Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
1da177e4 LT |
2 | * Anticipatory & deadline i/o scheduler. |
3 | * | |
4 | * Copyright (C) 2002 Jens Axboe <axboe@suse.de> | |
f5b3db00 | 5 | * Nick Piggin <nickpiggin@yahoo.com.au> |
1da177e4 LT |
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> | |
1da177e4 LT |
13 | #include <linux/module.h> |
14 | #include <linux/slab.h> | |
15 | #include <linux/init.h> | |
16 | #include <linux/compiler.h> | |
1da177e4 LT |
17 | #include <linux/rbtree.h> |
18 | #include <linux/interrupt.h> | |
19 | ||
20 | #define REQ_SYNC 1 | |
21 | #define REQ_ASYNC 0 | |
22 | ||
23 | /* | |
24 | * See Documentation/block/as-iosched.txt | |
25 | */ | |
26 | ||
27 | /* | |
28 | * max time before a read is submitted. | |
29 | */ | |
30 | #define default_read_expire (HZ / 8) | |
31 | ||
32 | /* | |
33 | * ditto for writes, these limits are not hard, even | |
34 | * if the disk is capable of satisfying them. | |
35 | */ | |
36 | #define default_write_expire (HZ / 4) | |
37 | ||
38 | /* | |
39 | * read_batch_expire describes how long we will allow a stream of reads to | |
40 | * persist before looking to see whether it is time to switch over to writes. | |
41 | */ | |
42 | #define default_read_batch_expire (HZ / 2) | |
43 | ||
44 | /* | |
45 | * write_batch_expire describes how long we want a stream of writes to run for. | |
46 | * This is not a hard limit, but a target we set for the auto-tuning thingy. | |
47 | * See, the problem is: we can send a lot of writes to disk cache / TCQ in | |
48 | * a short amount of time... | |
49 | */ | |
50 | #define default_write_batch_expire (HZ / 8) | |
51 | ||
52 | /* | |
53 | * max time we may wait to anticipate a read (default around 6ms) | |
54 | */ | |
55 | #define default_antic_expire ((HZ / 150) ? HZ / 150 : 1) | |
56 | ||
57 | /* | |
58 | * Keep track of up to 20ms thinktimes. We can go as big as we like here, | |
59 | * however huge values tend to interfere and not decay fast enough. A program | |
60 | * might be in a non-io phase of operation. Waiting on user input for example, | |
61 | * or doing a lengthy computation. A small penalty can be justified there, and | |
62 | * will still catch out those processes that constantly have large thinktimes. | |
63 | */ | |
64 | #define MAX_THINKTIME (HZ/50UL) | |
65 | ||
66 | /* Bits in as_io_context.state */ | |
67 | enum as_io_states { | |
f5b3db00 | 68 | AS_TASK_RUNNING=0, /* Process has not exited */ |
1da177e4 LT |
69 | AS_TASK_IOSTARTED, /* Process has started some IO */ |
70 | AS_TASK_IORUNNING, /* Process has completed some IO */ | |
71 | }; | |
72 | ||
73 | enum anticipation_status { | |
74 | ANTIC_OFF=0, /* Not anticipating (normal operation) */ | |
75 | ANTIC_WAIT_REQ, /* The last read has not yet completed */ | |
76 | ANTIC_WAIT_NEXT, /* Currently anticipating a request vs | |
77 | last read (which has completed) */ | |
78 | ANTIC_FINISHED, /* Anticipating but have found a candidate | |
79 | * or timed out */ | |
80 | }; | |
81 | ||
82 | struct as_data { | |
83 | /* | |
84 | * run time data | |
85 | */ | |
86 | ||
87 | struct request_queue *q; /* the "owner" queue */ | |
88 | ||
89 | /* | |
90 | * requests (as_rq s) are present on both sort_list and fifo_list | |
91 | */ | |
92 | struct rb_root sort_list[2]; | |
93 | struct list_head fifo_list[2]; | |
94 | ||
95 | struct as_rq *next_arq[2]; /* next in sort order */ | |
96 | sector_t last_sector[2]; /* last REQ_SYNC & REQ_ASYNC sectors */ | |
1da177e4 LT |
97 | |
98 | unsigned long exit_prob; /* probability a task will exit while | |
99 | being waited on */ | |
f5b3db00 NP |
100 | unsigned long exit_no_coop; /* probablility an exited task will |
101 | not be part of a later cooperating | |
102 | request */ | |
1da177e4 LT |
103 | unsigned long new_ttime_total; /* mean thinktime on new proc */ |
104 | unsigned long new_ttime_mean; | |
105 | u64 new_seek_total; /* mean seek on new proc */ | |
106 | sector_t new_seek_mean; | |
107 | ||
108 | unsigned long current_batch_expires; | |
109 | unsigned long last_check_fifo[2]; | |
110 | int changed_batch; /* 1: waiting for old batch to end */ | |
111 | int new_batch; /* 1: waiting on first read complete */ | |
112 | int batch_data_dir; /* current batch REQ_SYNC / REQ_ASYNC */ | |
113 | int write_batch_count; /* max # of reqs in a write batch */ | |
114 | int current_write_count; /* how many requests left this batch */ | |
115 | int write_batch_idled; /* has the write batch gone idle? */ | |
116 | mempool_t *arq_pool; | |
117 | ||
118 | enum anticipation_status antic_status; | |
119 | unsigned long antic_start; /* jiffies: when it started */ | |
120 | struct timer_list antic_timer; /* anticipatory scheduling timer */ | |
121 | struct work_struct antic_work; /* Deferred unplugging */ | |
122 | struct io_context *io_context; /* Identify the expected process */ | |
123 | int ioc_finished; /* IO associated with io_context is finished */ | |
124 | int nr_dispatched; | |
125 | ||
126 | /* | |
127 | * settings that change how the i/o scheduler behaves | |
128 | */ | |
129 | unsigned long fifo_expire[2]; | |
130 | unsigned long batch_expire[2]; | |
131 | unsigned long antic_expire; | |
132 | }; | |
133 | ||
134 | #define list_entry_fifo(ptr) list_entry((ptr), struct as_rq, fifo) | |
135 | ||
136 | /* | |
137 | * per-request data. | |
138 | */ | |
139 | enum arq_state { | |
140 | AS_RQ_NEW=0, /* New - not referenced and not on any lists */ | |
141 | AS_RQ_QUEUED, /* In the request queue. It belongs to the | |
142 | scheduler */ | |
143 | AS_RQ_DISPATCHED, /* On the dispatch list. It belongs to the | |
144 | driver now */ | |
145 | AS_RQ_PRESCHED, /* Debug poisoning for requests being used */ | |
146 | AS_RQ_REMOVED, | |
147 | AS_RQ_MERGED, | |
148 | AS_RQ_POSTSCHED, /* when they shouldn't be */ | |
149 | }; | |
150 | ||
151 | struct as_rq { | |
152 | /* | |
153 | * rbtree index, key is the starting offset | |
154 | */ | |
155 | struct rb_node rb_node; | |
156 | sector_t rb_key; | |
157 | ||
158 | struct request *request; | |
159 | ||
160 | struct io_context *io_context; /* The submitting task */ | |
161 | ||
1da177e4 LT |
162 | /* |
163 | * expire fifo | |
164 | */ | |
165 | struct list_head fifo; | |
166 | unsigned long expires; | |
167 | ||
168 | unsigned int is_sync; | |
169 | enum arq_state state; | |
170 | }; | |
171 | ||
172 | #define RQ_DATA(rq) ((struct as_rq *) (rq)->elevator_private) | |
173 | ||
