Merge branch 'mv-merge'
[linux-2.6] / fs / ocfs2 / dlm / dlmrecovery.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * dlmrecovery.c
5  *
6  * recovery stuff
7  *
8  * Copyright (C) 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  *
25  */
26
27
28 #include <linux/module.h>
29 #include <linux/fs.h>
30 #include <linux/types.h>
31 #include <linux/slab.h>
32 #include <linux/highmem.h>
33 #include <linux/utsname.h>
34 #include <linux/init.h>
35 #include <linux/sysctl.h>
36 #include <linux/random.h>
37 #include <linux/blkdev.h>
38 #include <linux/socket.h>
39 #include <linux/inet.h>
40 #include <linux/timer.h>
41 #include <linux/kthread.h>
42 #include <linux/delay.h>
43
44
45 #include "cluster/heartbeat.h"
46 #include "cluster/nodemanager.h"
47 #include "cluster/tcp.h"
48
49 #include "dlmapi.h"
50 #include "dlmcommon.h"
51 #include "dlmdomain.h"
52
53 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY)
54 #include "cluster/masklog.h"
55
56 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node);
57
58 static int dlm_recovery_thread(void *data);
59 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
60 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
61 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm);
62 static int dlm_do_recovery(struct dlm_ctxt *dlm);
63
64 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm);
65 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node);
66 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
67 static int dlm_request_all_locks(struct dlm_ctxt *dlm,
68                                  u8 request_from, u8 dead_node);
69 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
70
71 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res);
72 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
73                                         const char *lockname, int namelen,
74                                         int total_locks, u64 cookie,
75                                         u8 flags, u8 master);
76 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
77                                     struct dlm_migratable_lockres *mres,
78                                     u8 send_to,
79                                     struct dlm_lock_resource *res,
80                                     int total_locks);
81 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
82                                      struct dlm_lock_resource *res,
83                                      struct dlm_migratable_lockres *mres);
84 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm);
85 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm,
86                                  u8 dead_node, u8 send_to);
87 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node);
88 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
89                                         struct list_head *list, u8 dead_node);
90 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
91                                               u8 dead_node, u8 new_master);
92 static void dlm_reco_ast(void *astdata);
93 static void dlm_reco_bast(void *astdata, int blocked_type);
94 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st);
95 static void dlm_request_all_locks_worker(struct dlm_work_item *item,
96                                          void *data);
97 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data);
98
99 static u64 dlm_get_next_mig_cookie(void);
100
101 static spinlock_t dlm_reco_state_lock = SPIN_LOCK_UNLOCKED;
102 static spinlock_t dlm_mig_cookie_lock = SPIN_LOCK_UNLOCKED;
103 static u64 dlm_mig_cookie = 1;
104
105 static u64 dlm_get_next_mig_cookie(void)
106 {
107         u64 c;
108         spin_lock(&dlm_mig_cookie_lock);
109         c = dlm_mig_cookie;
110         if (dlm_mig_cookie == (~0ULL))
111                 dlm_mig_cookie = 1;
112         else
113                 dlm_mig_cookie++;
114         spin_unlock(&dlm_mig_cookie_lock);
115         return c;
116 }
117
118 static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
119 {
120         spin_lock(&dlm->spinlock);
121         clear_bit(dlm->reco.dead_node, dlm->recovery_map);
122         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
123         dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
124         spin_unlock(&dlm->spinlock);
125 }
126
127 /* Worker function used during recovery. */
128 void dlm_dispatch_work(void *data)
129 {
130         struct dlm_ctxt *dlm = (struct dlm_ctxt *)data;
131         LIST_HEAD(tmp_list);
132         struct list_head *iter, *iter2;
133         struct dlm_work_item *item;
134         dlm_workfunc_t *workfunc;
135
136         spin_lock(&dlm->work_lock);
137         list_splice_init(&dlm->work_list, &tmp_list);
138         spin_unlock(&dlm->work_lock);
139
140         list_for_each_safe(iter, iter2, &tmp_list) {
141                 item = list_entry(iter, struct dlm_work_item, list);
142                 workfunc = item->func;
143                 list_del_init(&item->list);
144
145                 /* already have ref on dlm to avoid having
146                  * it disappear.  just double-check. */
147                 BUG_ON(item->dlm != dlm);
148
149                 /* this is allowed to sleep and
150                  * call network stuff */
151                 workfunc(item, item->data);
152
153                 dlm_put(dlm);
154                 kfree(item);
155         }
156 }
157
158 /*
159  * RECOVERY THREAD
160  */
161
162 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
163 {
164         /* wake the recovery thread
165          * this will wake the reco thread in one of three places
166          * 1) sleeping with no recovery happening
167          * 2) sleeping with recovery mastered elsewhere
168          * 3) recovery mastered here, waiting on reco data */
169
170         wake_up(&dlm->dlm_reco_thread_wq);
171 }
172
173 /* Launch the recovery thread */
174 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
175 {
176         mlog(0, "starting dlm recovery thread...\n");
177
178         dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
179                                                 "dlm_reco_thread");
180         if (IS_ERR(dlm->dlm_reco_thread_task)) {
181                 mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
182                 dlm->dlm_reco_thread_task = NULL;
183                 return -EINVAL;
184         }
185
186         return 0;
187 }
188
189 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
190 {
191         if (dlm->dlm_reco_thread_task) {
192                 mlog(0, "waiting for dlm recovery thread to exit\n");
193                 kthread_stop(dlm->dlm_reco_thread_task);
194                 dlm->dlm_reco_thread_task = NULL;
195         }
196 }
197
198
199
200 /*
201  * this is lame, but here's how recovery works...
202  * 1) all recovery threads cluster wide will work on recovering
203  *    ONE node at a time
204  * 2) negotiate who will take over all the locks for the dead node.
205  *    thats right... ALL the locks.
206  * 3) once a new master is chosen, everyone scans all locks
207  *    and moves aside those mastered by the dead guy
208  * 4) each of these locks should be locked until recovery is done
209  * 5) the new master collects up all of secondary lock queue info
210  *    one lock at a time, forcing each node to communicate back
211  *    before continuing
212  * 6) each secondary lock queue responds with the full known lock info
213  * 7) once the new master has run all its locks, it sends a ALLDONE!
214  *    message to everyone
215  * 8) upon receiving this message, the secondary queue node unlocks
216  *    and responds to the ALLDONE
217  * 9) once the new master gets responses from everyone, he unlocks
218  *    everything and recovery for this dead node is done
219  *10) go back to 2) while there are still dead nodes
220  *
221  */
222
223
224 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
225
226 static int dlm_recovery_thread(void *data)
227 {
228         int status;
229         struct dlm_ctxt *dlm = data;
230         unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
231
232         mlog(0, "dlm thread running for %s...\n", dlm->name);
233
234         while (!kthread_should_stop()) {
235                 if (dlm_joined(dlm)) {
236                         status = dlm_do_recovery(dlm);
237                         if (status == -EAGAIN) {
238                                 /* do not sleep, recheck immediately. */
239                                 continue;
240                         }
241                         if (status < 0)
242                                 mlog_errno(status);
243                 }
244
245                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
246                                                  kthread_should_stop(),
247                                                  timeout);
248         }
249
250         mlog(0, "quitting DLM recovery thread\n");
251         return 0;
252 }
253
254 /* returns true when the recovery master has contacted us */
255 static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
256 {
257         int ready;
258         spin_lock(&dlm->spinlock);
259         ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
260         spin_unlock(&dlm->spinlock);
261         return ready;
262 }
263
264 /* returns true if node is no longer in the domain
265  * could be dead or just not joined */
266 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
267 {
268         int dead;
269         spin_lock(&dlm->spinlock);
270         dead = test_bit(node, dlm->domain_map);
271         spin_unlock(&dlm->spinlock);
272         return dead;
273 }
274
275 int dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout)
276 {
277         if (timeout) {
278                 mlog(ML_NOTICE, "%s: waiting %dms for notification of "
279                      "death of node %u\n", dlm->name, timeout, node);
280                 wait_event_timeout(dlm->dlm_reco_thread_wq,
281                            dlm_is_node_dead(dlm, node),
282                            msecs_to_jiffies(timeout));
283         } else {
284                 mlog(ML_NOTICE, "%s: waiting indefinitely for notification "
285                      "of death of node %u\n", dlm->name, node);
286                 wait_event(dlm->dlm_reco_thread_wq,
287                            dlm_is_node_dead(dlm, node));
288         }
289         /* for now, return 0 */
290         return 0;
291 }
292
293 /* callers of the top-level api calls (dlmlock/dlmunlock) should
294  * block on the dlm->reco.event when recovery is in progress.
295  * the dlm recovery thread will set this state when it begins
296  * recovering a dead node (as the new master or not) and clear
297  * the state and wake as soon as all affected lock resources have
298  * been marked with the RECOVERY flag */
299 static int dlm_in_recovery(struct dlm_ctxt *dlm)
300 {
301         int in_recovery;
302         spin_lock(&dlm->spinlock);
303         in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
304         spin_unlock(&dlm->spinlock);
305         return in_recovery;
306 }
307
308
309 void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
310 {
311         wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
312 }
313
314 static void dlm_begin_recovery(struct dlm_ctxt *dlm)
315 {
316         spin_lock(&dlm->spinlock);
317         BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
318         dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
319         spin_unlock(&dlm->spinlock);
320 }
321
322 static void dlm_end_recovery(struct dlm_ctxt *dlm)
323 {
324         spin_lock(&dlm->spinlock);
325         BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
326         dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
327         spin_unlock(&dlm->spinlock);
328         wake_up(&dlm->reco.event);
329 }
330
331 static int dlm_do_recovery(struct dlm_ctxt *dlm)
332 {
333         int status = 0;
334         int ret;
335
336         spin_lock(&dlm->spinlock);
337
338         /* check to see if the new master has died */
339         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
340             test_bit(dlm->reco.new_master, dlm->recovery_map)) {
341                 mlog(0, "new master %u died while recovering %u!\n",
342                      dlm->reco.new_master, dlm->reco.dead_node);
343                 /* unset the new_master, leave dead_node */
344                 dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
345         }
346
347         /* select a target to recover */
348         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
349                 int bit;
350
351                 bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES+1, 0);
352                 if (bit >= O2NM_MAX_NODES || bit < 0)
353                         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
354                 else
355                         dlm->reco.dead_node = bit;
356         } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
357                 /* BUG? */
358                 mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
359                      dlm->reco.dead_node);
360                 dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
361         }
362
363         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
364                 // mlog(0, "nothing to recover!  sleeping now!\n");
365                 spin_unlock(&dlm->spinlock);
366                 /* return to main thread loop and sleep. */
367                 return 0;
368         }
369         mlog(0, "recovery thread found node %u in the recovery map!\n",
370              dlm->reco.dead_node);
371         spin_unlock(&dlm->spinlock);
372
373         /* take write barrier */
374         /* (stops the list reshuffling thread, proxy ast handling) */
375         dlm_begin_recovery(dlm);
376
377         if (dlm->reco.new_master == dlm->node_num)
378                 goto master_here;
379
380         if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
381                 /* choose a new master, returns 0 if this node
382                  * is the master, -EEXIST if it's another node.
