Commit | Line | Data |
---|---|---|
6714d8e8 KH |
1 | /* -*- mode: c; c-basic-offset: 8; -*- |
2 | * vim: noexpandtab sw=8 ts=8 sts=0: | |
3 | * | |
4 | * dlmmod.c | |
5 | * | |
6 | * standalone DLM module | |
7 | * | |
8 | * Copyright (C) 2004 Oracle. All rights reserved. | |
9 | * | |
10 | * This program is free software; you can redistribute it and/or | |
11 | * modify it under the terms of the GNU General Public | |
12 | * License as published by the Free Software Foundation; either | |
13 | * version 2 of the License, or (at your option) any later version. | |
14 | * | |
15 | * This program is distributed in the hope that it will be useful, | |
16 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
17 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
18 | * General Public License for more details. | |
19 | * | |
20 | * You should have received a copy of the GNU General Public | |
21 | * License along with this program; if not, write to the | |
22 | * Free Software Foundation, Inc., 59 Temple Place - Suite 330, | |
23 | * Boston, MA 021110-1307, USA. | |
24 | * | |
25 | */ | |
26 | ||
27 | ||
28 | #include <linux/module.h> | |
29 | #include <linux/fs.h> | |
30 | #include <linux/types.h> | |
31 | #include <linux/slab.h> | |
32 | #include <linux/highmem.h> | |
33 | #include <linux/utsname.h> | |
34 | #include <linux/init.h> | |
35 | #include <linux/sysctl.h> | |
36 | #include <linux/random.h> | |
37 | #include <linux/blkdev.h> | |
38 | #include <linux/socket.h> | |
39 | #include <linux/inet.h> | |
40 | #include <linux/spinlock.h> | |
41 | #include <linux/delay.h> | |
42 | ||
43 | ||
44 | #include "cluster/heartbeat.h" | |
45 | #include "cluster/nodemanager.h" | |
46 | #include "cluster/tcp.h" | |
47 | ||
48 | #include "dlmapi.h" | |
49 | #include "dlmcommon.h" | |
82353b59 | 50 | #include "dlmdomain.h" |
6714d8e8 KH |
51 | |
52 | #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER) | |
53 | #include "cluster/masklog.h" | |
54 | ||
55 | enum dlm_mle_type { | |
56 | DLM_MLE_BLOCK, | |
57 | DLM_MLE_MASTER, | |
58 | DLM_MLE_MIGRATION | |
59 | }; | |
60 | ||
61 | struct dlm_lock_name | |
62 | { | |
63 | u8 len; | |
64 | u8 name[DLM_LOCKID_NAME_MAX]; | |
65 | }; | |
66 | ||
67 | struct dlm_master_list_entry | |
68 | { | |
69 | struct list_head list; | |
70 | struct list_head hb_events; | |
71 | struct dlm_ctxt *dlm; | |
72 | spinlock_t spinlock; | |
73 | wait_queue_head_t wq; | |
74 | atomic_t woken; | |
75 | struct kref mle_refs; | |
a2bf0477 | 76 | int inuse; |
6714d8e8 KH |
77 | unsigned long maybe_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; |
78 | unsigned long vote_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; | |
79 | unsigned long response_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; | |
80 | unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; | |
81 | u8 master; | |
82 | u8 new_master; | |
83 | enum dlm_mle_type type; | |
84 | struct o2hb_callback_func mle_hb_up; | |
85 | struct o2hb_callback_func mle_hb_down; | |
86 | union { | |
87 | struct dlm_lock_resource *res; | |
88 | struct dlm_lock_name name; | |
89 | } u; | |
90 | }; | |
91 | ||
92 | static void dlm_mle_node_down(struct dlm_ctxt *dlm, | |
93 | struct dlm_master_list_entry *mle, | |
94 | struct o2nm_node *node, | |
95 | int idx); | |
96 | static void dlm_mle_node_up(struct dlm_ctxt *dlm, | |
97 | struct dlm_master_list_entry *mle, | |
98 | struct o2nm_node *node, | |
99 | int idx); | |
100 | ||
101 | static void dlm_assert_master_worker(struct dlm_work_item *item, void *data); | |
ba2bf218 KH |
102 | static int dlm_do_assert_master(struct dlm_ctxt *dlm, |
103 | struct dlm_lock_resource *res, | |
104 | void *nodemap, u32 flags); | |
f3f85464 | 105 | static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data); |
6714d8e8 KH |
106 | |
107 | static inline int dlm_mle_equal(struct dlm_ctxt *dlm, | |
108 | struct dlm_master_list_entry *mle, | |
109 | const char *name, | |
110 | unsigned int namelen) | |
111 | { | |
112 | struct dlm_lock_resource *res; | |
113 | ||
114 | if (dlm != mle->dlm) | |
115 | return 0; | |
116 | ||
117 | if (mle->type == DLM_MLE_BLOCK || | |
118 | mle->type == DLM_MLE_MIGRATION) { | |
119 | if (namelen != mle->u.name.len || | |
120 | memcmp(name, mle->u.name.name, namelen)!=0) | |
121 | return 0; | |
122 | } else { | |
123 | res = mle->u.res; | |
124 | if (namelen != res->lockname.len || | |
125 | memcmp(res->lockname.name, name, namelen) != 0) | |
126 | return 0; | |
127 | } | |
128 | return 1; | |
129 | } | |
130 | ||
95883719 | 131 | #define dlm_print_nodemap(m) _dlm_print_nodemap(m,#m) |
8a9343fa | 132 | static void _dlm_print_nodemap(unsigned long *map, const char *mapname) |
95883719 KH |
133 | { |
134 | int i; | |
135 | printk("%s=[ ", mapname); | |
136 | for (i=0; i<O2NM_MAX_NODES; i++) | |
137 | if (test_bit(i, map)) | |
138 | printk("%d ", i); | |
139 | printk("]"); | |
140 | } | |
141 | ||
8a9343fa | 142 | static void dlm_print_one_mle(struct dlm_master_list_entry *mle) |
6714d8e8 | 143 | { |
95883719 | 144 | int refs; |
6714d8e8 KH |
145 | char *type; |
146 | char attached; | |
147 | u8 master; | |
148 | unsigned int namelen; | |
149 | const char *name; | |
150 | struct kref *k; | |
95883719 KH |
151 | unsigned long *maybe = mle->maybe_map, |
152 | *vote = mle->vote_map, | |
153 | *resp = mle->response_map, | |
154 | *node = mle->node_map; | |
6714d8e8 KH |
155 | |
156 | k = &mle->mle_refs; | |
157 | if (mle->type == DLM_MLE_BLOCK) | |
158 | type = "BLK"; | |
159 | else if (mle->type == DLM_MLE_MASTER) | |
160 | type = "MAS"; | |
161 | else | |
162 | type = "MIG"; | |
163 | refs = atomic_read(&k->refcount); | |
164 | master = mle->master; | |
165 | attached = (list_empty(&mle->hb_events) ? 'N' : 'Y'); | |
166 | ||
167 | if (mle->type != DLM_MLE_MASTER) { | |
168 | namelen = mle->u.name.len; | |
169 | name = mle->u.name.name; | |
170 | } else { | |
171 | namelen = mle->u.res->lockname.len; | |
172 | name = mle->u.res->lockname.name; | |
173 | } | |
174 | ||
95883719 KH |
175 | mlog(ML_NOTICE, "%.*s: %3s refs=%3d mas=%3u new=%3u evt=%c inuse=%d ", |
176 | namelen, name, type, refs, master, mle->new_master, attached, | |
177 | mle->inuse); | |
178 | dlm_print_nodemap(maybe); | |
179 | printk(", "); | |
180 | dlm_print_nodemap(vote); | |
181 | printk(", "); | |
182 | dlm_print_nodemap(resp); | |
183 | printk(", "); | |
184 | dlm_print_nodemap(node); | |
185 | printk(", "); | |
186 | printk("\n"); | |
6714d8e8 KH |
187 | } |
188 | ||
8a9343fa MF |
189 | #if 0 |
190 | /* Code here is included but defined out as it aids debugging */ | |
191 | ||
6714d8e8 KH |
192 | static void dlm_dump_mles(struct dlm_ctxt *dlm) |
193 | { | |
194 | struct dlm_master_list_entry *mle; | |
6714d8e8 KH |
195 | |
196 | mlog(ML_NOTICE, "dumping all mles for domain %s:\n", dlm->name); | |
6714d8e8 | 197 | spin_lock(&dlm->master_lock); |
800deef3 | 198 | list_for_each_entry(mle, &dlm->master_list, list) |
6714d8e8 | 199 | dlm_print_one_mle(mle); |
6714d8e8 KH |
200 | spin_unlock(&dlm->master_lock); |
201 | } | |
202 | ||
6714d8e8 KH |
203 | int dlm_dump_all_mles(const char __user *data, unsigned int len) |
204 | { | |
6714d8e8 KH |
205 | struct dlm_ctxt *dlm; |
206 | ||
207 | spin_lock(&dlm_domain_lock); | |
800deef3 | 208 | list_for_each_entry(dlm, &dlm_domains, list) { |
6714d8e8 KH |
209 | mlog(ML_NOTICE, "found dlm: %p, name=%s\n", dlm, dlm->name); |
210 | dlm_dump_mles(dlm); | |
211 | } | |
212 | spin_unlock(&dlm_domain_lock); | |
213 | return len; | |
214 | } | |
215 | EXPORT_SYMBOL_GPL(dlm_dump_all_mles); | |
216 | ||
217 | #endif /* 0 */ | |
218 | ||
219 | ||
e18b890b | 220 | static struct kmem_cache *dlm_mle_cache = NULL; |
6714d8e8 KH |
221 | |
222 | ||
223 | static void dlm_mle_release(struct kref *kref); | |
224 | static void dlm_init_mle(struct dlm_master_list_entry *mle, | |
225 | enum dlm_mle_type type, | |
226 | struct dlm_ctxt *dlm, | |
227 | struct dlm_lock_resource *res, | |
228 | const char *name, | |
229 | unsigned int namelen); | |
230 | static void dlm_put_mle(struct dlm_master_list_entry *mle); | |
231 | static void __dlm_put_mle(struct dlm_master_list_entry *mle); | |
232 | static int dlm_find_mle(struct dlm_ctxt *dlm, | |
233 | struct dlm_master_list_entry **mle, | |
234 | char *name, unsigned int namelen); | |
235 | ||
ba2bf218 KH |
236 | static int dlm_do_master_request(struct dlm_lock_resource *res, |
237 | struct dlm_master_list_entry *mle, int to); | |
6714d8e8 KH |
238 | |
239 | ||
240 | static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm, | |
241 | struct dlm_lock_resource *res, | |
242 | struct dlm_master_list_entry *mle, | |
243 | int *blocked); | |
244 | static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm, | |
245 | struct dlm_lock_resource *res, | |
246 | struct dlm_master_list_entry *mle, | |
247 | int blocked); | |
248 | static int dlm_add_migration_mle(struct dlm_ctxt *dlm, | |
249 | struct dlm_lock_resource *res, | |
250 | struct dlm_master_list_entry *mle, | |
251 | struct dlm_master_list_entry **oldmle, | |
252 | const char *name, unsigned int namelen, | |
253 | u8 new_master, u8 master); | |
254 | ||
255 | static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm, | |
256 | struct dlm_lock_resource *res); | |
257 | static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm, | |
258 | struct dlm_lock_resource *res); | |
259 | static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm, | |
260 | struct dlm_lock_resource *res, | |
261 | u8 target); | |
c03872f5 KH |
262 | static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm, |
263 | struct dlm_lock_resource *res); | |
6714d8e8 KH |
264 | |
265 | ||
266 | int dlm_is_host_down(int errno) | |
267 | { | |
268 | switch (errno) { | |
269 | case -EBADF: | |
270 | case -ECONNREFUSED: | |
271 | case -ENOTCONN: | |
272 | case -ECONNRESET: | |
273 | case -EPIPE: | |
274 | case -EHOSTDOWN: | |
275 | case -EHOSTUNREACH: | |
276 | case -ETIMEDOUT: | |
277 | case -ECONNABORTED: | |
278 | case -ENETDOWN: | |
279 | case -ENETUNREACH: | |
280 | case -ENETRESET: | |
281 | case -ESHUTDOWN: | |
282 | case -ENOPROTOOPT: | |
283 | case -EINVAL: /* if returned from our tcp code, | |
284 | this means there is no socket */ | |
285 | return 1; | |
286 | } | |
287 | return 0; | |
288 | } | |
289 | ||
290 | ||
291 | /* | |
292 | * MASTER LIST FUNCTIONS | |
293 | */ | |
294 | ||
295 | ||
296 | /* | |
297 | * regarding master list entries and heartbeat callbacks: | |
298 | * | |
299 | * in order to avoid sleeping and allocation that occurs in | |
300 | * heartbeat, master list entries are simply attached to the | |
301 | * dlm's established heartbeat callbacks. the mle is attached | |
302 | * when it is created, and since the dlm->spinlock is held at | |
303 | * that time, any heartbeat event will be properly discovered | |
304 | * by the mle. the mle needs to be detached from the | |
305 | * dlm->mle_hb_events list as soon as heartbeat events are no | |
306 | * longer useful to the mle, and before the mle is freed. | |
307 | * | |
308 | * as a general rule, heartbeat events are no longer needed by | |
309 | * the mle once an "answer" regarding the lock master has been | |
310 | * received. | |
311 | */ | |
312 | static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm, | |
313 | struct dlm_master_list_entry *mle) | |
314 | { | |
315 | assert_spin_locked(&dlm->spinlock); | |
316 | ||
317 | list_add_tail(&mle->hb_events, &dlm->mle_hb_events); | |
318 | } | |
319 | ||
320 | ||
321 | static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm, | |
322 | struct dlm_master_list_entry *mle) | |
323 | { | |
324 | if (!list_empty(&mle->hb_events)) | |
325 | list_del_init(&mle->hb_events); | |
326 | } | |
327 | ||
328 | ||
329 | static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm, | |
330 | struct dlm_master_list_entry *mle) | |
331 | { | |
332 | spin_lock(&dlm->spinlock); | |
333 | __dlm_mle_detach_hb_events(dlm, mle); | |
334 | spin_unlock(&dlm->spinlock); | |
335 | } | |
336 | ||
a2bf0477 KH |
337 | static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle) |
338 | { | |
339 | struct dlm_ctxt *dlm; | |
340 | dlm = mle->dlm; | |
341 | ||
342 | assert_spin_locked(&dlm->spinlock); | |
343 | assert_spin_locked(&dlm->master_lock); | |
344 | mle->inuse++; | |
345 | kref_get(&mle->mle_refs); | |
346 | } | |
347 | ||
348 | static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle) | |
349 | { | |
350 | struct dlm_ctxt *dlm; | |
351 | dlm = mle->dlm; | |
352 | ||
353 | spin_lock(&dlm->spinlock); | |
354 | spin_lock(&dlm->master_lock); | |
355 | mle->inuse--; | |
356 | __dlm_put_mle(mle); | |
357 | spin_unlock(&dlm->master_lock); | |
358 | spin_unlock(&dlm->spinlock); | |
359 | ||
360 | } | |
361 | ||
6714d8e8 KH |
362 | /* remove from list and free */ |
363 | static void __dlm_put_mle(struct dlm_master_list_entry *mle) | |
364 | { | |
365 | struct dlm_ctxt *dlm; | |
366 | dlm = mle->dlm; | |
367 | ||
368 | assert_spin_locked(&dlm->spinlock); | |
369 | assert_spin_locked(&dlm->master_lock); | |
aa852354 KH |
370 | if (!atomic_read(&mle->mle_refs.refcount)) { |
371 | /* this may or may not crash, but who cares. | |
372 | * it's a BUG. */ | |
373 | mlog(ML_ERROR, "bad mle: %p\n", mle); | |
374 | dlm_print_one_mle(mle); | |
375 | BUG(); | |
376 | } else | |
377 | kref_put(&mle->mle_refs, dlm_mle_release); | |
6714d8e8 KH |
378 | } |
379 | ||
380 | ||
381 | /* must not have any spinlocks coming in */ | |
382 | static void dlm_put_mle(struct dlm_master_list_entry *mle) | |
383 | { | |
384 | struct dlm_ctxt *dlm; | |
385 | dlm = mle->dlm; | |
386 | ||
387 | spin_lock(&dlm->spinlock); | |
388 | spin_lock(&dlm->master_lock); | |
389 | __dlm_put_mle(mle); | |
390 | spin_unlock(&dlm->master_lock); | |
391 | spin_unlock(&dlm->spinlock); | |
392 | } | |
393 | ||
394 | static inline void dlm_get_mle(struct dlm_master_list_entry *mle) | |
395 | { | |
396 | kref_get(&mle->mle_refs); | |
397 | } | |
398 | ||
399 | static void dlm_init_mle(struct dlm_master_list_entry *mle, | |
400 | enum dlm_mle_type type, | |
401 | struct dlm_ctxt *dlm, | |
402 | struct dlm_lock_resource *res, | |
403 | const char *name, | |
404 | unsigned int namelen) | |
405 | { | |
406 | assert_spin_locked(&dlm->spinlock); | |
407 | ||
408 | mle->dlm = dlm; | |
409 | mle->type = type; | |
410 | INIT_LIST_HEAD(&mle->list); | |
411 | INIT_LIST_HEAD(&mle->hb_events); | |
412 | memset(mle->maybe_map, 0, sizeof(mle->maybe_map)); | |
413 | spin_lock_init(&mle->spinlock); | |
414 | init_waitqueue_head(&mle->wq); | |
415 | atomic_set(&mle->woken, 0); | |
416 | kref_init(&mle->mle_refs); | |
417 | memset(mle->response_map, 0, sizeof(mle->response_map)); | |
418 | mle->master = O2NM_MAX_NODES; | |
419 | mle->new_master = O2NM_MAX_NODES; | |
a2bf0477 | 420 | mle->inuse = 0; |
6714d8e8 KH |
421 | |
422 | if (mle->type == DLM_MLE_MASTER) { | |
423 | BUG_ON(!res); | |
424 | mle->u.res = res; | |
425 | } else if (mle->type == DLM_MLE_BLOCK) { | |
426 | BUG_ON(!name); | |
427 | memcpy(mle->u.name.name, name, namelen); | |
428 | mle->u.name.len = namelen; | |
429 | } else /* DLM_MLE_MIGRATION */ { | |
430 | BUG_ON(!name); | |
431 | memcpy(mle->u.name.name, name, namelen); | |
432 | mle->u.name.len = namelen; | |
433 | } | |
434 | ||
435 | /* copy off the node_map and register hb callbacks on our copy */ | |
436 | memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map)); | |
437 | memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map)); | |
438 | clear_bit(dlm->node_num, mle->vote_map); | |
439 | clear_bit(dlm->node_num, mle->node_map); | |
440 | ||
441 | /* attach the mle to the domain node up/down events */ | |
442 | __dlm_mle_attach_hb_events(dlm, mle); | |
443 | } | |
444 | ||
445 | ||
446 | /* returns 1 if found, 0 if not */ | |
447 | static int dlm_find_mle(struct dlm_ctxt *dlm, | |
448 | struct dlm_master_list_entry **mle, | |
449 | char *name, unsigned int namelen) | |
450 | { | |
451 | struct dlm_master_list_entry *tmpmle; | |
6714d8e8 KH |
452 | |
453 | assert_spin_locked(&dlm->master_lock); | |
454 | ||
800deef3 | 455 | list_for_each_entry(tmpmle, &dlm->master_list, list) { |
6714d8e8 KH |
456 | if (!dlm_mle_equal(dlm, tmpmle, name, namelen)) |
457 | continue; | |
458 | dlm_get_mle(tmpmle); | |
