807 return -EINVAL;
1808
1809 name = res->lockname.name;
1810 namelen = res->lockname.len;
1811
1812 mlog(0, "migrating %.*s to %u\n", namelen, name, target);
1813
1814 /*
1815 * ensure this lockres is a proper candidate for migration
1816 */
1817 spin_lock(&res->spinlock);
1818 if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1819 mlog(0, "cannot migrate lockres with unknown owner!\n");
1820 spin_unlock(&res->spinlock);
1821 goto leave;
1822 }
1823 if (res->owner != dlm->node_num) {
1824 mlog(0, "cannot migrate lockres this node doesn't own!\n");
1825 spin_unlock(&res->spinlock);
1826 goto leave;
1827 }
1828 mlog(0, "checking queues...\n");
1829 queue = &res->granted;
1830 for (i=0; i<3; i++) {
1831 list_for_each(iter, queue) {
1832 lock = list_entry (iter, struct dlm_lock, list);
1833 empty = 0;
1834 if (lock->ml.node == dlm->node_num) {
1835 mlog(0, "found a lock owned by this node "
1836 "still on the %s queue! will not "
1837 "migrate this lockres\n",
1838 i==0 ? "granted" :
1839 (i==1 ? "converting" : "blocked"));
1840 spin_unlock(&res->spinlock);
1841 ret = -ENOTEMPTY;
1842 goto leave;
1843 }
1844 }
1845 queue++;
1846 }
1847 mlog(0, "all locks on this lockres are nonlocal. continuing\n");
1848 spin_unlock(&res->spinlock);
1849
1850 /* no work to do */
1851 if (empty) {
1852 mlog(0, "no locks were found on this lockres! done!\n");
1853 ret = 0;
1854 goto leave;
1855 }
1856
1857 /*
1858 * preallocate up front
1859 * if this fails, abort
1860 */
1861
1862 ret = -ENOMEM;
1863 mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_KERNEL);
1864 if (!mres) {
1865 mlog_errno(ret);
1866 goto leave;
1867 }
1868
1869 mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
1870 GFP_KERNEL);
1871 if (!mle) {
1872 mlog_errno(ret);
1873 goto leave;
1874 }
1875 ret = 0;
1876
1877 /*
1878 * find a node to migrate the lockres to
1879 */
1880
1881 mlog(0, "picking a migration node\n");
1882 spin_lock(&dlm->spinlock);
1883 /* pick a new node */
1884 if (!test_bit(target, dlm->domain_map) ||
1885 target >= O2NM_MAX_NODES) {
1886 target = dlm_pick_migration_target(dlm, res);
1887 }
1888 mlog(0, "node %u chosen for migration\n", target);
1889
1890 if (target >= O2NM_MAX_NODES ||
1891 !test_bit(target, dlm->domain_map)) {
1892 /* target chosen is not alive */
1893 ret = -EINVAL;
1894 }
1895
1896 if (ret) {
1897 spin_unlock(&dlm->spinlock);
1898 goto fail;
1899 }
1900
1901 mlog(0, "continuing with target = %u\n", target);
1902
1903 /*
1904 * clear any existing master requests and
1905 * add the migration mle to the list
1906 */
1907 spin_lock(&dlm->master_lock);
1908 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
1909 namelen, target, dlm->node_num);
1910 spin_unlock(&dlm->master_lock);
1911 spin_unlock(&dlm->spinlock);
1912
1913 if (ret == -EEXIST) {
1914 mlog(0, "another process is already migrating it\n");
1915 goto fail;
1916 }
1917 mle_added = 1;
1918
1919 /*
1920 * set the MIGRATING flag and flush asts
1921 * if we fail after this we need to re-dirty the lockres
1922 */
1923 if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
1924 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
1925 "the target went down.\n", res->lockname.len,
1926 res->lockname.name, target);
1927 spin_lock(&res->spinlock);
1928 res->state &= ~DLM_LOCK_RES_MIGRATING;
1929 spin_unlock(&res->spinlock);
1930 ret = -EINVAL;
1931 }
1932
1933fail:
1934 if (oldmle) {
1935 /* master is known, detach if not already detached */
1936 dlm_mle_detach_hb_events(dlm, oldmle);
1937 dlm_put_mle(oldmle);
1938 }
1939
1940 if (ret < 0) {
1941 if (mle_added) {
1942 dlm_mle_detach_hb_events(dlm, mle);
1943 dlm_put_mle(mle);
1944 } else if (mle) {
1945 kmem_cache_free(dlm_mle_cache, mle);
1946 }
1947 goto leave;
1948 }
1949
1950 /*
1951 * at this point, we have a migration target, an mle
1952 * in the master list, and the MIGRATING flag set on
1953 * the lockres
1954 */
1955
1956
1957 /* get an extra reference on the mle.
