]>
Commit | Line | Data |
---|---|---|
328970de | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
6714d8e8 KH |
2 | /* -*- mode: c; c-basic-offset: 8; -*- |
3 | * vim: noexpandtab sw=8 ts=8 sts=0: | |
4 | * | |
5 | * dlmthread.c | |
6 | * | |
7 | * standalone DLM module | |
8 | * | |
9 | * Copyright (C) 2004 Oracle. All rights reserved. | |
6714d8e8 KH |
10 | */ |
11 | ||
12 | ||
13 | #include <linux/module.h> | |
14 | #include <linux/fs.h> | |
15 | #include <linux/types.h> | |
6714d8e8 | 16 | #include <linux/highmem.h> |
6714d8e8 KH |
17 | #include <linux/init.h> |
18 | #include <linux/sysctl.h> | |
19 | #include <linux/random.h> | |
20 | #include <linux/blkdev.h> | |
21 | #include <linux/socket.h> | |
22 | #include <linux/inet.h> | |
23 | #include <linux/timer.h> | |
24 | #include <linux/kthread.h> | |
8d79d088 | 25 | #include <linux/delay.h> |
6714d8e8 KH |
26 | |
27 | ||
28 | #include "cluster/heartbeat.h" | |
29 | #include "cluster/nodemanager.h" | |
30 | #include "cluster/tcp.h" | |
31 | ||
32 | #include "dlmapi.h" | |
33 | #include "dlmcommon.h" | |
34 | #include "dlmdomain.h" | |
35 | ||
36 | #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD) | |
37 | #include "cluster/masklog.h" | |
38 | ||
6714d8e8 | 39 | static int dlm_thread(void *data); |
6714d8e8 KH |
40 | static void dlm_flush_asts(struct dlm_ctxt *dlm); |
41 | ||
42 | #define dlm_lock_is_remote(dlm, lock) ((lock)->ml.node != (dlm)->node_num) | |
43 | ||
44 | /* will exit holding res->spinlock, but may drop in function */ | |
45 | /* waits until flags are cleared on res->state */ | |
46 | void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags) | |
47 | { | |
48 | DECLARE_WAITQUEUE(wait, current); | |
49 | ||
50 | assert_spin_locked(&res->spinlock); | |
51 | ||
52 | add_wait_queue(&res->wq, &wait); | |
53 | repeat: | |
54 | set_current_state(TASK_UNINTERRUPTIBLE); | |
55 | if (res->state & flags) { | |
56 | spin_unlock(&res->spinlock); | |
57 | schedule(); | |
58 | spin_lock(&res->spinlock); | |
59 | goto repeat; | |
60 | } | |
61 | remove_wait_queue(&res->wq, &wait); | |
5c2c9d38 | 62 | __set_current_state(TASK_RUNNING); |
6714d8e8 KH |
63 | } |
64 | ||
ba2bf218 | 65 | int __dlm_lockres_has_locks(struct dlm_lock_resource *res) |
6714d8e8 KH |
66 | { |
67 | if (list_empty(&res->granted) && | |
68 | list_empty(&res->converting) && | |
ba2bf218 KH |
69 | list_empty(&res->blocked)) |
70 | return 0; | |
71 | return 1; | |
72 | } | |
73 | ||
74 | /* "unused": the lockres has no locks, is not on the dirty list, | |
75 | * has no inflight locks (in the gap between mastery and acquiring | |
76 | * the first lock), and has no bits in its refmap. | |
77 | * truly ready to be freed. */ | |
78 | int __dlm_lockres_unused(struct dlm_lock_resource *res) | |
79 | { | |
a524812b WW |
80 | int bit; |
81 | ||
ff0a522e SM |
82 | assert_spin_locked(&res->spinlock); |
83 | ||
a524812b WW |
84 | if (__dlm_lockres_has_locks(res)) |
85 | return 0; | |
86 | ||
ff0a522e SM |
87 | /* Locks are in the process of being created */ |
88 | if (res->inflight_locks) | |
89 | return 0; | |
90 | ||
a524812b WW |
91 | if (!list_empty(&res->dirty) || res->state & DLM_LOCK_RES_DIRTY) |
92 | return 0; | |
93 | ||
814ce694 JX |
94 | if (res->state & (DLM_LOCK_RES_RECOVERING| |
95 | DLM_LOCK_RES_RECOVERY_WAITING)) | |
a524812b WW |
96 | return 0; |
97 | ||
ff0a522e | 98 | /* Another node has this resource with this node as the master */ |
a524812b WW |
99 | bit = find_next_bit(res->refmap, O2NM_MAX_NODES, 0); |
100 | if (bit < O2NM_MAX_NODES) | |
101 | return 0; | |
102 | ||
a524812b | 103 | return 1; |
6714d8e8 KH |
104 | } |
105 | ||
106 | ||
107 | /* Call whenever you may have added or deleted something from one of | |
108 | * the lockres queue's. This will figure out whether it belongs on the | |