383                  * this does not return until a new master is chosen
384                  * or recovery completes entirely. */
385                 ret = dlm_pick_recovery_master(dlm);
386                 if (!ret) {
387                         /* already notified everyone.  go. */
388                         goto master_here;
389                 }
390                 mlog(0, "another node will master this recovery session.\n");
391         }
392         mlog(0, "dlm=%s, new_master=%u, this node=%u, dead_node=%u\n",
393              dlm->name, dlm->reco.new_master,
394              dlm->node_num, dlm->reco.dead_node);
395
396         /* it is safe to start everything back up here
397          * because all of the dead node's lock resources
398          * have been marked as in-recovery */
399         dlm_end_recovery(dlm);
400
401         /* sleep out in main dlm_recovery_thread loop. */
402         return 0;
403
404 master_here:
405         mlog(0, "mastering recovery of %s:%u here(this=%u)!\n",
406              dlm->name, dlm->reco.dead_node, dlm->node_num);
407
408         status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
409         if (status < 0) {
410                 mlog(ML_ERROR, "error %d remastering locks for node %u, "
411                      "retrying.\n", status, dlm->reco.dead_node);
412                 /* yield a bit to allow any final network messages
413                  * to get handled on remaining nodes */
414                 msleep(100);
415         } else {
416                 /* success!  see if any other nodes need recovery */
417                 mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
418                      dlm->name, dlm->reco.dead_node, dlm->node_num);
419                 dlm_reset_recovery(dlm);
420         }
421         dlm_end_recovery(dlm);
422
423         /* continue and look for another dead node */
424         return -EAGAIN;
425 }
426
427 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
428 {
429         int status = 0;
430         struct dlm_reco_node_data *ndata;
431         struct list_head *iter;
432         int all_nodes_done;
433         int destroy = 0;
434         int pass = 0;
435
436         status = dlm_init_recovery_area(dlm, dead_node);
437         if (status < 0)
438                 goto leave;
439
440         /* safe to access the node data list without a lock, since this
441          * process is the only one to change the list */
442         list_for_each(iter, &dlm->reco.node_data) {
443                 ndata = list_entry (iter, struct dlm_reco_node_data, list);
444                 BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
445                 ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
446
447                 mlog(0, "requesting lock info from node %u\n",
448                      ndata->node_num);
449
450                 if (ndata->node_num == dlm->node_num) {
451                         ndata->state = DLM_RECO_NODE_DATA_DONE;
452                         continue;
453                 }
454
455                 status = dlm_request_all_locks(dlm, ndata->node_num, dead_node);
456                 if (status < 0) {
457                         mlog_errno(status);
458                         if (dlm_is_host_down(status))
459                                 ndata->state = DLM_RECO_NODE_DATA_DEAD;
460                         else {
461                                 destroy = 1;
462                                 goto leave;
463                         }
464                 }
465
466                 switch (ndata->state) {
467                         case DLM_RECO_NODE_DATA_INIT:
468                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
469                         case DLM_RECO_NODE_DATA_REQUESTED:
470                                 BUG();
471                                 break;
472                         case DLM_RECO_NODE_DATA_DEAD:
473                                 mlog(0, "node %u died after requesting "
474                                      "recovery info for node %u\n",
475                                      ndata->node_num, dead_node);
476                                 // start all over
477                                 destroy = 1;
478                                 status = -EAGAIN;
479                                 goto leave;
480                         case DLM_RECO_NODE_DATA_REQUESTING:
481                                 ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
482                                 mlog(0, "now receiving recovery data from "
483                                      "node %u for dead node %u\n",
484                                      ndata->node_num, dead_node);
485                                 break;
486                         case DLM_RECO_NODE_DATA_RECEIVING:
487                                 mlog(0, "already receiving recovery data from "
488                                      "node %u for dead node %u\n",
489                                      ndata->node_num, dead_node);
490                                 break;
491                         case DLM_RECO_NODE_DATA_DONE:
492                                 mlog(0, "already DONE receiving recovery data "
493                                      "from node %u for dead node %u\n",
494                                      ndata->node_num, dead_node);
495                                 break;
496                 }
497         }
498
499         mlog(0, "done requesting all lock info\n");
500
501         /* nodes should be sending reco data now
502          * just need to wait */
503
504         while (1) {
505                 /* check all the nodes now to see if we are
506                  * done, or if anyone died */
507                 all_nodes_done = 1;
508                 spin_lock(&dlm_reco_state_lock);
509                 list_for_each(iter, &dlm->reco.node_data) {
510                         ndata = list_entry (iter, struct dlm_reco_node_data, list);
511
512                         mlog(0, "checking recovery state of node %u\n",
513                              ndata->node_num);
514                         switch (ndata->state) {
515                                 case DLM_RECO_NODE_DATA_INIT:
516                                 case DLM_RECO_NODE_DATA_REQUESTING:
517                                         mlog(ML_ERROR, "bad ndata state for "
518                                              "node %u: state=%d\n",
519                                              ndata->node_num, ndata->state);
520                                         BUG();
521                                         break;
522                                 case DLM_RECO_NODE_DATA_DEAD:
523                                         mlog(ML_NOTICE, "node %u died after "
524                                              "requesting recovery info for "
525                                              "node %u\n", ndata->node_num,
526                                              dead_node);
527                                         spin_unlock(&dlm_reco_state_lock);
528                                         // start all over
529                                         destroy = 1;
530                                         status = -EAGAIN;
531                                         /* instead of spinning like crazy here,
532                                          * wait for the domain map to catch up
533                                          * with the network state.  otherwise this
534                                          * can be hit hundreds of times before
535                                          * the node is really seen as dead. */
536                                         wait_event_timeout(dlm->dlm_reco_thread_wq,
537                                                            dlm_is_node_dead(dlm,
538                                                                 ndata->node_num),
539                                                            msecs_to_jiffies(1000));
540                                         mlog(0, "waited 1 sec for %u, "
541                                              "dead? %s\n", ndata->node_num,
542                                              dlm_is_node_dead(dlm, ndata->node_num) ?
543                                              "yes" : "no");
544                                         goto leave;
545                                 case DLM_RECO_NODE_DATA_RECEIVING:
546                                 case DLM_RECO_NODE_DATA_REQUESTED:
547                                         all_nodes_done = 0;
548                                         break;
549                                 case DLM_RECO_NODE_DATA_DONE:
550                                         break;
551                                 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
552                                         break;
553                         }
554                 }
555                 spin_unlock(&dlm_reco_state_lock);
556
557                 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
558                      all_nodes_done?"yes":"no");
559                 if (all_nodes_done) {
560                         int ret;
561
562                         /* all nodes are now in DLM_RECO_NODE_DATA_DONE state
563                          * just send a finalize message to everyone and
564                          * clean up */
565                         mlog(0, "all nodes are done! send finalize\n");
566                         ret = dlm_send_finalize_reco_message(dlm);
567                         if (ret < 0)
568                                 mlog_errno(ret);
569
570                         spin_lock(&dlm->spinlock);
571                         dlm_finish_local_lockres_recovery(dlm, dead_node,
572                                                           dlm->node_num);
573                         spin_unlock(&dlm->spinlock);
574                         mlog(0, "should be done with recovery!\n");
575
576                         mlog(0, "finishing recovery of %s at %lu, "
577                              "dead=%u, this=%u, new=%u\n", dlm->name,
578                              jiffies, dlm->reco.dead_node,
579                              dlm->node_num, dlm->reco.new_master);
580                         destroy = 1;
581                         status = ret;
582                         /* rescan everything marked dirty along the way */
583                         dlm_kick_thread(dlm, NULL);
584                         break;
585                 }
586                 /* wait to be signalled, with periodic timeout
587                  * to check for node death */
588                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
589                                          kthread_should_stop(),
590                                          msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
591
592         }
593
594 leave:
595         if (destroy)
596                 dlm_destroy_recovery_area(dlm, dead_node);
597
598         mlog_exit(status);
599         return status;
600 }
601
602 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
603 {
604         int num=0;
605         struct dlm_reco_node_data *ndata;
606
607         spin_lock(&dlm->spinlock);
608         memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
609         /* nodes can only be removed (by dying) after dropping
610          * this lock, and death will be trapped later, so this should do */
611         spin_unlock(&dlm->spinlock);
612
613         while (1) {
614                 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
615                 if (num >= O2NM_MAX_NODES) {
616                         break;
617                 }
618                 BUG_ON(num == dead_node);
619
620                 ndata = kcalloc(1, sizeof(*ndata), GFP_KERNEL);
621                 if (!ndata) {
622                         dlm_destroy_recovery_area(dlm, dead_node);
623                         return -ENOMEM;
624                 }
625                 ndata->node_num = num;
626                 ndata->state = DLM_RECO_NODE_DATA_INIT;
627                 spin_lock(&dlm_reco_state_lock);
628                 list_add_tail(&ndata->list, &dlm->reco.node_data);
629                 spin_unlock(&dlm_reco_state_lock);
630                 num++;
631         }
632
633         return 0;
634 }
635
636 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
637 {
638         struct list_head *iter, *iter2;
639         struct dlm_reco_node_data *ndata;
640         LIST_HEAD(tmplist);
641
642         spin_lock(&dlm_reco_state_lock);
643         list_splice_init(&dlm->reco.node_data, &tmplist);
644         spin_unlock(&dlm_reco_state_lock);
645
646         list_for_each_safe(iter, iter2, &tmplist) {
647                 ndata = list_entry (iter, struct dlm_reco_node_data, list);
648                 list_del_init(&ndata->list);
649                 kfree(ndata);
650         }
651 }
652
653 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
654                                  u8 dead_node)
