dlmthread.c 21 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756
  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/slab.h>
  30. #include <linux/highmem.h>
  31. #include <linux/utsname.h>
  32. #include <linux/init.h>
  33. #include <linux/sysctl.h>
  34. #include <linux/random.h>
  35. #include <linux/blkdev.h>
  36. #include <linux/socket.h>
  37. #include <linux/inet.h>
  38. #include <linux/timer.h>
  39. #include <linux/kthread.h>
  40. #include <linux/delay.h>
  41. #include "cluster/heartbeat.h"
  42. #include "cluster/nodemanager.h"
  43. #include "cluster/tcp.h"
  44. #include "dlmapi.h"
  45. #include "dlmcommon.h"
  46. #include "dlmdomain.h"
  47. #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
  48. #include "cluster/masklog.h"
  49. static int dlm_thread(void *data);
  50. static void dlm_flush_asts(struct dlm_ctxt *dlm);
  51. #define dlm_lock_is_remote(dlm, lock) ((lock)->ml.node != (dlm)->node_num)
  52. /* will exit holding res->spinlock, but may drop in function */
  53. /* waits until flags are cleared on res->state */
  54. void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
  55. {
  56. DECLARE_WAITQUEUE(wait, current);
  57. assert_spin_locked(&res->spinlock);
  58. add_wait_queue(&res->wq, &wait);
  59. repeat:
  60. set_current_state(TASK_UNINTERRUPTIBLE);
  61. if (res->state & flags) {
  62. spin_unlock(&res->spinlock);
  63. schedule();
  64. spin_lock(&res->spinlock);
  65. goto repeat;
  66. }
  67. remove_wait_queue(&res->wq, &wait);
  68. __set_current_state(TASK_RUNNING);
  69. }
  70. int __dlm_lockres_has_locks(struct dlm_lock_resource *res)
  71. {
  72. if (list_empty(&res->granted) &&
  73. list_empty(&res->converting) &&
  74. list_empty(&res->blocked))
  75. return 0;
  76. return 1;
  77. }
  78. /* "unused": the lockres has no locks, is not on the dirty list,
  79. * has no inflight locks (in the gap between mastery and acquiring
  80. * the first lock), and has no bits in its refmap.
  81. * truly ready to be freed. */
  82. int __dlm_lockres_unused(struct dlm_lock_resource *res)
  83. {
  84. if (!__dlm_lockres_has_locks(res) &&
  85. (list_empty(&res->dirty) && !(res->state & DLM_LOCK_RES_DIRTY))) {
  86. /* try not to scan the bitmap unless the first two
  87. * conditions are already true */
  88. int bit = find_next_bit(res->refmap, O2NM_MAX_NODES, 0);
  89. if (bit >= O2NM_MAX_NODES) {
  90. /* since the bit for dlm->node_num is not
  91. * set, inflight_locks better be zero */
  92. BUG_ON(res->inflight_locks != 0);
  93. return 1;
  94. }
  95. }
  96. return 0;
  97. }
  98. /* Call whenever you may have added or deleted something from one of
  99. * the lockres queue's. This will figure out whether it belongs on the
  100. * unused list or not and does the appropriate thing. */
  101. void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
  102. struct dlm_lock_resource *res)
  103. {
  104. mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
  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, "putting lockres %.*s:%p onto purge list\n",
  110. res->lockname.len, res->lockname.name, res);
  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, "removing lockres %.*s:%p from purge list, owner=%u\n",
  118. res->lockname.len, res->lockname.name, res, res->owner);
  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. mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
  128. spin_lock(&dlm->spinlock);
  129. spin_lock(&res->spinlock);
  130. __dlm_lockres_calc_usage(dlm, res);