1958 * otherwise the assert_master from the new
1959 * master will destroy this.
1960 * also, make sure that all callers of dlm_get_mle
1961 * take both dlm->spinlock and dlm->master_lock */
1962 spin_lock(&dlm->spinlock);
1963 spin_lock(&dlm->master_lock);
1964 dlm_get_mle(mle);
1965 spin_unlock(&dlm->master_lock);
1966 spin_unlock(&dlm->spinlock);
1967
1968 /* notify new node and send all lock state */
1969 /* call send_one_lockres with migration flag.
1970 * this serves as notice to the target node that a
1971 * migration is starting. */
1972 ret = dlm_send_one_lockres(dlm, res, mres, target,
1973 DLM_MRES_MIGRATION);
1974
1975 if (ret < 0) {
1976 mlog(0, "migration to node %u failed with %d\n",
1977 target, ret);
1978 /* migration failed, detach and clean up mle */
1979 dlm_mle_detach_hb_events(dlm, mle);
1980 dlm_put_mle(mle);
1981 dlm_put_mle(mle);
1982 goto leave;
1983 }
1984
1985 /* at this point, the target sends a message to all nodes,
1986 * (using dlm_do_migrate_request). this node is skipped since
1987 * we had to put an mle in the list to begin the process. this
1988 * node now waits for target to do an assert master. this node
1989 * will be the last one notified, ensuring that the migration
1990 * is complete everywhere. if the target dies while this is
1991 * going on, some nodes could potentially see the target as the
1992 * master, so it is important that my recovery finds the migration
1993 * mle and sets the master to UNKNONWN. */
1994
1995
1996 /* wait for new node to assert master */
1997 while (1) {
1998 ret = wait_event_interruptible_timeout(mle->wq,
1999 (atomic_read(&mle->woken) == 1),
2000 msecs_to_jiffies(5000));
2001
2002 if (ret >= 0) {
2003 if (atomic_read(&mle->woken) == 1 ||
2004 res->owner == target)
2005 break;
2006
2007 mlog(0, "timed out during migration\n");
2008 }
2009 if (ret == -ERESTARTSYS) {
2010 /* migration failed, detach and clean up mle */
2011 dlm_mle_detach_hb_events(dlm, mle);
2012 dlm_put_mle(mle);
2013 dlm_put_mle(mle);
2014 goto leave;
2015 }
2016 /* TODO: if node died: stop, clean up, return error */
2017 }
2018
2019 /* all done, set the owner, clear the flag */
2020 spin_lock(&res->spinlock);
2021 dlm_set_lockres_owner(dlm, res, target);
2022 res->state &= ~DLM_LOCK_RES_MIGRATING;
2023 dlm_remove_nonlocal_locks(dlm, res);
2024 spin_unlock(&res->spinlock);
2025 wake_up(&res->wq);
2026
2027 /* master is known, detach if not already detached */
2028 dlm_mle_detach_hb_events(dlm, mle);
2029 dlm_put_mle(mle);
2030 ret = 0;
2031
2032 dlm_lockres_calc_usage(dlm, res);
2033
2034leave:
2035 /* re-dirty the lockres if we failed */
2036 if (ret < 0)
2037 dlm_kick_thread(dlm, res);
2038
2039 /* TODO: cleanup */
2040 if (mres)
2041 free_page((unsigned long)mres);
2042
2043 dlm_put(dlm);
2044
2045 mlog(0, "returning %d\n", ret);
2046 return ret;
2047}
2048EXPORT_SYMBOL_GPL(dlm_migrate_lockres);
2049
2050int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2051{
2052 int ret;
2053 spin_lock(&dlm->ast_lock);
2054 spin_lock(&lock->spinlock);
2055 ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2056 spin_unlock(&lock->spinlock);
2057 spin_unlock(&dlm->ast_lock);
2058 return ret;
2059}
2060
2061static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2062 struct dlm_lock_resource *res,
2063 u8 mig_target)
2064{
2065 int can_proceed;
2066 spin_lock(&res->spinlock);
2067 can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2068 spin_unlock(&res->spinlock);
2069
2070 /* target has died, so make the caller break out of the
2071 * wait_event, but caller must recheck the domain_map */
2072 spin_lock(&dlm->spinlock);
2073 if (!test_bit(mig_target, dlm->domain_map))
2074 can_proceed = 1;
2075 spin_unlock(&dlm->spinlock);
2076 return can_proceed;
