Merge branch 'master' of /pub/scm/linux/kernel/git/torvalds/linux-2.6
[sfrench/cifs-2.6.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 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
99                                       struct dlm_lock_resource *res,
100                                       u8 *real_master);
101
102 static u64 dlm_get_next_mig_cookie(void);
103
104 static DEFINE_SPINLOCK(dlm_reco_state_lock);
105 static DEFINE_SPINLOCK(dlm_mig_cookie_lock);
106 static u64 dlm_mig_cookie = 1;
107
108 static u64 dlm_get_next_mig_cookie(void)
109 {
110         u64 c;
111         spin_lock(&dlm_mig_cookie_lock);
112         c = dlm_mig_cookie;
113         if (dlm_mig_cookie == (~0ULL))
114                 dlm_mig_cookie = 1;
115         else
116                 dlm_mig_cookie++;
117         spin_unlock(&dlm_mig_cookie_lock);
118         return c;
119 }
120
121 static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm,
122                                           u8 dead_node)
123 {
124         assert_spin_locked(&dlm->spinlock);
125         if (dlm->reco.dead_node != dead_node)
126                 mlog(0, "%s: changing dead_node from %u to %u\n",
127                      dlm->name, dlm->reco.dead_node, dead_node);
128         dlm->reco.dead_node = dead_node;
129 }
130
131 static inline void dlm_set_reco_master(struct dlm_ctxt *dlm,
132                                        u8 master)
133 {
134         assert_spin_locked(&dlm->spinlock);
135         mlog(0, "%s: changing new_master from %u to %u\n",
136              dlm->name, dlm->reco.new_master, master);
137         dlm->reco.new_master = master;
138 }
139
140 static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm)
141 {
142         assert_spin_locked(&dlm->spinlock);
143         clear_bit(dlm->reco.dead_node, dlm->recovery_map);
144         dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
145         dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
146 }
147
148 static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
149 {
150         spin_lock(&dlm->spinlock);
151         __dlm_reset_recovery(dlm);
152         spin_unlock(&dlm->spinlock);
153 }
154
155 /* Worker function used during recovery. */
156 void dlm_dispatch_work(struct work_struct *work)
157 {
158         struct dlm_ctxt *dlm =
159                 container_of(work, struct dlm_ctxt, dispatched_work);
160         LIST_HEAD(tmp_list);
161         struct dlm_work_item *item, *next;
162         dlm_workfunc_t *workfunc;
163         int tot=0;
164
165         spin_lock(&dlm->work_lock);
166         list_splice_init(&dlm->work_list, &tmp_list);
167         spin_unlock(&dlm->work_lock);
168
169         list_for_each_entry(item, &tmp_list, list) {
170                 tot++;
171         }
172         mlog(0, "%s: work thread has %d work items\n", dlm->name, tot);
173
174         list_for_each_entry_safe(item, next, &tmp_list, list) {
175                 workfunc = item->func;
176                 list_del_init(&item->list);
177
178                 /* already have ref on dlm to avoid having
179                  * it disappear.  just double-check. */
180                 BUG_ON(item->dlm != dlm);
181
182                 /* this is allowed to sleep and
183                  * call network stuff */
184                 workfunc(item, item->data);
185
186                 dlm_put(dlm);
187                 kfree(item);
188         }
189 }
190
191 /*
192  * RECOVERY THREAD
193  */
194
195 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
196 {
197         /* wake the recovery thread
198          * this will wake the reco thread in one of three places
199          * 1) sleeping with no recovery happening
200          * 2) sleeping with recovery mastered elsewhere
201          * 3) recovery mastered here, waiting on reco data */
202
203         wake_up(&dlm->dlm_reco_thread_wq);
204 }
205
206 /* Launch the recovery thread */
207 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
208 {
209         mlog(0, "starting dlm recovery thread...\n");
210
211         dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
212                                                 "dlm_reco_thread");
213         if (IS_ERR(dlm->dlm_reco_thread_task)) {
214                 mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
215                 dlm->dlm_reco_thread_task = NULL;
216                 return -EINVAL;
217         }
218
219         return 0;
220 }
221
222 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
223 {
224         if (dlm->dlm_reco_thread_task) {
225                 mlog(0, "waiting for dlm recovery thread to exit\n");
226                 kthread_stop(dlm->dlm_reco_thread_task);
227                 dlm->dlm_reco_thread_task = NULL;
228         }
229 }
230
231
232
233 /*
234  * this is lame, but here's how recovery works...
235  * 1) all recovery threads cluster wide will work on recovering
236  *    ONE node at a time
237  * 2) negotiate who will take over all the locks for the dead node.
238  *    thats right... ALL the locks.
239  * 3) once a new master is chosen, everyone scans all locks
240  *    and moves aside those mastered by the dead guy
241  * 4) each of these locks should be locked until recovery is done
242  * 5) the new master collects up all of secondary lock queue info
243  *    one lock at a time, forcing each node to communicate back
244  *    before continuing
245  * 6) each secondary lock queue responds with the full known lock info
246  * 7) once the new master has run all its locks, it sends a ALLDONE!
247  *    message to everyone
248  * 8) upon receiving this message, the secondary queue node unlocks
249  *    and responds to the ALLDONE
250  * 9) once the new master gets responses from everyone, he unlocks
251  *    everything and recovery for this dead node is done
252  *10) go back to 2) while there are still dead nodes
253  *
254  */
255
256 static void dlm_print_reco_node_status(struct dlm_ctxt *dlm)
257 {
258         struct dlm_reco_node_data *ndata;
259         struct dlm_lock_resource *res;
260
261         mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n",
262              dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
263              dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive",
264              dlm->reco.dead_node, dlm->reco.new_master);
265
266         list_for_each_entry(ndata, &dlm->reco.node_data, list) {
267                 char *st = "unknown";
268                 switch (ndata->state) {
269                         case DLM_RECO_NODE_DATA_INIT:
270                                 st = "init";
271                                 break;
272                         case DLM_RECO_NODE_DATA_REQUESTING:
273                                 st = "requesting";
274                                 break;
275                         case DLM_RECO_NODE_DATA_DEAD:
276                                 st = "dead";
277                                 break;
278                         case DLM_RECO_NODE_DATA_RECEIVING:
279                                 st = "receiving";
280                                 break;
281                         case DLM_RECO_NODE_DATA_REQUESTED:
282                                 st = "requested";
283                                 break;
284                         case DLM_RECO_NODE_DATA_DONE:
285                                 st = "done";
286                                 break;
287                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
288                                 st = "finalize-sent";
289                                 break;
290                         default:
291                                 st = "bad";
292                                 break;
293                 }
294                 mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n",
295                      dlm->name, ndata->node_num, st);
296         }
297         list_for_each_entry(res, &dlm->reco.resources, recovering) {
298                 mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n",
299                      dlm->name, res->lockname.len, res->lockname.name);
300         }
301 }
302
303 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
304
305 static int dlm_recovery_thread(void *data)
306 {
307         int status;
308         struct dlm_ctxt *dlm = data;
309         unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
310
311         mlog(0, "dlm thread running for %s...\n", dlm->name);
312
313         while (!kthread_should_stop()) {
314                 if (dlm_joined(dlm)) {
315                         status = dlm_do_recovery(dlm);
316                         if (status == -EAGAIN) {
317                                 /* do not sleep, recheck immediately. */
318                                 continue;
319                         }
320                         if (status < 0)
321                                 mlog_errno(status);
322                 }
323
324                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
325                                                  kthread_should_stop(),
326                                                  timeout);
327         }
328
329         mlog(0, "quitting DLM recovery thread\n");
330         return 0;
331 }
332
333 /* returns true when the recovery master has contacted us */
334 static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
335 {
336         int ready;
337         spin_lock(&dlm->spinlock);
338         ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
339         spin_unlock(&dlm->spinlock);
340         return ready;
341 }
342
343 /* returns true if node is no longer in the domain
344  * could be dead or just not joined */
345 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
346 {
347         int dead;
348         spin_lock(&dlm->spinlock);
349         dead = !test_bit(node, dlm->domain_map);
350         spin_unlock(&dlm->spinlock);
351         return dead;
352 }
353
354 /* returns true if node is no longer in the domain
355  * could be dead or just not joined */
356 static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node)
357 {
358         int recovered;
359         spin_lock(&dlm->spinlock);
360         recovered = !test_bit(node, dlm->recovery_map);
361         spin_unlock(&dlm->spinlock);
362         return recovered;
363 }
364
365
366 int dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout)
367 {
368         if (timeout) {
369                 mlog(ML_NOTICE, "%s: waiting %dms for notification of "
370                      "death of node %u\n", dlm->name, timeout, node);
371                 wait_event_timeout(dlm->dlm_reco_thread_wq,
372                            dlm_is_node_dead(dlm, node),
373                            msecs_to_jiffies(timeout));
374         } else {
375                 mlog(ML_NOTICE, "%s: waiting indefinitely for notification "
376                      "of death of node %u\n", dlm->name, node);
377                 wait_event(dlm->dlm_reco_thread_wq,
378                            dlm_is_node_dead(dlm, node));
379         }
380         /* for now, return 0 */
381         return 0;
382 }
383
384 int dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout)
385 {
386         if (timeout) {
387                 mlog(0, "%s: waiting %dms for notification of "
388                      "recovery of node %u\n", dlm->name, timeout, node);
389                 wait_event_timeout(dlm->dlm_reco_thread_wq,
390                            dlm_is_node_recovered(dlm, node),
391                            msecs_to_jiffies(timeout));
392         } else {
393                 mlog(0, "%s: waiting indefinitely for notification "
394                      "of recovery of node %u\n", dlm->name, node);
395                 wait_event(dlm->dlm_reco_thread_wq,
396                            dlm_is_node_recovered(dlm, node));
397         }
398         /* for now, return 0 */
399         return 0;
400 }
401
402 /* callers of the top-level api calls (dlmlock/dlmunlock) should
403  * block on the dlm->reco.event when recovery is in progress.
404  * the dlm recovery thread will set this state when it begins
405  * recovering a dead node (as the new master or not) and clear
406  * the state and wake as soon as all affected lock resources have
407  * been marked with the RECOVERY flag */
408 static int dlm_in_recovery(struct dlm_ctxt *dlm)
409 {
410         int in_recovery;
411         spin_lock(&dlm->spinlock);
412         in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
413         spin_unlock(&dlm->spinlock);
414         return in_recovery;
415 }
416
417
418 void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
419 {
420         if (dlm_in_recovery(dlm)) {
421                 mlog(0, "%s: reco thread %d in recovery: "
422                      "state=%d, master=%u, dead=%u\n",
423                      dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
424                      dlm->reco.state, dlm->reco.new_master,
425                      dlm->reco.dead_node);
426         }
427         wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
428 }
429
430 static void dlm_begin_recovery(struct dlm_ctxt *dlm)
431 {
432         spin_lock(&dlm->spinlock);
433         BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
434         dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
435         spin_unlock(&dlm->spinlock);
436 }
437
438 static void dlm_end_recovery(struct dlm_ctxt *dlm)
439 {
440         spin_lock(&dlm->spinlock);
441         BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
442         dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
443         spin_unlock(&dlm->spinlock);
444         wake_up(&dlm->reco.event);
445 }
446
447 static int dlm_do_recovery(struct dlm_ctxt *dlm)
448 {
449         int status = 0;
450         int ret;
451
452         spin_lock(&dlm->spinlock);
453
454         /* check to see if the new master has died */
455         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
456             test_bit(dlm->reco.new_master, dlm->recovery_map)) {
457                 mlog(0, "new master %u died while recovering %u!\n",
458                      dlm->reco.new_master, dlm->reco.dead_node);
459                 /* unset the new_master, leave dead_node */
460                 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
461         }
462
463         /* select a target to recover */
464         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
465                 int bit;
466
467                 bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES+1, 0);
468                 if (bit >= O2NM_MAX_NODES || bit < 0)
469                         dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
470                 else
471                         dlm_set_reco_dead_node(dlm, bit);
472         } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
473                 /* BUG? */
474                 mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
475                      dlm->reco.dead_node);
476                 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
477         }
478
479         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
480                 // mlog(0, "nothing to recover!  sleeping now!\n");
481                 spin_unlock(&dlm->spinlock);
482                 /* return to main thread loop and sleep. */
483                 return 0;
484         }
485         mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n",
486              dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
487              dlm->reco.dead_node);
488         spin_unlock(&dlm->spinlock);
489
490         /* take write barrier */
491         /* (stops the list reshuffling thread, proxy ast handling) */
492         dlm_begin_recovery(dlm);
493
494         if (dlm->reco.new_master == dlm->node_num)
495                 goto master_here;
496
497         if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
498                 /* choose a new master, returns 0 if this node
499                  * is the master, -EEXIST if it's another node.
