GNU Linux-libre 6.1.86-gnu
[releases.git] / fs / ocfs2 / dlm / dlmrecovery.c
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * dlmrecovery.c
4  *
5  * recovery stuff
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/timer.h>
23 #include <linux/kthread.h>
24 #include <linux/delay.h>
25
26
27 #include "../cluster/heartbeat.h"
28 #include "../cluster/nodemanager.h"
29 #include "../cluster/tcp.h"
30
31 #include "dlmapi.h"
32 #include "dlmcommon.h"
33 #include "dlmdomain.h"
34
35 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY)
36 #include "../cluster/masklog.h"
37
38 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node);
39
40 static int dlm_recovery_thread(void *data);
41 static int dlm_do_recovery(struct dlm_ctxt *dlm);
42
43 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm);
44 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node);
45 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
46 static int dlm_request_all_locks(struct dlm_ctxt *dlm,
47                                  u8 request_from, u8 dead_node);
48 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm);
49
50 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res);
51 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
52                                         const char *lockname, int namelen,
53                                         int total_locks, u64 cookie,
54                                         u8 flags, u8 master);
55 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
56                                     struct dlm_migratable_lockres *mres,
57                                     u8 send_to,
58                                     struct dlm_lock_resource *res,
59                                     int total_locks);
60 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
61                                      struct dlm_lock_resource *res,
62                                      struct dlm_migratable_lockres *mres);
63 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm);
64 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm,
65                                  u8 dead_node, u8 send_to);
66 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node);
67 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
68                                         struct list_head *list, u8 dead_node);
69 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
70                                               u8 dead_node, u8 new_master);
71 static void dlm_reco_ast(void *astdata);
72 static void dlm_reco_bast(void *astdata, int blocked_type);
73 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st);
74 static void dlm_request_all_locks_worker(struct dlm_work_item *item,
75                                          void *data);
76 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data);
77 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
78                                       struct dlm_lock_resource *res,
79                                       u8 *real_master);
80
81 static u64 dlm_get_next_mig_cookie(void);
82
83 static DEFINE_SPINLOCK(dlm_reco_state_lock);
84 static DEFINE_SPINLOCK(dlm_mig_cookie_lock);
85 static u64 dlm_mig_cookie = 1;
86
87 static u64 dlm_get_next_mig_cookie(void)
88 {
89         u64 c;
90         spin_lock(&dlm_mig_cookie_lock);
91         c = dlm_mig_cookie;
92         if (dlm_mig_cookie == (~0ULL))
93                 dlm_mig_cookie = 1;
94         else
95                 dlm_mig_cookie++;
96         spin_unlock(&dlm_mig_cookie_lock);
97         return c;
98 }
99
100 static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm,
101                                           u8 dead_node)
102 {
103         assert_spin_locked(&dlm->spinlock);
104         if (dlm->reco.dead_node != dead_node)
105                 mlog(0, "%s: changing dead_node from %u to %u\n",
106                      dlm->name, dlm->reco.dead_node, dead_node);
107         dlm->reco.dead_node = dead_node;
108 }
109
110 static inline void dlm_set_reco_master(struct dlm_ctxt *dlm,
111                                        u8 master)
112 {
113         assert_spin_locked(&dlm->spinlock);
114         mlog(0, "%s: changing new_master from %u to %u\n",
115              dlm->name, dlm->reco.new_master, master);
116         dlm->reco.new_master = master;
117 }
118
119 static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm)
120 {
121         assert_spin_locked(&dlm->spinlock);
122         clear_bit(dlm->reco.dead_node, dlm->recovery_map);
123         dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
124         dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
125 }
126
127 /* Worker function used during recovery. */
128 void dlm_dispatch_work(struct work_struct *work)
129 {
130         struct dlm_ctxt *dlm =
131                 container_of(work, struct dlm_ctxt, dispatched_work);
132         LIST_HEAD(tmp_list);
133         struct dlm_work_item *item, *next;
134         dlm_workfunc_t *workfunc;
135         int tot=0;
136
137         spin_lock(&dlm->work_lock);
138         list_splice_init(&dlm->work_list, &tmp_list);
139         spin_unlock(&dlm->work_lock);
140
141         list_for_each_entry(item, &tmp_list, list) {
142                 tot++;
143         }
144         mlog(0, "%s: work thread has %d work items\n", dlm->name, tot);
145
146         list_for_each_entry_safe(item, next, &tmp_list, list) {
147                 workfunc = item->func;
148                 list_del_init(&item->list);
149
150                 /* already have ref on dlm to avoid having
151                  * it disappear.  just double-check. */
152                 BUG_ON(item->dlm != dlm);
153
154                 /* this is allowed to sleep and
155                  * call network stuff */
156                 workfunc(item, item->data);
157
158                 dlm_put(dlm);
159                 kfree(item);
160         }
161 }
162
163 /*
164  * RECOVERY THREAD
165  */
166
167 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
168 {
169         /* wake the recovery thread
170          * this will wake the reco thread in one of three places
171          * 1) sleeping with no recovery happening
172          * 2) sleeping with recovery mastered elsewhere
173          * 3) recovery mastered here, waiting on reco data */
174
175         wake_up(&dlm->dlm_reco_thread_wq);
176 }
177
178 /* Launch the recovery thread */
179 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
180 {
181         mlog(0, "starting dlm recovery thread...\n");
182
183         dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
184                         "dlm_reco-%s", dlm->name);
185         if (IS_ERR(dlm->dlm_reco_thread_task)) {
186                 mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
187                 dlm->dlm_reco_thread_task = NULL;
188                 return -EINVAL;
189         }
190
191         return 0;
192 }
193
194 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
195 {
196         if (dlm->dlm_reco_thread_task) {
197                 mlog(0, "waiting for dlm recovery thread to exit\n");
198                 kthread_stop(dlm->dlm_reco_thread_task);
199                 dlm->dlm_reco_thread_task = NULL;
200         }
201 }
202
203
204
205 /*
206  * this is lame, but here's how recovery works...
207  * 1) all recovery threads cluster wide will work on recovering
208  *    ONE node at a time
209  * 2) negotiate who will take over all the locks for the dead node.
210  *    thats right... ALL the locks.
211  * 3) once a new master is chosen, everyone scans all locks
212  *    and moves aside those mastered by the dead guy
213  * 4) each of these locks should be locked until recovery is done
214  * 5) the new master collects up all of secondary lock queue info
215  *    one lock at a time, forcing each node to communicate back
216  *    before continuing
217  * 6) each secondary lock queue responds with the full known lock info
218  * 7) once the new master has run all its locks, it sends a ALLDONE!
219  *    message to everyone
220  * 8) upon receiving this message, the secondary queue node unlocks
221  *    and responds to the ALLDONE
222  * 9) once the new master gets responses from everyone, he unlocks
223  *    everything and recovery for this dead node is done
224  *10) go back to 2) while there are still dead nodes
225  *
226  */
227
228 static void dlm_print_reco_node_status(struct dlm_ctxt *dlm)
229 {
230         struct dlm_reco_node_data *ndata;
231         struct dlm_lock_resource *res;
232
233         mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n",
234              dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
235              dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive",
236              dlm->reco.dead_node, dlm->reco.new_master);
237
238         list_for_each_entry(ndata, &dlm->reco.node_data, list) {
239                 char *st = "unknown";
240                 switch (ndata->state) {
241                         case DLM_RECO_NODE_DATA_INIT:
242                                 st = "init";
243                                 break;
244                         case DLM_RECO_NODE_DATA_REQUESTING:
245                                 st = "requesting";
246                                 break;
247                         case DLM_RECO_NODE_DATA_DEAD:
248                                 st = "dead";
249                                 break;
250                         case DLM_RECO_NODE_DATA_RECEIVING:
251                                 st = "receiving";
252                                 break;
253                         case DLM_RECO_NODE_DATA_REQUESTED:
254                                 st = "requested";
255                                 break;
256                         case DLM_RECO_NODE_DATA_DONE:
257                                 st = "done";
258                                 break;
259                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
260                                 st = "finalize-sent";
261                                 break;
262                         default:
263                                 st = "bad";
264                                 break;
265                 }
266                 mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n",
267                      dlm->name, ndata->node_num, st);
268         }
269         list_for_each_entry(res, &dlm->reco.resources, recovering) {
270                 mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n",
271                      dlm->name, res->lockname.len, res->lockname.name);
272         }
273 }
274
275 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
276
277 static int dlm_recovery_thread(void *data)
278 {
279         int status;
280         struct dlm_ctxt *dlm = data;
281         unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
282
283         mlog(0, "dlm thread running for %s...\n", dlm->name);
284
285         while (!kthread_should_stop()) {
286                 if (dlm_domain_fully_joined(dlm)) {
287                         status = dlm_do_recovery(dlm);
288                         if (status == -EAGAIN) {
289                                 /* do not sleep, recheck immediately. */
290                                 continue;
291                         }
292                         if (status < 0)
293                                 mlog_errno(status);
294                 }
295
296                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
297                                                  kthread_should_stop(),
298                                                  timeout);
299         }
300
301         mlog(0, "quitting DLM recovery thread\n");
302         return 0;
303 }
304
305 /* returns true when the recovery master has contacted us */
306 static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
307 {
308         int ready;
309         spin_lock(&dlm->spinlock);
310         ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
311         spin_unlock(&dlm->spinlock);
312         return ready;
313 }
314
315 /* returns true if node is no longer in the domain
316  * could be dead or just not joined */
317 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
318 {
319         int dead;
320         spin_lock(&dlm->spinlock);
321         dead = !test_bit(node, dlm->domain_map);
322         spin_unlock(&dlm->spinlock);
323         return dead;
324 }
325
326 /* returns true if node is no longer in the domain
327  * could be dead or just not joined */
328 static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node)
329 {
330         int recovered;
331         spin_lock(&dlm->spinlock);
332         recovered = !test_bit(node, dlm->recovery_map);
333         spin_unlock(&dlm->spinlock);
334         return recovered;
335 }
336
337
338 void dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout)
339 {
340         if (dlm_is_node_dead(dlm, node))
341                 return;
342
343         printk(KERN_NOTICE "o2dlm: Waiting on the death of node %u in "
344                "domain %s\n", node, dlm->name);
345
346         if (timeout)
347                 wait_event_timeout(dlm->dlm_reco_thread_wq,
348                                    dlm_is_node_dead(dlm, node),
349                                    msecs_to_jiffies(timeout));
350         else
351                 wait_event(dlm->dlm_reco_thread_wq,
352                            dlm_is_node_dead(dlm, node));
353 }
354
355 void dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout)
356 {
357         if (dlm_is_node_recovered(dlm, node))
358                 return;
359
360         printk(KERN_NOTICE "o2dlm: Waiting on the recovery of node %u in "
361                "domain %s\n", node, dlm->name);
362
363         if (timeout)
364                 wait_event_timeout(dlm->dlm_reco_thread_wq,
365                                    dlm_is_node_recovered(dlm, node),
366                                    msecs_to_jiffies(timeout));
367         else
368                 wait_event(dlm->dlm_reco_thread_wq,
369                            dlm_is_node_recovered(dlm, node));
370 }
371
372 /* callers of the top-level api calls (dlmlock/dlmunlock) should
373  * block on the dlm->reco.event when recovery is in progress.
374  * the dlm recovery thread will set this state when it begins
375  * recovering a dead node (as the new master or not) and clear
376  * the state and wake as soon as all affected lock resources have
377  * been marked with the RECOVERY flag */
378 static int dlm_in_recovery(struct dlm_ctxt *dlm)
379 {
380         int in_recovery;
381         spin_lock(&dlm->spinlock);
382         in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
383         spin_unlock(&dlm->spinlock);
384         return in_recovery;
385 }
386
387
388 void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
389 {
390         if (dlm_in_recovery(dlm)) {
391                 mlog(0, "%s: reco thread %d in recovery: "
392                      "state=%d, master=%u, dead=%u\n",
393                      dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
394                      dlm->reco.state, dlm->reco.new_master,
395                      dlm->reco.dead_node);
396         }
397         wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
398 }
399
400 static void dlm_begin_recovery(struct dlm_ctxt *dlm)
401 {
402         assert_spin_locked(&dlm->spinlock);
403         BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
404         printk(KERN_NOTICE "o2dlm: Begin recovery on domain %s for node %u\n",
405                dlm->name, dlm->reco.dead_node);
406         dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
407 }
408
409 static void dlm_end_recovery(struct dlm_ctxt *dlm)
410 {
411         spin_lock(&dlm->spinlock);
412         BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
413         dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
414         spin_unlock(&dlm->spinlock);
415         printk(KERN_NOTICE "o2dlm: End recovery on domain %s\n", dlm->name);
416         wake_up(&dlm->reco.event);
417 }
418
419 static void dlm_print_recovery_master(struct dlm_ctxt *dlm)
420 {
421         printk(KERN_NOTICE "o2dlm: Node %u (%s) is the Recovery Master for the "
422                "dead node %u in domain %s\n", dlm->reco.new_master,
423                (dlm->node_num == dlm->reco.new_master ? "me" : "he"),
424                dlm->reco.dead_node, dlm->name);
425 }
426
427 static int dlm_do_recovery(struct dlm_ctxt *dlm)
428 {
429         int status = 0;
430         int ret;
431
432         spin_lock(&dlm->spinlock);
433
434         if (dlm->migrate_done) {
435                 mlog(0, "%s: no need do recovery after migrating all "
436                      "lock resources\n", dlm->name);
437                 spin_unlock(&dlm->spinlock);
438                 return 0;
439         }
440
441         /* check to see if the new master has died */
442         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
443             test_bit(dlm->reco.new_master, dlm->recovery_map)) {
444                 mlog(0, "new master %u died while recovering %u!\n",
445                      dlm->reco.new_master, dlm->reco.dead_node);
446                 /* unset the new_master, leave dead_node */
447                 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
448         }
449
450         /* select a target to recover */
451         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
452                 int bit;
453
454                 bit = find_first_bit(dlm->recovery_map, O2NM_MAX_NODES);
455                 if (bit >= O2NM_MAX_NODES || bit < 0)
456                         dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
457                 else
458                         dlm_set_reco_dead_node(dlm, bit);
459         } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
460                 /* BUG? */
461                 mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
462                      dlm->reco.dead_node);
463                 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
464         }
465
466         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
467                 // mlog(0, "nothing to recover!  sleeping now!\n");
468                 spin_unlock(&dlm->spinlock);
469                 /* return to main thread loop and sleep. */
470                 return 0;
471         }
472         mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n",
473              dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
474              dlm->reco.dead_node);
475
476         /* take write barrier */
477         /* (stops the list reshuffling thread, proxy ast handling) */
478         dlm_begin_recovery(dlm);
479
480         spin_unlock(&dlm->spinlock);
481
482         if (dlm->reco.new_master == dlm->node_num)
483                 goto master_here;
484
485         if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
486                 /* choose a new master, returns 0 if this node
487                  * is the master, -EEXIST if it's another node.
