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 | 70 | mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n", |
f5a923d1 MF |
71 | pages, (unsigned long)DLM_HASH_PAGES, |
72 | (unsigned long)DLM_BUCKETS_PER_PAGE); | |
03d864c0 DP |
73 | return vec; |
74 | out_free: | |
75 | dlm_free_pagevec(vec, i); | |
76 | return NULL; | |
77 | } | |
78 | ||
6714d8e8 KH |
79 | /* |
80 | * | |
81 | * spinlock lock ordering: if multiple locks are needed, obey this ordering: | |
82 | * dlm_domain_lock | |
83 | * struct dlm_ctxt->spinlock | |
84 | * struct dlm_lock_resource->spinlock | |
85 | * struct dlm_ctxt->master_lock | |
86 | * struct dlm_ctxt->ast_lock | |
87 | * dlm_master_list_entry->spinlock | |
88 | * dlm_lock->spinlock | |
89 | * | |
90 | */ | |
91 | ||
34af946a | 92 | DEFINE_SPINLOCK(dlm_domain_lock); |
6714d8e8 KH |
93 | LIST_HEAD(dlm_domains); |
94 | static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events); | |
95 | ||
96 | #define DLM_DOMAIN_BACKOFF_MS 200 | |
97 | ||
98 | static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data); | |
99 | static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data); | |
100 | static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data); | |
101 | static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data); | |
102 | ||
103 | static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm); | |
104 | ||
105 | void __dlm_unhash_lockres(struct dlm_lock_resource *lockres) | |
106 | { | |
81f2094a | 107 | hlist_del_init(&lockres->hash_node); |
6714d8e8 KH |
108 | dlm_lockres_put(lockres); |
109 | } | |
110 | ||
111 | void __dlm_insert_lockres(struct dlm_ctxt *dlm, | |
112 | struct dlm_lock_resource *res) | |
113 | { | |
81f2094a | 114 | struct hlist_head *bucket; |
6714d8e8 KH |
115 | struct qstr *q; |
116 | ||
117 | assert_spin_locked(&dlm->spinlock); | |
118 | ||
119 | q = &res->lockname; | |
03d864c0 | 120 | bucket = dlm_lockres_hash(dlm, q->hash); |
6714d8e8 KH |
121 | |
122 | /* get a reference for our hashtable */ | |
123 | dlm_lockres_get(res); | |
124 | ||
81f2094a | 125 | hlist_add_head(&res->hash_node, bucket); |
6714d8e8 KH |
126 | } |
127 | ||
128 | struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm, | |
a3d33291 MF |
129 | const char *name, |
130 | unsigned int len, | |
131 | unsigned int hash) | |
6714d8e8 | 132 | { |
81f2094a | 133 | struct hlist_head *bucket; |
4198985f | 134 | struct hlist_node *list; |
6714d8e8 KH |
135 | |
136 | mlog_entry("%.*s\n", len, name); | |
137 | ||
138 | assert_spin_locked(&dlm->spinlock); | |
139 | ||
03d864c0 DP |
140 | bucket = dlm_lockres_hash(dlm, hash); |
141 | ||
4198985f DP |
142 | hlist_for_each(list, bucket) { |
143 | struct dlm_lock_resource *res = hlist_entry(list, | |
144 | struct dlm_lock_resource, hash_node); | |
145 | if (res->lockname.name[0] != name[0]) | |
146 | continue; | |
147 | if (unlikely(res->lockname.len != len)) | |
148 | continue; | |
149 | if (memcmp(res->lockname.name + 1, name + 1, len - 1)) | |
150 | continue; | |
151 | dlm_lockres_get(res); | |
152 | return res; | |
6714d8e8 | 153 | } |
4198985f | 154 | return NULL; |
6714d8e8 KH |
155 | } |
156 | ||
157 | struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm, | |
158 | const char *name, | |
159 | unsigned int len) | |
160 | { | |
161 | struct dlm_lock_resource *res; | |
a3d33291 | 162 | unsigned int hash = dlm_lockid_hash(name, len); |
6714d8e8 KH |
163 | |
164 | spin_lock(&dlm->spinlock); | |
a3d33291 | 165 | res = __dlm_lookup_lockres(dlm, name, len, hash); |
6714d8e8 KH |
166 | spin_unlock(&dlm->spinlock); |
167 | return res; | |
168 | } | |
169 | ||
170 | static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len) | |
171 | { | |
172 | struct dlm_ctxt *tmp = NULL; | |
173 | struct list_head *iter; | |
174 | ||
175 | assert_spin_locked(&dlm_domain_lock); | |
176 | ||
177 | /* tmp->name here is always NULL terminated, | |
178 | * but domain may not be! */ | |
179 | list_for_each(iter, &dlm_domains) { | |
180 | tmp = list_entry (iter, struct dlm_ctxt, list); | |
181 | if (strlen(tmp->name) == len && | |
182 | memcmp(tmp->name, domain, len)==0) | |
183 | break; | |
184 | tmp = NULL; | |
185 | } | |
186 | ||
187 | return tmp; | |
188 | } | |
189 | ||
190 | /* For null terminated domain strings ONLY */ | |
191 | static struct dlm_ctxt * __dlm_lookup_domain(const char *domain) | |
192 | { | |
193 | assert_spin_locked(&dlm_domain_lock); | |
194 | ||
195 | return __dlm_lookup_domain_full(domain, strlen(domain)); | |
196 | } | |
197 | ||
198 | ||
199 | /* returns true on one of two conditions: | |
200 | * 1) the domain does not exist | |
201 | * 2) the domain exists and it's state is "joined" */ | |
202 | static int dlm_wait_on_domain_helper(const char *domain) | |
203 | { | |
204 | int ret = 0; | |
205 | struct dlm_ctxt *tmp = NULL; | |
206 | ||
207 | spin_lock(&dlm_domain_lock); | |
208 | ||
209 | tmp = __dlm_lookup_domain(domain); | |
210 | if (!tmp) | |
211 | ret = 1; | |
212 | else if (tmp->dlm_state == DLM_CTXT_JOINED) | |
213 | ret = 1; | |
214 | ||
215 | spin_unlock(&dlm_domain_lock); | |
216 | return ret; | |
217 | } | |
218 | ||
219 | static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm) | |
220 | { | |
81f2094a | 221 | if (dlm->lockres_hash) |
03d864c0 | 222 | dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES); |
6714d8e8 KH |
223 | |
224 | if (dlm->name) | |
225 | kfree(dlm->name); | |
226 | ||
227 | kfree(dlm); | |
228 | } | |
229 | ||
230 | /* A little strange - this function will be called while holding | |
231 | * dlm_domain_lock and is expected to be holding it on the way out. We | |
232 | * will however drop and reacquire it multiple times */ | |
233 | static void dlm_ctxt_release(struct kref *kref) | |
234 | { | |
235 | struct dlm_ctxt *dlm; | |
236 | ||
237 | dlm = container_of(kref, struct dlm_ctxt, dlm_refs); | |
238 | ||
239 | BUG_ON(dlm->num_joins); | |
240 | BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED); | |
241 | ||
242 | /* we may still be in the list if we hit an error during join. */ | |
243 | list_del_init(&dlm->list); | |
244 | ||
245 | spin_unlock(&dlm_domain_lock); | |
246 | ||