655 {
656         struct dlm_lock_request lr;
657         enum dlm_status ret;
658
659         mlog(0, "\n");
660
661
662         mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
663                   "to %u\n", dead_node, request_from);
664
665         memset(&lr, 0, sizeof(lr));
666         lr.node_idx = dlm->node_num;
667         lr.dead_node = dead_node;
668
669         // send message
670         ret = DLM_NOLOCKMGR;
671         ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
672                                  &lr, sizeof(lr), request_from, NULL);
673
674         /* negative status is handled by caller */
675         if (ret < 0)
676                 mlog_errno(ret);
677
678         // return from here, then
679         // sleep until all received or error
680         return ret;
681
682 }
683
684 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data)
685 {
686         struct dlm_ctxt *dlm = data;
687         struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
688         char *buf = NULL;
689         struct dlm_work_item *item = NULL;
690
691         if (!dlm_grab(dlm))
692                 return -EINVAL;
693
694         BUG_ON(lr->dead_node != dlm->reco.dead_node);
695
696         item = kcalloc(1, sizeof(*item), GFP_KERNEL);
697         if (!item) {
698                 dlm_put(dlm);
699                 return -ENOMEM;
700         }
701
702         /* this will get freed by dlm_request_all_locks_worker */
703         buf = (char *) __get_free_page(GFP_KERNEL);
704         if (!buf) {
705                 kfree(item);
706                 dlm_put(dlm);
707                 return -ENOMEM;
708         }
709
710         /* queue up work for dlm_request_all_locks_worker */
711         dlm_grab(dlm);  /* get an extra ref for the work item */
712         dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
713         item->u.ral.reco_master = lr->node_idx;
714         item->u.ral.dead_node = lr->dead_node;
715         spin_lock(&dlm->work_lock);
716         list_add_tail(&item->list, &dlm->work_list);
717         spin_unlock(&dlm->work_lock);
718         schedule_work(&dlm->dispatched_work);
719
720         dlm_put(dlm);
721         return 0;
722 }
723
724 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
725 {
726         struct dlm_migratable_lockres *mres;
727         struct dlm_lock_resource *res;
728         struct dlm_ctxt *dlm;
729         LIST_HEAD(resources);
730         struct list_head *iter;
731         int ret;
732         u8 dead_node, reco_master;
733
734         dlm = item->dlm;
735         dead_node = item->u.ral.dead_node;
736         reco_master = item->u.ral.reco_master;
737         mres = (struct dlm_migratable_lockres *)data;
738
739         if (dead_node != dlm->reco.dead_node ||
740             reco_master != dlm->reco.new_master) {
741                 /* show extra debug info if the recovery state is messed */
742                 mlog(ML_ERROR, "%s: bad reco state: reco(dead=%u, master=%u), "
743                      "request(dead=%u, master=%u)\n",
744                      dlm->name, dlm->reco.dead_node, dlm->reco.new_master,
745                      dead_node, reco_master);
746                 mlog(ML_ERROR, "%s: name=%.*s master=%u locks=%u/%u flags=%u "
747                      "entry[0]={c=%u:%llu,l=%u,f=%u,t=%d,ct=%d,hb=%d,n=%u}\n",
748                      dlm->name, mres->lockname_len, mres->lockname, mres->master,
749                      mres->num_locks, mres->total_locks, mres->flags,
750                      dlm_get_lock_cookie_node(mres->ml[0].cookie),
751                      dlm_get_lock_cookie_seq(mres->ml[0].cookie),
752                      mres->ml[0].list, mres->ml[0].flags,
753                      mres->ml[0].type, mres->ml[0].convert_type,
754                      mres->ml[0].highest_blocked, mres->ml[0].node);
755                 BUG();
756         }
757         BUG_ON(dead_node != dlm->reco.dead_node);
758         BUG_ON(reco_master != dlm->reco.new_master);
759
760         /* lock resources should have already been moved to the
761          * dlm->reco.resources list.  now move items from that list
762          * to a temp list if the dead owner matches.  note that the
763          * whole cluster recovers only one node at a time, so we
764          * can safely move UNKNOWN lock resources for each recovery
765          * session. */
766         dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
767
768         /* now we can begin blasting lockreses without the dlm lock */
769         list_for_each(iter, &resources) {
770                 res = list_entry (iter, struct dlm_lock_resource, recovering);
771                 ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
772                                         DLM_MRES_RECOVERY);
773                 if (ret < 0)
774                         mlog_errno(ret);
775         }
776
777         /* move the resources back to the list */
778         spin_lock(&dlm->spinlock);
779         list_splice_init(&resources, &dlm->reco.resources);
780         spin_unlock(&dlm->spinlock);
781
782         ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
783         if (ret < 0)
784                 mlog_errno(ret);
785
786         free_page((unsigned long)data);
787 }
788
789
790 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
791 {
792         int ret, tmpret;
793         struct dlm_reco_data_done done_msg;
794
795         memset(&done_msg, 0, sizeof(done_msg));
796         done_msg.node_idx = dlm->node_num;
797         done_msg.dead_node = dead_node;
798         mlog(0, "sending DATA DONE message to %u, "
799              "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
800              done_msg.dead_node);
801
802         ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
803                                  sizeof(done_msg), send_to, &tmpret);
804         /* negative status is ignored by the caller */
805         if (ret >= 0)
806                 ret = tmpret;
807         return ret;
808 }
809
810
811 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data)
812 {
813         struct dlm_ctxt *dlm = data;
814         struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
815         struct list_head *iter;
816         struct dlm_reco_node_data *ndata = NULL;
817         int ret = -EINVAL;
818
819         if (!dlm_grab(dlm))
820                 return -EINVAL;
821
822         mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
823              "node_idx=%u, this node=%u\n", done->dead_node,
824              dlm->reco.dead_node, done->node_idx, dlm->node_num);
825         BUG_ON(done->dead_node != dlm->reco.dead_node);
826
827         spin_lock(&dlm_reco_state_lock);
828         list_for_each(iter, &dlm->reco.node_data) {
829                 ndata = list_entry (iter, struct dlm_reco_node_data, list);
830                 if (ndata->node_num != done->node_idx)
831                         continue;
832
833                 switch (ndata->state) {
834                         /* should have moved beyond INIT but not to FINALIZE yet */
835                         case DLM_RECO_NODE_DATA_INIT:
836                         case DLM_RECO_NODE_DATA_DEAD:
837                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
838                                 mlog(ML_ERROR, "bad ndata state for node %u:"
839                                      " state=%d\n", ndata->node_num,
840                                      ndata->state);
841                                 BUG();
842                                 break;
843                         /* these states are possible at this point, anywhere along
844                          * the line of recovery */
845                         case DLM_RECO_NODE_DATA_DONE:
846                         case DLM_RECO_NODE_DATA_RECEIVING:
847                         case DLM_RECO_NODE_DATA_REQUESTED:
848                         case DLM_RECO_NODE_DATA_REQUESTING:
849                                 mlog(0, "node %u is DONE sending "
850                                           "recovery data!\n",
851                                           ndata->node_num);
852
853                                 ndata->state = DLM_RECO_NODE_DATA_DONE;
854                                 ret = 0;
855                                 break;
856                 }
857         }
858         spin_unlock(&dlm_reco_state_lock);
859
860         /* wake the recovery thread, some node is done */
861         if (!ret)
862                 dlm_kick_recovery_thread(dlm);
863
864         if (ret < 0)
865                 mlog(ML_ERROR, "failed to find recovery node data for node "
866                      "%u\n", done->node_idx);
867         dlm_put(dlm);
868
869         mlog(0, "leaving reco data done handler, ret=%d\n", ret);
870         return ret;
871 }
872
873 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
874                                         struct list_head *list,
875                                         u8 dead_node)
876 {
877         struct dlm_lock_resource *res;
878         struct list_head *iter, *iter2;
879         struct dlm_lock *lock;
880
881         spin_lock(&dlm->spinlock);
882         list_for_each_safe(iter, iter2, &dlm->reco.resources) {
883                 res = list_entry (iter, struct dlm_lock_resource, recovering);
884                 /* always prune any $RECOVERY entries for dead nodes,
885                  * otherwise hangs can occur during later recovery */
886                 if (dlm_is_recovery_lock(res->lockname.name,
887                                          res->lockname.len)) {
888                         spin_lock(&res->spinlock);
889                         list_for_each_entry(lock, &res->granted, list) {
890                                 if (lock->ml.node == dead_node) {
891                                         mlog(0, "AHA! there was "
892                                              "a $RECOVERY lock for dead "
893                                              "node %u (%s)!\n", 
894                                              dead_node, dlm->name);
895                                         list_del_init(&lock->list);
896                                         dlm_lock_put(lock);
897                                         break;
898                                 }
899                         }
900                         spin_unlock(&res->spinlock);
901                         continue;
902                 }
903
904                 if (res->owner == dead_node) {
905                         mlog(0, "found lockres owned by dead node while "
906                                   "doing recovery for node %u. sending it.\n",
907                                   dead_node);
908                         list_del_init(&res->recovering);
909                         list_add_tail(&res->recovering, list);
910                 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
911                         mlog(0, "found UNKNOWN owner while doing recovery "
912                                   "for node %u. sending it.\n", dead_node);
913                         list_del_init(&res->recovering);
914                         list_add_tail(&res->recovering, list);
915                 }
916         }
917         spin_unlock(&dlm->spinlock);
918 }
919
920 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
921 {
922         int total_locks = 0;
923         struct list_head *iter, *queue = &res->granted;
924         int i;
925
926         for (i=0; i<3; i++) {
927                 list_for_each(iter, queue)
928                         total_locks++;
929                 queue++;
930         }
931         return total_locks;
932 }
933
934
935 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
936                                       struct dlm_migratable_lockres *mres,
937                                       u8 send_to,
938                                       struct dlm_lock_resource *res,
939                                       int total_locks)
940 {
941         u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
942         int mres_total_locks = be32_to_cpu(mres->total_locks);
943         int sz, ret = 0, status = 0;
944         u8 orig_flags = mres->flags,
945            orig_master = mres->master;
946
947         BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
948         if (!mres->num_locks)
949                 return 0;
950
951         sz = sizeof(struct dlm_migratable_lockres) +
952                 (mres->num_locks * sizeof(struct dlm_migratable_lock));
953
954         /* add an all-done flag if we reached the last lock */
955         orig_flags = mres->flags;
956         BUG_ON(total_locks > mres_total_locks);
957         if (total_locks == mres_total_locks)
958                 mres->flags |= DLM_MRES_ALL_DONE;
959
960         /* send it */
961         ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
962                                  sz, send_to, &status);
963         if (ret < 0) {
964                 /* XXX: negative status is not handled.
