Commit | Line | Data |
---|---|---|
6714d8e8 KH |
1 | /* -*- mode: c; c-basic-offset: 8; -*- |
2 | * vim: noexpandtab sw=8 ts=8 sts=0: | |
3 | * | |
4 | * dlmdomain.c | |
5 | * | |
6 | * defines domain join / leave apis | |
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 | #include <linux/module.h> | |
28 | #include <linux/types.h> | |
29 | #include <linux/slab.h> | |
30 | #include <linux/highmem.h> | |
31 | #include <linux/utsname.h> | |
32 | #include <linux/init.h> | |
33 | #include <linux/spinlock.h> | |
34 | #include <linux/delay.h> | |
35 | #include <linux/err.h> | |
36 | ||
37 | #include "cluster/heartbeat.h" | |
38 | #include "cluster/nodemanager.h" | |
39 | #include "cluster/tcp.h" | |
40 | ||
41 | #include "dlmapi.h" | |
42 | #include "dlmcommon.h" | |
43 | ||
44 | #include "dlmdebug.h" | |
45 | #include "dlmdomain.h" | |
46 | ||
47 | #include "dlmver.h" | |
48 | ||
49 | #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN) | |
50 | #include "cluster/masklog.h" | |
51 | ||
52 | /* | |
53 | * | |
54 | * spinlock lock ordering: if multiple locks are needed, obey this ordering: | |
55 | * dlm_domain_lock | |
56 | * struct dlm_ctxt->spinlock | |
57 | * struct dlm_lock_resource->spinlock | |
58 | * struct dlm_ctxt->master_lock | |
59 | * struct dlm_ctxt->ast_lock | |
60 | * dlm_master_list_entry->spinlock | |
61 | * dlm_lock->spinlock | |
62 | * | |
63 | */ | |
64 | ||
65 | spinlock_t dlm_domain_lock = SPIN_LOCK_UNLOCKED; | |
66 | LIST_HEAD(dlm_domains); | |
67 | static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events); | |
68 | ||
69 | #define DLM_DOMAIN_BACKOFF_MS 200 | |
70 | ||
71 | static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data); | |
72 | static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data); | |
73 | static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data); | |
74 | static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data); | |
75 | ||
76 | static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm); | |
77 | ||
78 | void __dlm_unhash_lockres(struct dlm_lock_resource *lockres) | |
79 | { | |
81f2094a | 80 | hlist_del_init(&lockres->hash_node); |
6714d8e8 KH |
81 | dlm_lockres_put(lockres); |
82 | } | |
83 | ||
84 | void __dlm_insert_lockres(struct dlm_ctxt *dlm, | |
85 | struct dlm_lock_resource *res) | |
86 | { | |
81f2094a | 87 | struct hlist_head *bucket; |
6714d8e8 KH |
88 | struct qstr *q; |
89 | ||
90 | assert_spin_locked(&dlm->spinlock); | |
91 | ||
92 | q = &res->lockname; | |
93 | q->hash = full_name_hash(q->name, q->len); | |
81f2094a | 94 | bucket = &(dlm->lockres_hash[q->hash % DLM_HASH_BUCKETS]); |
6714d8e8 KH |
95 | |
96 | /* get a reference for our hashtable */ | |
97 | dlm_lockres_get(res); | |
98 | ||
81f2094a | 99 | hlist_add_head(&res->hash_node, bucket); |
6714d8e8 KH |
100 | } |
101 | ||
102 | struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm, | |
103 | const char *name, | |
104 | unsigned int len) | |
105 | { | |
106 | unsigned int hash; | |
81f2094a | 107 | struct hlist_node *iter; |
6714d8e8 | 108 | struct dlm_lock_resource *tmpres=NULL; |
81f2094a | 109 | struct hlist_head *bucket; |
6714d8e8 KH |
110 | |
111 | mlog_entry("%.*s\n", len, name); | |
112 | ||
113 | assert_spin_locked(&dlm->spinlock); | |
114 | ||
115 | hash = full_name_hash(name, len); | |
116 | ||
81f2094a | 117 | bucket = &(dlm->lockres_hash[hash % DLM_HASH_BUCKETS]); |
6714d8e8 KH |
118 | |
119 | /* check for pre-existing lock */ | |
81f2094a MF |
120 | hlist_for_each(iter, bucket) { |
121 | tmpres = hlist_entry(iter, struct dlm_lock_resource, hash_node); | |
6714d8e8 KH |
122 | if (tmpres->lockname.len == len && |
123 | memcmp(tmpres->lockname.name, name, len) == 0) { | |
124 | dlm_lockres_get(tmpres); | |
125 | break; | |
126 | } | |
127 | ||
128 | tmpres = NULL; | |
129 | } | |
130 | return tmpres; | |
131 | } | |
132 | ||
133 | struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm, | |
134 | const char *name, | |
135 | unsigned int len) | |
136 | { | |
137 | struct dlm_lock_resource *res; | |
138 | ||
139 | spin_lock(&dlm->spinlock); | |
140 | res = __dlm_lookup_lockres(dlm, name, len); | |
141 | spin_unlock(&dlm->spinlock); | |
142 | return res; | |
143 | } | |
144 | ||
145 | static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len) | |
146 | { | |
147 | struct dlm_ctxt *tmp = NULL; | |
148 | struct list_head *iter; | |
149 | ||
150 | assert_spin_locked(&dlm_domain_lock); | |
151 | ||
152 | /* tmp->name here is always NULL terminated, | |
153 | * but domain may not be! */ | |
154 | list_for_each(iter, &dlm_domains) { | |
155 | tmp = list_entry (iter, struct dlm_ctxt, list); | |
156 | if (strlen(tmp->name) == len && | |
157 | memcmp(tmp->name, domain, len)==0) | |
158 | break; | |
159 | tmp = NULL; | |
160 | } | |
161 | ||
162 | return tmp; | |
163 | } | |
164 | ||
165 | /* For null terminated domain strings ONLY */ | |
166 | static struct dlm_ctxt * __dlm_lookup_domain(const char *domain) | |
167 | { | |
168 | assert_spin_locked(&dlm_domain_lock); | |
169 | ||
170 | return __dlm_lookup_domain_full(domain, strlen(domain)); | |
171 | } | |
172 | ||
173 | ||
174 | /* returns true on one of two conditions: | |
175 | * 1) the domain does not exist | |
176 | * 2) the domain exists and it's state is "joined" */ | |
177 | static int dlm_wait_on_domain_helper(const char *domain) | |
178 | { | |
179 | int ret = 0; | |
180 | struct dlm_ctxt *tmp = NULL; | |
181 | ||
182 | spin_lock(&dlm_domain_lock); | |
183 | ||
184 | tmp = __dlm_lookup_domain(domain); | |
185 | if (!tmp) | |
186 | ret = 1; | |
187 | else if (tmp->dlm_state == DLM_CTXT_JOINED) | |
188 | ret = 1; | |
189 | ||
190 | spin_unlock(&dlm_domain_lock); | |
191 | return ret; | |
192 | } | |
193 | ||
194 | static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm) | |
195 | { | |
81f2094a MF |
196 | if (dlm->lockres_hash) |
197 | free_page((unsigned long) dlm->lockres_hash); | |
6714d8e8 KH |
198 | |
199 | if (dlm->name) | |
200 | kfree(dlm->name); | |
201 | ||
202 | kfree(dlm); | |
203 | } | |
204 | ||
205 | /* A little strange - this function will be called while holding | |