2077}
2078
2079int dlm_lockres_is_dirty(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2080{
2081 int ret;
2082 spin_lock(&res->spinlock);
2083 ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2084 spin_unlock(&res->spinlock);
2085 return ret;
2086}
2087
2088
2089static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2090 struct dlm_lock_resource *res,
2091 u8 target)
2092{
2093 int ret = 0;
2094
2095 mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2096 res->lockname.len, res->lockname.name, dlm->node_num,
2097 target);
2098 /* need to set MIGRATING flag on lockres. this is done by
2099 * ensuring that all asts have been flushed for this lockres. */
2100 spin_lock(&res->spinlock);
2101 BUG_ON(res->migration_pending);
2102 res->migration_pending = 1;
2103 /* strategy is to reserve an extra ast then release
2104 * it below, letting the release do all of the work */
2105 __dlm_lockres_reserve_ast(res);
2106 spin_unlock(&res->spinlock);
2107
2108 /* now flush all the pending asts.. hang out for a bit */
2109 dlm_kick_thread(dlm, res);
2110 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2111 dlm_lockres_release_ast(dlm, res);
2112
2113 mlog(0, "about to wait on migration_wq, dirty=%s\n",
2114 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2115 /* if the extra ref we just put was the final one, this
2116 * will pass thru immediately. otherwise, we need to wait
2117 * for the last ast to finish. */
2118again:
2119 ret = wait_event_interruptible_timeout(dlm->migration_wq,
2120 dlm_migration_can_proceed(dlm, res, target),
2121 msecs_to_jiffies(1000));
2122 if (ret < 0) {
2123 mlog(0, "woken again: migrating? %s, dead? %s\n",
2124 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2125 test_bit(target, dlm->domain_map) ? "no":"yes");
2126 } else {
2127 mlog(0, "all is well: migrating? %s, dead? %s\n",
2128 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2129 test_bit(target, dlm->domain_map) ? "no":"yes");
2130 }
2131 if (!dlm_migration_can_proceed(dlm, res, target)) {
2132 mlog(0, "trying again...\n");
2133 goto again;
2134 }
2135
2136 /* did the target go down or die? */
2137 spin_lock(&dlm->spinlock);
2138 if (!test_bit(target, dlm->domain_map)) {
2139 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2140 target);
2141 ret = -EHOSTDOWN;
2142 }
2143 spin_unlock(&dlm->spinlock);
2144
2145 /*
2146 * at this point:
2147 *
2148 * o the DLM_LOCK_RES_MIGRATING flag is set
2149 * o there are no pending asts on this lockres
2150 * o all processes trying to reserve an ast on this
2151 * lockres must wait for the MIGRATING flag to clear
2152 */
2153 return ret;
2154}
2155
2156/* last step in the migration process.
2157 * original master calls this to free all of the dlm_lock
2158 * structures that used to be for other nodes. */
2159static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2160 struct dlm_lock_resource *res)
2161{
2162 struct list_head *iter, *iter2;
2163 struct list_head *queue = &res->granted;
2164 int i;
2165 struct dlm_lock *lock;
2166
2167 assert_spin_locked(&res->spinlock);
2168
2169 BUG_ON(res->owner == dlm->node_num);
2170
2171 for (i=0; i<3; i++) {
2172 list_for_each_safe(iter, iter2, queue) {
2173 lock = list_entry (iter, struct dlm_lock, list);
2174 if (lock->ml.node != dlm->node_num) {
2175 mlog(0, "putting lock for node %u\n",
2176 lock->ml.node);
2177 /* be extra careful */
2178 BUG_ON(!list_empty(&lock->ast_list));
2179 BUG_ON(!list_empty(&lock->bast_list));
2180 BUG_ON(lock->ast_pending);
2181 BUG_ON(lock->bast_pending);
2182 list_del_init(&lock->list);
2183 dlm_lock_put(lock);
2184 }
2185 }
2186 queue++;
2187 }
2188}
2189
2190/* for now this is not too intelligent. we will
2191 * need stats to make this do the right thing.