247 | mlog(0, "freeing memory from domain %s\n", dlm->name); | |
248 | ||
249 | wake_up(&dlm_domain_events); | |
250 | ||
251 | dlm_free_ctxt_mem(dlm); | |
252 | ||
253 | spin_lock(&dlm_domain_lock); | |
254 | } | |
255 | ||
256 | void dlm_put(struct dlm_ctxt *dlm) | |
257 | { | |
258 | spin_lock(&dlm_domain_lock); | |
259 | kref_put(&dlm->dlm_refs, dlm_ctxt_release); | |
260 | spin_unlock(&dlm_domain_lock); | |
261 | } | |
262 | ||
263 | static void __dlm_get(struct dlm_ctxt *dlm) | |
264 | { | |
265 | kref_get(&dlm->dlm_refs); | |
266 | } | |
267 | ||
268 | /* given a questionable reference to a dlm object, gets a reference if | |
269 | * it can find it in the list, otherwise returns NULL in which case | |
270 | * you shouldn't trust your pointer. */ | |
271 | struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm) | |
272 | { | |
273 | struct list_head *iter; | |
274 | struct dlm_ctxt *target = NULL; | |
275 | ||
276 | spin_lock(&dlm_domain_lock); | |
277 | ||
278 | list_for_each(iter, &dlm_domains) { | |
279 | target = list_entry (iter, struct dlm_ctxt, list); | |
280 | ||
281 | if (target == dlm) { | |
282 | __dlm_get(target); | |
283 | break; | |
284 | } | |
285 | ||
286 | target = NULL; | |
287 | } | |
288 | ||
289 | spin_unlock(&dlm_domain_lock); | |
290 | ||
291 | return target; | |
292 | } | |
293 | ||
294 | int dlm_domain_fully_joined(struct dlm_ctxt *dlm) | |
295 | { | |
296 | int ret; | |
297 | ||
298 | spin_lock(&dlm_domain_lock); | |
299 | ret = (dlm->dlm_state == DLM_CTXT_JOINED) || | |
300 | (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN); | |
301 | spin_unlock(&dlm_domain_lock); | |
302 | ||
303 | return ret; | |
304 | } | |
305 | ||
3156d267 KH |
306 | static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm) |
307 | { | |
308 | if (dlm->dlm_worker) { | |
309 | flush_workqueue(dlm->dlm_worker); | |
310 | destroy_workqueue(dlm->dlm_worker); | |
311 | dlm->dlm_worker = NULL; | |
312 | } | |
313 | } | |
314 | ||
6714d8e8 KH |
315 | static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm) |
316 | { | |
317 | dlm_unregister_domain_handlers(dlm); | |
318 | dlm_complete_thread(dlm); | |
319 | dlm_complete_recovery_thread(dlm); | |
3156d267 | 320 | dlm_destroy_dlm_worker(dlm); |
6714d8e8 KH |
321 | |
322 | /* We've left the domain. Now we can take ourselves out of the | |
323 | * list and allow the kref stuff to help us free the | |
324 | * memory. */ | |
325 | spin_lock(&dlm_domain_lock); | |
326 | list_del_init(&dlm->list); | |
327 | spin_unlock(&dlm_domain_lock); | |
328 | ||
329 | /* Wake up anyone waiting for us to remove this domain */ | |
330 | wake_up(&dlm_domain_events); | |
331 | } | |
332 | ||
333 | static void dlm_migrate_all_locks(struct dlm_ctxt *dlm) | |
334 | { | |
335 | int i; | |
336 | struct dlm_lock_resource *res; | |
337 | ||
338 | mlog(0, "Migrating locks from domain %s\n", dlm->name); | |
339 | restart: | |
340 | spin_lock(&dlm->spinlock); | |
81f2094a | 341 | for (i = 0; i < DLM_HASH_BUCKETS; i++) { |
03d864c0 DP |
342 | while (!hlist_empty(dlm_lockres_hash(dlm, i))) { |
343 | res = hlist_entry(dlm_lockres_hash(dlm, i)->first, | |
81f2094a | 344 | struct dlm_lock_resource, hash_node); |
6714d8e8 KH |
345 | /* need reference when manually grabbing lockres */ |
346 | dlm_lockres_get(res); | |
347 | /* this should unhash the lockres | |
348 | * and exit with dlm->spinlock */ | |
349 | mlog(0, "purging res=%p\n", res); | |
350 | if (dlm_lockres_is_dirty(dlm, res)) { | |
351 | /* HACK! this should absolutely go. | |
352 | * need to figure out why some empty | |
353 | * lockreses are still marked dirty */ | |
354 | mlog(ML_ERROR, "lockres %.*s dirty!\n", | |
355 | res->lockname.len, res->lockname.name); | |
356 | ||
357 | spin_unlock(&dlm->spinlock); | |
358 | dlm_kick_thread(dlm, res); | |
359 | wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res)); | |
360 | dlm_lockres_put(res); | |
361 | goto restart; | |
362 | } | |
363 | dlm_purge_lockres(dlm, res); | |
364 | dlm_lockres_put(res); | |
365 | } | |
366 | } | |
367 | spin_unlock(&dlm->spinlock); | |
368 | ||
369 | mlog(0, "DONE Migrating locks from domain %s\n", dlm->name); | |
370 | } | |
371 | ||
372 | static int dlm_no_joining_node(struct dlm_ctxt *dlm) | |
373 | { | |
374 | int ret; | |
375 | ||
376 | spin_lock(&dlm->spinlock); | |
377 | ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN; | |
378 | spin_unlock(&dlm->spinlock); | |
379 | ||
380 | return ret; | |
381 | } | |
382 | ||
383 | static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm) | |
384 | { | |
385 | /* Yikes, a double spinlock! I need domain_lock for the dlm | |
386 | * state and the dlm spinlock for join state... Sorry! */ | |
387 | again: | |
388 | spin_lock(&dlm_domain_lock); | |
389 | spin_lock(&dlm->spinlock); | |
390 | ||
391 | if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) { | |
392 | mlog(0, "Node %d is joining, we wait on it.\n", | |
393 | dlm->joining_node); | |
394 | spin_unlock(&dlm->spinlock); | |
395 | spin_unlock(&dlm_domain_lock); | |
396 | ||
397 | wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm)); | |
398 | goto again; | |
399 | } | |
400 | ||
401 | dlm->dlm_state = DLM_CTXT_LEAVING; | |
402 | spin_unlock(&dlm->spinlock); | |
403 | spin_unlock(&dlm_domain_lock); | |
404 | } | |
405 | ||
406 | static void __dlm_print_nodes(struct dlm_ctxt *dlm) | |
407 | { | |
408 | int node = -1; | |
409 | ||
410 | assert_spin_locked(&dlm->spinlock); | |
411 | ||
781ee3e2 | 412 | printk(KERN_INFO "ocfs2_dlm: Nodes in domain (\"%s\"): ", dlm->name); |
6714d8e8 KH |
413 | |
414 | while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, | |
415 | node + 1)) < O2NM_MAX_NODES) { | |
781ee3e2 | 416 | printk("%d ", node); |
6714d8e8 | 417 | } |
781ee3e2 | 418 | printk("\n"); |
6714d8e8 KH |
419 | } |
420 | ||
421 | static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data) | |
422 | { | |
423 | struct dlm_ctxt *dlm = data; | |
424 | unsigned int node; | |
425 | struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf; | |
426 | ||
427 | mlog_entry("%p %u %p", msg, len, data); | |
428 | ||
429 | if (!dlm_grab(dlm)) | |
430 | return 0; | |
431 | ||
