c699a450b0f21a1f3ac244c605969b5c148bf088
[linux-3.10.git] / fs / ocfs2 / dlm / dlmrecovery.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * dlmrecovery.c
5  *
6  * recovery stuff
7  *
8  * Copyright (C) 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  *
25  */
26
27
28 #include <linux/module.h>
29 #include <linux/fs.h>
30 #include <linux/types.h>
31 #include <linux/slab.h>
32 #include <linux/highmem.h>
33 #include <linux/utsname.h>
34 #include <linux/init.h>
35 #include <linux/sysctl.h>
36 #include <linux/random.h>
37 #include <linux/blkdev.h>
38 #include <linux/socket.h>
39 #include <linux/inet.h>
40 #include <linux/timer.h>
41 #include <linux/kthread.h>
42 #include <linux/delay.h>
43
44
45 #include "cluster/heartbeat.h"
46 #include "cluster/nodemanager.h"
47 #include "cluster/tcp.h"
48
49 #include "dlmapi.h"
50 #include "dlmcommon.h"
51 #include "dlmdomain.h"
52
53 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY)
54 #include "cluster/masklog.h"
55
56 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node);
57
58 static int dlm_recovery_thread(void *data);
59 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
60 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
61 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm);
62 static int dlm_do_recovery(struct dlm_ctxt *dlm);
63
64 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm);
65 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node);
66 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
67 static int dlm_request_all_locks(struct dlm_ctxt *dlm,
68                                  u8 request_from, u8 dead_node);
69 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
70
71 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res);
72 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
73                                         const char *lockname, int namelen,
74                                         int total_locks, u64 cookie,
75                                         u8 flags, u8 master);
76 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
77                                     struct dlm_migratable_lockres *mres,
78                                     u8 send_to,
79                                     struct dlm_lock_resource *res,
80                                     int total_locks);
81 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
82                                      struct dlm_lock_resource *res,
83                                      struct dlm_migratable_lockres *mres);
84 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm);
85 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm,
86                                  u8 dead_node, u8 send_to);
87 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node);
88 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
89                                         struct list_head *list, u8 dead_node);
90 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
91                                               u8 dead_node, u8 new_master);
92 static void dlm_reco_ast(void *astdata);
93 static void dlm_reco_bast(void *astdata, int blocked_type);
94 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st);
95 static void dlm_request_all_locks_worker(struct dlm_work_item *item,
96                                          void *data);
97 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data);
98
99 static u64 dlm_get_next_mig_cookie(void);
100
101 static spinlock_t dlm_reco_state_lock = SPIN_LOCK_UNLOCKED;
102 static spinlock_t dlm_mig_cookie_lock = SPIN_LOCK_UNLOCKED;
103 static u64 dlm_mig_cookie = 1;
104
105 static u64 dlm_get_next_mig_cookie(void)
106 {
107         u64 c;
108         spin_lock(&dlm_mig_cookie_lock);
109         c = dlm_mig_cookie;
110         if (dlm_mig_cookie == (~0ULL))
111                 dlm_mig_cookie = 1;
112         else
113                 dlm_mig_cookie++;
114         spin_unlock(&dlm_mig_cookie_lock);
115         return c;
116 }
117
118 static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm,
119                                           u8 dead_node)
120 {
121         assert_spin_locked(&dlm->spinlock);
122         if (dlm->reco.dead_node != dead_node)
123                 mlog(0, "%s: changing dead_node from %u to %u\n",
124                      dlm->name, dlm->reco.dead_node, dead_node);
125         dlm->reco.dead_node = dead_node;
126 }
127
128 static inline void dlm_set_reco_master(struct dlm_ctxt *dlm,
129                                        u8 master)
130 {
131         assert_spin_locked(&dlm->spinlock);
132         mlog(0, "%s: changing new_master from %u to %u\n",
133              dlm->name, dlm->reco.new_master, master);
134         dlm->reco.new_master = master;
135 }
136
137 static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm)
138 {
139         assert_spin_locked(&dlm->spinlock);
140         clear_bit(dlm->reco.dead_node, dlm->recovery_map);
141         dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
142         dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
143 }
144
145 static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
146 {
147         spin_lock(&dlm->spinlock);
148         __dlm_reset_recovery(dlm);
149         spin_unlock(&dlm->spinlock);
150 }
151
152 /* Worker function used during recovery. */
153 void dlm_dispatch_work(void *data)
154 {
155         struct dlm_ctxt *dlm = (struct dlm_ctxt *)data;
156         LIST_HEAD(tmp_list);
157         struct list_head *iter, *iter2;
158         struct dlm_work_item *item;
159         dlm_workfunc_t *workfunc;
160
161         spin_lock(&dlm->work_lock);
162         list_splice_init(&dlm->work_list, &tmp_list);
163         spin_unlock(&dlm->work_lock);
164
165         list_for_each_safe(iter, iter2, &tmp_list) {
166                 item = list_entry(iter, struct dlm_work_item, list);
167                 workfunc = item->func;
168                 list_del_init(&item->list);
169
170                 /* already have ref on dlm to avoid having
171                  * it disappear.  just double-check. */
172                 BUG_ON(item->dlm != dlm);
173
174                 /* this is allowed to sleep and
175                  * call network stuff */
176                 workfunc(item, item->data);
177
178                 dlm_put(dlm);
179                 kfree(item);
180         }
181 }
182
183 /*
184  * RECOVERY THREAD
185  */
186
187 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
188 {
189         /* wake the recovery thread
190          * this will wake the reco thread in one of three places
191          * 1) sleeping with no recovery happening
192          * 2) sleeping with recovery mastered elsewhere
193          * 3) recovery mastered here, waiting on reco data */
194
195         wake_up(&dlm->dlm_reco_thread_wq);
196 }
197
198 /* Launch the recovery thread */
199 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
200 {
201         mlog(0, "starting dlm recovery thread...\n");
202
203         dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
204                                                 "dlm_reco_thread");
205         if (IS_ERR(dlm->dlm_reco_thread_task)) {
206                 mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
207                 dlm->dlm_reco_thread_task = NULL;
208                 return -EINVAL;
209         }
210
211         return 0;
212 }
213
214 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
215 {
216         if (dlm->dlm_reco_thread_task) {
217                 mlog(0, "waiting for dlm recovery thread to exit\n");
218                 kthread_stop(dlm->dlm_reco_thread_task);
219                 dlm->dlm_reco_thread_task = NULL;
220         }
221 }
222
223
224
225 /*
226  * this is lame, but here's how recovery works...
227  * 1) all recovery threads cluster wide will work on recovering
228  *    ONE node at a time
229  * 2) negotiate who will take over all the locks for the dead node.
230  *    thats right... ALL the locks.
231  * 3) once a new master is chosen, everyone scans all locks
232  *    and moves aside those mastered by the dead guy
233  * 4) each of these locks should be locked until recovery is done
234  * 5) the new master collects up all of secondary lock queue info
235  *    one lock at a time, forcing each node to communicate back
236  *    before continuing
237  * 6) each secondary lock queue responds with the full known lock info
238  * 7) once the new master has run all its locks, it sends a ALLDONE!
239  *    message to everyone
240  * 8) upon receiving this message, the secondary queue node unlocks
241  *    and responds to the ALLDONE
242  * 9) once the new master gets responses from everyone, he unlocks
243  *    everything and recovery for this dead node is done
244  *10) go back to 2) while there are still dead nodes
245  *
246  */
247
248 static void dlm_print_reco_node_status(struct dlm_ctxt *dlm)
249 {
250         struct dlm_reco_node_data *ndata;
251         struct dlm_lock_resource *res;
252
253         mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n",
254              dlm->name, dlm->dlm_reco_thread_task->pid,
255              dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive",
256              dlm->reco.dead_node, dlm->reco.new_master);
257
258         list_for_each_entry(ndata, &dlm->reco.node_data, list) {
259                 char *st = "unknown";
260                 switch (ndata->state) {
261                         case DLM_RECO_NODE_DATA_INIT:
262                                 st = "init";
263                                 break;
264                         case DLM_RECO_NODE_DATA_REQUESTING:
265                                 st = "requesting";
266                                 break;
267                         case DLM_RECO_NODE_DATA_DEAD:
268                                 st = "dead";
269                                 break;
270                         case DLM_RECO_NODE_DATA_RECEIVING:
271                                 st = "receiving";
272                                 break;
273                         case DLM_RECO_NODE_DATA_REQUESTED:
274                                 st = "requested";
275                                 break;
276                         case DLM_RECO_NODE_DATA_DONE:
277                                 st = "done";
278                                 break;
279                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
280                                 st = "finalize-sent";
281                                 break;
282                         default:
283                                 st = "bad";
284                                 break;
285                 }
286                 mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n",
287                      dlm->name, ndata->node_num, st);
288         }
289         list_for_each_entry(res, &dlm->reco.resources, recovering) {
290                 mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n",
291                      dlm->name, res->lockname.len, res->lockname.name);
292         }
293 }
294
295 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
296
297 static int dlm_recovery_thread(void *data)
298 {
299         int status;
300         struct dlm_ctxt *dlm = data;
301         unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
302
303         mlog(0, "dlm thread running for %s...\n", dlm->name);
304
305         while (!kthread_should_stop()) {
306                 if (dlm_joined(dlm)) {
307                         status = dlm_do_recovery(dlm);
308                         if (status == -EAGAIN) {
309                                 /* do not sleep, recheck immediately. */
310                                 continue;
311                         }
312                         if (status < 0)
313                                 mlog_errno(status);
314                 }
315
316                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
317                                                  kthread_should_stop(),
318                                                  timeout);
319         }
320
321         mlog(0, "quitting DLM recovery thread\n");
322         return 0;
323 }
324
325 /* returns true when the recovery master has contacted us */
326 static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
327 {
328         int ready;
329         spin_lock(&dlm->spinlock);
330         ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
331         spin_unlock(&dlm->spinlock);
332         return ready;
333 }
334
335 /* returns true if node is no longer in the domain
336  * could be dead or just not joined */
337 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
338 {
339         int dead;
340         spin_lock(&dlm->spinlock);
341         dead = !test_bit(node, dlm->domain_map);
342         spin_unlock(&dlm->spinlock);
343         return dead;
344 }
345
346 int dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout)
347 {
348         if (timeout) {
349                 mlog(ML_NOTICE, "%s: waiting %dms for notification of "
350                      "death of node %u\n", dlm->name, timeout, node);
351                 wait_event_timeout(dlm->dlm_reco_thread_wq,
352                            dlm_is_node_dead(dlm, node),
353                            msecs_to_jiffies(timeout));
354         } else {
355                 mlog(ML_NOTICE, "%s: waiting indefinitely for notification "
356                      "of death of node %u\n", dlm->name, node);
357                 wait_event(dlm->dlm_reco_thread_wq,
358                            dlm_is_node_dead(dlm, node));
359         }
360         /* for now, return 0 */
361         return 0;
362 }
363
364 /* callers of the top-level api calls (dlmlock/dlmunlock) should
365  * block on the dlm->reco.event when recovery is in progress.
