Merge git://git.kernel.org/pub/scm/linux/kernel/git/steve/gfs2-2.6-nmw
[linux-2.6.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/utsname.h>
34 #include <linux/init.h>
35 #include <linux/sysctl.h>
36 #include <linux/random.h>
37 #include <linux/blkdev.h>
38 #include <linux/socket.h>
39 #include <linux/inet.h>
40 #include <linux/spinlock.h>
41 #include <linux/delay.h>
42
43
44 #include "cluster/heartbeat.h"
45 #include "cluster/nodemanager.h"
46 #include "cluster/tcp.h"
47
48 #include "dlmapi.h"
49 #include "dlmcommon.h"
50 #include "dlmdomain.h"
51
52 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
53 #include "cluster/masklog.h"
54
55 enum dlm_mle_type {
56         DLM_MLE_BLOCK,
57         DLM_MLE_MASTER,
58         DLM_MLE_MIGRATION
59 };
60
61 struct dlm_lock_name
62 {
63         u8 len;
64         u8 name[DLM_LOCKID_NAME_MAX];
65 };
66
67 struct dlm_master_list_entry
68 {
69         struct list_head list;
70         struct list_head hb_events;
71         struct dlm_ctxt *dlm;
72         spinlock_t spinlock;
73         wait_queue_head_t wq;
74         atomic_t woken;
75         struct kref mle_refs;
76         int inuse;
77         unsigned long maybe_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
78         unsigned long vote_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
79         unsigned long response_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
80         unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
81         u8 master;
82         u8 new_master;
83         enum dlm_mle_type type;
84         struct o2hb_callback_func mle_hb_up;
85         struct o2hb_callback_func mle_hb_down;
86         union {
87                 struct dlm_lock_resource *res;
88                 struct dlm_lock_name name;
89         } u;
90 };
91
92 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
93                               struct dlm_master_list_entry *mle,
94                               struct o2nm_node *node,
95                               int idx);
96 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
97                             struct dlm_master_list_entry *mle,
98                             struct o2nm_node *node,
99                             int idx);
100
101 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
102 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
103                                 struct dlm_lock_resource *res,
104                                 void *nodemap, u32 flags);
105 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
106
107 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
108                                 struct dlm_master_list_entry *mle,
109                                 const char *name,
110                                 unsigned int namelen)
111 {
112         struct dlm_lock_resource *res;
113
114         if (dlm != mle->dlm)
115                 return 0;
116
117         if (mle->type == DLM_MLE_BLOCK ||
118             mle->type == DLM_MLE_MIGRATION) {
119                 if (namelen != mle->u.name.len ||
120                     memcmp(name, mle->u.name.name, namelen)!=0)
121                         return 0;
122         } else {
123                 res = mle->u.res;
124                 if (namelen != res->lockname.len ||
125                     memcmp(res->lockname.name, name, namelen) != 0)
126                         return 0;
127         }
128         return 1;
129 }
130
131 #define dlm_print_nodemap(m)  _dlm_print_nodemap(m,#m)
132 static void _dlm_print_nodemap(unsigned long *map, const char *mapname)
133 {
134         int i;
135         printk("%s=[ ", mapname);
136         for (i=0; i<O2NM_MAX_NODES; i++)
137                 if (test_bit(i, map))
138                         printk("%d ", i);
139         printk("]");
140 }
141
142 static void dlm_print_one_mle(struct dlm_master_list_entry *mle)
143 {
144         int refs;
145         char *type;
146         char attached;
147         u8 master;
148         unsigned int namelen;
149         const char *name;
150         struct kref *k;
151         unsigned long *maybe = mle->maybe_map,
152                       *vote = mle->vote_map,
153                       *resp = mle->response_map,
154                       *node = mle->node_map;
155
156         k = &mle->mle_refs;
157         if (mle->type == DLM_MLE_BLOCK)
158                 type = "BLK";
159         else if (mle->type == DLM_MLE_MASTER)
160                 type = "MAS";
161         else
162                 type = "MIG";
163         refs = atomic_read(&k->refcount);
164         master = mle->master;
165         attached = (list_empty(&mle->hb_events) ? 'N' : 'Y');
166
167         if (mle->type != DLM_MLE_MASTER) {
168                 namelen = mle->u.name.len;
169                 name = mle->u.name.name;
170         } else {
171                 namelen = mle->u.res->lockname.len;
172                 name = mle->u.res->lockname.name;
173         }
174
175         mlog(ML_NOTICE, "%.*s: %3s refs=%3d mas=%3u new=%3u evt=%c inuse=%d ",
176                   namelen, name, type, refs, master, mle->new_master, attached,
177                   mle->inuse);
178         dlm_print_nodemap(maybe);
179         printk(", ");
180         dlm_print_nodemap(vote);
181         printk(", ");
182         dlm_print_nodemap(resp);
183         printk(", ");
184         dlm_print_nodemap(node);
185         printk(", ");
186         printk("\n");
187 }
188
189 #if 0
190 /* Code here is included but defined out as it aids debugging */
191
192 static void dlm_dump_mles(struct dlm_ctxt *dlm)
193 {
194         struct dlm_master_list_entry *mle;
195         
196         mlog(ML_NOTICE, "dumping all mles for domain %s:\n", dlm->name);
197         spin_lock(&dlm->master_lock);
198         list_for_each_entry(mle, &dlm->master_list, list)
199                 dlm_print_one_mle(mle);
200         spin_unlock(&dlm->master_lock);
201 }
202
203 int dlm_dump_all_mles(const char __user *data, unsigned int len)
204 {
205         struct dlm_ctxt *dlm;
206
207         spin_lock(&dlm_domain_lock);
208         list_for_each_entry(dlm, &dlm_domains, list) {
209                 mlog(ML_NOTICE, "found dlm: %p, name=%s\n", dlm, dlm->name);
210                 dlm_dump_mles(dlm);
211         }
212         spin_unlock(&dlm_domain_lock);
213         return len;
214 }
215 EXPORT_SYMBOL_GPL(dlm_dump_all_mles);
216
217 #endif  /*  0  */
218
219
220 static struct kmem_cache *dlm_mle_cache = NULL;
221
222
223 static void dlm_mle_release(struct kref *kref);
224 static void dlm_init_mle(struct dlm_master_list_entry *mle,
225                         enum dlm_mle_type type,
226                         struct dlm_ctxt *dlm,
227                         struct dlm_lock_resource *res,
228                         const char *name,
229                         unsigned int namelen);
230 static void dlm_put_mle(struct dlm_master_list_entry *mle);
231 static void __dlm_put_mle(struct dlm_master_list_entry *mle);
232 static int dlm_find_mle(struct dlm_ctxt *dlm,
233                         struct dlm_master_list_entry **mle,
234                         char *name, unsigned int namelen);
235
236 static int dlm_do_master_request(struct dlm_lock_resource *res,
237                                  struct dlm_master_list_entry *mle, int to);
238
239
240 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
241                                      struct dlm_lock_resource *res,
242                                      struct dlm_master_list_entry *mle,
243                                      int *blocked);
244 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
245                                     struct dlm_lock_resource *res,
246                                     struct dlm_master_list_entry *mle,
247                                     int blocked);
248 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
249                                  struct dlm_lock_resource *res,
250                                  struct dlm_master_list_entry *mle,
251                                  struct dlm_master_list_entry **oldmle,
252                                  const char *name, unsigned int namelen,
253                                  u8 new_master, u8 master);
254
255 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
256                                     struct dlm_lock_resource *res);
257 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
258                                       struct dlm_lock_resource *res);
259 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
260                                        struct dlm_lock_resource *res,
261                                        u8 target);
262 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
263                                        struct dlm_lock_resource *res);
264
265
266 int dlm_is_host_down(int errno)
267 {
268         switch (errno) {
269                 case -EBADF:
270                 case -ECONNREFUSED:
271                 case -ENOTCONN:
272                 case -ECONNRESET:
273                 case -EPIPE:
274                 case -EHOSTDOWN:
275                 case -EHOSTUNREACH:
276                 case -ETIMEDOUT:
277                 case -ECONNABORTED:
278                 case -ENETDOWN:
279                 case -ENETUNREACH:
280                 case -ENETRESET:
281                 case -ESHUTDOWN:
282                 case -ENOPROTOOPT:
283                 case -EINVAL:   /* if returned from our tcp code,
284                                    this means there is no socket */
285                         return 1;
286         }
287         return 0;
288 }
289
290
291 /*
292  * MASTER LIST FUNCTIONS
293  */
294
295
296 /*
297  * regarding master list entries and heartbeat callbacks:
298  *
299  * in order to avoid sleeping and allocation that occurs in
300  * heartbeat, master list entries are simply attached to the
301  * dlm's established heartbeat callbacks.  the mle is attached
302  * when it is created, and since the dlm->spinlock is held at
303  * that time, any heartbeat event will be properly discovered
304  * by the mle.  the mle needs to be detached from the
305  * dlm->mle_hb_events list as soon as heartbeat events are no
306  * longer useful to the mle, and before the mle is freed.
307  *
308  * as a general rule, heartbeat events are no longer needed by
309  * the mle once an "answer" regarding the lock master has been
310  * received.
311  */
312 static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
313                                               struct dlm_master_list_entry *mle)
314 {
315         assert_spin_locked(&dlm->spinlock);
316
317         list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
318 }
319
320
321 static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
322                                               struct dlm_master_list_entry *mle)
323 {
324         if (!list_empty(&mle->hb_events))
325                 list_del_init(&mle->hb_events);
326 }
327
328
329 static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
330                                             struct dlm_master_list_entry *mle)
331 {
332         spin_lock(&dlm->spinlock);
333         __dlm_mle_detach_hb_events(dlm, mle);
334         spin_unlock(&dlm->spinlock);
335 }
336
337 static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
338 {
339         struct dlm_ctxt *dlm;
340         dlm = mle->dlm;
341
342         assert_spin_locked(&dlm->spinlock);
343         assert_spin_locked(&dlm->master_lock);
344         mle->inuse++;
345         kref_get(&mle->mle_refs);
346 }
347
348 static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
349 {
350         struct dlm_ctxt *dlm;
351         dlm = mle->dlm;
352
353         spin_lock(&dlm->spinlock);
354         spin_lock(&dlm->master_lock);
355         mle->inuse--;
356         __dlm_put_mle(mle);
357         spin_unlock(&dlm->master_lock);
358         spin_unlock(&dlm->spinlock);
359
360 }
361
362 /* remove from list and free */
363 static void __dlm_put_mle(struct dlm_master_list_entry *mle)
364 {
365         struct dlm_ctxt *dlm;
366         dlm = mle->dlm;
367
368         assert_spin_locked(&dlm->spinlock);
369         assert_spin_locked(&dlm->master_lock);
370         if (!atomic_read(&mle->mle_refs.refcount)) {
371                 /* this may or may not crash, but who cares.
372                  * it's a BUG. */
373                 mlog(ML_ERROR, "bad mle: %p\n", mle);
374                 dlm_print_one_mle(mle);
375                 BUG();
376         } else
377                 kref_put(&mle->mle_refs, dlm_mle_release);
378 }
379
380
381 /* must not have any spinlocks coming in */
382 static void dlm_put_mle(struct dlm_master_list_entry *mle)
383 {
384         struct dlm_ctxt *dlm;
385         dlm = mle->dlm;
386
387         spin_lock(&dlm->spinlock);
388         spin_lock(&dlm->master_lock);
389         __dlm_put_mle(mle);
390         spin_unlock(&dlm->master_lock);
391         spin_unlock(&dlm->spinlock);
392 }
393
394 static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
395 {
396         kref_get(&mle->mle_refs);
397 }
398
399 static void dlm_init_mle(struct dlm_master_list_entry *mle,
400                         enum dlm_mle_type type,
401                         struct dlm_ctxt *dlm,
402                         struct dlm_lock_resource *res,
403                         const char *name,
404                         unsigned int namelen)
405 {
406         assert_spin_locked(&dlm->spinlock);
407
408         mle->dlm = dlm;
409         mle->type = type;
410         INIT_LIST_HEAD(&mle->list);
411         INIT_LIST_HEAD(&mle->hb_events);
412         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
413         spin_lock_init(&mle->spinlock);
414         init_waitqueue_head(&mle->wq);
415         atomic_set(&mle->woken, 0);
416         kref_init(&mle->mle_refs);
417         memset(mle->response_map, 0, sizeof(mle->response_map));
418         mle->master = O2NM_MAX_NODES;
419         mle->new_master = O2NM_MAX_NODES;
420         mle->inuse = 0;
421
422         if (mle->type == DLM_MLE_MASTER) {
423                 BUG_ON(!res);
424                 mle->u.res = res;
425         } else if (mle->type == DLM_MLE_BLOCK) {
426                 BUG_ON(!name);
427                 memcpy(mle->u.name.name, name, namelen);
428                 mle->u.name.len = namelen;
429         } else /* DLM_MLE_MIGRATION */ {
430                 BUG_ON(!name);
431                 memcpy(mle->u.name.name, name, namelen);
432                 mle->u.name.len = namelen;
433         }
434
435         /* copy off the node_map and register hb callbacks on our copy */
436         memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
437         memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
438         clear_bit(dlm->node_num, mle->vote_map);
439         clear_bit(dlm->node_num, mle->node_map);
440
441         /* attach the mle to the domain node up/down events */
442         __dlm_mle_attach_hb_events(dlm, mle);
443 }
444
445
446 /* returns 1 if found, 0 if not */
447 static int dlm_find_mle(struct dlm_ctxt *dlm,
448                         struct dlm_master_list_entry **mle,
449                         char *name, unsigned int namelen)
450 {
451         struct dlm_master_list_entry *tmpmle;
452
453         assert_spin_locked(&dlm->master_lock);
454
455         list_for_each_entry(tmpmle, &dlm->master_list, list) {
456                 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
457                         continue;
458                 dlm_get_mle(tmpmle);
459                 *mle = tmpmle;
460                 return 1;
461         }
462         return 0;
463 }
464
465 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
466 {
467         struct dlm_master_list_entry *mle;
468
469         assert_spin_locked(&dlm->spinlock);
470         
471         list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
472                 if (node_up)
473                         dlm_mle_node_up(dlm, mle, NULL, idx);
474                 else
475                         dlm_mle_node_down(dlm, mle, NULL, idx);
476         }
477 }
478
479 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
480                               struct dlm_master_list_entry *mle,
481                               struct o2nm_node *node, int idx)
482 {
483         spin_lock(&mle->spinlock);
484
485         if (!test_bit(idx, mle->node_map))
486                 mlog(0, "node %u already removed from nodemap!\n", idx);
487         else
488                 clear_bit(idx, mle->node_map);
489
490         spin_unlock(&mle->spinlock);
491 }
492
493 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
494                             struct dlm_master_list_entry *mle,
495                             struct o2nm_node *node, int idx)
496 {
497         spin_lock(&mle->spinlock);
498
499         if (test_bit(idx, mle->node_map))
500                 mlog(0, "node %u already in node map!\n", idx);
501         else
502                 set_bit(idx, mle->node_map);
503
504         spin_unlock(&mle->spinlock);
505 }
506
507
508 int dlm_init_mle_cache(void)
509 {
510         dlm_mle_cache = kmem_cache_create("dlm_mle_cache",
511                                           sizeof(struct dlm_master_list_entry),
512                                           0, SLAB_HWCACHE_ALIGN,
513                                           NULL);
514         if (dlm_mle_cache == NULL)
515                 return -ENOMEM;
516         return 0;
517 }
518
519 void dlm_destroy_mle_cache(void)
520 {
521         if (dlm_mle_cache)
522                 kmem_cache_destroy(dlm_mle_cache);
523 }
524
525 static void dlm_mle_release(struct kref *kref)
526 {
527         struct dlm_master_list_entry *mle;
528         struct dlm_ctxt *dlm;
529
530         mlog_entry_void();
531
532         mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
533         dlm = mle->dlm;
534
535         if (mle->type != DLM_MLE_MASTER) {
536                 mlog(0, "calling mle_release for %.*s, type %d\n",
537                      mle->u.name.len, mle->u.name.name, mle->type);
538         } else {
539                 mlog(0, "calling mle_release for %.*s, type %d\n",
540                      mle->u.res->lockname.len,
541                      mle->u.res->lockname.name, mle->type);
542         }
543         assert_spin_locked(&dlm->spinlock);
544         assert_spin_locked(&dlm->master_lock);
545
546         /* remove from list if not already */
547         if (!list_empty(&mle->list))
548                 list_del_init(&mle->list);
549
550         /* detach the mle from the domain node up/down events */
551         __dlm_mle_detach_hb_events(dlm, mle);
552
553         /* NOTE: kfree under spinlock here.
