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 | ||
6714d8e8 KH |
44 | #include "dlmdomain.h" |
45 | ||
46 | #include "dlmver.h" | |
47 | ||
48 | #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN) | |
49 | #include "cluster/masklog.h" | |
50 | ||
03d864c0 DP |
51 | static void dlm_free_pagevec(void **vec, int pages) |
52 | { | |
53 | while (pages--) | |
54 | free_page((unsigned long)vec[pages]); | |
55 | kfree(vec); | |
56 | } | |
57 | ||
58 | static void **dlm_alloc_pagevec(int pages) | |
59 | { | |
60 | void **vec = kmalloc(pages * sizeof(void *), GFP_KERNEL); | |
61 | int i; | |
62 | ||
63 | if (!vec) | |
64 | return NULL; | |
65 | ||
66 | for (i = 0; i < pages; i++) | |
67 | if (!(vec[i] = (void *)__get_free_page(GFP_KERNEL))) | |
68 | goto out_free; | |
c8f33b6e | 69 | |
685f1adb MF |
70 | mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n", |
71 | pages, DLM_HASH_PAGES, (unsigned long)DLM_BUCKETS_PER_PAGE); | |
03d864c0 DP |
72 | return vec; |
73 | out_free: | |
74 | dlm_free_pagevec(vec, i); | |
75 | return NULL; | |
76 | } | |
77 | ||
6714d8e8 KH |
78 | /* |
79 | * | |
80 | * spinlock lock ordering: if multiple locks are needed, obey this ordering: | |
81 | * dlm_domain_lock | |
82 | * struct dlm_ctxt->spinlock | |
83 | * struct dlm_lock_resource->spinlock | |
84 | * struct dlm_ctxt->master_lock | |
85 | * struct dlm_ctxt->ast_lock | |
86 | * dlm_master_list_entry->spinlock | |
87 | * dlm_lock->spinlock | |
88 | * | |
89 | */ | |
90 | ||
34af946a | 91 | DEFINE_SPINLOCK(dlm_domain_lock); |
6714d8e8 KH |
92 | LIST_HEAD(dlm_domains); |
93 | static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events); | |
94 | ||
95 | #define DLM_DOMAIN_BACKOFF_MS 200 | |
96 | ||
97 | static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data); | |
98 | static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data); | |
99 | static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data); | |
100 | static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data); | |
101 | ||
102 | static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm); | |
103 | ||
104 | void __dlm_unhash_lockres(struct dlm_lock_resource *lockres) | |
105 | { | |
81f2094a | 106 | hlist_del_init(&lockres->hash_node); |
6714d8e8 KH |
107 | dlm_lockres_put(lockres); |
108 | } | |
109 | ||
110 | void __dlm_insert_lockres(struct dlm_ctxt *dlm, | |
111 | struct dlm_lock_resource *res) | |
112 | { | |
81f2094a | 113 | struct hlist_head *bucket; |
6714d8e8 KH |
114 | struct qstr *q; |
115 | ||
116 | assert_spin_locked(&dlm->spinlock); | |
117 | ||
118 | q = &res->lockname; | |
03d864c0 | 119 | bucket = dlm_lockres_hash(dlm, q->hash); |
6714d8e8 KH |
120 | |
121 | /* get a reference for our hashtable */ | |
122 | dlm_lockres_get(res); | |
123 | ||
81f2094a | 124 | hlist_add_head(&res->hash_node, bucket); |
6714d8e8 KH |
125 | } |
126 | ||
127 | struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm, | |
a3d33291 MF |
128 | const char *name, |
129 | unsigned int len, | |
130 | unsigned int hash) | |
6714d8e8 | 131 | { |
81f2094a | 132 | struct hlist_head *bucket; |
4198985f | 133 | struct hlist_node *list; |
6714d8e8 KH |
134 | |
135 | mlog_entry("%.*s\n", len, name); | |
136 | ||
137 | assert_spin_locked(&dlm->spinlock); | |
138 | ||
03d864c0 DP |
139 | bucket = dlm_lockres_hash(dlm, hash); |
140 | ||
4198985f DP |
141 | hlist_for_each(list, bucket) { |
142 | struct dlm_lock_resource *res = hlist_entry(list, | |
143 | struct dlm_lock_resource, hash_node); | |
144 | if (res->lockname.name[0] != name[0]) | |
145 | continue; | |
146 | if (unlikely(res->lockname.len != len)) | |
147 | continue; | |
148 | if (memcmp(res->lockname.name + 1, name + 1, len - 1)) | |
149 | continue; | |
150 | dlm_lockres_get(res); | |
151 | return res; | |
6714d8e8 | 152 | } |
4198985f | 153 | return NULL; |
6714d8e8 KH |
154 | } |
155 | ||
156 | struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm, | |
157 | const char *name, | |
158 | unsigned int len) | |
159 | { | |
160 | struct dlm_lock_resource *res; | |
a3d33291 | 161 | unsigned int hash = dlm_lockid_hash(name, len); |
6714d8e8 KH |
162 | |
163 | spin_lock(&dlm->spinlock); | |
a3d33291 | 164 | res = __dlm_lookup_lockres(dlm, name, len, hash); |
6714d8e8 KH |
165 | spin_unlock(&dlm->spinlock); |
166 | return res; | |
167 | } | |
168 | ||
169 | static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len) | |
170 | { | |
171 | struct dlm_ctxt *tmp = NULL; | |
172 | struct list_head *iter; | |
173 | ||
174 | assert_spin_locked(&dlm_domain_lock); | |
175 | ||
176 | /* tmp->name here is always NULL terminated, | |
177 | * but domain may not be! */ | |
178 | list_for_each(iter, &dlm_domains) { | |
179 | tmp = list_entry (iter, struct dlm_ctxt, list); | |
180 | if (strlen(tmp->name) == len && | |
181 | memcmp(tmp->name, domain, len)==0) | |
182 | break; | |
183 | tmp = NULL; | |
184 | } | |
185 | ||
186 | return tmp; | |
187 | } | |
188 | ||
189 | /* For null terminated domain strings ONLY */ | |
190 | static struct dlm_ctxt * __dlm_lookup_domain(const char *domain) | |
191 | { | |
192 | assert_spin_locked(&dlm_domain_lock); | |
193 | ||
194 | return __dlm_lookup_domain_full(domain, strlen(domain)); | |
195 | } | |
196 | ||
197 | ||
198 | /* returns true on one of two conditions: | |
199 | * 1) the domain does not exist | |
200 | * 2) the domain exists and it's state is "joined" */ | |
201 | static int dlm_wait_on_domain_helper(const char *domain) | |
202 | { | |
203 | int ret = 0; | |
204 | struct dlm_ctxt *tmp = NULL; | |
205 | ||
206 | spin_lock(&dlm_domain_lock); | |
207 | ||
208 | tmp = __dlm_lookup_domain(domain); | |
209 | if (!tmp) | |
210 | ret = 1; | |
211 | else if (tmp->dlm_state == DLM_CTXT_JOINED) | |
212 | ret = 1; | |
213 | ||
214 | spin_unlock(&dlm_domain_lock); | |
215 | return ret; | |
216 | } | |
217 | ||
218 | static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm) | |
219 | { | |
81f2094a | 220 | if (dlm->lockres_hash) |
03d864c0 | 221 | dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES); |
6714d8e8 KH |
222 | |
223 | if (dlm->name) | |
224 | kfree(dlm->name); | |
225 | ||
226 | kfree(dlm); | |
227 | } | |
228 | ||