366  * the dlm recovery thread will set this state when it begins
367  * recovering a dead node (as the new master or not) and clear
368  * the state and wake as soon as all affected lock resources have
369  * been marked with the RECOVERY flag */
370 static int dlm_in_recovery(struct dlm_ctxt *dlm)
371 {
372         int in_recovery;
373         spin_lock(&dlm->spinlock);
374         in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
375         spin_unlock(&dlm->spinlock);
376         return in_recovery;
377 }
378
379
380 void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
381 {
382         wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
383 }
384
385 static void dlm_begin_recovery(struct dlm_ctxt *dlm)
386 {
387         spin_lock(&dlm->spinlock);
388         BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
389         dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
390         spin_unlock(&dlm->spinlock);
391 }
392
393 static void dlm_end_recovery(struct dlm_ctxt *dlm)
394 {
395         spin_lock(&dlm->spinlock);
396         BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
397         dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
398         spin_unlock(&dlm->spinlock);
399         wake_up(&dlm->reco.event);
400 }
401
402 static int dlm_do_recovery(struct dlm_ctxt *dlm)
403 {
404         int status = 0;
405         int ret;
406
407         spin_lock(&dlm->spinlock);
408
409         /* check to see if the new master has died */
410         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
411             test_bit(dlm->reco.new_master, dlm->recovery_map)) {
412                 mlog(0, "new master %u died while recovering %u!\n",
413                      dlm->reco.new_master, dlm->reco.dead_node);
414                 /* unset the new_master, leave dead_node */
415                 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
416         }
417
418         /* select a target to recover */
419         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
420                 int bit;
421
422                 bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES+1, 0);
423                 if (bit >= O2NM_MAX_NODES || bit < 0)
424                         dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
425                 else
426                         dlm_set_reco_dead_node(dlm, bit);
427         } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
428                 /* BUG? */
429                 mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
430                      dlm->reco.dead_node);
431                 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
432         }
433
434         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
435                 // mlog(0, "nothing to recover!  sleeping now!\n");
436                 spin_unlock(&dlm->spinlock);
437                 /* return to main thread loop and sleep. */
438                 return 0;
439         }
440         mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n",
441              dlm->name, dlm->dlm_reco_thread_task->pid,
442              dlm->reco.dead_node);
443         spin_unlock(&dlm->spinlock);
444
445         /* take write barrier */
446         /* (stops the list reshuffling thread, proxy ast handling) */
447         dlm_begin_recovery(dlm);
448
449         if (dlm->reco.new_master == dlm->node_num)
450                 goto master_here;
451
452         if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
453                 /* choose a new master, returns 0 if this node
454                  * is the master, -EEXIST if it's another node.
455                  * this does not return until a new master is chosen
456                  * or recovery completes entirely. */
457                 ret = dlm_pick_recovery_master(dlm);
458                 if (!ret) {
459                         /* already notified everyone.  go. */
460                         goto master_here;
461                 }
462                 mlog(0, "another node will master this recovery session.\n");
463         }
464         mlog(0, "dlm=%s (%d), new_master=%u, this node=%u, dead_node=%u\n",
465              dlm->name, dlm->dlm_reco_thread_task->pid, dlm->reco.new_master,
466              dlm->node_num, dlm->reco.dead_node);
467
468         /* it is safe to start everything back up here
469          * because all of the dead node's lock resources
470          * have been marked as in-recovery */
471         dlm_end_recovery(dlm);
472
473         /* sleep out in main dlm_recovery_thread loop. */
474         return 0;
475
476 master_here:
477         mlog(0, "(%d) mastering recovery of %s:%u here(this=%u)!\n",
478              dlm->dlm_reco_thread_task->pid,
479              dlm->name, dlm->reco.dead_node, dlm->node_num);
480
481         status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
482         if (status < 0) {
483                 /* we should never hit this anymore */
484                 mlog(ML_ERROR, "error %d remastering locks for node %u, "
485                      "retrying.\n", status, dlm->reco.dead_node);
486                 /* yield a bit to allow any final network messages
487                  * to get handled on remaining nodes */
488                 msleep(100);
489         } else {
490                 /* success!  see if any other nodes need recovery */
491                 mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
492                      dlm->name, dlm->reco.dead_node, dlm->node_num);
493                 dlm_reset_recovery(dlm);
494         }
495         dlm_end_recovery(dlm);
496
497         /* continue and look for another dead node */
498         return -EAGAIN;
499 }
500
501 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
502 {
503         int status = 0;
504         struct dlm_reco_node_data *ndata;
505         struct list_head *iter;
506         int all_nodes_done;
507         int destroy = 0;
508         int pass = 0;
509
510         do {
511                 /* we have become recovery master.  there is no escaping
512                  * this, so just keep trying until we get it. */
513                 status = dlm_init_recovery_area(dlm, dead_node);
514                 if (status < 0) {
515                         mlog(ML_ERROR, "%s: failed to alloc recovery area, "
516                              "retrying\n", dlm->name);
517                         msleep(1000);
518                 }
519         } while (status != 0);
520
521         /* safe to access the node data list without a lock, since this
522          * process is the only one to change the list */
523         list_for_each(iter, &dlm->reco.node_data) {
524                 ndata = list_entry (iter, struct dlm_reco_node_data, list);
525                 BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
526                 ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
527
528                 mlog(0, "requesting lock info from node %u\n",
529                      ndata->node_num);
530
531                 if (ndata->node_num == dlm->node_num) {
532                         ndata->state = DLM_RECO_NODE_DATA_DONE;
533                         continue;
534                 }
535
536                 do {
537                         status = dlm_request_all_locks(dlm, ndata->node_num,
538                                                        dead_node);
539                         if (status < 0) {
540                                 mlog_errno(status);
541                                 if (dlm_is_host_down(status)) {
542                                         /* node died, ignore it for recovery */
543                                         status = 0;
544                                         ndata->state = DLM_RECO_NODE_DATA_DEAD;
545                                         /* wait for the domain map to catch up
546                                          * with the network state. */
547                                         wait_event_timeout(dlm->dlm_reco_thread_wq,
548                                                            dlm_is_node_dead(dlm,
549                                                                 ndata->node_num),
550                                                            msecs_to_jiffies(1000));
551                                         mlog(0, "waited 1 sec for %u, "
552                                              "dead? %s\n", ndata->node_num,
553                                              dlm_is_node_dead(dlm, ndata->node_num) ?
554                                              "yes" : "no");
555                                 } else {
556                                         /* -ENOMEM on the other node */
557                                         mlog(0, "%s: node %u returned "
558                                              "%d during recovery, retrying "
559                                              "after a short wait\n",
560                                              dlm->name, ndata->node_num,
561                                              status);
562                                         msleep(100);
563                                 }
564                         }
565                 } while (status != 0);
566
567                 switch (ndata->state) {
568                         case DLM_RECO_NODE_DATA_INIT:
569                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
570                         case DLM_RECO_NODE_DATA_REQUESTED:
571                                 BUG();
572                                 break;
573                         case DLM_RECO_NODE_DATA_DEAD:
574                                 mlog(0, "node %u died after requesting "
575                                      "recovery info for node %u\n",
576                                      ndata->node_num, dead_node);
577                                 /* fine.  don't need this node's info.
578                                  * continue without it. */
579                                 break;
580                         case DLM_RECO_NODE_DATA_REQUESTING:
581                                 ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
582                                 mlog(0, "now receiving recovery data from "
583                                      "node %u for dead node %u\n",
584                                      ndata->node_num, dead_node);
585                                 break;
586                         case DLM_RECO_NODE_DATA_RECEIVING:
587                                 mlog(0, "already receiving recovery data from "
588                                      "node %u for dead node %u\n",
589                                      ndata->node_num, dead_node);
590                                 break;
591                         case DLM_RECO_NODE_DATA_DONE:
592                                 mlog(0, "already DONE receiving recovery data "
593                                      "from node %u for dead node %u\n",
594                                      ndata->node_num, dead_node);
595                                 break;
596                 }
597         }
598
599         mlog(0, "done requesting all lock info\n");
600
601         /* nodes should be sending reco data now
602          * just need to wait */
603
604         while (1) {
605                 /* check all the nodes now to see if we are
606                  * done, or if anyone died */
607                 all_nodes_done = 1;
608                 spin_lock(&dlm_reco_state_lock);
609                 list_for_each(iter, &dlm->reco.node_data) {
610                         ndata = list_entry (iter, struct dlm_reco_node_data, list);
611
612                         mlog(0, "checking recovery state of node %u\n",
613                              ndata->node_num);
614                         switch (ndata->state) {
615                                 case DLM_RECO_NODE_DATA_INIT:
616                                 case DLM_RECO_NODE_DATA_REQUESTING:
617                                         mlog(ML_ERROR, "bad ndata state for "
618                                              "node %u: state=%d\n",
619                                              ndata->node_num, ndata->state);
620                                         BUG();
621                                         break;
622                                 case DLM_RECO_NODE_DATA_DEAD:
623                                         mlog(0, "node %u died after "
624                                              "requesting recovery info for "
625                                              "node %u\n", ndata->node_num,
626                                              dead_node);
627                                         break;
628                                 case DLM_RECO_NODE_DATA_RECEIVING:
629                                 case DLM_RECO_NODE_DATA_REQUESTED:
630                                         mlog(0, "%s: node %u still in state %s\n",
631                                              dlm->name, ndata->node_num,
632                                              ndata->state==DLM_RECO_NODE_DATA_RECEIVING ?