554          * if this is bad, we can move this to a freelist. */
555         kmem_cache_free(dlm_mle_cache, mle);
556 }
557
558
559 /*
560  * LOCK RESOURCE FUNCTIONS
561  */
562
563 static void dlm_set_lockres_owner(struct dlm_ctxt *dlm,
564                                   struct dlm_lock_resource *res,
565                                   u8 owner)
566 {
567         assert_spin_locked(&res->spinlock);
568
569         mlog_entry("%.*s, %u\n", res->lockname.len, res->lockname.name, owner);
570
571         if (owner == dlm->node_num)
572                 atomic_inc(&dlm->local_resources);
573         else if (owner == DLM_LOCK_RES_OWNER_UNKNOWN)
574                 atomic_inc(&dlm->unknown_resources);
575         else
576                 atomic_inc(&dlm->remote_resources);
577
578         res->owner = owner;
579 }
580
581 void dlm_change_lockres_owner(struct dlm_ctxt *dlm,
582                               struct dlm_lock_resource *res, u8 owner)
583 {
584         assert_spin_locked(&res->spinlock);
585
586         if (owner == res->owner)
587                 return;
588
589         if (res->owner == dlm->node_num)
590                 atomic_dec(&dlm->local_resources);
591         else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN)
592                 atomic_dec(&dlm->unknown_resources);
593         else
594                 atomic_dec(&dlm->remote_resources);
595
596         dlm_set_lockres_owner(dlm, res, owner);
597 }
598
599
600 static void dlm_lockres_release(struct kref *kref)
601 {
602         struct dlm_lock_resource *res;
603
604         res = container_of(kref, struct dlm_lock_resource, refs);
605
606         /* This should not happen -- all lockres' have a name
607          * associated with them at init time. */
608         BUG_ON(!res->lockname.name);
609
610         mlog(0, "destroying lockres %.*s\n", res->lockname.len,
611              res->lockname.name);
612
613         if (!hlist_unhashed(&res->hash_node) ||
614             !list_empty(&res->granted) ||
615             !list_empty(&res->converting) ||
616             !list_empty(&res->blocked) ||
617             !list_empty(&res->dirty) ||
618             !list_empty(&res->recovering) ||
619             !list_empty(&res->purge)) {
620                 mlog(ML_ERROR,
621                      "Going to BUG for resource %.*s."
622                      "  We're on a list! [%c%c%c%c%c%c%c]\n",
623                      res->lockname.len, res->lockname.name,
624                      !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
625                      !list_empty(&res->granted) ? 'G' : ' ',
626                      !list_empty(&res->converting) ? 'C' : ' ',
627                      !list_empty(&res->blocked) ? 'B' : ' ',
628                      !list_empty(&res->dirty) ? 'D' : ' ',
629                      !list_empty(&res->recovering) ? 'R' : ' ',
630                      !list_empty(&res->purge) ? 'P' : ' ');
631
632                 dlm_print_one_lock_resource(res);
633         }
634
635         /* By the time we're ready to blow this guy away, we shouldn't
636          * be on any lists. */
637         BUG_ON(!hlist_unhashed(&res->hash_node));
638         BUG_ON(!list_empty(&res->granted));
639         BUG_ON(!list_empty(&res->converting));
640         BUG_ON(!list_empty(&res->blocked));
641         BUG_ON(!list_empty(&res->dirty));
642         BUG_ON(!list_empty(&res->recovering));
643         BUG_ON(!list_empty(&res->purge));
644
645         kfree(res->lockname.name);
646
647         kfree(res);
648 }
649
650 void dlm_lockres_put(struct dlm_lock_resource *res)
651 {
652         kref_put(&res->refs, dlm_lockres_release);
653 }
654
655 static void dlm_init_lockres(struct dlm_ctxt *dlm,
656                              struct dlm_lock_resource *res,
657                              const char *name, unsigned int namelen)
658 {
659         char *qname;
660
661         /* If we memset here, we lose our reference to the kmalloc'd
662          * res->lockname.name, so be sure to init every field
663          * correctly! */
664
665         qname = (char *) res->lockname.name;
666         memcpy(qname, name, namelen);
667
668         res->lockname.len = namelen;
669         res->lockname.hash = dlm_lockid_hash(name, namelen);
670
671         init_waitqueue_head(&res->wq);
672         spin_lock_init(&res->spinlock);
673         INIT_HLIST_NODE(&res->hash_node);
674         INIT_LIST_HEAD(&res->granted);
675         INIT_LIST_HEAD(&res->converting);
676         INIT_LIST_HEAD(&res->blocked);
677         INIT_LIST_HEAD(&res->dirty);
678         INIT_LIST_HEAD(&res->recovering);
679         INIT_LIST_HEAD(&res->purge);
680         atomic_set(&res->asts_reserved, 0);
681         res->migration_pending = 0;
682         res->inflight_locks = 0;
683
684         kref_init(&res->refs);
685
686         /* just for consistency */
687         spin_lock(&res->spinlock);
688         dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
689         spin_unlock(&res->spinlock);
690
691         res->state = DLM_LOCK_RES_IN_PROGRESS;
692
693         res->last_used = 0;
694
695         memset(res->lvb, 0, DLM_LVB_LEN);
696         memset(res->refmap, 0, sizeof(res->refmap));
697 }
698
699 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
700                                    const char *name,
701                                    unsigned int namelen)
702 {
703         struct dlm_lock_resource *res;
704
705         res = kmalloc(sizeof(struct dlm_lock_resource), GFP_NOFS);
706         if (!res)
707                 return NULL;
708
709         res->lockname.name = kmalloc(namelen, GFP_NOFS);
710         if (!res->lockname.name) {
711                 kfree(res);
712                 return NULL;
713         }
714
715         dlm_init_lockres(dlm, res, name, namelen);
716         return res;
717 }
718
719 void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
720                                    struct dlm_lock_resource *res,
721                                    int new_lockres,
722                                    const char *file,
723                                    int line)
724 {
725         if (!new_lockres)
726                 assert_spin_locked(&res->spinlock);
727
728         if (!test_bit(dlm->node_num, res->refmap)) {
729                 BUG_ON(res->inflight_locks != 0);
730                 dlm_lockres_set_refmap_bit(dlm->node_num, res);
731         }
732         res->inflight_locks++;
733         mlog(0, "%s:%.*s: inflight++: now %u\n",
734              dlm->name, res->lockname.len, res->lockname.name,
735              res->inflight_locks);
736 }
737
738 void __dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
739                                    struct dlm_lock_resource *res,
740                                    const char *file,
741                                    int line)
742 {
743         assert_spin_locked(&res->spinlock);
744
745         BUG_ON(res->inflight_locks == 0);
746         res->inflight_locks--;
747         mlog(0, "%s:%.*s: inflight--: now %u\n",
748              dlm->name, res->lockname.len, res->lockname.name,
749              res->inflight_locks);
750         if (res->inflight_locks == 0)
751                 dlm_lockres_clear_refmap_bit(dlm->node_num, res);
752         wake_up(&res->wq);
753 }
754
755 /*
756  * lookup a lock resource by name.
757  * may already exist in the hashtable.
758  * lockid is null terminated
759  *
760  * if not, allocate enough for the lockres and for
761  * the temporary structure used in doing the mastering.
762  *
763  * also, do a lookup in the dlm->master_list to see
764  * if another node has begun mastering the same lock.
765  * if so, there should be a block entry in there
766  * for this name, and we should *not* attempt to master
767  * the lock here.   need to wait around for that node
768  * to assert_master (or die).
769  *
770  */
771 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
772                                           const char *lockid,
773                                           int namelen,
774                                           int flags)
775 {
776         struct dlm_lock_resource *tmpres=NULL, *res=NULL;
777         struct dlm_master_list_entry *mle = NULL;
778         struct dlm_master_list_entry *alloc_mle = NULL;
779         int blocked = 0;
780         int ret, nodenum;
781         struct dlm_node_iter iter;
782         unsigned int hash;
783         int tries = 0;
784         int bit, wait_on_recovery = 0;
785         int drop_inflight_if_nonlocal = 0;
786
787         BUG_ON(!lockid);
788
789         hash = dlm_lockid_hash(lockid, namelen);
790
791         mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
792
793 lookup:
794         spin_lock(&dlm->spinlock);
795         tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
796         if (tmpres) {
797                 int dropping_ref = 0;
798
799                 spin_lock(&tmpres->spinlock);
800                 if (tmpres->owner == dlm->node_num) {
801                         BUG_ON(tmpres->state & DLM_LOCK_RES_DROPPING_REF);
802                         dlm_lockres_grab_inflight_ref(dlm, tmpres);
803                 } else if (tmpres->state & DLM_LOCK_RES_DROPPING_REF)
804                         dropping_ref = 1;
805                 spin_unlock(&tmpres->spinlock);
806                 spin_unlock(&dlm->spinlock);
807
808                 /* wait until done messaging the master, drop our ref to allow
809                  * the lockres to be purged, start over. */
810                 if (dropping_ref) {
811                         spin_lock(&tmpres->spinlock);
812                         __dlm_wait_on_lockres_flags(tmpres, DLM_LOCK_RES_DROPPING_REF);
813                         spin_unlock(&tmpres->spinlock);
814                         dlm_lockres_put(tmpres);
815                         tmpres = NULL;
816                         goto lookup;
817                 }
818
819                 mlog(0, "found in hash!\n");
820                 if (res)
821                         dlm_lockres_put(res);
822                 res = tmpres;
823                 goto leave;
824         }
825
826         if (!res) {
827                 spin_unlock(&dlm->spinlock);
828                 mlog(0, "allocating a new resource\n");
829                 /* nothing found and we need to allocate one. */
830                 alloc_mle = (struct dlm_master_list_entry *)
831                         kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
832                 if (!alloc_mle)
833                         goto leave;
834                 res = dlm_new_lockres(dlm, lockid, namelen);
835                 if (!res)
836                         goto leave;
837                 goto lookup;
838         }
839
840         mlog(0, "no lockres found, allocated our own: %p\n", res);
841
842         if (flags & LKM_LOCAL) {
843                 /* caller knows it's safe to assume it's not mastered elsewhere
844                  * DONE!  return right away */
845                 spin_lock(&res->spinlock);
846                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
847                 __dlm_insert_lockres(dlm, res);
848                 dlm_lockres_grab_inflight_ref(dlm, res);
849                 spin_unlock(&res->spinlock);
850                 spin_unlock(&dlm->spinlock);
851                 /* lockres still marked IN_PROGRESS */
852                 goto wake_waiters;
853         }
854
855         /* check master list to see if another node has started mastering it */
856         spin_lock(&dlm->master_lock);
857
858         /* if we found a block, wait for lock to be mastered by another node */
859         blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
860         if (blocked) {
861                 int mig;
862                 if (mle->type == DLM_MLE_MASTER) {
863                         mlog(ML_ERROR, "master entry for nonexistent lock!\n");
864                         BUG();
865                 }
866                 mig = (mle->type == DLM_MLE_MIGRATION);
867                 /* if there is a migration in progress, let the migration
868                  * finish before continuing.  we can wait for the absence
869                  * of the MIGRATION mle: either the migrate finished or
870                  * one of the nodes died and the mle was cleaned up.
871                  * if there is a BLOCK here, but it already has a master
872                  * set, we are too late.  the master does not have a ref
873                  * for us in the refmap.  detach the mle and drop it.