459 | *mle = tmpmle; | |
460 | return 1; | |
461 | } | |
462 | return 0; | |
463 | } | |
464 | ||
465 | void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up) | |
466 | { | |
467 | struct dlm_master_list_entry *mle; | |
6714d8e8 KH |
468 | |
469 | assert_spin_locked(&dlm->spinlock); | |
470 | ||
800deef3 | 471 | list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) { |
6714d8e8 KH |
472 | if (node_up) |
473 | dlm_mle_node_up(dlm, mle, NULL, idx); | |
474 | else | |
475 | dlm_mle_node_down(dlm, mle, NULL, idx); | |
476 | } | |
477 | } | |
478 | ||
479 | static void dlm_mle_node_down(struct dlm_ctxt *dlm, | |
480 | struct dlm_master_list_entry *mle, | |
481 | struct o2nm_node *node, int idx) | |
482 | { | |
483 | spin_lock(&mle->spinlock); | |
484 | ||
485 | if (!test_bit(idx, mle->node_map)) | |
486 | mlog(0, "node %u already removed from nodemap!\n", idx); | |
487 | else | |
488 | clear_bit(idx, mle->node_map); | |
489 | ||
490 | spin_unlock(&mle->spinlock); | |
491 | } | |
492 | ||
493 | static void dlm_mle_node_up(struct dlm_ctxt *dlm, | |
494 | struct dlm_master_list_entry *mle, | |
495 | struct o2nm_node *node, int idx) | |
496 | { | |
497 | spin_lock(&mle->spinlock); | |
498 | ||
499 | if (test_bit(idx, mle->node_map)) | |
500 | mlog(0, "node %u already in node map!\n", idx); | |
501 | else | |
502 | set_bit(idx, mle->node_map); | |
503 | ||
504 | spin_unlock(&mle->spinlock); | |
505 | } | |
506 | ||
507 | ||
508 | int dlm_init_mle_cache(void) | |
509 | { | |
510 | dlm_mle_cache = kmem_cache_create("dlm_mle_cache", | |
511 | sizeof(struct dlm_master_list_entry), | |
512 | 0, SLAB_HWCACHE_ALIGN, | |
20c2df83 | 513 | NULL); |
6714d8e8 KH |
514 | if (dlm_mle_cache == NULL) |
515 | return -ENOMEM; | |
516 | return 0; | |
517 | } | |
518 | ||
519 | void dlm_destroy_mle_cache(void) | |
520 | { | |
521 | if (dlm_mle_cache) | |
522 | kmem_cache_destroy(dlm_mle_cache); | |
523 | } | |
524 | ||
525 | static void dlm_mle_release(struct kref *kref) | |
526 | { | |
527 | struct dlm_master_list_entry *mle; | |
528 | struct dlm_ctxt *dlm; | |
529 | ||
530 | mlog_entry_void(); | |
531 | ||
532 | mle = container_of(kref, struct dlm_master_list_entry, mle_refs); | |
533 | dlm = mle->dlm; | |
534 | ||
535 | if (mle->type != DLM_MLE_MASTER) { | |
536 | mlog(0, "calling mle_release for %.*s, type %d\n", | |
537 | mle->u.name.len, mle->u.name.name, mle->type); | |
538 | } else { | |
539 | mlog(0, "calling mle_release for %.*s, type %d\n", | |
540 | mle->u.res->lockname.len, | |
541 | mle->u.res->lockname.name, mle->type); | |
542 | } | |
543 | assert_spin_locked(&dlm->spinlock); | |
544 | assert_spin_locked(&dlm->master_lock); | |
545 | ||
546 | /* remove from list if not already */ | |
547 | if (!list_empty(&mle->list)) | |
548 | list_del_init(&mle->list); | |
549 | ||
550 | /* detach the mle from the domain node up/down events */ | |
551 | __dlm_mle_detach_hb_events(dlm, mle); | |
552 | ||
553 | /* NOTE: kfree under spinlock here. | |
554 | * if this is bad, we can move this to a freelist. */ | |
555 | kmem_cache_free(dlm_mle_cache, mle); | |
556 | } | |
557 | ||
558 | ||
559 | /* | |
560 | * LOCK RESOURCE FUNCTIONS | |
561 | */ | |
562 | ||
563 | static void dlm_set_lockres_owner(struct dlm_ctxt *dlm, | |
564 | struct dlm_lock_resource *res, | |
565 | u8 owner) | |
566 | { | |
567 | assert_spin_locked(&res->spinlock); | |
568 | ||
569 | mlog_entry("%.*s, %u\n", res->lockname.len, res->lockname.name, owner); | |
570 | ||
571 | if (owner == dlm->node_num) | |
572 | atomic_inc(&dlm->local_resources); | |
573 | else if (owner == DLM_LOCK_RES_OWNER_UNKNOWN) | |
574 | atomic_inc(&dlm->unknown_resources); | |
575 | else | |
576 | atomic_inc(&dlm->remote_resources); | |
577 | ||
578 | res->owner = owner; | |
579 | } | |
580 | ||
581 | void dlm_change_lockres_owner(struct dlm_ctxt *dlm, | |
582 | struct dlm_lock_resource *res, u8 owner) | |
583 | { | |
584 | assert_spin_locked(&res->spinlock); | |
585 | ||
586 | if (owner == res->owner) | |
587 | return; | |
588 | ||
589 | if (res->owner == dlm->node_num) | |
590 | atomic_dec(&dlm->local_resources); | |
591 | else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) | |
592 | atomic_dec(&dlm->unknown_resources); | |
593 | else | |
594 | atomic_dec(&dlm->remote_resources); | |
595 | ||
596 | dlm_set_lockres_owner(dlm, res, owner); | |
597 | } | |
598 | ||
599 | ||
600 | static void dlm_lockres_release(struct kref *kref) | |
601 | { | |
602 | struct dlm_lock_resource *res; | |
603 | ||
604 | res = container_of(kref, struct dlm_lock_resource, refs); | |
605 | ||
606 | /* This should not happen -- all lockres' have a name | |
607 | * associated with them at init time. */ | |
608 | BUG_ON(!res->lockname.name); | |
609 | ||
610 | mlog(0, "destroying lockres %.*s\n", res->lockname.len, | |
611 | res->lockname.name); | |
612 | ||
a7f90d83 KH |
613 | if (!hlist_unhashed(&res->hash_node) || |
614 | !list_empty(&res->granted) || | |
615 | !list_empty(&res->converting) || | |
616 | !list_empty(&res->blocked) || | |
617 | !list_empty(&res->dirty) || | |
618 | !list_empty(&res->recovering) || | |
619 | !list_empty(&res->purge)) { | |
620 | mlog(ML_ERROR, | |
621 | "Going to BUG for resource %.*s." | |
622 | " We're on a list! [%c%c%c%c%c%c%c]\n", | |
623 | res->lockname.len, res->lockname.name, | |
624 | !hlist_unhashed(&res->hash_node) ? 'H' : ' ', | |
625 | !list_empty(&res->granted) ? 'G' : ' ', | |
626 | !list_empty(&res->converting) ? 'C' : ' ', | |
627 | !list_empty(&res->blocked) ? 'B' : ' ', | |
628 | !list_empty(&res->dirty) ? 'D' : ' ', | |
629 | !list_empty(&res->recovering) ? 'R' : ' ', | |
630 | !list_empty(&res->purge) ? 'P' : ' '); | |
631 | ||
632 | dlm_print_one_lock_resource(res); | |
633 | } | |
634 | ||
6714d8e8 KH |
635 | /* By the time we're ready to blow this guy away, we shouldn't |
636 | * be on any lists. */ | |
81f2094a | 637 | BUG_ON(!hlist_unhashed(&res->hash_node)); |
6714d8e8 KH |
638 | BUG_ON(!list_empty(&res->granted)); |
639 | BUG_ON(!list_empty(&res->converting)); | |
640 | BUG_ON(!list_empty(&res->blocked)); | |
641 | BUG_ON(!list_empty(&res->dirty)); | |
642 | BUG_ON(!list_empty(&res->recovering)); | |
643 | BUG_ON(!list_empty(&res->purge)); | |
644 | ||
645 | kfree(res->lockname.name); | |
646 | ||
647 | kfree(res); | |
648 | } | |
649 | ||
6714d8e8 KH |
650 | void dlm_lockres_put(struct dlm_lock_resource *res) |
651 | { | |
652 | kref_put(&res->refs, dlm_lockres_release); | |
653 | } | |
654 | ||
655 | static void dlm_init_lockres(struct dlm_ctxt *dlm, | |
656 | struct dlm_lock_resource *res, | |
657 | const char *name, unsigned int namelen) | |
658 | { | |
659 | char *qname; | |
660 | ||
661 | /* If we memset here, we lose our reference to the kmalloc'd | |
662 | * res->lockname.name, so be sure to init every field | |
663 | * correctly! */ | |
664 | ||
665 | qname = (char *) res->lockname.name; | |
666 | memcpy(qname, name, namelen); | |
667 | ||
668 | res->lockname.len = namelen; | |
a3d33291 | 669 | res->lockname.hash = dlm_lockid_hash(name, namelen); |
6714d8e8 KH |
670 | |
671 | init_waitqueue_head(&res->wq); | |
672 | spin_lock_init(&res->spinlock); | |
81f2094a | 673 | INIT_HLIST_NODE(&res->hash_node); |
6714d8e8 KH |
674 | INIT_LIST_HEAD(&res->granted); |
675 | INIT_LIST_HEAD(&res->converting); | |
676 | INIT_LIST_HEAD(&res->blocked); | |
677 | INIT_LIST_HEAD(&res->dirty); | |
678 | INIT_LIST_HEAD(&res->recovering); | |
679 | INIT_LIST_HEAD(&res->purge); | |
680 | atomic_set(&res->asts_reserved, 0); | |
681 | res->migration_pending = 0; | |
ba2bf218 | 682 | res->inflight_locks = 0; |
6714d8e8 KH |
683 | |
684 | kref_init(&res->refs); | |
685 | ||
686 | /* just for consistency */ | |
687 | spin_lock(&res->spinlock); | |
688 | dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN); | |
689 | spin_unlock(&res->spinlock); | |
690 | ||
691 | res->state = DLM_LOCK_RES_IN_PROGRESS; | |
692 | ||
693 | res->last_used = 0; | |
694 | ||
695 | memset(res->lvb, 0, DLM_LVB_LEN); | |
ba2bf218 | 696 | memset(res->refmap, 0, sizeof(res->refmap)); |
6714d8e8 KH |
697 | } |
698 | ||
699 | struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm, | |
700 | const char *name, | |
701 | unsigned int namelen) | |
702 | { | |
703 | struct dlm_lock_resource *res; | |
704 | ||
ad8100e0 | 705 | res = kmalloc(sizeof(struct dlm_lock_resource), GFP_NOFS); |
6714d8e8 KH |
706 | if (!res) |
707 | return NULL; | |
708 | ||
ad8100e0 | 709 | res->lockname.name = kmalloc(namelen, GFP_NOFS); |
6714d8e8 KH |
710 | if (!res->lockname.name) { |
711 | kfree(res); | |
712 | return NULL; | |
713 | } | |
714 | ||
715 | dlm_init_lockres(dlm, res, name, namelen); | |
716 | return res; | |
717 | } | |
718 | ||
ba2bf218 KH |
719 | void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm, |
720 | struct dlm_lock_resource *res, | |
721 | int new_lockres, | |
722 | const char *file, | |
723 | int line) | |
724 | { | |
725 | if (!new_lockres) | |
726 | assert_spin_locked(&res->spinlock); | |
727 | ||
728 | if (!test_bit(dlm->node_num, res->refmap)) { | |
729 | BUG_ON(res->inflight_locks != 0); | |
730 | dlm_lockres_set_refmap_bit(dlm->node_num, res); | |
731 | } | |
732 | res->inflight_locks++; | |
733 | mlog(0, "%s:%.*s: inflight++: now %u\n", | |
734 | dlm->name, res->lockname.len, res->lockname.name, | |
735 | res->inflight_locks); | |
736 | } | |
737 | ||
738 | void __dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm, | |
739 | struct dlm_lock_resource *res, | |
740 | const char *file, | |
741 | int line) | |
742 | { | |
743 | assert_spin_locked(&res->spinlock); | |
744 | ||
745 | BUG_ON(res->inflight_locks == 0); | |
746 | res->inflight_locks--; | |
747 | mlog(0, "%s:%.*s: inflight--: now %u\n", | |
748 | dlm->name, res->lockname.len, res->lockname.name, | |
749 | res->inflight_locks); | |
750 | if (res->inflight_locks == 0) | |
751 | dlm_lockres_clear_refmap_bit(dlm->node_num, res); | |
752 | wake_up(&res->wq); | |
753 | } | |
754 | ||
6714d8e8 KH |
755 | /* |
756 | * lookup a lock resource by name. | |
757 | * may already exist in the hashtable. | |
758 | * lockid is null terminated | |
759 | * | |
760 | * if not, allocate enough for the lockres and for | |
761 | * the temporary structure used in doing the mastering. | |
762 | * | |
763 | * also, do a lookup in the dlm->master_list to see | |
764 | * if another node has begun mastering the same lock. | |
765 | * if so, there should be a block entry in there | |
766 | * for this name, and we should *not* attempt to master | |
767 | * the lock here. need to wait around for that node | |
768 | * to assert_master (or die). | |
769 | * | |
770 | */ | |
771 | struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm, | |
772 | const char *lockid, | |
3384f3df | 773 | int namelen, |
6714d8e8 KH |
774 | int flags) |
775 | { | |
776 | struct dlm_lock_resource *tmpres=NULL, *res=NULL; | |
777 | struct dlm_master_list_entry *mle = NULL; | |
778 | struct dlm_master_list_entry *alloc_mle = NULL; | |
779 | int blocked = 0; | |
780 | int ret, nodenum; | |
781 | struct dlm_node_iter iter; | |
3384f3df | 782 | unsigned int hash; |
6714d8e8 | 783 | int tries = 0; |
c03872f5 | 784 | int bit, wait_on_recovery = 0; |
ba2bf218 | 785 | int drop_inflight_if_nonlocal = 0; |
6714d8e8 KH |
786 | |
787 | BUG_ON(!lockid); | |
788 | ||
a3d33291 | 789 | hash = dlm_lockid_hash(lockid, namelen); |
6714d8e8 KH |
790 | |
791 | mlog(0, "get lockres %s (len %d)\n", lockid, namelen); | |
792 | ||
793 | lookup: | |
794 | spin_lock(&dlm->spinlock); | |
ba2bf218 | 795 | tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash); |
6714d8e8 | 796 | if (tmpres) { |
ba2bf218 KH |
797 | int dropping_ref = 0; |
798 | ||
799 | spin_lock(&tmpres->spinlock); | |
800 | if (tmpres->owner == dlm->node_num) { | |
801 | BUG_ON(tmpres->state & DLM_LOCK_RES_DROPPING_REF); | |
802 | dlm_lockres_grab_inflight_ref(dlm, tmpres); | |
803 | } else if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) | |
804 | dropping_ref = 1; | |
805 | spin_unlock(&tmpres->spinlock); | |
6714d8e8 | 806 | spin_unlock(&dlm->spinlock); |
ba2bf218 KH |
807 | |
808 | /* wait until done messaging the master, drop our ref to allow | |
809 | * the lockres to be purged, start over. */ | |
810 | if (dropping_ref) { | |
811 | spin_lock(&tmpres->spinlock); | |
812 | __dlm_wait_on_lockres_flags(tmpres, DLM_LOCK_RES_DROPPING_REF); | |
813 | spin_unlock(&tmpres->spinlock); | |
814 | dlm_lockres_put(tmpres); | |
815 | tmpres = NULL; | |
816 | goto lookup; | |
817 | } | |
818 | ||
6714d8e8 KH |
819 | mlog(0, "found in hash!\n"); |
820 | if (res) | |
821 | dlm_lockres_put(res); | |
822 | res = tmpres; | |
823 | goto leave; | |
824 | } | |
825 | ||
826 | if (!res) { | |
827 | spin_unlock(&dlm->spinlock); | |
828 | mlog(0, "allocating a new resource\n"); | |
829 | /* nothing found and we need to allocate one. */ | |
830 | alloc_mle = (struct dlm_master_list_entry *) | |
ad8100e0 | 831 | kmem_cache_alloc(dlm_mle_cache, GFP_NOFS); |
6714d8e8 KH |
832 | if (!alloc_mle) |
833 | goto leave; | |
834 | res = dlm_new_lockres(dlm, lockid, namelen); | |
835 | if (!res) | |
836 | goto leave; | |
837 | goto lookup; | |
838 | } | |
839 | ||
840 | mlog(0, "no lockres found, allocated our own: %p\n", res); | |
841 | ||
842 | if (flags & LKM_LOCAL) { | |
843 | /* caller knows it's safe to assume it's not mastered elsewhere | |
844 | * DONE! return right away */ | |
845 | spin_lock(&res->spinlock); | |
846 | dlm_change_lockres_owner(dlm, res, dlm->node_num); | |
847 | __dlm_insert_lockres(dlm, res); | |
ba2bf218 | 848 | dlm_lockres_grab_inflight_ref(dlm, res); |
6714d8e8 KH |
849 | spin_unlock(&res->spinlock); |
850 | spin_unlock(&dlm->spinlock); | |
851 | /* lockres still marked IN_PROGRESS */ | |
852 | goto wake_waiters; | |
853 | } | |
854 | ||
855 | /* check master list to see if another node has started mastering it */ | |
856 | spin_lock(&dlm->master_lock); | |
857 | ||
858 | /* if we found a block, wait for lock to be mastered by another node */ | |
859 | blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen); | |
860 | if (blocked) { | |
ba2bf218 | 861 | int mig; |
6714d8e8 KH |
862 | if (mle->type == DLM_MLE_MASTER) { |
863 | mlog(ML_ERROR, "master entry for nonexistent lock!\n"); | |
864 | BUG(); | |
ba2bf218 KH |
865 | } |
866 | mig = (mle->type == DLM_MLE_MIGRATION); | |
867 | /* if there is a migration in progress, let the migration | |
868 | * finish before continuing. we can wait for the absence | |
869 | * of the MIGRATION mle: either the migrate finished or | |
870 | * one of the nodes died and the mle was cleaned up. | |
871 | * if there is a BLOCK here, but it already has a master | |
872 | * set, we are too late. the master does not have a ref | |
873 | * for us in the refmap. detach the mle and drop it. | |
874 | * either way, go back to the top and start over. */ | |
875 | if (mig || mle->master != O2NM_MAX_NODES) { | |
876 | BUG_ON(mig && mle->master == dlm->node_num); | |
877 | /* we arrived too late. the master does not | |
878 | * have a ref for us. retry. */ | |
879 | mlog(0, "%s:%.*s: late on %s\n", | |
880 | dlm->name, namelen, lockid, | |
881 | mig ? "MIGRATION" : "BLOCK"); | |
6714d8e8 | 882 | spin_unlock(&dlm->master_lock); |
6714d8e8 KH |
883 | spin_unlock(&dlm->spinlock); |
884 | ||
885 | /* master is known, detach */ | |
ba2bf218 KH |
886 | if (!mig) |
887 | dlm_mle_detach_hb_events(dlm, mle); | |
6714d8e8 KH |
888 | dlm_put_mle(mle); |
889 | mle = NULL; | |
ba2bf218 KH |
890 | /* this is lame, but we cant wait on either |
891 | * the mle or lockres waitqueue here */ | |
892 | if (mig) | |
893 | msleep(100); | |
894 | goto lookup; | |
6714d8e8 KH |
895 | } |
896 | } else { | |
897 | /* go ahead and try to master lock on this node */ | |
898 | mle = alloc_mle; | |
899 | /* make sure this does not get freed below */ | |
900 | alloc_mle = NULL; | |
901 | dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0); | |
902 | set_bit(dlm->node_num, mle->maybe_map); | |
903 | list_add(&mle->list, &dlm->master_list); | |
c03872f5 KH |
904 | |
905 | /* still holding the dlm spinlock, check the recovery map | |
906 | * to see if there are any nodes that still need to be | |
907 | * considered. these will not appear in the mle nodemap | |
