dlmthread.c 20 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738
  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. spin_lock(&res->spinlock);
  154. /* This ensures that clear refmap is sent after the set */
  155. __dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
  156. spin_unlock(&res->spinlock);
  157. /* drop spinlock to do messaging, retake below */
  158. spin_unlock(&dlm->spinlock);
  159. /* clear our bit from the master's refmap, ignore errors */
  160. ret = dlm_drop_lockres_ref(dlm, res);
  161. if (ret < 0) {
  162. mlog_errno(ret);
  163. if (!dlm_is_host_down(ret))
  164. BUG();
  165. }
  166. mlog(0, "%s:%.*s: dlm_deref_lockres returned %d\n",
  167. dlm->name, res->lockname.len, res->lockname.name, ret);
  168. spin_lock(&dlm->spinlock);
  169. }
  170. if (!list_empty(&res->purge)) {
  171. mlog(0, "removing lockres %.*s:%p from purgelist, "
  172. "master = %d\n", res->lockname.len, res->lockname.name,
  173. res, master);
  174. list_del_init(&res->purge);
  175. dlm_lockres_put(res);
  176. dlm->purge_count--;
  177. }
  178. __dlm_unhash_lockres(res);
  179. /* lockres is not in the hash now. drop the flag and wake up
  180. * any processes waiting in dlm_get_lock_resource. */
  181. if (!master) {
  182. spin_lock(&res->spinlock);
  183. res->state &= ~DLM_LOCK_RES_DROPPING_REF;
  184. spin_unlock(&res->spinlock);
  185. wake_up(&res->wq);
  186. }
  187. return 0;
  188. }
  189. static void dlm_run_purge_list(struct dlm_ctxt *dlm,
  190. int purge_now)
  191. {
  192. unsigned int run_max, unused;
  193. unsigned long purge_jiffies;
  194. struct dlm_lock_resource *lockres;
  195. spin_lock(&dlm->spinlock);
  196. run_max = dlm->purge_count;
  197. while(run_max && !list_empty(&dlm->purge_list)) {
  198. run_max--;
  199. lockres = list_entry(dlm->purge_list.next,
  200. struct dlm_lock_resource, purge);
  201. /* Status of the lockres *might* change so double
  202. * check. If the lockres is unused, holding the dlm
  203. * spinlock will prevent people from getting and more
  204. * refs on it -- there's no need to keep the lockres
  205. * spinlock. */
  206. spin_lock(&lockres->spinlock);
  207. unused = __dlm_lockres_unused(lockres);
  208. spin_unlock(&lockres->spinlock);
  209. if (!unused)
  210. continue;
  211. purge_jiffies = lockres->last_used +
  212. msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
  213. /* Make sure that we want to be processing this guy at
  214. * this time. */
  215. if (!purge_now && time_after(purge_jiffies, jiffies)) {
  216. /* Since resources are added to the purge list
  217. * in tail order, we can stop at the first
  218. * unpurgable resource -- anyone added after
  219. * him will have a greater last_used value */
  220. break;
  221. }
  222. dlm_lockres_get(lockres);
  223. /* This may drop and reacquire the dlm spinlock if it
  224. * has to do migration. */
  225. if (dlm_purge_lockres(dlm, lockres))
  226. BUG();
  227. dlm_lockres_put(lockres);
  228. /* Avoid adding any scheduling latencies */
  229. cond_resched_lock(&dlm->spinlock);
  230. }
  231. spin_unlock(&dlm->spinlock);
  232. }
  233. static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
  234. struct dlm_lock_resource *res)
  235. {
  236. struct dlm_lock *lock, *target;
  237. struct list_head *iter;
  238. struct list_head *head;
  239. int can_grant = 1;
  240. //mlog(0, "res->lockname.len=%d\n", res->lockname.len);
  241. //mlog(0, "res->lockname.name=%p\n", res->lockname.name);
  242. //mlog(0, "shuffle res %.*s\n", res->lockname.len,
  243. // res->lockname.name);
  244. /* because this function is called with the lockres
  245. * spinlock, and because we know that it is not migrating/
  246. * recovering/in-progress, it is fine to reserve asts and