500                  * this does not return until a new master is chosen
501                  * or recovery completes entirely. */
502                 ret = dlm_pick_recovery_master(dlm);
503                 if (!ret) {
504                         /* already notified everyone.  go. */
505                         goto master_here;
506                 }
507                 mlog(0, "another node will master this recovery session.\n");
508         }
509         mlog(0, "dlm=%s (%d), new_master=%u, this node=%u, dead_node=%u\n",
510              dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), dlm->reco.new_master,
511              dlm->node_num, dlm->reco.dead_node);
512
513         /* it is safe to start everything back up here
514          * because all of the dead node's lock resources
515          * have been marked as in-recovery */
516         dlm_end_recovery(dlm);
517
518         /* sleep out in main dlm_recovery_thread loop. */
519         return 0;
520
521 master_here:
522         mlog(ML_NOTICE, "(%d) Node %u is the Recovery Master for the Dead Node "
523              "%u for Domain %s\n", task_pid_nr(dlm->dlm_reco_thread_task),
524              dlm->node_num, dlm->reco.dead_node, dlm->name);
525
526         status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
527         if (status < 0) {
528                 /* we should never hit this anymore */
529                 mlog(ML_ERROR, "error %d remastering locks for node %u, "
530                      "retrying.\n", status, dlm->reco.dead_node);
531                 /* yield a bit to allow any final network messages
532                  * to get handled on remaining nodes */
533                 msleep(100);
534         } else {
535                 /* success!  see if any other nodes need recovery */
536                 mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
537                      dlm->name, dlm->reco.dead_node, dlm->node_num);
538                 dlm_reset_recovery(dlm);
539         }
540         dlm_end_recovery(dlm);
541
542         /* continue and look for another dead node */
543         return -EAGAIN;
544 }
545
546 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
547 {
548         int status = 0;
549         struct dlm_reco_node_data *ndata;
550         int all_nodes_done;
551         int destroy = 0;
552         int pass = 0;
553
554         do {
555                 /* we have become recovery master.  there is no escaping
556                  * this, so just keep trying until we get it. */
557                 status = dlm_init_recovery_area(dlm, dead_node);
558                 if (status < 0) {
559                         mlog(ML_ERROR, "%s: failed to alloc recovery area, "
560                              "retrying\n", dlm->name);
561                         msleep(1000);
562                 }
563         } while (status != 0);
564
565         /* safe to access the node data list without a lock, since this
566          * process is the only one to change the list */
567         list_for_each_entry(ndata, &dlm->reco.node_data, list) {
568                 BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
569                 ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
570
571                 mlog(0, "requesting lock info from node %u\n",
572                      ndata->node_num);
573
574                 if (ndata->node_num == dlm->node_num) {
575                         ndata->state = DLM_RECO_NODE_DATA_DONE;
576                         continue;
577                 }
578
579                 do {
580                         status = dlm_request_all_locks(dlm, ndata->node_num,
581                                                        dead_node);
582                         if (status < 0) {
583                                 mlog_errno(status);
584                                 if (dlm_is_host_down(status)) {
585                                         /* node died, ignore it for recovery */
586                                         status = 0;
587                                         ndata->state = DLM_RECO_NODE_DATA_DEAD;
588                                         /* wait for the domain map to catch up
589                                          * with the network state. */
590                                         wait_event_timeout(dlm->dlm_reco_thread_wq,
591                                                            dlm_is_node_dead(dlm,
592                                                                 ndata->node_num),
593                                                            msecs_to_jiffies(1000));
594                                         mlog(0, "waited 1 sec for %u, "
595                                              "dead? %s\n", ndata->node_num,
596                                              dlm_is_node_dead(dlm, ndata->node_num) ?
597                                              "yes" : "no");
598                                 } else {
599                                         /* -ENOMEM on the other node */
600                                         mlog(0, "%s: node %u returned "
601                                              "%d during recovery, retrying "
602                                              "after a short wait\n",
603                                              dlm->name, ndata->node_num,
604                                              status);
605                                         msleep(100);
606                                 }
607                         }
608                 } while (status != 0);
609
610                 spin_lock(&dlm_reco_state_lock);
611                 switch (ndata->state) {
612                         case DLM_RECO_NODE_DATA_INIT:
613                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
614                         case DLM_RECO_NODE_DATA_REQUESTED:
615                                 BUG();
616                                 break;
617                         case DLM_RECO_NODE_DATA_DEAD:
618                                 mlog(0, "node %u died after requesting "
619                                      "recovery info for node %u\n",
620                                      ndata->node_num, dead_node);
621                                 /* fine.  don't need this node's info.
622                                  * continue without it. */
623                                 break;
624                         case DLM_RECO_NODE_DATA_REQUESTING:
625                                 ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
626                                 mlog(0, "now receiving recovery data from "
627                                      "node %u for dead node %u\n",
628                                      ndata->node_num, dead_node);
629                                 break;
630                         case DLM_RECO_NODE_DATA_RECEIVING:
631                                 mlog(0, "already receiving recovery data from "
632                                      "node %u for dead node %u\n",
633                                      ndata->node_num, dead_node);
634                                 break;
635                         case DLM_RECO_NODE_DATA_DONE:
636                                 mlog(0, "already DONE receiving recovery data "
637                                      "from node %u for dead node %u\n",
638                                      ndata->node_num, dead_node);
639                                 break;
640                 }
641                 spin_unlock(&dlm_reco_state_lock);
642         }
643
644         mlog(0, "done requesting all lock info\n");
645
646         /* nodes should be sending reco data now
647          * just need to wait */
648
649         while (1) {
650                 /* check all the nodes now to see if we are
651                  * done, or if anyone died */
652                 all_nodes_done = 1;
653                 spin_lock(&dlm_reco_state_lock);
654                 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
655                         mlog(0, "checking recovery state of node %u\n",
656                              ndata->node_num);
657                         switch (ndata->state) {
658                                 case DLM_RECO_NODE_DATA_INIT:
659                                 case DLM_RECO_NODE_DATA_REQUESTING:
660                                         mlog(ML_ERROR, "bad ndata state for "
661                                              "node %u: state=%d\n",
662                                              ndata->node_num, ndata->state);
663                                         BUG();
664                                         break;
665                                 case DLM_RECO_NODE_DATA_DEAD:
666                                         mlog(0, "node %u died after "
667                                              "requesting recovery info for "
668                                              "node %u\n", ndata->node_num,
669                                              dead_node);
670                                         break;
671                                 case DLM_RECO_NODE_DATA_RECEIVING:
672                                 case DLM_RECO_NODE_DATA_REQUESTED:
673                                         mlog(0, "%s: node %u still in state %s\n",
674                                              dlm->name, ndata->node_num,
675                                              ndata->state==DLM_RECO_NODE_DATA_RECEIVING ?
676                                              "receiving" : "requested");
677                                         all_nodes_done = 0;
678                                         break;
679                                 case DLM_RECO_NODE_DATA_DONE:
680                                         mlog(0, "%s: node %u state is done\n",
681                                              dlm->name, ndata->node_num);
682                                         break;
683                                 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
684                                         mlog(0, "%s: node %u state is finalize\n",
685                                              dlm->name, ndata->node_num);
686                                         break;
687                         }
688                 }
689                 spin_unlock(&dlm_reco_state_lock);
690
691                 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
692                      all_nodes_done?"yes":"no");
693                 if (all_nodes_done) {
694                         int ret;
695
696                         /* all nodes are now in DLM_RECO_NODE_DATA_DONE state
697                          * just send a finalize message to everyone and
698                          * clean up */
699                         mlog(0, "all nodes are done! send finalize\n");
700                         ret = dlm_send_finalize_reco_message(dlm);
701                         if (ret < 0)
702                                 mlog_errno(ret);
703
704                         spin_lock(&dlm->spinlock);
705                         dlm_finish_local_lockres_recovery(dlm, dead_node,
706                                                           dlm->node_num);
707                         spin_unlock(&dlm->spinlock);
708                         mlog(0, "should be done with recovery!\n");
709
710                         mlog(0, "finishing recovery of %s at %lu, "
711                              "dead=%u, this=%u, new=%u\n", dlm->name,
712                              jiffies, dlm->reco.dead_node,
713                              dlm->node_num, dlm->reco.new_master);
714                         destroy = 1;
715                         status = 0;
716                         /* rescan everything marked dirty along the way */
717                         dlm_kick_thread(dlm, NULL);
718                         break;
719                 }
720                 /* wait to be signalled, with periodic timeout
721                  * to check for node death */
722                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
723                                          kthread_should_stop(),
724                                          msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
725
726         }
727
728         if (destroy)
729                 dlm_destroy_recovery_area(dlm, dead_node);
730
731         mlog_exit(status);
732         return status;
733 }
734
735 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
736 {
737         int num=0;
738         struct dlm_reco_node_data *ndata;
739
740         spin_lock(&dlm->spinlock);
741         memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
742         /* nodes can only be removed (by dying) after dropping
743          * this lock, and death will be trapped later, so this should do */
744         spin_unlock(&dlm->spinlock);
745
746         while (1) {
747                 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
748                 if (num >= O2NM_MAX_NODES) {
749                         break;
750                 }
751                 BUG_ON(num == dead_node);
752
753                 ndata = kzalloc(sizeof(*ndata), GFP_NOFS);
754                 if (!ndata) {
755                         dlm_destroy_recovery_area(dlm, dead_node);
756                         return -ENOMEM;
757                 }
758                 ndata->node_num = num;
759                 ndata->state = DLM_RECO_NODE_DATA_INIT;
760                 spin_lock(&dlm_reco_state_lock);
761                 list_add_tail(&ndata->list, &dlm->reco.node_data);
762                 spin_unlock(&dlm_reco_state_lock);
763                 num++;
764         }
765
766         return 0;
767 }
768
769 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
770 {
771         struct dlm_reco_node_data *ndata, *next;
772         LIST_HEAD(tmplist);
773
774         spin_lock(&dlm_reco_state_lock);
775         list_splice_init(&dlm->reco.node_data, &tmplist);
776         spin_unlock(&dlm_reco_state_lock);
777
778         list_for_each_entry_safe(ndata, next, &tmplist, list) {
779                 list_del_init(&ndata->list);
780                 kfree(ndata);
781         }
782 }
783
784 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
785                                  u8 dead_node)
786 {
787         struct dlm_lock_request lr;
788         enum dlm_status ret;
789
790         mlog(0, "\n");
791
792
793         mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
794                   "to %u\n", dead_node, request_from);