633                                              "receiving" : "requested");
634                                         all_nodes_done = 0;
635                                         break;
636                                 case DLM_RECO_NODE_DATA_DONE:
637                                         mlog(0, "%s: node %u state is done\n",
638                                              dlm->name, ndata->node_num);
639                                         break;
640                                 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
641                                         mlog(0, "%s: node %u state is finalize\n",
642                                              dlm->name, ndata->node_num);
643                                         break;
644                         }
645                 }
646                 spin_unlock(&dlm_reco_state_lock);
647
648                 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
649                      all_nodes_done?"yes":"no");
650                 if (all_nodes_done) {
651                         int ret;
652
653                         /* all nodes are now in DLM_RECO_NODE_DATA_DONE state
654                          * just send a finalize message to everyone and
655                          * clean up */
656                         mlog(0, "all nodes are done! send finalize\n");
657                         ret = dlm_send_finalize_reco_message(dlm);
658                         if (ret < 0)
659                                 mlog_errno(ret);
660
661                         spin_lock(&dlm->spinlock);
662                         dlm_finish_local_lockres_recovery(dlm, dead_node,
663                                                           dlm->node_num);
664                         spin_unlock(&dlm->spinlock);
665                         mlog(0, "should be done with recovery!\n");
666
667                         mlog(0, "finishing recovery of %s at %lu, "
668                              "dead=%u, this=%u, new=%u\n", dlm->name,
669                              jiffies, dlm->reco.dead_node,
670                              dlm->node_num, dlm->reco.new_master);
671                         destroy = 1;
672                         status = 0;
673                         /* rescan everything marked dirty along the way */
674                         dlm_kick_thread(dlm, NULL);
675                         break;
676                 }
677                 /* wait to be signalled, with periodic timeout
678                  * to check for node death */
679                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
680                                          kthread_should_stop(),
681                                          msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
682
683         }
684
685         if (destroy)
686                 dlm_destroy_recovery_area(dlm, dead_node);
687
688         mlog_exit(status);
689         return status;
690 }
691
692 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
693 {
694         int num=0;
695         struct dlm_reco_node_data *ndata;
696
697         spin_lock(&dlm->spinlock);
698         memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
699         /* nodes can only be removed (by dying) after dropping
700          * this lock, and death will be trapped later, so this should do */
701         spin_unlock(&dlm->spinlock);
702
703         while (1) {
704                 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
705                 if (num >= O2NM_MAX_NODES) {
706                         break;
707                 }
708                 BUG_ON(num == dead_node);
709
710                 ndata = kcalloc(1, sizeof(*ndata), GFP_KERNEL);
711                 if (!ndata) {
712                         dlm_destroy_recovery_area(dlm, dead_node);
713                         return -ENOMEM;
714                 }
715                 ndata->node_num = num;
716                 ndata->state = DLM_RECO_NODE_DATA_INIT;
717                 spin_lock(&dlm_reco_state_lock);
718                 list_add_tail(&ndata->list, &dlm->reco.node_data);
719                 spin_unlock(&dlm_reco_state_lock);
720                 num++;
721         }
722
723         return 0;
724 }
725
726 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
727 {
728         struct list_head *iter, *iter2;
729         struct dlm_reco_node_data *ndata;
730         LIST_HEAD(tmplist);
731
732         spin_lock(&dlm_reco_state_lock);
733         list_splice_init(&dlm->reco.node_data, &tmplist);
734         spin_unlock(&dlm_reco_state_lock);
735
736         list_for_each_safe(iter, iter2, &tmplist) {
737                 ndata = list_entry (iter, struct dlm_reco_node_data, list);
738                 list_del_init(&ndata->list);
739                 kfree(ndata);
740         }
741 }
742
743 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
744                                  u8 dead_node)
745 {
746         struct dlm_lock_request lr;
747         enum dlm_status ret;
748
749         mlog(0, "\n");
750
751
752         mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
753                   "to %u\n", dead_node, request_from);
754
755         memset(&lr, 0, sizeof(lr));
756         lr.node_idx = dlm->node_num;
757         lr.dead_node = dead_node;
758
759         // send message
760         ret = DLM_NOLOCKMGR;
761         ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
762                                  &lr, sizeof(lr), request_from, NULL);
763
764         /* negative status is handled by caller */
765         if (ret < 0)
766                 mlog_errno(ret);
767
768         // return from here, then
769         // sleep until all received or error
770         return ret;
771
772 }
773
774 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data)
775 {
776         struct dlm_ctxt *dlm = data;
777         struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
778         char *buf = NULL;
779         struct dlm_work_item *item = NULL;
780
781         if (!dlm_grab(dlm))
782                 return -EINVAL;
783
784         if (lr->dead_node != dlm->reco.dead_node) {
785                 mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local "
786                      "dead_node is %u\n", dlm->name, lr->node_idx,
787                      lr->dead_node, dlm->reco.dead_node);
788                 dlm_print_reco_node_status(dlm);
789                 /* this is a hack */
790                 dlm_put(dlm);
791                 return -ENOMEM;
792         }
793         BUG_ON(lr->dead_node != dlm->reco.dead_node);
794
795         item = kcalloc(1, sizeof(*item), GFP_KERNEL);
796         if (!item) {
797                 dlm_put(dlm);
798                 return -ENOMEM;
799         }
800
801         /* this will get freed by dlm_request_all_locks_worker */
802         buf = (char *) __get_free_page(GFP_KERNEL);
803         if (!buf) {
804                 kfree(item);
805                 dlm_put(dlm);
806                 return -ENOMEM;
807         }
808
809         /* queue up work for dlm_request_all_locks_worker */
810         dlm_grab(dlm);  /* get an extra ref for the work item */
811         dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
812         item->u.ral.reco_master = lr->node_idx;
813         item->u.ral.dead_node = lr->dead_node;
814         spin_lock(&dlm->work_lock);
815         list_add_tail(&item->list, &dlm->work_list);
816         spin_unlock(&dlm->work_lock);
817         schedule_work(&dlm->dispatched_work);
818
819         dlm_put(dlm);
820         return 0;
821 }
822
823 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
824 {
825         struct dlm_migratable_lockres *mres;
826         struct dlm_lock_resource *res;
827         struct dlm_ctxt *dlm;
828         LIST_HEAD(resources);
829         struct list_head *iter;
830         int ret;
831         u8 dead_node, reco_master;
832         int skip_all_done = 0;
833
834         dlm = item->dlm;
835         dead_node = item->u.ral.dead_node;
836         reco_master = item->u.ral.reco_master;
837         mres = (struct dlm_migratable_lockres *)data;
838
839         mlog(0, "%s: recovery worker started, dead=%u, master=%u\n",
840              dlm->name, dead_node, reco_master);
841
842         if (dead_node != dlm->reco.dead_node ||
843             reco_master != dlm->reco.new_master) {
844                 /* worker could have been created before the recovery master
845                  * died.  if so, do not continue, but do not error. */
846                 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
847                         mlog(ML_NOTICE, "%s: will not send recovery state, "
848                              "recovery master %u died, thread=(dead=%u,mas=%u)"
849                              " current=(dead=%u,mas=%u)\n", dlm->name,
850                              reco_master, dead_node, reco_master,
851                              dlm->reco.dead_node, dlm->reco.new_master);
852                 } else {
853                         mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, "
854                              "master=%u), request(dead=%u, master=%u)\n",
855                              dlm->name, dlm->reco.dead_node,
856                              dlm->reco.new_master, dead_node, reco_master);
857                 }
858                 goto leave;
859         }
860
861         /* lock resources should have already been moved to the
862          * dlm->reco.resources list.  now move items from that list
863          * to a temp list if the dead owner matches.  note that the
864          * whole cluster recovers only one node at a time, so we
865          * can safely move UNKNOWN lock resources for each recovery
866          * session. */
867         dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
868
869         /* now we can begin blasting lockreses without the dlm lock */
870
871         /* any errors returned will be due to the new_master dying,
872          * the dlm_reco_thread should detect this */
873         list_for_each(iter, &resources) {
874                 res = list_entry (iter, struct dlm_lock_resource, recovering);
875                 ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
876                                         DLM_MRES_RECOVERY);
877                 if (ret < 0) {
878                         mlog(ML_ERROR, "%s: node %u went down while sending "
879                              "recovery state for dead node %u, ret=%d\n", dlm->name,
880                              reco_master, dead_node, ret);
881                         skip_all_done = 1;
882                         break;
883                 }
884         }
885
886         /* move the resources back to the list */
887         spin_lock(&dlm->spinlock);
888         list_splice_init(&resources, &dlm->reco.resources);
889         spin_unlock(&dlm->spinlock);
890
891         if (!skip_all_done) {
892                 ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
893                 if (ret < 0) {
894                         mlog(ML_ERROR, "%s: node %u went down while sending "
895                              "recovery all-done for dead node %u, ret=%d\n",
896                              dlm->name, reco_master, dead_node, ret);
897                 }
898         }
899 leave:
900         free_page((unsigned long)data);
901 }
902
903
904 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
905 {
906         int ret, tmpret;
907         struct dlm_reco_data_done done_msg;
908
909         memset(&done_msg, 0, sizeof(done_msg));
910         done_msg.node_idx = dlm->node_num;
911         done_msg.dead_node = dead_node;
912         mlog(0, "sending DATA DONE message to %u, "
913              "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
914              done_msg.dead_node);
915
916         ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
917                                  sizeof(done_msg), send_to, &tmpret);
918         if (ret < 0) {
919                 if (!dlm_is_host_down(ret)) {
920                         mlog_errno(ret);
921                         mlog(ML_ERROR, "%s: unknown error sending data-done "
922                              "to %u\n", dlm->name, send_to);
923                         BUG();
924                 }
925         } else
926                 ret = tmpret;
927         return ret;
928 }
929
930
931 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data)
932 {
933         struct dlm_ctxt *dlm = data;
934         struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
935         struct list_head *iter;
936         struct dlm_reco_node_data *ndata = NULL;
937         int ret = -EINVAL;
938
939         if (!dlm_grab(dlm))
940                 return -EINVAL;
941
942         mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
943              "node_idx=%u, this node=%u\n", done->dead_node,
944              dlm->reco.dead_node, done->node_idx, dlm->node_num);
945
946         mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node),
947                         "Got DATA DONE: dead_node=%u, reco.dead_node=%u, "
948                         "node_idx=%u, this node=%u\n", done->dead_node,
949                         dlm->reco.dead_node, done->node_idx, dlm->node_num);
950
951         spin_lock(&dlm_reco_state_lock);
952         list_for_each(iter, &dlm->reco.node_data) {
953                 ndata = list_entry (iter, struct dlm_reco_node_data, list);
954                 if (ndata->node_num != done->node_idx)
955                         continue;
956
957                 switch (ndata->state) {
958                         /* should have moved beyond INIT but not to FINALIZE yet */
959                         case DLM_RECO_NODE_DATA_INIT:
960                         case DLM_RECO_NODE_DATA_DEAD:
961                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
962                                 mlog(ML_ERROR, "bad ndata state for node %u:"
963                                      " state=%d\n", ndata->node_num,
964                                      ndata->state);
965                                 BUG();
966                                 break;
967                         /* these states are possible at this point, anywhere along
968                          * the line of recovery */
969                         case DLM_RECO_NODE_DATA_DONE:
970                         case DLM_RECO_NODE_DATA_RECEIVING:
971                         case DLM_RECO_NODE_DATA_REQUESTED:
972                         case DLM_RECO_NODE_DATA_REQUESTING:
973                                 mlog(0, "node %u is DONE sending "
974                                           "recovery data!\n",
975                                           ndata->node_num);
976
977                                 ndata->state = DLM_RECO_NODE_DATA_DONE;
978                                 ret = 0;
979                                 break;
980                 }
981         }
982         spin_unlock(&dlm_reco_state_lock);
983
984         /* wake the recovery thread, some node is done */
985         if (!ret)
986                 dlm_kick_recovery_thread(dlm);
987
988         if (ret < 0)
989                 mlog(ML_ERROR, "failed to find recovery node data for node "
990                      "%u\n", done->node_idx);
991         dlm_put(dlm);
992
993         mlog(0, "leaving reco data done handler, ret=%d\n", ret);
994         return ret;
995 }
996
997 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
998                                         struct list_head *list,
999                                         u8 dead_node)
1000 {
1001         struct dlm_lock_resource *res;
1002         struct list_head *iter, *iter2;
1003         struct dlm_lock *lock;
1004
1005         spin_lock(&dlm->spinlock);
1006         list_for_each_safe(iter, iter2, &dlm->reco.resources) {
1007                 res = list_entry (iter, struct dlm_lock_resource, recovering);
1008                 /* always prune any $RECOVERY entries for dead nodes,
1009                  * otherwise hangs can occur during later recovery */
1010                 if (dlm_is_recovery_lock(res->lockname.name,
1011                                          res->lockname.len)) {
1012                         spin_lock(&res->spinlock);
1013                         list_for_each_entry(lock, &res->granted, list) {
1014                                 if (lock->ml.node == dead_node) {
1015                                         mlog(0, "AHA! there was "
1016                                              "a $RECOVERY lock for dead "
1017                                              "node %u (%s)!\n", 
1018                                              dead_node, dlm->name);
1019                                         list_del_init(&lock->list);
1020                                         dlm_lock_put(lock);
1021                                         break;
1022                                 }
1023                         }
1024                         spin_unlock(&res->spinlock);
1025                         continue;
1026                 }
1027
1028                 if (res->owner == dead_node) {
1029                         mlog(0, "found lockres owned by dead node while "
1030                                   "doing recovery for node %u. sending it.\n",
1031                                   dead_node);
1032                         list_move_tail(&res->recovering, list);
1033                 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1034                         mlog(0, "found UNKNOWN owner while doing recovery "
1035                                   "for node %u. sending it.\n", dead_node);
1036                         list_move_tail(&res->recovering, list);
1037                 }
1038         }
1039         spin_unlock(&dlm->spinlock);
1040 }
1041
1042 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
1043 {
1044         int total_locks = 0;
1045         struct list_head *iter, *queue = &res->granted;
1046         int i;
1047
1048         for (i=0; i<3; i++) {
1049                 list_for_each(iter, queue)
1050                         total_locks++;
1051                 queue++;
1052         }
1053         return total_locks;
1054 }
1055
1056
1057 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
1058                                       struct dlm_migratable_lockres *mres,
1059                                       u8 send_to,
1060                                       struct dlm_lock_resource *res,
1061                                       int total_locks)
1062 {
1063         u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
1064         int mres_total_locks = be32_to_cpu(mres->total_locks);
1065         int sz, ret = 0, status = 0;
1066         u8 orig_flags = mres->flags,
1067            orig_master = mres->master;
1068
1069         BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
1070         if (!mres->num_locks)
1071                 return 0;
1072
1073         sz = sizeof(struct dlm_migratable_lockres) +
1074                 (mres->num_locks * sizeof(struct dlm_migratable_lock));
1075
1076         /* add an all-done flag if we reached the last lock */
1077         orig_flags = mres->flags;
1078         BUG_ON(total_locks > mres_total_locks);
1079         if (total_locks == mres_total_locks)
1080                 mres->flags |= DLM_MRES_ALL_DONE;
1081
1082         /* send it */
1083         ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
1084                                  sz, send_to, &status);
1085         if (ret < 0) {
1086                 /* XXX: negative status is not handled.