  247. * basts right before queueing them all throughout */
  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, "res %.*s has locks on a convert queue\n", res->lockname.len,
  256. 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: converting a lock with no "
  260. "convert_type!\n", 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, "calling ast for converting lock: %.*s, have: %d, "
  304. "granting: %d, node: %u\n", res->lockname.len,
  305. res->lockname.name, target->ml.type,
  306. target->ml.convert_type, target->ml.node);
  307. target->ml.type = target->ml.convert_type;
  308. target->ml.convert_type = LKM_IVMODE;
  309. list_move_tail(&target->list, &res->granted);
  310. BUG_ON(!target->lksb);
  311. target->lksb->status = DLM_NORMAL;
  312. spin_unlock(&target->spinlock);
  313. __dlm_lockres_reserve_ast(res);
  314. dlm_queue_ast(dlm, target);
  315. /* go back and check for more */
  316. goto converting;
  317. }
  318. blocked:
  319. if (list_empty(&res->blocked))
  320. goto leave;
  321. target = list_entry(res->blocked.next, struct dlm_lock, list);
  322. head = &res->granted;
  323. list_for_each(iter, head) {
  324. lock = list_entry(iter, struct dlm_lock, list);
  325. if (lock==target)
  326. continue;
  327. if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
  328. can_grant = 0;
  329. if (lock->ml.highest_blocked == LKM_IVMODE) {
  330. __dlm_lockres_reserve_ast(res);
  331. dlm_queue_bast(dlm, lock);
  332. }
  333. if (lock->ml.highest_blocked < target->ml.type)
  334. lock->ml.highest_blocked = target->ml.type;
  335. }
  336. }
  337. head = &res->converting;
  338. list_for_each(iter, head) {
  339. lock = list_entry(iter, struct dlm_lock, list);
  340. if (lock==target)
  341. continue;
  342. if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
  343. can_grant = 0;
  344. if (lock->ml.highest_blocked == LKM_IVMODE) {
  345. __dlm_lockres_reserve_ast(res);
  346. dlm_queue_bast(dlm, lock);
  347. }
  348. if (lock->ml.highest_blocked < target->ml.type)
  349. lock->ml.highest_blocked = target->ml.type;
  350. }
  351. }
  352. /* we can grant the blocked lock (only
  353. * possible if converting list empty) */
  354. if (can_grant) {
  355. spin_lock(&target->spinlock);
  356. BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
  357. mlog(0, "calling ast for blocked lock: %.*s, granting: %d, "
  358. "node: %u\n", res->lockname.len, res->lockname.name,
  359. target->ml.type, target->ml.node);
  360. // target->ml.type is already correct
  361. list_move_tail(&target->list, &res->granted);
  362. BUG_ON(!target->lksb);
  363. target->lksb->status = DLM_NORMAL;
  364. spin_unlock(&target->spinlock);
  365. __dlm_lockres_reserve_ast(res);
  366. dlm_queue_ast(dlm, target);
  367. /* go back and check for more */
  368. goto converting;
  369. }
  370. leave:
  371. return;
  372. }
  373. /* must have NO locks when calling this with res !=NULL * */
  374. void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
  375. {
  376. mlog_entry("dlm=%p, res=%p\n", dlm, res);
  377. if (res) {
  378. spin_lock(&dlm->spinlock);
  379. spin_lock(&res->spinlock);
  380. __dlm_dirty_lockres(dlm, res);
  381. spin_unlock(&res->spinlock);
  382. spin_unlock(&dlm->spinlock);
  383. }
  384. wake_up(&dlm->dlm_thread_wq);
  385. }
  386. void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
  387. {
  388. mlog_entry("dlm=%p, res=%p\n", dlm, res);
  389. assert_spin_locked(&dlm->spinlock);
  390. assert_spin_locked(&res->spinlock);
  391. /* don't shuffle secondary queues */
  392. if ((res->owner == dlm->node_num)) {
  393. if (res->state & (DLM_LOCK_RES_MIGRATING |
  394. DLM_LOCK_RES_BLOCK_DIRTY))
  395. return;
  396. if (list_empty(&res->dirty)) {
  397. /* ref for dirty_list */