109 | * unused list or not and does the appropriate thing. */ | |
110 | void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm, | |
111 | struct dlm_lock_resource *res) | |
112 | { | |
6714d8e8 KH |
113 | assert_spin_locked(&dlm->spinlock); |
114 | assert_spin_locked(&res->spinlock); | |
115 | ||
116 | if (__dlm_lockres_unused(res)){ | |
117 | if (list_empty(&res->purge)) { | |
8e17d16f SM |
118 | mlog(0, "%s: Adding res %.*s to purge list\n", |
119 | dlm->name, res->lockname.len, res->lockname.name); | |
6714d8e8 KH |
120 | |
121 | res->last_used = jiffies; | |
ba2bf218 | 122 | dlm_lockres_get(res); |
6714d8e8 KH |
123 | list_add_tail(&res->purge, &dlm->purge_list); |
124 | dlm->purge_count++; | |
125 | } | |
126 | } else if (!list_empty(&res->purge)) { | |
8e17d16f SM |
127 | mlog(0, "%s: Removing res %.*s from purge list\n", |
128 | dlm->name, res->lockname.len, res->lockname.name); | |
6714d8e8 KH |
129 | |
130 | list_del_init(&res->purge); | |
ba2bf218 | 131 | dlm_lockres_put(res); |
6714d8e8 KH |
132 | dlm->purge_count--; |
133 | } | |
134 | } | |
135 | ||
136 | void dlm_lockres_calc_usage(struct dlm_ctxt *dlm, | |
137 | struct dlm_lock_resource *res) | |
138 | { | |
6714d8e8 KH |
139 | spin_lock(&dlm->spinlock); |
140 | spin_lock(&res->spinlock); | |
141 | ||
142 | __dlm_lockres_calc_usage(dlm, res); | |
143 | ||
144 | spin_unlock(&res->spinlock); | |
145 | spin_unlock(&dlm->spinlock); | |
146 | } | |
147 | ||
ee8f7fcb | 148 | /* |
149 | * Do the real purge work: | |
150 | * unhash the lockres, and | |
151 | * clear flag DLM_LOCK_RES_DROPPING_REF. | |
152 | * It requires dlm and lockres spinlock to be taken. | |
153 | */ | |
154 | void __dlm_do_purge_lockres(struct dlm_ctxt *dlm, | |
155 | struct dlm_lock_resource *res) | |
156 | { | |
157 | assert_spin_locked(&dlm->spinlock); | |
158 | assert_spin_locked(&res->spinlock); | |
159 | ||
160 | if (!list_empty(&res->purge)) { | |
161 | mlog(0, "%s: Removing res %.*s from purgelist\n", | |
162 | dlm->name, res->lockname.len, res->lockname.name); | |
163 | list_del_init(&res->purge); | |
164 | dlm_lockres_put(res); | |
165 | dlm->purge_count--; | |
166 | } | |
167 | ||
168 | if (!__dlm_lockres_unused(res)) { | |
169 | mlog(ML_ERROR, "%s: res %.*s in use after deref\n", | |
170 | dlm->name, res->lockname.len, res->lockname.name); | |
171 | __dlm_print_one_lock_resource(res); | |
172 | BUG(); | |
173 | } | |
174 | ||
175 | __dlm_unhash_lockres(dlm, res); | |
176 | ||
177 | spin_lock(&dlm->track_lock); | |
178 | if (!list_empty(&res->tracking)) | |
179 | list_del_init(&res->tracking); | |
180 | else { | |
181 | mlog(ML_ERROR, "%s: Resource %.*s not on the Tracking list\n", | |
182 | dlm->name, res->lockname.len, res->lockname.name); | |
183 | __dlm_print_one_lock_resource(res); | |
184 | } | |
185 | spin_unlock(&dlm->track_lock); | |
186 | ||
187 | /* | |
188 | * lockres is not in the hash now. drop the flag and wake up | |
189 | * any processes waiting in dlm_get_lock_resource. | |
190 | */ | |
191 | res->state &= ~DLM_LOCK_RES_DROPPING_REF; | |
192 | } | |
193 | ||
7beaf243 | 194 | static void dlm_purge_lockres(struct dlm_ctxt *dlm, |
faf0ec9f | 195 | struct dlm_lock_resource *res) |
6714d8e8 KH |
196 | { |
197 | int master; | |
ba2bf218 | 198 | int ret = 0; |
6714d8e8 | 199 | |
7beaf243 SE |
200 | assert_spin_locked(&dlm->spinlock); |
201 | assert_spin_locked(&res->spinlock); | |
516b7e52 | 202 | |
ba2bf218 | 203 | master = (res->owner == dlm->node_num); |
516b7e52 | 204 | |
8e17d16f SM |
205 | mlog(0, "%s: Purging res %.*s, master %d\n", dlm->name, |
206 | res->lockname.len, res->lockname.name, master); | |
ba2bf218 KH |
207 | |
208 | if (!master) { | |
309e9191 | 209 | if (res->state & DLM_LOCK_RES_DROPPING_REF) { |
ee8f7fcb | 210 | mlog(ML_NOTICE, "%s: res %.*s already in DLM_LOCK_RES_DROPPING_REF state\n", |
