dlmthread.c 20 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741
  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. spin_unlock(&res->spinlock);
  142. mlog(0, "%s:%.*s: tried to purge but not unused\n",
  143. dlm->name, res->lockname.len, res->lockname.name);
  144. return -ENOTEMPTY;
  145. }
  146. master = (res->owner == dlm->node_num);
  147. if (!master)
  148. res->state |= DLM_LOCK_RES_DROPPING_REF;
  149. spin_unlock(&res->spinlock);
  150. mlog(0, "purging lockres %.*s, master = %d\n", res->lockname.len,
  151. res->lockname.name, master);
  152. if (!master) {
  153. /* drop spinlock... retake below */
  154. spin_unlock(&dlm->spinlock);
  155. spin_lock(&res->spinlock);
  156. /* This ensures that clear refmap is sent after the set */
  157. __dlm_wait_on_lockres_flags(res, (DLM_LOCK_RES_SETREF_INPROG |
  158. DLM_LOCK_RES_MIGRATING));
  159. spin_unlock(&res->spinlock);
  160. /* clear our bit from the master's refmap, ignore errors */
  161. ret = dlm_drop_lockres_ref(dlm, res);
  162. if (ret < 0) {
  163. mlog_errno(ret);
  164. if (!dlm_is_host_down(ret))
  165. BUG();
  166. }
  167. mlog(0, "%s:%.*s: dlm_deref_lockres returned %d\n",
  168. dlm->name, res->lockname.len, res->lockname.name, ret);
  169. spin_lock(&dlm->spinlock);
  170. }
  171. if (!list_empty(&res->purge)) {
  172. mlog(0, "removing lockres %.*s:%p from purgelist, "
  173. "master = %d\n", res->lockname.len, res->lockname.name,
  174. res, master);
  175. list_del_init(&res->purge);
  176. dlm_lockres_put(res);
  177. dlm->purge_count--;
  178. }
  179. __dlm_unhash_lockres(res);
  180. /* lockres is not in the hash now. drop the flag and wake up
  181. * any processes waiting in dlm_get_lock_resource. */
  182. if (!master) {
  183. spin_lock(&res->spinlock);
  184. res->state &= ~DLM_LOCK_RES_DROPPING_REF;
  185. spin_unlock(&res->spinlock);
  186. wake_up(&res->wq);
  187. }
  188. return 0;
  189. }
  190. static void dlm_run_purge_list(struct dlm_ctxt *dlm,
  191. int purge_now)
  192. {
  193. unsigned int run_max, unused;
  194. unsigned long purge_jiffies;
  195. struct dlm_lock_resource *lockres;
  196. spin_lock(&dlm->spinlock);
  197. run_max = dlm->purge_count;
  198. while(run_max && !list_empty(&dlm->purge_list)) {
  199. run_max--;
  200. lockres = list_entry(dlm->purge_list.next,
  201. struct dlm_lock_resource, purge);
  202. /* Status of the lockres *might* change so double
  203. * check. If the lockres is unused, holding the dlm
  204. * spinlock will prevent people from getting and more
  205. * refs on it -- there's no need to keep the lockres
  206. * spinlock. */
  207. spin_lock(&lockres->spinlock);
  208. unused = __dlm_lockres_unused(lockres);
  209. spin_unlock(&lockres->spinlock);
  210. if (!unused)
  211. continue;
  212. purge_jiffies = lockres->last_used +
  213. msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
  214. /* Make sure that we want to be processing this guy at
  215. * this time. */
  216. if (!purge_now && time_after(purge_jiffies, jiffies)) {
  217. /* Since resources are added to the purge list
  218. * in tail order, we can stop at the first
  219. * unpurgable resource -- anyone added after
  220. * him will have a greater last_used value */
  221. break;
  222. }
  223. dlm_lockres_get(lockres);