432 | node = exit_msg->node_idx; | |
433 | ||
781ee3e2 | 434 | printk(KERN_INFO "ocfs2_dlm: Node %u leaves domain %s\n", node, dlm->name); |
6714d8e8 KH |
435 | |
436 | spin_lock(&dlm->spinlock); | |
437 | clear_bit(node, dlm->domain_map); | |
438 | __dlm_print_nodes(dlm); | |
439 | ||
440 | /* notify anything attached to the heartbeat events */ | |
441 | dlm_hb_event_notify_attached(dlm, node, 0); | |
442 | ||
443 | spin_unlock(&dlm->spinlock); | |
444 | ||
445 | dlm_put(dlm); | |
446 | ||
447 | return 0; | |
448 | } | |
449 | ||
450 | static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm, | |
451 | unsigned int node) | |
452 | { | |
453 | int status; | |
454 | struct dlm_exit_domain leave_msg; | |
455 | ||
456 | mlog(0, "Asking node %u if we can leave the domain %s me = %u\n", | |
457 | node, dlm->name, dlm->node_num); | |
458 | ||
459 | memset(&leave_msg, 0, sizeof(leave_msg)); | |
460 | leave_msg.node_idx = dlm->node_num; | |
461 | ||
462 | status = o2net_send_message(DLM_EXIT_DOMAIN_MSG, dlm->key, | |
463 | &leave_msg, sizeof(leave_msg), node, | |
464 | NULL); | |
465 | ||
466 | mlog(0, "status return %d from o2net_send_message\n", status); | |
467 | ||
468 | return status; | |
469 | } | |
470 | ||
471 | ||
472 | static void dlm_leave_domain(struct dlm_ctxt *dlm) | |
473 | { | |
474 | int node, clear_node, status; | |
475 | ||
476 | /* At this point we've migrated away all our locks and won't | |
477 | * accept mastership of new ones. The dlm is responsible for | |
478 | * almost nothing now. We make sure not to confuse any joining | |
479 | * nodes and then commence shutdown procedure. */ | |
480 | ||
481 | spin_lock(&dlm->spinlock); | |
482 | /* Clear ourselves from the domain map */ | |
483 | clear_bit(dlm->node_num, dlm->domain_map); | |
484 | while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, | |
485 | 0)) < O2NM_MAX_NODES) { | |
486 | /* Drop the dlm spinlock. This is safe wrt the domain_map. | |
487 | * -nodes cannot be added now as the | |
488 | * query_join_handlers knows to respond with OK_NO_MAP | |
489 | * -we catch the right network errors if a node is | |
490 | * removed from the map while we're sending him the | |
491 | * exit message. */ | |
492 | spin_unlock(&dlm->spinlock); | |
493 | ||
494 | clear_node = 1; | |
495 | ||
496 | status = dlm_send_one_domain_exit(dlm, node); | |
497 | if (status < 0 && | |
498 | status != -ENOPROTOOPT && | |
499 | status != -ENOTCONN) { | |
500 | mlog(ML_NOTICE, "Error %d sending domain exit message " | |
501 | "to node %d\n", status, node); | |
502 | ||
503 | /* Not sure what to do here but lets sleep for | |
504 | * a bit in case this was a transient | |
505 | * error... */ | |
506 | msleep(DLM_DOMAIN_BACKOFF_MS); | |
507 | clear_node = 0; | |
508 | } | |
509 | ||
510 | spin_lock(&dlm->spinlock); | |
511 | /* If we're not clearing the node bit then we intend | |
512 | * to loop back around to try again. */ | |
513 | if (clear_node) | |
514 | clear_bit(node, dlm->domain_map); | |
515 | } | |
516 | spin_unlock(&dlm->spinlock); | |
517 | } | |
518 | ||
519 | int dlm_joined(struct dlm_ctxt *dlm) | |
520 | { | |
521 | int ret = 0; | |
522 | ||
523 | spin_lock(&dlm_domain_lock); | |
524 | ||
525 | if (dlm->dlm_state == DLM_CTXT_JOINED) | |
526 | ret = 1; | |
527 | ||
528 | spin_unlock(&dlm_domain_lock); | |
529 | ||
530 | return ret; | |
531 | } | |
532 | ||
533 | int dlm_shutting_down(struct dlm_ctxt *dlm) | |
534 | { | |
535 | int ret = 0; | |
536 | ||
537 | spin_lock(&dlm_domain_lock); | |
538 | ||
539 | if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN) | |
540 | ret = 1; | |
541 | ||
542 | spin_unlock(&dlm_domain_lock); | |
543 | ||
544 | return ret; | |
545 | } | |
546 | ||
547 | void dlm_unregister_domain(struct dlm_ctxt *dlm) | |
548 | { | |
549 | int leave = 0; | |
550 | ||
551 | spin_lock(&dlm_domain_lock); | |
552 | BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED); | |
553 | BUG_ON(!dlm->num_joins); | |
554 | ||
555 | dlm->num_joins--; | |
556 | if (!dlm->num_joins) { | |
557 | /* We mark it "in shutdown" now so new register | |
558 | * requests wait until we've completely left the | |
559 | * domain. Don't use DLM_CTXT_LEAVING yet as we still | |
560 | * want new domain joins to communicate with us at | |
561 | * least until we've completed migration of our | |
562 | * resources. */ | |
563 | dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN; | |
564 | leave = 1; | |
565 | } | |
566 | spin_unlock(&dlm_domain_lock); | |
567 | ||
568 | if (leave) { | |
569 | mlog(0, "shutting down domain %s\n", dlm->name); | |
570 | ||
571 | /* We changed dlm state, notify the thread */ | |
572 | dlm_kick_thread(dlm, NULL); | |
573 | ||
574 | dlm_migrate_all_locks(dlm); | |
575 | dlm_mark_domain_leaving(dlm); | |
576 | dlm_leave_domain(dlm); | |
577 | dlm_complete_dlm_shutdown(dlm); | |
578 | } | |
579 | dlm_put(dlm); | |
580 | } | |
581 | EXPORT_SYMBOL_GPL(dlm_unregister_domain); | |
582 | ||
583 | static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data) | |
584 | { | |
585 | struct dlm_query_join_request *query; | |
586 | enum dlm_query_join_response response; | |
587 | struct dlm_ctxt *dlm = NULL; | |
588 | ||
589 | query = (struct dlm_query_join_request *) msg->buf; | |
590 | ||
591 | mlog(0, "node %u wants to join domain %s\n", query->node_idx, | |
592 | query->domain); | |
593 | ||
594 | /* | |
595 | * If heartbeat doesn't consider the node live, tell it | |
596 | * to back off and try again. This gives heartbeat a chance | |
597 | * to catch up. | |
598 | */ | |
599 | if (!o2hb_check_node_heartbeating(query->node_idx)) { | |
600 | mlog(0, "node %u is not in our live map yet\n", | |
601 | query->node_idx); | |
602 | ||
603 | response = JOIN_DISALLOW; | |
604 | goto respond; | |
605 | } | |
606 | ||
607 | response = JOIN_OK_NO_MAP; | |
608 | ||
609 | spin_lock(&dlm_domain_lock); | |
610 | dlm = __dlm_lookup_domain_full(query->domain, query->name_len); | |
611 | /* Once the dlm ctxt is marked as leaving then we don't want | |
e2faea4c KH |
612 | * to be put in someone's domain map. |
613 | * Also, explicitly disallow joining at certain troublesome | |