488                  * this does not return until a new master is chosen
489                  * or recovery completes entirely. */
490                 ret = dlm_pick_recovery_master(dlm);
491                 if (!ret) {
492                         /* already notified everyone.  go. */
493                         goto master_here;
494                 }
495                 mlog(0, "another node will master this recovery session.\n");
496         }
497
498         dlm_print_recovery_master(dlm);
499
500         /* it is safe to start everything back up here
501          * because all of the dead node's lock resources
502          * have been marked as in-recovery */
503         dlm_end_recovery(dlm);
504
505         /* sleep out in main dlm_recovery_thread loop. */
506         return 0;
507
508 master_here:
509         dlm_print_recovery_master(dlm);
510
511         status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
512         if (status < 0) {
513                 /* we should never hit this anymore */
514                 mlog(ML_ERROR, "%s: Error %d remastering locks for node %u, "
515                      "retrying.\n", dlm->name, status, dlm->reco.dead_node);
516                 /* yield a bit to allow any final network messages
517                  * to get handled on remaining nodes */
518                 msleep(100);
519         } else {
520                 /* success!  see if any other nodes need recovery */
521                 mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
522                      dlm->name, dlm->reco.dead_node, dlm->node_num);
523                 spin_lock(&dlm->spinlock);
524                 __dlm_reset_recovery(dlm);
525                 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
526                 spin_unlock(&dlm->spinlock);
527         }
528         dlm_end_recovery(dlm);
529
530         /* continue and look for another dead node */
531         return -EAGAIN;
532 }
533
534 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
535 {
536         int status = 0;
537         struct dlm_reco_node_data *ndata;
538         int all_nodes_done;
539         int destroy = 0;
540         int pass = 0;
541
542         do {
543                 /* we have become recovery master.  there is no escaping
544                  * this, so just keep trying until we get it. */
545                 status = dlm_init_recovery_area(dlm, dead_node);
546                 if (status < 0) {
547                         mlog(ML_ERROR, "%s: failed to alloc recovery area, "
548                              "retrying\n", dlm->name);
549                         msleep(1000);
550                 }
551         } while (status != 0);
552
553         /* safe to access the node data list without a lock, since this
554          * process is the only one to change the list */
555         list_for_each_entry(ndata, &dlm->reco.node_data, list) {
556                 BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
557                 ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
558
559                 mlog(0, "%s: Requesting lock info from node %u\n", dlm->name,
560                      ndata->node_num);
561
562                 if (ndata->node_num == dlm->node_num) {
563                         ndata->state = DLM_RECO_NODE_DATA_DONE;
564                         continue;
565                 }
566
567                 do {
568                         status = dlm_request_all_locks(dlm, ndata->node_num,
569                                                        dead_node);
570                         if (status < 0) {
571                                 mlog_errno(status);
572                                 if (dlm_is_host_down(status)) {
573                                         /* node died, ignore it for recovery */
574                                         status = 0;
575                                         ndata->state = DLM_RECO_NODE_DATA_DEAD;
576                                         /* wait for the domain map to catch up
577                                          * with the network state. */
578                                         wait_event_timeout(dlm->dlm_reco_thread_wq,
579                                                            dlm_is_node_dead(dlm,
580                                                                 ndata->node_num),
581                                                            msecs_to_jiffies(1000));
582                                         mlog(0, "waited 1 sec for %u, "
583                                              "dead? %s\n", ndata->node_num,
584                                              dlm_is_node_dead(dlm, ndata->node_num) ?
585                                              "yes" : "no");
586                                 } else {
587                                         /* -ENOMEM on the other node */
588                                         mlog(0, "%s: node %u returned "
589                                              "%d during recovery, retrying "
590                                              "after a short wait\n",
591                                              dlm->name, ndata->node_num,
592                                              status);
593                                         msleep(100);
594                                 }
595                         }
596                 } while (status != 0);
597
598                 spin_lock(&dlm_reco_state_lock);
599                 switch (ndata->state) {
600                         case DLM_RECO_NODE_DATA_INIT:
601                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
602                         case DLM_RECO_NODE_DATA_REQUESTED:
603                                 BUG();
604                                 break;
605                         case DLM_RECO_NODE_DATA_DEAD:
606                                 mlog(0, "node %u died after requesting "
607                                      "recovery info for node %u\n",
608                                      ndata->node_num, dead_node);
609                                 /* fine.  don't need this node's info.
610                                  * continue without it. */
611                                 break;
612                         case DLM_RECO_NODE_DATA_REQUESTING:
613                                 ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
614                                 mlog(0, "now receiving recovery data from "
615                                      "node %u for dead node %u\n",
616                                      ndata->node_num, dead_node);
617                                 break;
618                         case DLM_RECO_NODE_DATA_RECEIVING:
619                                 mlog(0, "already receiving recovery data from "
620                                      "node %u for dead node %u\n",
621                                      ndata->node_num, dead_node);
622                                 break;
623                         case DLM_RECO_NODE_DATA_DONE:
624                                 mlog(0, "already DONE receiving recovery data "
625                                      "from node %u for dead node %u\n",
626                                      ndata->node_num, dead_node);
627                                 break;
628                 }
629                 spin_unlock(&dlm_reco_state_lock);
630         }
631
632         mlog(0, "%s: Done requesting all lock info\n", dlm->name);
633
634         /* nodes should be sending reco data now
635          * just need to wait */
636
637         while (1) {
638                 /* check all the nodes now to see if we are
639                  * done, or if anyone died */
640                 all_nodes_done = 1;
641                 spin_lock(&dlm_reco_state_lock);
642                 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
643                         mlog(0, "checking recovery state of node %u\n",
644                              ndata->node_num);
645                         switch (ndata->state) {
646                                 case DLM_RECO_NODE_DATA_INIT:
647                                 case DLM_RECO_NODE_DATA_REQUESTING:
648                                         mlog(ML_ERROR, "bad ndata state for "
649                                              "node %u: state=%d\n",
650                                              ndata->node_num, ndata->state);
651                                         BUG();
652                                         break;
653                                 case DLM_RECO_NODE_DATA_DEAD:
654                                         mlog(0, "node %u died after "
655                                              "requesting recovery info for "
656                                              "node %u\n", ndata->node_num,
657                                              dead_node);
658                                         break;
659                                 case DLM_RECO_NODE_DATA_RECEIVING:
660                                 case DLM_RECO_NODE_DATA_REQUESTED:
661                                         mlog(0, "%s: node %u still in state %s\n",
662                                              dlm->name, ndata->node_num,
663                                              ndata->state==DLM_RECO_NODE_DATA_RECEIVING ?
664                                              "receiving" : "requested");
665                                         all_nodes_done = 0;
666                                         break;
667                                 case DLM_RECO_NODE_DATA_DONE:
668                                         mlog(0, "%s: node %u state is done\n",
669                                              dlm->name, ndata->node_num);
670                                         break;
671                                 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
672                                         mlog(0, "%s: node %u state is finalize\n",
673                                              dlm->name, ndata->node_num);
674                                         break;
675                         }
676                 }
677                 spin_unlock(&dlm_reco_state_lock);
678
679                 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
680                      all_nodes_done?"yes":"no");
681                 if (all_nodes_done) {
682                         int ret;
683
684                         /* Set this flag on recovery master to avoid
685                          * a new recovery for another dead node start
686                          * before the recovery is not done. That may
687                          * cause recovery hung.*/
688                         spin_lock(&dlm->spinlock);
689                         dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
690                         spin_unlock(&dlm->spinlock);
691
692                         /* all nodes are now in DLM_RECO_NODE_DATA_DONE state
693                          * just send a finalize message to everyone and
694                          * clean up */
695                         mlog(0, "all nodes are done! send finalize\n");
696                         ret = dlm_send_finalize_reco_message(dlm);
697                         if (ret < 0)
698                                 mlog_errno(ret);
699
700                         spin_lock(&dlm->spinlock);
701                         dlm_finish_local_lockres_recovery(dlm, dead_node,
702                                                           dlm->node_num);
703                         spin_unlock(&dlm->spinlock);
704                         mlog(0, "should be done with recovery!\n");
705
706                         mlog(0, "finishing recovery of %s at %lu, "
707                              "dead=%u, this=%u, new=%u\n", dlm->name,
708                              jiffies, dlm->reco.dead_node,
709                              dlm->node_num, dlm->reco.new_master);
710                         destroy = 1;
711                         status = 0;
712                         /* rescan everything marked dirty along the way */
713                         dlm_kick_thread(dlm, NULL);
714                         break;
715                 }
716                 /* wait to be signalled, with periodic timeout
717                  * to check for node death */
718                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
719                                          kthread_should_stop(),
720                                          msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
721
722         }
723
724         if (destroy)
725                 dlm_destroy_recovery_area(dlm);
726
727         return status;
728 }
729
730 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
731 {
732         int num=0;
733         struct dlm_reco_node_data *ndata;
734
735         spin_lock(&dlm->spinlock);
736         memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
737         /* nodes can only be removed (by dying) after dropping
738          * this lock, and death will be trapped later, so this should do */
739         spin_unlock(&dlm->spinlock);
740
741         while (1) {
742                 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
743                 if (num >= O2NM_MAX_NODES) {
744                         break;
745                 }
746                 BUG_ON(num == dead_node);
747
748                 ndata = kzalloc(sizeof(*ndata), GFP_NOFS);
749                 if (!ndata) {
750                         dlm_destroy_recovery_area(dlm);
751                         return -ENOMEM;
752                 }
753                 ndata->node_num = num;
754                 ndata->state = DLM_RECO_NODE_DATA_INIT;
755                 spin_lock(&dlm_reco_state_lock);
756                 list_add_tail(&ndata->list, &dlm->reco.node_data);
757                 spin_unlock(&dlm_reco_state_lock);
758                 num++;
759         }
760
761         return 0;
762 }
763
764 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm)
765 {
766         struct dlm_reco_node_data *ndata, *next;
767         LIST_HEAD(tmplist);
768
769         spin_lock(&dlm_reco_state_lock);
770         list_splice_init(&dlm->reco.node_data, &tmplist);
771         spin_unlock(&dlm_reco_state_lock);
772
773         list_for_each_entry_safe(ndata, next, &tmplist, list) {
774                 list_del_init(&ndata->list);
775                 kfree(ndata);
776         }
777 }
778
779 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
780                                  u8 dead_node)
781 {
782         struct dlm_lock_request lr;
783         int ret;
784         int status;
785
786         mlog(0, "\n");
787
788
789         mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
790                   "to %u\n", dead_node, request_from);
791
792         memset(&lr, 0, sizeof(lr));
793         lr.node_idx = dlm->node_num;
794         lr.dead_node = dead_node;
795
796         // send message