  224. /* This may drop and reacquire the dlm spinlock if it
  225. * has to do migration. */
  226. if (dlm_purge_lockres(dlm, lockres))
  227. BUG();
  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. //mlog(0, "res->lockname.len=%d\n", res->lockname.len);
  242. //mlog(0, "res->lockname.name=%p\n", res->lockname.name);
  243. //mlog(0, "shuffle res %.*s\n", res->lockname.len,
  244. // res->lockname.name);
  245. /* because this function is called with the lockres
  246. * spinlock, and because we know that it is not migrating/
  247. * recovering/in-progress, it is fine to reserve asts and
  248. * basts right before queueing them all throughout */
  249. assert_spin_locked(&res->spinlock);
  250. BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
  251. DLM_LOCK_RES_RECOVERING|
  252. DLM_LOCK_RES_IN_PROGRESS)));
  253. converting:
  254. if (list_empty(&res->converting))
  255. goto blocked;
  256. mlog(0, "res %.*s has locks on a convert queue\n", res->lockname.len,
  257. res->lockname.name);
  258. target = list_entry(res->converting.next, struct dlm_lock, list);
  259. if (target->ml.convert_type == LKM_IVMODE) {
  260. mlog(ML_ERROR, "%.*s: converting a lock with no "
  261. "convert_type!\n", res->lockname.len, res->lockname.name);
  262. BUG();
  263. }
  264. head = &res->granted;
  265. list_for_each(iter, head) {
  266. lock = list_entry(iter, struct dlm_lock, list);
  267. if (lock==target)
  268. continue;
  269. if (!dlm_lock_compatible(lock->ml.type,
  270. target->ml.convert_type)) {
  271. can_grant = 0;
  272. /* queue the BAST if not already */
  273. if (lock->ml.highest_blocked == LKM_IVMODE) {
  274. __dlm_lockres_reserve_ast(res);
  275. dlm_queue_bast(dlm, lock);
  276. }
  277. /* update the highest_blocked if needed */
  278. if (lock->ml.highest_blocked < target->ml.convert_type)
  279. lock->ml.highest_blocked =
  280. target->ml.convert_type;
  281. }
  282. }
  283. head = &res->converting;
  284. list_for_each(iter, head) {
  285. lock = list_entry(iter, struct dlm_lock, list);
  286. if (lock==target)
  287. continue;
  288. if (!dlm_lock_compatible(lock->ml.type,
  289. target->ml.convert_type)) {
  290. can_grant = 0;
  291. if (lock->ml.highest_blocked == LKM_IVMODE) {
  292. __dlm_lockres_reserve_ast(res);
  293. dlm_queue_bast(dlm, lock);
  294. }
  295. if (lock->ml.highest_blocked < target->ml.convert_type)
  296. lock->ml.highest_blocked =
  297. target->ml.convert_type;
  298. }
  299. }
  300. /* we can convert the lock */
  301. if (can_grant) {
  302. spin_lock(&target->spinlock);
  303. BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
  304. mlog(0, "calling ast for converting lock: %.*s, have: %d, "
  305. "granting: %d, node: %u\n", res->lockname.len,
  306. res->lockname.name, target->ml.type,
  307. target->ml.convert_type, target->ml.node);
  308. target->ml.type = target->ml.convert_type;
  309. target->ml.convert_type = LKM_IVMODE;
  310. list_move_tail(&target->list, &res->granted);
  311. BUG_ON(!target->lksb);
  312. target->lksb->status = DLM_NORMAL;
  313. spin_unlock(&target->spinlock);
  314. __dlm_lockres_reserve_ast(res);
  315. dlm_queue_ast(dlm, target);
  316. /* go back and check for more */
  317. goto converting;
  318. }
  319. blocked:
  320. if (list_empty(&res->blocked))
  321. goto leave;
  322. target = list_entry(res->blocked.next, struct dlm_lock, list);
  323. head = &res->granted;
  324. list_for_each(iter, head) {
  325. lock = list_entry(iter, struct dlm_lock, list);
  326. if (lock==target)
  327. continue;
  328. if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
  329. can_grant = 0;
  330. if (lock->ml.highest_blocked == LKM_IVMODE) {
  331. __dlm_lockres_reserve_ast(res);
  332. dlm_queue_bast(dlm, lock);
  333. }
  334. if (lock->ml.highest_blocked < target->ml.type)
  335. lock->ml.highest_blocked = target->ml.type;
  336. }
  337. }
  338. head = &res->converting;
  339. list_for_each(iter, head) {
  340. lock = list_entry(iter, struct dlm_lock, list);
  341. if (lock==target)
  342. continue;
  343. if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
  344. can_grant = 0;
  345. if (lock->ml.highest_blocked == LKM_IVMODE) {
  346. __dlm_lockres_reserve_ast(res);
  347. dlm_queue_bast(dlm, lock);
  348. }
  349. if (lock->ml.highest_blocked < target->ml.type)
  350. lock->ml.highest_blocked = target->ml.type;
  351. }
  352. }
  353. /* we can grant the blocked lock (only
  354. * possible if converting list empty) */
  355. if (can_grant) {
  356. spin_lock(&target->spinlock);
  357. BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
  358. mlog(0, "calling ast for blocked lock: %.*s, granting: %d, "
  359. "node: %u\n", res->lockname.len, res->lockname.name,
  360. target->ml.type, target->ml.node);
  361. // target->ml.type is already correct
  362. list_move_tail(&target->list, &res->granted);
  363. BUG_ON(!target->lksb);
  364. target->lksb->status = DLM_NORMAL;
  365. spin_unlock(&target->spinlock);
  366. __dlm_lockres_reserve_ast(res);
  367. dlm_queue_ast(dlm, target);
  368. /* go back and check for more */
  369. goto converting;
  370. }
  371. leave:
  372. return;
  373. }
  374. /* must have NO locks when calling this with res !=NULL * */
  375. void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
  376. {
  377. mlog_entry("dlm=%p, res=%p\n", dlm, res);
  378. if (res) {
  379. spin_lock(&dlm->spinlock);
  380. spin_lock(&res->spinlock);
  381. __dlm_dirty_lockres(dlm, res);
  382. spin_unlock(&res->spinlock);
  383. spin_unlock(&dlm->spinlock);
  384. }
  385. wake_up(&dlm->dlm_thread_wq);
  386. }
  387. void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
  388. {
  389. mlog_entry("dlm=%p, res=%p\n", dlm, res);
  390. assert_spin_locked(&dlm->spinlock);
  391. assert_spin_locked(&res->spinlock);
  392. /* don't shuffle secondary queues */
  393. if ((res->owner == dlm->node_num)) {
  394. if (res->state & (DLM_LOCK_RES_MIGRATING |
  395. DLM_LOCK_RES_BLOCK_DIRTY))
  396. return;
  397. if (list_empty(&res->dirty)) {
  398. /* ref for dirty_list */
  399. dlm_lockres_get(res);
  400. list_add_tail(&res->dirty, &dlm->dirty_list);
  401. res->state |= DLM_LOCK_RES_DIRTY;
  402. }
  403. }
  404. }
  405. /* Launch the NM thread for the mounted volume */
  406. int dlm_launch_thread(struct dlm_ctxt *dlm)
  407. {
  408. mlog(0, "starting dlm thread...\n");
  409. dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
  410. if (IS_ERR(dlm->dlm_thread_task)) {
  411. mlog_errno(PTR_ERR(dlm->dlm_thread_task));