614 | * times (ie. during recovery). */ | |
6714d8e8 | 615 | if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) { |
e2faea4c | 616 | int bit = query->node_idx; |
6714d8e8 KH |
617 | spin_lock(&dlm->spinlock); |
618 | ||
619 | if (dlm->dlm_state == DLM_CTXT_NEW && | |
620 | dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) { | |
621 | /*If this is a brand new context and we | |
622 | * haven't started our join process yet, then | |
623 | * the other node won the race. */ | |
624 | response = JOIN_OK_NO_MAP; | |
625 | } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) { | |
626 | /* Disallow parallel joins. */ | |
627 | response = JOIN_DISALLOW; | |
e2faea4c KH |
628 | } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) { |
629 | mlog(ML_NOTICE, "node %u trying to join, but recovery " | |
630 | "is ongoing.\n", bit); | |
631 | response = JOIN_DISALLOW; | |
632 | } else if (test_bit(bit, dlm->recovery_map)) { | |
633 | mlog(ML_NOTICE, "node %u trying to join, but it " | |
634 | "still needs recovery.\n", bit); | |
635 | response = JOIN_DISALLOW; | |
636 | } else if (test_bit(bit, dlm->domain_map)) { | |
637 | mlog(ML_NOTICE, "node %u trying to join, but it " | |
638 | "is still in the domain! needs recovery?\n", | |
639 | bit); | |
640 | response = JOIN_DISALLOW; | |
6714d8e8 KH |
641 | } else { |
642 | /* Alright we're fully a part of this domain | |
643 | * so we keep some state as to who's joining | |
644 | * and indicate to him that needs to be fixed | |
645 | * up. */ | |
646 | response = JOIN_OK; | |
647 | __dlm_set_joining_node(dlm, query->node_idx); | |
648 | } | |
649 | ||
650 | spin_unlock(&dlm->spinlock); | |
651 | } | |
652 | spin_unlock(&dlm_domain_lock); | |
653 | ||
654 | respond: | |
655 | mlog(0, "We respond with %u\n", response); | |
656 | ||
657 | return response; | |
658 | } | |
659 | ||
660 | static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data) | |
661 | { | |
662 | struct dlm_assert_joined *assert; | |
663 | struct dlm_ctxt *dlm = NULL; | |
664 | ||
665 | assert = (struct dlm_assert_joined *) msg->buf; | |
666 | ||
667 | mlog(0, "node %u asserts join on domain %s\n", assert->node_idx, | |
668 | assert->domain); | |
669 | ||
670 | spin_lock(&dlm_domain_lock); | |
671 | dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len); | |
672 | /* XXX should we consider no dlm ctxt an error? */ | |
673 | if (dlm) { | |
674 | spin_lock(&dlm->spinlock); | |
675 | ||
676 | /* Alright, this node has officially joined our | |
677 | * domain. Set him in the map and clean up our | |
678 | * leftover join state. */ | |
679 | BUG_ON(dlm->joining_node != assert->node_idx); | |
680 | set_bit(assert->node_idx, dlm->domain_map); | |
681 | __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); | |
682 | ||
781ee3e2 SM |
683 | printk(KERN_INFO "ocfs2_dlm: Node %u joins domain %s\n", |
684 | assert->node_idx, dlm->name); | |
6714d8e8 KH |
685 | __dlm_print_nodes(dlm); |
686 | ||
687 | /* notify anything attached to the heartbeat events */ | |
688 | dlm_hb_event_notify_attached(dlm, assert->node_idx, 1); | |
689 | ||
690 | spin_unlock(&dlm->spinlock); | |
691 | } | |
692 | spin_unlock(&dlm_domain_lock); | |
693 | ||
694 | return 0; | |
695 | } | |
696 | ||
697 | static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data) | |
698 | { | |
699 | struct dlm_cancel_join *cancel; | |
700 | struct dlm_ctxt *dlm = NULL; | |
701 | ||
702 | cancel = (struct dlm_cancel_join *) msg->buf; | |
703 | ||
704 | mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx, | |
705 | cancel->domain); | |
706 | ||
707 | spin_lock(&dlm_domain_lock); | |
708 | dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len); | |
709 | ||
710 | if (dlm) { | |
711 | spin_lock(&dlm->spinlock); | |
712 | ||
713 | /* Yikes, this guy wants to cancel his join. No | |
714 | * problem, we simply cleanup our join state. */ | |
715 | BUG_ON(dlm->joining_node != cancel->node_idx); | |
716 | __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); | |
717 | ||
718 | spin_unlock(&dlm->spinlock); | |
719 | } | |
720 | spin_unlock(&dlm_domain_lock); | |
721 | ||
722 | return 0; | |
723 | } | |
724 | ||
725 | static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm, | |
726 | unsigned int node) | |
727 | { | |
728 | int status; | |
729 | struct dlm_cancel_join cancel_msg; | |
730 | ||
731 | memset(&cancel_msg, 0, sizeof(cancel_msg)); | |
732 | cancel_msg.node_idx = dlm->node_num; | |
733 | cancel_msg.name_len = strlen(dlm->name); | |
734 | memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len); | |
735 | ||
736 | status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY, | |
737 | &cancel_msg, sizeof(cancel_msg), node, | |
738 | NULL); | |
739 | if (status < 0) { | |
740 | mlog_errno(status); | |
741 | goto bail; | |
742 | } | |
743 | ||
744 | bail: | |
745 | return status; | |
746 | } | |
747 | ||
748 | /* map_size should be in bytes. */ | |
749 | static int dlm_send_join_cancels(struct dlm_ctxt *dlm, | |
750 | unsigned long *node_map, | |
751 | unsigned int map_size) | |
752 | { | |
753 | int status, tmpstat; | |
754 | unsigned int node; | |
755 | ||
756 | if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) * | |
757 | sizeof(unsigned long))) { | |
758 | mlog(ML_ERROR, | |
759 | "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n", | |
760 | map_size, BITS_TO_LONGS(O2NM_MAX_NODES)); | |
761 | return -EINVAL; | |
762 | } | |
763 | ||
764 | status = 0; | |
765 | node = -1; | |
766 | while ((node = find_next_bit(node_map, O2NM_MAX_NODES, | |
767 | node + 1)) < O2NM_MAX_NODES) { | |
768 | if (node == dlm->node_num) | |
769 | continue; | |
770 | ||
771 | tmpstat = dlm_send_one_join_cancel(dlm, node); | |
772 | if (tmpstat) { | |
773 | mlog(ML_ERROR, "Error return %d cancelling join on " | |
774 | "node %d\n", tmpstat, node); | |
775 | if (!status) | |
776 | status = tmpstat; | |
777 | } | |
778 | } | |
779 | ||
780 | if (status) | |
781 | mlog_errno(status); | |
782 | return status; | |
783 | } | |
784 | ||
785 | static int dlm_request_join(struct dlm_ctxt *dlm, | |
786 | int node, | |
787 | enum dlm_query_join_response *response) | |
788 | { | |
