projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
ocfs2/dlm: fix deadlock when dispatch assert master
[linux-2.6-block.git] / fs / ocfs2 / dlm / dlmrecovery.c
CommitLineData
6714d8e8
KH		 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>
6714d8e8
KH		 33#include <linux/init.h>
34#include <linux/sysctl.h>
35#include <linux/random.h>
36#include <linux/blkdev.h>
37#include <linux/socket.h>
38#include <linux/inet.h>
39#include <linux/timer.h>
40#include <linux/kthread.h>
b4c7f538 41#include <linux/delay.h>
6714d8e8
KH		 42
43
44#include "cluster/heartbeat.h"
45#include "cluster/nodemanager.h"
46#include "cluster/tcp.h"
47
48#include "dlmapi.h"
49#include "dlmcommon.h"
50#include "dlmdomain.h"
51
52#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY)
53#include "cluster/masklog.h"
54
55static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node);
56
57static int dlm_recovery_thread(void *data);
6714d8e8
KH		 58static int dlm_do_recovery(struct dlm_ctxt *dlm);
59
60static int dlm_pick_recovery_master(struct dlm_ctxt *dlm);
61static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node);
62static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
63static int dlm_request_all_locks(struct dlm_ctxt *dlm,
64 u8 request_from, u8 dead_node);
65static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
66
67static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res);
68static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
69 const char *lockname, int namelen,
70 int total_locks, u64 cookie,
71 u8 flags, u8 master);
72static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
73 struct dlm_migratable_lockres *mres,
74 u8 send_to,
75 struct dlm_lock_resource *res,
76 int total_locks);
6714d8e8
KH		 77static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
78 struct dlm_lock_resource *res,
79 struct dlm_migratable_lockres *mres);
6714d8e8
KH		 80static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm);
81static int dlm_send_all_done_msg(struct dlm_ctxt *dlm,
82 u8 dead_node, u8 send_to);
83static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node);
84static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
85 struct list_head *list, u8 dead_node);
86static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
87 u8 dead_node, u8 new_master);
88static void dlm_reco_ast(void *astdata);
89static void dlm_reco_bast(void *astdata, int blocked_type);
90static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st);
91static void dlm_request_all_locks_worker(struct dlm_work_item *item,
92 void *data);
93static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data);
8169cae5
AB		 94static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
95 struct dlm_lock_resource *res,
96 u8 *real_master);
6714d8e8
KH		 97
98static u64 dlm_get_next_mig_cookie(void);
99
34af946a
IM		 100static DEFINE_SPINLOCK(dlm_reco_state_lock);
101static DEFINE_SPINLOCK(dlm_mig_cookie_lock);
6714d8e8
KH		 102static u64 dlm_mig_cookie = 1;
103
104static u64 dlm_get_next_mig_cookie(void)
105{
106 u64 c;
107 spin_lock(&dlm_mig_cookie_lock);
108 c = dlm_mig_cookie;
109 if (dlm_mig_cookie == (~0ULL))
110 dlm_mig_cookie = 1;
111 else
112 dlm_mig_cookie++;
113 spin_unlock(&dlm_mig_cookie_lock);
114 return c;
115}
116
ab27eb6f
KH		 117static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm,
118 u8 dead_node)
119{
120 assert_spin_locked(&dlm->spinlock);
121 if (dlm->reco.dead_node != dead_node)
122 mlog(0, "%s: changing dead_node from %u to %u\n",
123 dlm->name, dlm->reco.dead_node, dead_node);
124 dlm->reco.dead_node = dead_node;
125}
126
127static inline void dlm_set_reco_master(struct dlm_ctxt *dlm,
128 u8 master)
129{
130 assert_spin_locked(&dlm->spinlock);
131 mlog(0, "%s: changing new_master from %u to %u\n",
132 dlm->name, dlm->reco.new_master, master);
133 dlm->reco.new_master = master;
134}
135
466d1a45 136static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm)
6714d8e8 137{
466d1a45 138 assert_spin_locked(&dlm->spinlock);
6714d8e8 139 clear_bit(dlm->reco.dead_node, dlm->recovery_map);
ab27eb6f
KH		 140 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
141 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
466d1a45
KH		 142}
143
144static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
145{
146 spin_lock(&dlm->spinlock);
147 __dlm_reset_recovery(dlm);
6714d8e8
KH		 148 spin_unlock(&dlm->spinlock);
149}
150
151/* Worker function used during recovery. */
c4028958 152void dlm_dispatch_work(struct work_struct *work)
6714d8e8 153{
c4028958
DH		 154 struct dlm_ctxt *dlm =
155 container_of(work, struct dlm_ctxt, dispatched_work);
6714d8e8 156 LIST_HEAD(tmp_list);
800deef3 157 struct dlm_work_item *item, *next;
6714d8e8 158 dlm_workfunc_t *workfunc;
3156d267
KH		 159 int tot=0;
160
6714d8e8
KH		 161 spin_lock(&dlm->work_lock);
162 list_splice_init(&dlm->work_list, &tmp_list);
163 spin_unlock(&dlm->work_lock);
164
800deef3 165 list_for_each_entry(item, &tmp_list, list) {
3156d267
KH		 166 tot++;
167 }
168 mlog(0, "%s: work thread has %d work items\n", dlm->name, tot);
169
800deef3 170 list_for_each_entry_safe(item, next, &tmp_list, list) {
6714d8e8
KH		 171 workfunc = item->func;
172 list_del_init(&item->list);
173
174 /* already have ref on dlm to avoid having
175 * it disappear. just double-check. */
176 BUG_ON(item->dlm != dlm);
177
178 /* this is allowed to sleep and
179 * call network stuff */
180 workfunc(item, item->data);
181
182 dlm_put(dlm);
183 kfree(item);
184 }
185}
186
187/*
188 * RECOVERY THREAD
189 */
190
c03872f5 191void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
6714d8e8
KH		 192{
193 /* wake the recovery thread
194 * this will wake the reco thread in one of three places
195 * 1) sleeping with no recovery happening
196 * 2) sleeping with recovery mastered elsewhere
197 * 3) recovery mastered here, waiting on reco data */
198
199 wake_up(&dlm->dlm_reco_thread_wq);
200}
201
202/* Launch the recovery thread */
203int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
204{
205 mlog(0, "starting dlm recovery thread...\n");
206
207 dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
208 "dlm_reco_thread");
209 if (IS_ERR(dlm->dlm_reco_thread_task)) {
210 mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
211 dlm->dlm_reco_thread_task = NULL;
212 return -EINVAL;
213 }
214
215 return 0;
216}
217
218void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
219{
220 if (dlm->dlm_reco_thread_task) {
221 mlog(0, "waiting for dlm recovery thread to exit\n");
222 kthread_stop(dlm->dlm_reco_thread_task);
223 dlm->dlm_reco_thread_task = NULL;
224 }
225}
226
227
228
229/*
230 * this is lame, but here's how recovery works...
231 * 1) all recovery threads cluster wide will work on recovering
232 * ONE node at a time
233 * 2) negotiate who will take over all the locks for the dead node.
234 * thats right... ALL the locks.
235 * 3) once a new master is chosen, everyone scans all locks
236 * and moves aside those mastered by the dead guy
237 * 4) each of these locks should be locked until recovery is done
238 * 5) the new master collects up all of secondary lock queue info
239 * one lock at a time, forcing each node to communicate back
240 * before continuing
241 * 6) each secondary lock queue responds with the full known lock info
242 * 7) once the new master has run all its locks, it sends a ALLDONE!
243 * message to everyone
244 * 8) upon receiving this message, the secondary queue node unlocks
245 * and responds to the ALLDONE
246 * 9) once the new master gets responses from everyone, he unlocks
247 * everything and recovery for this dead node is done
248 *10) go back to 2) while there are still dead nodes
249 *
250 */
251
d6dea6e9
KH		 252static void dlm_print_reco_node_status(struct dlm_ctxt *dlm)
253{
254 struct dlm_reco_node_data *ndata;
255 struct dlm_lock_resource *res;
256
257 mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n",
ba25f9dc 258 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
d6dea6e9
KH		 259 dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive",
260 dlm->reco.dead_node, dlm->reco.new_master);
261
262 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
263 char *st = "unknown";
264 switch (ndata->state) {
265 case DLM_RECO_NODE_DATA_INIT:
266 st = "init";
267 break;
268 case DLM_RECO_NODE_DATA_REQUESTING:
269 st = "requesting";
270 break;
271 case DLM_RECO_NODE_DATA_DEAD:
272 st = "dead";
273 break;
274 case DLM_RECO_NODE_DATA_RECEIVING:
275 st = "receiving";
276 break;
277 case DLM_RECO_NODE_DATA_REQUESTED:
278 st = "requested";
279 break;
280 case DLM_RECO_NODE_DATA_DONE:
281 st = "done";
282 break;
283 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
284 st = "finalize-sent";
285 break;
286 default:
287 st = "bad";
288 break;
289 }
290 mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n",
291 dlm->name, ndata->node_num, st);
292 }
293 list_for_each_entry(res, &dlm->reco.resources, recovering) {
294 mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n",
295 dlm->name, res->lockname.len, res->lockname.name);
296 }
297}
6714d8e8
KH		 298
299#define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
300
301static int dlm_recovery_thread(void *data)
302{
303 int status;
304 struct dlm_ctxt *dlm = data;
305 unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
306
307 mlog(0, "dlm thread running for %s...\n", dlm->name);
308
309 while (!kthread_should_stop()) {
bc9838c4 310 if (dlm_domain_fully_joined(dlm)) {
6714d8e8
KH		 311 status = dlm_do_recovery(dlm);
312 if (status == -EAGAIN) {
313 /* do not sleep, recheck immediately. */
314 continue;
315 }
316 if (status < 0)
317 mlog_errno(status);
318 }
319
320 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
321 kthread_should_stop(),
322 timeout);
323 }
324
325 mlog(0, "quitting DLM recovery thread\n");
326 return 0;
327}
328
e2faea4c
KH		 329/* returns true when the recovery master has contacted us */
330static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
331{
332 int ready;
333 spin_lock(&dlm->spinlock);
334 ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
335 spin_unlock(&dlm->spinlock);
336 return ready;
337}
338
339/* returns true if node is no longer in the domain
340 * could be dead or just not joined */
341int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
342{
343 int dead;
344 spin_lock(&dlm->spinlock);
aba9aac7 345 dead = !test_bit(node, dlm->domain_map);
e2faea4c
KH		 346 spin_unlock(&dlm->spinlock);
347 return dead;
348}
349
b7084ab5
KH		 350/* returns true if node is no longer in the domain
351 * could be dead or just not joined */
3fb5a989 352static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node)
b7084ab5
KH		 353{
354 int recovered;
355 spin_lock(&dlm->spinlock);
356 recovered = !test_bit(node, dlm->recovery_map);
357 spin_unlock(&dlm->spinlock);
358 return recovered;
359}
360
361
ed8625c6 362void dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout)
44465a7d 363{
ed8625c6
SM		 364 if (dlm_is_node_dead(dlm, node))
365 return;
366
367 printk(KERN_NOTICE "o2dlm: Waiting on the death of node %u in "
368 "domain %s\n", node, dlm->name);
369
370 if (timeout)
44465a7d 371 wait_event_timeout(dlm->dlm_reco_thread_wq,
ed8625c6
SM		 372 dlm_is_node_dead(dlm, node),
373 msecs_to_jiffies(timeout));
374 else
44465a7d
KH		 375 wait_event(dlm->dlm_reco_thread_wq,
376 dlm_is_node_dead(dlm, node));
44465a7d
KH		 377}
378
ed8625c6 379void dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout)
b7084ab5 380{
ed8625c6
SM		 381 if (dlm_is_node_recovered(dlm, node))
382 return;
383
384 printk(KERN_NOTICE "o2dlm: Waiting on the recovery of node %u in "
385 "domain %s\n", node, dlm->name);
386
387 if (timeout)
b7084ab5 388 wait_event_timeout(dlm->dlm_reco_thread_wq,
ed8625c6
SM		 389 dlm_is_node_recovered(dlm, node),
390 msecs_to_jiffies(timeout));
391 else
b7084ab5
KH		 392 wait_event(dlm->dlm_reco_thread_wq,
393 dlm_is_node_recovered(dlm, node));
b7084ab5
KH		 394}
395
6714d8e8
KH		 396/* callers of the top-level api calls (dlmlock/dlmunlock) should
397 * block on the dlm->reco.event when recovery is in progress.