797         ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
798                                  &lr, sizeof(lr), request_from, &status);
799
800         /* negative status is handled by caller */
801         if (ret < 0)
802                 mlog(ML_ERROR, "%s: Error %d send LOCK_REQUEST to node %u "
803                      "to recover dead node %u\n", dlm->name, ret,
804                      request_from, dead_node);
805         else
806                 ret = status;
807         // return from here, then
808         // sleep until all received or error
809         return ret;
810
811 }
812
813 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
814                                   void **ret_data)
815 {
816         struct dlm_ctxt *dlm = data;
817         struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
818         char *buf = NULL;
819         struct dlm_work_item *item = NULL;
820
821         if (!dlm_grab(dlm))
822                 return -EINVAL;
823
824         if (lr->dead_node != dlm->reco.dead_node) {
825                 mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local "
826                      "dead_node is %u\n", dlm->name, lr->node_idx,
827                      lr->dead_node, dlm->reco.dead_node);
828                 dlm_print_reco_node_status(dlm);
829                 /* this is a hack */
830                 dlm_put(dlm);
831                 return -ENOMEM;
832         }
833         BUG_ON(lr->dead_node != dlm->reco.dead_node);
834
835         item = kzalloc(sizeof(*item), GFP_NOFS);
836         if (!item) {
837                 dlm_put(dlm);
838                 return -ENOMEM;
839         }
840
841         /* this will get freed by dlm_request_all_locks_worker */
842         buf = (char *) __get_free_page(GFP_NOFS);
843         if (!buf) {
844                 kfree(item);
845                 dlm_put(dlm);
846                 return -ENOMEM;
847         }
848
849         /* queue up work for dlm_request_all_locks_worker */
850         dlm_grab(dlm);  /* get an extra ref for the work item */
851         dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
852         item->u.ral.reco_master = lr->node_idx;
853         item->u.ral.dead_node = lr->dead_node;
854         spin_lock(&dlm->work_lock);
855         list_add_tail(&item->list, &dlm->work_list);
856         spin_unlock(&dlm->work_lock);
857         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
858
859         dlm_put(dlm);
860         return 0;
861 }
862
863 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
864 {
865         struct dlm_migratable_lockres *mres;
866         struct dlm_lock_resource *res;
867         struct dlm_ctxt *dlm;
868         LIST_HEAD(resources);
869         int ret;
870         u8 dead_node, reco_master;
871         int skip_all_done = 0;
872
873         dlm = item->dlm;
874         dead_node = item->u.ral.dead_node;
875         reco_master = item->u.ral.reco_master;
876         mres = (struct dlm_migratable_lockres *)data;
877
878         mlog(0, "%s: recovery worker started, dead=%u, master=%u\n",
879              dlm->name, dead_node, reco_master);
880
881         if (dead_node != dlm->reco.dead_node ||
882             reco_master != dlm->reco.new_master) {
883                 /* worker could have been created before the recovery master
884                  * died.  if so, do not continue, but do not error. */
885                 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
886                         mlog(ML_NOTICE, "%s: will not send recovery state, "
887                              "recovery master %u died, thread=(dead=%u,mas=%u)"
888                              " current=(dead=%u,mas=%u)\n", dlm->name,
889                              reco_master, dead_node, reco_master,
890                              dlm->reco.dead_node, dlm->reco.new_master);
891                 } else {
892                         mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, "
893                              "master=%u), request(dead=%u, master=%u)\n",
894                              dlm->name, dlm->reco.dead_node,
895                              dlm->reco.new_master, dead_node, reco_master);
896                 }
897                 goto leave;
898         }
899
900         /* lock resources should have already been moved to the
901          * dlm->reco.resources list.  now move items from that list
902          * to a temp list if the dead owner matches.  note that the
903          * whole cluster recovers only one node at a time, so we
904          * can safely move UNKNOWN lock resources for each recovery
905          * session. */
906         dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
907
908         /* now we can begin blasting lockreses without the dlm lock */
909
910         /* any errors returned will be due to the new_master dying,
911          * the dlm_reco_thread should detect this */
912         list_for_each_entry(res, &resources, recovering) {
913                 ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
914                                         DLM_MRES_RECOVERY);
915                 if (ret < 0) {
916                         mlog(ML_ERROR, "%s: node %u went down while sending "
917                              "recovery state for dead node %u, ret=%d\n", dlm->name,
918                              reco_master, dead_node, ret);
919                         skip_all_done = 1;
920                         break;
921                 }
922         }
923
924         /* move the resources back to the list */
925         spin_lock(&dlm->spinlock);
926         list_splice_init(&resources, &dlm->reco.resources);
927         spin_unlock(&dlm->spinlock);
928
929         if (!skip_all_done) {
930                 ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
931                 if (ret < 0) {
932                         mlog(ML_ERROR, "%s: node %u went down while sending "
933                              "recovery all-done for dead node %u, ret=%d\n",
934                              dlm->name, reco_master, dead_node, ret);
935                 }
936         }
937 leave:
938         free_page((unsigned long)data);
939 }
940
941
942 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
943 {
944         int ret, tmpret;
945         struct dlm_reco_data_done done_msg;
946
947         memset(&done_msg, 0, sizeof(done_msg));
948         done_msg.node_idx = dlm->node_num;
949         done_msg.dead_node = dead_node;
950         mlog(0, "sending DATA DONE message to %u, "
951              "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
952              done_msg.dead_node);
953
954         ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
955                                  sizeof(done_msg), send_to, &tmpret);
956         if (ret < 0) {
957                 mlog(ML_ERROR, "%s: Error %d send RECO_DATA_DONE to node %u "
958                      "to recover dead node %u\n", dlm->name, ret, send_to,
959                      dead_node);
960                 if (!dlm_is_host_down(ret)) {
961                         BUG();
962                 }
963         } else
964                 ret = tmpret;
965         return ret;
966 }
967
968
969 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
970                                void **ret_data)
971 {
972         struct dlm_ctxt *dlm = data;
973         struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
974         struct dlm_reco_node_data *ndata = NULL;
975         int ret = -EINVAL;
976
977         if (!dlm_grab(dlm))
978                 return -EINVAL;
979
980         mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
981              "node_idx=%u, this node=%u\n", done->dead_node,
982              dlm->reco.dead_node, done->node_idx, dlm->node_num);
983
984         mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node),
985                         "Got DATA DONE: dead_node=%u, reco.dead_node=%u, "
986                         "node_idx=%u, this node=%u\n", done->dead_node,
987                         dlm->reco.dead_node, done->node_idx, dlm->node_num);
988
989         spin_lock(&dlm_reco_state_lock);
990         list_for_each_entry(ndata, &dlm->reco.node_data, list) {
991                 if (ndata->node_num != done->node_idx)
992                         continue;
993
994                 switch (ndata->state) {
995                         /* should have moved beyond INIT but not to FINALIZE yet */
996                         case DLM_RECO_NODE_DATA_INIT:
997                         case DLM_RECO_NODE_DATA_DEAD:
998                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
999                                 mlog(ML_ERROR, "bad ndata state for node %u:"
1000                                      " state=%d\n", ndata->node_num,
1001                                      ndata->state);
1002                                 BUG();
1003                                 break;
1004                         /* these states are possible at this point, anywhere along
1005                          * the line of recovery */
1006                         case DLM_RECO_NODE_DATA_DONE:
1007                         case DLM_RECO_NODE_DATA_RECEIVING:
1008                         case DLM_RECO_NODE_DATA_REQUESTED:
1009                         case DLM_RECO_NODE_DATA_REQUESTING:
1010                                 mlog(0, "node %u is DONE sending "
1011                                           "recovery data!\n",
1012                                           ndata->node_num);
1013
1014                                 ndata->state = DLM_RECO_NODE_DATA_DONE;
1015                                 ret = 0;
1016                                 break;
1017                 }
1018         }
1019         spin_unlock(&dlm_reco_state_lock);
1020
1021         /* wake the recovery thread, some node is done */
1022         if (!ret)
1023                 dlm_kick_recovery_thread(dlm);
1024
1025         if (ret < 0)
1026                 mlog(ML_ERROR, "failed to find recovery node data for node "
1027                      "%u\n", done->node_idx);
1028         dlm_put(dlm);
1029
1030         mlog(0, "leaving reco data done handler, ret=%d\n", ret);
1031         return ret;
1032 }
1033
1034 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
1035                                         struct list_head *list,
1036                                         u8 dead_node)
1037 {
1038         struct dlm_lock_resource *res, *next;
1039         struct dlm_lock *lock;
1040
1041         spin_lock(&dlm->spinlock);
1042         list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
1043                 /* always prune any $RECOVERY entries for dead nodes,
1044                  * otherwise hangs can occur during later recovery */
1045                 if (dlm_is_recovery_lock(res->lockname.name,
1046                                          res->lockname.len)) {
1047                         spin_lock(&res->spinlock);
1048                         list_for_each_entry(lock, &res->granted, list) {
1049                                 if (lock->ml.node == dead_node) {
1050                                         mlog(0, "AHA! there was "
1051                                              "a $RECOVERY lock for dead "
1052                                              "node %u (%s)!\n",
1053                                              dead_node, dlm->name);
1054                                         list_del_init(&lock->list);
1055                                         dlm_lock_put(lock);
1056                                         /* Can't schedule DLM_UNLOCK_FREE_LOCK
1057                                          * - do manually */
1058                                         dlm_lock_put(lock);
1059                                         break;
1060                                 }
1061                         }
1062                         spin_unlock(&res->spinlock);
1063                         continue;
1064                 }
1065
1066                 if (res->owner == dead_node) {
1067                         mlog(0, "found lockres owned by dead node while "
1068                                   "doing recovery for node %u. sending it.\n",
1069                                   dead_node);
1070                         list_move_tail(&res->recovering, list);
1071                 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1072                         mlog(0, "found UNKNOWN owner while doing recovery "
1073                                   "for node %u. sending it.\n", dead_node);
1074                         list_move_tail(&res->recovering, list);
1075                 }
1076         }
1077         spin_unlock(&dlm->spinlock);
1078 }
1079
1080 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
1081 {
1082         int total_locks = 0;
1083         struct list_head *iter, *queue = &res->granted;
1084         int i;
1085
1086         for (i=0; i<3; i++) {
1087                 list_for_each(iter, queue)
1088                         total_locks++;
1089                 queue++;
1090         }
1091         return total_locks;
1092 }
1093
1094
1095 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
1096                                       struct dlm_migratable_lockres *mres,
1097                                       u8 send_to,
1098                                       struct dlm_lock_resource *res,
1099                                       int total_locks)
1100 {
1101         u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
1102         int mres_total_locks = be32_to_cpu(mres->total_locks);
1103         int ret = 0, status = 0;
1104         u8 orig_flags = mres->flags,
1105            orig_master = mres->master;
1106
1107         BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
1108         if (!mres->num_locks)
1109                 return 0;
1110
1111         /* add an all-done flag if we reached the last lock */
1112         orig_flags = mres->flags;
1113         BUG_ON(total_locks > mres_total_locks);
1114         if (total_locks == mres_total_locks)
1115                 mres->flags |= DLM_MRES_ALL_DONE;
1116
1117         mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n",
1118              dlm->name, res->lockname.len, res->lockname.name,
1119              orig_flags & DLM_MRES_MIGRATION ? "migration" : "recovery",
1120              send_to);
1121
1122         /* send it */
1123         ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
1124                                  struct_size(mres, ml, mres->num_locks),
1125                                  send_to, &status);
1126         if (ret < 0) {
1127                 /* XXX: negative status is not handled.
1128                  * this will end up killing this node. */
1129                 mlog(ML_ERROR, "%s: res %.*s, Error %d send MIG_LOCKRES to "
1130                      "node %u (%s)\n", dlm->name, mres->lockname_len,
1131                      mres->lockname, ret, send_to,
1132                      (orig_flags & DLM_MRES_MIGRATION ?