211 | dlm->name, res->lockname.len, res->lockname.name); | |
309e9191 | 212 | spin_unlock(&res->spinlock); |
213 | return; | |
214 | } | |
215 | ||
7beaf243 | 216 | res->state |= DLM_LOCK_RES_DROPPING_REF; |
c824c3c7 | 217 | /* drop spinlock... retake below */ |
7beaf243 | 218 | spin_unlock(&res->spinlock); |
c824c3c7 SM |
219 | spin_unlock(&dlm->spinlock); |
220 | ||
3b8118cf KH |
221 | spin_lock(&res->spinlock); |
222 | /* This ensures that clear refmap is sent after the set */ | |
7dc102b7 | 223 | __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG); |
3b8118cf | 224 | spin_unlock(&res->spinlock); |
c824c3c7 | 225 | |
ba2bf218 KH |
226 | /* clear our bit from the master's refmap, ignore errors */ |
227 | ret = dlm_drop_lockres_ref(dlm, res); | |
228 | if (ret < 0) { | |
ba2bf218 KH |
229 | if (!dlm_is_host_down(ret)) |
230 | BUG(); | |
231 | } | |
ba2bf218 | 232 | spin_lock(&dlm->spinlock); |
7beaf243 | 233 | spin_lock(&res->spinlock); |
6714d8e8 KH |
234 | } |
235 | ||
ba2bf218 | 236 | if (!list_empty(&res->purge)) { |
8e17d16f SM |
237 | mlog(0, "%s: Removing res %.*s from purgelist, master %d\n", |
238 | dlm->name, res->lockname.len, res->lockname.name, master); | |
ba2bf218 KH |
239 | list_del_init(&res->purge); |
240 | dlm_lockres_put(res); | |
6714d8e8 | 241 | dlm->purge_count--; |
7beaf243 SE |
242 | } |
243 | ||
309e9191 | 244 | if (!master && ret == DLM_DEREF_RESPONSE_INPROG) { |
245 | mlog(0, "%s: deref %.*s in progress\n", | |
842b90b6 | 246 | dlm->name, res->lockname.len, res->lockname.name); |
247 | spin_unlock(&res->spinlock); | |
248 | return; | |
249 | } | |
250 | ||
7beaf243 | 251 | if (!__dlm_lockres_unused(res)) { |
8e17d16f | 252 | mlog(ML_ERROR, "%s: res %.*s in use after deref\n", |
7beaf243 SE |
253 | dlm->name, res->lockname.len, res->lockname.name); |
254 | __dlm_print_one_lock_resource(res); | |
255 | BUG(); | |
256 | } | |
83e32d90 | 257 | |
e9f0b6a6 | 258 | __dlm_unhash_lockres(dlm, res); |
6714d8e8 | 259 | |
f57a22dd YJ |
260 | spin_lock(&dlm->track_lock); |
261 | if (!list_empty(&res->tracking)) | |
262 | list_del_init(&res->tracking); | |
263 | else { | |
264 | mlog(ML_ERROR, "Resource %.*s not on the Tracking list\n", | |
265 | res->lockname.len, res->lockname.name); | |
266 | __dlm_print_one_lock_resource(res); | |
267 | } | |
268 | spin_unlock(&dlm->track_lock); | |
269 | ||
ba2bf218 KH |
270 | /* lockres is not in the hash now. drop the flag and wake up |
271 | * any processes waiting in dlm_get_lock_resource. */ | |
272 | if (!master) { | |
ba2bf218 KH |
273 | res->state &= ~DLM_LOCK_RES_DROPPING_REF; |
274 | spin_unlock(&res->spinlock); | |
275 | wake_up(&res->wq); | |
7beaf243 SE |
276 | } else |
277 | spin_unlock(&res->spinlock); | |
8b219809 KH |
278 | } |
279 | ||
6714d8e8 KH |
280 | static void dlm_run_purge_list(struct dlm_ctxt *dlm, |
281 | int purge_now) | |
282 | { | |
283 | unsigned int run_max, unused; | |
284 | unsigned long purge_jiffies; | |
285 | struct dlm_lock_resource *lockres; | |
286 | ||
287 | spin_lock(&dlm->spinlock); | |
288 | run_max = dlm->purge_count; | |
289 | ||
290 | while(run_max && !list_empty(&dlm->purge_list)) { | |
291 | run_max--; | |
292 | ||
293 | lockres = list_entry(dlm->purge_list.next, | |
294 | struct dlm_lock_resource, purge); | |
295 | ||
6714d8e8 | 296 | spin_lock(&lockres->spinlock); |
6714d8e8 KH |
297 | |
298 | purge_jiffies = lockres->last_used + | |
299 | msecs_to_jiffies(DLM_PURGE_INTERVAL_MS); | |
300 | ||
301 | /* Make sure that we want to be processing this guy at | |
302 | * this time. */ | |
303 | if (!purge_now && time_after(purge_jiffies, jiffies)) { | |
304 | /* Since resources are added to the purge list | |
305 | * in tail order, we can stop at the first | |
306 | * unpurgable resource -- anyone added after | |