229 | /* A little strange - this function will be called while holding | |
230 | * dlm_domain_lock and is expected to be holding it on the way out. We | |
231 | * will however drop and reacquire it multiple times */ | |
232 | static void dlm_ctxt_release(struct kref *kref) | |
233 | { | |
234 | struct dlm_ctxt *dlm; | |
235 | ||
236 | dlm = container_of(kref, struct dlm_ctxt, dlm_refs); | |
237 | ||
238 | BUG_ON(dlm->num_joins); | |
239 | BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED); | |
240 | ||
241 | /* we may still be in the list if we hit an error during join. */ | |
242 | list_del_init(&dlm->list); | |
243 | ||
244 | spin_unlock(&dlm_domain_lock); | |
245 | ||
246 | mlog(0, "freeing memory from domain %s\n", dlm->name); | |
247 | ||
248 | wake_up(&dlm_domain_events); | |
249 | ||
250 | dlm_free_ctxt_mem(dlm); | |
251 | ||
252 | spin_lock(&dlm_domain_lock); | |
253 | } | |
254 | ||
255 | void dlm_put(struct dlm_ctxt *dlm) | |
256 | { | |
257 | spin_lock(&dlm_domain_lock); | |
258 | kref_put(&dlm->dlm_refs, dlm_ctxt_release); | |
259 | spin_unlock(&dlm_domain_lock); | |
260 | } | |
261 | ||
262 | static void __dlm_get(struct dlm_ctxt *dlm) | |
263 | { | |
264 | kref_get(&dlm->dlm_refs); | |
265 | } | |
266 | ||
267 | /* given a questionable reference to a dlm object, gets a reference if | |
268 | * it can find it in the list, otherwise returns NULL in which case | |
269 | * you shouldn't trust your pointer. */ | |
270 | struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm) | |
271 | { | |
272 | struct list_head *iter; | |
273 | struct dlm_ctxt *target = NULL; | |
274 | ||
275 | spin_lock(&dlm_domain_lock); | |
276 | ||
277 | list_for_each(iter, &dlm_domains) { | |
278 | target = list_entry (iter, struct dlm_ctxt, list); | |
279 | ||
280 | if (target == dlm) { | |
281 | __dlm_get(target); | |
282 | break; | |
283 | } | |
284 | ||
285 | target = NULL; | |
286 | } | |
287 | ||
288 | spin_unlock(&dlm_domain_lock); | |
289 | ||
290 | return target; | |
291 | } | |
292 | ||
293 | int dlm_domain_fully_joined(struct dlm_ctxt *dlm) | |
294 | { | |
295 | int ret; | |
296 | ||
297 | spin_lock(&dlm_domain_lock); | |
298 | ret = (dlm->dlm_state == DLM_CTXT_JOINED) || | |
299 | (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN); | |
300 | spin_unlock(&dlm_domain_lock); | |
301 | ||
302 | return ret; | |
303 | } | |
304 | ||
3156d267 KH |
305 | static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm) |
306 | { | |
307 | if (dlm->dlm_worker) { | |
308 | flush_workqueue(dlm->dlm_worker); | |
309 | destroy_workqueue(dlm->dlm_worker); | |
310 | dlm->dlm_worker = NULL; | |
311 | } | |
312 | } | |
313 | ||
6714d8e8 KH |
314 | static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm) |
315 | { | |
316 | dlm_unregister_domain_handlers(dlm); | |
317 | dlm_complete_thread(dlm); | |
318 | dlm_complete_recovery_thread(dlm); | |
3156d267 | 319 | dlm_destroy_dlm_worker(dlm); |
6714d8e8 KH |
320 | |
321 | /* We've left the domain. Now we can take ourselves out of the | |
322 | * list and allow the kref stuff to help us free the | |
323 | * memory. */ | |
324 | spin_lock(&dlm_domain_lock); | |
325 | list_del_init(&dlm->list); | |
326 | spin_unlock(&dlm_domain_lock); | |
327 | ||
328 | /* Wake up anyone waiting for us to remove this domain */ | |
329 | wake_up(&dlm_domain_events); | |
330 | } | |
331 | ||
332 | static void dlm_migrate_all_locks(struct dlm_ctxt *dlm) | |
333 | { | |
334 | int i; | |
335 | struct dlm_lock_resource *res; | |
336 | ||
337 | mlog(0, "Migrating locks from domain %s\n", dlm->name); | |
338 | restart: | |
339 | spin_lock(&dlm->spinlock); | |
81f2094a | 340 | for (i = 0; i < DLM_HASH_BUCKETS; i++) { |
03d864c0 DP |
341 | while (!hlist_empty(dlm_lockres_hash(dlm, i))) { |
342 | res = hlist_entry(dlm_lockres_hash(dlm, i)->first, | |
81f2094a | 343 | struct dlm_lock_resource, hash_node); |
6714d8e8 KH |
344 | /* need reference when manually grabbing lockres */ |
345 | dlm_lockres_get(res); | |
346 | /* this should unhash the lockres | |
347 | * and exit with dlm->spinlock */ | |
348 | mlog(0, "purging res=%p\n", res); | |
349 | if (dlm_lockres_is_dirty(dlm, res)) { | |
350 | /* HACK! this should absolutely go. | |
351 | * need to figure out why some empty | |
352 | * lockreses are still marked dirty */ | |
353 | mlog(ML_ERROR, "lockres %.*s dirty!\n", | |
354 | res->lockname.len, res->lockname.name); | |
355 | ||
356 | spin_unlock(&dlm->spinlock); | |
357 | dlm_kick_thread(dlm, res); | |
358 | wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res)); | |
359 | dlm_lockres_put(res); | |
360 | goto restart; | |
361 | } | |
362 | dlm_purge_lockres(dlm, res); | |
363 | dlm_lockres_put(res); | |
364 | } | |
365 | } | |
366 | spin_unlock(&dlm->spinlock); | |
367 | ||
368 | mlog(0, "DONE Migrating locks from domain %s\n", dlm->name); | |
369 | } | |
370 | ||
371 | static int dlm_no_joining_node(struct dlm_ctxt *dlm) | |
372 | { | |
373 | int ret; | |
374 | ||
375 | spin_lock(&dlm->spinlock); | |
376 | ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN; | |
377 | spin_unlock(&dlm->spinlock); | |
378 | ||
379 | return ret; | |
380 | } | |
381 | ||
382 | static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm) | |
383 | { | |
384 | /* Yikes, a double spinlock! I need domain_lock for the dlm | |
385 | * state and the dlm spinlock for join state... Sorry! */ | |
386 | again: | |
387 | spin_lock(&dlm_domain_lock); | |
388 | spin_lock(&dlm->spinlock); | |
389 | ||
390 | if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) { | |
391 | mlog(0, "Node %d is joining, we wait on it.\n", | |
392 | dlm->joining_node); | |
393 | spin_unlock(&dlm->spinlock); | |
394 | spin_unlock(&dlm_domain_lock); | |
395 | ||
396 | wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm)); | |
397 | goto again; | |
398 | } | |
399 | ||
400 | dlm->dlm_state = DLM_CTXT_LEAVING; | |
401 | spin_unlock(&dlm->spinlock); | |
402 | spin_unlock(&dlm_domain_lock); | |
403 | } | |
404 | ||
405 | static void __dlm_print_nodes(struct dlm_ctxt *dlm) | |
406 | { | |
407 | int node = -1; | |
408 | ||
409 | assert_spin_locked(&dlm->spinlock); | |
410 | ||
411 | mlog(ML_NOTICE, "Nodes in my domain (\"%s\"):\n", dlm->name); | |
412 | ||