1087                  * this will end up killing this node. */
1088                 mlog_errno(ret);
1089         } else {
1090                 /* might get an -ENOMEM back here */
1091                 ret = status;
1092                 if (ret < 0) {
1093                         mlog_errno(ret);
1094
1095                         if (ret == -EFAULT) {
1096                                 mlog(ML_ERROR, "node %u told me to kill "
1097                                      "myself!\n", send_to);
1098                                 BUG();
1099                         }
1100                 }
1101         }
1102
1103         /* zero and reinit the message buffer */
1104         dlm_init_migratable_lockres(mres, res->lockname.name,
1105                                     res->lockname.len, mres_total_locks,
1106                                     mig_cookie, orig_flags, orig_master);
1107         return ret;
1108 }
1109
1110 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
1111                                         const char *lockname, int namelen,
1112                                         int total_locks, u64 cookie,
1113                                         u8 flags, u8 master)
1114 {
1115         /* mres here is one full page */
1116         memset(mres, 0, PAGE_SIZE);
1117         mres->lockname_len = namelen;
1118         memcpy(mres->lockname, lockname, namelen);
1119         mres->num_locks = 0;
1120         mres->total_locks = cpu_to_be32(total_locks);
1121         mres->mig_cookie = cpu_to_be64(cookie);
1122         mres->flags = flags;
1123         mres->master = master;
1124 }
1125
1126
1127 /* returns 1 if this lock fills the network structure,
1128  * 0 otherwise */
1129 static int dlm_add_lock_to_array(struct dlm_lock *lock,
1130                                  struct dlm_migratable_lockres *mres, int queue)
1131 {
1132         struct dlm_migratable_lock *ml;
1133         int lock_num = mres->num_locks;
1134
1135         ml = &(mres->ml[lock_num]);
1136         ml->cookie = lock->ml.cookie;
1137         ml->type = lock->ml.type;
1138         ml->convert_type = lock->ml.convert_type;
1139         ml->highest_blocked = lock->ml.highest_blocked;
1140         ml->list = queue;
1141         if (lock->lksb) {
1142                 ml->flags = lock->lksb->flags;
1143                 /* send our current lvb */
1144                 if (ml->type == LKM_EXMODE ||
1145                     ml->type == LKM_PRMODE) {
1146                         /* if it is already set, this had better be a PR
1147                          * and it has to match */
1148                         if (!dlm_lvb_is_empty(mres->lvb) &&
1149                             (ml->type == LKM_EXMODE ||
1150                              memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))) {
1151                                 mlog(ML_ERROR, "mismatched lvbs!\n");
1152                                 __dlm_print_one_lock_resource(lock->lockres);
1153                                 BUG();
1154                         }
1155                         memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1156                 }
1157         }
1158         ml->node = lock->ml.node;
1159         mres->num_locks++;
1160         /* we reached the max, send this network message */
1161         if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1162                 return 1;
1163         return 0;
1164 }
1165
1166
1167 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1168                          struct dlm_migratable_lockres *mres,
1169                          u8 send_to, u8 flags)
1170 {
1171         struct list_head *queue, *iter;
1172         int total_locks, i;
1173         u64 mig_cookie = 0;
1174         struct dlm_lock *lock;
1175         int ret = 0;
1176
1177         BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1178
1179         mlog(0, "sending to %u\n", send_to);
1180
1181         total_locks = dlm_num_locks_in_lockres(res);
1182         if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1183                 /* rare, but possible */
1184                 mlog(0, "argh.  lockres has %d locks.  this will "
1185                           "require more than one network packet to "
1186                           "migrate\n", total_locks);
1187                 mig_cookie = dlm_get_next_mig_cookie();
1188         }
1189
1190         dlm_init_migratable_lockres(mres, res->lockname.name,
1191                                     res->lockname.len, total_locks,
1192                                     mig_cookie, flags, res->owner);
1193
1194         total_locks = 0;
1195         for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1196                 queue = dlm_list_idx_to_ptr(res, i);
1197                 list_for_each(iter, queue) {
1198                         lock = list_entry (iter, struct dlm_lock, list);
1199
1200                         /* add another lock. */
1201                         total_locks++;
1202                         if (!dlm_add_lock_to_array(lock, mres, i))
1203                                 continue;
1204
1205                         /* this filled the lock message,
1206                          * we must send it immediately. */
1207                         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1208                                                        res, total_locks);
1209                         if (ret < 0)
1210                                 goto error;
1211                 }
1212         }
1213         /* flush any remaining locks */
1214         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1215         if (ret < 0)
1216                 goto error;
1217         return ret;
1218
1219 error:
1220         mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n",
1221              dlm->name, ret);
1222         if (!dlm_is_host_down(ret))
1223                 BUG();
1224         mlog(0, "%s: node %u went down while sending %s "
1225              "lockres %.*s\n", dlm->name, send_to,
1226              flags & DLM_MRES_RECOVERY ?  "recovery" : "migration",
1227              res->lockname.len, res->lockname.name);
1228         return ret;
1229 }
1230
1231
1232
1233 /*
1234  * this message will contain no more than one page worth of
1235  * recovery data, and it will work on only one lockres.
1236  * there may be many locks in this page, and we may need to wait
1237  * for additional packets to complete all the locks (rare, but
1238  * possible).
1239  */
1240 /*
1241  * NOTE: the allocation error cases here are scary
1242  * we really cannot afford to fail an alloc in recovery
1243  * do we spin?  returning an error only delays the problem really
1244  */
1245
1246 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data)
1247 {
1248         struct dlm_ctxt *dlm = data;
1249         struct dlm_migratable_lockres *mres =
1250                 (struct dlm_migratable_lockres *)msg->buf;
1251         int ret = 0;
1252         u8 real_master;
1253         char *buf = NULL;
1254         struct dlm_work_item *item = NULL;
1255         struct dlm_lock_resource *res = NULL;
1256
1257         if (!dlm_grab(dlm))
1258                 return -EINVAL;
1259
1260         BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1261
1262         real_master = mres->master;
1263         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1264                 /* cannot migrate a lockres with no master */
1265                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1266         }
1267
1268         mlog(0, "%s message received from node %u\n",
1269                   (mres->flags & DLM_MRES_RECOVERY) ?
1270                   "recovery" : "migration", mres->master);
1271         if (mres->flags & DLM_MRES_ALL_DONE)
1272                 mlog(0, "all done flag.  all lockres data received!\n");
1273
1274         ret = -ENOMEM;
1275         buf = kmalloc(be16_to_cpu(msg->data_len), GFP_KERNEL);
1276         item = kcalloc(1, sizeof(*item), GFP_KERNEL);
1277         if (!buf || !item)
1278                 goto leave;
1279
1280         /* lookup the lock to see if we have a secondary queue for this
1281          * already...  just add the locks in and this will have its owner
1282          * and RECOVERY flag changed when it completes. */
1283         res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len);
1284         if (res) {
1285                 /* this will get a ref on res */
1286                 /* mark it as recovering/migrating and hash it */
1287                 spin_lock(&res->spinlock);
1288                 if (mres->flags & DLM_MRES_RECOVERY) {
1289                         res->state |= DLM_LOCK_RES_RECOVERING;
1290                 } else {
1291                         if (res->state & DLM_LOCK_RES_MIGRATING) {
1292                                 /* this is at least the second
1293                                  * lockres message */
1294                                 mlog(0, "lock %.*s is already migrating\n",
1295                                           mres->lockname_len,
1296                                           mres->lockname);
1297                         } else if (res->state & DLM_LOCK_RES_RECOVERING) {
1298                                 /* caller should BUG */
1299                                 mlog(ML_ERROR, "node is attempting to migrate "
1300                                      "lock %.*s, but marked as recovering!\n",
1301                                      mres->lockname_len, mres->lockname);
1302                                 ret = -EFAULT;
1303                                 spin_unlock(&res->spinlock);
1304                                 goto leave;
1305                         }
1306                         res->state |= DLM_LOCK_RES_MIGRATING;
1307                 }
1308                 spin_unlock(&res->spinlock);
1309         } else {
1310                 /* need to allocate, just like if it was
1311                  * mastered here normally  */
1312                 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1313                 if (!res)
1314                         goto leave;
1315
1316                 /* to match the ref that we would have gotten if
1317                  * dlm_lookup_lockres had succeeded */
1318                 dlm_lockres_get(res);
1319
1320                 /* mark it as recovering/migrating and hash it */
1321                 if (mres->flags & DLM_MRES_RECOVERY)
1322                         res->state |= DLM_LOCK_RES_RECOVERING;
1323                 else
1324                         res->state |= DLM_LOCK_RES_MIGRATING;
1325
1326                 spin_lock(&dlm->spinlock);
1327                 __dlm_insert_lockres(dlm, res);
1328                 spin_unlock(&dlm->spinlock);
1329
1330                 /* now that the new lockres is inserted,
1331                  * make it usable by other processes */
1332                 spin_lock(&res->spinlock);
1333                 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1334                 spin_unlock(&res->spinlock);
1335
1336                 /* add an extra ref for just-allocated lockres 
1337                  * otherwise the lockres will be purged immediately */
1338                 dlm_lockres_get(res);
1339
1340         }
1341
1342         /* at this point we have allocated everything we need,
1343          * and we have a hashed lockres with an extra ref and
1344          * the proper res->state flags. */
1345         ret = 0;
1346         if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1347                 /* migration cannot have an unknown master */
1348                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1349                 mlog(0, "recovery has passed me a lockres with an "
1350                           "unknown owner.. will need to requery: "
1351                           "%.*s\n", mres->lockname_len, mres->lockname);
1352         } else {
1353                 spin_lock(&res->spinlock);
1354                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
1355                 spin_unlock(&res->spinlock);
1356         }
1357
1358         /* queue up work for dlm_mig_lockres_worker */
1359         dlm_grab(dlm);  /* get an extra ref for the work item */
1360         memcpy(buf, msg->buf, be16_to_cpu(msg->data_len));  /* copy the whole message */
1361         dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1362         item->u.ml.lockres = res; /* already have a ref */
1363         item->u.ml.real_master = real_master;
1364         spin_lock(&dlm->work_lock);
1365         list_add_tail(&item->list, &dlm->work_list);
1366         spin_unlock(&dlm->work_lock);
1367         schedule_work(&dlm->dispatched_work);
1368
1369 leave:
1370         dlm_put(dlm);
1371         if (ret < 0) {
1372                 if (buf)
1373                         kfree(buf);
1374                 if (item)
1375                         kfree(item);
1376         }
1377
1378         mlog_exit(ret);
1379         return ret;
1380 }
1381
1382
1383 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1384 {
1385         struct dlm_ctxt *dlm = data;
1386         struct dlm_migratable_lockres *mres;
1387         int ret = 0;
1388         struct dlm_lock_resource *res;
1389         u8 real_master;
1390
1391         dlm = item->dlm;
1392         mres = (struct dlm_migratable_lockres *)data;
1393
1394         res = item->u.ml.lockres;
1395         real_master = item->u.ml.real_master;
1396
1397         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1398                 /* this case is super-rare. only occurs if
1399                  * node death happens during migration. */
1400 again:
1401                 ret = dlm_lockres_master_requery(dlm, res, &real_master);
1402                 if (ret < 0) {
1403                         mlog(0, "dlm_lockres_master_requery ret=%d\n",
1404                                   ret);
1405                         goto again;
1406                 }
1407                 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1408                         mlog(0, "lockres %.*s not claimed.  "
1409                                    "this node will take it.\n",
1410                                    res->lockname.len, res->lockname.name);
1411                 } else {
1412                         mlog(0, "master needs to respond to sender "
1413                                   "that node %u still owns %.*s\n",
1414                                   real_master, res->lockname.len,
1415                                   res->lockname.name);
1416                         /* cannot touch this lockres */
1417                         goto leave;
1418                 }
1419         }
1420
1421         ret = dlm_process_recovery_data(dlm, res, mres);
1422         if (ret < 0)
1423                 mlog(0, "dlm_process_recovery_data returned  %d\n", ret);
1424         else
1425                 mlog(0, "dlm_process_recovery_data succeeded\n");
1426
1427         if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1428                            (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1429                 ret = dlm_finish_migration(dlm, res, mres->master);
1430                 if (ret < 0)
1431                         mlog_errno(ret);
1432         }
1433
1434 leave:
1435         kfree(data);
1436         mlog_exit(ret);
1437 }
1438
1439
1440
1441 int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1442                                struct dlm_lock_resource *res, u8 *real_master)
1443 {
1444         struct dlm_node_iter iter;
1445         int nodenum;
1446         int ret = 0;
1447
1448         *real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1449
1450         /* we only reach here if one of the two nodes in a
1451          * migration died while the migration was in progress.