  412. dlm->dlm_thread_task = NULL;
  413. return -EINVAL;
  414. }
  415. return 0;
  416. }
  417. void dlm_complete_thread(struct dlm_ctxt *dlm)
  418. {
  419. if (dlm->dlm_thread_task) {
  420. mlog(ML_KTHREAD, "waiting for dlm thread to exit\n");
  421. kthread_stop(dlm->dlm_thread_task);
  422. dlm->dlm_thread_task = NULL;
  423. }
  424. }
  425. static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
  426. {
  427. int empty;
  428. spin_lock(&dlm->spinlock);
  429. empty = list_empty(&dlm->dirty_list);
  430. spin_unlock(&dlm->spinlock);
  431. return empty;
  432. }
  433. static void dlm_flush_asts(struct dlm_ctxt *dlm)
  434. {
  435. int ret;
  436. struct dlm_lock *lock;
  437. struct dlm_lock_resource *res;
  438. u8 hi;
  439. spin_lock(&dlm->ast_lock);
  440. while (!list_empty(&dlm->pending_asts)) {
  441. lock = list_entry(dlm->pending_asts.next,
  442. struct dlm_lock, ast_list);
  443. /* get an extra ref on lock */
  444. dlm_lock_get(lock);
  445. res = lock->lockres;
  446. mlog(0, "delivering an ast for this lockres\n");
  447. BUG_ON(!lock->ast_pending);
  448. /* remove from list (including ref) */
  449. list_del_init(&lock->ast_list);
  450. dlm_lock_put(lock);
  451. spin_unlock(&dlm->ast_lock);
  452. if (lock->ml.node != dlm->node_num) {
  453. ret = dlm_do_remote_ast(dlm, res, lock);
  454. if (ret < 0)
  455. mlog_errno(ret);
  456. } else
  457. dlm_do_local_ast(dlm, res, lock);
  458. spin_lock(&dlm->ast_lock);
  459. /* possible that another ast was queued while
  460. * we were delivering the last one */
  461. if (!list_empty(&lock->ast_list)) {
  462. mlog(0, "aha another ast got queued while "
  463. "we were finishing the last one. will "
  464. "keep the ast_pending flag set.\n");
  465. } else
  466. lock->ast_pending = 0;
  467. /* drop the extra ref.
  468. * this may drop it completely. */
  469. dlm_lock_put(lock);
  470. dlm_lockres_release_ast(dlm, res);
  471. }
  472. while (!list_empty(&dlm->pending_basts)) {
  473. lock = list_entry(dlm->pending_basts.next,
  474. struct dlm_lock, bast_list);
  475. /* get an extra ref on lock */
  476. dlm_lock_get(lock);
  477. res = lock->lockres;
  478. BUG_ON(!lock->bast_pending);
  479. /* get the highest blocked lock, and reset */
  480. spin_lock(&lock->spinlock);
  481. BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
  482. hi = lock->ml.highest_blocked;
  483. lock->ml.highest_blocked = LKM_IVMODE;
  484. spin_unlock(&lock->spinlock);
  485. /* remove from list (including ref) */
  486. list_del_init(&lock->bast_list);
  487. dlm_lock_put(lock);
  488. spin_unlock(&dlm->ast_lock);
  489. mlog(0, "delivering a bast for this lockres "
  490. "(blocked = %d\n", hi);
  491. if (lock->ml.node != dlm->node_num) {
  492. ret = dlm_send_proxy_bast(dlm, res, lock, hi);
  493. if (ret < 0)
  494. mlog_errno(ret);
  495. } else
  496. dlm_do_local_bast(dlm, res, lock, hi);
  497. spin_lock(&dlm->ast_lock);
  498. /* possible that another bast was queued while
  499. * we were delivering the last one */
  500. if (!list_empty(&lock->bast_list)) {
  501. mlog(0, "aha another bast got queued while "
  502. "we were finishing the last one. will "
  503. "keep the bast_pending flag set.\n");
  504. } else
  505. lock->bast_pending = 0;
  506. /* drop the extra ref.
