GNU Linux-libre 6.1.86-gnu
[releases.git] / fs / ocfs2 / dlm / dlmmaster.c
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * dlmmod.c
4  *
5  * standalone DLM module
6  *
7  * Copyright (C) 2004 Oracle.  All rights reserved.
8  */
9
10
11 #include <linux/module.h>
12 #include <linux/fs.h>
13 #include <linux/types.h>
14 #include <linux/slab.h>
15 #include <linux/highmem.h>
16 #include <linux/init.h>
17 #include <linux/sysctl.h>
18 #include <linux/random.h>
19 #include <linux/blkdev.h>
20 #include <linux/socket.h>
21 #include <linux/inet.h>
22 #include <linux/spinlock.h>
23 #include <linux/delay.h>
24
25
26 #include "../cluster/heartbeat.h"
27 #include "../cluster/nodemanager.h"
28 #include "../cluster/tcp.h"
29
30 #include "dlmapi.h"
31 #include "dlmcommon.h"
32 #include "dlmdomain.h"
33 #include "dlmdebug.h"
34
35 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
36 #include "../cluster/masklog.h"
37
38 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
39                               struct dlm_master_list_entry *mle,
40                               struct o2nm_node *node,
41                               int idx);
42 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
43                             struct dlm_master_list_entry *mle,
44                             struct o2nm_node *node,
45                             int idx);
46
47 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
48 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
49                                 struct dlm_lock_resource *res,
50                                 void *nodemap, u32 flags);
51 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
52
53 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
54                                 struct dlm_master_list_entry *mle,
55                                 const char *name,
56                                 unsigned int namelen)
57 {
58         if (dlm != mle->dlm)
59                 return 0;
60
61         if (namelen != mle->mnamelen ||
62             memcmp(name, mle->mname, namelen) != 0)
63                 return 0;
64
65         return 1;
66 }
67
68 static struct kmem_cache *dlm_lockres_cache;
69 static struct kmem_cache *dlm_lockname_cache;
70 static struct kmem_cache *dlm_mle_cache;
71
72 static void dlm_mle_release(struct kref *kref);
73 static void dlm_init_mle(struct dlm_master_list_entry *mle,
74                         enum dlm_mle_type type,
75                         struct dlm_ctxt *dlm,
76                         struct dlm_lock_resource *res,
77                         const char *name,
78                         unsigned int namelen);
79 static void dlm_put_mle(struct dlm_master_list_entry *mle);
80 static void __dlm_put_mle(struct dlm_master_list_entry *mle);
81 static int dlm_find_mle(struct dlm_ctxt *dlm,
82                         struct dlm_master_list_entry **mle,
83                         char *name, unsigned int namelen);
84
85 static int dlm_do_master_request(struct dlm_lock_resource *res,
86                                  struct dlm_master_list_entry *mle, int to);
87
88
89 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
90                                      struct dlm_lock_resource *res,
91                                      struct dlm_master_list_entry *mle,
92                                      int *blocked);
93 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
94                                     struct dlm_lock_resource *res,
95                                     struct dlm_master_list_entry *mle,
96                                     int blocked);
97 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
98                                  struct dlm_lock_resource *res,
99                                  struct dlm_master_list_entry *mle,
100                                  struct dlm_master_list_entry **oldmle,
101                                  const char *name, unsigned int namelen,
102                                  u8 new_master, u8 master);
103
104 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
105                                     struct dlm_lock_resource *res);
106 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
107                                       struct dlm_lock_resource *res);
108 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
109                                        struct dlm_lock_resource *res,
110                                        u8 target);
111 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
112                                        struct dlm_lock_resource *res);
113
114
115 int dlm_is_host_down(int errno)
116 {
117         switch (errno) {
118                 case -EBADF:
119                 case -ECONNREFUSED:
120                 case -ENOTCONN:
121                 case -ECONNRESET:
122                 case -EPIPE:
123                 case -EHOSTDOWN:
124                 case -EHOSTUNREACH:
125                 case -ETIMEDOUT:
126                 case -ECONNABORTED:
127                 case -ENETDOWN:
128                 case -ENETUNREACH:
129                 case -ENETRESET:
130                 case -ESHUTDOWN:
131                 case -ENOPROTOOPT:
132                 case -EINVAL:   /* if returned from our tcp code,
133                                    this means there is no socket */
134                         return 1;
135         }
136         return 0;
137 }
138
139
140 /*
141  * MASTER LIST FUNCTIONS
142  */
143
144
145 /*
146  * regarding master list entries and heartbeat callbacks:
147  *
148  * in order to avoid sleeping and allocation that occurs in
149  * heartbeat, master list entries are simply attached to the
150  * dlm's established heartbeat callbacks.  the mle is attached
151  * when it is created, and since the dlm->spinlock is held at
152  * that time, any heartbeat event will be properly discovered
153  * by the mle.  the mle needs to be detached from the
154  * dlm->mle_hb_events list as soon as heartbeat events are no
155  * longer useful to the mle, and before the mle is freed.
156  *
157  * as a general rule, heartbeat events are no longer needed by
158  * the mle once an "answer" regarding the lock master has been
159  * received.
160  */
161 static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
162                                               struct dlm_master_list_entry *mle)
163 {
164         assert_spin_locked(&dlm->spinlock);
165
166         list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
167 }
168
169
170 static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
171                                               struct dlm_master_list_entry *mle)
172 {
173         if (!list_empty(&mle->hb_events))
174                 list_del_init(&mle->hb_events);
175 }
176
177
178 static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
179                                             struct dlm_master_list_entry *mle)
180 {
181         spin_lock(&dlm->spinlock);
182         __dlm_mle_detach_hb_events(dlm, mle);
183         spin_unlock(&dlm->spinlock);
184 }
185
186 static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
187 {
188         struct dlm_ctxt *dlm;
189         dlm = mle->dlm;
190
191         assert_spin_locked(&dlm->spinlock);
192         assert_spin_locked(&dlm->master_lock);
193         mle->inuse++;
194         kref_get(&mle->mle_refs);
195 }
196
197 static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
198 {
199         struct dlm_ctxt *dlm;
200         dlm = mle->dlm;
201
202         spin_lock(&dlm->spinlock);
203         spin_lock(&dlm->master_lock);
204         mle->inuse--;
205         __dlm_put_mle(mle);
206         spin_unlock(&dlm->master_lock);
207         spin_unlock(&dlm->spinlock);
208
209 }
210
211 /* remove from list and free */
212 static void __dlm_put_mle(struct dlm_master_list_entry *mle)
213 {
214         struct dlm_ctxt *dlm;
215         dlm = mle->dlm;
216
217         assert_spin_locked(&dlm->spinlock);
218         assert_spin_locked(&dlm->master_lock);
219         if (!kref_read(&mle->mle_refs)) {
220                 /* this may or may not crash, but who cares.
221                  * it's a BUG. */
222                 mlog(ML_ERROR, "bad mle: %p\n", mle);
223                 dlm_print_one_mle(mle);
224                 BUG();
225         } else
226                 kref_put(&mle->mle_refs, dlm_mle_release);
227 }
228
229
230 /* must not have any spinlocks coming in */
231 static void dlm_put_mle(struct dlm_master_list_entry *mle)
232 {
233         struct dlm_ctxt *dlm;
234         dlm = mle->dlm;
235
236         spin_lock(&dlm->spinlock);
237         spin_lock(&dlm->master_lock);
238         __dlm_put_mle(mle);
239         spin_unlock(&dlm->master_lock);
240         spin_unlock(&dlm->spinlock);
241 }
242
243 static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
244 {
245         kref_get(&mle->mle_refs);
246 }
247
248 static void dlm_init_mle(struct dlm_master_list_entry *mle,
249                         enum dlm_mle_type type,
250                         struct dlm_ctxt *dlm,
251                         struct dlm_lock_resource *res,
252                         const char *name,
253                         unsigned int namelen)
254 {
255         assert_spin_locked(&dlm->spinlock);
256
257         mle->dlm = dlm;
258         mle->type = type;
259         INIT_HLIST_NODE(&mle->master_hash_node);
260         INIT_LIST_HEAD(&mle->hb_events);
261         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
262         spin_lock_init(&mle->spinlock);
263         init_waitqueue_head(&mle->wq);
264         atomic_set(&mle->woken, 0);
265         kref_init(&mle->mle_refs);
266         memset(mle->response_map, 0, sizeof(mle->response_map));
267         mle->master = O2NM_MAX_NODES;
268         mle->new_master = O2NM_MAX_NODES;
269         mle->inuse = 0;
270
271         BUG_ON(mle->type != DLM_MLE_BLOCK &&
272                mle->type != DLM_MLE_MASTER &&
273                mle->type != DLM_MLE_MIGRATION);
274
275         if (mle->type == DLM_MLE_MASTER) {
276                 BUG_ON(!res);
277                 mle->mleres = res;
278                 memcpy(mle->mname, res->lockname.name, res->lockname.len);
279                 mle->mnamelen = res->lockname.len;
280                 mle->mnamehash = res->lockname.hash;
281         } else {
282                 BUG_ON(!name);
283                 mle->mleres = NULL;
284                 memcpy(mle->mname, name, namelen);
285                 mle->mnamelen = namelen;
286                 mle->mnamehash = dlm_lockid_hash(name, namelen);
287         }
288
289         atomic_inc(&dlm->mle_tot_count[mle->type]);
290         atomic_inc(&dlm->mle_cur_count[mle->type]);
291
292         /* copy off the node_map and register hb callbacks on our copy */
293         memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
294         memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
295         clear_bit(dlm->node_num, mle->vote_map);
296         clear_bit(dlm->node_num, mle->node_map);
297
298         /* attach the mle to the domain node up/down events */
299         __dlm_mle_attach_hb_events(dlm, mle);
300 }
301
302 void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
303 {
304         assert_spin_locked(&dlm->spinlock);
305         assert_spin_locked(&dlm->master_lock);
306
307         if (!hlist_unhashed(&mle->master_hash_node))
308                 hlist_del_init(&mle->master_hash_node);
309 }
310
311 void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
312 {
313         struct hlist_head *bucket;
314
315         assert_spin_locked(&dlm->master_lock);
316
317         bucket = dlm_master_hash(dlm, mle->mnamehash);
318         hlist_add_head(&mle->master_hash_node, bucket);
319 }
320
321 /* returns 1 if found, 0 if not */
322 static int dlm_find_mle(struct dlm_ctxt *dlm,
323                         struct dlm_master_list_entry **mle,
324                         char *name, unsigned int namelen)
325 {
326         struct dlm_master_list_entry *tmpmle;
327         struct hlist_head *bucket;
328         unsigned int hash;
329
330         assert_spin_locked(&dlm->master_lock);
331
332         hash = dlm_lockid_hash(name, namelen);
333         bucket = dlm_master_hash(dlm, hash);
334         hlist_for_each_entry(tmpmle, bucket, master_hash_node) {
335                 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
336                         continue;
337                 dlm_get_mle(tmpmle);
338                 *mle = tmpmle;
339                 return 1;
340         }
341         return 0;
342 }
343
344 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
345 {
346         struct dlm_master_list_entry *mle;
347
348         assert_spin_locked(&dlm->spinlock);
349
350         list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
351                 if (node_up)
352                         dlm_mle_node_up(dlm, mle, NULL, idx);
353                 else
354                         dlm_mle_node_down(dlm, mle, NULL, idx);
355         }
356 }
357
358 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
359                               struct dlm_master_list_entry *mle,
360                               struct o2nm_node *node, int idx)
361 {
362         spin_lock(&mle->spinlock);
363
364         if (!test_bit(idx, mle->node_map))
365                 mlog(0, "node %u already removed from nodemap!\n", idx);
366         else
367                 clear_bit(idx, mle->node_map);
368
369         spin_unlock(&mle->spinlock);
370 }
371
372 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
373                             struct dlm_master_list_entry *mle,
374                             struct o2nm_node *node, int idx)
375 {
376         spin_lock(&mle->spinlock);
377
378         if (test_bit(idx, mle->node_map))
379                 mlog(0, "node %u already in node map!\n", idx);
380         else
381                 set_bit(idx, mle->node_map);
382
383         spin_unlock(&mle->spinlock);
384 }
385
386
387 int dlm_init_mle_cache(void)
388 {
389         dlm_mle_cache = kmem_cache_create("o2dlm_mle",
390                                           sizeof(struct dlm_master_list_entry),
391                                           0, SLAB_HWCACHE_ALIGN,
392                                           NULL);
393         if (dlm_mle_cache == NULL)
394                 return -ENOMEM;
395         return 0;
396 }
397
398 void dlm_destroy_mle_cache(void)
399 {
400         kmem_cache_destroy(dlm_mle_cache);
401 }
402
403 static void dlm_mle_release(struct kref *kref)
404 {
405         struct dlm_master_list_entry *mle;
406         struct dlm_ctxt *dlm;
407
408         mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
409         dlm = mle->dlm;
410
411         assert_spin_locked(&dlm->spinlock);
412         assert_spin_locked(&dlm->master_lock);
413
414         mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
415              mle->type);
416
417         /* remove from list if not already */
418         __dlm_unlink_mle(dlm, mle);
419
420         /* detach the mle from the domain node up/down events */
421         __dlm_mle_detach_hb_events(dlm, mle);
422
423         atomic_dec(&dlm->mle_cur_count[mle->type]);
424
425         /* NOTE: kfree under spinlock here.
426          * if this is bad, we can move this to a freelist. */
427         kmem_cache_free(dlm_mle_cache, mle);
428 }
429
430
431 /*
432  * LOCK RESOURCE FUNCTIONS
433  */
434
435 int dlm_init_master_caches(void)
436 {
437         dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
438                                               sizeof(struct dlm_lock_resource),
439                                               0, SLAB_HWCACHE_ALIGN, NULL);
440         if (!dlm_lockres_cache)
441                 goto bail;
442
443         dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
444                                                DLM_LOCKID_NAME_MAX, 0,
445                                                SLAB_HWCACHE_ALIGN, NULL);
446         if (!dlm_lockname_cache)
447                 goto bail;
448
449         return 0;
450 bail:
451         dlm_destroy_master_caches();
452         return -ENOMEM;
453 }
454
455 void dlm_destroy_master_caches(void)
456 {
457         kmem_cache_destroy(dlm_lockname_cache);
458         dlm_lockname_cache = NULL;
459
460         kmem_cache_destroy(dlm_lockres_cache);
461         dlm_lockres_cache = NULL;
462 }
463
464 static void dlm_lockres_release(struct kref *kref)
465 {
466         struct dlm_lock_resource *res;
467         struct dlm_ctxt *dlm;
468
469         res = container_of(kref, struct dlm_lock_resource, refs);
470         dlm = res->dlm;
471
472         /* This should not happen -- all lockres' have a name
473          * associated with them at init time. */
474         BUG_ON(!res->lockname.name);
475
476         mlog(0, "destroying lockres %.*s\n", res->lockname.len,
477              res->lockname.name);
478
479         atomic_dec(&dlm->res_cur_count);
480
481         if (!hlist_unhashed(&res->hash_node) ||
482             !list_empty(&res->granted) ||
483             !list_empty(&res->converting) ||
484             !list_empty(&res->blocked) ||
485             !list_empty(&res->dirty) ||
486             !list_empty(&res->recovering) ||
487             !list_empty(&res->purge)) {
488                 mlog(ML_ERROR,
489                      "Going to BUG for resource %.*s."