908 | * but they might own this lockres. wait on them. */ | |
909 | bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0); | |
910 | if (bit < O2NM_MAX_NODES) { | |
2759236f | 911 | mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to " |
c03872f5 KH |
912 | "recover before lock mastery can begin\n", |
913 | dlm->name, namelen, (char *)lockid, bit); | |
914 | wait_on_recovery = 1; | |
915 | } | |
6714d8e8 KH |
916 | } |
917 | ||
918 | /* at this point there is either a DLM_MLE_BLOCK or a | |
919 | * DLM_MLE_MASTER on the master list, so it's safe to add the | |
920 | * lockres to the hashtable. anyone who finds the lock will | |
921 | * still have to wait on the IN_PROGRESS. */ | |
922 | ||
923 | /* finally add the lockres to its hash bucket */ | |
924 | __dlm_insert_lockres(dlm, res); | |
ba2bf218 KH |
925 | /* since this lockres is new it doesnt not require the spinlock */ |
926 | dlm_lockres_grab_inflight_ref_new(dlm, res); | |
927 | ||
928 | /* if this node does not become the master make sure to drop | |
929 | * this inflight reference below */ | |
930 | drop_inflight_if_nonlocal = 1; | |
931 | ||
6714d8e8 KH |
932 | /* get an extra ref on the mle in case this is a BLOCK |
933 | * if so, the creator of the BLOCK may try to put the last | |
934 | * ref at this time in the assert master handler, so we | |
935 | * need an extra one to keep from a bad ptr deref. */ | |
a2bf0477 | 936 | dlm_get_mle_inuse(mle); |
6714d8e8 KH |
937 | spin_unlock(&dlm->master_lock); |
938 | spin_unlock(&dlm->spinlock); | |
939 | ||
e7e69eb3 | 940 | redo_request: |
c03872f5 KH |
941 | while (wait_on_recovery) { |
942 | /* any cluster changes that occurred after dropping the | |
943 | * dlm spinlock would be detectable be a change on the mle, | |
944 | * so we only need to clear out the recovery map once. */ | |
945 | if (dlm_is_recovery_lock(lockid, namelen)) { | |
946 | mlog(ML_NOTICE, "%s: recovery map is not empty, but " | |
947 | "must master $RECOVERY lock now\n", dlm->name); | |
948 | if (!dlm_pre_master_reco_lockres(dlm, res)) | |
949 | wait_on_recovery = 0; | |
950 | else { | |
951 | mlog(0, "%s: waiting 500ms for heartbeat state " | |
952 | "change\n", dlm->name); | |
953 | msleep(500); | |
954 | } | |
955 | continue; | |
956 | } | |
957 | ||
958 | dlm_kick_recovery_thread(dlm); | |
aa087b84 | 959 | msleep(1000); |
c03872f5 KH |
960 | dlm_wait_for_recovery(dlm); |
961 | ||
962 | spin_lock(&dlm->spinlock); | |
963 | bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0); | |
964 | if (bit < O2NM_MAX_NODES) { | |
2759236f | 965 | mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to " |
c03872f5 KH |
966 | "recover before lock mastery can begin\n", |
967 | dlm->name, namelen, (char *)lockid, bit); | |
968 | wait_on_recovery = 1; | |
969 | } else | |
970 | wait_on_recovery = 0; | |
971 | spin_unlock(&dlm->spinlock); | |
b7084ab5 KH |
972 | |
973 | if (wait_on_recovery) | |
974 | dlm_wait_for_node_recovery(dlm, bit, 10000); | |
c03872f5 KH |
975 | } |
976 | ||
6714d8e8 KH |
977 | /* must wait for lock to be mastered elsewhere */ |
978 | if (blocked) | |
979 | goto wait; | |
980 | ||
6714d8e8 KH |
981 | ret = -EINVAL; |
982 | dlm_node_iter_init(mle->vote_map, &iter); | |
983 | while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { | |
ba2bf218 | 984 | ret = dlm_do_master_request(res, mle, nodenum); |
6714d8e8 KH |
985 | if (ret < 0) |
986 | mlog_errno(ret); | |
987 | if (mle->master != O2NM_MAX_NODES) { | |
988 | /* found a master ! */ | |
9c6510a5 KH |
989 | if (mle->master <= nodenum) |
990 | break; | |
991 | /* if our master request has not reached the master | |
992 | * yet, keep going until it does. this is how the | |
993 | * master will know that asserts are needed back to | |
994 | * the lower nodes. */ | |
995 | mlog(0, "%s:%.*s: requests only up to %u but master " | |
996 | "is %u, keep going\n", dlm->name, namelen, | |
997 | lockid, nodenum, mle->master); | |
6714d8e8 KH |
998 | } |
999 | } | |
1000 | ||
1001 | wait: | |
1002 | /* keep going until the response map includes all nodes */ | |
1003 | ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked); | |
1004 | if (ret < 0) { | |
e7e69eb3 | 1005 | wait_on_recovery = 1; |
6714d8e8 KH |
1006 | mlog(0, "%s:%.*s: node map changed, redo the " |
1007 | "master request now, blocked=%d\n", | |
1008 | dlm->name, res->lockname.len, | |
1009 | res->lockname.name, blocked); | |
1010 | if (++tries > 20) { | |
1011 | mlog(ML_ERROR, "%s:%.*s: spinning on " | |
1012 | "dlm_wait_for_lock_mastery, blocked=%d\n", | |
1013 | dlm->name, res->lockname.len, | |
1014 | res->lockname.name, blocked); | |
1015 | dlm_print_one_lock_resource(res); | |
8a9343fa | 1016 | dlm_print_one_mle(mle); |
6714d8e8 KH |
1017 | tries = 0; |
1018 | } | |
1019 | goto redo_request; | |
1020 | } | |
1021 | ||
1022 | mlog(0, "lockres mastered by %u\n", res->owner); | |
1023 | /* make sure we never continue without this */ | |
1024 | BUG_ON(res->owner == O2NM_MAX_NODES); | |
1025 | ||
1026 | /* master is known, detach if not already detached */ | |
1027 | dlm_mle_detach_hb_events(dlm, mle); | |
1028 | dlm_put_mle(mle); | |
1029 | /* put the extra ref */ | |
a2bf0477 | 1030 | dlm_put_mle_inuse(mle); |
6714d8e8 KH |
1031 | |
1032 | wake_waiters: | |
1033 | spin_lock(&res->spinlock); | |
ba2bf218 KH |
1034 | if (res->owner != dlm->node_num && drop_inflight_if_nonlocal) |
1035 | dlm_lockres_drop_inflight_ref(dlm, res); | |
6714d8e8 KH |
1036 | res->state &= ~DLM_LOCK_RES_IN_PROGRESS; |
1037 | spin_unlock(&res->spinlock); | |
1038 | wake_up(&res->wq); | |
1039 | ||
1040 | leave: | |
1041 | /* need to free the unused mle */ | |
1042 | if (alloc_mle) | |
1043 | kmem_cache_free(dlm_mle_cache, alloc_mle); | |
1044 | ||
1045 | return res; | |
1046 | } | |
1047 | ||
1048 | ||
1049 | #define DLM_MASTERY_TIMEOUT_MS 5000 | |
1050 | ||
1051 | static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm, | |
1052 | struct dlm_lock_resource *res, | |
1053 | struct dlm_master_list_entry *mle, | |
1054 | int *blocked) | |
1055 | { | |
1056 | u8 m; | |
1057 | int ret, bit; | |
1058 | int map_changed, voting_done; | |
1059 | int assert, sleep; | |
1060 | ||
1061 | recheck: | |
1062 | ret = 0; | |
1063 | assert = 0; | |
1064 | ||
1065 | /* check if another node has already become the owner */ | |
1066 | spin_lock(&res->spinlock); | |
1067 | if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) { | |
9c6510a5 KH |
1068 | mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name, |
1069 | res->lockname.len, res->lockname.name, res->owner); | |
6714d8e8 | 1070 | spin_unlock(&res->spinlock); |
9c6510a5 KH |
1071 | /* this will cause the master to re-assert across |
1072 | * the whole cluster, freeing up mles */ | |
588e0090 | 1073 | if (res->owner != dlm->node_num) { |
ba2bf218 | 1074 | ret = dlm_do_master_request(res, mle, res->owner); |
588e0090 KH |
1075 | if (ret < 0) { |
1076 | /* give recovery a chance to run */ | |
1077 | mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret); | |
1078 | msleep(500); | |
1079 | goto recheck; | |
1080 | } | |
9c6510a5 KH |
1081 | } |
1082 | ret = 0; | |
6714d8e8 KH |
1083 | goto leave; |
1084 | } | |
1085 | spin_unlock(&res->spinlock); | |
1086 | ||
1087 | spin_lock(&mle->spinlock); | |
1088 | m = mle->master; | |
1089 | map_changed = (memcmp(mle->vote_map, mle->node_map, | |
1090 | sizeof(mle->vote_map)) != 0); | |
1091 | voting_done = (memcmp(mle->vote_map, mle->response_map, | |
1092 | sizeof(mle->vote_map)) == 0); | |
1093 | ||
1094 | /* restart if we hit any errors */ | |
1095 | if (map_changed) { | |
1096 | int b; | |
1097 | mlog(0, "%s: %.*s: node map changed, restarting\n", | |
1098 | dlm->name, res->lockname.len, res->lockname.name); | |
1099 | ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked); | |
1100 | b = (mle->type == DLM_MLE_BLOCK); | |
1101 | if ((*blocked && !b) || (!*blocked && b)) { | |
1102 | mlog(0, "%s:%.*s: status change: old=%d new=%d\n", | |
1103 | dlm->name, res->lockname.len, res->lockname.name, | |
1104 | *blocked, b); | |
1105 | *blocked = b; | |
1106 | } | |
1107 | spin_unlock(&mle->spinlock); | |
1108 | if (ret < 0) { | |
1109 | mlog_errno(ret); | |
1110 | goto leave; | |
1111 | } | |
1112 | mlog(0, "%s:%.*s: restart lock mastery succeeded, " | |
1113 | "rechecking now\n", dlm->name, res->lockname.len, | |
1114 | res->lockname.name); | |
1115 | goto recheck; | |
aa852354 KH |
1116 | } else { |
1117 | if (!voting_done) { | |
1118 | mlog(0, "map not changed and voting not done " | |
1119 | "for %s:%.*s\n", dlm->name, res->lockname.len, | |
1120 | res->lockname.name); | |
1121 | } | |
6714d8e8 KH |
1122 | } |
1123 | ||
1124 | if (m != O2NM_MAX_NODES) { | |
1125 | /* another node has done an assert! | |
1126 | * all done! */ | |
1127 | sleep = 0; | |
1128 | } else { | |
1129 | sleep = 1; | |
1130 | /* have all nodes responded? */ | |
1131 | if (voting_done && !*blocked) { | |
1132 | bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0); | |
1133 | if (dlm->node_num <= bit) { | |
1134 | /* my node number is lowest. | |
1135 | * now tell other nodes that I am | |
1136 | * mastering this. */ | |
1137 | mle->master = dlm->node_num; | |
ba2bf218 KH |
1138 | /* ref was grabbed in get_lock_resource |
1139 | * will be dropped in dlmlock_master */ | |
6714d8e8 KH |
1140 | assert = 1; |
1141 | sleep = 0; | |
1142 | } | |
1143 | /* if voting is done, but we have not received | |
1144 | * an assert master yet, we must sleep */ | |
1145 | } | |
1146 | } | |
1147 | ||
1148 | spin_unlock(&mle->spinlock); | |
1149 | ||
1150 | /* sleep if we haven't finished voting yet */ | |
1151 | if (sleep) { | |
1152 | unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS); | |
1153 | ||
1154 | /* | |
1155 | if (atomic_read(&mle->mle_refs.refcount) < 2) | |
1156 | mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle, | |
1157 | atomic_read(&mle->mle_refs.refcount), | |
1158 | res->lockname.len, res->lockname.name); | |
1159 | */ | |
1160 | atomic_set(&mle->woken, 0); | |
1161 | (void)wait_event_timeout(mle->wq, | |
1162 | (atomic_read(&mle->woken) == 1), | |
1163 | timeo); | |
1164 | if (res->owner == O2NM_MAX_NODES) { | |
ba2bf218 KH |
1165 | mlog(0, "%s:%.*s: waiting again\n", dlm->name, |
1166 | res->lockname.len, res->lockname.name); | |
6714d8e8 KH |
1167 | goto recheck; |
1168 | } | |
1169 | mlog(0, "done waiting, master is %u\n", res->owner); | |
1170 | ret = 0; | |
1171 | goto leave; | |
1172 | } | |
1173 | ||
1174 | ret = 0; /* done */ | |
1175 | if (assert) { | |
1176 | m = dlm->node_num; | |
1177 | mlog(0, "about to master %.*s here, this=%u\n", | |
1178 | res->lockname.len, res->lockname.name, m); | |
ba2bf218 | 1179 | ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0); |
6714d8e8 KH |
1180 | if (ret) { |
1181 | /* This is a failure in the network path, | |
1182 | * not in the response to the assert_master | |
1183 | * (any nonzero response is a BUG on this node). | |
1184 | * Most likely a socket just got disconnected | |
1185 | * due to node death. */ | |
1186 | mlog_errno(ret); | |
1187 | } | |
1188 | /* no longer need to restart lock mastery. | |
1189 | * all living nodes have been contacted. */ | |
1190 | ret = 0; | |
1191 | } | |
1192 | ||
1193 | /* set the lockres owner */ | |
1194 | spin_lock(&res->spinlock); | |
ba2bf218 KH |
1195 | /* mastery reference obtained either during |
1196 | * assert_master_handler or in get_lock_resource */ | |
6714d8e8 KH |
1197 | dlm_change_lockres_owner(dlm, res, m); |
1198 | spin_unlock(&res->spinlock); | |
1199 | ||
1200 | leave: | |
1201 | return ret; | |
1202 | } | |
1203 | ||
1204 | struct dlm_bitmap_diff_iter | |
1205 | { | |
1206 | int curnode; | |
1207 | unsigned long *orig_bm; | |
1208 | unsigned long *cur_bm; | |
1209 | unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)]; | |
1210 | }; | |
1211 | ||
1212 | enum dlm_node_state_change | |
1213 | { | |
1214 | NODE_DOWN = -1, | |
1215 | NODE_NO_CHANGE = 0, | |
1216 | NODE_UP | |
1217 | }; | |
1218 | ||
1219 | static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter, | |
1220 | unsigned long *orig_bm, | |
1221 | unsigned long *cur_bm) | |
1222 | { | |
1223 | unsigned long p1, p2; | |
1224 | int i; | |
1225 | ||
1226 | iter->curnode = -1; | |
1227 | iter->orig_bm = orig_bm; | |
1228 | iter->cur_bm = cur_bm; | |
1229 | ||
1230 | for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) { | |
1231 | p1 = *(iter->orig_bm + i); | |
1232 | p2 = *(iter->cur_bm + i); | |
1233 | iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1); | |
1234 | } | |
1235 | } | |
1236 | ||
1237 | static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter, | |
1238 | enum dlm_node_state_change *state) | |
1239 | { | |
1240 | int bit; | |
1241 | ||
1242 | if (iter->curnode >= O2NM_MAX_NODES) | |
1243 | return -ENOENT; | |
1244 | ||
1245 | bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES, | |
1246 | iter->curnode+1); | |
1247 | if (bit >= O2NM_MAX_NODES) { | |
1248 | iter->curnode = O2NM_MAX_NODES; | |
1249 | return -ENOENT; | |
1250 | } | |
1251 | ||
1252 | /* if it was there in the original then this node died */ | |
1253 | if (test_bit(bit, iter->orig_bm)) | |
1254 | *state = NODE_DOWN; | |
1255 | else | |
1256 | *state = NODE_UP; | |
1257 | ||
1258 | iter->curnode = bit; | |
1259 | return bit; | |
1260 | } | |
1261 | ||
1262 | ||
1263 | static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm, | |
1264 | struct dlm_lock_resource *res, | |
1265 | struct dlm_master_list_entry *mle, | |
1266 | int blocked) | |
1267 | { | |
1268 | struct dlm_bitmap_diff_iter bdi; | |
1269 | enum dlm_node_state_change sc; | |
1270 | int node; | |
1271 | int ret = 0; | |
1272 | ||
1273 | mlog(0, "something happened such that the " | |
1274 | "master process may need to be restarted!\n"); | |
1275 | ||
1276 | assert_spin_locked(&mle->spinlock); | |
1277 | ||
1278 | dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map); | |
1279 | node = dlm_bitmap_diff_iter_next(&bdi, &sc); | |
1280 | while (node >= 0) { | |
1281 | if (sc == NODE_UP) { | |
e2faea4c KH |
1282 | /* a node came up. clear any old vote from |
1283 | * the response map and set it in the vote map | |
1284 | * then restart the mastery. */ | |
1285 | mlog(ML_NOTICE, "node %d up while restarting\n", node); | |
6714d8e8 KH |
1286 | |
1287 | /* redo the master request, but only for the new node */ | |
1288 | mlog(0, "sending request to new node\n"); | |
1289 | clear_bit(node, mle->response_map); | |
1290 | set_bit(node, mle->vote_map); | |
1291 | } else { | |
1292 | mlog(ML_ERROR, "node down! %d\n", node); | |
6714d8e8 KH |
1293 | if (blocked) { |
1294 | int lowest = find_next_bit(mle->maybe_map, | |
1295 | O2NM_MAX_NODES, 0); | |
1296 | ||
1297 | /* act like it was never there */ | |
1298 | clear_bit(node, mle->maybe_map); | |
1299 | ||
e7e69eb3 KH |
1300 | if (node == lowest) { |
1301 | mlog(0, "expected master %u died" | |
1302 | " while this node was blocked " | |
1303 | "waiting on it!\n", node); | |
1304 | lowest = find_next_bit(mle->maybe_map, | |
1305 | O2NM_MAX_NODES, | |
1306 | lowest+1); | |
1307 | if (lowest < O2NM_MAX_NODES) { | |
1308 | mlog(0, "%s:%.*s:still " | |
1309 | "blocked. waiting on %u " | |
1310 | "now\n", dlm->name, | |
1311 | res->lockname.len, | |
1312 | res->lockname.name, | |
1313 | lowest); | |
1314 | } else { | |
1315 | /* mle is an MLE_BLOCK, but | |
1316 | * there is now nothing left to | |
1317 | * block on. we need to return | |
1318 | * all the way back out and try | |
1319 | * again with an MLE_MASTER. | |
1320 | * dlm_do_local_recovery_cleanup | |
1321 | * has already run, so the mle | |
1322 | * refcount is ok */ | |
1323 | mlog(0, "%s:%.*s: no " | |
1324 | "longer blocking. try to " | |
1325 | "master this here\n", | |
1326 | dlm->name, | |
1327 | res->lockname.len, | |
1328 | res->lockname.name); | |
1329 | mle->type = DLM_MLE_MASTER; | |
1330 | mle->u.res = res; | |
1331 | } | |
6714d8e8 | 1332 | } |
6714d8e8 KH |
1333 | } |
1334 | ||
e7e69eb3 KH |
1335 | /* now blank out everything, as if we had never |
1336 | * contacted anyone */ | |
1337 | memset(mle->maybe_map, 0, sizeof(mle->maybe_map)); | |
1338 | memset(mle->response_map, 0, sizeof(mle->response_map)); | |
1339 | /* reset the vote_map to the current node_map */ | |