307 | * him will have a greater last_used value */ | |
7beaf243 | 308 | spin_unlock(&lockres->spinlock); |
6714d8e8 KH |
309 | break; |
310 | } | |
311 | ||
7beaf243 SE |
312 | /* Status of the lockres *might* change so double |
313 | * check. If the lockres is unused, holding the dlm | |
314 | * spinlock will prevent people from getting and more | |
315 | * refs on it. */ | |
316 | unused = __dlm_lockres_unused(lockres); | |
317 | if (!unused || | |
ac4fef4d X |
318 | (lockres->state & DLM_LOCK_RES_MIGRATING) || |
319 | (lockres->inflight_assert_workers != 0)) { | |
8e17d16f | 320 | mlog(0, "%s: res %.*s is in use or being remastered, " |
ac4fef4d X |
321 | "used %d, state %d, assert master workers %u\n", |
322 | dlm->name, lockres->lockname.len, | |
323 | lockres->lockname.name, | |
324 | !unused, lockres->state, | |
325 | lockres->inflight_assert_workers); | |
a270c6d3 | 326 | list_move_tail(&lockres->purge, &dlm->purge_list); |
7beaf243 SE |
327 | spin_unlock(&lockres->spinlock); |
328 | continue; | |
329 | } | |
330 | ||
78062cb2 | 331 | dlm_lockres_get(lockres); |
6714d8e8 | 332 | |
7beaf243 | 333 | dlm_purge_lockres(dlm, lockres); |
78062cb2 | 334 | |
3fca0894 | 335 | dlm_lockres_put(lockres); |
6714d8e8 KH |
336 | |
337 | /* Avoid adding any scheduling latencies */ | |
338 | cond_resched_lock(&dlm->spinlock); | |
339 | } | |
340 | ||
341 | spin_unlock(&dlm->spinlock); | |
342 | } | |
343 | ||
344 | static void dlm_shuffle_lists(struct dlm_ctxt *dlm, | |
345 | struct dlm_lock_resource *res) | |
346 | { | |
347 | struct dlm_lock *lock, *target; | |
6714d8e8 KH |
348 | int can_grant = 1; |
349 | ||
8e17d16f SM |
350 | /* |
351 | * Because this function is called with the lockres | |
6714d8e8 KH |
352 | * spinlock, and because we know that it is not migrating/ |
353 | * recovering/in-progress, it is fine to reserve asts and | |
8e17d16f SM |
354 | * basts right before queueing them all throughout |
355 | */ | |
d9ef7522 | 356 | assert_spin_locked(&dlm->ast_lock); |
6714d8e8 KH |
357 | assert_spin_locked(&res->spinlock); |
358 | BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING| | |
359 | DLM_LOCK_RES_RECOVERING| | |
360 | DLM_LOCK_RES_IN_PROGRESS))); | |
361 | ||
362 | converting: | |
363 | if (list_empty(&res->converting)) | |
364 | goto blocked; | |
8e17d16f SM |
365 | mlog(0, "%s: res %.*s has locks on the convert queue\n", dlm->name, |
366 | res->lockname.len, res->lockname.name); | |
6714d8e8 KH |
367 | |
368 | target = list_entry(res->converting.next, struct dlm_lock, list); | |
369 | if (target->ml.convert_type == LKM_IVMODE) { | |
8e17d16f SM |
370 | mlog(ML_ERROR, "%s: res %.*s converting lock to invalid mode\n", |
371 | dlm->name, res->lockname.len, res->lockname.name); | |
6714d8e8 KH |
372 | BUG(); |
373 | } | |
df53cd3b | 374 | list_for_each_entry(lock, &res->granted, list) { |
6714d8e8 KH |
375 | if (lock==target) |
376 | continue; | |
377 | if (!dlm_lock_compatible(lock->ml.type, | |
378 | target->ml.convert_type)) { | |
379 | can_grant = 0; | |
380 | /* queue the BAST if not already */ | |
381 | if (lock->ml.highest_blocked == LKM_IVMODE) { | |
382 | __dlm_lockres_reserve_ast(res); | |
d9ef7522 | 383 | __dlm_queue_bast(dlm, lock); |
6714d8e8 KH |
384 | } |
385 | /* update the highest_blocked if needed */ | |
386 | if (lock->ml.highest_blocked < target->ml.convert_type) | |
387 | lock->ml.highest_blocked = | |
388 | target->ml.convert_type; | |
389 | } | |
390 | } | |
df53cd3b DF |
391 | |
392 | list_for_each_entry(lock, &res->converting, list) { | |
6714d8e8 KH |
393 | if (lock==target) |
394 | continue; | |
395 | if (!dlm_lock_compatible(lock->ml.type, | |
396 | target->ml.convert_type)) { | |
397 | can_grant = 0; | |
398 | if (lock->ml.highest_blocked == LKM_IVMODE) { | |
399 | __dlm_lockres_reserve_ast(res); | |