  398. dlm_lockres_get(res);
  399. list_add_tail(&res->dirty, &dlm->dirty_list);
  400. res->state |= DLM_LOCK_RES_DIRTY;
  401. }
  402. }
  403. }
  404. /* Launch the NM thread for the mounted volume */
  405. int dlm_launch_thread(struct dlm_ctxt *dlm)
  406. {
  407. mlog(0, "starting dlm thread...\n");
  408. dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
  409. if (IS_ERR(dlm->dlm_thread_task)) {
  410. mlog_errno(PTR_ERR(dlm->dlm_thread_task));
  411. dlm->dlm_thread_task = NULL;
  412. return -EINVAL;
  413. }
  414. return 0;
  415. }
  416. void dlm_complete_thread(struct dlm_ctxt *dlm)
  417. {
  418. if (dlm->dlm_thread_task) {
  419. mlog(ML_KTHREAD, "waiting for dlm thread to exit\n");
  420. kthread_stop(dlm->dlm_thread_task);
  421. dlm->dlm_thread_task = NULL;
  422. }
  423. }
  424. static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
  425. {
  426. int empty;
  427. spin_lock(&dlm->spinlock);
  428. empty = list_empty(&dlm->dirty_list);
  429. spin_unlock(&dlm->spinlock);
  430. return empty;
  431. }
  432. static void dlm_flush_asts(struct dlm_ctxt *dlm)
  433. {
  434. int ret;
  435. struct dlm_lock *lock;
  436. struct dlm_lock_resource *res;
  437. u8 hi;
  438. spin_lock(&dlm->ast_lock);
  439. while (!list_empty(&dlm->pending_asts)) {
  440. lock = list_entry(dlm->pending_asts.next,
  441. struct dlm_lock, ast_list);
  442. /* get an extra ref on lock */
  443. dlm_lock_get(lock);
  444. res = lock->lockres;
  445. mlog(0, "delivering an ast for this lockres\n");
  446. BUG_ON(!lock->ast_pending);
  447. /* remove from list (including ref) */
  448. list_del_init(&lock->ast_list);
  449. dlm_lock_put(lock);
  450. spin_unlock(&dlm->ast_lock);
  451. if (lock->ml.node != dlm->node_num) {
  452. ret = dlm_do_remote_ast(dlm, res, lock);
  453. if (ret < 0)
  454. mlog_errno(ret);
  455. } else
  456. dlm_do_local_ast(dlm, res, lock);
  457. spin_lock(&dlm->ast_lock);
  458. /* possible that another ast was queued while
  459. * we were delivering the last one */
  460. if (!list_empty(&lock->ast_list)) {
  461. mlog(0, "aha another ast got queued while "
  462. "we were finishing the last one. will "
  463. "keep the ast_pending flag set.\n");
  464. } else
  465. lock->ast_pending = 0;
  466. /* drop the extra ref.
  467. * this may drop it completely. */
  468. dlm_lock_put(lock);
  469. dlm_lockres_release_ast(dlm, res);
  470. }
  471. while (!list_empty(&dlm->pending_basts)) {
  472. lock = list_entry(dlm->pending_basts.next,
  473. struct dlm_lock, bast_list);
  474. /* get an extra ref on lock */
  475. dlm_lock_get(lock);
  476. res = lock->lockres;
  477. BUG_ON(!lock->bast_pending);
  478. /* get the highest blocked lock, and reset */
  479. spin_lock(&lock->spinlock);
  480. BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
  481. hi = lock->ml.highest_blocked;
  482. lock->ml.highest_blocked = LKM_IVMODE;
  483. spin_unlock(&lock->spinlock);
  484. /* remove from list (including ref) */
  485. list_del_init(&lock->bast_list);
  486. dlm_lock_put(lock);
  487. spin_unlock(&dlm->ast_lock);
  488. mlog(0, "delivering a bast for this lockres "
  489. "(blocked = %d\n", hi);
  490. if (lock->ml.node != dlm->node_num) {
  491. ret = dlm_send_proxy_bast(dlm, res, lock, hi);
  492. if (ret < 0)
  493. mlog_errno(ret);
  494. } else
  495. dlm_do_local_bast(dlm, res, lock, hi);
  496. spin_lock(&dlm->ast_lock);
  497. /* possible that another bast was queued while
  498. * we were delivering the last one */
  499. if (!list_empty(&lock->bast_list)) {
  500. mlog(0, "aha another bast got queued while "
  501. "we were finishing the last one. will "
  502. "keep the bast_pending flag set.\n");
  503. } else
  504. lock->bast_pending = 0;
  505. /* drop the extra ref.
