as-iosched.c 38 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485
  1. /*
  2. * Anticipatory & deadline i/o scheduler.
  3. *
  4. * Copyright (C) 2002 Jens Axboe <axboe@kernel.dk>
  5. * Nick Piggin <nickpiggin@yahoo.com.au>
  6. *
  7. */
  8. #include <linux/kernel.h>
  9. #include <linux/fs.h>
  10. #include <linux/blkdev.h>
  11. #include <linux/elevator.h>
  12. #include <linux/bio.h>
  13. #include <linux/module.h>
  14. #include <linux/slab.h>
  15. #include <linux/init.h>
  16. #include <linux/compiler.h>
  17. #include <linux/rbtree.h>
  18. #include <linux/interrupt.h>
  19. #define REQ_SYNC 1
  20. #define REQ_ASYNC 0
  21. /*
  22. * See Documentation/block/as-iosched.txt
  23. */
  24. /*
  25. * max time before a read is submitted.
  26. */
  27. #define default_read_expire (HZ / 8)
  28. /*
  29. * ditto for writes, these limits are not hard, even
  30. * if the disk is capable of satisfying them.
  31. */
  32. #define default_write_expire (HZ / 4)
  33. /*
  34. * read_batch_expire describes how long we will allow a stream of reads to
  35. * persist before looking to see whether it is time to switch over to writes.
  36. */
  37. #define default_read_batch_expire (HZ / 2)
  38. /*
  39. * write_batch_expire describes how long we want a stream of writes to run for.
  40. * This is not a hard limit, but a target we set for the auto-tuning thingy.
  41. * See, the problem is: we can send a lot of writes to disk cache / TCQ in
  42. * a short amount of time...
  43. */
  44. #define default_write_batch_expire (HZ / 8)
  45. /*
  46. * max time we may wait to anticipate a read (default around 6ms)
  47. */
  48. #define default_antic_expire ((HZ / 150) ? HZ / 150 : 1)
  49. /*
  50. * Keep track of up to 20ms thinktimes. We can go as big as we like here,
  51. * however huge values tend to interfere and not decay fast enough. A program
  52. * might be in a non-io phase of operation. Waiting on user input for example,
  53. * or doing a lengthy computation. A small penalty can be justified there, and
  54. * will still catch out those processes that constantly have large thinktimes.
  55. */
  56. #define MAX_THINKTIME (HZ/50UL)
  57. /* Bits in as_io_context.state */
  58. enum as_io_states {
  59. AS_TASK_RUNNING=0, /* Process has not exited */
  60. AS_TASK_IOSTARTED, /* Process has started some IO */
  61. AS_TASK_IORUNNING, /* Process has completed some IO */
  62. };
  63. enum anticipation_status {
  64. ANTIC_OFF=0, /* Not anticipating (normal operation) */
  65. ANTIC_WAIT_REQ, /* The last read has not yet completed */
  66. ANTIC_WAIT_NEXT, /* Currently anticipating a request vs
  67. last read (which has completed) */
  68. ANTIC_FINISHED, /* Anticipating but have found a candidate
  69. * or timed out */
  70. };
  71. struct as_data {
  72. /*
  73. * run time data
  74. */
  75. struct request_queue *q; /* the "owner" queue */
  76. /*
  77. * requests (as_rq s) are present on both sort_list and fifo_list
  78. */
  79. struct rb_root sort_list[2];
  80. struct list_head fifo_list[2];
  81. struct request *next_rq[2]; /* next in sort order */
  82. sector_t last_sector[2]; /* last REQ_SYNC & REQ_ASYNC sectors */
  83. unsigned long exit_prob; /* probability a task will exit while
  84. being waited on */
  85. unsigned long exit_no_coop; /* probablility an exited task will
  86. not be part of a later cooperating
  87. request */
  88. unsigned long new_ttime_total; /* mean thinktime on new proc */
  89. unsigned long new_ttime_mean;
  90. u64 new_seek_total; /* mean seek on new proc */
  91. sector_t new_seek_mean;
  92. unsigned long current_batch_expires;
  93. unsigned long last_check_fifo[2];
  94. int changed_batch; /* 1: waiting for old batch to end */
  95. int new_batch; /* 1: waiting on first read complete */
  96. int batch_data_dir; /* current batch REQ_SYNC / REQ_ASYNC */
  97. int write_batch_count; /* max # of reqs in a write batch */
  98. int current_write_count; /* how many requests left this batch */
  99. int write_batch_idled; /* has the write batch gone idle? */
  100. enum anticipation_status antic_status;
  101. unsigned long antic_start; /* jiffies: when it started */
  102. struct timer_list antic_timer; /* anticipatory scheduling timer */
  103. struct work_struct antic_work; /* Deferred unplugging */
  104. struct io_context *io_context; /* Identify the expected process */
  105. int ioc_finished; /* IO associated with io_context is finished */
  106. int nr_dispatched;
  107. /*
  108. * settings that change how the i/o scheduler behaves
  109. */
  110. unsigned long fifo_expire[2];
  111. unsigned long batch_expire[2];
  112. unsigned long antic_expire;
  113. };
  114. /*
  115. * per-request data.