1133                       "migration" : "recovery"));
1134         } else {
1135                 /* might get an -ENOMEM back here */
1136                 ret = status;
1137                 if (ret < 0) {
1138                         mlog_errno(ret);
1139
1140                         if (ret == -EFAULT) {
1141                                 mlog(ML_ERROR, "node %u told me to kill "
1142                                      "myself!\n", send_to);
1143                                 BUG();
1144                         }
1145                 }
1146         }
1147
1148         /* zero and reinit the message buffer */
1149         dlm_init_migratable_lockres(mres, res->lockname.name,
1150                                     res->lockname.len, mres_total_locks,
1151                                     mig_cookie, orig_flags, orig_master);
1152         return ret;
1153 }
1154
1155 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
1156                                         const char *lockname, int namelen,
1157                                         int total_locks, u64 cookie,
1158                                         u8 flags, u8 master)
1159 {
1160         /* mres here is one full page */
1161         clear_page(mres);
1162         mres->lockname_len = namelen;
1163         memcpy(mres->lockname, lockname, namelen);
1164         mres->num_locks = 0;
1165         mres->total_locks = cpu_to_be32(total_locks);
1166         mres->mig_cookie = cpu_to_be64(cookie);
1167         mres->flags = flags;
1168         mres->master = master;
1169 }
1170
1171 static void dlm_prepare_lvb_for_migration(struct dlm_lock *lock,
1172                                           struct dlm_migratable_lockres *mres,
1173                                           int queue)
1174 {
1175         if (!lock->lksb)
1176                return;
1177
1178         /* Ignore lvb in all locks in the blocked list */
1179         if (queue == DLM_BLOCKED_LIST)
1180                 return;
1181
1182         /* Only consider lvbs in locks with granted EX or PR lock levels */
1183         if (lock->ml.type != LKM_EXMODE && lock->ml.type != LKM_PRMODE)
1184                 return;
1185
1186         if (dlm_lvb_is_empty(mres->lvb)) {
1187                 memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1188                 return;
1189         }
1190
1191         /* Ensure the lvb copied for migration matches in other valid locks */
1192         if (!memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))
1193                 return;
1194
1195         mlog(ML_ERROR, "Mismatched lvb in lock cookie=%u:%llu, name=%.*s, "
1196              "node=%u\n",
1197              dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
1198              dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
1199              lock->lockres->lockname.len, lock->lockres->lockname.name,
1200              lock->ml.node);
1201         dlm_print_one_lock_resource(lock->lockres);
1202         BUG();
1203 }
1204
1205 /* returns 1 if this lock fills the network structure,
1206  * 0 otherwise */
1207 static int dlm_add_lock_to_array(struct dlm_lock *lock,
1208                                  struct dlm_migratable_lockres *mres, int queue)
1209 {
1210         struct dlm_migratable_lock *ml;
1211         int lock_num = mres->num_locks;
1212
1213         ml = &(mres->ml[lock_num]);
1214         ml->cookie = lock->ml.cookie;
1215         ml->type = lock->ml.type;
1216         ml->convert_type = lock->ml.convert_type;
1217         ml->highest_blocked = lock->ml.highest_blocked;
1218         ml->list = queue;
1219         if (lock->lksb) {
1220                 ml->flags = lock->lksb->flags;
1221                 dlm_prepare_lvb_for_migration(lock, mres, queue);
1222         }
1223         ml->node = lock->ml.node;
1224         mres->num_locks++;
1225         /* we reached the max, send this network message */
1226         if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1227                 return 1;
1228         return 0;
1229 }
1230
1231 static void dlm_add_dummy_lock(struct dlm_ctxt *dlm,
1232                                struct dlm_migratable_lockres *mres)
1233 {
1234         struct dlm_lock dummy;
1235         memset(&dummy, 0, sizeof(dummy));
1236         dummy.ml.cookie = 0;
1237         dummy.ml.type = LKM_IVMODE;
1238         dummy.ml.convert_type = LKM_IVMODE;
1239         dummy.ml.highest_blocked = LKM_IVMODE;
1240         dummy.lksb = NULL;
1241         dummy.ml.node = dlm->node_num;
1242         dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST);
1243 }
1244
1245 static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm,
1246                                     struct dlm_migratable_lock *ml,
1247                                     u8 *nodenum)
1248 {
1249         if (unlikely(ml->cookie == 0 &&
1250             ml->type == LKM_IVMODE &&
1251             ml->convert_type == LKM_IVMODE &&
1252             ml->highest_blocked == LKM_IVMODE &&
1253             ml->list == DLM_BLOCKED_LIST)) {
1254                 *nodenum = ml->node;
1255                 return 1;
1256         }
1257         return 0;
1258 }
1259
1260 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1261                          struct dlm_migratable_lockres *mres,
1262                          u8 send_to, u8 flags)
1263 {
1264         struct list_head *queue;
1265         int total_locks, i;
1266         u64 mig_cookie = 0;
1267         struct dlm_lock *lock;
1268         int ret = 0;
1269
1270         BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1271
1272         mlog(0, "sending to %u\n", send_to);
1273
1274         total_locks = dlm_num_locks_in_lockres(res);
1275         if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1276                 /* rare, but possible */
1277                 mlog(0, "argh.  lockres has %d locks.  this will "
1278                           "require more than one network packet to "
1279                           "migrate\n", total_locks);
1280                 mig_cookie = dlm_get_next_mig_cookie();
1281         }
1282
1283         dlm_init_migratable_lockres(mres, res->lockname.name,
1284                                     res->lockname.len, total_locks,
1285                                     mig_cookie, flags, res->owner);
1286
1287         total_locks = 0;
1288         for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1289                 queue = dlm_list_idx_to_ptr(res, i);
1290                 list_for_each_entry(lock, queue, list) {
1291                         /* add another lock. */
1292                         total_locks++;
1293                         if (!dlm_add_lock_to_array(lock, mres, i))
1294                                 continue;
1295
1296                         /* this filled the lock message,
1297                          * we must send it immediately. */
1298                         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1299                                                        res, total_locks);
1300                         if (ret < 0)
1301                                 goto error;
1302                 }
1303         }
1304         if (total_locks == 0) {
1305                 /* send a dummy lock to indicate a mastery reference only */
1306                 mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n",
1307                      dlm->name, res->lockname.len, res->lockname.name,
1308                      send_to, flags & DLM_MRES_RECOVERY ? "recovery" :
1309                      "migration");
1310                 dlm_add_dummy_lock(dlm, mres);
1311         }
1312         /* flush any remaining locks */
1313         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1314         if (ret < 0)
1315                 goto error;
1316         return ret;
1317
1318 error:
1319         mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n",
1320              dlm->name, ret);
1321         if (!dlm_is_host_down(ret))
1322                 BUG();
1323         mlog(0, "%s: node %u went down while sending %s "
1324              "lockres %.*s\n", dlm->name, send_to,
1325              flags & DLM_MRES_RECOVERY ?  "recovery" : "migration",
1326              res->lockname.len, res->lockname.name);
1327         return ret;
1328 }
1329
1330
1331
1332 /*
1333  * this message will contain no more than one page worth of
1334  * recovery data, and it will work on only one lockres.
1335  * there may be many locks in this page, and we may need to wait
1336  * for additional packets to complete all the locks (rare, but
1337  * possible).
1338  */
1339 /*
1340  * NOTE: the allocation error cases here are scary
1341  * we really cannot afford to fail an alloc in recovery
1342  * do we spin?  returning an error only delays the problem really
1343  */
1344
1345 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
1346                             void **ret_data)
1347 {
1348         struct dlm_ctxt *dlm = data;
1349         struct dlm_migratable_lockres *mres =
1350                 (struct dlm_migratable_lockres *)msg->buf;
1351         int ret = 0;
1352         u8 real_master;
1353         u8 extra_refs = 0;
1354         char *buf = NULL;
1355         struct dlm_work_item *item = NULL;
1356         struct dlm_lock_resource *res = NULL;
1357         unsigned int hash;
1358
1359         if (!dlm_grab(dlm))
1360                 return -EINVAL;
1361
1362         if (!dlm_joined(dlm)) {
1363                 mlog(ML_ERROR, "Domain %s not joined! "
1364                           "lockres %.*s, master %u\n",
1365                           dlm->name, mres->lockname_len,
1366                           mres->lockname, mres->master);
1367                 dlm_put(dlm);
1368                 return -EINVAL;
1369         }
1370
1371         BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1372
1373         real_master = mres->master;
1374         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1375                 /* cannot migrate a lockres with no master */
1376                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1377         }
1378
1379         mlog(0, "%s message received from node %u\n",
1380                   (mres->flags & DLM_MRES_RECOVERY) ?
1381                   "recovery" : "migration", mres->master);
1382         if (mres->flags & DLM_MRES_ALL_DONE)
1383                 mlog(0, "all done flag.  all lockres data received!\n");
1384
1385         ret = -ENOMEM;
1386         buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS);
1387         item = kzalloc(sizeof(*item), GFP_NOFS);
1388         if (!buf || !item)
1389                 goto leave;
1390
1391         /* lookup the lock to see if we have a secondary queue for this
1392          * already...  just add the locks in and this will have its owner
1393          * and RECOVERY flag changed when it completes. */
1394         hash = dlm_lockid_hash(mres->lockname, mres->lockname_len);
1395         spin_lock(&dlm->spinlock);
1396         res = __dlm_lookup_lockres_full(dlm, mres->lockname, mres->lockname_len,
1397                         hash);
1398         if (res) {
1399                 /* this will get a ref on res */
1400                 /* mark it as recovering/migrating and hash it */
1401                 spin_lock(&res->spinlock);
1402                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
1403                         mlog(0, "%s: node is attempting to migrate "
1404                                 "lockres %.*s, but marked as dropping "
1405                                 " ref!\n", dlm->name,
1406                                 mres->lockname_len, mres->lockname);
1407                         ret = -EINVAL;
1408                         spin_unlock(&res->spinlock);
1409                         spin_unlock(&dlm->spinlock);
1410                         dlm_lockres_put(res);
1411                         goto leave;
1412                 }
1413
1414                 if (mres->flags & DLM_MRES_RECOVERY) {
1415                         res->state |= DLM_LOCK_RES_RECOVERING;
1416                 } else {
1417                         if (res->state & DLM_LOCK_RES_MIGRATING) {
1418                                 /* this is at least the second
1419                                  * lockres message */
1420                                 mlog(0, "lock %.*s is already migrating\n",
1421                                           mres->lockname_len,
1422                                           mres->lockname);
1423                         } else if (res->state & DLM_LOCK_RES_RECOVERING) {
1424                                 /* caller should BUG */
1425                                 mlog(ML_ERROR, "node is attempting to migrate "
1426                                      "lock %.*s, but marked as recovering!\n",
1427                                      mres->lockname_len, mres->lockname);
1428                                 ret = -EFAULT;
1429                                 spin_unlock(&res->spinlock);
1430                                 spin_unlock(&dlm->spinlock);
1431                                 dlm_lockres_put(res);
1432                                 goto leave;
1433                         }
1434                         res->state |= DLM_LOCK_RES_MIGRATING;
1435                 }
1436                 spin_unlock(&res->spinlock);
1437                 spin_unlock(&dlm->spinlock);
1438         } else {
1439                 spin_unlock(&dlm->spinlock);
1440                 /* need to allocate, just like if it was
1441                  * mastered here normally  */
1442                 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1443                 if (!res)
1444                         goto leave;
1445
1446                 /* to match the ref that we would have gotten if
1447                  * dlm_lookup_lockres had succeeded */
1448                 dlm_lockres_get(res);
1449
1450                 /* mark it as recovering/migrating and hash it */
1451                 if (mres->flags & DLM_MRES_RECOVERY)
1452                         res->state |= DLM_LOCK_RES_RECOVERING;
1453                 else
1454                         res->state |= DLM_LOCK_RES_MIGRATING;
1455
1456                 spin_lock(&dlm->spinlock);
1457                 __dlm_insert_lockres(dlm, res);
1458                 spin_unlock(&dlm->spinlock);
1459
1460                 /* Add an extra ref for this lock-less lockres lest the
1461                  * dlm_thread purges it before we get the chance to add
1462                  * locks to it */
1463                 dlm_lockres_get(res);
1464
1465                 /* There are three refs that need to be put.
1466                  * 1. Taken above.
1467                  * 2. kref_init in dlm_new_lockres()->dlm_init_lockres().
1468                  * 3. dlm_lookup_lockres()
1469                  * The first one is handled at the end of this function. The
1470                  * other two are handled in the worker thread after locks have
1471                  * been attached. Yes, we don't wait for purge time to match
1472                  * kref_init. The lockres will still have atleast one ref
1473                  * added because it is in the hash __dlm_insert_lockres() */
1474                 extra_refs++;
1475
1476                 /* now that the new lockres is inserted,
1477                  * make it usable by other processes */
1478                 spin_lock(&res->spinlock);
1479                 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1480                 spin_unlock(&res->spinlock);
1481                 wake_up(&res->wq);
1482         }
1483
1484         /* at this point we have allocated everything we need,
1485          * and we have a hashed lockres with an extra ref and
1486          * the proper res->state flags. */
1487         ret = 0;
1488         spin_lock(&res->spinlock);
1489         /* drop this either when master requery finds a different master
1490          * or when a lock is added by the recovery worker */
1491         dlm_lockres_grab_inflight_ref(dlm, res);
1492         if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1493                 /* migration cannot have an unknown master */
1494                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1495                 mlog(0, "recovery has passed me a lockres with an "
1496                           "unknown owner.. will need to requery: "
1497                           "%.*s\n", mres->lockname_len, mres->lockname);
1498         } else {
1499                 /* take a reference now to pin the lockres, drop it
1500                  * when locks are added in the worker */
1501                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
1502         }
1503         spin_unlock(&res->spinlock);
1504
1505         /* queue up work for dlm_mig_lockres_worker */
1506         dlm_grab(dlm);  /* get an extra ref for the work item */
1507         memcpy(buf, msg->buf, be16_to_cpu(msg->data_len));  /* copy the whole message */
1508         dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1509         item->u.ml.lockres = res; /* already have a ref */
1510         item->u.ml.real_master = real_master;
1511         item->u.ml.extra_ref = extra_refs;
1512         spin_lock(&dlm->work_lock);
1513         list_add_tail(&item->list, &dlm->work_list);
1514         spin_unlock(&dlm->work_lock);
1515         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
1516
1517 leave:
1518         /* One extra ref taken needs to be put here */
1519         if (extra_refs)
1520                 dlm_lockres_put(res);
1521
1522         dlm_put(dlm);
1523         if (ret < 0) {
1524                 kfree(buf);
1525                 kfree(item);
1526                 mlog_errno(ret);
1527         }
1528
1529         return ret;
1530 }
1531
1532
1533 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1534 {
1535         struct dlm_ctxt *dlm;
1536         struct dlm_migratable_lockres *mres;
1537         int ret = 0;
1538         struct dlm_lock_resource *res;
1539         u8 real_master;
1540         u8 extra_ref;
1541
1542         dlm = item->dlm;
1543         mres = (struct dlm_migratable_lockres *)data;
1544
1545         res = item->u.ml.lockres;
1546         real_master = item->u.ml.real_master;
1547         extra_ref = item->u.ml.extra_ref;
1548
1549         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1550                 /* this case is super-rare. only occurs if
1551                  * node death happens during migration. */
1552 again:
1553                 ret = dlm_lockres_master_requery(dlm, res, &real_master);
1554                 if (ret < 0) {
1555                         mlog(0, "dlm_lockres_master_requery ret=%d\n",
1556                                   ret);
1557                         goto again;
1558                 }
1559                 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1560                         mlog(0, "lockres %.*s not claimed.  "
1561                                    "this node will take it.\n",
1562                                    res->lockname.len, res->lockname.name);
1563                 } else {
1564                         spin_lock(&res->spinlock);
1565                         dlm_lockres_drop_inflight_ref(dlm, res);
1566                         spin_unlock(&res->spinlock);
1567                         mlog(0, "master needs to respond to sender "
1568                                   "that node %u still owns %.*s\n",
1569                                   real_master, res->lockname.len,
1570                                   res->lockname.name);
1571                         /* cannot touch this lockres */
1572                         goto leave;
1573                 }
1574         }
1575
1576         ret = dlm_process_recovery_data(dlm, res, mres);
1577         if (ret < 0)
1578                 mlog(0, "dlm_process_recovery_data returned  %d\n", ret);
1579         else
1580                 mlog(0, "dlm_process_recovery_data succeeded\n");
1581
1582         if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1583                            (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1584                 ret = dlm_finish_migration(dlm, res, mres->master);
1585                 if (ret < 0)
1586                         mlog_errno(ret);
1587         }
1588
1589 leave:
1590         /* See comment in dlm_mig_lockres_handler() */
1591         if (res) {
1592                 if (extra_ref)
1593                         dlm_lockres_put(res);
1594                 dlm_lockres_put(res);
1595         }
1596         kfree(data);
1597 }
1598
1599
1600
1601 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1602                                       struct dlm_lock_resource *res,
1603                                       u8 *real_master)
1604 {
1605         struct dlm_node_iter iter;
1606         int nodenum;
1607         int ret = 0;
1608
1609         *real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1610
1611         /* we only reach here if one of the two nodes in a
1612          * migration died while the migration was in progress.