795
796         memset(&lr, 0, sizeof(lr));
797         lr.node_idx = dlm->node_num;
798         lr.dead_node = dead_node;
799
800         // send message
801         ret = DLM_NOLOCKMGR;
802         ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
803                                  &lr, sizeof(lr), request_from, NULL);
804
805         /* negative status is handled by caller */
806         if (ret < 0)
807                 mlog_errno(ret);
808
809         // return from here, then
810         // sleep until all received or error
811         return ret;
812
813 }
814
815 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
816                                   void **ret_data)
817 {
818         struct dlm_ctxt *dlm = data;
819         struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
820         char *buf = NULL;
821         struct dlm_work_item *item = NULL;
822
823         if (!dlm_grab(dlm))
824                 return -EINVAL;
825
826         if (lr->dead_node != dlm->reco.dead_node) {
827                 mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local "
828                      "dead_node is %u\n", dlm->name, lr->node_idx,
829                      lr->dead_node, dlm->reco.dead_node);
830                 dlm_print_reco_node_status(dlm);
831                 /* this is a hack */
832                 dlm_put(dlm);
833                 return -ENOMEM;
834         }
835         BUG_ON(lr->dead_node != dlm->reco.dead_node);
836
837         item = kzalloc(sizeof(*item), GFP_NOFS);
838         if (!item) {
839                 dlm_put(dlm);
840                 return -ENOMEM;
841         }
842
843         /* this will get freed by dlm_request_all_locks_worker */
844         buf = (char *) __get_free_page(GFP_NOFS);
845         if (!buf) {
846                 kfree(item);
847                 dlm_put(dlm);
848                 return -ENOMEM;
849         }
850
851         /* queue up work for dlm_request_all_locks_worker */
852         dlm_grab(dlm);  /* get an extra ref for the work item */
853         dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
854         item->u.ral.reco_master = lr->node_idx;
855         item->u.ral.dead_node = lr->dead_node;
856         spin_lock(&dlm->work_lock);
857         list_add_tail(&item->list, &dlm->work_list);
858         spin_unlock(&dlm->work_lock);
859         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
860
861         dlm_put(dlm);
862         return 0;
863 }
864
865 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
866 {
867         struct dlm_migratable_lockres *mres;
868         struct dlm_lock_resource *res;
869         struct dlm_ctxt *dlm;
870         LIST_HEAD(resources);
871         int ret;
872         u8 dead_node, reco_master;
873         int skip_all_done = 0;
874
875         dlm = item->dlm;
876         dead_node = item->u.ral.dead_node;
877         reco_master = item->u.ral.reco_master;
878         mres = (struct dlm_migratable_lockres *)data;
879
880         mlog(0, "%s: recovery worker started, dead=%u, master=%u\n",
881              dlm->name, dead_node, reco_master);
882
883         if (dead_node != dlm->reco.dead_node ||
884             reco_master != dlm->reco.new_master) {
885                 /* worker could have been created before the recovery master
886                  * died.  if so, do not continue, but do not error. */
887                 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
888                         mlog(ML_NOTICE, "%s: will not send recovery state, "
889                              "recovery master %u died, thread=(dead=%u,mas=%u)"
890                              " current=(dead=%u,mas=%u)\n", dlm->name,
891                              reco_master, dead_node, reco_master,
892                              dlm->reco.dead_node, dlm->reco.new_master);
893                 } else {
894                         mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, "
895                              "master=%u), request(dead=%u, master=%u)\n",
896                              dlm->name, dlm->reco.dead_node,
897                              dlm->reco.new_master, dead_node, reco_master);
898                 }
899                 goto leave;
900         }
901
902         /* lock resources should have already been moved to the
903          * dlm->reco.resources list.  now move items from that list
904          * to a temp list if the dead owner matches.  note that the
905          * whole cluster recovers only one node at a time, so we
906          * can safely move UNKNOWN lock resources for each recovery
907          * session. */
908         dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
909
910         /* now we can begin blasting lockreses without the dlm lock */
911
912         /* any errors returned will be due to the new_master dying,
913          * the dlm_reco_thread should detect this */
914         list_for_each_entry(res, &resources, recovering) {
915                 ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
916                                         DLM_MRES_RECOVERY);
917                 if (ret < 0) {
918                         mlog(ML_ERROR, "%s: node %u went down while sending "
919                              "recovery state for dead node %u, ret=%d\n", dlm->name,
920                              reco_master, dead_node, ret);
921                         skip_all_done = 1;
922                         break;
923                 }
924         }
925
926         /* move the resources back to the list */
927         spin_lock(&dlm->spinlock);
928         list_splice_init(&resources, &dlm->reco.resources);
929         spin_unlock(&dlm->spinlock);
930
931         if (!skip_all_done) {
932                 ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
933                 if (ret < 0) {
934                         mlog(ML_ERROR, "%s: node %u went down while sending "
935                              "recovery all-done for dead node %u, ret=%d\n",
936                              dlm->name, reco_master, dead_node, ret);
937                 }
938         }
939 leave:
940         free_page((unsigned long)data);
941 }
942
943
944 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
945 {
946         int ret, tmpret;
947         struct dlm_reco_data_done done_msg;
948
949         memset(&done_msg, 0, sizeof(done_msg));
950         done_msg.node_idx = dlm->node_num;
951         done_msg.dead_node = dead_node;
952         mlog(0, "sending DATA DONE message to %u, "
953              "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
954              done_msg.dead_node);
955
956         ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
957                                  sizeof(done_msg), send_to, &tmpret);
958         if (ret < 0) {
959                 if (!dlm_is_host_down(ret)) {
960                         mlog_errno(ret);
961                         mlog(ML_ERROR, "%s: unknown error sending data-done "
962                              "to %u\n", dlm->name, send_to);
963                         BUG();
964                 }
965         } else
966                 ret = tmpret;
967         return ret;
968 }
969
970
971 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
972                                void **ret_data)
973 {
974         struct dlm_ctxt *dlm = data;
975         struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
976         struct dlm_reco_node_data *ndata = NULL;
977         int ret = -EINVAL;
978
979         if (!dlm_grab(dlm))
980                 return -EINVAL;
981
982         mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
983              "node_idx=%u, this node=%u\n", done->dead_node,
984              dlm->reco.dead_node, done->node_idx, dlm->node_num);
985
986         mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node),
987                         "Got DATA DONE: dead_node=%u, reco.dead_node=%u, "
988                         "node_idx=%u, this node=%u\n", done->dead_node,
989                         dlm->reco.dead_node, done->node_idx, dlm->node_num);
990
991         spin_lock(&dlm_reco_state_lock);
992         list_for_each_entry(ndata, &dlm->reco.node_data, list) {
993                 if (ndata->node_num != done->node_idx)
994                         continue;
995
996                 switch (ndata->state) {
997                         /* should have moved beyond INIT but not to FINALIZE yet */
998                         case DLM_RECO_NODE_DATA_INIT:
999                         case DLM_RECO_NODE_DATA_DEAD:
1000                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
1001                                 mlog(ML_ERROR, "bad ndata state for node %u:"
1002                                      " state=%d\n", ndata->node_num,
1003                                      ndata->state);
1004                                 BUG();
1005                                 break;
1006                         /* these states are possible at this point, anywhere along
1007                          * the line of recovery */
1008                         case DLM_RECO_NODE_DATA_DONE:
1009                         case DLM_RECO_NODE_DATA_RECEIVING:
1010                         case DLM_RECO_NODE_DATA_REQUESTED:
1011                         case DLM_RECO_NODE_DATA_REQUESTING:
1012                                 mlog(0, "node %u is DONE sending "
1013                                           "recovery data!\n",
1014                                           ndata->node_num);
1015
1016                                 ndata->state = DLM_RECO_NODE_DATA_DONE;
1017                                 ret = 0;
1018                                 break;
1019                 }
1020         }
1021         spin_unlock(&dlm_reco_state_lock);
1022
1023         /* wake the recovery thread, some node is done */
1024         if (!ret)
1025                 dlm_kick_recovery_thread(dlm);
1026
1027         if (ret < 0)
1028                 mlog(ML_ERROR, "failed to find recovery node data for node "
1029                      "%u\n", done->node_idx);
1030         dlm_put(dlm);
1031
1032         mlog(0, "leaving reco data done handler, ret=%d\n", ret);
1033         return ret;
1034 }
1035
1036 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
1037                                         struct list_head *list,
1038                                         u8 dead_node)
1039 {
1040         struct dlm_lock_resource *res, *next;
1041         struct dlm_lock *lock;
1042
1043         spin_lock(&dlm->spinlock);
1044         list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
1045                 /* always prune any $RECOVERY entries for dead nodes,
1046                  * otherwise hangs can occur during later recovery */
1047                 if (dlm_is_recovery_lock(res->lockname.name,
1048                                          res->lockname.len)) {
1049                         spin_lock(&res->spinlock);
1050                         list_for_each_entry(lock, &res->granted, list) {
1051                                 if (lock->ml.node == dead_node) {
1052                                         mlog(0, "AHA! there was "
1053                                              "a $RECOVERY lock for dead "
1054                                              "node %u (%s)!\n", 
1055                                              dead_node, dlm->name);
1056                                         list_del_init(&lock->list);
1057                                         dlm_lock_put(lock);
1058                                         break;
1059                                 }
1060                         }
1061                         spin_unlock(&res->spinlock);
1062                         continue;
1063                 }
1064
1065                 if (res->owner == dead_node) {
1066                         mlog(0, "found lockres owned by dead node while "
1067                                   "doing recovery for node %u. sending it.\n",
1068                                   dead_node);
1069                         list_move_tail(&res->recovering, list);
1070                 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1071                         mlog(0, "found UNKNOWN owner while doing recovery "
1072                                   "for node %u. sending it.\n", dead_node);
1073                         list_move_tail(&res->recovering, list);
1074                 }
1075         }
1076         spin_unlock(&dlm->spinlock);
1077 }
1078
1079 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
1080 {
1081         int total_locks = 0;
1082         struct list_head *iter, *queue = &res->granted;
1083         int i;
1084
1085         for (i=0; i<3; i++) {
1086                 list_for_each(iter, queue)
1087                         total_locks++;
1088                 queue++;
1089         }
1090         return total_locks;
1091 }
1092
1093
1094 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
1095                                       struct dlm_migratable_lockres *mres,
1096                                       u8 send_to,
1097                                       struct dlm_lock_resource *res,
1098                                       int total_locks)
1099 {
1100         u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
1101         int mres_total_locks = be32_to_cpu(mres->total_locks);
1102         int sz, ret = 0, status = 0;
1103         u8 orig_flags = mres->flags,
1104            orig_master = mres->master;
1105
1106         BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
1107         if (!mres->num_locks)
1108                 return 0;
1109
1110         sz = sizeof(struct dlm_migratable_lockres) +
1111                 (mres->num_locks * sizeof(struct dlm_migratable_lock));
1112
1113         /* add an all-done flag if we reached the last lock */
1114         orig_flags = mres->flags;
1115         BUG_ON(total_locks > mres_total_locks);
1116         if (total_locks == mres_total_locks)
1117                 mres->flags |= DLM_MRES_ALL_DONE;
1118
1119         mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n",
1120              dlm->name, res->lockname.len, res->lockname.name,
1121              orig_flags & DLM_MRES_MIGRATION ? "migrate" : "recovery",
1122              send_to);
1123
1124         /* send it */
1125         ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
1126                                  sz, send_to, &status);
1127         if (ret < 0) {
1128                 /* XXX: negative status is not handled.