  116. */
  117. enum arq_state {
  118. AS_RQ_NEW=0, /* New - not referenced and not on any lists */
  119. AS_RQ_QUEUED, /* In the request queue. It belongs to the
  120. scheduler */
  121. AS_RQ_DISPATCHED, /* On the dispatch list. It belongs to the
  122. driver now */
  123. AS_RQ_PRESCHED, /* Debug poisoning for requests being used */
  124. AS_RQ_REMOVED,
  125. AS_RQ_MERGED,
  126. AS_RQ_POSTSCHED, /* when they shouldn't be */
  127. };
  128. #define RQ_IOC(rq) ((struct io_context *) (rq)->elevator_private)
  129. #define RQ_STATE(rq) ((enum arq_state)(rq)->elevator_private2)
  130. #define RQ_SET_STATE(rq, state) ((rq)->elevator_private2 = (void *) state)
  131. static DEFINE_PER_CPU(unsigned long, ioc_count);
  132. static struct completion *ioc_gone;
  133. static void as_move_to_dispatch(struct as_data *ad, struct request *rq);
  134. static void as_antic_stop(struct as_data *ad);
  135. /*
  136. * IO Context helper functions
  137. */
  138. /* Called to deallocate the as_io_context */
  139. static void free_as_io_context(struct as_io_context *aic)
  140. {
  141. kfree(aic);
  142. elv_ioc_count_dec(ioc_count);
  143. if (ioc_gone && !elv_ioc_count_read(ioc_count))
  144. complete(ioc_gone);
  145. }
  146. static void as_trim(struct io_context *ioc)
  147. {
  148. if (ioc->aic)
  149. free_as_io_context(ioc->aic);
  150. ioc->aic = NULL;
  151. }
  152. /* Called when the task exits */
  153. static void exit_as_io_context(struct as_io_context *aic)
  154. {
  155. WARN_ON(!test_bit(AS_TASK_RUNNING, &aic->state));
  156. clear_bit(AS_TASK_RUNNING, &aic->state);
  157. }
  158. static struct as_io_context *alloc_as_io_context(void)
  159. {
  160. struct as_io_context *ret;
  161. ret = kmalloc(sizeof(*ret), GFP_ATOMIC);
  162. if (ret) {
  163. ret->dtor = free_as_io_context;
  164. ret->exit = exit_as_io_context;
  165. ret->state = 1 << AS_TASK_RUNNING;
  166. atomic_set(&ret->nr_queued, 0);
  167. atomic_set(&ret->nr_dispatched, 0);
  168. spin_lock_init(&ret->lock);
  169. ret->ttime_total = 0;
  170. ret->ttime_samples = 0;
  171. ret->ttime_mean = 0;
  172. ret->seek_total = 0;
  173. ret->seek_samples = 0;
  174. ret->seek_mean = 0;
  175. elv_ioc_count_inc(ioc_count);
  176. }
  177. return ret;
  178. }
  179. /*
  180. * If the current task has no AS IO context then create one and initialise it.
  181. * Then take a ref on the task's io context and return it.
  182. */
  183. static struct io_context *as_get_io_context(int node)
  184. {
  185. struct io_context *ioc = get_io_context(GFP_ATOMIC, node);
  186. if (ioc && !ioc->aic) {
  187. ioc->aic = alloc_as_io_context();
  188. if (!ioc->aic) {
  189. put_io_context(ioc);
  190. ioc = NULL;
  191. }
  192. }
  193. return ioc;
  194. }
  195. static void as_put_io_context(struct request *rq)
  196. {
  197. struct as_io_context *aic;
  198. if (unlikely(!RQ_IOC(rq)))
  199. return;
  200. aic = RQ_IOC(rq)->aic;
  201. if (rq_is_sync(rq) && aic) {
  202. spin_lock(&aic->lock);
  203. set_bit(AS_TASK_IORUNNING, &aic->state);
  204. aic->last_end_request = jiffies;
  205. spin_unlock(&aic->lock);
  206. }
  207. put_io_context(RQ_IOC(rq));
  208. }
  209. /*
  210. * rb tree support functions
  211. */
  212. #define RQ_RB_ROOT(ad, rq) (&(ad)->sort_list[rq_is_sync((rq))])
  213. static void as_add_rq_rb(struct as_data *ad, struct request *rq)
  214. {
  215. struct request *alias;
  216. while ((unlikely(alias = elv_rb_add(RQ_RB_ROOT(ad, rq), rq)))) {
  217. as_move_to_dispatch(ad, alias);
  218. as_antic_stop(ad);
  219. }
  220. }
  221. static inline void as_del_rq_rb(struct as_data *ad, struct request *rq)
  222. {
  223. elv_rb_del(RQ_RB_ROOT(ad, rq), rq);
  224. }
  225. /*
  226. * IO Scheduler proper
  227. */
  228. #define MAXBACK (1024 * 1024) /*
  229. * Maximum distance the disk will go backward
  230. * for a request.
  231. */
  232. #define BACK_PENALTY 2
  233. /*
  234. * as_choose_req selects the preferred one of two requests of the same data_dir
  235. * ignoring time - eg. timeouts, which is the job of as_dispatch_request
  236. */
  237. static struct request *
  238. as_choose_req(struct as_data *ad, struct request *rq1, struct request *rq2)
  239. {
  240. int data_dir;
  241. sector_t last, s1, s2, d1, d2;
  242. int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */
  243. const sector_t maxback = MAXBACK;
  244. if (rq1 == NULL || rq1 == rq2)
  245. return rq2;
  246. if (rq2 == NULL)
  247. return rq1;
  248. data_dir = rq_is_sync(rq1);
  249. last = ad->last_sector[data_dir];
  250. s1 = rq1->sector;
  251. s2 = rq2->sector;
  252. BUG_ON(data_dir != rq_is_sync(rq2));
  253. /*
  254. * Strict one way elevator _except_ in the case where we allow
  255. * short backward seeks which are biased as twice the cost of a
  256. * similar forward seek.
  257. */
  258. if (s1 >= last)
  259. d1 = s1 - last;
  260. else if (s1+maxback >= last)
  261. d1 = (last - s1)*BACK_PENALTY;
  262. else {
  263. r1_wrap = 1;
  264. d1 = 0; /* shut up, gcc */
  265. }
  266. if (s2 >= last)
  267. d2 = s2 - last;
  268. else if (s2+maxback >= last)
  269. d2 = (last - s2)*BACK_PENALTY;
  270. else {
  271. r2_wrap = 1;
  272. d2 = 0;
  273. }
  274. /* Found required data */
  275. if (!r1_wrap && r2_wrap)
  276. return rq1;
  277. else if (!r2_wrap && r1_wrap)
  278. return rq2;
  279. else if (r1_wrap && r2_wrap) {
  280. /* both behind the head */
  281. if (s1 <= s2)
  282. return rq1;
  283. else
  284. return rq2;
  285. }
  286. /* Both requests in front of the head */
  287. if (d1 < d2)
  288. return rq1;
  289. else if (d2 < d1)
  290. return rq2;
  291. else {
  292. if (s1 >= s2)
  293. return rq1;
  294. else
  295. return rq2;
  296. }
  297. }
  298. /*
  299. * as_find_next_rq finds the next request after @prev in elevator order.
  300. * this with as_choose_req form the basis for how the scheduler chooses
  301. * what request to process next. Anticipation works on top of this.
  302. */
  303. static struct request *
  304. as_find_next_rq(struct as_data *ad, struct request *last)
  305. {
  306. struct rb_node *rbnext = rb_next(&last->rb_node);
  307. struct rb_node *rbprev = rb_prev(&last->rb_node);
  308. struct request *next = NULL, *prev = NULL;
  309. BUG_ON(RB_EMPTY_NODE(&last->rb_node));
  310. if (rbprev)
  311. prev = rb_entry_rq(rbprev);
  312. if (rbnext)
  313. next = rb_entry_rq(rbnext);
  314. else {
  315. const int data_dir = rq_is_sync(last);
  316. rbnext = rb_first(&ad->sort_list[data_dir]);
  317. if (rbnext && rbnext != &last->rb_node)
  318. next = rb_entry_rq(rbnext);
  319. }
  320. return as_choose_req(ad, next, prev);
  321. }
  322. /*
  323. * anticipatory scheduling functions follow
  324. */
  325. /*
  326. * as_antic_expired tells us when we have anticipated too long.