1613          * at this point we need to requery the master.  we
1614          * know that the new_master got as far as creating
1615          * an mle on at least one node, but we do not know
1616          * if any nodes had actually cleared the mle and set
1617          * the master to the new_master.  the old master
1618          * is supposed to set the owner to UNKNOWN in the
1619          * event of a new_master death, so the only possible
1620          * responses that we can get from nodes here are
1621          * that the master is new_master, or that the master
1622          * is UNKNOWN.
1623          * if all nodes come back with UNKNOWN then we know
1624          * the lock needs remastering here.
1625          * if any node comes back with a valid master, check
1626          * to see if that master is the one that we are
1627          * recovering.  if so, then the new_master died and
1628          * we need to remaster this lock.  if not, then the
1629          * new_master survived and that node will respond to
1630          * other nodes about the owner.
1631          * if there is an owner, this node needs to dump this
1632          * lockres and alert the sender that this lockres
1633          * was rejected. */
1634         spin_lock(&dlm->spinlock);
1635         dlm_node_iter_init(dlm->domain_map, &iter);
1636         spin_unlock(&dlm->spinlock);
1637
1638         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1639                 /* do not send to self */
1640                 if (nodenum == dlm->node_num)
1641                         continue;
1642                 ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1643                 if (ret < 0) {
1644                         mlog_errno(ret);
1645                         if (!dlm_is_host_down(ret))
1646                                 BUG();
1647                         /* host is down, so answer for that node would be
1648                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
1649                 }
1650                 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1651                         mlog(0, "lock master is %u\n", *real_master);
1652                         break;
1653                 }
1654         }
1655         return ret;
1656 }
1657
1658
1659 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1660                           u8 nodenum, u8 *real_master)
1661 {
1662         int ret;
1663         struct dlm_master_requery req;
1664         int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1665
1666         memset(&req, 0, sizeof(req));
1667         req.node_idx = dlm->node_num;
1668         req.namelen = res->lockname.len;
1669         memcpy(req.name, res->lockname.name, res->lockname.len);
1670
1671 resend:
1672         ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1673                                  &req, sizeof(req), nodenum, &status);
1674         if (ret < 0)
1675                 mlog(ML_ERROR, "Error %d when sending message %u (key "
1676                      "0x%x) to node %u\n", ret, DLM_MASTER_REQUERY_MSG,
1677                      dlm->key, nodenum);
1678         else if (status == -ENOMEM) {
1679                 mlog_errno(status);
1680                 msleep(50);
1681                 goto resend;
1682         } else {
1683                 BUG_ON(status < 0);
1684                 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1685                 *real_master = (u8) (status & 0xff);
1686                 mlog(0, "node %u responded to master requery with %u\n",
1687                           nodenum, *real_master);
1688                 ret = 0;
1689         }
1690         return ret;
1691 }
1692
1693
1694 /* this function cannot error, so unless the sending
1695  * or receiving of the message failed, the owner can
1696  * be trusted */
1697 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
1698                                void **ret_data)
1699 {
1700         struct dlm_ctxt *dlm = data;
1701         struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1702         struct dlm_lock_resource *res = NULL;
1703         unsigned int hash;
1704         int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1705         u32 flags = DLM_ASSERT_MASTER_REQUERY;
1706         int dispatched = 0;
1707
1708         if (!dlm_grab(dlm)) {
1709                 /* since the domain has gone away on this
1710                  * node, the proper response is UNKNOWN */
1711                 return master;
1712         }
1713
1714         hash = dlm_lockid_hash(req->name, req->namelen);
1715
1716         spin_lock(&dlm->spinlock);
1717         res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash);
1718         if (res) {
1719                 spin_lock(&res->spinlock);
1720                 master = res->owner;
1721                 if (master == dlm->node_num) {
1722                         int ret = dlm_dispatch_assert_master(dlm, res,
1723                                                              0, 0, flags);
1724                         if (ret < 0) {
1725                                 mlog_errno(ret);
1726                                 spin_unlock(&res->spinlock);
1727                                 dlm_lockres_put(res);
1728                                 spin_unlock(&dlm->spinlock);
1729                                 dlm_put(dlm);
1730                                 /* sender will take care of this and retry */
1731                                 return ret;
1732                         } else {
1733                                 dispatched = 1;
1734                                 __dlm_lockres_grab_inflight_worker(dlm, res);
1735                                 spin_unlock(&res->spinlock);
1736                         }
1737                 } else {
1738                         /* put.. incase we are not the master */
1739                         spin_unlock(&res->spinlock);
1740                         dlm_lockres_put(res);
1741                 }
1742         }
1743         spin_unlock(&dlm->spinlock);
1744
1745         if (!dispatched)
1746                 dlm_put(dlm);
1747         return master;
1748 }
1749
1750 static inline struct list_head *
1751 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1752 {
1753         struct list_head *ret;
1754         BUG_ON(list_num < 0);
1755         BUG_ON(list_num > 2);
1756         ret = &(res->granted);
1757         ret += list_num;
1758         return ret;
1759 }
1760 /* TODO: do ast flush business
1761  * TODO: do MIGRATING and RECOVERING spinning
1762  */
1763
1764 /*
1765 * NOTE about in-flight requests during migration:
1766 *
1767 * Before attempting the migrate, the master has marked the lockres as
1768 * MIGRATING and then flushed all of its pending ASTS.  So any in-flight
1769 * requests either got queued before the MIGRATING flag got set, in which
1770 * case the lock data will reflect the change and a return message is on
1771 * the way, or the request failed to get in before MIGRATING got set.  In
1772 * this case, the caller will be told to spin and wait for the MIGRATING
1773 * flag to be dropped, then recheck the master.
1774 * This holds true for the convert, cancel and unlock cases, and since lvb
1775 * updates are tied to these same messages, it applies to lvb updates as
1776 * well.  For the lock case, there is no way a lock can be on the master
1777 * queue and not be on the secondary queue since the lock is always added
1778 * locally first.  This means that the new target node will never be sent
1779 * a lock that he doesn't already have on the list.
1780 * In total, this means that the local lock is correct and should not be
1781 * updated to match the one sent by the master.  Any messages sent back
1782 * from the master before the MIGRATING flag will bring the lock properly
1783 * up-to-date, and the change will be ordered properly for the waiter.
1784 * We will *not* attempt to modify the lock underneath the waiter.
1785 */
1786
1787 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1788                                      struct dlm_lock_resource *res,
1789                                      struct dlm_migratable_lockres *mres)
1790 {
1791         struct dlm_migratable_lock *ml;
1792         struct list_head *queue, *iter;
1793         struct list_head *tmpq = NULL;
1794         struct dlm_lock *newlock = NULL;
1795         struct dlm_lockstatus *lksb = NULL;
1796         int ret = 0;
1797         int i, j, bad;
1798         struct dlm_lock *lock;
1799         u8 from = O2NM_MAX_NODES;
1800         __be64 c;
1801
1802         mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1803         for (i=0; i<mres->num_locks; i++) {
1804                 ml = &(mres->ml[i]);
1805
1806                 if (dlm_is_dummy_lock(dlm, ml, &from)) {
1807                         /* placeholder, just need to set the refmap bit */
1808                         BUG_ON(mres->num_locks != 1);
1809                         mlog(0, "%s:%.*s: dummy lock for %u\n",
1810                              dlm->name, mres->lockname_len, mres->lockname,
1811                              from);
1812                         spin_lock(&res->spinlock);
1813                         dlm_lockres_set_refmap_bit(dlm, res, from);
1814                         spin_unlock(&res->spinlock);
1815                         break;
1816                 }
1817                 BUG_ON(ml->highest_blocked != LKM_IVMODE);
1818                 newlock = NULL;
1819                 lksb = NULL;
1820
1821                 queue = dlm_list_num_to_pointer(res, ml->list);
1822                 tmpq = NULL;
1823
1824                 /* if the lock is for the local node it needs to
1825                  * be moved to the proper location within the queue.
1826                  * do not allocate a new lock structure. */
1827                 if (ml->node == dlm->node_num) {
1828                         /* MIGRATION ONLY! */
1829                         BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1830
1831                         lock = NULL;
1832                         spin_lock(&res->spinlock);
1833                         for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) {
1834                                 tmpq = dlm_list_idx_to_ptr(res, j);
1835                                 list_for_each(iter, tmpq) {
1836                                         lock = list_entry(iter,
1837                                                   struct dlm_lock, list);
1838                                         if (lock->ml.cookie == ml->cookie)
1839                                                 break;
1840                                         lock = NULL;
1841                                 }
1842                                 if (lock)
1843                                         break;
1844                         }
1845
1846                         /* lock is always created locally first, and
1847                          * destroyed locally last.  it must be on the list */
1848                         if (!lock) {
1849                                 c = ml->cookie;
1850                                 mlog(ML_ERROR, "Could not find local lock "
1851                                                "with cookie %u:%llu, node %u, "
1852                                                "list %u, flags 0x%x, type %d, "
1853                                                "conv %d, highest blocked %d\n",
1854                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1855                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1856                                      ml->node, ml->list, ml->flags, ml->type,
1857                                      ml->convert_type, ml->highest_blocked);
1858                                 __dlm_print_one_lock_resource(res);
1859                                 BUG();
1860                         }
1861
1862                         if (lock->ml.node != ml->node) {
1863                                 c = lock->ml.cookie;
1864                                 mlog(ML_ERROR, "Mismatched node# in lock "
1865                                      "cookie %u:%llu, name %.*s, node %u\n",
1866                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1867                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1868                                      res->lockname.len, res->lockname.name,
1869                                      lock->ml.node);
1870                                 c = ml->cookie;
1871                                 mlog(ML_ERROR, "Migrate lock cookie %u:%llu, "
1872                                      "node %u, list %u, flags 0x%x, type %d, "
1873                                      "conv %d, highest blocked %d\n",
1874                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1875                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1876                                      ml->node, ml->list, ml->flags, ml->type,
1877                                      ml->convert_type, ml->highest_blocked);
1878                                 __dlm_print_one_lock_resource(res);
1879                                 BUG();
1880                         }
1881
1882                         if (tmpq != queue) {
1883                                 c = ml->cookie;
1884                                 mlog(0, "Lock cookie %u:%llu was on list %u "
1885                                      "instead of list %u for %.*s\n",
1886                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1887                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1888                                      j, ml->list, res->lockname.len,
1889                                      res->lockname.name);
1890                                 __dlm_print_one_lock_resource(res);
1891                                 spin_unlock(&res->spinlock);
1892                                 continue;
1893                         }
1894
1895                         /* see NOTE above about why we do not update
1896                          * to match the master here */
1897
1898                         /* move the lock to its proper place */
1899                         /* do not alter lock refcount.  switching lists. */
1900                         list_move_tail(&lock->list, queue);
1901                         spin_unlock(&res->spinlock);
1902
1903                         mlog(0, "just reordered a local lock!\n");
1904                         continue;
1905                 }
1906
1907                 /* lock is for another node. */
1908                 newlock = dlm_new_lock(ml->type, ml->node,
1909                                        be64_to_cpu(ml->cookie), NULL);
1910                 if (!newlock) {
1911                         ret = -ENOMEM;
1912                         goto leave;
1913                 }
1914                 lksb = newlock->lksb;
1915                 dlm_lock_attach_lockres(newlock, res);
1916
1917                 if (ml->convert_type != LKM_IVMODE) {
1918                         BUG_ON(queue != &res->converting);
1919                         newlock->ml.convert_type = ml->convert_type;
1920                 }
1921                 lksb->flags |= (ml->flags &
1922                                 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1923
1924                 if (ml->type == LKM_NLMODE)
1925                         goto skip_lvb;
1926
1927                 /*
1928                  * If the lock is in the blocked list it can't have a valid lvb,
1929                  * so skip it
1930                  */
1931                 if (ml->list == DLM_BLOCKED_LIST)
1932                         goto skip_lvb;
1933
1934                 if (!dlm_lvb_is_empty(mres->lvb)) {
1935                         if (lksb->flags & DLM_LKSB_PUT_LVB) {
1936                                 /* other node was trying to update
1937                                  * lvb when node died.  recreate the
1938                                  * lksb with the updated lvb. */
1939                                 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1940                                 /* the lock resource lvb update must happen
1941                                  * NOW, before the spinlock is dropped.
