dlmthread.c 21 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763
  1. /* -*- mode: c; c-basic-offset: 8; -*-
  2. * vim: noexpandtab sw=8 ts=8 sts=0:
  3. *
  4. * dlmthread.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. #include <linux/module.h>
  27. #include <linux/fs.h>
  28. #include <linux/types.h>
  29. #include <linux/highmem.h>
  30. #include <linux/init.h>
  31. #include <linux/sysctl.h>
  32. #include <linux/random.h>
  33. #include <linux/blkdev.h>
  34. #include <linux/socket.h>
  35. #include <linux/inet.h>
  36. #include <linux/timer.h>
  37. #include <linux/kthread.h>
  38. #include <linux/delay.h>
  39. #include "cluster/heartbeat.h"
  40. #include "cluster/nodemanager.h"
  41. #include "cluster/tcp.h"
  42. #include "dlmapi.h"
  43. #include "dlmcommon.h"
  44. #include "dlmdomain.h"
  45. #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
  46. #include "cluster/masklog.h"
  47. static int dlm_thread(void *data);
  48. static void dlm_flush_asts(struct dlm_ctxt *dlm);
  49. #define dlm_lock_is_remote(dlm, lock) ((lock)->ml.node != (dlm)->node_num)
  50. /* will exit holding res->spinlock, but may drop in function */
  51. /* waits until flags are cleared on res->state */
  52. void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
  53. {
  54. DECLARE_WAITQUEUE(wait, current);
  55. assert_spin_locked(&res->spinlock);
  56. add_wait_queue(&res->wq, &wait);
  57. repeat:
  58. set_current_state(TASK_UNINTERRUPTIBLE);
  59. if (res->state & flags) {
  60. spin_unlock(&res->spinlock);
  61. schedule();
  62. spin_lock(&res->spinlock);
  63. goto repeat;
  64. }
  65. remove_wait_queue(&res->wq, &wait);
  66. __set_current_state(TASK_RUNNING);
  67. }
  68. int __dlm_lockres_has_locks(struct dlm_lock_resource *res)
  69. {
  70. if (list_empty(&res->granted) &&
  71. list_empty(&res->converting) &&
  72. list_empty(&res->blocked))
  73. return 0;
  74. return 1;
  75. }
  76. /* "unused": the lockres has no locks, is not on the dirty list,
  77. * has no inflight locks (in the gap between mastery and acquiring
  78. * the first lock), and has no bits in its refmap.
  79. * truly ready to be freed. */
  80. int __dlm_lockres_unused(struct dlm_lock_resource *res)
  81. {
  82. int bit;
  83. if (__dlm_lockres_has_locks(res))
  84. return 0;
  85. if (!list_empty(&res->dirty) || res->state & DLM_LOCK_RES_DIRTY)
  86. return 0;
  87. if (res->state & DLM_LOCK_RES_RECOVERING)
  88. return 0;
  89. bit = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
  90. if (bit < O2NM_MAX_NODES)
  91. return 0;
  92. /*
  93. * since the bit for dlm->node_num is not set, inflight_locks better
  94. * be zero
  95. */
  96. BUG_ON(res->inflight_locks != 0);
  97. return 1;
  98. }
  99. /* Call whenever you may have added or deleted something from one of
  100. * the lockres queue's. This will figure out whether it belongs on the
  101. * unused list or not and does the appropriate thing. */
  102. void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
  103. struct dlm_lock_resource *res)
  104. {
  105. assert_spin_locked(&dlm->spinlock);
  106. assert_spin_locked(&res->spinlock);
  107. if (__dlm_lockres_unused(res)){
  108. if (list_empty(&res->purge)) {
  109. mlog(0, "%s: Adding res %.*s to purge list\n",
  110. dlm->name, res->lockname.len, res->lockname.name);
  111. res->last_used = jiffies;
  112. dlm_lockres_get(res);
  113. list_add_tail(&res->purge, &dlm->purge_list);
  114. dlm->purge_count++;
  115. }
  116. } else if (!list_empty(&res->purge)) {
  117. mlog(0, "%s: Removing res %.*s from purge list\n",
  118. dlm->name, res->lockname.len, res->lockname.name);
  119. list_del_init(&res->purge);