  327. * The funny "absolute difference" math on the elapsed time is to handle
  328. * jiffy wraps, and disks which have been idle for 0x80000000 jiffies.
  329. */
  330. static int as_antic_expired(struct as_data *ad)
  331. {
  332. long delta_jif;
  333. delta_jif = jiffies - ad->antic_start;
  334. if (unlikely(delta_jif < 0))
  335. delta_jif = -delta_jif;
  336. if (delta_jif < ad->antic_expire)
  337. return 0;
  338. return 1;
  339. }
  340. /*
  341. * as_antic_waitnext starts anticipating that a nice request will soon be
  342. * submitted. See also as_antic_waitreq
  343. */
  344. static void as_antic_waitnext(struct as_data *ad)
  345. {
  346. unsigned long timeout;
  347. BUG_ON(ad->antic_status != ANTIC_OFF
  348. && ad->antic_status != ANTIC_WAIT_REQ);
  349. timeout = ad->antic_start + ad->antic_expire;
  350. mod_timer(&ad->antic_timer, timeout);
  351. ad->antic_status = ANTIC_WAIT_NEXT;
  352. }
  353. /*
  354. * as_antic_waitreq starts anticipating. We don't start timing the anticipation
  355. * until the request that we're anticipating on has finished. This means we
  356. * are timing from when the candidate process wakes up hopefully.
  357. */
  358. static void as_antic_waitreq(struct as_data *ad)
  359. {
  360. BUG_ON(ad->antic_status == ANTIC_FINISHED);
  361. if (ad->antic_status == ANTIC_OFF) {
  362. if (!ad->io_context || ad->ioc_finished)
  363. as_antic_waitnext(ad);
  364. else
  365. ad->antic_status = ANTIC_WAIT_REQ;
  366. }
  367. }
  368. /*
  369. * This is called directly by the functions in this file to stop anticipation.
  370. * We kill the timer and schedule a call to the request_fn asap.
  371. */
  372. static void as_antic_stop(struct as_data *ad)
  373. {
  374. int status = ad->antic_status;
  375. if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) {
  376. if (status == ANTIC_WAIT_NEXT)
  377. del_timer(&ad->antic_timer);
  378. ad->antic_status = ANTIC_FINISHED;
  379. /* see as_work_handler */
  380. kblockd_schedule_work(&ad->antic_work);
  381. }
  382. }
  383. /*
  384. * as_antic_timeout is the timer function set by as_antic_waitnext.
  385. */
  386. static void as_antic_timeout(unsigned long data)
  387. {
  388. struct request_queue *q = (struct request_queue *)data;
  389. struct as_data *ad = q->elevator->elevator_data;
  390. unsigned long flags;
  391. spin_lock_irqsave(q->queue_lock, flags);
  392. if (ad->antic_status == ANTIC_WAIT_REQ
  393. || ad->antic_status == ANTIC_WAIT_NEXT) {
  394. struct as_io_context *aic = ad->io_context->aic;
  395. ad->antic_status = ANTIC_FINISHED;
  396. kblockd_schedule_work(&ad->antic_work);
  397. if (aic->ttime_samples == 0) {
  398. /* process anticipated on has exited or timed out*/
  399. ad->exit_prob = (7*ad->exit_prob + 256)/8;
  400. }
  401. if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
  402. /* process not "saved" by a cooperating request */
  403. ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8;
  404. }
  405. }
  406. spin_unlock_irqrestore(q->queue_lock, flags);
  407. }
  408. static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic,
  409. unsigned long ttime)
  410. {
  411. /* fixed point: 1.0 == 1<<8 */
  412. if (aic->ttime_samples == 0) {
  413. ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8;
  414. ad->new_ttime_mean = ad->new_ttime_total / 256;
  415. ad->exit_prob = (7*ad->exit_prob)/8;
  416. }
  417. aic->ttime_samples = (7*aic->ttime_samples + 256) / 8;
  418. aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8;
  419. aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples;
  420. }
  421. static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic,
  422. sector_t sdist)
  423. {
  424. u64 total;
  425. if (aic->seek_samples == 0) {
  426. ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8;
  427. ad->new_seek_mean = ad->new_seek_total / 256;
  428. }
  429. /*
  430. * Don't allow the seek distance to get too large from the
  431. * odd fragment, pagein, etc
  432. */
  433. if (aic->seek_samples <= 60) /* second&third seek */
  434. sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024);
  435. else
  436. sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64);
  437. aic->seek_samples = (7*aic->seek_samples + 256) / 8;
  438. aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8;
  439. total = aic->seek_total + (aic->seek_samples/2);
  440. do_div(total, aic->seek_samples);
  441. aic->seek_mean = (sector_t)total;
  442. }
  443. /*
  444. * as_update_iohist keeps a decaying histogram of IO thinktimes, and
  445. * updates @aic->ttime_mean based on that. It is called when a new
  446. * request is queued.
  447. */
  448. static void as_update_iohist(struct as_data *ad, struct as_io_context *aic,
  449. struct request *rq)
  450. {
  451. int data_dir = rq_is_sync(rq);
  452. unsigned long thinktime = 0;
  453. sector_t seek_dist;
  454. if (aic == NULL)
  455. return;
  456. if (data_dir == REQ_SYNC) {
  457. unsigned long in_flight = atomic_read(&aic->nr_queued)
  458. + atomic_read(&aic->nr_dispatched);
  459. spin_lock(&aic->lock);
  460. if (test_bit(AS_TASK_IORUNNING, &aic->state) ||
  461. test_bit(AS_TASK_IOSTARTED, &aic->state)) {
  462. /* Calculate read -> read thinktime */
  463. if (test_bit(AS_TASK_IORUNNING, &aic->state)
  464. && in_flight == 0) {
  465. thinktime = jiffies - aic->last_end_request;
  466. thinktime = min(thinktime, MAX_THINKTIME-1);
  467. }
  468. as_update_thinktime(ad, aic, thinktime);
  469. /* Calculate read -> read seek distance */
  470. if (aic->last_request_pos < rq->sector)
  471. seek_dist = rq->sector - aic->last_request_pos;
  472. else
  473. seek_dist = aic->last_request_pos - rq->sector;
  474. as_update_seekdist(ad, aic, seek_dist);
  475. }
  476. aic->last_request_pos = rq->sector + rq->nr_sectors;
  477. set_bit(AS_TASK_IOSTARTED, &aic->state);
  478. spin_unlock(&aic->lock);
  479. }
  480. }
  481. /*
  482. * as_close_req decides if one request is considered "close" to the
  483. * previous one issued.