413 | while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, | |
414 | node + 1)) < O2NM_MAX_NODES) { | |
415 | mlog(ML_NOTICE, " node %d\n", node); | |
416 | } | |
417 | } | |
418 | ||
419 | static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data) | |
420 | { | |
421 | struct dlm_ctxt *dlm = data; | |
422 | unsigned int node; | |
423 | struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf; | |
424 | ||
425 | mlog_entry("%p %u %p", msg, len, data); | |
426 | ||
427 | if (!dlm_grab(dlm)) | |
428 | return 0; | |
429 | ||
430 | node = exit_msg->node_idx; | |
431 | ||
432 | mlog(0, "Node %u leaves domain %s\n", node, dlm->name); | |
433 | ||
434 | spin_lock(&dlm->spinlock); | |
435 | clear_bit(node, dlm->domain_map); | |
436 | __dlm_print_nodes(dlm); | |
437 | ||
438 | /* notify anything attached to the heartbeat events */ | |
439 | dlm_hb_event_notify_attached(dlm, node, 0); | |
440 | ||
441 | spin_unlock(&dlm->spinlock); | |
442 | ||
443 | dlm_put(dlm); | |
444 | ||
445 | return 0; | |
446 | } | |
447 | ||
448 | static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm, | |
449 | unsigned int node) | |
450 | { | |
451 | int status; | |
452 | struct dlm_exit_domain leave_msg; | |
453 | ||
454 | mlog(0, "Asking node %u if we can leave the domain %s me = %u\n", | |
455 | node, dlm->name, dlm->node_num); | |
456 | ||
457 | memset(&leave_msg, 0, sizeof(leave_msg)); | |
458 | leave_msg.node_idx = dlm->node_num; | |
459 | ||
460 | status = o2net_send_message(DLM_EXIT_DOMAIN_MSG, dlm->key, | |
461 | &leave_msg, sizeof(leave_msg), node, | |
462 | NULL); | |
463 | ||
464 | mlog(0, "status return %d from o2net_send_message\n", status); | |
465 | ||
466 | return status; | |
467 | } | |
468 | ||
469 | ||
470 | static void dlm_leave_domain(struct dlm_ctxt *dlm) | |
471 | { | |
472 | int node, clear_node, status; | |
473 | ||
474 | /* At this point we've migrated away all our locks and won't | |
475 | * accept mastership of new ones. The dlm is responsible for | |
476 | * almost nothing now. We make sure not to confuse any joining | |
477 | * nodes and then commence shutdown procedure. */ | |
478 | ||
479 | spin_lock(&dlm->spinlock); | |
480 | /* Clear ourselves from the domain map */ | |
481 | clear_bit(dlm->node_num, dlm->domain_map); | |
482 | while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, | |
483 | 0)) < O2NM_MAX_NODES) { | |
484 | /* Drop the dlm spinlock. This is safe wrt the domain_map. | |
485 | * -nodes cannot be added now as the | |
486 | * query_join_handlers knows to respond with OK_NO_MAP | |
487 | * -we catch the right network errors if a node is | |
488 | * removed from the map while we're sending him the | |
489 | * exit message. */ | |
490 | spin_unlock(&dlm->spinlock); | |
491 | ||
492 | clear_node = 1; | |
493 | ||
494 | status = dlm_send_one_domain_exit(dlm, node); | |
495 | if (status < 0 && | |
496 | status != -ENOPROTOOPT && | |
497 | status != -ENOTCONN) { | |
498 | mlog(ML_NOTICE, "Error %d sending domain exit message " | |
499 | "to node %d\n", status, node); | |
500 | ||
501 | /* Not sure what to do here but lets sleep for | |
502 | * a bit in case this was a transient | |
503 | * error... */ | |
504 | msleep(DLM_DOMAIN_BACKOFF_MS); | |
505 | clear_node = 0; | |
506 | } | |
507 | ||
508 | spin_lock(&dlm->spinlock); | |
509 | /* If we're not clearing the node bit then we intend | |
510 | * to loop back around to try again. */ | |
511 | if (clear_node) | |
512 | clear_bit(node, dlm->domain_map); | |
513 | } | |
514 | spin_unlock(&dlm->spinlock); | |
515 | } | |
516 | ||
517 | int dlm_joined(struct dlm_ctxt *dlm) | |
518 | { | |
519 | int ret = 0; | |
520 | ||
521 | spin_lock(&dlm_domain_lock); | |
522 | ||
523 | if (dlm->dlm_state == DLM_CTXT_JOINED) | |
524 | ret = 1; | |
525 | ||
526 | spin_unlock(&dlm_domain_lock); | |
527 | ||
528 | return ret; | |
529 | } | |
530 | ||
531 | int dlm_shutting_down(struct dlm_ctxt *dlm) | |
532 | { | |
533 | int ret = 0; | |
534 | ||
535 | spin_lock(&dlm_domain_lock); | |
536 | ||
537 | if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN) | |
538 | ret = 1; | |
539 | ||
540 | spin_unlock(&dlm_domain_lock); | |
541 | ||
542 | return ret; | |
543 | } | |
544 | ||
545 | void dlm_unregister_domain(struct dlm_ctxt *dlm) | |
546 | { | |
547 | int leave = 0; | |
548 | ||
549 | spin_lock(&dlm_domain_lock); | |
550 | BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED); | |
551 | BUG_ON(!dlm->num_joins); | |
552 | ||
553 | dlm->num_joins--; | |
554 | if (!dlm->num_joins) { | |
555 | /* We mark it "in shutdown" now so new register | |
556 | * requests wait until we've completely left the | |
557 | * domain. Don't use DLM_CTXT_LEAVING yet as we still | |
558 | * want new domain joins to communicate with us at | |
559 | * least until we've completed migration of our | |
560 | * resources. */ | |
561 | dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN; | |
562 | leave = 1; | |
563 | } | |
564 | spin_unlock(&dlm_domain_lock); | |
565 | ||
566 | if (leave) { | |
567 | mlog(0, "shutting down domain %s\n", dlm->name); | |
568 | ||
569 | /* We changed dlm state, notify the thread */ | |
570 | dlm_kick_thread(dlm, NULL); | |
571 | ||
572 | dlm_migrate_all_locks(dlm); | |
573 | dlm_mark_domain_leaving(dlm); | |
574 | dlm_leave_domain(dlm); | |
575 | dlm_complete_dlm_shutdown(dlm); | |
576 | } | |
577 | dlm_put(dlm); | |
578 | } | |
579 | EXPORT_SYMBOL_GPL(dlm_unregister_domain); | |
580 | ||
581 | static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data) | |
582 | { | |
583 | struct dlm_query_join_request *query; | |
584 | enum dlm_query_join_response response; | |
585 | struct dlm_ctxt *dlm = NULL; | |
586 | ||
587 | query = (struct dlm_query_join_request *) msg->buf; | |
588 | ||
589 | mlog(0, "node %u wants to join domain %s\n", query->node_idx, | |
590 | query->domain); | |
591 | ||
592 | /* | |
593 | * If heartbeat doesn't consider the node live, tell it | |
594 | * to back off and try again. This gives heartbeat a chance | |
595 | * to catch up. | |
596 | */ | |
597 | if (!o2hb_check_node_heartbeating(query->node_idx)) { | |
598 | mlog(0, "node %u is not in our live map yet\n", | |
599 | query->node_idx); | |