490                      "  We're on a list! [%c%c%c%c%c%c%c]\n",
491                      res->lockname.len, res->lockname.name,
492                      !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
493                      !list_empty(&res->granted) ? 'G' : ' ',
494                      !list_empty(&res->converting) ? 'C' : ' ',
495                      !list_empty(&res->blocked) ? 'B' : ' ',
496                      !list_empty(&res->dirty) ? 'D' : ' ',
497                      !list_empty(&res->recovering) ? 'R' : ' ',
498                      !list_empty(&res->purge) ? 'P' : ' ');
499
500                 dlm_print_one_lock_resource(res);
501         }
502
503         /* By the time we're ready to blow this guy away, we shouldn't
504          * be on any lists. */
505         BUG_ON(!hlist_unhashed(&res->hash_node));
506         BUG_ON(!list_empty(&res->granted));
507         BUG_ON(!list_empty(&res->converting));
508         BUG_ON(!list_empty(&res->blocked));
509         BUG_ON(!list_empty(&res->dirty));
510         BUG_ON(!list_empty(&res->recovering));
511         BUG_ON(!list_empty(&res->purge));
512
513         kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
514
515         kmem_cache_free(dlm_lockres_cache, res);
516 }
517
518 void dlm_lockres_put(struct dlm_lock_resource *res)
519 {
520         kref_put(&res->refs, dlm_lockres_release);
521 }
522
523 static void dlm_init_lockres(struct dlm_ctxt *dlm,
524                              struct dlm_lock_resource *res,
525                              const char *name, unsigned int namelen)
526 {
527         char *qname;
528
529         /* If we memset here, we lose our reference to the kmalloc'd
530          * res->lockname.name, so be sure to init every field
531          * correctly! */
532
533         qname = (char *) res->lockname.name;
534         memcpy(qname, name, namelen);
535
536         res->lockname.len = namelen;
537         res->lockname.hash = dlm_lockid_hash(name, namelen);
538
539         init_waitqueue_head(&res->wq);
540         spin_lock_init(&res->spinlock);
541         INIT_HLIST_NODE(&res->hash_node);
542         INIT_LIST_HEAD(&res->granted);
543         INIT_LIST_HEAD(&res->converting);
544         INIT_LIST_HEAD(&res->blocked);
545         INIT_LIST_HEAD(&res->dirty);
546         INIT_LIST_HEAD(&res->recovering);
547         INIT_LIST_HEAD(&res->purge);
548         INIT_LIST_HEAD(&res->tracking);
549         atomic_set(&res->asts_reserved, 0);
550         res->migration_pending = 0;
551         res->inflight_locks = 0;
552         res->inflight_assert_workers = 0;
553
554         res->dlm = dlm;
555
556         kref_init(&res->refs);
557
558         atomic_inc(&dlm->res_tot_count);
559         atomic_inc(&dlm->res_cur_count);
560
561         /* just for consistency */
562         spin_lock(&res->spinlock);
563         dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
564         spin_unlock(&res->spinlock);
565
566         res->state = DLM_LOCK_RES_IN_PROGRESS;
567
568         res->last_used = 0;
569
570         spin_lock(&dlm->track_lock);
571         list_add_tail(&res->tracking, &dlm->tracking_list);
572         spin_unlock(&dlm->track_lock);
573
574         memset(res->lvb, 0, DLM_LVB_LEN);
575         memset(res->refmap, 0, sizeof(res->refmap));
576 }
577
578 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
579                                    const char *name,
580                                    unsigned int namelen)
581 {
582         struct dlm_lock_resource *res = NULL;
583
584         res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
585         if (!res)
586                 goto error;
587
588         res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
589         if (!res->lockname.name)
590                 goto error;
591
592         dlm_init_lockres(dlm, res, name, namelen);
593         return res;
594
595 error:
596         if (res)
597                 kmem_cache_free(dlm_lockres_cache, res);
598         return NULL;
599 }
600
601 void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm,
602                                 struct dlm_lock_resource *res, int bit)
603 {
604         assert_spin_locked(&res->spinlock);
605
606         mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len,
607              res->lockname.name, bit, __builtin_return_address(0));
608
609         set_bit(bit, res->refmap);
610 }
611
612 void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm,
613                                   struct dlm_lock_resource *res, int bit)
614 {
615         assert_spin_locked(&res->spinlock);
616
617         mlog(0, "res %.*s, clr node %u, %ps()\n", res->lockname.len,
618              res->lockname.name, bit, __builtin_return_address(0));
619
620         clear_bit(bit, res->refmap);
621 }
622
623 static void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
624                                    struct dlm_lock_resource *res)
625 {
626         res->inflight_locks++;
627
628         mlog(0, "%s: res %.*s, inflight++: now %u, %ps()\n", dlm->name,
629              res->lockname.len, res->lockname.name, res->inflight_locks,
630              __builtin_return_address(0));
631 }
632
633 void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
634                                    struct dlm_lock_resource *res)
635 {
636         assert_spin_locked(&res->spinlock);
637         __dlm_lockres_grab_inflight_ref(dlm, res);
638 }
639
640 void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
641                                    struct dlm_lock_resource *res)
642 {
643         assert_spin_locked(&res->spinlock);
644
645         BUG_ON(res->inflight_locks == 0);
646
647         res->inflight_locks--;
648
649         mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm->name,
650              res->lockname.len, res->lockname.name, res->inflight_locks,
651              __builtin_return_address(0));
652
653         wake_up(&res->wq);
654 }
655
656 void __dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm,
657                 struct dlm_lock_resource *res)
658 {
659         assert_spin_locked(&res->spinlock);
660         res->inflight_assert_workers++;
661         mlog(0, "%s:%.*s: inflight assert worker++: now %u\n",
662                         dlm->name, res->lockname.len, res->lockname.name,
663                         res->inflight_assert_workers);
664 }
665
666 static void __dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
667                 struct dlm_lock_resource *res)
668 {
669         assert_spin_locked(&res->spinlock);
670         BUG_ON(res->inflight_assert_workers == 0);
671         res->inflight_assert_workers--;
672         mlog(0, "%s:%.*s: inflight assert worker--: now %u\n",
673                         dlm->name, res->lockname.len, res->lockname.name,
674                         res->inflight_assert_workers);
675 }
676
677 static void dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
678                 struct dlm_lock_resource *res)
679 {
680         spin_lock(&res->spinlock);
681         __dlm_lockres_drop_inflight_worker(dlm, res);
682         spin_unlock(&res->spinlock);
683 }
684
685 /*
686  * lookup a lock resource by name.
687  * may already exist in the hashtable.
688  * lockid is null terminated
689  *
690  * if not, allocate enough for the lockres and for
691  * the temporary structure used in doing the mastering.
692  *
693  * also, do a lookup in the dlm->master_list to see
694  * if another node has begun mastering the same lock.
695  * if so, there should be a block entry in there
696  * for this name, and we should *not* attempt to master
697  * the lock here.   need to wait around for that node
698  * to assert_master (or die).
699  *
700  */
701 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
702                                           const char *lockid,
703                                           int namelen,
704                                           int flags)
705 {
706         struct dlm_lock_resource *tmpres=NULL, *res=NULL;
707         struct dlm_master_list_entry *mle = NULL;
708         struct dlm_master_list_entry *alloc_mle = NULL;
709         int blocked = 0;
710         int ret, nodenum;
711         struct dlm_node_iter iter;
712         unsigned int hash;
713         int tries = 0;
714         int bit, wait_on_recovery = 0;
715
716         BUG_ON(!lockid);
717
718         hash = dlm_lockid_hash(lockid, namelen);
719
720         mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
721
722 lookup:
723         spin_lock(&dlm->spinlock);
724         tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
725         if (tmpres) {
726                 spin_unlock(&dlm->spinlock);
727                 spin_lock(&tmpres->spinlock);
728
729                 /*
730                  * Right after dlm spinlock was released, dlm_thread could have
731                  * purged the lockres. Check if lockres got unhashed. If so
732                  * start over.
733                  */
734                 if (hlist_unhashed(&tmpres->hash_node)) {
735                         spin_unlock(&tmpres->spinlock);
736                         dlm_lockres_put(tmpres);
737                         tmpres = NULL;
738                         goto lookup;
739                 }
740
741                 /* Wait on the thread that is mastering the resource */
742                 if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
743                         __dlm_wait_on_lockres(tmpres);
744                         BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
745                         spin_unlock(&tmpres->spinlock);
746                         dlm_lockres_put(tmpres);
747                         tmpres = NULL;
748                         goto lookup;
749                 }
750
751                 /* Wait on the resource purge to complete before continuing */
752                 if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) {
753                         BUG_ON(tmpres->owner == dlm->node_num);
754                         __dlm_wait_on_lockres_flags(tmpres,
755                                                     DLM_LOCK_RES_DROPPING_REF);
756                         spin_unlock(&tmpres->spinlock);
757                         dlm_lockres_put(tmpres);
758                         tmpres = NULL;
759                         goto lookup;
760                 }
761
762                 /* Grab inflight ref to pin the resource */
763                 dlm_lockres_grab_inflight_ref(dlm, tmpres);
764
765                 spin_unlock(&tmpres->spinlock);
766                 if (res) {
767                         spin_lock(&dlm->track_lock);
768                         if (!list_empty(&res->tracking))
769                                 list_del_init(&res->tracking);
770                         else
771                                 mlog(ML_ERROR, "Resource %.*s not "
772                                                 "on the Tracking list\n",
773                                                 res->lockname.len,
774                                                 res->lockname.name);
775                         spin_unlock(&dlm->track_lock);
776                         dlm_lockres_put(res);
777                 }
778                 res = tmpres;
779                 goto leave;
780         }
781
782         if (!res) {
783                 spin_unlock(&dlm->spinlock);
784                 mlog(0, "allocating a new resource\n");
785                 /* nothing found and we need to allocate one. */
786                 alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
787                 if (!alloc_mle)
788                         goto leave;
789                 res = dlm_new_lockres(dlm, lockid, namelen);
790                 if (!res)
791                         goto leave;
792                 goto lookup;
793         }
794
795         mlog(0, "no lockres found, allocated our own: %p\n", res);
796
797         if (flags & LKM_LOCAL) {
798                 /* caller knows it's safe to assume it's not mastered elsewhere
799                  * DONE!  return right away */
800                 spin_lock(&res->spinlock);
801                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
802                 __dlm_insert_lockres(dlm, res);
803                 dlm_lockres_grab_inflight_ref(dlm, res);
804                 spin_unlock(&res->spinlock);
805                 spin_unlock(&dlm->spinlock);
806                 /* lockres still marked IN_PROGRESS */
807                 goto wake_waiters;
808         }
809
810         /* check master list to see if another node has started mastering it */
811         spin_lock(&dlm->master_lock);
812
813         /* if we found a block, wait for lock to be mastered by another node */
814         blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
815         if (blocked) {
816                 int mig;
817                 if (mle->type == DLM_MLE_MASTER) {
818                         mlog(ML_ERROR, "master entry for nonexistent lock!\n");
819                         BUG();
820                 }
821                 mig = (mle->type == DLM_MLE_MIGRATION);
822                 /* if there is a migration in progress, let the migration
823                  * finish before continuing.  we can wait for the absence
824                  * of the MIGRATION mle: either the migrate finished or
825                  * one of the nodes died and the mle was cleaned up.
826                  * if there is a BLOCK here, but it already has a master
827                  * set, we are too late.  the master does not have a ref
828                  * for us in the refmap.  detach the mle and drop it.
