]>
Commit | Line | Data |
---|---|---|
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> | |
33 | #include <linux/utsname.h> | |
34 | #include <linux/init.h> | |
35 | #include <linux/sysctl.h> | |
36 | #include <linux/random.h> | |
37 | #include <linux/blkdev.h> | |
38 | #include <linux/socket.h> | |
39 | #include <linux/inet.h> | |
40 | #include <linux/timer.h> | |
41 | #include <linux/kthread.h> | |
b4c7f538 | 42 | #include <linux/delay.h> |
6714d8e8 KH |
43 | |
44 | ||
45 | #include "cluster/heartbeat.h" | |
46 | #include "cluster/nodemanager.h" | |
47 | #include "cluster/tcp.h" | |
48 | ||
49 | #include "dlmapi.h" | |
50 | #include "dlmcommon.h" | |
51 | #include "dlmdomain.h" | |
52 | ||
53 | #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY) | |
54 | #include "cluster/masklog.h" | |
55 | ||
56 | static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node); | |
57 | ||
58 | static int dlm_recovery_thread(void *data); | |
59 | void dlm_complete_recovery_thread(struct dlm_ctxt *dlm); | |
60 | int dlm_launch_recovery_thread(struct dlm_ctxt *dlm); | |
c03872f5 | 61 | void dlm_kick_recovery_thread(struct dlm_ctxt *dlm); |
6714d8e8 KH |
62 | static int dlm_do_recovery(struct dlm_ctxt *dlm); |
63 | ||
64 | static int dlm_pick_recovery_master(struct dlm_ctxt *dlm); | |
65 | static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node); | |
66 | static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node); | |
67 | static int dlm_request_all_locks(struct dlm_ctxt *dlm, | |
68 | u8 request_from, u8 dead_node); | |
69 | static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node); | |
70 | ||
71 | static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res); | |
72 | static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres, | |
73 | const char *lockname, int namelen, | |
74 | int total_locks, u64 cookie, | |
75 | u8 flags, u8 master); | |
76 | static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm, | |
77 | struct dlm_migratable_lockres *mres, | |
78 | u8 send_to, | |
79 | struct dlm_lock_resource *res, | |
80 | int total_locks); | |
6714d8e8 KH |
81 | static int dlm_process_recovery_data(struct dlm_ctxt *dlm, |
82 | struct dlm_lock_resource *res, | |
83 | struct dlm_migratable_lockres *mres); | |
6714d8e8 KH |
84 | static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm); |
85 | static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, | |
86 | u8 dead_node, u8 send_to); | |
87 | static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node); | |
88 | static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm, | |
89 | struct list_head *list, u8 dead_node); | |
90 | static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm, | |
91 | u8 dead_node, u8 new_master); | |
92 | static void dlm_reco_ast(void *astdata); | |
93 | static void dlm_reco_bast(void *astdata, int blocked_type); | |
94 | static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st); | |
95 | static void dlm_request_all_locks_worker(struct dlm_work_item *item, | |
96 | void *data); | |
97 | static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data); | |
8169cae5 AB |
98 | static int dlm_lockres_master_requery(struct dlm_ctxt *dlm, |
99 | struct dlm_lock_resource *res, | |
100 | u8 *real_master); | |
6714d8e8 KH |
101 | |
102 | static u64 dlm_get_next_mig_cookie(void); | |
103 | ||
34af946a IM |
104 | static DEFINE_SPINLOCK(dlm_reco_state_lock); |
105 | static DEFINE_SPINLOCK(dlm_mig_cookie_lock); | |
6714d8e8 KH |
106 | static u64 dlm_mig_cookie = 1; |
107 | ||
108 | static u64 dlm_get_next_mig_cookie(void) | |
109 | { | |
110 | u64 c; | |
111 | spin_lock(&dlm_mig_cookie_lock); | |
112 | c = dlm_mig_cookie; | |
113 | if (dlm_mig_cookie == (~0ULL)) | |
114 | dlm_mig_cookie = 1; | |
115 | else | |
116 | dlm_mig_cookie++; | |
117 | spin_unlock(&dlm_mig_cookie_lock); | |
118 | return c; | |
119 | } | |
120 | ||
ab27eb6f KH |
121 | static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm, |
122 | u8 dead_node) | |
123 | { | |
124 | assert_spin_locked(&dlm->spinlock); | |
125 | if (dlm->reco.dead_node != dead_node) | |
126 | mlog(0, "%s: changing dead_node from %u to %u\n", | |
127 | dlm->name, dlm->reco.dead_node, dead_node); | |
128 | dlm->reco.dead_node = dead_node; | |
129 | } | |
130 | ||
131 | static inline void dlm_set_reco_master(struct dlm_ctxt *dlm, | |
132 | u8 master) | |
133 | { | |
134 | assert_spin_locked(&dlm->spinlock); | |
135 | mlog(0, "%s: changing new_master from %u to %u\n", | |
136 | dlm->name, dlm->reco.new_master, master); | |
137 | dlm->reco.new_master = master; | |
138 | } | |
139 | ||
466d1a45 | 140 | static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm) |
6714d8e8 | 141 | { |
466d1a45 | 142 | assert_spin_locked(&dlm->spinlock); |
6714d8e8 | 143 | clear_bit(dlm->reco.dead_node, dlm->recovery_map); |
ab27eb6f KH |
144 | dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); |
145 | dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM); | |
466d1a45 KH |
146 | } |
147 | ||
148 | static inline void dlm_reset_recovery(struct dlm_ctxt *dlm) | |
149 | { | |
150 | spin_lock(&dlm->spinlock); | |
151 | __dlm_reset_recovery(dlm); | |
6714d8e8 KH |
152 | spin_unlock(&dlm->spinlock); |
153 | } | |
154 | ||
155 | /* Worker function used during recovery. */ | |
c4028958 | 156 | void dlm_dispatch_work(struct work_struct *work) |
6714d8e8 | 157 | { |
c4028958 DH |
158 | struct dlm_ctxt *dlm = |
159 | container_of(work, struct dlm_ctxt, dispatched_work); | |
6714d8e8 KH |
160 | LIST_HEAD(tmp_list); |
161 | struct list_head *iter, *iter2; | |
162 | struct dlm_work_item *item; | |
163 | dlm_workfunc_t *workfunc; | |
3156d267 KH |
164 | int tot=0; |
165 | ||
6714d8e8 KH |
166 | spin_lock(&dlm->work_lock); |
167 | list_splice_init(&dlm->work_list, &tmp_list); | |
168 | spin_unlock(&dlm->work_lock); | |
169 | ||
3156d267 KH |
170 | list_for_each_safe(iter, iter2, &tmp_list) { |
171 | tot++; | |
172 | } | |
173 | mlog(0, "%s: work thread has %d work items\n", dlm->name, tot); | |
174 | ||
6714d8e8 KH |
175 | list_for_each_safe(iter, iter2, &tmp_list) { |
176 | item = list_entry(iter, struct dlm_work_item, list); | |
177 | workfunc = item->func; | |
178 | list_del_init(&item->list); | |
179 | ||
180 | /* already have ref on dlm to avoid having | |
181 | * it disappear. just double-check. */ | |
182 | BUG_ON(item->dlm != dlm); | |
183 | ||
184 | /* this is allowed to sleep and | |
185 | * call network stuff */ | |
186 | workfunc(item, item->data); | |
187 | ||
188 | dlm_put(dlm); | |
189 | kfree(item); | |
190 | } | |
191 | } | |
192 | ||
193 | /* | |
194 | * RECOVERY THREAD | |
195 | */ | |
196 | ||
c03872f5 | 197 | void dlm_kick_recovery_thread(struct dlm_ctxt *dlm) |
6714d8e8 KH |
198 | { |
199 | /* wake the recovery thread | |
200 | * this will wake the reco thread in one of three places | |
201 | * 1) sleeping with no recovery happening | |
202 | * 2) sleeping with recovery mastered elsewhere | |
203 | * 3) recovery mastered here, waiting on reco data */ | |
204 | ||
205 | wake_up(&dlm->dlm_reco_thread_wq); | |
206 | } | |
207 | ||
208 | /* Launch the recovery thread */ | |
209 | int dlm_launch_recovery_thread(struct dlm_ctxt *dlm) | |
210 | { | |
211 | mlog(0, "starting dlm recovery thread...\n"); | |
212 | ||
213 | dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm, | |
214 | "dlm_reco_thread"); | |
215 | if (IS_ERR(dlm->dlm_reco_thread_task)) { | |
216 | mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task)); | |
217 | dlm->dlm_reco_thread_task = NULL; | |
218 | return -EINVAL; | |
219 | } | |
220 | ||
221 | return 0; | |
222 | } | |
223 | ||
224 | void dlm_complete_recovery_thread(struct dlm_ctxt *dlm) | |
225 | { | |
226 | if (dlm->dlm_reco_thread_task) { | |
227 | mlog(0, "waiting for dlm recovery thread to exit\n"); | |
228 | kthread_stop(dlm->dlm_reco_thread_task); | |
229 | dlm->dlm_reco_thread_task = NULL; | |
230 | } | |
231 | } | |
232 | ||
233 | ||
234 | ||
235 | /* | |
236 | * this is lame, but here's how recovery works... | |
237 | * 1) all recovery threads cluster wide will work on recovering | |
238 | * ONE node at a time | |
239 | * 2) negotiate who will take over all the locks for the dead node. | |
240 | * thats right... ALL the locks. | |
241 | * 3) once a new master is chosen, everyone scans all locks | |
242 | * and moves aside those mastered by the dead guy | |
243 | * 4) each of these locks should be locked until recovery is done | |
244 | * 5) the new master collects up all of secondary lock queue info | |
245 | * one lock at a time, forcing each node to communicate back | |
246 | * before continuing | |
247 | * 6) each secondary lock queue responds with the full known lock info | |
248 | * 7) once the new master has run all its locks, it sends a ALLDONE! | |
249 | * message to everyone | |
250 | * 8) upon receiving this message, the secondary queue node unlocks | |
251 | * and responds to the ALLDONE | |
252 | * 9) once the new master gets responses from everyone, he unlocks | |
253 | * everything and recovery for this dead node is done | |
254 | *10) go back to 2) while there are still dead nodes | |
255 | * | |
256 | */ | |
257 | ||
d6dea6e9 KH |
258 | static void dlm_print_reco_node_status(struct dlm_ctxt *dlm) |
259 | { | |
260 | struct dlm_reco_node_data *ndata; | |
261 | struct dlm_lock_resource *res; | |
262 | ||
263 | mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n", | |
264 | dlm->name, dlm->dlm_reco_thread_task->pid, | |
265 | dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive", | |
266 | dlm->reco.dead_node, dlm->reco.new_master); | |
267 | ||
268 | list_for_each_entry(ndata, &dlm->reco.node_data, list) { | |
269 | char *st = "unknown"; | |
270 | switch (ndata->state) { | |
271 | case DLM_RECO_NODE_DATA_INIT: | |
272 | st = "init"; | |
273 | break; | |
274 | case DLM_RECO_NODE_DATA_REQUESTING: | |
275 | st = "requesting"; | |
276 | break; | |
277 | case DLM_RECO_NODE_DATA_DEAD: | |
278 | st = "dead"; | |
279 | break; | |
280 | case DLM_RECO_NODE_DATA_RECEIVING: | |
281 | st = "receiving"; | |
282 | break; | |
283 | case DLM_RECO_NODE_DATA_REQUESTED: | |
284 | st = "requested"; | |
285 | break; | |
286 | case DLM_RECO_NODE_DATA_DONE: | |
287 | st = "done"; | |
288 | break; | |
289 | case DLM_RECO_NODE_DATA_FINALIZE_SENT: | |
290 | st = "finalize-sent"; | |
291 | break; | |
292 | default: | |
293 | st = "bad"; | |
294 | break; | |
295 | } | |
296 | mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n", | |
297 | dlm->name, ndata->node_num, st); | |
298 | } | |
299 | list_for_each_entry(res, &dlm->reco.resources, recovering) { | |
300 | mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n", | |
301 | dlm->name, res->lockname.len, res->lockname.name); | |
302 | } | |
303 | } | |
6714d8e8 KH |
304 | |
305 | #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000) | |
306 | ||
307 | static int dlm_recovery_thread(void *data) | |
308 | { | |
309 | int status; | |
310 | struct dlm_ctxt *dlm = data; | |
311 | unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS); | |
312 | ||
313 | mlog(0, "dlm thread running for %s...\n", dlm->name); | |
314 | ||
315 | while (!kthread_should_stop()) { | |
316 | if (dlm_joined(dlm)) { | |
317 | status = dlm_do_recovery(dlm); | |
318 | if (status == -EAGAIN) { | |
319 | /* do not sleep, recheck immediately. */ | |
320 | continue; | |
321 | } | |
322 | if (status < 0) | |
323 | mlog_errno(status); | |
324 | } | |
325 | ||
326 | wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq, | |
327 | kthread_should_stop(), | |
328 | timeout); | |
329 | } | |
330 | ||
331 | mlog(0, "quitting DLM recovery thread\n"); | |
332 | return 0; | |
333 | } | |
334 | ||
e2faea4c KH |
335 | /* returns true when the recovery master has contacted us */ |
336 | static int dlm_reco_master_ready(struct dlm_ctxt *dlm) | |
337 | { | |
338 | int ready; | |
339 | spin_lock(&dlm->spinlock); | |
340 | ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM); | |
341 | spin_unlock(&dlm->spinlock); | |
342 | return ready; | |
343 | } | |
344 | ||
345 | /* returns true if node is no longer in the domain | |
346 | * could be dead or just not joined */ | |
347 | int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node) | |
348 | { | |
349 | int dead; | |
350 | spin_lock(&dlm->spinlock); | |
aba9aac7 | 351 | dead = !test_bit(node, dlm->domain_map); |
e2faea4c KH |
352 | spin_unlock(&dlm->spinlock); |
353 | return dead; | |
354 | } | |
355 | ||
b7084ab5 KH |
356 | /* returns true if node is no longer in the domain |
357 | * could be dead or just not joined */ | |
3fb5a989 | 358 | static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node) |
b7084ab5 KH |
359 | { |
360 | int recovered; | |
361 | spin_lock(&dlm->spinlock); | |
362 | recovered = !test_bit(node, dlm->recovery_map); | |
363 | spin_unlock(&dlm->spinlock); | |
364 | return recovered; | |
365 | } | |
366 | ||
367 | ||
44465a7d KH |
368 | int dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout) |
369 | { | |
370 | if (timeout) { | |
371 | mlog(ML_NOTICE, "%s: waiting %dms for notification of " | |
372 | "death of node %u\n", dlm->name, timeout, node); | |
373 | wait_event_timeout(dlm->dlm_reco_thread_wq, | |
374 | dlm_is_node_dead(dlm, node), | |
375 | msecs_to_jiffies(timeout)); | |
376 | } else { | |
377 | mlog(ML_NOTICE, "%s: waiting indefinitely for notification " | |
378 | "of death of node %u\n", dlm->name, node); | |
379 | wait_event(dlm->dlm_reco_thread_wq, | |
380 | dlm_is_node_dead(dlm, node)); | |
381 | } | |
382 | /* for now, return 0 */ | |
383 | return 0; | |
384 | } | |
385 | ||
b7084ab5 KH |
386 | int dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout) |