1942                                  * we no longer wait for the AST to update
1943                                  * the lvb. */
1944                                 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1945                         } else {
1946                                 /* otherwise, the node is sending its
1947                                  * most recent valid lvb info */
1948                                 BUG_ON(ml->type != LKM_EXMODE &&
1949                                        ml->type != LKM_PRMODE);
1950                                 if (!dlm_lvb_is_empty(res->lvb) &&
1951                                     (ml->type == LKM_EXMODE ||
1952                                      memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1953                                         int i;
1954                                         mlog(ML_ERROR, "%s:%.*s: received bad "
1955                                              "lvb! type=%d\n", dlm->name,
1956                                              res->lockname.len,
1957                                              res->lockname.name, ml->type);
1958                                         printk("lockres lvb=[");
1959                                         for (i=0; i<DLM_LVB_LEN; i++)
1960                                                 printk("%02x", res->lvb[i]);
1961                                         printk("]\nmigrated lvb=[");
1962                                         for (i=0; i<DLM_LVB_LEN; i++)
1963                                                 printk("%02x", mres->lvb[i]);
1964                                         printk("]\n");
1965                                         dlm_print_one_lock_resource(res);
1966                                         BUG();
1967                                 }
1968                                 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1969                         }
1970                 }
1971 skip_lvb:
1972
1973                 /* NOTE:
1974                  * wrt lock queue ordering and recovery:
1975                  *    1. order of locks on granted queue is
1976                  *       meaningless.
1977                  *    2. order of locks on converting queue is
1978                  *       LOST with the node death.  sorry charlie.
1979                  *    3. order of locks on the blocked queue is
1980                  *       also LOST.
1981                  * order of locks does not affect integrity, it
1982                  * just means that a lock request may get pushed
1983                  * back in line as a result of the node death.
1984                  * also note that for a given node the lock order
1985                  * for its secondary queue locks is preserved
1986                  * relative to each other, but clearly *not*
1987                  * preserved relative to locks from other nodes.
1988                  */
1989                 bad = 0;
1990                 spin_lock(&res->spinlock);
1991                 list_for_each_entry(lock, queue, list) {
1992                         if (lock->ml.cookie == ml->cookie) {
1993                                 c = lock->ml.cookie;
1994                                 mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already "
1995                                      "exists on this lockres!\n", dlm->name,
1996                                      res->lockname.len, res->lockname.name,
1997                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1998                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)));
1999
2000                                 mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, "
2001                                      "node=%u, cookie=%u:%llu, queue=%d\n",
2002                                      ml->type, ml->convert_type, ml->node,
2003                                      dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)),
2004                                      dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)),
2005                                      ml->list);
2006
2007                                 __dlm_print_one_lock_resource(res);
2008                                 bad = 1;
2009                                 break;
2010                         }
2011                 }
2012                 if (!bad) {
2013                         dlm_lock_get(newlock);
2014                         if (mres->flags & DLM_MRES_RECOVERY &&
2015                                         ml->list == DLM_CONVERTING_LIST &&
2016                                         newlock->ml.type >
2017                                         newlock->ml.convert_type) {
2018                                 /* newlock is doing downconvert, add it to the
2019                                  * head of converting list */
2020                                 list_add(&newlock->list, queue);
2021                         } else
2022                                 list_add_tail(&newlock->list, queue);
2023                         mlog(0, "%s:%.*s: added lock for node %u, "
2024                              "setting refmap bit\n", dlm->name,
2025                              res->lockname.len, res->lockname.name, ml->node);
2026                         dlm_lockres_set_refmap_bit(dlm, res, ml->node);
2027                 }
2028                 spin_unlock(&res->spinlock);
2029         }
2030         mlog(0, "done running all the locks\n");
2031
2032 leave:
2033         /* balance the ref taken when the work was queued */
2034         spin_lock(&res->spinlock);
2035         dlm_lockres_drop_inflight_ref(dlm, res);
2036         spin_unlock(&res->spinlock);
2037
2038         if (ret < 0)
2039                 mlog_errno(ret);
2040
2041         return ret;
2042 }
2043
2044 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
2045                                        struct dlm_lock_resource *res)
2046 {
2047         int i;
2048         struct list_head *queue;
2049         struct dlm_lock *lock, *next;
2050
2051         assert_spin_locked(&dlm->spinlock);
2052         assert_spin_locked(&res->spinlock);
2053         res->state |= DLM_LOCK_RES_RECOVERING;
2054         if (!list_empty(&res->recovering)) {
2055                 mlog(0,
2056                      "Recovering res %s:%.*s, is already on recovery list!\n",
2057                      dlm->name, res->lockname.len, res->lockname.name);
2058                 list_del_init(&res->recovering);
2059                 dlm_lockres_put(res);
2060         }
2061         /* We need to hold a reference while on the recovery list */
2062         dlm_lockres_get(res);
2063         list_add_tail(&res->recovering, &dlm->reco.resources);
2064
2065         /* find any pending locks and put them back on proper list */
2066         for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
2067                 queue = dlm_list_idx_to_ptr(res, i);
2068                 list_for_each_entry_safe(lock, next, queue, list) {
2069                         dlm_lock_get(lock);
2070                         if (lock->convert_pending) {
2071                                 /* move converting lock back to granted */
2072                                 mlog(0, "node died with convert pending "
2073                                      "on %.*s. move back to granted list.\n",
2074                                      res->lockname.len, res->lockname.name);
2075                                 dlm_revert_pending_convert(res, lock);
2076                                 lock->convert_pending = 0;
2077                         } else if (lock->lock_pending) {
2078                                 /* remove pending lock requests completely */
2079                                 BUG_ON(i != DLM_BLOCKED_LIST);
2080                                 mlog(0, "node died with lock pending "
2081                                      "on %.*s. remove from blocked list and skip.\n",
2082                                      res->lockname.len, res->lockname.name);
2083                                 /* lock will be floating until ref in
2084                                  * dlmlock_remote is freed after the network
2085                                  * call returns.  ok for it to not be on any
2086                                  * list since no ast can be called
2087                                  * (the master is dead). */
2088                                 dlm_revert_pending_lock(res, lock);
2089                                 lock->lock_pending = 0;
2090                         } else if (lock->unlock_pending) {
2091                                 /* if an unlock was in progress, treat as
2092                                  * if this had completed successfully
2093                                  * before sending this lock state to the
2094                                  * new master.  note that the dlm_unlock
2095                                  * call is still responsible for calling
2096                                  * the unlockast.  that will happen after
2097                                  * the network call times out.  for now,
2098                                  * just move lists to prepare the new
2099                                  * recovery master.  */
2100                                 BUG_ON(i != DLM_GRANTED_LIST);
2101                                 mlog(0, "node died with unlock pending "
2102                                      "on %.*s. remove from blocked list and skip.\n",
2103                                      res->lockname.len, res->lockname.name);
2104                                 dlm_commit_pending_unlock(res, lock);
2105                                 lock->unlock_pending = 0;
2106                         } else if (lock->cancel_pending) {
2107                                 /* if a cancel was in progress, treat as
2108                                  * if this had completed successfully
2109                                  * before sending this lock state to the
2110                                  * new master */
2111                                 BUG_ON(i != DLM_CONVERTING_LIST);
2112                                 mlog(0, "node died with cancel pending "
2113                                      "on %.*s. move back to granted list.\n",
2114                                      res->lockname.len, res->lockname.name);
2115                                 dlm_commit_pending_cancel(res, lock);
2116                                 lock->cancel_pending = 0;
2117                         }
2118                         dlm_lock_put(lock);
2119                 }
2120         }
2121 }
2122
2123
2124
2125 /* removes all recovered locks from the recovery list.
2126  * sets the res->owner to the new master.
2127  * unsets the RECOVERY flag and wakes waiters. */
2128 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
2129                                               u8 dead_node, u8 new_master)
2130 {
2131         int i;
2132         struct hlist_head *bucket;
2133         struct dlm_lock_resource *res, *next;
2134
2135         assert_spin_locked(&dlm->spinlock);
2136
2137         list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
2138                 if (res->owner == dead_node) {
2139                         mlog(0, "%s: res %.*s, Changing owner from %u to %u\n",
2140                              dlm->name, res->lockname.len, res->lockname.name,
2141                              res->owner, new_master);
2142                         list_del_init(&res->recovering);
2143                         spin_lock(&res->spinlock);
2144                         /* new_master has our reference from
2145                          * the lock state sent during recovery */
2146                         dlm_change_lockres_owner(dlm, res, new_master);
2147                         res->state &= ~DLM_LOCK_RES_RECOVERING;
2148                         if (__dlm_lockres_has_locks(res))
2149                                 __dlm_dirty_lockres(dlm, res);
2150                         spin_unlock(&res->spinlock);
2151                         wake_up(&res->wq);
2152                         dlm_lockres_put(res);
2153                 }
2154         }
2155
2156         /* this will become unnecessary eventually, but
2157          * for now we need to run the whole hash, clear
2158          * the RECOVERING state and set the owner
2159          * if necessary */
2160         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2161                 bucket = dlm_lockres_hash(dlm, i);
2162                 hlist_for_each_entry(res, bucket, hash_node) {
2163                         if (res->state & DLM_LOCK_RES_RECOVERY_WAITING) {
2164                                 spin_lock(&res->spinlock);
2165                                 res->state &= ~DLM_LOCK_RES_RECOVERY_WAITING;
2166                                 spin_unlock(&res->spinlock);
2167                                 wake_up(&res->wq);
2168                         }
2169
2170                         if (!(res->state & DLM_LOCK_RES_RECOVERING))
2171                                 continue;
2172
2173                         if (res->owner != dead_node &&
2174                             res->owner != dlm->node_num)
2175                                 continue;
2176
2177                         if (!list_empty(&res->recovering)) {
2178                                 list_del_init(&res->recovering);
2179                                 dlm_lockres_put(res);
2180                         }
2181
2182                         /* new_master has our reference from
2183                          * the lock state sent during recovery */
2184                         mlog(0, "%s: res %.*s, Changing owner from %u to %u\n",
2185                              dlm->name, res->lockname.len, res->lockname.name,
2186                              res->owner, new_master);
2187                         spin_lock(&res->spinlock);
2188                         dlm_change_lockres_owner(dlm, res, new_master);
2189                         res->state &= ~DLM_LOCK_RES_RECOVERING;
2190                         if (__dlm_lockres_has_locks(res))
2191                                 __dlm_dirty_lockres(dlm, res);
2192                         spin_unlock(&res->spinlock);
2193                         wake_up(&res->wq);
2194                 }
2195         }
2196 }
2197
2198 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
2199 {
2200         if (local) {
2201                 if (lock->ml.type != LKM_EXMODE &&
2202                     lock->ml.type != LKM_PRMODE)
2203                         return 1;
2204         } else if (lock->ml.type == LKM_EXMODE)
2205                 return 1;
2206         return 0;
2207 }
2208
2209 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
2210                                struct dlm_lock_resource *res, u8 dead_node)
2211 {
2212         struct list_head *queue;
2213         struct dlm_lock *lock;
2214         int blank_lvb = 0, local = 0;
2215         int i;
2216         u8 search_node;
2217
2218         assert_spin_locked(&dlm->spinlock);
2219         assert_spin_locked(&res->spinlock);
2220
2221         if (res->owner == dlm->node_num)
2222                 /* if this node owned the lockres, and if the dead node
2223                  * had an EX when he died, blank out the lvb */
2224                 search_node = dead_node;
2225         else {
2226                 /* if this is a secondary lockres, and we had no EX or PR
2227                  * locks granted, we can no longer trust the lvb */
2228                 search_node = dlm->node_num;
2229                 local = 1;  /* check local state for valid lvb */
2230         }
2231
2232         for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
2233                 queue = dlm_list_idx_to_ptr(res, i);
2234                 list_for_each_entry(lock, queue, list) {
2235                         if (lock->ml.node == search_node) {
2236                                 if (dlm_lvb_needs_invalidation(lock, local)) {
2237                                         /* zero the lksb lvb and lockres lvb */
2238                                         blank_lvb = 1;
2239                                         memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
2240                                 }
2241                         }
2242                 }
2243         }
2244
2245         if (blank_lvb) {
2246                 mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
2247                      res->lockname.len, res->lockname.name, dead_node);
2248                 memset(res->lvb, 0, DLM_LVB_LEN);
2249         }
2250 }
2251
2252 static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
2253                                 struct dlm_lock_resource *res, u8 dead_node)
2254 {
2255         struct dlm_lock *lock, *next;
2256         unsigned int freed = 0;
2257
2258         /* this node is the lockres master:
2259          * 1) remove any stale locks for the dead node
2260          * 2) if the dead node had an EX when he died, blank out the lvb
2261          */
2262         assert_spin_locked(&dlm->spinlock);
2263         assert_spin_locked(&res->spinlock);
2264
2265         /* We do two dlm_lock_put(). One for removing from list and the other is
2266          * to force the DLM_UNLOCK_FREE_LOCK action so as to free the locks */
2267
2268         /* TODO: check pending_asts, pending_basts here */
2269         list_for_each_entry_safe(lock, next, &res->granted, list) {
2270                 if (lock->ml.node == dead_node) {
2271                         list_del_init(&lock->list);
2272                         dlm_lock_put(lock);
2273                         /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2274                         dlm_lock_put(lock);
2275                         freed++;
2276                 }
2277         }
2278         list_for_each_entry_safe(lock, next, &res->converting, list) {
2279                 if (lock->ml.node == dead_node) {
2280                         list_del_init(&lock->list);
2281                         dlm_lock_put(lock);
2282                         /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2283                         dlm_lock_put(lock);
2284                         freed++;
2285                 }
2286         }
2287         list_for_each_entry_safe(lock, next, &res->blocked, list) {
2288                 if (lock->ml.node == dead_node) {
2289                         list_del_init(&lock->list);
2290                         dlm_lock_put(lock);
2291                         /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2292                         dlm_lock_put(lock);
2293                         freed++;
2294                 }
2295         }
2296
2297         if (freed) {
2298                 mlog(0, "%s:%.*s: freed %u locks for dead node %u, "
2299                      "dropping ref from lockres\n", dlm->name,
2300                      res->lockname.len, res->lockname.name, freed, dead_node);
2301                 if(!test_bit(dead_node, res->refmap)) {
2302                         mlog(ML_ERROR, "%s:%.*s: freed %u locks for dead node %u, "
2303                              "but ref was not set\n", dlm->name,
2304                              res->lockname.len, res->lockname.name, freed, dead_node);
2305                         __dlm_print_one_lock_resource(res);
2306                 }
2307                 res->state |= DLM_LOCK_RES_RECOVERY_WAITING;
2308                 dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2309         } else if (test_bit(dead_node, res->refmap)) {
2310                 mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2311                      "no locks and had not purged before dying\n", dlm->name,
2312                      res->lockname.len, res->lockname.name, dead_node);
2313                 dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2314         }
2315
2316         /* do not kick thread yet */
2317         __dlm_dirty_lockres(dlm, res);
2318 }
2319
2320 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
2321 {
2322         struct dlm_lock_resource *res;
2323         int i;
2324         struct hlist_head *bucket;
2325         struct hlist_node *tmp;
2326         struct dlm_lock *lock;
2327
2328
2329         /* purge any stale mles */
2330         dlm_clean_master_list(dlm, dead_node);
2331
2332         /*
2333          * now clean up all lock resources.  there are two rules:
2334          *
2335          * 1) if the dead node was the master, move the lockres
2336          *    to the recovering list.  set the RECOVERING flag.