874                  * either way, go back to the top and start over. */
875                 if (mig || mle->master != O2NM_MAX_NODES) {
876                         BUG_ON(mig && mle->master == dlm->node_num);
877                         /* we arrived too late.  the master does not
878                          * have a ref for us. retry. */
879                         mlog(0, "%s:%.*s: late on %s\n",
880                              dlm->name, namelen, lockid,
881                              mig ?  "MIGRATION" : "BLOCK");
882                         spin_unlock(&dlm->master_lock);
883                         spin_unlock(&dlm->spinlock);
884
885                         /* master is known, detach */
886                         if (!mig)
887                                 dlm_mle_detach_hb_events(dlm, mle);
888                         dlm_put_mle(mle);
889                         mle = NULL;
890                         /* this is lame, but we cant wait on either
891                          * the mle or lockres waitqueue here */
892                         if (mig)
893                                 msleep(100);
894                         goto lookup;
895                 }
896         } else {
897                 /* go ahead and try to master lock on this node */
898                 mle = alloc_mle;
899                 /* make sure this does not get freed below */
900                 alloc_mle = NULL;
901                 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
902                 set_bit(dlm->node_num, mle->maybe_map);
903                 list_add(&mle->list, &dlm->master_list);
904
905                 /* still holding the dlm spinlock, check the recovery map
906                  * to see if there are any nodes that still need to be 
907                  * considered.  these will not appear in the mle nodemap
908                  * but they might own this lockres.  wait on them. */
909                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
910                 if (bit < O2NM_MAX_NODES) {
911                         mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to "
912                              "recover before lock mastery can begin\n",
913                              dlm->name, namelen, (char *)lockid, bit);
914                         wait_on_recovery = 1;
915                 }
916         }
917
918         /* at this point there is either a DLM_MLE_BLOCK or a
919          * DLM_MLE_MASTER on the master list, so it's safe to add the
920          * lockres to the hashtable.  anyone who finds the lock will
921          * still have to wait on the IN_PROGRESS. */
922
923         /* finally add the lockres to its hash bucket */
924         __dlm_insert_lockres(dlm, res);
925         /* since this lockres is new it doesnt not require the spinlock */
926         dlm_lockres_grab_inflight_ref_new(dlm, res);
927
928         /* if this node does not become the master make sure to drop
929          * this inflight reference below */
930         drop_inflight_if_nonlocal = 1;
931
932         /* get an extra ref on the mle in case this is a BLOCK
933          * if so, the creator of the BLOCK may try to put the last
934          * ref at this time in the assert master handler, so we
935          * need an extra one to keep from a bad ptr deref. */
936         dlm_get_mle_inuse(mle);
937         spin_unlock(&dlm->master_lock);
938         spin_unlock(&dlm->spinlock);
939
940 redo_request:
941         while (wait_on_recovery) {
942                 /* any cluster changes that occurred after dropping the
943                  * dlm spinlock would be detectable be a change on the mle,
944                  * so we only need to clear out the recovery map once. */
945                 if (dlm_is_recovery_lock(lockid, namelen)) {
946                         mlog(ML_NOTICE, "%s: recovery map is not empty, but "
947                              "must master $RECOVERY lock now\n", dlm->name);
948                         if (!dlm_pre_master_reco_lockres(dlm, res))
949                                 wait_on_recovery = 0;
950                         else {
951                                 mlog(0, "%s: waiting 500ms for heartbeat state "
952                                     "change\n", dlm->name);
953                                 msleep(500);
954                         }
955                         continue;
956                 } 
957
958                 dlm_kick_recovery_thread(dlm);
959                 msleep(1000);
960                 dlm_wait_for_recovery(dlm);
961
962                 spin_lock(&dlm->spinlock);
963                 bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
964                 if (bit < O2NM_MAX_NODES) {
965                         mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to "
966                              "recover before lock mastery can begin\n",
967                              dlm->name, namelen, (char *)lockid, bit);
968                         wait_on_recovery = 1;
969                 } else
970                         wait_on_recovery = 0;
971                 spin_unlock(&dlm->spinlock);
972
973                 if (wait_on_recovery)
974                         dlm_wait_for_node_recovery(dlm, bit, 10000);
975         }
976
977         /* must wait for lock to be mastered elsewhere */
978         if (blocked)
979                 goto wait;
980
981         ret = -EINVAL;
982         dlm_node_iter_init(mle->vote_map, &iter);
983         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
984                 ret = dlm_do_master_request(res, mle, nodenum);
985                 if (ret < 0)
986                         mlog_errno(ret);
987                 if (mle->master != O2NM_MAX_NODES) {
988                         /* found a master ! */
989                         if (mle->master <= nodenum)
990                                 break;
991                         /* if our master request has not reached the master
992                          * yet, keep going until it does.  this is how the
993                          * master will know that asserts are needed back to
994                          * the lower nodes. */
995                         mlog(0, "%s:%.*s: requests only up to %u but master "
996                              "is %u, keep going\n", dlm->name, namelen,
997                              lockid, nodenum, mle->master);
998                 }
999         }
1000
1001 wait:
1002         /* keep going until the response map includes all nodes */
1003         ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
1004         if (ret < 0) {
1005                 wait_on_recovery = 1;
1006                 mlog(0, "%s:%.*s: node map changed, redo the "
1007                      "master request now, blocked=%d\n",
1008                      dlm->name, res->lockname.len,
1009                      res->lockname.name, blocked);
1010                 if (++tries > 20) {
1011                         mlog(ML_ERROR, "%s:%.*s: spinning on "
1012                              "dlm_wait_for_lock_mastery, blocked=%d\n", 
1013                              dlm->name, res->lockname.len, 
1014                              res->lockname.name, blocked);
1015                         dlm_print_one_lock_resource(res);
1016                         dlm_print_one_mle(mle);
1017                         tries = 0;
1018                 }
1019                 goto redo_request;
1020         }
1021
1022         mlog(0, "lockres mastered by %u\n", res->owner);
1023         /* make sure we never continue without this */
1024         BUG_ON(res->owner == O2NM_MAX_NODES);
1025
1026         /* master is known, detach if not already detached */
1027         dlm_mle_detach_hb_events(dlm, mle);
1028         dlm_put_mle(mle);
1029         /* put the extra ref */
1030         dlm_put_mle_inuse(mle);
1031
1032 wake_waiters:
1033         spin_lock(&res->spinlock);
1034         if (res->owner != dlm->node_num && drop_inflight_if_nonlocal)
1035                 dlm_lockres_drop_inflight_ref(dlm, res);
1036         res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1037         spin_unlock(&res->spinlock);
1038         wake_up(&res->wq);
1039
1040 leave:
1041         /* need to free the unused mle */
1042         if (alloc_mle)
1043                 kmem_cache_free(dlm_mle_cache, alloc_mle);
1044
1045         return res;
1046 }
1047
1048
1049 #define DLM_MASTERY_TIMEOUT_MS   5000
1050
1051 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
1052                                      struct dlm_lock_resource *res,
1053                                      struct dlm_master_list_entry *mle,
1054                                      int *blocked)
1055 {
1056         u8 m;
1057         int ret, bit;
1058         int map_changed, voting_done;
1059         int assert, sleep;
1060
1061 recheck:
1062         ret = 0;
1063         assert = 0;
1064
1065         /* check if another node has already become the owner */
1066         spin_lock(&res->spinlock);
1067         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1068                 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1069                      res->lockname.len, res->lockname.name, res->owner);
1070                 spin_unlock(&res->spinlock);
1071                 /* this will cause the master to re-assert across
1072                  * the whole cluster, freeing up mles */
1073                 if (res->owner != dlm->node_num) {
1074                         ret = dlm_do_master_request(res, mle, res->owner);
1075                         if (ret < 0) {
1076                                 /* give recovery a chance to run */
1077                                 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1078                                 msleep(500);
1079                                 goto recheck;
1080                         }
1081                 }
1082                 ret = 0;
1083                 goto leave;
1084         }
1085         spin_unlock(&res->spinlock);
1086
1087         spin_lock(&mle->spinlock);
1088         m = mle->master;
1089         map_changed = (memcmp(mle->vote_map, mle->node_map,
1090                               sizeof(mle->vote_map)) != 0);
1091         voting_done = (memcmp(mle->vote_map, mle->response_map,
1092                              sizeof(mle->vote_map)) == 0);
1093
1094         /* restart if we hit any errors */
1095         if (map_changed) {
1096                 int b;
1097                 mlog(0, "%s: %.*s: node map changed, restarting\n",
1098                      dlm->name, res->lockname.len, res->lockname.name);
1099                 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1100                 b = (mle->type == DLM_MLE_BLOCK);
1101                 if ((*blocked && !b) || (!*blocked && b)) {
1102                         mlog(0, "%s:%.*s: status change: old=%d new=%d\n", 
1103                              dlm->name, res->lockname.len, res->lockname.name,
1104                              *blocked, b);
1105                         *blocked = b;
1106                 }
1107                 spin_unlock(&mle->spinlock);
1108                 if (ret < 0) {
1109                         mlog_errno(ret);
1110                         goto leave;
1111                 }
1112                 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1113                      "rechecking now\n", dlm->name, res->lockname.len,
1114                      res->lockname.name);
1115                 goto recheck;
1116         } else {
1117                 if (!voting_done) {
1118                         mlog(0, "map not changed and voting not done "
1119                              "for %s:%.*s\n", dlm->name, res->lockname.len,
1120                              res->lockname.name);
1121                 }
1122         }
1123
1124         if (m != O2NM_MAX_NODES) {
1125                 /* another node has done an assert!
1126                  * all done! */
1127                 sleep = 0;
1128         } else {
1129                 sleep = 1;
1130                 /* have all nodes responded? */
1131                 if (voting_done && !*blocked) {
1132                         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
1133                         if (dlm->node_num <= bit) {
1134                                 /* my node number is lowest.
1135                                  * now tell other nodes that I am
1136                                  * mastering this. */
1137                                 mle->master = dlm->node_num;
1138                                 /* ref was grabbed in get_lock_resource
1139                                  * will be dropped in dlmlock_master */
1140                                 assert = 1;
1141                                 sleep = 0;
1142                         }
1143                         /* if voting is done, but we have not received
1144                          * an assert master yet, we must sleep */
1145                 }
1146         }
1147
1148         spin_unlock(&mle->spinlock);
1149
1150         /* sleep if we haven't finished voting yet */
1151         if (sleep) {
1152                 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1153
1154                 /*
1155                 if (atomic_read(&mle->mle_refs.refcount) < 2)
1156                         mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
1157                         atomic_read(&mle->mle_refs.refcount),
1158                         res->lockname.len, res->lockname.name);
1159                 */
1160                 atomic_set(&mle->woken, 0);
1161                 (void)wait_event_timeout(mle->wq,
1162                                          (atomic_read(&mle->woken) == 1),
1163                                          timeo);
1164                 if (res->owner == O2NM_MAX_NODES) {
1165                         mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1166                              res->lockname.len, res->lockname.name);
1167                         goto recheck;
1168                 }
1169                 mlog(0, "done waiting, master is %u\n", res->owner);
1170                 ret = 0;
1171                 goto leave;
1172         }
1173
1174         ret = 0;   /* done */
1175         if (assert) {
1176                 m = dlm->node_num;
1177                 mlog(0, "about to master %.*s here, this=%u\n",
1178                      res->lockname.len, res->lockname.name, m);
1179                 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1180                 if (ret) {
1181                         /* This is a failure in the network path,
1182                          * not in the response to the assert_master
1183                          * (any nonzero response is a BUG on this node).
1184                          * Most likely a socket just got disconnected
1185                          * due to node death. */
1186                         mlog_errno(ret);
1187                 }
1188                 /* no longer need to restart lock mastery.
1189                  * all living nodes have been contacted. */
1190                 ret = 0;
1191         }
1192
1193         /* set the lockres owner */
1194         spin_lock(&res->spinlock);
1195         /* mastery reference obtained either during
1196          * assert_master_handler or in get_lock_resource */
1197         dlm_change_lockres_owner(dlm, res, m);
1198         spin_unlock(&res->spinlock);
1199
1200 leave:
1201         return ret;
1202 }
1203
1204 struct dlm_bitmap_diff_iter
1205 {
1206         int curnode;
1207         unsigned long *orig_bm;
1208         unsigned long *cur_bm;
1209         unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1210 };
1211
1212 enum dlm_node_state_change
1213 {
1214         NODE_DOWN = -1,
1215         NODE_NO_CHANGE = 0,
1216         NODE_UP
1217 };
1218
1219 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1220                                       unsigned long *orig_bm,
1221                                       unsigned long *cur_bm)
1222 {
1223         unsigned long p1, p2;
1224         int i;
1225
1226         iter->curnode = -1;
1227         iter->orig_bm = orig_bm;
1228         iter->cur_bm = cur_bm;
1229
1230         for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1231                 p1 = *(iter->orig_bm + i);
1232                 p2 = *(iter->cur_bm + i);
1233                 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1234         }
1235 }
1236
1237 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1238                                      enum dlm_node_state_change *state)
1239 {
1240         int bit;
1241
1242         if (iter->curnode >= O2NM_MAX_NODES)
1243                 return -ENOENT;
1244
1245         bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1246                             iter->curnode+1);
1247         if (bit >= O2NM_MAX_NODES) {
1248                 iter->curnode = O2NM_MAX_NODES;
1249                 return -ENOENT;
1250         }
1251
1252         /* if it was there in the original then this node died */
1253         if (test_bit(bit, iter->orig_bm))
1254                 *state = NODE_DOWN;
1255         else
1256                 *state = NODE_UP;
1257
1258         iter->curnode = bit;
1259         return bit;
1260 }
1261
1262
1263 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1264                                     struct dlm_lock_resource *res,
1265                                     struct dlm_master_list_entry *mle,
1266                                     int blocked)
1267 {
1268         struct dlm_bitmap_diff_iter bdi;
1269         enum dlm_node_state_change sc;
1270         int node;
1271         int ret = 0;
1272
1273         mlog(0, "something happened such that the "
1274              "master process may need to be restarted!\n");
1275
1276         assert_spin_locked(&mle->spinlock);
1277
1278         dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1279         node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1280         while (node >= 0) {
1281                 if (sc == NODE_UP) {
1282                         /* a node came up.  clear any old vote from
1283                          * the response map and set it in the vote map
1284                          * then restart the mastery. */
1285                         mlog(ML_NOTICE, "node %d up while restarting\n", node);
1286
1287                         /* redo the master request, but only for the new node */
1288                         mlog(0, "sending request to new node\n");
1289                         clear_bit(node, mle->response_map);
1290                         set_bit(node, mle->vote_map);
1291                 } else {
1292                         mlog(ML_ERROR, "node down! %d\n", node);
1293                         if (blocked) {
1294                                 int lowest = find_next_bit(mle->maybe_map,
1295                                                        O2NM_MAX_NODES, 0);
1296
1297                                 /* act like it was never there */
1298                                 clear_bit(node, mle->maybe_map);
1299
1300                                 if (node == lowest) {
1301                                         mlog(0, "expected master %u died"
1302                                             " while this node was blocked "
1303                                             "waiting on it!\n", node);
1304                                         lowest = find_next_bit(mle->maybe_map,
1305                                                         O2NM_MAX_NODES,
1306                                                         lowest+1);
1307                                         if (lowest < O2NM_MAX_NODES) {
1308                                                 mlog(0, "%s:%.*s:still "
1309                                                      "blocked. waiting on %u "
1310                                                      "now\n", dlm->name,
1311                                                      res->lockname.len,
1312                                                      res->lockname.name,
1313                                                      lowest);
1314                                         } else {
1315                                                 /* mle is an MLE_BLOCK, but
1316                                                  * there is now nothing left to
1317                                                  * block on.  we need to return
1318                                                  * all the way back out and try
1319                                                  * again with an MLE_MASTER.