1340 | memcpy(mle->vote_map, mle->node_map, | |
1341 | sizeof(mle->node_map)); | |
1342 | /* put myself into the maybe map */ | |
1343 | if (mle->type != DLM_MLE_BLOCK) | |
1344 | set_bit(dlm->node_num, mle->maybe_map); | |
6714d8e8 KH |
1345 | } |
1346 | ret = -EAGAIN; | |
6714d8e8 KH |
1347 | node = dlm_bitmap_diff_iter_next(&bdi, &sc); |
1348 | } | |
1349 | return ret; | |
1350 | } | |
1351 | ||
1352 | ||
1353 | /* | |
1354 | * DLM_MASTER_REQUEST_MSG | |
1355 | * | |
1356 | * returns: 0 on success, | |
1357 | * -errno on a network error | |
1358 | * | |
1359 | * on error, the caller should assume the target node is "dead" | |
1360 | * | |
1361 | */ | |
1362 | ||
ba2bf218 KH |
1363 | static int dlm_do_master_request(struct dlm_lock_resource *res, |
1364 | struct dlm_master_list_entry *mle, int to) | |
6714d8e8 KH |
1365 | { |
1366 | struct dlm_ctxt *dlm = mle->dlm; | |
1367 | struct dlm_master_request request; | |
1368 | int ret, response=0, resend; | |
1369 | ||
1370 | memset(&request, 0, sizeof(request)); | |
1371 | request.node_idx = dlm->node_num; | |
1372 | ||
1373 | BUG_ON(mle->type == DLM_MLE_MIGRATION); | |
1374 | ||
1375 | if (mle->type != DLM_MLE_MASTER) { | |
1376 | request.namelen = mle->u.name.len; | |
1377 | memcpy(request.name, mle->u.name.name, request.namelen); | |
1378 | } else { | |
1379 | request.namelen = mle->u.res->lockname.len; | |
1380 | memcpy(request.name, mle->u.res->lockname.name, | |
1381 | request.namelen); | |
1382 | } | |
1383 | ||
1384 | again: | |
1385 | ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request, | |
1386 | sizeof(request), to, &response); | |
1387 | if (ret < 0) { | |
1388 | if (ret == -ESRCH) { | |
1389 | /* should never happen */ | |
1390 | mlog(ML_ERROR, "TCP stack not ready!\n"); | |
1391 | BUG(); | |
1392 | } else if (ret == -EINVAL) { | |
1393 | mlog(ML_ERROR, "bad args passed to o2net!\n"); | |
1394 | BUG(); | |
1395 | } else if (ret == -ENOMEM) { | |
1396 | mlog(ML_ERROR, "out of memory while trying to send " | |
1397 | "network message! retrying\n"); | |
1398 | /* this is totally crude */ | |
1399 | msleep(50); | |
1400 | goto again; | |
1401 | } else if (!dlm_is_host_down(ret)) { | |
1402 | /* not a network error. bad. */ | |
1403 | mlog_errno(ret); | |
1404 | mlog(ML_ERROR, "unhandled error!"); | |
1405 | BUG(); | |
1406 | } | |
1407 | /* all other errors should be network errors, | |
1408 | * and likely indicate node death */ | |
1409 | mlog(ML_ERROR, "link to %d went down!\n", to); | |
1410 | goto out; | |
1411 | } | |
1412 | ||
1413 | ret = 0; | |
1414 | resend = 0; | |
1415 | spin_lock(&mle->spinlock); | |
1416 | switch (response) { | |
1417 | case DLM_MASTER_RESP_YES: | |
1418 | set_bit(to, mle->response_map); | |
1419 | mlog(0, "node %u is the master, response=YES\n", to); | |
ba2bf218 KH |
1420 | mlog(0, "%s:%.*s: master node %u now knows I have a " |
1421 | "reference\n", dlm->name, res->lockname.len, | |
1422 | res->lockname.name, to); | |
6714d8e8 KH |
1423 | mle->master = to; |
1424 | break; | |
1425 | case DLM_MASTER_RESP_NO: | |
1426 | mlog(0, "node %u not master, response=NO\n", to); | |
1427 | set_bit(to, mle->response_map); | |
1428 | break; | |
1429 | case DLM_MASTER_RESP_MAYBE: | |
1430 | mlog(0, "node %u not master, response=MAYBE\n", to); | |
1431 | set_bit(to, mle->response_map); | |
1432 | set_bit(to, mle->maybe_map); | |
1433 | break; | |
1434 | case DLM_MASTER_RESP_ERROR: | |
1435 | mlog(0, "node %u hit an error, resending\n", to); | |
1436 | resend = 1; | |
1437 | response = 0; | |
1438 | break; | |
1439 | default: | |
1440 | mlog(ML_ERROR, "bad response! %u\n", response); | |
1441 | BUG(); | |
1442 | } | |
1443 | spin_unlock(&mle->spinlock); | |
1444 | if (resend) { | |
1445 | /* this is also totally crude */ | |
1446 | msleep(50); | |
1447 | goto again; | |
1448 | } | |
1449 | ||
1450 | out: | |
1451 | return ret; | |
1452 | } | |
1453 | ||
1454 | /* | |
1455 | * locks that can be taken here: | |
1456 | * dlm->spinlock | |
1457 | * res->spinlock | |
1458 | * mle->spinlock | |
1459 | * dlm->master_list | |
1460 | * | |
1461 | * if possible, TRIM THIS DOWN!!! | |
1462 | */ | |
d74c9803 KH |
1463 | int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data, |
1464 | void **ret_data) | |
6714d8e8 KH |
1465 | { |
1466 | u8 response = DLM_MASTER_RESP_MAYBE; | |
1467 | struct dlm_ctxt *dlm = data; | |
9c6510a5 | 1468 | struct dlm_lock_resource *res = NULL; |
6714d8e8 KH |
1469 | struct dlm_master_request *request = (struct dlm_master_request *) msg->buf; |
1470 | struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL; | |
1471 | char *name; | |
a3d33291 | 1472 | unsigned int namelen, hash; |
6714d8e8 KH |
1473 | int found, ret; |
1474 | int set_maybe; | |
9c6510a5 | 1475 | int dispatch_assert = 0; |
6714d8e8 KH |
1476 | |
1477 | if (!dlm_grab(dlm)) | |
1478 | return DLM_MASTER_RESP_NO; | |
1479 | ||
1480 | if (!dlm_domain_fully_joined(dlm)) { | |
1481 | response = DLM_MASTER_RESP_NO; | |
1482 | goto send_response; | |
1483 | } | |
1484 | ||
1485 | name = request->name; | |
1486 | namelen = request->namelen; | |
a3d33291 | 1487 | hash = dlm_lockid_hash(name, namelen); |
6714d8e8 KH |
1488 | |
1489 | if (namelen > DLM_LOCKID_NAME_MAX) { | |
1490 | response = DLM_IVBUFLEN; | |
1491 | goto send_response; | |
1492 | } | |
1493 | ||
1494 | way_up_top: | |
1495 | spin_lock(&dlm->spinlock); | |
a3d33291 | 1496 | res = __dlm_lookup_lockres(dlm, name, namelen, hash); |
6714d8e8 KH |
1497 | if (res) { |
1498 | spin_unlock(&dlm->spinlock); | |
1499 | ||
1500 | /* take care of the easy cases up front */ | |
1501 | spin_lock(&res->spinlock); | |
1cd04dbe KH |
1502 | if (res->state & (DLM_LOCK_RES_RECOVERING| |
1503 | DLM_LOCK_RES_MIGRATING)) { | |
6714d8e8 KH |
1504 | spin_unlock(&res->spinlock); |
1505 | mlog(0, "returning DLM_MASTER_RESP_ERROR since res is " | |
1cd04dbe | 1506 | "being recovered/migrated\n"); |
6714d8e8 KH |
1507 | response = DLM_MASTER_RESP_ERROR; |
1508 | if (mle) | |
1509 | kmem_cache_free(dlm_mle_cache, mle); | |
1510 | goto send_response; | |
1511 | } | |
1512 | ||
1513 | if (res->owner == dlm->node_num) { | |
ba2bf218 KH |
1514 | mlog(0, "%s:%.*s: setting bit %u in refmap\n", |
1515 | dlm->name, namelen, name, request->node_idx); | |
1516 | dlm_lockres_set_refmap_bit(request->node_idx, res); | |
6714d8e8 | 1517 | spin_unlock(&res->spinlock); |
6714d8e8 KH |
1518 | response = DLM_MASTER_RESP_YES; |
1519 | if (mle) | |
1520 | kmem_cache_free(dlm_mle_cache, mle); | |
1521 | ||
1522 | /* this node is the owner. | |
1523 | * there is some extra work that needs to | |
1524 | * happen now. the requesting node has | |
1525 | * caused all nodes up to this one to | |
1526 | * create mles. this node now needs to | |
1527 | * go back and clean those up. */ | |
9c6510a5 | 1528 | dispatch_assert = 1; |
6714d8e8 KH |
1529 | goto send_response; |
1530 | } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) { | |
1531 | spin_unlock(&res->spinlock); | |
1532 | // mlog(0, "node %u is the master\n", res->owner); | |
1533 | response = DLM_MASTER_RESP_NO; | |
1534 | if (mle) | |
1535 | kmem_cache_free(dlm_mle_cache, mle); | |
1536 | goto send_response; | |
1537 | } | |
1538 | ||
1539 | /* ok, there is no owner. either this node is | |
1540 | * being blocked, or it is actively trying to | |
1541 | * master this lock. */ | |
1542 | if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) { | |
1543 | mlog(ML_ERROR, "lock with no owner should be " | |
1544 | "in-progress!\n"); | |
1545 | BUG(); | |
1546 | } | |
1547 | ||
1548 | // mlog(0, "lockres is in progress...\n"); | |
1549 | spin_lock(&dlm->master_lock); | |
1550 | found = dlm_find_mle(dlm, &tmpmle, name, namelen); | |
1551 | if (!found) { | |
1552 | mlog(ML_ERROR, "no mle found for this lock!\n"); | |
1553 | BUG(); | |
1554 | } | |
1555 | set_maybe = 1; | |
1556 | spin_lock(&tmpmle->spinlock); | |
1557 | if (tmpmle->type == DLM_MLE_BLOCK) { | |
1558 | // mlog(0, "this node is waiting for " | |
1559 | // "lockres to be mastered\n"); | |
1560 | response = DLM_MASTER_RESP_NO; | |
1561 | } else if (tmpmle->type == DLM_MLE_MIGRATION) { | |
1562 | mlog(0, "node %u is master, but trying to migrate to " | |
1563 | "node %u.\n", tmpmle->master, tmpmle->new_master); | |
1564 | if (tmpmle->master == dlm->node_num) { | |
6714d8e8 KH |
1565 | mlog(ML_ERROR, "no owner on lockres, but this " |
1566 | "node is trying to migrate it to %u?!\n", | |
1567 | tmpmle->new_master); | |
1568 | BUG(); | |
1569 | } else { | |
1570 | /* the real master can respond on its own */ | |
1571 | response = DLM_MASTER_RESP_NO; | |
1572 | } | |
1573 | } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) { | |
1574 | set_maybe = 0; | |
9c6510a5 | 1575 | if (tmpmle->master == dlm->node_num) { |
6714d8e8 | 1576 | response = DLM_MASTER_RESP_YES; |
9c6510a5 KH |
1577 | /* this node will be the owner. |
1578 | * go back and clean the mles on any | |
1579 | * other nodes */ | |
1580 | dispatch_assert = 1; | |
ba2bf218 KH |
1581 | dlm_lockres_set_refmap_bit(request->node_idx, res); |
1582 | mlog(0, "%s:%.*s: setting bit %u in refmap\n", | |
1583 | dlm->name, namelen, name, | |
1584 | request->node_idx); | |
9c6510a5 | 1585 | } else |
6714d8e8 KH |
1586 | response = DLM_MASTER_RESP_NO; |
1587 | } else { | |
1588 | // mlog(0, "this node is attempting to " | |
1589 | // "master lockres\n"); | |
1590 | response = DLM_MASTER_RESP_MAYBE; | |
1591 | } | |
1592 | if (set_maybe) | |
1593 | set_bit(request->node_idx, tmpmle->maybe_map); | |
1594 | spin_unlock(&tmpmle->spinlock); | |
1595 | ||
1596 | spin_unlock(&dlm->master_lock); | |
1597 | spin_unlock(&res->spinlock); | |
1598 | ||
1599 | /* keep the mle attached to heartbeat events */ | |
1600 | dlm_put_mle(tmpmle); | |
1601 | if (mle) | |
1602 | kmem_cache_free(dlm_mle_cache, mle); | |
1603 | goto send_response; | |
1604 | } | |
1605 | ||
1606 | /* | |
1607 | * lockres doesn't exist on this node | |
1608 | * if there is an MLE_BLOCK, return NO | |
1609 | * if there is an MLE_MASTER, return MAYBE | |
1610 | * otherwise, add an MLE_BLOCK, return NO | |
1611 | */ | |
1612 | spin_lock(&dlm->master_lock); | |
1613 | found = dlm_find_mle(dlm, &tmpmle, name, namelen); | |
1614 | if (!found) { | |
1615 | /* this lockid has never been seen on this node yet */ | |
1616 | // mlog(0, "no mle found\n"); | |
1617 | if (!mle) { | |
1618 | spin_unlock(&dlm->master_lock); | |
1619 | spin_unlock(&dlm->spinlock); | |
1620 | ||
1621 | mle = (struct dlm_master_list_entry *) | |
ad8100e0 | 1622 | kmem_cache_alloc(dlm_mle_cache, GFP_NOFS); |
6714d8e8 | 1623 | if (!mle) { |
6714d8e8 | 1624 | response = DLM_MASTER_RESP_ERROR; |
9c6510a5 | 1625 | mlog_errno(-ENOMEM); |
6714d8e8 KH |
1626 | goto send_response; |
1627 | } | |
6714d8e8 KH |
1628 | goto way_up_top; |
1629 | } | |
1630 | ||
1631 | // mlog(0, "this is second time thru, already allocated, " | |
1632 | // "add the block.\n"); | |
41b8c8a1 | 1633 | dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen); |
6714d8e8 KH |
1634 | set_bit(request->node_idx, mle->maybe_map); |
1635 | list_add(&mle->list, &dlm->master_list); | |
1636 | response = DLM_MASTER_RESP_NO; | |
1637 | } else { | |
1638 | // mlog(0, "mle was found\n"); | |
1639 | set_maybe = 1; | |
1640 | spin_lock(&tmpmle->spinlock); | |
9c6510a5 KH |
1641 | if (tmpmle->master == dlm->node_num) { |
1642 | mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n"); | |
1643 | BUG(); | |
1644 | } | |
6714d8e8 KH |
1645 | if (tmpmle->type == DLM_MLE_BLOCK) |
1646 | response = DLM_MASTER_RESP_NO; | |
1647 | else if (tmpmle->type == DLM_MLE_MIGRATION) { | |
1648 | mlog(0, "migration mle was found (%u->%u)\n", | |
1649 | tmpmle->master, tmpmle->new_master); | |
6714d8e8 KH |
1650 | /* real master can respond on its own */ |
1651 | response = DLM_MASTER_RESP_NO; | |
9c6510a5 KH |
1652 | } else |
1653 | response = DLM_MASTER_RESP_MAYBE; | |
6714d8e8 KH |
1654 | if (set_maybe) |
1655 | set_bit(request->node_idx, tmpmle->maybe_map); | |
1656 | spin_unlock(&tmpmle->spinlock); | |
1657 | } | |
1658 | spin_unlock(&dlm->master_lock); | |
1659 | spin_unlock(&dlm->spinlock); | |
1660 | ||
1661 | if (found) { | |
1662 | /* keep the mle attached to heartbeat events */ | |
1663 | dlm_put_mle(tmpmle); | |
1664 | } | |
1665 | send_response: | |
9c6510a5 KH |
1666 | |
1667 | if (dispatch_assert) { | |
1668 | if (response != DLM_MASTER_RESP_YES) | |
1669 | mlog(ML_ERROR, "invalid response %d\n", response); | |
1670 | if (!res) { | |
1671 | mlog(ML_ERROR, "bad lockres while trying to assert!\n"); | |
1672 | BUG(); | |
1673 | } | |
1674 | mlog(0, "%u is the owner of %.*s, cleaning everyone else\n", | |
1675 | dlm->node_num, res->lockname.len, res->lockname.name); | |
1676 | ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx, | |
1677 | DLM_ASSERT_MASTER_MLE_CLEANUP); | |
1678 | if (ret < 0) { | |
1679 | mlog(ML_ERROR, "failed to dispatch assert master work\n"); | |
1680 | response = DLM_MASTER_RESP_ERROR; | |
1681 | } | |
1682 | } | |
1683 | ||
6714d8e8 KH |
1684 | dlm_put(dlm); |
1685 | return response; | |
1686 | } | |
1687 | ||
1688 | /* | |
1689 | * DLM_ASSERT_MASTER_MSG | |
1690 | */ | |
1691 | ||
1692 | ||
1693 | /* | |
1694 | * NOTE: this can be used for debugging | |
1695 | * can periodically run all locks owned by this node | |
1696 | * and re-assert across the cluster... | |
1697 | */ | |
05488bbe AB |
1698 | static int dlm_do_assert_master(struct dlm_ctxt *dlm, |
1699 | struct dlm_lock_resource *res, | |
1700 | void *nodemap, u32 flags) | |
6714d8e8 KH |
1701 | { |
1702 | struct dlm_assert_master assert; | |
1703 | int to, tmpret; | |
1704 | struct dlm_node_iter iter; | |
1705 | int ret = 0; | |
9c6510a5 | 1706 | int reassert; |
ba2bf218 KH |
1707 | const char *lockname = res->lockname.name; |
1708 | unsigned int namelen = res->lockname.len; | |
6714d8e8 KH |
1709 | |
1710 | BUG_ON(namelen > O2NM_MAX_NAME_LEN); | |
f3f85464 SM |
1711 | |
1712 | spin_lock(&res->spinlock); | |
1713 | res->state |= DLM_LOCK_RES_SETREF_INPROG; | |
1714 | spin_unlock(&res->spinlock); | |
1715 | ||
9c6510a5 KH |
1716 | again: |
1717 | reassert = 0; | |
6714d8e8 KH |
1718 | |
1719 | /* note that if this nodemap is empty, it returns 0 */ | |
1720 | dlm_node_iter_init(nodemap, &iter); | |
1721 | while ((to = dlm_node_iter_next(&iter)) >= 0) { | |
1722 | int r = 0; | |
a9ee4c8a KH |
1723 | struct dlm_master_list_entry *mle = NULL; |
1724 | ||
6714d8e8 KH |
1725 | mlog(0, "sending assert master to %d (%.*s)\n", to, |
1726 | namelen, lockname); | |
1727 | memset(&assert, 0, sizeof(assert)); | |
1728 | assert.node_idx = dlm->node_num; | |
1729 | assert.namelen = namelen; | |
1730 | memcpy(assert.name, lockname, namelen); | |
1731 | assert.flags = cpu_to_be32(flags); | |
1732 | ||
1733 | tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key, | |
1734 | &assert, sizeof(assert), to, &r); | |
1735 | if (tmpret < 0) { | |
3b3b84a8 | 1736 | mlog(0, "assert_master returned %d!\n", tmpret); |
6714d8e8 | 1737 | if (!dlm_is_host_down(tmpret)) { |
3b3b84a8 | 1738 | mlog(ML_ERROR, "unhandled error=%d!\n", tmpret); |
6714d8e8 KH |
1739 | BUG(); |
1740 | } | |
1741 | /* a node died. finish out the rest of the nodes. */ | |
3b3b84a8 | 1742 | mlog(0, "link to %d went down!\n", to); |
6714d8e8 KH |
1743 | /* any nonzero status return will do */ |
1744 | ret = tmpret; | |
ba2bf218 | 1745 | r = 0; |
6714d8e8 KH |
1746 | } else if (r < 0) { |
1747 | /* ok, something horribly messed. kill thyself. */ | |
1748 | mlog(ML_ERROR,"during assert master of %.*s to %u, " | |
1749 | "got %d.\n", namelen, lockname, to, r); | |
a9ee4c8a KH |
1750 | spin_lock(&dlm->spinlock); |
1751 | spin_lock(&dlm->master_lock); | |
1752 | if (dlm_find_mle(dlm, &mle, (char *)lockname, | |
1753 | namelen)) { | |
1754 | dlm_print_one_mle(mle); | |
1755 | __dlm_put_mle(mle); | |
1756 | } | |
1757 | spin_unlock(&dlm->master_lock); | |
1758 | spin_unlock(&dlm->spinlock); | |
6714d8e8 | 1759 | BUG(); |
ba2bf218 KH |
1760 | } |
1761 | ||
1762 | if (r & DLM_ASSERT_RESPONSE_REASSERT && | |
1763 | !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) { | |
1764 | mlog(ML_ERROR, "%.*s: very strange, " | |
1765 | "master MLE but no lockres on %u\n", | |
1766 | namelen, lockname, to); | |
1767 | } | |
1768 | ||
1769 | if (r & DLM_ASSERT_RESPONSE_REASSERT) { | |