398 * the dlm recovery thread will set this state when it begins
399 * recovering a dead node (as the new master or not) and clear
400 * the state and wake as soon as all affected lock resources have
401 * been marked with the RECOVERY flag */
402static int dlm_in_recovery(struct dlm_ctxt *dlm)
403{
404 int in_recovery;
405 spin_lock(&dlm->spinlock);
406 in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
407 spin_unlock(&dlm->spinlock);
408 return in_recovery;
409}
410
411
412void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
413{
56a7c104 414 if (dlm_in_recovery(dlm)) {
3b3b84a8 415 mlog(0, "%s: reco thread %d in recovery: "
56a7c104 416 "state=%d, master=%u, dead=%u\n",
ba25f9dc 417 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
56a7c104
KH		 418 dlm->reco.state, dlm->reco.new_master,
419 dlm->reco.dead_node);
420 }
6714d8e8
KH		 421 wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
422}
423
424static void dlm_begin_recovery(struct dlm_ctxt *dlm)
425{
426 spin_lock(&dlm->spinlock);
427 BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
8decab3c
SM		 428 printk(KERN_NOTICE "o2dlm: Begin recovery on domain %s for node %u\n",
429 dlm->name, dlm->reco.dead_node);
6714d8e8
KH		 430 dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
431 spin_unlock(&dlm->spinlock);
432}
433
434static void dlm_end_recovery(struct dlm_ctxt *dlm)
435{
436 spin_lock(&dlm->spinlock);
437 BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
438 dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
439 spin_unlock(&dlm->spinlock);
8decab3c 440 printk(KERN_NOTICE "o2dlm: End recovery on domain %s\n", dlm->name);
6714d8e8
KH		 441 wake_up(&dlm->reco.event);
442}
443
8decab3c
SM		 444static void dlm_print_recovery_master(struct dlm_ctxt *dlm)
445{
446 printk(KERN_NOTICE "o2dlm: Node %u (%s) is the Recovery Master for the "
447 "dead node %u in domain %s\n", dlm->reco.new_master,
448 (dlm->node_num == dlm->reco.new_master ? "me" : "he"),
449 dlm->reco.dead_node, dlm->name);
450}
451
6714d8e8
KH		 452static int dlm_do_recovery(struct dlm_ctxt *dlm)
453{
454 int status = 0;
e2faea4c 455 int ret;
6714d8e8
KH		 456
457 spin_lock(&dlm->spinlock);
458
459 /* check to see if the new master has died */
460 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
461 test_bit(dlm->reco.new_master, dlm->recovery_map)) {
462 mlog(0, "new master %u died while recovering %u!\n",
463 dlm->reco.new_master, dlm->reco.dead_node);
464 /* unset the new_master, leave dead_node */
ab27eb6f 465 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
6714d8e8
KH		 466 }
467
468 /* select a target to recover */
469 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
470 int bit;
471
f471c9df 472 bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES, 0);
6714d8e8 473 if (bit >= O2NM_MAX_NODES || bit < 0)
ab27eb6f 474 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
6714d8e8 475 else
ab27eb6f 476 dlm_set_reco_dead_node(dlm, bit);
6714d8e8
KH		 477 } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
478 /* BUG? */
479 mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
480 dlm->reco.dead_node);
ab27eb6f 481 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
6714d8e8
KH		 482 }
483
484 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
485 // mlog(0, "nothing to recover! sleeping now!\n");
486 spin_unlock(&dlm->spinlock);
487 /* return to main thread loop and sleep. */
488 return 0;
489 }
d6dea6e9 490 mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n",
ba25f9dc 491 dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
6714d8e8
KH		 492 dlm->reco.dead_node);
493 spin_unlock(&dlm->spinlock);
494
495 /* take write barrier */
496 /* (stops the list reshuffling thread, proxy ast handling) */
497 dlm_begin_recovery(dlm);
498
499 if (dlm->reco.new_master == dlm->node_num)
500 goto master_here;
501
502 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
e2faea4c
KH		 503 /* choose a new master, returns 0 if this node
504 * is the master, -EEXIST if it's another node.
505 * this does not return until a new master is chosen
506 * or recovery completes entirely. */
507 ret = dlm_pick_recovery_master(dlm);
508 if (!ret) {
6714d8e8 509 /* already notified everyone. go. */
6714d8e8
KH		 510 goto master_here;
511 }
512 mlog(0, "another node will master this recovery session.\n");
513 }
8decab3c
SM		 514
515 dlm_print_recovery_master(dlm);
6714d8e8
KH		 516
517 /* it is safe to start everything back up here
518 * because all of the dead node's lock resources
519 * have been marked as in-recovery */
520 dlm_end_recovery(dlm);
521
522 /* sleep out in main dlm_recovery_thread loop. */
523 return 0;
524
525master_here:
8decab3c 526 dlm_print_recovery_master(dlm);
6714d8e8
KH		 527
528 status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
529 if (status < 0) {
6a413211 530 /* we should never hit this anymore */
8decab3c
SM		 531 mlog(ML_ERROR, "%s: Error %d remastering locks for node %u, "
532 "retrying.\n", dlm->name, status, dlm->reco.dead_node);
e2faea4c
KH		 533 /* yield a bit to allow any final network messages
534 * to get handled on remaining nodes */
535 msleep(100);
6714d8e8
KH		 536 } else {
537 /* success! see if any other nodes need recovery */
e2faea4c
KH		 538 mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
539 dlm->name, dlm->reco.dead_node, dlm->node_num);
ded2cf71
JB		 540 spin_lock(&dlm->spinlock);
541 __dlm_reset_recovery(dlm);
542 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
543 spin_unlock(&dlm->spinlock);
6714d8e8
KH		 544 }
545 dlm_end_recovery(dlm);
546
547 /* continue and look for another dead node */
548 return -EAGAIN;
549}
550
551static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
552{
553 int status = 0;
554 struct dlm_reco_node_data *ndata;
6714d8e8
KH		 555 int all_nodes_done;
556 int destroy = 0;
557 int pass = 0;
558
6a413211
KH		 559 do {
560 /* we have become recovery master. there is no escaping
561 * this, so just keep trying until we get it. */
562 status = dlm_init_recovery_area(dlm, dead_node);
563 if (status < 0) {
564 mlog(ML_ERROR, "%s: failed to alloc recovery area, "
565 "retrying\n", dlm->name);
566 msleep(1000);
567 }
568 } while (status != 0);
6714d8e8
KH		 569
570 /* safe to access the node data list without a lock, since this
571 * process is the only one to change the list */
800deef3 572 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
6714d8e8
KH		 573 BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
574 ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
575
8decab3c 576 mlog(0, "%s: Requesting lock info from node %u\n", dlm->name,
6714d8e8
KH		 577 ndata->node_num);
578
579 if (ndata->node_num == dlm->node_num) {
580 ndata->state = DLM_RECO_NODE_DATA_DONE;
581 continue;
582 }
583
6a413211
KH		 584 do {
585 status = dlm_request_all_locks(dlm, ndata->node_num,
586 dead_node);
587 if (status < 0) {
588 mlog_errno(status);
589 if (dlm_is_host_down(status)) {
590 /* node died, ignore it for recovery */
591 status = 0;
592 ndata->state = DLM_RECO_NODE_DATA_DEAD;
593 /* wait for the domain map to catch up
594 * with the network state. */
595 wait_event_timeout(dlm->dlm_reco_thread_wq,
596 dlm_is_node_dead(dlm,
597 ndata->node_num),
598 msecs_to_jiffies(1000));
599 mlog(0, "waited 1 sec for %u, "
600 "dead? %s\n", ndata->node_num,
601 dlm_is_node_dead(dlm, ndata->node_num) ?
602 "yes" : "no");
603 } else {
604 /* -ENOMEM on the other node */
605 mlog(0, "%s: node %u returned "
606 "%d during recovery, retrying "
607 "after a short wait\n",
608 dlm->name, ndata->node_num,
609 status);
610 msleep(100);
611 }
6714d8e8 612 }
6a413211 613 } while (status != 0);
6714d8e8 614
756a1501 615 spin_lock(&dlm_reco_state_lock);
6714d8e8
KH		 616 switch (ndata->state) {
617 case DLM_RECO_NODE_DATA_INIT:
618 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
619 case DLM_RECO_NODE_DATA_REQUESTED:
620 BUG();
621 break;
622 case DLM_RECO_NODE_DATA_DEAD:
623 mlog(0, "node %u died after requesting "
624 "recovery info for node %u\n",
625 ndata->node_num, dead_node);
6a413211
KH		 626 /* fine. don't need this node's info.
627 * continue without it. */
628 break;
6714d8e8
KH		 629 case DLM_RECO_NODE_DATA_REQUESTING:
630 ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
631 mlog(0, "now receiving recovery data from "
632 "node %u for dead node %u\n",
633 ndata->node_num, dead_node);
634 break;
635 case DLM_RECO_NODE_DATA_RECEIVING:
636 mlog(0, "already receiving recovery data from "
637 "node %u for dead node %u\n",
638 ndata->node_num, dead_node);
639 break;
640 case DLM_RECO_NODE_DATA_DONE:
641 mlog(0, "already DONE receiving recovery data "
642 "from node %u for dead node %u\n",
643 ndata->node_num, dead_node);
644 break;
645 }
756a1501 646 spin_unlock(&dlm_reco_state_lock);
6714d8e8
KH		 647 }
648
8decab3c 649 mlog(0, "%s: Done requesting all lock info\n", dlm->name);
6714d8e8
KH		 650
651 /* nodes should be sending reco data now
652 * just need to wait */
653
654 while (1) {
655 /* check all the nodes now to see if we are
656 * done, or if anyone died */
657 all_nodes_done = 1;
658 spin_lock(&dlm_reco_state_lock);
800deef3 659 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
6714d8e8
KH		 660 mlog(0, "checking recovery state of node %u\n",
661 ndata->node_num);
662 switch (ndata->state) {
663 case DLM_RECO_NODE_DATA_INIT:
664 case DLM_RECO_NODE_DATA_REQUESTING:
665 mlog(ML_ERROR, "bad ndata state for "
666 "node %u: state=%d\n",
667 ndata->node_num, ndata->state);
668 BUG();
669 break;
670 case DLM_RECO_NODE_DATA_DEAD:
6a413211 671 mlog(0, "node %u died after "
6714d8e8
KH		 672 "requesting recovery info for "
673 "node %u\n", ndata->node_num,
674 dead_node);
6a413211 675 break;
6714d8e8
KH		 676 case DLM_RECO_NODE_DATA_RECEIVING:
677 case DLM_RECO_NODE_DATA_REQUESTED:
d6dea6e9
KH		 678 mlog(0, "%s: node %u still in state %s\n",
679 dlm->name, ndata->node_num,
680 ndata->state==DLM_RECO_NODE_DATA_RECEIVING ?
681 "receiving" : "requested");
6714d8e8
KH		 682 all_nodes_done = 0;
683 break;
684 case DLM_RECO_NODE_DATA_DONE:
d6dea6e9
KH		 685 mlog(0, "%s: node %u state is done\n",
686 dlm->name, ndata->node_num);
6714d8e8
KH		 687 break;
688 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
d6dea6e9
KH		 689 mlog(0, "%s: node %u state is finalize\n",
690 dlm->name, ndata->node_num);
6714d8e8
KH		 691 break;
692 }
693 }
694 spin_unlock(&dlm_reco_state_lock);
695
696 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
697 all_nodes_done?"yes":"no");
698 if (all_nodes_done) {
699 int ret;
700
ded2cf71
JB		 701 /* Set this flag on recovery master to avoid
702 * a new recovery for another dead node start
703 * before the recovery is not done. That may
704 * cause recovery hung.*/
705 spin_lock(&dlm->spinlock);
706 dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
707 spin_unlock(&dlm->spinlock);
708
6714d8e8
KH		 709 /* all nodes are now in DLM_RECO_NODE_DATA_DONE state
710 * just send a finalize message to everyone and
711 * clean up */
712 mlog(0, "all nodes are done! send finalize\n");
713 ret = dlm_send_finalize_reco_message(dlm);
714 if (ret < 0)
715 mlog_errno(ret);
716
717 spin_lock(&dlm->spinlock);
718 dlm_finish_local_lockres_recovery(dlm, dead_node,
719 dlm->node_num);
720 spin_unlock(&dlm->spinlock);
721 mlog(0, "should be done with recovery!\n");
722
723 mlog(0, "finishing recovery of %s at %lu, "
724 "dead=%u, this=%u, new=%u\n", dlm->name,
725 jiffies, dlm->reco.dead_node,
726 dlm->node_num, dlm->reco.new_master);
727 destroy = 1;
6a413211 728 status = 0;
6714d8e8
KH		 729 /* rescan everything marked dirty along the way */
730 dlm_kick_thread(dlm, NULL);
731 break;
732 }
733 /* wait to be signalled, with periodic timeout
734 * to check for node death */
735 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
736 kthread_should_stop(),
737 msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
738
739 }
740
6714d8e8
KH		 741 if (destroy)
742 dlm_destroy_recovery_area(dlm, dead_node);
743
6714d8e8
KH		 744 return status;
745}
746
747static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
748{
749 int num=0;
750 struct dlm_reco_node_data *ndata;
751
752 spin_lock(&dlm->spinlock);
753 memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
754 /* nodes can only be removed (by dying) after dropping
755 * this lock, and death will be trapped later, so this should do */
756 spin_unlock(&dlm->spinlock);
757
758 while (1) {
759 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
760 if (num >= O2NM_MAX_NODES) {
761 break;
762 }
763 BUG_ON(num == dead_node);
764
cd861280 765 ndata = kzalloc(sizeof(*ndata), GFP_NOFS);
6714d8e8
KH		 766 if (!ndata) {
767 dlm_destroy_recovery_area(dlm, dead_node);
768 return -ENOMEM;