  484. */
  485. static int as_close_req(struct as_data *ad, struct as_io_context *aic,
  486. struct request *rq)
  487. {
  488. unsigned long delay; /* milliseconds */
  489. sector_t last = ad->last_sector[ad->batch_data_dir];
  490. sector_t next = rq->sector;
  491. sector_t delta; /* acceptable close offset (in sectors) */
  492. sector_t s;
  493. if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished)
  494. delay = 0;
  495. else
  496. delay = ((jiffies - ad->antic_start) * 1000) / HZ;
  497. if (delay == 0)
  498. delta = 8192;
  499. else if (delay <= 20 && delay <= ad->antic_expire)
  500. delta = 8192 << delay;
  501. else
  502. return 1;
  503. if ((last <= next + (delta>>1)) && (next <= last + delta))
  504. return 1;
  505. if (last < next)
  506. s = next - last;
  507. else
  508. s = last - next;
  509. if (aic->seek_samples == 0) {
  510. /*
  511. * Process has just started IO. Use past statistics to
  512. * gauge success possibility
  513. */
  514. if (ad->new_seek_mean > s) {
  515. /* this request is better than what we're expecting */
  516. return 1;
  517. }
  518. } else {
  519. if (aic->seek_mean > s) {
  520. /* this request is better than what we're expecting */
  521. return 1;
  522. }
  523. }
  524. return 0;
  525. }
  526. /*
  527. * as_can_break_anticipation returns true if we have been anticipating this
  528. * request.
  529. *
  530. * It also returns true if the process against which we are anticipating
  531. * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to
  532. * dispatch it ASAP, because we know that application will not be submitting
  533. * any new reads.
  534. *
  535. * If the task which has submitted the request has exited, break anticipation.
  536. *
  537. * If this task has queued some other IO, do not enter enticipation.
  538. */
  539. static int as_can_break_anticipation(struct as_data *ad, struct request *rq)
  540. {
  541. struct io_context *ioc;
  542. struct as_io_context *aic;
  543. ioc = ad->io_context;
  544. BUG_ON(!ioc);
  545. if (rq && ioc == RQ_IOC(rq)) {
  546. /* request from same process */
  547. return 1;
  548. }
  549. if (ad->ioc_finished && as_antic_expired(ad)) {
  550. /*
  551. * In this situation status should really be FINISHED,
  552. * however the timer hasn't had the chance to run yet.
  553. */
  554. return 1;
  555. }
  556. aic = ioc->aic;
  557. if (!aic)
  558. return 0;
  559. if (atomic_read(&aic->nr_queued) > 0) {
  560. /* process has more requests queued */
  561. return 1;
  562. }
  563. if (atomic_read(&aic->nr_dispatched) > 0) {
  564. /* process has more requests dispatched */
  565. return 1;
  566. }
  567. if (rq && rq_is_sync(rq) && as_close_req(ad, aic, rq)) {
  568. /*
  569. * Found a close request that is not one of ours.
  570. *
  571. * This makes close requests from another process update
  572. * our IO history. Is generally useful when there are
  573. * two or more cooperating processes working in the same
  574. * area.
  575. */
  576. if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
  577. if (aic->ttime_samples == 0)
  578. ad->exit_prob = (7*ad->exit_prob + 256)/8;
  579. ad->exit_no_coop = (7*ad->exit_no_coop)/8;
  580. }
  581. as_update_iohist(ad, aic, rq);
  582. return 1;
  583. }
  584. if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
  585. /* process anticipated on has exited */
  586. if (aic->ttime_samples == 0)
  587. ad->exit_prob = (7*ad->exit_prob + 256)/8;
  588. if (ad->exit_no_coop > 128)
  589. return 1;
  590. }
  591. if (aic->ttime_samples == 0) {
  592. if (ad->new_ttime_mean > ad->antic_expire)
  593. return 1;
  594. if (ad->exit_prob * ad->exit_no_coop > 128*256)
  595. return 1;
  596. } else if (aic->ttime_mean > ad->antic_expire) {
  597. /* the process thinks too much between requests */
  598. return 1;
  599. }
  600. return 0;
  601. }
  602. /*
  603. * as_can_anticipate indicates whether we should either run rq
  604. * or keep anticipating a better request.
  605. */
  606. static int as_can_anticipate(struct as_data *ad, struct request *rq)
  607. {
  608. if (!ad->io_context)
  609. /*
  610. * Last request submitted was a write
  611. */
  612. return 0;
  613. if (ad->antic_status == ANTIC_FINISHED)
  614. /*
  615. * Don't restart if we have just finished. Run the next request
  616. */
  617. return 0;
  618. if (as_can_break_anticipation(ad, rq))
  619. /*
  620. * This request is a good candidate. Don't keep anticipating,
  621. * run it.
  622. */
  623. return 0;
  624. /*
  625. * OK from here, we haven't finished, and don't have a decent request!
  626. * Status is either ANTIC_OFF so start waiting,
  627. * ANTIC_WAIT_REQ so continue waiting for request to finish
  628. * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request.
  629. */
  630. return 1;
  631. }
  632. /*
  633. * as_update_rq must be called whenever a request (rq) is added to
  634. * the sort_list. This function keeps caches up to date, and checks if the
  635. * request might be one we are "anticipating"
  636. */
  637. static void as_update_rq(struct as_data *ad, struct request *rq)
  638. {
  639. const int data_dir = rq_is_sync(rq);
  640. /* keep the next_rq cache up to date */
  641. ad->next_rq[data_dir] = as_choose_req(ad, rq, ad->next_rq[data_dir]);
  642. /*
  643. * have we been anticipating this request?
  644. * or does it come from the same process as the one we are anticipating
  645. * for?