789 | int status, retval; | |
790 | struct dlm_query_join_request join_msg; | |
791 | ||
792 | mlog(0, "querying node %d\n", node); | |
793 | ||
794 | memset(&join_msg, 0, sizeof(join_msg)); | |
795 | join_msg.node_idx = dlm->node_num; | |
796 | join_msg.name_len = strlen(dlm->name); | |
797 | memcpy(join_msg.domain, dlm->name, join_msg.name_len); | |
798 | ||
799 | status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg, | |
800 | sizeof(join_msg), node, &retval); | |
801 | if (status < 0 && status != -ENOPROTOOPT) { | |
802 | mlog_errno(status); | |
803 | goto bail; | |
804 | } | |
805 | ||
806 | /* -ENOPROTOOPT from the net code means the other side isn't | |
807 | listening for our message type -- that's fine, it means | |
808 | his dlm isn't up, so we can consider him a 'yes' but not | |
809 | joined into the domain. */ | |
810 | if (status == -ENOPROTOOPT) { | |
811 | status = 0; | |
812 | *response = JOIN_OK_NO_MAP; | |
813 | } else if (retval == JOIN_DISALLOW || | |
814 | retval == JOIN_OK || | |
815 | retval == JOIN_OK_NO_MAP) { | |
816 | *response = retval; | |
817 | } else { | |
818 | status = -EINVAL; | |
819 | mlog(ML_ERROR, "invalid response %d from node %u\n", retval, | |
820 | node); | |
821 | } | |
822 | ||
823 | mlog(0, "status %d, node %d response is %d\n", status, node, | |
824 | *response); | |
825 | ||
826 | bail: | |
827 | return status; | |
828 | } | |
829 | ||
830 | static int dlm_send_one_join_assert(struct dlm_ctxt *dlm, | |
831 | unsigned int node) | |
832 | { | |
833 | int status; | |
834 | struct dlm_assert_joined assert_msg; | |
835 | ||
836 | mlog(0, "Sending join assert to node %u\n", node); | |
837 | ||
838 | memset(&assert_msg, 0, sizeof(assert_msg)); | |
839 | assert_msg.node_idx = dlm->node_num; | |
840 | assert_msg.name_len = strlen(dlm->name); | |
841 | memcpy(assert_msg.domain, dlm->name, assert_msg.name_len); | |
842 | ||
843 | status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY, | |
844 | &assert_msg, sizeof(assert_msg), node, | |
845 | NULL); | |
846 | if (status < 0) | |
847 | mlog_errno(status); | |
848 | ||
849 | return status; | |
850 | } | |
851 | ||
852 | static void dlm_send_join_asserts(struct dlm_ctxt *dlm, | |
853 | unsigned long *node_map) | |
854 | { | |
855 | int status, node, live; | |
856 | ||
857 | status = 0; | |
858 | node = -1; | |
859 | while ((node = find_next_bit(node_map, O2NM_MAX_NODES, | |
860 | node + 1)) < O2NM_MAX_NODES) { | |
861 | if (node == dlm->node_num) | |
862 | continue; | |
863 | ||
864 | do { | |
865 | /* It is very important that this message be | |
866 | * received so we spin until either the node | |
867 | * has died or it gets the message. */ | |
868 | status = dlm_send_one_join_assert(dlm, node); | |
869 | ||
870 | spin_lock(&dlm->spinlock); | |
871 | live = test_bit(node, dlm->live_nodes_map); | |
872 | spin_unlock(&dlm->spinlock); | |
873 | ||
874 | if (status) { | |
875 | mlog(ML_ERROR, "Error return %d asserting " | |
876 | "join on node %d\n", status, node); | |
877 | ||
878 | /* give us some time between errors... */ | |
879 | if (live) | |
880 | msleep(DLM_DOMAIN_BACKOFF_MS); | |
881 | } | |
882 | } while (status && live); | |
883 | } | |
884 | } | |
885 | ||
886 | struct domain_join_ctxt { | |
887 | unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; | |
888 | unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; | |
889 | }; | |
890 | ||
891 | static int dlm_should_restart_join(struct dlm_ctxt *dlm, | |
892 | struct domain_join_ctxt *ctxt, | |
893 | enum dlm_query_join_response response) | |
894 | { | |
895 | int ret; | |
896 | ||
897 | if (response == JOIN_DISALLOW) { | |
898 | mlog(0, "Latest response of disallow -- should restart\n"); | |
899 | return 1; | |
900 | } | |
901 | ||
902 | spin_lock(&dlm->spinlock); | |
903 | /* For now, we restart the process if the node maps have | |
904 | * changed at all */ | |
905 | ret = memcmp(ctxt->live_map, dlm->live_nodes_map, | |
906 | sizeof(dlm->live_nodes_map)); | |
907 | spin_unlock(&dlm->spinlock); | |
908 | ||
909 | if (ret) | |
910 | mlog(0, "Node maps changed -- should restart\n"); | |
911 | ||
912 | return ret; | |
913 | } | |
914 | ||
915 | static int dlm_try_to_join_domain(struct dlm_ctxt *dlm) | |
916 | { | |
917 | int status = 0, tmpstat, node; | |
918 | struct domain_join_ctxt *ctxt; | |
919 | enum dlm_query_join_response response; | |
920 | ||
921 | mlog_entry("%p", dlm); | |
922 | ||
cd861280 | 923 | ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL); |
6714d8e8 KH |
924 | if (!ctxt) { |
925 | status = -ENOMEM; | |
926 | mlog_errno(status); | |
927 | goto bail; | |
928 | } | |
929 | ||
930 | /* group sem locking should work for us here -- we're already | |
931 | * registered for heartbeat events so filling this should be | |
932 | * atomic wrt getting those handlers called. */ | |
933 | o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map)); | |
934 | ||
935 | spin_lock(&dlm->spinlock); | |
936 | memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map)); | |
937 | ||
938 | __dlm_set_joining_node(dlm, dlm->node_num); | |
939 | ||
940 | spin_unlock(&dlm->spinlock); | |
941 | ||
942 | node = -1; | |
943 | while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES, | |
944 | node + 1)) < O2NM_MAX_NODES) { | |
945 | if (node == dlm->node_num) | |
946 | continue; | |
947 | ||
948 | status = dlm_request_join(dlm, node, &response); | |
949 | if (status < 0) { | |
950 | mlog_errno(status); | |
951 | goto bail; | |
952 | } | |
953 | ||
954 | /* Ok, either we got a response or the node doesn't have a | |
955 | * dlm up. */ | |
956 | if (response == JOIN_OK) | |
957 | set_bit(node, ctxt->yes_resp_map); | |
958 | ||
959 | if (dlm_should_restart_join(dlm, ctxt, response)) { | |
960 | status = -EAGAIN; | |
961 | goto bail; | |
962 | } | |
963 | } | |
964 | ||
965 | mlog(0, "Yay, done querying nodes!\n"); | |
966 | ||
967 | /* Yay, everyone agree's we can join the domain. My domain is | |
968 | * comprised of all nodes who were put in the | |
969 | * yes_resp_map. Copy that into our domain map and send a join | |
970 | * assert message to clean up everyone elses state. */ | |