769 }
770 ndata->node_num = num;
771 ndata->state = DLM_RECO_NODE_DATA_INIT;
772 spin_lock(&dlm_reco_state_lock);
773 list_add_tail(&ndata->list, &dlm->reco.node_data);
774 spin_unlock(&dlm_reco_state_lock);
775 num++;
776 }
777
778 return 0;
779}
780
781static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
782{
800deef3 783 struct dlm_reco_node_data *ndata, *next;
6714d8e8
KH		 784 LIST_HEAD(tmplist);
785
786 spin_lock(&dlm_reco_state_lock);
787 list_splice_init(&dlm->reco.node_data, &tmplist);
788 spin_unlock(&dlm_reco_state_lock);
789
800deef3 790 list_for_each_entry_safe(ndata, next, &tmplist, list) {
6714d8e8
KH		 791 list_del_init(&ndata->list);
792 kfree(ndata);
793 }
794}
795
796static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
797 u8 dead_node)
798{
799 struct dlm_lock_request lr;
22ab9014 800 int ret;
98ac9125 801 int status;
6714d8e8
KH		 802
803 mlog(0, "\n");
804
805
806 mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
807 "to %u\n", dead_node, request_from);
808
809 memset(&lr, 0, sizeof(lr));
810 lr.node_idx = dlm->node_num;
811 lr.dead_node = dead_node;
812
813 // send message
6714d8e8 814 ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
98ac9125 815 &lr, sizeof(lr), request_from, &status);
6714d8e8
KH		 816
817 /* negative status is handled by caller */
818 if (ret < 0)
8decab3c
SM		 819 mlog(ML_ERROR, "%s: Error %d send LOCK_REQUEST to node %u "
820 "to recover dead node %u\n", dlm->name, ret,
821 request_from, dead_node);
98ac9125
X		 822 else
823 ret = status;
6714d8e8
KH		 824 // return from here, then
825 // sleep until all received or error
826 return ret;
827
828}
829
d74c9803
KH		 830int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
831 void **ret_data)
6714d8e8
KH		 832{
833 struct dlm_ctxt *dlm = data;
834 struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
835 char *buf = NULL;
836 struct dlm_work_item *item = NULL;
837
838 if (!dlm_grab(dlm))
839 return -EINVAL;
840
c3187ce5
KH		 841 if (lr->dead_node != dlm->reco.dead_node) {
842 mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local "
843 "dead_node is %u\n", dlm->name, lr->node_idx,
844 lr->dead_node, dlm->reco.dead_node);
d6dea6e9 845 dlm_print_reco_node_status(dlm);
c3187ce5
KH		 846 /* this is a hack */
847 dlm_put(dlm);
848 return -ENOMEM;
849 }
6714d8e8
KH		 850 BUG_ON(lr->dead_node != dlm->reco.dead_node);
851
cd861280 852 item = kzalloc(sizeof(*item), GFP_NOFS);
6714d8e8
KH		 853 if (!item) {
854 dlm_put(dlm);
855 return -ENOMEM;
856 }
857
858 /* this will get freed by dlm_request_all_locks_worker */
ad8100e0 859 buf = (char *) __get_free_page(GFP_NOFS);
6714d8e8
KH		 860 if (!buf) {
861 kfree(item);
862 dlm_put(dlm);
863 return -ENOMEM;
864 }
865
866 /* queue up work for dlm_request_all_locks_worker */
867 dlm_grab(dlm); /* get an extra ref for the work item */
868 dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
869 item->u.ral.reco_master = lr->node_idx;
870 item->u.ral.dead_node = lr->dead_node;
871 spin_lock(&dlm->work_lock);
872 list_add_tail(&item->list, &dlm->work_list);
873 spin_unlock(&dlm->work_lock);
3156d267 874 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
6714d8e8
KH		 875
876 dlm_put(dlm);
877 return 0;
878}
879
880static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
881{
882 struct dlm_migratable_lockres *mres;
883 struct dlm_lock_resource *res;
884 struct dlm_ctxt *dlm;
885 LIST_HEAD(resources);
6714d8e8
KH		 886 int ret;
887 u8 dead_node, reco_master;
29c0fa0f 888 int skip_all_done = 0;
6714d8e8
KH		 889
890 dlm = item->dlm;
891 dead_node = item->u.ral.dead_node;
892 reco_master = item->u.ral.reco_master;
e2faea4c
KH		 893 mres = (struct dlm_migratable_lockres *)data;
894
d6dea6e9
KH		 895 mlog(0, "%s: recovery worker started, dead=%u, master=%u\n",
896 dlm->name, dead_node, reco_master);
897
e2faea4c
KH		 898 if (dead_node != dlm->reco.dead_node ||
899 reco_master != dlm->reco.new_master) {
6a413211
KH		 900 /* worker could have been created before the recovery master
901 * died. if so, do not continue, but do not error. */
902 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
903 mlog(ML_NOTICE, "%s: will not send recovery state, "
904 "recovery master %u died, thread=(dead=%u,mas=%u)"
905 " current=(dead=%u,mas=%u)\n", dlm->name,
906 reco_master, dead_node, reco_master,
907 dlm->reco.dead_node, dlm->reco.new_master);
908 } else {
909 mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, "
910 "master=%u), request(dead=%u, master=%u)\n",
911 dlm->name, dlm->reco.dead_node,
912 dlm->reco.new_master, dead_node, reco_master);
913 }
914 goto leave;
e2faea4c 915 }
6714d8e8 916
6714d8e8
KH		 917 /* lock resources should have already been moved to the
918 * dlm->reco.resources list. now move items from that list
919 * to a temp list if the dead owner matches. note that the
920 * whole cluster recovers only one node at a time, so we
921 * can safely move UNKNOWN lock resources for each recovery
922 * session. */
923 dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
924
925 /* now we can begin blasting lockreses without the dlm lock */
29c0fa0f
KH		 926
927 /* any errors returned will be due to the new_master dying,
928 * the dlm_reco_thread should detect this */
800deef3 929 list_for_each_entry(res, &resources, recovering) {
6714d8e8
KH		 930 ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
931 DLM_MRES_RECOVERY);
29c0fa0f 932 if (ret < 0) {
d6dea6e9
KH		 933 mlog(ML_ERROR, "%s: node %u went down while sending "
934 "recovery state for dead node %u, ret=%d\n", dlm->name,
935 reco_master, dead_node, ret);
29c0fa0f
KH		 936 skip_all_done = 1;
937 break;
938 }
6714d8e8
KH		 939 }
940
941 /* move the resources back to the list */
942 spin_lock(&dlm->spinlock);
943 list_splice_init(&resources, &dlm->reco.resources);
944 spin_unlock(&dlm->spinlock);
945
29c0fa0f
KH		 946 if (!skip_all_done) {
947 ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
948 if (ret < 0) {
d6dea6e9
KH		 949 mlog(ML_ERROR, "%s: node %u went down while sending "
950 "recovery all-done for dead node %u, ret=%d\n",
951 dlm->name, reco_master, dead_node, ret);
29c0fa0f
KH		 952 }
953 }
6a413211 954leave:
6714d8e8
KH		 955 free_page((unsigned long)data);
956}
957
958
959static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
960{
961 int ret, tmpret;
962 struct dlm_reco_data_done done_msg;
963
964 memset(&done_msg, 0, sizeof(done_msg));
965 done_msg.node_idx = dlm->node_num;
966 done_msg.dead_node = dead_node;
967 mlog(0, "sending DATA DONE message to %u, "
968 "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
969 done_msg.dead_node);
970
971 ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
972 sizeof(done_msg), send_to, &tmpret);
29c0fa0f 973 if (ret < 0) {
8decab3c
SM		 974 mlog(ML_ERROR, "%s: Error %d send RECO_DATA_DONE to node %u "
975 "to recover dead node %u\n", dlm->name, ret, send_to,
976 dead_node);
29c0fa0f 977 if (!dlm_is_host_down(ret)) {
29c0fa0f
KH		 978 BUG();
979 }
980 } else
6714d8e8
KH		 981 ret = tmpret;
982 return ret;
983}
984
985
d74c9803
KH		 986int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
987 void **ret_data)
6714d8e8
KH		 988{
989 struct dlm_ctxt *dlm = data;
990 struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
6714d8e8
KH		 991 struct dlm_reco_node_data *ndata = NULL;
992 int ret = -EINVAL;
993
994 if (!dlm_grab(dlm))
995 return -EINVAL;
996
997 mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
998 "node_idx=%u, this node=%u\n", done->dead_node,
999 dlm->reco.dead_node, done->node_idx, dlm->node_num);
d6dea6e9
KH		 1000
1001 mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node),
1002 "Got DATA DONE: dead_node=%u, reco.dead_node=%u, "
1003 "node_idx=%u, this node=%u\n", done->dead_node,
1004 dlm->reco.dead_node, done->node_idx, dlm->node_num);
6714d8e8
KH		 1005
1006 spin_lock(&dlm_reco_state_lock);
800deef3 1007 list_for_each_entry(ndata, &dlm->reco.node_data, list) {
6714d8e8
KH		 1008 if (ndata->node_num != done->node_idx)
1009 continue;
1010
1011 switch (ndata->state) {
e2faea4c 1012 /* should have moved beyond INIT but not to FINALIZE yet */
6714d8e8
KH		 1013 case DLM_RECO_NODE_DATA_INIT:
1014 case DLM_RECO_NODE_DATA_DEAD:
6714d8e8
KH		 1015 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
1016 mlog(ML_ERROR, "bad ndata state for node %u:"
1017 " state=%d\n", ndata->node_num,
1018 ndata->state);
1019 BUG();
1020 break;
e2faea4c
KH		 1021 /* these states are possible at this point, anywhere along
1022 * the line of recovery */
1023 case DLM_RECO_NODE_DATA_DONE:
6714d8e8
KH		 1024 case DLM_RECO_NODE_DATA_RECEIVING:
1025 case DLM_RECO_NODE_DATA_REQUESTED:
1026 case DLM_RECO_NODE_DATA_REQUESTING:
1027 mlog(0, "node %u is DONE sending "
1028 "recovery data!\n",
1029 ndata->node_num);
1030
1031 ndata->state = DLM_RECO_NODE_DATA_DONE;
1032 ret = 0;
1033 break;
1034 }
1035 }
1036 spin_unlock(&dlm_reco_state_lock);
1037
1038 /* wake the recovery thread, some node is done */
1039 if (!ret)
1040 dlm_kick_recovery_thread(dlm);
1041
1042 if (ret < 0)
1043 mlog(ML_ERROR, "failed to find recovery node data for node "
1044 "%u\n", done->node_idx);
1045 dlm_put(dlm);
1046
1047 mlog(0, "leaving reco data done handler, ret=%d\n", ret);
1048 return ret;
1049}
1050
1051static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
1052 struct list_head *list,
1053 u8 dead_node)
1054{
800deef3 1055 struct dlm_lock_resource *res, *next;
e2faea4c 1056 struct dlm_lock *lock;
6714d8e8
KH		 1057
1058 spin_lock(&dlm->spinlock);
800deef3 1059 list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
e2faea4c
KH		 1060 /* always prune any $RECOVERY entries for dead nodes,
1061 * otherwise hangs can occur during later recovery */
6714d8e8 1062 if (dlm_is_recovery_lock(res->lockname.name,
e2faea4c
KH		 1063 res->lockname.len)) {
1064 spin_lock(&res->spinlock);
1065 list_for_each_entry(lock, &res->granted, list) {
1066 if (lock->ml.node == dead_node) {
1067 mlog(0, "AHA! there was "
1068 "a $RECOVERY lock for dead "
2bd63216 1069 "node %u (%s)!\n",
e2faea4c
KH		 1070 dead_node, dlm->name);
1071 list_del_init(&lock->list);
1072 dlm_lock_put(lock);
b934beaf
X		 1073 /* Can't schedule DLM_UNLOCK_FREE_LOCK
1074 * - do manually */
1075 dlm_lock_put(lock);
e2faea4c
KH		 1076 break;
1077 }
1078 }
1079 spin_unlock(&res->spinlock);
6714d8e8 1080 continue;
e2faea4c
KH		 1081 }
1082
6714d8e8
KH		 1083 if (res->owner == dead_node) {
1084 mlog(0, "found lockres owned by dead node while "
1085 "doing recovery for node %u. sending it.\n",
1086 dead_node);
f116629d 1087 list_move_tail(&res->recovering, list);
6714d8e8
KH		 1088 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1089 mlog(0, "found UNKNOWN owner while doing recovery "
1090 "for node %u. sending it.\n", dead_node);
f116629d 1091 list_move_tail(&res->recovering, list);
6714d8e8
KH		 1092 }
1093 }
1094 spin_unlock(&dlm->spinlock);
1095}
1096
1097static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
1098{
1099 int total_locks = 0;
1100 struct list_head *iter, *queue = &res->granted;
1101 int i;
1102
1103 for (i=0; i<3; i++) {
1104 list_for_each(iter, queue)
1105 total_locks++;
1106 queue++;
1107 }
1108 return total_locks;
1109}
1110
1111
1112static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
1113 struct dlm_migratable_lockres *mres,
1114 u8 send_to,
1115 struct dlm_lock_resource *res,
1116 int total_locks)
1117{
1118 u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
1119 int mres_total_locks = be32_to_cpu(mres->total_locks);
1120 int sz, ret = 0, status = 0;
1121 u8 orig_flags = mres->flags,
1122 orig_master = mres->master;
1123
1124 BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
1125 if (!mres->num_locks)
1126 return 0;
1127
1128 sz = sizeof(struct dlm_migratable_lockres) +
1129 (mres->num_locks * sizeof(struct dlm_migratable_lock));
1130
1131 /* add an all-done flag if we reached the last lock */
1132 orig_flags = mres->flags;
1133 BUG_ON(total_locks > mres_total_locks);
1134 if (total_locks == mres_total_locks)
1135 mres->flags |= DLM_MRES_ALL_DONE;
1136
ba2bf218
KH		 1137 mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n",
1138 dlm->name, res->lockname.len, res->lockname.name,
17ae26b6 1139 orig_flags & DLM_MRES_MIGRATION ? "migration" : "recovery",
ba2bf218
KH		 1140 send_to);
1141
6714d8e8
KH		 1142 /* send it */
1143 ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
1144 sz, send_to, &status);
1145 if (ret < 0) {
1146 /* XXX: negative status is not handled.
1147 * this will end up killing this node. */
8decab3c
SM		 1148 mlog(ML_ERROR, "%s: res %.*s, Error %d send MIG_LOCKRES to "
1149 "node %u (%s)\n", dlm->name, mres->lockname_len,
1150 mres->lockname, ret, send_to,
1151 (orig_flags & DLM_MRES_MIGRATION ?