174 | static kmem_cache_t *arq_pool; | |
175 | ||
334e94de AV |
176 | static atomic_t ioc_count = ATOMIC_INIT(0); |
177 | static struct completion *ioc_gone; | |
178 | ||
ef9be1d3 TH |
179 | static void as_move_to_dispatch(struct as_data *ad, struct as_rq *arq); |
180 | static void as_antic_stop(struct as_data *ad); | |
181 | ||
1da177e4 LT |
182 | /* |
183 | * IO Context helper functions | |
184 | */ | |
185 | ||
186 | /* Called to deallocate the as_io_context */ | |
187 | static void free_as_io_context(struct as_io_context *aic) | |
188 | { | |
189 | kfree(aic); | |
334e94de AV |
190 | if (atomic_dec_and_test(&ioc_count) && ioc_gone) |
191 | complete(ioc_gone); | |
1da177e4 LT |
192 | } |
193 | ||
e17a9489 AV |
194 | static void as_trim(struct io_context *ioc) |
195 | { | |
334e94de AV |
196 | if (ioc->aic) |
197 | free_as_io_context(ioc->aic); | |
e17a9489 AV |
198 | ioc->aic = NULL; |
199 | } | |
200 | ||
1da177e4 LT |
201 | /* Called when the task exits */ |
202 | static void exit_as_io_context(struct as_io_context *aic) | |
203 | { | |
204 | WARN_ON(!test_bit(AS_TASK_RUNNING, &aic->state)); | |
205 | clear_bit(AS_TASK_RUNNING, &aic->state); | |
206 | } | |
207 | ||
208 | static struct as_io_context *alloc_as_io_context(void) | |
209 | { | |
210 | struct as_io_context *ret; | |
211 | ||
212 | ret = kmalloc(sizeof(*ret), GFP_ATOMIC); | |
213 | if (ret) { | |
214 | ret->dtor = free_as_io_context; | |
215 | ret->exit = exit_as_io_context; | |
216 | ret->state = 1 << AS_TASK_RUNNING; | |
217 | atomic_set(&ret->nr_queued, 0); | |
218 | atomic_set(&ret->nr_dispatched, 0); | |
219 | spin_lock_init(&ret->lock); | |
220 | ret->ttime_total = 0; | |
221 | ret->ttime_samples = 0; | |
222 | ret->ttime_mean = 0; | |
223 | ret->seek_total = 0; | |
224 | ret->seek_samples = 0; | |
225 | ret->seek_mean = 0; | |
334e94de | 226 | atomic_inc(&ioc_count); |
1da177e4 LT |
227 | } |
228 | ||
229 | return ret; | |
230 | } | |
231 | ||
232 | /* | |
233 | * If the current task has no AS IO context then create one and initialise it. | |
234 | * Then take a ref on the task's io context and return it. | |
235 | */ | |
236 | static struct io_context *as_get_io_context(void) | |
237 | { | |
238 | struct io_context *ioc = get_io_context(GFP_ATOMIC); | |
239 | if (ioc && !ioc->aic) { | |
240 | ioc->aic = alloc_as_io_context(); | |
241 | if (!ioc->aic) { | |
242 | put_io_context(ioc); | |
243 | ioc = NULL; | |
244 | } | |
245 | } | |
246 | return ioc; | |
247 | } | |
248 | ||
b4878f24 JA |
249 | static void as_put_io_context(struct as_rq *arq) |
250 | { | |
251 | struct as_io_context *aic; | |
252 | ||
253 | if (unlikely(!arq->io_context)) | |
254 | return; | |
255 | ||
256 | aic = arq->io_context->aic; | |
257 | ||
258 | if (arq->is_sync == REQ_SYNC && aic) { | |
259 | spin_lock(&aic->lock); | |
260 | set_bit(AS_TASK_IORUNNING, &aic->state); | |
261 | aic->last_end_request = jiffies; | |
262 | spin_unlock(&aic->lock); | |
263 | } | |
264 | ||
265 | put_io_context(arq->io_context); | |
266 | } | |
267 | ||
1da177e4 LT |
268 | /* |
269 | * rb tree support functions | |
270 | */ | |
1da177e4 LT |
271 | #define rb_entry_arq(node) rb_entry((node), struct as_rq, rb_node) |
272 | #define ARQ_RB_ROOT(ad, arq) (&(ad)->sort_list[(arq)->is_sync]) | |
273 | #define rq_rb_key(rq) (rq)->sector | |
274 | ||
275 | /* | |
276 | * as_find_first_arq finds the first (lowest sector numbered) request | |
277 | * for the specified data_dir. Used to sweep back to the start of the disk | |
278 | * (1-way elevator) after we process the last (highest sector) request. | |
279 | */ | |
280 | static struct as_rq *as_find_first_arq(struct as_data *ad, int data_dir) | |
281 | { | |
282 | struct rb_node *n = ad->sort_list[data_dir].rb_node; | |
283 | ||
284 | if (n == NULL) | |
285 | return NULL; | |
286 | ||
287 | for (;;) { | |
288 | if (n->rb_left == NULL) | |
289 | return rb_entry_arq(n); | |
290 | ||
291 | n = n->rb_left; | |
292 | } | |
293 | } | |
294 | ||
295 | /* | |
296 | * Add the request to the rb tree if it is unique. If there is an alias (an | |
297 | * existing request against the same sector), which can happen when using | |
298 | * direct IO, then return the alias. | |
299 | */ | |
ef9be1d3 | 300 | static struct as_rq *__as_add_arq_rb(struct as_data *ad, struct as_rq *arq) |
1da177e4 LT |
301 | { |
302 | struct rb_node **p = &ARQ_RB_ROOT(ad, arq)->rb_node; | |
303 | struct rb_node *parent = NULL; | |
304 | struct as_rq *__arq; | |
305 | struct request *rq = arq->request; | |
306 | ||
307 | arq->rb_key = rq_rb_key(rq); | |
308 | ||
309 | while (*p) { | |
310 | parent = *p; | |
311 | __arq = rb_entry_arq(parent); | |
312 | ||
313 | if (arq->rb_key < __arq->rb_key) | |
314 | p = &(*p)->rb_left; | |
315 | else if (arq->rb_key > __arq->rb_key) | |
316 | p = &(*p)->rb_right; | |
317 | else | |
318 | return __arq; | |
319 | } | |
320 | ||
321 | rb_link_node(&arq->rb_node, parent, p); | |
322 | rb_insert_color(&arq->rb_node, ARQ_RB_ROOT(ad, arq)); | |
323 | ||
324 | return NULL; | |
325 | } | |
326 | ||
ef9be1d3 TH |
327 | static void as_add_arq_rb(struct as_data *ad, struct as_rq *arq) |
328 | { | |
329 | struct as_rq *alias; | |
330 | ||
331 | while ((unlikely(alias = __as_add_arq_rb(ad, arq)))) { | |
332 | as_move_to_dispatch(ad, alias); | |
333 | as_antic_stop(ad); | |
334 | } | |
335 | } | |
336 | ||
1da177e4 LT |
337 | static inline void as_del_arq_rb(struct as_data *ad, struct as_rq *arq) |
338 | { | |
dd67d051 | 339 | if (!RB_EMPTY_NODE(&arq->rb_node)) { |
1da177e4 LT |
340 | WARN_ON(1); |
341 | return; | |
342 | } | |
343 | ||
344 | rb_erase(&arq->rb_node, ARQ_RB_ROOT(ad, arq)); | |
dd67d051 | 345 | RB_CLEAR_NODE(&arq->rb_node); |
1da177e4 LT |
346 | } |
347 | ||
348 | static struct request * | |
349 | as_find_arq_rb(struct as_data *ad, sector_t sector, int data_dir) | |
350 | { | |
351 | struct rb_node *n = ad->sort_list[data_dir].rb_node; | |
352 | struct as_rq *arq; | |
353 | ||
354 | while (n) { | |
355 | arq = rb_entry_arq(n); | |
356 | ||
357 | if (sector < arq->rb_key) | |
358 | n = n->rb_left; | |
359 | else if (sector > arq->rb_key) | |
360 | n = n->rb_right; | |
361 | else | |
362 | return arq->request; | |
363 | } | |
364 | ||
365 | return NULL; | |
366 | } | |
367 | ||
368 | /* | |
369 | * IO Scheduler proper | |
370 | */ | |
371 | ||
372 | #define MAXBACK (1024 * 1024) /* | |
373 | * Maximum distance the disk will go backward | |
374 | * for a request. | |
375 | */ | |