2192 * this just finds the first lock on one of the
2193 * queues and uses that node as the target. */
2194static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2195 struct dlm_lock_resource *res)
2196{
2197 int i;
2198 struct list_head *queue = &res->granted;
2199 struct list_head *iter;
2200 struct dlm_lock *lock;
2201 int nodenum;
2202
2203 assert_spin_locked(&dlm->spinlock);
2204
2205 spin_lock(&res->spinlock);
2206 for (i=0; i<3; i++) {
2207 list_for_each(iter, queue) {
2208 /* up to the caller to make sure this node
2209 * is alive */
2210 lock = list_entry (iter, struct dlm_lock, list);
2211 if (lock->ml.node != dlm->node_num) {
2212 spin_unlock(&res->spinlock);
2213 return lock->ml.node;
2214 }
2215 }
2216 queue++;
2217 }
2218 spin_unlock(&res->spinlock);
2219 mlog(0, "have not found a suitable target yet! checking domain map\n");
2220
2221 /* ok now we're getting desperate. pick anyone alive. */
2222 nodenum = -1;
2223 while (1) {
2224 nodenum = find_next_bit(dlm->domain_map,
2225 O2NM_MAX_NODES, nodenum+1);
2226 mlog(0, "found %d in domain map\n", nodenum);
2227 if (nodenum >= O2NM_MAX_NODES)
2228 break;
2229 if (nodenum != dlm->node_num) {
2230 mlog(0, "picking %d\n", nodenum);
2231 return nodenum;
2232 }
2233 }
2234
2235 mlog(0, "giving up. no master to migrate to\n");
2236 return DLM_LOCK_RES_OWNER_UNKNOWN;
2237}
2238
2239
2240
2241/* this is called by the new master once all lockres
2242 * data has been received */
2243static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2244 struct dlm_lock_resource *res,
2245 u8 master, u8 new_master,
2246 struct dlm_node_iter *iter)
2247{
2248 struct dlm_migrate_request migrate;
2249 int ret, status = 0;
2250 int nodenum;
2251
2252 memset(&migrate, 0, sizeof(migrate));
2253 migrate.namelen = res->lockname.len;
2254 memcpy(migrate.name, res->lockname.name, migrate.namelen);
2255 migrate.new_master = new_master;
2256 migrate.master = master;
2257
2258 ret = 0;
2259
2260 /* send message to all nodes, except the master and myself */
2261 while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2262 if (nodenum == master ||
2263 nodenum == new_master)
2264 continue;
2265
2266 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2267 &migrate, sizeof(migrate), nodenum,
2268 &status);
2269 if (ret < 0)
2270 mlog_errno(ret);
2271 else if (status < 0) {
2272 mlog(0, "migrate request (node %u) returned %d!\n",
2273 nodenum, status);
2274 ret = status;
2275 }
2276 }
2277
2278 if (ret < 0)
2279 mlog_errno(ret);
2280
2281 mlog(0, "returning ret=%d\n", ret);
2282 return ret;
2283}
2284
2285
2286/* if there is an existing mle for this lockres, we now know who the master is.
2287 * (the one who sent us *this* message) we can clear it up right away.