  646. */
  647. if (ad->antic_status == ANTIC_WAIT_REQ
  648. || ad->antic_status == ANTIC_WAIT_NEXT) {
  649. if (as_can_break_anticipation(ad, rq))
  650. as_antic_stop(ad);
  651. }
  652. }
  653. /*
  654. * Gathers timings and resizes the write batch automatically
  655. */
  656. static void update_write_batch(struct as_data *ad)
  657. {
  658. unsigned long batch = ad->batch_expire[REQ_ASYNC];
  659. long write_time;
  660. write_time = (jiffies - ad->current_batch_expires) + batch;
  661. if (write_time < 0)
  662. write_time = 0;
  663. if (write_time > batch && !ad->write_batch_idled) {
  664. if (write_time > batch * 3)
  665. ad->write_batch_count /= 2;
  666. else
  667. ad->write_batch_count--;
  668. } else if (write_time < batch && ad->current_write_count == 0) {
  669. if (batch > write_time * 3)
  670. ad->write_batch_count *= 2;
  671. else
  672. ad->write_batch_count++;
  673. }
  674. if (ad->write_batch_count < 1)
  675. ad->write_batch_count = 1;
  676. }
  677. /*
  678. * as_completed_request is to be called when a request has completed and
  679. * returned something to the requesting process, be it an error or data.
  680. */
  681. static void as_completed_request(request_queue_t *q, struct request *rq)
  682. {
  683. struct as_data *ad = q->elevator->elevator_data;
  684. WARN_ON(!list_empty(&rq->queuelist));
  685. if (RQ_STATE(rq) != AS_RQ_REMOVED) {
  686. printk("rq->state %d\n", RQ_STATE(rq));
  687. WARN_ON(1);
  688. goto out;
  689. }
  690. if (ad->changed_batch && ad->nr_dispatched == 1) {
  691. kblockd_schedule_work(&ad->antic_work);
  692. ad->changed_batch = 0;
  693. if (ad->batch_data_dir == REQ_SYNC)
  694. ad->new_batch = 1;
  695. }
  696. WARN_ON(ad->nr_dispatched == 0);
  697. ad->nr_dispatched--;
  698. /*
  699. * Start counting the batch from when a request of that direction is
  700. * actually serviced. This should help devices with big TCQ windows
  701. * and writeback caches
  702. */
  703. if (ad->new_batch && ad->batch_data_dir == rq_is_sync(rq)) {
  704. update_write_batch(ad);
  705. ad->current_batch_expires = jiffies +
  706. ad->batch_expire[REQ_SYNC];
  707. ad->new_batch = 0;
  708. }
  709. if (ad->io_context == RQ_IOC(rq) && ad->io_context) {
  710. ad->antic_start = jiffies;
  711. ad->ioc_finished = 1;
  712. if (ad->antic_status == ANTIC_WAIT_REQ) {
  713. /*
  714. * We were waiting on this request, now anticipate
  715. * the next one
  716. */
  717. as_antic_waitnext(ad);
  718. }
  719. }
  720. as_put_io_context(rq);
  721. out:
  722. RQ_SET_STATE(rq, AS_RQ_POSTSCHED);
  723. }
  724. /*
  725. * as_remove_queued_request removes a request from the pre dispatch queue
  726. * without updating refcounts. It is expected the caller will drop the
  727. * reference unless it replaces the request at somepart of the elevator
  728. * (ie. the dispatch queue)
  729. */
  730. static void as_remove_queued_request(request_queue_t *q, struct request *rq)
  731. {
  732. const int data_dir = rq_is_sync(rq);
  733. struct as_data *ad = q->elevator->elevator_data;
  734. struct io_context *ioc;
  735. WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED);
  736. ioc = RQ_IOC(rq);
  737. if (ioc && ioc->aic) {
  738. BUG_ON(!atomic_read(&ioc->aic->nr_queued));
  739. atomic_dec(&ioc->aic->nr_queued);
  740. }
  741. /*
  742. * Update the "next_rq" cache if we are about to remove its
  743. * entry
  744. */
  745. if (ad->next_rq[data_dir] == rq)
  746. ad->next_rq[data_dir] = as_find_next_rq(ad, rq);
  747. rq_fifo_clear(rq);
  748. as_del_rq_rb(ad, rq);
  749. }
  750. /*
  751. * as_fifo_expired returns 0 if there are no expired reads on the fifo,
  752. * 1 otherwise. It is ratelimited so that we only perform the check once per
  753. * `fifo_expire' interval. Otherwise a large number of expired requests
  754. * would create a hopeless seekstorm.
  755. *
  756. * See as_antic_expired comment.
  757. */
  758. static int as_fifo_expired(struct as_data *ad, int adir)
  759. {
  760. struct request *rq;
  761. long delta_jif;
  762. delta_jif = jiffies - ad->last_check_fifo[adir];
  763. if (unlikely(delta_jif < 0))
  764. delta_jif = -delta_jif;
  765. if (delta_jif < ad->fifo_expire[adir])
  766. return 0;
  767. ad->last_check_fifo[adir] = jiffies;
  768. if (list_empty(&ad->fifo_list[adir]))
  769. return 0;
  770. rq = rq_entry_fifo(ad->fifo_list[adir].next);
  771. return time_after(jiffies, rq_fifo_time(rq));
  772. }
  773. /*
  774. * as_batch_expired returns true if the current batch has expired. A batch
  775. * is a set of reads or a set of writes.
  776. */
  777. static inline int as_batch_expired(struct as_data *ad)
  778. {
  779. if (ad->changed_batch || ad->new_batch)
  780. return 0;
  781. if (ad->batch_data_dir == REQ_SYNC)
  782. /* TODO! add a check so a complete fifo gets written? */
  783. return time_after(jiffies, ad->current_batch_expires);
  784. return time_after(jiffies, ad->current_batch_expires)
  785. || ad->current_write_count == 0;
  786. }
  787. /*
  788. * move an entry to dispatch queue
  789. */
  790. static void as_move_to_dispatch(struct as_data *ad, struct request *rq)
  791. {
  792. const int data_dir = rq_is_sync(rq);
  793. BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
  794. as_antic_stop(ad);
  795. ad->antic_status = ANTIC_OFF;
  796. /*
  797. * This has to be set in order to be correctly updated by
  798. * as_find_next_rq
  799. */
  800. ad->last_sector[data_dir] = rq->sector + rq->nr_sectors;
  801. if (data_dir == REQ_SYNC) {
  802. struct io_context *ioc = RQ_IOC(rq);
  803. /* In case we have to anticipate after this */
  804. copy_io_context(&ad->io_context, &ioc);
  805. } else {
  806. if (ad->io_context) {
  807. put_io_context(ad->io_context);
  808. ad->io_context = NULL;
  809. }
  810. if (ad->current_write_count != 0)
  811. ad->current_write_count--;
  812. }
  813. ad->ioc_finished = 0;
  814. ad->next_rq[data_dir] = as_find_next_rq(ad, rq);
  815. /*
  816. * take it off the sort and fifo list, add to dispatch queue
  817. */
  818. as_remove_queued_request(ad->q, rq);
  819. WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED);
  820. elv_dispatch_sort(ad->q, rq);
  821. RQ_SET_STATE(rq, AS_RQ_DISPATCHED);
  822. if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
  823. atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched);
  824. ad->nr_dispatched++;
  825. }
  826. /*
  827. * as_dispatch_request selects the best request according to
  828. * read/write expire, batch expire, etc, and moves it to the dispatch
  829. * queue. Returns 1 if a request was found, 0 otherwise.