376 | ||
377 | #define BACK_PENALTY 2 | |
378 | ||
379 | /* | |
380 | * as_choose_req selects the preferred one of two requests of the same data_dir | |
381 | * ignoring time - eg. timeouts, which is the job of as_dispatch_request | |
382 | */ | |
383 | static struct as_rq * | |
384 | as_choose_req(struct as_data *ad, struct as_rq *arq1, struct as_rq *arq2) | |
385 | { | |
386 | int data_dir; | |
387 | sector_t last, s1, s2, d1, d2; | |
388 | int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */ | |
389 | const sector_t maxback = MAXBACK; | |
390 | ||
391 | if (arq1 == NULL || arq1 == arq2) | |
392 | return arq2; | |
393 | if (arq2 == NULL) | |
394 | return arq1; | |
395 | ||
396 | data_dir = arq1->is_sync; | |
397 | ||
398 | last = ad->last_sector[data_dir]; | |
399 | s1 = arq1->request->sector; | |
400 | s2 = arq2->request->sector; | |
401 | ||
402 | BUG_ON(data_dir != arq2->is_sync); | |
403 | ||
404 | /* | |
405 | * Strict one way elevator _except_ in the case where we allow | |
406 | * short backward seeks which are biased as twice the cost of a | |
407 | * similar forward seek. | |
408 | */ | |
409 | if (s1 >= last) | |
410 | d1 = s1 - last; | |
411 | else if (s1+maxback >= last) | |
412 | d1 = (last - s1)*BACK_PENALTY; | |
413 | else { | |
414 | r1_wrap = 1; | |
415 | d1 = 0; /* shut up, gcc */ | |
416 | } | |
417 | ||
418 | if (s2 >= last) | |
419 | d2 = s2 - last; | |
420 | else if (s2+maxback >= last) | |
421 | d2 = (last - s2)*BACK_PENALTY; | |
422 | else { | |
423 | r2_wrap = 1; | |
424 | d2 = 0; | |
425 | } | |
426 | ||
427 | /* Found required data */ | |
428 | if (!r1_wrap && r2_wrap) | |
429 | return arq1; | |
430 | else if (!r2_wrap && r1_wrap) | |
431 | return arq2; | |
432 | else if (r1_wrap && r2_wrap) { | |
433 | /* both behind the head */ | |
434 | if (s1 <= s2) | |
435 | return arq1; | |
436 | else | |
437 | return arq2; | |
438 | } | |
439 | ||
440 | /* Both requests in front of the head */ | |
441 | if (d1 < d2) | |
442 | return arq1; | |
443 | else if (d2 < d1) | |
444 | return arq2; | |
445 | else { | |
446 | if (s1 >= s2) | |
447 | return arq1; | |
448 | else | |
449 | return arq2; | |
450 | } | |
451 | } | |
452 | ||
453 | /* | |
454 | * as_find_next_arq finds the next request after @prev in elevator order. | |
455 | * this with as_choose_req form the basis for how the scheduler chooses | |
456 | * what request to process next. Anticipation works on top of this. | |
457 | */ | |
458 | static struct as_rq *as_find_next_arq(struct as_data *ad, struct as_rq *last) | |
459 | { | |
460 | const int data_dir = last->is_sync; | |
461 | struct as_rq *ret; | |
462 | struct rb_node *rbnext = rb_next(&last->rb_node); | |
463 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
464 | struct as_rq *arq_next, *arq_prev; | |
465 | ||
dd67d051 | 466 | BUG_ON(!RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
467 | |
468 | if (rbprev) | |
469 | arq_prev = rb_entry_arq(rbprev); | |
470 | else | |
471 | arq_prev = NULL; | |
472 | ||
473 | if (rbnext) | |
474 | arq_next = rb_entry_arq(rbnext); | |
475 | else { | |
476 | arq_next = as_find_first_arq(ad, data_dir); | |
477 | if (arq_next == last) | |
478 | arq_next = NULL; | |
479 | } | |
480 | ||
481 | ret = as_choose_req(ad, arq_next, arq_prev); | |
482 | ||
483 | return ret; | |
484 | } | |
485 | ||
486 | /* | |
487 | * anticipatory scheduling functions follow | |
488 | */ | |
489 | ||
490 | /* | |
491 | * as_antic_expired tells us when we have anticipated too long. | |
492 | * The funny "absolute difference" math on the elapsed time is to handle | |
493 | * jiffy wraps, and disks which have been idle for 0x80000000 jiffies. | |
494 | */ | |
495 | static int as_antic_expired(struct as_data *ad) | |
496 | { | |
497 | long delta_jif; | |
498 | ||
499 | delta_jif = jiffies - ad->antic_start; | |
500 | if (unlikely(delta_jif < 0)) | |
501 | delta_jif = -delta_jif; | |
502 | if (delta_jif < ad->antic_expire) | |
503 | return 0; | |
504 | ||
505 | return 1; | |
506 | } | |
507 | ||
508 | /* | |
509 | * as_antic_waitnext starts anticipating that a nice request will soon be | |
510 | * submitted. See also as_antic_waitreq | |
511 | */ | |
512 | static void as_antic_waitnext(struct as_data *ad) | |
513 | { | |
514 | unsigned long timeout; | |
515 | ||
516 | BUG_ON(ad->antic_status != ANTIC_OFF | |
517 | && ad->antic_status != ANTIC_WAIT_REQ); | |
518 | ||
519 | timeout = ad->antic_start + ad->antic_expire; | |
520 | ||
521 | mod_timer(&ad->antic_timer, timeout); | |
522 | ||
523 | ad->antic_status = ANTIC_WAIT_NEXT; | |
524 | } | |
525 | ||
526 | /* | |
527 | * as_antic_waitreq starts anticipating. We don't start timing the anticipation | |
528 | * until the request that we're anticipating on has finished. This means we | |
529 | * are timing from when the candidate process wakes up hopefully. | |
530 | */ | |
531 | static void as_antic_waitreq(struct as_data *ad) | |
532 | { | |
533 | BUG_ON(ad->antic_status == ANTIC_FINISHED); | |
534 | if (ad->antic_status == ANTIC_OFF) { | |
535 | if (!ad->io_context || ad->ioc_finished) | |
536 | as_antic_waitnext(ad); | |
537 | else | |
538 | ad->antic_status = ANTIC_WAIT_REQ; | |
539 | } | |
540 | } | |
541 | ||
542 | /* | |
543 | * This is called directly by the functions in this file to stop anticipation. | |
544 | * We kill the timer and schedule a call to the request_fn asap. | |
545 | */ | |
546 | static void as_antic_stop(struct as_data *ad) | |
547 | { | |
548 | int status = ad->antic_status; | |
549 | ||
550 | if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) { | |
551 | if (status == ANTIC_WAIT_NEXT) | |
552 | del_timer(&ad->antic_timer); | |
553 | ad->antic_status = ANTIC_FINISHED; | |
554 | /* see as_work_handler */ | |
555 | kblockd_schedule_work(&ad->antic_work); | |
556 | } | |
557 | } | |
558 | ||
559 | /* | |
560 | * as_antic_timeout is the timer function set by as_antic_waitnext. | |
561 | */ | |
562 | static void as_antic_timeout(unsigned long data) | |
563 | { | |
564 | struct request_queue *q = (struct request_queue *)data; | |
565 | struct as_data *ad = q->elevator->elevator_data; | |
566 | unsigned long flags; | |
567 | ||
568 | spin_lock_irqsave(q->queue_lock, flags); | |
569 | if (ad->antic_status == ANTIC_WAIT_REQ | |
570 | || ad->antic_status == ANTIC_WAIT_NEXT) { | |
571 | struct as_io_context *aic = ad->io_context->aic; | |
572 | ||
573 | ad->antic_status = ANTIC_FINISHED; | |