1152 "migration" : "recovery"));
6714d8e8
KH		 1153 } else {
1154 /* might get an -ENOMEM back here */
1155 ret = status;
1156 if (ret < 0) {
1157 mlog_errno(ret);
1158
1159 if (ret == -EFAULT) {
1160 mlog(ML_ERROR, "node %u told me to kill "
1161 "myself!\n", send_to);
1162 BUG();
1163 }
1164 }
1165 }
1166
1167 /* zero and reinit the message buffer */
1168 dlm_init_migratable_lockres(mres, res->lockname.name,
1169 res->lockname.len, mres_total_locks,
1170 mig_cookie, orig_flags, orig_master);
1171 return ret;
1172}
1173
1174static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
1175 const char *lockname, int namelen,
1176 int total_locks, u64 cookie,
1177 u8 flags, u8 master)
1178{
1179 /* mres here is one full page */
5fb0f7f0 1180 clear_page(mres);
6714d8e8
KH		 1181 mres->lockname_len = namelen;
1182 memcpy(mres->lockname, lockname, namelen);
1183 mres->num_locks = 0;
1184 mres->total_locks = cpu_to_be32(total_locks);
1185 mres->mig_cookie = cpu_to_be64(cookie);
1186 mres->flags = flags;
1187 mres->master = master;
1188}
1189
71656fa6
SM		 1190static void dlm_prepare_lvb_for_migration(struct dlm_lock *lock,
1191 struct dlm_migratable_lockres *mres,
1192 int queue)
1193{
1194 if (!lock->lksb)
1195 return;
1196
1197 /* Ignore lvb in all locks in the blocked list */
1198 if (queue == DLM_BLOCKED_LIST)
1199 return;
1200
1201 /* Only consider lvbs in locks with granted EX or PR lock levels */
1202 if (lock->ml.type != LKM_EXMODE && lock->ml.type != LKM_PRMODE)
1203 return;
1204
1205 if (dlm_lvb_is_empty(mres->lvb)) {
1206 memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1207 return;
1208 }
1209
1210 /* Ensure the lvb copied for migration matches in other valid locks */
1211 if (!memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))
1212 return;
1213
1214 mlog(ML_ERROR, "Mismatched lvb in lock cookie=%u:%llu, name=%.*s, "
1215 "node=%u\n",
1216 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
1217 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
1218 lock->lockres->lockname.len, lock->lockres->lockname.name,
1219 lock->ml.node);
1220 dlm_print_one_lock_resource(lock->lockres);
1221 BUG();
1222}
6714d8e8
KH		 1223
1224/* returns 1 if this lock fills the network structure,
1225 * 0 otherwise */
1226static int dlm_add_lock_to_array(struct dlm_lock *lock,
1227 struct dlm_migratable_lockres *mres, int queue)
1228{
1229 struct dlm_migratable_lock *ml;
1230 int lock_num = mres->num_locks;
1231
1232 ml = &(mres->ml[lock_num]);
1233 ml->cookie = lock->ml.cookie;
1234 ml->type = lock->ml.type;
1235 ml->convert_type = lock->ml.convert_type;
1236 ml->highest_blocked = lock->ml.highest_blocked;
1237 ml->list = queue;
1238 if (lock->lksb) {
1239 ml->flags = lock->lksb->flags;
71656fa6 1240 dlm_prepare_lvb_for_migration(lock, mres, queue);
6714d8e8
KH		 1241 }
1242 ml->node = lock->ml.node;
1243 mres->num_locks++;
1244 /* we reached the max, send this network message */
1245 if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1246 return 1;
1247 return 0;
1248}
1249
ba2bf218
KH		 1250static void dlm_add_dummy_lock(struct dlm_ctxt *dlm,
1251 struct dlm_migratable_lockres *mres)
1252{
1253 struct dlm_lock dummy;
1254 memset(&dummy, 0, sizeof(dummy));
1255 dummy.ml.cookie = 0;
1256 dummy.ml.type = LKM_IVMODE;
1257 dummy.ml.convert_type = LKM_IVMODE;
1258 dummy.ml.highest_blocked = LKM_IVMODE;
1259 dummy.lksb = NULL;
1260 dummy.ml.node = dlm->node_num;
1261 dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST);
1262}
1263
1264static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm,
1265 struct dlm_migratable_lock *ml,
1266 u8 *nodenum)
1267{
1268 if (unlikely(ml->cookie == 0 &&
1269 ml->type == LKM_IVMODE &&
1270 ml->convert_type == LKM_IVMODE &&
1271 ml->highest_blocked == LKM_IVMODE &&
1272 ml->list == DLM_BLOCKED_LIST)) {
1273 *nodenum = ml->node;
1274 return 1;
1275 }
1276 return 0;
1277}
6714d8e8
KH		 1278
1279int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1280 struct dlm_migratable_lockres *mres,
1281 u8 send_to, u8 flags)
1282{
800deef3 1283 struct list_head *queue;
6714d8e8
KH		 1284 int total_locks, i;
1285 u64 mig_cookie = 0;
1286 struct dlm_lock *lock;
1287 int ret = 0;
1288
1289 BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1290
1291 mlog(0, "sending to %u\n", send_to);
1292
1293 total_locks = dlm_num_locks_in_lockres(res);
1294 if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1295 /* rare, but possible */
1296 mlog(0, "argh. lockres has %d locks. this will "
1297 "require more than one network packet to "
1298 "migrate\n", total_locks);
1299 mig_cookie = dlm_get_next_mig_cookie();
1300 }
1301
1302 dlm_init_migratable_lockres(mres, res->lockname.name,
1303 res->lockname.len, total_locks,
1304 mig_cookie, flags, res->owner);
1305
1306 total_locks = 0;
1307 for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1308 queue = dlm_list_idx_to_ptr(res, i);
800deef3 1309 list_for_each_entry(lock, queue, list) {
6714d8e8
KH		 1310 /* add another lock. */
1311 total_locks++;
1312 if (!dlm_add_lock_to_array(lock, mres, i))
1313 continue;
1314
1315 /* this filled the lock message,
1316 * we must send it immediately. */
1317 ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1318 res, total_locks);
29c0fa0f
KH		 1319 if (ret < 0)
1320 goto error;
6714d8e8
KH		 1321 }
1322 }
ba2bf218
KH		 1323 if (total_locks == 0) {
1324 /* send a dummy lock to indicate a mastery reference only */
1325 mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n",
1326 dlm->name, res->lockname.len, res->lockname.name,
1327 send_to, flags & DLM_MRES_RECOVERY ? "recovery" :
1328 "migration");
1329 dlm_add_dummy_lock(dlm, mres);
1330 }
6714d8e8
KH		 1331 /* flush any remaining locks */
1332 ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
29c0fa0f
KH		 1333 if (ret < 0)
1334 goto error;
1335 return ret;
1336
1337error:
1338 mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n",
1339 dlm->name, ret);
1340 if (!dlm_is_host_down(ret))
6714d8e8 1341 BUG();
29c0fa0f
KH		 1342 mlog(0, "%s: node %u went down while sending %s "
1343 "lockres %.*s\n", dlm->name, send_to,
1344 flags & DLM_MRES_RECOVERY ? "recovery" : "migration",
1345 res->lockname.len, res->lockname.name);
6714d8e8
KH		 1346 return ret;
1347}
1348
1349
1350
1351/*
1352 * this message will contain no more than one page worth of
1353 * recovery data, and it will work on only one lockres.
1354 * there may be many locks in this page, and we may need to wait
1355 * for additional packets to complete all the locks (rare, but
1356 * possible).
1357 */
1358/*
1359 * NOTE: the allocation error cases here are scary
1360 * we really cannot afford to fail an alloc in recovery
1361 * do we spin? returning an error only delays the problem really
1362 */
1363
d74c9803
KH		 1364int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
1365 void **ret_data)
6714d8e8
KH		 1366{
1367 struct dlm_ctxt *dlm = data;
1368 struct dlm_migratable_lockres *mres =
1369 (struct dlm_migratable_lockres *)msg->buf;
1370 int ret = 0;
1371 u8 real_master;
52987e2a 1372 u8 extra_refs = 0;
6714d8e8
KH		 1373 char *buf = NULL;
1374 struct dlm_work_item *item = NULL;
1375 struct dlm_lock_resource *res = NULL;
1376
1377 if (!dlm_grab(dlm))
1378 return -EINVAL;
1379
1380 BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1381
1382 real_master = mres->master;
1383 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1384 /* cannot migrate a lockres with no master */
1385 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1386 }
1387
1388 mlog(0, "%s message received from node %u\n",
1389 (mres->flags & DLM_MRES_RECOVERY) ?
1390 "recovery" : "migration", mres->master);
1391 if (mres->flags & DLM_MRES_ALL_DONE)
1392 mlog(0, "all done flag. all lockres data received!\n");
1393
1394 ret = -ENOMEM;
ad8100e0 1395 buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS);
cd861280 1396 item = kzalloc(sizeof(*item), GFP_NOFS);
6714d8e8
KH		 1397 if (!buf || !item)
1398 goto leave;
1399
1400 /* lookup the lock to see if we have a secondary queue for this
1401 * already... just add the locks in and this will have its owner
1402 * and RECOVERY flag changed when it completes. */
1403 res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len);
1404 if (res) {
1405 /* this will get a ref on res */
1406 /* mark it as recovering/migrating and hash it */
1407 spin_lock(&res->spinlock);
1408 if (mres->flags & DLM_MRES_RECOVERY) {
1409 res->state |= DLM_LOCK_RES_RECOVERING;
1410 } else {
1411 if (res->state & DLM_LOCK_RES_MIGRATING) {
1412 /* this is at least the second
1413 * lockres message */
1414 mlog(0, "lock %.*s is already migrating\n",
1415 mres->lockname_len,
1416 mres->lockname);
1417 } else if (res->state & DLM_LOCK_RES_RECOVERING) {
1418 /* caller should BUG */
1419 mlog(ML_ERROR, "node is attempting to migrate "
1420 "lock %.*s, but marked as recovering!\n",
1421 mres->lockname_len, mres->lockname);
1422 ret = -EFAULT;
1423 spin_unlock(&res->spinlock);
27749f2f 1424 dlm_lockres_put(res);
6714d8e8
KH		 1425 goto leave;
1426 }
1427 res->state |= DLM_LOCK_RES_MIGRATING;
1428 }
1429 spin_unlock(&res->spinlock);
1430 } else {
1431 /* need to allocate, just like if it was
1432 * mastered here normally */
1433 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1434 if (!res)
1435 goto leave;
1436
1437 /* to match the ref that we would have gotten if
1438 * dlm_lookup_lockres had succeeded */
1439 dlm_lockres_get(res);
1440
1441 /* mark it as recovering/migrating and hash it */
1442 if (mres->flags & DLM_MRES_RECOVERY)
1443 res->state |= DLM_LOCK_RES_RECOVERING;
1444 else
1445 res->state |= DLM_LOCK_RES_MIGRATING;
1446
1447 spin_lock(&dlm->spinlock);
1448 __dlm_insert_lockres(dlm, res);
1449 spin_unlock(&dlm->spinlock);
1450
52987e2a
SM		 1451 /* Add an extra ref for this lock-less lockres lest the
1452 * dlm_thread purges it before we get the chance to add
1453 * locks to it */
1454 dlm_lockres_get(res);
1455
1456 /* There are three refs that need to be put.
1457 * 1. Taken above.
1458 * 2. kref_init in dlm_new_lockres()->dlm_init_lockres().
1459 * 3. dlm_lookup_lockres()
1460 * The first one is handled at the end of this function. The
1461 * other two are handled in the worker thread after locks have
1462 * been attached. Yes, we don't wait for purge time to match
1463 * kref_init. The lockres will still have atleast one ref
1464 * added because it is in the hash __dlm_insert_lockres() */
1465 extra_refs++;
1466
6714d8e8
KH		 1467 /* now that the new lockres is inserted,
1468 * make it usable by other processes */
1469 spin_lock(&res->spinlock);
1470 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1471 spin_unlock(&res->spinlock);
a6fa3640 1472 wake_up(&res->wq);
6714d8e8
KH		 1473 }
1474
1475 /* at this point we have allocated everything we need,
1476 * and we have a hashed lockres with an extra ref and
1477 * the proper res->state flags. */
1478 ret = 0;
ba2bf218
KH		 1479 spin_lock(&res->spinlock);
1480 /* drop this either when master requery finds a different master
1481 * or when a lock is added by the recovery worker */
1482 dlm_lockres_grab_inflight_ref(dlm, res);
6714d8e8
KH		 1483 if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1484 /* migration cannot have an unknown master */
1485 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1486 mlog(0, "recovery has passed me a lockres with an "
1487 "unknown owner.. will need to requery: "
1488 "%.*s\n", mres->lockname_len, mres->lockname);
1489 } else {
ba2bf218
KH		 1490 /* take a reference now to pin the lockres, drop it
1491 * when locks are added in the worker */
6714d8e8 1492 dlm_change_lockres_owner(dlm, res, dlm->node_num);
6714d8e8 1493 }
ba2bf218 1494 spin_unlock(&res->spinlock);
6714d8e8
KH		 1495
1496 /* queue up work for dlm_mig_lockres_worker */
1497 dlm_grab(dlm); /* get an extra ref for the work item */
1498 memcpy(buf, msg->buf, be16_to_cpu(msg->data_len)); /* copy the whole message */
1499 dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1500 item->u.ml.lockres = res; /* already have a ref */
1501 item->u.ml.real_master = real_master;
52987e2a 1502 item->u.ml.extra_ref = extra_refs;
6714d8e8
KH		 1503 spin_lock(&dlm->work_lock);
1504 list_add_tail(&item->list, &dlm->work_list);
1505 spin_unlock(&dlm->work_lock);
3156d267 1506 queue_work(dlm->dlm_worker, &dlm->dispatched_work);
6714d8e8
KH		 1507
1508leave:
52987e2a
SM		 1509 /* One extra ref taken needs to be put here */
1510 if (extra_refs)
1511 dlm_lockres_put(res);
1512
6714d8e8
KH		 1513 dlm_put(dlm);
1514 if (ret < 0) {
7cfa74d1
SK		 1515 kfree(buf);
1516 kfree(item);
c1e8d35e 1517 mlog_errno(ret);
6714d8e8
KH		 1518 }
1519
6714d8e8
KH		 1520 return ret;
1521}
1522
1523
1524static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1525{
52987e2a 1526 struct dlm_ctxt *dlm;
6714d8e8
KH		 1527 struct dlm_migratable_lockres *mres;
1528 int ret = 0;
1529 struct dlm_lock_resource *res;
1530 u8 real_master;
52987e2a 1531 u8 extra_ref;
6714d8e8
KH		 1532
1533 dlm = item->dlm;
1534 mres = (struct dlm_migratable_lockres *)data;
1535
1536 res = item->u.ml.lockres;
1537 real_master = item->u.ml.real_master;
52987e2a 1538 extra_ref = item->u.ml.extra_ref;
6714d8e8
KH		 1539
1540 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1541 /* this case is super-rare. only occurs if
1542 * node death happens during migration. */
1543again:
1544 ret = dlm_lockres_master_requery(dlm, res, &real_master);
1545 if (ret < 0) {
e2faea4c 1546 mlog(0, "dlm_lockres_master_requery ret=%d\n",
6714d8e8
KH		 1547 ret);
1548 goto again;
1549 }
1550 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1551 mlog(0, "lockres %.*s not claimed. "
1552 "this node will take it.\n",
1553 res->lockname.len, res->lockname.name);
1554 } else {
ba2bf218
KH		 1555 spin_lock(&res->spinlock);
1556 dlm_lockres_drop_inflight_ref(dlm, res);
1557 spin_unlock(&res->spinlock);
6714d8e8
KH		 1558 mlog(0, "master needs to respond to sender "
1559 "that node %u still owns %.*s\n",
1560 real_master, res->lockname.len,
1561 res->lockname.name);
1562 /* cannot touch this lockres */
1563 goto leave;
1564 }
1565 }
1566
1567 ret = dlm_process_recovery_data(dlm, res, mres);
1568 if (ret < 0)
1569 mlog(0, "dlm_process_recovery_data returned %d\n", ret);
1570 else
1571 mlog(0, "dlm_process_recovery_data succeeded\n");
1572
1573 if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1574 (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1575 ret = dlm_finish_migration(dlm, res, mres->master);
1576 if (ret < 0)
1577 mlog_errno(ret);
1578 }
1579
1580leave:
52987e2a
SM		 1581 /* See comment in dlm_mig_lockres_handler() */
1582 if (res) {
1583 if (extra_ref)
1584 dlm_lockres_put(res);
1585 dlm_lockres_put(res);
1586 }
6714d8e8 1587 kfree(data);
6714d8e8
KH		 1588}
1589
1590
1591
8169cae5
AB		 1592static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1593 struct dlm_lock_resource *res,
1594 u8 *real_master)
6714d8e8
KH		 1595{
1596 struct dlm_node_iter iter;
1597 int nodenum;
1598 int ret = 0;
1599
1600 *real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1601
1602 /* we only reach here if one of the two nodes in a
1603 * migration died while the migration was in progress.
