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