9c6510a5 KH |
1770 | mlog(0, "%.*s: node %u create mles on other " |
1771 | "nodes and requests a re-assert\n", | |
1772 | namelen, lockname, to); | |
1773 | reassert = 1; | |
6714d8e8 | 1774 | } |
ba2bf218 KH |
1775 | if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) { |
1776 | mlog(0, "%.*s: node %u has a reference to this " | |
1777 | "lockres, set the bit in the refmap\n", | |
1778 | namelen, lockname, to); | |
1779 | spin_lock(&res->spinlock); | |
1780 | dlm_lockres_set_refmap_bit(to, res); | |
1781 | spin_unlock(&res->spinlock); | |
1782 | } | |
6714d8e8 KH |
1783 | } |
1784 | ||
9c6510a5 KH |
1785 | if (reassert) |
1786 | goto again; | |
1787 | ||
f3f85464 SM |
1788 | spin_lock(&res->spinlock); |
1789 | res->state &= ~DLM_LOCK_RES_SETREF_INPROG; | |
1790 | spin_unlock(&res->spinlock); | |
1791 | wake_up(&res->wq); | |
1792 | ||
6714d8e8 KH |
1793 | return ret; |
1794 | } | |
1795 | ||
1796 | /* | |
1797 | * locks that can be taken here: | |
1798 | * dlm->spinlock | |
1799 | * res->spinlock | |
1800 | * mle->spinlock | |
1801 | * dlm->master_list | |
1802 | * | |
1803 | * if possible, TRIM THIS DOWN!!! | |
1804 | */ | |
d74c9803 KH |
1805 | int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data, |
1806 | void **ret_data) | |
6714d8e8 KH |
1807 | { |
1808 | struct dlm_ctxt *dlm = data; | |
1809 | struct dlm_master_list_entry *mle = NULL; | |
1810 | struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf; | |
1811 | struct dlm_lock_resource *res = NULL; | |
1812 | char *name; | |
a3d33291 | 1813 | unsigned int namelen, hash; |
6714d8e8 | 1814 | u32 flags; |
ba2bf218 | 1815 | int master_request = 0, have_lockres_ref = 0; |
9c6510a5 | 1816 | int ret = 0; |
6714d8e8 KH |
1817 | |
1818 | if (!dlm_grab(dlm)) | |
1819 | return 0; | |
1820 | ||
1821 | name = assert->name; | |
1822 | namelen = assert->namelen; | |
a3d33291 | 1823 | hash = dlm_lockid_hash(name, namelen); |
6714d8e8 KH |
1824 | flags = be32_to_cpu(assert->flags); |
1825 | ||
1826 | if (namelen > DLM_LOCKID_NAME_MAX) { | |
1827 | mlog(ML_ERROR, "Invalid name length!"); | |
1828 | goto done; | |
1829 | } | |
1830 | ||
1831 | spin_lock(&dlm->spinlock); | |
1832 | ||
1833 | if (flags) | |
1834 | mlog(0, "assert_master with flags: %u\n", flags); | |
1835 | ||
1836 | /* find the MLE */ | |
1837 | spin_lock(&dlm->master_lock); | |
1838 | if (!dlm_find_mle(dlm, &mle, name, namelen)) { | |
1839 | /* not an error, could be master just re-asserting */ | |
1840 | mlog(0, "just got an assert_master from %u, but no " | |
1841 | "MLE for it! (%.*s)\n", assert->node_idx, | |
1842 | namelen, name); | |
1843 | } else { | |
1844 | int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0); | |
1845 | if (bit >= O2NM_MAX_NODES) { | |
1846 | /* not necessarily an error, though less likely. | |
1847 | * could be master just re-asserting. */ | |
aa852354 | 1848 | mlog(0, "no bits set in the maybe_map, but %u " |
6714d8e8 KH |
1849 | "is asserting! (%.*s)\n", assert->node_idx, |
1850 | namelen, name); | |
1851 | } else if (bit != assert->node_idx) { | |
1852 | if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) { | |
1853 | mlog(0, "master %u was found, %u should " | |
1854 | "back off\n", assert->node_idx, bit); | |
1855 | } else { | |
1856 | /* with the fix for bug 569, a higher node | |
1857 | * number winning the mastery will respond | |
1858 | * YES to mastery requests, but this node | |
1859 | * had no way of knowing. let it pass. */ | |
aa852354 | 1860 | mlog(0, "%u is the lowest node, " |
6714d8e8 KH |
1861 | "%u is asserting. (%.*s) %u must " |
1862 | "have begun after %u won.\n", bit, | |
1863 | assert->node_idx, namelen, name, bit, | |
1864 | assert->node_idx); | |
1865 | } | |
1866 | } | |
2d1a868c KH |
1867 | if (mle->type == DLM_MLE_MIGRATION) { |
1868 | if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) { | |
1869 | mlog(0, "%s:%.*s: got cleanup assert" | |
1870 | " from %u for migration\n", | |
1871 | dlm->name, namelen, name, | |
1872 | assert->node_idx); | |
1873 | } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) { | |
1874 | mlog(0, "%s:%.*s: got unrelated assert" | |
1875 | " from %u for migration, ignoring\n", | |
1876 | dlm->name, namelen, name, | |
1877 | assert->node_idx); | |
1878 | __dlm_put_mle(mle); | |
1879 | spin_unlock(&dlm->master_lock); | |
1880 | spin_unlock(&dlm->spinlock); | |
1881 | goto done; | |
1882 | } | |
1883 | } | |
6714d8e8 KH |
1884 | } |
1885 | spin_unlock(&dlm->master_lock); | |
1886 | ||
1887 | /* ok everything checks out with the MLE | |
1888 | * now check to see if there is a lockres */ | |
a3d33291 | 1889 | res = __dlm_lookup_lockres(dlm, name, namelen, hash); |
6714d8e8 KH |
1890 | if (res) { |
1891 | spin_lock(&res->spinlock); | |
1892 | if (res->state & DLM_LOCK_RES_RECOVERING) { | |
1893 | mlog(ML_ERROR, "%u asserting but %.*s is " | |
1894 | "RECOVERING!\n", assert->node_idx, namelen, name); | |
1895 | goto kill; | |
1896 | } | |
1897 | if (!mle) { | |
dc2ed195 KH |
1898 | if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN && |
1899 | res->owner != assert->node_idx) { | |
6714d8e8 KH |
1900 | mlog(ML_ERROR, "assert_master from " |
1901 | "%u, but current owner is " | |
1902 | "%u! (%.*s)\n", | |
1903 | assert->node_idx, res->owner, | |
1904 | namelen, name); | |
1905 | goto kill; | |
1906 | } | |
1907 | } else if (mle->type != DLM_MLE_MIGRATION) { | |
1908 | if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) { | |
1909 | /* owner is just re-asserting */ | |
1910 | if (res->owner == assert->node_idx) { | |
1911 | mlog(0, "owner %u re-asserting on " | |
1912 | "lock %.*s\n", assert->node_idx, | |
1913 | namelen, name); | |
1914 | goto ok; | |
1915 | } | |
1916 | mlog(ML_ERROR, "got assert_master from " | |
1917 | "node %u, but %u is the owner! " | |
1918 | "(%.*s)\n", assert->node_idx, | |
1919 | res->owner, namelen, name); | |
1920 | goto kill; | |
1921 | } | |
1922 | if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) { | |
1923 | mlog(ML_ERROR, "got assert from %u, but lock " | |
1924 | "with no owner should be " | |
1925 | "in-progress! (%.*s)\n", | |
1926 | assert->node_idx, | |
1927 | namelen, name); | |
1928 | goto kill; | |
1929 | } | |
1930 | } else /* mle->type == DLM_MLE_MIGRATION */ { | |
1931 | /* should only be getting an assert from new master */ | |
1932 | if (assert->node_idx != mle->new_master) { | |
1933 | mlog(ML_ERROR, "got assert from %u, but " | |
1934 | "new master is %u, and old master " | |
1935 | "was %u (%.*s)\n", | |
1936 | assert->node_idx, mle->new_master, | |
1937 | mle->master, namelen, name); | |
1938 | goto kill; | |
1939 | } | |
1940 | ||
1941 | } | |
1942 | ok: | |
1943 | spin_unlock(&res->spinlock); | |
1944 | } | |
1945 | spin_unlock(&dlm->spinlock); | |
1946 | ||
1947 | // mlog(0, "woo! got an assert_master from node %u!\n", | |
1948 | // assert->node_idx); | |
1949 | if (mle) { | |
9c6510a5 KH |
1950 | int extra_ref = 0; |
1951 | int nn = -1; | |
a2bf0477 | 1952 | int rr, err = 0; |
6714d8e8 KH |
1953 | |
1954 | spin_lock(&mle->spinlock); | |
9c6510a5 KH |
1955 | if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION) |
1956 | extra_ref = 1; | |
1957 | else { | |
1958 | /* MASTER mle: if any bits set in the response map | |
1959 | * then the calling node needs to re-assert to clear | |
1960 | * up nodes that this node contacted */ | |
1961 | while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES, | |
1962 | nn+1)) < O2NM_MAX_NODES) { | |
1963 | if (nn != dlm->node_num && nn != assert->node_idx) | |
1964 | master_request = 1; | |
1965 | } | |
1966 | } | |
6714d8e8 KH |
1967 | mle->master = assert->node_idx; |
1968 | atomic_set(&mle->woken, 1); | |
1969 | wake_up(&mle->wq); | |
1970 | spin_unlock(&mle->spinlock); | |
1971 | ||
a2bf0477 | 1972 | if (res) { |
a6fa3640 | 1973 | int wake = 0; |
6714d8e8 | 1974 | spin_lock(&res->spinlock); |
a2bf0477 KH |
1975 | if (mle->type == DLM_MLE_MIGRATION) { |
1976 | mlog(0, "finishing off migration of lockres %.*s, " | |
1977 | "from %u to %u\n", | |
1978 | res->lockname.len, res->lockname.name, | |
1979 | dlm->node_num, mle->new_master); | |
1980 | res->state &= ~DLM_LOCK_RES_MIGRATING; | |
a6fa3640 | 1981 | wake = 1; |
a2bf0477 KH |
1982 | dlm_change_lockres_owner(dlm, res, mle->new_master); |
1983 | BUG_ON(res->state & DLM_LOCK_RES_DIRTY); | |
1984 | } else { | |
1985 | dlm_change_lockres_owner(dlm, res, mle->master); | |
1986 | } | |
6714d8e8 | 1987 | spin_unlock(&res->spinlock); |
ba2bf218 | 1988 | have_lockres_ref = 1; |
a6fa3640 KH |
1989 | if (wake) |
1990 | wake_up(&res->wq); | |
6714d8e8 | 1991 | } |
a2bf0477 KH |
1992 | |
1993 | /* master is known, detach if not already detached. | |
1994 | * ensures that only one assert_master call will happen | |
1995 | * on this mle. */ | |
1996 | spin_lock(&dlm->spinlock); | |
1997 | spin_lock(&dlm->master_lock); | |
1998 | ||
1999 | rr = atomic_read(&mle->mle_refs.refcount); | |
2000 | if (mle->inuse > 0) { | |
2001 | if (extra_ref && rr < 3) | |
2002 | err = 1; | |
2003 | else if (!extra_ref && rr < 2) | |
2004 | err = 1; | |
2005 | } else { | |
2006 | if (extra_ref && rr < 2) | |
2007 | err = 1; | |
2008 | else if (!extra_ref && rr < 1) | |
2009 | err = 1; | |
2010 | } | |
2011 | if (err) { | |
2012 | mlog(ML_ERROR, "%s:%.*s: got assert master from %u " | |
2013 | "that will mess up this node, refs=%d, extra=%d, " | |
2014 | "inuse=%d\n", dlm->name, namelen, name, | |
2015 | assert->node_idx, rr, extra_ref, mle->inuse); | |
2016 | dlm_print_one_mle(mle); | |
2017 | } | |
2018 | list_del_init(&mle->list); | |
2019 | __dlm_mle_detach_hb_events(dlm, mle); | |
2020 | __dlm_put_mle(mle); | |
6714d8e8 KH |
2021 | if (extra_ref) { |
2022 | /* the assert master message now balances the extra | |
2023 | * ref given by the master / migration request message. | |
2024 | * if this is the last put, it will be removed | |
2025 | * from the list. */ | |
a2bf0477 KH |
2026 | __dlm_put_mle(mle); |
2027 | } | |
2028 | spin_unlock(&dlm->master_lock); | |
2029 | spin_unlock(&dlm->spinlock); | |
2030 | } else if (res) { | |
2031 | if (res->owner != assert->node_idx) { | |
2032 | mlog(0, "assert_master from %u, but current " | |
2033 | "owner is %u (%.*s), no mle\n", assert->node_idx, | |
2034 | res->owner, namelen, name); | |
6714d8e8 KH |
2035 | } |
2036 | } | |
2037 | ||
2038 | done: | |
9c6510a5 | 2039 | ret = 0; |
3b8118cf KH |
2040 | if (res) { |
2041 | spin_lock(&res->spinlock); | |
2042 | res->state |= DLM_LOCK_RES_SETREF_INPROG; | |
2043 | spin_unlock(&res->spinlock); | |
2044 | *ret_data = (void *)res; | |
2045 | } | |
6714d8e8 | 2046 | dlm_put(dlm); |
9c6510a5 KH |
2047 | if (master_request) { |
2048 | mlog(0, "need to tell master to reassert\n"); | |
ba2bf218 KH |
2049 | /* positive. negative would shoot down the node. */ |
2050 | ret |= DLM_ASSERT_RESPONSE_REASSERT; | |
2051 | if (!have_lockres_ref) { | |
2052 | mlog(ML_ERROR, "strange, got assert from %u, MASTER " | |
2053 | "mle present here for %s:%.*s, but no lockres!\n", | |
2054 | assert->node_idx, dlm->name, namelen, name); | |
2055 | } | |
2056 | } | |
2057 | if (have_lockres_ref) { | |
2058 | /* let the master know we have a reference to the lockres */ | |
2059 | ret |= DLM_ASSERT_RESPONSE_MASTERY_REF; | |
2060 | mlog(0, "%s:%.*s: got assert from %u, need a ref\n", | |
2061 | dlm->name, namelen, name, assert->node_idx); | |
9c6510a5 KH |
2062 | } |
2063 | return ret; | |
6714d8e8 KH |
2064 | |
2065 | kill: | |
2066 | /* kill the caller! */ | |
a9ee4c8a KH |
2067 | mlog(ML_ERROR, "Bad message received from another node. Dumping state " |
2068 | "and killing the other node now! This node is OK and can continue.\n"); | |
2069 | __dlm_print_one_lock_resource(res); | |
6714d8e8 KH |
2070 | spin_unlock(&res->spinlock); |
2071 | spin_unlock(&dlm->spinlock); | |
3b8118cf | 2072 | *ret_data = (void *)res; |
6714d8e8 KH |
2073 | dlm_put(dlm); |
2074 | return -EINVAL; | |
2075 | } | |
2076 | ||
3b8118cf KH |
2077 | void dlm_assert_master_post_handler(int status, void *data, void *ret_data) |
2078 | { | |
2079 | struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data; | |
2080 | ||
2081 | if (ret_data) { | |
2082 | spin_lock(&res->spinlock); | |
2083 | res->state &= ~DLM_LOCK_RES_SETREF_INPROG; | |
2084 | spin_unlock(&res->spinlock); | |
2085 | wake_up(&res->wq); | |
2086 | dlm_lockres_put(res); | |
2087 | } | |
2088 | return; | |
2089 | } | |
2090 | ||
6714d8e8 KH |
2091 | int dlm_dispatch_assert_master(struct dlm_ctxt *dlm, |
2092 | struct dlm_lock_resource *res, | |
2093 | int ignore_higher, u8 request_from, u32 flags) | |
2094 | { | |
2095 | struct dlm_work_item *item; | |
cd861280 | 2096 | item = kzalloc(sizeof(*item), GFP_NOFS); |
6714d8e8 KH |
2097 | if (!item) |
2098 | return -ENOMEM; | |
2099 | ||
2100 | ||
2101 | /* queue up work for dlm_assert_master_worker */ | |
2102 | dlm_grab(dlm); /* get an extra ref for the work item */ | |
2103 | dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL); | |
2104 | item->u.am.lockres = res; /* already have a ref */ | |
2105 | /* can optionally ignore node numbers higher than this node */ | |
2106 | item->u.am.ignore_higher = ignore_higher; | |
2107 | item->u.am.request_from = request_from; | |
2108 | item->u.am.flags = flags; | |
2109 | ||
9c6510a5 KH |
2110 | if (ignore_higher) |
2111 | mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len, | |
2112 | res->lockname.name); | |
2113 | ||
6714d8e8 KH |
2114 | spin_lock(&dlm->work_lock); |
2115 | list_add_tail(&item->list, &dlm->work_list); | |
2116 | spin_unlock(&dlm->work_lock); | |
2117 | ||
3156d267 | 2118 | queue_work(dlm->dlm_worker, &dlm->dispatched_work); |
6714d8e8 KH |
2119 | return 0; |
2120 | } | |
2121 | ||
2122 | static void dlm_assert_master_worker(struct dlm_work_item *item, void *data) | |
2123 | { | |
2124 | struct dlm_ctxt *dlm = data; | |
2125 | int ret = 0; | |
2126 | struct dlm_lock_resource *res; | |
2127 | unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)]; | |
2128 | int ignore_higher; | |
2129 | int bit; | |
2130 | u8 request_from; | |
2131 | u32 flags; | |
2132 | ||
2133 | dlm = item->dlm; | |
2134 | res = item->u.am.lockres; | |
2135 | ignore_higher = item->u.am.ignore_higher; | |
2136 | request_from = item->u.am.request_from; | |
2137 | flags = item->u.am.flags; | |
2138 | ||
2139 | spin_lock(&dlm->spinlock); | |
2140 | memcpy(nodemap, dlm->domain_map, sizeof(nodemap)); | |
2141 | spin_unlock(&dlm->spinlock); | |
2142 | ||
2143 | clear_bit(dlm->node_num, nodemap); | |
2144 | if (ignore_higher) { | |
2145 | /* if is this just to clear up mles for nodes below | |
2146 | * this node, do not send the message to the original | |
2147 | * caller or any node number higher than this */ | |
2148 | clear_bit(request_from, nodemap); | |
2149 | bit = dlm->node_num; | |
2150 | while (1) { | |
2151 | bit = find_next_bit(nodemap, O2NM_MAX_NODES, | |
2152 | bit+1); | |
2153 | if (bit >= O2NM_MAX_NODES) | |
2154 | break; | |
2155 | clear_bit(bit, nodemap); | |
2156 | } | |
2157 | } | |
2158 | ||
36407488 KH |
2159 | /* |
2160 | * If we're migrating this lock to someone else, we are no | |
2161 | * longer allowed to assert out own mastery. OTOH, we need to | |
2162 | * prevent migration from starting while we're still asserting | |
2163 | * our dominance. The reserved ast delays migration. | |
2164 | */ | |
2165 | spin_lock(&res->spinlock); | |
2166 | if (res->state & DLM_LOCK_RES_MIGRATING) { | |
2167 | mlog(0, "Someone asked us to assert mastery, but we're " | |
2168 | "in the middle of migration. Skipping assert, " | |
2169 | "the new master will handle that.\n"); | |
2170 | spin_unlock(&res->spinlock); | |
2171 | goto put; | |
2172 | } else | |
2173 | __dlm_lockres_reserve_ast(res); | |
2174 | spin_unlock(&res->spinlock); | |
2175 | ||
6714d8e8 KH |
2176 | /* this call now finishes out the nodemap |
2177 | * even if one or more nodes die */ | |
2178 | mlog(0, "worker about to master %.*s here, this=%u\n", | |
2179 | res->lockname.len, res->lockname.name, dlm->node_num); | |
ba2bf218 | 2180 | ret = dlm_do_assert_master(dlm, res, nodemap, flags); |
6714d8e8 KH |
2181 | if (ret < 0) { |
2182 | /* no need to restart, we are done */ | |
3b3b84a8 KH |
2183 | if (!dlm_is_host_down(ret)) |
2184 | mlog_errno(ret); | |