206 | * dlm_domain_lock and is expected to be holding it on the way out. We | |
207 | * will however drop and reacquire it multiple times */ | |
208 | static void dlm_ctxt_release(struct kref *kref) | |
209 | { | |
210 | struct dlm_ctxt *dlm; | |
211 | ||
212 | dlm = container_of(kref, struct dlm_ctxt, dlm_refs); | |
213 | ||
214 | BUG_ON(dlm->num_joins); | |
215 | BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED); | |
216 | ||
217 | /* we may still be in the list if we hit an error during join. */ | |
218 | list_del_init(&dlm->list); | |
219 | ||
220 | spin_unlock(&dlm_domain_lock); | |
221 | ||
222 | mlog(0, "freeing memory from domain %s\n", dlm->name); | |
223 | ||
224 | wake_up(&dlm_domain_events); | |
225 | ||
226 | dlm_free_ctxt_mem(dlm); | |
227 | ||
228 | spin_lock(&dlm_domain_lock); | |
229 | } | |
230 | ||
231 | void dlm_put(struct dlm_ctxt *dlm) | |
232 | { | |
233 | spin_lock(&dlm_domain_lock); | |
234 | kref_put(&dlm->dlm_refs, dlm_ctxt_release); | |
235 | spin_unlock(&dlm_domain_lock); | |
236 | } | |
237 | ||
238 | static void __dlm_get(struct dlm_ctxt *dlm) | |
239 | { | |
240 | kref_get(&dlm->dlm_refs); | |
241 | } | |
242 | ||
243 | /* given a questionable reference to a dlm object, gets a reference if | |
244 | * it can find it in the list, otherwise returns NULL in which case | |
245 | * you shouldn't trust your pointer. */ | |
246 | struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm) | |
247 | { | |
248 | struct list_head *iter; | |
249 | struct dlm_ctxt *target = NULL; | |
250 | ||
251 | spin_lock(&dlm_domain_lock); | |
252 | ||
253 | list_for_each(iter, &dlm_domains) { | |
254 | target = list_entry (iter, struct dlm_ctxt, list); | |
255 | ||
256 | if (target == dlm) { | |
257 | __dlm_get(target); | |
258 | break; | |
259 | } | |
260 | ||
261 | target = NULL; | |
262 | } | |
263 | ||
264 | spin_unlock(&dlm_domain_lock); | |
265 | ||
266 | return target; | |
267 | } | |
268 | ||
269 | int dlm_domain_fully_joined(struct dlm_ctxt *dlm) | |
270 | { | |
271 | int ret; | |
272 | ||
273 | spin_lock(&dlm_domain_lock); | |
274 | ret = (dlm->dlm_state == DLM_CTXT_JOINED) || | |
275 | (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN); | |
276 | spin_unlock(&dlm_domain_lock); | |
277 | ||
278 | return ret; | |
279 | } | |
280 | ||
281 | static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm) | |
282 | { | |
283 | dlm_unregister_domain_handlers(dlm); | |
284 | dlm_complete_thread(dlm); | |
285 | dlm_complete_recovery_thread(dlm); | |
286 | ||
287 | /* We've left the domain. Now we can take ourselves out of the | |
288 | * list and allow the kref stuff to help us free the | |
289 | * memory. */ | |
290 | spin_lock(&dlm_domain_lock); | |
291 | list_del_init(&dlm->list); | |
292 | spin_unlock(&dlm_domain_lock); | |
293 | ||
294 | /* Wake up anyone waiting for us to remove this domain */ | |
295 | wake_up(&dlm_domain_events); | |
296 | } | |
297 | ||
298 | static void dlm_migrate_all_locks(struct dlm_ctxt *dlm) | |
299 | { | |
300 | int i; | |
301 | struct dlm_lock_resource *res; | |
302 | ||
303 | mlog(0, "Migrating locks from domain %s\n", dlm->name); | |
304 | restart: | |
305 | spin_lock(&dlm->spinlock); | |
81f2094a MF |
306 | for (i = 0; i < DLM_HASH_BUCKETS; i++) { |
307 | while (!hlist_empty(&dlm->lockres_hash[i])) { | |
308 | res = hlist_entry(dlm->lockres_hash[i].first, | |
309 | struct dlm_lock_resource, hash_node); | |
6714d8e8 KH |
310 | /* need reference when manually grabbing lockres */ |
311 | dlm_lockres_get(res); | |
312 | /* this should unhash the lockres | |
313 | * and exit with dlm->spinlock */ | |
314 | mlog(0, "purging res=%p\n", res); | |
315 | if (dlm_lockres_is_dirty(dlm, res)) { | |
316 | /* HACK! this should absolutely go. | |
317 | * need to figure out why some empty | |
318 | * lockreses are still marked dirty */ | |
319 | mlog(ML_ERROR, "lockres %.*s dirty!\n", | |
320 | res->lockname.len, res->lockname.name); | |
321 | ||
322 | spin_unlock(&dlm->spinlock); | |
323 | dlm_kick_thread(dlm, res); | |
324 | wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res)); | |
325 | dlm_lockres_put(res); | |
326 | goto restart; | |
327 | } | |
328 | dlm_purge_lockres(dlm, res); | |
329 | dlm_lockres_put(res); | |
330 | } | |
331 | } | |
332 | spin_unlock(&dlm->spinlock); | |
333 | ||
334 | mlog(0, "DONE Migrating locks from domain %s\n", dlm->name); | |
335 | } | |
336 | ||
337 | static int dlm_no_joining_node(struct dlm_ctxt *dlm) | |
338 | { | |
339 | int ret; | |
340 | ||
341 | spin_lock(&dlm->spinlock); | |
342 | ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN; | |
343 | spin_unlock(&dlm->spinlock); | |
344 | ||
345 | return ret; | |
346 | } | |
347 | ||
348 | static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm) | |
349 | { | |
350 | /* Yikes, a double spinlock! I need domain_lock for the dlm | |
351 | * state and the dlm spinlock for join state... Sorry! */ | |
352 | again: | |
353 | spin_lock(&dlm_domain_lock); | |
354 | spin_lock(&dlm->spinlock); | |
355 | ||
356 | if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) { | |
357 | mlog(0, "Node %d is joining, we wait on it.\n", | |
358 | dlm->joining_node); | |
359 | spin_unlock(&dlm->spinlock); | |
360 | spin_unlock(&dlm_domain_lock); | |
361 | ||
362 | wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm)); | |
363 | goto again; | |
364 | } | |
365 | ||
366 | dlm->dlm_state = DLM_CTXT_LEAVING; | |
367 | spin_unlock(&dlm->spinlock); | |
368 | spin_unlock(&dlm_domain_lock); | |
369 | } | |
370 | ||
371 | static void __dlm_print_nodes(struct dlm_ctxt *dlm) | |
372 | { | |
373 | int node = -1; | |
374 | ||
375 | assert_spin_locked(&dlm->spinlock); | |
376 | ||
377 | mlog(ML_NOTICE, "Nodes in my domain (\"%s\"):\n", dlm->name); | |
378 | ||
379 | while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, | |
380 | node + 1)) < O2NM_MAX_NODES) { | |
381 | mlog(ML_NOTICE, " node %d\n", node); | |
382 | } | |
383 | } | |
384 | ||
385 | static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data) | |
386 | { | |