387 | { | |
388 | if (timeout) { | |
389 | mlog(0, "%s: waiting %dms for notification of " | |
390 | "recovery of node %u\n", dlm->name, timeout, node); | |
391 | wait_event_timeout(dlm->dlm_reco_thread_wq, | |
392 | dlm_is_node_recovered(dlm, node), | |
393 | msecs_to_jiffies(timeout)); | |
394 | } else { | |
395 | mlog(0, "%s: waiting indefinitely for notification " | |
396 | "of recovery of node %u\n", dlm->name, node); | |
397 | wait_event(dlm->dlm_reco_thread_wq, | |
398 | dlm_is_node_recovered(dlm, node)); | |
399 | } | |
400 | /* for now, return 0 */ | |
401 | return 0; | |
402 | } | |
403 | ||
6714d8e8 KH |
404 | /* callers of the top-level api calls (dlmlock/dlmunlock) should |
405 | * block on the dlm->reco.event when recovery is in progress. | |
406 | * the dlm recovery thread will set this state when it begins | |
407 | * recovering a dead node (as the new master or not) and clear | |
408 | * the state and wake as soon as all affected lock resources have | |
409 | * been marked with the RECOVERY flag */ | |
410 | static int dlm_in_recovery(struct dlm_ctxt *dlm) | |
411 | { | |
412 | int in_recovery; | |
413 | spin_lock(&dlm->spinlock); | |
414 | in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE); | |
415 | spin_unlock(&dlm->spinlock); | |
416 | return in_recovery; | |
417 | } | |
418 | ||
419 | ||
420 | void dlm_wait_for_recovery(struct dlm_ctxt *dlm) | |
421 | { | |
56a7c104 | 422 | if (dlm_in_recovery(dlm)) { |
3b3b84a8 | 423 | mlog(0, "%s: reco thread %d in recovery: " |
56a7c104 KH |
424 | "state=%d, master=%u, dead=%u\n", |
425 | dlm->name, dlm->dlm_reco_thread_task->pid, | |
426 | dlm->reco.state, dlm->reco.new_master, | |
427 | dlm->reco.dead_node); | |
428 | } | |
6714d8e8 KH |
429 | wait_event(dlm->reco.event, !dlm_in_recovery(dlm)); |
430 | } | |
431 | ||
432 | static void dlm_begin_recovery(struct dlm_ctxt *dlm) | |
433 | { | |
434 | spin_lock(&dlm->spinlock); | |
435 | BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE); | |
436 | dlm->reco.state |= DLM_RECO_STATE_ACTIVE; | |
437 | spin_unlock(&dlm->spinlock); | |
438 | } | |
439 | ||
440 | static void dlm_end_recovery(struct dlm_ctxt *dlm) | |
441 | { | |
442 | spin_lock(&dlm->spinlock); | |
443 | BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE)); | |
444 | dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE; | |
445 | spin_unlock(&dlm->spinlock); | |
446 | wake_up(&dlm->reco.event); | |
447 | } | |
448 | ||
449 | static int dlm_do_recovery(struct dlm_ctxt *dlm) | |
450 | { | |
451 | int status = 0; | |
e2faea4c | 452 | int ret; |
6714d8e8 KH |
453 | |
454 | spin_lock(&dlm->spinlock); | |
455 | ||
456 | /* check to see if the new master has died */ | |
457 | if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM && | |
458 | test_bit(dlm->reco.new_master, dlm->recovery_map)) { | |
459 | mlog(0, "new master %u died while recovering %u!\n", | |
460 | dlm->reco.new_master, dlm->reco.dead_node); | |
461 | /* unset the new_master, leave dead_node */ | |
ab27eb6f | 462 | dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM); |
6714d8e8 KH |
463 | } |
464 | ||
465 | /* select a target to recover */ | |
466 | if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { | |
467 | int bit; | |
468 | ||
469 | bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES+1, 0); | |
470 | if (bit >= O2NM_MAX_NODES || bit < 0) | |
ab27eb6f | 471 | dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); |
6714d8e8 | 472 | else |
ab27eb6f | 473 | dlm_set_reco_dead_node(dlm, bit); |
6714d8e8 KH |
474 | } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) { |
475 | /* BUG? */ | |
476 | mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n", | |
477 | dlm->reco.dead_node); | |
ab27eb6f | 478 | dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); |
6714d8e8 KH |
479 | } |
480 | ||
481 | if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { | |
482 | // mlog(0, "nothing to recover! sleeping now!\n"); | |
483 | spin_unlock(&dlm->spinlock); | |
484 | /* return to main thread loop and sleep. */ | |
485 | return 0; | |
486 | } | |
d6dea6e9 KH |
487 | mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n", |
488 | dlm->name, dlm->dlm_reco_thread_task->pid, | |
6714d8e8 KH |
489 | dlm->reco.dead_node); |
490 | spin_unlock(&dlm->spinlock); | |
491 | ||
492 | /* take write barrier */ | |
493 | /* (stops the list reshuffling thread, proxy ast handling) */ | |
494 | dlm_begin_recovery(dlm); | |
495 | ||
496 | if (dlm->reco.new_master == dlm->node_num) | |
497 | goto master_here; | |
498 | ||
499 | if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) { | |
e2faea4c KH |
500 | /* choose a new master, returns 0 if this node |
501 | * is the master, -EEXIST if it's another node. | |
502 | * this does not return until a new master is chosen | |
503 | * or recovery completes entirely. */ | |
504 | ret = dlm_pick_recovery_master(dlm); | |
505 | if (!ret) { | |
6714d8e8 | 506 | /* already notified everyone. go. */ |
6714d8e8 KH |
507 | goto master_here; |
508 | } | |
509 | mlog(0, "another node will master this recovery session.\n"); | |
510 | } | |
d6dea6e9 KH |
511 | mlog(0, "dlm=%s (%d), new_master=%u, this node=%u, dead_node=%u\n", |
512 | dlm->name, dlm->dlm_reco_thread_task->pid, dlm->reco.new_master, | |
6714d8e8 KH |
513 | dlm->node_num, dlm->reco.dead_node); |
514 | ||
515 | /* it is safe to start everything back up here | |
516 | * because all of the dead node's lock resources | |
517 | * have been marked as in-recovery */ | |
518 | dlm_end_recovery(dlm); | |
519 | ||
520 | /* sleep out in main dlm_recovery_thread loop. */ | |
521 | return 0; | |
522 | ||
523 | master_here: | |
d6dea6e9 KH |
524 | mlog(0, "(%d) mastering recovery of %s:%u here(this=%u)!\n", |
525 | dlm->dlm_reco_thread_task->pid, | |
6714d8e8 KH |
526 | dlm->name, dlm->reco.dead_node, dlm->node_num); |
527 | ||
528 | status = dlm_remaster_locks(dlm, dlm->reco.dead_node); | |
529 | if (status < 0) { | |
6a413211 | 530 | /* we should never hit this anymore */ |
6714d8e8 KH |
531 | mlog(ML_ERROR, "error %d remastering locks for node %u, " |
532 | "retrying.\n", 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); | |
6714d8e8 KH |
540 | dlm_reset_recovery(dlm); |
541 | } | |
542 | dlm_end_recovery(dlm); | |
543 | ||
544 | /* continue and look for another dead node */ | |
545 | return -EAGAIN; | |
546 | } | |
547 | ||
548 | static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node) | |
549 | { | |
550 | int status = 0; | |
551 | struct dlm_reco_node_data *ndata; | |
552 | struct list_head *iter; | |
553 | int all_nodes_done; | |
554 | int destroy = 0; | |
555 | int pass = 0; | |
556 | ||
6a413211 KH |
557 | do { |
558 | /* we have become recovery master. there is no escaping | |
559 | * this, so just keep trying until we get it. */ | |
560 | status = dlm_init_recovery_area(dlm, dead_node); | |
561 | if (status < 0) { | |
562 | mlog(ML_ERROR, "%s: failed to alloc recovery area, " | |
563 | "retrying\n", dlm->name); | |
564 | msleep(1000); | |
565 | } | |
566 | } while (status != 0); | |
6714d8e8 KH |
567 | |
568 | /* safe to access the node data list without a lock, since this | |
569 | * process is the only one to change the list */ | |
570 | list_for_each(iter, &dlm->reco.node_data) { | |
571 | ndata = list_entry (iter, struct dlm_reco_node_data, list); | |
572 | BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT); | |
573 | ndata->state = DLM_RECO_NODE_DATA_REQUESTING; | |
574 | ||
575 | mlog(0, "requesting lock info from node %u\n", | |
576 | ndata->node_num); | |
577 | ||
578 | if (ndata->node_num == dlm->node_num) { | |
579 | ndata->state = DLM_RECO_NODE_DATA_DONE; | |
580 | continue; | |
581 | } | |
582 | ||
6a413211 KH |
583 | do { |
584 | status = dlm_request_all_locks(dlm, ndata->node_num, | |
585 | dead_node); | |
586 | if (status < 0) { | |
587 | mlog_errno(status); | |
588 | if (dlm_is_host_down(status)) { | |
589 | /* node died, ignore it for recovery */ | |
590 | status = 0; | |
591 | ndata->state = DLM_RECO_NODE_DATA_DEAD; | |
592 | /* wait for the domain map to catch up | |
593 | * with the network state. */ | |
594 | wait_event_timeout(dlm->dlm_reco_thread_wq, | |
595 | dlm_is_node_dead(dlm, | |
596 | ndata->node_num), | |
597 | msecs_to_jiffies(1000)); | |
598 | mlog(0, "waited 1 sec for %u, " | |
599 | "dead? %s\n", ndata->node_num, | |
600 | dlm_is_node_dead(dlm, ndata->node_num) ? | |
601 | "yes" : "no"); | |
602 | } else { | |
603 | /* -ENOMEM on the other node */ | |
604 | mlog(0, "%s: node %u returned " | |
605 | "%d during recovery, retrying " | |
606 | "after a short wait\n", | |
607 | dlm->name, ndata->node_num, | |
608 | status); | |
609 | msleep(100); | |
610 | } | |
6714d8e8 | 611 | } |
6a413211 | 612 | } while (status != 0); |
6714d8e8 | 613 | |
756a1501 | 614 | spin_lock(&dlm_reco_state_lock); |
6714d8e8 KH |
615 | switch (ndata->state) { |
616 | case DLM_RECO_NODE_DATA_INIT: | |
617 | case DLM_RECO_NODE_DATA_FINALIZE_SENT: | |
618 | case DLM_RECO_NODE_DATA_REQUESTED: | |
619 | BUG(); | |
620 | break; | |
621 | case DLM_RECO_NODE_DATA_DEAD: | |
622 | mlog(0, "node %u died after requesting " | |
623 | "recovery info for node %u\n", | |
624 | ndata->node_num, dead_node); | |
6a413211 KH |
625 | /* fine. don't need this node's info. |
626 | * continue without it. */ | |
627 | break; | |
6714d8e8 KH |
628 | case DLM_RECO_NODE_DATA_REQUESTING: |
629 | ndata->state = DLM_RECO_NODE_DATA_REQUESTED; | |
630 | mlog(0, "now receiving recovery data from " | |
631 | "node %u for dead node %u\n", | |
632 | ndata->node_num, dead_node); | |
633 | break; | |
634 | case DLM_RECO_NODE_DATA_RECEIVING: | |
635 | mlog(0, "already receiving recovery data from " | |
636 | "node %u for dead node %u\n", | |
637 | ndata->node_num, dead_node); | |
638 | break; | |
639 | case DLM_RECO_NODE_DATA_DONE: | |
640 | mlog(0, "already DONE receiving recovery data " | |
641 | "from node %u for dead node %u\n", | |
642 | ndata->node_num, dead_node); | |
643 | break; | |
644 | } | |
756a1501 | 645 | spin_unlock(&dlm_reco_state_lock); |
6714d8e8 KH |
646 | } |
647 | ||
648 | mlog(0, "done requesting all lock info\n"); | |
649 | ||
650 | /* nodes should be sending reco data now | |
651 | * just need to wait */ | |
652 | ||
653 | while (1) { | |
654 | /* check all the nodes now to see if we are | |
655 | * done, or if anyone died */ | |
656 | all_nodes_done = 1; | |
657 | spin_lock(&dlm_reco_state_lock); | |
658 | list_for_each(iter, &dlm->reco.node_data) { | |
659 | ndata = list_entry (iter, struct dlm_reco_node_data, list); | |
660 | ||
661 | mlog(0, "checking recovery state of node %u\n", | |
662 | ndata->node_num); | |
663 | switch (ndata->state) { | |
664 | case DLM_RECO_NODE_DATA_INIT: | |
665 | case DLM_RECO_NODE_DATA_REQUESTING: | |
666 | mlog(ML_ERROR, "bad ndata state for " | |
667 | "node %u: state=%d\n", | |
668 | ndata->node_num, ndata->state); | |
669 | BUG(); | |
670 | break; | |
671 | case DLM_RECO_NODE_DATA_DEAD: | |
6a413211 | 672 | mlog(0, "node %u died after " |
6714d8e8 KH |
673 | "requesting recovery info for " |
674 | "node %u\n", ndata->node_num, | |
675 | dead_node); | |
6a413211 | 676 | break; |
6714d8e8 KH |
677 | case DLM_RECO_NODE_DATA_RECEIVING: |
678 | case DLM_RECO_NODE_DATA_REQUESTED: | |
d6dea6e9 KH |
679 | mlog(0, "%s: node %u still in state %s\n", |
680 | dlm->name, ndata->node_num, | |
681 | ndata->state==DLM_RECO_NODE_DATA_RECEIVING ? | |
682 | "receiving" : "requested"); | |
6714d8e8 KH |
683 | all_nodes_done = 0; |
684 | break; | |
685 | case DLM_RECO_NODE_DATA_DONE: | |
d6dea6e9 KH |
686 | mlog(0, "%s: node %u state is done\n", |
687 | dlm->name, ndata->node_num); | |
6714d8e8 KH |
688 | break; |
689 | case DLM_RECO_NODE_DATA_FINALIZE_SENT: | |
d6dea6e9 KH |
690 | mlog(0, "%s: node %u state is finalize\n", |
691 | dlm->name, ndata->node_num); | |
6714d8e8 KH |
692 | break; |
693 | } | |
694 | } | |
695 | spin_unlock(&dlm_reco_state_lock); | |
696 | ||
697 | mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass, | |
698 | all_nodes_done?"yes":"no"); | |
699 | if (all_nodes_done) { | |
700 | int ret; | |
701 | ||
702 | /* all nodes are now in DLM_RECO_NODE_DATA_DONE state | |
703 | * just send a finalize message to everyone and | |
704 | * clean up */ | |
705 | mlog(0, "all nodes are done! send finalize\n"); | |
706 | ret = dlm_send_finalize_reco_message(dlm); | |
707 | if (ret < 0) | |
708 | mlog_errno(ret); | |
709 | ||
710 | spin_lock(&dlm->spinlock); | |
711 | dlm_finish_local_lockres_recovery(dlm, dead_node, | |
712 | dlm->node_num); | |
713 | spin_unlock(&dlm->spinlock); | |
714 | mlog(0, "should be done with recovery!\n"); | |
715 | ||
716 | mlog(0, "finishing recovery of %s at %lu, " | |
717 | "dead=%u, this=%u, new=%u\n", dlm->name, | |
718 | jiffies, dlm->reco.dead_node, | |
719 | dlm->node_num, dlm->reco.new_master); | |
720 | destroy = 1; | |
6a413211 | 721 | status = 0; |
6714d8e8 KH |
722 | /* rescan everything marked dirty along the way */ |
723 | dlm_kick_thread(dlm, NULL); | |
724 | break; | |
725 | } | |
726 | /* wait to be signalled, with periodic timeout | |
727 | * to check for node death */ | |
728 | wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq, | |
729 | kthread_should_stop(), | |
730 | msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS)); | |
731 | ||
732 | } | |
733 | ||
6714d8e8 KH |
734 | if (destroy) |
735 | dlm_destroy_recovery_area(dlm, dead_node); | |
736 | ||
737 | mlog_exit(status); | |
738 | return status; | |
739 | } | |
740 | ||
741 | static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node) | |
742 | { | |
743 | int num=0; | |
744 | struct dlm_reco_node_data *ndata; | |
745 | ||
746 | spin_lock(&dlm->spinlock); | |
