deadline-iosched.c 11 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469
  1. /*
  2. * Deadline i/o scheduler.
  3. *
  4. * Copyright (C) 2002 Jens Axboe <axboe@kernel.dk>
  5. */
  6. #include <linux/kernel.h>
  7. #include <linux/fs.h>
  8. #include <linux/blkdev.h>
  9. #include <linux/elevator.h>
  10. #include <linux/bio.h>
  11. #include <linux/module.h>
  12. #include <linux/slab.h>
  13. #include <linux/init.h>
  14. #include <linux/compiler.h>
  15. #include <linux/rbtree.h>
  16. /*
  17. * See Documentation/block/deadline-iosched.txt
  18. */
  19. static const int read_expire = HZ / 2; /* max time before a read is submitted. */
  20. static const int write_expire = 5 * HZ; /* ditto for writes, these limits are SOFT! */
  21. static const int writes_starved = 2; /* max times reads can starve a write */
  22. static const int fifo_batch = 16; /* # of sequential requests treated as one
  23. by the above parameters. For throughput. */
  24. struct deadline_data {
  25. /*
  26. * run time data
  27. */
  28. /*
  29. * requests (deadline_rq s) are present on both sort_list and fifo_list
  30. */
  31. struct rb_root sort_list[2];
  32. struct list_head fifo_list[2];
  33. /*
  34. * next in sort order. read, write or both are NULL
  35. */
  36. struct request *next_rq[2];
  37. unsigned int batching; /* number of sequential requests made */
  38. sector_t last_sector; /* head position */
  39. unsigned int starved; /* times reads have starved writes */
  40. /*
  41. * settings that change how the i/o scheduler behaves
  42. */
  43. int fifo_expire[2];
  44. int fifo_batch;
  45. int writes_starved;
  46. int front_merges;
  47. };
  48. static void deadline_move_request(struct deadline_data *, struct request *);
  49. static inline struct rb_root *
  50. deadline_rb_root(struct deadline_data *dd, struct request *rq)
  51. {
  52. return &dd->sort_list[rq_data_dir(rq)];
  53. }
  54. /*
  55. * get the request after `rq' in sector-sorted order
  56. */
  57. static inline struct request *
  58. deadline_latter_request(struct request *rq)
  59. {
  60. struct rb_node *node = rb_next(&rq->rb_node);
  61. if (node)
  62. return rb_entry_rq(node);
  63. return NULL;
  64. }
  65. static void
  66. deadline_add_rq_rb(struct deadline_data *dd, struct request *rq)
  67. {
  68. struct rb_root *root = deadline_rb_root(dd, rq);
  69. struct request *__alias;
  70. while (unlikely(__alias = elv_rb_add(root, rq)))
  71. deadline_move_request(dd, __alias);
  72. }
  73. static inline void
  74. deadline_del_rq_rb(struct deadline_data *dd, struct request *rq)
  75. {
  76. const int data_dir = rq_data_dir(rq);
  77. if (dd->next_rq[data_dir] == rq)
  78. dd->next_rq[data_dir] = deadline_latter_request(rq);
  79. elv_rb_del(deadline_rb_root(dd, rq), rq);
  80. }
  81. /*
  82. * add rq to rbtree and fifo
  83. */
  84. static void
  85. deadline_add_request(struct request_queue *q, struct request *rq)
  86. {
  87. struct deadline_data *dd = q->elevator->elevator_data;
  88. const int data_dir = rq_data_dir(rq);
  89. deadline_add_rq_rb(dd, rq);
  90. /*
  91. * set expire time and add to fifo list
  92. */
  93. rq_set_fifo_time(rq, jiffies + dd->fifo_expire[data_dir]);
  94. list_add_tail(&rq->queuelist, &dd->fifo_list[data_dir]);
  95. }
  96. /*
  97. * remove rq from rbtree and fifo.