387 | struct dlm_ctxt *dlm = data; | |
388 | unsigned int node; | |
389 | struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf; | |
390 | ||
391 | mlog_entry("%p %u %p", msg, len, data); | |
392 | ||
393 | if (!dlm_grab(dlm)) | |
394 | return 0; | |
395 | ||
396 | node = exit_msg->node_idx; | |
397 | ||
398 | mlog(0, "Node %u leaves domain %s\n", node, dlm->name); | |
399 | ||
400 | spin_lock(&dlm->spinlock); | |
401 | clear_bit(node, dlm->domain_map); | |
402 | __dlm_print_nodes(dlm); | |
403 | ||
404 | /* notify anything attached to the heartbeat events */ | |
405 | dlm_hb_event_notify_attached(dlm, node, 0); | |
406 | ||
407 | spin_unlock(&dlm->spinlock); | |
408 | ||
409 | dlm_put(dlm); | |
410 | ||
411 | return 0; | |
412 | } | |
413 | ||
414 | static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm, | |
415 | unsigned int node) | |
416 | { | |
417 | int status; | |
418 | struct dlm_exit_domain leave_msg; | |
419 | ||
420 | mlog(0, "Asking node %u if we can leave the domain %s me = %u\n", | |
421 | node, dlm->name, dlm->node_num); | |
422 | ||
423 | memset(&leave_msg, 0, sizeof(leave_msg)); | |
424 | leave_msg.node_idx = dlm->node_num; | |
425 | ||
426 | status = o2net_send_message(DLM_EXIT_DOMAIN_MSG, dlm->key, | |
427 | &leave_msg, sizeof(leave_msg), node, | |
428 | NULL); | |
429 | ||
430 | mlog(0, "status return %d from o2net_send_message\n", status); | |
431 | ||
432 | return status; | |
433 | } | |
434 | ||
435 | ||
436 | static void dlm_leave_domain(struct dlm_ctxt *dlm) | |
437 | { | |
438 | int node, clear_node, status; | |
439 | ||
440 | /* At this point we've migrated away all our locks and won't | |
441 | * accept mastership of new ones. The dlm is responsible for | |
442 | * almost nothing now. We make sure not to confuse any joining | |
443 | * nodes and then commence shutdown procedure. */ | |
444 | ||
445 | spin_lock(&dlm->spinlock); | |
446 | /* Clear ourselves from the domain map */ | |
447 | clear_bit(dlm->node_num, dlm->domain_map); | |
448 | while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, | |
449 | 0)) < O2NM_MAX_NODES) { | |
450 | /* Drop the dlm spinlock. This is safe wrt the domain_map. | |
451 | * -nodes cannot be added now as the | |
452 | * query_join_handlers knows to respond with OK_NO_MAP | |
453 | * -we catch the right network errors if a node is | |
454 | * removed from the map while we're sending him the | |
455 | * exit message. */ | |
456 | spin_unlock(&dlm->spinlock); | |
457 | ||
458 | clear_node = 1; | |
459 | ||
460 | status = dlm_send_one_domain_exit(dlm, node); | |
461 | if (status < 0 && | |
462 | status != -ENOPROTOOPT && | |
463 | status != -ENOTCONN) { | |
464 | mlog(ML_NOTICE, "Error %d sending domain exit message " | |
465 | "to node %d\n", status, node); | |
466 | ||
467 | /* Not sure what to do here but lets sleep for | |
468 | * a bit in case this was a transient | |
469 | * error... */ | |
470 | msleep(DLM_DOMAIN_BACKOFF_MS); | |
471 | clear_node = 0; | |
472 | } | |
473 | ||
474 | spin_lock(&dlm->spinlock); | |
475 | /* If we're not clearing the node bit then we intend | |
476 | * to loop back around to try again. */ | |
477 | if (clear_node) | |
478 | clear_bit(node, dlm->domain_map); | |
479 | } | |
480 | spin_unlock(&dlm->spinlock); | |
481 | } | |
482 | ||
483 | int dlm_joined(struct dlm_ctxt *dlm) | |
484 | { | |
485 | int ret = 0; | |
486 | ||
487 | spin_lock(&dlm_domain_lock); | |
488 | ||
489 | if (dlm->dlm_state == DLM_CTXT_JOINED) | |
490 | ret = 1; | |
491 | ||
492 | spin_unlock(&dlm_domain_lock); | |
493 | ||
494 | return ret; | |
495 | } | |
496 | ||
497 | int dlm_shutting_down(struct dlm_ctxt *dlm) | |
498 | { | |
499 | int ret = 0; | |
500 | ||
501 | spin_lock(&dlm_domain_lock); | |
502 | ||
503 | if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN) | |
504 | ret = 1; | |
505 | ||
506 | spin_unlock(&dlm_domain_lock); | |
507 | ||
508 | return ret; | |
509 | } | |
510 | ||
511 | void dlm_unregister_domain(struct dlm_ctxt *dlm) | |
512 | { | |
513 | int leave = 0; | |
514 | ||
515 | spin_lock(&dlm_domain_lock); | |
516 | BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED); | |
517 | BUG_ON(!dlm->num_joins); | |
518 | ||
519 | dlm->num_joins--; | |
520 | if (!dlm->num_joins) { | |
521 | /* We mark it "in shutdown" now so new register | |
522 | * requests wait until we've completely left the | |
523 | * domain. Don't use DLM_CTXT_LEAVING yet as we still | |
524 | * want new domain joins to communicate with us at | |
525 | * least until we've completed migration of our | |
526 | * resources. */ | |
527 | dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN; | |
528 | leave = 1; | |
529 | } | |
530 | spin_unlock(&dlm_domain_lock); | |
531 | ||
532 | if (leave) { | |
533 | mlog(0, "shutting down domain %s\n", dlm->name); | |
534 | ||
535 | /* We changed dlm state, notify the thread */ | |
536 | dlm_kick_thread(dlm, NULL); | |
537 | ||
538 | dlm_migrate_all_locks(dlm); | |
539 | dlm_mark_domain_leaving(dlm); | |
540 | dlm_leave_domain(dlm); | |
541 | dlm_complete_dlm_shutdown(dlm); | |
542 | } | |
543 | dlm_put(dlm); | |
544 | } | |
545 | EXPORT_SYMBOL_GPL(dlm_unregister_domain); | |
546 | ||
547 | static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data) | |
548 | { | |
549 | struct dlm_query_join_request *query; | |
550 | enum dlm_query_join_response response; | |
551 | struct dlm_ctxt *dlm = NULL; | |
552 | ||
553 | query = (struct dlm_query_join_request *) msg->buf; | |
554 | ||
555 | mlog(0, "node %u wants to join domain %s\n", query->node_idx, | |
556 | query->domain); | |
557 | ||
558 | /* | |
559 | * If heartbeat doesn't consider the node live, tell it | |
560 | * to back off and try again. This gives heartbeat a chance | |
561 | * to catch up. | |
562 | */ | |
563 | if (!o2hb_check_node_heartbeating(query->node_idx)) { | |
564 | mlog(0, "node %u is not in our live map yet\n", | |
565 | query->node_idx); | |
566 | ||
567 | response = JOIN_DISALLOW; | |
568 | goto respond; | |
569 | } | |
570 | ||
571 | response = JOIN_OK_NO_MAP; | |
572 | ||
573 | spin_lock(&dlm_domain_lock); | |