2288 * since the process that put the mle on the list still has a reference to it,
2289 * we can unhash it now, set the master and wake the process. as a result,
2290 * we will have no mle in the list to start with. now we can add an mle for
2291 * the migration and this should be the only one found for those scanning the
2292 * list. */
2293int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data)
2294{
2295 struct dlm_ctxt *dlm = data;
2296 struct dlm_lock_resource *res = NULL;
2297 struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
2298 struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
2299 const char *name;
2300 unsigned int namelen;
2301 int ret = 0;
2302
2303 if (!dlm_grab(dlm))
2304 return -EINVAL;
2305
2306 name = migrate->name;
2307 namelen = migrate->namelen;
2308
2309 /* preallocate.. if this fails, abort */
2310 mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
2311 GFP_KERNEL);
2312
2313 if (!mle) {
2314 ret = -ENOMEM;
2315 goto leave;
2316 }
2317
2318 /* check for pre-existing lock */
2319 spin_lock(&dlm->spinlock);
2320 res = __dlm_lookup_lockres(dlm, name, namelen);
2321 spin_lock(&dlm->master_lock);
2322
2323 if (res) {
2324 spin_lock(&res->spinlock);
2325 if (res->state & DLM_LOCK_RES_RECOVERING) {
2326 /* if all is working ok, this can only mean that we got
2327 * a migrate request from a node that we now see as
2328 * dead. what can we do here? drop it to the floor? */
2329 spin_unlock(&res->spinlock);
2330 mlog(ML_ERROR, "Got a migrate request, but the "
2331 "lockres is marked as recovering!");
2332 kmem_cache_free(dlm_mle_cache, mle);
2333 ret = -EINVAL; /* need a better solution */
2334 goto unlock;
2335 }
2336 res->state |= DLM_LOCK_RES_MIGRATING;
2337 spin_unlock(&res->spinlock);
2338 }
2339
2340 /* ignore status. only nonzero status would BUG. */
2341 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
2342 name, namelen,
2343 migrate->new_master,
2344 migrate->master);
2345
2346unlock:
2347 spin_unlock(&dlm->master_lock);
2348 spin_unlock(&dlm->spinlock);
2349
2350 if (oldmle) {
2351 /* master is known, detach if not already detached */
2352 dlm_mle_detach_hb_events(dlm, oldmle);
2353 dlm_put_mle(oldmle);
2354 }
2355
2356 if (res)
2357 dlm_lockres_put(res);
2358leave:
2359 dlm_put(dlm);
2360 return ret;
2361}
2362
2363/* must be holding dlm->spinlock and dlm->master_lock
2364 * when adding a migration mle, we can clear any other mles
2365 * in the master list because we know with certainty that
2366 * the master is "master". so we remove any old mle from
2367 * the list after setting it's master field, and then add
2368 * the new migration mle. this way we can hold with the rule
2369 * of having only one mle for a given lock name at all times. */
2370static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
2371 struct dlm_lock_resource *res,
2372 struct dlm_master_list_entry *mle,
2373 struct dlm_master_list_entry **oldmle,
2374 const char *name, unsigned int namelen,
2375 u8 new_master, u8 master)
2376{
2377 int found;
2378 int ret = 0;
2379
2380 *oldmle = NULL;
2381
2382 mlog_entry_void();
2383
2384 assert_spin_locked(&dlm->spinlock);
2385 assert_spin_locked(&dlm->master_lock);
2386
2387 /* caller is responsible for any ref taken here on oldmle */
2388 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
2389 if (found) {
2390 struct dlm_master_list_entry *tmp = *oldmle;
2391 spin_lock(&tmp->spinlock);
2392 if (tmp->type == DLM_MLE_MIGRATION) {
2393 if (master == dlm->node_num) {
2394 /* ah another process raced me to it */
2395 mlog(0, "tried to migrate %.*s, but some "
2396 "process beat me to it\n",
2397 namelen, name);
2398 ret = -EEXIST;
2399 } else {
2400 /* bad. 2 NODES are trying to migrate! */
2401 mlog(ML_ERROR, "migration error mle: "
2402 "master=%u new_master=%u // request: "
2403 "master=%u new_master=%u // "
2404 "lockres=%.*s\n",
2405 tmp->master, tmp->new_master,
2406 master, new_master,
2407 namelen, name);
2408 BUG();
2409 }
2410 } else {
2411 /* this is essentially what assert_master does */
2412 tmp->master = master;
2413 atomic_set(&tmp->woken, 1);
2414 wake_up(&tmp->wq);
2415 /* remove it from the list so that only one
2416 * mle will be found */
2417 list_del_init(&tmp->list);
2418 }
2419 spin_unlock(&tmp->spinlock);
2420 }
2421
2422 /* now add a migration mle to the tail of the list */
2423 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
2424 mle->new_master = new_master;
2425 mle->master = master;
2426 /* do this for consistency with other mle types */
2427 set_bit(new_master, mle->maybe_map);
2428 list_add(&mle->list, &dlm->master_list);
2429
2430 return ret;
2431}
2432
2433
2434void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
2435{
2436 struct list_head *iter, *iter2;
2437 struct dlm_master_list_entry *mle;
2438 struct dlm_lock_resource *res;
2439
2440 mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node);
2441top:
2442 assert_spin_locked(&dlm->spinlock);
2443
2444 /* clean the master list */
2445 spin_lock(&dlm->master_lock);
2446 list_for_each_safe(iter, iter2, &dlm->master_list) {
2447 mle = list_entry(iter, struct dlm_master_list_entry, list);
2448
2449 BUG_ON(mle->type != DLM_MLE_BLOCK &&
2450 mle->type != DLM_MLE_MASTER &&
2451 mle->type != DLM_MLE_MIGRATION);
2452
2453 /* MASTER mles are initiated locally. the waiting
2454 * process will notice the node map change
2455 * shortly. let that happen as normal. */
2456 if (mle->type == DLM_MLE_MASTER)
2457 continue;
2458
2459
2460 /* BLOCK mles are initiated by other nodes.