1320                                                  * dlm_do_local_recovery_cleanup
1321                                                  * has already run, so the mle
1322                                                  * refcount is ok */
1323                                                 mlog(0, "%s:%.*s: no "
1324                                                      "longer blocking. try to "
1325                                                      "master this here\n",
1326                                                      dlm->name,
1327                                                      res->lockname.len,
1328                                                      res->lockname.name);
1329                                                 mle->type = DLM_MLE_MASTER;
1330                                                 mle->u.res = res;
1331                                         }
1332                                 }
1333                         }
1334
1335                         /* now blank out everything, as if we had never
1336                          * contacted anyone */
1337                         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1338                         memset(mle->response_map, 0, sizeof(mle->response_map));
1339                         /* reset the vote_map to the current node_map */
1340                         memcpy(mle->vote_map, mle->node_map,
1341                                sizeof(mle->node_map));
1342                         /* put myself into the maybe map */
1343                         if (mle->type != DLM_MLE_BLOCK)
1344                                 set_bit(dlm->node_num, mle->maybe_map);
1345                 }
1346                 ret = -EAGAIN;
1347                 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1348         }
1349         return ret;
1350 }
1351
1352
1353 /*
1354  * DLM_MASTER_REQUEST_MSG
1355  *
1356  * returns: 0 on success,
1357  *          -errno on a network error
1358  *
1359  * on error, the caller should assume the target node is "dead"
1360  *
1361  */
1362
1363 static int dlm_do_master_request(struct dlm_lock_resource *res,
1364                                  struct dlm_master_list_entry *mle, int to)
1365 {
1366         struct dlm_ctxt *dlm = mle->dlm;
1367         struct dlm_master_request request;
1368         int ret, response=0, resend;
1369
1370         memset(&request, 0, sizeof(request));
1371         request.node_idx = dlm->node_num;
1372
1373         BUG_ON(mle->type == DLM_MLE_MIGRATION);
1374
1375         if (mle->type != DLM_MLE_MASTER) {
1376                 request.namelen = mle->u.name.len;
1377                 memcpy(request.name, mle->u.name.name, request.namelen);
1378         } else {
1379                 request.namelen = mle->u.res->lockname.len;
1380                 memcpy(request.name, mle->u.res->lockname.name,
1381                         request.namelen);
1382         }
1383
1384 again:
1385         ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1386                                  sizeof(request), to, &response);
1387         if (ret < 0)  {
1388                 if (ret == -ESRCH) {
1389                         /* should never happen */
1390                         mlog(ML_ERROR, "TCP stack not ready!\n");
1391                         BUG();
1392                 } else if (ret == -EINVAL) {
1393                         mlog(ML_ERROR, "bad args passed to o2net!\n");
1394                         BUG();
1395                 } else if (ret == -ENOMEM) {
1396                         mlog(ML_ERROR, "out of memory while trying to send "
1397                              "network message!  retrying\n");
1398                         /* this is totally crude */
1399                         msleep(50);
1400                         goto again;
1401                 } else if (!dlm_is_host_down(ret)) {
1402                         /* not a network error. bad. */
1403                         mlog_errno(ret);
1404                         mlog(ML_ERROR, "unhandled error!");
1405                         BUG();
1406                 }
1407                 /* all other errors should be network errors,
1408                  * and likely indicate node death */
1409                 mlog(ML_ERROR, "link to %d went down!\n", to);
1410                 goto out;
1411         }
1412
1413         ret = 0;
1414         resend = 0;
1415         spin_lock(&mle->spinlock);
1416         switch (response) {
1417                 case DLM_MASTER_RESP_YES:
1418                         set_bit(to, mle->response_map);
1419                         mlog(0, "node %u is the master, response=YES\n", to);
1420                         mlog(0, "%s:%.*s: master node %u now knows I have a "
1421                              "reference\n", dlm->name, res->lockname.len,
1422                              res->lockname.name, to);
1423                         mle->master = to;
1424                         break;
1425                 case DLM_MASTER_RESP_NO:
1426                         mlog(0, "node %u not master, response=NO\n", to);
1427                         set_bit(to, mle->response_map);
1428                         break;
1429                 case DLM_MASTER_RESP_MAYBE:
1430                         mlog(0, "node %u not master, response=MAYBE\n", to);
1431                         set_bit(to, mle->response_map);
1432                         set_bit(to, mle->maybe_map);
1433                         break;
1434                 case DLM_MASTER_RESP_ERROR:
1435                         mlog(0, "node %u hit an error, resending\n", to);
1436                         resend = 1;
1437                         response = 0;
1438                         break;
1439                 default:
1440                         mlog(ML_ERROR, "bad response! %u\n", response);
1441                         BUG();
1442         }
1443         spin_unlock(&mle->spinlock);
1444         if (resend) {
1445                 /* this is also totally crude */
1446                 msleep(50);
1447                 goto again;
1448         }
1449
1450 out:
1451         return ret;
1452 }
1453
1454 /*
1455  * locks that can be taken here:
1456  * dlm->spinlock
1457  * res->spinlock
1458  * mle->spinlock
1459  * dlm->master_list
1460  *
1461  * if possible, TRIM THIS DOWN!!!
1462  */
1463 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1464                                void **ret_data)
1465 {
1466         u8 response = DLM_MASTER_RESP_MAYBE;
1467         struct dlm_ctxt *dlm = data;
1468         struct dlm_lock_resource *res = NULL;
1469         struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1470         struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1471         char *name;
1472         unsigned int namelen, hash;
1473         int found, ret;
1474         int set_maybe;
1475         int dispatch_assert = 0;
1476
1477         if (!dlm_grab(dlm))
1478                 return DLM_MASTER_RESP_NO;
1479
1480         if (!dlm_domain_fully_joined(dlm)) {
1481                 response = DLM_MASTER_RESP_NO;
1482                 goto send_response;
1483         }
1484
1485         name = request->name;
1486         namelen = request->namelen;
1487         hash = dlm_lockid_hash(name, namelen);
1488
1489         if (namelen > DLM_LOCKID_NAME_MAX) {
1490                 response = DLM_IVBUFLEN;
1491                 goto send_response;
1492         }
1493
1494 way_up_top:
1495         spin_lock(&dlm->spinlock);
1496         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1497         if (res) {
1498                 spin_unlock(&dlm->spinlock);
1499
1500                 /* take care of the easy cases up front */
1501                 spin_lock(&res->spinlock);
1502                 if (res->state & (DLM_LOCK_RES_RECOVERING|
1503                                   DLM_LOCK_RES_MIGRATING)) {
1504                         spin_unlock(&res->spinlock);
1505                         mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1506                              "being recovered/migrated\n");
1507                         response = DLM_MASTER_RESP_ERROR;
1508                         if (mle)
1509                                 kmem_cache_free(dlm_mle_cache, mle);
1510                         goto send_response;
1511                 }
1512
1513                 if (res->owner == dlm->node_num) {
1514                         mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1515                              dlm->name, namelen, name, request->node_idx);
1516                         dlm_lockres_set_refmap_bit(request->node_idx, res);
1517                         spin_unlock(&res->spinlock);
1518                         response = DLM_MASTER_RESP_YES;
1519                         if (mle)
1520                                 kmem_cache_free(dlm_mle_cache, mle);
1521
1522                         /* this node is the owner.
1523                          * there is some extra work that needs to
1524                          * happen now.  the requesting node has
1525                          * caused all nodes up to this one to
1526                          * create mles.  this node now needs to
1527                          * go back and clean those up. */
1528                         dispatch_assert = 1;
1529                         goto send_response;
1530                 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1531                         spin_unlock(&res->spinlock);
1532                         // mlog(0, "node %u is the master\n", res->owner);
1533                         response = DLM_MASTER_RESP_NO;
1534                         if (mle)
1535                                 kmem_cache_free(dlm_mle_cache, mle);
1536                         goto send_response;
1537                 }
1538
1539                 /* ok, there is no owner.  either this node is
1540                  * being blocked, or it is actively trying to
1541                  * master this lock. */
1542                 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1543                         mlog(ML_ERROR, "lock with no owner should be "
1544                              "in-progress!\n");
1545                         BUG();
1546                 }
1547
1548                 // mlog(0, "lockres is in progress...\n");
1549                 spin_lock(&dlm->master_lock);
1550                 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1551                 if (!found) {
1552                         mlog(ML_ERROR, "no mle found for this lock!\n");
1553                         BUG();
1554                 }
1555                 set_maybe = 1;
1556                 spin_lock(&tmpmle->spinlock);
1557                 if (tmpmle->type == DLM_MLE_BLOCK) {
1558                         // mlog(0, "this node is waiting for "
1559                         // "lockres to be mastered\n");
1560                         response = DLM_MASTER_RESP_NO;
1561                 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1562                         mlog(0, "node %u is master, but trying to migrate to "
1563                              "node %u.\n", tmpmle->master, tmpmle->new_master);
1564                         if (tmpmle->master == dlm->node_num) {
1565                                 mlog(ML_ERROR, "no owner on lockres, but this "
1566                                      "node is trying to migrate it to %u?!\n",
1567                                      tmpmle->new_master);
1568                                 BUG();
1569                         } else {
1570                                 /* the real master can respond on its own */
1571                                 response = DLM_MASTER_RESP_NO;
1572                         }
1573                 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1574                         set_maybe = 0;
1575                         if (tmpmle->master == dlm->node_num) {
1576                                 response = DLM_MASTER_RESP_YES;
1577                                 /* this node will be the owner.
1578                                  * go back and clean the mles on any
1579                                  * other nodes */
1580                                 dispatch_assert = 1;
1581                                 dlm_lockres_set_refmap_bit(request->node_idx, res);
1582                                 mlog(0, "%s:%.*s: setting bit %u in refmap\n",
1583                                      dlm->name, namelen, name,
1584                                      request->node_idx);
1585                         } else
1586                                 response = DLM_MASTER_RESP_NO;
1587                 } else {
1588                         // mlog(0, "this node is attempting to "
1589                         // "master lockres\n");
1590                         response = DLM_MASTER_RESP_MAYBE;
1591                 }
1592                 if (set_maybe)
1593                         set_bit(request->node_idx, tmpmle->maybe_map);
1594                 spin_unlock(&tmpmle->spinlock);
1595
1596                 spin_unlock(&dlm->master_lock);
1597                 spin_unlock(&res->spinlock);
1598
1599                 /* keep the mle attached to heartbeat events */
1600                 dlm_put_mle(tmpmle);
1601                 if (mle)
1602                         kmem_cache_free(dlm_mle_cache, mle);
1603                 goto send_response;
1604         }
1605
1606         /*
1607          * lockres doesn't exist on this node
1608          * if there is an MLE_BLOCK, return NO
1609          * if there is an MLE_MASTER, return MAYBE
1610          * otherwise, add an MLE_BLOCK, return NO
1611          */
1612         spin_lock(&dlm->master_lock);
1613         found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1614         if (!found) {
1615                 /* this lockid has never been seen on this node yet */
1616                 // mlog(0, "no mle found\n");
1617                 if (!mle) {
1618                         spin_unlock(&dlm->master_lock);
1619                         spin_unlock(&dlm->spinlock);
1620
1621                         mle = (struct dlm_master_list_entry *)
1622                                 kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1623                         if (!mle) {
1624                                 response = DLM_MASTER_RESP_ERROR;
1625                                 mlog_errno(-ENOMEM);
1626                                 goto send_response;
1627                         }
1628                         goto way_up_top;
1629                 }
1630
1631                 // mlog(0, "this is second time thru, already allocated, "
1632                 // "add the block.\n");
1633                 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1634                 set_bit(request->node_idx, mle->maybe_map);
1635                 list_add(&mle->list, &dlm->master_list);
1636                 response = DLM_MASTER_RESP_NO;
1637         } else {
1638                 // mlog(0, "mle was found\n");
1639                 set_maybe = 1;
1640                 spin_lock(&tmpmle->spinlock);
1641                 if (tmpmle->master == dlm->node_num) {
1642                         mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1643                         BUG();
1644                 }
1645                 if (tmpmle->type == DLM_MLE_BLOCK)
1646                         response = DLM_MASTER_RESP_NO;
1647                 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1648                         mlog(0, "migration mle was found (%u->%u)\n",
1649                              tmpmle->master, tmpmle->new_master);
1650                         /* real master can respond on its own */
1651                         response = DLM_MASTER_RESP_NO;
1652                 } else
1653                         response = DLM_MASTER_RESP_MAYBE;
1654                 if (set_maybe)
1655                         set_bit(request->node_idx, tmpmle->maybe_map);
1656                 spin_unlock(&tmpmle->spinlock);
1657         }
1658         spin_unlock(&dlm->master_lock);
1659         spin_unlock(&dlm->spinlock);
1660
1661         if (found) {
1662                 /* keep the mle attached to heartbeat events */
1663                 dlm_put_mle(tmpmle);
1664         }
1665 send_response:
1666         /*
1667          * __dlm_lookup_lockres() grabbed a reference to this lockres.
1668          * The reference is released by dlm_assert_master_worker() under
1669          * the call to dlm_dispatch_assert_master().  If
1670          * dlm_assert_master_worker() isn't called, we drop it here.
1671          */
1672         if (dispatch_assert) {
1673                 if (response != DLM_MASTER_RESP_YES)
1674                         mlog(ML_ERROR, "invalid response %d\n", response);
1675                 if (!res) {
1676                         mlog(ML_ERROR, "bad lockres while trying to assert!\n");
1677                         BUG();
1678                 }
1679                 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1680                              dlm->node_num, res->lockname.len, res->lockname.name);
1681                 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx, 
1682                                                  DLM_ASSERT_MASTER_MLE_CLEANUP);
1683                 if (ret < 0) {
1684                         mlog(ML_ERROR, "failed to dispatch assert master work\n");
1685                         response = DLM_MASTER_RESP_ERROR;
1686                         dlm_lockres_put(res);
1687                 }
1688         } else {
1689                 if (res)
1690                         dlm_lockres_put(res);
1691         }
1692
1693         dlm_put(dlm);
1694         return response;
1695 }
1696
1697 /*
1698  * DLM_ASSERT_MASTER_MSG
1699  */
1700
1701
1702 /*
1703  * NOTE: this can be used for debugging
1704  * can periodically run all locks owned by this node
1705  * and re-assert across the cluster...
1706  */
1707 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1708                                 struct dlm_lock_resource *res,
1709                                 void *nodemap, u32 flags)
1710 {
1711         struct dlm_assert_master assert;
1712         int to, tmpret;
1713         struct dlm_node_iter iter;
1714         int ret = 0;
1715         int reassert;
1716         const char *lockname = res->lockname.name;
1717         unsigned int namelen = res->lockname.len;
1718
1719         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1720
1721         spin_lock(&res->spinlock);
1722         res->state |= DLM_LOCK_RES_SETREF_INPROG;
1723         spin_unlock(&res->spinlock);
1724
1725 again:
1726         reassert = 0;
1727
1728         /* note that if this nodemap is empty, it returns 0 */
1729         dlm_node_iter_init(nodemap, &iter);
1730         while ((to = dlm_node_iter_next(&iter)) >= 0) {
1731                 int r = 0;
1732                 struct dlm_master_list_entry *mle = NULL;
1733
1734                 mlog(0, "sending assert master to %d (%.*s)\n", to,
1735                      namelen, lockname);
1736                 memset(&assert, 0, sizeof(assert));
1737                 assert.node_idx = dlm->node_num;
1738                 assert.namelen = namelen;
1739                 memcpy(assert.name, lockname, namelen);
1740                 assert.flags = cpu_to_be32(flags);
1741
1742                 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1743                                             &assert, sizeof(assert), to, &r);
1744                 if (tmpret < 0) {
1745                         mlog(0, "assert_master returned %d!\n", tmpret);
1746                         if (!dlm_is_host_down(tmpret)) {
1747                                 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1748                                 BUG();
1749                         }
1750                         /* a node died.  finish out the rest of the nodes. */
1751                         mlog(0, "link to %d went down!\n", to);
1752                         /* any nonzero status return will do */
1753                         ret = tmpret;
1754                         r = 0;
1755                 } else if (r < 0) {
1756                         /* ok, something horribly messed.  kill thyself. */
1757                         mlog(ML_ERROR,"during assert master of %.*s to %u, "
1758                              "got %d.\n", namelen, lockname, to, r);
1759                         spin_lock(&dlm->spinlock);
1760                         spin_lock(&dlm->master_lock);
1761                         if (dlm_find_mle(dlm, &mle, (char *)lockname,
1762                                          namelen)) {
1763                                 dlm_print_one_mle(mle);
1764                                 __dlm_put_mle(mle);
1765                         }
1766                         spin_unlock(&dlm->master_lock);
1767                         spin_unlock(&dlm->spinlock);
1768                         BUG();
1769                 }
1770
1771                 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1772                     !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1773                                 mlog(ML_ERROR, "%.*s: very strange, "
1774                                      "master MLE but no lockres on %u\n",
1775                                      namelen, lockname, to);
1776                 }
1777
1778                 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1779                         mlog(0, "%.*s: node %u create mles on other "
1780                              "nodes and requests a re-assert\n", 
1781                              namelen, lockname, to);
1782                         reassert = 1;
1783                 }
1784                 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1785                         mlog(0, "%.*s: node %u has a reference to this "
1786                              "lockres, set the bit in the refmap\n",
1787                              namelen, lockname, to);
1788                         spin_lock(&res->spinlock);
1789                         dlm_lockres_set_refmap_bit(to, res);
1790                         spin_unlock(&res->spinlock);
1791                 }
1792         }
1793
1794         if (reassert)
1795                 goto again;
1796
1797         spin_lock(&res->spinlock);
1798         res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1799         spin_unlock(&res->spinlock);
1800         wake_up(&res->wq);
1801
1802         return ret;
1803 }
1804
1805 /*
1806  * locks that can be taken here:
1807  * dlm->spinlock
1808  * res->spinlock
1809  * mle->spinlock
1810  * dlm->master_list
1811  *
1812  * if possible, TRIM THIS DOWN!!!