1129                  * this will end up killing this node. */
1130                 mlog_errno(ret);
1131         } else {
1132                 /* might get an -ENOMEM back here */
1133                 ret = status;
1134                 if (ret < 0) {
1135                         mlog_errno(ret);
1136
1137                         if (ret == -EFAULT) {
1138                                 mlog(ML_ERROR, "node %u told me to kill "
1139                                      "myself!\n", send_to);
1140                                 BUG();
1141                         }
1142                 }
1143         }
1144
1145         /* zero and reinit the message buffer */
1146         dlm_init_migratable_lockres(mres, res->lockname.name,
1147                                     res->lockname.len, mres_total_locks,
1148                                     mig_cookie, orig_flags, orig_master);
1149         return ret;
1150 }
1151
1152 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
1153                                         const char *lockname, int namelen,
1154                                         int total_locks, u64 cookie,
1155                                         u8 flags, u8 master)
1156 {
1157         /* mres here is one full page */
1158         clear_page(mres);
1159         mres->lockname_len = namelen;
1160         memcpy(mres->lockname, lockname, namelen);
1161         mres->num_locks = 0;
1162         mres->total_locks = cpu_to_be32(total_locks);
1163         mres->mig_cookie = cpu_to_be64(cookie);
1164         mres->flags = flags;
1165         mres->master = master;
1166 }
1167
1168
1169 /* returns 1 if this lock fills the network structure,
1170  * 0 otherwise */
1171 static int dlm_add_lock_to_array(struct dlm_lock *lock,
1172                                  struct dlm_migratable_lockres *mres, int queue)
1173 {
1174         struct dlm_migratable_lock *ml;
1175         int lock_num = mres->num_locks;
1176
1177         ml = &(mres->ml[lock_num]);
1178         ml->cookie = lock->ml.cookie;
1179         ml->type = lock->ml.type;
1180         ml->convert_type = lock->ml.convert_type;
1181         ml->highest_blocked = lock->ml.highest_blocked;
1182         ml->list = queue;
1183         if (lock->lksb) {
1184                 ml->flags = lock->lksb->flags;
1185                 /* send our current lvb */
1186                 if (ml->type == LKM_EXMODE ||
1187                     ml->type == LKM_PRMODE) {
1188                         /* if it is already set, this had better be a PR
1189                          * and it has to match */
1190                         if (!dlm_lvb_is_empty(mres->lvb) &&
1191                             (ml->type == LKM_EXMODE ||
1192                              memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))) {
1193                                 mlog(ML_ERROR, "mismatched lvbs!\n");
1194                                 dlm_print_one_lock_resource(lock->lockres);
1195                                 BUG();
1196                         }
1197                         memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1198                 }
1199         }
1200         ml->node = lock->ml.node;
1201         mres->num_locks++;
1202         /* we reached the max, send this network message */
1203         if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1204                 return 1;
1205         return 0;
1206 }
1207
1208 static void dlm_add_dummy_lock(struct dlm_ctxt *dlm,
1209                                struct dlm_migratable_lockres *mres)
1210 {
1211         struct dlm_lock dummy;
1212         memset(&dummy, 0, sizeof(dummy));
1213         dummy.ml.cookie = 0;
1214         dummy.ml.type = LKM_IVMODE;
1215         dummy.ml.convert_type = LKM_IVMODE;
1216         dummy.ml.highest_blocked = LKM_IVMODE;
1217         dummy.lksb = NULL;
1218         dummy.ml.node = dlm->node_num;
1219         dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST);
1220 }
1221
1222 static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm,
1223                                     struct dlm_migratable_lock *ml,
1224                                     u8 *nodenum)
1225 {
1226         if (unlikely(ml->cookie == 0 &&
1227             ml->type == LKM_IVMODE &&
1228             ml->convert_type == LKM_IVMODE &&
1229             ml->highest_blocked == LKM_IVMODE &&
1230             ml->list == DLM_BLOCKED_LIST)) {
1231                 *nodenum = ml->node;
1232                 return 1;
1233         }
1234         return 0;
1235 }
1236
1237 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1238                          struct dlm_migratable_lockres *mres,
1239                          u8 send_to, u8 flags)
1240 {
1241         struct list_head *queue;
1242         int total_locks, i;
1243         u64 mig_cookie = 0;
1244         struct dlm_lock *lock;
1245         int ret = 0;
1246
1247         BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1248
1249         mlog(0, "sending to %u\n", send_to);
1250
1251         total_locks = dlm_num_locks_in_lockres(res);
1252         if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1253                 /* rare, but possible */
1254                 mlog(0, "argh.  lockres has %d locks.  this will "
1255                           "require more than one network packet to "
1256                           "migrate\n", total_locks);
1257                 mig_cookie = dlm_get_next_mig_cookie();
1258         }
1259
1260         dlm_init_migratable_lockres(mres, res->lockname.name,
1261                                     res->lockname.len, total_locks,
1262                                     mig_cookie, flags, res->owner);
1263
1264         total_locks = 0;
1265         for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1266                 queue = dlm_list_idx_to_ptr(res, i);
1267                 list_for_each_entry(lock, queue, list) {
1268                         /* add another lock. */
1269                         total_locks++;
1270                         if (!dlm_add_lock_to_array(lock, mres, i))
1271                                 continue;
1272
1273                         /* this filled the lock message,
1274                          * we must send it immediately. */
1275                         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1276                                                        res, total_locks);
1277                         if (ret < 0)
1278                                 goto error;
1279                 }
1280         }
1281         if (total_locks == 0) {
1282                 /* send a dummy lock to indicate a mastery reference only */
1283                 mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n",
1284                      dlm->name, res->lockname.len, res->lockname.name,
1285                      send_to, flags & DLM_MRES_RECOVERY ? "recovery" :
1286                      "migration");
1287                 dlm_add_dummy_lock(dlm, mres);
1288         }
1289         /* flush any remaining locks */
1290         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1291         if (ret < 0)
1292                 goto error;
1293         return ret;
1294
1295 error:
1296         mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n",
1297              dlm->name, ret);
1298         if (!dlm_is_host_down(ret))
1299                 BUG();
1300         mlog(0, "%s: node %u went down while sending %s "
1301              "lockres %.*s\n", dlm->name, send_to,
1302              flags & DLM_MRES_RECOVERY ?  "recovery" : "migration",
1303              res->lockname.len, res->lockname.name);
1304         return ret;
1305 }
1306
1307
1308
1309 /*
1310  * this message will contain no more than one page worth of
1311  * recovery data, and it will work on only one lockres.
1312  * there may be many locks in this page, and we may need to wait
1313  * for additional packets to complete all the locks (rare, but
1314  * possible).
1315  */
1316 /*
1317  * NOTE: the allocation error cases here are scary
1318  * we really cannot afford to fail an alloc in recovery
1319  * do we spin?  returning an error only delays the problem really
1320  */
1321
1322 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
1323                             void **ret_data)
1324 {
1325         struct dlm_ctxt *dlm = data;
1326         struct dlm_migratable_lockres *mres =
1327                 (struct dlm_migratable_lockres *)msg->buf;
1328         int ret = 0;
1329         u8 real_master;
1330         u8 extra_refs = 0;
1331         char *buf = NULL;
1332         struct dlm_work_item *item = NULL;
1333         struct dlm_lock_resource *res = NULL;
1334
1335         if (!dlm_grab(dlm))
1336                 return -EINVAL;
1337
1338         BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1339
1340         real_master = mres->master;
1341         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1342                 /* cannot migrate a lockres with no master */
1343                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1344         }
1345
1346         mlog(0, "%s message received from node %u\n",
1347                   (mres->flags & DLM_MRES_RECOVERY) ?
1348                   "recovery" : "migration", mres->master);
1349         if (mres->flags & DLM_MRES_ALL_DONE)
1350                 mlog(0, "all done flag.  all lockres data received!\n");
1351
1352         ret = -ENOMEM;
1353         buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS);
1354         item = kzalloc(sizeof(*item), GFP_NOFS);
1355         if (!buf || !item)
1356                 goto leave;
1357
1358         /* lookup the lock to see if we have a secondary queue for this
1359          * already...  just add the locks in and this will have its owner
1360          * and RECOVERY flag changed when it completes. */
1361         res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len);
1362         if (res) {
1363                 /* this will get a ref on res */
1364                 /* mark it as recovering/migrating and hash it */
1365                 spin_lock(&res->spinlock);
1366                 if (mres->flags & DLM_MRES_RECOVERY) {
1367                         res->state |= DLM_LOCK_RES_RECOVERING;
1368                 } else {
1369                         if (res->state & DLM_LOCK_RES_MIGRATING) {
1370                                 /* this is at least the second
1371                                  * lockres message */
1372                                 mlog(0, "lock %.*s is already migrating\n",
1373                                           mres->lockname_len,
1374                                           mres->lockname);
1375                         } else if (res->state & DLM_LOCK_RES_RECOVERING) {
1376                                 /* caller should BUG */
1377                                 mlog(ML_ERROR, "node is attempting to migrate "
1378                                      "lock %.*s, but marked as recovering!\n",
1379                                      mres->lockname_len, mres->lockname);
1380                                 ret = -EFAULT;
1381                                 spin_unlock(&res->spinlock);
1382                                 goto leave;
1383                         }
1384                         res->state |= DLM_LOCK_RES_MIGRATING;
1385                 }
1386                 spin_unlock(&res->spinlock);
1387         } else {
1388                 /* need to allocate, just like if it was
1389                  * mastered here normally  */
1390                 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1391                 if (!res)
1392                         goto leave;
1393
1394                 /* to match the ref that we would have gotten if
1395                  * dlm_lookup_lockres had succeeded */
1396                 dlm_lockres_get(res);
1397
1398                 /* mark it as recovering/migrating and hash it */
1399                 if (mres->flags & DLM_MRES_RECOVERY)
1400                         res->state |= DLM_LOCK_RES_RECOVERING;
1401                 else
1402                         res->state |= DLM_LOCK_RES_MIGRATING;
1403
1404                 spin_lock(&dlm->spinlock);
1405                 __dlm_insert_lockres(dlm, res);
1406                 spin_unlock(&dlm->spinlock);
1407
1408                 /* Add an extra ref for this lock-less lockres lest the
1409                  * dlm_thread purges it before we get the chance to add
1410                  * locks to it */
1411                 dlm_lockres_get(res);
1412
1413                 /* There are three refs that need to be put.
1414                  * 1. Taken above.
1415                  * 2. kref_init in dlm_new_lockres()->dlm_init_lockres().
1416                  * 3. dlm_lookup_lockres()
1417                  * The first one is handled at the end of this function. The
1418                  * other two are handled in the worker thread after locks have
1419                  * been attached. Yes, we don't wait for purge time to match
1420                  * kref_init. The lockres will still have atleast one ref
1421                  * added because it is in the hash __dlm_insert_lockres() */
1422                 extra_refs++;
1423
1424                 /* now that the new lockres is inserted,
1425                  * make it usable by other processes */
1426                 spin_lock(&res->spinlock);
1427                 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1428                 spin_unlock(&res->spinlock);
1429                 wake_up(&res->wq);
1430         }
1431
1432         /* at this point we have allocated everything we need,
1433          * and we have a hashed lockres with an extra ref and
1434          * the proper res->state flags. */
1435         ret = 0;
1436         spin_lock(&res->spinlock);
1437         /* drop this either when master requery finds a different master
1438          * or when a lock is added by the recovery worker */
1439         dlm_lockres_grab_inflight_ref(dlm, res);
1440         if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1441                 /* migration cannot have an unknown master */
1442                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1443                 mlog(0, "recovery has passed me a lockres with an "
1444                           "unknown owner.. will need to requery: "
1445                           "%.*s\n", mres->lockname_len, mres->lockname);
1446         } else {
1447                 /* take a reference now to pin the lockres, drop it
1448                  * when locks are added in the worker */
1449                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
1450         }
1451         spin_unlock(&res->spinlock);
1452
1453         /* queue up work for dlm_mig_lockres_worker */
1454         dlm_grab(dlm);  /* get an extra ref for the work item */
1455         memcpy(buf, msg->buf, be16_to_cpu(msg->data_len));  /* copy the whole message */
1456         dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1457         item->u.ml.lockres = res; /* already have a ref */
1458         item->u.ml.real_master = real_master;
1459         item->u.ml.extra_ref = extra_refs;
1460         spin_lock(&dlm->work_lock);
1461         list_add_tail(&item->list, &dlm->work_list);
1462         spin_unlock(&dlm->work_lock);
1463         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
1464
1465 leave:
1466         /* One extra ref taken needs to be put here */
1467         if (extra_refs)
1468                 dlm_lockres_put(res);
1469
1470         dlm_put(dlm);
1471         if (ret < 0) {
1472                 if (buf)
1473                         kfree(buf);
1474                 if (item)
1475                         kfree(item);
1476         }
1477
1478         mlog_exit(ret);
1479         return ret;
1480 }
1481
1482
1483 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1484 {
1485         struct dlm_ctxt *dlm;
1486         struct dlm_migratable_lockres *mres;
1487         int ret = 0;
1488         struct dlm_lock_resource *res;
1489         u8 real_master;
1490         u8 extra_ref;
1491
1492         dlm = item->dlm;
1493         mres = (struct dlm_migratable_lockres *)data;
1494
1495         res = item->u.ml.lockres;
1496         real_master = item->u.ml.real_master;
1497         extra_ref = item->u.ml.extra_ref;
1498
1499         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1500                 /* this case is super-rare. only occurs if
1501                  * node death happens during migration. */
1502 again:
1503                 ret = dlm_lockres_master_requery(dlm, res, &real_master);
1504                 if (ret < 0) {
1505                         mlog(0, "dlm_lockres_master_requery ret=%d\n",
1506                                   ret);
1507                         goto again;
1508                 }
1509                 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1510                         mlog(0, "lockres %.*s not claimed.  "
1511                                    "this node will take it.\n",
1512                                    res->lockname.len, res->lockname.name);
1513                 } else {
1514                         spin_lock(&res->spinlock);
1515                         dlm_lockres_drop_inflight_ref(dlm, res);
1516                         spin_unlock(&res->spinlock);
1517                         mlog(0, "master needs to respond to sender "
1518                                   "that node %u still owns %.*s\n",
1519                                   real_master, res->lockname.len,
1520                                   res->lockname.name);
1521                         /* cannot touch this lockres */
1522                         goto leave;
1523                 }
1524         }
1525
1526         ret = dlm_process_recovery_data(dlm, res, mres);
1527         if (ret < 0)
1528                 mlog(0, "dlm_process_recovery_data returned  %d\n", ret);
1529         else
1530                 mlog(0, "dlm_process_recovery_data succeeded\n");
1531
1532         if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1533                            (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1534                 ret = dlm_finish_migration(dlm, res, mres->master);
1535                 if (ret < 0)
1536                         mlog_errno(ret);
1537         }
1538
1539 leave:
1540         /* See comment in dlm_mig_lockres_handler() */
1541         if (res) {
1542                 if (extra_ref)
1543                         dlm_lockres_put(res);
1544                 dlm_lockres_put(res);
1545         }
1546         kfree(data);
1547         mlog_exit(ret);
1548 }
1549
1550
1551
1552 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1553                                       struct dlm_lock_resource *res,
1554                                       u8 *real_master)
1555 {
1556         struct dlm_node_iter iter;
1557         int nodenum;
1558         int ret = 0;
1559
1560         *real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1561
1562         /* we only reach here if one of the two nodes in a
1563          * migration died while the migration was in progress.