574 | kblockd_schedule_work(&ad->antic_work); | |
575 | ||
576 | if (aic->ttime_samples == 0) { | |
f5b3db00 | 577 | /* process anticipated on has exited or timed out*/ |
1da177e4 LT |
578 | ad->exit_prob = (7*ad->exit_prob + 256)/8; |
579 | } | |
f5b3db00 NP |
580 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
581 | /* process not "saved" by a cooperating request */ | |
582 | ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8; | |
583 | } | |
1da177e4 LT |
584 | } |
585 | spin_unlock_irqrestore(q->queue_lock, flags); | |
586 | } | |
587 | ||
f5b3db00 NP |
588 | static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic, |
589 | unsigned long ttime) | |
590 | { | |
591 | /* fixed point: 1.0 == 1<<8 */ | |
592 | if (aic->ttime_samples == 0) { | |
593 | ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8; | |
594 | ad->new_ttime_mean = ad->new_ttime_total / 256; | |
595 | ||
596 | ad->exit_prob = (7*ad->exit_prob)/8; | |
597 | } | |
598 | aic->ttime_samples = (7*aic->ttime_samples + 256) / 8; | |
599 | aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8; | |
600 | aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples; | |
601 | } | |
602 | ||
603 | static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic, | |
604 | sector_t sdist) | |
605 | { | |
606 | u64 total; | |
607 | ||
608 | if (aic->seek_samples == 0) { | |
609 | ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8; | |
610 | ad->new_seek_mean = ad->new_seek_total / 256; | |
611 | } | |
612 | ||
613 | /* | |
614 | * Don't allow the seek distance to get too large from the | |
615 | * odd fragment, pagein, etc | |
616 | */ | |
617 | if (aic->seek_samples <= 60) /* second&third seek */ | |
618 | sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024); | |
619 | else | |
620 | sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64); | |
621 | ||
622 | aic->seek_samples = (7*aic->seek_samples + 256) / 8; | |
623 | aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8; | |
624 | total = aic->seek_total + (aic->seek_samples/2); | |
625 | do_div(total, aic->seek_samples); | |
626 | aic->seek_mean = (sector_t)total; | |
627 | } | |
628 | ||
629 | /* | |
630 | * as_update_iohist keeps a decaying histogram of IO thinktimes, and | |
631 | * updates @aic->ttime_mean based on that. It is called when a new | |
632 | * request is queued. | |
633 | */ | |
634 | static void as_update_iohist(struct as_data *ad, struct as_io_context *aic, | |
635 | struct request *rq) | |
636 | { | |
637 | struct as_rq *arq = RQ_DATA(rq); | |
638 | int data_dir = arq->is_sync; | |
639 | unsigned long thinktime = 0; | |
640 | sector_t seek_dist; | |
641 | ||
642 | if (aic == NULL) | |
643 | return; | |
644 | ||
645 | if (data_dir == REQ_SYNC) { | |
646 | unsigned long in_flight = atomic_read(&aic->nr_queued) | |
647 | + atomic_read(&aic->nr_dispatched); | |
648 | spin_lock(&aic->lock); | |
649 | if (test_bit(AS_TASK_IORUNNING, &aic->state) || | |
650 | test_bit(AS_TASK_IOSTARTED, &aic->state)) { | |
651 | /* Calculate read -> read thinktime */ | |
652 | if (test_bit(AS_TASK_IORUNNING, &aic->state) | |
653 | && in_flight == 0) { | |
654 | thinktime = jiffies - aic->last_end_request; | |
655 | thinktime = min(thinktime, MAX_THINKTIME-1); | |
656 | } | |
657 | as_update_thinktime(ad, aic, thinktime); | |
658 | ||
659 | /* Calculate read -> read seek distance */ | |
660 | if (aic->last_request_pos < rq->sector) | |
661 | seek_dist = rq->sector - aic->last_request_pos; | |
662 | else | |
663 | seek_dist = aic->last_request_pos - rq->sector; | |
664 | as_update_seekdist(ad, aic, seek_dist); | |
665 | } | |
666 | aic->last_request_pos = rq->sector + rq->nr_sectors; | |
667 | set_bit(AS_TASK_IOSTARTED, &aic->state); | |
668 | spin_unlock(&aic->lock); | |
669 | } | |
670 | } | |
671 | ||
1da177e4 LT |
672 | /* |
673 | * as_close_req decides if one request is considered "close" to the | |
674 | * previous one issued. | |
675 | */ | |
f5b3db00 NP |
676 | static int as_close_req(struct as_data *ad, struct as_io_context *aic, |
677 | struct as_rq *arq) | |
1da177e4 LT |
678 | { |
679 | unsigned long delay; /* milliseconds */ | |
680 | sector_t last = ad->last_sector[ad->batch_data_dir]; | |
681 | sector_t next = arq->request->sector; | |
682 | sector_t delta; /* acceptable close offset (in sectors) */ | |
f5b3db00 | 683 | sector_t s; |
1da177e4 LT |
684 | |
685 | if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished) | |
686 | delay = 0; | |
687 | else | |
688 | delay = ((jiffies - ad->antic_start) * 1000) / HZ; | |
689 | ||
f5b3db00 NP |
690 | if (delay == 0) |
691 | delta = 8192; | |
1da177e4 | 692 | else if (delay <= 20 && delay <= ad->antic_expire) |
f5b3db00 | 693 | delta = 8192 << delay; |
1da177e4 LT |
694 | else |
695 | return 1; | |
696 | ||
f5b3db00 NP |
697 | if ((last <= next + (delta>>1)) && (next <= last + delta)) |
698 | return 1; | |
699 | ||
700 | if (last < next) | |
701 | s = next - last; | |
702 | else | |
703 | s = last - next; | |
704 | ||
705 | if (aic->seek_samples == 0) { | |
706 | /* | |
707 | * Process has just started IO. Use past statistics to | |
708 | * gauge success possibility | |
709 | */ | |
710 | if (ad->new_seek_mean > s) { | |
711 | /* this request is better than what we're expecting */ | |
712 | return 1; | |
713 | } | |
714 | ||
715 | } else { | |
716 | if (aic->seek_mean > s) { | |
717 | /* this request is better than what we're expecting */ | |
718 | return 1; | |
719 | } | |
720 | } | |
721 | ||
722 | return 0; | |
1da177e4 LT |
723 | } |
724 | ||
725 | /* | |
726 | * as_can_break_anticipation returns true if we have been anticipating this | |
727 | * request. | |
728 | * | |
729 | * It also returns true if the process against which we are anticipating | |
730 | * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to | |
731 | * dispatch it ASAP, because we know that application will not be submitting | |
732 | * any new reads. | |
733 | * | |
f5b3db00 | 734 | * If the task which has submitted the request has exited, break anticipation. |
1da177e4 LT |
735 | * |
736 | * If this task has queued some other IO, do not enter enticipation. | |
737 | */ | |
738 | static int as_can_break_anticipation(struct as_data *ad, struct as_rq *arq) | |
739 | { | |
740 | struct io_context *ioc; | |
741 | struct as_io_context *aic; | |
1da177e4 LT |
742 | |
743 | ioc = ad->io_context; | |
744 | BUG_ON(!ioc); | |
745 | ||
746 | if (arq && ioc == arq->io_context) { | |