  507. * this may drop it completely. */
  508. dlm_lock_put(lock);
  509. dlm_lockres_release_ast(dlm, res);
  510. }
  511. wake_up(&dlm->ast_wq);
  512. spin_unlock(&dlm->ast_lock);
  513. }
  514. #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
  515. #define DLM_THREAD_MAX_DIRTY 100
  516. #define DLM_THREAD_MAX_ASTS 10
  517. static int dlm_thread(void *data)
  518. {
  519. struct dlm_lock_resource *res;
  520. struct dlm_ctxt *dlm = data;
  521. unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
  522. mlog(0, "dlm thread running for %s...\n", dlm->name);
  523. while (!kthread_should_stop()) {
  524. int n = DLM_THREAD_MAX_DIRTY;
  525. /* dlm_shutting_down is very point-in-time, but that
  526. * doesn't matter as we'll just loop back around if we
  527. * get false on the leading edge of a state
  528. * transition. */
  529. dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
  530. /* We really don't want to hold dlm->spinlock while
  531. * calling dlm_shuffle_lists on each lockres that
  532. * needs to have its queues adjusted and AST/BASTs
  533. * run. So let's pull each entry off the dirty_list
  534. * and drop dlm->spinlock ASAP. Once off the list,
  535. * res->spinlock needs to be taken again to protect
  536. * the queues while calling dlm_shuffle_lists. */
  537. spin_lock(&dlm->spinlock);
  538. while (!list_empty(&dlm->dirty_list)) {
  539. int delay = 0;
  540. res = list_entry(dlm->dirty_list.next,
  541. struct dlm_lock_resource, dirty);
  542. /* peel a lockres off, remove it from the list,
  543. * unset the dirty flag and drop the dlm lock */
  544. BUG_ON(!res);
  545. dlm_lockres_get(res);
  546. spin_lock(&res->spinlock);
  547. /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
  548. list_del_init(&res->dirty);
  549. spin_unlock(&res->spinlock);
  550. spin_unlock(&dlm->spinlock);
  551. /* Drop dirty_list ref */
  552. dlm_lockres_put(res);
  553. /* lockres can be re-dirtied/re-added to the
  554. * dirty_list in this gap, but that is ok */
  555. spin_lock(&res->spinlock);
  556. if (res->owner != dlm->node_num) {
  557. __dlm_print_one_lock_resource(res);
  558. mlog(ML_ERROR, "inprog:%s, mig:%s, reco:%s, dirty:%s\n",
  559. res->state & DLM_LOCK_RES_IN_PROGRESS ? "yes" : "no",
  560. res->state & DLM_LOCK_RES_MIGRATING ? "yes" : "no",
  561. res->state & DLM_LOCK_RES_RECOVERING ? "yes" : "no",
  562. res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
  563. }
  564. BUG_ON(res->owner != dlm->node_num);
  565. /* it is now ok to move lockreses in these states
  566. * to the dirty list, assuming that they will only be
  567. * dirty for a short while. */
  568. BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
  569. if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
  570. DLM_LOCK_RES_RECOVERING)) {
  571. /* move it to the tail and keep going */
  572. res->state &= ~DLM_LOCK_RES_DIRTY;
  573. spin_unlock(&res->spinlock);
  574. mlog(0, "delaying list shuffling for in-"
  575. "progress lockres %.*s, state=%d\n",
  576. res->lockname.len, res->lockname.name,
  577. res->state);
  578. delay = 1;
  579. goto in_progress;
  580. }
  581. /* at this point the lockres is not migrating/
  582. * recovering/in-progress. we have the lockres
  583. * spinlock and do NOT have the dlm lock.
  584. * safe to reserve/queue asts and run the lists. */
  585. mlog(0, "calling dlm_shuffle_lists with dlm=%s, "
  586. "res=%.*s\n", dlm->name,
  587. res->lockname.len, res->lockname.name);
  588. /* called while holding lockres lock */
  589. dlm_shuffle_lists(dlm, res);
  590. res->state &= ~DLM_LOCK_RES_DIRTY;
  591. spin_unlock(&res->spinlock);
  592. dlm_lockres_calc_usage(dlm, res);
  593. in_progress:
  594. spin_lock(&dlm->spinlock);
  595. /* if the lock was in-progress, stick
  596. * it on the back of the list */
  597. if (delay) {
  598. spin_lock(&res->spinlock);
  599. __dlm_dirty_lockres(dlm, res);
  600. spin_unlock(&res->spinlock);
  601. }
  602. dlm_lockres_put(res);
  603. /* unlikely, but we may need to give time to
  604. * other tasks */
  605. if (!--n) {
  606. mlog(0, "throttling dlm_thread\n");
  607. break;
  608. }
  609. }
  610. spin_unlock(&dlm->spinlock);
  611. dlm_flush_asts(dlm);
  612. /* yield and continue right away if there is more work to do */
  613. if (!n) {
  614. cond_resched();
  615. continue;
  616. }
  617. wait_event_interruptible_timeout(dlm->dlm_thread_wq,
  618. !dlm_dirty_list_empty(dlm) ||
  619. kthread_should_stop(),
  620. timeout);
  621. }
  622. mlog(0, "quitting DLM thread\n");
  623. return 0;
  624. }