  131. spin_unlock(&res->spinlock);
  132. spin_unlock(&dlm->spinlock);
  133. }
  134. static int dlm_purge_lockres(struct dlm_ctxt *dlm,
  135. struct dlm_lock_resource *res)
  136. {
  137. int master;
  138. int ret = 0;
  139. spin_lock(&res->spinlock);
  140. if (!__dlm_lockres_unused(res)) {
  141. mlog(0, "%s:%.*s: tried to purge but not unused\n",
  142. dlm->name, res->lockname.len, res->lockname.name);
  143. __dlm_print_one_lock_resource(res);
  144. spin_unlock(&res->spinlock);
  145. BUG();
  146. }
  147. if (res->state & DLM_LOCK_RES_MIGRATING) {
  148. mlog(0, "%s:%.*s: Delay dropref as this lockres is "
  149. "being remastered\n", dlm->name, res->lockname.len,
  150. res->lockname.name);
  151. /* Re-add the lockres to the end of the purge list */
  152. if (!list_empty(&res->purge)) {
  153. list_del_init(&res->purge);
  154. list_add_tail(&res->purge, &dlm->purge_list);
  155. }
  156. spin_unlock(&res->spinlock);
  157. return 0;
  158. }
  159. master = (res->owner == dlm->node_num);
  160. if (!master)
  161. res->state |= DLM_LOCK_RES_DROPPING_REF;
  162. spin_unlock(&res->spinlock);
  163. mlog(0, "purging lockres %.*s, master = %d\n", res->lockname.len,
  164. res->lockname.name, master);
  165. if (!master) {
  166. /* drop spinlock... retake below */
  167. spin_unlock(&dlm->spinlock);
  168. spin_lock(&res->spinlock);
  169. /* This ensures that clear refmap is sent after the set */
  170. __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
  171. spin_unlock(&res->spinlock);
  172. /* clear our bit from the master's refmap, ignore errors */
  173. ret = dlm_drop_lockres_ref(dlm, res);
  174. if (ret < 0) {
  175. mlog_errno(ret);
  176. if (!dlm_is_host_down(ret))
  177. BUG();
  178. }
  179. mlog(0, "%s:%.*s: dlm_deref_lockres returned %d\n",
  180. dlm->name, res->lockname.len, res->lockname.name, ret);
  181. spin_lock(&dlm->spinlock);
  182. }
  183. if (!list_empty(&res->purge)) {
  184. mlog(0, "removing lockres %.*s:%p from purgelist, "
  185. "master = %d\n", res->lockname.len, res->lockname.name,
  186. res, master);
  187. list_del_init(&res->purge);
  188. dlm_lockres_put(res);
  189. dlm->purge_count--;
  190. }
  191. __dlm_unhash_lockres(res);
  192. /* lockres is not in the hash now. drop the flag and wake up
  193. * any processes waiting in dlm_get_lock_resource. */
  194. if (!master) {
  195. spin_lock(&res->spinlock);
  196. res->state &= ~DLM_LOCK_RES_DROPPING_REF;
  197. spin_unlock(&res->spinlock);
  198. wake_up(&res->wq);
  199. }
  200. return 0;
  201. }
  202. static void dlm_run_purge_list(struct dlm_ctxt *dlm,
  203. int purge_now)
  204. {
  205. unsigned int run_max, unused;
  206. unsigned long purge_jiffies;
  207. struct dlm_lock_resource *lockres;
  208. spin_lock(&dlm->spinlock);
  209. run_max = dlm->purge_count;
  210. while(run_max && !list_empty(&dlm->purge_list)) {
  211. run_max--;
  212. lockres = list_entry(dlm->purge_list.next,
  213. struct dlm_lock_resource, purge);
  214. /* Status of the lockres *might* change so double
  215. * check. If the lockres is unused, holding the dlm
  216. * spinlock will prevent people from getting and more
  217. * refs on it -- there's no need to keep the lockres
  218. * spinlock. */
  219. spin_lock(&lockres->spinlock);
  220. unused = __dlm_lockres_unused(lockres);
  221. spin_unlock(&lockres->spinlock);
  222. if (!unused)
  223. continue;
  224. purge_jiffies = lockres->last_used +
  225. msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
  226. /* Make sure that we want to be processing this guy at