2461 * need to clean up if the dead node would have
2462 * been the master. */
2463 if (mle->type == DLM_MLE_BLOCK) {
2464 int bit;
2465
2466 spin_lock(&mle->spinlock);
2467 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
2468 if (bit != dead_node) {
2469 mlog(0, "mle found, but dead node %u would "
2470 "not have been master\n", dead_node);
2471 spin_unlock(&mle->spinlock);
2472 } else {
2473 /* must drop the refcount by one since the
2474 * assert_master will never arrive. this
2475 * may result in the mle being unlinked and
2476 * freed, but there may still be a process
2477 * waiting in the dlmlock path which is fine. */
2478 mlog(ML_ERROR, "node %u was expected master\n",
2479 dead_node);
2480 atomic_set(&mle->woken, 1);
2481 spin_unlock(&mle->spinlock);
2482 wake_up(&mle->wq);
2483 /* final put will take care of list removal */
2484 __dlm_put_mle(mle);
2485 }
2486 continue;
2487 }
2488
2489 /* everything else is a MIGRATION mle */
2490
2491 /* the rule for MIGRATION mles is that the master
2492 * becomes UNKNOWN if *either* the original or
2493 * the new master dies. all UNKNOWN lockreses
2494 * are sent to whichever node becomes the recovery
2495 * master. the new master is responsible for
2496 * determining if there is still a master for
2497 * this lockres, or if he needs to take over
2498 * mastery. either way, this node should expect
2499 * another message to resolve this. */
2500 if (mle->master != dead_node &&
2501 mle->new_master != dead_node)
2502 continue;
2503
2504 /* if we have reached this point, this mle needs to
2505 * be removed from the list and freed. */
2506
2507 /* remove from the list early. NOTE: unlinking
2508 * list_head while in list_for_each_safe */
2509 spin_lock(&mle->spinlock);
2510 list_del_init(&mle->list);
2511 atomic_set(&mle->woken, 1);
2512 spin_unlock(&mle->spinlock);
2513 wake_up(&mle->wq);
2514
2515 mlog(0, "node %u died during migration from "
2516 "%u to %u!\n", dead_node,
2517 mle->master, mle->new_master);
2518 /* if there is a lockres associated with this
2519 * mle, find it and set its owner to UNKNOWN */
2520 res = __dlm_lookup_lockres(dlm, mle->u.name.name,
2521 mle->u.name.len);
2522 if (res) {
2523 /* unfortunately if we hit this rare case, our
2524 * lock ordering is messed. we need to drop
2525 * the master lock so that we can take the
2526 * lockres lock, meaning that we will have to
2527 * restart from the head of list. */
2528 spin_unlock(&dlm->master_lock);
2529
2530 /* move lockres onto recovery list */
2531 spin_lock(&res->spinlock);
2532 dlm_set_lockres_owner(dlm, res,
2533 DLM_LOCK_RES_OWNER_UNKNOWN);
2534 dlm_move_lockres_to_recovery_list(dlm, res);
2535 spin_unlock(&res->spinlock);
2536 dlm_lockres_put(res);
2537
2538 /* dump the mle */
2539 spin_lock(&dlm->master_lock);
2540 __dlm_put_mle(mle);
2541 spin_unlock(&dlm->master_lock);
2542
2543 /* restart */
2544 goto top;
2545 }
2546
2547 /* this may be the last reference */
2548 __dlm_put_mle(mle);
2549 }
2550 spin_unlock(&dlm->master_lock);
2551}
2552
2553
2554int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
2555 u8 old_master)
2556{
2557 struct dlm_node_iter iter;
2558 int ret = 0;
2559
2560 spin_lock(&dlm->spinlock);
2561 dlm_node_iter_init(dlm->domain_map, &iter);
2562 clear_bit(old_master, iter.node_map);
2563 clear_bit(dlm->node_num, iter.node_map);
2564 spin_unlock(&dlm->spinlock);
2565
2566 mlog(0, "now time to do a migrate request to other nodes\n");