d9ef7522 | 400 | __dlm_queue_bast(dlm, lock); |
6714d8e8 KH |
401 | } |
402 | if (lock->ml.highest_blocked < target->ml.convert_type) | |
403 | lock->ml.highest_blocked = | |
404 | target->ml.convert_type; | |
405 | } | |
406 | } | |
407 | ||
408 | /* we can convert the lock */ | |
409 | if (can_grant) { | |
410 | spin_lock(&target->spinlock); | |
411 | BUG_ON(target->ml.highest_blocked != LKM_IVMODE); | |
412 | ||
8e17d16f SM |
413 | mlog(0, "%s: res %.*s, AST for Converting lock %u:%llu, type " |
414 | "%d => %d, node %u\n", dlm->name, res->lockname.len, | |
415 | res->lockname.name, | |
416 | dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)), | |
417 | dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)), | |
418 | target->ml.type, | |
6714d8e8 KH |
419 | target->ml.convert_type, target->ml.node); |
420 | ||
421 | target->ml.type = target->ml.convert_type; | |
422 | target->ml.convert_type = LKM_IVMODE; | |
f116629d | 423 | list_move_tail(&target->list, &res->granted); |
6714d8e8 KH |
424 | |
425 | BUG_ON(!target->lksb); | |
426 | target->lksb->status = DLM_NORMAL; | |
427 | ||
428 | spin_unlock(&target->spinlock); | |
429 | ||
430 | __dlm_lockres_reserve_ast(res); | |
d9ef7522 | 431 | __dlm_queue_ast(dlm, target); |
6714d8e8 KH |
432 | /* go back and check for more */ |
433 | goto converting; | |
434 | } | |
435 | ||
436 | blocked: | |
437 | if (list_empty(&res->blocked)) | |
438 | goto leave; | |
439 | target = list_entry(res->blocked.next, struct dlm_lock, list); | |
440 | ||
df53cd3b | 441 | list_for_each_entry(lock, &res->granted, list) { |
6714d8e8 KH |
442 | if (lock==target) |
443 | continue; | |
444 | if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) { | |
445 | can_grant = 0; | |
446 | if (lock->ml.highest_blocked == LKM_IVMODE) { | |
447 | __dlm_lockres_reserve_ast(res); | |
d9ef7522 | 448 | __dlm_queue_bast(dlm, lock); |
6714d8e8 KH |
449 | } |
450 | if (lock->ml.highest_blocked < target->ml.type) | |
451 | lock->ml.highest_blocked = target->ml.type; | |
452 | } | |
453 | } | |
454 | ||
df53cd3b | 455 | list_for_each_entry(lock, &res->converting, list) { |
6714d8e8 KH |
456 | if (lock==target) |
457 | continue; | |
458 | if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) { | |
459 | can_grant = 0; | |
460 | if (lock->ml.highest_blocked == LKM_IVMODE) { | |
461 | __dlm_lockres_reserve_ast(res); | |
d9ef7522 | 462 | __dlm_queue_bast(dlm, lock); |
6714d8e8 KH |
463 | } |
464 | if (lock->ml.highest_blocked < target->ml.type) | |
465 | lock->ml.highest_blocked = target->ml.type; | |
466 | } | |
467 | } | |
468 | ||
469 | /* we can grant the blocked lock (only | |
470 | * possible if converting list empty) */ | |
471 | if (can_grant) { | |
472 | spin_lock(&target->spinlock); | |
473 | BUG_ON(target->ml.highest_blocked != LKM_IVMODE); | |
474 | ||
8e17d16f SM |
475 | mlog(0, "%s: res %.*s, AST for Blocked lock %u:%llu, type %d, " |
476 | "node %u\n", dlm->name, res->lockname.len, | |
477 | res->lockname.name, | |
478 | dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)), | |
479 | dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)), | |
6714d8e8 KH |
480 | target->ml.type, target->ml.node); |
481 | ||
8e17d16f | 482 | /* target->ml.type is already correct */ |
f116629d | 483 | list_move_tail(&target->list, &res->granted); |
6714d8e8 KH |
484 | |
485 | BUG_ON(!target->lksb); | |
486 | target->lksb->status = DLM_NORMAL; | |
487 | ||
488 | spin_unlock(&target->spinlock); | |
489 | ||
490 | __dlm_lockres_reserve_ast(res); | |
d9ef7522 | 491 | __dlm_queue_ast(dlm, target); |
6714d8e8 KH |
492 | /* go back and check for more */ |
493 | goto converting; | |
494 | } | |
495 | ||
496 | leave: | |
497 | return; | |
498 | } | |
499 | ||
500 | /* must have NO locks when calling this with res !=NULL * */ | |