  227. * this time. */
  228. if (!purge_now && time_after(purge_jiffies, jiffies)) {
  229. /* Since resources are added to the purge list
  230. * in tail order, we can stop at the first
  231. * unpurgable resource -- anyone added after
  232. * him will have a greater last_used value */
  233. break;
  234. }
  235. dlm_lockres_get(lockres);
  236. /* This may drop and reacquire the dlm spinlock if it
  237. * has to do migration. */
  238. if (dlm_purge_lockres(dlm, lockres))
  239. BUG();
  240. dlm_lockres_put(lockres);
  241. /* Avoid adding any scheduling latencies */
  242. cond_resched_lock(&dlm->spinlock);
  243. }
  244. spin_unlock(&dlm->spinlock);
  245. }
  246. static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
  247. struct dlm_lock_resource *res)
  248. {
  249. struct dlm_lock *lock, *target;
  250. struct list_head *iter;
  251. struct list_head *head;
  252. int can_grant = 1;
  253. //mlog(0, "res->lockname.len=%d\n", res->lockname.len);
  254. //mlog(0, "res->lockname.name=%p\n", res->lockname.name);
  255. //mlog(0, "shuffle res %.*s\n", res->lockname.len,
  256. // res->lockname.name);
  257. /* because this function is called with the lockres
  258. * spinlock, and because we know that it is not migrating/
  259. * recovering/in-progress, it is fine to reserve asts and
  260. * basts right before queueing them all throughout */
  261. assert_spin_locked(&res->spinlock);
  262. BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
  263. DLM_LOCK_RES_RECOVERING|
  264. DLM_LOCK_RES_IN_PROGRESS)));
  265. converting:
  266. if (list_empty(&res->converting))
  267. goto blocked;
  268. mlog(0, "res %.*s has locks on a convert queue\n", res->lockname.len,
  269. res->lockname.name);
  270. target = list_entry(res->converting.next, struct dlm_lock, list);
  271. if (target->ml.convert_type == LKM_IVMODE) {
  272. mlog(ML_ERROR, "%.*s: converting a lock with no "
  273. "convert_type!\n", res->lockname.len, res->lockname.name);
  274. BUG();
  275. }
  276. head = &res->granted;
  277. list_for_each(iter, head) {
  278. lock = list_entry(iter, struct dlm_lock, list);
  279. if (lock==target)
  280. continue;
  281. if (!dlm_lock_compatible(lock->ml.type,
  282. target->ml.convert_type)) {
  283. can_grant = 0;
  284. /* queue the BAST if not already */
  285. if (lock->ml.highest_blocked == LKM_IVMODE) {
  286. __dlm_lockres_reserve_ast(res);
  287. dlm_queue_bast(dlm, lock);
  288. }
  289. /* update the highest_blocked if needed */
  290. if (lock->ml.highest_blocked < target->ml.convert_type)
  291. lock->ml.highest_blocked =
  292. target->ml.convert_type;
  293. }
  294. }
  295. head = &res->converting;
  296. list_for_each(iter, head) {
  297. lock = list_entry(iter, struct dlm_lock, list);
  298. if (lock==target)
  299. continue;
  300. if (!dlm_lock_compatible(lock->ml.type,
  301. target->ml.convert_type)) {
  302. can_grant = 0;
  303. if (lock->ml.highest_blocked == LKM_IVMODE) {
  304. __dlm_lockres_reserve_ast(res);
  305. dlm_queue_bast(dlm, lock);
  306. }
  307. if (lock->ml.highest_blocked < target->ml.convert_type)
  308. lock->ml.highest_blocked =
  309. target->ml.convert_type;
  310. }
  311. }
  312. /* we can convert the lock */
  313. if (can_grant) {
  314. spin_lock(&target->spinlock);
  315. BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
  316. mlog(0, "calling ast for converting lock: %.*s, have: %d, "
  317. "granting: %d, node: %u\n", res->lockname.len,
  318. res->lockname.name, target->ml.type,