747 | memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map)); | |
748 | /* nodes can only be removed (by dying) after dropping | |
749 | * this lock, and death will be trapped later, so this should do */ | |
750 | spin_unlock(&dlm->spinlock); | |
751 | ||
752 | while (1) { | |
753 | num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num); | |
754 | if (num >= O2NM_MAX_NODES) { | |
755 | break; | |
756 | } | |
757 | BUG_ON(num == dead_node); | |
758 | ||
cd861280 | 759 | ndata = kzalloc(sizeof(*ndata), GFP_NOFS); |
6714d8e8 KH |
760 | if (!ndata) { |
761 | dlm_destroy_recovery_area(dlm, dead_node); | |
762 | return -ENOMEM; | |
763 | } | |
764 | ndata->node_num = num; | |
765 | ndata->state = DLM_RECO_NODE_DATA_INIT; | |
766 | spin_lock(&dlm_reco_state_lock); | |
767 | list_add_tail(&ndata->list, &dlm->reco.node_data); | |
768 | spin_unlock(&dlm_reco_state_lock); | |
769 | num++; | |
770 | } | |
771 | ||
772 | return 0; | |
773 | } | |
774 | ||
775 | static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node) | |
776 | { | |
777 | struct list_head *iter, *iter2; | |
778 | struct dlm_reco_node_data *ndata; | |
779 | LIST_HEAD(tmplist); | |
780 | ||
781 | spin_lock(&dlm_reco_state_lock); | |
782 | list_splice_init(&dlm->reco.node_data, &tmplist); | |
783 | spin_unlock(&dlm_reco_state_lock); | |
784 | ||
785 | list_for_each_safe(iter, iter2, &tmplist) { | |
786 | ndata = list_entry (iter, struct dlm_reco_node_data, list); | |
787 | list_del_init(&ndata->list); | |
788 | kfree(ndata); | |
789 | } | |
790 | } | |
791 | ||
792 | static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from, | |
793 | u8 dead_node) | |
794 | { | |
795 | struct dlm_lock_request lr; | |
796 | enum dlm_status ret; | |
797 | ||
798 | mlog(0, "\n"); | |
799 | ||
800 | ||
801 | mlog(0, "dlm_request_all_locks: dead node is %u, sending request " | |
802 | "to %u\n", dead_node, request_from); | |
803 | ||
804 | memset(&lr, 0, sizeof(lr)); | |
805 | lr.node_idx = dlm->node_num; | |
806 | lr.dead_node = dead_node; | |
807 | ||
808 | // send message | |
809 | ret = DLM_NOLOCKMGR; | |
810 | ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key, | |
811 | &lr, sizeof(lr), request_from, NULL); | |
812 | ||
813 | /* negative status is handled by caller */ | |
814 | if (ret < 0) | |
815 | mlog_errno(ret); | |
816 | ||
817 | // return from here, then | |
818 | // sleep until all received or error | |
819 | return ret; | |
820 | ||
821 | } | |
822 | ||
d74c9803 KH |
823 | int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data, |
824 | void **ret_data) | |
6714d8e8 KH |
825 | { |
826 | struct dlm_ctxt *dlm = data; | |
827 | struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf; | |
828 | char *buf = NULL; | |
829 | struct dlm_work_item *item = NULL; | |
830 | ||
831 | if (!dlm_grab(dlm)) | |
832 | return -EINVAL; | |
833 | ||
c3187ce5 KH |
834 | if (lr->dead_node != dlm->reco.dead_node) { |
835 | mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local " | |
836 | "dead_node is %u\n", dlm->name, lr->node_idx, | |
837 | lr->dead_node, dlm->reco.dead_node); | |
d6dea6e9 | 838 | dlm_print_reco_node_status(dlm); |
c3187ce5 KH |
839 | /* this is a hack */ |
840 | dlm_put(dlm); | |
841 | return -ENOMEM; | |
842 | } | |
6714d8e8 KH |
843 | BUG_ON(lr->dead_node != dlm->reco.dead_node); |
844 | ||
cd861280 | 845 | item = kzalloc(sizeof(*item), GFP_NOFS); |
6714d8e8 KH |
846 | if (!item) { |
847 | dlm_put(dlm); | |
848 | return -ENOMEM; | |
849 | } | |
850 | ||
851 | /* this will get freed by dlm_request_all_locks_worker */ | |
ad8100e0 | 852 | buf = (char *) __get_free_page(GFP_NOFS); |
6714d8e8 KH |
853 | if (!buf) { |
854 | kfree(item); | |
855 | dlm_put(dlm); | |
856 | return -ENOMEM; | |
857 | } | |
858 | ||
859 | /* queue up work for dlm_request_all_locks_worker */ | |
860 | dlm_grab(dlm); /* get an extra ref for the work item */ | |
861 | dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf); | |
862 | item->u.ral.reco_master = lr->node_idx; | |
863 | item->u.ral.dead_node = lr->dead_node; | |
864 | spin_lock(&dlm->work_lock); | |
865 | list_add_tail(&item->list, &dlm->work_list); | |
866 | spin_unlock(&dlm->work_lock); | |
3156d267 | 867 | queue_work(dlm->dlm_worker, &dlm->dispatched_work); |
6714d8e8 KH |
868 | |
869 | dlm_put(dlm); | |
870 | return 0; | |
871 | } | |
872 | ||
873 | static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data) | |
874 | { | |
875 | struct dlm_migratable_lockres *mres; | |
876 | struct dlm_lock_resource *res; | |
877 | struct dlm_ctxt *dlm; | |
878 | LIST_HEAD(resources); | |
879 | struct list_head *iter; | |
880 | int ret; | |
881 | u8 dead_node, reco_master; | |
29c0fa0f | 882 | int skip_all_done = 0; |
6714d8e8 KH |
883 | |
884 | dlm = item->dlm; | |
885 | dead_node = item->u.ral.dead_node; | |
886 | reco_master = item->u.ral.reco_master; | |
e2faea4c KH |
887 | mres = (struct dlm_migratable_lockres *)data; |
888 | ||
d6dea6e9 KH |
889 | mlog(0, "%s: recovery worker started, dead=%u, master=%u\n", |
890 | dlm->name, dead_node, reco_master); | |
891 | ||
e2faea4c KH |
892 | if (dead_node != dlm->reco.dead_node || |
893 | reco_master != dlm->reco.new_master) { | |
6a413211 KH |
894 | /* worker could have been created before the recovery master |
895 | * died. if so, do not continue, but do not error. */ | |
896 | if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) { | |
897 | mlog(ML_NOTICE, "%s: will not send recovery state, " | |
898 | "recovery master %u died, thread=(dead=%u,mas=%u)" | |
899 | " current=(dead=%u,mas=%u)\n", dlm->name, | |
900 | reco_master, dead_node, reco_master, | |
901 | dlm->reco.dead_node, dlm->reco.new_master); | |
902 | } else { | |
903 | mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, " | |
904 | "master=%u), request(dead=%u, master=%u)\n", | |
905 | dlm->name, dlm->reco.dead_node, | |
906 | dlm->reco.new_master, dead_node, reco_master); | |
907 | } | |
908 | goto leave; | |
e2faea4c | 909 | } |
6714d8e8 | 910 | |
6714d8e8 KH |
911 | /* lock resources should have already been moved to the |
912 | * dlm->reco.resources list. now move items from that list | |
913 | * to a temp list if the dead owner matches. note that the | |
914 | * whole cluster recovers only one node at a time, so we | |
915 | * can safely move UNKNOWN lock resources for each recovery | |
916 | * session. */ | |
917 | dlm_move_reco_locks_to_list(dlm, &resources, dead_node); | |
918 | ||
919 | /* now we can begin blasting lockreses without the dlm lock */ | |
29c0fa0f KH |
920 | |
921 | /* any errors returned will be due to the new_master dying, | |
922 | * the dlm_reco_thread should detect this */ | |
6714d8e8 KH |
923 | list_for_each(iter, &resources) { |
924 | res = list_entry (iter, struct dlm_lock_resource, recovering); | |
925 | ret = dlm_send_one_lockres(dlm, res, mres, reco_master, | |
926 | DLM_MRES_RECOVERY); | |
29c0fa0f | 927 | if (ret < 0) { |
d6dea6e9 KH |
928 | mlog(ML_ERROR, "%s: node %u went down while sending " |
929 | "recovery state for dead node %u, ret=%d\n", dlm->name, | |
930 | reco_master, dead_node, ret); | |
29c0fa0f KH |
931 | skip_all_done = 1; |
932 | break; | |
933 | } | |
6714d8e8 KH |
934 | } |
935 | ||
936 | /* move the resources back to the list */ | |
937 | spin_lock(&dlm->spinlock); | |
938 | list_splice_init(&resources, &dlm->reco.resources); | |
939 | spin_unlock(&dlm->spinlock); | |
940 | ||
29c0fa0f KH |
941 | if (!skip_all_done) { |
942 | ret = dlm_send_all_done_msg(dlm, dead_node, reco_master); | |
943 | if (ret < 0) { | |
d6dea6e9 KH |
944 | mlog(ML_ERROR, "%s: node %u went down while sending " |
945 | "recovery all-done for dead node %u, ret=%d\n", | |
946 | dlm->name, reco_master, dead_node, ret); | |
29c0fa0f KH |
947 | } |
948 | } | |
6a413211 | 949 | leave: |
6714d8e8 KH |
950 | free_page((unsigned long)data); |
951 | } | |
952 | ||
953 | ||
954 | static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to) | |
955 | { | |
956 | int ret, tmpret; | |
957 | struct dlm_reco_data_done done_msg; | |
958 | ||
959 | memset(&done_msg, 0, sizeof(done_msg)); | |
960 | done_msg.node_idx = dlm->node_num; | |
961 | done_msg.dead_node = dead_node; | |
962 | mlog(0, "sending DATA DONE message to %u, " | |
963 | "my node=%u, dead node=%u\n", send_to, done_msg.node_idx, | |
964 | done_msg.dead_node); | |
965 | ||
966 | ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg, | |
967 | sizeof(done_msg), send_to, &tmpret); | |
29c0fa0f KH |
968 | if (ret < 0) { |
969 | if (!dlm_is_host_down(ret)) { | |
970 | mlog_errno(ret); | |
971 | mlog(ML_ERROR, "%s: unknown error sending data-done " | |
972 | "to %u\n", dlm->name, send_to); | |
973 | BUG(); | |
974 | } | |
975 | } else | |
6714d8e8 KH |
976 | ret = tmpret; |
977 | return ret; | |
978 | } | |
979 | ||
980 | ||
d74c9803 KH |
981 | int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data, |
982 | void **ret_data) | |
6714d8e8 KH |
983 | { |
984 | struct dlm_ctxt *dlm = data; | |
985 | struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf; | |
986 | struct list_head *iter; | |
987 | struct dlm_reco_node_data *ndata = NULL; | |
988 | int ret = -EINVAL; | |
989 | ||
990 | if (!dlm_grab(dlm)) | |
991 | return -EINVAL; | |
992 | ||
993 | mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, " | |
994 | "node_idx=%u, this node=%u\n", done->dead_node, | |
995 | dlm->reco.dead_node, done->node_idx, dlm->node_num); | |
d6dea6e9 KH |
996 | |
997 | mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node), | |
998 | "Got DATA DONE: dead_node=%u, reco.dead_node=%u, " | |
999 | "node_idx=%u, this node=%u\n", done->dead_node, | |
1000 | dlm->reco.dead_node, done->node_idx, dlm->node_num); | |
6714d8e8 KH |
1001 | |
1002 | spin_lock(&dlm_reco_state_lock); | |
1003 | list_for_each(iter, &dlm->reco.node_data) { | |
1004 | ndata = list_entry (iter, struct dlm_reco_node_data, list); | |
1005 | if (ndata->node_num != done->node_idx) | |
1006 | continue; | |
1007 | ||
1008 | switch (ndata->state) { | |
e2faea4c | 1009 | /* should have moved beyond INIT but not to FINALIZE yet */ |
6714d8e8 KH |
1010 | case DLM_RECO_NODE_DATA_INIT: |
1011 | case DLM_RECO_NODE_DATA_DEAD: | |
6714d8e8 KH |
1012 | case DLM_RECO_NODE_DATA_FINALIZE_SENT: |
1013 | mlog(ML_ERROR, "bad ndata state for node %u:" | |
1014 | " state=%d\n", ndata->node_num, | |
1015 | ndata->state); | |
1016 | BUG(); | |
1017 | break; | |
e2faea4c KH |
1018 | /* these states are possible at this point, anywhere along |
1019 | * the line of recovery */ | |
1020 | case DLM_RECO_NODE_DATA_DONE: | |
6714d8e8 KH |
1021 | case DLM_RECO_NODE_DATA_RECEIVING: |
1022 | case DLM_RECO_NODE_DATA_REQUESTED: | |
1023 | case DLM_RECO_NODE_DATA_REQUESTING: | |
1024 | mlog(0, "node %u is DONE sending " | |
1025 | "recovery data!\n", | |
1026 | ndata->node_num); | |
1027 | ||
1028 | ndata->state = DLM_RECO_NODE_DATA_DONE; | |
1029 | ret = 0; | |
1030 | break; | |
1031 | } | |
1032 | } | |
1033 | spin_unlock(&dlm_reco_state_lock); | |
1034 | ||
1035 | /* wake the recovery thread, some node is done */ | |
1036 | if (!ret) | |
1037 | dlm_kick_recovery_thread(dlm); | |
1038 | ||
1039 | if (ret < 0) | |
1040 | mlog(ML_ERROR, "failed to find recovery node data for node " | |
1041 | "%u\n", done->node_idx); | |
1042 | dlm_put(dlm); | |
1043 | ||
1044 | mlog(0, "leaving reco data done handler, ret=%d\n", ret); | |
1045 | return ret; | |
1046 | } | |
1047 | ||
1048 | static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm, | |
1049 | struct list_head *list, | |
1050 | u8 dead_node) | |
1051 | { | |
1052 | struct dlm_lock_resource *res; | |
1053 | struct list_head *iter, *iter2; | |
e2faea4c | 1054 | struct dlm_lock *lock; |
6714d8e8 KH |
1055 | |
1056 | spin_lock(&dlm->spinlock); | |
1057 | list_for_each_safe(iter, iter2, &dlm->reco.resources) { | |
1058 | res = list_entry (iter, struct dlm_lock_resource, recovering); | |
e2faea4c KH |
1059 | /* always prune any $RECOVERY entries for dead nodes, |
1060 | * otherwise hangs can occur during later recovery */ | |
6714d8e8 | 1061 | if (dlm_is_recovery_lock(res->lockname.name, |
e2faea4c KH |
1062 | res->lockname.len)) { |
1063 | spin_lock(&res->spinlock); | |
1064 | list_for_each_entry(lock, &res->granted, list) { | |
1065 | if (lock->ml.node == dead_node) { | |
1066 | mlog(0, "AHA! there was " | |
1067 | "a $RECOVERY lock for dead " | |
1068 | "node %u (%s)!\n", | |
1069 | dead_node, dlm->name); | |
1070 | list_del_init(&lock->list); | |
1071 | dlm_lock_put(lock); | |
1072 | break; | |
1073 | } | |
1074 | } | |
1075 | spin_unlock(&res->spinlock); | |
6714d8e8 | 1076 | continue; |
e2faea4c KH |
1077 | } |
1078 | ||
6714d8e8 KH |
1079 | if (res->owner == dead_node) { |
1080 | mlog(0, "found lockres owned by dead node while " | |
1081 | "doing recovery for node %u. sending it.\n", | |
1082 | dead_node); | |
f116629d | 1083 | list_move_tail(&res->recovering, list); |
6714d8e8 KH |
1084 | } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) { |
1085 | mlog(0, "found UNKNOWN owner while doing recovery " | |
1086 | "for node %u. sending it.\n", dead_node); | |
f116629d | 1087 | list_move_tail(&res->recovering, list); |
6714d8e8 KH |
1088 | } |
1089 | } | |
1090 | spin_unlock(&dlm->spinlock); | |
1091 | } | |
1092 | ||
1093 | static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res) | |
1094 | { | |
1095 | int total_locks = 0; | |
1096 | struct list_head *iter, *queue = &res->granted; | |
1097 | int i; | |
1098 | ||
1099 | for (i=0; i<3; i++) { | |
1100 | list_for_each(iter, queue) | |
1101 | total_locks++; | |
1102 | queue++; | |
1103 | } | |
1104 | return total_locks; | |
1105 | } | |
1106 | ||
1107 | ||
1108 | static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm, | |