  98. */
  99. static void deadline_remove_request(struct request_queue *q, struct request *rq)
  100. {
  101. struct deadline_data *dd = q->elevator->elevator_data;
  102. rq_fifo_clear(rq);
  103. deadline_del_rq_rb(dd, rq);
  104. }
  105. static int
  106. deadline_merge(struct request_queue *q, struct request **req, struct bio *bio)
  107. {
  108. struct deadline_data *dd = q->elevator->elevator_data;
  109. struct request *__rq;
  110. int ret;
  111. /*
  112. * check for front merge
  113. */
  114. if (dd->front_merges) {
  115. sector_t sector = bio->bi_sector + bio_sectors(bio);
  116. __rq = elv_rb_find(&dd->sort_list[bio_data_dir(bio)], sector);
  117. if (__rq) {
  118. BUG_ON(sector != blk_rq_pos(__rq));
  119. if (elv_rq_merge_ok(__rq, bio)) {
  120. ret = ELEVATOR_FRONT_MERGE;
  121. goto out;
  122. }
  123. }
  124. }
  125. return ELEVATOR_NO_MERGE;
  126. out:
  127. *req = __rq;
  128. return ret;
  129. }
  130. static void deadline_merged_request(struct request_queue *q,
  131. struct request *req, int type)
  132. {
  133. struct deadline_data *dd = q->elevator->elevator_data;
  134. /*
  135. * if the merge was a front merge, we need to reposition request
  136. */
  137. if (type == ELEVATOR_FRONT_MERGE) {
  138. elv_rb_del(deadline_rb_root(dd, req), req);
  139. deadline_add_rq_rb(dd, req);
  140. }
  141. }
  142. static void
  143. deadline_merged_requests(struct request_queue *q, struct request *req,
  144. struct request *next)
  145. {
  146. /*
  147. * if next expires before rq, assign its expire time to rq
  148. * and move into next position (next will be deleted) in fifo
  149. */
  150. if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
  151. if (time_before(rq_fifo_time(next), rq_fifo_time(req))) {
  152. list_move(&req->queuelist, &next->queuelist);
  153. rq_set_fifo_time(req, rq_fifo_time(next));
  154. }
  155. }
  156. /*
  157. * kill knowledge of next, this one is a goner
  158. */
  159. deadline_remove_request(q, next);
  160. }
  161. /*
  162. * move request from sort list to dispatch queue.
  163. */
  164. static inline void
  165. deadline_move_to_dispatch(struct deadline_data *dd, struct request *rq)
  166. {
  167. struct request_queue *q = rq->q;
  168. deadline_remove_request(q, rq);
  169. elv_dispatch_add_tail(q, rq);
  170. }
  171. /*
  172. * move an entry to dispatch queue
  173. */
  174. static void
  175. deadline_move_request(struct deadline_data *dd, struct request *rq)
  176. {
  177. const int data_dir = rq_data_dir(rq);
  178. dd->next_rq[READ] = NULL;
  179. dd->next_rq[WRITE] = NULL;
  180. dd->next_rq[data_dir] = deadline_latter_request(rq);
  181. dd->last_sector = rq_end_sector(rq);
  182. /*
  183. * take it off the sort and fifo list, move
  184. * to dispatch queue
  185. */
  186. deadline_move_to_dispatch(dd, rq);
  187. }
  188. /*
  189. * deadline_check_fifo returns 0 if there are no expired requests on the fifo,
  190. * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
  191. */
  192. static inline int deadline_check_fifo(struct deadline_data *dd, int ddir)
  193. {
  194. struct request *rq = rq_entry_fifo(dd->fifo_list[ddir].next);
  195. /*
  196. * rq is expired!