1813  */
1814 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1815                               void **ret_data)
1816 {
1817         struct dlm_ctxt *dlm = data;
1818         struct dlm_master_list_entry *mle = NULL;
1819         struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1820         struct dlm_lock_resource *res = NULL;
1821         char *name;
1822         unsigned int namelen, hash;
1823         u32 flags;
1824         int master_request = 0, have_lockres_ref = 0;
1825         int ret = 0;
1826
1827         if (!dlm_grab(dlm))
1828                 return 0;
1829
1830         name = assert->name;
1831         namelen = assert->namelen;
1832         hash = dlm_lockid_hash(name, namelen);
1833         flags = be32_to_cpu(assert->flags);
1834
1835         if (namelen > DLM_LOCKID_NAME_MAX) {
1836                 mlog(ML_ERROR, "Invalid name length!");
1837                 goto done;
1838         }
1839
1840         spin_lock(&dlm->spinlock);
1841
1842         if (flags)
1843                 mlog(0, "assert_master with flags: %u\n", flags);
1844
1845         /* find the MLE */
1846         spin_lock(&dlm->master_lock);
1847         if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1848                 /* not an error, could be master just re-asserting */
1849                 mlog(0, "just got an assert_master from %u, but no "
1850                      "MLE for it! (%.*s)\n", assert->node_idx,
1851                      namelen, name);
1852         } else {
1853                 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1854                 if (bit >= O2NM_MAX_NODES) {
1855                         /* not necessarily an error, though less likely.
1856                          * could be master just re-asserting. */
1857                         mlog(0, "no bits set in the maybe_map, but %u "
1858                              "is asserting! (%.*s)\n", assert->node_idx,
1859                              namelen, name);
1860                 } else if (bit != assert->node_idx) {
1861                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1862                                 mlog(0, "master %u was found, %u should "
1863                                      "back off\n", assert->node_idx, bit);
1864                         } else {
1865                                 /* with the fix for bug 569, a higher node
1866                                  * number winning the mastery will respond
1867                                  * YES to mastery requests, but this node
1868                                  * had no way of knowing.  let it pass. */
1869                                 mlog(0, "%u is the lowest node, "
1870                                      "%u is asserting. (%.*s)  %u must "
1871                                      "have begun after %u won.\n", bit,
1872                                      assert->node_idx, namelen, name, bit,
1873                                      assert->node_idx);
1874                         }
1875                 }
1876                 if (mle->type == DLM_MLE_MIGRATION) {
1877                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1878                                 mlog(0, "%s:%.*s: got cleanup assert"
1879                                      " from %u for migration\n",
1880                                      dlm->name, namelen, name,
1881                                      assert->node_idx);
1882                         } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1883                                 mlog(0, "%s:%.*s: got unrelated assert"
1884                                      " from %u for migration, ignoring\n",
1885                                      dlm->name, namelen, name,
1886                                      assert->node_idx);
1887                                 __dlm_put_mle(mle);
1888                                 spin_unlock(&dlm->master_lock);
1889                                 spin_unlock(&dlm->spinlock);
1890                                 goto done;
1891                         }       
1892                 }
1893         }
1894         spin_unlock(&dlm->master_lock);
1895
1896         /* ok everything checks out with the MLE
1897          * now check to see if there is a lockres */
1898         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1899         if (res) {
1900                 spin_lock(&res->spinlock);
1901                 if (res->state & DLM_LOCK_RES_RECOVERING)  {
1902                         mlog(ML_ERROR, "%u asserting but %.*s is "
1903                              "RECOVERING!\n", assert->node_idx, namelen, name);
1904                         goto kill;
1905                 }
1906                 if (!mle) {
1907                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1908                             res->owner != assert->node_idx) {
1909                                 mlog(ML_ERROR, "assert_master from "
1910                                           "%u, but current owner is "
1911                                           "%u! (%.*s)\n",
1912                                        assert->node_idx, res->owner,
1913                                        namelen, name);
1914                                 goto kill;
1915                         }
1916                 } else if (mle->type != DLM_MLE_MIGRATION) {
1917                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1918                                 /* owner is just re-asserting */
1919                                 if (res->owner == assert->node_idx) {
1920                                         mlog(0, "owner %u re-asserting on "
1921                                              "lock %.*s\n", assert->node_idx,
1922                                              namelen, name);
1923                                         goto ok;
1924                                 }
1925                                 mlog(ML_ERROR, "got assert_master from "
1926                                      "node %u, but %u is the owner! "
1927                                      "(%.*s)\n", assert->node_idx,
1928                                      res->owner, namelen, name);
1929                                 goto kill;
1930                         }
1931                         if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1932                                 mlog(ML_ERROR, "got assert from %u, but lock "
1933                                      "with no owner should be "
1934                                      "in-progress! (%.*s)\n",
1935                                      assert->node_idx,
1936                                      namelen, name);
1937                                 goto kill;
1938                         }
1939                 } else /* mle->type == DLM_MLE_MIGRATION */ {
1940                         /* should only be getting an assert from new master */
1941                         if (assert->node_idx != mle->new_master) {
1942                                 mlog(ML_ERROR, "got assert from %u, but "
1943                                      "new master is %u, and old master "
1944                                      "was %u (%.*s)\n",
1945                                      assert->node_idx, mle->new_master,
1946                                      mle->master, namelen, name);
1947                                 goto kill;
1948                         }
1949
1950                 }
1951 ok:
1952                 spin_unlock(&res->spinlock);
1953         }
1954         spin_unlock(&dlm->spinlock);
1955
1956         // mlog(0, "woo!  got an assert_master from node %u!\n",
1957         //           assert->node_idx);
1958         if (mle) {
1959                 int extra_ref = 0;
1960                 int nn = -1;
1961                 int rr, err = 0;
1962                 
1963                 spin_lock(&mle->spinlock);
1964                 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1965                         extra_ref = 1;
1966                 else {
1967                         /* MASTER mle: if any bits set in the response map
1968                          * then the calling node needs to re-assert to clear
1969                          * up nodes that this node contacted */
1970                         while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES, 
1971                                                     nn+1)) < O2NM_MAX_NODES) {
1972                                 if (nn != dlm->node_num && nn != assert->node_idx)
1973                                         master_request = 1;
1974                         }
1975                 }
1976                 mle->master = assert->node_idx;
1977                 atomic_set(&mle->woken, 1);
1978                 wake_up(&mle->wq);
1979                 spin_unlock(&mle->spinlock);
1980
1981                 if (res) {
1982                         int wake = 0;
1983                         spin_lock(&res->spinlock);
1984                         if (mle->type == DLM_MLE_MIGRATION) {
1985                                 mlog(0, "finishing off migration of lockres %.*s, "
1986                                         "from %u to %u\n",
1987                                         res->lockname.len, res->lockname.name,
1988                                         dlm->node_num, mle->new_master);
1989                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
1990                                 wake = 1;
1991                                 dlm_change_lockres_owner(dlm, res, mle->new_master);
1992                                 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1993                         } else {
1994                                 dlm_change_lockres_owner(dlm, res, mle->master);
1995                         }
1996                         spin_unlock(&res->spinlock);
1997                         have_lockres_ref = 1;
1998                         if (wake)
1999                                 wake_up(&res->wq);
2000                 }
2001
2002                 /* master is known, detach if not already detached.
2003                  * ensures that only one assert_master call will happen
2004                  * on this mle. */
2005                 spin_lock(&dlm->spinlock);
2006                 spin_lock(&dlm->master_lock);
2007
2008                 rr = atomic_read(&mle->mle_refs.refcount);
2009                 if (mle->inuse > 0) {
2010                         if (extra_ref && rr < 3)
2011                                 err = 1;
2012                         else if (!extra_ref && rr < 2)
2013                                 err = 1;
2014                 } else {
2015                         if (extra_ref && rr < 2)
2016                                 err = 1;
2017                         else if (!extra_ref && rr < 1)
2018                                 err = 1;
2019                 }
2020                 if (err) {
2021                         mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
2022                              "that will mess up this node, refs=%d, extra=%d, "
2023                              "inuse=%d\n", dlm->name, namelen, name,
2024                              assert->node_idx, rr, extra_ref, mle->inuse);
2025                         dlm_print_one_mle(mle);
2026                 }
2027                 list_del_init(&mle->list);
2028                 __dlm_mle_detach_hb_events(dlm, mle);
2029                 __dlm_put_mle(mle);
2030                 if (extra_ref) {
2031                         /* the assert master message now balances the extra
2032                          * ref given by the master / migration request message.
2033                          * if this is the last put, it will be removed
2034                          * from the list. */
2035                         __dlm_put_mle(mle);
2036                 }
2037                 spin_unlock(&dlm->master_lock);
2038                 spin_unlock(&dlm->spinlock);
2039         } else if (res) {
2040                 if (res->owner != assert->node_idx) {
2041                         mlog(0, "assert_master from %u, but current "
2042                              "owner is %u (%.*s), no mle\n", assert->node_idx,
2043                              res->owner, namelen, name);
2044                 }
2045         }
2046
2047 done:
2048         ret = 0;
2049         if (res) {
2050                 spin_lock(&res->spinlock);
2051                 res->state |= DLM_LOCK_RES_SETREF_INPROG;
2052                 spin_unlock(&res->spinlock);
2053                 *ret_data = (void *)res;
2054         }
2055         dlm_put(dlm);
2056         if (master_request) {
2057                 mlog(0, "need to tell master to reassert\n");
2058                 /* positive. negative would shoot down the node. */
2059                 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2060                 if (!have_lockres_ref) {
2061                         mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2062                              "mle present here for %s:%.*s, but no lockres!\n",
2063                              assert->node_idx, dlm->name, namelen, name);
2064                 }
2065         }
2066         if (have_lockres_ref) {
2067                 /* let the master know we have a reference to the lockres */
2068                 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2069                 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2070                      dlm->name, namelen, name, assert->node_idx);
2071         }
2072         return ret;
2073
2074 kill:
2075         /* kill the caller! */
2076         mlog(ML_ERROR, "Bad message received from another node.  Dumping state "
2077              "and killing the other node now!  This node is OK and can continue.\n");
2078         __dlm_print_one_lock_resource(res);
2079         spin_unlock(&res->spinlock);
2080         spin_unlock(&dlm->spinlock);
2081         *ret_data = (void *)res; 
2082         dlm_put(dlm);
2083         return -EINVAL;
2084 }
2085
2086 void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2087 {
2088         struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2089
2090         if (ret_data) {
2091                 spin_lock(&res->spinlock);
2092                 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2093                 spin_unlock(&res->spinlock);
2094                 wake_up(&res->wq);
2095                 dlm_lockres_put(res);
2096         }
2097         return;
2098 }
2099
2100 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2101                                struct dlm_lock_resource *res,
2102                                int ignore_higher, u8 request_from, u32 flags)
2103 {
2104         struct dlm_work_item *item;
2105         item = kzalloc(sizeof(*item), GFP_NOFS);
2106         if (!item)
2107                 return -ENOMEM;
2108
2109
2110         /* queue up work for dlm_assert_master_worker */
2111         dlm_grab(dlm);  /* get an extra ref for the work item */
2112         dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2113         item->u.am.lockres = res; /* already have a ref */
2114         /* can optionally ignore node numbers higher than this node */
2115         item->u.am.ignore_higher = ignore_higher;
2116         item->u.am.request_from = request_from;
2117         item->u.am.flags = flags;
2118
2119         if (ignore_higher) 
2120                 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len, 
2121                      res->lockname.name);
2122                 
2123         spin_lock(&dlm->work_lock);
2124         list_add_tail(&item->list, &dlm->work_list);
2125         spin_unlock(&dlm->work_lock);
2126
2127         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2128         return 0;
2129 }
2130
2131 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2132 {
2133         struct dlm_ctxt *dlm = data;
2134         int ret = 0;
2135         struct dlm_lock_resource *res;
2136         unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2137         int ignore_higher;
2138         int bit;
2139         u8 request_from;
2140         u32 flags;
2141
2142         dlm = item->dlm;
2143         res = item->u.am.lockres;
2144         ignore_higher = item->u.am.ignore_higher;
2145         request_from = item->u.am.request_from;
2146         flags = item->u.am.flags;
2147
2148         spin_lock(&dlm->spinlock);
2149         memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2150         spin_unlock(&dlm->spinlock);
2151
2152         clear_bit(dlm->node_num, nodemap);
2153         if (ignore_higher) {
2154                 /* if is this just to clear up mles for nodes below
2155                  * this node, do not send the message to the original
2156                  * caller or any node number higher than this */
2157                 clear_bit(request_from, nodemap);
2158                 bit = dlm->node_num;
2159                 while (1) {
2160                         bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2161                                             bit+1);
2162                         if (bit >= O2NM_MAX_NODES)
2163                                 break;
2164                         clear_bit(bit, nodemap);
2165                 }
2166         }
2167
2168         /*
2169          * If we're migrating this lock to someone else, we are no
2170          * longer allowed to assert out own mastery.  OTOH, we need to
2171          * prevent migration from starting while we're still asserting
2172          * our dominance.  The reserved ast delays migration.