965                  * this will end up killing this node. */
966                 mlog_errno(ret);
967         } else {
968                 /* might get an -ENOMEM back here */
969                 ret = status;
970                 if (ret < 0) {
971                         mlog_errno(ret);
972
973                         if (ret == -EFAULT) {
974                                 mlog(ML_ERROR, "node %u told me to kill "
975                                      "myself!\n", send_to);
976                                 BUG();
977                         }
978                 }
979         }
980
981         /* zero and reinit the message buffer */
982         dlm_init_migratable_lockres(mres, res->lockname.name,
983                                     res->lockname.len, mres_total_locks,
984                                     mig_cookie, orig_flags, orig_master);
985         return ret;
986 }
987
988 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
989                                         const char *lockname, int namelen,
990                                         int total_locks, u64 cookie,
991                                         u8 flags, u8 master)
992 {
993         /* mres here is one full page */
994         memset(mres, 0, PAGE_SIZE);
995         mres->lockname_len = namelen;
996         memcpy(mres->lockname, lockname, namelen);
997         mres->num_locks = 0;
998         mres->total_locks = cpu_to_be32(total_locks);
999         mres->mig_cookie = cpu_to_be64(cookie);
1000         mres->flags = flags;
1001         mres->master = master;
1002 }
1003
1004
1005 /* returns 1 if this lock fills the network structure,
1006  * 0 otherwise */
1007 static int dlm_add_lock_to_array(struct dlm_lock *lock,
1008                                  struct dlm_migratable_lockres *mres, int queue)
1009 {
1010         struct dlm_migratable_lock *ml;
1011         int lock_num = mres->num_locks;
1012
1013         ml = &(mres->ml[lock_num]);
1014         ml->cookie = lock->ml.cookie;
1015         ml->type = lock->ml.type;
1016         ml->convert_type = lock->ml.convert_type;
1017         ml->highest_blocked = lock->ml.highest_blocked;
1018         ml->list = queue;
1019         if (lock->lksb) {
1020                 ml->flags = lock->lksb->flags;
1021                 /* send our current lvb */
1022                 if (ml->type == LKM_EXMODE ||
1023                     ml->type == LKM_PRMODE) {
1024                         /* if it is already set, this had better be a PR
1025                          * and it has to match */
1026                         if (mres->lvb[0] && (ml->type == LKM_EXMODE ||
1027                             memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))) {
1028                                 mlog(ML_ERROR, "mismatched lvbs!\n");
1029                                 __dlm_print_one_lock_resource(lock->lockres);
1030                                 BUG();
1031                         }
1032                         memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1033                 }
1034         }
1035         ml->node = lock->ml.node;
1036         mres->num_locks++;
1037         /* we reached the max, send this network message */
1038         if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1039                 return 1;
1040         return 0;
1041 }
1042
1043
1044 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1045                          struct dlm_migratable_lockres *mres,
1046                          u8 send_to, u8 flags)
1047 {
1048         struct list_head *queue, *iter;
1049         int total_locks, i;
1050         u64 mig_cookie = 0;
1051         struct dlm_lock *lock;
1052         int ret = 0;
1053
1054         BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1055
1056         mlog(0, "sending to %u\n", send_to);
1057
1058         total_locks = dlm_num_locks_in_lockres(res);
1059         if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1060                 /* rare, but possible */
1061                 mlog(0, "argh.  lockres has %d locks.  this will "
1062                           "require more than one network packet to "
1063                           "migrate\n", total_locks);
1064                 mig_cookie = dlm_get_next_mig_cookie();
1065         }
1066
1067         dlm_init_migratable_lockres(mres, res->lockname.name,
1068                                     res->lockname.len, total_locks,
1069                                     mig_cookie, flags, res->owner);
1070
1071         total_locks = 0;
1072         for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1073                 queue = dlm_list_idx_to_ptr(res, i);
1074                 list_for_each(iter, queue) {
1075                         lock = list_entry (iter, struct dlm_lock, list);
1076
1077                         /* add another lock. */
1078                         total_locks++;
1079                         if (!dlm_add_lock_to_array(lock, mres, i))
1080                                 continue;
1081
1082                         /* this filled the lock message,
1083                          * we must send it immediately. */
1084                         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1085                                                        res, total_locks);
1086                         if (ret < 0) {
1087                                 // TODO
1088                                 mlog(ML_ERROR, "dlm_send_mig_lockres_msg "
1089                                      "returned %d, TODO\n", ret);
1090                                 BUG();
1091                         }
1092                 }
1093         }
1094         /* flush any remaining locks */
1095         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1096         if (ret < 0) {
1097                 // TODO
1098                 mlog(ML_ERROR, "dlm_send_mig_lockres_msg returned %d, "
1099                      "TODO\n", ret);
1100                 BUG();
1101         }
1102         return ret;
1103 }
1104
1105
1106
1107 /*
1108  * this message will contain no more than one page worth of
1109  * recovery data, and it will work on only one lockres.
1110  * there may be many locks in this page, and we may need to wait
1111  * for additional packets to complete all the locks (rare, but
1112  * possible).
1113  */
1114 /*
1115  * NOTE: the allocation error cases here are scary
1116  * we really cannot afford to fail an alloc in recovery
1117  * do we spin?  returning an error only delays the problem really
1118  */
1119
1120 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data)
1121 {
1122         struct dlm_ctxt *dlm = data;
1123         struct dlm_migratable_lockres *mres =
1124                 (struct dlm_migratable_lockres *)msg->buf;
1125         int ret = 0;
1126         u8 real_master;
1127         char *buf = NULL;
1128         struct dlm_work_item *item = NULL;
1129         struct dlm_lock_resource *res = NULL;
1130
1131         if (!dlm_grab(dlm))
1132                 return -EINVAL;
1133
1134         BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1135
1136         real_master = mres->master;
1137         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1138                 /* cannot migrate a lockres with no master */
1139                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1140         }
1141
1142         mlog(0, "%s message received from node %u\n",
1143                   (mres->flags & DLM_MRES_RECOVERY) ?
1144                   "recovery" : "migration", mres->master);
1145         if (mres->flags & DLM_MRES_ALL_DONE)
1146                 mlog(0, "all done flag.  all lockres data received!\n");
1147
1148         ret = -ENOMEM;
1149         buf = kmalloc(be16_to_cpu(msg->data_len), GFP_KERNEL);
1150         item = kcalloc(1, sizeof(*item), GFP_KERNEL);
1151         if (!buf || !item)
1152                 goto leave;
1153
1154         /* lookup the lock to see if we have a secondary queue for this
1155          * already...  just add the locks in and this will have its owner
1156          * and RECOVERY flag changed when it completes. */
1157         res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len);
1158         if (res) {
1159                 /* this will get a ref on res */
1160                 /* mark it as recovering/migrating and hash it */
1161                 spin_lock(&res->spinlock);
1162                 if (mres->flags & DLM_MRES_RECOVERY) {
1163                         res->state |= DLM_LOCK_RES_RECOVERING;
1164                 } else {
1165                         if (res->state & DLM_LOCK_RES_MIGRATING) {
1166                                 /* this is at least the second
1167                                  * lockres message */
1168                                 mlog(0, "lock %.*s is already migrating\n",
1169                                           mres->lockname_len,
1170                                           mres->lockname);
1171                         } else if (res->state & DLM_LOCK_RES_RECOVERING) {
1172                                 /* caller should BUG */
1173                                 mlog(ML_ERROR, "node is attempting to migrate "
1174                                      "lock %.*s, but marked as recovering!\n",
1175                                      mres->lockname_len, mres->lockname);
1176                                 ret = -EFAULT;
1177                                 spin_unlock(&res->spinlock);
1178                                 goto leave;
1179                         }
1180                         res->state |= DLM_LOCK_RES_MIGRATING;
1181                 }
1182                 spin_unlock(&res->spinlock);
1183         } else {
1184                 /* need to allocate, just like if it was
1185                  * mastered here normally  */
1186                 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1187                 if (!res)
1188                         goto leave;
1189
1190                 /* to match the ref that we would have gotten if
1191                  * dlm_lookup_lockres had succeeded */
1192                 dlm_lockres_get(res);
1193
1194                 /* mark it as recovering/migrating and hash it */
1195                 if (mres->flags & DLM_MRES_RECOVERY)
1196                         res->state |= DLM_LOCK_RES_RECOVERING;
1197                 else
1198                         res->state |= DLM_LOCK_RES_MIGRATING;
1199
1200                 spin_lock(&dlm->spinlock);
1201                 __dlm_insert_lockres(dlm, res);
1202                 spin_unlock(&dlm->spinlock);
1203
1204                 /* now that the new lockres is inserted,
1205                  * make it usable by other processes */
1206                 spin_lock(&res->spinlock);
1207                 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1208                 spin_unlock(&res->spinlock);
1209
1210                 /* add an extra ref for just-allocated lockres 
1211                  * otherwise the lockres will be purged immediately */
1212                 dlm_lockres_get(res);
1213
1214         }
1215
1216         /* at this point we have allocated everything we need,
1217          * and we have a hashed lockres with an extra ref and
1218          * the proper res->state flags. */
1219         ret = 0;
1220         if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1221                 /* migration cannot have an unknown master */
1222                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1223                 mlog(0, "recovery has passed me a lockres with an "
1224                           "unknown owner.. will need to requery: "
1225                           "%.*s\n", mres->lockname_len, mres->lockname);
1226         } else {
1227                 spin_lock(&res->spinlock);
1228                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
1229                 spin_unlock(&res->spinlock);
1230         }
1231
1232         /* queue up work for dlm_mig_lockres_worker */
1233         dlm_grab(dlm);  /* get an extra ref for the work item */
1234         memcpy(buf, msg->buf, be16_to_cpu(msg->data_len));  /* copy the whole message */
1235         dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1236         item->u.ml.lockres = res; /* already have a ref */
1237         item->u.ml.real_master = real_master;
1238         spin_lock(&dlm->work_lock);
1239         list_add_tail(&item->list, &dlm->work_list);
1240         spin_unlock(&dlm->work_lock);
1241         schedule_work(&dlm->dispatched_work);
1242
1243 leave:
1244         dlm_put(dlm);
1245         if (ret < 0) {
1246                 if (buf)
1247                         kfree(buf);
1248                 if (item)
1249                         kfree(item);
1250         }
1251
1252         mlog_exit(ret);
1253         return ret;
1254 }
1255
1256
1257 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1258 {
1259         struct dlm_ctxt *dlm = data;
1260         struct dlm_migratable_lockres *mres;
1261         int ret = 0;
1262         struct dlm_lock_resource *res;
1263         u8 real_master;
1264
1265         dlm = item->dlm;
1266         mres = (struct dlm_migratable_lockres *)data;
1267
1268         res = item->u.ml.lockres;
1269         real_master = item->u.ml.real_master;
1270
1271         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1272                 /* this case is super-rare. only occurs if
1273                  * node death happens during migration. */
1274 again:
1275                 ret = dlm_lockres_master_requery(dlm, res, &real_master);
1276                 if (ret < 0) {
1277                         mlog(0, "dlm_lockres_master_requery ret=%d\n",
1278                                   ret);
1279                         goto again;
1280                 }
1281                 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1282                         mlog(0, "lockres %.*s not claimed.  "
1283                                    "this node will take it.\n",
1284                                    res->lockname.len, res->lockname.name);
1285                 } else {
1286                         mlog(0, "master needs to respond to sender "
1287                                   "that node %u still owns %.*s\n",
1288                                   real_master, res->lockname.len,
1289                                   res->lockname.name);
1290                         /* cannot touch this lockres */
1291                         goto leave;
1292                 }
1293         }
1294
1295         ret = dlm_process_recovery_data(dlm, res, mres);
1296         if (ret < 0)
1297                 mlog(0, "dlm_process_recovery_data returned  %d\n", ret);
1298         else
1299                 mlog(0, "dlm_process_recovery_data succeeded\n");
1300
1301         if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1302                            (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1303                 ret = dlm_finish_migration(dlm, res, mres->master);
1304                 if (ret < 0)
1305                         mlog_errno(ret);
1306         }
1307
1308 leave:
1309         kfree(data);
1310         mlog_exit(ret);
1311 }
1312
1313
1314
1315 int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1316                                struct dlm_lock_resource *res, u8 *real_master)
1317 {
1318         struct dlm_node_iter iter;
1319         int nodenum;
1320         int ret = 0;
1321
1322         *real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1323
1324         /* we only reach here if one of the two nodes in a
1325          * migration died while the migration was in progress.