1109 | struct dlm_migratable_lockres *mres, | |
1110 | u8 send_to, | |
1111 | struct dlm_lock_resource *res, | |
1112 | int total_locks) | |
1113 | { | |
1114 | u64 mig_cookie = be64_to_cpu(mres->mig_cookie); | |
1115 | int mres_total_locks = be32_to_cpu(mres->total_locks); | |
1116 | int sz, ret = 0, status = 0; | |
1117 | u8 orig_flags = mres->flags, | |
1118 | orig_master = mres->master; | |
1119 | ||
1120 | BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS); | |
1121 | if (!mres->num_locks) | |
1122 | return 0; | |
1123 | ||
1124 | sz = sizeof(struct dlm_migratable_lockres) + | |
1125 | (mres->num_locks * sizeof(struct dlm_migratable_lock)); | |
1126 | ||
1127 | /* add an all-done flag if we reached the last lock */ | |
1128 | orig_flags = mres->flags; | |
1129 | BUG_ON(total_locks > mres_total_locks); | |
1130 | if (total_locks == mres_total_locks) | |
1131 | mres->flags |= DLM_MRES_ALL_DONE; | |
1132 | ||
ba2bf218 KH |
1133 | mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n", |
1134 | dlm->name, res->lockname.len, res->lockname.name, | |
1135 | orig_flags & DLM_MRES_MIGRATION ? "migrate" : "recovery", | |
1136 | send_to); | |
1137 | ||
6714d8e8 KH |
1138 | /* send it */ |
1139 | ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres, | |
1140 | sz, send_to, &status); | |
1141 | if (ret < 0) { | |
1142 | /* XXX: negative status is not handled. | |
1143 | * this will end up killing this node. */ | |
1144 | mlog_errno(ret); | |
1145 | } else { | |
1146 | /* might get an -ENOMEM back here */ | |
1147 | ret = status; | |
1148 | if (ret < 0) { | |
1149 | mlog_errno(ret); | |
1150 | ||
1151 | if (ret == -EFAULT) { | |
1152 | mlog(ML_ERROR, "node %u told me to kill " | |
1153 | "myself!\n", send_to); | |
1154 | BUG(); | |
1155 | } | |
1156 | } | |
1157 | } | |
1158 | ||
1159 | /* zero and reinit the message buffer */ | |
1160 | dlm_init_migratable_lockres(mres, res->lockname.name, | |
1161 | res->lockname.len, mres_total_locks, | |
1162 | mig_cookie, orig_flags, orig_master); | |
1163 | return ret; | |
1164 | } | |
1165 | ||
1166 | static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres, | |
1167 | const char *lockname, int namelen, | |
1168 | int total_locks, u64 cookie, | |
1169 | u8 flags, u8 master) | |
1170 | { | |
1171 | /* mres here is one full page */ | |
1172 | memset(mres, 0, PAGE_SIZE); | |
1173 | mres->lockname_len = namelen; | |
1174 | memcpy(mres->lockname, lockname, namelen); | |
1175 | mres->num_locks = 0; | |
1176 | mres->total_locks = cpu_to_be32(total_locks); | |
1177 | mres->mig_cookie = cpu_to_be64(cookie); | |
1178 | mres->flags = flags; | |
1179 | mres->master = master; | |
1180 | } | |
1181 | ||
1182 | ||
1183 | /* returns 1 if this lock fills the network structure, | |
1184 | * 0 otherwise */ | |
1185 | static int dlm_add_lock_to_array(struct dlm_lock *lock, | |
1186 | struct dlm_migratable_lockres *mres, int queue) | |
1187 | { | |
1188 | struct dlm_migratable_lock *ml; | |
1189 | int lock_num = mres->num_locks; | |
1190 | ||
1191 | ml = &(mres->ml[lock_num]); | |
1192 | ml->cookie = lock->ml.cookie; | |
1193 | ml->type = lock->ml.type; | |
1194 | ml->convert_type = lock->ml.convert_type; | |
1195 | ml->highest_blocked = lock->ml.highest_blocked; | |
1196 | ml->list = queue; | |
1197 | if (lock->lksb) { | |
1198 | ml->flags = lock->lksb->flags; | |
1199 | /* send our current lvb */ | |
1200 | if (ml->type == LKM_EXMODE || | |
1201 | ml->type == LKM_PRMODE) { | |
1202 | /* if it is already set, this had better be a PR | |
1203 | * and it has to match */ | |
8bc674cb KH |
1204 | if (!dlm_lvb_is_empty(mres->lvb) && |
1205 | (ml->type == LKM_EXMODE || | |
1206 | memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))) { | |
6714d8e8 KH |
1207 | mlog(ML_ERROR, "mismatched lvbs!\n"); |
1208 | __dlm_print_one_lock_resource(lock->lockres); | |
1209 | BUG(); | |
1210 | } | |
1211 | memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN); | |
1212 | } | |
1213 | } | |
1214 | ml->node = lock->ml.node; | |
1215 | mres->num_locks++; | |
1216 | /* we reached the max, send this network message */ | |
1217 | if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS) | |
1218 | return 1; | |
1219 | return 0; | |
1220 | } | |
1221 | ||
ba2bf218 KH |
1222 | static void dlm_add_dummy_lock(struct dlm_ctxt *dlm, |
1223 | struct dlm_migratable_lockres *mres) | |
1224 | { | |
1225 | struct dlm_lock dummy; | |
1226 | memset(&dummy, 0, sizeof(dummy)); | |
1227 | dummy.ml.cookie = 0; | |
1228 | dummy.ml.type = LKM_IVMODE; | |
1229 | dummy.ml.convert_type = LKM_IVMODE; | |
1230 | dummy.ml.highest_blocked = LKM_IVMODE; | |
1231 | dummy.lksb = NULL; | |
1232 | dummy.ml.node = dlm->node_num; | |
1233 | dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST); | |
1234 | } | |
1235 | ||
1236 | static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm, | |
1237 | struct dlm_migratable_lock *ml, | |
1238 | u8 *nodenum) | |
1239 | { | |
1240 | if (unlikely(ml->cookie == 0 && | |
1241 | ml->type == LKM_IVMODE && | |
1242 | ml->convert_type == LKM_IVMODE && | |
1243 | ml->highest_blocked == LKM_IVMODE && | |
1244 | ml->list == DLM_BLOCKED_LIST)) { | |
1245 | *nodenum = ml->node; | |
1246 | return 1; | |
1247 | } | |
1248 | return 0; | |
1249 | } | |
6714d8e8 KH |
1250 | |
1251 | int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, | |
1252 | struct dlm_migratable_lockres *mres, | |
1253 | u8 send_to, u8 flags) | |
1254 | { | |
1255 | struct list_head *queue, *iter; | |
1256 | int total_locks, i; | |
1257 | u64 mig_cookie = 0; | |
1258 | struct dlm_lock *lock; | |
1259 | int ret = 0; | |
1260 | ||
1261 | BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION))); | |
1262 | ||
1263 | mlog(0, "sending to %u\n", send_to); | |
1264 | ||
1265 | total_locks = dlm_num_locks_in_lockres(res); | |
1266 | if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) { | |
1267 | /* rare, but possible */ | |
1268 | mlog(0, "argh. lockres has %d locks. this will " | |
1269 | "require more than one network packet to " | |
1270 | "migrate\n", total_locks); | |
1271 | mig_cookie = dlm_get_next_mig_cookie(); | |
1272 | } | |
1273 | ||
1274 | dlm_init_migratable_lockres(mres, res->lockname.name, | |
1275 | res->lockname.len, total_locks, | |
1276 | mig_cookie, flags, res->owner); | |
1277 | ||
1278 | total_locks = 0; | |
1279 | for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) { | |
1280 | queue = dlm_list_idx_to_ptr(res, i); | |
1281 | list_for_each(iter, queue) { | |
1282 | lock = list_entry (iter, struct dlm_lock, list); | |
1283 | ||
1284 | /* add another lock. */ | |
1285 | total_locks++; | |
1286 | if (!dlm_add_lock_to_array(lock, mres, i)) | |
1287 | continue; | |
1288 | ||
1289 | /* this filled the lock message, | |
1290 | * we must send it immediately. */ | |
1291 | ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, | |
1292 | res, total_locks); | |
29c0fa0f KH |
1293 | if (ret < 0) |
1294 | goto error; | |
6714d8e8 KH |
1295 | } |
1296 | } | |
ba2bf218 KH |
1297 | if (total_locks == 0) { |
1298 | /* send a dummy lock to indicate a mastery reference only */ | |
1299 | mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n", | |
1300 | dlm->name, res->lockname.len, res->lockname.name, | |
1301 | send_to, flags & DLM_MRES_RECOVERY ? "recovery" : | |
1302 | "migration"); | |
1303 | dlm_add_dummy_lock(dlm, mres); | |
1304 | } | |
6714d8e8 KH |
1305 | /* flush any remaining locks */ |
1306 | ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks); | |
29c0fa0f KH |
1307 | if (ret < 0) |
1308 | goto error; | |
1309 | return ret; | |
1310 | ||
1311 | error: | |
1312 | mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n", | |
1313 | dlm->name, ret); | |
1314 | if (!dlm_is_host_down(ret)) | |
6714d8e8 | 1315 | BUG(); |
29c0fa0f KH |
1316 | mlog(0, "%s: node %u went down while sending %s " |
1317 | "lockres %.*s\n", dlm->name, send_to, | |
1318 | flags & DLM_MRES_RECOVERY ? "recovery" : "migration", | |
1319 | res->lockname.len, res->lockname.name); | |
6714d8e8 KH |
1320 | return ret; |
1321 | } | |
1322 | ||
1323 | ||
1324 | ||
1325 | /* | |
1326 | * this message will contain no more than one page worth of | |
1327 | * recovery data, and it will work on only one lockres. | |
1328 | * there may be many locks in this page, and we may need to wait | |
1329 | * for additional packets to complete all the locks (rare, but | |
1330 | * possible). | |
1331 | */ | |
1332 | /* | |
1333 | * NOTE: the allocation error cases here are scary | |
1334 | * we really cannot afford to fail an alloc in recovery | |
1335 | * do we spin? returning an error only delays the problem really | |
1336 | */ | |
1337 | ||
d74c9803 KH |
1338 | int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data, |
1339 | void **ret_data) | |
6714d8e8 KH |
1340 | { |
1341 | struct dlm_ctxt *dlm = data; | |
1342 | struct dlm_migratable_lockres *mres = | |
1343 | (struct dlm_migratable_lockres *)msg->buf; | |
1344 | int ret = 0; | |
1345 | u8 real_master; | |
1346 | char *buf = NULL; | |
1347 | struct dlm_work_item *item = NULL; | |
1348 | struct dlm_lock_resource *res = NULL; | |
1349 | ||
1350 | if (!dlm_grab(dlm)) | |
1351 | return -EINVAL; | |
1352 | ||
1353 | BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION))); | |
1354 | ||
1355 | real_master = mres->master; | |
1356 | if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { | |
1357 | /* cannot migrate a lockres with no master */ | |
1358 | BUG_ON(!(mres->flags & DLM_MRES_RECOVERY)); | |
1359 | } | |
1360 | ||
1361 | mlog(0, "%s message received from node %u\n", | |
1362 | (mres->flags & DLM_MRES_RECOVERY) ? | |
1363 | "recovery" : "migration", mres->master); | |
1364 | if (mres->flags & DLM_MRES_ALL_DONE) | |
1365 | mlog(0, "all done flag. all lockres data received!\n"); | |
1366 | ||
1367 | ret = -ENOMEM; | |
ad8100e0 | 1368 | buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS); |
cd861280 | 1369 | item = kzalloc(sizeof(*item), GFP_NOFS); |
6714d8e8 KH |
1370 | if (!buf || !item) |
1371 | goto leave; | |
1372 | ||
1373 | /* lookup the lock to see if we have a secondary queue for this | |
1374 | * already... just add the locks in and this will have its owner | |
1375 | * and RECOVERY flag changed when it completes. */ | |
1376 | res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len); | |
1377 | if (res) { | |
1378 | /* this will get a ref on res */ | |
1379 | /* mark it as recovering/migrating and hash it */ | |
1380 | spin_lock(&res->spinlock); | |
1381 | if (mres->flags & DLM_MRES_RECOVERY) { | |
1382 | res->state |= DLM_LOCK_RES_RECOVERING; | |
1383 | } else { | |
1384 | if (res->state & DLM_LOCK_RES_MIGRATING) { | |
1385 | /* this is at least the second | |
1386 | * lockres message */ | |
1387 | mlog(0, "lock %.*s is already migrating\n", | |
1388 | mres->lockname_len, | |
1389 | mres->lockname); | |
1390 | } else if (res->state & DLM_LOCK_RES_RECOVERING) { | |
1391 | /* caller should BUG */ | |
1392 | mlog(ML_ERROR, "node is attempting to migrate " | |
1393 | "lock %.*s, but marked as recovering!\n", | |
1394 | mres->lockname_len, mres->lockname); | |
1395 | ret = -EFAULT; | |
1396 | spin_unlock(&res->spinlock); | |
1397 | goto leave; | |
1398 | } | |
1399 | res->state |= DLM_LOCK_RES_MIGRATING; | |
1400 | } | |
1401 | spin_unlock(&res->spinlock); | |
1402 | } else { | |
1403 | /* need to allocate, just like if it was | |
1404 | * mastered here normally */ | |
1405 | res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len); | |
1406 | if (!res) | |
1407 | goto leave; | |
1408 | ||
1409 | /* to match the ref that we would have gotten if | |
1410 | * dlm_lookup_lockres had succeeded */ | |
1411 | dlm_lockres_get(res); | |
1412 | ||
1413 | /* mark it as recovering/migrating and hash it */ | |
1414 | if (mres->flags & DLM_MRES_RECOVERY) | |
1415 | res->state |= DLM_LOCK_RES_RECOVERING; | |
1416 | else | |
1417 | res->state |= DLM_LOCK_RES_MIGRATING; | |
1418 | ||
1419 | spin_lock(&dlm->spinlock); | |
1420 | __dlm_insert_lockres(dlm, res); | |
1421 | spin_unlock(&dlm->spinlock); | |
1422 | ||
1423 | /* now that the new lockres is inserted, | |
1424 | * make it usable by other processes */ | |
1425 | spin_lock(&res->spinlock); | |
1426 | res->state &= ~DLM_LOCK_RES_IN_PROGRESS; | |
1427 | spin_unlock(&res->spinlock); | |
a6fa3640 | 1428 | wake_up(&res->wq); |
6714d8e8 KH |
1429 | |
1430 | /* add an extra ref for just-allocated lockres | |
1431 | * otherwise the lockres will be purged immediately */ | |
1432 | dlm_lockres_get(res); | |
6714d8e8 KH |
1433 | } |
1434 | ||
1435 | /* at this point we have allocated everything we need, | |
1436 | * and we have a hashed lockres with an extra ref and | |
1437 | * the proper res->state flags. */ | |
1438 | ret = 0; | |
ba2bf218 KH |
1439 | spin_lock(&res->spinlock); |
1440 | /* drop this either when master requery finds a different master | |
1441 | * or when a lock is added by the recovery worker */ | |
1442 | dlm_lockres_grab_inflight_ref(dlm, res); | |
6714d8e8 KH |
1443 | if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) { |
1444 | /* migration cannot have an unknown master */ | |
1445 | BUG_ON(!(mres->flags & DLM_MRES_RECOVERY)); | |
1446 | mlog(0, "recovery has passed me a lockres with an " | |
1447 | "unknown owner.. will need to requery: " | |
1448 | "%.*s\n", mres->lockname_len, mres->lockname); | |
1449 | } else { | |
ba2bf218 KH |
1450 | /* take a reference now to pin the lockres, drop it |
1451 | * when locks are added in the worker */ | |
6714d8e8 | 1452 | dlm_change_lockres_owner(dlm, res, dlm->node_num); |
6714d8e8 | 1453 | } |
ba2bf218 | 1454 | spin_unlock(&res->spinlock); |
6714d8e8 KH |
1455 | |