600 | ||
601 | response = JOIN_DISALLOW; | |
602 | goto respond; | |
603 | } | |
604 | ||
605 | response = JOIN_OK_NO_MAP; | |
606 | ||
607 | spin_lock(&dlm_domain_lock); | |
608 | dlm = __dlm_lookup_domain_full(query->domain, query->name_len); | |
609 | /* Once the dlm ctxt is marked as leaving then we don't want | |
e2faea4c KH |
610 | * to be put in someone's domain map. |
611 | * Also, explicitly disallow joining at certain troublesome | |
612 | * times (ie. during recovery). */ | |
6714d8e8 | 613 | if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) { |
e2faea4c | 614 | int bit = query->node_idx; |
6714d8e8 KH |
615 | spin_lock(&dlm->spinlock); |
616 | ||
617 | if (dlm->dlm_state == DLM_CTXT_NEW && | |
618 | dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) { | |
619 | /*If this is a brand new context and we | |
620 | * haven't started our join process yet, then | |
621 | * the other node won the race. */ | |
622 | response = JOIN_OK_NO_MAP; | |
623 | } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) { | |
624 | /* Disallow parallel joins. */ | |
625 | response = JOIN_DISALLOW; | |
e2faea4c KH |
626 | } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) { |
627 | mlog(ML_NOTICE, "node %u trying to join, but recovery " | |
628 | "is ongoing.\n", bit); | |
629 | response = JOIN_DISALLOW; | |
630 | } else if (test_bit(bit, dlm->recovery_map)) { | |
631 | mlog(ML_NOTICE, "node %u trying to join, but it " | |
632 | "still needs recovery.\n", bit); | |
633 | response = JOIN_DISALLOW; | |
634 | } else if (test_bit(bit, dlm->domain_map)) { | |
635 | mlog(ML_NOTICE, "node %u trying to join, but it " | |
636 | "is still in the domain! needs recovery?\n", | |
637 | bit); | |
638 | response = JOIN_DISALLOW; | |
6714d8e8 KH |
639 | } else { |
640 | /* Alright we're fully a part of this domain | |
641 | * so we keep some state as to who's joining | |
642 | * and indicate to him that needs to be fixed | |
643 | * up. */ | |
644 | response = JOIN_OK; | |
645 | __dlm_set_joining_node(dlm, query->node_idx); | |
646 | } | |
647 | ||
648 | spin_unlock(&dlm->spinlock); | |
649 | } | |
650 | spin_unlock(&dlm_domain_lock); | |
651 | ||
652 | respond: | |
653 | mlog(0, "We respond with %u\n", response); | |
654 | ||
655 | return response; | |
656 | } | |
657 | ||
658 | static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data) | |
659 | { | |
660 | struct dlm_assert_joined *assert; | |
661 | struct dlm_ctxt *dlm = NULL; | |
662 | ||
663 | assert = (struct dlm_assert_joined *) msg->buf; | |
664 | ||
665 | mlog(0, "node %u asserts join on domain %s\n", assert->node_idx, | |
666 | assert->domain); | |
667 | ||
668 | spin_lock(&dlm_domain_lock); | |
669 | dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len); | |
670 | /* XXX should we consider no dlm ctxt an error? */ | |
671 | if (dlm) { | |
672 | spin_lock(&dlm->spinlock); | |
673 | ||
674 | /* Alright, this node has officially joined our | |
675 | * domain. Set him in the map and clean up our | |
676 | * leftover join state. */ | |
677 | BUG_ON(dlm->joining_node != assert->node_idx); | |
678 | set_bit(assert->node_idx, dlm->domain_map); | |
679 | __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); | |
680 | ||
681 | __dlm_print_nodes(dlm); | |
682 | ||
683 | /* notify anything attached to the heartbeat events */ | |
684 | dlm_hb_event_notify_attached(dlm, assert->node_idx, 1); | |
685 | ||
686 | spin_unlock(&dlm->spinlock); | |
687 | } | |
688 | spin_unlock(&dlm_domain_lock); | |
689 | ||
690 | return 0; | |
691 | } | |
692 | ||
693 | static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data) | |
694 | { | |
695 | struct dlm_cancel_join *cancel; | |
696 | struct dlm_ctxt *dlm = NULL; | |
697 | ||
698 | cancel = (struct dlm_cancel_join *) msg->buf; | |
699 | ||
700 | mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx, | |
701 | cancel->domain); | |
702 | ||
703 | spin_lock(&dlm_domain_lock); | |
704 | dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len); | |
705 | ||
706 | if (dlm) { | |
707 | spin_lock(&dlm->spinlock); | |
708 | ||
709 | /* Yikes, this guy wants to cancel his join. No | |
710 | * problem, we simply cleanup our join state. */ | |
711 | BUG_ON(dlm->joining_node != cancel->node_idx); | |
712 | __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); | |
713 | ||
714 | spin_unlock(&dlm->spinlock); | |
715 | } | |
716 | spin_unlock(&dlm_domain_lock); | |
717 | ||
718 | return 0; | |
719 | } | |
720 | ||
721 | static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm, | |
722 | unsigned int node) | |
723 | { | |
724 | int status; | |
725 | struct dlm_cancel_join cancel_msg; | |
726 | ||
727 | memset(&cancel_msg, 0, sizeof(cancel_msg)); | |
728 | cancel_msg.node_idx = dlm->node_num; | |
729 | cancel_msg.name_len = strlen(dlm->name); | |
730 | memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len); | |
731 | ||
732 | status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY, | |
733 | &cancel_msg, sizeof(cancel_msg), node, | |
734 | NULL); | |
735 | if (status < 0) { | |
736 | mlog_errno(status); | |
737 | goto bail; | |
738 | } | |
739 | ||
740 | bail: | |
741 | return status; | |
742 | } | |
743 | ||
744 | /* map_size should be in bytes. */ | |
745 | static int dlm_send_join_cancels(struct dlm_ctxt *dlm, | |
746 | unsigned long *node_map, | |
747 | unsigned int map_size) | |
748 | { | |
749 | int status, tmpstat; | |
750 | unsigned int node; | |
751 | ||
752 | if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) * | |
753 | sizeof(unsigned long))) { | |
754 | mlog(ML_ERROR, | |
755 | "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n", | |
756 | map_size, BITS_TO_LONGS(O2NM_MAX_NODES)); | |
757 | return -EINVAL; | |
758 | } | |
759 | ||
760 | status = 0; | |
761 | node = -1; | |
762 | while ((node = find_next_bit(node_map, O2NM_MAX_NODES, | |
763 | node + 1)) < O2NM_MAX_NODES) { | |
764 | if (node == dlm->node_num) | |
765 | continue; | |
766 | ||
767 | tmpstat = dlm_send_one_join_cancel(dlm, node); | |
768 | if (tmpstat) { | |
769 | mlog(ML_ERROR, "Error return %d cancelling join on " | |
770 | "node %d\n", tmpstat, node); | |
771 | if (!status) | |
772 | status = tmpstat; | |