2337          *    this lockres needs to be cleaned up before it can
2338          *    be used further.
2339          *
2340          * 2) if this node was the master, remove all locks from
2341          *    each of the lockres queues that were owned by the
2342          *    dead node.  once recovery finishes, the dlm thread
2343          *    can be kicked again to see if any ASTs or BASTs
2344          *    need to be fired as a result.
2345          */
2346         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2347                 bucket = dlm_lockres_hash(dlm, i);
2348                 hlist_for_each_entry_safe(res, tmp, bucket, hash_node) {
2349                         /* always prune any $RECOVERY entries for dead nodes,
2350                          * otherwise hangs can occur during later recovery */
2351                         if (dlm_is_recovery_lock(res->lockname.name,
2352                                                  res->lockname.len)) {
2353                                 spin_lock(&res->spinlock);
2354                                 list_for_each_entry(lock, &res->granted, list) {
2355                                         if (lock->ml.node == dead_node) {
2356                                                 mlog(0, "AHA! there was "
2357                                                      "a $RECOVERY lock for dead "
2358                                                      "node %u (%s)!\n",
2359                                                      dead_node, dlm->name);
2360                                                 list_del_init(&lock->list);
2361                                                 dlm_lock_put(lock);
2362                                                 /* Can't schedule
2363                                                  * DLM_UNLOCK_FREE_LOCK
2364                                                  * - do manually */
2365                                                 dlm_lock_put(lock);
2366                                                 break;
2367                                         }
2368                                 }
2369
2370                                 if ((res->owner == dead_node) &&
2371                                                         (res->state & DLM_LOCK_RES_DROPPING_REF)) {
2372                                         dlm_lockres_get(res);
2373                                         __dlm_do_purge_lockres(dlm, res);
2374                                         spin_unlock(&res->spinlock);
2375                                         wake_up(&res->wq);
2376                                         dlm_lockres_put(res);
2377                                         continue;
2378                                 } else if (res->owner == dlm->node_num)
2379                                         dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2380                                 spin_unlock(&res->spinlock);
2381                                 continue;
2382                         }
2383                         spin_lock(&res->spinlock);
2384                         /* zero the lvb if necessary */
2385                         dlm_revalidate_lvb(dlm, res, dead_node);
2386                         if (res->owner == dead_node) {
2387                                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
2388                                         mlog(0, "%s:%.*s: owned by "
2389                                                 "dead node %u, this node was "
2390                                                 "dropping its ref when master died. "
2391                                                 "continue, purging the lockres.\n",
2392                                                 dlm->name, res->lockname.len,
2393                                                 res->lockname.name, dead_node);
2394                                         dlm_lockres_get(res);
2395                                         __dlm_do_purge_lockres(dlm, res);
2396                                         spin_unlock(&res->spinlock);
2397                                         wake_up(&res->wq);
2398                                         dlm_lockres_put(res);
2399                                         continue;
2400                                 }
2401                                 dlm_move_lockres_to_recovery_list(dlm, res);
2402                         } else if (res->owner == dlm->node_num) {
2403                                 dlm_free_dead_locks(dlm, res, dead_node);
2404                                 __dlm_lockres_calc_usage(dlm, res);
2405                         } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
2406                                 if (test_bit(dead_node, res->refmap)) {
2407                                         mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2408                                                 "no locks and had not purged before dying\n",
2409                                                 dlm->name, res->lockname.len,
2410                                                 res->lockname.name, dead_node);
2411                                         dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2412                                 }
2413                         }
2414                         spin_unlock(&res->spinlock);
2415                 }
2416         }
2417
2418 }
2419
2420 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
2421 {
2422         assert_spin_locked(&dlm->spinlock);
2423
2424         if (dlm->reco.new_master == idx) {
2425                 mlog(0, "%s: recovery master %d just died\n",
2426                      dlm->name, idx);
2427                 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2428                         /* finalize1 was reached, so it is safe to clear
2429                          * the new_master and dead_node.  that recovery
2430                          * is complete. */
2431                         mlog(0, "%s: dead master %d had reached "
2432                              "finalize1 state, clearing\n", dlm->name, idx);
2433                         dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2434                         __dlm_reset_recovery(dlm);
2435                 }
2436         }
2437
2438         /* Clean up join state on node death. */
2439         if (dlm->joining_node == idx) {
2440                 mlog(0, "Clearing join state for node %u\n", idx);
2441                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
2442         }
2443
2444         /* check to see if the node is already considered dead */
2445         if (!test_bit(idx, dlm->live_nodes_map)) {
2446                 mlog(0, "for domain %s, node %d is already dead. "
2447                      "another node likely did recovery already.\n",
2448                      dlm->name, idx);
2449                 return;
2450         }
2451
2452         /* check to see if we do not care about this node */
2453         if (!test_bit(idx, dlm->domain_map)) {
2454                 /* This also catches the case that we get a node down
2455                  * but haven't joined the domain yet. */
2456                 mlog(0, "node %u already removed from domain!\n", idx);
2457                 return;
2458         }
2459
2460         clear_bit(idx, dlm->live_nodes_map);
2461
2462         /* make sure local cleanup occurs before the heartbeat events */
2463         if (!test_bit(idx, dlm->recovery_map))
2464                 dlm_do_local_recovery_cleanup(dlm, idx);
2465
2466         /* notify anything attached to the heartbeat events */
2467         dlm_hb_event_notify_attached(dlm, idx, 0);
2468
2469         mlog(0, "node %u being removed from domain map!\n", idx);
2470         clear_bit(idx, dlm->domain_map);
2471         clear_bit(idx, dlm->exit_domain_map);
2472         /* wake up migration waiters if a node goes down.
2473          * perhaps later we can genericize this for other waiters. */
2474         wake_up(&dlm->migration_wq);
2475
2476         set_bit(idx, dlm->recovery_map);
2477 }
2478
2479 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
2480 {
2481         struct dlm_ctxt *dlm = data;
2482
2483         if (!dlm_grab(dlm))
2484                 return;
2485
2486         /*
2487          * This will notify any dlm users that a node in our domain
2488          * went away without notifying us first.
2489          */
2490         if (test_bit(idx, dlm->domain_map))
2491                 dlm_fire_domain_eviction_callbacks(dlm, idx);
2492
2493         spin_lock(&dlm->spinlock);
2494         __dlm_hb_node_down(dlm, idx);
2495         spin_unlock(&dlm->spinlock);
2496
2497         dlm_put(dlm);
2498 }
2499
2500 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
2501 {
2502         struct dlm_ctxt *dlm = data;
2503
2504         if (!dlm_grab(dlm))
2505                 return;
2506
2507         spin_lock(&dlm->spinlock);
2508         set_bit(idx, dlm->live_nodes_map);
2509         /* do NOT notify mle attached to the heartbeat events.