1456 | /* queue up work for dlm_mig_lockres_worker */ | |
1457 | dlm_grab(dlm); /* get an extra ref for the work item */ | |
1458 | memcpy(buf, msg->buf, be16_to_cpu(msg->data_len)); /* copy the whole message */ | |
1459 | dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf); | |
1460 | item->u.ml.lockres = res; /* already have a ref */ | |
1461 | item->u.ml.real_master = real_master; | |
1462 | spin_lock(&dlm->work_lock); | |
1463 | list_add_tail(&item->list, &dlm->work_list); | |
1464 | spin_unlock(&dlm->work_lock); | |
3156d267 | 1465 | queue_work(dlm->dlm_worker, &dlm->dispatched_work); |
6714d8e8 KH |
1466 | |
1467 | leave: | |
1468 | dlm_put(dlm); | |
1469 | if (ret < 0) { | |
1470 | if (buf) | |
1471 | kfree(buf); | |
1472 | if (item) | |
1473 | kfree(item); | |
1474 | } | |
1475 | ||
1476 | mlog_exit(ret); | |
1477 | return ret; | |
1478 | } | |
1479 | ||
1480 | ||
1481 | static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data) | |
1482 | { | |
1483 | struct dlm_ctxt *dlm = data; | |
1484 | struct dlm_migratable_lockres *mres; | |
1485 | int ret = 0; | |
1486 | struct dlm_lock_resource *res; | |
1487 | u8 real_master; | |
1488 | ||
1489 | dlm = item->dlm; | |
1490 | mres = (struct dlm_migratable_lockres *)data; | |
1491 | ||
1492 | res = item->u.ml.lockres; | |
1493 | real_master = item->u.ml.real_master; | |
1494 | ||
1495 | if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { | |
1496 | /* this case is super-rare. only occurs if | |
1497 | * node death happens during migration. */ | |
1498 | again: | |
1499 | ret = dlm_lockres_master_requery(dlm, res, &real_master); | |
1500 | if (ret < 0) { | |
e2faea4c | 1501 | mlog(0, "dlm_lockres_master_requery ret=%d\n", |
6714d8e8 KH |
1502 | ret); |
1503 | goto again; | |
1504 | } | |
1505 | if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { | |
1506 | mlog(0, "lockres %.*s not claimed. " | |
1507 | "this node will take it.\n", | |
1508 | res->lockname.len, res->lockname.name); | |
1509 | } else { | |
ba2bf218 KH |
1510 | spin_lock(&res->spinlock); |
1511 | dlm_lockres_drop_inflight_ref(dlm, res); | |
1512 | spin_unlock(&res->spinlock); | |
6714d8e8 KH |
1513 | mlog(0, "master needs to respond to sender " |
1514 | "that node %u still owns %.*s\n", | |
1515 | real_master, res->lockname.len, | |
1516 | res->lockname.name); | |
1517 | /* cannot touch this lockres */ | |
1518 | goto leave; | |
1519 | } | |
1520 | } | |
1521 | ||
1522 | ret = dlm_process_recovery_data(dlm, res, mres); | |
1523 | if (ret < 0) | |
1524 | mlog(0, "dlm_process_recovery_data returned %d\n", ret); | |
1525 | else | |
1526 | mlog(0, "dlm_process_recovery_data succeeded\n"); | |
1527 | ||
1528 | if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) == | |
1529 | (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) { | |
1530 | ret = dlm_finish_migration(dlm, res, mres->master); | |
1531 | if (ret < 0) | |
1532 | mlog_errno(ret); | |
1533 | } | |
1534 | ||
1535 | leave: | |
1536 | kfree(data); | |
1537 | mlog_exit(ret); | |
1538 | } | |
1539 | ||
1540 | ||
1541 | ||
8169cae5 AB |
1542 | static int dlm_lockres_master_requery(struct dlm_ctxt *dlm, |
1543 | struct dlm_lock_resource *res, | |
1544 | u8 *real_master) | |
6714d8e8 KH |
1545 | { |
1546 | struct dlm_node_iter iter; | |
1547 | int nodenum; | |
1548 | int ret = 0; | |
1549 | ||
1550 | *real_master = DLM_LOCK_RES_OWNER_UNKNOWN; | |
1551 | ||
1552 | /* we only reach here if one of the two nodes in a | |
1553 | * migration died while the migration was in progress. | |
1554 | * at this point we need to requery the master. we | |
1555 | * know that the new_master got as far as creating | |
1556 | * an mle on at least one node, but we do not know | |
1557 | * if any nodes had actually cleared the mle and set | |
1558 | * the master to the new_master. the old master | |
1559 | * is supposed to set the owner to UNKNOWN in the | |
1560 | * event of a new_master death, so the only possible | |
1561 | * responses that we can get from nodes here are | |
1562 | * that the master is new_master, or that the master | |
1563 | * is UNKNOWN. | |
1564 | * if all nodes come back with UNKNOWN then we know | |
1565 | * the lock needs remastering here. | |
1566 | * if any node comes back with a valid master, check | |
1567 | * to see if that master is the one that we are | |
1568 | * recovering. if so, then the new_master died and | |
1569 | * we need to remaster this lock. if not, then the | |
1570 | * new_master survived and that node will respond to | |
1571 | * other nodes about the owner. | |
1572 | * if there is an owner, this node needs to dump this | |
1573 | * lockres and alert the sender that this lockres | |
1574 | * was rejected. */ | |
1575 | spin_lock(&dlm->spinlock); | |
1576 | dlm_node_iter_init(dlm->domain_map, &iter); | |
1577 | spin_unlock(&dlm->spinlock); | |
1578 | ||
1579 | while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { | |
1580 | /* do not send to self */ | |
1581 | if (nodenum == dlm->node_num) | |
1582 | continue; | |
1583 | ret = dlm_do_master_requery(dlm, res, nodenum, real_master); | |
1584 | if (ret < 0) { | |
1585 | mlog_errno(ret); | |
c03872f5 KH |
1586 | if (!dlm_is_host_down(ret)) |
1587 | BUG(); | |
1588 | /* host is down, so answer for that node would be | |
1589 | * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */ | |
6714d8e8 KH |
1590 | } |
1591 | if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) { | |
1592 | mlog(0, "lock master is %u\n", *real_master); | |
1593 | break; | |
1594 | } | |
1595 | } | |
1596 | return ret; | |
1597 | } | |
1598 | ||
1599 | ||
c03872f5 KH |
1600 | int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, |
1601 | u8 nodenum, u8 *real_master) | |
6714d8e8 KH |
1602 | { |
1603 | int ret = -EINVAL; | |
1604 | struct dlm_master_requery req; | |
1605 | int status = DLM_LOCK_RES_OWNER_UNKNOWN; | |
1606 | ||
1607 | memset(&req, 0, sizeof(req)); | |
1608 | req.node_idx = dlm->node_num; | |
1609 | req.namelen = res->lockname.len; | |
1610 | memcpy(req.name, res->lockname.name, res->lockname.len); | |
1611 | ||
1612 | ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key, | |
1613 | &req, sizeof(req), nodenum, &status); | |
1614 | /* XXX: negative status not handled properly here. */ | |
1615 | if (ret < 0) | |
1616 | mlog_errno(ret); | |
1617 | else { | |
1618 | BUG_ON(status < 0); | |
1619 | BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN); | |
1620 | *real_master = (u8) (status & 0xff); | |
1621 | mlog(0, "node %u responded to master requery with %u\n", | |
1622 | nodenum, *real_master); | |
1623 | ret = 0; | |
1624 | } | |
1625 | return ret; | |
1626 | } | |
1627 | ||
1628 | ||
1629 | /* this function cannot error, so unless the sending | |
1630 | * or receiving of the message failed, the owner can | |
1631 | * be trusted */ | |
d74c9803 KH |
1632 | int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data, |
1633 | void **ret_data) | |
6714d8e8 KH |
1634 | { |
1635 | struct dlm_ctxt *dlm = data; | |
1636 | struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf; | |
1637 | struct dlm_lock_resource *res = NULL; | |
a3d33291 | 1638 | unsigned int hash; |
6714d8e8 KH |
1639 | int master = DLM_LOCK_RES_OWNER_UNKNOWN; |
1640 | u32 flags = DLM_ASSERT_MASTER_REQUERY; | |
1641 | ||
1642 | if (!dlm_grab(dlm)) { | |
1643 | /* since the domain has gone away on this | |
1644 | * node, the proper response is UNKNOWN */ | |
1645 | return master; | |
1646 | } | |
1647 | ||
a3d33291 MF |
1648 | hash = dlm_lockid_hash(req->name, req->namelen); |
1649 | ||
6714d8e8 | 1650 | spin_lock(&dlm->spinlock); |
a3d33291 | 1651 | res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash); |
6714d8e8 KH |
1652 | if (res) { |
1653 | spin_lock(&res->spinlock); | |
1654 | master = res->owner; | |
1655 | if (master == dlm->node_num) { | |
1656 | int ret = dlm_dispatch_assert_master(dlm, res, | |
1657 | 0, 0, flags); | |
1658 | if (ret < 0) { | |
1659 | mlog_errno(-ENOMEM); | |
1660 | /* retry!? */ | |
1661 | BUG(); | |
1662 | } | |
1663 | } | |
1664 | spin_unlock(&res->spinlock); | |
1665 | } | |
1666 | spin_unlock(&dlm->spinlock); | |
1667 | ||
1668 | dlm_put(dlm); | |
1669 | return master; | |
1670 | } | |
1671 | ||
1672 | static inline struct list_head * | |
1673 | dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num) | |
1674 | { | |
1675 | struct list_head *ret; | |
1676 | BUG_ON(list_num < 0); | |
1677 | BUG_ON(list_num > 2); | |
1678 | ret = &(res->granted); | |
1679 | ret += list_num; | |
1680 | return ret; | |
1681 | } | |
1682 | /* TODO: do ast flush business | |
1683 | * TODO: do MIGRATING and RECOVERING spinning | |
1684 | */ | |
1685 | ||
1686 | /* | |
1687 | * NOTE about in-flight requests during migration: | |
1688 | * | |
1689 | * Before attempting the migrate, the master has marked the lockres as | |
1690 | * MIGRATING and then flushed all of its pending ASTS. So any in-flight | |
1691 | * requests either got queued before the MIGRATING flag got set, in which | |
1692 | * case the lock data will reflect the change and a return message is on | |
1693 | * the way, or the request failed to get in before MIGRATING got set. In | |
1694 | * this case, the caller will be told to spin and wait for the MIGRATING | |
1695 | * flag to be dropped, then recheck the master. | |
1696 | * This holds true for the convert, cancel and unlock cases, and since lvb | |
1697 | * updates are tied to these same messages, it applies to lvb updates as | |
1698 | * well. For the lock case, there is no way a lock can be on the master | |
1699 | * queue and not be on the secondary queue since the lock is always added | |
1700 | * locally first. This means that the new target node will never be sent | |
1701 | * a lock that he doesn't already have on the list. | |
1702 | * In total, this means that the local lock is correct and should not be | |
1703 | * updated to match the one sent by the master. Any messages sent back | |
1704 | * from the master before the MIGRATING flag will bring the lock properly | |
1705 | * up-to-date, and the change will be ordered properly for the waiter. | |
1706 | * We will *not* attempt to modify the lock underneath the waiter. | |
1707 | */ | |
1708 | ||
1709 | static int dlm_process_recovery_data(struct dlm_ctxt *dlm, | |
1710 | struct dlm_lock_resource *res, | |
1711 | struct dlm_migratable_lockres *mres) | |
1712 | { | |
1713 | struct dlm_migratable_lock *ml; | |
1714 | struct list_head *queue; | |
e17e75ec | 1715 | struct list_head *tmpq = NULL; |
6714d8e8 KH |
1716 | struct dlm_lock *newlock = NULL; |
1717 | struct dlm_lockstatus *lksb = NULL; | |
1718 | int ret = 0; | |
e17e75ec | 1719 | int i, j, bad; |
6714d8e8 KH |
1720 | struct list_head *iter; |
1721 | struct dlm_lock *lock = NULL; | |
ba2bf218 KH |
1722 | u8 from = O2NM_MAX_NODES; |
1723 | unsigned int added = 0; | |
6714d8e8 KH |
1724 | |
1725 | mlog(0, "running %d locks for this lockres\n", mres->num_locks); | |
1726 | for (i=0; i<mres->num_locks; i++) { | |
1727 | ml = &(mres->ml[i]); | |
ba2bf218 KH |
1728 | |
1729 | if (dlm_is_dummy_lock(dlm, ml, &from)) { | |
1730 | /* placeholder, just need to set the refmap bit */ | |
1731 | BUG_ON(mres->num_locks != 1); | |
1732 | mlog(0, "%s:%.*s: dummy lock for %u\n", | |
1733 | dlm->name, mres->lockname_len, mres->lockname, | |
1734 | from); | |
1735 | spin_lock(&res->spinlock); | |
1736 | dlm_lockres_set_refmap_bit(from, res); | |
1737 | spin_unlock(&res->spinlock); | |
1738 | added++; | |
1739 | break; | |
1740 | } | |
6714d8e8 KH |
1741 | BUG_ON(ml->highest_blocked != LKM_IVMODE); |
1742 | newlock = NULL; | |
1743 | lksb = NULL; | |
1744 | ||
1745 | queue = dlm_list_num_to_pointer(res, ml->list); | |
e17e75ec | 1746 | tmpq = NULL; |
6714d8e8 KH |
1747 | |
1748 | /* if the lock is for the local node it needs to | |
1749 | * be moved to the proper location within the queue. | |
1750 | * do not allocate a new lock structure. */ | |
1751 | if (ml->node == dlm->node_num) { | |
1752 | /* MIGRATION ONLY! */ | |
1753 | BUG_ON(!(mres->flags & DLM_MRES_MIGRATION)); | |
1754 | ||
1755 | spin_lock(&res->spinlock); | |
e17e75ec KH |
1756 | for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) { |
1757 | tmpq = dlm_list_idx_to_ptr(res, j); | |
1758 | list_for_each(iter, tmpq) { | |
1759 | lock = list_entry (iter, struct dlm_lock, list); | |
1760 | if (lock->ml.cookie != ml->cookie) | |
1761 | lock = NULL; | |
1762 | else | |
1763 | break; | |
1764 | } | |
1765 | if (lock) | |
6714d8e8 KH |
1766 | break; |
1767 | } | |
1768 | ||
1769 | /* lock is always created locally first, and | |
1770 | * destroyed locally last. it must be on the list */ | |
1771 | if (!lock) { | |
a7d25539 | 1772 | __be64 c = ml->cookie; |
6714d8e8 | 1773 | mlog(ML_ERROR, "could not find local lock " |
29004858 | 1774 | "with cookie %u:%llu!\n", |
74aa2585 KH |
1775 | dlm_get_lock_cookie_node(be64_to_cpu(c)), |
1776 | dlm_get_lock_cookie_seq(be64_to_cpu(c))); | |
71ac1062 | 1777 | __dlm_print_one_lock_resource(res); |
6714d8e8 KH |
1778 | BUG(); |
1779 | } | |
1780 | BUG_ON(lock->ml.node != ml->node); | |
1781 | ||
e17e75ec KH |
1782 | if (tmpq != queue) { |
1783 | mlog(0, "lock was on %u instead of %u for %.*s\n", | |
1784 | j, ml->list, res->lockname.len, res->lockname.name); | |
1785 | spin_unlock(&res->spinlock); | |
1786 | continue; | |
1787 | } | |
1788 | ||
6714d8e8 KH |
1789 | /* see NOTE above about why we do not update |
1790 | * to match the master here */ | |
1791 | ||
1792 | /* move the lock to its proper place */ | |
1793 | /* do not alter lock refcount. switching lists. */ | |
f116629d | 1794 | list_move_tail(&lock->list, queue); |
6714d8e8 | 1795 | spin_unlock(&res->spinlock); |
ba2bf218 | 1796 | added++; |