  120. dlm_lockres_put(res);
  121. dlm->purge_count--;
  122. }
  123. }
  124. void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
  125. struct dlm_lock_resource *res)
  126. {
  127. spin_lock(&dlm->spinlock);
  128. spin_lock(&res->spinlock);
  129. __dlm_lockres_calc_usage(dlm, res);
  130. spin_unlock(&res->spinlock);
  131. spin_unlock(&dlm->spinlock);
  132. }
  133. static void dlm_purge_lockres(struct dlm_ctxt *dlm,
  134. struct dlm_lock_resource *res)
  135. {
  136. int master;
  137. int ret = 0;
  138. assert_spin_locked(&dlm->spinlock);
  139. assert_spin_locked(&res->spinlock);
  140. master = (res->owner == dlm->node_num);
  141. mlog(0, "%s: Purging res %.*s, master %d\n", dlm->name,
  142. res->lockname.len, res->lockname.name, master);
  143. if (!master) {
  144. res->state |= DLM_LOCK_RES_DROPPING_REF;
  145. /* drop spinlock... retake below */
  146. spin_unlock(&res->spinlock);
  147. spin_unlock(&dlm->spinlock);
  148. spin_lock(&res->spinlock);
  149. /* This ensures that clear refmap is sent after the set */
  150. __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
  151. spin_unlock(&res->spinlock);
  152. /* clear our bit from the master's refmap, ignore errors */
  153. ret = dlm_drop_lockres_ref(dlm, res);
  154. if (ret < 0) {
  155. mlog(ML_ERROR, "%s: deref %.*s failed %d\n", dlm->name,
  156. res->lockname.len, res->lockname.name, ret);
  157. if (!dlm_is_host_down(ret))
  158. BUG();
  159. }
  160. spin_lock(&dlm->spinlock);
  161. spin_lock(&res->spinlock);
  162. }
  163. if (!list_empty(&res->purge)) {
  164. mlog(0, "%s: Removing res %.*s from purgelist, master %d\n",
  165. dlm->name, res->lockname.len, res->lockname.name, master);
  166. list_del_init(&res->purge);
  167. dlm_lockres_put(res);
  168. dlm->purge_count--;
  169. }
  170. if (!__dlm_lockres_unused(res)) {
  171. mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
  172. dlm->name, res->lockname.len, res->lockname.name);
  173. __dlm_print_one_lock_resource(res);
  174. BUG();
  175. }
  176. __dlm_unhash_lockres(res);
  177. /* lockres is not in the hash now. drop the flag and wake up
  178. * any processes waiting in dlm_get_lock_resource. */
  179. if (!master) {
  180. res->state &= ~DLM_LOCK_RES_DROPPING_REF;
  181. spin_unlock(&res->spinlock);
  182. wake_up(&res->wq);
  183. } else
  184. spin_unlock(&res->spinlock);
  185. }
  186. static void dlm_run_purge_list(struct dlm_ctxt *dlm,
  187. int purge_now)
  188. {
  189. unsigned int run_max, unused;
  190. unsigned long purge_jiffies;
  191. struct dlm_lock_resource *lockres;
  192. spin_lock(&dlm->spinlock);
  193. run_max = dlm->purge_count;
  194. while(run_max && !list_empty(&dlm->purge_list)) {
  195. run_max--;
  196. lockres = list_entry(dlm->purge_list.next,
  197. struct dlm_lock_resource, purge);
  198. spin_lock(&lockres->spinlock);
  199. purge_jiffies = lockres->last_used +
  200. msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
  201. /* Make sure that we want to be processing this guy at
  202. * this time. */
  203. if (!purge_now && time_after(purge_jiffies, jiffies)) {
  204. /* Since resources are added to the purge list
  205. * in tail order, we can stop at the first
  206. * unpurgable resource -- anyone added after
  207. * him will have a greater last_used value */
  208. spin_unlock(&lockres->spinlock);
  209. break;
  210. }
  211. /* Status of the lockres *might* change so double
  212. * check. If the lockres is unused, holding the dlm
  213. * spinlock will prevent people from getting and more
  214. * refs on it. */
  215. unused = __dlm_lockres_unused(lockres);
  216. if (!unused ||
  217. (lockres->state & DLM_LOCK_RES_MIGRATING)) {