2510          * new nodes are not interesting in mastery until joined. */
2511         spin_unlock(&dlm->spinlock);
2512
2513         dlm_put(dlm);
2514 }
2515
2516 static void dlm_reco_ast(void *astdata)
2517 {
2518         struct dlm_ctxt *dlm = astdata;
2519         mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2520              dlm->node_num, dlm->name);
2521 }
2522 static void dlm_reco_bast(void *astdata, int blocked_type)
2523 {
2524         struct dlm_ctxt *dlm = astdata;
2525         mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2526              dlm->node_num, dlm->name);
2527 }
2528 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
2529 {
2530         mlog(0, "unlockast for recovery lock fired!\n");
2531 }
2532
2533 /*
2534  * dlm_pick_recovery_master will continually attempt to use
2535  * dlmlock() on the special "$RECOVERY" lockres with the
2536  * LKM_NOQUEUE flag to get an EX.  every thread that enters
2537  * this function on each node racing to become the recovery
2538  * master will not stop attempting this until either:
2539  * a) this node gets the EX (and becomes the recovery master),
2540  * or b) dlm->reco.new_master gets set to some nodenum
2541  * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2542  * so each time a recovery master is needed, the entire cluster
2543  * will sync at this point.  if the new master dies, that will
2544  * be detected in dlm_do_recovery */
2545 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2546 {
2547         enum dlm_status ret;
2548         struct dlm_lockstatus lksb;
2549         int status = -EINVAL;
2550
2551         mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2552              dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2553 again:
2554         memset(&lksb, 0, sizeof(lksb));
2555
2556         ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
2557                       DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN,
2558                       dlm_reco_ast, dlm, dlm_reco_bast);
2559
2560         mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2561              dlm->name, ret, lksb.status);
2562
2563         if (ret == DLM_NORMAL) {
2564                 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2565                      dlm->name, dlm->node_num);
2566
2567                 /* got the EX lock.  check to see if another node
2568                  * just became the reco master */
2569                 if (dlm_reco_master_ready(dlm)) {
2570                         mlog(0, "%s: got reco EX lock, but %u will "
2571                              "do the recovery\n", dlm->name,
2572                              dlm->reco.new_master);
2573                         status = -EEXIST;
2574                 } else {
2575                         status = 0;
2576
2577                         /* see if recovery was already finished elsewhere */
2578                         spin_lock(&dlm->spinlock);
2579                         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
2580                                 status = -EINVAL;
2581                                 mlog(0, "%s: got reco EX lock, but "
2582                                      "node got recovered already\n", dlm->name);
2583                                 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2584                                         mlog(ML_ERROR, "%s: new master is %u "
2585                                              "but no dead node!\n",
2586                                              dlm->name, dlm->reco.new_master);
2587                                         BUG();
2588                                 }
2589                         }
2590                         spin_unlock(&dlm->spinlock);
2591                 }
2592
2593                 /* if this node has actually become the recovery master,
2594                  * set the master and send the messages to begin recovery */
2595                 if (!status) {
2596                         mlog(0, "%s: dead=%u, this=%u, sending "
2597                              "begin_reco now\n", dlm->name,
2598                              dlm->reco.dead_node, dlm->node_num);
2599                         status = dlm_send_begin_reco_message(dlm,
2600                                       dlm->reco.dead_node);
2601                         /* this always succeeds */
2602                         BUG_ON(status);
2603
2604                         /* set the new_master to this node */
2605                         spin_lock(&dlm->spinlock);
2606                         dlm_set_reco_master(dlm, dlm->node_num);
2607                         spin_unlock(&dlm->spinlock);
2608                 }
2609
2610                 /* recovery lock is a special case.  ast will not get fired,
2611                  * so just go ahead and unlock it. */
2612                 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
2613                 if (ret == DLM_DENIED) {
2614                         mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2615                         ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2616                 }
2617                 if (ret != DLM_NORMAL) {
2618                         /* this would really suck. this could only happen
2619                          * if there was a network error during the unlock
2620                          * because of node death.  this means the unlock
2621                          * is actually "done" and the lock structure is
2622                          * even freed.  we can continue, but only
2623                          * because this specific lock name is special. */
2624                         mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
2625                 }
2626         } else if (ret == DLM_NOTQUEUED) {
2627                 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2628                      dlm->name, dlm->node_num);
2629                 /* another node is master. wait on
2630                  * reco.new_master != O2NM_INVALID_NODE_NUM
2631                  * for at most one second */
2632                 wait_event_timeout(dlm->dlm_reco_thread_wq,
2633                                          dlm_reco_master_ready(dlm),
2634                                          msecs_to_jiffies(1000));
2635                 if (!dlm_reco_master_ready(dlm)) {
2636                         mlog(0, "%s: reco master taking awhile\n",
2637                              dlm->name);
2638                         goto again;
2639                 }
2640                 /* another node has informed this one that it is reco master */
2641                 mlog(0, "%s: reco master %u is ready to recover %u\n",
2642                      dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
2643                 status = -EEXIST;
2644         } else if (ret == DLM_RECOVERING) {
2645                 mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n",
2646                      dlm->name, dlm->node_num);
2647                 goto again;
2648         } else {
2649                 struct dlm_lock_resource *res;
2650
2651                 /* dlmlock returned something other than NOTQUEUED or NORMAL */
2652                 mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2653                      "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2654                      dlm_errname(lksb.status));
2655                 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2656                                          DLM_RECOVERY_LOCK_NAME_LEN);
2657                 if (res) {
2658                         dlm_print_one_lock_resource(res);
2659                         dlm_lockres_put(res);
2660                 } else {
2661                         mlog(ML_ERROR, "recovery lock not found\n");
2662                 }
2663                 BUG();
2664         }
2665
2666         return status;
2667 }
2668
2669 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2670 {
2671         struct dlm_begin_reco br;
2672         int ret = 0;
2673         struct dlm_node_iter iter;
2674         int nodenum;
2675         int status;
2676
2677         mlog(0, "%s: dead node is %u\n", dlm->name, dead_node);
2678
2679         spin_lock(&dlm->spinlock);
2680         dlm_node_iter_init(dlm->domain_map, &iter);
2681         spin_unlock(&dlm->spinlock);
2682
2683         clear_bit(dead_node, iter.node_map);
2684
2685         memset(&br, 0, sizeof(br));
2686         br.node_idx = dlm->node_num;
2687         br.dead_node = dead_node;
2688
2689         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2690                 ret = 0;
2691                 if (nodenum == dead_node) {
2692                         mlog(0, "not sending begin reco to dead node "
2693                                   "%u\n", dead_node);
2694                         continue;
2695                 }
2696                 if (nodenum == dlm->node_num) {
2697                         mlog(0, "not sending begin reco to self\n");
2698                         continue;
2699                 }
2700 retry:
2701                 mlog(0, "attempting to send begin reco msg to %d\n",
2702                           nodenum);
2703                 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2704                                          &br, sizeof(br), nodenum, &status);
2705                 /* negative status is handled ok by caller here */
2706                 if (ret >= 0)
2707                         ret = status;
2708                 if (dlm_is_host_down(ret)) {
2709                         /* node is down.  not involved in recovery
2710                          * so just keep going */
2711                         mlog(ML_NOTICE, "%s: node %u was down when sending "
2712                              "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2713                         ret = 0;
2714                 }
2715
2716                 /*
2717                  * Prior to commit aad1b15310b9bcd59fa81ab8f2b1513b59553ea8,
2718                  * dlm_begin_reco_handler() returned EAGAIN and not -EAGAIN.
2719                  * We are handling both for compatibility reasons.
2720                  */
2721                 if (ret == -EAGAIN || ret == EAGAIN) {
2722                         mlog(0, "%s: trying to start recovery of node "
2723                              "%u, but node %u is waiting for last recovery "
2724                              "to complete, backoff for a bit\n", dlm->name,
2725                              dead_node, nodenum);
2726                         msleep(100);
2727                         goto retry;
2728                 }
2729                 if (ret < 0) {
2730                         struct dlm_lock_resource *res;
2731
2732                         /* this is now a serious problem, possibly ENOMEM
2733                          * in the network stack.  must retry */
2734                         mlog_errno(ret);
2735                         mlog(ML_ERROR, "begin reco of dlm %s to node %u "
2736                              "returned %d\n", dlm->name, nodenum, ret);
2737                         res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2738                                                  DLM_RECOVERY_LOCK_NAME_LEN);
2739                         if (res) {
2740                                 dlm_print_one_lock_resource(res);
2741                                 dlm_lockres_put(res);
2742                         } else {
2743                                 mlog(ML_ERROR, "recovery lock not found\n");
2744                         }
2745                         /* sleep for a bit in hopes that we can avoid
2746                          * another ENOMEM */
2747                         msleep(100);
2748                         goto retry;
2749                 }
2750         }
2751
2752         return ret;
2753 }
2754
2755 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2756                            void **ret_data)
2757 {
2758         struct dlm_ctxt *dlm = data;
2759         struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2760
2761         /* ok to return 0, domain has gone away */
2762         if (!dlm_grab(dlm))
2763                 return 0;
2764
2765         spin_lock(&dlm->spinlock);
2766         if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2767                 mlog(0, "%s: node %u wants to recover node %u (%u:%u) "
2768                      "but this node is in finalize state, waiting on finalize2\n",
2769                      dlm->name, br->node_idx, br->dead_node,
2770                      dlm->reco.dead_node, dlm->reco.new_master);
2771                 spin_unlock(&dlm->spinlock);
2772                 dlm_put(dlm);
2773                 return -EAGAIN;
2774         }
2775         spin_unlock(&dlm->spinlock);
2776
2777         mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n",
2778              dlm->name, br->node_idx, br->dead_node,
2779              dlm->reco.dead_node, dlm->reco.new_master);
2780
2781         dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2782
2783         spin_lock(&dlm->spinlock);
2784         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2785                 if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2786                         mlog(0, "%s: new_master %u died, changing "
2787                              "to %u\n", dlm->name, dlm->reco.new_master,
2788                              br->node_idx);
2789                 } else {
2790                         mlog(0, "%s: new_master %u NOT DEAD, changing "
2791                              "to %u\n", dlm->name, dlm->reco.new_master,
2792                              br->node_idx);
2793                         /* may not have seen the new master as dead yet */
2794                 }
2795         }
2796         if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2797                 mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2798                      "node %u changing it to %u\n", dlm->name,
2799                      dlm->reco.dead_node, br->node_idx, br->dead_node);
2800         }
2801         dlm_set_reco_master(dlm, br->node_idx);
2802         dlm_set_reco_dead_node(dlm, br->dead_node);
2803         if (!test_bit(br->dead_node, dlm->recovery_map)) {
2804                 mlog(0, "recovery master %u sees %u as dead, but this "
2805                      "node has not yet.  marking %u as dead\n",
2806                      br->node_idx, br->dead_node, br->dead_node);
2807                 if (!test_bit(br->dead_node, dlm->domain_map) ||
2808                     !test_bit(br->dead_node, dlm->live_nodes_map))
2809                         mlog(0, "%u not in domain/live_nodes map "
2810                              "so setting it in reco map manually\n",
2811                              br->dead_node);
2812                 /* force the recovery cleanup in __dlm_hb_node_down
2813                  * both of these will be cleared in a moment */
2814                 set_bit(br->dead_node, dlm->domain_map);
2815                 set_bit(br->dead_node, dlm->live_nodes_map);
2816                 __dlm_hb_node_down(dlm, br->dead_node);
2817         }
2818         spin_unlock(&dlm->spinlock);
2819
2820         dlm_kick_recovery_thread(dlm);
2821
2822         mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n",
2823              dlm->name, br->node_idx, br->dead_node,
2824              dlm->reco.dead_node, dlm->reco.new_master);
2825
2826         dlm_put(dlm);
2827         return 0;
2828 }
2829
2830 #define DLM_FINALIZE_STAGE2  0x01
2831 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2832 {
2833         int ret = 0;
2834         struct dlm_finalize_reco fr;
2835         struct dlm_node_iter iter;
2836         int nodenum;
2837         int status;
2838         int stage = 1;
2839
2840         mlog(0, "finishing recovery for node %s:%u, "
2841              "stage %d\n", dlm->name, dlm->reco.dead_node, stage);
2842
2843         spin_lock(&dlm->spinlock);
2844         dlm_node_iter_init(dlm->domain_map, &iter);
2845         spin_unlock(&dlm->spinlock);
2846
2847 stage2:
2848         memset(&fr, 0, sizeof(fr));
2849         fr.node_idx = dlm->node_num;
2850         fr.dead_node = dlm->reco.dead_node;
2851         if (stage == 2)
2852                 fr.flags |= DLM_FINALIZE_STAGE2;
2853
2854         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2855                 if (nodenum == dlm->node_num)
2856                         continue;
2857                 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2858                                          &fr, sizeof(fr), nodenum, &status);
2859                 if (ret >= 0)
2860                         ret = status;
2861                 if (ret < 0) {
2862                         mlog(ML_ERROR, "Error %d when sending message %u (key "
2863                              "0x%x) to node %u\n", ret, DLM_FINALIZE_RECO_MSG,
2864                              dlm->key, nodenum);
2865                         if (dlm_is_host_down(ret)) {
2866                                 /* this has no effect on this recovery
2867                                  * session, so set the status to zero to
2868                                  * finish out the last recovery */
2869                                 mlog(ML_ERROR, "node %u went down after this "
2870                                      "node finished recovery.\n", nodenum);
2871                                 ret = 0;
2872                                 continue;
2873                         }
2874                         break;
2875                 }
2876         }
2877         if (stage == 1) {
2878                 /* reset the node_iter back to the top and send finalize2 */
2879                 iter.curnode = -1;
2880                 stage = 2;
2881                 goto stage2;
2882         }
2883
2884         return ret;
2885 }
2886
2887 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2888                               void **ret_data)
2889 {
2890         struct dlm_ctxt *dlm = data;
2891         struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2892         int stage = 1;
2893
2894         /* ok to return 0, domain has gone away */
2895         if (!dlm_grab(dlm))
2896                 return 0;
2897
2898         if (fr->flags & DLM_FINALIZE_STAGE2)
2899                 stage = 2;
2900
2901         mlog(0, "%s: node %u finalizing recovery stage%d of "
2902              "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage,
2903              fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master);
2904
2905         spin_lock(&dlm->spinlock);
2906
2907         if (dlm->reco.new_master != fr->node_idx) {
2908                 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2909                      "%u is supposed to be the new master, dead=%u\n",
2910                      fr->node_idx, dlm->reco.new_master, fr->dead_node);
2911                 BUG();
2912         }
2913         if (dlm->reco.dead_node != fr->dead_node) {
2914                 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2915                      "node %u, but node %u is supposed to be dead\n",
2916                      fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2917                 BUG();
2918         }
2919
2920         switch (stage) {
2921                 case 1:
2922                         dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2923                         if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2924                                 mlog(ML_ERROR, "%s: received finalize1 from "
2925                                      "new master %u for dead node %u, but "
2926                                      "this node has already received it!\n",
2927                                      dlm->name, fr->node_idx, fr->dead_node);
2928                                 dlm_print_reco_node_status(dlm);
2929                                 BUG();
2930                         }
2931                         dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
2932                         spin_unlock(&dlm->spinlock);
2933                         break;
2934                 case 2:
2935                         if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) {
2936                                 mlog(ML_ERROR, "%s: received finalize2 from "
2937                                      "new master %u for dead node %u, but "
2938                                      "this node did not have finalize1!\n",
2939                                      dlm->name, fr->node_idx, fr->dead_node);
2940                                 dlm_print_reco_node_status(dlm);
2941                                 BUG();
2942                         }
2943                         dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2944                         __dlm_reset_recovery(dlm);
2945                         spin_unlock(&dlm->spinlock);
2946                         dlm_kick_recovery_thread(dlm);
2947                         break;
2948         }
2949
2950         mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n",
2951              dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master);
2952
2953         dlm_put(dlm);
2954         return 0;
2955 }