1604 * at this point we need to requery the master. we
1605 * know that the new_master got as far as creating
1606 * an mle on at least one node, but we do not know
1607 * if any nodes had actually cleared the mle and set
1608 * the master to the new_master. the old master
1609 * is supposed to set the owner to UNKNOWN in the
1610 * event of a new_master death, so the only possible
1611 * responses that we can get from nodes here are
1612 * that the master is new_master, or that the master
1613 * is UNKNOWN.
1614 * if all nodes come back with UNKNOWN then we know
1615 * the lock needs remastering here.
1616 * if any node comes back with a valid master, check
1617 * to see if that master is the one that we are
1618 * recovering. if so, then the new_master died and
1619 * we need to remaster this lock. if not, then the
1620 * new_master survived and that node will respond to
1621 * other nodes about the owner.
1622 * if there is an owner, this node needs to dump this
1623 * lockres and alert the sender that this lockres
1624 * was rejected. */
1625 spin_lock(&dlm->spinlock);
1626 dlm_node_iter_init(dlm->domain_map, &iter);
1627 spin_unlock(&dlm->spinlock);
1628
1629 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1630 /* do not send to self */
1631 if (nodenum == dlm->node_num)
1632 continue;
1633 ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1634 if (ret < 0) {
1635 mlog_errno(ret);
c03872f5
KH		 1636 if (!dlm_is_host_down(ret))
1637 BUG();
1638 /* host is down, so answer for that node would be
1639 * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
6714d8e8
KH		 1640 }
1641 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1642 mlog(0, "lock master is %u\n", *real_master);
1643 break;
1644 }
1645 }
1646 return ret;
1647}
1648
1649
c03872f5
KH		 1650int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1651 u8 nodenum, u8 *real_master)
6714d8e8
KH		 1652{
1653 int ret = -EINVAL;
1654 struct dlm_master_requery req;
1655 int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1656
1657 memset(&req, 0, sizeof(req));
1658 req.node_idx = dlm->node_num;
1659 req.namelen = res->lockname.len;
1660 memcpy(req.name, res->lockname.name, res->lockname.len);
1661
f08736bd 1662resend:
6714d8e8
KH		 1663 ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1664 &req, sizeof(req), nodenum, &status);
6714d8e8 1665 if (ret < 0)
a5196ec5
WW		 1666 mlog(ML_ERROR, "Error %d when sending message %u (key "
1667 "0x%x) to node %u\n", ret, DLM_MASTER_REQUERY_MSG,
1668 dlm->key, nodenum);
f08736bd
JQ		 1669 else if (status == -ENOMEM) {
1670 mlog_errno(status);
1671 msleep(50);
1672 goto resend;
1673 } else {
6714d8e8
KH		 1674 BUG_ON(status < 0);
1675 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1676 *real_master = (u8) (status & 0xff);
1677 mlog(0, "node %u responded to master requery with %u\n",
1678 nodenum, *real_master);
1679 ret = 0;
1680 }
1681 return ret;
1682}
1683
1684
1685/* this function cannot error, so unless the sending
1686 * or receiving of the message failed, the owner can
1687 * be trusted */
d74c9803
KH		 1688int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
1689 void **ret_data)
6714d8e8
KH		 1690{
1691 struct dlm_ctxt *dlm = data;
1692 struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1693 struct dlm_lock_resource *res = NULL;
a3d33291 1694 unsigned int hash;
6714d8e8
KH		 1695 int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1696 u32 flags = DLM_ASSERT_MASTER_REQUERY;
012572d4 1697 int dispatched = 0;
6714d8e8
KH		 1698
1699 if (!dlm_grab(dlm)) {
1700 /* since the domain has gone away on this
1701 * node, the proper response is UNKNOWN */
1702 return master;
1703 }
1704
a3d33291
MF		 1705 hash = dlm_lockid_hash(req->name, req->namelen);
1706
6714d8e8 1707 spin_lock(&dlm->spinlock);
a3d33291 1708 res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash);
6714d8e8
KH		 1709 if (res) {
1710 spin_lock(&res->spinlock);
1711 master = res->owner;
1712 if (master == dlm->node_num) {
1713 int ret = dlm_dispatch_assert_master(dlm, res,
1714 0, 0, flags);
1715 if (ret < 0) {
f08736bd
JQ		 1716 mlog_errno(ret);
1717 spin_unlock(&res->spinlock);
1718 dlm_lockres_put(res);
1719 spin_unlock(&dlm->spinlock);
1720 dlm_put(dlm);
1721 /* sender will take care of this and retry */
1722 return ret;
012572d4
JQ		 1723 } else {
1724 dispatched = 1;
ac4fef4d 1725 __dlm_lockres_grab_inflight_worker(dlm, res);
012572d4 1726 }
9a7e6b5a 1727 spin_unlock(&res->spinlock);
1728 } else {
1729 /* put.. incase we are not the master */
1730 spin_unlock(&res->spinlock);
52987e2a 1731 dlm_lockres_put(res);
9a7e6b5a 1732 }
6714d8e8
KH		 1733 }
1734 spin_unlock(&dlm->spinlock);
1735
012572d4
JQ		 1736 if (!dispatched)
1737 dlm_put(dlm);
6714d8e8
KH		 1738 return master;
1739}
1740
1741static inline struct list_head *
1742dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1743{
1744 struct list_head *ret;
1745 BUG_ON(list_num < 0);
1746 BUG_ON(list_num > 2);
1747 ret = &(res->granted);
1748 ret += list_num;
1749 return ret;
1750}
1751/* TODO: do ast flush business
1752 * TODO: do MIGRATING and RECOVERING spinning
1753 */
1754
1755/*
1756* NOTE about in-flight requests during migration:
1757*
1758* Before attempting the migrate, the master has marked the lockres as
1759* MIGRATING and then flushed all of its pending ASTS. So any in-flight
1760* requests either got queued before the MIGRATING flag got set, in which
1761* case the lock data will reflect the change and a return message is on
1762* the way, or the request failed to get in before MIGRATING got set. In
1763* this case, the caller will be told to spin and wait for the MIGRATING
1764* flag to be dropped, then recheck the master.
1765* This holds true for the convert, cancel and unlock cases, and since lvb
1766* updates are tied to these same messages, it applies to lvb updates as
1767* well. For the lock case, there is no way a lock can be on the master
1768* queue and not be on the secondary queue since the lock is always added
1769* locally first. This means that the new target node will never be sent
1770* a lock that he doesn't already have on the list.
1771* In total, this means that the local lock is correct and should not be
1772* updated to match the one sent by the master. Any messages sent back
1773* from the master before the MIGRATING flag will bring the lock properly
1774* up-to-date, and the change will be ordered properly for the waiter.
1775* We will *not* attempt to modify the lock underneath the waiter.
1776*/
1777
1778static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1779 struct dlm_lock_resource *res,
1780 struct dlm_migratable_lockres *mres)
1781{
1782 struct dlm_migratable_lock *ml;
e527b22c 1783 struct list_head *queue, *iter;
e17e75ec 1784 struct list_head *tmpq = NULL;
6714d8e8
KH		 1785 struct dlm_lock *newlock = NULL;
1786 struct dlm_lockstatus *lksb = NULL;
1787 int ret = 0;
e17e75ec 1788 int i, j, bad;
34aa8dac 1789 struct dlm_lock *lock;
ba2bf218
KH		 1790 u8 from = O2NM_MAX_NODES;
1791 unsigned int added = 0;
26636bf6 1792 __be64 c;
6714d8e8
KH		 1793
1794 mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1795 for (i=0; i<mres->num_locks; i++) {
1796 ml = &(mres->ml[i]);
ba2bf218
KH		 1797
1798 if (dlm_is_dummy_lock(dlm, ml, &from)) {
1799 /* placeholder, just need to set the refmap bit */
1800 BUG_ON(mres->num_locks != 1);
1801 mlog(0, "%s:%.*s: dummy lock for %u\n",
1802 dlm->name, mres->lockname_len, mres->lockname,
1803 from);
1804 spin_lock(&res->spinlock);
8d400b81 1805 dlm_lockres_set_refmap_bit(dlm, res, from);
ba2bf218
KH		 1806 spin_unlock(&res->spinlock);
1807 added++;
1808 break;
1809 }
6714d8e8
KH		 1810 BUG_ON(ml->highest_blocked != LKM_IVMODE);
1811 newlock = NULL;
1812 lksb = NULL;
1813
1814 queue = dlm_list_num_to_pointer(res, ml->list);
e17e75ec 1815 tmpq = NULL;
6714d8e8
KH		 1816
1817 /* if the lock is for the local node it needs to
1818 * be moved to the proper location within the queue.
1819 * do not allocate a new lock structure. */
1820 if (ml->node == dlm->node_num) {
1821 /* MIGRATION ONLY! */
1822 BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1823
34aa8dac 1824 lock = NULL;
6714d8e8 1825 spin_lock(&res->spinlock);
e17e75ec
KH		 1826 for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) {
1827 tmpq = dlm_list_idx_to_ptr(res, j);
e527b22c
AM		 1828 list_for_each(iter, tmpq) {
1829 lock = list_entry(iter,
1830 struct dlm_lock, list);
34aa8dac 1831 if (lock->ml.cookie == ml->cookie)
e17e75ec 1832 break;
34aa8dac 1833 lock = NULL;
e17e75ec
KH		 1834 }
1835 if (lock)
6714d8e8
KH		 1836 break;
1837 }
1838
1839 /* lock is always created locally first, and
1840 * destroyed locally last. it must be on the list */
1841 if (!lock) {
26636bf6
SM		 1842 c = ml->cookie;
1843 mlog(ML_ERROR, "Could not find local lock "
1844 "with cookie %u:%llu, node %u, "
1845 "list %u, flags 0x%x, type %d, "
1846 "conv %d, highest blocked %d\n",
74aa2585 1847 dlm_get_lock_cookie_node(be64_to_cpu(c)),
26636bf6
SM		 1848 dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1849 ml->node, ml->list, ml->flags, ml->type,
1850 ml->convert_type, ml->highest_blocked);
1851 __dlm_print_one_lock_resource(res);
1852 BUG();
1853 }
1854
1855 if (lock->ml.node != ml->node) {
1856 c = lock->ml.cookie;
1857 mlog(ML_ERROR, "Mismatched node# in lock "
1858 "cookie %u:%llu, name %.*s, node %u\n",
1859 dlm_get_lock_cookie_node(be64_to_cpu(c)),
1860 dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1861 res->lockname.len, res->lockname.name,
1862 lock->ml.node);
1863 c = ml->cookie;
1864 mlog(ML_ERROR, "Migrate lock cookie %u:%llu, "
1865 "node %u, list %u, flags 0x%x, type %d, "
1866 "conv %d, highest blocked %d\n",
1867 dlm_get_lock_cookie_node(be64_to_cpu(c)),
1868 dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1869 ml->node, ml->list, ml->flags, ml->type,
1870 ml->convert_type, ml->highest_blocked);
71ac1062 1871 __dlm_print_one_lock_resource(res);
6714d8e8
KH		 1872 BUG();
1873 }
6714d8e8 1874
e17e75ec 1875 if (tmpq != queue) {
26636bf6
SM		 1876 c = ml->cookie;
1877 mlog(0, "Lock cookie %u:%llu was on list %u "
1878 "instead of list %u for %.*s\n",
1879 dlm_get_lock_cookie_node(be64_to_cpu(c)),
1880 dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1881 j, ml->list, res->lockname.len,
1882 res->lockname.name);
1883 __dlm_print_one_lock_resource(res);
e17e75ec
KH		 1884 spin_unlock(&res->spinlock);
1885 continue;
1886 }
1887
6714d8e8
KH		 1888 /* see NOTE above about why we do not update
1889 * to match the master here */
1890
1891 /* move the lock to its proper place */
1892 /* do not alter lock refcount. switching lists. */
f116629d 1893 list_move_tail(&lock->list, queue);
6714d8e8 1894 spin_unlock(&res->spinlock);
ba2bf218 1895 added++;
6714d8e8
KH		 1896
1897 mlog(0, "just reordered a local lock!\n");
1898 continue;
1899 }
1900
1901 /* lock is for another node. */
1902 newlock = dlm_new_lock(ml->type, ml->node,
1903 be64_to_cpu(ml->cookie), NULL);
1904 if (!newlock) {
1905 ret = -ENOMEM;
1906 goto leave;
1907 }
1908 lksb = newlock->lksb;
1909 dlm_lock_attach_lockres(newlock, res);
1910
1911 if (ml->convert_type != LKM_IVMODE) {
1912 BUG_ON(queue != &res->converting);
1913 newlock->ml.convert_type = ml->convert_type;
1914 }
1915 lksb->flags |= (ml->flags &
1916 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
ccd8b1f9
KH		 1917
1918 if (ml->type == LKM_NLMODE)
1919 goto skip_lvb;
1920
750e3c65
X		 1921 /*
1922 * If the lock is in the blocked list it can't have a valid lvb,
1923 * so skip it
1924 */
1925 if (ml->list == DLM_BLOCKED_LIST)
1926 goto skip_lvb;
1927
8bc674cb 1928 if (!dlm_lvb_is_empty(mres->lvb)) {
6714d8e8
KH		 1929 if (lksb->flags & DLM_LKSB_PUT_LVB) {
1930 /* other node was trying to update
1931 * lvb when node died. recreate the
1932 * lksb with the updated lvb. */
1933 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
ccd8b1f9
KH		 1934 /* the lock resource lvb update must happen
1935 * NOW, before the spinlock is dropped.