574 | dlm = __dlm_lookup_domain_full(query->domain, query->name_len); | |
575 | /* Once the dlm ctxt is marked as leaving then we don't want | |
e2faea4c KH |
576 | * to be put in someone's domain map. |
577 | * Also, explicitly disallow joining at certain troublesome | |
578 | * times (ie. during recovery). */ | |
6714d8e8 | 579 | if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) { |
e2faea4c | 580 | int bit = query->node_idx; |
6714d8e8 KH |
581 | spin_lock(&dlm->spinlock); |
582 | ||
583 | if (dlm->dlm_state == DLM_CTXT_NEW && | |
584 | dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) { | |
585 | /*If this is a brand new context and we | |
586 | * haven't started our join process yet, then | |
587 | * the other node won the race. */ | |
588 | response = JOIN_OK_NO_MAP; | |
589 | } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) { | |
590 | /* Disallow parallel joins. */ | |
591 | response = JOIN_DISALLOW; | |
e2faea4c KH |
592 | } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) { |
593 | mlog(ML_NOTICE, "node %u trying to join, but recovery " | |
594 | "is ongoing.\n", bit); | |
595 | response = JOIN_DISALLOW; | |
596 | } else if (test_bit(bit, dlm->recovery_map)) { | |
597 | mlog(ML_NOTICE, "node %u trying to join, but it " | |
598 | "still needs recovery.\n", bit); | |
599 | response = JOIN_DISALLOW; | |
600 | } else if (test_bit(bit, dlm->domain_map)) { | |
601 | mlog(ML_NOTICE, "node %u trying to join, but it " | |
602 | "is still in the domain! needs recovery?\n", | |
603 | bit); | |
604 | response = JOIN_DISALLOW; | |
6714d8e8 KH |
605 | } else { |
606 | /* Alright we're fully a part of this domain | |
607 | * so we keep some state as to who's joining | |
608 | * and indicate to him that needs to be fixed | |
609 | * up. */ | |
610 | response = JOIN_OK; | |
611 | __dlm_set_joining_node(dlm, query->node_idx); | |
612 | } | |
613 | ||
614 | spin_unlock(&dlm->spinlock); | |
615 | } | |
616 | spin_unlock(&dlm_domain_lock); | |
617 | ||
618 | respond: | |
619 | mlog(0, "We respond with %u\n", response); | |
620 | ||
621 | return response; | |
622 | } | |
623 | ||
624 | static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data) | |
625 | { | |
626 | struct dlm_assert_joined *assert; | |
627 | struct dlm_ctxt *dlm = NULL; | |
628 | ||
629 | assert = (struct dlm_assert_joined *) msg->buf; | |
630 | ||
631 | mlog(0, "node %u asserts join on domain %s\n", assert->node_idx, | |
632 | assert->domain); | |
633 | ||
634 | spin_lock(&dlm_domain_lock); | |
635 | dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len); | |
636 | /* XXX should we consider no dlm ctxt an error? */ | |
637 | if (dlm) { | |
638 | spin_lock(&dlm->spinlock); | |
639 | ||
640 | /* Alright, this node has officially joined our | |
641 | * domain. Set him in the map and clean up our | |
642 | * leftover join state. */ | |
643 | BUG_ON(dlm->joining_node != assert->node_idx); | |
644 | set_bit(assert->node_idx, dlm->domain_map); | |
645 | __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); | |
646 | ||
647 | __dlm_print_nodes(dlm); | |
648 | ||
649 | /* notify anything attached to the heartbeat events */ | |
650 | dlm_hb_event_notify_attached(dlm, assert->node_idx, 1); | |
651 | ||
652 | spin_unlock(&dlm->spinlock); | |
653 | } | |
654 | spin_unlock(&dlm_domain_lock); | |
655 | ||
656 | return 0; | |
657 | } | |
658 | ||
659 | static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data) | |
660 | { | |
661 | struct dlm_cancel_join *cancel; | |
662 | struct dlm_ctxt *dlm = NULL; | |
663 | ||
664 | cancel = (struct dlm_cancel_join *) msg->buf; | |
665 | ||
666 | mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx, | |
667 | cancel->domain); | |
668 | ||
669 | spin_lock(&dlm_domain_lock); | |
670 | dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len); | |
671 | ||
672 | if (dlm) { | |
673 | spin_lock(&dlm->spinlock); | |
674 | ||
675 | /* Yikes, this guy wants to cancel his join. No | |
676 | * problem, we simply cleanup our join state. */ | |
677 | BUG_ON(dlm->joining_node != cancel->node_idx); | |
678 | __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); | |
679 | ||
680 | spin_unlock(&dlm->spinlock); | |
681 | } | |
682 | spin_unlock(&dlm_domain_lock); | |
683 | ||
684 | return 0; | |
685 | } | |
686 | ||
687 | static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm, | |
688 | unsigned int node) | |
689 | { | |
690 | int status; | |
691 | struct dlm_cancel_join cancel_msg; | |
692 | ||
693 | memset(&cancel_msg, 0, sizeof(cancel_msg)); | |
694 | cancel_msg.node_idx = dlm->node_num; | |
695 | cancel_msg.name_len = strlen(dlm->name); | |
696 | memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len); | |
697 | ||
698 | status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY, | |
699 | &cancel_msg, sizeof(cancel_msg), node, | |
700 | NULL); | |
701 | if (status < 0) { | |
702 | mlog_errno(status); | |
703 | goto bail; | |
704 | } | |
705 | ||
706 | bail: | |
707 | return status; | |
708 | } | |
709 | ||
710 | /* map_size should be in bytes. */ | |
711 | static int dlm_send_join_cancels(struct dlm_ctxt *dlm, | |
712 | unsigned long *node_map, | |
713 | unsigned int map_size) | |
714 | { | |
715 | int status, tmpstat; | |
716 | unsigned int node; | |
717 | ||
718 | if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) * | |
719 | sizeof(unsigned long))) { | |
720 | mlog(ML_ERROR, | |
721 | "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n", | |
722 | map_size, BITS_TO_LONGS(O2NM_MAX_NODES)); | |
723 | return -EINVAL; | |
724 | } | |
725 | ||
726 | status = 0; | |
727 | node = -1; | |
728 | while ((node = find_next_bit(node_map, O2NM_MAX_NODES, | |
729 | node + 1)) < O2NM_MAX_NODES) { | |
730 | if (node == dlm->node_num) | |
731 | continue; | |
732 | ||
733 | tmpstat = dlm_send_one_join_cancel(dlm, node); | |
734 | if (tmpstat) { | |
735 | mlog(ML_ERROR, "Error return %d cancelling join on " | |
736 | "node %d\n", tmpstat, node); | |
737 | if (!status) | |
738 | status = tmpstat; | |
739 | } | |
740 | } | |
741 | ||
742 | if (status) | |