1452          * at this point we need to requery the master.  we
1453          * know that the new_master got as far as creating
1454          * an mle on at least one node, but we do not know
1455          * if any nodes had actually cleared the mle and set
1456          * the master to the new_master.  the old master
1457          * is supposed to set the owner to UNKNOWN in the
1458          * event of a new_master death, so the only possible
1459          * responses that we can get from nodes here are
1460          * that the master is new_master, or that the master
1461          * is UNKNOWN.
1462          * if all nodes come back with UNKNOWN then we know
1463          * the lock needs remastering here.
1464          * if any node comes back with a valid master, check
1465          * to see if that master is the one that we are
1466          * recovering.  if so, then the new_master died and
1467          * we need to remaster this lock.  if not, then the
1468          * new_master survived and that node will respond to
1469          * other nodes about the owner.
1470          * if there is an owner, this node needs to dump this
1471          * lockres and alert the sender that this lockres
1472          * was rejected. */
1473         spin_lock(&dlm->spinlock);
1474         dlm_node_iter_init(dlm->domain_map, &iter);
1475         spin_unlock(&dlm->spinlock);
1476
1477         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1478                 /* do not send to self */
1479                 if (nodenum == dlm->node_num)
1480                         continue;
1481                 ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1482                 if (ret < 0) {
1483                         mlog_errno(ret);
1484                         if (!dlm_is_host_down(ret))
1485                                 BUG();
1486                         /* host is down, so answer for that node would be
1487                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
1488                 }
1489                 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1490                         mlog(0, "lock master is %u\n", *real_master);
1491                         break;
1492                 }
1493         }
1494         return ret;
1495 }
1496
1497
1498 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1499                           u8 nodenum, u8 *real_master)
1500 {
1501         int ret = -EINVAL;
1502         struct dlm_master_requery req;
1503         int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1504
1505         memset(&req, 0, sizeof(req));
1506         req.node_idx = dlm->node_num;
1507         req.namelen = res->lockname.len;
1508         memcpy(req.name, res->lockname.name, res->lockname.len);
1509
1510         ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1511                                  &req, sizeof(req), nodenum, &status);
1512         /* XXX: negative status not handled properly here. */
1513         if (ret < 0)
1514                 mlog_errno(ret);
1515         else {
1516                 BUG_ON(status < 0);
1517                 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1518                 *real_master = (u8) (status & 0xff);
1519                 mlog(0, "node %u responded to master requery with %u\n",
1520                           nodenum, *real_master);
1521                 ret = 0;
1522         }
1523         return ret;
1524 }
1525
1526
1527 /* this function cannot error, so unless the sending
1528  * or receiving of the message failed, the owner can
1529  * be trusted */
1530 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data)
1531 {
1532         struct dlm_ctxt *dlm = data;
1533         struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1534         struct dlm_lock_resource *res = NULL;
1535         unsigned int hash;
1536         int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1537         u32 flags = DLM_ASSERT_MASTER_REQUERY;
1538
1539         if (!dlm_grab(dlm)) {
1540                 /* since the domain has gone away on this
1541                  * node, the proper response is UNKNOWN */
1542                 return master;
1543         }
1544
1545         hash = dlm_lockid_hash(req->name, req->namelen);
1546
1547         spin_lock(&dlm->spinlock);
1548         res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash);
1549         if (res) {
1550                 spin_lock(&res->spinlock);
1551                 master = res->owner;
1552                 if (master == dlm->node_num) {
1553                         int ret = dlm_dispatch_assert_master(dlm, res,
1554                                                              0, 0, flags);
1555                         if (ret < 0) {
1556                                 mlog_errno(-ENOMEM);
1557                                 /* retry!? */
1558                                 BUG();
1559                         }
1560                 }
1561                 spin_unlock(&res->spinlock);
1562         }
1563         spin_unlock(&dlm->spinlock);
1564
1565         dlm_put(dlm);
1566         return master;
1567 }
1568
1569 static inline struct list_head *
1570 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1571 {
1572         struct list_head *ret;
1573         BUG_ON(list_num < 0);
1574         BUG_ON(list_num > 2);
1575         ret = &(res->granted);
1576         ret += list_num;
1577         return ret;
1578 }
1579 /* TODO: do ast flush business
1580  * TODO: do MIGRATING and RECOVERING spinning
1581  */
1582
1583 /*
1584 * NOTE about in-flight requests during migration:
1585 *
1586 * Before attempting the migrate, the master has marked the lockres as
1587 * MIGRATING and then flushed all of its pending ASTS.  So any in-flight
1588 * requests either got queued before the MIGRATING flag got set, in which
1589 * case the lock data will reflect the change and a return message is on
1590 * the way, or the request failed to get in before MIGRATING got set.  In
1591 * this case, the caller will be told to spin and wait for the MIGRATING
1592 * flag to be dropped, then recheck the master.
1593 * This holds true for the convert, cancel and unlock cases, and since lvb
1594 * updates are tied to these same messages, it applies to lvb updates as
1595 * well.  For the lock case, there is no way a lock can be on the master
1596 * queue and not be on the secondary queue since the lock is always added
1597 * locally first.  This means that the new target node will never be sent
1598 * a lock that he doesn't already have on the list.
1599 * In total, this means that the local lock is correct and should not be
1600 * updated to match the one sent by the master.  Any messages sent back
1601 * from the master before the MIGRATING flag will bring the lock properly
1602 * up-to-date, and the change will be ordered properly for the waiter.
1603 * We will *not* attempt to modify the lock underneath the waiter.
1604 */
1605
1606 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1607                                      struct dlm_lock_resource *res,
1608                                      struct dlm_migratable_lockres *mres)
1609 {
1610         struct dlm_migratable_lock *ml;
1611         struct list_head *queue;
1612         struct dlm_lock *newlock = NULL;
1613         struct dlm_lockstatus *lksb = NULL;
1614         int ret = 0;
1615         int i, bad;
1616         struct list_head *iter;
1617         struct dlm_lock *lock = NULL;
1618
1619         mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1620         for (i=0; i<mres->num_locks; i++) {
1621                 ml = &(mres->ml[i]);
1622                 BUG_ON(ml->highest_blocked != LKM_IVMODE);
1623                 newlock = NULL;
1624                 lksb = NULL;
1625
1626                 queue = dlm_list_num_to_pointer(res, ml->list);
1627
1628                 /* if the lock is for the local node it needs to
1629                  * be moved to the proper location within the queue.
1630                  * do not allocate a new lock structure. */
1631                 if (ml->node == dlm->node_num) {
1632                         /* MIGRATION ONLY! */
1633                         BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1634
1635                         spin_lock(&res->spinlock);
1636                         list_for_each(iter, queue) {
1637                                 lock = list_entry (iter, struct dlm_lock, list);
1638                                 if (lock->ml.cookie != ml->cookie)
1639                                         lock = NULL;
1640                                 else
1641                                         break;
1642                         }
1643
1644                         /* lock is always created locally first, and
1645                          * destroyed locally last.  it must be on the list */
1646                         if (!lock) {
1647                                 u64 c = ml->cookie;
1648                                 mlog(ML_ERROR, "could not find local lock "
1649                                                "with cookie %u:%llu!\n",
1650                                                dlm_get_lock_cookie_node(c),
1651                                                dlm_get_lock_cookie_seq(c));
1652                                 BUG();
1653                         }
1654                         BUG_ON(lock->ml.node != ml->node);
1655
1656                         /* see NOTE above about why we do not update
1657                          * to match the master here */
1658
1659                         /* move the lock to its proper place */
1660                         /* do not alter lock refcount.  switching lists. */
1661                         list_move_tail(&lock->list, queue);
1662                         spin_unlock(&res->spinlock);
1663
1664                         mlog(0, "just reordered a local lock!\n");
1665                         continue;
1666                 }
1667
1668                 /* lock is for another node. */
1669                 newlock = dlm_new_lock(ml->type, ml->node,
1670                                        be64_to_cpu(ml->cookie), NULL);
1671                 if (!newlock) {
1672                         ret = -ENOMEM;
1673                         goto leave;
1674                 }
1675                 lksb = newlock->lksb;
1676                 dlm_lock_attach_lockres(newlock, res);
1677
1678                 if (ml->convert_type != LKM_IVMODE) {
1679                         BUG_ON(queue != &res->converting);
1680                         newlock->ml.convert_type = ml->convert_type;
1681                 }
1682                 lksb->flags |= (ml->flags &
1683                                 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1684
1685                 if (ml->type == LKM_NLMODE)
1686                         goto skip_lvb;
1687
1688                 if (!dlm_lvb_is_empty(mres->lvb)) {
1689                         if (lksb->flags & DLM_LKSB_PUT_LVB) {
1690                                 /* other node was trying to update
1691                                  * lvb when node died.  recreate the
1692                                  * lksb with the updated lvb. */
1693                                 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1694                                 /* the lock resource lvb update must happen
1695                                  * NOW, before the spinlock is dropped.