971 | spin_lock(&dlm->spinlock); | |
972 | memcpy(dlm->domain_map, ctxt->yes_resp_map, | |
973 | sizeof(ctxt->yes_resp_map)); | |
974 | set_bit(dlm->node_num, dlm->domain_map); | |
975 | spin_unlock(&dlm->spinlock); | |
976 | ||
977 | dlm_send_join_asserts(dlm, ctxt->yes_resp_map); | |
978 | ||
979 | /* Joined state *must* be set before the joining node | |
980 | * information, otherwise the query_join handler may read no | |
981 | * current joiner but a state of NEW and tell joining nodes | |
982 | * we're not in the domain. */ | |
983 | spin_lock(&dlm_domain_lock); | |
984 | dlm->dlm_state = DLM_CTXT_JOINED; | |
985 | dlm->num_joins++; | |
986 | spin_unlock(&dlm_domain_lock); | |
987 | ||
988 | bail: | |
989 | spin_lock(&dlm->spinlock); | |
990 | __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); | |
991 | if (!status) | |
992 | __dlm_print_nodes(dlm); | |
993 | spin_unlock(&dlm->spinlock); | |
994 | ||
995 | if (ctxt) { | |
996 | /* Do we need to send a cancel message to any nodes? */ | |
997 | if (status < 0) { | |
998 | tmpstat = dlm_send_join_cancels(dlm, | |
999 | ctxt->yes_resp_map, | |
1000 | sizeof(ctxt->yes_resp_map)); | |
1001 | if (tmpstat < 0) | |
1002 | mlog_errno(tmpstat); | |
1003 | } | |
1004 | kfree(ctxt); | |
1005 | } | |
1006 | ||
1007 | mlog(0, "returning %d\n", status); | |
1008 | return status; | |
1009 | } | |
1010 | ||
1011 | static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm) | |
1012 | { | |
1013 | o2hb_unregister_callback(&dlm->dlm_hb_up); | |
1014 | o2hb_unregister_callback(&dlm->dlm_hb_down); | |
1015 | o2net_unregister_handler_list(&dlm->dlm_domain_handlers); | |
1016 | } | |
1017 | ||
1018 | static int dlm_register_domain_handlers(struct dlm_ctxt *dlm) | |
1019 | { | |
1020 | int status; | |
1021 | ||
1022 | mlog(0, "registering handlers.\n"); | |
1023 | ||
1024 | o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB, | |
1025 | dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI); | |
1026 | status = o2hb_register_callback(&dlm->dlm_hb_down); | |
1027 | if (status) | |
1028 | goto bail; | |
1029 | ||
1030 | o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB, | |
1031 | dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI); | |
1032 | status = o2hb_register_callback(&dlm->dlm_hb_up); | |
1033 | if (status) | |
1034 | goto bail; | |
1035 | ||
1036 | status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key, | |
1037 | sizeof(struct dlm_master_request), | |
1038 | dlm_master_request_handler, | |
1039 | dlm, &dlm->dlm_domain_handlers); | |
1040 | if (status) | |
1041 | goto bail; | |
1042 | ||
1043 | status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key, | |
1044 | sizeof(struct dlm_assert_master), | |
1045 | dlm_assert_master_handler, | |
1046 | dlm, &dlm->dlm_domain_handlers); | |
1047 | if (status) | |
1048 | goto bail; | |
1049 | ||
1050 | status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key, | |
1051 | sizeof(struct dlm_create_lock), | |
1052 | dlm_create_lock_handler, | |
1053 | dlm, &dlm->dlm_domain_handlers); | |
1054 | if (status) | |
1055 | goto bail; | |
1056 | ||
1057 | status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key, | |
1058 | DLM_CONVERT_LOCK_MAX_LEN, | |
1059 | dlm_convert_lock_handler, | |
1060 | dlm, &dlm->dlm_domain_handlers); | |
1061 | if (status) | |
1062 | goto bail; | |
1063 | ||
1064 | status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key, | |
1065 | DLM_UNLOCK_LOCK_MAX_LEN, | |
1066 | dlm_unlock_lock_handler, | |
1067 | dlm, &dlm->dlm_domain_handlers); | |
1068 | if (status) | |
1069 | goto bail; | |
1070 | ||
1071 | status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key, | |
1072 | DLM_PROXY_AST_MAX_LEN, | |
1073 | dlm_proxy_ast_handler, | |
1074 | dlm, &dlm->dlm_domain_handlers); | |
1075 | if (status) | |
1076 | goto bail; | |
1077 | ||
1078 | status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key, | |
1079 | sizeof(struct dlm_exit_domain), | |
1080 | dlm_exit_domain_handler, | |
1081 | dlm, &dlm->dlm_domain_handlers); | |
1082 | if (status) | |
1083 | goto bail; | |
1084 | ||
1085 | status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key, | |
1086 | sizeof(struct dlm_migrate_request), | |
1087 | dlm_migrate_request_handler, | |
1088 | dlm, &dlm->dlm_domain_handlers); | |
1089 | if (status) | |
1090 | goto bail; | |
1091 | ||
1092 | status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key, | |
1093 | DLM_MIG_LOCKRES_MAX_LEN, | |
1094 | dlm_mig_lockres_handler, | |
1095 | dlm, &dlm->dlm_domain_handlers); | |
1096 | if (status) | |
1097 | goto bail; | |
1098 | ||
1099 | status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key, | |
1100 | sizeof(struct dlm_master_requery), | |
1101 | dlm_master_requery_handler, | |
1102 | dlm, &dlm->dlm_domain_handlers); | |
1103 | if (status) | |
1104 | goto bail; | |
1105 | ||
1106 | status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key, | |
1107 | sizeof(struct dlm_lock_request), | |
1108 | dlm_request_all_locks_handler, | |
1109 | dlm, &dlm->dlm_domain_handlers); | |
1110 | if (status) | |
1111 | goto bail; | |
1112 | ||
1113 | status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key, | |
1114 | sizeof(struct dlm_reco_data_done), | |
1115 | dlm_reco_data_done_handler, | |
1116 | dlm, &dlm->dlm_domain_handlers); | |
1117 | if (status) | |
1118 | goto bail; | |
1119 | ||
1120 | status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key, | |
1121 | sizeof(struct dlm_begin_reco), | |
1122 | dlm_begin_reco_handler, | |
1123 | dlm, &dlm->dlm_domain_handlers); | |
1124 | if (status) | |
1125 | goto bail; | |
1126 | ||
1127 | status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key, | |
1128 | sizeof(struct dlm_finalize_reco), | |
1129 | dlm_finalize_reco_handler, | |
1130 | dlm, &dlm->dlm_domain_handlers); | |
1131 | if (status) | |
1132 | goto bail; | |
1133 | ||
1134 | bail: | |
1135 | if (status) | |
1136 | dlm_unregister_domain_handlers(dlm); | |
1137 | ||
1138 | return status; | |
1139 | } | |
1140 | ||
1141 | static int dlm_join_domain(struct dlm_ctxt *dlm) | |
1142 | { | |
1143 | int status; | |
1144 | ||