6714d8e8 KH |
1797 | |
1798 | mlog(0, "just reordered a local lock!\n"); | |
1799 | continue; | |
1800 | } | |
1801 | ||
1802 | /* lock is for another node. */ | |
1803 | newlock = dlm_new_lock(ml->type, ml->node, | |
1804 | be64_to_cpu(ml->cookie), NULL); | |
1805 | if (!newlock) { | |
1806 | ret = -ENOMEM; | |
1807 | goto leave; | |
1808 | } | |
1809 | lksb = newlock->lksb; | |
1810 | dlm_lock_attach_lockres(newlock, res); | |
1811 | ||
1812 | if (ml->convert_type != LKM_IVMODE) { | |
1813 | BUG_ON(queue != &res->converting); | |
1814 | newlock->ml.convert_type = ml->convert_type; | |
1815 | } | |
1816 | lksb->flags |= (ml->flags & | |
1817 | (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB)); | |
ccd8b1f9 KH |
1818 | |
1819 | if (ml->type == LKM_NLMODE) | |
1820 | goto skip_lvb; | |
1821 | ||
8bc674cb | 1822 | if (!dlm_lvb_is_empty(mres->lvb)) { |
6714d8e8 KH |
1823 | if (lksb->flags & DLM_LKSB_PUT_LVB) { |
1824 | /* other node was trying to update | |
1825 | * lvb when node died. recreate the | |
1826 | * lksb with the updated lvb. */ | |
1827 | memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN); | |
ccd8b1f9 KH |
1828 | /* the lock resource lvb update must happen |
1829 | * NOW, before the spinlock is dropped. | |
1830 | * we no longer wait for the AST to update | |
1831 | * the lvb. */ | |
1832 | memcpy(res->lvb, mres->lvb, DLM_LVB_LEN); | |
6714d8e8 KH |
1833 | } else { |
1834 | /* otherwise, the node is sending its | |
1835 | * most recent valid lvb info */ | |
1836 | BUG_ON(ml->type != LKM_EXMODE && | |
1837 | ml->type != LKM_PRMODE); | |
8bc674cb | 1838 | if (!dlm_lvb_is_empty(res->lvb) && |
ccd8b1f9 KH |
1839 | (ml->type == LKM_EXMODE || |
1840 | memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) { | |
1841 | int i; | |
1842 | mlog(ML_ERROR, "%s:%.*s: received bad " | |
1843 | "lvb! type=%d\n", dlm->name, | |
1844 | res->lockname.len, | |
1845 | res->lockname.name, ml->type); | |
1846 | printk("lockres lvb=["); | |
1847 | for (i=0; i<DLM_LVB_LEN; i++) | |
1848 | printk("%02x", res->lvb[i]); | |
1849 | printk("]\nmigrated lvb=["); | |
1850 | for (i=0; i<DLM_LVB_LEN; i++) | |
1851 | printk("%02x", mres->lvb[i]); | |
1852 | printk("]\n"); | |
1853 | dlm_print_one_lock_resource(res); | |
1854 | BUG(); | |
6714d8e8 KH |
1855 | } |
1856 | memcpy(res->lvb, mres->lvb, DLM_LVB_LEN); | |
1857 | } | |
1858 | } | |
ccd8b1f9 | 1859 | skip_lvb: |
6714d8e8 KH |
1860 | |
1861 | /* NOTE: | |
1862 | * wrt lock queue ordering and recovery: | |
1863 | * 1. order of locks on granted queue is | |
1864 | * meaningless. | |
1865 | * 2. order of locks on converting queue is | |
1866 | * LOST with the node death. sorry charlie. | |
1867 | * 3. order of locks on the blocked queue is | |
1868 | * also LOST. | |
1869 | * order of locks does not affect integrity, it | |
1870 | * just means that a lock request may get pushed | |
1871 | * back in line as a result of the node death. | |
1872 | * also note that for a given node the lock order | |
1873 | * for its secondary queue locks is preserved | |
1874 | * relative to each other, but clearly *not* | |
1875 | * preserved relative to locks from other nodes. | |
1876 | */ | |
c3187ce5 | 1877 | bad = 0; |
6714d8e8 | 1878 | spin_lock(&res->spinlock); |
c3187ce5 KH |
1879 | list_for_each_entry(lock, queue, list) { |
1880 | if (lock->ml.cookie == ml->cookie) { | |
a7d25539 | 1881 | __be64 c = lock->ml.cookie; |
c3187ce5 KH |
1882 | mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already " |
1883 | "exists on this lockres!\n", dlm->name, | |
1884 | res->lockname.len, res->lockname.name, | |
74aa2585 KH |
1885 | dlm_get_lock_cookie_node(be64_to_cpu(c)), |
1886 | dlm_get_lock_cookie_seq(be64_to_cpu(c))); | |
c3187ce5 KH |
1887 | |
1888 | mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, " | |
1889 | "node=%u, cookie=%u:%llu, queue=%d\n", | |
1890 | ml->type, ml->convert_type, ml->node, | |
74aa2585 KH |
1891 | dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)), |
1892 | dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)), | |
c3187ce5 KH |
1893 | ml->list); |
1894 | ||
1895 | __dlm_print_one_lock_resource(res); | |
1896 | bad = 1; | |
1897 | break; | |
1898 | } | |
1899 | } | |
1900 | if (!bad) { | |
1901 | dlm_lock_get(newlock); | |
1902 | list_add_tail(&newlock->list, queue); | |
ba2bf218 KH |
1903 | mlog(0, "%s:%.*s: added lock for node %u, " |
1904 | "setting refmap bit\n", dlm->name, | |
1905 | res->lockname.len, res->lockname.name, ml->node); | |
1906 | dlm_lockres_set_refmap_bit(ml->node, res); | |
1907 | added++; | |
c3187ce5 | 1908 | } |
6714d8e8 KH |
1909 | spin_unlock(&res->spinlock); |
1910 | } | |
1911 | mlog(0, "done running all the locks\n"); | |
1912 | ||
1913 | leave: | |
ba2bf218 | 1914 | /* balance the ref taken when the work was queued */ |
50635f15 KH |
1915 | spin_lock(&res->spinlock); |
1916 | dlm_lockres_drop_inflight_ref(dlm, res); | |
1917 | spin_unlock(&res->spinlock); | |
ba2bf218 | 1918 | |
6714d8e8 KH |
1919 | if (ret < 0) { |
1920 | mlog_errno(ret); | |
1921 | if (newlock) | |
1922 | dlm_lock_put(newlock); | |
1923 | } | |
1924 | ||
1925 | mlog_exit(ret); | |
1926 | return ret; | |
1927 | } | |
1928 | ||
1929 | void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm, | |
1930 | struct dlm_lock_resource *res) | |
1931 | { | |
1932 | int i; | |
1933 | struct list_head *queue, *iter, *iter2; | |
1934 | struct dlm_lock *lock; | |
1935 | ||
1936 | res->state |= DLM_LOCK_RES_RECOVERING; | |
69d72b06 KH |
1937 | if (!list_empty(&res->recovering)) { |
1938 | mlog(0, | |
1939 | "Recovering res %s:%.*s, is already on recovery list!\n", | |
1940 | dlm->name, res->lockname.len, res->lockname.name); | |
6714d8e8 | 1941 | list_del_init(&res->recovering); |
69d72b06 KH |
1942 | } |
1943 | /* We need to hold a reference while on the recovery list */ | |
1944 | dlm_lockres_get(res); | |
6714d8e8 KH |
1945 | list_add_tail(&res->recovering, &dlm->reco.resources); |
1946 | ||
1947 | /* find any pending locks and put them back on proper list */ | |
1948 | for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) { | |
1949 | queue = dlm_list_idx_to_ptr(res, i); | |
1950 | list_for_each_safe(iter, iter2, queue) { | |
1951 | lock = list_entry (iter, struct dlm_lock, list); | |
1952 | dlm_lock_get(lock); | |
1953 | if (lock->convert_pending) { | |
1954 | /* move converting lock back to granted */ | |
1955 | BUG_ON(i != DLM_CONVERTING_LIST); | |
1956 | mlog(0, "node died with convert pending " | |
1957 | "on %.*s. move back to granted list.\n", | |
1958 | res->lockname.len, res->lockname.name); | |
1959 | dlm_revert_pending_convert(res, lock); | |
1960 | lock->convert_pending = 0; | |
1961 | } else if (lock->lock_pending) { | |
1962 | /* remove pending lock requests completely */ | |
1963 | BUG_ON(i != DLM_BLOCKED_LIST); | |
1964 | mlog(0, "node died with lock pending " | |
1965 | "on %.*s. remove from blocked list and skip.\n", | |
1966 | res->lockname.len, res->lockname.name); | |
1967 | /* lock will be floating until ref in | |
1968 | * dlmlock_remote is freed after the network | |
1969 | * call returns. ok for it to not be on any | |
1970 | * list since no ast can be called | |
1971 | * (the master is dead). */ | |
1972 | dlm_revert_pending_lock(res, lock); | |
1973 | lock->lock_pending = 0; | |
1974 | } else if (lock->unlock_pending) { | |
1975 | /* if an unlock was in progress, treat as | |
1976 | * if this had completed successfully | |
1977 | * before sending this lock state to the | |
1978 | * new master. note that the dlm_unlock | |
1979 | * call is still responsible for calling | |
1980 | * the unlockast. that will happen after | |
1981 | * the network call times out. for now, | |
1982 | * just move lists to prepare the new | |
1983 | * recovery master. */ | |
1984 | BUG_ON(i != DLM_GRANTED_LIST); | |
1985 | mlog(0, "node died with unlock pending " | |
1986 | "on %.*s. remove from blocked list and skip.\n", | |
1987 | res->lockname.len, res->lockname.name); | |
1988 | dlm_commit_pending_unlock(res, lock); | |
1989 | lock->unlock_pending = 0; | |
1990 | } else if (lock->cancel_pending) { | |
1991 | /* if a cancel was in progress, treat as | |
1992 | * if this had completed successfully | |
1993 | * before sending this lock state to the | |
1994 | * new master */ | |
1995 | BUG_ON(i != DLM_CONVERTING_LIST); | |
1996 | mlog(0, "node died with cancel pending " | |
1997 | "on %.*s. move back to granted list.\n", | |
1998 | res->lockname.len, res->lockname.name); | |
1999 | dlm_commit_pending_cancel(res, lock); | |
2000 | lock->cancel_pending = 0; | |
2001 | } | |
2002 | dlm_lock_put(lock); | |
2003 | } | |
2004 | } | |
2005 | } | |
2006 | ||
2007 | ||
2008 | ||
2009 | /* removes all recovered locks from the recovery list. | |
2010 | * sets the res->owner to the new master. | |
2011 | * unsets the RECOVERY flag and wakes waiters. */ | |
2012 | static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm, | |
2013 | u8 dead_node, u8 new_master) | |
2014 | { | |
2015 | int i; | |
81f2094a MF |
2016 | struct list_head *iter, *iter2; |
2017 | struct hlist_node *hash_iter; | |
2018 | struct hlist_head *bucket; | |
2019 | ||
6714d8e8 KH |
2020 | struct dlm_lock_resource *res; |
2021 | ||
2022 | mlog_entry_void(); | |
2023 | ||
2024 | assert_spin_locked(&dlm->spinlock); | |
2025 | ||
2026 | list_for_each_safe(iter, iter2, &dlm->reco.resources) { | |
2027 | res = list_entry (iter, struct dlm_lock_resource, recovering); | |
2028 | if (res->owner == dead_node) { | |
2029 | list_del_init(&res->recovering); | |
2030 | spin_lock(&res->spinlock); | |
ba2bf218 KH |
2031 | /* new_master has our reference from |
2032 | * the lock state sent during recovery */ | |
6714d8e8 KH |
2033 | dlm_change_lockres_owner(dlm, res, new_master); |
2034 | res->state &= ~DLM_LOCK_RES_RECOVERING; | |
ba2bf218 | 2035 | if (__dlm_lockres_has_locks(res)) |
69d72b06 | 2036 | __dlm_dirty_lockres(dlm, res); |
6714d8e8 KH |
2037 | spin_unlock(&res->spinlock); |
2038 | wake_up(&res->wq); | |
69d72b06 | 2039 | dlm_lockres_put(res); |
6714d8e8 KH |
2040 | } |
2041 | } | |
2042 | ||
2043 | /* this will become unnecessary eventually, but | |
2044 | * for now we need to run the whole hash, clear | |
2045 | * the RECOVERING state and set the owner | |
2046 | * if necessary */ | |
81f2094a | 2047 | for (i = 0; i < DLM_HASH_BUCKETS; i++) { |
03d864c0 | 2048 | bucket = dlm_lockres_hash(dlm, i); |
81f2094a | 2049 | hlist_for_each_entry(res, hash_iter, bucket, hash_node) { |
6714d8e8 KH |
2050 | if (res->state & DLM_LOCK_RES_RECOVERING) { |
2051 | if (res->owner == dead_node) { | |
2052 | mlog(0, "(this=%u) res %.*s owner=%u " | |
2053 | "was not on recovering list, but " | |
2054 | "clearing state anyway\n", | |
2055 | dlm->node_num, res->lockname.len, | |
2056 | res->lockname.name, new_master); | |
2057 | } else if (res->owner == dlm->node_num) { | |
2058 | mlog(0, "(this=%u) res %.*s owner=%u " | |
2059 | "was not on recovering list, " | |
2060 | "owner is THIS node, clearing\n", | |
2061 | dlm->node_num, res->lockname.len, | |
2062 | res->lockname.name, new_master); | |
2063 | } else | |
2064 | continue; | |
2065 | ||
c03872f5 KH |
2066 | if (!list_empty(&res->recovering)) { |
2067 | mlog(0, "%s:%.*s: lockres was " | |
2068 | "marked RECOVERING, owner=%u\n", | |
2069 | dlm->name, res->lockname.len, | |
2070 | res->lockname.name, res->owner); | |
2071 | list_del_init(&res->recovering); | |
69d72b06 | 2072 | dlm_lockres_put(res); |
c03872f5 | 2073 | } |
6714d8e8 | 2074 | spin_lock(&res->spinlock); |
ba2bf218 KH |
2075 | /* new_master has our reference from |
2076 | * the lock state sent during recovery */ | |
6714d8e8 KH |
2077 | dlm_change_lockres_owner(dlm, res, new_master); |
2078 | res->state &= ~DLM_LOCK_RES_RECOVERING; | |
ba2bf218 | 2079 | if (__dlm_lockres_has_locks(res)) |
69d72b06 | 2080 | __dlm_dirty_lockres(dlm, res); |
6714d8e8 KH |
2081 | spin_unlock(&res->spinlock); |
2082 | wake_up(&res->wq); | |
2083 | } | |
2084 | } | |
2085 | } | |
2086 | } | |
2087 | ||
2088 | static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local) | |
2089 | { | |
2090 | if (local) { | |
2091 | if (lock->ml.type != LKM_EXMODE && | |
2092 | lock->ml.type != LKM_PRMODE) | |
2093 | return 1; | |
2094 | } else if (lock->ml.type == LKM_EXMODE) | |
2095 | return 1; | |
2096 | return 0; | |
2097 | } | |
2098 | ||
2099 | static void dlm_revalidate_lvb(struct dlm_ctxt *dlm, | |
2100 | struct dlm_lock_resource *res, u8 dead_node) | |
2101 | { | |
2102 | struct list_head *iter, *queue; | |
2103 | struct dlm_lock *lock; | |
2104 | int blank_lvb = 0, local = 0; | |
2105 | int i; | |
2106 | u8 search_node; | |
2107 | ||
2108 | assert_spin_locked(&dlm->spinlock); | |
2109 | assert_spin_locked(&res->spinlock); | |
2110 | ||
2111 | if (res->owner == dlm->node_num) | |
2112 | /* if this node owned the lockres, and if the dead node | |
2113 | * had an EX when he died, blank out the lvb */ | |
2114 | search_node = dead_node; | |
2115 | else { | |
2116 | /* if this is a secondary lockres, and we had no EX or PR | |
2117 | * locks granted, we can no longer trust the lvb */ | |
2118 | search_node = dlm->node_num; | |
2119 | local = 1; /* check local state for valid lvb */ | |
2120 | } | |
2121 | ||
2122 | for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) { | |
2123 | queue = dlm_list_idx_to_ptr(res, i); | |
2124 | list_for_each(iter, queue) { | |
2125 | lock = list_entry (iter, struct dlm_lock, list); | |
2126 | if (lock->ml.node == search_node) { | |
2127 | if (dlm_lvb_needs_invalidation(lock, local)) { | |
2128 | /* zero the lksb lvb and lockres lvb */ | |
2129 | blank_lvb = 1; | |
2130 | memset(lock->lksb->lvb, 0, DLM_LVB_LEN); | |