1564          * at this point we need to requery the master.  we
1565          * know that the new_master got as far as creating
1566          * an mle on at least one node, but we do not know
1567          * if any nodes had actually cleared the mle and set
1568          * the master to the new_master.  the old master
1569          * is supposed to set the owner to UNKNOWN in the
1570          * event of a new_master death, so the only possible
1571          * responses that we can get from nodes here are
1572          * that the master is new_master, or that the master
1573          * is UNKNOWN.
1574          * if all nodes come back with UNKNOWN then we know
1575          * the lock needs remastering here.
1576          * if any node comes back with a valid master, check
1577          * to see if that master is the one that we are
1578          * recovering.  if so, then the new_master died and
1579          * we need to remaster this lock.  if not, then the
1580          * new_master survived and that node will respond to
1581          * other nodes about the owner.
1582          * if there is an owner, this node needs to dump this
1583          * lockres and alert the sender that this lockres
1584          * was rejected. */
1585         spin_lock(&dlm->spinlock);
1586         dlm_node_iter_init(dlm->domain_map, &iter);
1587         spin_unlock(&dlm->spinlock);
1588
1589         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1590                 /* do not send to self */
1591                 if (nodenum == dlm->node_num)
1592                         continue;
1593                 ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1594                 if (ret < 0) {
1595                         mlog_errno(ret);
1596                         if (!dlm_is_host_down(ret))
1597                                 BUG();
1598                         /* host is down, so answer for that node would be
1599                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
1600                 }
1601                 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1602                         mlog(0, "lock master is %u\n", *real_master);
1603                         break;
1604                 }
1605         }
1606         return ret;
1607 }
1608
1609
1610 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1611                           u8 nodenum, u8 *real_master)
1612 {
1613         int ret = -EINVAL;
1614         struct dlm_master_requery req;
1615         int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1616
1617         memset(&req, 0, sizeof(req));
1618         req.node_idx = dlm->node_num;
1619         req.namelen = res->lockname.len;
1620         memcpy(req.name, res->lockname.name, res->lockname.len);
1621
1622         ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1623                                  &req, sizeof(req), nodenum, &status);
1624         /* XXX: negative status not handled properly here. */
1625         if (ret < 0)
1626                 mlog_errno(ret);
1627         else {
1628                 BUG_ON(status < 0);
1629                 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1630                 *real_master = (u8) (status & 0xff);
1631                 mlog(0, "node %u responded to master requery with %u\n",
1632                           nodenum, *real_master);
1633                 ret = 0;
1634         }
1635         return ret;
1636 }
1637
1638
1639 /* this function cannot error, so unless the sending
1640  * or receiving of the message failed, the owner can
1641  * be trusted */
1642 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
1643                                void **ret_data)
1644 {
1645         struct dlm_ctxt *dlm = data;
1646         struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1647         struct dlm_lock_resource *res = NULL;
1648         unsigned int hash;
1649         int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1650         u32 flags = DLM_ASSERT_MASTER_REQUERY;
1651
1652         if (!dlm_grab(dlm)) {
1653                 /* since the domain has gone away on this
1654                  * node, the proper response is UNKNOWN */
1655                 return master;
1656         }
1657
1658         hash = dlm_lockid_hash(req->name, req->namelen);
1659
1660         spin_lock(&dlm->spinlock);
1661         res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash);
1662         if (res) {
1663                 spin_lock(&res->spinlock);
1664                 master = res->owner;
1665                 if (master == dlm->node_num) {
1666                         int ret = dlm_dispatch_assert_master(dlm, res,
1667                                                              0, 0, flags);
1668                         if (ret < 0) {
1669                                 mlog_errno(-ENOMEM);
1670                                 /* retry!? */
1671                                 BUG();
1672                         }
1673                 } else /* put.. incase we are not the master */
1674                         dlm_lockres_put(res);
1675                 spin_unlock(&res->spinlock);
1676         }
1677         spin_unlock(&dlm->spinlock);
1678
1679         dlm_put(dlm);
1680         return master;
1681 }
1682
1683 static inline struct list_head *
1684 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1685 {
1686         struct list_head *ret;
1687         BUG_ON(list_num < 0);
1688         BUG_ON(list_num > 2);
1689         ret = &(res->granted);
1690         ret += list_num;
1691         return ret;
1692 }
1693 /* TODO: do ast flush business
1694  * TODO: do MIGRATING and RECOVERING spinning
1695  */
1696
1697 /*
1698 * NOTE about in-flight requests during migration:
1699 *
1700 * Before attempting the migrate, the master has marked the lockres as
1701 * MIGRATING and then flushed all of its pending ASTS.  So any in-flight
1702 * requests either got queued before the MIGRATING flag got set, in which
1703 * case the lock data will reflect the change and a return message is on
1704 * the way, or the request failed to get in before MIGRATING got set.  In
1705 * this case, the caller will be told to spin and wait for the MIGRATING
1706 * flag to be dropped, then recheck the master.
1707 * This holds true for the convert, cancel and unlock cases, and since lvb
1708 * updates are tied to these same messages, it applies to lvb updates as
1709 * well.  For the lock case, there is no way a lock can be on the master
1710 * queue and not be on the secondary queue since the lock is always added
1711 * locally first.  This means that the new target node will never be sent
1712 * a lock that he doesn't already have on the list.
1713 * In total, this means that the local lock is correct and should not be
1714 * updated to match the one sent by the master.  Any messages sent back
1715 * from the master before the MIGRATING flag will bring the lock properly
1716 * up-to-date, and the change will be ordered properly for the waiter.
1717 * We will *not* attempt to modify the lock underneath the waiter.
1718 */
1719
1720 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1721                                      struct dlm_lock_resource *res,
1722                                      struct dlm_migratable_lockres *mres)
1723 {
1724         struct dlm_migratable_lock *ml;
1725         struct list_head *queue;
1726         struct list_head *tmpq = NULL;
1727         struct dlm_lock *newlock = NULL;
1728         struct dlm_lockstatus *lksb = NULL;
1729         int ret = 0;
1730         int i, j, bad;
1731         struct dlm_lock *lock = NULL;
1732         u8 from = O2NM_MAX_NODES;
1733         unsigned int added = 0;
1734
1735         mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1736         for (i=0; i<mres->num_locks; i++) {
1737                 ml = &(mres->ml[i]);
1738
1739                 if (dlm_is_dummy_lock(dlm, ml, &from)) {
1740                         /* placeholder, just need to set the refmap bit */
1741                         BUG_ON(mres->num_locks != 1);
1742                         mlog(0, "%s:%.*s: dummy lock for %u\n",
1743                              dlm->name, mres->lockname_len, mres->lockname,
1744                              from);
1745                         spin_lock(&res->spinlock);
1746                         dlm_lockres_set_refmap_bit(from, res);
1747                         spin_unlock(&res->spinlock);
1748                         added++;
1749                         break;
1750                 }
1751                 BUG_ON(ml->highest_blocked != LKM_IVMODE);
1752                 newlock = NULL;
1753                 lksb = NULL;
1754
1755                 queue = dlm_list_num_to_pointer(res, ml->list);
1756                 tmpq = NULL;
1757
1758                 /* if the lock is for the local node it needs to
1759                  * be moved to the proper location within the queue.
1760                  * do not allocate a new lock structure. */
1761                 if (ml->node == dlm->node_num) {
1762                         /* MIGRATION ONLY! */
1763                         BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1764
1765                         spin_lock(&res->spinlock);
1766                         for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) {
1767                                 tmpq = dlm_list_idx_to_ptr(res, j);
1768                                 list_for_each_entry(lock, tmpq, list) {
1769                                         if (lock->ml.cookie != ml->cookie)
1770                                                 lock = NULL;
1771                                         else
1772                                                 break;
1773                                 }
1774                                 if (lock)
1775                                         break;
1776                         }
1777
1778                         /* lock is always created locally first, and
1779                          * destroyed locally last.  it must be on the list */
1780                         if (!lock) {
1781                                 __be64 c = ml->cookie;
1782                                 mlog(ML_ERROR, "could not find local lock "
1783                                                "with cookie %u:%llu!\n",
1784                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1785                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)));
1786                                 __dlm_print_one_lock_resource(res);
1787                                 BUG();
1788                         }
1789                         BUG_ON(lock->ml.node != ml->node);
1790
1791                         if (tmpq != queue) {
1792                                 mlog(0, "lock was on %u instead of %u for %.*s\n",
1793                                      j, ml->list, res->lockname.len, res->lockname.name);
1794                                 spin_unlock(&res->spinlock);
1795                                 continue;
1796                         }
1797
1798                         /* see NOTE above about why we do not update
1799                          * to match the master here */
1800
1801                         /* move the lock to its proper place */
1802                         /* do not alter lock refcount.  switching lists. */
1803                         list_move_tail(&lock->list, queue);
1804                         spin_unlock(&res->spinlock);
1805                         added++;
1806
1807                         mlog(0, "just reordered a local lock!\n");
1808                         continue;
1809                 }
1810
1811                 /* lock is for another node. */
1812                 newlock = dlm_new_lock(ml->type, ml->node,
1813                                        be64_to_cpu(ml->cookie), NULL);
1814                 if (!newlock) {
1815                         ret = -ENOMEM;
1816                         goto leave;
1817                 }
1818                 lksb = newlock->lksb;
1819                 dlm_lock_attach_lockres(newlock, res);
1820
1821                 if (ml->convert_type != LKM_IVMODE) {
1822                         BUG_ON(queue != &res->converting);
1823                         newlock->ml.convert_type = ml->convert_type;
1824                 }
1825                 lksb->flags |= (ml->flags &
1826                                 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1827
1828                 if (ml->type == LKM_NLMODE)
1829                         goto skip_lvb;
1830
1831                 if (!dlm_lvb_is_empty(mres->lvb)) {
1832                         if (lksb->flags & DLM_LKSB_PUT_LVB) {
1833                                 /* other node was trying to update
1834                                  * lvb when node died.  recreate the
1835                                  * lksb with the updated lvb. */
1836                                 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1837                                 /* the lock resource lvb update must happen
1838                                  * NOW, before the spinlock is dropped.