1936 * we no longer wait for the AST to update
1937 * the lvb. */
1938 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
6714d8e8 1939 } else {
2bd63216 1940 /* otherwise, the node is sending its
6714d8e8
KH		 1941 * most recent valid lvb info */
1942 BUG_ON(ml->type != LKM_EXMODE &&
1943 ml->type != LKM_PRMODE);
8bc674cb 1944 if (!dlm_lvb_is_empty(res->lvb) &&
ccd8b1f9
KH		 1945 (ml->type == LKM_EXMODE ||
1946 memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1947 int i;
1948 mlog(ML_ERROR, "%s:%.*s: received bad "
1949 "lvb! type=%d\n", dlm->name,
1950 res->lockname.len,
1951 res->lockname.name, ml->type);
1952 printk("lockres lvb=[");
1953 for (i=0; i<DLM_LVB_LEN; i++)
1954 printk("%02x", res->lvb[i]);
1955 printk("]\nmigrated lvb=[");
1956 for (i=0; i<DLM_LVB_LEN; i++)
1957 printk("%02x", mres->lvb[i]);
1958 printk("]\n");
1959 dlm_print_one_lock_resource(res);
1960 BUG();
6714d8e8
KH		 1961 }
1962 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1963 }
1964 }
ccd8b1f9 1965skip_lvb:
6714d8e8
KH		 1966
1967 /* NOTE:
1968 * wrt lock queue ordering and recovery:
1969 * 1. order of locks on granted queue is
1970 * meaningless.
1971 * 2. order of locks on converting queue is
1972 * LOST with the node death. sorry charlie.
1973 * 3. order of locks on the blocked queue is
1974 * also LOST.
1975 * order of locks does not affect integrity, it
1976 * just means that a lock request may get pushed
1977 * back in line as a result of the node death.
1978 * also note that for a given node the lock order
1979 * for its secondary queue locks is preserved
1980 * relative to each other, but clearly *not*
1981 * preserved relative to locks from other nodes.
1982 */
c3187ce5 1983 bad = 0;
6714d8e8 1984 spin_lock(&res->spinlock);
c3187ce5
KH		 1985 list_for_each_entry(lock, queue, list) {
1986 if (lock->ml.cookie == ml->cookie) {
26636bf6 1987 c = lock->ml.cookie;
c3187ce5
KH		 1988 mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already "
1989 "exists on this lockres!\n", dlm->name,
1990 res->lockname.len, res->lockname.name,
74aa2585
KH		 1991 dlm_get_lock_cookie_node(be64_to_cpu(c)),
1992 dlm_get_lock_cookie_seq(be64_to_cpu(c)));
c3187ce5
KH		 1993
1994 mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, "
1995 "node=%u, cookie=%u:%llu, queue=%d\n",
1996 ml->type, ml->convert_type, ml->node,
74aa2585
KH		 1997 dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)),
1998 dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)),
c3187ce5
KH		 1999 ml->list);
2000
2001 __dlm_print_one_lock_resource(res);
2002 bad = 1;
2003 break;
2004 }
2005 }
2006 if (!bad) {
2007 dlm_lock_get(newlock);
6718cb5e
X		 2008 if (mres->flags & DLM_MRES_RECOVERY &&
2009 ml->list == DLM_CONVERTING_LIST &&
2010 newlock->ml.type >
2011 newlock->ml.convert_type) {
2012 /* newlock is doing downconvert, add it to the
2013 * head of converting list */
2014 list_add(&newlock->list, queue);
2015 } else
2016 list_add_tail(&newlock->list, queue);
ba2bf218
KH		 2017 mlog(0, "%s:%.*s: added lock for node %u, "
2018 "setting refmap bit\n", dlm->name,
2019 res->lockname.len, res->lockname.name, ml->node);
8d400b81 2020 dlm_lockres_set_refmap_bit(dlm, res, ml->node);
ba2bf218 2021 added++;
c3187ce5 2022 }
6714d8e8
KH		 2023 spin_unlock(&res->spinlock);
2024 }
2025 mlog(0, "done running all the locks\n");
2026
2027leave:
ba2bf218 2028 /* balance the ref taken when the work was queued */
50635f15
KH		 2029 spin_lock(&res->spinlock);
2030 dlm_lockres_drop_inflight_ref(dlm, res);
2031 spin_unlock(&res->spinlock);
ba2bf218 2032
eb4f73b4 2033 if (ret < 0)
6714d8e8 2034 mlog_errno(ret);
6714d8e8 2035
6714d8e8
KH		 2036 return ret;
2037}
2038
2039void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
2040 struct dlm_lock_resource *res)
2041{
2042 int i;
800deef3
CH		 2043 struct list_head *queue;
2044 struct dlm_lock *lock, *next;
6714d8e8 2045
a524812b
WW		 2046 assert_spin_locked(&dlm->spinlock);
2047 assert_spin_locked(&res->spinlock);
6714d8e8 2048 res->state |= DLM_LOCK_RES_RECOVERING;
69d72b06
KH		 2049 if (!list_empty(&res->recovering)) {
2050 mlog(0,
2051 "Recovering res %s:%.*s, is already on recovery list!\n",
2052 dlm->name, res->lockname.len, res->lockname.name);
6714d8e8 2053 list_del_init(&res->recovering);
52987e2a 2054 dlm_lockres_put(res);
69d72b06
KH		 2055 }
2056 /* We need to hold a reference while on the recovery list */
2057 dlm_lockres_get(res);
6714d8e8
KH		 2058 list_add_tail(&res->recovering, &dlm->reco.resources);
2059
2060 /* find any pending locks and put them back on proper list */
2061 for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
2062 queue = dlm_list_idx_to_ptr(res, i);
800deef3 2063 list_for_each_entry_safe(lock, next, queue, list) {
6714d8e8
KH		 2064 dlm_lock_get(lock);
2065 if (lock->convert_pending) {
2066 /* move converting lock back to granted */
2067 BUG_ON(i != DLM_CONVERTING_LIST);
2068 mlog(0, "node died with convert pending "
2069 "on %.*s. move back to granted list.\n",
2070 res->lockname.len, res->lockname.name);
2071 dlm_revert_pending_convert(res, lock);
2072 lock->convert_pending = 0;
2073 } else if (lock->lock_pending) {
2074 /* remove pending lock requests completely */
2075 BUG_ON(i != DLM_BLOCKED_LIST);
2076 mlog(0, "node died with lock pending "
2077 "on %.*s. remove from blocked list and skip.\n",
2078 res->lockname.len, res->lockname.name);
2079 /* lock will be floating until ref in
2080 * dlmlock_remote is freed after the network
2081 * call returns. ok for it to not be on any
2082 * list since no ast can be called
2083 * (the master is dead). */
2084 dlm_revert_pending_lock(res, lock);
2085 lock->lock_pending = 0;
2086 } else if (lock->unlock_pending) {
2087 /* if an unlock was in progress, treat as
2088 * if this had completed successfully
2089 * before sending this lock state to the
2090 * new master. note that the dlm_unlock
2091 * call is still responsible for calling
2092 * the unlockast. that will happen after
2093 * the network call times out. for now,
2094 * just move lists to prepare the new
2095 * recovery master. */
2096 BUG_ON(i != DLM_GRANTED_LIST);
2097 mlog(0, "node died with unlock pending "
2098 "on %.*s. remove from blocked list and skip.\n",
2099 res->lockname.len, res->lockname.name);
2100 dlm_commit_pending_unlock(res, lock);
2101 lock->unlock_pending = 0;
2102 } else if (lock->cancel_pending) {
2103 /* if a cancel was in progress, treat as
2104 * if this had completed successfully
2105 * before sending this lock state to the
2106 * new master */
2107 BUG_ON(i != DLM_CONVERTING_LIST);
2108 mlog(0, "node died with cancel pending "
2109 "on %.*s. move back to granted list.\n",
2110 res->lockname.len, res->lockname.name);
2111 dlm_commit_pending_cancel(res, lock);
2112 lock->cancel_pending = 0;
2113 }
2114 dlm_lock_put(lock);
2115 }
2116 }
2117}
2118
2119
2120
2121/* removes all recovered locks from the recovery list.
2122 * sets the res->owner to the new master.
2123 * unsets the RECOVERY flag and wakes waiters. */
2124static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
2125 u8 dead_node, u8 new_master)
2126{
2127 int i;
81f2094a 2128 struct hlist_head *bucket;
800deef3 2129 struct dlm_lock_resource *res, *next;
6714d8e8 2130
6714d8e8
KH		 2131 assert_spin_locked(&dlm->spinlock);
2132
800deef3 2133 list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
6714d8e8 2134 if (res->owner == dead_node) {
0afbba13
SM		 2135 mlog(0, "%s: res %.*s, Changing owner from %u to %u\n",
2136 dlm->name, res->lockname.len, res->lockname.name,
2137 res->owner, new_master);
6714d8e8
KH		 2138 list_del_init(&res->recovering);
2139 spin_lock(&res->spinlock);
ba2bf218
KH		 2140 /* new_master has our reference from
2141 * the lock state sent during recovery */
6714d8e8
KH		 2142 dlm_change_lockres_owner(dlm, res, new_master);
2143 res->state &= ~DLM_LOCK_RES_RECOVERING;
ba2bf218 2144 if (__dlm_lockres_has_locks(res))
69d72b06 2145 __dlm_dirty_lockres(dlm, res);
6714d8e8
KH		 2146 spin_unlock(&res->spinlock);
2147 wake_up(&res->wq);
69d72b06 2148 dlm_lockres_put(res);
6714d8e8
KH		 2149 }
2150 }
2151
2152 /* this will become unnecessary eventually, but
2153 * for now we need to run the whole hash, clear
2154 * the RECOVERING state and set the owner
2155 * if necessary */
81f2094a 2156 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
03d864c0 2157 bucket = dlm_lockres_hash(dlm, i);
b67bfe0d 2158 hlist_for_each_entry(res, bucket, hash_node) {
0afbba13
SM		 2159 if (!(res->state & DLM_LOCK_RES_RECOVERING))
2160 continue;
2161
2162 if (res->owner != dead_node &&
2163 res->owner != dlm->node_num)
2164 continue;
2165
2166 if (!list_empty(&res->recovering)) {
2167 list_del_init(&res->recovering);
2168 dlm_lockres_put(res);
6714d8e8 2169 }
0afbba13
SM		 2170
2171 /* new_master has our reference from
2172 * the lock state sent during recovery */
2173 mlog(0, "%s: res %.*s, Changing owner from %u to %u\n",
2174 dlm->name, res->lockname.len, res->lockname.name,
2175 res->owner, new_master);
2176 spin_lock(&res->spinlock);
2177 dlm_change_lockres_owner(dlm, res, new_master);
2178 res->state &= ~DLM_LOCK_RES_RECOVERING;
2179 if (__dlm_lockres_has_locks(res))
2180 __dlm_dirty_lockres(dlm, res);
2181 spin_unlock(&res->spinlock);
2182 wake_up(&res->wq);
6714d8e8
KH		 2183 }
2184 }
2185}
2186
2187static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
2188{
2189 if (local) {
2190 if (lock->ml.type != LKM_EXMODE &&
2191 lock->ml.type != LKM_PRMODE)
2192 return 1;
2193 } else if (lock->ml.type == LKM_EXMODE)
2194 return 1;
2195 return 0;
2196}
2197
2198static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
2199 struct dlm_lock_resource *res, u8 dead_node)
2200{
800deef3 2201 struct list_head *queue;
6714d8e8
KH		 2202 struct dlm_lock *lock;
2203 int blank_lvb = 0, local = 0;
2204 int i;
2205 u8 search_node;
2206
2207 assert_spin_locked(&dlm->spinlock);
2208 assert_spin_locked(&res->spinlock);
2209
2210 if (res->owner == dlm->node_num)
2bd63216 2211 /* if this node owned the lockres, and if the dead node
6714d8e8
KH		 2212 * had an EX when he died, blank out the lvb */
2213 search_node = dead_node;
2214 else {
2215 /* if this is a secondary lockres, and we had no EX or PR
2216 * locks granted, we can no longer trust the lvb */
2217 search_node = dlm->node_num;
2218 local = 1; /* check local state for valid lvb */
2219 }
2220
2221 for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
2222 queue = dlm_list_idx_to_ptr(res, i);
800deef3 2223 list_for_each_entry(lock, queue, list) {
6714d8e8
KH		 2224 if (lock->ml.node == search_node) {
2225 if (dlm_lvb_needs_invalidation(lock, local)) {
2226 /* zero the lksb lvb and lockres lvb */
2227 blank_lvb = 1;
2228 memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
2229 }
2230 }
2231 }
2232 }
2233
2234 if (blank_lvb) {
2235 mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
2236 res->lockname.len, res->lockname.name, dead_node);
2237 memset(res->lvb, 0, DLM_LVB_LEN);
2238 }
2239}
2240
2241static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
2242 struct dlm_lock_resource *res, u8 dead_node)
2243{
800deef3 2244 struct dlm_lock *lock, *next;
ba2bf218 2245 unsigned int freed = 0;
6714d8e8
KH		 2246
2247 /* this node is the lockres master:
2248 * 1) remove any stale locks for the dead node
2bd63216 2249 * 2) if the dead node had an EX when he died, blank out the lvb
6714d8e8
KH		 2250 */
2251 assert_spin_locked(&dlm->spinlock);
2252 assert_spin_locked(&res->spinlock);
2253
2c5c54ac
SM		 2254 /* We do two dlm_lock_put(). One for removing from list and the other is
2255 * to force the DLM_UNLOCK_FREE_LOCK action so as to free the locks */
2256
6714d8e8 2257 /* TODO: check pending_asts, pending_basts here */
800deef3 2258 list_for_each_entry_safe(lock, next, &res->granted, list) {
6714d8e8
KH		 2259 if (lock->ml.node == dead_node) {
2260 list_del_init(&lock->list);
2261 dlm_lock_put(lock);
2c5c54ac
SM		 2262 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2263 dlm_lock_put(lock);
ba2bf218 2264 freed++;
6714d8e8
KH		 2265 }
2266 }
800deef3 2267 list_for_each_entry_safe(lock, next, &res->converting, list) {
6714d8e8
KH		 2268 if (lock->ml.node == dead_node) {
2269 list_del_init(&lock->list);
2270 dlm_lock_put(lock);
2c5c54ac
SM		 2271 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2272 dlm_lock_put(lock);
ba2bf218 2273 freed++;
6714d8e8
KH		 2274 }
2275 }
800deef3 2276 list_for_each_entry_safe(lock, next, &res->blocked, list) {
6714d8e8
KH		 2277 if (lock->ml.node == dead_node) {
2278 list_del_init(&lock->list);
2279 dlm_lock_put(lock);
2c5c54ac
SM		 2280 /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2281 dlm_lock_put(lock);
ba2bf218 2282 freed++;
6714d8e8
KH		 2283 }
2284 }
2285
ba2bf218
KH		 2286 if (freed) {
2287 mlog(0, "%s:%.*s: freed %u locks for dead node %u, "
2288 "dropping ref from lockres\n", dlm->name,
2289 res->lockname.len, res->lockname.name, freed, dead_node);
cda70ba8
SM		 2290 if(!test_bit(dead_node, res->refmap)) {
2291 mlog(ML_ERROR, "%s:%.*s: freed %u locks for dead node %u, "
2292 "but ref was not set\n", dlm->name,
2293 res->lockname.len, res->lockname.name, freed, dead_node);
2294 __dlm_print_one_lock_resource(res);
2295 }
8d400b81 2296 dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
ba2bf218
KH		 2297 } else if (test_bit(dead_node, res->refmap)) {
2298 mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2299 "no locks and had not purged before dying\n", dlm->name,
2300 res->lockname.len, res->lockname.name, dead_node);
8d400b81 2301 dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
ba2bf218
KH		 2302 }
2303
6714d8e8
KH		 2304 /* do not kick thread yet */
2305 __dlm_dirty_lockres(dlm, res);
2306}
2307
2308/* if this node is the recovery master, and there are no
2309 * locks for a given lockres owned by this node that are in
2310 * either PR or EX mode, zero out the lvb before requesting.