  197. */
  198. if (time_after(jiffies, rq_fifo_time(rq)))
  199. return 1;
  200. return 0;
  201. }
  202. /*
  203. * deadline_dispatch_requests selects the best request according to
  204. * read/write expire, fifo_batch, etc
  205. */
  206. static int deadline_dispatch_requests(struct request_queue *q, int force)
  207. {
  208. struct deadline_data *dd = q->elevator->elevator_data;
  209. const int reads = !list_empty(&dd->fifo_list[READ]);
  210. const int writes = !list_empty(&dd->fifo_list[WRITE]);
  211. struct request *rq;
  212. int data_dir;
  213. /*
  214. * batches are currently reads XOR writes
  215. */
  216. if (dd->next_rq[WRITE])
  217. rq = dd->next_rq[WRITE];
  218. else
  219. rq = dd->next_rq[READ];
  220. if (rq && dd->batching < dd->fifo_batch)
  221. /* we have a next request are still entitled to batch */
  222. goto dispatch_request;
  223. /*
  224. * at this point we are not running a batch. select the appropriate
  225. * data direction (read / write)
  226. */
  227. if (reads) {
  228. BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[READ]));
  229. if (writes && (dd->starved++ >= dd->writes_starved))
  230. goto dispatch_writes;
  231. data_dir = READ;
  232. goto dispatch_find_request;
  233. }
  234. /*
  235. * there are either no reads or writes have been starved
  236. */
  237. if (writes) {
  238. dispatch_writes:
  239. BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[WRITE]));
  240. dd->starved = 0;
  241. data_dir = WRITE;
  242. goto dispatch_find_request;
  243. }
  244. return 0;
  245. dispatch_find_request:
  246. /*
  247. * we are not running a batch, find best request for selected data_dir
  248. */
  249. if (deadline_check_fifo(dd, data_dir) || !dd->next_rq[data_dir]) {
  250. /*
  251. * A deadline has expired, the last request was in the other
  252. * direction, or we have run out of higher-sectored requests.
  253. * Start again from the request with the earliest expiry time.
  254. */
  255. rq = rq_entry_fifo(dd->fifo_list[data_dir].next);
  256. } else {
  257. /*
  258. * The last req was the same dir and we have a next request in
  259. * sort order. No expired requests so continue on from here.
  260. */
  261. rq = dd->next_rq[data_dir];
  262. }
  263. dd->batching = 0;
  264. dispatch_request:
  265. /*
  266. * rq is the selected appropriate request.
  267. */
  268. dd->batching++;
  269. deadline_move_request(dd, rq);
  270. return 1;
  271. }
  272. static void deadline_exit_queue(struct elevator_queue *e)
  273. {
  274. struct deadline_data *dd = e->elevator_data;
  275. BUG_ON(!list_empty(&dd->fifo_list[READ]));
  276. BUG_ON(!list_empty(&dd->fifo_list[WRITE]));
  277. kfree(dd);
  278. }
  279. /*
  280. * initialize elevator private data (deadline_data).
  281. */
  282. static void *deadline_init_queue(struct request_queue *q)
  283. {
  284. struct deadline_data *dd;
  285. dd = kmalloc_node(sizeof(*dd), GFP_KERNEL | __GFP_ZERO, q->node);
  286. if (!dd)
  287. return NULL;
  288. INIT_LIST_HEAD(&dd->fifo_list[READ]);
  289. INIT_LIST_HEAD(&dd->fifo_list[WRITE]);
  290. dd->sort_list[READ] = RB_ROOT;
  291. dd->sort_list[WRITE] = RB_ROOT;
  292. dd->fifo_expire[READ] = read_expire;
  293. dd->fifo_expire[WRITE] = write_expire;
  294. dd->writes_starved = writes_starved;
  295. dd->front_merges = 1;
  296. dd->fifo_batch = fifo_batch;
  297. return dd;
  298. }
  299. /*
  300. * sysfs parts below
  301. */
  302. static ssize_t