501 | void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) | |
502 | { | |
6714d8e8 KH |
503 | if (res) { |
504 | spin_lock(&dlm->spinlock); | |
505 | spin_lock(&res->spinlock); | |
506 | __dlm_dirty_lockres(dlm, res); | |
507 | spin_unlock(&res->spinlock); | |
508 | spin_unlock(&dlm->spinlock); | |
509 | } | |
510 | wake_up(&dlm->dlm_thread_wq); | |
511 | } | |
512 | ||
513 | void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res) | |
514 | { | |
6714d8e8 KH |
515 | assert_spin_locked(&dlm->spinlock); |
516 | assert_spin_locked(&res->spinlock); | |
517 | ||
518 | /* don't shuffle secondary queues */ | |
32c1b90d | 519 | if (res->owner == dlm->node_num) { |
ddc09c8d KH |
520 | if (res->state & (DLM_LOCK_RES_MIGRATING | |
521 | DLM_LOCK_RES_BLOCK_DIRTY)) | |
522 | return; | |
523 | ||
524 | if (list_empty(&res->dirty)) { | |
525 | /* ref for dirty_list */ | |
526 | dlm_lockres_get(res); | |
527 | list_add_tail(&res->dirty, &dlm->dirty_list); | |
528 | res->state |= DLM_LOCK_RES_DIRTY; | |
529 | } | |
6714d8e8 | 530 | } |
8e17d16f SM |
531 | |
532 | mlog(0, "%s: res %.*s\n", dlm->name, res->lockname.len, | |
533 | res->lockname.name); | |
6714d8e8 KH |
534 | } |
535 | ||
536 | ||
537 | /* Launch the NM thread for the mounted volume */ | |
538 | int dlm_launch_thread(struct dlm_ctxt *dlm) | |
539 | { | |
8e17d16f | 540 | mlog(0, "Starting dlm_thread...\n"); |
6714d8e8 | 541 | |
5afc44e2 JQ |
542 | dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm-%s", |
543 | dlm->name); | |
6714d8e8 KH |
544 | if (IS_ERR(dlm->dlm_thread_task)) { |
545 | mlog_errno(PTR_ERR(dlm->dlm_thread_task)); | |
546 | dlm->dlm_thread_task = NULL; | |
547 | return -EINVAL; | |
548 | } | |
549 | ||
550 | return 0; | |
551 | } | |
552 | ||
553 | void dlm_complete_thread(struct dlm_ctxt *dlm) | |
554 | { | |
555 | if (dlm->dlm_thread_task) { | |
8e17d16f | 556 | mlog(ML_KTHREAD, "Waiting for dlm thread to exit\n"); |
6714d8e8 KH |
557 | kthread_stop(dlm->dlm_thread_task); |
558 | dlm->dlm_thread_task = NULL; | |
559 | } | |
560 | } | |
561 | ||
562 | static int dlm_dirty_list_empty(struct dlm_ctxt *dlm) | |
563 | { | |
564 | int empty; | |
565 | ||
566 | spin_lock(&dlm->spinlock); | |
567 | empty = list_empty(&dlm->dirty_list); | |
568 | spin_unlock(&dlm->spinlock); | |
569 | ||
570 | return empty; | |
571 | } | |
572 | ||
573 | static void dlm_flush_asts(struct dlm_ctxt *dlm) | |
574 | { | |
575 | int ret; | |
576 | struct dlm_lock *lock; | |
577 | struct dlm_lock_resource *res; | |
578 | u8 hi; | |
579 | ||
580 | spin_lock(&dlm->ast_lock); | |
581 | while (!list_empty(&dlm->pending_asts)) { | |
582 | lock = list_entry(dlm->pending_asts.next, | |
583 | struct dlm_lock, ast_list); | |
584 | /* get an extra ref on lock */ | |
585 | dlm_lock_get(lock); | |
586 | res = lock->lockres; | |
8e17d16f SM |
587 | mlog(0, "%s: res %.*s, Flush AST for lock %u:%llu, type %d, " |
588 | "node %u\n", dlm->name, res->lockname.len, | |
589 | res->lockname.name, | |
590 | dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), | |
591 | dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), | |
592 | lock->ml.type, lock->ml.node); | |
6714d8e8 KH |
593 | |
594 | BUG_ON(!lock->ast_pending); | |
595 | ||
596 | /* remove from list (including ref) */ | |
597 | list_del_init(&lock->ast_list); | |
598 | dlm_lock_put(lock); | |
599 | spin_unlock(&dlm->ast_lock); | |
600 | ||
601 | if (lock->ml.node != dlm->node_num) { | |
602 | ret = dlm_do_remote_ast(dlm, res, lock); | |
603 | if (ret < 0) | |
604 | mlog_errno(ret); | |
605 | } else | |
606 | dlm_do_local_ast(dlm, res, lock); | |
607 | ||
608 | spin_lock(&dlm->ast_lock); | |
609 | ||
610 | /* possible that another ast was queued while | |
611 | * we were delivering the last one */ | |
612 | if (!list_empty(&lock->ast_list)) { | |
8e17d16f SM |