  319. target->ml.convert_type, target->ml.node);
  320. target->ml.type = target->ml.convert_type;
  321. target->ml.convert_type = LKM_IVMODE;
  322. list_move_tail(&target->list, &res->granted);
  323. BUG_ON(!target->lksb);
  324. target->lksb->status = DLM_NORMAL;
  325. spin_unlock(&target->spinlock);
  326. __dlm_lockres_reserve_ast(res);
  327. dlm_queue_ast(dlm, target);
  328. /* go back and check for more */
  329. goto converting;
  330. }
  331. blocked:
  332. if (list_empty(&res->blocked))
  333. goto leave;
  334. target = list_entry(res->blocked.next, struct dlm_lock, list);
  335. head = &res->granted;
  336. list_for_each(iter, head) {
  337. lock = list_entry(iter, struct dlm_lock, list);
  338. if (lock==target)
  339. continue;
  340. if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
  341. can_grant = 0;
  342. if (lock->ml.highest_blocked == LKM_IVMODE) {
  343. __dlm_lockres_reserve_ast(res);
  344. dlm_queue_bast(dlm, lock);
  345. }
  346. if (lock->ml.highest_blocked < target->ml.type)
  347. lock->ml.highest_blocked = target->ml.type;
  348. }
  349. }
  350. head = &res->converting;
  351. list_for_each(iter, head) {
  352. lock = list_entry(iter, struct dlm_lock, list);
  353. if (lock==target)
  354. continue;
  355. if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
  356. can_grant = 0;
  357. if (lock->ml.highest_blocked == LKM_IVMODE) {
  358. __dlm_lockres_reserve_ast(res);
  359. dlm_queue_bast(dlm, lock);
  360. }
  361. if (lock->ml.highest_blocked < target->ml.type)
  362. lock->ml.highest_blocked = target->ml.type;
  363. }
  364. }
  365. /* we can grant the blocked lock (only
  366. * possible if converting list empty) */
  367. if (can_grant) {
  368. spin_lock(&target->spinlock);
  369. BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
  370. mlog(0, "calling ast for blocked lock: %.*s, granting: %d, "
  371. "node: %u\n", res->lockname.len, res->lockname.name,
  372. target->ml.type, target->ml.node);
  373. // target->ml.type is already correct
  374. list_move_tail(&target->list, &res->granted);
  375. BUG_ON(!target->lksb);
  376. target->lksb->status = DLM_NORMAL;
  377. spin_unlock(&target->spinlock);
  378. __dlm_lockres_reserve_ast(res);
  379. dlm_queue_ast(dlm, target);
  380. /* go back and check for more */
  381. goto converting;
  382. }
  383. leave:
  384. return;
  385. }
  386. /* must have NO locks when calling this with res !=NULL * */
  387. void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
  388. {
  389. mlog_entry("dlm=%p, res=%p\n", dlm, res);
  390. if (res) {
  391. spin_lock(&dlm->spinlock);
  392. spin_lock(&res->spinlock);
  393. __dlm_dirty_lockres(dlm, res);
  394. spin_unlock(&res->spinlock);
  395. spin_unlock(&dlm->spinlock);
  396. }
  397. wake_up(&dlm->dlm_thread_wq);
  398. }
  399. void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
  400. {
  401. mlog_entry("dlm=%p, res=%p\n", dlm, res);
  402. assert_spin_locked(&dlm->spinlock);
  403. assert_spin_locked(&res->spinlock);
  404. /* don't shuffle secondary queues */
  405. if ((res->owner == dlm->node_num)) {
  406. if (res->state & (DLM_LOCK_RES_MIGRATING |
  407. DLM_LOCK_RES_BLOCK_DIRTY))
  408. return;
  409. if (list_empty(&res->dirty)) {
  410. /* ref for dirty_list */
  411. dlm_lockres_get(res);
  412. list_add_tail(&res->dirty, &dlm->dirty_list);
  413. res->state |= DLM_LOCK_RES_DIRTY;
  414. }
  415. }
  416. }
  417. /* Launch the NM thread for the mounted volume */
  418. int dlm_launch_thread(struct dlm_ctxt *dlm)