747 | /* request from same process */ | |
748 | return 1; | |
749 | } | |
750 | ||
751 | if (ad->ioc_finished && as_antic_expired(ad)) { | |
752 | /* | |
753 | * In this situation status should really be FINISHED, | |
754 | * however the timer hasn't had the chance to run yet. | |
755 | */ | |
756 | return 1; | |
757 | } | |
758 | ||
759 | aic = ioc->aic; | |
760 | if (!aic) | |
761 | return 0; | |
762 | ||
1da177e4 LT |
763 | if (atomic_read(&aic->nr_queued) > 0) { |
764 | /* process has more requests queued */ | |
765 | return 1; | |
766 | } | |
767 | ||
768 | if (atomic_read(&aic->nr_dispatched) > 0) { | |
769 | /* process has more requests dispatched */ | |
770 | return 1; | |
771 | } | |
772 | ||
f5b3db00 | 773 | if (arq && arq->is_sync == REQ_SYNC && as_close_req(ad, aic, arq)) { |
1da177e4 LT |
774 | /* |
775 | * Found a close request that is not one of ours. | |
776 | * | |
f5b3db00 NP |
777 | * This makes close requests from another process update |
778 | * our IO history. Is generally useful when there are | |
1da177e4 LT |
779 | * two or more cooperating processes working in the same |
780 | * area. | |
781 | */ | |
f5b3db00 NP |
782 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
783 | if (aic->ttime_samples == 0) | |
784 | ad->exit_prob = (7*ad->exit_prob + 256)/8; | |
785 | ||
786 | ad->exit_no_coop = (7*ad->exit_no_coop)/8; | |
787 | } | |
788 | ||
789 | as_update_iohist(ad, aic, arq->request); | |
1da177e4 LT |
790 | return 1; |
791 | } | |
792 | ||
f5b3db00 NP |
793 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
794 | /* process anticipated on has exited */ | |
795 | if (aic->ttime_samples == 0) | |
796 | ad->exit_prob = (7*ad->exit_prob + 256)/8; | |
797 | ||
798 | if (ad->exit_no_coop > 128) | |
799 | return 1; | |
800 | } | |
1da177e4 LT |
801 | |
802 | if (aic->ttime_samples == 0) { | |
803 | if (ad->new_ttime_mean > ad->antic_expire) | |
804 | return 1; | |
f5b3db00 | 805 | if (ad->exit_prob * ad->exit_no_coop > 128*256) |
1da177e4 LT |
806 | return 1; |
807 | } else if (aic->ttime_mean > ad->antic_expire) { | |
808 | /* the process thinks too much between requests */ | |
809 | return 1; | |
810 | } | |
811 | ||
1da177e4 LT |
812 | return 0; |
813 | } | |
814 | ||
815 | /* | |
d6e05edc | 816 | * as_can_anticipate indicates whether we should either run arq |
1da177e4 LT |
817 | * or keep anticipating a better request. |
818 | */ | |
819 | static int as_can_anticipate(struct as_data *ad, struct as_rq *arq) | |
820 | { | |
821 | if (!ad->io_context) | |
822 | /* | |
823 | * Last request submitted was a write | |
824 | */ | |
825 | return 0; | |
826 | ||
827 | if (ad->antic_status == ANTIC_FINISHED) | |
828 | /* | |
829 | * Don't restart if we have just finished. Run the next request | |
830 | */ | |
831 | return 0; | |
832 | ||
833 | if (as_can_break_anticipation(ad, arq)) | |
834 | /* | |
835 | * This request is a good candidate. Don't keep anticipating, | |
836 | * run it. | |
837 | */ | |
838 | return 0; | |
839 | ||
840 | /* | |
841 | * OK from here, we haven't finished, and don't have a decent request! | |
842 | * Status is either ANTIC_OFF so start waiting, | |
843 | * ANTIC_WAIT_REQ so continue waiting for request to finish | |
844 | * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request. | |
1da177e4 LT |
845 | */ |
846 | ||
847 | return 1; | |
848 | } | |
849 | ||
1da177e4 LT |
850 | /* |
851 | * as_update_arq must be called whenever a request (arq) is added to | |
852 | * the sort_list. This function keeps caches up to date, and checks if the | |
853 | * request might be one we are "anticipating" | |
854 | */ | |
855 | static void as_update_arq(struct as_data *ad, struct as_rq *arq) | |
856 | { | |
857 | const int data_dir = arq->is_sync; | |
858 | ||
859 | /* keep the next_arq cache up to date */ | |
860 | ad->next_arq[data_dir] = as_choose_req(ad, arq, ad->next_arq[data_dir]); | |
861 | ||
862 | /* | |
863 | * have we been anticipating this request? | |
864 | * or does it come from the same process as the one we are anticipating | |
865 | * for? | |
866 | */ | |
867 | if (ad->antic_status == ANTIC_WAIT_REQ | |
868 | || ad->antic_status == ANTIC_WAIT_NEXT) { | |
869 | if (as_can_break_anticipation(ad, arq)) | |
870 | as_antic_stop(ad); | |
871 | } | |
872 | } | |
873 | ||
874 | /* | |
875 | * Gathers timings and resizes the write batch automatically | |
876 | */ | |
877 | static void update_write_batch(struct as_data *ad) | |
878 | { | |
879 | unsigned long batch = ad->batch_expire[REQ_ASYNC]; | |
880 | long write_time; | |
881 | ||
882 | write_time = (jiffies - ad->current_batch_expires) + batch; | |
883 | if (write_time < 0) | |
884 | write_time = 0; | |
885 | ||
886 | if (write_time > batch && !ad->write_batch_idled) { | |
887 | if (write_time > batch * 3) | |
888 | ad->write_batch_count /= 2; | |
889 | else | |
890 | ad->write_batch_count--; | |
891 | } else if (write_time < batch && ad->current_write_count == 0) { | |
892 | if (batch > write_time * 3) | |
893 | ad->write_batch_count *= 2; | |
894 | else | |
895 | ad->write_batch_count++; | |
896 | } | |
897 | ||
898 | if (ad->write_batch_count < 1) | |
899 | ad->write_batch_count = 1; | |
900 | } | |
901 | ||
902 | /* | |
903 | * as_completed_request is to be called when a request has completed and | |
904 | * returned something to the requesting process, be it an error or data. | |
905 | */ | |
906 | static void as_completed_request(request_queue_t *q, struct request *rq) | |
907 | { | |
908 | struct as_data *ad = q->elevator->elevator_data; | |
909 | struct as_rq *arq = RQ_DATA(rq); | |
910 | ||
911 | WARN_ON(!list_empty(&rq->queuelist)); | |
912 | ||
1da177e4 LT |
913 | if (arq->state != AS_RQ_REMOVED) { |
914 | printk("arq->state %d\n", arq->state); | |
915 | WARN_ON(1); | |
916 | goto out; | |
917 | } | |
918 | ||
1da177e4 LT |
919 | if (ad->changed_batch && ad->nr_dispatched == 1) { |
920 | kblockd_schedule_work(&ad->antic_work); | |
921 | ad->changed_batch = 0; | |
922 | ||
923 | if (ad->batch_data_dir == REQ_SYNC) | |
924 | ad->new_batch = 1; | |
925 | } | |
926 | WARN_ON(ad->nr_dispatched == 0); | |
927 | ad->nr_dispatched--; | |
928 | ||
929 | /* | |
930 | * Start counting the batch from when a request of that direction is | |
931 | * actually serviced. This should help devices with big TCQ windows | |
932 | * and writeback caches | |
933 | */ | |
934 | if (ad->new_batch && ad->batch_data_dir == arq->is_sync) { | |