1145 | BUG_ON(!dlm); | |
1146 | ||
1147 | mlog(0, "Join domain %s\n", dlm->name); | |
1148 | ||
1149 | status = dlm_register_domain_handlers(dlm); | |
1150 | if (status) { | |
1151 | mlog_errno(status); | |
1152 | goto bail; | |
1153 | } | |
1154 | ||
1155 | status = dlm_launch_thread(dlm); | |
1156 | if (status < 0) { | |
1157 | mlog_errno(status); | |
1158 | goto bail; | |
1159 | } | |
1160 | ||
1161 | status = dlm_launch_recovery_thread(dlm); | |
1162 | if (status < 0) { | |
1163 | mlog_errno(status); | |
1164 | goto bail; | |
1165 | } | |
1166 | ||
3156d267 KH |
1167 | dlm->dlm_worker = create_singlethread_workqueue("dlm_wq"); |
1168 | if (!dlm->dlm_worker) { | |
1169 | status = -ENOMEM; | |
1170 | mlog_errno(status); | |
1171 | goto bail; | |
1172 | } | |
1173 | ||
6714d8e8 KH |
1174 | do { |
1175 | unsigned int backoff; | |
1176 | status = dlm_try_to_join_domain(dlm); | |
1177 | ||
1178 | /* If we're racing another node to the join, then we | |
1179 | * need to back off temporarily and let them | |
1180 | * complete. */ | |
1181 | if (status == -EAGAIN) { | |
1182 | if (signal_pending(current)) { | |
1183 | status = -ERESTARTSYS; | |
1184 | goto bail; | |
1185 | } | |
1186 | ||
1187 | /* | |
1188 | * <chip> After you! | |
1189 | * <dale> No, after you! | |
1190 | * <chip> I insist! | |
1191 | * <dale> But you first! | |
1192 | * ... | |
1193 | */ | |
1194 | backoff = (unsigned int)(jiffies & 0x3); | |
1195 | backoff *= DLM_DOMAIN_BACKOFF_MS; | |
1196 | mlog(0, "backoff %d\n", backoff); | |
1197 | msleep(backoff); | |
1198 | } | |
1199 | } while (status == -EAGAIN); | |
1200 | ||
1201 | if (status < 0) { | |
1202 | mlog_errno(status); | |
1203 | goto bail; | |
1204 | } | |
1205 | ||
1206 | status = 0; | |
1207 | bail: | |
1208 | wake_up(&dlm_domain_events); | |
1209 | ||
1210 | if (status) { | |
1211 | dlm_unregister_domain_handlers(dlm); | |
1212 | dlm_complete_thread(dlm); | |
1213 | dlm_complete_recovery_thread(dlm); | |
3156d267 | 1214 | dlm_destroy_dlm_worker(dlm); |
6714d8e8 KH |
1215 | } |
1216 | ||
1217 | return status; | |
1218 | } | |
1219 | ||
1220 | static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain, | |
1221 | u32 key) | |
1222 | { | |
1223 | int i; | |
1224 | struct dlm_ctxt *dlm = NULL; | |
1225 | ||
cd861280 | 1226 | dlm = kzalloc(sizeof(*dlm), GFP_KERNEL); |
6714d8e8 KH |
1227 | if (!dlm) { |
1228 | mlog_errno(-ENOMEM); | |
1229 | goto leave; | |
1230 | } | |
1231 | ||
1232 | dlm->name = kmalloc(strlen(domain) + 1, GFP_KERNEL); | |
1233 | if (dlm->name == NULL) { | |
1234 | mlog_errno(-ENOMEM); | |
1235 | kfree(dlm); | |
1236 | dlm = NULL; | |
1237 | goto leave; | |
1238 | } | |
1239 | ||
03d864c0 | 1240 | dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES); |
81f2094a | 1241 | if (!dlm->lockres_hash) { |
6714d8e8 KH |
1242 | mlog_errno(-ENOMEM); |
1243 | kfree(dlm->name); | |
1244 | kfree(dlm); | |
1245 | dlm = NULL; | |
1246 | goto leave; | |
1247 | } | |
6714d8e8 | 1248 | |
03d864c0 DP |
1249 | for (i = 0; i < DLM_HASH_BUCKETS; i++) |
1250 | INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i)); | |
6714d8e8 KH |
1251 | |
1252 | strcpy(dlm->name, domain); | |
1253 | dlm->key = key; | |
1254 | dlm->node_num = o2nm_this_node(); | |
1255 | ||
1256 | spin_lock_init(&dlm->spinlock); | |
1257 | spin_lock_init(&dlm->master_lock); | |
1258 | spin_lock_init(&dlm->ast_lock); | |
1259 | INIT_LIST_HEAD(&dlm->list); | |
1260 | INIT_LIST_HEAD(&dlm->dirty_list); | |
1261 | INIT_LIST_HEAD(&dlm->reco.resources); | |
1262 | INIT_LIST_HEAD(&dlm->reco.received); | |
1263 | INIT_LIST_HEAD(&dlm->reco.node_data); | |
1264 | INIT_LIST_HEAD(&dlm->purge_list); | |
1265 | INIT_LIST_HEAD(&dlm->dlm_domain_handlers); | |
1266 | dlm->reco.state = 0; | |
1267 | ||
1268 | INIT_LIST_HEAD(&dlm->pending_asts); | |
1269 | INIT_LIST_HEAD(&dlm->pending_basts); | |
1270 | ||
1271 | mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n", | |
1272 | dlm->recovery_map, &(dlm->recovery_map[0])); | |
1273 | ||
1274 | memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map)); | |
1275 | memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map)); | |
1276 | memset(dlm->domain_map, 0, sizeof(dlm->domain_map)); | |
1277 | ||
1278 | dlm->dlm_thread_task = NULL; | |
1279 | dlm->dlm_reco_thread_task = NULL; | |
3156d267 | 1280 | dlm->dlm_worker = NULL; |
6714d8e8 KH |
1281 | init_waitqueue_head(&dlm->dlm_thread_wq); |
1282 | init_waitqueue_head(&dlm->dlm_reco_thread_wq); | |
1283 | init_waitqueue_head(&dlm->reco.event); | |
1284 | init_waitqueue_head(&dlm->ast_wq); | |
1285 | init_waitqueue_head(&dlm->migration_wq); | |
1286 | INIT_LIST_HEAD(&dlm->master_list); | |
1287 | INIT_LIST_HEAD(&dlm->mle_hb_events); | |
1288 | ||
1289 | dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN; | |
1290 | init_waitqueue_head(&dlm->dlm_join_events); | |
1291 | ||
1292 | dlm->reco.new_master = O2NM_INVALID_NODE_NUM; | |
1293 | dlm->reco.dead_node = O2NM_INVALID_NODE_NUM; | |
1294 | atomic_set(&dlm->local_resources, 0); | |
1295 | atomic_set(&dlm->remote_resources, 0); | |
1296 | atomic_set(&dlm->unknown_resources, 0); | |
1297 | ||
1298 | spin_lock_init(&dlm->work_lock); | |
1299 | INIT_LIST_HEAD(&dlm->work_list); | |
c4028958 | 1300 | INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work); |
6714d8e8 KH |
1301 | |
1302 | kref_init(&dlm->dlm_refs); | |
1303 | dlm->dlm_state = DLM_CTXT_NEW; | |
1304 | ||
1305 | INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks); | |
1306 | ||
1307 | mlog(0, "context init: refcount %u\n", | |
1308 | atomic_read(&dlm->dlm_refs.refcount)); | |
1309 | ||
1310 | leave: | |
1311 | return dlm; | |
1312 | } | |
1313 | ||
1314 | /* | |
1315 | * dlm_register_domain: one-time setup per "domain" | |
1316 | */ | |
1317 | struct dlm_ctxt * dlm_register_domain(const char *domain, | |
1318 | u32 key) | |
1319 | { | |
1320 | int ret; | |
1321 | struct dlm_ctxt *dlm = NULL; | |
1322 | struct dlm_ctxt *new_ctxt = NULL; | |
1323 | ||
1324 | if (strlen(domain) > O2NM_MAX_NAME_LEN) { | |
1325 | ret = -ENAMETOOLONG; | |
1326 | mlog(ML_ERROR, "domain name length too long\n"); | |
1327 | goto leave; | |
1328 | } | |
1329 | ||