743 | mlog_errno(status); | |
744 | return status; | |
745 | } | |
746 | ||
747 | static int dlm_request_join(struct dlm_ctxt *dlm, | |
748 | int node, | |
749 | enum dlm_query_join_response *response) | |
750 | { | |
751 | int status, retval; | |
752 | struct dlm_query_join_request join_msg; | |
753 | ||
754 | mlog(0, "querying node %d\n", node); | |
755 | ||
756 | memset(&join_msg, 0, sizeof(join_msg)); | |
757 | join_msg.node_idx = dlm->node_num; | |
758 | join_msg.name_len = strlen(dlm->name); | |
759 | memcpy(join_msg.domain, dlm->name, join_msg.name_len); | |
760 | ||
761 | status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg, | |
762 | sizeof(join_msg), node, &retval); | |
763 | if (status < 0 && status != -ENOPROTOOPT) { | |
764 | mlog_errno(status); | |
765 | goto bail; | |
766 | } | |
767 | ||
768 | /* -ENOPROTOOPT from the net code means the other side isn't | |
769 | listening for our message type -- that's fine, it means | |
770 | his dlm isn't up, so we can consider him a 'yes' but not | |
771 | joined into the domain. */ | |
772 | if (status == -ENOPROTOOPT) { | |
773 | status = 0; | |
774 | *response = JOIN_OK_NO_MAP; | |
775 | } else if (retval == JOIN_DISALLOW || | |
776 | retval == JOIN_OK || | |
777 | retval == JOIN_OK_NO_MAP) { | |
778 | *response = retval; | |
779 | } else { | |
780 | status = -EINVAL; | |
781 | mlog(ML_ERROR, "invalid response %d from node %u\n", retval, | |
782 | node); | |
783 | } | |
784 | ||
785 | mlog(0, "status %d, node %d response is %d\n", status, node, | |
786 | *response); | |
787 | ||
788 | bail: | |
789 | return status; | |
790 | } | |
791 | ||
792 | static int dlm_send_one_join_assert(struct dlm_ctxt *dlm, | |
793 | unsigned int node) | |
794 | { | |
795 | int status; | |
796 | struct dlm_assert_joined assert_msg; | |
797 | ||
798 | mlog(0, "Sending join assert to node %u\n", node); | |
799 | ||
800 | memset(&assert_msg, 0, sizeof(assert_msg)); | |
801 | assert_msg.node_idx = dlm->node_num; | |
802 | assert_msg.name_len = strlen(dlm->name); | |
803 | memcpy(assert_msg.domain, dlm->name, assert_msg.name_len); | |
804 | ||
805 | status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY, | |
806 | &assert_msg, sizeof(assert_msg), node, | |
807 | NULL); | |
808 | if (status < 0) | |
809 | mlog_errno(status); | |
810 | ||
811 | return status; | |
812 | } | |
813 | ||
814 | static void dlm_send_join_asserts(struct dlm_ctxt *dlm, | |
815 | unsigned long *node_map) | |
816 | { | |
817 | int status, node, live; | |
818 | ||
819 | status = 0; | |
820 | node = -1; | |
821 | while ((node = find_next_bit(node_map, O2NM_MAX_NODES, | |
822 | node + 1)) < O2NM_MAX_NODES) { | |
823 | if (node == dlm->node_num) | |
824 | continue; | |
825 | ||
826 | do { | |
827 | /* It is very important that this message be | |
828 | * received so we spin until either the node | |
829 | * has died or it gets the message. */ | |
830 | status = dlm_send_one_join_assert(dlm, node); | |
831 | ||
832 | spin_lock(&dlm->spinlock); | |
833 | live = test_bit(node, dlm->live_nodes_map); | |
834 | spin_unlock(&dlm->spinlock); | |
835 | ||
836 | if (status) { | |
837 | mlog(ML_ERROR, "Error return %d asserting " | |
838 | "join on node %d\n", status, node); | |
839 | ||
840 | /* give us some time between errors... */ | |
841 | if (live) | |
842 | msleep(DLM_DOMAIN_BACKOFF_MS); | |
843 | } | |
844 | } while (status && live); | |
845 | } | |
846 | } | |
847 | ||
848 | struct domain_join_ctxt { | |
849 | unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; | |
850 | unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; | |
851 | }; | |
852 | ||
853 | static int dlm_should_restart_join(struct dlm_ctxt *dlm, | |
854 | struct domain_join_ctxt *ctxt, | |
855 | enum dlm_query_join_response response) | |
856 | { | |
857 | int ret; | |
858 | ||
859 | if (response == JOIN_DISALLOW) { | |
860 | mlog(0, "Latest response of disallow -- should restart\n"); | |
861 | return 1; | |
862 | } | |
863 | ||
864 | spin_lock(&dlm->spinlock); | |
865 | /* For now, we restart the process if the node maps have | |
866 | * changed at all */ | |
867 | ret = memcmp(ctxt->live_map, dlm->live_nodes_map, | |
868 | sizeof(dlm->live_nodes_map)); | |
869 | spin_unlock(&dlm->spinlock); | |
870 | ||
871 | if (ret) | |
872 | mlog(0, "Node maps changed -- should restart\n"); | |
873 | ||
874 | return ret; | |
875 | } | |
876 | ||
877 | static int dlm_try_to_join_domain(struct dlm_ctxt *dlm) | |
878 | { | |
879 | int status = 0, tmpstat, node; | |
880 | struct domain_join_ctxt *ctxt; | |
881 | enum dlm_query_join_response response; | |
882 | ||
883 | mlog_entry("%p", dlm); | |
884 | ||
885 | ctxt = kcalloc(1, sizeof(*ctxt), GFP_KERNEL); | |
886 | if (!ctxt) { | |
887 | status = -ENOMEM; | |
888 | mlog_errno(status); | |
889 | goto bail; | |
890 | } | |
891 | ||
892 | /* group sem locking should work for us here -- we're already | |
893 | * registered for heartbeat events so filling this should be | |
894 | * atomic wrt getting those handlers called. */ | |
895 | o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map)); | |
896 | ||
897 | spin_lock(&dlm->spinlock); | |
898 | memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map)); | |
899 | ||
900 | __dlm_set_joining_node(dlm, dlm->node_num); | |
901 | ||
902 | spin_unlock(&dlm->spinlock); | |
903 | ||
904 | node = -1; | |
905 | while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES, | |
906 | node + 1)) < O2NM_MAX_NODES) { | |
907 | if (node == dlm->node_num) | |
908 | continue; | |
909 | ||
910 | status = dlm_request_join(dlm, node, &response); | |
911 | if (status < 0) { | |
912 | mlog_errno(status); | |
913 | goto bail; | |
914 | } | |
915 | ||
916 | /* Ok, either we got a response or the node doesn't have a | |
917 | * dlm up. */ | |
918 | if (response == JOIN_OK) | |
919 | set_bit(node, ctxt->yes_resp_map); | |
920 | ||
921 | if (dlm_should_restart_join(dlm, ctxt, response)) { | |
922 | status = -EAGAIN; | |
923 | goto bail; | |
924 | } | |
925 | } | |
926 | ||
927 | mlog(0, "Yay, done querying nodes!\n"); | |
928 | ||