2131 | } | |
2132 | } | |
2133 | } | |
2134 | } | |
2135 | ||
2136 | if (blank_lvb) { | |
2137 | mlog(0, "clearing %.*s lvb, dead node %u had EX\n", | |
2138 | res->lockname.len, res->lockname.name, dead_node); | |
2139 | memset(res->lvb, 0, DLM_LVB_LEN); | |
2140 | } | |
2141 | } | |
2142 | ||
2143 | static void dlm_free_dead_locks(struct dlm_ctxt *dlm, | |
2144 | struct dlm_lock_resource *res, u8 dead_node) | |
2145 | { | |
2146 | struct list_head *iter, *tmpiter; | |
2147 | struct dlm_lock *lock; | |
ba2bf218 | 2148 | unsigned int freed = 0; |
6714d8e8 KH |
2149 | |
2150 | /* this node is the lockres master: | |
2151 | * 1) remove any stale locks for the dead node | |
2152 | * 2) if the dead node had an EX when he died, blank out the lvb | |
2153 | */ | |
2154 | assert_spin_locked(&dlm->spinlock); | |
2155 | assert_spin_locked(&res->spinlock); | |
2156 | ||
2157 | /* TODO: check pending_asts, pending_basts here */ | |
2158 | list_for_each_safe(iter, tmpiter, &res->granted) { | |
2159 | lock = list_entry (iter, struct dlm_lock, list); | |
2160 | if (lock->ml.node == dead_node) { | |
2161 | list_del_init(&lock->list); | |
2162 | dlm_lock_put(lock); | |
ba2bf218 | 2163 | freed++; |
6714d8e8 KH |
2164 | } |
2165 | } | |
2166 | list_for_each_safe(iter, tmpiter, &res->converting) { | |
2167 | lock = list_entry (iter, struct dlm_lock, list); | |
2168 | if (lock->ml.node == dead_node) { | |
2169 | list_del_init(&lock->list); | |
2170 | dlm_lock_put(lock); | |
ba2bf218 | 2171 | freed++; |
6714d8e8 KH |
2172 | } |
2173 | } | |
2174 | list_for_each_safe(iter, tmpiter, &res->blocked) { | |
2175 | lock = list_entry (iter, struct dlm_lock, list); | |
2176 | if (lock->ml.node == dead_node) { | |
2177 | list_del_init(&lock->list); | |
2178 | dlm_lock_put(lock); | |
ba2bf218 | 2179 | freed++; |
6714d8e8 KH |
2180 | } |
2181 | } | |
2182 | ||
ba2bf218 KH |
2183 | if (freed) { |
2184 | mlog(0, "%s:%.*s: freed %u locks for dead node %u, " | |
2185 | "dropping ref from lockres\n", dlm->name, | |
2186 | res->lockname.len, res->lockname.name, freed, dead_node); | |
2187 | BUG_ON(!test_bit(dead_node, res->refmap)); | |
2188 | dlm_lockres_clear_refmap_bit(dead_node, res); | |
2189 | } else if (test_bit(dead_node, res->refmap)) { | |
2190 | mlog(0, "%s:%.*s: dead node %u had a ref, but had " | |
2191 | "no locks and had not purged before dying\n", dlm->name, | |
2192 | res->lockname.len, res->lockname.name, dead_node); | |
2193 | dlm_lockres_clear_refmap_bit(dead_node, res); | |
2194 | } | |
2195 | ||
6714d8e8 KH |
2196 | /* do not kick thread yet */ |
2197 | __dlm_dirty_lockres(dlm, res); | |
2198 | } | |
2199 | ||
2200 | /* if this node is the recovery master, and there are no | |
2201 | * locks for a given lockres owned by this node that are in | |
2202 | * either PR or EX mode, zero out the lvb before requesting. | |
2203 | * | |
2204 | */ | |
2205 | ||
2206 | ||
2207 | static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node) | |
2208 | { | |
81f2094a | 2209 | struct hlist_node *iter; |
6714d8e8 KH |
2210 | struct dlm_lock_resource *res; |
2211 | int i; | |
81f2094a | 2212 | struct hlist_head *bucket; |
e2faea4c | 2213 | struct dlm_lock *lock; |
6714d8e8 KH |
2214 | |
2215 | ||
2216 | /* purge any stale mles */ | |
2217 | dlm_clean_master_list(dlm, dead_node); | |
2218 | ||
2219 | /* | |
2220 | * now clean up all lock resources. there are two rules: | |
2221 | * | |
2222 | * 1) if the dead node was the master, move the lockres | |
2223 | * to the recovering list. set the RECOVERING flag. | |
2224 | * this lockres needs to be cleaned up before it can | |
2225 | * be used further. | |
2226 | * | |
2227 | * 2) if this node was the master, remove all locks from | |
2228 | * each of the lockres queues that were owned by the | |
2229 | * dead node. once recovery finishes, the dlm thread | |
2230 | * can be kicked again to see if any ASTs or BASTs | |
2231 | * need to be fired as a result. | |
2232 | */ | |
81f2094a | 2233 | for (i = 0; i < DLM_HASH_BUCKETS; i++) { |
03d864c0 | 2234 | bucket = dlm_lockres_hash(dlm, i); |
81f2094a | 2235 | hlist_for_each_entry(res, iter, bucket, hash_node) { |
e2faea4c KH |
2236 | /* always prune any $RECOVERY entries for dead nodes, |
2237 | * otherwise hangs can occur during later recovery */ | |
6714d8e8 | 2238 | if (dlm_is_recovery_lock(res->lockname.name, |
e2faea4c KH |
2239 | res->lockname.len)) { |
2240 | spin_lock(&res->spinlock); | |
2241 | list_for_each_entry(lock, &res->granted, list) { | |
2242 | if (lock->ml.node == dead_node) { | |
2243 | mlog(0, "AHA! there was " | |
2244 | "a $RECOVERY lock for dead " | |
2245 | "node %u (%s)!\n", | |
2246 | dead_node, dlm->name); | |
2247 | list_del_init(&lock->list); | |
2248 | dlm_lock_put(lock); | |
2249 | break; | |
2250 | } | |
2251 | } | |
2252 | spin_unlock(&res->spinlock); | |
6714d8e8 | 2253 | continue; |
e2faea4c | 2254 | } |
6714d8e8 KH |
2255 | spin_lock(&res->spinlock); |
2256 | /* zero the lvb if necessary */ | |
2257 | dlm_revalidate_lvb(dlm, res, dead_node); | |
ba2bf218 KH |
2258 | if (res->owner == dead_node) { |
2259 | if (res->state & DLM_LOCK_RES_DROPPING_REF) | |
2260 | mlog(0, "%s:%.*s: owned by " | |
2261 | "dead node %u, this node was " | |
2262 | "dropping its ref when it died. " | |
2263 | "continue, dropping the flag.\n", | |
2264 | dlm->name, res->lockname.len, | |
2265 | res->lockname.name, dead_node); | |
2266 | ||
2267 | /* the wake_up for this will happen when the | |
2268 | * RECOVERING flag is dropped later */ | |
2269 | res->state &= ~DLM_LOCK_RES_DROPPING_REF; | |
2270 | ||
6714d8e8 | 2271 | dlm_move_lockres_to_recovery_list(dlm, res); |
ba2bf218 | 2272 | } else if (res->owner == dlm->node_num) { |
6714d8e8 KH |
2273 | dlm_free_dead_locks(dlm, res, dead_node); |
2274 | __dlm_lockres_calc_usage(dlm, res); | |
2275 | } | |
2276 | spin_unlock(&res->spinlock); | |
2277 | } | |
2278 | } | |
2279 | ||
2280 | } | |
2281 | ||
2282 | static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx) | |
2283 | { | |
2284 | assert_spin_locked(&dlm->spinlock); | |
2285 | ||
466d1a45 KH |
2286 | if (dlm->reco.new_master == idx) { |
2287 | mlog(0, "%s: recovery master %d just died\n", | |
2288 | dlm->name, idx); | |
2289 | if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { | |
2290 | /* finalize1 was reached, so it is safe to clear | |
2291 | * the new_master and dead_node. that recovery | |
2292 | * is complete. */ | |
2293 | mlog(0, "%s: dead master %d had reached " | |
2294 | "finalize1 state, clearing\n", dlm->name, idx); | |
2295 | dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE; | |
2296 | __dlm_reset_recovery(dlm); | |
2297 | } | |
2298 | } | |
2299 | ||
6714d8e8 KH |
2300 | /* check to see if the node is already considered dead */ |
2301 | if (!test_bit(idx, dlm->live_nodes_map)) { | |
2302 | mlog(0, "for domain %s, node %d is already dead. " | |
2303 | "another node likely did recovery already.\n", | |
2304 | dlm->name, idx); | |
2305 | return; | |
2306 | } | |
2307 | ||
2308 | /* check to see if we do not care about this node */ | |
2309 | if (!test_bit(idx, dlm->domain_map)) { | |
2310 | /* This also catches the case that we get a node down | |
2311 | * but haven't joined the domain yet. */ | |
2312 | mlog(0, "node %u already removed from domain!\n", idx); | |
2313 | return; | |
2314 | } | |
2315 | ||
2316 | clear_bit(idx, dlm->live_nodes_map); | |
2317 | ||
2318 | /* Clean up join state on node death. */ | |
2319 | if (dlm->joining_node == idx) { | |
2320 | mlog(0, "Clearing join state for node %u\n", idx); | |
2321 | __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); | |
2322 | } | |
2323 | ||
2324 | /* make sure local cleanup occurs before the heartbeat events */ | |
2325 | if (!test_bit(idx, dlm->recovery_map)) | |
2326 | dlm_do_local_recovery_cleanup(dlm, idx); | |
2327 | ||
2328 | /* notify anything attached to the heartbeat events */ | |
2329 | dlm_hb_event_notify_attached(dlm, idx, 0); | |
2330 | ||
2331 | mlog(0, "node %u being removed from domain map!\n", idx); | |
2332 | clear_bit(idx, dlm->domain_map); | |
2333 | /* wake up migration waiters if a node goes down. | |
2334 | * perhaps later we can genericize this for other waiters. */ | |
2335 | wake_up(&dlm->migration_wq); | |
2336 | ||
2337 | if (test_bit(idx, dlm->recovery_map)) | |
2338 | mlog(0, "domain %s, node %u already added " | |
2339 | "to recovery map!\n", dlm->name, idx); | |
2340 | else | |
2341 | set_bit(idx, dlm->recovery_map); | |
2342 | } | |
2343 | ||
2344 | void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data) | |
2345 | { | |
2346 | struct dlm_ctxt *dlm = data; | |
2347 | ||
2348 | if (!dlm_grab(dlm)) | |
2349 | return; | |
2350 | ||
2351 | spin_lock(&dlm->spinlock); | |
2352 | __dlm_hb_node_down(dlm, idx); | |
2353 | spin_unlock(&dlm->spinlock); | |
2354 | ||
2355 | dlm_put(dlm); | |
2356 | } | |
2357 | ||
2358 | void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data) | |
2359 | { | |
2360 | struct dlm_ctxt *dlm = data; | |
2361 | ||
2362 | if (!dlm_grab(dlm)) | |
2363 | return; | |
2364 | ||
2365 | spin_lock(&dlm->spinlock); | |
6714d8e8 | 2366 | set_bit(idx, dlm->live_nodes_map); |
e2faea4c KH |
2367 | /* do NOT notify mle attached to the heartbeat events. |
2368 | * new nodes are not interesting in mastery until joined. */ | |
6714d8e8 KH |
2369 | spin_unlock(&dlm->spinlock); |
2370 | ||
2371 | dlm_put(dlm); | |
2372 | } | |
2373 | ||
2374 | static void dlm_reco_ast(void *astdata) | |
2375 | { | |
2376 | struct dlm_ctxt *dlm = astdata; | |
2377 | mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n", | |
2378 | dlm->node_num, dlm->name); | |
2379 | } | |
2380 | static void dlm_reco_bast(void *astdata, int blocked_type) | |
2381 | { | |
2382 | struct dlm_ctxt *dlm = astdata; | |
2383 | mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n", | |
2384 | dlm->node_num, dlm->name); | |
2385 | } | |
2386 | static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st) | |
2387 | { | |
2388 | mlog(0, "unlockast for recovery lock fired!\n"); | |
2389 | } | |
2390 | ||
e2faea4c KH |
2391 | /* |
2392 | * dlm_pick_recovery_master will continually attempt to use | |
2393 | * dlmlock() on the special "$RECOVERY" lockres with the | |
2394 | * LKM_NOQUEUE flag to get an EX. every thread that enters | |
2395 | * this function on each node racing to become the recovery | |
2396 | * master will not stop attempting this until either: | |
2397 | * a) this node gets the EX (and becomes the recovery master), | |
2398 | * or b) dlm->reco.new_master gets set to some nodenum | |
2399 | * != O2NM_INVALID_NODE_NUM (another node will do the reco). | |
2400 | * so each time a recovery master is needed, the entire cluster | |
2401 | * will sync at this point. if the new master dies, that will | |
2402 | * be detected in dlm_do_recovery */ | |
6714d8e8 KH |
2403 | static int dlm_pick_recovery_master(struct dlm_ctxt *dlm) |
2404 | { | |
2405 | enum dlm_status ret; | |
2406 | struct dlm_lockstatus lksb; | |
2407 | int status = -EINVAL; | |
2408 | ||
2409 | mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n", | |
2410 | dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num); | |
e2faea4c | 2411 | again: |
6714d8e8 KH |
2412 | memset(&lksb, 0, sizeof(lksb)); |
2413 | ||
2414 | ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY, | |
3384f3df MF |
2415 | DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN, |
2416 | dlm_reco_ast, dlm, dlm_reco_bast); | |
6714d8e8 | 2417 | |
e2faea4c KH |
2418 | mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n", |
2419 | dlm->name, ret, lksb.status); | |
2420 | ||
6714d8e8 KH |
2421 | if (ret == DLM_NORMAL) { |
2422 | mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n", | |
2423 | dlm->name, dlm->node_num); | |
e2faea4c KH |
2424 | |
2425 | /* got the EX lock. check to see if another node | |
2426 | * just became the reco master */ | |
2427 | if (dlm_reco_master_ready(dlm)) { | |
2428 | mlog(0, "%s: got reco EX lock, but %u will " | |
2429 | "do the recovery\n", dlm->name, | |
2430 | dlm->reco.new_master); | |
2431 | status = -EEXIST; | |
2432 | } else { | |
898effac KH |
2433 | status = 0; |
2434 | ||
2435 | /* see if recovery was already finished elsewhere */ | |
2436 | spin_lock(&dlm->spinlock); | |
2437 | if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { | |
2438 | status = -EINVAL; | |
2439 | mlog(0, "%s: got reco EX lock, but " | |
2440 | "node got recovered already\n", dlm->name); | |
2441 | if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) { | |
2442 | mlog(ML_ERROR, "%s: new master is %u " | |
2443 | "but no dead node!\n", | |
2444 | dlm->name, dlm->reco.new_master); | |
2445 | BUG(); | |
2446 | } | |
2447 | } | |
2448 | spin_unlock(&dlm->spinlock); | |
2449 | } | |
2450 | ||
2451 | /* if this node has actually become the recovery master, | |
2452 | * set the master and send the messages to begin recovery */ | |
2453 | if (!status) { | |
2454 | mlog(0, "%s: dead=%u, this=%u, sending " | |
2455 | "begin_reco now\n", dlm->name, | |
2456 | dlm->reco.dead_node, dlm->node_num); | |
e2faea4c KH |
2457 | status = dlm_send_begin_reco_message(dlm, |
2458 | dlm->reco.dead_node); | |
2459 | /* this always succeeds */ | |
2460 | BUG_ON(status); | |
2461 | ||
2462 | /* set the new_master to this node */ | |
2463 | spin_lock(&dlm->spinlock); | |
ab27eb6f | 2464 | dlm_set_reco_master(dlm, dlm->node_num); |
e2faea4c KH |
2465 | spin_unlock(&dlm->spinlock); |
2466 | } | |
6714d8e8 KH |
2467 | |
2468 | /* recovery lock is a special case. ast will not get fired, | |
2469 | * so just go ahead and unlock it. */ | |
2470 | ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm); | |
e2faea4c KH |