1696                                  * we no longer wait for the AST to update
1697                                  * the lvb. */
1698                                 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1699                         } else {
1700                                 /* otherwise, the node is sending its 
1701                                  * most recent valid lvb info */
1702                                 BUG_ON(ml->type != LKM_EXMODE &&
1703                                        ml->type != LKM_PRMODE);
1704                                 if (!dlm_lvb_is_empty(res->lvb) &&
1705                                     (ml->type == LKM_EXMODE ||
1706                                      memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1707                                         int i;
1708                                         mlog(ML_ERROR, "%s:%.*s: received bad "
1709                                              "lvb! type=%d\n", dlm->name,
1710                                              res->lockname.len,
1711                                              res->lockname.name, ml->type);
1712                                         printk("lockres lvb=[");
1713                                         for (i=0; i<DLM_LVB_LEN; i++)
1714                                                 printk("%02x", res->lvb[i]);
1715                                         printk("]\nmigrated lvb=[");
1716                                         for (i=0; i<DLM_LVB_LEN; i++)
1717                                                 printk("%02x", mres->lvb[i]);
1718                                         printk("]\n");
1719                                         dlm_print_one_lock_resource(res);
1720                                         BUG();
1721                                 }
1722                                 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1723                         }
1724                 }
1725 skip_lvb:
1726
1727                 /* NOTE:
1728                  * wrt lock queue ordering and recovery:
1729                  *    1. order of locks on granted queue is
1730                  *       meaningless.
1731                  *    2. order of locks on converting queue is
1732                  *       LOST with the node death.  sorry charlie.
1733                  *    3. order of locks on the blocked queue is
1734                  *       also LOST.
1735                  * order of locks does not affect integrity, it
1736                  * just means that a lock request may get pushed
1737                  * back in line as a result of the node death.
1738                  * also note that for a given node the lock order
1739                  * for its secondary queue locks is preserved
1740                  * relative to each other, but clearly *not*
1741                  * preserved relative to locks from other nodes.
1742                  */
1743                 bad = 0;
1744                 spin_lock(&res->spinlock);
1745                 list_for_each_entry(lock, queue, list) {
1746                         if (lock->ml.cookie == ml->cookie) {
1747                                 u64 c = lock->ml.cookie;
1748                                 mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already "
1749                                      "exists on this lockres!\n", dlm->name,
1750                                      res->lockname.len, res->lockname.name,
1751                                      dlm_get_lock_cookie_node(c),
1752                                      dlm_get_lock_cookie_seq(c));
1753
1754                                 mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, "
1755                                      "node=%u, cookie=%u:%llu, queue=%d\n",
1756                                      ml->type, ml->convert_type, ml->node,
1757                                      dlm_get_lock_cookie_node(ml->cookie),
1758                                      dlm_get_lock_cookie_seq(ml->cookie),
1759                                      ml->list);
1760
1761                                 __dlm_print_one_lock_resource(res);
1762                                 bad = 1;
1763                                 break;
1764                         }
1765                 }
1766                 if (!bad) {
1767                         dlm_lock_get(newlock);
1768                         list_add_tail(&newlock->list, queue);
1769                 }
1770                 spin_unlock(&res->spinlock);
1771         }
1772         mlog(0, "done running all the locks\n");
1773
1774 leave:
1775         if (ret < 0) {
1776                 mlog_errno(ret);
1777                 if (newlock)
1778                         dlm_lock_put(newlock);
1779         }
1780
1781         mlog_exit(ret);
1782         return ret;
1783 }
1784
1785 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
1786                                        struct dlm_lock_resource *res)
1787 {
1788         int i;
1789         struct list_head *queue, *iter, *iter2;
1790         struct dlm_lock *lock;
1791
1792         res->state |= DLM_LOCK_RES_RECOVERING;
1793         if (!list_empty(&res->recovering)) {
1794                 mlog(0,
1795                      "Recovering res %s:%.*s, is already on recovery list!\n",
1796                      dlm->name, res->lockname.len, res->lockname.name);
1797                 list_del_init(&res->recovering);
1798         }
1799         /* We need to hold a reference while on the recovery list */
1800         dlm_lockres_get(res);
1801         list_add_tail(&res->recovering, &dlm->reco.resources);
1802
1803         /* find any pending locks and put them back on proper list */
1804         for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
1805                 queue = dlm_list_idx_to_ptr(res, i);
1806                 list_for_each_safe(iter, iter2, queue) {
1807                         lock = list_entry (iter, struct dlm_lock, list);
1808                         dlm_lock_get(lock);
1809                         if (lock->convert_pending) {
1810                                 /* move converting lock back to granted */
1811                                 BUG_ON(i != DLM_CONVERTING_LIST);
1812                                 mlog(0, "node died with convert pending "
1813                                      "on %.*s. move back to granted list.\n",
1814                                      res->lockname.len, res->lockname.name);
1815                                 dlm_revert_pending_convert(res, lock);
1816                                 lock->convert_pending = 0;
1817                         } else if (lock->lock_pending) {
1818                                 /* remove pending lock requests completely */
1819                                 BUG_ON(i != DLM_BLOCKED_LIST);
1820                                 mlog(0, "node died with lock pending "
1821                                      "on %.*s. remove from blocked list and skip.\n",
1822                                      res->lockname.len, res->lockname.name);
1823                                 /* lock will be floating until ref in
1824                                  * dlmlock_remote is freed after the network
1825                                  * call returns.  ok for it to not be on any
1826                                  * list since no ast can be called
1827                                  * (the master is dead). */
1828                                 dlm_revert_pending_lock(res, lock);
1829                                 lock->lock_pending = 0;
1830                         } else if (lock->unlock_pending) {
1831                                 /* if an unlock was in progress, treat as
1832                                  * if this had completed successfully
1833                                  * before sending this lock state to the
1834                                  * new master.  note that the dlm_unlock
1835                                  * call is still responsible for calling
1836                                  * the unlockast.  that will happen after
1837                                  * the network call times out.  for now,
1838                                  * just move lists to prepare the new
1839                                  * recovery master.  */
1840                                 BUG_ON(i != DLM_GRANTED_LIST);
1841                                 mlog(0, "node died with unlock pending "
1842                                      "on %.*s. remove from blocked list and skip.\n",
1843                                      res->lockname.len, res->lockname.name);
1844                                 dlm_commit_pending_unlock(res, lock);
1845                                 lock->unlock_pending = 0;
1846                         } else if (lock->cancel_pending) {
1847                                 /* if a cancel was in progress, treat as
1848                                  * if this had completed successfully
1849                                  * before sending this lock state to the
1850                                  * new master */
1851                                 BUG_ON(i != DLM_CONVERTING_LIST);
1852                                 mlog(0, "node died with cancel pending "
1853                                      "on %.*s. move back to granted list.\n",
1854                                      res->lockname.len, res->lockname.name);
1855                                 dlm_commit_pending_cancel(res, lock);
1856                                 lock->cancel_pending = 0;
1857                         }
1858                         dlm_lock_put(lock);
1859                 }
1860         }
1861 }
1862
1863
1864
1865 /* removes all recovered locks from the recovery list.
1866  * sets the res->owner to the new master.
1867  * unsets the RECOVERY flag and wakes waiters. */
1868 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
1869                                               u8 dead_node, u8 new_master)
1870 {
1871         int i;
1872         struct list_head *iter, *iter2;
1873         struct hlist_node *hash_iter;
1874         struct hlist_head *bucket;
1875
1876         struct dlm_lock_resource *res;
1877
1878         mlog_entry_void();
1879
1880         assert_spin_locked(&dlm->spinlock);
1881
1882         list_for_each_safe(iter, iter2, &dlm->reco.resources) {
1883                 res = list_entry (iter, struct dlm_lock_resource, recovering);
1884                 if (res->owner == dead_node) {
1885                         list_del_init(&res->recovering);
1886                         spin_lock(&res->spinlock);
1887                         dlm_change_lockres_owner(dlm, res, new_master);
1888                         res->state &= ~DLM_LOCK_RES_RECOVERING;
1889                         if (!__dlm_lockres_unused(res))
1890                                 __dlm_dirty_lockres(dlm, res);
1891                         spin_unlock(&res->spinlock);
1892                         wake_up(&res->wq);
1893                         dlm_lockres_put(res);
1894                 }
1895         }
1896
1897         /* this will become unnecessary eventually, but
1898          * for now we need to run the whole hash, clear
1899          * the RECOVERING state and set the owner
1900          * if necessary */
1901         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
1902                 bucket = dlm_lockres_hash(dlm, i);
1903                 hlist_for_each_entry(res, hash_iter, bucket, hash_node) {
1904                         if (res->state & DLM_LOCK_RES_RECOVERING) {
1905                                 if (res->owner == dead_node) {
1906                                         mlog(0, "(this=%u) res %.*s owner=%u "
1907                                              "was not on recovering list, but "
1908                                              "clearing state anyway\n",
1909                                              dlm->node_num, res->lockname.len,
1910                                              res->lockname.name, new_master);
1911                                 } else if (res->owner == dlm->node_num) {
1912                                         mlog(0, "(this=%u) res %.*s owner=%u "
1913                                              "was not on recovering list, "
1914                                              "owner is THIS node, clearing\n",
1915                                              dlm->node_num, res->lockname.len,
1916                                              res->lockname.name, new_master);
1917                                 } else
1918                                         continue;
1919
1920                                 if (!list_empty(&res->recovering)) {
1921                                         mlog(0, "%s:%.*s: lockres was "
1922                                              "marked RECOVERING, owner=%u\n",
1923                                              dlm->name, res->lockname.len,
1924                                              res->lockname.name, res->owner);
1925                                         list_del_init(&res->recovering);
1926                                         dlm_lockres_put(res);
1927                                 }
1928                                 spin_lock(&res->spinlock);
1929                                 dlm_change_lockres_owner(dlm, res, new_master);
1930                                 res->state &= ~DLM_LOCK_RES_RECOVERING;
1931                                 if (!__dlm_lockres_unused(res))
1932                                         __dlm_dirty_lockres(dlm, res);
1933                                 spin_unlock(&res->spinlock);
1934                                 wake_up(&res->wq);
1935                         }
1936                 }
1937         }
1938 }
1939
1940 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
1941 {
1942         if (local) {
1943                 if (lock->ml.type != LKM_EXMODE &&
1944                     lock->ml.type != LKM_PRMODE)
1945                         return 1;
1946         } else if (lock->ml.type == LKM_EXMODE)
1947                 return 1;
1948         return 0;
1949 }
1950
1951 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
1952                                struct dlm_lock_resource *res, u8 dead_node)
1953 {
1954         struct list_head *iter, *queue;
1955         struct dlm_lock *lock;
1956         int blank_lvb = 0, local = 0;
1957         int i;
1958         u8 search_node;
1959
1960         assert_spin_locked(&dlm->spinlock);
1961         assert_spin_locked(&res->spinlock);
1962
1963         if (res->owner == dlm->node_num)
1964                 /* if this node owned the lockres, and if the dead node 
1965                  * had an EX when he died, blank out the lvb */
1966                 search_node = dead_node;
1967         else {
1968                 /* if this is a secondary lockres, and we had no EX or PR
1969                  * locks granted, we can no longer trust the lvb */
1970                 search_node = dlm->node_num;
1971                 local = 1;  /* check local state for valid lvb */
1972         }
1973
1974         for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
1975                 queue = dlm_list_idx_to_ptr(res, i);
1976                 list_for_each(iter, queue) {
1977                         lock = list_entry (iter, struct dlm_lock, list);
1978                         if (lock->ml.node == search_node) {
1979                                 if (dlm_lvb_needs_invalidation(lock, local)) {
1980                                         /* zero the lksb lvb and lockres lvb */
1981                                         blank_lvb = 1;
1982                                         memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
1983                                 }
1984                         }
1985                 }
1986         }
1987
1988         if (blank_lvb) {
1989                 mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
1990                      res->lockname.len, res->lockname.name, dead_node);
1991                 memset(res->lvb, 0, DLM_LVB_LEN);
1992         }
1993 }
1994
1995 static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
1996                                 struct dlm_lock_resource *res, u8 dead_node)
1997 {
1998         struct list_head *iter, *tmpiter;
1999         struct dlm_lock *lock;
2000
2001         /* this node is the lockres master:
2002          * 1) remove any stale locks for the dead node
2003          * 2) if the dead node had an EX when he died, blank out the lvb 
2004          */
2005         assert_spin_locked(&dlm->spinlock);
2006         assert_spin_locked(&res->spinlock);
2007
2008         /* TODO: check pending_asts, pending_basts here */
2009         list_for_each_safe(iter, tmpiter, &res->granted) {
2010                 lock = list_entry (iter, struct dlm_lock, list);
2011                 if (lock->ml.node == dead_node) {
2012                         list_del_init(&lock->list);
2013                         dlm_lock_put(lock);
2014                 }
2015         }
2016         list_for_each_safe(iter, tmpiter, &res->converting) {
2017                 lock = list_entry (iter, struct dlm_lock, list);
2018                 if (lock->ml.node == dead_node) {
2019                         list_del_init(&lock->list);
2020                         dlm_lock_put(lock);
2021                 }
2022         }
2023         list_for_each_safe(iter, tmpiter, &res->blocked) {
2024                 lock = list_entry (iter, struct dlm_lock, list);
2025                 if (lock->ml.node == dead_node) {
2026                         list_del_init(&lock->list);
2027                         dlm_lock_put(lock);
2028                 }
2029         }
2030
2031         /* do not kick thread yet */
2032         __dlm_dirty_lockres(dlm, res);
2033 }
2034
2035 /* if this node is the recovery master, and there are no
2036  * locks for a given lockres owned by this node that are in
2037  * either PR or EX mode, zero out the lvb before requesting.