  830. */
  831. static int as_dispatch_request(request_queue_t *q, int force)
  832. {
  833. struct as_data *ad = q->elevator->elevator_data;
  834. const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]);
  835. const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]);
  836. struct request *rq;
  837. if (unlikely(force)) {
  838. /*
  839. * Forced dispatch, accounting is useless. Reset
  840. * accounting states and dump fifo_lists. Note that
  841. * batch_data_dir is reset to REQ_SYNC to avoid
  842. * screwing write batch accounting as write batch
  843. * accounting occurs on W->R transition.
  844. */
  845. int dispatched = 0;
  846. ad->batch_data_dir = REQ_SYNC;
  847. ad->changed_batch = 0;
  848. ad->new_batch = 0;
  849. while (ad->next_rq[REQ_SYNC]) {
  850. as_move_to_dispatch(ad, ad->next_rq[REQ_SYNC]);
  851. dispatched++;
  852. }
  853. ad->last_check_fifo[REQ_SYNC] = jiffies;
  854. while (ad->next_rq[REQ_ASYNC]) {
  855. as_move_to_dispatch(ad, ad->next_rq[REQ_ASYNC]);
  856. dispatched++;
  857. }
  858. ad->last_check_fifo[REQ_ASYNC] = jiffies;
  859. return dispatched;
  860. }
  861. /* Signal that the write batch was uncontended, so we can't time it */
  862. if (ad->batch_data_dir == REQ_ASYNC && !reads) {
  863. if (ad->current_write_count == 0 || !writes)
  864. ad->write_batch_idled = 1;
  865. }
  866. if (!(reads || writes)
  867. || ad->antic_status == ANTIC_WAIT_REQ
  868. || ad->antic_status == ANTIC_WAIT_NEXT
  869. || ad->changed_batch)
  870. return 0;
  871. if (!(reads && writes && as_batch_expired(ad))) {
  872. /*
  873. * batch is still running or no reads or no writes
  874. */
  875. rq = ad->next_rq[ad->batch_data_dir];
  876. if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) {
  877. if (as_fifo_expired(ad, REQ_SYNC))
  878. goto fifo_expired;
  879. if (as_can_anticipate(ad, rq)) {
  880. as_antic_waitreq(ad);
  881. return 0;
  882. }
  883. }
  884. if (rq) {
  885. /* we have a "next request" */
  886. if (reads && !writes)
  887. ad->current_batch_expires =
  888. jiffies + ad->batch_expire[REQ_SYNC];
  889. goto dispatch_request;
  890. }
  891. }
  892. /*
  893. * at this point we are not running a batch. select the appropriate
  894. * data direction (read / write)
  895. */
  896. if (reads) {
  897. BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_SYNC]));
  898. if (writes && ad->batch_data_dir == REQ_SYNC)
  899. /*
  900. * Last batch was a read, switch to writes
  901. */
  902. goto dispatch_writes;
  903. if (ad->batch_data_dir == REQ_ASYNC) {
  904. WARN_ON(ad->new_batch);
  905. ad->changed_batch = 1;
  906. }
  907. ad->batch_data_dir = REQ_SYNC;
  908. rq = rq_entry_fifo(ad->fifo_list[REQ_SYNC].next);
  909. ad->last_check_fifo[ad->batch_data_dir] = jiffies;
  910. goto dispatch_request;
  911. }
  912. /*
  913. * the last batch was a read
  914. */
  915. if (writes) {
  916. dispatch_writes:
  917. BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_ASYNC]));
  918. if (ad->batch_data_dir == REQ_SYNC) {
  919. ad->changed_batch = 1;
  920. /*
  921. * new_batch might be 1 when the queue runs out of
  922. * reads. A subsequent submission of a write might
  923. * cause a change of batch before the read is finished.
  924. */
  925. ad->new_batch = 0;
  926. }
  927. ad->batch_data_dir = REQ_ASYNC;
  928. ad->current_write_count = ad->write_batch_count;
  929. ad->write_batch_idled = 0;
  930. rq = ad->next_rq[ad->batch_data_dir];
  931. goto dispatch_request;
  932. }
  933. BUG();
  934. return 0;
  935. dispatch_request:
  936. /*
  937. * If a request has expired, service it.
  938. */
  939. if (as_fifo_expired(ad, ad->batch_data_dir)) {
  940. fifo_expired:
  941. rq = rq_entry_fifo(ad->fifo_list[ad->batch_data_dir].next);
  942. }
  943. if (ad->changed_batch) {
  944. WARN_ON(ad->new_batch);
  945. if (ad->nr_dispatched)
  946. return 0;
  947. if (ad->batch_data_dir == REQ_ASYNC)
  948. ad->current_batch_expires = jiffies +
  949. ad->batch_expire[REQ_ASYNC];
  950. else
  951. ad->new_batch = 1;
  952. ad->changed_batch = 0;
  953. }
  954. /*
  955. * rq is the selected appropriate request.