2173          */
2174         spin_lock(&res->spinlock);
2175         if (res->state & DLM_LOCK_RES_MIGRATING) {
2176                 mlog(0, "Someone asked us to assert mastery, but we're "
2177                      "in the middle of migration.  Skipping assert, "
2178                      "the new master will handle that.\n");
2179                 spin_unlock(&res->spinlock);
2180                 goto put;
2181         } else
2182                 __dlm_lockres_reserve_ast(res);
2183         spin_unlock(&res->spinlock);
2184
2185         /* this call now finishes out the nodemap
2186          * even if one or more nodes die */
2187         mlog(0, "worker about to master %.*s here, this=%u\n",
2188                      res->lockname.len, res->lockname.name, dlm->node_num);
2189         ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2190         if (ret < 0) {
2191                 /* no need to restart, we are done */
2192                 if (!dlm_is_host_down(ret))
2193                         mlog_errno(ret);
2194         }
2195
2196         /* Ok, we've asserted ourselves.  Let's let migration start. */
2197         dlm_lockres_release_ast(dlm, res);
2198
2199 put:
2200         dlm_lockres_put(res);
2201
2202         mlog(0, "finished with dlm_assert_master_worker\n");
2203 }
2204
2205 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2206  * We cannot wait for node recovery to complete to begin mastering this
2207  * lockres because this lockres is used to kick off recovery! ;-)
2208  * So, do a pre-check on all living nodes to see if any of those nodes
2209  * think that $RECOVERY is currently mastered by a dead node.  If so,
2210  * we wait a short time to allow that node to get notified by its own
2211  * heartbeat stack, then check again.  All $RECOVERY lock resources
2212  * mastered by dead nodes are purged when the hearbeat callback is 
2213  * fired, so we can know for sure that it is safe to continue once
2214  * the node returns a live node or no node.  */
2215 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2216                                        struct dlm_lock_resource *res)
2217 {
2218         struct dlm_node_iter iter;
2219         int nodenum;
2220         int ret = 0;
2221         u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2222
2223         spin_lock(&dlm->spinlock);
2224         dlm_node_iter_init(dlm->domain_map, &iter);
2225         spin_unlock(&dlm->spinlock);
2226
2227         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2228                 /* do not send to self */
2229                 if (nodenum == dlm->node_num)
2230                         continue;
2231                 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2232                 if (ret < 0) {
2233                         mlog_errno(ret);
2234                         if (!dlm_is_host_down(ret))
2235                                 BUG();
2236                         /* host is down, so answer for that node would be
2237                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
2238                         ret = 0;
2239                 }
2240
2241                 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2242                         /* check to see if this master is in the recovery map */
2243                         spin_lock(&dlm->spinlock);
2244                         if (test_bit(master, dlm->recovery_map)) {
2245                                 mlog(ML_NOTICE, "%s: node %u has not seen "
2246                                      "node %u go down yet, and thinks the "
2247                                      "dead node is mastering the recovery "
2248                                      "lock.  must wait.\n", dlm->name,
2249                                      nodenum, master);
2250                                 ret = -EAGAIN;
2251                         }
2252                         spin_unlock(&dlm->spinlock);
2253                         mlog(0, "%s: reco lock master is %u\n", dlm->name, 
2254                              master);
2255                         break;
2256                 }
2257         }
2258         return ret;
2259 }
2260
2261 /*
2262  * DLM_DEREF_LOCKRES_MSG
2263  */
2264
2265 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2266 {
2267         struct dlm_deref_lockres deref;
2268         int ret = 0, r;
2269         const char *lockname;
2270         unsigned int namelen;
2271
2272         lockname = res->lockname.name;
2273         namelen = res->lockname.len;
2274         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2275
2276         mlog(0, "%s:%.*s: sending deref to %d\n",
2277              dlm->name, namelen, lockname, res->owner);
2278         memset(&deref, 0, sizeof(deref));
2279         deref.node_idx = dlm->node_num;
2280         deref.namelen = namelen;
2281         memcpy(deref.name, lockname, namelen);
2282
2283         ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2284                                  &deref, sizeof(deref), res->owner, &r);
2285         if (ret < 0)
2286                 mlog_errno(ret);
2287         else if (r < 0) {
2288                 /* BAD.  other node says I did not have a ref. */
2289                 mlog(ML_ERROR,"while dropping ref on %s:%.*s "
2290                     "(master=%u) got %d.\n", dlm->name, namelen,
2291                     lockname, res->owner, r);
2292                 dlm_print_one_lock_resource(res);
2293                 BUG();
2294         }
2295         return ret;
2296 }
2297
2298 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2299                               void **ret_data)
2300 {
2301         struct dlm_ctxt *dlm = data;
2302         struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2303         struct dlm_lock_resource *res = NULL;
2304         char *name;
2305         unsigned int namelen;
2306         int ret = -EINVAL;
2307         u8 node;
2308         unsigned int hash;
2309         struct dlm_work_item *item;
2310         int cleared = 0;
2311         int dispatch = 0;
2312
2313         if (!dlm_grab(dlm))
2314                 return 0;
2315
2316         name = deref->name;
2317         namelen = deref->namelen;
2318         node = deref->node_idx;
2319
2320         if (namelen > DLM_LOCKID_NAME_MAX) {
2321                 mlog(ML_ERROR, "Invalid name length!");
2322                 goto done;
2323         }
2324         if (deref->node_idx >= O2NM_MAX_NODES) {
2325                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2326                 goto done;
2327         }
2328
2329         hash = dlm_lockid_hash(name, namelen);
2330
2331         spin_lock(&dlm->spinlock);
2332         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2333         if (!res) {
2334                 spin_unlock(&dlm->spinlock);
2335                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2336                      dlm->name, namelen, name);
2337                 goto done;
2338         }
2339         spin_unlock(&dlm->spinlock);
2340
2341         spin_lock(&res->spinlock);
2342         if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2343                 dispatch = 1;
2344         else {
2345                 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2346                 if (test_bit(node, res->refmap)) {
2347                         dlm_lockres_clear_refmap_bit(node, res);
2348                         cleared = 1;
2349                 }
2350         }
2351         spin_unlock(&res->spinlock);
2352
2353         if (!dispatch) {
2354                 if (cleared)
2355                         dlm_lockres_calc_usage(dlm, res);
2356                 else {
2357                         mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2358                         "but it is already dropped!\n", dlm->name,
2359                         res->lockname.len, res->lockname.name, node);
2360                         dlm_print_one_lock_resource(res);
2361                 }
2362                 ret = 0;
2363                 goto done;
2364         }
2365
2366         item = kzalloc(sizeof(*item), GFP_NOFS);
2367         if (!item) {
2368                 ret = -ENOMEM;
2369                 mlog_errno(ret);
2370                 goto done;
2371         }
2372
2373         dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2374         item->u.dl.deref_res = res;
2375         item->u.dl.deref_node = node;
2376
2377         spin_lock(&dlm->work_lock);
2378         list_add_tail(&item->list, &dlm->work_list);
2379         spin_unlock(&dlm->work_lock);
2380
2381         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2382         return 0;
2383
2384 done:
2385         if (res)
2386                 dlm_lockres_put(res);
2387         dlm_put(dlm);
2388
2389         return ret;
2390 }
2391
2392 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2393 {
2394         struct dlm_ctxt *dlm;
2395         struct dlm_lock_resource *res;
2396         u8 node;
2397         u8 cleared = 0;
2398
2399         dlm = item->dlm;
2400         res = item->u.dl.deref_res;
2401         node = item->u.dl.deref_node;
2402
2403         spin_lock(&res->spinlock);
2404         BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2405         if (test_bit(node, res->refmap)) {
2406                 __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2407                 dlm_lockres_clear_refmap_bit(node, res);
2408                 cleared = 1;
2409         }
2410         spin_unlock(&res->spinlock);
2411
2412         if (cleared) {
2413                 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2414                      dlm->name, res->lockname.len, res->lockname.name, node);
2415                 dlm_lockres_calc_usage(dlm, res);
2416         } else {
2417                 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2418                      "but it is already dropped!\n", dlm->name,
2419                      res->lockname.len, res->lockname.name, node);
2420                 dlm_print_one_lock_resource(res);
2421         }
2422
2423         dlm_lockres_put(res);
2424 }
2425
2426 /* Checks whether the lockres can be migrated. Returns 0 if yes, < 0
2427  * if not. If 0, numlocks is set to the number of locks in the lockres.
2428  */
2429 static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm,
2430                                       struct dlm_lock_resource *res,
2431                                       int *numlocks)
2432 {
2433         int ret;
2434         int i;
2435         int count = 0;
2436         struct list_head *queue;
2437         struct dlm_lock *lock;
2438
2439         assert_spin_locked(&res->spinlock);
2440
2441         ret = -EINVAL;
2442         if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
2443                 mlog(0, "cannot migrate lockres with unknown owner!\n");
2444                 goto leave;
2445         }
2446
2447         if (res->owner != dlm->node_num) {
2448                 mlog(0, "cannot migrate lockres this node doesn't own!\n");
2449                 goto leave;
2450         }
2451
2452         ret = 0;
2453         queue = &res->granted;
2454         for (i = 0; i < 3; i++) {
2455                 list_for_each_entry(lock, queue, list) {
2456                         ++count;
2457                         if (lock->ml.node == dlm->node_num) {
2458                                 mlog(0, "found a lock owned by this node still "
2459                                      "on the %s queue!  will not migrate this "
2460                                      "lockres\n", (i == 0 ? "granted" :
2461                                                    (i == 1 ? "converting" :
2462                                                     "blocked")));
2463                                 ret = -ENOTEMPTY;
2464                                 goto leave;
2465                         }
2466                 }
2467                 queue++;
2468         }
2469
2470         *numlocks = count;
2471         mlog(0, "migrateable lockres having %d locks\n", *numlocks);
2472
2473 leave:
2474         return ret;
2475 }
2476
2477 /*
2478  * DLM_MIGRATE_LOCKRES
2479  */
2480
2481
2482 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2483                                struct dlm_lock_resource *res,
2484                                u8 target)
2485 {
2486         struct dlm_master_list_entry *mle = NULL;
2487         struct dlm_master_list_entry *oldmle = NULL;
2488         struct dlm_migratable_lockres *mres = NULL;
2489         int ret = 0;
2490         const char *name;
2491         unsigned int namelen;
2492         int mle_added = 0;
2493         int numlocks;
2494         int wake = 0;
2495
2496         if (!dlm_grab(dlm))
2497                 return -EINVAL;
2498
2499         name = res->lockname.name;
2500         namelen = res->lockname.len;
2501
2502         mlog(0, "migrating %.*s to %u\n", namelen, name, target);
2503
2504         /*
2505          * ensure this lockres is a proper candidate for migration
2506          */
2507         spin_lock(&res->spinlock);
2508         ret = dlm_is_lockres_migrateable(dlm, res, &numlocks);
2509         if (ret < 0) {
2510                 spin_unlock(&res->spinlock);
2511                 goto leave;
2512         }
2513         spin_unlock(&res->spinlock);
2514
2515         /* no work to do */
2516         if (numlocks == 0) {
2517                 mlog(0, "no locks were found on this lockres! done!\n");
2518                 goto leave;
2519         }
2520
2521         /*
2522          * preallocate up front
2523          * if this fails, abort
2524          */
2525
2526         ret = -ENOMEM;
2527         mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2528         if (!mres) {
2529                 mlog_errno(ret);
2530                 goto leave;
2531         }
2532
2533         mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
2534                                                                 GFP_NOFS);
2535         if (!mle) {
2536                 mlog_errno(ret);
2537                 goto leave;
2538         }
2539         ret = 0;
2540
2541         /*
2542          * find a node to migrate the lockres to
2543          */
2544
2545         mlog(0, "picking a migration node\n");
2546         spin_lock(&dlm->spinlock);
2547         /* pick a new node */
2548         if (!test_bit(target, dlm->domain_map) ||
2549             target >= O2NM_MAX_NODES) {
2550                 target = dlm_pick_migration_target(dlm, res);
2551         }
2552         mlog(0, "node %u chosen for migration\n", target);
2553
2554         if (target >= O2NM_MAX_NODES ||
2555             !test_bit(target, dlm->domain_map)) {
2556                 /* target chosen is not alive */
2557                 ret = -EINVAL;
2558         }
2559
2560         if (ret) {
2561                 spin_unlock(&dlm->spinlock);
2562                 goto fail;
2563         }
2564
2565         mlog(0, "continuing with target = %u\n", target);
2566
2567         /*
2568          * clear any existing master requests and
2569          * add the migration mle to the list
2570          */
2571         spin_lock(&dlm->master_lock);
2572         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2573                                     namelen, target, dlm->node_num);
2574         spin_unlock(&dlm->master_lock);
2575         spin_unlock(&dlm->spinlock);
2576
2577         if (ret == -EEXIST) {
2578                 mlog(0, "another process is already migrating it\n");
2579                 goto fail;
2580         }
2581         mle_added = 1;
2582
2583         /*
2584          * set the MIGRATING flag and flush asts
2585          * if we fail after this we need to re-dirty the lockres
2586          */
2587         if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2588                 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2589                      "the target went down.\n", res->lockname.len,
2590                      res->lockname.name, target);
2591                 spin_lock(&res->spinlock);
2592                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2593                 wake = 1;
2594                 spin_unlock(&res->spinlock);
2595                 ret = -EINVAL;
2596         }
2597
2598 fail:
2599         if (oldmle) {
2600                 /* master is known, detach if not already detached */
2601                 dlm_mle_detach_hb_events(dlm, oldmle);
2602                 dlm_put_mle(oldmle);
2603         }
2604
2605         if (ret < 0) {
2606                 if (mle_added) {
2607                         dlm_mle_detach_hb_events(dlm, mle);
2608                         dlm_put_mle(mle);
2609                 } else if (mle) {
2610                         kmem_cache_free(dlm_mle_cache, mle);
2611                 }
2612                 goto leave;
2613         }
2614
2615         /*
2616          * at this point, we have a migration target, an mle
2617          * in the master list, and the MIGRATING flag set on
2618          * the lockres
2619          */
2620
2621         /* now that remote nodes are spinning on the MIGRATING flag,
2622          * ensure that all assert_master work is flushed. */
2623         flush_workqueue(dlm->dlm_worker);
2624
2625         /* get an extra reference on the mle.
2626          * otherwise the assert_master from the new
2627          * master will destroy this.
2628          * also, make sure that all callers of dlm_get_mle
2629          * take both dlm->spinlock and dlm->master_lock */
2630         spin_lock(&dlm->spinlock);
2631         spin_lock(&dlm->master_lock);
2632         dlm_get_mle_inuse(mle);
2633         spin_unlock(&dlm->master_lock);
2634         spin_unlock(&dlm->spinlock);
2635
2636         /* notify new node and send all lock state */
2637         /* call send_one_lockres with migration flag.