1326          * at this point we need to requery the master.  we
1327          * know that the new_master got as far as creating
1328          * an mle on at least one node, but we do not know
1329          * if any nodes had actually cleared the mle and set
1330          * the master to the new_master.  the old master
1331          * is supposed to set the owner to UNKNOWN in the
1332          * event of a new_master death, so the only possible
1333          * responses that we can get from nodes here are
1334          * that the master is new_master, or that the master
1335          * is UNKNOWN.
1336          * if all nodes come back with UNKNOWN then we know
1337          * the lock needs remastering here.
1338          * if any node comes back with a valid master, check
1339          * to see if that master is the one that we are
1340          * recovering.  if so, then the new_master died and
1341          * we need to remaster this lock.  if not, then the
1342          * new_master survived and that node will respond to
1343          * other nodes about the owner.
1344          * if there is an owner, this node needs to dump this
1345          * lockres and alert the sender that this lockres
1346          * was rejected. */
1347         spin_lock(&dlm->spinlock);
1348         dlm_node_iter_init(dlm->domain_map, &iter);
1349         spin_unlock(&dlm->spinlock);
1350
1351         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1352                 /* do not send to self */
1353                 if (nodenum == dlm->node_num)
1354                         continue;
1355                 ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1356                 if (ret < 0) {
1357                         mlog_errno(ret);
1358                         if (!dlm_is_host_down(ret))
1359                                 BUG();
1360                         /* host is down, so answer for that node would be
1361                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
1362                 }
1363                 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1364                         mlog(0, "lock master is %u\n", *real_master);
1365                         break;
1366                 }
1367         }
1368         return ret;
1369 }
1370
1371
1372 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1373                           u8 nodenum, u8 *real_master)
1374 {
1375         int ret = -EINVAL;
1376         struct dlm_master_requery req;
1377         int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1378
1379         memset(&req, 0, sizeof(req));
1380         req.node_idx = dlm->node_num;
1381         req.namelen = res->lockname.len;
1382         memcpy(req.name, res->lockname.name, res->lockname.len);
1383
1384         ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1385                                  &req, sizeof(req), nodenum, &status);
1386         /* XXX: negative status not handled properly here. */
1387         if (ret < 0)
1388                 mlog_errno(ret);
1389         else {
1390                 BUG_ON(status < 0);
1391                 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1392                 *real_master = (u8) (status & 0xff);
1393                 mlog(0, "node %u responded to master requery with %u\n",
1394                           nodenum, *real_master);
1395                 ret = 0;
1396         }
1397         return ret;
1398 }
1399
1400
1401 /* this function cannot error, so unless the sending
1402  * or receiving of the message failed, the owner can
1403  * be trusted */
1404 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data)
1405 {
1406         struct dlm_ctxt *dlm = data;
1407         struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1408         struct dlm_lock_resource *res = NULL;
1409         int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1410         u32 flags = DLM_ASSERT_MASTER_REQUERY;
1411
1412         if (!dlm_grab(dlm)) {
1413                 /* since the domain has gone away on this
1414                  * node, the proper response is UNKNOWN */
1415                 return master;
1416         }
1417
1418         spin_lock(&dlm->spinlock);
1419         res = __dlm_lookup_lockres(dlm, req->name, req->namelen);
1420         if (res) {
1421                 spin_lock(&res->spinlock);
1422                 master = res->owner;
1423                 if (master == dlm->node_num) {
1424                         int ret = dlm_dispatch_assert_master(dlm, res,
1425                                                              0, 0, flags);
1426                         if (ret < 0) {
1427                                 mlog_errno(-ENOMEM);
1428                                 /* retry!? */
1429                                 BUG();
1430                         }
1431                 }
1432                 spin_unlock(&res->spinlock);
1433         }
1434         spin_unlock(&dlm->spinlock);
1435
1436         dlm_put(dlm);
1437         return master;
1438 }
1439
1440 static inline struct list_head *
1441 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1442 {
1443         struct list_head *ret;
1444         BUG_ON(list_num < 0);
1445         BUG_ON(list_num > 2);
1446         ret = &(res->granted);
1447         ret += list_num;
1448         return ret;
1449 }
1450 /* TODO: do ast flush business
1451  * TODO: do MIGRATING and RECOVERING spinning
1452  */
1453
1454 /*
1455 * NOTE about in-flight requests during migration:
1456 *
1457 * Before attempting the migrate, the master has marked the lockres as
1458 * MIGRATING and then flushed all of its pending ASTS.  So any in-flight
1459 * requests either got queued before the MIGRATING flag got set, in which
1460 * case the lock data will reflect the change and a return message is on
1461 * the way, or the request failed to get in before MIGRATING got set.  In
1462 * this case, the caller will be told to spin and wait for the MIGRATING
1463 * flag to be dropped, then recheck the master.
1464 * This holds true for the convert, cancel and unlock cases, and since lvb
1465 * updates are tied to these same messages, it applies to lvb updates as
1466 * well.  For the lock case, there is no way a lock can be on the master
1467 * queue and not be on the secondary queue since the lock is always added
1468 * locally first.  This means that the new target node will never be sent
1469 * a lock that he doesn't already have on the list.
1470 * In total, this means that the local lock is correct and should not be
1471 * updated to match the one sent by the master.  Any messages sent back
1472 * from the master before the MIGRATING flag will bring the lock properly
1473 * up-to-date, and the change will be ordered properly for the waiter.
1474 * We will *not* attempt to modify the lock underneath the waiter.
1475 */
1476
1477 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1478                                      struct dlm_lock_resource *res,
1479                                      struct dlm_migratable_lockres *mres)
1480 {
1481         struct dlm_migratable_lock *ml;
1482         struct list_head *queue;
1483         struct dlm_lock *newlock = NULL;
1484         struct dlm_lockstatus *lksb = NULL;
1485         int ret = 0;
1486         int i;
1487         struct list_head *iter;
1488         struct dlm_lock *lock = NULL;
1489
1490         mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1491         for (i=0; i<mres->num_locks; i++) {
1492                 ml = &(mres->ml[i]);
1493                 BUG_ON(ml->highest_blocked != LKM_IVMODE);
1494                 newlock = NULL;
1495                 lksb = NULL;
1496
1497                 queue = dlm_list_num_to_pointer(res, ml->list);
1498
1499                 /* if the lock is for the local node it needs to
1500                  * be moved to the proper location within the queue.