773 | } | |
774 | } | |
775 | ||
776 | if (status) | |
777 | mlog_errno(status); | |
778 | return status; | |
779 | } | |
780 | ||
781 | static int dlm_request_join(struct dlm_ctxt *dlm, | |
782 | int node, | |
783 | enum dlm_query_join_response *response) | |
784 | { | |
785 | int status, retval; | |
786 | struct dlm_query_join_request join_msg; | |
787 | ||
788 | mlog(0, "querying node %d\n", node); | |
789 | ||
790 | memset(&join_msg, 0, sizeof(join_msg)); | |
791 | join_msg.node_idx = dlm->node_num; | |
792 | join_msg.name_len = strlen(dlm->name); | |
793 | memcpy(join_msg.domain, dlm->name, join_msg.name_len); | |
794 | ||
795 | status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg, | |
796 | sizeof(join_msg), node, &retval); | |
797 | if (status < 0 && status != -ENOPROTOOPT) { | |
798 | mlog_errno(status); | |
799 | goto bail; | |
800 | } | |
801 | ||
802 | /* -ENOPROTOOPT from the net code means the other side isn't | |
803 | listening for our message type -- that's fine, it means | |
804 | his dlm isn't up, so we can consider him a 'yes' but not | |
805 | joined into the domain. */ | |
806 | if (status == -ENOPROTOOPT) { | |
807 | status = 0; | |
808 | *response = JOIN_OK_NO_MAP; | |
809 | } else if (retval == JOIN_DISALLOW || | |
810 | retval == JOIN_OK || | |
811 | retval == JOIN_OK_NO_MAP) { | |
812 | *response = retval; | |
813 | } else { | |
814 | status = -EINVAL; | |
815 | mlog(ML_ERROR, "invalid response %d from node %u\n", retval, | |
816 | node); | |
817 | } | |
818 | ||
819 | mlog(0, "status %d, node %d response is %d\n", status, node, | |
820 | *response); | |
821 | ||
822 | bail: | |
823 | return status; | |
824 | } | |
825 | ||
826 | static int dlm_send_one_join_assert(struct dlm_ctxt *dlm, | |
827 | unsigned int node) | |
828 | { | |
829 | int status; | |
830 | struct dlm_assert_joined assert_msg; | |
831 | ||
832 | mlog(0, "Sending join assert to node %u\n", node); | |
833 | ||
834 | memset(&assert_msg, 0, sizeof(assert_msg)); | |
835 | assert_msg.node_idx = dlm->node_num; | |
836 | assert_msg.name_len = strlen(dlm->name); | |
837 | memcpy(assert_msg.domain, dlm->name, assert_msg.name_len); | |
838 | ||
839 | status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY, | |
840 | &assert_msg, sizeof(assert_msg), node, | |
841 | NULL); | |
842 | if (status < 0) | |
843 | mlog_errno(status); | |
844 | ||
845 | return status; | |
846 | } | |
847 | ||
848 | static void dlm_send_join_asserts(struct dlm_ctxt *dlm, | |
849 | unsigned long *node_map) | |
850 | { | |
851 | int status, node, live; | |
852 | ||
853 | status = 0; | |
854 | node = -1; | |
855 | while ((node = find_next_bit(node_map, O2NM_MAX_NODES, | |
856 | node + 1)) < O2NM_MAX_NODES) { | |
857 | if (node == dlm->node_num) | |
858 | continue; | |
859 | ||
860 | do { | |
861 | /* It is very important that this message be | |
862 | * received so we spin until either the node | |
863 | * has died or it gets the message. */ | |
864 | status = dlm_send_one_join_assert(dlm, node); | |
865 | ||
866 | spin_lock(&dlm->spinlock); | |
867 | live = test_bit(node, dlm->live_nodes_map); | |
868 | spin_unlock(&dlm->spinlock); | |
869 | ||
870 | if (status) { | |
871 | mlog(ML_ERROR, "Error return %d asserting " | |
872 | "join on node %d\n", status, node); | |
873 | ||
874 | /* give us some time between errors... */ | |
875 | if (live) | |
876 | msleep(DLM_DOMAIN_BACKOFF_MS); | |
877 | } | |
878 | } while (status && live); | |
879 | } | |
880 | } | |
881 | ||
882 | struct domain_join_ctxt { | |
883 | unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; | |
884 | unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; | |
885 | }; | |
886 | ||
887 | static int dlm_should_restart_join(struct dlm_ctxt *dlm, | |
888 | struct domain_join_ctxt *ctxt, | |
889 | enum dlm_query_join_response response) | |
890 | { | |
891 | int ret; | |
892 | ||
893 | if (response == JOIN_DISALLOW) { | |
894 | mlog(0, "Latest response of disallow -- should restart\n"); | |
895 | return 1; | |
896 | } | |
897 | ||
898 | spin_lock(&dlm->spinlock); | |
899 | /* For now, we restart the process if the node maps have | |
900 | * changed at all */ | |
901 | ret = memcmp(ctxt->live_map, dlm->live_nodes_map, | |
902 | sizeof(dlm->live_nodes_map)); | |
903 | spin_unlock(&dlm->spinlock); | |
904 | ||
905 | if (ret) | |
906 | mlog(0, "Node maps changed -- should restart\n"); | |
907 | ||
908 | return ret; | |
909 | } | |
910 | ||
911 | static int dlm_try_to_join_domain(struct dlm_ctxt *dlm) | |
912 | { | |
913 | int status = 0, tmpstat, node; | |
914 | struct domain_join_ctxt *ctxt; | |
915 | enum dlm_query_join_response response; | |
916 | ||
917 | mlog_entry("%p", dlm); | |
918 | ||
919 | ctxt = kcalloc(1, sizeof(*ctxt), GFP_KERNEL); | |
920 | if (!ctxt) { | |
921 | status = -ENOMEM; | |
922 | mlog_errno(status); | |
923 | goto bail; | |
924 | } | |
925 | ||
926 | /* group sem locking should work for us here -- we're already | |
927 | * registered for heartbeat events so filling this should be | |
928 | * atomic wrt getting those handlers called. */ | |
929 | o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map)); | |
930 | ||
931 | spin_lock(&dlm->spinlock); | |
932 | memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map)); | |
933 | ||
934 | __dlm_set_joining_node(dlm, dlm->node_num); | |
935 | ||
936 | spin_unlock(&dlm->spinlock); | |
937 | ||
938 | node = -1; | |
939 | while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES, | |
940 | node + 1)) < O2NM_MAX_NODES) { | |
941 | if (node == dlm->node_num) | |
942 | continue; | |
943 | ||
944 | status = dlm_request_join(dlm, node, &response); | |
945 | if (status < 0) { | |
946 | mlog_errno(status); | |
947 | goto bail; | |
948 | } | |
949 | ||
950 | /* Ok, either we got a response or the node doesn't have a | |
951 | * dlm up. */ | |
952 | if (response == JOIN_OK) | |
953 | set_bit(node, ctxt->yes_resp_map); | |
954 | ||
955 | if (dlm_should_restart_join(dlm, ctxt, response)) { | |
956 | status = -EAGAIN; | |
957 | goto bail; | |
958 | } | |
959 | } | |
960 | ||
961 | mlog(0, "Yay, done querying nodes!\n"); | |
962 | ||
963 | /* Yay, everyone agree's we can join the domain. My domain is | |