  218. mlog(0, "%s: res %.*s is in use or being remastered, "
  219. "used %d, state %d\n", dlm->name,
  220. lockres->lockname.len, lockres->lockname.name,
  221. !unused, lockres->state);
  222. list_move_tail(&dlm->purge_list, &lockres->purge);
  223. spin_unlock(&lockres->spinlock);
  224. continue;
  225. }
  226. dlm_lockres_get(lockres);
  227. dlm_purge_lockres(dlm, lockres);
  228. dlm_lockres_put(lockres);
  229. /* Avoid adding any scheduling latencies */
  230. cond_resched_lock(&dlm->spinlock);
  231. }
  232. spin_unlock(&dlm->spinlock);
  233. }
  234. static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
  235. struct dlm_lock_resource *res)
  236. {
  237. struct dlm_lock *lock, *target;
  238. struct list_head *iter;
  239. struct list_head *head;
  240. int can_grant = 1;
  241. /*
  242. * Because this function is called with the lockres
  243. * spinlock, and because we know that it is not migrating/
  244. * recovering/in-progress, it is fine to reserve asts and
  245. * basts right before queueing them all throughout
  246. */
  247. assert_spin_locked(&dlm->ast_lock);
  248. assert_spin_locked(&res->spinlock);
  249. BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
  250. DLM_LOCK_RES_RECOVERING|
  251. DLM_LOCK_RES_IN_PROGRESS)));
  252. converting:
  253. if (list_empty(&res->converting))
  254. goto blocked;
  255. mlog(0, "%s: res %.*s has locks on the convert queue\n", dlm->name,
  256. res->lockname.len, res->lockname.name);
  257. target = list_entry(res->converting.next, struct dlm_lock, list);
  258. if (target->ml.convert_type == LKM_IVMODE) {
  259. mlog(ML_ERROR, "%s: res %.*s converting lock to invalid mode\n",
  260. dlm->name, res->lockname.len, res->lockname.name);
  261. BUG();
  262. }
  263. head = &res->granted;
  264. list_for_each(iter, head) {
  265. lock = list_entry(iter, struct dlm_lock, list);
  266. if (lock==target)
  267. continue;
  268. if (!dlm_lock_compatible(lock->ml.type,
  269. target->ml.convert_type)) {
  270. can_grant = 0;
  271. /* queue the BAST if not already */
  272. if (lock->ml.highest_blocked == LKM_IVMODE) {
  273. __dlm_lockres_reserve_ast(res);
  274. __dlm_queue_bast(dlm, lock);
  275. }
  276. /* update the highest_blocked if needed */
  277. if (lock->ml.highest_blocked < target->ml.convert_type)
  278. lock->ml.highest_blocked =
  279. target->ml.convert_type;
  280. }
  281. }
  282. head = &res->converting;
  283. list_for_each(iter, head) {
  284. lock = list_entry(iter, struct dlm_lock, list);
  285. if (lock==target)
  286. continue;
  287. if (!dlm_lock_compatible(lock->ml.type,
  288. target->ml.convert_type)) {
  289. can_grant = 0;
  290. if (lock->ml.highest_blocked == LKM_IVMODE) {
  291. __dlm_lockres_reserve_ast(res);
  292. __dlm_queue_bast(dlm, lock);
  293. }
  294. if (lock->ml.highest_blocked < target->ml.convert_type)
  295. lock->ml.highest_blocked =
  296. target->ml.convert_type;
  297. }
  298. }
  299. /* we can convert the lock */
  300. if (can_grant) {
  301. spin_lock(&target->spinlock);
  302. BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
  303. mlog(0, "%s: res %.*s, AST for Converting lock %u:%llu, type "
  304. "%d => %d, node %u\n", dlm->name, res->lockname.len,
  305. res->lockname.name,
  306. dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
  307. dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
  308. target->ml.type,
  309. target->ml.convert_type, target->ml.node);
  310. target->ml.type = target->ml.convert_type;
  311. target->ml.convert_type = LKM_IVMODE;
  312. list_move_tail(&target->list, &res->granted);
  313. BUG_ON(!target->lksb);
  314. target->lksb->status = DLM_NORMAL;