  956. */
  957. as_move_to_dispatch(ad, rq);
  958. return 1;
  959. }
  960. /*
  961. * add rq to rbtree and fifo
  962. */
  963. static void as_add_request(request_queue_t *q, struct request *rq)
  964. {
  965. struct as_data *ad = q->elevator->elevator_data;
  966. int data_dir;
  967. RQ_SET_STATE(rq, AS_RQ_NEW);
  968. data_dir = rq_is_sync(rq);
  969. rq->elevator_private = as_get_io_context(q->node);
  970. if (RQ_IOC(rq)) {
  971. as_update_iohist(ad, RQ_IOC(rq)->aic, rq);
  972. atomic_inc(&RQ_IOC(rq)->aic->nr_queued);
  973. }
  974. as_add_rq_rb(ad, rq);
  975. /*
  976. * set expire time (only used for reads) and add to fifo list
  977. */
  978. rq_set_fifo_time(rq, jiffies + ad->fifo_expire[data_dir]);
  979. list_add_tail(&rq->queuelist, &ad->fifo_list[data_dir]);
  980. as_update_rq(ad, rq); /* keep state machine up to date */
  981. RQ_SET_STATE(rq, AS_RQ_QUEUED);
  982. }
  983. static void as_activate_request(request_queue_t *q, struct request *rq)
  984. {
  985. WARN_ON(RQ_STATE(rq) != AS_RQ_DISPATCHED);
  986. RQ_SET_STATE(rq, AS_RQ_REMOVED);
  987. if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
  988. atomic_dec(&RQ_IOC(rq)->aic->nr_dispatched);
  989. }
  990. static void as_deactivate_request(request_queue_t *q, struct request *rq)
  991. {
  992. WARN_ON(RQ_STATE(rq) != AS_RQ_REMOVED);
  993. RQ_SET_STATE(rq, AS_RQ_DISPATCHED);
  994. if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
  995. atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched);
  996. }
  997. /*
  998. * as_queue_empty tells us if there are requests left in the device. It may
  999. * not be the case that a driver can get the next request even if the queue
  1000. * is not empty - it is used in the block layer to check for plugging and
  1001. * merging opportunities
  1002. */
  1003. static int as_queue_empty(request_queue_t *q)
  1004. {
  1005. struct as_data *ad = q->elevator->elevator_data;
  1006. return list_empty(&ad->fifo_list[REQ_ASYNC])
  1007. && list_empty(&ad->fifo_list[REQ_SYNC]);
  1008. }
  1009. static int
  1010. as_merge(request_queue_t *q, struct request **req, struct bio *bio)
  1011. {
  1012. struct as_data *ad = q->elevator->elevator_data;
  1013. sector_t rb_key = bio->bi_sector + bio_sectors(bio);
  1014. struct request *__rq;
  1015. /*
  1016. * check for front merge
  1017. */
  1018. __rq = elv_rb_find(&ad->sort_list[bio_data_dir(bio)], rb_key);
  1019. if (__rq && elv_rq_merge_ok(__rq, bio)) {
  1020. *req = __rq;
  1021. return ELEVATOR_FRONT_MERGE;
  1022. }
  1023. return ELEVATOR_NO_MERGE;
  1024. }
  1025. static void as_merged_request(request_queue_t *q, struct request *req, int type)
  1026. {
  1027. struct as_data *ad = q->elevator->elevator_data;
  1028. /*
  1029. * if the merge was a front merge, we need to reposition request
  1030. */
  1031. if (type == ELEVATOR_FRONT_MERGE) {
  1032. as_del_rq_rb(ad, req);
  1033. as_add_rq_rb(ad, req);
  1034. /*
  1035. * Note! At this stage of this and the next function, our next
  1036. * request may not be optimal - eg the request may have "grown"
  1037. * behind the disk head. We currently don't bother adjusting.
  1038. */
  1039. }
  1040. }
  1041. static void as_merged_requests(request_queue_t *q, struct request *req,
  1042. struct request *next)
  1043. {
  1044. /*
  1045. * if next expires before rq, assign its expire time to arq
  1046. * and move into next position (next will be deleted) in fifo
  1047. */
  1048. if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
  1049. if (time_before(rq_fifo_time(next), rq_fifo_time(req))) {
  1050. struct io_context *rioc = RQ_IOC(req);
  1051. struct io_context *nioc = RQ_IOC(next);
  1052. list_move(&req->queuelist, &next->queuelist);
  1053. rq_set_fifo_time(req, rq_fifo_time(next));
  1054. /*
  1055. * Don't copy here but swap, because when anext is
  1056. * removed below, it must contain the unused context
  1057. */
  1058. swap_io_context(&rioc, &nioc);
  1059. }
  1060. }
  1061. /*
  1062. * kill knowledge of next, this one is a goner
  1063. */
  1064. as_remove_queued_request(q, next);
  1065. as_put_io_context(next);
  1066. RQ_SET_STATE(next, AS_RQ_MERGED);
  1067. }
  1068. /*
  1069. * This is executed in a "deferred" process context, by kblockd. It calls the
  1070. * driver's request_fn so the driver can submit that request.
  1071. *
  1072. * IMPORTANT! This guy will reenter the elevator, so set up all queue global
  1073. * state before calling, and don't rely on any state over calls.
  1074. *
  1075. * FIXME! dispatch queue is not a queue at all!
  1076. */
  1077. static void as_work_handler(struct work_struct *work)
  1078. {
  1079. struct as_data *ad = container_of(work, struct as_data, antic_work);
  1080. struct request_queue *q = ad->q;
  1081. unsigned long flags;
  1082. spin_lock_irqsave(q->queue_lock, flags);
  1083. blk_start_queueing(q);
  1084. spin_unlock_irqrestore(q->queue_lock, flags);
  1085. }
  1086. static int as_may_queue(request_queue_t *q, int rw)
  1087. {
  1088. int ret = ELV_MQUEUE_MAY;
  1089. struct as_data *ad = q->elevator->elevator_data;
  1090. struct io_context *ioc;
  1091. if (ad->antic_status == ANTIC_WAIT_REQ ||
  1092. ad->antic_status == ANTIC_WAIT_NEXT) {
  1093. ioc = as_get_io_context(q->node);
  1094. if (ad->io_context == ioc)
  1095. ret = ELV_MQUEUE_MUST;
  1096. put_io_context(ioc);
  1097. }
  1098. return ret;
  1099. }
  1100. static void as_exit_queue(elevator_t *e)
  1101. {
  1102. struct as_data *ad = e->elevator_data;
  1103. del_timer_sync(&ad->antic_timer);
  1104. kblockd_flush();
  1105. BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC]));
  1106. BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC]));
  1107. put_io_context(ad->io_context);
  1108. kfree(ad);
  1109. }
  1110. /*
  1111. * initialize elevator private data (as_data).