2311 *
2312 */
2313
2314
2315static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
2316{
6714d8e8
KH		 2317 struct dlm_lock_resource *res;
2318 int i;
81f2094a 2319 struct hlist_head *bucket;
e2faea4c 2320 struct dlm_lock *lock;
6714d8e8
KH		 2321
2322
2323 /* purge any stale mles */
2324 dlm_clean_master_list(dlm, dead_node);
2325
2326 /*
2327 * now clean up all lock resources. there are two rules:
2328 *
2329 * 1) if the dead node was the master, move the lockres
2330 * to the recovering list. set the RECOVERING flag.
2331 * this lockres needs to be cleaned up before it can
2332 * be used further.
2333 *
2334 * 2) if this node was the master, remove all locks from
2335 * each of the lockres queues that were owned by the
2336 * dead node. once recovery finishes, the dlm thread
2337 * can be kicked again to see if any ASTs or BASTs
2338 * need to be fired as a result.
2339 */
81f2094a 2340 for (i = 0; i < DLM_HASH_BUCKETS; i++) {
03d864c0 2341 bucket = dlm_lockres_hash(dlm, i);
b67bfe0d 2342 hlist_for_each_entry(res, bucket, hash_node) {
e2faea4c
KH		 2343 /* always prune any $RECOVERY entries for dead nodes,
2344 * otherwise hangs can occur during later recovery */
6714d8e8 2345 if (dlm_is_recovery_lock(res->lockname.name,
e2faea4c
KH		 2346 res->lockname.len)) {
2347 spin_lock(&res->spinlock);
2348 list_for_each_entry(lock, &res->granted, list) {
2349 if (lock->ml.node == dead_node) {
2350 mlog(0, "AHA! there was "
2351 "a $RECOVERY lock for dead "
2352 "node %u (%s)!\n",
2353 dead_node, dlm->name);
2354 list_del_init(&lock->list);
2355 dlm_lock_put(lock);
b934beaf
X		 2356 /* Can't schedule
2357 * DLM_UNLOCK_FREE_LOCK
2358 * - do manually */
2359 dlm_lock_put(lock);
e2faea4c
KH		 2360 break;
2361 }
2362 }
2363 spin_unlock(&res->spinlock);
6714d8e8 2364 continue;
2bd63216 2365 }
6714d8e8
KH		 2366 spin_lock(&res->spinlock);
2367 /* zero the lvb if necessary */
2368 dlm_revalidate_lvb(dlm, res, dead_node);
ba2bf218 2369 if (res->owner == dead_node) {
a524812b 2370 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
8decab3c 2371 mlog(ML_NOTICE, "%s: res %.*s, Skip "
a524812b 2372 "recovery as it is being freed\n",
8decab3c 2373 dlm->name, res->lockname.len,
a524812b
WW		 2374 res->lockname.name);
2375 } else
2376 dlm_move_lockres_to_recovery_list(dlm,
2377 res);
ba2bf218 2378
ba2bf218 2379 } else if (res->owner == dlm->node_num) {
6714d8e8
KH		 2380 dlm_free_dead_locks(dlm, res, dead_node);
2381 __dlm_lockres_calc_usage(dlm, res);
69b2bd16
X		 2382 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
2383 if (test_bit(dead_node, res->refmap)) {
2384 mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2385 "no locks and had not purged before dying\n",
2386 dlm->name, res->lockname.len,
2387 res->lockname.name, dead_node);
2388 dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2389 }
6714d8e8
KH		 2390 }
2391 spin_unlock(&res->spinlock);
2392 }
2393 }
2394
2395}
2396
2397static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
2398{
2399 assert_spin_locked(&dlm->spinlock);
2400
466d1a45
KH		 2401 if (dlm->reco.new_master == idx) {
2402 mlog(0, "%s: recovery master %d just died\n",
2403 dlm->name, idx);
2404 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2405 /* finalize1 was reached, so it is safe to clear
2406 * the new_master and dead_node. that recovery
2407 * is complete. */
2408 mlog(0, "%s: dead master %d had reached "
2409 "finalize1 state, clearing\n", dlm->name, idx);
2410 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2411 __dlm_reset_recovery(dlm);
2412 }
2413 }
2414
2d4b1cbb
TM		 2415 /* Clean up join state on node death. */
2416 if (dlm->joining_node == idx) {
2417 mlog(0, "Clearing join state for node %u\n", idx);
2418 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
2419 }
2420
6714d8e8
KH		 2421 /* check to see if the node is already considered dead */
2422 if (!test_bit(idx, dlm->live_nodes_map)) {
2423 mlog(0, "for domain %s, node %d is already dead. "
2424 "another node likely did recovery already.\n",
2425 dlm->name, idx);
2426 return;
2427 }
2428
2429 /* check to see if we do not care about this node */
2430 if (!test_bit(idx, dlm->domain_map)) {
2431 /* This also catches the case that we get a node down
2432 * but haven't joined the domain yet. */
2433 mlog(0, "node %u already removed from domain!\n", idx);
2434 return;
2435 }
2436
2437 clear_bit(idx, dlm->live_nodes_map);
2438
6714d8e8
KH		 2439 /* make sure local cleanup occurs before the heartbeat events */
2440 if (!test_bit(idx, dlm->recovery_map))
2441 dlm_do_local_recovery_cleanup(dlm, idx);
2442
2443 /* notify anything attached to the heartbeat events */
2444 dlm_hb_event_notify_attached(dlm, idx, 0);
2445
2446 mlog(0, "node %u being removed from domain map!\n", idx);
2447 clear_bit(idx, dlm->domain_map);
bddefdee 2448 clear_bit(idx, dlm->exit_domain_map);
6714d8e8
KH		 2449 /* wake up migration waiters if a node goes down.
2450 * perhaps later we can genericize this for other waiters. */
2451 wake_up(&dlm->migration_wq);
2452
2453 if (test_bit(idx, dlm->recovery_map))
2454 mlog(0, "domain %s, node %u already added "
2455 "to recovery map!\n", dlm->name, idx);
2456 else
2457 set_bit(idx, dlm->recovery_map);
2458}
2459
2460void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
2461{
2462 struct dlm_ctxt *dlm = data;
2463
2464 if (!dlm_grab(dlm))
2465 return;
2466
6561168c
MF		 2467 /*
2468 * This will notify any dlm users that a node in our domain
2469 * went away without notifying us first.
2470 */
2471 if (test_bit(idx, dlm->domain_map))
2472 dlm_fire_domain_eviction_callbacks(dlm, idx);
2473
6714d8e8
KH		 2474 spin_lock(&dlm->spinlock);
2475 __dlm_hb_node_down(dlm, idx);
2476 spin_unlock(&dlm->spinlock);
2477
2478 dlm_put(dlm);
2479}
2480
2481void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
2482{
2483 struct dlm_ctxt *dlm = data;
2484
2485 if (!dlm_grab(dlm))
2486 return;
2487
2488 spin_lock(&dlm->spinlock);
6714d8e8 2489 set_bit(idx, dlm->live_nodes_map);
e2faea4c
KH		 2490 /* do NOT notify mle attached to the heartbeat events.