1501                  * do not allocate a new lock structure. */
1502                 if (ml->node == dlm->node_num) {
1503                         /* MIGRATION ONLY! */
1504                         BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1505
1506                         spin_lock(&res->spinlock);
1507                         list_for_each(iter, queue) {
1508                                 lock = list_entry (iter, struct dlm_lock, list);
1509                                 if (lock->ml.cookie != ml->cookie)
1510                                         lock = NULL;
1511                                 else
1512                                         break;
1513                         }
1514
1515                         /* lock is always created locally first, and
1516                          * destroyed locally last.  it must be on the list */
1517                         if (!lock) {
1518                                 u64 c = ml->cookie;
1519                                 mlog(ML_ERROR, "could not find local lock "
1520                                                "with cookie %u:%llu!\n",
1521                                                dlm_get_lock_cookie_node(c),
1522                                                dlm_get_lock_cookie_seq(c));
1523                                 BUG();
1524                         }
1525                         BUG_ON(lock->ml.node != ml->node);
1526
1527                         /* see NOTE above about why we do not update
1528                          * to match the master here */
1529
1530                         /* move the lock to its proper place */
1531                         /* do not alter lock refcount.  switching lists. */
1532                         list_del_init(&lock->list);
1533                         list_add_tail(&lock->list, queue);
1534                         spin_unlock(&res->spinlock);
1535
1536                         mlog(0, "just reordered a local lock!\n");
1537                         continue;
1538                 }
1539
1540                 /* lock is for another node. */
1541                 newlock = dlm_new_lock(ml->type, ml->node,
1542                                        be64_to_cpu(ml->cookie), NULL);
1543                 if (!newlock) {
1544                         ret = -ENOMEM;
1545                         goto leave;
1546                 }
1547                 lksb = newlock->lksb;
1548                 dlm_lock_attach_lockres(newlock, res);
1549
1550                 if (ml->convert_type != LKM_IVMODE) {
1551                         BUG_ON(queue != &res->converting);
1552                         newlock->ml.convert_type = ml->convert_type;
1553                 }
1554                 lksb->flags |= (ml->flags &
1555                                 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1556                         
1557                 if (mres->lvb[0]) {
1558                         if (lksb->flags & DLM_LKSB_PUT_LVB) {
1559                                 /* other node was trying to update
1560                                  * lvb when node died.  recreate the
1561                                  * lksb with the updated lvb. */
1562                                 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1563                         } else {
1564                                 /* otherwise, the node is sending its 
1565                                  * most recent valid lvb info */
1566                                 BUG_ON(ml->type != LKM_EXMODE &&
1567                                        ml->type != LKM_PRMODE);
1568                                 if (res->lvb[0] && (ml->type == LKM_EXMODE ||
1569                                     memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1570                                         mlog(ML_ERROR, "received bad lvb!\n");
1571                                         __dlm_print_one_lock_resource(res);
1572                                         BUG();
1573                                 }
1574                                 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1575                         }
1576                 }
1577
1578
1579                 /* NOTE:
1580                  * wrt lock queue ordering and recovery:
1581                  *    1. order of locks on granted queue is
1582                  *       meaningless.
1583                  *    2. order of locks on converting queue is
1584                  *       LOST with the node death.  sorry charlie.
1585                  *    3. order of locks on the blocked queue is
1586                  *       also LOST.
1587                  * order of locks does not affect integrity, it
1588                  * just means that a lock request may get pushed
1589                  * back in line as a result of the node death.
1590                  * also note that for a given node the lock order
1591                  * for its secondary queue locks is preserved
1592                  * relative to each other, but clearly *not*
1593                  * preserved relative to locks from other nodes.
1594                  */
1595                 spin_lock(&res->spinlock);
1596                 dlm_lock_get(newlock);
1597                 list_add_tail(&newlock->list, queue);
1598                 spin_unlock(&res->spinlock);
1599         }
1600         mlog(0, "done running all the locks\n");
1601
1602 leave:
1603         if (ret < 0) {
1604                 mlog_errno(ret);
1605                 if (newlock)
1606                         dlm_lock_put(newlock);
1607         }
1608
1609         mlog_exit(ret);
1610         return ret;
1611 }
1612
1613 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
1614                                        struct dlm_lock_resource *res)
1615 {
1616         int i;
1617         struct list_head *queue, *iter, *iter2;
1618         struct dlm_lock *lock;
1619
1620         res->state |= DLM_LOCK_RES_RECOVERING;
1621         if (!list_empty(&res->recovering))
1622                 list_del_init(&res->recovering);
1623         list_add_tail(&res->recovering, &dlm->reco.resources);
1624
1625         /* find any pending locks and put them back on proper list */
1626         for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
1627                 queue = dlm_list_idx_to_ptr(res, i);
1628                 list_for_each_safe(iter, iter2, queue) {
1629                         lock = list_entry (iter, struct dlm_lock, list);
1630                         dlm_lock_get(lock);
1631                         if (lock->convert_pending) {
1632                                 /* move converting lock back to granted */
1633                                 BUG_ON(i != DLM_CONVERTING_LIST);
1634                                 mlog(0, "node died with convert pending "
1635                                      "on %.*s. move back to granted list.\n",
1636                                      res->lockname.len, res->lockname.name);
1637                                 dlm_revert_pending_convert(res, lock);
1638                                 lock->convert_pending = 0;
1639                         } else if (lock->lock_pending) {
1640                                 /* remove pending lock requests completely */
1641                                 BUG_ON(i != DLM_BLOCKED_LIST);
1642                                 mlog(0, "node died with lock pending "
1643                                      "on %.*s. remove from blocked list and skip.\n",
1644                                      res->lockname.len, res->lockname.name);
1645                                 /* lock will be floating until ref in
1646                                  * dlmlock_remote is freed after the network
1647                                  * call returns.  ok for it to not be on any
1648                                  * list since no ast can be called
1649                                  * (the master is dead). */
1650                                 dlm_revert_pending_lock(res, lock);
1651                                 lock->lock_pending = 0;
1652                         } else if (lock->unlock_pending) {
1653                                 /* if an unlock was in progress, treat as
1654                                  * if this had completed successfully
1655                                  * before sending this lock state to the
1656                                  * new master.  note that the dlm_unlock
1657                                  * call is still responsible for calling
1658                                  * the unlockast.  that will happen after
1659                                  * the network call times out.  for now,
1660                                  * just move lists to prepare the new
1661                                  * recovery master.  */
1662                                 BUG_ON(i != DLM_GRANTED_LIST);
1663                                 mlog(0, "node died with unlock pending "
1664                                      "on %.*s. remove from blocked list and skip.\n",
1665                                      res->lockname.len, res->lockname.name);
1666                                 dlm_commit_pending_unlock(res, lock);
1667                                 lock->unlock_pending = 0;
1668                         } else if (lock->cancel_pending) {
1669                                 /* if a cancel was in progress, treat as
1670                                  * if this had completed successfully
1671                                  * before sending this lock state to the
1672                                  * new master */
1673                                 BUG_ON(i != DLM_CONVERTING_LIST);
1674                                 mlog(0, "node died with cancel pending "
1675                                      "on %.*s. move back to granted list.\n",
1676                                      res->lockname.len, res->lockname.name);
1677                                 dlm_commit_pending_cancel(res, lock);
1678                                 lock->cancel_pending = 0;
1679                         }
1680                         dlm_lock_put(lock);
1681                 }
1682         }
1683 }
1684
1685
1686
1687 /* removes all recovered locks from the recovery list.
1688  * sets the res->owner to the new master.
1689  * unsets the RECOVERY flag and wakes waiters. */
1690 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
1691                                               u8 dead_node, u8 new_master)
1692 {
1693         int i;
1694         struct list_head *iter, *iter2;
1695         struct hlist_node *hash_iter;
1696         struct hlist_head *bucket;
1697
1698         struct dlm_lock_resource *res;
1699
1700         mlog_entry_void();
1701
1702         assert_spin_locked(&dlm->spinlock);
1703
1704         list_for_each_safe(iter, iter2, &dlm->reco.resources) {
1705                 res = list_entry (iter, struct dlm_lock_resource, recovering);
1706                 if (res->owner == dead_node) {
1707                         list_del_init(&res->recovering);
1708                         spin_lock(&res->spinlock);
1709                         dlm_change_lockres_owner(dlm, res, new_master);
1710                         res->state &= ~DLM_LOCK_RES_RECOVERING;
1711                         __dlm_dirty_lockres(dlm, res);
1712                         spin_unlock(&res->spinlock);
1713                         wake_up(&res->wq);
1714                 }
1715         }
1716
1717         /* this will become unnecessary eventually, but
1718          * for now we need to run the whole hash, clear
1719          * the RECOVERING state and set the owner
1720          * if necessary */
1721         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
1722                 bucket = &(dlm->lockres_hash[i]);
1723                 hlist_for_each_entry(res, hash_iter, bucket, hash_node) {
1724                         if (res->state & DLM_LOCK_RES_RECOVERING) {
1725                                 if (res->owner == dead_node) {
1726                                         mlog(0, "(this=%u) res %.*s owner=%u "
1727                                              "was not on recovering list, but "
1728                                              "clearing state anyway\n",
1729                                              dlm->node_num, res->lockname.len,
1730                                              res->lockname.name, new_master);
1731                                 } else if (res->owner == dlm->node_num) {
1732                                         mlog(0, "(this=%u) res %.*s owner=%u "
1733                                              "was not on recovering list, "
1734                                              "owner is THIS node, clearing\n",
1735                                              dlm->node_num, res->lockname.len,
1736                                              res->lockname.name, new_master);
1737                                 } else
1738                                         continue;
1739
1740                                 if (!list_empty(&res->recovering)) {
1741                                         mlog(0, "%s:%.*s: lockres was "
1742                                              "marked RECOVERING, owner=%u\n",
1743                                              dlm->name, res->lockname.len,
1744                                              res->lockname.name, res->owner);
1745                                         list_del_init(&res->recovering);
1746                                 }
1747                                 spin_lock(&res->spinlock);
1748                                 dlm_change_lockres_owner(dlm, res, new_master);
1749                                 res->state &= ~DLM_LOCK_RES_RECOVERING;
1750                                 __dlm_dirty_lockres(dlm, res);
1751                                 spin_unlock(&res->spinlock);
1752                                 wake_up(&res->wq);
1753                         }
1754                 }
1755         }
1756 }
1757
1758 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
1759 {
1760         if (local) {
1761                 if (lock->ml.type != LKM_EXMODE &&
1762                     lock->ml.type != LKM_PRMODE)
1763                         return 1;
1764         } else if (lock->ml.type == LKM_EXMODE)
1765                 return 1;
1766         return 0;
1767 }
1768
1769 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
1770                                struct dlm_lock_resource *res, u8 dead_node)
1771 {
1772         struct list_head *iter, *queue;
1773         struct dlm_lock *lock;
1774         int blank_lvb = 0, local = 0;
1775         int i;
1776         u8 search_node;
1777
1778         assert_spin_locked(&dlm->spinlock);
1779         assert_spin_locked(&res->spinlock);
1780
1781         if (res->owner == dlm->node_num)
1782                 /* if this node owned the lockres, and if the dead node 
1783                  * had an EX when he died, blank out the lvb */
1784                 search_node = dead_node;
1785         else {
1786                 /* if this is a secondary lockres, and we had no EX or PR
1787                  * locks granted, we can no longer trust the lvb */
1788                 search_node = dlm->node_num;
1789                 local = 1;  /* check local state for valid lvb */
1790         }
1791
1792         for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
1793                 queue = dlm_list_idx_to_ptr(res, i);
1794                 list_for_each(iter, queue) {
1795                         lock = list_entry (iter, struct dlm_lock, list);
1796                         if (lock->ml.node == search_node) {
1797                                 if (dlm_lvb_needs_invalidation(lock, local)) {
1798                                         /* zero the lksb lvb and lockres lvb */
1799                                         blank_lvb = 1;
1800                                         memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
1801                                 }
1802                         }
1803                 }
1804         }
1805
1806         if (blank_lvb) {
1807                 mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
1808                      res->lockname.len, res->lockname.name, dead_node);
1809                 memset(res->lvb, 0, DLM_LVB_LEN);
1810         }
1811 }
1812
1813 static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
1814                                 struct dlm_lock_resource *res, u8 dead_node)
1815 {
1816         struct list_head *iter, *tmpiter;
1817         struct dlm_lock *lock;
1818
1819         /* this node is the lockres master:
1820          * 1) remove any stale locks for the dead node
1821          * 2) if the dead node had an EX when he died, blank out the lvb 
1822          */
1823         assert_spin_locked(&dlm->spinlock);
1824         assert_spin_locked(&res->spinlock);
1825
1826         /* TODO: check pending_asts, pending_basts here */
1827         list_for_each_safe(iter, tmpiter, &res->granted) {
1828                 lock = list_entry (iter, struct dlm_lock, list);
1829                 if (lock->ml.node == dead_node) {
1830                         list_del_init(&lock->list);
1831                         dlm_lock_put(lock);
1832                 }
1833         }
1834         list_for_each_safe(iter, tmpiter, &res->converting) {
1835                 lock = list_entry (iter, struct dlm_lock, list);
1836                 if (lock->ml.node == dead_node) {
1837                         list_del_init(&lock->list);
1838                         dlm_lock_put(lock);
1839                 }
1840         }
1841         list_for_each_safe(iter, tmpiter, &res->blocked) {
1842                 lock = list_entry (iter, struct dlm_lock, list);
1843                 if (lock->ml.node == dead_node) {
1844                         list_del_init(&lock->list);
1845                         dlm_lock_put(lock);
1846                 }
1847         }
1848
1849         /* do not kick thread yet */
1850         __dlm_dirty_lockres(dlm, res);
1851 }
1852
1853 /* if this node is the recovery master, and there are no
1854  * locks for a given lockres owned by this node that are in
1855  * either PR or EX mode, zero out the lvb before requesting.