6714d8e8 KH |
2185 | } |
2186 | ||
36407488 KH |
2187 | /* Ok, we've asserted ourselves. Let's let migration start. */ |
2188 | dlm_lockres_release_ast(dlm, res); | |
2189 | ||
2190 | put: | |
6714d8e8 KH |
2191 | dlm_lockres_put(res); |
2192 | ||
2193 | mlog(0, "finished with dlm_assert_master_worker\n"); | |
2194 | } | |
2195 | ||
c03872f5 KH |
2196 | /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread. |
2197 | * We cannot wait for node recovery to complete to begin mastering this | |
2198 | * lockres because this lockres is used to kick off recovery! ;-) | |
2199 | * So, do a pre-check on all living nodes to see if any of those nodes | |
2200 | * think that $RECOVERY is currently mastered by a dead node. If so, | |
2201 | * we wait a short time to allow that node to get notified by its own | |
2202 | * heartbeat stack, then check again. All $RECOVERY lock resources | |
2203 | * mastered by dead nodes are purged when the hearbeat callback is | |
2204 | * fired, so we can know for sure that it is safe to continue once | |
2205 | * the node returns a live node or no node. */ | |
2206 | static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm, | |
2207 | struct dlm_lock_resource *res) | |
2208 | { | |
2209 | struct dlm_node_iter iter; | |
2210 | int nodenum; | |
2211 | int ret = 0; | |
2212 | u8 master = DLM_LOCK_RES_OWNER_UNKNOWN; | |
2213 | ||
2214 | spin_lock(&dlm->spinlock); | |
2215 | dlm_node_iter_init(dlm->domain_map, &iter); | |
2216 | spin_unlock(&dlm->spinlock); | |
2217 | ||
2218 | while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { | |
2219 | /* do not send to self */ | |
2220 | if (nodenum == dlm->node_num) | |
2221 | continue; | |
2222 | ret = dlm_do_master_requery(dlm, res, nodenum, &master); | |
2223 | if (ret < 0) { | |
2224 | mlog_errno(ret); | |
2225 | if (!dlm_is_host_down(ret)) | |
2226 | BUG(); | |
2227 | /* host is down, so answer for that node would be | |
2228 | * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */ | |
f42a100b | 2229 | ret = 0; |
c03872f5 KH |
2230 | } |
2231 | ||
2232 | if (master != DLM_LOCK_RES_OWNER_UNKNOWN) { | |
2233 | /* check to see if this master is in the recovery map */ | |
2234 | spin_lock(&dlm->spinlock); | |
2235 | if (test_bit(master, dlm->recovery_map)) { | |
2236 | mlog(ML_NOTICE, "%s: node %u has not seen " | |
2237 | "node %u go down yet, and thinks the " | |
2238 | "dead node is mastering the recovery " | |
2239 | "lock. must wait.\n", dlm->name, | |
2240 | nodenum, master); | |
2241 | ret = -EAGAIN; | |
2242 | } | |
2243 | spin_unlock(&dlm->spinlock); | |
2244 | mlog(0, "%s: reco lock master is %u\n", dlm->name, | |
2245 | master); | |
2246 | break; | |
2247 | } | |
2248 | } | |
2249 | return ret; | |
2250 | } | |
2251 | ||
ba2bf218 KH |
2252 | /* |
2253 | * DLM_DEREF_LOCKRES_MSG | |
2254 | */ | |
2255 | ||
2256 | int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) | |
2257 | { | |
2258 | struct dlm_deref_lockres deref; | |
2259 | int ret = 0, r; | |
2260 | const char *lockname; | |
2261 | unsigned int namelen; | |
2262 | ||
2263 | lockname = res->lockname.name; | |
2264 | namelen = res->lockname.len; | |
2265 | BUG_ON(namelen > O2NM_MAX_NAME_LEN); | |
2266 | ||
2267 | mlog(0, "%s:%.*s: sending deref to %d\n", | |
2268 | dlm->name, namelen, lockname, res->owner); | |
2269 | memset(&deref, 0, sizeof(deref)); | |
2270 | deref.node_idx = dlm->node_num; | |
2271 | deref.namelen = namelen; | |
2272 | memcpy(deref.name, lockname, namelen); | |
2273 | ||
2274 | ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key, | |
2275 | &deref, sizeof(deref), res->owner, &r); | |
2276 | if (ret < 0) | |
2277 | mlog_errno(ret); | |
2278 | else if (r < 0) { | |
2279 | /* BAD. other node says I did not have a ref. */ | |
2280 | mlog(ML_ERROR,"while dropping ref on %s:%.*s " | |
2281 | "(master=%u) got %d.\n", dlm->name, namelen, | |
2282 | lockname, res->owner, r); | |
2283 | dlm_print_one_lock_resource(res); | |
2284 | BUG(); | |
2285 | } | |
2286 | return ret; | |
2287 | } | |
2288 | ||
d74c9803 KH |
2289 | int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data, |
2290 | void **ret_data) | |
ba2bf218 KH |
2291 | { |
2292 | struct dlm_ctxt *dlm = data; | |
2293 | struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf; | |
2294 | struct dlm_lock_resource *res = NULL; | |
2295 | char *name; | |
2296 | unsigned int namelen; | |
2297 | int ret = -EINVAL; | |
2298 | u8 node; | |
2299 | unsigned int hash; | |
f3f85464 SM |
2300 | struct dlm_work_item *item; |
2301 | int cleared = 0; | |
2302 | int dispatch = 0; | |
ba2bf218 KH |
2303 | |
2304 | if (!dlm_grab(dlm)) | |
2305 | return 0; | |
2306 | ||
2307 | name = deref->name; | |
2308 | namelen = deref->namelen; | |
2309 | node = deref->node_idx; | |
2310 | ||
2311 | if (namelen > DLM_LOCKID_NAME_MAX) { | |
2312 | mlog(ML_ERROR, "Invalid name length!"); | |
2313 | goto done; | |
2314 | } | |
2315 | if (deref->node_idx >= O2NM_MAX_NODES) { | |
2316 | mlog(ML_ERROR, "Invalid node number: %u\n", node); | |
2317 | goto done; | |
2318 | } | |
2319 | ||
2320 | hash = dlm_lockid_hash(name, namelen); | |
2321 | ||
2322 | spin_lock(&dlm->spinlock); | |
2323 | res = __dlm_lookup_lockres_full(dlm, name, namelen, hash); | |
2324 | if (!res) { | |
2325 | spin_unlock(&dlm->spinlock); | |
2326 | mlog(ML_ERROR, "%s:%.*s: bad lockres name\n", | |
2327 | dlm->name, namelen, name); | |
2328 | goto done; | |
2329 | } | |
2330 | spin_unlock(&dlm->spinlock); | |
2331 | ||
2332 | spin_lock(&res->spinlock); | |
f3f85464 SM |
2333 | if (res->state & DLM_LOCK_RES_SETREF_INPROG) |
2334 | dispatch = 1; | |
2335 | else { | |
2336 | BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF); | |
2337 | if (test_bit(node, res->refmap)) { | |
2338 | dlm_lockres_clear_refmap_bit(node, res); | |
2339 | cleared = 1; | |
2340 | } | |
ba2bf218 KH |
2341 | } |
2342 | spin_unlock(&res->spinlock); | |
2343 | ||
f3f85464 SM |
2344 | if (!dispatch) { |
2345 | if (cleared) | |
2346 | dlm_lockres_calc_usage(dlm, res); | |
2347 | else { | |
2348 | mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref " | |
2349 | "but it is already dropped!\n", dlm->name, | |
2350 | res->lockname.len, res->lockname.name, node); | |
2351 | __dlm_print_one_lock_resource(res); | |
2352 | } | |
2353 | ret = 0; | |
2354 | goto done; | |
2355 | } | |
2356 | ||
2357 | item = kzalloc(sizeof(*item), GFP_NOFS); | |
2358 | if (!item) { | |
2359 | ret = -ENOMEM; | |
2360 | mlog_errno(ret); | |
2361 | goto done; | |
2362 | } | |
2363 | ||
2364 | dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL); | |
2365 | item->u.dl.deref_res = res; | |
2366 | item->u.dl.deref_node = node; | |
2367 | ||
2368 | spin_lock(&dlm->work_lock); | |
2369 | list_add_tail(&item->list, &dlm->work_list); | |
2370 | spin_unlock(&dlm->work_lock); | |
2371 | ||
2372 | queue_work(dlm->dlm_worker, &dlm->dispatched_work); | |
2373 | return 0; | |
2374 | ||
ba2bf218 KH |
2375 | done: |
2376 | if (res) | |
2377 | dlm_lockres_put(res); | |
2378 | dlm_put(dlm); | |
f3f85464 | 2379 | |
ba2bf218 KH |
2380 | return ret; |
2381 | } | |
2382 | ||
f3f85464 SM |
2383 | static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data) |
2384 | { | |
2385 | struct dlm_ctxt *dlm; | |
2386 | struct dlm_lock_resource *res; | |
2387 | u8 node; | |
2388 | u8 cleared = 0; | |
2389 | ||
2390 | dlm = item->dlm; | |
2391 | res = item->u.dl.deref_res; | |
2392 | node = item->u.dl.deref_node; | |
2393 | ||
2394 | spin_lock(&res->spinlock); | |
2395 | BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF); | |
2396 | if (test_bit(node, res->refmap)) { | |
2397 | __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG); | |
2398 | dlm_lockres_clear_refmap_bit(node, res); | |
2399 | cleared = 1; | |
2400 | } | |
2401 | spin_unlock(&res->spinlock); | |
2402 | ||
2403 | if (cleared) { | |
2404 | mlog(0, "%s:%.*s node %u ref dropped in dispatch\n", | |
2405 | dlm->name, res->lockname.len, res->lockname.name, node); | |
2406 | dlm_lockres_calc_usage(dlm, res); | |
2407 | } else { | |
2408 | mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref " | |
2409 | "but it is already dropped!\n", dlm->name, | |
2410 | res->lockname.len, res->lockname.name, node); | |
2411 | __dlm_print_one_lock_resource(res); | |
2412 | } | |
2413 | ||
2414 | dlm_lockres_put(res); | |
2415 | } | |
2416 | ||
2f5bf1f2 SM |
2417 | /* Checks whether the lockres can be migrated. Returns 0 if yes, < 0 |
2418 | * if not. If 0, numlocks is set to the number of locks in the lockres. | |
2419 | */ | |
2420 | static int dlm_is_lockres_migrateable(struct dlm_ctxt *dlm, | |
2421 | struct dlm_lock_resource *res, | |
2422 | int *numlocks) | |
2423 | { | |
2424 | int ret; | |
2425 | int i; | |
2426 | int count = 0; | |
800deef3 | 2427 | struct list_head *queue; |
2f5bf1f2 SM |
2428 | struct dlm_lock *lock; |
2429 | ||
2430 | assert_spin_locked(&res->spinlock); | |
2431 | ||
2432 | ret = -EINVAL; | |
2433 | if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) { | |
2434 | mlog(0, "cannot migrate lockres with unknown owner!\n"); | |
2435 | goto leave; | |
2436 | } | |
2437 | ||
2438 | if (res->owner != dlm->node_num) { | |
2439 | mlog(0, "cannot migrate lockres this node doesn't own!\n"); | |
2440 | goto leave; | |
2441 | } | |
2442 | ||
2443 | ret = 0; | |
2444 | queue = &res->granted; | |
2445 | for (i = 0; i < 3; i++) { | |
800deef3 | 2446 | list_for_each_entry(lock, queue, list) { |
2f5bf1f2 SM |
2447 | ++count; |
2448 | if (lock->ml.node == dlm->node_num) { | |
2449 | mlog(0, "found a lock owned by this node still " | |
2450 | "on the %s queue! will not migrate this " | |
2451 | "lockres\n", (i == 0 ? "granted" : | |
2452 | (i == 1 ? "converting" : | |
2453 | "blocked"))); | |
2454 | ret = -ENOTEMPTY; | |
2455 | goto leave; | |
2456 | } | |
2457 | } | |
2458 | queue++; | |
2459 | } | |
2460 | ||
2461 | *numlocks = count; | |
2462 | mlog(0, "migrateable lockres having %d locks\n", *numlocks); | |
2463 | ||
2464 | leave: | |
2465 | return ret; | |
2466 | } | |
6714d8e8 KH |
2467 | |
2468 | /* | |
2469 | * DLM_MIGRATE_LOCKRES | |
2470 | */ | |
2471 | ||
2472 | ||
faf0ec9f AB |
2473 | static int dlm_migrate_lockres(struct dlm_ctxt *dlm, |
2474 | struct dlm_lock_resource *res, | |
2475 | u8 target) | |
6714d8e8 KH |
2476 | { |
2477 | struct dlm_master_list_entry *mle = NULL; | |
2478 | struct dlm_master_list_entry *oldmle = NULL; | |
2479 | struct dlm_migratable_lockres *mres = NULL; | |
2f5bf1f2 | 2480 | int ret = 0; |
6714d8e8 KH |
2481 | const char *name; |
2482 | unsigned int namelen; | |
2483 | int mle_added = 0; | |
2f5bf1f2 SM |
2484 | int numlocks; |
2485 | int wake = 0; | |
6714d8e8 KH |
2486 | |
2487 | if (!dlm_grab(dlm)) | |
2488 | return -EINVAL; | |
2489 | ||
2490 | name = res->lockname.name; | |
2491 | namelen = res->lockname.len; | |
2492 | ||
2493 | mlog(0, "migrating %.*s to %u\n", namelen, name, target); | |
2494 | ||
2495 | /* | |
2496 | * ensure this lockres is a proper candidate for migration | |
2497 | */ | |
2498 | spin_lock(&res->spinlock); | |
2f5bf1f2 SM |
2499 | ret = dlm_is_lockres_migrateable(dlm, res, &numlocks); |
2500 | if (ret < 0) { | |
6714d8e8 KH |
2501 | spin_unlock(&res->spinlock); |
2502 | goto leave; | |
2503 | } | |
6714d8e8 KH |
2504 | spin_unlock(&res->spinlock); |
2505 | ||
2506 | /* no work to do */ | |
2f5bf1f2 | 2507 | if (numlocks == 0) { |
6714d8e8 | 2508 | mlog(0, "no locks were found on this lockres! done!\n"); |
6714d8e8 KH |
2509 | goto leave; |
2510 | } | |
2511 | ||
2512 | /* | |
2513 | * preallocate up front | |
2514 | * if this fails, abort | |
2515 | */ | |
2516 | ||
2517 | ret = -ENOMEM; | |
ad8100e0 | 2518 | mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS); |
6714d8e8 KH |
2519 | if (!mres) { |
2520 | mlog_errno(ret); | |
2521 | goto leave; | |
2522 | } | |
2523 | ||
2524 | mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache, | |
ad8100e0 | 2525 | GFP_NOFS); |
6714d8e8 KH |
2526 | if (!mle) { |
2527 | mlog_errno(ret); | |
2528 | goto leave; | |
2529 | } | |
2530 | ret = 0; | |
2531 | ||
2532 | /* | |
2533 | * find a node to migrate the lockres to | |
2534 | */ | |
2535 | ||
2536 | mlog(0, "picking a migration node\n"); | |
2537 | spin_lock(&dlm->spinlock); | |
2538 | /* pick a new node */ | |
2539 | if (!test_bit(target, dlm->domain_map) || | |
2540 | target >= O2NM_MAX_NODES) { | |
2541 | target = dlm_pick_migration_target(dlm, res); | |
2542 | } | |
2543 | mlog(0, "node %u chosen for migration\n", target); | |
2544 | ||
2545 | if (target >= O2NM_MAX_NODES || | |
2546 | !test_bit(target, dlm->domain_map)) { | |
2547 | /* target chosen is not alive */ | |
2548 | ret = -EINVAL; | |
2549 | } | |
2550 | ||
2551 | if (ret) { | |
2552 | spin_unlock(&dlm->spinlock); | |
2553 | goto fail; | |
2554 | } | |
2555 | ||
2556 | mlog(0, "continuing with target = %u\n", target); | |
2557 | ||
2558 | /* | |
2559 | * clear any existing master requests and | |
2560 | * add the migration mle to the list | |
2561 | */ | |
2562 | spin_lock(&dlm->master_lock); | |
2563 | ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name, | |
2564 | namelen, target, dlm->node_num); | |
2565 | spin_unlock(&dlm->master_lock); | |
2566 | spin_unlock(&dlm->spinlock); | |
2567 | ||
2568 | if (ret == -EEXIST) { | |
2569 | mlog(0, "another process is already migrating it\n"); | |
2570 | goto fail; | |
2571 | } | |
2572 | mle_added = 1; | |
2573 | ||
2574 | /* | |
2575 | * set the MIGRATING flag and flush asts | |
2576 | * if we fail after this we need to re-dirty the lockres | |
2577 | */ | |
2578 | if (dlm_mark_lockres_migrating(dlm, res, target) < 0) { | |
2579 | mlog(ML_ERROR, "tried to migrate %.*s to %u, but " | |
2580 | "the target went down.\n", res->lockname.len, | |
2581 | res->lockname.name, target); | |
2582 | spin_lock(&res->spinlock); | |
2583 | res->state &= ~DLM_LOCK_RES_MIGRATING; | |
a6fa3640 | 2584 | wake = 1; |
6714d8e8 KH |
2585 | spin_unlock(&res->spinlock); |
2586 | ret = -EINVAL; | |
2587 | } | |
2588 | ||
2589 | fail: | |
2590 | if (oldmle) { | |
2591 | /* master is known, detach if not already detached */ | |
2592 | dlm_mle_detach_hb_events(dlm, oldmle); | |
2593 | dlm_put_mle(oldmle); | |
2594 | } | |
2595 | ||
2596 | if (ret < 0) { | |
2597 | if (mle_added) { | |
2598 | dlm_mle_detach_hb_events(dlm, mle); | |
2599 | dlm_put_mle(mle); | |
2600 | } else if (mle) { | |
2601 | kmem_cache_free(dlm_mle_cache, mle); | |
2602 | } | |
2603 | goto leave; | |
2604 | } | |
2605 | ||
2606 | /* | |
2607 | * at this point, we have a migration target, an mle | |
2608 | * in the master list, and the MIGRATING flag set on | |
2609 | * the lockres | |
2610 | */ | |
2611 | ||
1cd04dbe KH |
2612 | /* now that remote nodes are spinning on the MIGRATING flag, |
2613 | * ensure that all assert_master work is flushed. */ | |
2614 | flush_workqueue(dlm->dlm_worker); | |
6714d8e8 KH |
2615 | |
2616 | /* get an extra reference on the mle. | |
2617 | * otherwise the assert_master from the new | |
2618 | * master will destroy this. | |
2619 | * also, make sure that all callers of dlm_get_mle | |
2620 | * take both dlm->spinlock and dlm->master_lock */ | |
2621 | spin_lock(&dlm->spinlock); | |
2622 | spin_lock(&dlm->master_lock); | |
a2bf0477 | 2623 | dlm_get_mle_inuse(mle); |
6714d8e8 KH |
2624 | spin_unlock(&dlm->master_lock); |
2625 | spin_unlock(&dlm->spinlock); | |
2626 | ||
2627 | /* notify new node and send all lock state */ | |
2628 | /* call send_one_lockres with migration flag. | |
2629 | * this serves as notice to the target node that a | |
2630 | * migration is starting. */ | |
2631 | ret = dlm_send_one_lockres(dlm, res, mres, target, | |
2632 | DLM_MRES_MIGRATION); | |
2633 | ||
2634 | if (ret < 0) { | |
2635 | mlog(0, "migration to node %u failed with %d\n", | |
2636 | target, ret); | |
2637 | /* migration failed, detach and clean up mle */ | |
2638 | dlm_mle_detach_hb_events(dlm, mle); | |
2639 | dlm_put_mle(mle); | |
a2bf0477 KH |
2640 | dlm_put_mle_inuse(mle); |
2641 | spin_lock(&res->spinlock); | |
2642 | res->state &= ~DLM_LOCK_RES_MIGRATING; | |
a6fa3640 | 2643 | wake = 1; |
a2bf0477 | 2644 | spin_unlock(&res->spinlock); |
6714d8e8 KH |
2645 | goto leave; |
2646 | } | |
2647 | ||
2648 | /* at this point, the target sends a message to all nodes, | |