964 | * comprised of all nodes who were put in the | |
965 | * yes_resp_map. Copy that into our domain map and send a join | |
966 | * assert message to clean up everyone elses state. */ | |
967 | spin_lock(&dlm->spinlock); | |
968 | memcpy(dlm->domain_map, ctxt->yes_resp_map, | |
969 | sizeof(ctxt->yes_resp_map)); | |
970 | set_bit(dlm->node_num, dlm->domain_map); | |
971 | spin_unlock(&dlm->spinlock); | |
972 | ||
973 | dlm_send_join_asserts(dlm, ctxt->yes_resp_map); | |
974 | ||
975 | /* Joined state *must* be set before the joining node | |
976 | * information, otherwise the query_join handler may read no | |
977 | * current joiner but a state of NEW and tell joining nodes | |
978 | * we're not in the domain. */ | |
979 | spin_lock(&dlm_domain_lock); | |
980 | dlm->dlm_state = DLM_CTXT_JOINED; | |
981 | dlm->num_joins++; | |
982 | spin_unlock(&dlm_domain_lock); | |
983 | ||
984 | bail: | |
985 | spin_lock(&dlm->spinlock); | |
986 | __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); | |
987 | if (!status) | |
988 | __dlm_print_nodes(dlm); | |
989 | spin_unlock(&dlm->spinlock); | |
990 | ||
991 | if (ctxt) { | |
992 | /* Do we need to send a cancel message to any nodes? */ | |
993 | if (status < 0) { | |
994 | tmpstat = dlm_send_join_cancels(dlm, | |
995 | ctxt->yes_resp_map, | |
996 | sizeof(ctxt->yes_resp_map)); | |
997 | if (tmpstat < 0) | |
998 | mlog_errno(tmpstat); | |
999 | } | |
1000 | kfree(ctxt); | |
1001 | } | |
1002 | ||
1003 | mlog(0, "returning %d\n", status); | |
1004 | return status; | |
1005 | } | |
1006 | ||
1007 | static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm) | |
1008 | { | |
1009 | o2hb_unregister_callback(&dlm->dlm_hb_up); | |
1010 | o2hb_unregister_callback(&dlm->dlm_hb_down); | |
1011 | o2net_unregister_handler_list(&dlm->dlm_domain_handlers); | |
1012 | } | |
1013 | ||
1014 | static int dlm_register_domain_handlers(struct dlm_ctxt *dlm) | |
1015 | { | |
1016 | int status; | |
1017 | ||
1018 | mlog(0, "registering handlers.\n"); | |
1019 | ||
1020 | o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB, | |
1021 | dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI); | |
1022 | status = o2hb_register_callback(&dlm->dlm_hb_down); | |
1023 | if (status) | |
1024 | goto bail; | |
1025 | ||
1026 | o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB, | |
1027 | dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI); | |
1028 | status = o2hb_register_callback(&dlm->dlm_hb_up); | |
1029 | if (status) | |
1030 | goto bail; | |
1031 | ||
1032 | status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key, | |
1033 | sizeof(struct dlm_master_request), | |
1034 | dlm_master_request_handler, | |
1035 | dlm, &dlm->dlm_domain_handlers); | |
1036 | if (status) | |
1037 | goto bail; | |
1038 | ||
1039 | status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key, | |
1040 | sizeof(struct dlm_assert_master), | |
1041 | dlm_assert_master_handler, | |
1042 | dlm, &dlm->dlm_domain_handlers); | |
1043 | if (status) | |
1044 | goto bail; | |
1045 | ||
1046 | status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key, | |
1047 | sizeof(struct dlm_create_lock), | |
1048 | dlm_create_lock_handler, | |
1049 | dlm, &dlm->dlm_domain_handlers); | |
1050 | if (status) | |
1051 | goto bail; | |
1052 | ||
1053 | status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key, | |
1054 | DLM_CONVERT_LOCK_MAX_LEN, | |
1055 | dlm_convert_lock_handler, | |
1056 | dlm, &dlm->dlm_domain_handlers); | |
1057 | if (status) | |
1058 | goto bail; | |
1059 | ||
1060 | status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key, | |
1061 | DLM_UNLOCK_LOCK_MAX_LEN, | |
1062 | dlm_unlock_lock_handler, | |
1063 | dlm, &dlm->dlm_domain_handlers); | |
1064 | if (status) | |
1065 | goto bail; | |
1066 | ||
1067 | status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key, | |
1068 | DLM_PROXY_AST_MAX_LEN, | |
1069 | dlm_proxy_ast_handler, | |
1070 | dlm, &dlm->dlm_domain_handlers); | |
1071 | if (status) | |
1072 | goto bail; | |
1073 | ||
1074 | status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key, | |
1075 | sizeof(struct dlm_exit_domain), | |
1076 | dlm_exit_domain_handler, | |
1077 | dlm, &dlm->dlm_domain_handlers); | |
1078 | if (status) | |
1079 | goto bail; | |
1080 | ||
1081 | status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key, | |
1082 | sizeof(struct dlm_migrate_request), | |
1083 | dlm_migrate_request_handler, | |
1084 | dlm, &dlm->dlm_domain_handlers); | |
1085 | if (status) | |
1086 | goto bail; | |
1087 | ||
1088 | status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key, | |
1089 | DLM_MIG_LOCKRES_MAX_LEN, | |
1090 | dlm_mig_lockres_handler, | |
1091 | dlm, &dlm->dlm_domain_handlers); | |
1092 | if (status) | |
1093 | goto bail; | |
1094 | ||
1095 | status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key, | |
1096 | sizeof(struct dlm_master_requery), | |
1097 | dlm_master_requery_handler, | |
1098 | dlm, &dlm->dlm_domain_handlers); | |
1099 | if (status) | |
1100 | goto bail; | |
1101 | ||
1102 | status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key, | |
1103 | sizeof(struct dlm_lock_request), | |
1104 | dlm_request_all_locks_handler, | |
1105 | dlm, &dlm->dlm_domain_handlers); | |
1106 | if (status) | |
1107 | goto bail; | |
1108 | ||
1109 | status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key, | |
1110 | sizeof(struct dlm_reco_data_done), | |
1111 | dlm_reco_data_done_handler, | |
1112 | dlm, &dlm->dlm_domain_handlers); | |
1113 | if (status) | |
1114 | goto bail; | |
1115 | ||
1116 | status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key, | |
1117 | sizeof(struct dlm_begin_reco), | |
1118 | dlm_begin_reco_handler, | |
1119 | dlm, &dlm->dlm_domain_handlers); | |
1120 | if (status) | |
1121 | goto bail; | |
1122 | ||
1123 | status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key, | |
1124 | sizeof(struct dlm_finalize_reco), | |
1125 | dlm_finalize_reco_handler, | |
1126 | dlm, &dlm->dlm_domain_handlers); | |
1127 | if (status) | |
1128 | goto bail; | |
1129 | ||
1130 | bail: | |
1131 | if (status) | |
1132 | dlm_unregister_domain_handlers(dlm); | |
1133 | ||
1134 | return status; | |