829                  * either way, go back to the top and start over. */
830                 if (mig || mle->master != O2NM_MAX_NODES) {
831                         BUG_ON(mig && mle->master == dlm->node_num);
832                         /* we arrived too late.  the master does not
833                          * have a ref for us. retry. */
834                         mlog(0, "%s:%.*s: late on %s\n",
835                              dlm->name, namelen, lockid,
836                              mig ?  "MIGRATION" : "BLOCK");
837                         spin_unlock(&dlm->master_lock);
838                         spin_unlock(&dlm->spinlock);
839
840                         /* master is known, detach */
841                         if (!mig)
842                                 dlm_mle_detach_hb_events(dlm, mle);
843                         dlm_put_mle(mle);
844                         mle = NULL;
845                         /* this is lame, but we can't wait on either
846                          * the mle or lockres waitqueue here */
847                         if (mig)
848                                 msleep(100);
849                         goto lookup;
850                 }
851         } else {
852                 /* go ahead and try to master lock on this node */
853                 mle = alloc_mle;
854                 /* make sure this does not get freed below */
855                 alloc_mle = NULL;
856                 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
857                 set_bit(dlm->node_num, mle->maybe_map);
858                 __dlm_insert_mle(dlm, mle);
859
860                 /* still holding the dlm spinlock, check the recovery map
861                  * to see if there are any nodes that still need to be
862                  * considered.  these will not appear in the mle nodemap
863                  * but they might own this lockres.  wait on them. */
864                 bit = find_first_bit(dlm->recovery_map, O2NM_MAX_NODES);
865                 if (bit < O2NM_MAX_NODES) {
866                         mlog(0, "%s: res %.*s, At least one node (%d) "
867                              "to recover before lock mastery can begin\n",
868                              dlm->name, namelen, (char *)lockid, bit);
869                         wait_on_recovery = 1;
870                 }
871         }
872
873         /* at this point there is either a DLM_MLE_BLOCK or a
874          * DLM_MLE_MASTER on the master list, so it's safe to add the
875          * lockres to the hashtable.  anyone who finds the lock will
876          * still have to wait on the IN_PROGRESS. */
877
878         /* finally add the lockres to its hash bucket */
879         __dlm_insert_lockres(dlm, res);
880
881         /* since this lockres is new it doesn't not require the spinlock */
882         __dlm_lockres_grab_inflight_ref(dlm, res);
883
884         /* get an extra ref on the mle in case this is a BLOCK
885          * if so, the creator of the BLOCK may try to put the last
886          * ref at this time in the assert master handler, so we
887          * need an extra one to keep from a bad ptr deref. */
888         dlm_get_mle_inuse(mle);
889         spin_unlock(&dlm->master_lock);
890         spin_unlock(&dlm->spinlock);
891
892 redo_request:
893         while (wait_on_recovery) {
894                 /* any cluster changes that occurred after dropping the
895                  * dlm spinlock would be detectable be a change on the mle,
896                  * so we only need to clear out the recovery map once. */
897                 if (dlm_is_recovery_lock(lockid, namelen)) {
898                         mlog(0, "%s: Recovery map is not empty, but must "
899                              "master $RECOVERY lock now\n", dlm->name);
900                         if (!dlm_pre_master_reco_lockres(dlm, res))
901                                 wait_on_recovery = 0;
902                         else {
903                                 mlog(0, "%s: waiting 500ms for heartbeat state "
904                                     "change\n", dlm->name);
905                                 msleep(500);
906                         }
907                         continue;
908                 }
909
910                 dlm_kick_recovery_thread(dlm);
911                 msleep(1000);
912                 dlm_wait_for_recovery(dlm);
913
914                 spin_lock(&dlm->spinlock);
915                 bit = find_first_bit(dlm->recovery_map, O2NM_MAX_NODES);
916                 if (bit < O2NM_MAX_NODES) {
917                         mlog(0, "%s: res %.*s, At least one node (%d) "
918                              "to recover before lock mastery can begin\n",
919                              dlm->name, namelen, (char *)lockid, bit);
920                         wait_on_recovery = 1;
921                 } else
922                         wait_on_recovery = 0;
923                 spin_unlock(&dlm->spinlock);
924
925                 if (wait_on_recovery)
926                         dlm_wait_for_node_recovery(dlm, bit, 10000);
927         }
928
929         /* must wait for lock to be mastered elsewhere */
930         if (blocked)
931                 goto wait;
932
933         ret = -EINVAL;
934         dlm_node_iter_init(mle->vote_map, &iter);
935         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
936                 ret = dlm_do_master_request(res, mle, nodenum);
937                 if (ret < 0)
938                         mlog_errno(ret);
939                 if (mle->master != O2NM_MAX_NODES) {
940                         /* found a master ! */
941                         if (mle->master <= nodenum)
942                                 break;
943                         /* if our master request has not reached the master
944                          * yet, keep going until it does.  this is how the
945                          * master will know that asserts are needed back to
946                          * the lower nodes. */
947                         mlog(0, "%s: res %.*s, Requests only up to %u but "
948                              "master is %u, keep going\n", dlm->name, namelen,
949                              lockid, nodenum, mle->master);
950                 }
951         }
952
953 wait:
954         /* keep going until the response map includes all nodes */
955         ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
956         if (ret < 0) {
957                 wait_on_recovery = 1;
958                 mlog(0, "%s: res %.*s, Node map changed, redo the master "
959                      "request now, blocked=%d\n", dlm->name, res->lockname.len,
960                      res->lockname.name, blocked);
961                 if (++tries > 20) {
962                         mlog(ML_ERROR, "%s: res %.*s, Spinning on "
963                              "dlm_wait_for_lock_mastery, blocked = %d\n",
964                              dlm->name, res->lockname.len,
965                              res->lockname.name, blocked);
966                         dlm_print_one_lock_resource(res);
967                         dlm_print_one_mle(mle);
968                         tries = 0;
969                 }
970                 goto redo_request;
971         }
972
973         mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len,
974              res->lockname.name, res->owner);
975         /* make sure we never continue without this */
976         BUG_ON(res->owner == O2NM_MAX_NODES);
977
978         /* master is known, detach if not already detached */
979         dlm_mle_detach_hb_events(dlm, mle);
980         dlm_put_mle(mle);
981         /* put the extra ref */
982         dlm_put_mle_inuse(mle);
983
984 wake_waiters:
985         spin_lock(&res->spinlock);
986         res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
987         spin_unlock(&res->spinlock);
988         wake_up(&res->wq);
989
990 leave:
991         /* need to free the unused mle */
992         if (alloc_mle)
993                 kmem_cache_free(dlm_mle_cache, alloc_mle);
994
995         return res;
996 }
997
998
999 #define DLM_MASTERY_TIMEOUT_MS   5000
1000
1001 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
1002                                      struct dlm_lock_resource *res,
1003                                      struct dlm_master_list_entry *mle,
1004                                      int *blocked)
1005 {
1006         u8 m;
1007         int ret, bit;
1008         int map_changed, voting_done;
1009         int assert, sleep;
1010
1011 recheck:
1012         ret = 0;
1013         assert = 0;
1014
1015         /* check if another node has already become the owner */
1016         spin_lock(&res->spinlock);
1017         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1018                 mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1019                      res->lockname.len, res->lockname.name, res->owner);
1020                 spin_unlock(&res->spinlock);
1021                 /* this will cause the master to re-assert across
1022                  * the whole cluster, freeing up mles */
1023                 if (res->owner != dlm->node_num) {
1024                         ret = dlm_do_master_request(res, mle, res->owner);
1025                         if (ret < 0) {
1026                                 /* give recovery a chance to run */
1027                                 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1028                                 msleep(500);
1029                                 goto recheck;
1030                         }
1031                 }
1032                 ret = 0;
1033                 goto leave;
1034         }
1035         spin_unlock(&res->spinlock);
1036
1037         spin_lock(&mle->spinlock);
1038         m = mle->master;
1039         map_changed = (memcmp(mle->vote_map, mle->node_map,
1040                               sizeof(mle->vote_map)) != 0);
1041         voting_done = (memcmp(mle->vote_map, mle->response_map,
1042                              sizeof(mle->vote_map)) == 0);
1043
1044         /* restart if we hit any errors */
1045         if (map_changed) {
1046                 int b;
1047                 mlog(0, "%s: %.*s: node map changed, restarting\n",
1048                      dlm->name, res->lockname.len, res->lockname.name);
1049                 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1050                 b = (mle->type == DLM_MLE_BLOCK);
1051                 if ((*blocked && !b) || (!*blocked && b)) {
1052                         mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1053                              dlm->name, res->lockname.len, res->lockname.name,
1054                              *blocked, b);
1055                         *blocked = b;
1056                 }
1057                 spin_unlock(&mle->spinlock);
1058                 if (ret < 0) {
1059                         mlog_errno(ret);
1060                         goto leave;
1061                 }
1062                 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1063                      "rechecking now\n", dlm->name, res->lockname.len,
1064                      res->lockname.name);
1065                 goto recheck;
1066         } else {
1067                 if (!voting_done) {
1068                         mlog(0, "map not changed and voting not done "
1069                              "for %s:%.*s\n", dlm->name, res->lockname.len,
1070                              res->lockname.name);
1071                 }
1072         }
1073
1074         if (m != O2NM_MAX_NODES) {
1075                 /* another node has done an assert!
1076                  * all done! */
1077                 sleep = 0;
1078         } else {
1079                 sleep = 1;
1080                 /* have all nodes responded? */
1081                 if (voting_done && !*blocked) {
1082                         bit = find_first_bit(mle->maybe_map, O2NM_MAX_NODES);
1083                         if (dlm->node_num <= bit) {
1084                                 /* my node number is lowest.
1085                                  * now tell other nodes that I am
1086                                  * mastering this. */
1087                                 mle->master = dlm->node_num;
1088                                 /* ref was grabbed in get_lock_resource
1089                                  * will be dropped in dlmlock_master */
1090                                 assert = 1;
1091                                 sleep = 0;
1092                         }
1093                         /* if voting is done, but we have not received
1094                          * an assert master yet, we must sleep */
1095                 }
1096         }
1097
1098         spin_unlock(&mle->spinlock);
1099
1100         /* sleep if we haven't finished voting yet */
1101         if (sleep) {
1102                 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1103                 atomic_set(&mle->woken, 0);
1104                 (void)wait_event_timeout(mle->wq,
1105                                          (atomic_read(&mle->woken) == 1),
1106                                          timeo);
1107                 if (res->owner == O2NM_MAX_NODES) {
1108                         mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1109                              res->lockname.len, res->lockname.name);
1110                         goto recheck;
1111                 }
1112                 mlog(0, "done waiting, master is %u\n", res->owner);
1113                 ret = 0;
1114                 goto leave;
1115         }
1116
1117         ret = 0;   /* done */
1118         if (assert) {
1119                 m = dlm->node_num;
1120                 mlog(0, "about to master %.*s here, this=%u\n",
1121                      res->lockname.len, res->lockname.name, m);
1122                 ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1123                 if (ret) {
1124                         /* This is a failure in the network path,
1125                          * not in the response to the assert_master
1126                          * (any nonzero response is a BUG on this node).
1127                          * Most likely a socket just got disconnected
1128                          * due to node death. */
1129                         mlog_errno(ret);
1130                 }
1131                 /* no longer need to restart lock mastery.
1132                  * all living nodes have been contacted. */
1133                 ret = 0;
1134         }
1135
1136         /* set the lockres owner */
1137         spin_lock(&res->spinlock);
1138         /* mastery reference obtained either during
1139          * assert_master_handler or in get_lock_resource */
1140         dlm_change_lockres_owner(dlm, res, m);
1141         spin_unlock(&res->spinlock);
1142
1143 leave:
1144         return ret;
1145 }
1146
1147 struct dlm_bitmap_diff_iter
1148 {
1149         int curnode;
1150         unsigned long *orig_bm;
1151         unsigned long *cur_bm;
1152         unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1153 };
1154
1155 enum dlm_node_state_change
1156 {
1157         NODE_DOWN = -1,
1158         NODE_NO_CHANGE = 0,
1159         NODE_UP
1160 };
1161
1162 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1163                                       unsigned long *orig_bm,
1164                                       unsigned long *cur_bm)
1165 {
1166         unsigned long p1, p2;
1167         int i;
1168
1169         iter->curnode = -1;
1170         iter->orig_bm = orig_bm;
1171         iter->cur_bm = cur_bm;
1172
1173         for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1174                 p1 = *(iter->orig_bm + i);
1175                 p2 = *(iter->cur_bm + i);
1176                 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1177         }
1178 }
1179
1180 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1181                                      enum dlm_node_state_change *state)
1182 {
1183         int bit;
1184
1185         if (iter->curnode >= O2NM_MAX_NODES)
1186                 return -ENOENT;
1187
1188         bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1189                             iter->curnode+1);
1190         if (bit >= O2NM_MAX_NODES) {
1191                 iter->curnode = O2NM_MAX_NODES;
1192                 return -ENOENT;
1193         }
1194
1195         /* if it was there in the original then this node died */
1196         if (test_bit(bit, iter->orig_bm))
1197                 *state = NODE_DOWN;
1198         else
1199                 *state = NODE_UP;
1200
1201         iter->curnode = bit;
1202         return bit;
1203 }
1204
1205
1206 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1207                                     struct dlm_lock_resource *res,
1208                                     struct dlm_master_list_entry *mle,
1209                                     int blocked)
1210 {
1211         struct dlm_bitmap_diff_iter bdi;
1212         enum dlm_node_state_change sc;
1213         int node;
1214         int ret = 0;
1215
1216         mlog(0, "something happened such that the "
1217              "master process may need to be restarted!\n");
1218
1219         assert_spin_locked(&mle->spinlock);
1220
1221         dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1222         node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1223         while (node >= 0) {
1224                 if (sc == NODE_UP) {
1225                         /* a node came up.  clear any old vote from
1226                          * the response map and set it in the vote map
1227                          * then restart the mastery. */
1228                         mlog(ML_NOTICE, "node %d up while restarting\n", node);
1229
1230                         /* redo the master request, but only for the new node */
1231                         mlog(0, "sending request to new node\n");
1232                         clear_bit(node, mle->response_map);
1233                         set_bit(node, mle->vote_map);
1234                 } else {
1235                         mlog(ML_ERROR, "node down! %d\n", node);
1236                         if (blocked) {
1237                                 int lowest = find_first_bit(mle->maybe_map,
1238                                                        O2NM_MAX_NODES);
1239
1240                                 /* act like it was never there */
1241                                 clear_bit(node, mle->maybe_map);
1242
1243                                 if (node == lowest) {
1244                                         mlog(0, "expected master %u died"
1245                                             " while this node was blocked "
1246                                             "waiting on it!\n", node);
1247                                         lowest = find_next_bit(mle->maybe_map,
1248                                                         O2NM_MAX_NODES,
1249                                                         lowest+1);
1250                                         if (lowest < O2NM_MAX_NODES) {
1251                                                 mlog(0, "%s:%.*s:still "
1252                                                      "blocked. waiting on %u "
1253                                                      "now\n", dlm->name,
1254                                                      res->lockname.len,
1255                                                      res->lockname.name,
1256                                                      lowest);
1257                                         } else {
1258                                                 /* mle is an MLE_BLOCK, but
1259                                                  * there is now nothing left to
1260                                                  * block on.  we need to return
1261                                                  * all the way back out and try
1262                                                  * again with an MLE_MASTER.
1263                                                  * dlm_do_local_recovery_cleanup
1264                                                  * has already run, so the mle
1265                                                  * refcount is ok */
1266                                                 mlog(0, "%s:%.*s: no "
1267                                                      "longer blocking. try to "
1268                                                      "master this here\n",
1269                                                      dlm->name,
1270                                                      res->lockname.len,
1271                                                      res->lockname.name);
1272                                                 mle->type = DLM_MLE_MASTER;
1273                                                 mle->mleres = res;
1274                                         }
1275                                 }
1276                         }
1277
1278                         /* now blank out everything, as if we had never
1279                          * contacted anyone */
1280                         memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
1281                         memset(mle->response_map, 0, sizeof(mle->response_map));
1282                         /* reset the vote_map to the current node_map */
1283                         memcpy(mle->vote_map, mle->node_map,
1284                                sizeof(mle->node_map));
1285                         /* put myself into the maybe map */
1286                         if (mle->type != DLM_MLE_BLOCK)
1287                                 set_bit(dlm->node_num, mle->maybe_map);
1288                 }
1289                 ret = -EAGAIN;
1290                 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1291         }
1292         return ret;
1293 }
1294
1295
1296 /*
1297  * DLM_MASTER_REQUEST_MSG
1298  *
1299  * returns: 0 on success,
1300  *          -errno on a network error
1301  *
1302  * on error, the caller should assume the target node is "dead"
1303  *
1304  */
1305
1306 static int dlm_do_master_request(struct dlm_lock_resource *res,
1307                                  struct dlm_master_list_entry *mle, int to)
1308 {
1309         struct dlm_ctxt *dlm = mle->dlm;
1310         struct dlm_master_request request;
1311         int ret, response=0, resend;
1312
1313         memset(&request, 0, sizeof(request));
1314         request.node_idx = dlm->node_num;
1315
1316         BUG_ON(mle->type == DLM_MLE_MIGRATION);
1317
1318         request.namelen = (u8)mle->mnamelen;
1319         memcpy(request.name, mle->mname, request.namelen);
1320
1321 again:
1322         ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1323                                  sizeof(request), to, &response);
1324         if (ret < 0)  {
1325                 if (ret == -ESRCH) {
1326                         /* should never happen */
1327                         mlog(ML_ERROR, "TCP stack not ready!\n");
1328                         BUG();
1329                 } else if (ret == -EINVAL) {
1330                         mlog(ML_ERROR, "bad args passed to o2net!\n");
1331                         BUG();
1332                 } else if (ret == -ENOMEM) {
1333                         mlog(ML_ERROR, "out of memory while trying to send "
1334                              "network message!  retrying\n");
1335                         /* this is totally crude */
1336                         msleep(50);
1337                         goto again;
1338                 } else if (!dlm_is_host_down(ret)) {
1339                         /* not a network error. bad. */
1340                         mlog_errno(ret);
1341                         mlog(ML_ERROR, "unhandled error!");
1342                         BUG();
1343                 }
1344                 /* all other errors should be network errors,
1345                  * and likely indicate node death */
1346                 mlog(ML_ERROR, "link to %d went down!\n", to);
1347                 goto out;
1348         }
1349
1350         ret = 0;
1351         resend = 0;
1352         spin_lock(&mle->spinlock);
1353         switch (response) {
1354                 case DLM_MASTER_RESP_YES:
1355                         set_bit(to, mle->response_map);
1356                         mlog(0, "node %u is the master, response=YES\n", to);
1357                         mlog(0, "%s:%.*s: master node %u now knows I have a "
1358                              "reference\n", dlm->name, res->lockname.len,
1359                              res->lockname.name, to);
1360                         mle->master = to;
1361                         break;
1362                 case DLM_MASTER_RESP_NO:
1363                         mlog(0, "node %u not master, response=NO\n", to);
1364                         set_bit(to, mle->response_map);
1365                         break;
1366                 case DLM_MASTER_RESP_MAYBE:
1367                         mlog(0, "node %u not master, response=MAYBE\n", to);
1368                         set_bit(to, mle->response_map);
1369                         set_bit(to, mle->maybe_map);
1370                         break;
1371                 case DLM_MASTER_RESP_ERROR:
1372                         mlog(0, "node %u hit an error, resending\n", to);
1373                         resend = 1;
1374                         response = 0;
1375                         break;
1376                 default:
1377                         mlog(ML_ERROR, "bad response! %u\n", response);
1378                         BUG();
1379         }
1380         spin_unlock(&mle->spinlock);
1381         if (resend) {
1382                 /* this is also totally crude */
1383                 msleep(50);
1384                 goto again;
1385         }
1386
1387 out:
1388         return ret;
1389 }
1390
1391 /*
1392  * locks that can be taken here:
1393  * dlm->spinlock
1394  * res->spinlock
1395  * mle->spinlock
1396  * dlm->master_list
1397  *
1398  * if possible, TRIM THIS DOWN!!!