935 | update_write_batch(ad); | |
936 | ad->current_batch_expires = jiffies + | |
937 | ad->batch_expire[REQ_SYNC]; | |
938 | ad->new_batch = 0; | |
939 | } | |
940 | ||
941 | if (ad->io_context == arq->io_context && ad->io_context) { | |
942 | ad->antic_start = jiffies; | |
943 | ad->ioc_finished = 1; | |
944 | if (ad->antic_status == ANTIC_WAIT_REQ) { | |
945 | /* | |
946 | * We were waiting on this request, now anticipate | |
947 | * the next one | |
948 | */ | |
949 | as_antic_waitnext(ad); | |
950 | } | |
951 | } | |
952 | ||
b4878f24 | 953 | as_put_io_context(arq); |
1da177e4 LT |
954 | out: |
955 | arq->state = AS_RQ_POSTSCHED; | |
956 | } | |
957 | ||
958 | /* | |
959 | * as_remove_queued_request removes a request from the pre dispatch queue | |
960 | * without updating refcounts. It is expected the caller will drop the | |
961 | * reference unless it replaces the request at somepart of the elevator | |
962 | * (ie. the dispatch queue) | |
963 | */ | |
964 | static void as_remove_queued_request(request_queue_t *q, struct request *rq) | |
965 | { | |
966 | struct as_rq *arq = RQ_DATA(rq); | |
967 | const int data_dir = arq->is_sync; | |
968 | struct as_data *ad = q->elevator->elevator_data; | |
969 | ||
970 | WARN_ON(arq->state != AS_RQ_QUEUED); | |
971 | ||
972 | if (arq->io_context && arq->io_context->aic) { | |
973 | BUG_ON(!atomic_read(&arq->io_context->aic->nr_queued)); | |
974 | atomic_dec(&arq->io_context->aic->nr_queued); | |
975 | } | |
976 | ||
977 | /* | |
978 | * Update the "next_arq" cache if we are about to remove its | |
979 | * entry | |
980 | */ | |
981 | if (ad->next_arq[data_dir] == arq) | |
982 | ad->next_arq[data_dir] = as_find_next_arq(ad, arq); | |
983 | ||
984 | list_del_init(&arq->fifo); | |
1da177e4 LT |
985 | as_del_arq_rb(ad, arq); |
986 | } | |
987 | ||
1da177e4 LT |
988 | /* |
989 | * as_fifo_expired returns 0 if there are no expired reads on the fifo, | |
990 | * 1 otherwise. It is ratelimited so that we only perform the check once per | |
991 | * `fifo_expire' interval. Otherwise a large number of expired requests | |
992 | * would create a hopeless seekstorm. | |
993 | * | |
994 | * See as_antic_expired comment. | |
995 | */ | |
996 | static int as_fifo_expired(struct as_data *ad, int adir) | |
997 | { | |
998 | struct as_rq *arq; | |
999 | long delta_jif; | |
1000 | ||
1001 | delta_jif = jiffies - ad->last_check_fifo[adir]; | |
1002 | if (unlikely(delta_jif < 0)) | |
1003 | delta_jif = -delta_jif; | |
1004 | if (delta_jif < ad->fifo_expire[adir]) | |
1005 | return 0; | |
1006 | ||
1007 | ad->last_check_fifo[adir] = jiffies; | |
1008 | ||
1009 | if (list_empty(&ad->fifo_list[adir])) | |
1010 | return 0; | |
1011 | ||
1012 | arq = list_entry_fifo(ad->fifo_list[adir].next); | |
1013 | ||
1014 | return time_after(jiffies, arq->expires); | |
1015 | } | |
1016 | ||
1017 | /* | |
1018 | * as_batch_expired returns true if the current batch has expired. A batch | |
1019 | * is a set of reads or a set of writes. | |
1020 | */ | |
1021 | static inline int as_batch_expired(struct as_data *ad) | |
1022 | { | |
1023 | if (ad->changed_batch || ad->new_batch) | |
1024 | return 0; | |
1025 | ||
1026 | if (ad->batch_data_dir == REQ_SYNC) | |
1027 | /* TODO! add a check so a complete fifo gets written? */ | |
1028 | return time_after(jiffies, ad->current_batch_expires); | |
1029 | ||
1030 | return time_after(jiffies, ad->current_batch_expires) | |
1031 | || ad->current_write_count == 0; | |
1032 | } | |
1033 | ||
1034 | /* | |
1035 | * move an entry to dispatch queue | |
1036 | */ | |
1037 | static void as_move_to_dispatch(struct as_data *ad, struct as_rq *arq) | |
1038 | { | |
1039 | struct request *rq = arq->request; | |
1da177e4 LT |
1040 | const int data_dir = arq->is_sync; |
1041 | ||
dd67d051 | 1042 | BUG_ON(!RB_EMPTY_NODE(&arq->rb_node)); |
1da177e4 LT |
1043 | |
1044 | as_antic_stop(ad); | |
1045 | ad->antic_status = ANTIC_OFF; | |
1046 | ||
1047 | /* | |
1048 | * This has to be set in order to be correctly updated by | |
1049 | * as_find_next_arq | |
1050 | */ | |
1051 | ad->last_sector[data_dir] = rq->sector + rq->nr_sectors; | |
1052 | ||
1053 | if (data_dir == REQ_SYNC) { | |
1054 | /* In case we have to anticipate after this */ | |
1055 | copy_io_context(&ad->io_context, &arq->io_context); | |
1056 | } else { | |
1057 | if (ad->io_context) { | |
1058 | put_io_context(ad->io_context); | |
1059 | ad->io_context = NULL; | |
1060 | } | |
1061 | ||
1062 | if (ad->current_write_count != 0) | |
1063 | ad->current_write_count--; | |
1064 | } | |
1065 | ad->ioc_finished = 0; | |
1066 | ||
1067 | ad->next_arq[data_dir] = as_find_next_arq(ad, arq); | |
1068 | ||
1069 | /* | |
1070 | * take it off the sort and fifo list, add to dispatch queue | |
1071 | */ | |
1da177e4 LT |
1072 | as_remove_queued_request(ad->q, rq); |
1073 | WARN_ON(arq->state != AS_RQ_QUEUED); | |
1074 | ||
b4878f24 JA |
1075 | elv_dispatch_sort(ad->q, rq); |
1076 | ||
1da177e4 LT |
1077 | arq->state = AS_RQ_DISPATCHED; |
1078 | if (arq->io_context && arq->io_context->aic) | |
1079 | atomic_inc(&arq->io_context->aic->nr_dispatched); | |
1080 | ad->nr_dispatched++; | |
1081 | } | |
1082 | ||
1083 | /* | |
1084 | * as_dispatch_request selects the best request according to | |
1085 | * read/write expire, batch expire, etc, and moves it to the dispatch | |
1086 | * queue. Returns 1 if a request was found, 0 otherwise. | |
1087 | */ | |
b4878f24 | 1088 | static int as_dispatch_request(request_queue_t *q, int force) |
1da177e4 | 1089 | { |
b4878f24 | 1090 | struct as_data *ad = q->elevator->elevator_data; |
1da177e4 LT |
1091 | struct as_rq *arq; |
1092 | const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]); | |
1093 | const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]); | |
1094 | ||
b4878f24 JA |
1095 | if (unlikely(force)) { |
1096 | /* | |
1097 | * Forced dispatch, accounting is useless. Reset | |
1098 | * accounting states and dump fifo_lists. Note that | |
1099 | * batch_data_dir is reset to REQ_SYNC to avoid | |
1100 | * screwing write batch accounting as write batch | |
1101 | * accounting occurs on W->R transition. | |
1102 | */ | |
1103 | int dispatched = 0; | |
1104 | ||
1105 | ad->batch_data_dir = REQ_SYNC; | |
1106 | ad->changed_batch = 0; | |
1107 | ad->new_batch = 0; | |
1108 | ||
1109 | while (ad->next_arq[REQ_SYNC]) { | |
1110 | as_move_to_dispatch(ad, ad->next_arq[REQ_SYNC]); | |