1135 | } | |
1136 | ||
1137 | static int dlm_join_domain(struct dlm_ctxt *dlm) | |
1138 | { | |
1139 | int status; | |
1140 | ||
1141 | BUG_ON(!dlm); | |
1142 | ||
1143 | mlog(0, "Join domain %s\n", dlm->name); | |
1144 | ||
1145 | status = dlm_register_domain_handlers(dlm); | |
1146 | if (status) { | |
1147 | mlog_errno(status); | |
1148 | goto bail; | |
1149 | } | |
1150 | ||
1151 | status = dlm_launch_thread(dlm); | |
1152 | if (status < 0) { | |
1153 | mlog_errno(status); | |
1154 | goto bail; | |
1155 | } | |
1156 | ||
1157 | status = dlm_launch_recovery_thread(dlm); | |
1158 | if (status < 0) { | |
1159 | mlog_errno(status); | |
1160 | goto bail; | |
1161 | } | |
1162 | ||
3156d267 KH |
1163 | dlm->dlm_worker = create_singlethread_workqueue("dlm_wq"); |
1164 | if (!dlm->dlm_worker) { | |
1165 | status = -ENOMEM; | |
1166 | mlog_errno(status); | |
1167 | goto bail; | |
1168 | } | |
1169 | ||
6714d8e8 KH |
1170 | do { |
1171 | unsigned int backoff; | |
1172 | status = dlm_try_to_join_domain(dlm); | |
1173 | ||
1174 | /* If we're racing another node to the join, then we | |
1175 | * need to back off temporarily and let them | |
1176 | * complete. */ | |
1177 | if (status == -EAGAIN) { | |
1178 | if (signal_pending(current)) { | |
1179 | status = -ERESTARTSYS; | |
1180 | goto bail; | |
1181 | } | |
1182 | ||
1183 | /* | |
1184 | * <chip> After you! | |
1185 | * <dale> No, after you! | |
1186 | * <chip> I insist! | |
1187 | * <dale> But you first! | |
1188 | * ... | |
1189 | */ | |
1190 | backoff = (unsigned int)(jiffies & 0x3); | |
1191 | backoff *= DLM_DOMAIN_BACKOFF_MS; | |
1192 | mlog(0, "backoff %d\n", backoff); | |
1193 | msleep(backoff); | |
1194 | } | |
1195 | } while (status == -EAGAIN); | |
1196 | ||
1197 | if (status < 0) { | |
1198 | mlog_errno(status); | |
1199 | goto bail; | |
1200 | } | |
1201 | ||
1202 | status = 0; | |
1203 | bail: | |
1204 | wake_up(&dlm_domain_events); | |
1205 | ||
1206 | if (status) { | |
1207 | dlm_unregister_domain_handlers(dlm); | |
1208 | dlm_complete_thread(dlm); | |
1209 | dlm_complete_recovery_thread(dlm); | |
3156d267 | 1210 | dlm_destroy_dlm_worker(dlm); |
6714d8e8 KH |
1211 | } |
1212 | ||
1213 | return status; | |
1214 | } | |
1215 | ||
1216 | static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain, | |
1217 | u32 key) | |
1218 | { | |
1219 | int i; | |
1220 | struct dlm_ctxt *dlm = NULL; | |
1221 | ||
1222 | dlm = kcalloc(1, sizeof(*dlm), GFP_KERNEL); | |
1223 | if (!dlm) { | |
1224 | mlog_errno(-ENOMEM); | |
1225 | goto leave; | |
1226 | } | |
1227 | ||
1228 | dlm->name = kmalloc(strlen(domain) + 1, GFP_KERNEL); | |
1229 | if (dlm->name == NULL) { | |
1230 | mlog_errno(-ENOMEM); | |
1231 | kfree(dlm); | |
1232 | dlm = NULL; | |
1233 | goto leave; | |
1234 | } | |
1235 | ||
03d864c0 | 1236 | dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES); |
81f2094a | 1237 | if (!dlm->lockres_hash) { |
6714d8e8 KH |
1238 | mlog_errno(-ENOMEM); |
1239 | kfree(dlm->name); | |
1240 | kfree(dlm); | |
1241 | dlm = NULL; | |
1242 | goto leave; | |
1243 | } | |
6714d8e8 | 1244 | |
03d864c0 DP |
1245 | for (i = 0; i < DLM_HASH_BUCKETS; i++) |
1246 | INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i)); | |
6714d8e8 KH |
1247 | |
1248 | strcpy(dlm->name, domain); | |
1249 | dlm->key = key; | |
1250 | dlm->node_num = o2nm_this_node(); | |
1251 | ||
1252 | spin_lock_init(&dlm->spinlock); | |
1253 | spin_lock_init(&dlm->master_lock); | |
1254 | spin_lock_init(&dlm->ast_lock); | |
1255 | INIT_LIST_HEAD(&dlm->list); | |
1256 | INIT_LIST_HEAD(&dlm->dirty_list); | |
1257 | INIT_LIST_HEAD(&dlm->reco.resources); | |
1258 | INIT_LIST_HEAD(&dlm->reco.received); | |
1259 | INIT_LIST_HEAD(&dlm->reco.node_data); | |
1260 | INIT_LIST_HEAD(&dlm->purge_list); | |
1261 | INIT_LIST_HEAD(&dlm->dlm_domain_handlers); | |
1262 | dlm->reco.state = 0; | |
1263 | ||
1264 | INIT_LIST_HEAD(&dlm->pending_asts); | |
1265 | INIT_LIST_HEAD(&dlm->pending_basts); | |
1266 | ||
1267 | mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n", | |
1268 | dlm->recovery_map, &(dlm->recovery_map[0])); | |
1269 | ||
1270 | memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map)); | |
1271 | memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map)); | |
1272 | memset(dlm->domain_map, 0, sizeof(dlm->domain_map)); | |
1273 | ||
1274 | dlm->dlm_thread_task = NULL; | |
1275 | dlm->dlm_reco_thread_task = NULL; | |
3156d267 | 1276 | dlm->dlm_worker = NULL; |
6714d8e8 KH |
1277 | init_waitqueue_head(&dlm->dlm_thread_wq); |
1278 | init_waitqueue_head(&dlm->dlm_reco_thread_wq); | |
1279 | init_waitqueue_head(&dlm->reco.event); | |
1280 | init_waitqueue_head(&dlm->ast_wq); | |
1281 | init_waitqueue_head(&dlm->migration_wq); | |
1282 | INIT_LIST_HEAD(&dlm->master_list); | |
1283 | INIT_LIST_HEAD(&dlm->mle_hb_events); | |
1284 | ||
1285 | dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN; | |
1286 | init_waitqueue_head(&dlm->dlm_join_events); | |
1287 | ||
1288 | dlm->reco.new_master = O2NM_INVALID_NODE_NUM; | |
1289 | dlm->reco.dead_node = O2NM_INVALID_NODE_NUM; | |
1290 | atomic_set(&dlm->local_resources, 0); | |
1291 | atomic_set(&dlm->remote_resources, 0); | |
1292 | atomic_set(&dlm->unknown_resources, 0); | |
1293 | ||
1294 | spin_lock_init(&dlm->work_lock); | |
1295 | INIT_LIST_HEAD(&dlm->work_list); | |
1296 | INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work, dlm); | |
1297 | ||
1298 | kref_init(&dlm->dlm_refs); | |
1299 | dlm->dlm_state = DLM_CTXT_NEW; | |
1300 | ||
1301 | INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks); | |
1302 | ||
1303 | mlog(0, "context init: refcount %u\n", | |
1304 | atomic_read(&dlm->dlm_refs.refcount)); | |
1305 | ||
1306 | leave: | |
1307 | return dlm; | |
1308 | } | |
1309 | ||
1310 | /* | |
1311 | * dlm_register_domain: one-time setup per "domain" | |
1312 | */ | |
1313 | struct dlm_ctxt * dlm_register_domain(const char *domain, | |
1314 | u32 key) | |
1315 | { | |
1316 | int ret; | |
1317 | struct dlm_ctxt *dlm = NULL; | |
1318 | struct dlm_ctxt *new_ctxt = NULL; | |
1319 | ||
1320 | if (strlen(domain) > O2NM_MAX_NAME_LEN) { | |
1321 | ret = -ENAMETOOLONG; | |
1322 | mlog(ML_ERROR, "domain name length too long\n"); | |
1323 | goto leave; | |