1399  */
1400 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1401                                void **ret_data)
1402 {
1403         u8 response = DLM_MASTER_RESP_MAYBE;
1404         struct dlm_ctxt *dlm = data;
1405         struct dlm_lock_resource *res = NULL;
1406         struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1407         struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1408         char *name;
1409         unsigned int namelen, hash;
1410         int found, ret;
1411         int set_maybe;
1412         int dispatch_assert = 0;
1413         int dispatched = 0;
1414
1415         if (!dlm_grab(dlm))
1416                 return DLM_MASTER_RESP_NO;
1417
1418         if (!dlm_domain_fully_joined(dlm)) {
1419                 response = DLM_MASTER_RESP_NO;
1420                 goto send_response;
1421         }
1422
1423         name = request->name;
1424         namelen = request->namelen;
1425         hash = dlm_lockid_hash(name, namelen);
1426
1427         if (namelen > DLM_LOCKID_NAME_MAX) {
1428                 response = DLM_IVBUFLEN;
1429                 goto send_response;
1430         }
1431
1432 way_up_top:
1433         spin_lock(&dlm->spinlock);
1434         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1435         if (res) {
1436                 spin_unlock(&dlm->spinlock);
1437
1438                 /* take care of the easy cases up front */
1439                 spin_lock(&res->spinlock);
1440
1441                 /*
1442                  * Right after dlm spinlock was released, dlm_thread could have
1443                  * purged the lockres. Check if lockres got unhashed. If so
1444                  * start over.
1445                  */
1446                 if (hlist_unhashed(&res->hash_node)) {
1447                         spin_unlock(&res->spinlock);
1448                         dlm_lockres_put(res);
1449                         goto way_up_top;
1450                 }
1451
1452                 if (res->state & (DLM_LOCK_RES_RECOVERING|
1453                                   DLM_LOCK_RES_MIGRATING)) {
1454                         spin_unlock(&res->spinlock);
1455                         mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1456                              "being recovered/migrated\n");
1457                         response = DLM_MASTER_RESP_ERROR;
1458                         if (mle)
1459                                 kmem_cache_free(dlm_mle_cache, mle);
1460                         goto send_response;
1461                 }
1462
1463                 if (res->owner == dlm->node_num) {
1464                         dlm_lockres_set_refmap_bit(dlm, res, request->node_idx);
1465                         spin_unlock(&res->spinlock);
1466                         response = DLM_MASTER_RESP_YES;
1467                         if (mle)
1468                                 kmem_cache_free(dlm_mle_cache, mle);
1469
1470                         /* this node is the owner.
1471                          * there is some extra work that needs to
1472                          * happen now.  the requesting node has
1473                          * caused all nodes up to this one to
1474                          * create mles.  this node now needs to
1475                          * go back and clean those up. */
1476                         dispatch_assert = 1;
1477                         goto send_response;
1478                 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1479                         spin_unlock(&res->spinlock);
1480                         // mlog(0, "node %u is the master\n", res->owner);
1481                         response = DLM_MASTER_RESP_NO;
1482                         if (mle)
1483                                 kmem_cache_free(dlm_mle_cache, mle);
1484                         goto send_response;
1485                 }
1486
1487                 /* ok, there is no owner.  either this node is
1488                  * being blocked, or it is actively trying to
1489                  * master this lock. */
1490                 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1491                         mlog(ML_ERROR, "lock with no owner should be "
1492                              "in-progress!\n");
1493                         BUG();
1494                 }
1495
1496                 // mlog(0, "lockres is in progress...\n");
1497                 spin_lock(&dlm->master_lock);
1498                 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1499                 if (!found) {
1500                         mlog(ML_ERROR, "no mle found for this lock!\n");
1501                         BUG();
1502                 }
1503                 set_maybe = 1;
1504                 spin_lock(&tmpmle->spinlock);
1505                 if (tmpmle->type == DLM_MLE_BLOCK) {
1506                         // mlog(0, "this node is waiting for "
1507                         // "lockres to be mastered\n");
1508                         response = DLM_MASTER_RESP_NO;
1509                 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1510                         mlog(0, "node %u is master, but trying to migrate to "
1511                              "node %u.\n", tmpmle->master, tmpmle->new_master);
1512                         if (tmpmle->master == dlm->node_num) {
1513                                 mlog(ML_ERROR, "no owner on lockres, but this "
1514                                      "node is trying to migrate it to %u?!\n",
1515                                      tmpmle->new_master);
1516                                 BUG();
1517                         } else {
1518                                 /* the real master can respond on its own */
1519                                 response = DLM_MASTER_RESP_NO;
1520                         }
1521                 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1522                         set_maybe = 0;
1523                         if (tmpmle->master == dlm->node_num) {
1524                                 response = DLM_MASTER_RESP_YES;
1525                                 /* this node will be the owner.
1526                                  * go back and clean the mles on any
1527                                  * other nodes */
1528                                 dispatch_assert = 1;
1529                                 dlm_lockres_set_refmap_bit(dlm, res,
1530                                                            request->node_idx);
1531                         } else
1532                                 response = DLM_MASTER_RESP_NO;
1533                 } else {
1534                         // mlog(0, "this node is attempting to "
1535                         // "master lockres\n");
1536                         response = DLM_MASTER_RESP_MAYBE;
1537                 }
1538                 if (set_maybe)
1539                         set_bit(request->node_idx, tmpmle->maybe_map);
1540                 spin_unlock(&tmpmle->spinlock);
1541
1542                 spin_unlock(&dlm->master_lock);
1543                 spin_unlock(&res->spinlock);
1544
1545                 /* keep the mle attached to heartbeat events */
1546                 dlm_put_mle(tmpmle);
1547                 if (mle)
1548                         kmem_cache_free(dlm_mle_cache, mle);
1549                 goto send_response;
1550         }
1551
1552         /*
1553          * lockres doesn't exist on this node
1554          * if there is an MLE_BLOCK, return NO
1555          * if there is an MLE_MASTER, return MAYBE
1556          * otherwise, add an MLE_BLOCK, return NO
1557          */
1558         spin_lock(&dlm->master_lock);
1559         found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1560         if (!found) {
1561                 /* this lockid has never been seen on this node yet */
1562                 // mlog(0, "no mle found\n");
1563                 if (!mle) {
1564                         spin_unlock(&dlm->master_lock);
1565                         spin_unlock(&dlm->spinlock);
1566
1567                         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1568                         if (!mle) {
1569                                 response = DLM_MASTER_RESP_ERROR;
1570                                 mlog_errno(-ENOMEM);
1571                                 goto send_response;
1572                         }
1573                         goto way_up_top;
1574                 }
1575
1576                 // mlog(0, "this is second time thru, already allocated, "
1577                 // "add the block.\n");
1578                 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1579                 set_bit(request->node_idx, mle->maybe_map);
1580                 __dlm_insert_mle(dlm, mle);
1581                 response = DLM_MASTER_RESP_NO;
1582         } else {
1583                 spin_lock(&tmpmle->spinlock);
1584                 if (tmpmle->master == dlm->node_num) {
1585                         mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1586                         BUG();
1587                 }
1588                 if (tmpmle->type == DLM_MLE_BLOCK)
1589                         response = DLM_MASTER_RESP_NO;
1590                 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1591                         mlog(0, "migration mle was found (%u->%u)\n",
1592                              tmpmle->master, tmpmle->new_master);
1593                         /* real master can respond on its own */
1594                         response = DLM_MASTER_RESP_NO;
1595                 } else
1596                         response = DLM_MASTER_RESP_MAYBE;
1597                 set_bit(request->node_idx, tmpmle->maybe_map);
1598                 spin_unlock(&tmpmle->spinlock);
1599         }
1600         spin_unlock(&dlm->master_lock);
1601         spin_unlock(&dlm->spinlock);
1602
1603         if (found) {
1604                 /* keep the mle attached to heartbeat events */
1605                 dlm_put_mle(tmpmle);
1606         }
1607 send_response:
1608         /*
1609          * __dlm_lookup_lockres() grabbed a reference to this lockres.
1610          * The reference is released by dlm_assert_master_worker() under
1611          * the call to dlm_dispatch_assert_master().  If
1612          * dlm_assert_master_worker() isn't called, we drop it here.
1613          */
1614         if (dispatch_assert) {
1615                 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1616                              dlm->node_num, res->lockname.len, res->lockname.name);
1617                 spin_lock(&res->spinlock);
1618                 ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1619                                                  DLM_ASSERT_MASTER_MLE_CLEANUP);
1620                 if (ret < 0) {
1621                         mlog(ML_ERROR, "failed to dispatch assert master work\n");
1622                         response = DLM_MASTER_RESP_ERROR;
1623                         spin_unlock(&res->spinlock);
1624                         dlm_lockres_put(res);
1625                 } else {
1626                         dispatched = 1;
1627                         __dlm_lockres_grab_inflight_worker(dlm, res);
1628                         spin_unlock(&res->spinlock);
1629                 }
1630         } else {
1631                 if (res)
1632                         dlm_lockres_put(res);
1633         }
1634
1635         if (!dispatched)
1636                 dlm_put(dlm);
1637         return response;
1638 }
1639
1640 /*
1641  * DLM_ASSERT_MASTER_MSG
1642  */
1643
1644
1645 /*
1646  * NOTE: this can be used for debugging
1647  * can periodically run all locks owned by this node
1648  * and re-assert across the cluster...
1649  */
1650 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1651                                 struct dlm_lock_resource *res,
1652                                 void *nodemap, u32 flags)
1653 {
1654         struct dlm_assert_master assert;
1655         int to, tmpret;
1656         struct dlm_node_iter iter;
1657         int ret = 0;
1658         int reassert;
1659         const char *lockname = res->lockname.name;
1660         unsigned int namelen = res->lockname.len;
1661
1662         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1663
1664         spin_lock(&res->spinlock);
1665         res->state |= DLM_LOCK_RES_SETREF_INPROG;
1666         spin_unlock(&res->spinlock);
1667
1668 again:
1669         reassert = 0;
1670
1671         /* note that if this nodemap is empty, it returns 0 */
1672         dlm_node_iter_init(nodemap, &iter);
1673         while ((to = dlm_node_iter_next(&iter)) >= 0) {
1674                 int r = 0;
1675                 struct dlm_master_list_entry *mle = NULL;
1676
1677                 mlog(0, "sending assert master to %d (%.*s)\n", to,
1678                      namelen, lockname);
1679                 memset(&assert, 0, sizeof(assert));
1680                 assert.node_idx = dlm->node_num;
1681                 assert.namelen = namelen;
1682                 memcpy(assert.name, lockname, namelen);
1683                 assert.flags = cpu_to_be32(flags);
1684
1685                 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1686                                             &assert, sizeof(assert), to, &r);
1687                 if (tmpret < 0) {
1688                         mlog(ML_ERROR, "Error %d when sending message %u (key "
1689                              "0x%x) to node %u\n", tmpret,
1690                              DLM_ASSERT_MASTER_MSG, dlm->key, to);
1691                         if (!dlm_is_host_down(tmpret)) {
1692                                 mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1693                                 BUG();
1694                         }
1695                         /* a node died.  finish out the rest of the nodes. */
1696                         mlog(0, "link to %d went down!\n", to);
1697                         /* any nonzero status return will do */
1698                         ret = tmpret;
1699                         r = 0;
1700                 } else if (r < 0) {
1701                         /* ok, something horribly messed.  kill thyself. */
1702                         mlog(ML_ERROR,"during assert master of %.*s to %u, "
1703                              "got %d.\n", namelen, lockname, to, r);
1704                         spin_lock(&dlm->spinlock);
1705                         spin_lock(&dlm->master_lock);
1706                         if (dlm_find_mle(dlm, &mle, (char *)lockname,
1707                                          namelen)) {
1708                                 dlm_print_one_mle(mle);
1709                                 __dlm_put_mle(mle);
1710                         }
1711                         spin_unlock(&dlm->master_lock);
1712                         spin_unlock(&dlm->spinlock);
1713                         BUG();
1714                 }
1715
1716                 if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1717                     !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1718                                 mlog(ML_ERROR, "%.*s: very strange, "
1719                                      "master MLE but no lockres on %u\n",
1720                                      namelen, lockname, to);
1721                 }
1722
1723                 if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1724                         mlog(0, "%.*s: node %u create mles on other "
1725                              "nodes and requests a re-assert\n",
1726                              namelen, lockname, to);
1727                         reassert = 1;
1728                 }
1729                 if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1730                         mlog(0, "%.*s: node %u has a reference to this "
1731                              "lockres, set the bit in the refmap\n",
1732                              namelen, lockname, to);
1733                         spin_lock(&res->spinlock);
1734                         dlm_lockres_set_refmap_bit(dlm, res, to);
1735                         spin_unlock(&res->spinlock);
1736                 }
1737         }
1738
1739         if (reassert)
1740                 goto again;
1741
1742         spin_lock(&res->spinlock);
1743         res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1744         spin_unlock(&res->spinlock);
1745         wake_up(&res->wq);
1746
1747         return ret;
1748 }
1749
1750 /*
1751  * locks that can be taken here:
1752  * dlm->spinlock
1753  * res->spinlock
1754  * mle->spinlock
1755  * dlm->master_list
1756  *
1757  * if possible, TRIM THIS DOWN!!!
1758  */
1759 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1760                               void **ret_data)
1761 {
1762         struct dlm_ctxt *dlm = data;
1763         struct dlm_master_list_entry *mle = NULL;
1764         struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1765         struct dlm_lock_resource *res = NULL;
1766         char *name;
1767         unsigned int namelen, hash;
1768         u32 flags;
1769         int master_request = 0, have_lockres_ref = 0;
1770         int ret = 0;
1771
1772         if (!dlm_grab(dlm))
1773                 return 0;
1774
1775         name = assert->name;
1776         namelen = assert->namelen;
1777         hash = dlm_lockid_hash(name, namelen);
1778         flags = be32_to_cpu(assert->flags);
1779
1780         if (namelen > DLM_LOCKID_NAME_MAX) {
1781                 mlog(ML_ERROR, "Invalid name length!");
1782                 goto done;
1783         }
1784
1785         spin_lock(&dlm->spinlock);
1786
1787         if (flags)
1788                 mlog(0, "assert_master with flags: %u\n", flags);
1789
1790         /* find the MLE */
1791         spin_lock(&dlm->master_lock);
1792         if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1793                 /* not an error, could be master just re-asserting */
1794                 mlog(0, "just got an assert_master from %u, but no "
1795                      "MLE for it! (%.*s)\n", assert->node_idx,
1796                      namelen, name);
1797         } else {
1798                 int bit = find_first_bit(mle->maybe_map, O2NM_MAX_NODES);
1799                 if (bit >= O2NM_MAX_NODES) {
1800                         /* not necessarily an error, though less likely.