1111 | dispatched++; | |
1112 | } | |
1113 | ad->last_check_fifo[REQ_SYNC] = jiffies; | |
1114 | ||
1115 | while (ad->next_arq[REQ_ASYNC]) { | |
1116 | as_move_to_dispatch(ad, ad->next_arq[REQ_ASYNC]); | |
1117 | dispatched++; | |
1118 | } | |
1119 | ad->last_check_fifo[REQ_ASYNC] = jiffies; | |
1120 | ||
1121 | return dispatched; | |
1122 | } | |
1123 | ||
1da177e4 LT |
1124 | /* Signal that the write batch was uncontended, so we can't time it */ |
1125 | if (ad->batch_data_dir == REQ_ASYNC && !reads) { | |
1126 | if (ad->current_write_count == 0 || !writes) | |
1127 | ad->write_batch_idled = 1; | |
1128 | } | |
1129 | ||
1130 | if (!(reads || writes) | |
1131 | || ad->antic_status == ANTIC_WAIT_REQ | |
1132 | || ad->antic_status == ANTIC_WAIT_NEXT | |
1133 | || ad->changed_batch) | |
1134 | return 0; | |
1135 | ||
f5b3db00 | 1136 | if (!(reads && writes && as_batch_expired(ad))) { |
1da177e4 LT |
1137 | /* |
1138 | * batch is still running or no reads or no writes | |
1139 | */ | |
1140 | arq = ad->next_arq[ad->batch_data_dir]; | |
1141 | ||
1142 | if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) { | |
1143 | if (as_fifo_expired(ad, REQ_SYNC)) | |
1144 | goto fifo_expired; | |
1145 | ||
1146 | if (as_can_anticipate(ad, arq)) { | |
1147 | as_antic_waitreq(ad); | |
1148 | return 0; | |
1149 | } | |
1150 | } | |
1151 | ||
1152 | if (arq) { | |
1153 | /* we have a "next request" */ | |
1154 | if (reads && !writes) | |
1155 | ad->current_batch_expires = | |
1156 | jiffies + ad->batch_expire[REQ_SYNC]; | |
1157 | goto dispatch_request; | |
1158 | } | |
1159 | } | |
1160 | ||
1161 | /* | |
1162 | * at this point we are not running a batch. select the appropriate | |
1163 | * data direction (read / write) | |
1164 | */ | |
1165 | ||
1166 | if (reads) { | |
dd67d051 | 1167 | BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_SYNC])); |
1da177e4 LT |
1168 | |
1169 | if (writes && ad->batch_data_dir == REQ_SYNC) | |
1170 | /* | |
1171 | * Last batch was a read, switch to writes | |
1172 | */ | |
1173 | goto dispatch_writes; | |
1174 | ||
1175 | if (ad->batch_data_dir == REQ_ASYNC) { | |
1176 | WARN_ON(ad->new_batch); | |
1177 | ad->changed_batch = 1; | |
1178 | } | |
1179 | ad->batch_data_dir = REQ_SYNC; | |
1180 | arq = list_entry_fifo(ad->fifo_list[ad->batch_data_dir].next); | |
1181 | ad->last_check_fifo[ad->batch_data_dir] = jiffies; | |
1182 | goto dispatch_request; | |
1183 | } | |
1184 | ||
1185 | /* | |
1186 | * the last batch was a read | |
1187 | */ | |
1188 | ||
1189 | if (writes) { | |
1190 | dispatch_writes: | |
dd67d051 | 1191 | BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_ASYNC])); |
1da177e4 LT |
1192 | |
1193 | if (ad->batch_data_dir == REQ_SYNC) { | |
1194 | ad->changed_batch = 1; | |
1195 | ||
1196 | /* | |
1197 | * new_batch might be 1 when the queue runs out of | |
1198 | * reads. A subsequent submission of a write might | |
1199 | * cause a change of batch before the read is finished. | |
1200 | */ | |
1201 | ad->new_batch = 0; | |
1202 | } | |
1203 | ad->batch_data_dir = REQ_ASYNC; | |
1204 | ad->current_write_count = ad->write_batch_count; | |
1205 | ad->write_batch_idled = 0; | |
1206 | arq = ad->next_arq[ad->batch_data_dir]; | |
1207 | goto dispatch_request; | |
1208 | } | |
1209 | ||
1210 | BUG(); | |
1211 | return 0; | |
1212 | ||
1213 | dispatch_request: | |
1214 | /* | |
1215 | * If a request has expired, service it. | |
1216 | */ | |
1217 | ||
1218 | if (as_fifo_expired(ad, ad->batch_data_dir)) { | |
1219 | fifo_expired: | |
1220 | arq = list_entry_fifo(ad->fifo_list[ad->batch_data_dir].next); | |
1221 | BUG_ON(arq == NULL); | |
1222 | } | |
1223 | ||
1224 | if (ad->changed_batch) { | |
1225 | WARN_ON(ad->new_batch); | |
1226 | ||
1227 | if (ad->nr_dispatched) | |
1228 | return 0; | |
1229 | ||
1230 | if (ad->batch_data_dir == REQ_ASYNC) | |
1231 | ad->current_batch_expires = jiffies + | |
1232 | ad->batch_expire[REQ_ASYNC]; | |
1233 | else | |
1234 | ad->new_batch = 1; | |
1235 | ||
1236 | ad->changed_batch = 0; | |
1237 | } | |
1238 | ||
1239 | /* | |
1240 | * arq is the selected appropriate request. | |
1241 | */ | |
1242 | as_move_to_dispatch(ad, arq); | |
1243 | ||
1244 | return 1; | |
1245 | } | |
1246 | ||
1da177e4 LT |
1247 | /* |
1248 | * add arq to rbtree and fifo | |
1249 | */ | |
b4878f24 | 1250 | static void as_add_request(request_queue_t *q, struct request *rq) |
1da177e4 | 1251 | { |
b4878f24 JA |
1252 | struct as_data *ad = q->elevator->elevator_data; |
1253 | struct as_rq *arq = RQ_DATA(rq); | |
1da177e4 LT |
1254 | int data_dir; |
1255 | ||
b4878f24 JA |
1256 | arq->state = AS_RQ_NEW; |
1257 | ||
1da177e4 | 1258 | if (rq_data_dir(arq->request) == READ |
4aff5e23 | 1259 | || (arq->request->cmd_flags & REQ_RW_SYNC)) |
1da177e4 LT |
1260 | arq->is_sync = 1; |
1261 | else | |
1262 | arq->is_sync = 0; | |
1263 | data_dir = arq->is_sync; | |
1264 | ||
1265 | arq->io_context = as_get_io_context(); | |
1266 | ||
1267 | if (arq->io_context) { | |
1268 | as_update_iohist(ad, arq->io_context->aic, arq->request); | |
1269 | atomic_inc(&arq->io_context->aic->nr_queued); | |
1270 | } | |
1271 | ||
ef9be1d3 | 1272 | as_add_arq_rb(ad, arq); |
1da177e4 | 1273 | |
ef9be1d3 TH |
1274 | /* |
1275 | * set expire time (only used for reads) and add to fifo list | |
1276 | */ | |
1277 | arq->expires = jiffies + ad->fifo_expire[data_dir]; | |
1278 | list_add_tail(&arq->fifo, &ad->fifo_list[data_dir]); | |
1da177e4 | 1279 | |
ef9be1d3 | 1280 | as_update_arq(ad, arq); /* keep state machine up to date */ |
1da177e4 LT |
1281 | arq->state = AS_RQ_QUEUED; |
1282 | } | |
1283 | ||
b4878f24 | 1284 | static void as_activate_request(request_queue_t *q, struct request *rq) |
1da177e4 | 1285 | { |
1da177e4 LT |
1286 | struct as_rq *arq = RQ_DATA(rq); |
1287 | ||
b4878f24 JA |
1288 | WARN_ON(arq->state != AS_RQ_DISPATCHED); |
1289 | arq->state = AS_RQ_REMOVED; | |
1290 | if (arq->io_context && arq->io_context->aic) | |
1291 | atomic_dec(&arq->io_context->aic->nr_dispatched); | |
1da177e4 LT |
1292 | } |
1293 | ||
b4878f24 | 1294 | static void as_deactivate_request(request_queue_t *q, struct request *rq) |
1da177e4 | 1295 | { |
1da177e4 LT |
1296 | struct as_rq *arq = RQ_DATA(rq); |
1297 | ||
b4878f24 JA |
1298 | WARN_ON(arq->state != AS_RQ_REMOVED); |
1299 | arq->state = AS_RQ_DISPATCHED; | |
1300 | if (arq->io_context && arq->io_context->aic) | |
1301 | atomic_inc(&arq->io_context->aic->nr_dispatched); | |