613 | mlog(0, "%s: res %.*s, AST queued while flushing last " |
614 | "one\n", dlm->name, res->lockname.len, | |
615 | res->lockname.name); | |
6714d8e8 KH |
616 | } else |
617 | lock->ast_pending = 0; | |
618 | ||
619 | /* drop the extra ref. | |
620 | * this may drop it completely. */ | |
621 | dlm_lock_put(lock); | |
622 | dlm_lockres_release_ast(dlm, res); | |
623 | } | |
624 | ||
625 | while (!list_empty(&dlm->pending_basts)) { | |
626 | lock = list_entry(dlm->pending_basts.next, | |
627 | struct dlm_lock, bast_list); | |
628 | /* get an extra ref on lock */ | |
629 | dlm_lock_get(lock); | |
630 | res = lock->lockres; | |
631 | ||
632 | BUG_ON(!lock->bast_pending); | |
633 | ||
634 | /* get the highest blocked lock, and reset */ | |
635 | spin_lock(&lock->spinlock); | |
636 | BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE); | |
637 | hi = lock->ml.highest_blocked; | |
638 | lock->ml.highest_blocked = LKM_IVMODE; | |
639 | spin_unlock(&lock->spinlock); | |
640 | ||
641 | /* remove from list (including ref) */ | |
642 | list_del_init(&lock->bast_list); | |
643 | dlm_lock_put(lock); | |
644 | spin_unlock(&dlm->ast_lock); | |
645 | ||
8e17d16f SM |
646 | mlog(0, "%s: res %.*s, Flush BAST for lock %u:%llu, " |
647 | "blocked %d, node %u\n", | |
648 | dlm->name, res->lockname.len, res->lockname.name, | |
649 | dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), | |
650 | dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), | |
651 | hi, lock->ml.node); | |
6714d8e8 KH |
652 | |
653 | if (lock->ml.node != dlm->node_num) { | |
654 | ret = dlm_send_proxy_bast(dlm, res, lock, hi); | |
655 | if (ret < 0) | |
656 | mlog_errno(ret); | |
657 | } else | |
658 | dlm_do_local_bast(dlm, res, lock, hi); | |
659 | ||
660 | spin_lock(&dlm->ast_lock); | |
661 | ||
662 | /* possible that another bast was queued while | |
663 | * we were delivering the last one */ | |
664 | if (!list_empty(&lock->bast_list)) { | |
8e17d16f SM |
665 | mlog(0, "%s: res %.*s, BAST queued while flushing last " |
666 | "one\n", dlm->name, res->lockname.len, | |
667 | res->lockname.name); | |
6714d8e8 KH |
668 | } else |
669 | lock->bast_pending = 0; | |
670 | ||
671 | /* drop the extra ref. | |
672 | * this may drop it completely. */ | |
673 | dlm_lock_put(lock); | |
674 | dlm_lockres_release_ast(dlm, res); | |
675 | } | |
676 | wake_up(&dlm->ast_wq); | |
677 | spin_unlock(&dlm->ast_lock); | |
678 | } | |
679 | ||
680 | ||
681 | #define DLM_THREAD_TIMEOUT_MS (4 * 1000) | |
682 | #define DLM_THREAD_MAX_DIRTY 100 | |
683 | #define DLM_THREAD_MAX_ASTS 10 | |
684 | ||
685 | static int dlm_thread(void *data) | |
686 | { | |
687 | struct dlm_lock_resource *res; | |
688 | struct dlm_ctxt *dlm = data; | |
689 | unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS); | |
690 | ||
691 | mlog(0, "dlm thread running for %s...\n", dlm->name); | |
692 | ||
693 | while (!kthread_should_stop()) { | |
694 | int n = DLM_THREAD_MAX_DIRTY; | |
695 | ||
696 | /* dlm_shutting_down is very point-in-time, but that | |
697 | * doesn't matter as we'll just loop back around if we | |
698 | * get false on the leading edge of a state | |
699 | * transition. */ | |
700 | dlm_run_purge_list(dlm, dlm_shutting_down(dlm)); | |
701 | ||
702 | /* We really don't want to hold dlm->spinlock while | |
703 | * calling dlm_shuffle_lists on each lockres that | |
704 | * needs to have its queues adjusted and AST/BASTs | |
705 | * run. So let's pull each entry off the dirty_list | |
706 | * and drop dlm->spinlock ASAP. Once off the list, | |
707 | * res->spinlock needs to be taken again to protect | |
708 | * the queues while calling dlm_shuffle_lists. */ | |
709 | spin_lock(&dlm->spinlock); | |
710 | while (!list_empty(&dlm->dirty_list)) { | |
711 | int delay = 0; | |
712 | res = list_entry(dlm->dirty_list.next, | |