  1112. */
  1113. static void *as_init_queue(request_queue_t *q)
  1114. {
  1115. struct as_data *ad;
  1116. ad = kmalloc_node(sizeof(*ad), GFP_KERNEL, q->node);
  1117. if (!ad)
  1118. return NULL;
  1119. memset(ad, 0, sizeof(*ad));
  1120. ad->q = q; /* Identify what queue the data belongs to */
  1121. /* anticipatory scheduling helpers */
  1122. ad->antic_timer.function = as_antic_timeout;
  1123. ad->antic_timer.data = (unsigned long)q;
  1124. init_timer(&ad->antic_timer);
  1125. INIT_WORK(&ad->antic_work, as_work_handler);
  1126. INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]);
  1127. INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]);
  1128. ad->sort_list[REQ_SYNC] = RB_ROOT;
  1129. ad->sort_list[REQ_ASYNC] = RB_ROOT;
  1130. ad->fifo_expire[REQ_SYNC] = default_read_expire;
  1131. ad->fifo_expire[REQ_ASYNC] = default_write_expire;
  1132. ad->antic_expire = default_antic_expire;
  1133. ad->batch_expire[REQ_SYNC] = default_read_batch_expire;
  1134. ad->batch_expire[REQ_ASYNC] = default_write_batch_expire;
  1135. ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC];
  1136. ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10;
  1137. if (ad->write_batch_count < 2)
  1138. ad->write_batch_count = 2;
  1139. return ad;
  1140. }
  1141. /*
  1142. * sysfs parts below
  1143. */
  1144. static ssize_t
  1145. as_var_show(unsigned int var, char *page)
  1146. {
  1147. return sprintf(page, "%d\n", var);
  1148. }
  1149. static ssize_t
  1150. as_var_store(unsigned long *var, const char *page, size_t count)
  1151. {
  1152. char *p = (char *) page;
  1153. *var = simple_strtoul(p, &p, 10);
  1154. return count;
  1155. }
  1156. static ssize_t est_time_show(elevator_t *e, char *page)
  1157. {
  1158. struct as_data *ad = e->elevator_data;
  1159. int pos = 0;
  1160. pos += sprintf(page+pos, "%lu %% exit probability\n",
  1161. 100*ad->exit_prob/256);
  1162. pos += sprintf(page+pos, "%lu %% probability of exiting without a "
  1163. "cooperating process submitting IO\n",
  1164. 100*ad->exit_no_coop/256);
  1165. pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean);
  1166. pos += sprintf(page+pos, "%llu sectors new seek distance\n",
  1167. (unsigned long long)ad->new_seek_mean);
  1168. return pos;
  1169. }
  1170. #define SHOW_FUNCTION(__FUNC, __VAR) \
  1171. static ssize_t __FUNC(elevator_t *e, char *page) \
  1172. { \
  1173. struct as_data *ad = e->elevator_data; \
  1174. return as_var_show(jiffies_to_msecs((__VAR)), (page)); \
  1175. }
  1176. SHOW_FUNCTION(as_read_expire_show, ad->fifo_expire[REQ_SYNC]);
  1177. SHOW_FUNCTION(as_write_expire_show, ad->fifo_expire[REQ_ASYNC]);
  1178. SHOW_FUNCTION(as_antic_expire_show, ad->antic_expire);
  1179. SHOW_FUNCTION(as_read_batch_expire_show, ad->batch_expire[REQ_SYNC]);
  1180. SHOW_FUNCTION(as_write_batch_expire_show, ad->batch_expire[REQ_ASYNC]);
  1181. #undef SHOW_FUNCTION
  1182. #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
  1183. static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
  1184. { \
  1185. struct as_data *ad = e->elevator_data; \
  1186. int ret = as_var_store(__PTR, (page), count); \
  1187. if (*(__PTR) < (MIN)) \
  1188. *(__PTR) = (MIN); \
  1189. else if (*(__PTR) > (MAX)) \
  1190. *(__PTR) = (MAX); \
  1191. *(__PTR) = msecs_to_jiffies(*(__PTR)); \
  1192. return ret; \
  1193. }
  1194. STORE_FUNCTION(as_read_expire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX);
  1195. STORE_FUNCTION(as_write_expire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX);
  1196. STORE_FUNCTION(as_antic_expire_store, &ad->antic_expire, 0, INT_MAX);
  1197. STORE_FUNCTION(as_read_batch_expire_store,
  1198. &ad->batch_expire[REQ_SYNC], 0, INT_MAX);
  1199. STORE_FUNCTION(as_write_batch_expire_store,
  1200. &ad->batch_expire[REQ_ASYNC], 0, INT_MAX);
  1201. #undef STORE_FUNCTION
  1202. #define AS_ATTR(name) \
  1203. __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store)
  1204. static struct elv_fs_entry as_attrs[] = {
  1205. __ATTR_RO(est_time),
  1206. AS_ATTR(read_expire),
  1207. AS_ATTR(write_expire),
  1208. AS_ATTR(antic_expire),
  1209. AS_ATTR(read_batch_expire),
  1210. AS_ATTR(write_batch_expire),
  1211. __ATTR_NULL
  1212. };
  1213. static struct elevator_type iosched_as = {
  1214. .ops = {
  1215. .elevator_merge_fn = as_merge,
  1216. .elevator_merged_fn = as_merged_request,
  1217. .elevator_merge_req_fn = as_merged_requests,
  1218. .elevator_dispatch_fn = as_dispatch_request,
  1219. .elevator_add_req_fn = as_add_request,
  1220. .elevator_activate_req_fn = as_activate_request,
  1221. .elevator_deactivate_req_fn = as_deactivate_request,
  1222. .elevator_queue_empty_fn = as_queue_empty,
  1223. .elevator_completed_req_fn = as_completed_request,
  1224. .elevator_former_req_fn = elv_rb_former_request,
  1225. .elevator_latter_req_fn = elv_rb_latter_request,
  1226. .elevator_may_queue_fn = as_may_queue,
  1227. .elevator_init_fn = as_init_queue,
  1228. .elevator_exit_fn = as_exit_queue,
  1229. .trim = as_trim,
  1230. },
  1231. .elevator_attrs = as_attrs,
  1232. .elevator_name = "anticipatory",
  1233. .elevator_owner = THIS_MODULE,
  1234. };
  1235. static int __init as_init(void)
  1236. {
  1237. return elv_register(&iosched_as);
  1238. }
  1239. static void __exit as_exit(void)
  1240. {
  1241. DECLARE_COMPLETION_ONSTACK(all_gone);
  1242. elv_unregister(&iosched_as);
  1243. ioc_gone = &all_gone;
  1244. /* ioc_gone's update must be visible before reading ioc_count */
  1245. smp_wmb();
  1246. if (elv_ioc_count_read(ioc_count))
  1247. wait_for_completion(ioc_gone);
  1248. synchronize_rcu();
  1249. }
  1250. module_init(as_init);
  1251. module_exit(as_exit);
  1252. MODULE_AUTHOR("Nick Piggin");
  1253. MODULE_LICENSE("GPL");
  1254. MODULE_DESCRIPTION("anticipatory IO scheduler");