1856  *
1857  */
1858
1859
1860 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
1861 {
1862         struct hlist_node *iter;
1863         struct dlm_lock_resource *res;
1864         int i;
1865         struct hlist_head *bucket;
1866         struct dlm_lock *lock;
1867
1868
1869         /* purge any stale mles */
1870         dlm_clean_master_list(dlm, dead_node);
1871
1872         /*
1873          * now clean up all lock resources.  there are two rules:
1874          *
1875          * 1) if the dead node was the master, move the lockres
1876          *    to the recovering list.  set the RECOVERING flag.
1877          *    this lockres needs to be cleaned up before it can
1878          *    be used further.
1879          *
1880          * 2) if this node was the master, remove all locks from
1881          *    each of the lockres queues that were owned by the
1882          *    dead node.  once recovery finishes, the dlm thread
1883          *    can be kicked again to see if any ASTs or BASTs
1884          *    need to be fired as a result.
1885          */
1886         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
1887                 bucket = &(dlm->lockres_hash[i]);
1888                 hlist_for_each_entry(res, iter, bucket, hash_node) {
1889                         /* always prune any $RECOVERY entries for dead nodes,
1890                          * otherwise hangs can occur during later recovery */
1891                         if (dlm_is_recovery_lock(res->lockname.name,
1892                                                  res->lockname.len)) {
1893                                 spin_lock(&res->spinlock);
1894                                 list_for_each_entry(lock, &res->granted, list) {
1895                                         if (lock->ml.node == dead_node) {
1896                                                 mlog(0, "AHA! there was "
1897                                                      "a $RECOVERY lock for dead "
1898                                                      "node %u (%s)!\n",
1899                                                      dead_node, dlm->name);
1900                                                 list_del_init(&lock->list);
1901                                                 dlm_lock_put(lock);
1902                                                 break;
1903                                         }
1904                                 }
1905                                 spin_unlock(&res->spinlock);
1906                                 continue;
1907                         }                       
1908                         spin_lock(&res->spinlock);
1909                         /* zero the lvb if necessary */
1910                         dlm_revalidate_lvb(dlm, res, dead_node);
1911                         if (res->owner == dead_node)
1912                                 dlm_move_lockres_to_recovery_list(dlm, res);
1913                         else if (res->owner == dlm->node_num) {
1914                                 dlm_free_dead_locks(dlm, res, dead_node);
1915                                 __dlm_lockres_calc_usage(dlm, res);
1916                         }
1917                         spin_unlock(&res->spinlock);
1918                 }
1919         }
1920
1921 }
1922
1923 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
1924 {
1925         assert_spin_locked(&dlm->spinlock);
1926
1927         /* check to see if the node is already considered dead */
1928         if (!test_bit(idx, dlm->live_nodes_map)) {
1929                 mlog(0, "for domain %s, node %d is already dead. "
1930                      "another node likely did recovery already.\n",
1931                      dlm->name, idx);
1932                 return;
1933         }
1934
1935         /* check to see if we do not care about this node */
1936         if (!test_bit(idx, dlm->domain_map)) {
1937                 /* This also catches the case that we get a node down
1938                  * but haven't joined the domain yet. */
1939                 mlog(0, "node %u already removed from domain!\n", idx);
1940                 return;
1941         }
1942
1943         clear_bit(idx, dlm->live_nodes_map);
1944
1945         /* Clean up join state on node death. */
1946         if (dlm->joining_node == idx) {
1947                 mlog(0, "Clearing join state for node %u\n", idx);
1948                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1949         }
1950
1951         /* make sure local cleanup occurs before the heartbeat events */
1952         if (!test_bit(idx, dlm->recovery_map))
1953                 dlm_do_local_recovery_cleanup(dlm, idx);
1954
1955         /* notify anything attached to the heartbeat events */
1956         dlm_hb_event_notify_attached(dlm, idx, 0);
1957
1958         mlog(0, "node %u being removed from domain map!\n", idx);
1959         clear_bit(idx, dlm->domain_map);
1960         /* wake up migration waiters if a node goes down.
1961          * perhaps later we can genericize this for other waiters. */
1962         wake_up(&dlm->migration_wq);
1963
1964         if (test_bit(idx, dlm->recovery_map))
1965                 mlog(0, "domain %s, node %u already added "
1966                      "to recovery map!\n", dlm->name, idx);
1967         else
1968                 set_bit(idx, dlm->recovery_map);
1969 }
1970
1971 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
1972 {
1973         struct dlm_ctxt *dlm = data;
1974
1975         if (!dlm_grab(dlm))
1976                 return;
1977
1978         spin_lock(&dlm->spinlock);
1979         __dlm_hb_node_down(dlm, idx);
1980         spin_unlock(&dlm->spinlock);
1981
1982         dlm_put(dlm);
1983 }
1984
1985 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
1986 {
1987         struct dlm_ctxt *dlm = data;
1988
1989         if (!dlm_grab(dlm))
1990                 return;
1991
1992         spin_lock(&dlm->spinlock);
1993         set_bit(idx, dlm->live_nodes_map);
1994         /* do NOT notify mle attached to the heartbeat events.