  419. {
  420. mlog(0, "starting dlm thread...\n");
  421. dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
  422. if (IS_ERR(dlm->dlm_thread_task)) {
  423. mlog_errno(PTR_ERR(dlm->dlm_thread_task));
  424. dlm->dlm_thread_task = NULL;
  425. return -EINVAL;
  426. }
  427. return 0;
  428. }
  429. void dlm_complete_thread(struct dlm_ctxt *dlm)
  430. {
  431. if (dlm->dlm_thread_task) {
  432. mlog(ML_KTHREAD, "waiting for dlm thread to exit\n");
  433. kthread_stop(dlm->dlm_thread_task);
  434. dlm->dlm_thread_task = NULL;
  435. }
  436. }
  437. static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
  438. {
  439. int empty;
  440. spin_lock(&dlm->spinlock);
  441. empty = list_empty(&dlm->dirty_list);
  442. spin_unlock(&dlm->spinlock);
  443. return empty;
  444. }
  445. static void dlm_flush_asts(struct dlm_ctxt *dlm)
  446. {
  447. int ret;
  448. struct dlm_lock *lock;
  449. struct dlm_lock_resource *res;
  450. u8 hi;
  451. spin_lock(&dlm->ast_lock);
  452. while (!list_empty(&dlm->pending_asts)) {
  453. lock = list_entry(dlm->pending_asts.next,
  454. struct dlm_lock, ast_list);
  455. /* get an extra ref on lock */
  456. dlm_lock_get(lock);
  457. res = lock->lockres;
  458. mlog(0, "delivering an ast for this lockres\n");
  459. BUG_ON(!lock->ast_pending);
  460. /* remove from list (including ref) */
  461. list_del_init(&lock->ast_list);
  462. dlm_lock_put(lock);
  463. spin_unlock(&dlm->ast_lock);
  464. if (lock->ml.node != dlm->node_num) {
  465. ret = dlm_do_remote_ast(dlm, res, lock);
  466. if (ret < 0)
  467. mlog_errno(ret);
  468. } else
  469. dlm_do_local_ast(dlm, res, lock);
  470. spin_lock(&dlm->ast_lock);
  471. /* possible that another ast was queued while
  472. * we were delivering the last one */
  473. if (!list_empty(&lock->ast_list)) {
  474. mlog(0, "aha another ast got queued while "
  475. "we were finishing the last one. will "
  476. "keep the ast_pending flag set.\n");
  477. } else
  478. lock->ast_pending = 0;
  479. /* drop the extra ref.
  480. * this may drop it completely. */
  481. dlm_lock_put(lock);
  482. dlm_lockres_release_ast(dlm, res);
  483. }
  484. while (!list_empty(&dlm->pending_basts)) {
  485. lock = list_entry(dlm->pending_basts.next,
  486. struct dlm_lock, bast_list);
  487. /* get an extra ref on lock */
  488. dlm_lock_get(lock);
  489. res = lock->lockres;
  490. BUG_ON(!lock->bast_pending);
  491. /* get the highest blocked lock, and reset */
  492. spin_lock(&lock->spinlock);
  493. BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
  494. hi = lock->ml.highest_blocked;
  495. lock->ml.highest_blocked = LKM_IVMODE;
  496. spin_unlock(&lock->spinlock);
  497. /* remove from list (including ref) */
  498. list_del_init(&lock->bast_list);
  499. dlm_lock_put(lock);
  500. spin_unlock(&dlm->ast_lock);
  501. mlog(0, "delivering a bast for this lockres "
  502. "(blocked = %d\n", hi);
  503. if (lock->ml.node != dlm->node_num) {
  504. ret = dlm_send_proxy_bast(dlm, res, lock, hi);
  505. if (ret < 0)
  506. mlog_errno(ret);
  507. } else
  508. dlm_do_local_bast(dlm, res, lock, hi);
  509. spin_lock(&dlm->ast_lock);
  510. /* possible that another bast was queued while
  511. * we were delivering the last one */
  512. if (!list_empty(&lock->bast_list)) {
  513. mlog(0, "aha another bast got queued while "
  514. "we were finishing the last one. will "
  515. "keep the bast_pending flag set.\n");
  516. } else
  517. lock->bast_pending = 0;
  518. /* drop the extra ref.