  506. * this may drop it completely. */
  507. dlm_lock_put(lock);
  508. dlm_lockres_release_ast(dlm, res);
  509. }
  510. wake_up(&dlm->ast_wq);
  511. spin_unlock(&dlm->ast_lock);
  512. }
  513. #define DLM_THREAD_TIMEOUT_MS (4 * 1000)
  514. #define DLM_THREAD_MAX_DIRTY 100
  515. #define DLM_THREAD_MAX_ASTS 10
  516. static int dlm_thread(void *data)
  517. {
  518. struct dlm_lock_resource *res;
  519. struct dlm_ctxt *dlm = data;
  520. unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
  521. mlog(0, "dlm thread running for %s...\n", dlm->name);
  522. while (!kthread_should_stop()) {
  523. int n = DLM_THREAD_MAX_DIRTY;
  524. /* dlm_shutting_down is very point-in-time, but that
  525. * doesn't matter as we'll just loop back around if we
  526. * get false on the leading edge of a state
  527. * transition. */
  528. dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
  529. /* We really don't want to hold dlm->spinlock while
  530. * calling dlm_shuffle_lists on each lockres that
  531. * needs to have its queues adjusted and AST/BASTs
  532. * run. So let's pull each entry off the dirty_list
  533. * and drop dlm->spinlock ASAP. Once off the list,
  534. * res->spinlock needs to be taken again to protect
  535. * the queues while calling dlm_shuffle_lists. */
  536. spin_lock(&dlm->spinlock);
  537. while (!list_empty(&dlm->dirty_list)) {
  538. int delay = 0;
  539. res = list_entry(dlm->dirty_list.next,
  540. struct dlm_lock_resource, dirty);
  541. /* peel a lockres off, remove it from the list,
  542. * unset the dirty flag and drop the dlm lock */
  543. BUG_ON(!res);
  544. dlm_lockres_get(res);
  545. spin_lock(&res->spinlock);
  546. /* We clear the DLM_LOCK_RES_DIRTY state once we shuffle lists below */
  547. list_del_init(&res->dirty);
  548. spin_unlock(&res->spinlock);
  549. spin_unlock(&dlm->spinlock);
  550. /* Drop dirty_list ref */
  551. dlm_lockres_put(res);
  552. /* lockres can be re-dirtied/re-added to the
  553. * dirty_list in this gap, but that is ok */
  554. spin_lock(&res->spinlock);
  555. if (res->owner != dlm->node_num) {
  556. __dlm_print_one_lock_resource(res);
  557. mlog(ML_ERROR, "inprog:%s, mig:%s, reco:%s, dirty:%s\n",
  558. res->state & DLM_LOCK_RES_IN_PROGRESS ? "yes" : "no",
  559. res->state & DLM_LOCK_RES_MIGRATING ? "yes" : "no",
  560. res->state & DLM_LOCK_RES_RECOVERING ? "yes" : "no",
  561. res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
  562. }
  563. BUG_ON(res->owner != dlm->node_num);
  564. /* it is now ok to move lockreses in these states
  565. * to the dirty list, assuming that they will only be
  566. * dirty for a short while. */
  567. BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
  568. if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
  569. DLM_LOCK_RES_RECOVERING)) {
  570. /* move it to the tail and keep going */
  571. res->state &= ~DLM_LOCK_RES_DIRTY;
  572. spin_unlock(&res->spinlock);
  573. mlog(0, "delaying list shuffling for in-"
  574. "progress lockres %.*s, state=%d\n",
  575. res->lockname.len, res->lockname.name,
  576. res->state);
  577. delay = 1;
  578. goto in_progress;
  579. }
  580. /* at this point the lockres is not migrating/
  581. * recovering/in-progress. we have the lockres
  582. * spinlock and do NOT have the dlm lock.
  583. * safe to reserve/queue asts and run the lists. */
  584. mlog(0, "calling dlm_shuffle_lists with dlm=%s, "
  585. "res=%.*s\n", dlm->name,
  586. res->lockname.len, res->lockname.name);
  587. /* called while holding lockres lock */
  588. dlm_shuffle_lists(dlm, res);
  589. res->state &= ~DLM_LOCK_RES_DIRTY;
  590. spin_unlock(&res->spinlock);
  591. dlm_lockres_calc_usage(dlm, res);
  592. in_progress:
  593. spin_lock(&dlm->spinlock);
  594. /* if the lock was in-progress, stick
  595. * it on the back of the list */
  596. if (delay) {
  597. spin_lock(&res->spinlock);
  598. __dlm_dirty_lockres(dlm, res);
  599. spin_unlock(&res->spinlock);
  600. }
  601. dlm_lockres_put(res);
  602. /* unlikely, but we may need to give time to
  603. * other tasks */
  604. if (!--n) {
  605. mlog(0, "throttling dlm_thread\n");
  606. break;
  607. }
  608. }
  609. spin_unlock(&dlm->spinlock);
  610. dlm_flush_asts(dlm);
  611. /* yield and continue right away if there is more work to do */
  612. if (!n) {
  613. cond_resched();
  614. continue;
  615. }
  616. wait_event_interruptible_timeout(dlm->dlm_thread_wq,
  617. !dlm_dirty_list_empty(dlm) ||
  618. kthread_should_stop(),
  619. timeout);
  620. }
  621. mlog(0, "quitting DLM thread\n");
  622. return 0;
  623. }