  315. spin_unlock(&target->spinlock);
  316. __dlm_lockres_reserve_ast(res);
  317. __dlm_queue_ast(dlm, target);
  318. /* go back and check for more */
  319. goto converting;
  320. }
  321. blocked:
  322. if (list_empty(&res->blocked))
  323. goto leave;
  324. target = list_entry(res->blocked.next, struct dlm_lock, list);
  325. head = &res->granted;
  326. list_for_each(iter, head) {
  327. lock = list_entry(iter, struct dlm_lock, list);
  328. if (lock==target)
  329. continue;
  330. if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
  331. can_grant = 0;
  332. if (lock->ml.highest_blocked == LKM_IVMODE) {
  333. __dlm_lockres_reserve_ast(res);
  334. __dlm_queue_bast(dlm, lock);
  335. }
  336. if (lock->ml.highest_blocked < target->ml.type)
  337. lock->ml.highest_blocked = target->ml.type;
  338. }
  339. }
  340. head = &res->converting;
  341. list_for_each(iter, head) {
  342. lock = list_entry(iter, struct dlm_lock, list);
  343. if (lock==target)
  344. continue;
  345. if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
  346. can_grant = 0;
  347. if (lock->ml.highest_blocked == LKM_IVMODE) {
  348. __dlm_lockres_reserve_ast(res);
  349. __dlm_queue_bast(dlm, lock);
  350. }
  351. if (lock->ml.highest_blocked < target->ml.type)
  352. lock->ml.highest_blocked = target->ml.type;
  353. }
  354. }
  355. /* we can grant the blocked lock (only
  356. * possible if converting list empty) */
  357. if (can_grant) {
  358. spin_lock(&target->spinlock);
  359. BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
  360. mlog(0, "%s: res %.*s, AST for Blocked lock %u:%llu, type %d, "
  361. "node %u\n", dlm->name, res->lockname.len,
  362. res->lockname.name,
  363. dlm_get_lock_cookie_node(be64_to_cpu(target->ml.cookie)),
  364. dlm_get_lock_cookie_seq(be64_to_cpu(target->ml.cookie)),
  365. target->ml.type, target->ml.node);
  366. /* target->ml.type is already correct */
  367. list_move_tail(&target->list, &res->granted);
  368. BUG_ON(!target->lksb);
  369. target->lksb->status = DLM_NORMAL;
  370. spin_unlock(&target->spinlock);
  371. __dlm_lockres_reserve_ast(res);
  372. __dlm_queue_ast(dlm, target);
  373. /* go back and check for more */
  374. goto converting;
  375. }
  376. leave:
  377. return;
  378. }
  379. /* must have NO locks when calling this with res !=NULL * */
  380. void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
  381. {
  382. if (res) {
  383. spin_lock(&dlm->spinlock);
  384. spin_lock(&res->spinlock);
  385. __dlm_dirty_lockres(dlm, res);
  386. spin_unlock(&res->spinlock);
  387. spin_unlock(&dlm->spinlock);
  388. }
  389. wake_up(&dlm->dlm_thread_wq);
  390. }
  391. void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
  392. {
  393. assert_spin_locked(&dlm->spinlock);
  394. assert_spin_locked(&res->spinlock);
  395. /* don't shuffle secondary queues */
  396. if ((res->owner == dlm->node_num)) {
  397. if (res->state & (DLM_LOCK_RES_MIGRATING |
  398. DLM_LOCK_RES_BLOCK_DIRTY))
  399. return;
  400. if (list_empty(&res->dirty)) {
  401. /* ref for dirty_list */
  402. dlm_lockres_get(res);
  403. list_add_tail(&res->dirty, &dlm->dirty_list);
  404. res->state |= DLM_LOCK_RES_DIRTY;
  405. }
  406. }
  407. mlog(0, "%s: res %.*s\n", dlm->name, res->lockname.len,
  408. res->lockname.name);
  409. }
  410. /* Launch the NM thread for the mounted volume */
  411. int dlm_launch_thread(struct dlm_ctxt *dlm)
  412. {
  413. mlog(0, "Starting dlm_thread...\n");
  414. dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
  415. if (IS_ERR(dlm->dlm_thread_task)) {
  416. mlog_errno(PTR_ERR(dlm->dlm_thread_task));
  417. dlm->dlm_thread_task = NULL;
  418. return -EINVAL;
  419. }