1839                                  * we no longer wait for the AST to update
1840                                  * the lvb. */
1841                                 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1842                         } else {
1843                                 /* otherwise, the node is sending its 
1844                                  * most recent valid lvb info */
1845                                 BUG_ON(ml->type != LKM_EXMODE &&
1846                                        ml->type != LKM_PRMODE);
1847                                 if (!dlm_lvb_is_empty(res->lvb) &&
1848                                     (ml->type == LKM_EXMODE ||
1849                                      memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1850                                         int i;
1851                                         mlog(ML_ERROR, "%s:%.*s: received bad "
1852                                              "lvb! type=%d\n", dlm->name,
1853                                              res->lockname.len,
1854                                              res->lockname.name, ml->type);
1855                                         printk("lockres lvb=[");
1856                                         for (i=0; i<DLM_LVB_LEN; i++)
1857                                                 printk("%02x", res->lvb[i]);
1858                                         printk("]\nmigrated lvb=[");
1859                                         for (i=0; i<DLM_LVB_LEN; i++)
1860                                                 printk("%02x", mres->lvb[i]);
1861                                         printk("]\n");
1862                                         dlm_print_one_lock_resource(res);
1863                                         BUG();
1864                                 }
1865                                 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1866                         }
1867                 }
1868 skip_lvb:
1869
1870                 /* NOTE:
1871                  * wrt lock queue ordering and recovery:
1872                  *    1. order of locks on granted queue is
1873                  *       meaningless.
1874                  *    2. order of locks on converting queue is
1875                  *       LOST with the node death.  sorry charlie.
1876                  *    3. order of locks on the blocked queue is
1877                  *       also LOST.
1878                  * order of locks does not affect integrity, it
1879                  * just means that a lock request may get pushed
1880                  * back in line as a result of the node death.
1881                  * also note that for a given node the lock order
1882                  * for its secondary queue locks is preserved
1883                  * relative to each other, but clearly *not*
1884                  * preserved relative to locks from other nodes.
1885                  */
1886                 bad = 0;
1887                 spin_lock(&res->spinlock);
1888                 list_for_each_entry(lock, queue, list) {
1889                         if (lock->ml.cookie == ml->cookie) {
1890                                 __be64 c = lock->ml.cookie;
1891                                 mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already "
1892                                      "exists on this lockres!\n", dlm->name,
1893                                      res->lockname.len, res->lockname.name,
1894                                      dlm_get_lock_cookie_node(be64_to_cpu(c)),
1895                                      dlm_get_lock_cookie_seq(be64_to_cpu(c)));
1896
1897                                 mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, "
1898                                      "node=%u, cookie=%u:%llu, queue=%d\n",
1899                                      ml->type, ml->convert_type, ml->node,
1900                                      dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)),
1901                                      dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)),
1902                                      ml->list);
1903
1904                                 __dlm_print_one_lock_resource(res);
1905                                 bad = 1;
1906                                 break;
1907                         }
1908                 }
1909                 if (!bad) {
1910                         dlm_lock_get(newlock);
1911                         list_add_tail(&newlock->list, queue);
1912                         mlog(0, "%s:%.*s: added lock for node %u, "
1913                              "setting refmap bit\n", dlm->name,
1914                              res->lockname.len, res->lockname.name, ml->node);
1915                         dlm_lockres_set_refmap_bit(ml->node, res);
1916                         added++;
1917                 }
1918                 spin_unlock(&res->spinlock);
1919         }
1920         mlog(0, "done running all the locks\n");
1921
1922 leave:
1923         /* balance the ref taken when the work was queued */
1924         spin_lock(&res->spinlock);
1925         dlm_lockres_drop_inflight_ref(dlm, res);
1926         spin_unlock(&res->spinlock);
1927
1928         if (ret < 0) {
1929                 mlog_errno(ret);
1930                 if (newlock)
1931                         dlm_lock_put(newlock);
1932         }
1933
1934         mlog_exit(ret);
1935         return ret;
1936 }
1937
1938 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
1939                                        struct dlm_lock_resource *res)
1940 {
1941         int i;
1942         struct list_head *queue;
1943         struct dlm_lock *lock, *next;
1944
1945         res->state |= DLM_LOCK_RES_RECOVERING;
1946         if (!list_empty(&res->recovering)) {
1947                 mlog(0,
1948                      "Recovering res %s:%.*s, is already on recovery list!\n",
1949                      dlm->name, res->lockname.len, res->lockname.name);
1950                 list_del_init(&res->recovering);
1951                 dlm_lockres_put(res);
1952         }
1953         /* We need to hold a reference while on the recovery list */
1954         dlm_lockres_get(res);
1955         list_add_tail(&res->recovering, &dlm->reco.resources);
1956
1957         /* find any pending locks and put them back on proper list */
1958         for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
1959                 queue = dlm_list_idx_to_ptr(res, i);
1960                 list_for_each_entry_safe(lock, next, queue, list) {
1961                         dlm_lock_get(lock);
1962                         if (lock->convert_pending) {
1963                                 /* move converting lock back to granted */
1964                                 BUG_ON(i != DLM_CONVERTING_LIST);
1965                                 mlog(0, "node died with convert pending "
1966                                      "on %.*s. move back to granted list.\n",
1967                                      res->lockname.len, res->lockname.name);
1968                                 dlm_revert_pending_convert(res, lock);
1969                                 lock->convert_pending = 0;
1970                         } else if (lock->lock_pending) {
1971                                 /* remove pending lock requests completely */
1972                                 BUG_ON(i != DLM_BLOCKED_LIST);
1973                                 mlog(0, "node died with lock pending "
1974                                      "on %.*s. remove from blocked list and skip.\n",
1975                                      res->lockname.len, res->lockname.name);
1976                                 /* lock will be floating until ref in
1977                                  * dlmlock_remote is freed after the network
1978                                  * call returns.  ok for it to not be on any
1979                                  * list since no ast can be called
1980                                  * (the master is dead). */
1981                                 dlm_revert_pending_lock(res, lock);
1982                                 lock->lock_pending = 0;
1983                         } else if (lock->unlock_pending) {
1984                                 /* if an unlock was in progress, treat as
1985                                  * if this had completed successfully
1986                                  * before sending this lock state to the
1987                                  * new master.  note that the dlm_unlock
1988                                  * call is still responsible for calling
1989                                  * the unlockast.  that will happen after
1990                                  * the network call times out.  for now,
1991                                  * just move lists to prepare the new
1992                                  * recovery master.  */
1993                                 BUG_ON(i != DLM_GRANTED_LIST);
1994                                 mlog(0, "node died with unlock pending "
1995                                      "on %.*s. remove from blocked list and skip.\n",
1996                                      res->lockname.len, res->lockname.name);
1997                                 dlm_commit_pending_unlock(res, lock);
1998                                 lock->unlock_pending = 0;
1999                         } else if (lock->cancel_pending) {
2000                                 /* if a cancel was in progress, treat as
2001                                  * if this had completed successfully
2002                                  * before sending this lock state to the
2003                                  * new master */
2004                                 BUG_ON(i != DLM_CONVERTING_LIST);
2005                                 mlog(0, "node died with cancel pending "
2006                                      "on %.*s. move back to granted list.\n",
2007                                      res->lockname.len, res->lockname.name);
2008                                 dlm_commit_pending_cancel(res, lock);
2009                                 lock->cancel_pending = 0;
2010                         }
2011                         dlm_lock_put(lock);
2012                 }
2013         }
2014 }
2015
2016
2017
2018 /* removes all recovered locks from the recovery list.
2019  * sets the res->owner to the new master.
2020  * unsets the RECOVERY flag and wakes waiters. */
2021 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
2022                                               u8 dead_node, u8 new_master)
2023 {
2024         int i;
2025         struct hlist_node *hash_iter;
2026         struct hlist_head *bucket;
2027         struct dlm_lock_resource *res, *next;
2028
2029         mlog_entry_void();
2030
2031         assert_spin_locked(&dlm->spinlock);
2032
2033         list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
2034                 if (res->owner == dead_node) {
2035                         list_del_init(&res->recovering);
2036                         spin_lock(&res->spinlock);
2037                         /* new_master has our reference from
2038                          * the lock state sent during recovery */
2039                         dlm_change_lockres_owner(dlm, res, new_master);
2040                         res->state &= ~DLM_LOCK_RES_RECOVERING;
2041                         if (__dlm_lockres_has_locks(res))
2042                                 __dlm_dirty_lockres(dlm, res);
2043                         spin_unlock(&res->spinlock);
2044                         wake_up(&res->wq);
2045                         dlm_lockres_put(res);
2046                 }
2047         }
2048
2049         /* this will become unnecessary eventually, but
2050          * for now we need to run the whole hash, clear
2051          * the RECOVERING state and set the owner
2052          * if necessary */
2053         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2054                 bucket = dlm_lockres_hash(dlm, i);
2055                 hlist_for_each_entry(res, hash_iter, bucket, hash_node) {
2056                         if (res->state & DLM_LOCK_RES_RECOVERING) {
2057                                 if (res->owner == dead_node) {
2058                                         mlog(0, "(this=%u) res %.*s owner=%u "
2059                                              "was not on recovering list, but "
2060                                              "clearing state anyway\n",
2061                                              dlm->node_num, res->lockname.len,
2062                                              res->lockname.name, new_master);
2063                                 } else if (res->owner == dlm->node_num) {
2064                                         mlog(0, "(this=%u) res %.*s owner=%u "
2065                                              "was not on recovering list, "
2066                                              "owner is THIS node, clearing\n",
2067                                              dlm->node_num, res->lockname.len,
2068                                              res->lockname.name, new_master);
2069                                 } else
2070                                         continue;
2071
2072                                 if (!list_empty(&res->recovering)) {
2073                                         mlog(0, "%s:%.*s: lockres was "
2074                                              "marked RECOVERING, owner=%u\n",
2075                                              dlm->name, res->lockname.len,
2076                                              res->lockname.name, res->owner);
2077                                         list_del_init(&res->recovering);
2078                                         dlm_lockres_put(res);
2079                                 }
2080                                 spin_lock(&res->spinlock);
2081                                 /* new_master has our reference from
2082                                  * the lock state sent during recovery */
2083                                 dlm_change_lockres_owner(dlm, res, new_master);
2084                                 res->state &= ~DLM_LOCK_RES_RECOVERING;
2085                                 if (__dlm_lockres_has_locks(res))
2086                                         __dlm_dirty_lockres(dlm, res);
2087                                 spin_unlock(&res->spinlock);
2088                                 wake_up(&res->wq);
2089                         }
2090                 }
2091         }
2092 }
2093
2094 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
2095 {
2096         if (local) {
2097                 if (lock->ml.type != LKM_EXMODE &&
2098                     lock->ml.type != LKM_PRMODE)
2099                         return 1;
2100         } else if (lock->ml.type == LKM_EXMODE)
2101                 return 1;
2102         return 0;
2103 }
2104
2105 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
2106                                struct dlm_lock_resource *res, u8 dead_node)
2107 {
2108         struct list_head *queue;
2109         struct dlm_lock *lock;
2110         int blank_lvb = 0, local = 0;
2111         int i;
2112         u8 search_node;
2113
2114         assert_spin_locked(&dlm->spinlock);
2115         assert_spin_locked(&res->spinlock);
2116
2117         if (res->owner == dlm->node_num)
2118                 /* if this node owned the lockres, and if the dead node 
2119                  * had an EX when he died, blank out the lvb */
2120                 search_node = dead_node;
2121         else {
2122                 /* if this is a secondary lockres, and we had no EX or PR
2123                  * locks granted, we can no longer trust the lvb */
2124                 search_node = dlm->node_num;
2125                 local = 1;  /* check local state for valid lvb */
2126         }
2127
2128         for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
2129                 queue = dlm_list_idx_to_ptr(res, i);
2130                 list_for_each_entry(lock, queue, list) {
2131                         if (lock->ml.node == search_node) {
2132                                 if (dlm_lvb_needs_invalidation(lock, local)) {
2133                                         /* zero the lksb lvb and lockres lvb */
2134                                         blank_lvb = 1;
2135                                         memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
2136                                 }
2137                         }
2138                 }
2139         }
2140
2141         if (blank_lvb) {
2142                 mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
2143                      res->lockname.len, res->lockname.name, dead_node);
2144                 memset(res->lvb, 0, DLM_LVB_LEN);
2145         }
2146 }
2147
2148 static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
2149                                 struct dlm_lock_resource *res, u8 dead_node)
2150 {
2151         struct dlm_lock *lock, *next;
2152         unsigned int freed = 0;
2153
2154         /* this node is the lockres master:
2155          * 1) remove any stale locks for the dead node
2156          * 2) if the dead node had an EX when he died, blank out the lvb 
2157          */
2158         assert_spin_locked(&dlm->spinlock);
2159         assert_spin_locked(&res->spinlock);
2160
2161         /* We do two dlm_lock_put(). One for removing from list and the other is
2162          * to force the DLM_UNLOCK_FREE_LOCK action so as to free the locks */
2163
2164         /* TODO: check pending_asts, pending_basts here */
2165         list_for_each_entry_safe(lock, next, &res->granted, list) {
2166                 if (lock->ml.node == dead_node) {
2167                         list_del_init(&lock->list);
2168                         dlm_lock_put(lock);
2169                         /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2170                         dlm_lock_put(lock);
2171                         freed++;
2172                 }
2173         }
2174         list_for_each_entry_safe(lock, next, &res->converting, list) {
2175                 if (lock->ml.node == dead_node) {
2176                         list_del_init(&lock->list);
2177                         dlm_lock_put(lock);
2178                         /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2179                         dlm_lock_put(lock);
2180                         freed++;
2181                 }
2182         }
2183         list_for_each_entry_safe(lock, next, &res->blocked, list) {
2184                 if (lock->ml.node == dead_node) {
2185                         list_del_init(&lock->list);
2186                         dlm_lock_put(lock);
2187                         /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2188                         dlm_lock_put(lock);
2189                         freed++;
2190                 }
2191         }
2192
2193         if (freed) {
2194                 mlog(0, "%s:%.*s: freed %u locks for dead node %u, "
2195                      "dropping ref from lockres\n", dlm->name,
2196                      res->lockname.len, res->lockname.name, freed, dead_node);
2197                 BUG_ON(!test_bit(dead_node, res->refmap));
2198                 dlm_lockres_clear_refmap_bit(dead_node, res);
2199         } else if (test_bit(dead_node, res->refmap)) {
2200                 mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2201                      "no locks and had not purged before dying\n", dlm->name,
2202                      res->lockname.len, res->lockname.name, dead_node);
2203                 dlm_lockres_clear_refmap_bit(dead_node, res);
2204         }
2205
2206         /* do not kick thread yet */
2207         __dlm_dirty_lockres(dlm, res);
2208 }
2209
2210 /* if this node is the recovery master, and there are no
2211  * locks for a given lockres owned by this node that are in
2212  * either PR or EX mode, zero out the lvb before requesting.