713 | struct dlm_lock_resource, dirty); | |
714 | ||
715 | /* peel a lockres off, remove it from the list, | |
716 | * unset the dirty flag and drop the dlm lock */ | |
717 | BUG_ON(!res); | |
718 | dlm_lockres_get(res); | |
719 | ||
720 | spin_lock(&res->spinlock); | |
ddc09c8d | 721 | /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */ |
6714d8e8 KH |
722 | list_del_init(&res->dirty); |
723 | spin_unlock(&res->spinlock); | |
724 | spin_unlock(&dlm->spinlock); | |
6ff06a93 KH |
725 | /* Drop dirty_list ref */ |
726 | dlm_lockres_put(res); | |
6714d8e8 KH |
727 | |
728 | /* lockres can be re-dirtied/re-added to the | |
729 | * dirty_list in this gap, but that is ok */ | |
730 | ||
d9ef7522 | 731 | spin_lock(&dlm->ast_lock); |
6714d8e8 KH |
732 | spin_lock(&res->spinlock); |
733 | if (res->owner != dlm->node_num) { | |
734 | __dlm_print_one_lock_resource(res); | |
8e17d16f SM |
735 | mlog(ML_ERROR, "%s: inprog %d, mig %d, reco %d," |
736 | " dirty %d\n", dlm->name, | |
737 | !!(res->state & DLM_LOCK_RES_IN_PROGRESS), | |
738 | !!(res->state & DLM_LOCK_RES_MIGRATING), | |
739 | !!(res->state & DLM_LOCK_RES_RECOVERING), | |
740 | !!(res->state & DLM_LOCK_RES_DIRTY)); | |
6714d8e8 KH |
741 | } |
742 | BUG_ON(res->owner != dlm->node_num); | |
743 | ||
744 | /* it is now ok to move lockreses in these states | |
745 | * to the dirty list, assuming that they will only be | |
746 | * dirty for a short while. */ | |
ddc09c8d | 747 | BUG_ON(res->state & DLM_LOCK_RES_MIGRATING); |
6714d8e8 | 748 | if (res->state & (DLM_LOCK_RES_IN_PROGRESS | |
814ce694 JX |
749 | DLM_LOCK_RES_RECOVERING | |
750 | DLM_LOCK_RES_RECOVERY_WAITING)) { | |
6714d8e8 | 751 | /* move it to the tail and keep going */ |
ddc09c8d | 752 | res->state &= ~DLM_LOCK_RES_DIRTY; |
6714d8e8 | 753 | spin_unlock(&res->spinlock); |
d9ef7522 | 754 | spin_unlock(&dlm->ast_lock); |
8e17d16f SM |
755 | mlog(0, "%s: res %.*s, inprogress, delay list " |
756 | "shuffle, state %d\n", dlm->name, | |
6714d8e8 KH |
757 | res->lockname.len, res->lockname.name, |
758 | res->state); | |
759 | delay = 1; | |
760 | goto in_progress; | |
761 | } | |
762 | ||
763 | /* at this point the lockres is not migrating/ | |
764 | * recovering/in-progress. we have the lockres | |
765 | * spinlock and do NOT have the dlm lock. | |
766 | * safe to reserve/queue asts and run the lists. */ | |
767 | ||
6714d8e8 KH |
768 | /* called while holding lockres lock */ |
769 | dlm_shuffle_lists(dlm, res); | |
ddc09c8d | 770 | res->state &= ~DLM_LOCK_RES_DIRTY; |
6714d8e8 | 771 | spin_unlock(&res->spinlock); |
d9ef7522 | 772 | spin_unlock(&dlm->ast_lock); |
6714d8e8 KH |
773 | |
774 | dlm_lockres_calc_usage(dlm, res); | |
775 | ||
776 | in_progress: | |
777 | ||
778 | spin_lock(&dlm->spinlock); | |
779 | /* if the lock was in-progress, stick | |
780 | * it on the back of the list */ | |
781 | if (delay) { | |
782 | spin_lock(&res->spinlock); | |
ddc09c8d | 783 | __dlm_dirty_lockres(dlm, res); |
6714d8e8 KH |
784 | spin_unlock(&res->spinlock); |
785 | } | |
786 | dlm_lockres_put(res); | |
787 | ||
788 | /* unlikely, but we may need to give time to | |
789 | * other tasks */ | |
790 | if (!--n) { | |
8e17d16f SM |
791 | mlog(0, "%s: Throttling dlm thread\n", |
792 | dlm->name); | |
6714d8e8 KH |
793 | break; |
794 | } | |
795 | } | |
796 | ||
797 | spin_unlock(&dlm->spinlock); | |
798 | dlm_flush_asts(dlm); | |
799 | ||
800 | /* yield and continue right away if there is more work to do */ | |
801 | if (!n) { | |
f85cd47a | 802 | cond_resched(); |
6714d8e8 KH |
803 | continue; |
804 | } | |
805 | ||
806 | wait_event_interruptible_timeout(dlm->dlm_thread_wq, | |
807 | !dlm_dirty_list_empty(dlm) || | |
808 | kthread_should_stop(), | |
809 | timeout); | |
810 | } | |
811 | ||
812 | mlog(0, "quitting DLM thread\n"); | |
813 | return 0; | |
814 | } |