2471 | if (ret == DLM_DENIED) { |
2472 | mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n"); | |
2473 | ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm); | |
2474 | } | |
6714d8e8 KH |
2475 | if (ret != DLM_NORMAL) { |
2476 | /* this would really suck. this could only happen | |
2477 | * if there was a network error during the unlock | |
2478 | * because of node death. this means the unlock | |
2479 | * is actually "done" and the lock structure is | |
2480 | * even freed. we can continue, but only | |
2481 | * because this specific lock name is special. */ | |
e2faea4c | 2482 | mlog(ML_ERROR, "dlmunlock returned %d\n", ret); |
6714d8e8 KH |
2483 | } |
2484 | } else if (ret == DLM_NOTQUEUED) { | |
2485 | mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n", | |
2486 | dlm->name, dlm->node_num); | |
2487 | /* another node is master. wait on | |
e2faea4c KH |
2488 | * reco.new_master != O2NM_INVALID_NODE_NUM |
2489 | * for at most one second */ | |
2490 | wait_event_timeout(dlm->dlm_reco_thread_wq, | |
2491 | dlm_reco_master_ready(dlm), | |
2492 | msecs_to_jiffies(1000)); | |
2493 | if (!dlm_reco_master_ready(dlm)) { | |
2494 | mlog(0, "%s: reco master taking awhile\n", | |
2495 | dlm->name); | |
2496 | goto again; | |
2497 | } | |
2498 | /* another node has informed this one that it is reco master */ | |
2499 | mlog(0, "%s: reco master %u is ready to recover %u\n", | |
2500 | dlm->name, dlm->reco.new_master, dlm->reco.dead_node); | |
6714d8e8 | 2501 | status = -EEXIST; |
c8df412e KH |
2502 | } else if (ret == DLM_RECOVERING) { |
2503 | mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n", | |
2504 | dlm->name, dlm->node_num); | |
2505 | goto again; | |
e2faea4c KH |
2506 | } else { |
2507 | struct dlm_lock_resource *res; | |
2508 | ||
2509 | /* dlmlock returned something other than NOTQUEUED or NORMAL */ | |
2510 | mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), " | |
2511 | "lksb.status=%s\n", dlm->name, dlm_errname(ret), | |
2512 | dlm_errname(lksb.status)); | |
2513 | res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME, | |
2514 | DLM_RECOVERY_LOCK_NAME_LEN); | |
2515 | if (res) { | |
2516 | dlm_print_one_lock_resource(res); | |
2517 | dlm_lockres_put(res); | |
2518 | } else { | |
2519 | mlog(ML_ERROR, "recovery lock not found\n"); | |
2520 | } | |
2521 | BUG(); | |
6714d8e8 KH |
2522 | } |
2523 | ||
2524 | return status; | |
2525 | } | |
2526 | ||
2527 | static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node) | |
2528 | { | |
2529 | struct dlm_begin_reco br; | |
2530 | int ret = 0; | |
2531 | struct dlm_node_iter iter; | |
2532 | int nodenum; | |
2533 | int status; | |
2534 | ||
2535 | mlog_entry("%u\n", dead_node); | |
2536 | ||
d6dea6e9 | 2537 | mlog(0, "%s: dead node is %u\n", dlm->name, dead_node); |
6714d8e8 KH |
2538 | |
2539 | spin_lock(&dlm->spinlock); | |
2540 | dlm_node_iter_init(dlm->domain_map, &iter); | |
2541 | spin_unlock(&dlm->spinlock); | |
2542 | ||
2543 | clear_bit(dead_node, iter.node_map); | |
2544 | ||
2545 | memset(&br, 0, sizeof(br)); | |
2546 | br.node_idx = dlm->node_num; | |
2547 | br.dead_node = dead_node; | |
2548 | ||
2549 | while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { | |
2550 | ret = 0; | |
2551 | if (nodenum == dead_node) { | |
2552 | mlog(0, "not sending begin reco to dead node " | |
2553 | "%u\n", dead_node); | |
2554 | continue; | |
2555 | } | |
2556 | if (nodenum == dlm->node_num) { | |
2557 | mlog(0, "not sending begin reco to self\n"); | |
2558 | continue; | |
2559 | } | |
e2faea4c | 2560 | retry: |
6714d8e8 KH |
2561 | ret = -EINVAL; |
2562 | mlog(0, "attempting to send begin reco msg to %d\n", | |
2563 | nodenum); | |
2564 | ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key, | |
2565 | &br, sizeof(br), nodenum, &status); | |
2566 | /* negative status is handled ok by caller here */ | |
2567 | if (ret >= 0) | |
2568 | ret = status; | |
e2faea4c KH |
2569 | if (dlm_is_host_down(ret)) { |
2570 | /* node is down. not involved in recovery | |
2571 | * so just keep going */ | |
2572 | mlog(0, "%s: node %u was down when sending " | |
2573 | "begin reco msg (%d)\n", dlm->name, nodenum, ret); | |
2574 | ret = 0; | |
2575 | } | |
6714d8e8 KH |
2576 | if (ret < 0) { |
2577 | struct dlm_lock_resource *res; | |
e2faea4c KH |
2578 | /* this is now a serious problem, possibly ENOMEM |
2579 | * in the network stack. must retry */ | |
6714d8e8 KH |
2580 | mlog_errno(ret); |
2581 | mlog(ML_ERROR, "begin reco of dlm %s to node %u " | |
2582 | " returned %d\n", dlm->name, nodenum, ret); | |
2583 | res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME, | |
2584 | DLM_RECOVERY_LOCK_NAME_LEN); | |
2585 | if (res) { | |
2586 | dlm_print_one_lock_resource(res); | |
2587 | dlm_lockres_put(res); | |
2588 | } else { | |
2589 | mlog(ML_ERROR, "recovery lock not found\n"); | |
2590 | } | |
e2faea4c KH |
2591 | /* sleep for a bit in hopes that we can avoid |
2592 | * another ENOMEM */ | |
2593 | msleep(100); | |
2594 | goto retry; | |
466d1a45 KH |
2595 | } else if (ret == EAGAIN) { |
2596 | mlog(0, "%s: trying to start recovery of node " | |
2597 | "%u, but node %u is waiting for last recovery " | |
2598 | "to complete, backoff for a bit\n", dlm->name, | |
2599 | dead_node, nodenum); | |
2600 | /* TODO Look into replacing msleep with cond_resched() */ | |
2601 | msleep(100); | |
2602 | goto retry; | |
6714d8e8 KH |
2603 | } |
2604 | } | |
2605 | ||
2606 | return ret; | |
2607 | } | |
2608 | ||
d74c9803 KH |
2609 | int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data, |
2610 | void **ret_data) | |
6714d8e8 KH |
2611 | { |
2612 | struct dlm_ctxt *dlm = data; | |
2613 | struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf; | |
2614 | ||
2615 | /* ok to return 0, domain has gone away */ | |
2616 | if (!dlm_grab(dlm)) | |
2617 | return 0; | |
2618 | ||
466d1a45 KH |
2619 | spin_lock(&dlm->spinlock); |
2620 | if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { | |
2621 | mlog(0, "%s: node %u wants to recover node %u (%u:%u) " | |
2622 | "but this node is in finalize state, waiting on finalize2\n", | |
2623 | dlm->name, br->node_idx, br->dead_node, | |
2624 | dlm->reco.dead_node, dlm->reco.new_master); | |
2625 | spin_unlock(&dlm->spinlock); | |
2626 | return EAGAIN; | |
2627 | } | |
2628 | spin_unlock(&dlm->spinlock); | |
2629 | ||
d6dea6e9 KH |
2630 | mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n", |
2631 | dlm->name, br->node_idx, br->dead_node, | |
2632 | dlm->reco.dead_node, dlm->reco.new_master); | |
6714d8e8 KH |
2633 | |
2634 | dlm_fire_domain_eviction_callbacks(dlm, br->dead_node); | |
2635 | ||
2636 | spin_lock(&dlm->spinlock); | |
2637 | if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) { | |
e2faea4c KH |
2638 | if (test_bit(dlm->reco.new_master, dlm->recovery_map)) { |
2639 | mlog(0, "%s: new_master %u died, changing " | |
2640 | "to %u\n", dlm->name, dlm->reco.new_master, | |
2641 | br->node_idx); | |
2642 | } else { | |
2643 | mlog(0, "%s: new_master %u NOT DEAD, changing " | |
2644 | "to %u\n", dlm->name, dlm->reco.new_master, | |
2645 | br->node_idx); | |
2646 | /* may not have seen the new master as dead yet */ | |
2647 | } | |
6714d8e8 KH |
2648 | } |
2649 | if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) { | |
e2faea4c KH |
2650 | mlog(ML_NOTICE, "%s: dead_node previously set to %u, " |
2651 | "node %u changing it to %u\n", dlm->name, | |
2652 | dlm->reco.dead_node, br->node_idx, br->dead_node); | |
6714d8e8 | 2653 | } |
ab27eb6f KH |
2654 | dlm_set_reco_master(dlm, br->node_idx); |
2655 | dlm_set_reco_dead_node(dlm, br->dead_node); | |
6714d8e8 | 2656 | if (!test_bit(br->dead_node, dlm->recovery_map)) { |
e2faea4c | 2657 | mlog(0, "recovery master %u sees %u as dead, but this " |
6714d8e8 KH |
2658 | "node has not yet. marking %u as dead\n", |
2659 | br->node_idx, br->dead_node, br->dead_node); | |
e2faea4c KH |
2660 | if (!test_bit(br->dead_node, dlm->domain_map) || |
2661 | !test_bit(br->dead_node, dlm->live_nodes_map)) | |
2662 | mlog(0, "%u not in domain/live_nodes map " | |
2663 | "so setting it in reco map manually\n", | |
2664 | br->dead_node); | |
c03872f5 KH |
2665 | /* force the recovery cleanup in __dlm_hb_node_down |
2666 | * both of these will be cleared in a moment */ | |
2667 | set_bit(br->dead_node, dlm->domain_map); | |
2668 | set_bit(br->dead_node, dlm->live_nodes_map); | |
6714d8e8 KH |
2669 | __dlm_hb_node_down(dlm, br->dead_node); |
2670 | } | |
2671 | spin_unlock(&dlm->spinlock); | |
2672 | ||
2673 | dlm_kick_recovery_thread(dlm); | |
d6dea6e9 KH |
2674 | |
2675 | mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n", | |
2676 | dlm->name, br->node_idx, br->dead_node, | |
2677 | dlm->reco.dead_node, dlm->reco.new_master); | |
2678 | ||
6714d8e8 KH |
2679 | dlm_put(dlm); |
2680 | return 0; | |
2681 | } | |
2682 | ||
466d1a45 | 2683 | #define DLM_FINALIZE_STAGE2 0x01 |
6714d8e8 KH |
2684 | static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm) |
2685 | { | |
2686 | int ret = 0; | |
2687 | struct dlm_finalize_reco fr; | |
2688 | struct dlm_node_iter iter; | |
2689 | int nodenum; | |
2690 | int status; | |
466d1a45 | 2691 | int stage = 1; |
6714d8e8 | 2692 | |
466d1a45 KH |
2693 | mlog(0, "finishing recovery for node %s:%u, " |
2694 | "stage %d\n", dlm->name, dlm->reco.dead_node, stage); | |
6714d8e8 KH |
2695 | |
2696 | spin_lock(&dlm->spinlock); | |
2697 | dlm_node_iter_init(dlm->domain_map, &iter); | |
2698 | spin_unlock(&dlm->spinlock); | |
2699 | ||
466d1a45 | 2700 | stage2: |
6714d8e8 KH |
2701 | memset(&fr, 0, sizeof(fr)); |
2702 | fr.node_idx = dlm->node_num; | |
2703 | fr.dead_node = dlm->reco.dead_node; | |
466d1a45 KH |
2704 | if (stage == 2) |
2705 | fr.flags |= DLM_FINALIZE_STAGE2; | |
6714d8e8 KH |
2706 | |
2707 | while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { | |
2708 | if (nodenum == dlm->node_num) | |
2709 | continue; | |
2710 | ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key, | |
2711 | &fr, sizeof(fr), nodenum, &status); | |
466d1a45 | 2712 | if (ret >= 0) |
6714d8e8 | 2713 | ret = status; |
466d1a45 KH |
2714 | if (ret < 0) { |
2715 | mlog_errno(ret); | |
6714d8e8 KH |
2716 | if (dlm_is_host_down(ret)) { |
2717 | /* this has no effect on this recovery | |
2718 | * session, so set the status to zero to | |
2719 | * finish out the last recovery */ | |
2720 | mlog(ML_ERROR, "node %u went down after this " | |
2721 | "node finished recovery.\n", nodenum); | |
2722 | ret = 0; | |
c27069e6 | 2723 | continue; |
6714d8e8 | 2724 | } |
6714d8e8 KH |
2725 | break; |
2726 | } | |
2727 | } | |
466d1a45 KH |
2728 | if (stage == 1) { |
2729 | /* reset the node_iter back to the top and send finalize2 */ | |
2730 | iter.curnode = -1; | |
2731 | stage = 2; | |
2732 | goto stage2; | |
2733 | } | |
6714d8e8 KH |
2734 | |
2735 | return ret; | |
2736 | } | |
2737 | ||
d74c9803 KH |
2738 | int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data, |
2739 | void **ret_data) | |
6714d8e8 KH |
2740 | { |
2741 | struct dlm_ctxt *dlm = data; | |
2742 | struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf; | |
466d1a45 | 2743 | int stage = 1; |
6714d8e8 KH |
2744 | |
2745 | /* ok to return 0, domain has gone away */ | |
2746 | if (!dlm_grab(dlm)) | |
2747 | return 0; | |
2748 | ||
466d1a45 KH |
2749 | if (fr->flags & DLM_FINALIZE_STAGE2) |
2750 | stage = 2; | |
6714d8e8 | 2751 | |
466d1a45 KH |
2752 | mlog(0, "%s: node %u finalizing recovery stage%d of " |
2753 | "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage, | |
2754 | fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master); | |
2755 | ||
6714d8e8 KH |
2756 | spin_lock(&dlm->spinlock); |
2757 | ||
2758 | if (dlm->reco.new_master != fr->node_idx) { | |
2759 | mlog(ML_ERROR, "node %u sent recovery finalize msg, but node " | |
2760 | "%u is supposed to be the new master, dead=%u\n", | |
2761 | fr->node_idx, dlm->reco.new_master, fr->dead_node); | |
2762 | BUG(); | |
2763 | } | |
2764 | if (dlm->reco.dead_node != fr->dead_node) { | |
2765 | mlog(ML_ERROR, "node %u sent recovery finalize msg for dead " | |
2766 | "node %u, but node %u is supposed to be dead\n", | |
2767 | fr->node_idx, fr->dead_node, dlm->reco.dead_node); | |
2768 | BUG(); | |
2769 | } | |
2770 | ||
466d1a45 KH |
2771 | switch (stage) { |
2772 | case 1: | |
2773 | dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx); | |
2774 | if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { | |
2775 | mlog(ML_ERROR, "%s: received finalize1 from " | |
2776 | "new master %u for dead node %u, but " | |
2777 | "this node has already received it!\n", | |
2778 | dlm->name, fr->node_idx, fr->dead_node); | |
2779 | dlm_print_reco_node_status(dlm); | |
2780 | BUG(); | |
2781 | } | |
2782 | dlm->reco.state |= DLM_RECO_STATE_FINALIZE; | |
2783 | spin_unlock(&dlm->spinlock); | |
2784 | break; | |
2785 | case 2: | |
2786 | if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) { | |
2787 | mlog(ML_ERROR, "%s: received finalize2 from " | |
2788 | "new master %u for dead node %u, but " | |
2789 | "this node did not have finalize1!\n", | |
2790 | dlm->name, fr->node_idx, fr->dead_node); | |
2791 | dlm_print_reco_node_status(dlm); | |
2792 | BUG(); | |
2793 | } | |
2794 | dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE; | |
2795 | spin_unlock(&dlm->spinlock); | |
2796 | dlm_reset_recovery(dlm); | |
2797 | dlm_kick_recovery_thread(dlm); | |
2798 | break; | |
2799 | default: | |
2800 | BUG(); | |
2801 | } | |
6714d8e8 | 2802 | |
d6dea6e9 KH |
2803 | mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n", |
2804 | dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master); | |
2805 | ||
6714d8e8 KH |
2806 | dlm_put(dlm); |
2807 | return 0; | |
2808 | } |