2213  *
2214  */
2215
2216
2217 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
2218 {
2219         struct hlist_node *iter;
2220         struct dlm_lock_resource *res;
2221         int i;
2222         struct hlist_head *bucket;
2223         struct dlm_lock *lock;
2224
2225
2226         /* purge any stale mles */
2227         dlm_clean_master_list(dlm, dead_node);
2228
2229         /*
2230          * now clean up all lock resources.  there are two rules:
2231          *
2232          * 1) if the dead node was the master, move the lockres
2233          *    to the recovering list.  set the RECOVERING flag.
2234          *    this lockres needs to be cleaned up before it can
2235          *    be used further.
2236          *
2237          * 2) if this node was the master, remove all locks from
2238          *    each of the lockres queues that were owned by the
2239          *    dead node.  once recovery finishes, the dlm thread
2240          *    can be kicked again to see if any ASTs or BASTs
2241          *    need to be fired as a result.
2242          */
2243         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2244                 bucket = dlm_lockres_hash(dlm, i);
2245                 hlist_for_each_entry(res, iter, bucket, hash_node) {
2246                         /* always prune any $RECOVERY entries for dead nodes,
2247                          * otherwise hangs can occur during later recovery */
2248                         if (dlm_is_recovery_lock(res->lockname.name,
2249                                                  res->lockname.len)) {
2250                                 spin_lock(&res->spinlock);
2251                                 list_for_each_entry(lock, &res->granted, list) {
2252                                         if (lock->ml.node == dead_node) {
2253                                                 mlog(0, "AHA! there was "
2254                                                      "a $RECOVERY lock for dead "
2255                                                      "node %u (%s)!\n",
2256                                                      dead_node, dlm->name);
2257                                                 list_del_init(&lock->list);
2258                                                 dlm_lock_put(lock);
2259                                                 break;
2260                                         }
2261                                 }
2262                                 spin_unlock(&res->spinlock);
2263                                 continue;
2264                         }                       
2265                         spin_lock(&res->spinlock);
2266                         /* zero the lvb if necessary */
2267                         dlm_revalidate_lvb(dlm, res, dead_node);
2268                         if (res->owner == dead_node) {
2269                                 if (res->state & DLM_LOCK_RES_DROPPING_REF)
2270                                         mlog(0, "%s:%.*s: owned by "
2271                                              "dead node %u, this node was "
2272                                              "dropping its ref when it died. "
2273                                              "continue, dropping the flag.\n",
2274                                              dlm->name, res->lockname.len,
2275                                              res->lockname.name, dead_node);
2276
2277                                 /* the wake_up for this will happen when the
2278                                  * RECOVERING flag is dropped later */
2279                                 res->state &= ~DLM_LOCK_RES_DROPPING_REF;
2280
2281                                 dlm_move_lockres_to_recovery_list(dlm, res);
2282                         } else if (res->owner == dlm->node_num) {
2283                                 dlm_free_dead_locks(dlm, res, dead_node);
2284                                 __dlm_lockres_calc_usage(dlm, res);
2285                         }
2286                         spin_unlock(&res->spinlock);
2287                 }
2288         }
2289
2290 }
2291
2292 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
2293 {
2294         assert_spin_locked(&dlm->spinlock);
2295
2296         if (dlm->reco.new_master == idx) {
2297                 mlog(0, "%s: recovery master %d just died\n",
2298                      dlm->name, idx);
2299                 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2300                         /* finalize1 was reached, so it is safe to clear
2301                          * the new_master and dead_node.  that recovery
2302                          * is complete. */
2303                         mlog(0, "%s: dead master %d had reached "
2304                              "finalize1 state, clearing\n", dlm->name, idx);
2305                         dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2306                         __dlm_reset_recovery(dlm);
2307                 }
2308         }
2309
2310         /* Clean up join state on node death. */
2311         if (dlm->joining_node == idx) {
2312                 mlog(0, "Clearing join state for node %u\n", idx);
2313                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
2314         }
2315
2316         /* check to see if the node is already considered dead */
2317         if (!test_bit(idx, dlm->live_nodes_map)) {
2318                 mlog(0, "for domain %s, node %d is already dead. "
2319                      "another node likely did recovery already.\n",
2320                      dlm->name, idx);
2321                 return;
2322         }
2323
2324         /* check to see if we do not care about this node */
2325         if (!test_bit(idx, dlm->domain_map)) {
2326                 /* This also catches the case that we get a node down
2327                  * but haven't joined the domain yet. */
2328                 mlog(0, "node %u already removed from domain!\n", idx);
2329                 return;
2330         }
2331
2332         clear_bit(idx, dlm->live_nodes_map);
2333
2334         /* make sure local cleanup occurs before the heartbeat events */
2335         if (!test_bit(idx, dlm->recovery_map))
2336                 dlm_do_local_recovery_cleanup(dlm, idx);
2337
2338         /* notify anything attached to the heartbeat events */
2339         dlm_hb_event_notify_attached(dlm, idx, 0);
2340
2341         mlog(0, "node %u being removed from domain map!\n", idx);
2342         clear_bit(idx, dlm->domain_map);
2343         /* wake up migration waiters if a node goes down.
2344          * perhaps later we can genericize this for other waiters. */
2345         wake_up(&dlm->migration_wq);
2346
2347         if (test_bit(idx, dlm->recovery_map))
2348                 mlog(0, "domain %s, node %u already added "
2349                      "to recovery map!\n", dlm->name, idx);
2350         else
2351                 set_bit(idx, dlm->recovery_map);
2352 }
2353
2354 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
2355 {
2356         struct dlm_ctxt *dlm = data;
2357
2358         if (!dlm_grab(dlm))
2359                 return;
2360
2361         /*
2362          * This will notify any dlm users that a node in our domain
2363          * went away without notifying us first.
2364          */
2365         if (test_bit(idx, dlm->domain_map))
2366                 dlm_fire_domain_eviction_callbacks(dlm, idx);
2367
2368         spin_lock(&dlm->spinlock);
2369         __dlm_hb_node_down(dlm, idx);
2370         spin_unlock(&dlm->spinlock);
2371
2372         dlm_put(dlm);
2373 }
2374
2375 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
2376 {
2377         struct dlm_ctxt *dlm = data;
2378
2379         if (!dlm_grab(dlm))
2380                 return;
2381
2382         spin_lock(&dlm->spinlock);
2383         set_bit(idx, dlm->live_nodes_map);
2384         /* do NOT notify mle attached to the heartbeat events.
