deadline-iosched.c 11 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470
  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. unsigned int starved; /* times reads have starved writes */
  39. /*
  40. * settings that change how the i/o scheduler behaves
  41. */
  42. int fifo_expire[2];
  43. int fifo_batch;
  44. int writes_starved;
  45. int front_merges;
  46. };
  47. static void deadline_move_request(struct deadline_data *, struct request *);
  48. static inline struct rb_root *
  49. deadline_rb_root(struct deadline_data *dd, struct request *rq)
  50. {
  51. return &dd->sort_list[rq_data_dir(rq)];
  52. }
  53. /*
  54. * get the request after `rq' in sector-sorted order
  55. */
  56. static inline struct request *
  57. deadline_latter_request(struct request *rq)
  58. {
  59. struct rb_node *node = rb_next(&rq->rb_node);
  60. if (node)
  61. return rb_entry_rq(node);
  62. return NULL;
  63. }
  64. static void
  65. deadline_add_rq_rb(struct deadline_data *dd, struct request *rq)
  66. {
  67. struct rb_root *root = deadline_rb_root(dd, rq);
  68. elv_rb_add(root, rq);
  69. }
  70. static inline void
  71. deadline_del_rq_rb(struct deadline_data *dd, struct request *rq)
  72. {
  73. const int data_dir = rq_data_dir(rq);
  74. if (dd->next_rq[data_dir] == rq)
  75. dd->next_rq[data_dir] = deadline_latter_request(rq);
  76. elv_rb_del(deadline_rb_root(dd, rq), rq);
  77. }
  78. /*
  79. * add rq to rbtree and fifo
  80. */
  81. static void
  82. deadline_add_request(struct request_queue *q, struct request *rq)
  83. {
  84. struct deadline_data *dd = q->elevator->elevator_data;
  85. const int data_dir = rq_data_dir(rq);
  86. deadline_add_rq_rb(dd, rq);
  87. /*
  88. * set expire time and add to fifo list
  89. */
  90. rq->fifo_time = jiffies + dd->fifo_expire[data_dir];
  91. list_add_tail(&rq->queuelist, &dd->fifo_list[data_dir]);
  92. }
  93. /*
  94. * remove rq from rbtree and fifo.
  95. */
  96. static void deadline_remove_request(struct request_queue *q, struct request *rq)
  97. {
  98. struct deadline_data *dd = q->elevator->elevator_data;
  99. rq_fifo_clear(rq);
  100. deadline_del_rq_rb(dd, rq);
  101. }
  102. static enum elv_merge
  103. deadline_merge(struct request_queue *q, struct request **req, struct bio *bio)
  104. {
  105. struct deadline_data *dd = q->elevator->elevator_data;
  106. struct request *__rq;
  107. /*
  108. * check for front merge
  109. */
  110. if (dd->front_merges) {
  111. sector_t sector = bio_end_sector(bio);
  112. __rq = elv_rb_find(&dd->sort_list[bio_data_dir(bio)], sector);
  113. if (__rq) {
  114. BUG_ON(sector != blk_rq_pos(__rq));
  115. if (elv_bio_merge_ok(__rq, bio)) {
  116. *req = __rq;
  117. return ELEVATOR_FRONT_MERGE;
  118. }
  119. }
  120. }
  121. return ELEVATOR_NO_MERGE;
  122. }
  123. static void deadline_merged_request(struct request_queue *q,
  124. struct request *req, enum elv_merge type)
  125. {
  126. struct deadline_data *dd = q->elevator->elevator_data;
  127. /*
  128. * if the merge was a front merge, we need to reposition request
  129. */
  130. if (type == ELEVATOR_FRONT_MERGE) {
  131. elv_rb_del(deadline_rb_root(dd, req), req);
  132. deadline_add_rq_rb(dd, req);
  133. }
  134. }
  135. static void
  136. deadline_merged_requests(struct request_queue *q, struct request *req,
  137. struct request *next)
  138. {
  139. /*
  140. * if next expires before rq, assign its expire time to rq
  141. * and move into next position (next will be deleted) in fifo
  142. */
  143. if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
  144. if (time_before((unsigned long)next->fifo_time,
  145. (unsigned long)req->fifo_time)) {
  146. list_move(&req->queuelist, &next->queuelist);
  147. req->fifo_time = next->fifo_time;
  148. }
  149. }
  150. /*
  151. * kill knowledge of next, this one is a goner
  152. */
  153. deadline_remove_request(q, next);
  154. }
  155. /*
  156. * move request from sort list to dispatch queue.