1801                          * could be master just re-asserting. */
1802                         mlog(0, "no bits set in the maybe_map, but %u "
1803                              "is asserting! (%.*s)\n", assert->node_idx,
1804                              namelen, name);
1805                 } else if (bit != assert->node_idx) {
1806                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1807                                 mlog(0, "master %u was found, %u should "
1808                                      "back off\n", assert->node_idx, bit);
1809                         } else {
1810                                 /* with the fix for bug 569, a higher node
1811                                  * number winning the mastery will respond
1812                                  * YES to mastery requests, but this node
1813                                  * had no way of knowing.  let it pass. */
1814                                 mlog(0, "%u is the lowest node, "
1815                                      "%u is asserting. (%.*s)  %u must "
1816                                      "have begun after %u won.\n", bit,
1817                                      assert->node_idx, namelen, name, bit,
1818                                      assert->node_idx);
1819                         }
1820                 }
1821                 if (mle->type == DLM_MLE_MIGRATION) {
1822                         if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1823                                 mlog(0, "%s:%.*s: got cleanup assert"
1824                                      " from %u for migration\n",
1825                                      dlm->name, namelen, name,
1826                                      assert->node_idx);
1827                         } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1828                                 mlog(0, "%s:%.*s: got unrelated assert"
1829                                      " from %u for migration, ignoring\n",
1830                                      dlm->name, namelen, name,
1831                                      assert->node_idx);
1832                                 __dlm_put_mle(mle);
1833                                 spin_unlock(&dlm->master_lock);
1834                                 spin_unlock(&dlm->spinlock);
1835                                 goto done;
1836                         }
1837                 }
1838         }
1839         spin_unlock(&dlm->master_lock);
1840
1841         /* ok everything checks out with the MLE
1842          * now check to see if there is a lockres */
1843         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1844         if (res) {
1845                 spin_lock(&res->spinlock);
1846                 if (res->state & DLM_LOCK_RES_RECOVERING)  {
1847                         mlog(ML_ERROR, "%u asserting but %.*s is "
1848                              "RECOVERING!\n", assert->node_idx, namelen, name);
1849                         goto kill;
1850                 }
1851                 if (!mle) {
1852                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1853                             res->owner != assert->node_idx) {
1854                                 mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1855                                      "but current owner is %u! (%.*s)\n",
1856                                      assert->node_idx, res->owner, namelen,
1857                                      name);
1858                                 __dlm_print_one_lock_resource(res);
1859                                 BUG();
1860                         }
1861                 } else if (mle->type != DLM_MLE_MIGRATION) {
1862                         if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1863                                 /* owner is just re-asserting */
1864                                 if (res->owner == assert->node_idx) {
1865                                         mlog(0, "owner %u re-asserting on "
1866                                              "lock %.*s\n", assert->node_idx,
1867                                              namelen, name);
1868                                         goto ok;
1869                                 }
1870                                 mlog(ML_ERROR, "got assert_master from "
1871                                      "node %u, but %u is the owner! "
1872                                      "(%.*s)\n", assert->node_idx,
1873                                      res->owner, namelen, name);
1874                                 goto kill;
1875                         }
1876                         if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1877                                 mlog(ML_ERROR, "got assert from %u, but lock "
1878                                      "with no owner should be "
1879                                      "in-progress! (%.*s)\n",
1880                                      assert->node_idx,
1881                                      namelen, name);
1882                                 goto kill;
1883                         }
1884                 } else /* mle->type == DLM_MLE_MIGRATION */ {
1885                         /* should only be getting an assert from new master */
1886                         if (assert->node_idx != mle->new_master) {
1887                                 mlog(ML_ERROR, "got assert from %u, but "
1888                                      "new master is %u, and old master "
1889                                      "was %u (%.*s)\n",
1890                                      assert->node_idx, mle->new_master,
1891                                      mle->master, namelen, name);
1892                                 goto kill;
1893                         }
1894
1895                 }
1896 ok:
1897                 spin_unlock(&res->spinlock);
1898         }
1899
1900         // mlog(0, "woo!  got an assert_master from node %u!\n",
1901         //           assert->node_idx);
1902         if (mle) {
1903                 int extra_ref = 0;
1904                 int nn = -1;
1905                 int rr, err = 0;
1906
1907                 spin_lock(&mle->spinlock);
1908                 if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1909                         extra_ref = 1;
1910                 else {
1911                         /* MASTER mle: if any bits set in the response map
1912                          * then the calling node needs to re-assert to clear
1913                          * up nodes that this node contacted */
1914                         while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1915                                                     nn+1)) < O2NM_MAX_NODES) {
1916                                 if (nn != dlm->node_num && nn != assert->node_idx) {
1917                                         master_request = 1;
1918                                         break;
1919                                 }
1920                         }
1921                 }
1922                 mle->master = assert->node_idx;
1923                 atomic_set(&mle->woken, 1);
1924                 wake_up(&mle->wq);
1925                 spin_unlock(&mle->spinlock);
1926
1927                 if (res) {
1928                         int wake = 0;
1929                         spin_lock(&res->spinlock);
1930                         if (mle->type == DLM_MLE_MIGRATION) {
1931                                 mlog(0, "finishing off migration of lockres %.*s, "
1932                                         "from %u to %u\n",
1933                                         res->lockname.len, res->lockname.name,
1934                                         dlm->node_num, mle->new_master);
1935                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
1936                                 wake = 1;
1937                                 dlm_change_lockres_owner(dlm, res, mle->new_master);
1938                                 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1939                         } else {
1940                                 dlm_change_lockres_owner(dlm, res, mle->master);
1941                         }
1942                         spin_unlock(&res->spinlock);
1943                         have_lockres_ref = 1;
1944                         if (wake)
1945                                 wake_up(&res->wq);
1946                 }
1947
1948                 /* master is known, detach if not already detached.
1949                  * ensures that only one assert_master call will happen
1950                  * on this mle. */
1951                 spin_lock(&dlm->master_lock);
1952
1953                 rr = kref_read(&mle->mle_refs);
1954                 if (mle->inuse > 0) {
1955                         if (extra_ref && rr < 3)
1956                                 err = 1;
1957                         else if (!extra_ref && rr < 2)
1958                                 err = 1;
1959                 } else {
1960                         if (extra_ref && rr < 2)
1961                                 err = 1;
1962                         else if (!extra_ref && rr < 1)
1963                                 err = 1;
1964                 }
1965                 if (err) {
1966                         mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
1967                              "that will mess up this node, refs=%d, extra=%d, "
1968                              "inuse=%d\n", dlm->name, namelen, name,
1969                              assert->node_idx, rr, extra_ref, mle->inuse);
1970                         dlm_print_one_mle(mle);
1971                 }
1972                 __dlm_unlink_mle(dlm, mle);
1973                 __dlm_mle_detach_hb_events(dlm, mle);
1974                 __dlm_put_mle(mle);
1975                 if (extra_ref) {
1976                         /* the assert master message now balances the extra
1977                          * ref given by the master / migration request message.
1978                          * if this is the last put, it will be removed
1979                          * from the list. */
1980                         __dlm_put_mle(mle);
1981                 }
1982                 spin_unlock(&dlm->master_lock);
1983         } else if (res) {
1984                 if (res->owner != assert->node_idx) {
1985                         mlog(0, "assert_master from %u, but current "
1986                              "owner is %u (%.*s), no mle\n", assert->node_idx,
1987                              res->owner, namelen, name);
1988                 }
1989         }
1990         spin_unlock(&dlm->spinlock);
1991
1992 done:
1993         ret = 0;
1994         if (res) {
1995                 spin_lock(&res->spinlock);
1996                 res->state |= DLM_LOCK_RES_SETREF_INPROG;
1997                 spin_unlock(&res->spinlock);
1998                 *ret_data = (void *)res;
1999         }
2000         dlm_put(dlm);
2001         if (master_request) {
2002                 mlog(0, "need to tell master to reassert\n");
2003                 /* positive. negative would shoot down the node. */
2004                 ret |= DLM_ASSERT_RESPONSE_REASSERT;
2005                 if (!have_lockres_ref) {
2006                         mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2007                              "mle present here for %s:%.*s, but no lockres!\n",
2008                              assert->node_idx, dlm->name, namelen, name);
2009                 }
2010         }
2011         if (have_lockres_ref) {
2012                 /* let the master know we have a reference to the lockres */
2013                 ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2014                 mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2015                      dlm->name, namelen, name, assert->node_idx);
2016         }
2017         return ret;
2018
2019 kill:
2020         /* kill the caller! */
2021         mlog(ML_ERROR, "Bad message received from another node.  Dumping state "
2022              "and killing the other node now!  This node is OK and can continue.\n");
2023         __dlm_print_one_lock_resource(res);
2024         spin_unlock(&res->spinlock);
2025         spin_lock(&dlm->master_lock);
2026         if (mle)
2027                 __dlm_put_mle(mle);
2028         spin_unlock(&dlm->master_lock);
2029         spin_unlock(&dlm->spinlock);
2030         *ret_data = (void *)res;
2031         dlm_put(dlm);
2032         return -EINVAL;
2033 }
2034
2035 void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2036 {
2037         struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2038
2039         if (ret_data) {
2040                 spin_lock(&res->spinlock);
2041                 res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2042                 spin_unlock(&res->spinlock);
2043                 wake_up(&res->wq);
2044                 dlm_lockres_put(res);
2045         }
2046         return;
2047 }
2048
2049 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2050                                struct dlm_lock_resource *res,
2051                                int ignore_higher, u8 request_from, u32 flags)
2052 {
2053         struct dlm_work_item *item;
2054         item = kzalloc(sizeof(*item), GFP_ATOMIC);
2055         if (!item)
2056                 return -ENOMEM;
2057
2058
2059         /* queue up work for dlm_assert_master_worker */
2060         dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2061         item->u.am.lockres = res; /* already have a ref */
2062         /* can optionally ignore node numbers higher than this node */
2063         item->u.am.ignore_higher = ignore_higher;
2064         item->u.am.request_from = request_from;
2065         item->u.am.flags = flags;
2066
2067         if (ignore_higher)
2068                 mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2069                      res->lockname.name);
2070
2071         spin_lock(&dlm->work_lock);
2072         list_add_tail(&item->list, &dlm->work_list);
2073         spin_unlock(&dlm->work_lock);
2074
2075         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2076         return 0;
2077 }
2078
2079 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2080 {
2081         struct dlm_ctxt *dlm = data;
2082         int ret = 0;
2083         struct dlm_lock_resource *res;
2084         unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2085         int ignore_higher;
2086         int bit;
2087         u8 request_from;
2088         u32 flags;
2089
2090         dlm = item->dlm;
2091         res = item->u.am.lockres;
2092         ignore_higher = item->u.am.ignore_higher;
2093         request_from = item->u.am.request_from;
2094         flags = item->u.am.flags;
2095
2096         spin_lock(&dlm->spinlock);
2097         memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
2098         spin_unlock(&dlm->spinlock);
2099
2100         clear_bit(dlm->node_num, nodemap);
2101         if (ignore_higher) {
2102                 /* if is this just to clear up mles for nodes below
2103                  * this node, do not send the message to the original
2104                  * caller or any node number higher than this */
2105                 clear_bit(request_from, nodemap);
2106                 bit = dlm->node_num;
2107                 while (1) {
2108                         bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2109                                             bit+1);
2110                         if (bit >= O2NM_MAX_NODES)
2111                                 break;
2112                         clear_bit(bit, nodemap);
2113                 }
2114         }
2115
2116         /*
2117          * If we're migrating this lock to someone else, we are no
2118          * longer allowed to assert out own mastery.  OTOH, we need to
2119          * prevent migration from starting while we're still asserting
2120          * our dominance.  The reserved ast delays migration.