2567 ret = dlm_do_migrate_request(dlm, res, old_master,
2568 dlm->node_num, &iter);
2569 if (ret < 0) {
2570 mlog_errno(ret);
2571 goto leave;
2572 }
2573
2574 mlog(0, "doing assert master of %.*s to all except the original node\n",
2575 res->lockname.len, res->lockname.name);
2576 /* this call now finishes out the nodemap
2577 * even if one or more nodes die */
2578 ret = dlm_do_assert_master(dlm, res->lockname.name,
2579 res->lockname.len, iter.node_map,
2580 DLM_ASSERT_MASTER_FINISH_MIGRATION);
2581 if (ret < 0) {
2582 /* no longer need to retry. all living nodes contacted. */
2583 mlog_errno(ret);
2584 ret = 0;
2585 }
2586
2587 memset(iter.node_map, 0, sizeof(iter.node_map));
2588 set_bit(old_master, iter.node_map);
2589 mlog(0, "doing assert master of %.*s back to %u\n",
2590 res->lockname.len, res->lockname.name, old_master);
2591 ret = dlm_do_assert_master(dlm, res->lockname.name,
2592 res->lockname.len, iter.node_map,
2593 DLM_ASSERT_MASTER_FINISH_MIGRATION);
2594 if (ret < 0) {
2595 mlog(0, "assert master to original master failed "
2596 "with %d.\n", ret);
2597 /* the only nonzero status here would be because of
2598 * a dead original node. we're done. */
2599 ret = 0;
2600 }
2601
2602 /* all done, set the owner, clear the flag */
2603 spin_lock(&res->spinlock);
2604 dlm_set_lockres_owner(dlm, res, dlm->node_num);
2605 res->state &= ~DLM_LOCK_RES_MIGRATING;
2606 spin_unlock(&res->spinlock);
2607 /* re-dirty it on the new master */
2608 dlm_kick_thread(dlm, res);
2609 wake_up(&res->wq);
2610leave:
2611 return ret;
2612}
2613
2614/*
2615 * LOCKRES AST REFCOUNT
2616 * this is integral to migration
2617 */
2618
2619/* for future intent to call an ast, reserve one ahead of time.
2620 * this should be called only after waiting on the lockres
2621 * with dlm_wait_on_lockres, and while still holding the
2622 * spinlock after the call. */
2623void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
2624{
2625 assert_spin_locked(&res->spinlock);
2626 if (res->state & DLM_LOCK_RES_MIGRATING) {
2627 __dlm_print_one_lock_resource(res);
2628 }
2629 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
2630
2631 atomic_inc(&res->asts_reserved);
2632}
2633
2634/*
2635 * used to drop the reserved ast, either because it went unused,
2636 * or because the ast/bast was actually called.
2637 *
2638 * also, if there is a pending migration on this lockres,
2639 * and this was the last pending ast on the lockres,
2640 * atomically set the MIGRATING flag before we drop the lock.
2641 * this is how we ensure that migration can proceed with no
2642 * asts in progress. note that it is ok if the state of the
2643 * queues is such that a lock should be granted in the future
2644 * or that a bast should be fired, because the new master will
2645 * shuffle the lists on this lockres as soon as it is migrated.
2646 */
2647void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
2648 struct dlm_lock_resource *res)
2649{
2650 if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
2651 return;
2652
2653 if (!res->migration_pending) {
2654 spin_unlock(&res->spinlock);
2655 return;
2656 }
2657
2658 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
2659 res->migration_pending = 0;
2660 res->state |= DLM_LOCK_RES_MIGRATING;
2661 spin_unlock(&res->spinlock);
2662 wake_up(&res->wq);
2663 wake_up(&dlm->migration_wq);
2664}
generated by cgit v1.2.2 (git 2.25.0) at 2024-11-15 13:43:01 -0500