2638          * this serves as notice to the target node that a
2639          * migration is starting. */
2640         ret = dlm_send_one_lockres(dlm, res, mres, target,
2641                                    DLM_MRES_MIGRATION);
2642
2643         if (ret < 0) {
2644                 mlog(0, "migration to node %u failed with %d\n",
2645                      target, ret);
2646                 /* migration failed, detach and clean up mle */
2647                 dlm_mle_detach_hb_events(dlm, mle);
2648                 dlm_put_mle(mle);
2649                 dlm_put_mle_inuse(mle);
2650                 spin_lock(&res->spinlock);
2651                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2652                 wake = 1;
2653                 spin_unlock(&res->spinlock);
2654                 goto leave;
2655         }
2656
2657         /* at this point, the target sends a message to all nodes,
2658          * (using dlm_do_migrate_request).  this node is skipped since
2659          * we had to put an mle in the list to begin the process.  this
2660          * node now waits for target to do an assert master.  this node
2661          * will be the last one notified, ensuring that the migration
2662          * is complete everywhere.  if the target dies while this is
2663          * going on, some nodes could potentially see the target as the
2664          * master, so it is important that my recovery finds the migration
2665          * mle and sets the master to UNKNONWN. */
2666
2667
2668         /* wait for new node to assert master */
2669         while (1) {
2670                 ret = wait_event_interruptible_timeout(mle->wq,
2671                                         (atomic_read(&mle->woken) == 1),
2672                                         msecs_to_jiffies(5000));
2673
2674                 if (ret >= 0) {
2675                         if (atomic_read(&mle->woken) == 1 ||
2676                             res->owner == target)
2677                                 break;
2678
2679                         mlog(0, "%s:%.*s: timed out during migration\n",
2680                              dlm->name, res->lockname.len, res->lockname.name);
2681                         /* avoid hang during shutdown when migrating lockres 
2682                          * to a node which also goes down */
2683                         if (dlm_is_node_dead(dlm, target)) {
2684                                 mlog(0, "%s:%.*s: expected migration "
2685                                      "target %u is no longer up, restarting\n",
2686                                      dlm->name, res->lockname.len,
2687                                      res->lockname.name, target);
2688                                 ret = -EINVAL;
2689                                 /* migration failed, detach and clean up mle */
2690                                 dlm_mle_detach_hb_events(dlm, mle);
2691                                 dlm_put_mle(mle);
2692                                 dlm_put_mle_inuse(mle);
2693                                 spin_lock(&res->spinlock);
2694                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2695                                 wake = 1;
2696                                 spin_unlock(&res->spinlock);
2697                                 goto leave;
2698                         }
2699                 } else
2700                         mlog(0, "%s:%.*s: caught signal during migration\n",
2701                              dlm->name, res->lockname.len, res->lockname.name);
2702         }
2703
2704         /* all done, set the owner, clear the flag */
2705         spin_lock(&res->spinlock);
2706         dlm_set_lockres_owner(dlm, res, target);
2707         res->state &= ~DLM_LOCK_RES_MIGRATING;
2708         dlm_remove_nonlocal_locks(dlm, res);
2709         spin_unlock(&res->spinlock);
2710         wake_up(&res->wq);
2711
2712         /* master is known, detach if not already detached */
2713         dlm_mle_detach_hb_events(dlm, mle);
2714         dlm_put_mle_inuse(mle);
2715         ret = 0;
2716
2717         dlm_lockres_calc_usage(dlm, res);
2718
2719 leave:
2720         /* re-dirty the lockres if we failed */
2721         if (ret < 0)
2722                 dlm_kick_thread(dlm, res);
2723
2724         /* wake up waiters if the MIGRATING flag got set
2725          * but migration failed */
2726         if (wake)
2727                 wake_up(&res->wq);
2728
2729         /* TODO: cleanup */
2730         if (mres)
2731                 free_page((unsigned long)mres);
2732
2733         dlm_put(dlm);
2734
2735         mlog(0, "returning %d\n", ret);
2736         return ret;
2737 }
2738
2739 #define DLM_MIGRATION_RETRY_MS  100
2740
2741 /* Should be called only after beginning the domain leave process.
2742  * There should not be any remaining locks on nonlocal lock resources,
2743  * and there should be no local locks left on locally mastered resources.
2744  *
2745  * Called with the dlm spinlock held, may drop it to do migration, but
2746  * will re-acquire before exit.
2747  *
2748  * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped */
2749 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2750 {
2751         int ret;
2752         int lock_dropped = 0;
2753         int numlocks;
2754
2755         spin_lock(&res->spinlock);
2756         if (res->owner != dlm->node_num) {
2757                 if (!__dlm_lockres_unused(res)) {
2758                         mlog(ML_ERROR, "%s:%.*s: this node is not master, "
2759                              "trying to free this but locks remain\n",
2760                              dlm->name, res->lockname.len, res->lockname.name);
2761                 }
2762                 spin_unlock(&res->spinlock);
2763                 goto leave;
2764         }
2765
2766         /* No need to migrate a lockres having no locks */
2767         ret = dlm_is_lockres_migrateable(dlm, res, &numlocks);
2768         if (ret >= 0 && numlocks == 0) {
2769                 spin_unlock(&res->spinlock);
2770                 goto leave;
2771         }
2772         spin_unlock(&res->spinlock);
2773
2774         /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2775         spin_unlock(&dlm->spinlock);
2776         lock_dropped = 1;
2777         while (1) {
2778                 ret = dlm_migrate_lockres(dlm, res, O2NM_MAX_NODES);
2779                 if (ret >= 0)
2780                         break;
2781                 if (ret == -ENOTEMPTY) {
2782                         mlog(ML_ERROR, "lockres %.*s still has local locks!\n",
2783                                 res->lockname.len, res->lockname.name);
2784                         BUG();
2785                 }
2786
2787                 mlog(0, "lockres %.*s: migrate failed, "
2788                      "retrying\n", res->lockname.len,
2789                      res->lockname.name);
2790                 msleep(DLM_MIGRATION_RETRY_MS);
2791         }
2792         spin_lock(&dlm->spinlock);
2793 leave:
2794         return lock_dropped;
2795 }
2796
2797 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2798 {
2799         int ret;
2800         spin_lock(&dlm->ast_lock);
2801         spin_lock(&lock->spinlock);
2802         ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2803         spin_unlock(&lock->spinlock);
2804         spin_unlock(&dlm->ast_lock);
2805         return ret;
2806 }
2807
2808 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2809                                      struct dlm_lock_resource *res,
2810                                      u8 mig_target)
2811 {
2812         int can_proceed;
2813         spin_lock(&res->spinlock);
2814         can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2815         spin_unlock(&res->spinlock);
2816
2817         /* target has died, so make the caller break out of the 
2818          * wait_event, but caller must recheck the domain_map */
2819         spin_lock(&dlm->spinlock);
2820         if (!test_bit(mig_target, dlm->domain_map))
2821                 can_proceed = 1;
2822         spin_unlock(&dlm->spinlock);
2823         return can_proceed;
2824 }
2825
2826 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2827                                 struct dlm_lock_resource *res)
2828 {
2829         int ret;
2830         spin_lock(&res->spinlock);
2831         ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2832         spin_unlock(&res->spinlock);
2833         return ret;
2834 }
2835
2836
2837 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2838                                        struct dlm_lock_resource *res,
2839                                        u8 target)
2840 {
2841         int ret = 0;
2842
2843         mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2844                res->lockname.len, res->lockname.name, dlm->node_num,
2845                target);
2846         /* need to set MIGRATING flag on lockres.  this is done by
2847          * ensuring that all asts have been flushed for this lockres. */
2848         spin_lock(&res->spinlock);
2849         BUG_ON(res->migration_pending);
2850         res->migration_pending = 1;
2851         /* strategy is to reserve an extra ast then release
2852          * it below, letting the release do all of the work */
2853         __dlm_lockres_reserve_ast(res);
2854         spin_unlock(&res->spinlock);
2855
2856         /* now flush all the pending asts */
2857         dlm_kick_thread(dlm, res);
2858         /* before waiting on DIRTY, block processes which may
2859          * try to dirty the lockres before MIGRATING is set */
2860         spin_lock(&res->spinlock);
2861         BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2862         res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2863         spin_unlock(&res->spinlock);
2864         /* now wait on any pending asts and the DIRTY state */
2865         wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2866         dlm_lockres_release_ast(dlm, res);
2867
2868         mlog(0, "about to wait on migration_wq, dirty=%s\n",
2869                res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2870         /* if the extra ref we just put was the final one, this
2871          * will pass thru immediately.  otherwise, we need to wait
2872          * for the last ast to finish. */
2873 again:
2874         ret = wait_event_interruptible_timeout(dlm->migration_wq,
2875                    dlm_migration_can_proceed(dlm, res, target),
2876                    msecs_to_jiffies(1000));
2877         if (ret < 0) {
2878                 mlog(0, "woken again: migrating? %s, dead? %s\n",
2879                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2880                        test_bit(target, dlm->domain_map) ? "no":"yes");
2881         } else {
2882                 mlog(0, "all is well: migrating? %s, dead? %s\n",
2883                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2884                        test_bit(target, dlm->domain_map) ? "no":"yes");
2885         }
2886         if (!dlm_migration_can_proceed(dlm, res, target)) {
2887                 mlog(0, "trying again...\n");
2888                 goto again;
2889         }
2890         /* now that we are sure the MIGRATING state is there, drop
2891          * the unneded state which blocked threads trying to DIRTY */
2892         spin_lock(&res->spinlock);
2893         BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2894         BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2895         res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2896         spin_unlock(&res->spinlock);
2897
2898         /* did the target go down or die? */
2899         spin_lock(&dlm->spinlock);
2900         if (!test_bit(target, dlm->domain_map)) {
2901                 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2902                      target);
2903                 ret = -EHOSTDOWN;
2904         }
2905         spin_unlock(&dlm->spinlock);
2906
2907         /*
2908          * at this point:
2909          *
2910          *   o the DLM_LOCK_RES_MIGRATING flag is set
2911          *   o there are no pending asts on this lockres
2912          *   o all processes trying to reserve an ast on this
2913          *     lockres must wait for the MIGRATING flag to clear
2914          */
2915         return ret;
2916 }
2917
2918 /* last step in the migration process.
2919  * original master calls this to free all of the dlm_lock
2920  * structures that used to be for other nodes. */
2921 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2922                                       struct dlm_lock_resource *res)
2923 {
2924         struct list_head *queue = &res->granted;
2925         int i, bit;
2926         struct dlm_lock *lock, *next;
2927
2928         assert_spin_locked(&res->spinlock);
2929
2930         BUG_ON(res->owner == dlm->node_num);
2931
2932         for (i=0; i<3; i++) {
2933                 list_for_each_entry_safe(lock, next, queue, list) {
2934                         if (lock->ml.node != dlm->node_num) {
2935                                 mlog(0, "putting lock for node %u\n",
2936                                      lock->ml.node);
2937                                 /* be extra careful */
2938                                 BUG_ON(!list_empty(&lock->ast_list));
2939                                 BUG_ON(!list_empty(&lock->bast_list));
2940                                 BUG_ON(lock->ast_pending);
2941                                 BUG_ON(lock->bast_pending);
2942                                 dlm_lockres_clear_refmap_bit(lock->ml.node, res);
2943                                 list_del_init(&lock->list);
2944                                 dlm_lock_put(lock);
2945                                 /* In a normal unlock, we would have added a
2946                                  * DLM_UNLOCK_FREE_LOCK action. Force it. */
2947                                 dlm_lock_put(lock);
2948                         }
2949                 }
2950                 queue++;
2951         }
2952         bit = 0;
2953         while (1) {
2954                 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2955                 if (bit >= O2NM_MAX_NODES)
2956                         break;
2957                 /* do not clear the local node reference, if there is a
2958                  * process holding this, let it drop the ref itself */
2959                 if (bit != dlm->node_num) {
2960                         mlog(0, "%s:%.*s: node %u had a ref to this "
2961                              "migrating lockres, clearing\n", dlm->name,
2962                              res->lockname.len, res->lockname.name, bit);
2963                         dlm_lockres_clear_refmap_bit(bit, res);
2964                 }
2965                 bit++;
2966         }
2967 }
2968
2969 /* for now this is not too intelligent.  we will
2970  * need stats to make this do the right thing.
2971  * this just finds the first lock on one of the
2972  * queues and uses that node as the target. */
2973 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2974                                     struct dlm_lock_resource *res)
2975 {
2976         int i;
2977         struct list_head *queue = &res->granted;
2978         struct dlm_lock *lock;
2979         int nodenum;
2980
2981         assert_spin_locked(&dlm->spinlock);
2982
2983         spin_lock(&res->spinlock);
2984         for (i=0; i<3; i++) {
2985                 list_for_each_entry(lock, queue, list) {
2986                         /* up to the caller to make sure this node
2987                          * is alive */
2988                         if (lock->ml.node != dlm->node_num) {
2989                                 spin_unlock(&res->spinlock);
2990                                 return lock->ml.node;
2991                         }
2992                 }
2993                 queue++;
2994         }
2995         spin_unlock(&res->spinlock);
2996         mlog(0, "have not found a suitable target yet! checking domain map\n");
2997
2998         /* ok now we're getting desperate.  pick anyone alive. */
2999         nodenum = -1;
3000         while (1) {
3001                 nodenum = find_next_bit(dlm->domain_map,
3002                                         O2NM_MAX_NODES, nodenum+1);
3003                 mlog(0, "found %d in domain map\n", nodenum);
3004                 if (nodenum >= O2NM_MAX_NODES)
3005                         break;
3006                 if (nodenum != dlm->node_num) {
3007                         mlog(0, "picking %d\n", nodenum);
3008                         return nodenum;
3009                 }
3010         }
3011
3012         mlog(0, "giving up.  no master to migrate to\n");
3013         return DLM_LOCK_RES_OWNER_UNKNOWN;
3014 }
3015
3016
3017
3018 /* this is called by the new master once all lockres
3019  * data has been received */
3020 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
3021                                   struct dlm_lock_resource *res,
3022                                   u8 master, u8 new_master,
3023                                   struct dlm_node_iter *iter)
3024 {
3025         struct dlm_migrate_request migrate;
3026         int ret, status = 0;
3027         int nodenum;
3028
3029         memset(&migrate, 0, sizeof(migrate));
3030         migrate.namelen = res->lockname.len;
3031         memcpy(migrate.name, res->lockname.name, migrate.namelen);
3032         migrate.new_master = new_master;
3033         migrate.master = master;
3034
3035         ret = 0;
3036
3037         /* send message to all nodes, except the master and myself */
3038         while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
3039                 if (nodenum == master ||
3040                     nodenum == new_master)
3041                         continue;
3042
3043                 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
3044                                          &migrate, sizeof(migrate), nodenum,
3045                                          &status);
3046                 if (ret < 0)
3047                         mlog_errno(ret);
3048                 else if (status < 0) {
3049                         mlog(0, "migrate request (node %u) returned %d!\n",
3050                              nodenum, status);
3051                         ret = status;
3052                 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
3053                         /* during the migration request we short-circuited
3054                          * the mastery of the lockres.  make sure we have
3055                          * a mastery ref for nodenum */
3056                         mlog(0, "%s:%.*s: need ref for node %u\n",
3057                              dlm->name, res->lockname.len, res->lockname.name,
3058                              nodenum);
3059                         spin_lock(&res->spinlock);
3060                         dlm_lockres_set_refmap_bit(nodenum, res);
3061                         spin_unlock(&res->spinlock);
3062                 }
3063         }
3064
3065         if (ret < 0)
3066                 mlog_errno(ret);
3067
3068         mlog(0, "returning ret=%d\n", ret);
3069         return ret;
3070 }
3071
3072
3073 /* if there is an existing mle for this lockres, we now know who the master is.