2038  *
2039  */
2040
2041
2042 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
2043 {
2044         struct hlist_node *iter;
2045         struct dlm_lock_resource *res;
2046         int i;
2047         struct hlist_head *bucket;
2048         struct dlm_lock *lock;
2049
2050
2051         /* purge any stale mles */
2052         dlm_clean_master_list(dlm, dead_node);
2053
2054         /*
2055          * now clean up all lock resources.  there are two rules:
2056          *
2057          * 1) if the dead node was the master, move the lockres
2058          *    to the recovering list.  set the RECOVERING flag.
2059          *    this lockres needs to be cleaned up before it can
2060          *    be used further.
2061          *
2062          * 2) if this node was the master, remove all locks from
2063          *    each of the lockres queues that were owned by the
2064          *    dead node.  once recovery finishes, the dlm thread
2065          *    can be kicked again to see if any ASTs or BASTs
2066          *    need to be fired as a result.
2067          */
2068         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2069                 bucket = dlm_lockres_hash(dlm, i);
2070                 hlist_for_each_entry(res, iter, bucket, hash_node) {
2071                         /* always prune any $RECOVERY entries for dead nodes,
2072                          * otherwise hangs can occur during later recovery */
2073                         if (dlm_is_recovery_lock(res->lockname.name,
2074                                                  res->lockname.len)) {
2075                                 spin_lock(&res->spinlock);
2076                                 list_for_each_entry(lock, &res->granted, list) {
2077                                         if (lock->ml.node == dead_node) {
2078                                                 mlog(0, "AHA! there was "
2079                                                      "a $RECOVERY lock for dead "
2080                                                      "node %u (%s)!\n",
2081                                                      dead_node, dlm->name);
2082                                                 list_del_init(&lock->list);
2083                                                 dlm_lock_put(lock);
2084                                                 break;
2085                                         }
2086                                 }
2087                                 spin_unlock(&res->spinlock);
2088                                 continue;
2089                         }                       
2090                         spin_lock(&res->spinlock);
2091                         /* zero the lvb if necessary */
2092                         dlm_revalidate_lvb(dlm, res, dead_node);
2093                         if (res->owner == dead_node)
2094                                 dlm_move_lockres_to_recovery_list(dlm, res);
2095                         else if (res->owner == dlm->node_num) {
2096                                 dlm_free_dead_locks(dlm, res, dead_node);
2097                                 __dlm_lockres_calc_usage(dlm, res);
2098                         }
2099                         spin_unlock(&res->spinlock);
2100                 }
2101         }
2102
2103 }
2104
2105 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
2106 {
2107         assert_spin_locked(&dlm->spinlock);
2108
2109         if (dlm->reco.new_master == idx) {
2110                 mlog(0, "%s: recovery master %d just died\n",
2111                      dlm->name, idx);
2112                 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2113                         /* finalize1 was reached, so it is safe to clear
2114                          * the new_master and dead_node.  that recovery
2115                          * is complete. */
2116                         mlog(0, "%s: dead master %d had reached "
2117                              "finalize1 state, clearing\n", dlm->name, idx);
2118                         dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2119                         __dlm_reset_recovery(dlm);
2120                 }
2121         }
2122
2123         /* check to see if the node is already considered dead */
2124         if (!test_bit(idx, dlm->live_nodes_map)) {
2125                 mlog(0, "for domain %s, node %d is already dead. "
2126                      "another node likely did recovery already.\n",
2127                      dlm->name, idx);
2128                 return;
2129         }
2130
2131         /* check to see if we do not care about this node */
2132         if (!test_bit(idx, dlm->domain_map)) {
2133                 /* This also catches the case that we get a node down
2134                  * but haven't joined the domain yet. */
2135                 mlog(0, "node %u already removed from domain!\n", idx);
2136                 return;
2137         }
2138
2139         clear_bit(idx, dlm->live_nodes_map);
2140
2141         /* Clean up join state on node death. */
2142         if (dlm->joining_node == idx) {
2143                 mlog(0, "Clearing join state for node %u\n", idx);
2144                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
2145         }
2146
2147         /* make sure local cleanup occurs before the heartbeat events */
2148         if (!test_bit(idx, dlm->recovery_map))
2149                 dlm_do_local_recovery_cleanup(dlm, idx);
2150
2151         /* notify anything attached to the heartbeat events */
2152         dlm_hb_event_notify_attached(dlm, idx, 0);
2153
2154         mlog(0, "node %u being removed from domain map!\n", idx);
2155         clear_bit(idx, dlm->domain_map);
2156         /* wake up migration waiters if a node goes down.
2157          * perhaps later we can genericize this for other waiters. */
2158         wake_up(&dlm->migration_wq);
2159
2160         if (test_bit(idx, dlm->recovery_map))
2161                 mlog(0, "domain %s, node %u already added "
2162                      "to recovery map!\n", dlm->name, idx);
2163         else
2164                 set_bit(idx, dlm->recovery_map);
2165 }
2166
2167 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
2168 {
2169         struct dlm_ctxt *dlm = data;
2170
2171         if (!dlm_grab(dlm))
2172                 return;
2173
2174         spin_lock(&dlm->spinlock);
2175         __dlm_hb_node_down(dlm, idx);
2176         spin_unlock(&dlm->spinlock);
2177
2178         dlm_put(dlm);
2179 }
2180
2181 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
2182 {
2183         struct dlm_ctxt *dlm = data;
2184
2185         if (!dlm_grab(dlm))
2186                 return;
2187
2188         spin_lock(&dlm->spinlock);
2189         set_bit(idx, dlm->live_nodes_map);
2190         /* do NOT notify mle attached to the heartbeat events.
2191          * new nodes are not interesting in mastery until joined. */
2192         spin_unlock(&dlm->spinlock);
2193
2194         dlm_put(dlm);
2195 }
2196
2197 static void dlm_reco_ast(void *astdata)
2198 {
2199         struct dlm_ctxt *dlm = astdata;
2200         mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2201              dlm->node_num, dlm->name);
2202 }
2203 static void dlm_reco_bast(void *astdata, int blocked_type)
2204 {
2205         struct dlm_ctxt *dlm = astdata;
2206         mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2207              dlm->node_num, dlm->name);
2208 }
2209 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
2210 {
2211         mlog(0, "unlockast for recovery lock fired!\n");
2212 }
2213
2214 /*
2215  * dlm_pick_recovery_master will continually attempt to use
2216  * dlmlock() on the special "$RECOVERY" lockres with the
2217  * LKM_NOQUEUE flag to get an EX.  every thread that enters
2218  * this function on each node racing to become the recovery
2219  * master will not stop attempting this until either:
2220  * a) this node gets the EX (and becomes the recovery master),
2221  * or b) dlm->reco.new_master gets set to some nodenum 
2222  * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2223  * so each time a recovery master is needed, the entire cluster
2224  * will sync at this point.  if the new master dies, that will
2225  * be detected in dlm_do_recovery */
2226 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2227 {
2228         enum dlm_status ret;
2229         struct dlm_lockstatus lksb;
2230         int status = -EINVAL;
2231
2232         mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2233              dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2234 again:  
2235         memset(&lksb, 0, sizeof(lksb));
2236
2237         ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
2238                       DLM_RECOVERY_LOCK_NAME, dlm_reco_ast, dlm, dlm_reco_bast);
2239
2240         mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2241              dlm->name, ret, lksb.status);
2242
2243         if (ret == DLM_NORMAL) {
2244                 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2245                      dlm->name, dlm->node_num);
2246                 
2247                 /* got the EX lock.  check to see if another node 
2248                  * just became the reco master */
2249                 if (dlm_reco_master_ready(dlm)) {
2250                         mlog(0, "%s: got reco EX lock, but %u will "
2251                              "do the recovery\n", dlm->name,
2252                              dlm->reco.new_master);
2253                         status = -EEXIST;
2254                 } else {
2255                         status = 0;
2256
2257                         /* see if recovery was already finished elsewhere */
2258                         spin_lock(&dlm->spinlock);
2259                         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
2260                                 status = -EINVAL;       
2261                                 mlog(0, "%s: got reco EX lock, but "
2262                                      "node got recovered already\n", dlm->name);
2263                                 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2264                                         mlog(ML_ERROR, "%s: new master is %u "
2265                                              "but no dead node!\n", 
2266                                              dlm->name, dlm->reco.new_master);
2267                                         BUG();
2268                                 }
2269                         }
2270                         spin_unlock(&dlm->spinlock);
2271                 }
2272
2273                 /* if this node has actually become the recovery master,
2274                  * set the master and send the messages to begin recovery */
2275                 if (!status) {
2276                         mlog(0, "%s: dead=%u, this=%u, sending "
2277                              "begin_reco now\n", dlm->name, 
2278                              dlm->reco.dead_node, dlm->node_num);
2279                         status = dlm_send_begin_reco_message(dlm,
2280                                       dlm->reco.dead_node);
2281                         /* this always succeeds */
2282                         BUG_ON(status);
2283
2284                         /* set the new_master to this node */
2285                         spin_lock(&dlm->spinlock);
2286                         dlm_set_reco_master(dlm, dlm->node_num);
2287                         spin_unlock(&dlm->spinlock);
2288                 }
2289
2290                 /* recovery lock is a special case.  ast will not get fired,
2291                  * so just go ahead and unlock it. */
2292                 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
2293                 if (ret == DLM_DENIED) {
2294                         mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2295                         ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2296                 }
2297                 if (ret != DLM_NORMAL) {
2298                         /* this would really suck. this could only happen
2299                          * if there was a network error during the unlock
2300                          * because of node death.  this means the unlock
2301                          * is actually "done" and the lock structure is
2302                          * even freed.  we can continue, but only
2303                          * because this specific lock name is special. */
2304                         mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
2305                 }
2306         } else if (ret == DLM_NOTQUEUED) {
2307                 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2308                      dlm->name, dlm->node_num);
2309                 /* another node is master. wait on
2310                  * reco.new_master != O2NM_INVALID_NODE_NUM 
2311                  * for at most one second */
2312                 wait_event_timeout(dlm->dlm_reco_thread_wq,
2313                                          dlm_reco_master_ready(dlm),
2314                                          msecs_to_jiffies(1000));
2315                 if (!dlm_reco_master_ready(dlm)) {
2316                         mlog(0, "%s: reco master taking awhile\n",
2317                              dlm->name);
2318                         goto again;
2319                 }
2320                 /* another node has informed this one that it is reco master */
2321                 mlog(0, "%s: reco master %u is ready to recover %u\n",
2322                      dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
2323                 status = -EEXIST;
2324         } else if (ret == DLM_RECOVERING) {
2325                 mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n",
2326                      dlm->name, dlm->node_num);
2327                 goto again;
2328         } else {