2491 * new nodes are not interesting in mastery until joined. */
6714d8e8
KH		 2492 spin_unlock(&dlm->spinlock);
2493
2494 dlm_put(dlm);
2495}
2496
2497static void dlm_reco_ast(void *astdata)
2498{
2499 struct dlm_ctxt *dlm = astdata;
2500 mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2501 dlm->node_num, dlm->name);
2502}
2503static void dlm_reco_bast(void *astdata, int blocked_type)
2504{
2505 struct dlm_ctxt *dlm = astdata;
2506 mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2507 dlm->node_num, dlm->name);
2508}
2509static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
2510{
2511 mlog(0, "unlockast for recovery lock fired!\n");
2512}
2513
e2faea4c
KH		 2514/*
2515 * dlm_pick_recovery_master will continually attempt to use
2516 * dlmlock() on the special "$RECOVERY" lockres with the
2517 * LKM_NOQUEUE flag to get an EX. every thread that enters
2518 * this function on each node racing to become the recovery
2519 * master will not stop attempting this until either:
2520 * a) this node gets the EX (and becomes the recovery master),
2bd63216 2521 * or b) dlm->reco.new_master gets set to some nodenum
e2faea4c
KH		 2522 * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2523 * so each time a recovery master is needed, the entire cluster
2524 * will sync at this point. if the new master dies, that will
2525 * be detected in dlm_do_recovery */
6714d8e8
KH		 2526static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2527{
2528 enum dlm_status ret;
2529 struct dlm_lockstatus lksb;
2530 int status = -EINVAL;
2531
2532 mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2533 dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2bd63216 2534again:
6714d8e8
KH		 2535 memset(&lksb, 0, sizeof(lksb));
2536
2537 ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
3384f3df
MF		 2538 DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN,
2539 dlm_reco_ast, dlm, dlm_reco_bast);
6714d8e8 2540
e2faea4c
KH		 2541 mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2542 dlm->name, ret, lksb.status);
2543
6714d8e8
KH		 2544 if (ret == DLM_NORMAL) {
2545 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2546 dlm->name, dlm->node_num);
2bd63216
SM		 2547
2548 /* got the EX lock. check to see if another node
e2faea4c
KH		 2549 * just became the reco master */
2550 if (dlm_reco_master_ready(dlm)) {
2551 mlog(0, "%s: got reco EX lock, but %u will "
2552 "do the recovery\n", dlm->name,
2553 dlm->reco.new_master);
2554 status = -EEXIST;
2555 } else {
898effac
KH		 2556 status = 0;
2557
2558 /* see if recovery was already finished elsewhere */
2559 spin_lock(&dlm->spinlock);
2560 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
2bd63216 2561 status = -EINVAL;
898effac
KH		 2562 mlog(0, "%s: got reco EX lock, but "
2563 "node got recovered already\n", dlm->name);
2564 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2565 mlog(ML_ERROR, "%s: new master is %u "
2bd63216 2566 "but no dead node!\n",
898effac
KH		 2567 dlm->name, dlm->reco.new_master);
2568 BUG();
2569 }
2570 }
2571 spin_unlock(&dlm->spinlock);
2572 }
2573
2574 /* if this node has actually become the recovery master,
2575 * set the master and send the messages to begin recovery */
2576 if (!status) {
2577 mlog(0, "%s: dead=%u, this=%u, sending "
2bd63216 2578 "begin_reco now\n", dlm->name,
898effac 2579 dlm->reco.dead_node, dlm->node_num);
e2faea4c
KH		 2580 status = dlm_send_begin_reco_message(dlm,
2581 dlm->reco.dead_node);
2582 /* this always succeeds */
2583 BUG_ON(status);
2584
2585 /* set the new_master to this node */
2586 spin_lock(&dlm->spinlock);
ab27eb6f 2587 dlm_set_reco_master(dlm, dlm->node_num);
e2faea4c
KH		 2588 spin_unlock(&dlm->spinlock);
2589 }
6714d8e8
KH		 2590
2591 /* recovery lock is a special case. ast will not get fired,
2592 * so just go ahead and unlock it. */
2593 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
e2faea4c
KH		 2594 if (ret == DLM_DENIED) {
2595 mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2596 ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2597 }
6714d8e8
KH		 2598 if (ret != DLM_NORMAL) {
2599 /* this would really suck. this could only happen
2600 * if there was a network error during the unlock
2601 * because of node death. this means the unlock
2602 * is actually "done" and the lock structure is
2603 * even freed. we can continue, but only
2604 * because this specific lock name is special. */
e2faea4c 2605 mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
6714d8e8
KH		 2606 }
2607 } else if (ret == DLM_NOTQUEUED) {
2608 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2609 dlm->name, dlm->node_num);
2610 /* another node is master. wait on
2bd63216 2611 * reco.new_master != O2NM_INVALID_NODE_NUM
e2faea4c
KH		 2612 * for at most one second */
2613 wait_event_timeout(dlm->dlm_reco_thread_wq,
2614 dlm_reco_master_ready(dlm),
2615 msecs_to_jiffies(1000));
2616 if (!dlm_reco_master_ready(dlm)) {
2617 mlog(0, "%s: reco master taking awhile\n",
2618 dlm->name);
2619 goto again;
2620 }
2621 /* another node has informed this one that it is reco master */
2622 mlog(0, "%s: reco master %u is ready to recover %u\n",
2623 dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
6714d8e8 2624 status = -EEXIST;
c8df412e
KH		 2625 } else if (ret == DLM_RECOVERING) {
2626 mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n",
2627 dlm->name, dlm->node_num);
2628 goto again;
e2faea4c
KH		 2629 } else {
2630 struct dlm_lock_resource *res;
2631
2632 /* dlmlock returned something other than NOTQUEUED or NORMAL */
2633 mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2634 "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2635 dlm_errname(lksb.status));
2636 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2637 DLM_RECOVERY_LOCK_NAME_LEN);
2638 if (res) {
2639 dlm_print_one_lock_resource(res);
2640 dlm_lockres_put(res);
2641 } else {
2642 mlog(ML_ERROR, "recovery lock not found\n");
2643 }
2644 BUG();
6714d8e8
KH		 2645 }
2646
2647 return status;
2648}
2649
2650static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2651{
2652 struct dlm_begin_reco br;
2653 int ret = 0;
2654 struct dlm_node_iter iter;
2655 int nodenum;
2656 int status;
2657
d6dea6e9 2658 mlog(0, "%s: dead node is %u\n", dlm->name, dead_node);
6714d8e8
KH		 2659
2660 spin_lock(&dlm->spinlock);
2661 dlm_node_iter_init(dlm->domain_map, &iter);
2662 spin_unlock(&dlm->spinlock);
2663
2664 clear_bit(dead_node, iter.node_map);
2665
2666 memset(&br, 0, sizeof(br));
2667 br.node_idx = dlm->node_num;
2668 br.dead_node = dead_node;
2669
2670 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2671 ret = 0;
2672 if (nodenum == dead_node) {
2673 mlog(0, "not sending begin reco to dead node "
2674 "%u\n", dead_node);
2675 continue;
2676 }
2677 if (nodenum == dlm->node_num) {
2678 mlog(0, "not sending begin reco to self\n");
2679 continue;
2680 }
e2faea4c 2681retry:
6714d8e8
KH		 2682 ret = -EINVAL;
2683 mlog(0, "attempting to send begin reco msg to %d\n",
2684 nodenum);
2685 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2686 &br, sizeof(br), nodenum, &status);
2687 /* negative status is handled ok by caller here */
2688 if (ret >= 0)
2689 ret = status;
e2faea4c
KH		 2690 if (dlm_is_host_down(ret)) {
2691 /* node is down. not involved in recovery
2692 * so just keep going */
a5196ec5 2693 mlog(ML_NOTICE, "%s: node %u was down when sending "
e2faea4c
KH		 2694 "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2695 ret = 0;
2696 }
cd34edd8
SM		 2697
2698 /*
2699 * Prior to commit aad1b15310b9bcd59fa81ab8f2b1513b59553ea8,
2700 * dlm_begin_reco_handler() returned EAGAIN and not -EAGAIN.
2701 * We are handling both for compatibility reasons.
2702 */
2703 if (ret == -EAGAIN || ret == EAGAIN) {
aad1b153
TY		 2704 mlog(0, "%s: trying to start recovery of node "
2705 "%u, but node %u is waiting for last recovery "
2706 "to complete, backoff for a bit\n", dlm->name,
2707 dead_node, nodenum);
2708 msleep(100);
2709 goto retry;
2710 }
6714d8e8
KH		 2711 if (ret < 0) {
2712 struct dlm_lock_resource *res;
a5196ec5 2713
2bd63216 2714 /* this is now a serious problem, possibly ENOMEM
e2faea4c 2715 * in the network stack. must retry */
6714d8e8
KH		 2716 mlog_errno(ret);
2717 mlog(ML_ERROR, "begin reco of dlm %s to node %u "
a5196ec5 2718 "returned %d\n", dlm->name, nodenum, ret);
6714d8e8
KH		 2719 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2720 DLM_RECOVERY_LOCK_NAME_LEN);
2721 if (res) {
2722 dlm_print_one_lock_resource(res);
2723 dlm_lockres_put(res);
2724 } else {
2725 mlog(ML_ERROR, "recovery lock not found\n");
2726 }
2bd63216 2727 /* sleep for a bit in hopes that we can avoid
e2faea4c
KH		 2728 * another ENOMEM */
2729 msleep(100);
2730 goto retry;
6714d8e8
KH		 2731 }
2732 }
2733
2734 return ret;
2735}
2736
d74c9803
KH		 2737int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2738 void **ret_data)
6714d8e8
KH		 2739{
2740 struct dlm_ctxt *dlm = data;
2741 struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2742
2743 /* ok to return 0, domain has gone away */
2744 if (!dlm_grab(dlm))
2745 return 0;
2746
466d1a45
KH		 2747 spin_lock(&dlm->spinlock);
2748 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2749 mlog(0, "%s: node %u wants to recover node %u (%u:%u) "
2750 "but this node is in finalize state, waiting on finalize2\n",
2751 dlm->name, br->node_idx, br->dead_node,
2752 dlm->reco.dead_node, dlm->reco.new_master);
2753 spin_unlock(&dlm->spinlock);
40c7f2ea 2754 dlm_put(dlm);
aad1b153 2755 return -EAGAIN;
466d1a45
KH		 2756 }
2757 spin_unlock(&dlm->spinlock);
2758
d6dea6e9
KH		 2759 mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n",
2760 dlm->name, br->node_idx, br->dead_node,
2761 dlm->reco.dead_node, dlm->reco.new_master);
6714d8e8
KH		 2762
2763 dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2764
2765 spin_lock(&dlm->spinlock);
2766 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
e2faea4c
KH		 2767 if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2768 mlog(0, "%s: new_master %u died, changing "
2769 "to %u\n", dlm->name, dlm->reco.new_master,
2770 br->node_idx);
2771 } else {
2772 mlog(0, "%s: new_master %u NOT DEAD, changing "
2773 "to %u\n", dlm->name, dlm->reco.new_master,
2774 br->node_idx);
2775 /* may not have seen the new master as dead yet */
2776 }
6714d8e8
KH		 2777 }
2778 if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
e2faea4c 2779 mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2bd63216 2780 "node %u changing it to %u\n", dlm->name,
e2faea4c 2781 dlm->reco.dead_node, br->node_idx, br->dead_node);
6714d8e8 2782 }
ab27eb6f
KH		 2783 dlm_set_reco_master(dlm, br->node_idx);
2784 dlm_set_reco_dead_node(dlm, br->dead_node);
6714d8e8 2785 if (!test_bit(br->dead_node, dlm->recovery_map)) {
e2faea4c 2786 mlog(0, "recovery master %u sees %u as dead, but this "
6714d8e8
KH		 2787 "node has not yet. marking %u as dead\n",
2788 br->node_idx, br->dead_node, br->dead_node);
e2faea4c
KH		 2789 if (!test_bit(br->dead_node, dlm->domain_map) ||
2790 !test_bit(br->dead_node, dlm->live_nodes_map))
2791 mlog(0, "%u not in domain/live_nodes map "
2792 "so setting it in reco map manually\n",
2793 br->dead_node);
c03872f5
KH		 2794 /* force the recovery cleanup in __dlm_hb_node_down
2795 * both of these will be cleared in a moment */
2796 set_bit(br->dead_node, dlm->domain_map);
2797 set_bit(br->dead_node, dlm->live_nodes_map);
6714d8e8
KH		 2798 __dlm_hb_node_down(dlm, br->dead_node);
2799 }
2800 spin_unlock(&dlm->spinlock);
2801
2802 dlm_kick_recovery_thread(dlm);
d6dea6e9
KH		 2803
2804 mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n",
2805 dlm->name, br->node_idx, br->dead_node,
2806 dlm->reco.dead_node, dlm->reco.new_master);
2807
6714d8e8
KH		 2808 dlm_put(dlm);
2809 return 0;
2810}
2811
466d1a45 2812#define DLM_FINALIZE_STAGE2 0x01
6714d8e8
KH		 2813static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2814{
2815 int ret = 0;
2816 struct dlm_finalize_reco fr;
2817 struct dlm_node_iter iter;
2818 int nodenum;
2819 int status;
466d1a45 2820 int stage = 1;
6714d8e8 2821
466d1a45
KH		 2822 mlog(0, "finishing recovery for node %s:%u, "
2823 "stage %d\n", dlm->name, dlm->reco.dead_node, stage);
6714d8e8
KH		 2824
2825 spin_lock(&dlm->spinlock);
2826 dlm_node_iter_init(dlm->domain_map, &iter);
2827 spin_unlock(&dlm->spinlock);
2828
466d1a45 2829stage2:
6714d8e8
KH		 2830 memset(&fr, 0, sizeof(fr));
2831 fr.node_idx = dlm->node_num;
2832 fr.dead_node = dlm->reco.dead_node;
466d1a45
KH		 2833 if (stage == 2)
2834 fr.flags |= DLM_FINALIZE_STAGE2;
6714d8e8
KH		 2835
2836 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2837 if (nodenum == dlm->node_num)
2838 continue;
2839 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2840 &fr, sizeof(fr), nodenum, &status);
466d1a45 2841 if (ret >= 0)
6714d8e8 2842 ret = status;
466d1a45 2843 if (ret < 0) {
a5196ec5
WW		 2844 mlog(ML_ERROR, "Error %d when sending message %u (key "
2845 "0x%x) to node %u\n", ret, DLM_FINALIZE_RECO_MSG,
2846 dlm->key, nodenum);
6714d8e8 2847 if (dlm_is_host_down(ret)) {
2bd63216
SM		 2848 /* this has no effect on this recovery
2849 * session, so set the status to zero to
6714d8e8
KH		 2850 * finish out the last recovery */
2851 mlog(ML_ERROR, "node %u went down after this "
2852 "node finished recovery.\n", nodenum);
2853 ret = 0;
c27069e6 2854 continue;
6714d8e8 2855 }
6714d8e8
KH		 2856 break;
2857 }
2858 }
466d1a45
KH		 2859 if (stage == 1) {
2860 /* reset the node_iter back to the top and send finalize2 */
2861 iter.curnode = -1;
2862 stage = 2;
2863 goto stage2;
2864 }
6714d8e8
KH		 2865
2866 return ret;
2867}
2868
d74c9803
KH		 2869int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2870 void **ret_data)
6714d8e8
KH		 2871{
2872 struct dlm_ctxt *dlm = data;
2873 struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
466d1a45 2874 int stage = 1;
6714d8e8
KH		 2875
2876 /* ok to return 0, domain has gone away */
2877 if (!dlm_grab(dlm))
2878 return 0;
2879
466d1a45
KH		 2880 if (fr->flags & DLM_FINALIZE_STAGE2)
2881 stage = 2;
6714d8e8 2882
466d1a45
KH		 2883 mlog(0, "%s: node %u finalizing recovery stage%d of "
2884 "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage,
2885 fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master);
2bd63216 2886
6714d8e8
KH		 2887 spin_lock(&dlm->spinlock);
2888
2889 if (dlm->reco.new_master != fr->node_idx) {
2890 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2891 "%u is supposed to be the new master, dead=%u\n",
2892 fr->node_idx, dlm->reco.new_master, fr->dead_node);
2893 BUG();
2894 }
2895 if (dlm->reco.dead_node != fr->dead_node) {
2896 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2897 "node %u, but node %u is supposed to be dead\n",
2898 fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2899 BUG();
2900 }
2901
466d1a45
KH		 2902 switch (stage) {
2903 case 1:
2904 dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2905 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2906 mlog(ML_ERROR, "%s: received finalize1 from "
2907 "new master %u for dead node %u, but "
2908 "this node has already received it!\n",
2909 dlm->name, fr->node_idx, fr->dead_node);
2910 dlm_print_reco_node_status(dlm);
2911 BUG();
2912 }
2913 dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
2914 spin_unlock(&dlm->spinlock);
2915 break;
2916 case 2:
2917 if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) {
2918 mlog(ML_ERROR, "%s: received finalize2 from "
2919 "new master %u for dead node %u, but "
2920 "this node did not have finalize1!\n",
2921 dlm->name, fr->node_idx, fr->dead_node);
2922 dlm_print_reco_node_status(dlm);
2923 BUG();
2924 }
2925 dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
ded2cf71 2926 __dlm_reset_recovery(dlm);
466d1a45 2927 spin_unlock(&dlm->spinlock);
466d1a45
KH		 2928 dlm_kick_recovery_thread(dlm);
2929 break;
2930 default:
2931 BUG();
2932 }
6714d8e8 2933
d6dea6e9
KH		 2934 mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n",
2935 dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master);
2936
6714d8e8
KH		 2937 dlm_put(dlm);
2938 return 0;
2939}
Linux 6.x block layer and io_uring tree(s)