2121          */
2122         spin_lock(&res->spinlock);
2123         if (res->state & DLM_LOCK_RES_MIGRATING) {
2124                 mlog(0, "Someone asked us to assert mastery, but we're "
2125                      "in the middle of migration.  Skipping assert, "
2126                      "the new master will handle that.\n");
2127                 spin_unlock(&res->spinlock);
2128                 goto put;
2129         } else
2130                 __dlm_lockres_reserve_ast(res);
2131         spin_unlock(&res->spinlock);
2132
2133         /* this call now finishes out the nodemap
2134          * even if one or more nodes die */
2135         mlog(0, "worker about to master %.*s here, this=%u\n",
2136                      res->lockname.len, res->lockname.name, dlm->node_num);
2137         ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2138         if (ret < 0) {
2139                 /* no need to restart, we are done */
2140                 if (!dlm_is_host_down(ret))
2141                         mlog_errno(ret);
2142         }
2143
2144         /* Ok, we've asserted ourselves.  Let's let migration start. */
2145         dlm_lockres_release_ast(dlm, res);
2146
2147 put:
2148         dlm_lockres_drop_inflight_worker(dlm, res);
2149
2150         dlm_lockres_put(res);
2151
2152         mlog(0, "finished with dlm_assert_master_worker\n");
2153 }
2154
2155 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2156  * We cannot wait for node recovery to complete to begin mastering this
2157  * lockres because this lockres is used to kick off recovery! ;-)
2158  * So, do a pre-check on all living nodes to see if any of those nodes
2159  * think that $RECOVERY is currently mastered by a dead node.  If so,
2160  * we wait a short time to allow that node to get notified by its own
2161  * heartbeat stack, then check again.  All $RECOVERY lock resources
2162  * mastered by dead nodes are purged when the heartbeat callback is
2163  * fired, so we can know for sure that it is safe to continue once
2164  * the node returns a live node or no node.  */
2165 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2166                                        struct dlm_lock_resource *res)
2167 {
2168         struct dlm_node_iter iter;
2169         int nodenum;
2170         int ret = 0;
2171         u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2172
2173         spin_lock(&dlm->spinlock);
2174         dlm_node_iter_init(dlm->domain_map, &iter);
2175         spin_unlock(&dlm->spinlock);
2176
2177         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2178                 /* do not send to self */
2179                 if (nodenum == dlm->node_num)
2180                         continue;
2181                 ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2182                 if (ret < 0) {
2183                         mlog_errno(ret);
2184                         if (!dlm_is_host_down(ret))
2185                                 BUG();
2186                         /* host is down, so answer for that node would be
2187                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
2188                         ret = 0;
2189                 }
2190
2191                 if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2192                         /* check to see if this master is in the recovery map */
2193                         spin_lock(&dlm->spinlock);
2194                         if (test_bit(master, dlm->recovery_map)) {
2195                                 mlog(ML_NOTICE, "%s: node %u has not seen "
2196                                      "node %u go down yet, and thinks the "
2197                                      "dead node is mastering the recovery "
2198                                      "lock.  must wait.\n", dlm->name,
2199                                      nodenum, master);
2200                                 ret = -EAGAIN;
2201                         }
2202                         spin_unlock(&dlm->spinlock);
2203                         mlog(0, "%s: reco lock master is %u\n", dlm->name,
2204                              master);
2205                         break;
2206                 }
2207         }
2208         return ret;
2209 }
2210
2211 /*
2212  * DLM_DEREF_LOCKRES_MSG
2213  */
2214
2215 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2216 {
2217         struct dlm_deref_lockres deref;
2218         int ret = 0, r;
2219         const char *lockname;
2220         unsigned int namelen;
2221
2222         lockname = res->lockname.name;
2223         namelen = res->lockname.len;
2224         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2225
2226         memset(&deref, 0, sizeof(deref));
2227         deref.node_idx = dlm->node_num;
2228         deref.namelen = namelen;
2229         memcpy(deref.name, lockname, namelen);
2230
2231         ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2232                                  &deref, sizeof(deref), res->owner, &r);
2233         if (ret < 0)
2234                 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n",
2235                      dlm->name, namelen, lockname, ret, res->owner);
2236         else if (r < 0) {
2237                 /* BAD.  other node says I did not have a ref. */
2238                 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2239                      dlm->name, namelen, lockname, res->owner, r);
2240                 dlm_print_one_lock_resource(res);
2241                 if (r == -ENOMEM)
2242                         BUG();
2243         } else
2244                 ret = r;
2245
2246         return ret;
2247 }
2248
2249 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2250                               void **ret_data)
2251 {
2252         struct dlm_ctxt *dlm = data;
2253         struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2254         struct dlm_lock_resource *res = NULL;
2255         char *name;
2256         unsigned int namelen;
2257         int ret = -EINVAL;
2258         u8 node;
2259         unsigned int hash;
2260         struct dlm_work_item *item;
2261         int cleared = 0;
2262         int dispatch = 0;
2263
2264         if (!dlm_grab(dlm))
2265                 return 0;
2266
2267         name = deref->name;
2268         namelen = deref->namelen;
2269         node = deref->node_idx;
2270
2271         if (namelen > DLM_LOCKID_NAME_MAX) {
2272                 mlog(ML_ERROR, "Invalid name length!");
2273                 goto done;
2274         }
2275         if (deref->node_idx >= O2NM_MAX_NODES) {
2276                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2277                 goto done;
2278         }
2279
2280         hash = dlm_lockid_hash(name, namelen);
2281
2282         spin_lock(&dlm->spinlock);
2283         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2284         if (!res) {
2285                 spin_unlock(&dlm->spinlock);
2286                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2287                      dlm->name, namelen, name);
2288                 goto done;
2289         }
2290         spin_unlock(&dlm->spinlock);
2291
2292         spin_lock(&res->spinlock);
2293         if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2294                 dispatch = 1;
2295         else {
2296                 BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2297                 if (test_bit(node, res->refmap)) {
2298                         dlm_lockres_clear_refmap_bit(dlm, res, node);
2299                         cleared = 1;
2300                 }
2301         }
2302         spin_unlock(&res->spinlock);
2303
2304         if (!dispatch) {
2305                 if (cleared)
2306                         dlm_lockres_calc_usage(dlm, res);
2307                 else {
2308                         mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2309                         "but it is already dropped!\n", dlm->name,
2310                         res->lockname.len, res->lockname.name, node);
2311                         dlm_print_one_lock_resource(res);
2312                 }
2313                 ret = DLM_DEREF_RESPONSE_DONE;
2314                 goto done;
2315         }
2316
2317         item = kzalloc(sizeof(*item), GFP_NOFS);
2318         if (!item) {
2319                 ret = -ENOMEM;
2320                 mlog_errno(ret);
2321                 goto done;
2322         }
2323
2324         dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2325         item->u.dl.deref_res = res;
2326         item->u.dl.deref_node = node;
2327
2328         spin_lock(&dlm->work_lock);
2329         list_add_tail(&item->list, &dlm->work_list);
2330         spin_unlock(&dlm->work_lock);
2331
2332         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2333         return DLM_DEREF_RESPONSE_INPROG;
2334
2335 done:
2336         if (res)
2337                 dlm_lockres_put(res);
2338         dlm_put(dlm);
2339
2340         return ret;
2341 }
2342
2343 int dlm_deref_lockres_done_handler(struct o2net_msg *msg, u32 len, void *data,
2344                               void **ret_data)
2345 {
2346         struct dlm_ctxt *dlm = data;
2347         struct dlm_deref_lockres_done *deref
2348                         = (struct dlm_deref_lockres_done *)msg->buf;
2349         struct dlm_lock_resource *res = NULL;
2350         char *name;
2351         unsigned int namelen;
2352         int ret = -EINVAL;
2353         u8 node;
2354         unsigned int hash;
2355
2356         if (!dlm_grab(dlm))
2357                 return 0;
2358
2359         name = deref->name;
2360         namelen = deref->namelen;
2361         node = deref->node_idx;
2362
2363         if (namelen > DLM_LOCKID_NAME_MAX) {
2364                 mlog(ML_ERROR, "Invalid name length!");
2365                 goto done;
2366         }
2367         if (deref->node_idx >= O2NM_MAX_NODES) {
2368                 mlog(ML_ERROR, "Invalid node number: %u\n", node);
2369                 goto done;
2370         }
2371
2372         hash = dlm_lockid_hash(name, namelen);
2373
2374         spin_lock(&dlm->spinlock);
2375         res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2376         if (!res) {
2377                 spin_unlock(&dlm->spinlock);
2378                 mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2379                      dlm->name, namelen, name);
2380                 goto done;
2381         }
2382
2383         spin_lock(&res->spinlock);
2384         if (!(res->state & DLM_LOCK_RES_DROPPING_REF)) {
2385                 spin_unlock(&res->spinlock);
2386                 spin_unlock(&dlm->spinlock);
2387                 mlog(ML_NOTICE, "%s:%.*s: node %u sends deref done "
2388                         "but it is already derefed!\n", dlm->name,
2389                         res->lockname.len, res->lockname.name, node);
2390                 ret = 0;
2391                 goto done;
2392         }
2393
2394         __dlm_do_purge_lockres(dlm, res);
2395         spin_unlock(&res->spinlock);
2396         wake_up(&res->wq);
2397
2398         spin_unlock(&dlm->spinlock);
2399
2400         ret = 0;
2401 done:
2402         if (res)
2403                 dlm_lockres_put(res);
2404         dlm_put(dlm);
2405         return ret;
2406 }
2407
2408 static void dlm_drop_lockres_ref_done(struct dlm_ctxt *dlm,
2409                 struct dlm_lock_resource *res, u8 node)
2410 {
2411         struct dlm_deref_lockres_done deref;
2412         int ret = 0, r;
2413         const char *lockname;
2414         unsigned int namelen;
2415
2416         lockname = res->lockname.name;
2417         namelen = res->lockname.len;
2418         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2419
2420         memset(&deref, 0, sizeof(deref));
2421         deref.node_idx = dlm->node_num;
2422         deref.namelen = namelen;
2423         memcpy(deref.name, lockname, namelen);
2424
2425         ret = o2net_send_message(DLM_DEREF_LOCKRES_DONE, dlm->key,
2426                                  &deref, sizeof(deref), node, &r);
2427         if (ret < 0) {
2428                 mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF DONE "
2429                                 " to node %u\n", dlm->name, namelen,
2430                                 lockname, ret, node);
2431         } else if (r < 0) {
2432                 /* ignore the error */
2433                 mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2434                      dlm->name, namelen, lockname, node, r);
2435                 dlm_print_one_lock_resource(res);
2436         }
2437 }
2438
2439 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2440 {
2441         struct dlm_ctxt *dlm;
2442         struct dlm_lock_resource *res;
2443         u8 node;
2444         u8 cleared = 0;
2445
2446         dlm = item->dlm;
2447         res = item->u.dl.deref_res;
2448         node = item->u.dl.deref_node;
2449
2450         spin_lock(&res->spinlock);
2451         BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2452         __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2453         if (test_bit(node, res->refmap)) {
2454                 dlm_lockres_clear_refmap_bit(dlm, res, node);
2455                 cleared = 1;
2456         }
2457         spin_unlock(&res->spinlock);
2458
2459         dlm_drop_lockres_ref_done(dlm, res, node);
2460
2461         if (cleared) {
2462                 mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2463                      dlm->name, res->lockname.len, res->lockname.name, node);
2464                 dlm_lockres_calc_usage(dlm, res);
2465         } else {
2466                 mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2467                      "but it is already dropped!\n", dlm->name,
2468                      res->lockname.len, res->lockname.name, node);
2469                 dlm_print_one_lock_resource(res);
2470         }
2471
2472         dlm_lockres_put(res);
2473 }
2474
2475 /*
2476  * A migratable resource is one that is :
2477  * 1. locally mastered, and,
2478  * 2. zero local locks, and,
2479  * 3. one or more non-local locks, or, one or more references
2480  * Returns 1 if yes, 0 if not.
2481  */
2482 static int dlm_is_lockres_migratable(struct dlm_ctxt *dlm,
2483                                       struct dlm_lock_resource *res)
2484 {
2485         enum dlm_lockres_list idx;
2486         int nonlocal = 0, node_ref;
2487         struct list_head *queue;
2488         struct dlm_lock *lock;
2489         u64 cookie;
2490
2491         assert_spin_locked(&res->spinlock);
2492
2493         /* delay migration when the lockres is in MIGRATING state */
2494         if (res->state & DLM_LOCK_RES_MIGRATING)
2495                 return 0;
2496
2497         /* delay migration when the lockres is in RECOCERING state */
2498         if (res->state & (DLM_LOCK_RES_RECOVERING|
2499                         DLM_LOCK_RES_RECOVERY_WAITING))
2500                 return 0;
2501
2502         if (res->owner != dlm->node_num)
2503                 return 0;
2504
2505         for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2506                 queue = dlm_list_idx_to_ptr(res, idx);
2507                 list_for_each_entry(lock, queue, list) {
2508                         if (lock->ml.node != dlm->node_num) {
2509                                 nonlocal++;
2510                                 continue;
2511                         }
2512                         cookie = be64_to_cpu(lock->ml.cookie);
2513                         mlog(0, "%s: Not migratable res %.*s, lock %u:%llu on "
2514                              "%s list\n", dlm->name, res->lockname.len,
2515                              res->lockname.name,
2516                              dlm_get_lock_cookie_node(cookie),
2517                              dlm_get_lock_cookie_seq(cookie),
2518                              dlm_list_in_text(idx));
2519                         return 0;
2520                 }
2521         }
2522
2523         if (!nonlocal) {
2524                 node_ref = find_first_bit(res->refmap, O2NM_MAX_NODES);
2525                 if (node_ref >= O2NM_MAX_NODES)
2526                         return 0;
2527         }
2528
2529         mlog(0, "%s: res %.*s, Migratable\n", dlm->name, res->lockname.len,
2530              res->lockname.name);
2531
2532         return 1;
2533 }
2534
2535 /*
2536  * DLM_MIGRATE_LOCKRES
2537  */
2538
2539
2540 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2541                                struct dlm_lock_resource *res, u8 target)
2542 {
2543         struct dlm_master_list_entry *mle = NULL;
2544         struct dlm_master_list_entry *oldmle = NULL;
2545         struct dlm_migratable_lockres *mres = NULL;
2546         int ret = 0;
2547         const char *name;
2548         unsigned int namelen;
2549         int mle_added = 0;
2550         int wake = 0;
2551
2552         if (!dlm_grab(dlm))
2553                 return -EINVAL;
2554
2555         name = res->lockname.name;
2556         namelen = res->lockname.len;
2557
2558         mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name,
2559              target);
2560
2561         /* preallocate up front. if this fails, abort */
2562         ret = -ENOMEM;
2563         mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2564         if (!mres) {
2565                 mlog_errno(ret);
2566                 goto leave;
2567         }
2568
2569         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2570         if (!mle) {
2571                 mlog_errno(ret);
2572                 goto leave;
2573         }
2574         ret = 0;
2575
2576         /*
2577          * clear any existing master requests and
2578          * add the migration mle to the list
2579          */
2580         spin_lock(&dlm->spinlock);
2581         spin_lock(&dlm->master_lock);
2582         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2583                                     namelen, target, dlm->node_num);
2584         /* get an extra reference on the mle.
2585          * otherwise the assert_master from the new
2586          * master will destroy this.
2587          */
2588         if (ret != -EEXIST)
2589                 dlm_get_mle_inuse(mle);
2590
2591         spin_unlock(&dlm->master_lock);
2592         spin_unlock(&dlm->spinlock);
2593
2594         if (ret == -EEXIST) {
2595                 mlog(0, "another process is already migrating it\n");
2596                 goto fail;
2597         }
2598         mle_added = 1;
2599
2600         /*
2601          * set the MIGRATING flag and flush asts
2602          * if we fail after this we need to re-dirty the lockres
2603          */
2604         if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2605                 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2606                      "the target went down.\n", res->lockname.len,
2607                      res->lockname.name, target);
2608                 spin_lock(&res->spinlock);
2609                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2610                 wake = 1;
2611                 spin_unlock(&res->spinlock);
2612                 ret = -EINVAL;
2613         }
2614
2615 fail:
2616         if (ret != -EEXIST && oldmle) {
2617                 /* master is known, detach if not already detached */
2618                 dlm_mle_detach_hb_events(dlm, oldmle);
2619                 dlm_put_mle(oldmle);
2620         }
2621
2622         if (ret < 0) {
2623                 if (mle_added) {
2624                         dlm_mle_detach_hb_events(dlm, mle);
2625                         dlm_put_mle(mle);
2626                         dlm_put_mle_inuse(mle);
2627                 } else if (mle) {
2628                         kmem_cache_free(dlm_mle_cache, mle);
2629                         mle = NULL;
2630                 }
2631                 goto leave;
2632         }
2633
2634         /*
2635          * at this point, we have a migration target, an mle
2636          * in the master list, and the MIGRATING flag set on
2637          * the lockres
2638          */
2639
2640         /* now that remote nodes are spinning on the MIGRATING flag,
2641          * ensure that all assert_master work is flushed. */
2642         flush_workqueue(dlm->dlm_worker);
2643
2644         /* notify new node and send all lock state */
2645         /* call send_one_lockres with migration flag.
2646          * this serves as notice to the target node that a
2647          * migration is starting. */
2648         ret = dlm_send_one_lockres(dlm, res, mres, target,
2649                                    DLM_MRES_MIGRATION);
2650
2651         if (ret < 0) {
2652                 mlog(0, "migration to node %u failed with %d\n",
2653                      target, ret);
2654                 /* migration failed, detach and clean up mle */
2655                 dlm_mle_detach_hb_events(dlm, mle);
2656                 dlm_put_mle(mle);
2657                 dlm_put_mle_inuse(mle);
2658                 spin_lock(&res->spinlock);
2659                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2660                 wake = 1;
2661                 spin_unlock(&res->spinlock);
2662                 if (dlm_is_host_down(ret))
2663                         dlm_wait_for_node_death(dlm, target,
2664                                                 DLM_NODE_DEATH_WAIT_MAX);
2665                 goto leave;
2666         }
2667
2668         /* at this point, the target sends a message to all nodes,
2669          * (using dlm_do_migrate_request).  this node is skipped since
2670          * we had to put an mle in the list to begin the process.  this
2671          * node now waits for target to do an assert master.  this node
2672          * will be the last one notified, ensuring that the migration
2673          * is complete everywhere.  if the target dies while this is
2674          * going on, some nodes could potentially see the target as the
2675          * master, so it is important that my recovery finds the migration
2676          * mle and sets the master to UNKNOWN. */
2677
2678
2679         /* wait for new node to assert master */
2680         while (1) {
2681                 ret = wait_event_interruptible_timeout(mle->wq,
2682                                         (atomic_read(&mle->woken) == 1),
2683                                         msecs_to_jiffies(5000));
2684
2685                 if (ret >= 0) {
2686                         if (atomic_read(&mle->woken) == 1 ||
2687                             res->owner == target)
2688                                 break;
2689
2690                         mlog(0, "%s:%.*s: timed out during migration\n",
2691                              dlm->name, res->lockname.len, res->lockname.name);
2692                         /* avoid hang during shutdown when migrating lockres
2693                          * to a node which also goes down */
2694                         if (dlm_is_node_dead(dlm, target)) {
2695                                 mlog(0, "%s:%.*s: expected migration "
2696                                      "target %u is no longer up, restarting\n",
2697                                      dlm->name, res->lockname.len,
2698                                      res->lockname.name, target);
2699                                 ret = -EINVAL;
2700                                 /* migration failed, detach and clean up mle */
2701                                 dlm_mle_detach_hb_events(dlm, mle);
2702                                 dlm_put_mle(mle);
2703                                 dlm_put_mle_inuse(mle);
2704                                 spin_lock(&res->spinlock);
2705                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
2706                                 wake = 1;
2707                                 spin_unlock(&res->spinlock);
2708                                 goto leave;
2709                         }
2710                 } else
2711                         mlog(0, "%s:%.*s: caught signal during migration\n",
2712                              dlm->name, res->lockname.len, res->lockname.name);
2713         }
2714
2715         /* all done, set the owner, clear the flag */
2716         spin_lock(&res->spinlock);
2717         dlm_set_lockres_owner(dlm, res, target);
2718         res->state &= ~DLM_LOCK_RES_MIGRATING;
2719         dlm_remove_nonlocal_locks(dlm, res);
2720         spin_unlock(&res->spinlock);
2721         wake_up(&res->wq);
2722
2723         /* master is known, detach if not already detached */
2724         dlm_mle_detach_hb_events(dlm, mle);
2725         dlm_put_mle_inuse(mle);
2726         ret = 0;
2727
2728         dlm_lockres_calc_usage(dlm, res);
2729
2730 leave:
2731         /* re-dirty the lockres if we failed */
2732         if (ret < 0)
2733                 dlm_kick_thread(dlm, res);
2734
2735         /* wake up waiters if the MIGRATING flag got set
2736          * but migration failed */
2737         if (wake)
2738                 wake_up(&res->wq);
2739
2740         if (mres)
2741                 free_page((unsigned long)mres);
2742
2743         dlm_put(dlm);
2744
2745         mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen,
2746              name, target, ret);
2747         return ret;
2748 }
2749
2750 /*
2751  * Should be called only after beginning the domain leave process.