3074  * (the one who sent us *this* message) we can clear it up right away.
3075  * since the process that put the mle on the list still has a reference to it,
3076  * we can unhash it now, set the master and wake the process.  as a result,
3077  * we will have no mle in the list to start with.  now we can add an mle for
3078  * the migration and this should be the only one found for those scanning the
3079  * list.  */
3080 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
3081                                 void **ret_data)
3082 {
3083         struct dlm_ctxt *dlm = data;
3084         struct dlm_lock_resource *res = NULL;
3085         struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
3086         struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
3087         const char *name;
3088         unsigned int namelen, hash;
3089         int ret = 0;
3090
3091         if (!dlm_grab(dlm))
3092                 return -EINVAL;
3093
3094         name = migrate->name;
3095         namelen = migrate->namelen;
3096         hash = dlm_lockid_hash(name, namelen);
3097
3098         /* preallocate.. if this fails, abort */
3099         mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
3100                                                          GFP_NOFS);
3101
3102         if (!mle) {
3103                 ret = -ENOMEM;
3104                 goto leave;
3105         }
3106
3107         /* check for pre-existing lock */
3108         spin_lock(&dlm->spinlock);
3109         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
3110         spin_lock(&dlm->master_lock);
3111
3112         if (res) {
3113                 spin_lock(&res->spinlock);
3114                 if (res->state & DLM_LOCK_RES_RECOVERING) {
3115                         /* if all is working ok, this can only mean that we got
3116                         * a migrate request from a node that we now see as
3117                         * dead.  what can we do here?  drop it to the floor? */
3118                         spin_unlock(&res->spinlock);
3119                         mlog(ML_ERROR, "Got a migrate request, but the "
3120                              "lockres is marked as recovering!");
3121                         kmem_cache_free(dlm_mle_cache, mle);
3122                         ret = -EINVAL; /* need a better solution */
3123                         goto unlock;
3124                 }
3125                 res->state |= DLM_LOCK_RES_MIGRATING;
3126                 spin_unlock(&res->spinlock);
3127         }
3128
3129         /* ignore status.  only nonzero status would BUG. */
3130         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3131                                     name, namelen,
3132                                     migrate->new_master,
3133                                     migrate->master);
3134
3135 unlock:
3136         spin_unlock(&dlm->master_lock);
3137         spin_unlock(&dlm->spinlock);
3138
3139         if (oldmle) {
3140                 /* master is known, detach if not already detached */
3141                 dlm_mle_detach_hb_events(dlm, oldmle);
3142                 dlm_put_mle(oldmle);
3143         }
3144
3145         if (res)
3146                 dlm_lockres_put(res);
3147 leave:
3148         dlm_put(dlm);
3149         return ret;
3150 }
3151
3152 /* must be holding dlm->spinlock and dlm->master_lock
3153  * when adding a migration mle, we can clear any other mles
3154  * in the master list because we know with certainty that
3155  * the master is "master".  so we remove any old mle from
3156  * the list after setting it's master field, and then add
3157  * the new migration mle.  this way we can hold with the rule
3158  * of having only one mle for a given lock name at all times. */
3159 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3160                                  struct dlm_lock_resource *res,
3161                                  struct dlm_master_list_entry *mle,
3162                                  struct dlm_master_list_entry **oldmle,
3163                                  const char *name, unsigned int namelen,
3164                                  u8 new_master, u8 master)
3165 {
3166         int found;
3167         int ret = 0;
3168
3169         *oldmle = NULL;
3170
3171         mlog_entry_void();
3172
3173         assert_spin_locked(&dlm->spinlock);
3174         assert_spin_locked(&dlm->master_lock);
3175
3176         /* caller is responsible for any ref taken here on oldmle */
3177         found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3178         if (found) {
3179                 struct dlm_master_list_entry *tmp = *oldmle;
3180                 spin_lock(&tmp->spinlock);
3181                 if (tmp->type == DLM_MLE_MIGRATION) {
3182                         if (master == dlm->node_num) {
3183                                 /* ah another process raced me to it */
3184                                 mlog(0, "tried to migrate %.*s, but some "
3185                                      "process beat me to it\n",
3186                                      namelen, name);
3187                                 ret = -EEXIST;
3188                         } else {
3189                                 /* bad.  2 NODES are trying to migrate! */
3190                                 mlog(ML_ERROR, "migration error  mle: "
3191                                      "master=%u new_master=%u // request: "
3192                                      "master=%u new_master=%u // "
3193                                      "lockres=%.*s\n",
3194                                      tmp->master, tmp->new_master,
3195                                      master, new_master,
3196                                      namelen, name);
3197                                 BUG();
3198                         }
3199                 } else {
3200                         /* this is essentially what assert_master does */
3201                         tmp->master = master;
3202                         atomic_set(&tmp->woken, 1);
3203                         wake_up(&tmp->wq);
3204                         /* remove it from the list so that only one
3205                          * mle will be found */
3206                         list_del_init(&tmp->list);
3207                         /* this was obviously WRONG.  mle is uninited here.  should be tmp. */
3208                         __dlm_mle_detach_hb_events(dlm, tmp);
3209                         ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3210                         mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3211                             "telling master to get ref for cleared out mle "
3212                             "during migration\n", dlm->name, namelen, name,
3213                             master, new_master);
3214                 }
3215                 spin_unlock(&tmp->spinlock);
3216         }
3217
3218         /* now add a migration mle to the tail of the list */
3219         dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3220         mle->new_master = new_master;
3221         /* the new master will be sending an assert master for this.
3222          * at that point we will get the refmap reference */
3223         mle->master = master;
3224         /* do this for consistency with other mle types */
3225         set_bit(new_master, mle->maybe_map);
3226         list_add(&mle->list, &dlm->master_list);
3227
3228         return ret;
3229 }
3230
3231
3232 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3233 {
3234         struct dlm_master_list_entry *mle, *next;
3235         struct dlm_lock_resource *res;
3236         unsigned int hash;
3237
3238         mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node);
3239 top:
3240         assert_spin_locked(&dlm->spinlock);
3241
3242         /* clean the master list */
3243         spin_lock(&dlm->master_lock);
3244         list_for_each_entry_safe(mle, next, &dlm->master_list, list) {
3245                 BUG_ON(mle->type != DLM_MLE_BLOCK &&
3246                        mle->type != DLM_MLE_MASTER &&
3247                        mle->type != DLM_MLE_MIGRATION);
3248
3249                 /* MASTER mles are initiated locally.  the waiting
3250                  * process will notice the node map change
3251                  * shortly.  let that happen as normal. */
3252                 if (mle->type == DLM_MLE_MASTER)
3253                         continue;
3254
3255
3256                 /* BLOCK mles are initiated by other nodes.
3257                  * need to clean up if the dead node would have
3258                  * been the master. */
3259                 if (mle->type == DLM_MLE_BLOCK) {
3260                         int bit;
3261
3262                         spin_lock(&mle->spinlock);
3263                         bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
3264                         if (bit != dead_node) {
3265                                 mlog(0, "mle found, but dead node %u would "
3266                                      "not have been master\n", dead_node);
3267                                 spin_unlock(&mle->spinlock);
3268                         } else {
3269                                 /* must drop the refcount by one since the
3270                                  * assert_master will never arrive.  this
3271                                  * may result in the mle being unlinked and
3272                                  * freed, but there may still be a process
3273                                  * waiting in the dlmlock path which is fine. */
3274                                 mlog(0, "node %u was expected master\n",
3275                                      dead_node);
3276                                 atomic_set(&mle->woken, 1);
3277                                 spin_unlock(&mle->spinlock);
3278                                 wake_up(&mle->wq);
3279                                 /* do not need events any longer, so detach 
3280                                  * from heartbeat */
3281                                 __dlm_mle_detach_hb_events(dlm, mle);
3282                                 __dlm_put_mle(mle);
3283                         }
3284                         continue;
3285                 }
3286
3287                 /* everything else is a MIGRATION mle */
3288
3289                 /* the rule for MIGRATION mles is that the master
3290                  * becomes UNKNOWN if *either* the original or
3291                  * the new master dies.  all UNKNOWN lockreses
3292                  * are sent to whichever node becomes the recovery
3293                  * master.  the new master is responsible for
3294                  * determining if there is still a master for
3295                  * this lockres, or if he needs to take over
3296                  * mastery.  either way, this node should expect
3297                  * another message to resolve this. */
3298                 if (mle->master != dead_node &&
3299                     mle->new_master != dead_node)
3300                         continue;
3301
3302                 /* if we have reached this point, this mle needs to
3303                  * be removed from the list and freed. */
3304
3305                 /* remove from the list early.  NOTE: unlinking
3306                  * list_head while in list_for_each_safe */
3307                 __dlm_mle_detach_hb_events(dlm, mle);
3308                 spin_lock(&mle->spinlock);
3309                 list_del_init(&mle->list);
3310                 atomic_set(&mle->woken, 1);
3311                 spin_unlock(&mle->spinlock);
3312                 wake_up(&mle->wq);
3313
3314                 mlog(0, "%s: node %u died during migration from "
3315                      "%u to %u!\n", dlm->name, dead_node,
3316                      mle->master, mle->new_master);
3317                 /* if there is a lockres associated with this
3318                  * mle, find it and set its owner to UNKNOWN */
3319                 hash = dlm_lockid_hash(mle->u.name.name, mle->u.name.len);
3320                 res = __dlm_lookup_lockres(dlm, mle->u.name.name,
3321                                            mle->u.name.len, hash);
3322                 if (res) {
3323                         /* unfortunately if we hit this rare case, our
3324                          * lock ordering is messed.  we need to drop
3325                          * the master lock so that we can take the
3326                          * lockres lock, meaning that we will have to
3327                          * restart from the head of list. */
3328                         spin_unlock(&dlm->master_lock);
3329
3330                         /* move lockres onto recovery list */
3331                         spin_lock(&res->spinlock);
3332                         dlm_set_lockres_owner(dlm, res,
3333                                         DLM_LOCK_RES_OWNER_UNKNOWN);
3334                         dlm_move_lockres_to_recovery_list(dlm, res);
3335                         spin_unlock(&res->spinlock);
3336                         dlm_lockres_put(res);
3337
3338                         /* about to get rid of mle, detach from heartbeat */
3339                         __dlm_mle_detach_hb_events(dlm, mle);
3340
3341                         /* dump the mle */
3342                         spin_lock(&dlm->master_lock);
3343                         __dlm_put_mle(mle);
3344                         spin_unlock(&dlm->master_lock);
3345
3346                         /* restart */
3347                         goto top;
3348                 }
3349
3350                 /* this may be the last reference */
3351                 __dlm_put_mle(mle);
3352         }
3353         spin_unlock(&dlm->master_lock);
3354 }
3355
3356
3357 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3358                          u8 old_master)
3359 {
3360         struct dlm_node_iter iter;
3361         int ret = 0;
3362
3363         spin_lock(&dlm->spinlock);
3364         dlm_node_iter_init(dlm->domain_map, &iter);
3365         clear_bit(old_master, iter.node_map);
3366         clear_bit(dlm->node_num, iter.node_map);
3367         spin_unlock(&dlm->spinlock);
3368
3369         /* ownership of the lockres is changing.  account for the
3370          * mastery reference here since old_master will briefly have
3371          * a reference after the migration completes */
3372         spin_lock(&res->spinlock);
3373         dlm_lockres_set_refmap_bit(old_master, res);
3374         spin_unlock(&res->spinlock);
3375
3376         mlog(0, "now time to do a migrate request to other nodes\n");
3377         ret = dlm_do_migrate_request(dlm, res, old_master,
3378                                      dlm->node_num, &iter);
3379         if (ret < 0) {
3380                 mlog_errno(ret);
3381                 goto leave;
3382         }
3383
3384         mlog(0, "doing assert master of %.*s to all except the original node\n",
3385              res->lockname.len, res->lockname.name);
3386         /* this call now finishes out the nodemap
3387          * even if one or more nodes die */
3388         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3389                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3390         if (ret < 0) {
3391                 /* no longer need to retry.  all living nodes contacted. */
3392                 mlog_errno(ret);
3393                 ret = 0;
3394         }
3395
3396         memset(iter.node_map, 0, sizeof(iter.node_map));
3397         set_bit(old_master, iter.node_map);
3398         mlog(0, "doing assert master of %.*s back to %u\n",
3399              res->lockname.len, res->lockname.name, old_master);
3400         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3401                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3402         if (ret < 0) {
3403                 mlog(0, "assert master to original master failed "
3404                      "with %d.\n", ret);
3405                 /* the only nonzero status here would be because of
3406                  * a dead original node.  we're done. */
3407                 ret = 0;
3408         }
3409
3410         /* all done, set the owner, clear the flag */
3411         spin_lock(&res->spinlock);
3412         dlm_set_lockres_owner(dlm, res, dlm->node_num);
3413         res->state &= ~DLM_LOCK_RES_MIGRATING;
3414         spin_unlock(&res->spinlock);
3415         /* re-dirty it on the new master */
3416         dlm_kick_thread(dlm, res);
3417         wake_up(&res->wq);
3418 leave:
3419         return ret;
3420 }
3421
3422 /*
3423  * LOCKRES AST REFCOUNT
3424  * this is integral to migration
3425  */
3426
3427 /* for future intent to call an ast, reserve one ahead of time.
3428  * this should be called only after waiting on the lockres
3429  * with dlm_wait_on_lockres, and while still holding the
3430  * spinlock after the call. */
3431 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3432 {
3433         assert_spin_locked(&res->spinlock);
3434         if (res->state & DLM_LOCK_RES_MIGRATING) {
3435                 __dlm_print_one_lock_resource(res);
3436         }
3437         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3438
3439         atomic_inc(&res->asts_reserved);
3440 }
3441
3442 /*
3443  * used to drop the reserved ast, either because it went unused,
3444  * or because the ast/bast was actually called.
3445  *
3446  * also, if there is a pending migration on this lockres,
3447  * and this was the last pending ast on the lockres,
3448  * atomically set the MIGRATING flag before we drop the lock.
3449  * this is how we ensure that migration can proceed with no
3450  * asts in progress.  note that it is ok if the state of the
3451  * queues is such that a lock should be granted in the future
3452  * or that a bast should be fired, because the new master will
3453  * shuffle the lists on this lockres as soon as it is migrated.
3454  */
3455 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3456                              struct dlm_lock_resource *res)
3457 {
3458         if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3459                 return;
3460
3461         if (!res->migration_pending) {
3462                 spin_unlock(&res->spinlock);
3463                 return;
3464         }
3465
3466         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3467         res->migration_pending = 0;
3468         res->state |= DLM_LOCK_RES_MIGRATING;
3469         spin_unlock(&res->spinlock);
3470         wake_up(&res->wq);
3471         wake_up(&dlm->migration_wq);
3472 }