1995          * new nodes are not interesting in mastery until joined. */
1996         spin_unlock(&dlm->spinlock);
1997
1998         dlm_put(dlm);
1999 }
2000
2001 static void dlm_reco_ast(void *astdata)
2002 {
2003         struct dlm_ctxt *dlm = astdata;
2004         mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2005              dlm->node_num, dlm->name);
2006 }
2007 static void dlm_reco_bast(void *astdata, int blocked_type)
2008 {
2009         struct dlm_ctxt *dlm = astdata;
2010         mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2011              dlm->node_num, dlm->name);
2012 }
2013 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
2014 {
2015         mlog(0, "unlockast for recovery lock fired!\n");
2016 }
2017
2018 /*
2019  * dlm_pick_recovery_master will continually attempt to use
2020  * dlmlock() on the special "$RECOVERY" lockres with the
2021  * LKM_NOQUEUE flag to get an EX.  every thread that enters
2022  * this function on each node racing to become the recovery
2023  * master will not stop attempting this until either:
2024  * a) this node gets the EX (and becomes the recovery master),
2025  * or b) dlm->reco.new_master gets set to some nodenum 
2026  * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2027  * so each time a recovery master is needed, the entire cluster
2028  * will sync at this point.  if the new master dies, that will
2029  * be detected in dlm_do_recovery */
2030 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2031 {
2032         enum dlm_status ret;
2033         struct dlm_lockstatus lksb;
2034         int status = -EINVAL;
2035
2036         mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2037              dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2038 again:  
2039         memset(&lksb, 0, sizeof(lksb));
2040
2041         ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
2042                       DLM_RECOVERY_LOCK_NAME, dlm_reco_ast, dlm, dlm_reco_bast);
2043
2044         mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2045              dlm->name, ret, lksb.status);
2046
2047         if (ret == DLM_NORMAL) {
2048                 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2049                      dlm->name, dlm->node_num);
2050                 
2051                 /* got the EX lock.  check to see if another node 
2052                  * just became the reco master */
2053                 if (dlm_reco_master_ready(dlm)) {
2054                         mlog(0, "%s: got reco EX lock, but %u will "
2055                              "do the recovery\n", dlm->name,
2056                              dlm->reco.new_master);
2057                         status = -EEXIST;
2058                 } else {
2059                         status = 0;
2060
2061                         /* see if recovery was already finished elsewhere */
2062                         spin_lock(&dlm->spinlock);
2063                         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
2064                                 status = -EINVAL;       
2065                                 mlog(0, "%s: got reco EX lock, but "
2066                                      "node got recovered already\n", dlm->name);
2067                                 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2068                                         mlog(ML_ERROR, "%s: new master is %u "
2069                                              "but no dead node!\n", 
2070                                              dlm->name, dlm->reco.new_master);
2071                                         BUG();
2072                                 }
2073                         }
2074                         spin_unlock(&dlm->spinlock);
2075                 }
2076
2077                 /* if this node has actually become the recovery master,
2078                  * set the master and send the messages to begin recovery */
2079                 if (!status) {
2080                         mlog(0, "%s: dead=%u, this=%u, sending "
2081                              "begin_reco now\n", dlm->name, 
2082                              dlm->reco.dead_node, dlm->node_num);
2083                         status = dlm_send_begin_reco_message(dlm,
2084                                       dlm->reco.dead_node);
2085                         /* this always succeeds */
2086                         BUG_ON(status);
2087
2088                         /* set the new_master to this node */
2089                         spin_lock(&dlm->spinlock);
2090                         dlm->reco.new_master = dlm->node_num;
2091                         spin_unlock(&dlm->spinlock);
2092                 }
2093
2094                 /* recovery lock is a special case.  ast will not get fired,
2095                  * so just go ahead and unlock it. */
2096                 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
2097                 if (ret == DLM_DENIED) {
2098                         mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2099                         ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2100                 }
2101                 if (ret != DLM_NORMAL) {
2102                         /* this would really suck. this could only happen
2103                          * if there was a network error during the unlock
2104                          * because of node death.  this means the unlock
2105                          * is actually "done" and the lock structure is
2106                          * even freed.  we can continue, but only
2107                          * because this specific lock name is special. */
2108                         mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
2109                 }
2110         } else if (ret == DLM_NOTQUEUED) {
2111                 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2112                      dlm->name, dlm->node_num);
2113                 /* another node is master. wait on
2114                  * reco.new_master != O2NM_INVALID_NODE_NUM 
2115                  * for at most one second */
2116                 wait_event_timeout(dlm->dlm_reco_thread_wq,
2117                                          dlm_reco_master_ready(dlm),
2118                                          msecs_to_jiffies(1000));
2119                 if (!dlm_reco_master_ready(dlm)) {
2120                         mlog(0, "%s: reco master taking awhile\n",
2121                              dlm->name);
2122                         goto again;
2123                 }
2124                 /* another node has informed this one that it is reco master */
2125                 mlog(0, "%s: reco master %u is ready to recover %u\n",
2126                      dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
2127                 status = -EEXIST;
2128         } else {
2129                 struct dlm_lock_resource *res;
2130
2131                 /* dlmlock returned something other than NOTQUEUED or NORMAL */
2132                 mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2133                      "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2134                      dlm_errname(lksb.status));
2135                 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2136                                          DLM_RECOVERY_LOCK_NAME_LEN);
2137                 if (res) {
2138                         dlm_print_one_lock_resource(res);
2139                         dlm_lockres_put(res);
2140                 } else {
2141                         mlog(ML_ERROR, "recovery lock not found\n");
2142                 }
2143                 BUG();
2144         }
2145
2146         return status;
2147 }
2148
2149 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2150 {
2151         struct dlm_begin_reco br;
2152         int ret = 0;
2153         struct dlm_node_iter iter;
2154         int nodenum;
2155         int status;
2156
2157         mlog_entry("%u\n", dead_node);
2158
2159         mlog(0, "dead node is %u\n", dead_node);
2160
2161         spin_lock(&dlm->spinlock);
2162         dlm_node_iter_init(dlm->domain_map, &iter);
2163         spin_unlock(&dlm->spinlock);
2164
2165         clear_bit(dead_node, iter.node_map);
2166
2167         memset(&br, 0, sizeof(br));
2168         br.node_idx = dlm->node_num;
2169         br.dead_node = dead_node;
2170
2171         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2172                 ret = 0;
2173                 if (nodenum == dead_node) {
2174                         mlog(0, "not sending begin reco to dead node "
2175                                   "%u\n", dead_node);
2176                         continue;
2177                 }
2178                 if (nodenum == dlm->node_num) {
2179                         mlog(0, "not sending begin reco to self\n");
2180                         continue;
2181                 }
2182 retry:
2183                 ret = -EINVAL;
2184                 mlog(0, "attempting to send begin reco msg to %d\n",
2185                           nodenum);
2186                 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2187                                          &br, sizeof(br), nodenum, &status);
2188                 /* negative status is handled ok by caller here */
2189                 if (ret >= 0)
2190                         ret = status;
2191                 if (dlm_is_host_down(ret)) {
2192                         /* node is down.  not involved in recovery
2193                          * so just keep going */
2194                         mlog(0, "%s: node %u was down when sending "
2195                              "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2196                         ret = 0;
2197                 }
2198                 if (ret < 0) {
2199                         struct dlm_lock_resource *res;
2200                         /* this is now a serious problem, possibly ENOMEM 
2201                          * in the network stack.  must retry */
2202                         mlog_errno(ret);
2203                         mlog(ML_ERROR, "begin reco of dlm %s to node %u "
2204                             " returned %d\n", dlm->name, nodenum, ret);
2205                         res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2206                                                  DLM_RECOVERY_LOCK_NAME_LEN);
2207                         if (res) {
2208                                 dlm_print_one_lock_resource(res);
2209                                 dlm_lockres_put(res);
2210                         } else {
2211                                 mlog(ML_ERROR, "recovery lock not found\n");
2212                         }
2213                         /* sleep for a bit in hopes that we can avoid 
2214                          * another ENOMEM */
2215                         msleep(100);
2216                         goto retry;
2217                 }
2218         }
2219
2220         return ret;
2221 }
2222
2223 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data)
2224 {
2225         struct dlm_ctxt *dlm = data;
2226         struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2227
2228         /* ok to return 0, domain has gone away */
2229         if (!dlm_grab(dlm))
2230                 return 0;
2231
2232         mlog(0, "node %u wants to recover node %u\n",
2233                   br->node_idx, br->dead_node);
2234
2235         dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2236
2237         spin_lock(&dlm->spinlock);
2238         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2239                 if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2240                         mlog(0, "%s: new_master %u died, changing "
2241                              "to %u\n", dlm->name, dlm->reco.new_master,
2242                              br->node_idx);
2243                 } else {
2244                         mlog(0, "%s: new_master %u NOT DEAD, changing "
2245                              "to %u\n", dlm->name, dlm->reco.new_master,
2246                              br->node_idx);
2247                         /* may not have seen the new master as dead yet */
2248                 }
2249         }
2250         if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2251                 mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2252                      "node %u changing it to %u\n", dlm->name, 
2253                      dlm->reco.dead_node, br->node_idx, br->dead_node);
2254         }
2255         dlm->reco.new_master = br->node_idx;
2256         dlm->reco.dead_node = br->dead_node;
2257         if (!test_bit(br->dead_node, dlm->recovery_map)) {
2258                 mlog(0, "recovery master %u sees %u as dead, but this "
2259                      "node has not yet.  marking %u as dead\n",
2260                      br->node_idx, br->dead_node, br->dead_node);
2261                 if (!test_bit(br->dead_node, dlm->domain_map) ||
2262                     !test_bit(br->dead_node, dlm->live_nodes_map))
2263                         mlog(0, "%u not in domain/live_nodes map "
2264                              "so setting it in reco map manually\n",
2265                              br->dead_node);
2266                 /* force the recovery cleanup in __dlm_hb_node_down
2267                  * both of these will be cleared in a moment */
2268                 set_bit(br->dead_node, dlm->domain_map);
2269                 set_bit(br->dead_node, dlm->live_nodes_map);
2270                 __dlm_hb_node_down(dlm, br->dead_node);
2271         }
2272         spin_unlock(&dlm->spinlock);
2273
2274         dlm_kick_recovery_thread(dlm);
2275         dlm_put(dlm);
2276         return 0;
2277 }
2278
2279 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2280 {
2281         int ret = 0;
2282         struct dlm_finalize_reco fr;
2283         struct dlm_node_iter iter;
2284         int nodenum;
2285         int status;
2286
2287         mlog(0, "finishing recovery for node %s:%u\n",
2288              dlm->name, dlm->reco.dead_node);
2289
2290         spin_lock(&dlm->spinlock);
2291         dlm_node_iter_init(dlm->domain_map, &iter);
2292         spin_unlock(&dlm->spinlock);
2293
2294         memset(&fr, 0, sizeof(fr));
2295         fr.node_idx = dlm->node_num;
2296         fr.dead_node = dlm->reco.dead_node;
2297
2298         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2299                 if (nodenum == dlm->node_num)
2300                         continue;
2301                 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2302                                          &fr, sizeof(fr), nodenum, &status);
2303                 if (ret >= 0) {
2304                         ret = status;
2305                         if (dlm_is_host_down(ret)) {
2306                                 /* this has no effect on this recovery 
2307                                  * session, so set the status to zero to 
2308                                  * finish out the last recovery */
2309                                 mlog(ML_ERROR, "node %u went down after this "
2310                                      "node finished recovery.\n", nodenum);
2311                                 ret = 0;
2312                         }
2313                 }
2314                 if (ret < 0) {
2315                         mlog_errno(ret);
2316                         break;
2317                 }
2318         }
2319
2320         return ret;
2321 }
2322
2323 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data)
2324 {
2325         struct dlm_ctxt *dlm = data;
2326         struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2327
2328         /* ok to return 0, domain has gone away */
2329         if (!dlm_grab(dlm))
2330                 return 0;
2331
2332         mlog(0, "node %u finalizing recovery of node %u\n",
2333              fr->node_idx, fr->dead_node);
2334
2335         spin_lock(&dlm->spinlock);
2336
2337         if (dlm->reco.new_master != fr->node_idx) {
2338                 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2339                      "%u is supposed to be the new master, dead=%u\n",
2340                      fr->node_idx, dlm->reco.new_master, fr->dead_node);
2341                 BUG();
2342         }
2343         if (dlm->reco.dead_node != fr->dead_node) {
2344                 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2345                      "node %u, but node %u is supposed to be dead\n",
2346                      fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2347                 BUG();
2348         }
2349
2350         dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2351
2352         spin_unlock(&dlm->spinlock);
2353
2354         dlm_reset_recovery(dlm);
2355
2356         dlm_kick_recovery_thread(dlm);
2357         dlm_put(dlm);
2358         return 0;
2359 }