2752  * There should not be any remaining locks on nonlocal lock resources,
2753  * and there should be no local locks left on locally mastered resources.
2754  *
2755  * Called with the dlm spinlock held, may drop it to do migration, but
2756  * will re-acquire before exit.
2757  *
2758  * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2759  */
2760 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2761         __must_hold(&dlm->spinlock)
2762 {
2763         int ret;
2764         int lock_dropped = 0;
2765         u8 target = O2NM_MAX_NODES;
2766
2767         assert_spin_locked(&dlm->spinlock);
2768
2769         spin_lock(&res->spinlock);
2770         if (dlm_is_lockres_migratable(dlm, res))
2771                 target = dlm_pick_migration_target(dlm, res);
2772         spin_unlock(&res->spinlock);
2773
2774         if (target == O2NM_MAX_NODES)
2775                 goto leave;
2776
2777         /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2778         spin_unlock(&dlm->spinlock);
2779         lock_dropped = 1;
2780         ret = dlm_migrate_lockres(dlm, res, target);
2781         if (ret)
2782                 mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2783                      dlm->name, res->lockname.len, res->lockname.name,
2784                      target, ret);
2785         spin_lock(&dlm->spinlock);
2786 leave:
2787         return lock_dropped;
2788 }
2789
2790 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2791 {
2792         int ret;
2793         spin_lock(&dlm->ast_lock);
2794         spin_lock(&lock->spinlock);
2795         ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2796         spin_unlock(&lock->spinlock);
2797         spin_unlock(&dlm->ast_lock);
2798         return ret;
2799 }
2800
2801 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2802                                      struct dlm_lock_resource *res,
2803                                      u8 mig_target)
2804 {
2805         int can_proceed;
2806         spin_lock(&res->spinlock);
2807         can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2808         spin_unlock(&res->spinlock);
2809
2810         /* target has died, so make the caller break out of the
2811          * wait_event, but caller must recheck the domain_map */
2812         spin_lock(&dlm->spinlock);
2813         if (!test_bit(mig_target, dlm->domain_map))
2814                 can_proceed = 1;
2815         spin_unlock(&dlm->spinlock);
2816         return can_proceed;
2817 }
2818
2819 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2820                                 struct dlm_lock_resource *res)
2821 {
2822         int ret;
2823         spin_lock(&res->spinlock);
2824         ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2825         spin_unlock(&res->spinlock);
2826         return ret;
2827 }
2828
2829
2830 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2831                                        struct dlm_lock_resource *res,
2832                                        u8 target)
2833 {
2834         int ret = 0;
2835
2836         mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2837                res->lockname.len, res->lockname.name, dlm->node_num,
2838                target);
2839         /* need to set MIGRATING flag on lockres.  this is done by
2840          * ensuring that all asts have been flushed for this lockres. */
2841         spin_lock(&res->spinlock);
2842         BUG_ON(res->migration_pending);
2843         res->migration_pending = 1;
2844         /* strategy is to reserve an extra ast then release
2845          * it below, letting the release do all of the work */
2846         __dlm_lockres_reserve_ast(res);
2847         spin_unlock(&res->spinlock);
2848
2849         /* now flush all the pending asts */
2850         dlm_kick_thread(dlm, res);
2851         /* before waiting on DIRTY, block processes which may
2852          * try to dirty the lockres before MIGRATING is set */
2853         spin_lock(&res->spinlock);
2854         BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2855         res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2856         spin_unlock(&res->spinlock);
2857         /* now wait on any pending asts and the DIRTY state */
2858         wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2859         dlm_lockres_release_ast(dlm, res);
2860
2861         mlog(0, "about to wait on migration_wq, dirty=%s\n",
2862                res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2863         /* if the extra ref we just put was the final one, this
2864          * will pass thru immediately.  otherwise, we need to wait
2865          * for the last ast to finish. */
2866 again:
2867         ret = wait_event_interruptible_timeout(dlm->migration_wq,
2868                    dlm_migration_can_proceed(dlm, res, target),
2869                    msecs_to_jiffies(1000));
2870         if (ret < 0) {
2871                 mlog(0, "woken again: migrating? %s, dead? %s\n",
2872                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2873                        test_bit(target, dlm->domain_map) ? "no":"yes");
2874         } else {
2875                 mlog(0, "all is well: migrating? %s, dead? %s\n",
2876                        res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2877                        test_bit(target, dlm->domain_map) ? "no":"yes");
2878         }
2879         if (!dlm_migration_can_proceed(dlm, res, target)) {
2880                 mlog(0, "trying again...\n");
2881                 goto again;
2882         }
2883
2884         ret = 0;
2885         /* did the target go down or die? */
2886         spin_lock(&dlm->spinlock);
2887         if (!test_bit(target, dlm->domain_map)) {
2888                 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2889                      target);
2890                 ret = -EHOSTDOWN;
2891         }
2892         spin_unlock(&dlm->spinlock);
2893
2894         /*
2895          * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2896          * another try; otherwise, we are sure the MIGRATING state is there,
2897          * drop the unneeded state which blocked threads trying to DIRTY
2898          */
2899         spin_lock(&res->spinlock);
2900         BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2901         res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2902         if (!ret)
2903                 BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2904         else
2905                 res->migration_pending = 0;
2906         spin_unlock(&res->spinlock);
2907
2908         /*
2909          * at this point:
2910          *
2911          *   o the DLM_LOCK_RES_MIGRATING flag is set if target not down
2912          *   o there are no pending asts on this lockres
2913          *   o all processes trying to reserve an ast on this
2914          *     lockres must wait for the MIGRATING flag to clear
2915          */
2916         return ret;
2917 }
2918
2919 /* last step in the migration process.
2920  * original master calls this to free all of the dlm_lock
2921  * structures that used to be for other nodes. */
2922 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2923                                       struct dlm_lock_resource *res)
2924 {
2925         struct list_head *queue = &res->granted;
2926         int i, bit;
2927         struct dlm_lock *lock, *next;
2928
2929         assert_spin_locked(&res->spinlock);
2930
2931         BUG_ON(res->owner == dlm->node_num);
2932
2933         for (i=0; i<3; i++) {
2934                 list_for_each_entry_safe(lock, next, queue, list) {
2935                         if (lock->ml.node != dlm->node_num) {
2936                                 mlog(0, "putting lock for node %u\n",
2937                                      lock->ml.node);
2938                                 /* be extra careful */
2939                                 BUG_ON(!list_empty(&lock->ast_list));
2940                                 BUG_ON(!list_empty(&lock->bast_list));
2941                                 BUG_ON(lock->ast_pending);
2942                                 BUG_ON(lock->bast_pending);
2943                                 dlm_lockres_clear_refmap_bit(dlm, res,
2944                                                              lock->ml.node);
2945                                 list_del_init(&lock->list);
2946                                 dlm_lock_put(lock);
2947                                 /* In a normal unlock, we would have added a
2948                                  * DLM_UNLOCK_FREE_LOCK action. Force it. */
2949                                 dlm_lock_put(lock);
2950                         }
2951                 }
2952                 queue++;
2953         }
2954         bit = 0;
2955         while (1) {
2956                 bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2957                 if (bit >= O2NM_MAX_NODES)
2958                         break;
2959                 /* do not clear the local node reference, if there is a
2960                  * process holding this, let it drop the ref itself */
2961                 if (bit != dlm->node_num) {
2962                         mlog(0, "%s:%.*s: node %u had a ref to this "
2963                              "migrating lockres, clearing\n", dlm->name,
2964                              res->lockname.len, res->lockname.name, bit);
2965                         dlm_lockres_clear_refmap_bit(dlm, res, bit);
2966                 }
2967                 bit++;
2968         }
2969 }
2970
2971 /*
2972  * Pick a node to migrate the lock resource to. This function selects a
2973  * potential target based first on the locks and then on refmap. It skips
2974  * nodes that are in the process of exiting the domain.
2975  */
2976 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2977                                     struct dlm_lock_resource *res)
2978 {
2979         enum dlm_lockres_list idx;
2980         struct list_head *queue;
2981         struct dlm_lock *lock;
2982         int noderef;
2983         u8 nodenum = O2NM_MAX_NODES;
2984
2985         assert_spin_locked(&dlm->spinlock);
2986         assert_spin_locked(&res->spinlock);
2987
2988         /* Go through all the locks */
2989         for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2990                 queue = dlm_list_idx_to_ptr(res, idx);
2991                 list_for_each_entry(lock, queue, list) {
2992                         if (lock->ml.node == dlm->node_num)
2993                                 continue;
2994                         if (test_bit(lock->ml.node, dlm->exit_domain_map))
2995                                 continue;
2996                         nodenum = lock->ml.node;
2997                         goto bail;
2998                 }
2999         }
3000
3001         /* Go thru the refmap */
3002         noderef = -1;
3003         while (1) {
3004                 noderef = find_next_bit(res->refmap, O2NM_MAX_NODES,
3005                                         noderef + 1);
3006                 if (noderef >= O2NM_MAX_NODES)
3007                         break;
3008                 if (noderef == dlm->node_num)
3009                         continue;
3010                 if (test_bit(noderef, dlm->exit_domain_map))
3011                         continue;
3012                 nodenum = noderef;
3013                 goto bail;
3014         }
3015
3016 bail:
3017         return nodenum;
3018 }
3019
3020 /* this is called by the new master once all lockres
3021  * data has been received */
3022 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
3023                                   struct dlm_lock_resource *res,
3024                                   u8 master, u8 new_master,
3025                                   struct dlm_node_iter *iter)
3026 {
3027         struct dlm_migrate_request migrate;
3028         int ret, skip, status = 0;
3029         int nodenum;
3030
3031         memset(&migrate, 0, sizeof(migrate));
3032         migrate.namelen = res->lockname.len;
3033         memcpy(migrate.name, res->lockname.name, migrate.namelen);
3034         migrate.new_master = new_master;
3035         migrate.master = master;
3036
3037         ret = 0;
3038
3039         /* send message to all nodes, except the master and myself */
3040         while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
3041                 if (nodenum == master ||
3042                     nodenum == new_master)
3043                         continue;
3044
3045                 /* We could race exit domain. If exited, skip. */
3046                 spin_lock(&dlm->spinlock);
3047                 skip = (!test_bit(nodenum, dlm->domain_map));
3048                 spin_unlock(&dlm->spinlock);
3049                 if (skip) {
3050                         clear_bit(nodenum, iter->node_map);
3051                         continue;
3052                 }
3053
3054                 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
3055                                          &migrate, sizeof(migrate), nodenum,
3056                                          &status);
3057                 if (ret < 0) {
3058                         mlog(ML_ERROR, "%s: res %.*s, Error %d send "
3059                              "MIGRATE_REQUEST to node %u\n", dlm->name,
3060                              migrate.namelen, migrate.name, ret, nodenum);
3061                         if (!dlm_is_host_down(ret)) {
3062                                 mlog(ML_ERROR, "unhandled error=%d!\n", ret);
3063                                 BUG();
3064                         }
3065                         clear_bit(nodenum, iter->node_map);
3066                         ret = 0;
3067                 } else if (status < 0) {
3068                         mlog(0, "migrate request (node %u) returned %d!\n",
3069                              nodenum, status);
3070                         ret = status;
3071                 } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
3072                         /* during the migration request we short-circuited
3073                          * the mastery of the lockres.  make sure we have
3074                          * a mastery ref for nodenum */
3075                         mlog(0, "%s:%.*s: need ref for node %u\n",
3076                              dlm->name, res->lockname.len, res->lockname.name,
3077                              nodenum);
3078                         spin_lock(&res->spinlock);
3079                         dlm_lockres_set_refmap_bit(dlm, res, nodenum);
3080                         spin_unlock(&res->spinlock);
3081                 }
3082         }
3083
3084         if (ret < 0)
3085                 mlog_errno(ret);
3086
3087         mlog(0, "returning ret=%d\n", ret);
3088         return ret;
3089 }
3090
3091
3092 /* if there is an existing mle for this lockres, we now know who the master is.
3093  * (the one who sent us *this* message) we can clear it up right away.