929 | /* Yay, everyone agree's we can join the domain. My domain is | |
930 | * comprised of all nodes who were put in the | |
931 | * yes_resp_map. Copy that into our domain map and send a join | |
932 | * assert message to clean up everyone elses state. */ | |
933 | spin_lock(&dlm->spinlock); | |
934 | memcpy(dlm->domain_map, ctxt->yes_resp_map, | |
935 | sizeof(ctxt->yes_resp_map)); | |
936 | set_bit(dlm->node_num, dlm->domain_map); | |
937 | spin_unlock(&dlm->spinlock); | |
938 | ||
939 | dlm_send_join_asserts(dlm, ctxt->yes_resp_map); | |
940 | ||
941 | /* Joined state *must* be set before the joining node | |
942 | * information, otherwise the query_join handler may read no | |
943 | * current joiner but a state of NEW and tell joining nodes | |
944 | * we're not in the domain. */ | |
945 | spin_lock(&dlm_domain_lock); | |
946 | dlm->dlm_state = DLM_CTXT_JOINED; | |
947 | dlm->num_joins++; | |
948 | spin_unlock(&dlm_domain_lock); | |
949 | ||
950 | bail: | |
951 | spin_lock(&dlm->spinlock); | |
952 | __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); | |
953 | if (!status) | |
954 | __dlm_print_nodes(dlm); | |
955 | spin_unlock(&dlm->spinlock); | |
956 | ||
957 | if (ctxt) { | |
958 | /* Do we need to send a cancel message to any nodes? */ | |
959 | if (status < 0) { | |
960 | tmpstat = dlm_send_join_cancels(dlm, | |
961 | ctxt->yes_resp_map, | |
962 | sizeof(ctxt->yes_resp_map)); | |
963 | if (tmpstat < 0) | |
964 | mlog_errno(tmpstat); | |
965 | } | |
966 | kfree(ctxt); | |
967 | } | |
968 | ||
969 | mlog(0, "returning %d\n", status); | |
970 | return status; | |
971 | } | |
972 | ||
973 | static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm) | |
974 | { | |
975 | o2hb_unregister_callback(&dlm->dlm_hb_up); | |
976 | o2hb_unregister_callback(&dlm->dlm_hb_down); | |
977 | o2net_unregister_handler_list(&dlm->dlm_domain_handlers); | |
978 | } | |
979 | ||
980 | static int dlm_register_domain_handlers(struct dlm_ctxt *dlm) | |
981 | { | |
982 | int status; | |
983 | ||
984 | mlog(0, "registering handlers.\n"); | |
985 | ||
986 | o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB, | |
987 | dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI); | |
988 | status = o2hb_register_callback(&dlm->dlm_hb_down); | |
989 | if (status) | |
990 | goto bail; | |
991 | ||
992 | o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB, | |
993 | dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI); | |
994 | status = o2hb_register_callback(&dlm->dlm_hb_up); | |
995 | if (status) | |
996 | goto bail; | |
997 | ||
998 | status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key, | |
999 | sizeof(struct dlm_master_request), | |
1000 | dlm_master_request_handler, | |
1001 | dlm, &dlm->dlm_domain_handlers); | |
1002 | if (status) | |
1003 | goto bail; | |
1004 | ||
1005 | status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key, | |
1006 | sizeof(struct dlm_assert_master), | |
1007 | dlm_assert_master_handler, | |
1008 | dlm, &dlm->dlm_domain_handlers); | |
1009 | if (status) | |
1010 | goto bail; | |
1011 | ||
1012 | status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key, | |
1013 | sizeof(struct dlm_create_lock), | |
1014 | dlm_create_lock_handler, | |
1015 | dlm, &dlm->dlm_domain_handlers); | |
1016 | if (status) | |
1017 | goto bail; | |
1018 | ||
1019 | status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key, | |
1020 | DLM_CONVERT_LOCK_MAX_LEN, | |
1021 | dlm_convert_lock_handler, | |
1022 | dlm, &dlm->dlm_domain_handlers); | |
1023 | if (status) | |
1024 | goto bail; | |
1025 | ||
1026 | status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key, | |
1027 | DLM_UNLOCK_LOCK_MAX_LEN, | |
1028 | dlm_unlock_lock_handler, | |
1029 | dlm, &dlm->dlm_domain_handlers); | |
1030 | if (status) | |
1031 | goto bail; | |
1032 | ||
1033 | status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key, | |
1034 | DLM_PROXY_AST_MAX_LEN, | |
1035 | dlm_proxy_ast_handler, | |
1036 | dlm, &dlm->dlm_domain_handlers); | |
1037 | if (status) | |
1038 | goto bail; | |
1039 | ||
1040 | status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key, | |
1041 | sizeof(struct dlm_exit_domain), | |
1042 | dlm_exit_domain_handler, | |
1043 | dlm, &dlm->dlm_domain_handlers); | |
1044 | if (status) | |
1045 | goto bail; | |
1046 | ||
1047 | status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key, | |
1048 | sizeof(struct dlm_migrate_request), | |
1049 | dlm_migrate_request_handler, | |
1050 | dlm, &dlm->dlm_domain_handlers); | |
1051 | if (status) | |
1052 | goto bail; | |
1053 | ||
1054 | status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key, | |
1055 | DLM_MIG_LOCKRES_MAX_LEN, | |
1056 | dlm_mig_lockres_handler, | |
1057 | dlm, &dlm->dlm_domain_handlers); | |
1058 | if (status) | |
1059 | goto bail; | |
1060 | ||
1061 | status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key, | |
1062 | sizeof(struct dlm_master_requery), | |
1063 | dlm_master_requery_handler, | |
1064 | dlm, &dlm->dlm_domain_handlers); | |
1065 | if (status) | |
1066 | goto bail; | |
1067 | ||
1068 | status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key, | |
1069 | sizeof(struct dlm_lock_request), | |
1070 | dlm_request_all_locks_handler, | |
1071 | dlm, &dlm->dlm_domain_handlers); | |
1072 | if (status) | |
1073 | goto bail; | |
1074 | ||
1075 | status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key, | |
1076 | sizeof(struct dlm_reco_data_done), | |
1077 | dlm_reco_data_done_handler, | |
1078 | dlm, &dlm->dlm_domain_handlers); | |
1079 | if (status) | |
1080 | goto bail; | |
1081 | ||
1082 | status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key, | |
1083 | sizeof(struct dlm_begin_reco), | |
1084 | dlm_begin_reco_handler, | |
1085 | dlm, &dlm->dlm_domain_handlers); | |
1086 | if (status) | |
1087 | goto bail; | |
1088 | ||
1089 | status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key, | |
1090 | sizeof(struct dlm_finalize_reco), | |
1091 | dlm_finalize_reco_handler, | |
1092 | dlm, &dlm->dlm_domain_handlers); | |
1093 | if (status) | |
1094 | goto bail; | |