1330 | if (!o2hb_check_local_node_heartbeating()) { | |
1331 | mlog(ML_ERROR, "the local node has not been configured, or is " | |
1332 | "not heartbeating\n"); | |
1333 | ret = -EPROTO; | |
1334 | goto leave; | |
1335 | } | |
1336 | ||
1337 | mlog(0, "register called for domain \"%s\"\n", domain); | |
1338 | ||
1339 | retry: | |
1340 | dlm = NULL; | |
1341 | if (signal_pending(current)) { | |
1342 | ret = -ERESTARTSYS; | |
1343 | mlog_errno(ret); | |
1344 | goto leave; | |
1345 | } | |
1346 | ||
1347 | spin_lock(&dlm_domain_lock); | |
1348 | ||
1349 | dlm = __dlm_lookup_domain(domain); | |
1350 | if (dlm) { | |
1351 | if (dlm->dlm_state != DLM_CTXT_JOINED) { | |
1352 | spin_unlock(&dlm_domain_lock); | |
1353 | ||
1354 | mlog(0, "This ctxt is not joined yet!\n"); | |
1355 | wait_event_interruptible(dlm_domain_events, | |
1356 | dlm_wait_on_domain_helper( | |
1357 | domain)); | |
1358 | goto retry; | |
1359 | } | |
1360 | ||
1361 | __dlm_get(dlm); | |
1362 | dlm->num_joins++; | |
1363 | ||
1364 | spin_unlock(&dlm_domain_lock); | |
1365 | ||
1366 | ret = 0; | |
1367 | goto leave; | |
1368 | } | |
1369 | ||
1370 | /* doesn't exist */ | |
1371 | if (!new_ctxt) { | |
1372 | spin_unlock(&dlm_domain_lock); | |
1373 | ||
1374 | new_ctxt = dlm_alloc_ctxt(domain, key); | |
1375 | if (new_ctxt) | |
1376 | goto retry; | |
1377 | ||
1378 | ret = -ENOMEM; | |
1379 | mlog_errno(ret); | |
1380 | goto leave; | |
1381 | } | |
1382 | ||
1383 | /* a little variable switch-a-roo here... */ | |
1384 | dlm = new_ctxt; | |
1385 | new_ctxt = NULL; | |
1386 | ||
1387 | /* add the new domain */ | |
1388 | list_add_tail(&dlm->list, &dlm_domains); | |
1389 | spin_unlock(&dlm_domain_lock); | |
1390 | ||
1391 | ret = dlm_join_domain(dlm); | |
1392 | if (ret) { | |
1393 | mlog_errno(ret); | |
1394 | dlm_put(dlm); | |
1395 | goto leave; | |
1396 | } | |
1397 | ||
1398 | ret = 0; | |
1399 | leave: | |
1400 | if (new_ctxt) | |
1401 | dlm_free_ctxt_mem(new_ctxt); | |
1402 | ||
1403 | if (ret < 0) | |
1404 | dlm = ERR_PTR(ret); | |
1405 | ||
1406 | return dlm; | |
1407 | } | |
1408 | EXPORT_SYMBOL_GPL(dlm_register_domain); | |
1409 | ||
1410 | static LIST_HEAD(dlm_join_handlers); | |
1411 | ||
1412 | static void dlm_unregister_net_handlers(void) | |
1413 | { | |
1414 | o2net_unregister_handler_list(&dlm_join_handlers); | |
1415 | } | |
1416 | ||
1417 | static int dlm_register_net_handlers(void) | |
1418 | { | |
1419 | int status = 0; | |
1420 | ||
1421 | status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, | |
1422 | sizeof(struct dlm_query_join_request), | |
1423 | dlm_query_join_handler, | |
1424 | NULL, &dlm_join_handlers); | |
1425 | if (status) | |
1426 | goto bail; | |
1427 | ||
1428 | status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY, | |
1429 | sizeof(struct dlm_assert_joined), | |
1430 | dlm_assert_joined_handler, | |
1431 | NULL, &dlm_join_handlers); | |
1432 | if (status) | |
1433 | goto bail; | |
1434 | ||
1435 | status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY, | |
1436 | sizeof(struct dlm_cancel_join), | |
1437 | dlm_cancel_join_handler, | |
1438 | NULL, &dlm_join_handlers); | |
1439 | ||
1440 | bail: | |
1441 | if (status < 0) | |
1442 | dlm_unregister_net_handlers(); | |
1443 | ||
1444 | return status; | |
1445 | } | |
1446 | ||
1447 | /* Domain eviction callback handling. | |
1448 | * | |
1449 | * The file system requires notification of node death *before* the | |
1450 | * dlm completes it's recovery work, otherwise it may be able to | |
1451 | * acquire locks on resources requiring recovery. Since the dlm can | |
1452 | * evict a node from it's domain *before* heartbeat fires, a similar | |
1453 | * mechanism is required. */ | |
1454 | ||
1455 | /* Eviction is not expected to happen often, so a per-domain lock is | |
1456 | * not necessary. Eviction callbacks are allowed to sleep for short | |
1457 | * periods of time. */ | |
1458 | static DECLARE_RWSEM(dlm_callback_sem); | |
1459 | ||
1460 | void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm, | |
1461 | int node_num) | |
1462 | { | |
1463 | struct list_head *iter; | |
1464 | struct dlm_eviction_cb *cb; | |
1465 | ||
1466 | down_read(&dlm_callback_sem); | |
1467 | list_for_each(iter, &dlm->dlm_eviction_callbacks) { | |
1468 | cb = list_entry(iter, struct dlm_eviction_cb, ec_item); | |
1469 | ||
1470 | cb->ec_func(node_num, cb->ec_data); | |
1471 | } | |
1472 | up_read(&dlm_callback_sem); | |
1473 | } | |
1474 | ||
1475 | void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb, | |
1476 | dlm_eviction_func *f, | |
1477 | void *data) | |
1478 | { | |
1479 | INIT_LIST_HEAD(&cb->ec_item); | |
1480 | cb->ec_func = f; | |
1481 | cb->ec_data = data; | |
1482 | } | |
1483 | EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb); | |
1484 | ||
1485 | void dlm_register_eviction_cb(struct dlm_ctxt *dlm, | |
1486 | struct dlm_eviction_cb *cb) | |
1487 | { | |
1488 | down_write(&dlm_callback_sem); | |
1489 | list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks); | |
1490 | up_write(&dlm_callback_sem); | |
1491 | } | |
1492 | EXPORT_SYMBOL_GPL(dlm_register_eviction_cb); | |
1493 | ||
1494 | void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb) | |
1495 | { | |
1496 | down_write(&dlm_callback_sem); | |
1497 | list_del_init(&cb->ec_item); | |
1498 | up_write(&dlm_callback_sem); | |
1499 | } | |
1500 | EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb); | |
1501 | ||
1502 | static int __init dlm_init(void) | |
1503 | { | |
1504 | int status; | |
1505 | ||
1506 | dlm_print_version(); | |
1507 | ||
1508 | status = dlm_init_mle_cache(); | |
1509 | if (status) | |
1510 | return -1; | |
1511 | ||
1512 | status = dlm_register_net_handlers(); | |
1513 | if (status) { | |
1514 | dlm_destroy_mle_cache(); | |
1515 | return -1; | |
1516 | } | |
1517 | ||
1518 | return 0; | |
1519 | } | |
1520 | ||
1521 | static void __exit dlm_exit (void) | |
1522 | { | |
1523 | dlm_unregister_net_handlers(); | |
1524 | dlm_destroy_mle_cache(); | |
1525 | } | |
1526 | ||
1527 | MODULE_AUTHOR("Oracle"); | |
1528 | MODULE_LICENSE("GPL"); | |
1529 | ||
1530 | module_init(dlm_init); | |
1531 | module_exit(dlm_exit); |