3094  * since the process that put the mle on the list still has a reference to it,
3095  * we can unhash it now, set the master and wake the process.  as a result,
3096  * we will have no mle in the list to start with.  now we can add an mle for
3097  * the migration and this should be the only one found for those scanning the
3098  * list.  */
3099 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
3100                                 void **ret_data)
3101 {
3102         struct dlm_ctxt *dlm = data;
3103         struct dlm_lock_resource *res = NULL;
3104         struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
3105         struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
3106         const char *name;
3107         unsigned int namelen, hash;
3108         int ret = 0;
3109
3110         if (!dlm_grab(dlm))
3111                 return 0;
3112
3113         name = migrate->name;
3114         namelen = migrate->namelen;
3115         hash = dlm_lockid_hash(name, namelen);
3116
3117         /* preallocate.. if this fails, abort */
3118         mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
3119
3120         if (!mle) {
3121                 ret = -ENOMEM;
3122                 goto leave;
3123         }
3124
3125         /* check for pre-existing lock */
3126         spin_lock(&dlm->spinlock);
3127         res = __dlm_lookup_lockres(dlm, name, namelen, hash);
3128         if (res) {
3129                 spin_lock(&res->spinlock);
3130                 if (res->state & DLM_LOCK_RES_RECOVERING) {
3131                         /* if all is working ok, this can only mean that we got
3132                         * a migrate request from a node that we now see as
3133                         * dead.  what can we do here?  drop it to the floor? */
3134                         spin_unlock(&res->spinlock);
3135                         mlog(ML_ERROR, "Got a migrate request, but the "
3136                              "lockres is marked as recovering!");
3137                         kmem_cache_free(dlm_mle_cache, mle);
3138                         ret = -EINVAL; /* need a better solution */
3139                         goto unlock;
3140                 }
3141                 res->state |= DLM_LOCK_RES_MIGRATING;
3142                 spin_unlock(&res->spinlock);
3143         }
3144
3145         spin_lock(&dlm->master_lock);
3146         /* ignore status.  only nonzero status would BUG. */
3147         ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3148                                     name, namelen,
3149                                     migrate->new_master,
3150                                     migrate->master);
3151
3152         if (ret < 0)
3153                 kmem_cache_free(dlm_mle_cache, mle);
3154
3155         spin_unlock(&dlm->master_lock);
3156 unlock:
3157         spin_unlock(&dlm->spinlock);
3158
3159         if (oldmle) {
3160                 /* master is known, detach if not already detached */
3161                 dlm_mle_detach_hb_events(dlm, oldmle);
3162                 dlm_put_mle(oldmle);
3163         }
3164
3165         if (res)
3166                 dlm_lockres_put(res);
3167 leave:
3168         dlm_put(dlm);
3169         return ret;
3170 }
3171
3172 /* must be holding dlm->spinlock and dlm->master_lock
3173  * when adding a migration mle, we can clear any other mles
3174  * in the master list because we know with certainty that
3175  * the master is "master".  so we remove any old mle from
3176  * the list after setting it's master field, and then add
3177  * the new migration mle.  this way we can hold with the rule
3178  * of having only one mle for a given lock name at all times. */
3179 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3180                                  struct dlm_lock_resource *res,
3181                                  struct dlm_master_list_entry *mle,
3182                                  struct dlm_master_list_entry **oldmle,
3183                                  const char *name, unsigned int namelen,
3184                                  u8 new_master, u8 master)
3185 {
3186         int found;
3187         int ret = 0;
3188
3189         *oldmle = NULL;
3190
3191         assert_spin_locked(&dlm->spinlock);
3192         assert_spin_locked(&dlm->master_lock);
3193
3194         /* caller is responsible for any ref taken here on oldmle */
3195         found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3196         if (found) {
3197                 struct dlm_master_list_entry *tmp = *oldmle;
3198                 spin_lock(&tmp->spinlock);
3199                 if (tmp->type == DLM_MLE_MIGRATION) {
3200                         if (master == dlm->node_num) {
3201                                 /* ah another process raced me to it */
3202                                 mlog(0, "tried to migrate %.*s, but some "
3203                                      "process beat me to it\n",
3204                                      namelen, name);
3205                                 spin_unlock(&tmp->spinlock);
3206                                 return -EEXIST;
3207                         } else {
3208                                 /* bad.  2 NODES are trying to migrate! */
3209                                 mlog(ML_ERROR, "migration error  mle: "
3210                                      "master=%u new_master=%u // request: "
3211                                      "master=%u new_master=%u // "
3212                                      "lockres=%.*s\n",
3213                                      tmp->master, tmp->new_master,
3214                                      master, new_master,
3215                                      namelen, name);
3216                                 BUG();
3217                         }
3218                 } else {
3219                         /* this is essentially what assert_master does */
3220                         tmp->master = master;
3221                         atomic_set(&tmp->woken, 1);
3222                         wake_up(&tmp->wq);
3223                         /* remove it so that only one mle will be found */
3224                         __dlm_unlink_mle(dlm, tmp);
3225                         __dlm_mle_detach_hb_events(dlm, tmp);
3226                         if (tmp->type == DLM_MLE_MASTER) {
3227                                 ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3228                                 mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3229                                                 "telling master to get ref "
3230                                                 "for cleared out mle during "
3231                                                 "migration\n", dlm->name,
3232                                                 namelen, name, master,
3233                                                 new_master);
3234                         }
3235                 }
3236                 spin_unlock(&tmp->spinlock);
3237         }
3238
3239         /* now add a migration mle to the tail of the list */
3240         dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3241         mle->new_master = new_master;
3242         /* the new master will be sending an assert master for this.
3243          * at that point we will get the refmap reference */
3244         mle->master = master;
3245         /* do this for consistency with other mle types */
3246         set_bit(new_master, mle->maybe_map);
3247         __dlm_insert_mle(dlm, mle);
3248
3249         return ret;
3250 }
3251
3252 /*
3253  * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3254  */
3255 static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3256                                         struct dlm_master_list_entry *mle)
3257 {
3258         struct dlm_lock_resource *res;
3259
3260         /* Find the lockres associated to the mle and set its owner to UNK */
3261         res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3262                                    mle->mnamehash);
3263         if (res) {
3264                 spin_unlock(&dlm->master_lock);
3265
3266                 /* move lockres onto recovery list */
3267                 spin_lock(&res->spinlock);
3268                 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3269                 dlm_move_lockres_to_recovery_list(dlm, res);
3270                 spin_unlock(&res->spinlock);
3271                 dlm_lockres_put(res);
3272
3273                 /* about to get rid of mle, detach from heartbeat */
3274                 __dlm_mle_detach_hb_events(dlm, mle);
3275
3276                 /* dump the mle */
3277                 spin_lock(&dlm->master_lock);
3278                 __dlm_put_mle(mle);
3279                 spin_unlock(&dlm->master_lock);
3280         }
3281
3282         return res;
3283 }
3284
3285 static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3286                                     struct dlm_master_list_entry *mle)
3287 {
3288         __dlm_mle_detach_hb_events(dlm, mle);
3289
3290         spin_lock(&mle->spinlock);
3291         __dlm_unlink_mle(dlm, mle);
3292         atomic_set(&mle->woken, 1);
3293         spin_unlock(&mle->spinlock);
3294
3295         wake_up(&mle->wq);
3296 }
3297
3298 static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3299                                 struct dlm_master_list_entry *mle, u8 dead_node)
3300 {
3301         int bit;
3302
3303         BUG_ON(mle->type != DLM_MLE_BLOCK);
3304
3305         spin_lock(&mle->spinlock);
3306         bit = find_first_bit(mle->maybe_map, O2NM_MAX_NODES);
3307         if (bit != dead_node) {
3308                 mlog(0, "mle found, but dead node %u would not have been "
3309                      "master\n", dead_node);
3310                 spin_unlock(&mle->spinlock);
3311         } else {
3312                 /* Must drop the refcount by one since the assert_master will
3313                  * never arrive. This may result in the mle being unlinked and
3314                  * freed, but there may still be a process waiting in the
3315                  * dlmlock path which is fine. */
3316                 mlog(0, "node %u was expected master\n", dead_node);
3317                 atomic_set(&mle->woken, 1);
3318                 spin_unlock(&mle->spinlock);
3319                 wake_up(&mle->wq);
3320
3321                 /* Do not need events any longer, so detach from heartbeat */
3322                 __dlm_mle_detach_hb_events(dlm, mle);
3323                 __dlm_put_mle(mle);
3324         }
3325 }
3326
3327 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3328 {
3329         struct dlm_master_list_entry *mle;
3330         struct dlm_lock_resource *res;
3331         struct hlist_head *bucket;
3332         struct hlist_node *tmp;
3333         unsigned int i;
3334
3335         mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
3336 top:
3337         assert_spin_locked(&dlm->spinlock);
3338
3339         /* clean the master list */
3340         spin_lock(&dlm->master_lock);
3341         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3342                 bucket = dlm_master_hash(dlm, i);
3343                 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3344                         BUG_ON(mle->type != DLM_MLE_BLOCK &&
3345                                mle->type != DLM_MLE_MASTER &&
3346                                mle->type != DLM_MLE_MIGRATION);
3347
3348                         /* MASTER mles are initiated locally. The waiting
3349                          * process will notice the node map change shortly.
3350                          * Let that happen as normal. */
3351                         if (mle->type == DLM_MLE_MASTER)
3352                                 continue;
3353
3354                         /* BLOCK mles are initiated by other nodes. Need to
3355                          * clean up if the dead node would have been the
3356                          * master. */
3357                         if (mle->type == DLM_MLE_BLOCK) {
3358                                 dlm_clean_block_mle(dlm, mle, dead_node);
3359                                 continue;
3360                         }
3361
3362                         /* Everything else is a MIGRATION mle */
3363
3364                         /* The rule for MIGRATION mles is that the master
3365                          * becomes UNKNOWN if *either* the original or the new
3366                          * master dies. All UNKNOWN lockres' are sent to
3367                          * whichever node becomes the recovery master. The new
3368                          * master is responsible for determining if there is
3369                          * still a master for this lockres, or if he needs to
3370                          * take over mastery. Either way, this node should
3371                          * expect another message to resolve this. */
3372
3373                         if (mle->master != dead_node &&
3374                             mle->new_master != dead_node)
3375                                 continue;
3376
3377                         if (mle->new_master == dead_node && mle->inuse) {
3378                                 mlog(ML_NOTICE, "%s: target %u died during "
3379                                                 "migration from %u, the MLE is "
3380                                                 "still keep used, ignore it!\n",
3381                                                 dlm->name, dead_node,
3382                                                 mle->master);
3383                                 continue;
3384                         }
3385
3386                         /* If we have reached this point, this mle needs to be
3387                          * removed from the list and freed. */
3388                         dlm_clean_migration_mle(dlm, mle);
3389
3390                         mlog(0, "%s: node %u died during migration from "
3391                              "%u to %u!\n", dlm->name, dead_node, mle->master,
3392                              mle->new_master);
3393
3394                         /* If we find a lockres associated with the mle, we've
3395                          * hit this rare case that messes up our lock ordering.
3396                          * If so, we need to drop the master lock so that we can
3397                          * take the lockres lock, meaning that we will have to
3398                          * restart from the head of list. */
3399                         res = dlm_reset_mleres_owner(dlm, mle);
3400                         if (res)
3401                                 /* restart */
3402                                 goto top;
3403
3404                         /* This may be the last reference */
3405                         __dlm_put_mle(mle);
3406                 }
3407         }
3408         spin_unlock(&dlm->master_lock);
3409 }
3410
3411 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3412                          u8 old_master)
3413 {
3414         struct dlm_node_iter iter;
3415         int ret = 0;
3416
3417         spin_lock(&dlm->spinlock);
3418         dlm_node_iter_init(dlm->domain_map, &iter);
3419         clear_bit(old_master, iter.node_map);
3420         clear_bit(dlm->node_num, iter.node_map);
3421         spin_unlock(&dlm->spinlock);
3422
3423         /* ownership of the lockres is changing.  account for the
3424          * mastery reference here since old_master will briefly have
3425          * a reference after the migration completes */
3426         spin_lock(&res->spinlock);
3427         dlm_lockres_set_refmap_bit(dlm, res, old_master);
3428         spin_unlock(&res->spinlock);
3429
3430         mlog(0, "now time to do a migrate request to other nodes\n");
3431         ret = dlm_do_migrate_request(dlm, res, old_master,
3432                                      dlm->node_num, &iter);
3433         if (ret < 0) {
3434                 mlog_errno(ret);
3435                 goto leave;
3436         }
3437
3438         mlog(0, "doing assert master of %.*s to all except the original node\n",
3439              res->lockname.len, res->lockname.name);
3440         /* this call now finishes out the nodemap
3441          * even if one or more nodes die */
3442         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3443                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3444         if (ret < 0) {
3445                 /* no longer need to retry.  all living nodes contacted. */
3446                 mlog_errno(ret);
3447                 ret = 0;
3448         }
3449
3450         memset(iter.node_map, 0, sizeof(iter.node_map));
3451         set_bit(old_master, iter.node_map);
3452         mlog(0, "doing assert master of %.*s back to %u\n",
3453              res->lockname.len, res->lockname.name, old_master);
3454         ret = dlm_do_assert_master(dlm, res, iter.node_map,
3455                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
3456         if (ret < 0) {
3457                 mlog(0, "assert master to original master failed "
3458                      "with %d.\n", ret);
3459                 /* the only nonzero status here would be because of
3460                  * a dead original node.  we're done. */
3461                 ret = 0;
3462         }
3463
3464         /* all done, set the owner, clear the flag */
3465         spin_lock(&res->spinlock);
3466         dlm_set_lockres_owner(dlm, res, dlm->node_num);
3467         res->state &= ~DLM_LOCK_RES_MIGRATING;
3468         spin_unlock(&res->spinlock);
3469         /* re-dirty it on the new master */
3470         dlm_kick_thread(dlm, res);
3471         wake_up(&res->wq);
3472 leave:
3473         return ret;
3474 }
3475
3476 /*
3477  * LOCKRES AST REFCOUNT
3478  * this is integral to migration
3479  */
3480
3481 /* for future intent to call an ast, reserve one ahead of time.
3482  * this should be called only after waiting on the lockres
3483  * with dlm_wait_on_lockres, and while still holding the
3484  * spinlock after the call. */
3485 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3486 {
3487         assert_spin_locked(&res->spinlock);
3488         if (res->state & DLM_LOCK_RES_MIGRATING) {
3489                 __dlm_print_one_lock_resource(res);
3490         }
3491         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3492
3493         atomic_inc(&res->asts_reserved);
3494 }
3495
3496 /*
3497  * used to drop the reserved ast, either because it went unused,
3498  * or because the ast/bast was actually called.
3499  *
3500  * also, if there is a pending migration on this lockres,
3501  * and this was the last pending ast on the lockres,
3502  * atomically set the MIGRATING flag before we drop the lock.
3503  * this is how we ensure that migration can proceed with no
3504  * asts in progress.  note that it is ok if the state of the
3505  * queues is such that a lock should be granted in the future
3506  * or that a bast should be fired, because the new master will
3507  * shuffle the lists on this lockres as soon as it is migrated.
3508  */
3509 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3510                              struct dlm_lock_resource *res)
3511 {
3512         if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3513                 return;
3514
3515         if (!res->migration_pending) {
3516                 spin_unlock(&res->spinlock);
3517                 return;
3518         }
3519
3520         BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3521         res->migration_pending = 0;
3522         res->state |= DLM_LOCK_RES_MIGRATING;
3523         spin_unlock(&res->spinlock);
3524         wake_up(&res->wq);
3525         wake_up(&dlm->migration_wq);
3526 }
3527
3528 void dlm_force_free_mles(struct dlm_ctxt *dlm)
3529 {
3530         int i;
3531         struct hlist_head *bucket;
3532         struct dlm_master_list_entry *mle;
3533         struct hlist_node *tmp;
3534
3535         /*
3536          * We notified all other nodes that we are exiting the domain and
3537          * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3538          * around we force free them and wake any processes that are waiting
3539          * on the mles
3540          */
3541         spin_lock(&dlm->spinlock);
3542         spin_lock(&dlm->master_lock);
3543
3544         BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3545         BUG_ON((find_first_bit(dlm->domain_map, O2NM_MAX_NODES) < O2NM_MAX_NODES));
3546
3547         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3548                 bucket = dlm_master_hash(dlm, i);
3549                 hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3550                         if (mle->type != DLM_MLE_BLOCK) {
3551                                 mlog(ML_ERROR, "bad mle: %p\n", mle);
3552                                 dlm_print_one_mle(mle);
3553                         }
3554                         atomic_set(&mle->woken, 1);
3555                         wake_up(&mle->wq);
3556
3557                         __dlm_unlink_mle(dlm, mle);
3558                         __dlm_mle_detach_hb_events(dlm, mle);
3559                         __dlm_put_mle(mle);
3560                 }
3561         }
3562         spin_unlock(&dlm->master_lock);
3563         spin_unlock(&dlm->spinlock);
3564 }