1095 | ||
1096 | bail: | |
1097 | if (status) | |
1098 | dlm_unregister_domain_handlers(dlm); | |
1099 | ||
1100 | return status; | |
1101 | } | |
1102 | ||
1103 | static int dlm_join_domain(struct dlm_ctxt *dlm) | |
1104 | { | |
1105 | int status; | |
1106 | ||
1107 | BUG_ON(!dlm); | |
1108 | ||
1109 | mlog(0, "Join domain %s\n", dlm->name); | |
1110 | ||
1111 | status = dlm_register_domain_handlers(dlm); | |
1112 | if (status) { | |
1113 | mlog_errno(status); | |
1114 | goto bail; | |
1115 | } | |
1116 | ||
1117 | status = dlm_launch_thread(dlm); | |
1118 | if (status < 0) { | |
1119 | mlog_errno(status); | |
1120 | goto bail; | |
1121 | } | |
1122 | ||
1123 | status = dlm_launch_recovery_thread(dlm); | |
1124 | if (status < 0) { | |
1125 | mlog_errno(status); | |
1126 | goto bail; | |
1127 | } | |
1128 | ||
1129 | do { | |
1130 | unsigned int backoff; | |
1131 | status = dlm_try_to_join_domain(dlm); | |
1132 | ||
1133 | /* If we're racing another node to the join, then we | |
1134 | * need to back off temporarily and let them | |
1135 | * complete. */ | |
1136 | if (status == -EAGAIN) { | |
1137 | if (signal_pending(current)) { | |
1138 | status = -ERESTARTSYS; | |
1139 | goto bail; | |
1140 | } | |
1141 | ||
1142 | /* | |
1143 | * <chip> After you! | |
1144 | * <dale> No, after you! | |
1145 | * <chip> I insist! | |
1146 | * <dale> But you first! | |
1147 | * ... | |
1148 | */ | |
1149 | backoff = (unsigned int)(jiffies & 0x3); | |
1150 | backoff *= DLM_DOMAIN_BACKOFF_MS; | |
1151 | mlog(0, "backoff %d\n", backoff); | |
1152 | msleep(backoff); | |
1153 | } | |
1154 | } while (status == -EAGAIN); | |
1155 | ||
1156 | if (status < 0) { | |
1157 | mlog_errno(status); | |
1158 | goto bail; | |
1159 | } | |
1160 | ||
1161 | status = 0; | |
1162 | bail: | |
1163 | wake_up(&dlm_domain_events); | |
1164 | ||
1165 | if (status) { | |
1166 | dlm_unregister_domain_handlers(dlm); | |
1167 | dlm_complete_thread(dlm); | |
1168 | dlm_complete_recovery_thread(dlm); | |
1169 | } | |
1170 | ||
1171 | return status; | |
1172 | } | |
1173 | ||
1174 | static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain, | |
1175 | u32 key) | |
1176 | { | |
1177 | int i; | |
1178 | struct dlm_ctxt *dlm = NULL; | |
1179 | ||
1180 | dlm = kcalloc(1, sizeof(*dlm), GFP_KERNEL); | |
1181 | if (!dlm) { | |
1182 | mlog_errno(-ENOMEM); | |
1183 | goto leave; | |
1184 | } | |
1185 | ||
1186 | dlm->name = kmalloc(strlen(domain) + 1, GFP_KERNEL); | |
1187 | if (dlm->name == NULL) { | |
1188 | mlog_errno(-ENOMEM); | |
1189 | kfree(dlm); | |
1190 | dlm = NULL; | |
1191 | goto leave; | |
1192 | } | |
1193 | ||
81f2094a MF |
1194 | dlm->lockres_hash = (struct hlist_head *) __get_free_page(GFP_KERNEL); |
1195 | if (!dlm->lockres_hash) { | |
6714d8e8 KH |
1196 | mlog_errno(-ENOMEM); |
1197 | kfree(dlm->name); | |
1198 | kfree(dlm); | |
1199 | dlm = NULL; | |
1200 | goto leave; | |
1201 | } | |
6714d8e8 | 1202 | |
81f2094a MF |
1203 | for (i=0; i<DLM_HASH_BUCKETS; i++) |
1204 | INIT_HLIST_HEAD(&dlm->lockres_hash[i]); | |
6714d8e8 KH |
1205 | |
1206 | strcpy(dlm->name, domain); | |
1207 | dlm->key = key; | |
1208 | dlm->node_num = o2nm_this_node(); | |
1209 | ||
1210 | spin_lock_init(&dlm->spinlock); | |
1211 | spin_lock_init(&dlm->master_lock); | |
1212 | spin_lock_init(&dlm->ast_lock); | |
1213 | INIT_LIST_HEAD(&dlm->list); | |
1214 | INIT_LIST_HEAD(&dlm->dirty_list); | |
1215 | INIT_LIST_HEAD(&dlm->reco.resources); | |
1216 | INIT_LIST_HEAD(&dlm->reco.received); | |
1217 | INIT_LIST_HEAD(&dlm->reco.node_data); | |
1218 | INIT_LIST_HEAD(&dlm->purge_list); | |
1219 | INIT_LIST_HEAD(&dlm->dlm_domain_handlers); | |
1220 | dlm->reco.state = 0; | |
1221 | ||
1222 | INIT_LIST_HEAD(&dlm->pending_asts); | |
1223 | INIT_LIST_HEAD(&dlm->pending_basts); | |
1224 | ||
1225 | mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n", | |
1226 | dlm->recovery_map, &(dlm->recovery_map[0])); | |
1227 | ||
1228 | memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map)); | |
1229 | memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map)); | |
1230 | memset(dlm->domain_map, 0, sizeof(dlm->domain_map)); | |
1231 | ||
1232 | dlm->dlm_thread_task = NULL; | |
1233 | dlm->dlm_reco_thread_task = NULL; | |
1234 | init_waitqueue_head(&dlm->dlm_thread_wq); | |
1235 | init_waitqueue_head(&dlm->dlm_reco_thread_wq); | |
1236 | init_waitqueue_head(&dlm->reco.event); | |
1237 | init_waitqueue_head(&dlm->ast_wq); | |
1238 | init_waitqueue_head(&dlm->migration_wq); | |
1239 | INIT_LIST_HEAD(&dlm->master_list); | |
1240 | INIT_LIST_HEAD(&dlm->mle_hb_events); | |
1241 | ||
1242 | dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN; | |
1243 | init_waitqueue_head(&dlm->dlm_join_events); | |
1244 | ||
1245 | dlm->reco.new_master = O2NM_INVALID_NODE_NUM; | |
1246 | dlm->reco.dead_node = O2NM_INVALID_NODE_NUM; | |
1247 | atomic_set(&dlm->local_resources, 0); | |
1248 | atomic_set(&dlm->remote_resources, 0); | |
1249 | atomic_set(&dlm->unknown_resources, 0); | |
1250 | ||
1251 | spin_lock_init(&dlm->work_lock); | |
1252 | INIT_LIST_HEAD(&dlm->work_list); | |
1253 | INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work, dlm); | |
1254 | ||
1255 | kref_init(&dlm->dlm_refs); | |
1256 | dlm->dlm_state = DLM_CTXT_NEW; | |
1257 | ||
1258 | INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks); | |
1259 | ||
1260 | mlog(0, "context init: refcount %u\n", | |
1261 | atomic_read(&dlm->dlm_refs.refcount)); | |
1262 | ||
1263 | leave: | |
1264 | return dlm; | |
1265 | } | |
1266 | ||
1267 | /* | |
1268 | * dlm_register_domain: one-time setup per "domain" | |
1269 | */ | |
1270 | struct dlm_ctxt * dlm_register_domain(const char *domain, | |
1271 | u32 key) | |
1272 | { | |
1273 | int ret; | |
1274 | struct dlm_ctxt *dlm = NULL; | |
1275 | struct dlm_ctxt *new_ctxt = NULL; | |
1276 | ||
1277 | if (strlen(domain) > O2NM_MAX_NAME_LEN) { | |
1278 | ret = -ENAMETOOLONG; | |
1279 | mlog(ML_ERROR, "domain name length too long\n"); | |
1280 | goto leave; | |
1281 | } | |
1282 | ||
1283 | if (!o2hb_check_local_node_heartbeating()) { | |