  420. return 0;
  421. }
  422. void dlm_complete_thread(struct dlm_ctxt *dlm)
  423. {
  424. if (dlm->dlm_thread_task) {
  425. mlog(ML_KTHREAD, "Waiting for dlm thread to exit\n");
  426. kthread_stop(dlm->dlm_thread_task);
  427. dlm->dlm_thread_task = NULL;
  428. }
  429. }
  430. static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
  431. {
  432. int empty;
  433. spin_lock(&dlm->spinlock);
  434. empty = list_empty(&dlm->dirty_list);
  435. spin_unlock(&dlm->spinlock);
  436. return empty;
  437. }
  438. static void dlm_flush_asts(struct dlm_ctxt *dlm)
  439. {
  440. int ret;
  441. struct dlm_lock *lock;
  442. struct dlm_lock_resource *res;
  443. u8 hi;
  444. spin_lock(&dlm->ast_lock);
  445. while (!list_empty(&dlm->pending_asts)) {
  446. lock = list_entry(dlm->pending_asts.next,
  447. struct dlm_lock, ast_list);
  448. /* get an extra ref on lock */
  449. dlm_lock_get(lock);
  450. res = lock->lockres;
  451. mlog(0, "%s: res %.*s, Flush AST for lock %u:%llu, type %d, "
  452. "node %u\n", dlm->name, res->lockname.len,
  453. res->lockname.name,
  454. dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
  455. dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
  456. lock->ml.type, lock->ml.node);
  457. BUG_ON(!lock->ast_pending);
  458. /* remove from list (including ref) */
  459. list_del_init(&lock->ast_list);
  460. dlm_lock_put(lock);
  461. spin_unlock(&dlm->ast_lock);
  462. if (lock->ml.node != dlm->node_num) {
  463. ret = dlm_do_remote_ast(dlm, res, lock);
  464. if (ret < 0)
  465. mlog_errno(ret);
  466. } else
  467. dlm_do_local_ast(dlm, res, lock);
  468. spin_lock(&dlm->ast_lock);
  469. /* possible that another ast was queued while
  470. * we were delivering the last one */
  471. if (!list_empty(&lock->ast_list)) {
  472. mlog(0, "%s: res %.*s, AST queued while flushing last "
  473. "one\n", dlm->name, res->lockname.len,
  474. res->lockname.name);
  475. } else
  476. lock->ast_pending = 0;
  477. /* drop the extra ref.
  478. * this may drop it completely. */
  479. dlm_lock_put(lock);
  480. dlm_lockres_release_ast(dlm, res);
  481. }
  482. while (!list_empty(&dlm->pending_basts)) {
  483. lock = list_entry(dlm->pending_basts.next,
  484. struct dlm_lock, bast_list);
  485. /* get an extra ref on lock */
  486. dlm_lock_get(lock);
  487. res = lock->lockres;
  488. BUG_ON(!lock->bast_pending);
  489. /* get the highest blocked lock, and reset */
  490. spin_lock(&lock->spinlock);
  491. BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
  492. hi = lock->ml.highest_blocked;
  493. lock->ml.highest_blocked = LKM_IVMODE;
  494. spin_unlock(&lock->spinlock);
  495. /* remove from list (including ref) */
  496. list_del_init(&lock->bast_list);
  497. dlm_lock_put(lock);
  498. spin_unlock(&dlm->ast_lock);
  499. mlog(0, "%s: res %.*s, Flush BAST for lock %u:%llu, "
  500. "blocked %d, node %u\n",
  501. dlm->name, res->lockname.len, res->lockname.name,
  502. dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
  503. dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
  504. hi, lock->ml.node);
  505. if (lock->ml.node != dlm->node_num) {
  506. ret = dlm_send_proxy_bast(dlm, res, lock, hi);
  507. if (ret < 0)
  508. mlog_errno(ret);
  509. } else
  510. dlm_do_local_bast(dlm, res, lock, hi);
  511. spin_lock(&dlm->ast_lock);
  512. /* possible that another bast was queued while
  513. * we were delivering the last one */
  514. if (!list_empty(&lock->bast_list)) {
  515. mlog(0, "%s: res %.*s, BAST queued while flushing last "
  516. "one\n", dlm->name, res->lockname.len,
  517. res->lockname.name);
  518. } else
  519. lock->bast_pending = 0;
  520. /* drop the extra ref.