2649 | * (using dlm_do_migrate_request). this node is skipped since | |
2650 | * we had to put an mle in the list to begin the process. this | |
2651 | * node now waits for target to do an assert master. this node | |
2652 | * will be the last one notified, ensuring that the migration | |
2653 | * is complete everywhere. if the target dies while this is | |
2654 | * going on, some nodes could potentially see the target as the | |
2655 | * master, so it is important that my recovery finds the migration | |
2656 | * mle and sets the master to UNKNONWN. */ | |
2657 | ||
2658 | ||
2659 | /* wait for new node to assert master */ | |
2660 | while (1) { | |
2661 | ret = wait_event_interruptible_timeout(mle->wq, | |
2662 | (atomic_read(&mle->woken) == 1), | |
2663 | msecs_to_jiffies(5000)); | |
2664 | ||
2665 | if (ret >= 0) { | |
2666 | if (atomic_read(&mle->woken) == 1 || | |
2667 | res->owner == target) | |
2668 | break; | |
2669 | ||
1cd04dbe KH |
2670 | mlog(0, "%s:%.*s: timed out during migration\n", |
2671 | dlm->name, res->lockname.len, res->lockname.name); | |
e2faea4c KH |
2672 | /* avoid hang during shutdown when migrating lockres |
2673 | * to a node which also goes down */ | |
2674 | if (dlm_is_node_dead(dlm, target)) { | |
aa852354 KH |
2675 | mlog(0, "%s:%.*s: expected migration " |
2676 | "target %u is no longer up, restarting\n", | |
e2faea4c KH |
2677 | dlm->name, res->lockname.len, |
2678 | res->lockname.name, target); | |
1cd04dbe KH |
2679 | ret = -EINVAL; |
2680 | /* migration failed, detach and clean up mle */ | |
2681 | dlm_mle_detach_hb_events(dlm, mle); | |
2682 | dlm_put_mle(mle); | |
2683 | dlm_put_mle_inuse(mle); | |
2684 | spin_lock(&res->spinlock); | |
2685 | res->state &= ~DLM_LOCK_RES_MIGRATING; | |
a6fa3640 | 2686 | wake = 1; |
1cd04dbe KH |
2687 | spin_unlock(&res->spinlock); |
2688 | goto leave; | |
e2faea4c | 2689 | } |
1cd04dbe KH |
2690 | } else |
2691 | mlog(0, "%s:%.*s: caught signal during migration\n", | |
2692 | dlm->name, res->lockname.len, res->lockname.name); | |
6714d8e8 KH |
2693 | } |
2694 | ||
2695 | /* all done, set the owner, clear the flag */ | |
2696 | spin_lock(&res->spinlock); | |
2697 | dlm_set_lockres_owner(dlm, res, target); | |
2698 | res->state &= ~DLM_LOCK_RES_MIGRATING; | |
2699 | dlm_remove_nonlocal_locks(dlm, res); | |
2700 | spin_unlock(&res->spinlock); | |
2701 | wake_up(&res->wq); | |
2702 | ||
2703 | /* master is known, detach if not already detached */ | |
2704 | dlm_mle_detach_hb_events(dlm, mle); | |
a2bf0477 | 2705 | dlm_put_mle_inuse(mle); |
6714d8e8 KH |
2706 | ret = 0; |
2707 | ||
2708 | dlm_lockres_calc_usage(dlm, res); | |
2709 | ||
2710 | leave: | |
2711 | /* re-dirty the lockres if we failed */ | |
2712 | if (ret < 0) | |
2713 | dlm_kick_thread(dlm, res); | |
2714 | ||
a6fa3640 KH |
2715 | /* wake up waiters if the MIGRATING flag got set |
2716 | * but migration failed */ | |
2717 | if (wake) | |
2718 | wake_up(&res->wq); | |
2719 | ||
6714d8e8 KH |
2720 | /* TODO: cleanup */ |
2721 | if (mres) | |
2722 | free_page((unsigned long)mres); | |
2723 | ||
2724 | dlm_put(dlm); | |
2725 | ||
2726 | mlog(0, "returning %d\n", ret); | |
2727 | return ret; | |
2728 | } | |
6714d8e8 | 2729 | |
ba2bf218 KH |
2730 | #define DLM_MIGRATION_RETRY_MS 100 |
2731 | ||
2732 | /* Should be called only after beginning the domain leave process. | |
2733 | * There should not be any remaining locks on nonlocal lock resources, | |
2734 | * and there should be no local locks left on locally mastered resources. | |
2735 | * | |
2736 | * Called with the dlm spinlock held, may drop it to do migration, but | |
2737 | * will re-acquire before exit. | |
2738 | * | |
2739 | * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped */ | |
2740 | int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) | |
2741 | { | |
2742 | int ret; | |
2743 | int lock_dropped = 0; | |
2f5bf1f2 | 2744 | int numlocks; |
ba2bf218 | 2745 | |
b36c3f84 | 2746 | spin_lock(&res->spinlock); |
ba2bf218 KH |
2747 | if (res->owner != dlm->node_num) { |
2748 | if (!__dlm_lockres_unused(res)) { | |
2749 | mlog(ML_ERROR, "%s:%.*s: this node is not master, " | |
2750 | "trying to free this but locks remain\n", | |
2751 | dlm->name, res->lockname.len, res->lockname.name); | |
2752 | } | |
b36c3f84 | 2753 | spin_unlock(&res->spinlock); |
ba2bf218 KH |
2754 | goto leave; |
2755 | } | |
2f5bf1f2 SM |
2756 | |
2757 | /* No need to migrate a lockres having no locks */ | |
2758 | ret = dlm_is_lockres_migrateable(dlm, res, &numlocks); | |
2759 | if (ret >= 0 && numlocks == 0) { | |
2760 | spin_unlock(&res->spinlock); | |
2761 | goto leave; | |
2762 | } | |
b36c3f84 | 2763 | spin_unlock(&res->spinlock); |
ba2bf218 KH |
2764 | |
2765 | /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */ | |
2766 | spin_unlock(&dlm->spinlock); | |
2767 | lock_dropped = 1; | |
2768 | while (1) { | |
2769 | ret = dlm_migrate_lockres(dlm, res, O2NM_MAX_NODES); | |
2770 | if (ret >= 0) | |
2771 | break; | |
2772 | if (ret == -ENOTEMPTY) { | |
2773 | mlog(ML_ERROR, "lockres %.*s still has local locks!\n", | |
2774 | res->lockname.len, res->lockname.name); | |
2775 | BUG(); | |
2776 | } | |
2777 | ||
2778 | mlog(0, "lockres %.*s: migrate failed, " | |
2779 | "retrying\n", res->lockname.len, | |
2780 | res->lockname.name); | |
2781 | msleep(DLM_MIGRATION_RETRY_MS); | |
2782 | } | |
2783 | spin_lock(&dlm->spinlock); | |
2784 | leave: | |
2785 | return lock_dropped; | |
2786 | } | |
2787 | ||
6714d8e8 KH |
2788 | int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock) |
2789 | { | |
2790 | int ret; | |
2791 | spin_lock(&dlm->ast_lock); | |
2792 | spin_lock(&lock->spinlock); | |
2793 | ret = (list_empty(&lock->bast_list) && !lock->bast_pending); | |
2794 | spin_unlock(&lock->spinlock); | |
2795 | spin_unlock(&dlm->ast_lock); | |
2796 | return ret; | |
2797 | } | |
2798 | ||
2799 | static int dlm_migration_can_proceed(struct dlm_ctxt *dlm, | |
2800 | struct dlm_lock_resource *res, | |
2801 | u8 mig_target) | |
2802 | { | |
2803 | int can_proceed; | |
2804 | spin_lock(&res->spinlock); | |
2805 | can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING); | |
2806 | spin_unlock(&res->spinlock); | |
2807 | ||
2808 | /* target has died, so make the caller break out of the | |
2809 | * wait_event, but caller must recheck the domain_map */ | |
2810 | spin_lock(&dlm->spinlock); | |
2811 | if (!test_bit(mig_target, dlm->domain_map)) | |
2812 | can_proceed = 1; | |
2813 | spin_unlock(&dlm->spinlock); | |
2814 | return can_proceed; | |
2815 | } | |
2816 | ||
faf0ec9f AB |
2817 | static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm, |
2818 | struct dlm_lock_resource *res) | |
6714d8e8 KH |
2819 | { |
2820 | int ret; | |
2821 | spin_lock(&res->spinlock); | |
2822 | ret = !!(res->state & DLM_LOCK_RES_DIRTY); | |
2823 | spin_unlock(&res->spinlock); | |
2824 | return ret; | |
2825 | } | |
2826 | ||
2827 | ||
2828 | static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm, | |
2829 | struct dlm_lock_resource *res, | |
2830 | u8 target) | |
2831 | { | |
2832 | int ret = 0; | |
2833 | ||
2834 | mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n", | |
2835 | res->lockname.len, res->lockname.name, dlm->node_num, | |
2836 | target); | |
2837 | /* need to set MIGRATING flag on lockres. this is done by | |
2838 | * ensuring that all asts have been flushed for this lockres. */ | |
2839 | spin_lock(&res->spinlock); | |
2840 | BUG_ON(res->migration_pending); | |
2841 | res->migration_pending = 1; | |
2842 | /* strategy is to reserve an extra ast then release | |
2843 | * it below, letting the release do all of the work */ | |
2844 | __dlm_lockres_reserve_ast(res); | |
2845 | spin_unlock(&res->spinlock); | |
2846 | ||
ddc09c8d | 2847 | /* now flush all the pending asts */ |
6714d8e8 | 2848 | dlm_kick_thread(dlm, res); |
ddc09c8d KH |
2849 | /* before waiting on DIRTY, block processes which may |
2850 | * try to dirty the lockres before MIGRATING is set */ | |
2851 | spin_lock(&res->spinlock); | |
2852 | BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY); | |
2853 | res->state |= DLM_LOCK_RES_BLOCK_DIRTY; | |
2854 | spin_unlock(&res->spinlock); | |
2855 | /* now wait on any pending asts and the DIRTY state */ | |
6714d8e8 KH |
2856 | wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res)); |
2857 | dlm_lockres_release_ast(dlm, res); | |
2858 | ||
2859 | mlog(0, "about to wait on migration_wq, dirty=%s\n", | |
2860 | res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no"); | |
2861 | /* if the extra ref we just put was the final one, this | |
2862 | * will pass thru immediately. otherwise, we need to wait | |
2863 | * for the last ast to finish. */ | |
2864 | again: | |
2865 | ret = wait_event_interruptible_timeout(dlm->migration_wq, | |
2866 | dlm_migration_can_proceed(dlm, res, target), | |
2867 | msecs_to_jiffies(1000)); | |
2868 | if (ret < 0) { | |
2869 | mlog(0, "woken again: migrating? %s, dead? %s\n", | |
2870 | res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no", | |
2871 | test_bit(target, dlm->domain_map) ? "no":"yes"); | |
2872 | } else { | |
2873 | mlog(0, "all is well: migrating? %s, dead? %s\n", | |
2874 | res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no", | |
2875 | test_bit(target, dlm->domain_map) ? "no":"yes"); | |
2876 | } | |
2877 | if (!dlm_migration_can_proceed(dlm, res, target)) { | |
2878 | mlog(0, "trying again...\n"); | |
2879 | goto again; | |
2880 | } | |
ddc09c8d KH |
2881 | /* now that we are sure the MIGRATING state is there, drop |
2882 | * the unneded state which blocked threads trying to DIRTY */ | |
2883 | spin_lock(&res->spinlock); | |
2884 | BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY)); | |
2885 | BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING)); | |
2886 | res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY; | |
2887 | spin_unlock(&res->spinlock); | |
6714d8e8 KH |
2888 | |
2889 | /* did the target go down or die? */ | |
2890 | spin_lock(&dlm->spinlock); | |
2891 | if (!test_bit(target, dlm->domain_map)) { | |
2892 | mlog(ML_ERROR, "aha. migration target %u just went down\n", | |
2893 | target); | |
2894 | ret = -EHOSTDOWN; | |
2895 | } | |
2896 | spin_unlock(&dlm->spinlock); | |
2897 | ||
2898 | /* | |
2899 | * at this point: | |
2900 | * | |
2901 | * o the DLM_LOCK_RES_MIGRATING flag is set | |
2902 | * o there are no pending asts on this lockres | |
2903 | * o all processes trying to reserve an ast on this | |
2904 | * lockres must wait for the MIGRATING flag to clear | |
2905 | */ | |
2906 | return ret; | |
2907 | } | |
2908 | ||
2909 | /* last step in the migration process. | |
2910 | * original master calls this to free all of the dlm_lock | |
2911 | * structures that used to be for other nodes. */ | |
2912 | static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm, | |
2913 | struct dlm_lock_resource *res) | |
2914 | { | |
6714d8e8 | 2915 | struct list_head *queue = &res->granted; |
ba2bf218 | 2916 | int i, bit; |
800deef3 | 2917 | struct dlm_lock *lock, *next; |
6714d8e8 KH |
2918 | |
2919 | assert_spin_locked(&res->spinlock); | |
2920 | ||
2921 | BUG_ON(res->owner == dlm->node_num); | |
2922 | ||
2923 | for (i=0; i<3; i++) { | |
800deef3 | 2924 | list_for_each_entry_safe(lock, next, queue, list) { |
6714d8e8 KH |
2925 | if (lock->ml.node != dlm->node_num) { |
2926 | mlog(0, "putting lock for node %u\n", | |
2927 | lock->ml.node); | |
2928 | /* be extra careful */ | |
2929 | BUG_ON(!list_empty(&lock->ast_list)); | |
2930 | BUG_ON(!list_empty(&lock->bast_list)); | |
2931 | BUG_ON(lock->ast_pending); | |
2932 | BUG_ON(lock->bast_pending); | |
ba2bf218 | 2933 | dlm_lockres_clear_refmap_bit(lock->ml.node, res); |
6714d8e8 KH |
2934 | list_del_init(&lock->list); |
2935 | dlm_lock_put(lock); | |
2936 | } | |
2937 | } | |
2938 | queue++; | |
2939 | } | |
ba2bf218 KH |
2940 | bit = 0; |
2941 | while (1) { | |
2942 | bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit); | |
2943 | if (bit >= O2NM_MAX_NODES) | |
2944 | break; | |
2945 | /* do not clear the local node reference, if there is a | |
2946 | * process holding this, let it drop the ref itself */ | |
2947 | if (bit != dlm->node_num) { | |
2948 | mlog(0, "%s:%.*s: node %u had a ref to this " | |
2949 | "migrating lockres, clearing\n", dlm->name, | |
2950 | res->lockname.len, res->lockname.name, bit); | |
2951 | dlm_lockres_clear_refmap_bit(bit, res); | |
2952 | } | |
2953 | bit++; | |
2954 | } | |
6714d8e8 KH |
2955 | } |
2956 | ||
2957 | /* for now this is not too intelligent. we will | |
2958 | * need stats to make this do the right thing. | |
2959 | * this just finds the first lock on one of the | |
2960 | * queues and uses that node as the target. */ | |
2961 | static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm, | |
2962 | struct dlm_lock_resource *res) | |
2963 | { | |
2964 | int i; | |
2965 | struct list_head *queue = &res->granted; | |
6714d8e8 KH |
2966 | struct dlm_lock *lock; |
2967 | int nodenum; | |
2968 | ||
2969 | assert_spin_locked(&dlm->spinlock); | |
2970 | ||
2971 | spin_lock(&res->spinlock); | |
2972 | for (i=0; i<3; i++) { | |
800deef3 | 2973 | list_for_each_entry(lock, queue, list) { |
6714d8e8 KH |
2974 | /* up to the caller to make sure this node |
2975 | * is alive */ | |
6714d8e8 KH |
2976 | if (lock->ml.node != dlm->node_num) { |
2977 | spin_unlock(&res->spinlock); | |
2978 | return lock->ml.node; | |
2979 | } | |
2980 | } | |
2981 | queue++; | |
2982 | } | |
2983 | spin_unlock(&res->spinlock); | |
2984 | mlog(0, "have not found a suitable target yet! checking domain map\n"); | |
2985 | ||
2986 | /* ok now we're getting desperate. pick anyone alive. */ | |
2987 | nodenum = -1; | |
2988 | while (1) { | |
2989 | nodenum = find_next_bit(dlm->domain_map, | |
2990 | O2NM_MAX_NODES, nodenum+1); | |
2991 | mlog(0, "found %d in domain map\n", nodenum); | |
2992 | if (nodenum >= O2NM_MAX_NODES) | |
2993 | break; | |
2994 | if (nodenum != dlm->node_num) { | |
2995 | mlog(0, "picking %d\n", nodenum); | |
2996 | return nodenum; | |
2997 | } | |
2998 | } | |
2999 | ||
3000 | mlog(0, "giving up. no master to migrate to\n"); | |
3001 | return DLM_LOCK_RES_OWNER_UNKNOWN; | |
3002 | } | |
3003 | ||
3004 | ||
3005 | ||
3006 | /* this is called by the new master once all lockres | |
3007 | * data has been received */ | |
3008 | static int dlm_do_migrate_request(struct dlm_ctxt *dlm, | |
3009 | struct dlm_lock_resource *res, | |
3010 | u8 master, u8 new_master, | |
3011 | struct dlm_node_iter *iter) | |
3012 | { | |
3013 | struct dlm_migrate_request migrate; | |
3014 | int ret, status = 0; | |
3015 | int nodenum; | |
3016 | ||
3017 | memset(&migrate, 0, sizeof(migrate)); | |
3018 | migrate.namelen = res->lockname.len; | |
3019 | memcpy(migrate.name, res->lockname.name, migrate.namelen); | |
3020 | migrate.new_master = new_master; | |
3021 | migrate.master = master; | |
3022 | ||
3023 | ret = 0; | |
3024 | ||
3025 | /* send message to all nodes, except the master and myself */ | |
3026 | while ((nodenum = dlm_node_iter_next(iter)) >= 0) { | |
3027 | if (nodenum == master || | |
3028 | nodenum == new_master) | |
3029 | continue; | |
3030 | ||
3031 | ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key, | |
3032 | &migrate, sizeof(migrate), nodenum, | |
3033 | &status); | |
3034 | if (ret < 0) | |
3035 | mlog_errno(ret); | |
3036 | else if (status < 0) { | |
3037 | mlog(0, "migrate request (node %u) returned %d!\n", | |
3038 | nodenum, status); | |
3039 | ret = status; | |
ba2bf218 KH |
3040 | } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) { |
3041 | /* during the migration request we short-circuited | |
3042 | * the mastery of the lockres. make sure we have | |
3043 | * a mastery ref for nodenum */ | |
3044 | mlog(0, "%s:%.*s: need ref for node %u\n", | |
3045 | dlm->name, res->lockname.len, res->lockname.name, | |
3046 | nodenum); | |
3047 | spin_lock(&res->spinlock); | |
3048 | dlm_lockres_set_refmap_bit(nodenum, res); | |
3049 | spin_unlock(&res->spinlock); | |
6714d8e8 KH |
3050 | } |
3051 | } | |
3052 | ||
3053 | if (ret < 0) | |
3054 | mlog_errno(ret); | |
3055 | ||
3056 | mlog(0, "returning ret=%d\n", ret); | |
3057 | return ret; | |
3058 | } | |
3059 | ||
3060 | ||