  303. deadline_var_show(int var, char *page)
  304. {
  305. return sprintf(page, "%d\n", var);
  306. }
  307. static ssize_t
  308. deadline_var_store(int *var, const char *page, size_t count)
  309. {
  310. char *p = (char *) page;
  311. *var = simple_strtol(p, &p, 10);
  312. return count;
  313. }
  314. #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
  315. static ssize_t __FUNC(struct elevator_queue *e, char *page) \
  316. { \
  317. struct deadline_data *dd = e->elevator_data; \
  318. int __data = __VAR; \
  319. if (__CONV) \
  320. __data = jiffies_to_msecs(__data); \
  321. return deadline_var_show(__data, (page)); \
  322. }
  323. SHOW_FUNCTION(deadline_read_expire_show, dd->fifo_expire[READ], 1);
  324. SHOW_FUNCTION(deadline_write_expire_show, dd->fifo_expire[WRITE], 1);
  325. SHOW_FUNCTION(deadline_writes_starved_show, dd->writes_starved, 0);
  326. SHOW_FUNCTION(deadline_front_merges_show, dd->front_merges, 0);
  327. SHOW_FUNCTION(deadline_fifo_batch_show, dd->fifo_batch, 0);
  328. #undef SHOW_FUNCTION
  329. #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
  330. static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
  331. { \
  332. struct deadline_data *dd = e->elevator_data; \
  333. int __data; \
  334. int ret = deadline_var_store(&__data, (page), count); \
  335. if (__data < (MIN)) \
  336. __data = (MIN); \
  337. else if (__data > (MAX)) \
  338. __data = (MAX); \
  339. if (__CONV) \
  340. *(__PTR) = msecs_to_jiffies(__data); \
  341. else \
  342. *(__PTR) = __data; \
  343. return ret; \
  344. }
  345. STORE_FUNCTION(deadline_read_expire_store, &dd->fifo_expire[READ], 0, INT_MAX, 1);
  346. STORE_FUNCTION(deadline_write_expire_store, &dd->fifo_expire[WRITE], 0, INT_MAX, 1);
  347. STORE_FUNCTION(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX, 0);
  348. STORE_FUNCTION(deadline_front_merges_store, &dd->front_merges, 0, 1, 0);
  349. STORE_FUNCTION(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX, 0);
  350. #undef STORE_FUNCTION
  351. #define DD_ATTR(name) \
  352. __ATTR(name, S_IRUGO|S_IWUSR, deadline_##name##_show, \
  353. deadline_##name##_store)
  354. static struct elv_fs_entry deadline_attrs[] = {
  355. DD_ATTR(read_expire),
  356. DD_ATTR(write_expire),
  357. DD_ATTR(writes_starved),
  358. DD_ATTR(front_merges),
  359. DD_ATTR(fifo_batch),
  360. __ATTR_NULL
  361. };
  362. static struct elevator_type iosched_deadline = {
  363. .ops = {
  364. .elevator_merge_fn = deadline_merge,
  365. .elevator_merged_fn = deadline_merged_request,
  366. .elevator_merge_req_fn = deadline_merged_requests,
  367. .elevator_dispatch_fn = deadline_dispatch_requests,
  368. .elevator_add_req_fn = deadline_add_request,
  369. .elevator_former_req_fn = elv_rb_former_request,
  370. .elevator_latter_req_fn = elv_rb_latter_request,
  371. .elevator_init_fn = deadline_init_queue,
  372. .elevator_exit_fn = deadline_exit_queue,
  373. },
  374. .elevator_attrs = deadline_attrs,
  375. .elevator_name = "deadline",
  376. .elevator_owner = THIS_MODULE,
  377. };
  378. static int __init deadline_init(void)
  379. {
  380. elv_register(&iosched_deadline);
  381. return 0;
  382. }
  383. static void __exit deadline_exit(void)
  384. {
  385. elv_unregister(&iosched_deadline);
  386. }
  387. module_init(deadline_init);
  388. module_exit(deadline_exit);
  389. MODULE_AUTHOR("Jens Axboe");
  390. MODULE_LICENSE("GPL");
  391. MODULE_DESCRIPTION("deadline IO scheduler");