1284 | mlog(ML_ERROR, "the local node has not been configured, or is " | |
1285 | "not heartbeating\n"); | |
1286 | ret = -EPROTO; | |
1287 | goto leave; | |
1288 | } | |
1289 | ||
1290 | mlog(0, "register called for domain \"%s\"\n", domain); | |
1291 | ||
1292 | retry: | |
1293 | dlm = NULL; | |
1294 | if (signal_pending(current)) { | |
1295 | ret = -ERESTARTSYS; | |
1296 | mlog_errno(ret); | |
1297 | goto leave; | |
1298 | } | |
1299 | ||
1300 | spin_lock(&dlm_domain_lock); | |
1301 | ||
1302 | dlm = __dlm_lookup_domain(domain); | |
1303 | if (dlm) { | |
1304 | if (dlm->dlm_state != DLM_CTXT_JOINED) { | |
1305 | spin_unlock(&dlm_domain_lock); | |
1306 | ||
1307 | mlog(0, "This ctxt is not joined yet!\n"); | |
1308 | wait_event_interruptible(dlm_domain_events, | |
1309 | dlm_wait_on_domain_helper( | |
1310 | domain)); | |
1311 | goto retry; | |
1312 | } | |
1313 | ||
1314 | __dlm_get(dlm); | |
1315 | dlm->num_joins++; | |
1316 | ||
1317 | spin_unlock(&dlm_domain_lock); | |
1318 | ||
1319 | ret = 0; | |
1320 | goto leave; | |
1321 | } | |
1322 | ||
1323 | /* doesn't exist */ | |
1324 | if (!new_ctxt) { | |
1325 | spin_unlock(&dlm_domain_lock); | |
1326 | ||
1327 | new_ctxt = dlm_alloc_ctxt(domain, key); | |
1328 | if (new_ctxt) | |
1329 | goto retry; | |
1330 | ||
1331 | ret = -ENOMEM; | |
1332 | mlog_errno(ret); | |
1333 | goto leave; | |
1334 | } | |
1335 | ||
1336 | /* a little variable switch-a-roo here... */ | |
1337 | dlm = new_ctxt; | |
1338 | new_ctxt = NULL; | |
1339 | ||
1340 | /* add the new domain */ | |
1341 | list_add_tail(&dlm->list, &dlm_domains); | |
1342 | spin_unlock(&dlm_domain_lock); | |
1343 | ||
1344 | ret = dlm_join_domain(dlm); | |
1345 | if (ret) { | |
1346 | mlog_errno(ret); | |
1347 | dlm_put(dlm); | |
1348 | goto leave; | |
1349 | } | |
1350 | ||
1351 | ret = 0; | |
1352 | leave: | |
1353 | if (new_ctxt) | |
1354 | dlm_free_ctxt_mem(new_ctxt); | |
1355 | ||
1356 | if (ret < 0) | |
1357 | dlm = ERR_PTR(ret); | |
1358 | ||
1359 | return dlm; | |
1360 | } | |
1361 | EXPORT_SYMBOL_GPL(dlm_register_domain); | |
1362 | ||
1363 | static LIST_HEAD(dlm_join_handlers); | |
1364 | ||
1365 | static void dlm_unregister_net_handlers(void) | |
1366 | { | |
1367 | o2net_unregister_handler_list(&dlm_join_handlers); | |
1368 | } | |
1369 | ||
1370 | static int dlm_register_net_handlers(void) | |
1371 | { | |
1372 | int status = 0; | |
1373 | ||
1374 | status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, | |
1375 | sizeof(struct dlm_query_join_request), | |
1376 | dlm_query_join_handler, | |
1377 | NULL, &dlm_join_handlers); | |
1378 | if (status) | |
1379 | goto bail; | |
1380 | ||
1381 | status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY, | |
1382 | sizeof(struct dlm_assert_joined), | |
1383 | dlm_assert_joined_handler, | |
1384 | NULL, &dlm_join_handlers); | |
1385 | if (status) | |
1386 | goto bail; | |
1387 | ||
1388 | status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY, | |
1389 | sizeof(struct dlm_cancel_join), | |
1390 | dlm_cancel_join_handler, | |
1391 | NULL, &dlm_join_handlers); | |
1392 | ||
1393 | bail: | |
1394 | if (status < 0) | |
1395 | dlm_unregister_net_handlers(); | |
1396 | ||
1397 | return status; | |
1398 | } | |
1399 | ||
1400 | /* Domain eviction callback handling. | |
1401 | * | |
1402 | * The file system requires notification of node death *before* the | |
1403 | * dlm completes it's recovery work, otherwise it may be able to | |
1404 | * acquire locks on resources requiring recovery. Since the dlm can | |
1405 | * evict a node from it's domain *before* heartbeat fires, a similar | |
1406 | * mechanism is required. */ | |
1407 | ||
1408 | /* Eviction is not expected to happen often, so a per-domain lock is | |
1409 | * not necessary. Eviction callbacks are allowed to sleep for short | |
1410 | * periods of time. */ | |
1411 | static DECLARE_RWSEM(dlm_callback_sem); | |
1412 | ||
1413 | void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm, | |
1414 | int node_num) | |
1415 | { | |
1416 | struct list_head *iter; | |
1417 | struct dlm_eviction_cb *cb; | |
1418 | ||
1419 | down_read(&dlm_callback_sem); | |
1420 | list_for_each(iter, &dlm->dlm_eviction_callbacks) { | |
1421 | cb = list_entry(iter, struct dlm_eviction_cb, ec_item); | |
1422 | ||
1423 | cb->ec_func(node_num, cb->ec_data); | |
1424 | } | |
1425 | up_read(&dlm_callback_sem); | |
1426 | } | |
1427 | ||
1428 | void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb, | |
1429 | dlm_eviction_func *f, | |
1430 | void *data) | |
1431 | { | |
1432 | INIT_LIST_HEAD(&cb->ec_item); | |
1433 | cb->ec_func = f; | |
1434 | cb->ec_data = data; | |
1435 | } | |
1436 | EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb); | |
1437 | ||
1438 | void dlm_register_eviction_cb(struct dlm_ctxt *dlm, | |
1439 | struct dlm_eviction_cb *cb) | |
1440 | { | |
1441 | down_write(&dlm_callback_sem); | |
1442 | list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks); | |
1443 | up_write(&dlm_callback_sem); | |
1444 | } | |
1445 | EXPORT_SYMBOL_GPL(dlm_register_eviction_cb); | |
1446 | ||
1447 | void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb) | |
1448 | { | |
1449 | down_write(&dlm_callback_sem); | |
1450 | list_del_init(&cb->ec_item); | |
1451 | up_write(&dlm_callback_sem); | |
1452 | } | |
1453 | EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb); | |
1454 | ||
1455 | static int __init dlm_init(void) | |
1456 | { | |
1457 | int status; | |
1458 | ||
1459 | dlm_print_version(); | |
1460 | ||
1461 | status = dlm_init_mle_cache(); | |
1462 | if (status) | |
1463 | return -1; | |
1464 | ||
1465 | status = dlm_register_net_handlers(); | |
1466 | if (status) { | |
1467 | dlm_destroy_mle_cache(); | |
1468 | return -1; | |
1469 | } | |
1470 | ||
1471 | return 0; | |
1472 | } | |
1473 | ||
1474 | static void __exit dlm_exit (void) | |
1475 | { | |
1476 | dlm_unregister_net_handlers(); | |
1477 | dlm_destroy_mle_cache(); | |
1478 | } | |
1479 | ||
1480 | MODULE_AUTHOR("Oracle"); | |
1481 | MODULE_LICENSE("GPL"); | |
1482 | ||
1483 | module_init(dlm_init); | |
1484 | module_exit(dlm_exit); |