  157. */
  158. static inline void
  159. deadline_move_to_dispatch(struct deadline_data *dd, struct request *rq)
  160. {
  161. struct request_queue *q = rq->q;
  162. deadline_remove_request(q, rq);
  163. elv_dispatch_add_tail(q, rq);
  164. }
  165. /*
  166. * move an entry to dispatch queue
  167. */
  168. static void
  169. deadline_move_request(struct deadline_data *dd, struct request *rq)
  170. {
  171. const int data_dir = rq_data_dir(rq);
  172. dd->next_rq[READ] = NULL;
  173. dd->next_rq[WRITE] = NULL;
  174. dd->next_rq[data_dir] = deadline_latter_request(rq);
  175. /*
  176. * take it off the sort and fifo list, move
  177. * to dispatch queue
  178. */
  179. deadline_move_to_dispatch(dd, rq);
  180. }
  181. /*
  182. * deadline_check_fifo returns 0 if there are no expired requests on the fifo,
  183. * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
  184. */
  185. static inline int deadline_check_fifo(struct deadline_data *dd, int ddir)
  186. {
  187. struct request *rq = rq_entry_fifo(dd->fifo_list[ddir].next);
  188. /*
  189. * rq is expired!
  190. */
  191. if (time_after_eq(jiffies, (unsigned long)rq->fifo_time))
  192. return 1;
  193. return 0;
  194. }
  195. /*
  196. * deadline_dispatch_requests selects the best request according to
  197. * read/write expire, fifo_batch, etc
  198. */
  199. static int deadline_dispatch_requests(struct request_queue *q, int force)
  200. {
  201. struct deadline_data *dd = q->elevator->elevator_data;
  202. const int reads = !list_empty(&dd->fifo_list[READ]);
  203. const int writes = !list_empty(&dd->fifo_list[WRITE]);
  204. struct request *rq;
  205. int data_dir;
  206. /*
  207. * batches are currently reads XOR writes
  208. */
  209. if (dd->next_rq[WRITE])
  210. rq = dd->next_rq[WRITE];
  211. else
  212. rq = dd->next_rq[READ];
  213. if (rq && dd->batching < dd->fifo_batch)
  214. /* we have a next request are still entitled to batch */
  215. goto dispatch_request;
  216. /*
  217. * at this point we are not running a batch. select the appropriate
  218. * data direction (read / write)
  219. */
  220. if (reads) {
  221. BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[READ]));
  222. if (writes && (dd->starved++ >= dd->writes_starved))
  223. goto dispatch_writes;
  224. data_dir = READ;
  225. goto dispatch_find_request;
  226. }
  227. /*
  228. * there are either no reads or writes have been starved
  229. */
  230. if (writes) {
  231. dispatch_writes:
  232. BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[WRITE]));
  233. dd->starved = 0;
  234. data_dir = WRITE;
  235. goto dispatch_find_request;
  236. }
  237. return 0;
  238. dispatch_find_request:
  239. /*
  240. * we are not running a batch, find best request for selected data_dir
  241. */
  242. if (deadline_check_fifo(dd, data_dir) || !dd->next_rq[data_dir]) {
  243. /*
  244. * A deadline has expired, the last request was in the other
  245. * direction, or we have run out of higher-sectored requests.
  246. * Start again from the request with the earliest expiry time.
  247. */
  248. rq = rq_entry_fifo(dd->fifo_list[data_dir].next);
  249. } else {
  250. /*
  251. * The last req was the same dir and we have a next request in
  252. * sort order. No expired requests so continue on from here.
  253. */
  254. rq = dd->next_rq[data_dir];
  255. }
  256. dd->batching = 0;
  257. dispatch_request:
  258. /*
  259. * rq is the selected appropriate request.
  260. */
  261. dd->batching++;
  262. deadline_move_request(dd, rq);
  263. return 1;
  264. }
  265. static void deadline_exit_queue(struct elevator_queue *e)
  266. {
  267. struct deadline_data *dd = e->elevator_data;
  268. BUG_ON(!list_empty(&dd->fifo_list[READ]));
  269. BUG_ON(!list_empty(&dd->fifo_list[WRITE]));
  270. kfree(dd);
  271. }
  272. /*
  273. * initialize elevator private data (deadline_data).