2329                 struct dlm_lock_resource *res;
2330
2331                 /* dlmlock returned something other than NOTQUEUED or NORMAL */
2332                 mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2333                      "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2334                      dlm_errname(lksb.status));
2335                 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2336                                          DLM_RECOVERY_LOCK_NAME_LEN);
2337                 if (res) {
2338                         dlm_print_one_lock_resource(res);
2339                         dlm_lockres_put(res);
2340                 } else {
2341                         mlog(ML_ERROR, "recovery lock not found\n");
2342                 }
2343                 BUG();
2344         }
2345
2346         return status;
2347 }
2348
2349 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2350 {
2351         struct dlm_begin_reco br;
2352         int ret = 0;
2353         struct dlm_node_iter iter;
2354         int nodenum;
2355         int status;
2356
2357         mlog_entry("%u\n", dead_node);
2358
2359         mlog(0, "%s: dead node is %u\n", dlm->name, dead_node);
2360
2361         spin_lock(&dlm->spinlock);
2362         dlm_node_iter_init(dlm->domain_map, &iter);
2363         spin_unlock(&dlm->spinlock);
2364
2365         clear_bit(dead_node, iter.node_map);
2366
2367         memset(&br, 0, sizeof(br));
2368         br.node_idx = dlm->node_num;
2369         br.dead_node = dead_node;
2370
2371         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2372                 ret = 0;
2373                 if (nodenum == dead_node) {
2374                         mlog(0, "not sending begin reco to dead node "
2375                                   "%u\n", dead_node);
2376                         continue;
2377                 }
2378                 if (nodenum == dlm->node_num) {
2379                         mlog(0, "not sending begin reco to self\n");
2380                         continue;
2381                 }
2382 retry:
2383                 ret = -EINVAL;
2384                 mlog(0, "attempting to send begin reco msg to %d\n",
2385                           nodenum);
2386                 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2387                                          &br, sizeof(br), nodenum, &status);
2388                 /* negative status is handled ok by caller here */
2389                 if (ret >= 0)
2390                         ret = status;
2391                 if (dlm_is_host_down(ret)) {
2392                         /* node is down.  not involved in recovery
2393                          * so just keep going */
2394                         mlog(0, "%s: node %u was down when sending "
2395                              "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2396                         ret = 0;
2397                 }
2398                 if (ret < 0) {
2399                         struct dlm_lock_resource *res;
2400                         /* this is now a serious problem, possibly ENOMEM 
2401                          * in the network stack.  must retry */
2402                         mlog_errno(ret);
2403                         mlog(ML_ERROR, "begin reco of dlm %s to node %u "
2404                             " returned %d\n", dlm->name, nodenum, ret);
2405                         res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2406                                                  DLM_RECOVERY_LOCK_NAME_LEN);
2407                         if (res) {
2408                                 dlm_print_one_lock_resource(res);
2409                                 dlm_lockres_put(res);
2410                         } else {
2411                                 mlog(ML_ERROR, "recovery lock not found\n");
2412                         }
2413                         /* sleep for a bit in hopes that we can avoid 
2414                          * another ENOMEM */
2415                         msleep(100);
2416                         goto retry;
2417                 } else if (ret == EAGAIN) {
2418                         mlog(0, "%s: trying to start recovery of node "
2419                              "%u, but node %u is waiting for last recovery "
2420                              "to complete, backoff for a bit\n", dlm->name,
2421                              dead_node, nodenum);
2422                         /* TODO Look into replacing msleep with cond_resched() */
2423                         msleep(100);
2424                         goto retry;
2425                 }
2426         }
2427
2428         return ret;
2429 }
2430
2431 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data)
2432 {
2433         struct dlm_ctxt *dlm = data;
2434         struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2435
2436         /* ok to return 0, domain has gone away */
2437         if (!dlm_grab(dlm))
2438                 return 0;
2439
2440         spin_lock(&dlm->spinlock);
2441         if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2442                 mlog(0, "%s: node %u wants to recover node %u (%u:%u) "
2443                      "but this node is in finalize state, waiting on finalize2\n",
2444                      dlm->name, br->node_idx, br->dead_node,
2445                      dlm->reco.dead_node, dlm->reco.new_master);
2446                 spin_unlock(&dlm->spinlock);
2447                 return EAGAIN;
2448         }
2449         spin_unlock(&dlm->spinlock);
2450
2451         mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n",
2452              dlm->name, br->node_idx, br->dead_node,
2453              dlm->reco.dead_node, dlm->reco.new_master);
2454
2455         dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2456
2457         spin_lock(&dlm->spinlock);
2458         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2459                 if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2460                         mlog(0, "%s: new_master %u died, changing "
2461                              "to %u\n", dlm->name, dlm->reco.new_master,
2462                              br->node_idx);
2463                 } else {
2464                         mlog(0, "%s: new_master %u NOT DEAD, changing "
2465                              "to %u\n", dlm->name, dlm->reco.new_master,
2466                              br->node_idx);
2467                         /* may not have seen the new master as dead yet */
2468                 }
2469         }
2470         if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2471                 mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2472                      "node %u changing it to %u\n", dlm->name, 
2473                      dlm->reco.dead_node, br->node_idx, br->dead_node);
2474         }
2475         dlm_set_reco_master(dlm, br->node_idx);
2476         dlm_set_reco_dead_node(dlm, br->dead_node);
2477         if (!test_bit(br->dead_node, dlm->recovery_map)) {
2478                 mlog(0, "recovery master %u sees %u as dead, but this "
2479                      "node has not yet.  marking %u as dead\n",
2480                      br->node_idx, br->dead_node, br->dead_node);
2481                 if (!test_bit(br->dead_node, dlm->domain_map) ||
2482                     !test_bit(br->dead_node, dlm->live_nodes_map))
2483                         mlog(0, "%u not in domain/live_nodes map "
2484                              "so setting it in reco map manually\n",
2485                              br->dead_node);
2486                 /* force the recovery cleanup in __dlm_hb_node_down
2487                  * both of these will be cleared in a moment */
2488                 set_bit(br->dead_node, dlm->domain_map);
2489                 set_bit(br->dead_node, dlm->live_nodes_map);
2490                 __dlm_hb_node_down(dlm, br->dead_node);
2491         }
2492         spin_unlock(&dlm->spinlock);
2493
2494         dlm_kick_recovery_thread(dlm);
2495
2496         mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n",
2497              dlm->name, br->node_idx, br->dead_node,
2498              dlm->reco.dead_node, dlm->reco.new_master);
2499
2500         dlm_put(dlm);
2501         return 0;
2502 }
2503
2504 #define DLM_FINALIZE_STAGE2  0x01
2505 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2506 {
2507         int ret = 0;
2508         struct dlm_finalize_reco fr;
2509         struct dlm_node_iter iter;
2510         int nodenum;
2511         int status;
2512         int stage = 1;
2513
2514         mlog(0, "finishing recovery for node %s:%u, "
2515              "stage %d\n", dlm->name, dlm->reco.dead_node, stage);
2516
2517         spin_lock(&dlm->spinlock);
2518         dlm_node_iter_init(dlm->domain_map, &iter);
2519         spin_unlock(&dlm->spinlock);
2520
2521 stage2:
2522         memset(&fr, 0, sizeof(fr));
2523         fr.node_idx = dlm->node_num;
2524         fr.dead_node = dlm->reco.dead_node;
2525         if (stage == 2)
2526                 fr.flags |= DLM_FINALIZE_STAGE2;
2527
2528         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2529                 if (nodenum == dlm->node_num)
2530                         continue;
2531                 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2532                                          &fr, sizeof(fr), nodenum, &status);
2533                 if (ret >= 0)
2534                         ret = status;
2535                 if (ret < 0) {
2536                         mlog_errno(ret);
2537                         if (dlm_is_host_down(ret)) {
2538                                 /* this has no effect on this recovery 
2539                                  * session, so set the status to zero to 
2540                                  * finish out the last recovery */
2541                                 mlog(ML_ERROR, "node %u went down after this "
2542                                      "node finished recovery.\n", nodenum);
2543                                 ret = 0;
2544                                 continue;
2545                         }
2546                         break;
2547                 }
2548         }
2549         if (stage == 1) {
2550                 /* reset the node_iter back to the top and send finalize2 */
2551                 iter.curnode = -1;
2552                 stage = 2;
2553                 goto stage2;
2554         }
2555
2556         return ret;
2557 }
2558
2559 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data)
2560 {
2561         struct dlm_ctxt *dlm = data;
2562         struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2563         int stage = 1;
2564
2565         /* ok to return 0, domain has gone away */
2566         if (!dlm_grab(dlm))
2567                 return 0;
2568
2569         if (fr->flags & DLM_FINALIZE_STAGE2)
2570                 stage = 2;
2571
2572         mlog(0, "%s: node %u finalizing recovery stage%d of "
2573              "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage,
2574              fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master);
2575  
2576         spin_lock(&dlm->spinlock);
2577
2578         if (dlm->reco.new_master != fr->node_idx) {
2579                 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2580                      "%u is supposed to be the new master, dead=%u\n",
2581                      fr->node_idx, dlm->reco.new_master, fr->dead_node);
2582                 BUG();
2583         }
2584         if (dlm->reco.dead_node != fr->dead_node) {
2585                 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2586                      "node %u, but node %u is supposed to be dead\n",
2587                      fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2588                 BUG();
2589         }
2590
2591         switch (stage) {
2592                 case 1:
2593                         dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2594                         if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2595                                 mlog(ML_ERROR, "%s: received finalize1 from "
2596                                      "new master %u for dead node %u, but "
2597                                      "this node has already received it!\n",
2598                                      dlm->name, fr->node_idx, fr->dead_node);
2599                                 dlm_print_reco_node_status(dlm);
2600                                 BUG();
2601                         }
2602                         dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
2603                         spin_unlock(&dlm->spinlock);
2604                         break;
2605                 case 2:
2606                         if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) {
2607                                 mlog(ML_ERROR, "%s: received finalize2 from "
2608                                      "new master %u for dead node %u, but "
2609                                      "this node did not have finalize1!\n",
2610                                      dlm->name, fr->node_idx, fr->dead_node);
2611                                 dlm_print_reco_node_status(dlm);
2612                                 BUG();
2613                         }
2614                         dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2615                         spin_unlock(&dlm->spinlock);
2616                         dlm_reset_recovery(dlm);
2617                         dlm_kick_recovery_thread(dlm);
2618                         break;
2619                 default:
2620                         BUG();
2621         }
2622
2623         mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n",
2624              dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master);
2625
2626         dlm_put(dlm);
2627         return 0;
2628 }