2385          * new nodes are not interesting in mastery until joined. */
2386         spin_unlock(&dlm->spinlock);
2387
2388         dlm_put(dlm);
2389 }
2390
2391 static void dlm_reco_ast(void *astdata)
2392 {
2393         struct dlm_ctxt *dlm = astdata;
2394         mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2395              dlm->node_num, dlm->name);
2396 }
2397 static void dlm_reco_bast(void *astdata, int blocked_type)
2398 {
2399         struct dlm_ctxt *dlm = astdata;
2400         mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2401              dlm->node_num, dlm->name);
2402 }
2403 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
2404 {
2405         mlog(0, "unlockast for recovery lock fired!\n");
2406 }
2407
2408 /*
2409  * dlm_pick_recovery_master will continually attempt to use
2410  * dlmlock() on the special "$RECOVERY" lockres with the
2411  * LKM_NOQUEUE flag to get an EX.  every thread that enters
2412  * this function on each node racing to become the recovery
2413  * master will not stop attempting this until either:
2414  * a) this node gets the EX (and becomes the recovery master),
2415  * or b) dlm->reco.new_master gets set to some nodenum 
2416  * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2417  * so each time a recovery master is needed, the entire cluster
2418  * will sync at this point.  if the new master dies, that will
2419  * be detected in dlm_do_recovery */
2420 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2421 {
2422         enum dlm_status ret;
2423         struct dlm_lockstatus lksb;
2424         int status = -EINVAL;
2425
2426         mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2427              dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2428 again:  
2429         memset(&lksb, 0, sizeof(lksb));
2430
2431         ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
2432                       DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN,
2433                       dlm_reco_ast, dlm, dlm_reco_bast);
2434
2435         mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2436              dlm->name, ret, lksb.status);
2437
2438         if (ret == DLM_NORMAL) {
2439                 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2440                      dlm->name, dlm->node_num);
2441                 
2442                 /* got the EX lock.  check to see if another node 
2443                  * just became the reco master */
2444                 if (dlm_reco_master_ready(dlm)) {
2445                         mlog(0, "%s: got reco EX lock, but %u will "
2446                              "do the recovery\n", dlm->name,
2447                              dlm->reco.new_master);
2448                         status = -EEXIST;
2449                 } else {
2450                         status = 0;
2451
2452                         /* see if recovery was already finished elsewhere */
2453                         spin_lock(&dlm->spinlock);
2454                         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
2455                                 status = -EINVAL;       
2456                                 mlog(0, "%s: got reco EX lock, but "
2457                                      "node got recovered already\n", dlm->name);
2458                                 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2459                                         mlog(ML_ERROR, "%s: new master is %u "
2460                                              "but no dead node!\n", 
2461                                              dlm->name, dlm->reco.new_master);
2462                                         BUG();
2463                                 }
2464                         }
2465                         spin_unlock(&dlm->spinlock);
2466                 }
2467
2468                 /* if this node has actually become the recovery master,
2469                  * set the master and send the messages to begin recovery */
2470                 if (!status) {
2471                         mlog(0, "%s: dead=%u, this=%u, sending "
2472                              "begin_reco now\n", dlm->name, 
2473                              dlm->reco.dead_node, dlm->node_num);
2474                         status = dlm_send_begin_reco_message(dlm,
2475                                       dlm->reco.dead_node);
2476                         /* this always succeeds */
2477                         BUG_ON(status);
2478
2479                         /* set the new_master to this node */
2480                         spin_lock(&dlm->spinlock);
2481                         dlm_set_reco_master(dlm, dlm->node_num);
2482                         spin_unlock(&dlm->spinlock);
2483                 }
2484
2485                 /* recovery lock is a special case.  ast will not get fired,
2486                  * so just go ahead and unlock it. */
2487                 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
2488                 if (ret == DLM_DENIED) {
2489                         mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2490                         ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2491                 }
2492                 if (ret != DLM_NORMAL) {
2493                         /* this would really suck. this could only happen
2494                          * if there was a network error during the unlock
2495                          * because of node death.  this means the unlock
2496                          * is actually "done" and the lock structure is
2497                          * even freed.  we can continue, but only
2498                          * because this specific lock name is special. */
2499                         mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
2500                 }
2501         } else if (ret == DLM_NOTQUEUED) {
2502                 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2503                      dlm->name, dlm->node_num);
2504                 /* another node is master. wait on
2505                  * reco.new_master != O2NM_INVALID_NODE_NUM 
2506                  * for at most one second */
2507                 wait_event_timeout(dlm->dlm_reco_thread_wq,
2508                                          dlm_reco_master_ready(dlm),
2509                                          msecs_to_jiffies(1000));
2510                 if (!dlm_reco_master_ready(dlm)) {
2511                         mlog(0, "%s: reco master taking awhile\n",
2512                              dlm->name);
2513                         goto again;
2514                 }
2515                 /* another node has informed this one that it is reco master */
2516                 mlog(0, "%s: reco master %u is ready to recover %u\n",
2517                      dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
2518                 status = -EEXIST;
2519         } else if (ret == DLM_RECOVERING) {
2520                 mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n",
2521                      dlm->name, dlm->node_num);
2522                 goto again;
2523         } else {
2524                 struct dlm_lock_resource *res;
2525
2526                 /* dlmlock returned something other than NOTQUEUED or NORMAL */
2527                 mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2528                      "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2529                      dlm_errname(lksb.status));
2530                 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2531                                          DLM_RECOVERY_LOCK_NAME_LEN);
2532                 if (res) {
2533                         dlm_print_one_lock_resource(res);
2534                         dlm_lockres_put(res);
2535                 } else {
2536                         mlog(ML_ERROR, "recovery lock not found\n");
2537                 }
2538                 BUG();
2539         }
2540
2541         return status;
2542 }
2543
2544 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2545 {
2546         struct dlm_begin_reco br;
2547         int ret = 0;
2548         struct dlm_node_iter iter;
2549         int nodenum;
2550         int status;
2551
2552         mlog_entry("%u\n", dead_node);
2553
2554         mlog(0, "%s: dead node is %u\n", dlm->name, dead_node);
2555
2556         spin_lock(&dlm->spinlock);
2557         dlm_node_iter_init(dlm->domain_map, &iter);
2558         spin_unlock(&dlm->spinlock);
2559
2560         clear_bit(dead_node, iter.node_map);
2561
2562         memset(&br, 0, sizeof(br));
2563         br.node_idx = dlm->node_num;
2564         br.dead_node = dead_node;
2565
2566         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2567                 ret = 0;
2568                 if (nodenum == dead_node) {
2569                         mlog(0, "not sending begin reco to dead node "
2570                                   "%u\n", dead_node);
2571                         continue;
2572                 }
2573                 if (nodenum == dlm->node_num) {
2574                         mlog(0, "not sending begin reco to self\n");
2575                         continue;
2576                 }
2577 retry:
2578                 ret = -EINVAL;
2579                 mlog(0, "attempting to send begin reco msg to %d\n",
2580                           nodenum);
2581                 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2582                                          &br, sizeof(br), nodenum, &status);
2583                 /* negative status is handled ok by caller here */
2584                 if (ret >= 0)
2585                         ret = status;
2586                 if (dlm_is_host_down(ret)) {
2587                         /* node is down.  not involved in recovery
2588                          * so just keep going */
2589                         mlog(0, "%s: node %u was down when sending "
2590                              "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2591                         ret = 0;
2592                 }
2593                 if (ret < 0) {
2594                         struct dlm_lock_resource *res;
2595                         /* this is now a serious problem, possibly ENOMEM 
2596                          * in the network stack.  must retry */
2597                         mlog_errno(ret);
2598                         mlog(ML_ERROR, "begin reco of dlm %s to node %u "
2599                             " returned %d\n", dlm->name, nodenum, ret);
2600                         res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2601                                                  DLM_RECOVERY_LOCK_NAME_LEN);
2602                         if (res) {
2603                                 dlm_print_one_lock_resource(res);
2604                                 dlm_lockres_put(res);
2605                         } else {
2606                                 mlog(ML_ERROR, "recovery lock not found\n");
2607                         }
2608                         /* sleep for a bit in hopes that we can avoid 
2609                          * another ENOMEM */
2610                         msleep(100);
2611                         goto retry;
2612                 } else if (ret == EAGAIN) {
2613                         mlog(0, "%s: trying to start recovery of node "
2614                              "%u, but node %u is waiting for last recovery "
2615                              "to complete, backoff for a bit\n", dlm->name,
2616                              dead_node, nodenum);
2617                         /* TODO Look into replacing msleep with cond_resched() */
2618                         msleep(100);
2619                         goto retry;
2620                 }
2621         }
2622
2623         return ret;
2624 }
2625
2626 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2627                            void **ret_data)
2628 {
2629         struct dlm_ctxt *dlm = data;
2630         struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2631
2632         /* ok to return 0, domain has gone away */
2633         if (!dlm_grab(dlm))
2634                 return 0;
2635
2636         spin_lock(&dlm->spinlock);
2637         if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2638                 mlog(0, "%s: node %u wants to recover node %u (%u:%u) "
2639                      "but this node is in finalize state, waiting on finalize2\n",
2640                      dlm->name, br->node_idx, br->dead_node,
2641                      dlm->reco.dead_node, dlm->reco.new_master);
2642                 spin_unlock(&dlm->spinlock);
2643                 return EAGAIN;
2644         }
2645         spin_unlock(&dlm->spinlock);
2646
2647         mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n",
2648              dlm->name, br->node_idx, br->dead_node,
2649              dlm->reco.dead_node, dlm->reco.new_master);
2650
2651         dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2652
2653         spin_lock(&dlm->spinlock);
2654         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2655                 if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2656                         mlog(0, "%s: new_master %u died, changing "
2657                              "to %u\n", dlm->name, dlm->reco.new_master,
2658                              br->node_idx);
2659                 } else {
2660                         mlog(0, "%s: new_master %u NOT DEAD, changing "
2661                              "to %u\n", dlm->name, dlm->reco.new_master,
2662                              br->node_idx);
2663                         /* may not have seen the new master as dead yet */
2664                 }
2665         }
2666         if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2667                 mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2668                      "node %u changing it to %u\n", dlm->name, 
2669                      dlm->reco.dead_node, br->node_idx, br->dead_node);
2670         }
2671         dlm_set_reco_master(dlm, br->node_idx);
2672         dlm_set_reco_dead_node(dlm, br->dead_node);
2673         if (!test_bit(br->dead_node, dlm->recovery_map)) {
2674                 mlog(0, "recovery master %u sees %u as dead, but this "
2675                      "node has not yet.  marking %u as dead\n",
2676                      br->node_idx, br->dead_node, br->dead_node);
2677                 if (!test_bit(br->dead_node, dlm->domain_map) ||
2678                     !test_bit(br->dead_node, dlm->live_nodes_map))
2679                         mlog(0, "%u not in domain/live_nodes map "
2680                              "so setting it in reco map manually\n",
2681                              br->dead_node);
2682                 /* force the recovery cleanup in __dlm_hb_node_down
2683                  * both of these will be cleared in a moment */
2684                 set_bit(br->dead_node, dlm->domain_map);
2685                 set_bit(br->dead_node, dlm->live_nodes_map);
2686                 __dlm_hb_node_down(dlm, br->dead_node);
2687         }
2688         spin_unlock(&dlm->spinlock);
2689
2690         dlm_kick_recovery_thread(dlm);
2691
2692         mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n",
2693              dlm->name, br->node_idx, br->dead_node,
2694              dlm->reco.dead_node, dlm->reco.new_master);
2695
2696         dlm_put(dlm);
2697         return 0;
2698 }
2699
2700 #define DLM_FINALIZE_STAGE2  0x01
2701 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2702 {
2703         int ret = 0;
2704         struct dlm_finalize_reco fr;
2705         struct dlm_node_iter iter;
2706         int nodenum;
2707         int status;
2708         int stage = 1;
2709
2710         mlog(0, "finishing recovery for node %s:%u, "
2711              "stage %d\n", dlm->name, dlm->reco.dead_node, stage);
2712
2713         spin_lock(&dlm->spinlock);
2714         dlm_node_iter_init(dlm->domain_map, &iter);
2715         spin_unlock(&dlm->spinlock);
2716
2717 stage2:
2718         memset(&fr, 0, sizeof(fr));
2719         fr.node_idx = dlm->node_num;
2720         fr.dead_node = dlm->reco.dead_node;
2721         if (stage == 2)
2722                 fr.flags |= DLM_FINALIZE_STAGE2;
2723
2724         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2725                 if (nodenum == dlm->node_num)
2726                         continue;
2727                 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2728                                          &fr, sizeof(fr), nodenum, &status);
2729                 if (ret >= 0)
2730                         ret = status;
2731                 if (ret < 0) {
2732                         mlog_errno(ret);
2733                         if (dlm_is_host_down(ret)) {
2734                                 /* this has no effect on this recovery 
2735                                  * session, so set the status to zero to 
2736                                  * finish out the last recovery */
2737                                 mlog(ML_ERROR, "node %u went down after this "
2738                                      "node finished recovery.\n", nodenum);
2739                                 ret = 0;
2740                                 continue;
2741                         }
2742                         break;
2743                 }
2744         }
2745         if (stage == 1) {
2746                 /* reset the node_iter back to the top and send finalize2 */
2747                 iter.curnode = -1;
2748                 stage = 2;
2749                 goto stage2;
2750         }
2751
2752         return ret;
2753 }
2754
2755 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2756                               void **ret_data)
2757 {
2758         struct dlm_ctxt *dlm = data;
2759         struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2760         int stage = 1;
2761
2762         /* ok to return 0, domain has gone away */
2763         if (!dlm_grab(dlm))
2764                 return 0;
2765
2766         if (fr->flags & DLM_FINALIZE_STAGE2)
2767                 stage = 2;
2768
2769         mlog(0, "%s: node %u finalizing recovery stage%d of "
2770              "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage,
2771              fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master);
2772  
2773         spin_lock(&dlm->spinlock);
2774
2775         if (dlm->reco.new_master != fr->node_idx) {
2776                 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2777                      "%u is supposed to be the new master, dead=%u\n",
2778                      fr->node_idx, dlm->reco.new_master, fr->dead_node);
2779                 BUG();
2780         }
2781         if (dlm->reco.dead_node != fr->dead_node) {
2782                 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2783                      "node %u, but node %u is supposed to be dead\n",
2784                      fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2785                 BUG();
2786         }
2787
2788         switch (stage) {
2789                 case 1:
2790                         dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2791                         if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2792                                 mlog(ML_ERROR, "%s: received finalize1 from "
2793                                      "new master %u for dead node %u, but "
2794                                      "this node has already received it!\n",
2795                                      dlm->name, fr->node_idx, fr->dead_node);
2796                                 dlm_print_reco_node_status(dlm);
2797                                 BUG();
2798                         }
2799                         dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
2800                         spin_unlock(&dlm->spinlock);
2801                         break;
2802                 case 2:
2803                         if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) {
2804                                 mlog(ML_ERROR, "%s: received finalize2 from "
2805                                      "new master %u for dead node %u, but "
2806                                      "this node did not have finalize1!\n",
2807                                      dlm->name, fr->node_idx, fr->dead_node);
2808                                 dlm_print_reco_node_status(dlm);
2809                                 BUG();
2810                         }
2811                         dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2812                         spin_unlock(&dlm->spinlock);
2813                         dlm_reset_recovery(dlm);
2814                         dlm_kick_recovery_thread(dlm);
2815                         break;
2816                 default:
2817                         BUG();
2818         }
2819
2820         mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n",
2821              dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master);
2822
2823         dlm_put(dlm);
2824         return 0;
2825 }