3061 | /* if there is an existing mle for this lockres, we now know who the master is. | |
3062 | * (the one who sent us *this* message) we can clear it up right away. | |
3063 | * since the process that put the mle on the list still has a reference to it, | |
3064 | * we can unhash it now, set the master and wake the process. as a result, | |
3065 | * we will have no mle in the list to start with. now we can add an mle for | |
3066 | * the migration and this should be the only one found for those scanning the | |
3067 | * list. */ | |
d74c9803 KH |
3068 | int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data, |
3069 | void **ret_data) | |
6714d8e8 KH |
3070 | { |
3071 | struct dlm_ctxt *dlm = data; | |
3072 | struct dlm_lock_resource *res = NULL; | |
3073 | struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf; | |
3074 | struct dlm_master_list_entry *mle = NULL, *oldmle = NULL; | |
3075 | const char *name; | |
a3d33291 | 3076 | unsigned int namelen, hash; |
6714d8e8 KH |
3077 | int ret = 0; |
3078 | ||
3079 | if (!dlm_grab(dlm)) | |
3080 | return -EINVAL; | |
3081 | ||
3082 | name = migrate->name; | |
3083 | namelen = migrate->namelen; | |
a3d33291 | 3084 | hash = dlm_lockid_hash(name, namelen); |
6714d8e8 KH |
3085 | |
3086 | /* preallocate.. if this fails, abort */ | |
3087 | mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache, | |
ad8100e0 | 3088 | GFP_NOFS); |
6714d8e8 KH |
3089 | |
3090 | if (!mle) { | |
3091 | ret = -ENOMEM; | |
3092 | goto leave; | |
3093 | } | |
3094 | ||
3095 | /* check for pre-existing lock */ | |
3096 | spin_lock(&dlm->spinlock); | |
a3d33291 | 3097 | res = __dlm_lookup_lockres(dlm, name, namelen, hash); |
6714d8e8 KH |
3098 | spin_lock(&dlm->master_lock); |
3099 | ||
3100 | if (res) { | |
3101 | spin_lock(&res->spinlock); | |
3102 | if (res->state & DLM_LOCK_RES_RECOVERING) { | |
3103 | /* if all is working ok, this can only mean that we got | |
3104 | * a migrate request from a node that we now see as | |
3105 | * dead. what can we do here? drop it to the floor? */ | |
3106 | spin_unlock(&res->spinlock); | |
3107 | mlog(ML_ERROR, "Got a migrate request, but the " | |
3108 | "lockres is marked as recovering!"); | |
3109 | kmem_cache_free(dlm_mle_cache, mle); | |
3110 | ret = -EINVAL; /* need a better solution */ | |
3111 | goto unlock; | |
3112 | } | |
3113 | res->state |= DLM_LOCK_RES_MIGRATING; | |
3114 | spin_unlock(&res->spinlock); | |
3115 | } | |
3116 | ||
3117 | /* ignore status. only nonzero status would BUG. */ | |
3118 | ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, | |
3119 | name, namelen, | |
3120 | migrate->new_master, | |
3121 | migrate->master); | |
3122 | ||
3123 | unlock: | |
3124 | spin_unlock(&dlm->master_lock); | |
3125 | spin_unlock(&dlm->spinlock); | |
3126 | ||
3127 | if (oldmle) { | |
3128 | /* master is known, detach if not already detached */ | |
3129 | dlm_mle_detach_hb_events(dlm, oldmle); | |
3130 | dlm_put_mle(oldmle); | |
3131 | } | |
3132 | ||
3133 | if (res) | |
3134 | dlm_lockres_put(res); | |
3135 | leave: | |
3136 | dlm_put(dlm); | |
3137 | return ret; | |
3138 | } | |
3139 | ||
3140 | /* must be holding dlm->spinlock and dlm->master_lock | |
3141 | * when adding a migration mle, we can clear any other mles | |
3142 | * in the master list because we know with certainty that | |
3143 | * the master is "master". so we remove any old mle from | |
3144 | * the list after setting it's master field, and then add | |
3145 | * the new migration mle. this way we can hold with the rule | |
3146 | * of having only one mle for a given lock name at all times. */ | |
3147 | static int dlm_add_migration_mle(struct dlm_ctxt *dlm, | |
3148 | struct dlm_lock_resource *res, | |
3149 | struct dlm_master_list_entry *mle, | |
3150 | struct dlm_master_list_entry **oldmle, | |
3151 | const char *name, unsigned int namelen, | |
3152 | u8 new_master, u8 master) | |
3153 | { | |
3154 | int found; | |
3155 | int ret = 0; | |
3156 | ||
3157 | *oldmle = NULL; | |
3158 | ||
3159 | mlog_entry_void(); | |
3160 | ||
3161 | assert_spin_locked(&dlm->spinlock); | |
3162 | assert_spin_locked(&dlm->master_lock); | |
3163 | ||
3164 | /* caller is responsible for any ref taken here on oldmle */ | |
3165 | found = dlm_find_mle(dlm, oldmle, (char *)name, namelen); | |
3166 | if (found) { | |
3167 | struct dlm_master_list_entry *tmp = *oldmle; | |
3168 | spin_lock(&tmp->spinlock); | |
3169 | if (tmp->type == DLM_MLE_MIGRATION) { | |
3170 | if (master == dlm->node_num) { | |
3171 | /* ah another process raced me to it */ | |
3172 | mlog(0, "tried to migrate %.*s, but some " | |
3173 | "process beat me to it\n", | |
3174 | namelen, name); | |
3175 | ret = -EEXIST; | |
3176 | } else { | |
3177 | /* bad. 2 NODES are trying to migrate! */ | |
3178 | mlog(ML_ERROR, "migration error mle: " | |
3179 | "master=%u new_master=%u // request: " | |
3180 | "master=%u new_master=%u // " | |
3181 | "lockres=%.*s\n", | |
3182 | tmp->master, tmp->new_master, | |
3183 | master, new_master, | |
3184 | namelen, name); | |
3185 | BUG(); | |
3186 | } | |
3187 | } else { | |
3188 | /* this is essentially what assert_master does */ | |
3189 | tmp->master = master; | |
3190 | atomic_set(&tmp->woken, 1); | |
3191 | wake_up(&tmp->wq); | |
3192 | /* remove it from the list so that only one | |
3193 | * mle will be found */ | |
3194 | list_del_init(&tmp->list); | |
ba2bf218 KH |
3195 | /* this was obviously WRONG. mle is uninited here. should be tmp. */ |
3196 | __dlm_mle_detach_hb_events(dlm, tmp); | |
3197 | ret = DLM_MIGRATE_RESPONSE_MASTERY_REF; | |
3198 | mlog(0, "%s:%.*s: master=%u, newmaster=%u, " | |
3199 | "telling master to get ref for cleared out mle " | |
3200 | "during migration\n", dlm->name, namelen, name, | |
3201 | master, new_master); | |
6714d8e8 KH |
3202 | } |
3203 | spin_unlock(&tmp->spinlock); | |
3204 | } | |
3205 | ||
3206 | /* now add a migration mle to the tail of the list */ | |
3207 | dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen); | |
3208 | mle->new_master = new_master; | |
ba2bf218 KH |
3209 | /* the new master will be sending an assert master for this. |
3210 | * at that point we will get the refmap reference */ | |
6714d8e8 KH |
3211 | mle->master = master; |
3212 | /* do this for consistency with other mle types */ | |
3213 | set_bit(new_master, mle->maybe_map); | |
3214 | list_add(&mle->list, &dlm->master_list); | |
3215 | ||
3216 | return ret; | |
3217 | } | |
3218 | ||
3219 | ||
3220 | void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node) | |
3221 | { | |
800deef3 | 3222 | struct dlm_master_list_entry *mle, *next; |
6714d8e8 | 3223 | struct dlm_lock_resource *res; |
a3d33291 | 3224 | unsigned int hash; |
6714d8e8 KH |
3225 | |
3226 | mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node); | |
3227 | top: | |
3228 | assert_spin_locked(&dlm->spinlock); | |
3229 | ||
3230 | /* clean the master list */ | |
3231 | spin_lock(&dlm->master_lock); | |
800deef3 | 3232 | list_for_each_entry_safe(mle, next, &dlm->master_list, list) { |
6714d8e8 KH |
3233 | BUG_ON(mle->type != DLM_MLE_BLOCK && |
3234 | mle->type != DLM_MLE_MASTER && | |
3235 | mle->type != DLM_MLE_MIGRATION); | |
3236 | ||
3237 | /* MASTER mles are initiated locally. the waiting | |
3238 | * process will notice the node map change | |
3239 | * shortly. let that happen as normal. */ | |
3240 | if (mle->type == DLM_MLE_MASTER) | |
3241 | continue; | |
3242 | ||
3243 | ||
3244 | /* BLOCK mles are initiated by other nodes. | |
3245 | * need to clean up if the dead node would have | |
3246 | * been the master. */ | |
3247 | if (mle->type == DLM_MLE_BLOCK) { | |
3248 | int bit; | |
3249 | ||
3250 | spin_lock(&mle->spinlock); | |
3251 | bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0); | |
3252 | if (bit != dead_node) { | |
3253 | mlog(0, "mle found, but dead node %u would " | |
3254 | "not have been master\n", dead_node); | |
3255 | spin_unlock(&mle->spinlock); | |
3256 | } else { | |
3257 | /* must drop the refcount by one since the | |
3258 | * assert_master will never arrive. this | |
3259 | * may result in the mle being unlinked and | |
3260 | * freed, but there may still be a process | |
3261 | * waiting in the dlmlock path which is fine. */ | |
3b3b84a8 | 3262 | mlog(0, "node %u was expected master\n", |
6714d8e8 KH |
3263 | dead_node); |
3264 | atomic_set(&mle->woken, 1); | |
3265 | spin_unlock(&mle->spinlock); | |
3266 | wake_up(&mle->wq); | |
f671c09b KH |
3267 | /* do not need events any longer, so detach |
3268 | * from heartbeat */ | |
3269 | __dlm_mle_detach_hb_events(dlm, mle); | |
6714d8e8 KH |
3270 | __dlm_put_mle(mle); |
3271 | } | |
3272 | continue; | |
3273 | } | |
3274 | ||
3275 | /* everything else is a MIGRATION mle */ | |
3276 | ||
3277 | /* the rule for MIGRATION mles is that the master | |
3278 | * becomes UNKNOWN if *either* the original or | |
3279 | * the new master dies. all UNKNOWN lockreses | |
3280 | * are sent to whichever node becomes the recovery | |
3281 | * master. the new master is responsible for | |
3282 | * determining if there is still a master for | |
3283 | * this lockres, or if he needs to take over | |
3284 | * mastery. either way, this node should expect | |
3285 | * another message to resolve this. */ | |
3286 | if (mle->master != dead_node && | |
3287 | mle->new_master != dead_node) | |
3288 | continue; | |
3289 | ||
3290 | /* if we have reached this point, this mle needs to | |
3291 | * be removed from the list and freed. */ | |
3292 | ||
3293 | /* remove from the list early. NOTE: unlinking | |
3294 | * list_head while in list_for_each_safe */ | |
da01ad05 | 3295 | __dlm_mle_detach_hb_events(dlm, mle); |
6714d8e8 KH |
3296 | spin_lock(&mle->spinlock); |
3297 | list_del_init(&mle->list); | |
3298 | atomic_set(&mle->woken, 1); | |
3299 | spin_unlock(&mle->spinlock); | |
3300 | wake_up(&mle->wq); | |
3301 | ||
aa852354 KH |
3302 | mlog(0, "%s: node %u died during migration from " |
3303 | "%u to %u!\n", dlm->name, dead_node, | |
6714d8e8 KH |
3304 | mle->master, mle->new_master); |
3305 | /* if there is a lockres associated with this | |
3306 | * mle, find it and set its owner to UNKNOWN */ | |
a3d33291 | 3307 | hash = dlm_lockid_hash(mle->u.name.name, mle->u.name.len); |
6714d8e8 | 3308 | res = __dlm_lookup_lockres(dlm, mle->u.name.name, |
a3d33291 | 3309 | mle->u.name.len, hash); |
6714d8e8 KH |
3310 | if (res) { |
3311 | /* unfortunately if we hit this rare case, our | |
3312 | * lock ordering is messed. we need to drop | |
3313 | * the master lock so that we can take the | |
3314 | * lockres lock, meaning that we will have to | |
3315 | * restart from the head of list. */ | |
3316 | spin_unlock(&dlm->master_lock); | |
3317 | ||
3318 | /* move lockres onto recovery list */ | |
3319 | spin_lock(&res->spinlock); | |
3320 | dlm_set_lockres_owner(dlm, res, | |
3321 | DLM_LOCK_RES_OWNER_UNKNOWN); | |
3322 | dlm_move_lockres_to_recovery_list(dlm, res); | |
3323 | spin_unlock(&res->spinlock); | |
3324 | dlm_lockres_put(res); | |
3325 | ||
f671c09b KH |
3326 | /* about to get rid of mle, detach from heartbeat */ |
3327 | __dlm_mle_detach_hb_events(dlm, mle); | |
3328 | ||
6714d8e8 KH |
3329 | /* dump the mle */ |
3330 | spin_lock(&dlm->master_lock); | |
3331 | __dlm_put_mle(mle); | |
3332 | spin_unlock(&dlm->master_lock); | |
3333 | ||
3334 | /* restart */ | |
3335 | goto top; | |
3336 | } | |
3337 | ||
3338 | /* this may be the last reference */ | |
3339 | __dlm_put_mle(mle); | |
3340 | } | |
3341 | spin_unlock(&dlm->master_lock); | |
3342 | } | |
3343 | ||
3344 | ||
3345 | int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, | |
3346 | u8 old_master) | |
3347 | { | |
3348 | struct dlm_node_iter iter; | |
3349 | int ret = 0; | |
3350 | ||
3351 | spin_lock(&dlm->spinlock); | |
3352 | dlm_node_iter_init(dlm->domain_map, &iter); | |
3353 | clear_bit(old_master, iter.node_map); | |
3354 | clear_bit(dlm->node_num, iter.node_map); | |
3355 | spin_unlock(&dlm->spinlock); | |
3356 | ||
ba2bf218 KH |
3357 | /* ownership of the lockres is changing. account for the |
3358 | * mastery reference here since old_master will briefly have | |
3359 | * a reference after the migration completes */ | |
3360 | spin_lock(&res->spinlock); | |
3361 | dlm_lockres_set_refmap_bit(old_master, res); | |
3362 | spin_unlock(&res->spinlock); | |
3363 | ||
6714d8e8 KH |
3364 | mlog(0, "now time to do a migrate request to other nodes\n"); |
3365 | ret = dlm_do_migrate_request(dlm, res, old_master, | |
3366 | dlm->node_num, &iter); | |
3367 | if (ret < 0) { | |
3368 | mlog_errno(ret); | |
3369 | goto leave; | |
3370 | } | |
3371 | ||
3372 | mlog(0, "doing assert master of %.*s to all except the original node\n", | |
3373 | res->lockname.len, res->lockname.name); | |
3374 | /* this call now finishes out the nodemap | |
3375 | * even if one or more nodes die */ | |
ba2bf218 | 3376 | ret = dlm_do_assert_master(dlm, res, iter.node_map, |
6714d8e8 KH |
3377 | DLM_ASSERT_MASTER_FINISH_MIGRATION); |
3378 | if (ret < 0) { | |
3379 | /* no longer need to retry. all living nodes contacted. */ | |
3380 | mlog_errno(ret); | |
3381 | ret = 0; | |
3382 | } | |
3383 | ||
3384 | memset(iter.node_map, 0, sizeof(iter.node_map)); | |
3385 | set_bit(old_master, iter.node_map); | |
3386 | mlog(0, "doing assert master of %.*s back to %u\n", | |
3387 | res->lockname.len, res->lockname.name, old_master); | |
ba2bf218 | 3388 | ret = dlm_do_assert_master(dlm, res, iter.node_map, |
6714d8e8 KH |
3389 | DLM_ASSERT_MASTER_FINISH_MIGRATION); |
3390 | if (ret < 0) { | |
3391 | mlog(0, "assert master to original master failed " | |
3392 | "with %d.\n", ret); | |
3393 | /* the only nonzero status here would be because of | |
3394 | * a dead original node. we're done. */ | |
3395 | ret = 0; | |
3396 | } | |
3397 | ||
3398 | /* all done, set the owner, clear the flag */ | |
3399 | spin_lock(&res->spinlock); | |
3400 | dlm_set_lockres_owner(dlm, res, dlm->node_num); | |
3401 | res->state &= ~DLM_LOCK_RES_MIGRATING; | |
3402 | spin_unlock(&res->spinlock); | |
3403 | /* re-dirty it on the new master */ | |
3404 | dlm_kick_thread(dlm, res); | |
3405 | wake_up(&res->wq); | |
3406 | leave: | |
3407 | return ret; | |
3408 | } | |
3409 | ||
3410 | /* | |
3411 | * LOCKRES AST REFCOUNT | |
3412 | * this is integral to migration | |
3413 | */ | |
3414 | ||
3415 | /* for future intent to call an ast, reserve one ahead of time. | |
3416 | * this should be called only after waiting on the lockres | |
3417 | * with dlm_wait_on_lockres, and while still holding the | |
3418 | * spinlock after the call. */ | |
3419 | void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res) | |
3420 | { | |
3421 | assert_spin_locked(&res->spinlock); | |
3422 | if (res->state & DLM_LOCK_RES_MIGRATING) { | |
3423 | __dlm_print_one_lock_resource(res); | |
3424 | } | |
3425 | BUG_ON(res->state & DLM_LOCK_RES_MIGRATING); | |
3426 | ||
3427 | atomic_inc(&res->asts_reserved); | |
3428 | } | |
3429 | ||
3430 | /* | |
3431 | * used to drop the reserved ast, either because it went unused, | |
3432 | * or because the ast/bast was actually called. | |
3433 | * | |
3434 | * also, if there is a pending migration on this lockres, | |
3435 | * and this was the last pending ast on the lockres, | |
3436 | * atomically set the MIGRATING flag before we drop the lock. | |
3437 | * this is how we ensure that migration can proceed with no | |
3438 | * asts in progress. note that it is ok if the state of the | |
3439 | * queues is such that a lock should be granted in the future | |
3440 | * or that a bast should be fired, because the new master will | |
3441 | * shuffle the lists on this lockres as soon as it is migrated. | |
3442 | */ | |
3443 | void dlm_lockres_release_ast(struct dlm_ctxt *dlm, | |
3444 | struct dlm_lock_resource *res) | |
3445 | { | |
3446 | if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock)) | |
3447 | return; | |
3448 | ||
3449 | if (!res->migration_pending) { | |
3450 | spin_unlock(&res->spinlock); | |
3451 | return; | |
3452 | } | |
3453 | ||
3454 | BUG_ON(res->state & DLM_LOCK_RES_MIGRATING); | |
3455 | res->migration_pending = 0; | |
3456 | res->state |= DLM_LOCK_RES_MIGRATING; | |
3457 | spin_unlock(&res->spinlock); | |
3458 | wake_up(&res->wq); | |
3459 | wake_up(&dlm->migration_wq); | |
3460 | } |