  521. * this may drop it completely. */
  522. dlm_lock_put(lock);
  523. dlm_lockres_release_ast(dlm, res);
  524. }
  525. wake_up(&dlm->ast_wq);
  526. spin_unlock(&dlm->ast_lock);
  527. }
  528. #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
  529. #define DLM_THREAD_MAX_DIRTY 100
  530. #define DLM_THREAD_MAX_ASTS 10
  531. static int dlm_thread(void *data)
  532. {
  533. struct dlm_lock_resource *res;
  534. struct dlm_ctxt *dlm = data;
  535. unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
  536. mlog(0, "dlm thread running for %s...\n", dlm->name);
  537. while (!kthread_should_stop()) {
  538. int n = DLM_THREAD_MAX_DIRTY;
  539. /* dlm_shutting_down is very point-in-time, but that
  540. * doesn't matter as we'll just loop back around if we
  541. * get false on the leading edge of a state
  542. * transition. */
  543. dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
  544. /* We really don't want to hold dlm->spinlock while
  545. * calling dlm_shuffle_lists on each lockres that
  546. * needs to have its queues adjusted and AST/BASTs
  547. * run. So let's pull each entry off the dirty_list
  548. * and drop dlm->spinlock ASAP. Once off the list,
  549. * res->spinlock needs to be taken again to protect
  550. * the queues while calling dlm_shuffle_lists. */
  551. spin_lock(&dlm->spinlock);
  552. while (!list_empty(&dlm->dirty_list)) {
  553. int delay = 0;
  554. res = list_entry(dlm->dirty_list.next,
  555. struct dlm_lock_resource, dirty);
  556. /* peel a lockres off, remove it from the list,
  557. * unset the dirty flag and drop the dlm lock */
  558. BUG_ON(!res);
  559. dlm_lockres_get(res);
  560. spin_lock(&res->spinlock);
  561. /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
  562. list_del_init(&res->dirty);
  563. spin_unlock(&res->spinlock);
  564. spin_unlock(&dlm->spinlock);
  565. /* Drop dirty_list ref */
  566. dlm_lockres_put(res);
  567. /* lockres can be re-dirtied/re-added to the
  568. * dirty_list in this gap, but that is ok */
  569. spin_lock(&dlm->ast_lock);
  570. spin_lock(&res->spinlock);
  571. if (res->owner != dlm->node_num) {
  572. __dlm_print_one_lock_resource(res);
  573. mlog(ML_ERROR, "%s: inprog %d, mig %d, reco %d,"
  574. " dirty %d\n", dlm->name,
  575. !!(res->state & DLM_LOCK_RES_IN_PROGRESS),
  576. !!(res->state & DLM_LOCK_RES_MIGRATING),
  577. !!(res->state & DLM_LOCK_RES_RECOVERING),
  578. !!(res->state & DLM_LOCK_RES_DIRTY));
  579. }
  580. BUG_ON(res->owner != dlm->node_num);
  581. /* it is now ok to move lockreses in these states
  582. * to the dirty list, assuming that they will only be
  583. * dirty for a short while. */
  584. BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
  585. if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
  586. DLM_LOCK_RES_RECOVERING)) {
  587. /* move it to the tail and keep going */
  588. res->state &= ~DLM_LOCK_RES_DIRTY;
  589. spin_unlock(&res->spinlock);
  590. spin_unlock(&dlm->ast_lock);
  591. mlog(0, "%s: res %.*s, inprogress, delay list "
  592. "shuffle, state %d\n", dlm->name,
  593. res->lockname.len, res->lockname.name,
  594. res->state);
  595. delay = 1;
  596. goto in_progress;
  597. }
  598. /* at this point the lockres is not migrating/
  599. * recovering/in-progress. we have the lockres
  600. * spinlock and do NOT have the dlm lock.
  601. * safe to reserve/queue asts and run the lists. */
  602. /* called while holding lockres lock */
  603. dlm_shuffle_lists(dlm, res);
  604. res->state &= ~DLM_LOCK_RES_DIRTY;
  605. spin_unlock(&res->spinlock);
  606. spin_unlock(&dlm->ast_lock);
  607. dlm_lockres_calc_usage(dlm, res);
  608. in_progress:
  609. spin_lock(&dlm->spinlock);
  610. /* if the lock was in-progress, stick
  611. * it on the back of the list */
  612. if (delay) {
  613. spin_lock(&res->spinlock);
  614. __dlm_dirty_lockres(dlm, res);
  615. spin_unlock(&res->spinlock);
  616. }
  617. dlm_lockres_put(res);
  618. /* unlikely, but we may need to give time to
  619. * other tasks */
  620. if (!--n) {
  621. mlog(0, "%s: Throttling dlm thread\n",
  622. dlm->name);
  623. break;
  624. }
  625. }
  626. spin_unlock(&dlm->spinlock);
  627. dlm_flush_asts(dlm);
  628. /* yield and continue right away if there is more work to do */
  629. if (!n) {
  630. cond_resched();
  631. continue;
  632. }
  633. wait_event_interruptible_timeout(dlm->dlm_thread_wq,
  634. !dlm_dirty_list_empty(dlm) ||
  635. kthread_should_stop(),
  636. timeout);
  637. }
  638. mlog(0, "quitting DLM thread\n");
  639. return 0;
  640. }