1324 | } | |
1325 | ||
1326 | if (!o2hb_check_local_node_heartbeating()) { | |
1327 | mlog(ML_ERROR, "the local node has not been configured, or is " | |
1328 | "not heartbeating\n"); | |
1329 | ret = -EPROTO; | |
1330 | goto leave; | |
1331 | } | |
1332 | ||
1333 | mlog(0, "register called for domain \"%s\"\n", domain); | |
1334 | ||
1335 | retry: | |
1336 | dlm = NULL; | |
1337 | if (signal_pending(current)) { | |
1338 | ret = -ERESTARTSYS; | |
1339 | mlog_errno(ret); | |
1340 | goto leave; | |
1341 | } | |
1342 | ||
1343 | spin_lock(&dlm_domain_lock); | |
1344 | ||
1345 | dlm = __dlm_lookup_domain(domain); | |
1346 | if (dlm) { | |
1347 | if (dlm->dlm_state != DLM_CTXT_JOINED) { | |
1348 | spin_unlock(&dlm_domain_lock); | |
1349 | ||
1350 | mlog(0, "This ctxt is not joined yet!\n"); | |
1351 | wait_event_interruptible(dlm_domain_events, | |
1352 | dlm_wait_on_domain_helper( | |
1353 | domain)); | |
1354 | goto retry; | |
1355 | } | |
1356 | ||
1357 | __dlm_get(dlm); | |
1358 | dlm->num_joins++; | |
1359 | ||
1360 | spin_unlock(&dlm_domain_lock); | |
1361 | ||
1362 | ret = 0; | |
1363 | goto leave; | |
1364 | } | |
1365 | ||
1366 | /* doesn't exist */ | |
1367 | if (!new_ctxt) { | |
1368 | spin_unlock(&dlm_domain_lock); | |
1369 | ||
1370 | new_ctxt = dlm_alloc_ctxt(domain, key); | |
1371 | if (new_ctxt) | |
1372 | goto retry; | |
1373 | ||
1374 | ret = -ENOMEM; | |
1375 | mlog_errno(ret); | |
1376 | goto leave; | |
1377 | } | |
1378 | ||
1379 | /* a little variable switch-a-roo here... */ | |
1380 | dlm = new_ctxt; | |
1381 | new_ctxt = NULL; | |
1382 | ||
1383 | /* add the new domain */ | |
1384 | list_add_tail(&dlm->list, &dlm_domains); | |
1385 | spin_unlock(&dlm_domain_lock); | |
1386 | ||
1387 | ret = dlm_join_domain(dlm); | |
1388 | if (ret) { | |
1389 | mlog_errno(ret); | |
1390 | dlm_put(dlm); | |
1391 | goto leave; | |
1392 | } | |
1393 | ||
1394 | ret = 0; | |
1395 | leave: | |
1396 | if (new_ctxt) | |
1397 | dlm_free_ctxt_mem(new_ctxt); | |
1398 | ||
1399 | if (ret < 0) | |
1400 | dlm = ERR_PTR(ret); | |
1401 | ||
1402 | return dlm; | |
1403 | } | |
1404 | EXPORT_SYMBOL_GPL(dlm_register_domain); | |
1405 | ||
1406 | static LIST_HEAD(dlm_join_handlers); | |
1407 | ||
1408 | static void dlm_unregister_net_handlers(void) | |
1409 | { | |
1410 | o2net_unregister_handler_list(&dlm_join_handlers); | |
1411 | } | |
1412 | ||
1413 | static int dlm_register_net_handlers(void) | |
1414 | { | |
1415 | int status = 0; | |
1416 | ||
1417 | status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, | |
1418 | sizeof(struct dlm_query_join_request), | |
1419 | dlm_query_join_handler, | |
1420 | NULL, &dlm_join_handlers); | |
1421 | if (status) | |
1422 | goto bail; | |
1423 | ||
1424 | status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY, | |
1425 | sizeof(struct dlm_assert_joined), | |
1426 | dlm_assert_joined_handler, | |
1427 | NULL, &dlm_join_handlers); | |
1428 | if (status) | |
1429 | goto bail; | |
1430 | ||
1431 | status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY, | |
1432 | sizeof(struct dlm_cancel_join), | |
1433 | dlm_cancel_join_handler, | |
1434 | NULL, &dlm_join_handlers); | |
1435 | ||
1436 | bail: | |
1437 | if (status < 0) | |
1438 | dlm_unregister_net_handlers(); | |
1439 | ||
1440 | return status; | |
1441 | } | |
1442 | ||
1443 | /* Domain eviction callback handling. | |
1444 | * | |
1445 | * The file system requires notification of node death *before* the | |
1446 | * dlm completes it's recovery work, otherwise it may be able to | |
1447 | * acquire locks on resources requiring recovery. Since the dlm can | |
1448 | * evict a node from it's domain *before* heartbeat fires, a similar | |
1449 | * mechanism is required. */ | |
1450 | ||
1451 | /* Eviction is not expected to happen often, so a per-domain lock is | |
1452 | * not necessary. Eviction callbacks are allowed to sleep for short | |
1453 | * periods of time. */ | |
1454 | static DECLARE_RWSEM(dlm_callback_sem); | |
1455 | ||
1456 | void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm, | |
1457 | int node_num) | |
1458 | { | |
1459 | struct list_head *iter; | |
1460 | struct dlm_eviction_cb *cb; | |
1461 | ||
1462 | down_read(&dlm_callback_sem); | |
1463 | list_for_each(iter, &dlm->dlm_eviction_callbacks) { | |
1464 | cb = list_entry(iter, struct dlm_eviction_cb, ec_item); | |
1465 | ||
1466 | cb->ec_func(node_num, cb->ec_data); | |
1467 | } | |
1468 | up_read(&dlm_callback_sem); | |
1469 | } | |
1470 | ||
1471 | void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb, | |
1472 | dlm_eviction_func *f, | |
1473 | void *data) | |
1474 | { | |
1475 | INIT_LIST_HEAD(&cb->ec_item); | |
1476 | cb->ec_func = f; | |
1477 | cb->ec_data = data; | |
1478 | } | |
1479 | EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb); | |
1480 | ||
1481 | void dlm_register_eviction_cb(struct dlm_ctxt *dlm, | |
1482 | struct dlm_eviction_cb *cb) | |
1483 | { | |
1484 | down_write(&dlm_callback_sem); | |
1485 | list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks); | |
1486 | up_write(&dlm_callback_sem); | |
1487 | } | |
1488 | EXPORT_SYMBOL_GPL(dlm_register_eviction_cb); | |
1489 | ||
1490 | void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb) | |
1491 | { | |
1492 | down_write(&dlm_callback_sem); | |
1493 | list_del_init(&cb->ec_item); | |
1494 | up_write(&dlm_callback_sem); | |
1495 | } | |
1496 | EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb); | |
1497 | ||
1498 | static int __init dlm_init(void) | |
1499 | { | |
1500 | int status; | |
1501 | ||
1502 | dlm_print_version(); | |
1503 | ||
1504 | status = dlm_init_mle_cache(); | |
1505 | if (status) | |
1506 | return -1; | |
1507 | ||
1508 | status = dlm_register_net_handlers(); | |
1509 | if (status) { | |
1510 | dlm_destroy_mle_cache(); | |
1511 | return -1; | |
1512 | } | |
1513 | ||
1514 | return 0; | |
1515 | } | |
1516 | ||
1517 | static void __exit dlm_exit (void) | |
1518 | { | |
1519 | dlm_unregister_net_handlers(); | |
1520 | dlm_destroy_mle_cache(); | |
1521 | } | |
1522 | ||
1523 | MODULE_AUTHOR("Oracle"); | |
1524 | MODULE_LICENSE("GPL"); | |
1525 | ||
1526 | module_init(dlm_init); | |
1527 | module_exit(dlm_exit); |