  519. * this may drop it completely. */
  520. dlm_lock_put(lock);
  521. dlm_lockres_release_ast(dlm, res);
  522. }
  523. wake_up(&dlm->ast_wq);
  524. spin_unlock(&dlm->ast_lock);
  525. }
  526. #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
  527. #define DLM_THREAD_MAX_DIRTY 100
  528. #define DLM_THREAD_MAX_ASTS 10
  529. static int dlm_thread(void *data)
  530. {
  531. struct dlm_lock_resource *res;
  532. struct dlm_ctxt *dlm = data;
  533. unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
  534. mlog(0, "dlm thread running for %s...\n", dlm->name);
  535. while (!kthread_should_stop()) {
  536. int n = DLM_THREAD_MAX_DIRTY;
  537. /* dlm_shutting_down is very point-in-time, but that
  538. * doesn't matter as we'll just loop back around if we
  539. * get false on the leading edge of a state
  540. * transition. */
  541. dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
  542. /* We really don't want to hold dlm->spinlock while
  543. * calling dlm_shuffle_lists on each lockres that
  544. * needs to have its queues adjusted and AST/BASTs
  545. * run. So let's pull each entry off the dirty_list
  546. * and drop dlm->spinlock ASAP. Once off the list,
  547. * res->spinlock needs to be taken again to protect
  548. * the queues while calling dlm_shuffle_lists. */
  549. spin_lock(&dlm->spinlock);
  550. while (!list_empty(&dlm->dirty_list)) {
  551. int delay = 0;
  552. res = list_entry(dlm->dirty_list.next,
  553. struct dlm_lock_resource, dirty);
  554. /* peel a lockres off, remove it from the list,
  555. * unset the dirty flag and drop the dlm lock */
  556. BUG_ON(!res);
  557. dlm_lockres_get(res);
  558. spin_lock(&res->spinlock);
  559. /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
  560. list_del_init(&res->dirty);
  561. spin_unlock(&res->spinlock);
  562. spin_unlock(&dlm->spinlock);
  563. /* Drop dirty_list ref */
  564. dlm_lockres_put(res);
  565. /* lockres can be re-dirtied/re-added to the
  566. * dirty_list in this gap, but that is ok */
  567. spin_lock(&res->spinlock);
  568. if (res->owner != dlm->node_num) {
  569. __dlm_print_one_lock_resource(res);
  570. mlog(ML_ERROR, "inprog:%s, mig:%s, reco:%s, dirty:%s\n",
  571. res->state & DLM_LOCK_RES_IN_PROGRESS ? "yes" : "no",
  572. res->state & DLM_LOCK_RES_MIGRATING ? "yes" : "no",
  573. res->state & DLM_LOCK_RES_RECOVERING ? "yes" : "no",
  574. res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
  575. }
  576. BUG_ON(res->owner != dlm->node_num);
  577. /* it is now ok to move lockreses in these states
  578. * to the dirty list, assuming that they will only be
  579. * dirty for a short while. */
  580. BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
  581. if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
  582. DLM_LOCK_RES_RECOVERING)) {
  583. /* move it to the tail and keep going */
  584. res->state &= ~DLM_LOCK_RES_DIRTY;
  585. spin_unlock(&res->spinlock);
  586. mlog(0, "delaying list shuffling for in-"
  587. "progress lockres %.*s, state=%d\n",
  588. res->lockname.len, res->lockname.name,
  589. res->state);
  590. delay = 1;
  591. goto in_progress;
  592. }
  593. /* at this point the lockres is not migrating/
  594. * recovering/in-progress. we have the lockres
  595. * spinlock and do NOT have the dlm lock.
  596. * safe to reserve/queue asts and run the lists. */
  597. mlog(0, "calling dlm_shuffle_lists with dlm=%s, "
  598. "res=%.*s\n", dlm->name,
  599. res->lockname.len, res->lockname.name);
  600. /* called while holding lockres lock */
  601. dlm_shuffle_lists(dlm, res);
  602. res->state &= ~DLM_LOCK_RES_DIRTY;
  603. spin_unlock(&res->spinlock);
  604. dlm_lockres_calc_usage(dlm, res);
  605. in_progress:
  606. spin_lock(&dlm->spinlock);
  607. /* if the lock was in-progress, stick
  608. * it on the back of the list */
  609. if (delay) {
  610. spin_lock(&res->spinlock);
  611. __dlm_dirty_lockres(dlm, res);
  612. spin_unlock(&res->spinlock);
  613. }
  614. dlm_lockres_put(res);
  615. /* unlikely, but we may need to give time to
  616. * other tasks */
  617. if (!--n) {
  618. mlog(0, "throttling dlm_thread\n");
  619. break;
  620. }
  621. }
  622. spin_unlock(&dlm->spinlock);
  623. dlm_flush_asts(dlm);
  624. /* yield and continue right away if there is more work to do */
  625. if (!n) {
  626. cond_resched();
  627. continue;
  628. }
  629. wait_event_interruptible_timeout(dlm->dlm_thread_wq,
  630. !dlm_dirty_list_empty(dlm) ||
  631. kthread_should_stop(),
  632. timeout);
  633. }
  634. mlog(0, "quitting DLM thread\n");
  635. return 0;
  636. }