  274. */
  275. static int deadline_init_queue(struct request_queue *q, struct elevator_type *e)
  276. {
  277. struct deadline_data *dd;
  278. struct elevator_queue *eq;
  279. eq = elevator_alloc(q, e);
  280. if (!eq)
  281. return -ENOMEM;
  282. dd = kzalloc_node(sizeof(*dd), GFP_KERNEL, q->node);
  283. if (!dd) {
  284. kobject_put(&eq->kobj);
  285. return -ENOMEM;
  286. }
  287. eq->elevator_data = dd;
  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. spin_lock_irq(q->queue_lock);
  298. q->elevator = eq;
  299. spin_unlock_irq(q->queue_lock);
  300. return 0;
  301. }
  302. /*
  303. * sysfs parts below
  304. */
  305. static ssize_t
  306. deadline_var_show(int var, char *page)
  307. {
  308. return sprintf(page, "%d\n", var);
  309. }
  310. static void
  311. deadline_var_store(int *var, const char *page)
  312. {
  313. char *p = (char *) page;
  314. *var = simple_strtol(p, &p, 10);
  315. }
  316. #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
  317. static ssize_t __FUNC(struct elevator_queue *e, char *page) \
  318. { \
  319. struct deadline_data *dd = e->elevator_data; \
  320. int __data = __VAR; \
  321. if (__CONV) \
  322. __data = jiffies_to_msecs(__data); \
  323. return deadline_var_show(__data, (page)); \
  324. }
  325. SHOW_FUNCTION(deadline_read_expire_show, dd->fifo_expire[READ], 1);
  326. SHOW_FUNCTION(deadline_write_expire_show, dd->fifo_expire[WRITE], 1);
  327. SHOW_FUNCTION(deadline_writes_starved_show, dd->writes_starved, 0);
  328. SHOW_FUNCTION(deadline_front_merges_show, dd->front_merges, 0);
  329. SHOW_FUNCTION(deadline_fifo_batch_show, dd->fifo_batch, 0);
  330. #undef SHOW_FUNCTION
  331. #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
  332. static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
  333. { \
  334. struct deadline_data *dd = e->elevator_data; \
  335. int __data; \
  336. deadline_var_store(&__data, (page)); \
  337. if (__data < (MIN)) \
  338. __data = (MIN); \
  339. else if (__data > (MAX)) \
  340. __data = (MAX); \
  341. if (__CONV) \
  342. *(__PTR) = msecs_to_jiffies(__data); \
  343. else \
  344. *(__PTR) = __data; \
  345. return count; \
  346. }
  347. STORE_FUNCTION(deadline_read_expire_store, &dd->fifo_expire[READ], 0, INT_MAX, 1);
  348. STORE_FUNCTION(deadline_write_expire_store, &dd->fifo_expire[WRITE], 0, INT_MAX, 1);
  349. STORE_FUNCTION(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX, 0);
  350. STORE_FUNCTION(deadline_front_merges_store, &dd->front_merges, 0, 1, 0);
  351. STORE_FUNCTION(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX, 0);
  352. #undef STORE_FUNCTION
  353. #define DD_ATTR(name) \
  354. __ATTR(name, S_IRUGO|S_IWUSR, deadline_##name##_show, \
  355. deadline_##name##_store)
  356. static struct elv_fs_entry deadline_attrs[] = {
  357. DD_ATTR(read_expire),
  358. DD_ATTR(write_expire),
  359. DD_ATTR(writes_starved),
  360. DD_ATTR(front_merges),
  361. DD_ATTR(fifo_batch),
  362. __ATTR_NULL
  363. };
  364. static struct elevator_type iosched_deadline = {
  365. .ops.sq = {
  366. .elevator_merge_fn = deadline_merge,
  367. .elevator_merged_fn = deadline_merged_request,
  368. .elevator_merge_req_fn = deadline_merged_requests,
  369. .elevator_dispatch_fn = deadline_dispatch_requests,
  370. .elevator_add_req_fn = deadline_add_request,
  371. .elevator_former_req_fn = elv_rb_former_request,
  372. .elevator_latter_req_fn = elv_rb_latter_request,
  373. .elevator_init_fn = deadline_init_queue,
  374. .elevator_exit_fn = deadline_exit_queue,
  375. },
  376. .elevator_attrs = deadline_attrs,
  377. .elevator_name = "deadline",
  378. .elevator_owner = THIS_MODULE,
  379. };
  380. static int __init deadline_init(void)
  381. {
  382. return elv_register(&iosched_deadline);
  383. }
  384. static void __exit deadline_exit(void)
  385. {
  386. elv_unregister(&iosched_deadline);
  387. }
  388. module_init(deadline_init);
  389. module_exit(deadline_exit);
  390. MODULE_AUTHOR("Jens Axboe");
  391. MODULE_LICENSE("GPL");
  392. MODULE_DESCRIPTION("deadline IO scheduler");