1da177e4 LT |
1302 | } |
1303 | ||
1304 | /* | |
1305 | * as_queue_empty tells us if there are requests left in the device. It may | |
1306 | * not be the case that a driver can get the next request even if the queue | |
1307 | * is not empty - it is used in the block layer to check for plugging and | |
1308 | * merging opportunities | |
1309 | */ | |
1310 | static int as_queue_empty(request_queue_t *q) | |
1311 | { | |
1312 | struct as_data *ad = q->elevator->elevator_data; | |
1313 | ||
b4878f24 JA |
1314 | return list_empty(&ad->fifo_list[REQ_ASYNC]) |
1315 | && list_empty(&ad->fifo_list[REQ_SYNC]); | |
1da177e4 LT |
1316 | } |
1317 | ||
f5b3db00 NP |
1318 | static struct request *as_former_request(request_queue_t *q, |
1319 | struct request *rq) | |
1da177e4 LT |
1320 | { |
1321 | struct as_rq *arq = RQ_DATA(rq); | |
1322 | struct rb_node *rbprev = rb_prev(&arq->rb_node); | |
1323 | struct request *ret = NULL; | |
1324 | ||
1325 | if (rbprev) | |
1326 | ret = rb_entry_arq(rbprev)->request; | |
1327 | ||
1328 | return ret; | |
1329 | } | |
1330 | ||
f5b3db00 NP |
1331 | static struct request *as_latter_request(request_queue_t *q, |
1332 | struct request *rq) | |
1da177e4 LT |
1333 | { |
1334 | struct as_rq *arq = RQ_DATA(rq); | |
1335 | struct rb_node *rbnext = rb_next(&arq->rb_node); | |
1336 | struct request *ret = NULL; | |
1337 | ||
1338 | if (rbnext) | |
1339 | ret = rb_entry_arq(rbnext)->request; | |
1340 | ||
1341 | return ret; | |
1342 | } | |
1343 | ||
1344 | static int | |
1345 | as_merge(request_queue_t *q, struct request **req, struct bio *bio) | |
1346 | { | |
1347 | struct as_data *ad = q->elevator->elevator_data; | |
1348 | sector_t rb_key = bio->bi_sector + bio_sectors(bio); | |
1349 | struct request *__rq; | |
1da177e4 LT |
1350 | |
1351 | /* | |
1352 | * check for front merge | |
1353 | */ | |
1354 | __rq = as_find_arq_rb(ad, rb_key, bio_data_dir(bio)); | |
9817064b JA |
1355 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
1356 | *req = __rq; | |
1357 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
1358 | } |
1359 | ||
1360 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
1361 | } |
1362 | ||
1363 | static void as_merged_request(request_queue_t *q, struct request *req) | |
1364 | { | |
1365 | struct as_data *ad = q->elevator->elevator_data; | |
1366 | struct as_rq *arq = RQ_DATA(req); | |
1367 | ||
1da177e4 LT |
1368 | /* |
1369 | * if the merge was a front merge, we need to reposition request | |
1370 | */ | |
1371 | if (rq_rb_key(req) != arq->rb_key) { | |
1da177e4 | 1372 | as_del_arq_rb(ad, arq); |
ef9be1d3 | 1373 | as_add_arq_rb(ad, arq); |
1da177e4 LT |
1374 | /* |
1375 | * Note! At this stage of this and the next function, our next | |
1376 | * request may not be optimal - eg the request may have "grown" | |
1377 | * behind the disk head. We currently don't bother adjusting. | |
1378 | */ | |
1379 | } | |
1da177e4 LT |
1380 | } |
1381 | ||
f5b3db00 NP |
1382 | static void as_merged_requests(request_queue_t *q, struct request *req, |
1383 | struct request *next) | |
1da177e4 LT |
1384 | { |
1385 | struct as_data *ad = q->elevator->elevator_data; | |
1386 | struct as_rq *arq = RQ_DATA(req); | |
1387 | struct as_rq *anext = RQ_DATA(next); | |
1388 | ||
1389 | BUG_ON(!arq); | |
1390 | BUG_ON(!anext); | |
1391 | ||
1da177e4 | 1392 | if (rq_rb_key(req) != arq->rb_key) { |
1da177e4 | 1393 | as_del_arq_rb(ad, arq); |
ef9be1d3 | 1394 | as_add_arq_rb(ad, arq); |
1da177e4 LT |
1395 | } |
1396 | ||
1397 | /* | |
1398 | * if anext expires before arq, assign its expire time to arq | |
1399 | * and move into anext position (anext will be deleted) in fifo | |
1400 | */ | |
1401 | if (!list_empty(&arq->fifo) && !list_empty(&anext->fifo)) { | |
1402 | if (time_before(anext->expires, arq->expires)) { | |
1403 | list_move(&arq->fifo, &anext->fifo); | |
1404 | arq->expires = anext->expires; | |
1405 | /* | |
1406 | * Don't copy here but swap, because when anext is | |
1407 | * removed below, it must contain the unused context | |
1408 | */ | |
1409 | swap_io_context(&arq->io_context, &anext->io_context); | |
1410 | } | |
1411 | } | |
1412 | ||
1da177e4 LT |
1413 | /* |
1414 | * kill knowledge of next, this one is a goner | |
1415 | */ | |
1416 | as_remove_queued_request(q, next); | |
b4878f24 | 1417 | as_put_io_context(anext); |
1da177e4 LT |
1418 | |
1419 | anext->state = AS_RQ_MERGED; | |
1420 | } | |
1421 | ||
1422 | /* | |
1423 | * This is executed in a "deferred" process context, by kblockd. It calls the | |
1424 | * driver's request_fn so the driver can submit that request. | |
1425 | * | |
1426 | * IMPORTANT! This guy will reenter the elevator, so set up all queue global | |
1427 | * state before calling, and don't rely on any state over calls. | |
1428 | * | |
1429 | * FIXME! dispatch queue is not a queue at all! | |
1430 | */ | |
1431 | static void as_work_handler(void *data) | |
1432 | { | |
1433 | struct request_queue *q = data; | |
1434 | unsigned long flags; | |
1435 | ||
1436 | spin_lock_irqsave(q->queue_lock, flags); | |
b4878f24 | 1437 | if (!as_queue_empty(q)) |
1da177e4 LT |
1438 | q->request_fn(q); |
1439 | spin_unlock_irqrestore(q->queue_lock, flags); | |
1440 | } | |
1441 | ||
1442 | static void as_put_request(request_queue_t *q, struct request *rq) | |
1443 | { | |
1444 | struct as_data *ad = q->elevator->elevator_data; | |
1445 | struct as_rq *arq = RQ_DATA(rq); | |
1446 | ||
1447 | if (!arq) { | |
1448 | WARN_ON(1); | |
1449 | return; | |
1450 | } | |
1451 | ||
b4878f24 JA |
1452 | if (unlikely(arq->state != AS_RQ_POSTSCHED && |
1453 | arq->state != AS_RQ_PRESCHED && | |
1454 | arq->state != AS_RQ_MERGED)) { | |
1da177e4 LT |
1455 | printk("arq->state %d\n", arq->state); |
1456 | WARN_ON(1); | |
1457 | } | |
1458 | ||
1459 | mempool_free(arq, ad->arq_pool); | |
1460 | rq->elevator_private = NULL; | |
1461 | } | |
1462 | ||
22e2c507 | 1463 | static int as_set_request(request_queue_t *q, struct request *rq, |
8267e268 | 1464 | struct bio *bio, gfp_t gfp_mask) |
1da177e4 LT |
1465 | { |
1466 | struct as_data *ad = q->elevator->elevator_data; | |
1467 | struct as_rq *arq = mempool_alloc(ad->arq_pool, gfp_mask); | |
1468 | ||
1469 | if (arq) { | |
1470 | memset(arq, 0, sizeof(*arq)); | |
dd67d051 | 1471 | RB_CLEAR_NODE(&arq->rb_node); |
1da177e4 LT |
1472 | arq->request = rq; |
1473 | arq->state = AS_RQ_PRESCHED; | |
1474 | arq->io_context = NULL; | |
1da177e4 LT |
1475 | INIT_LIST_HEAD(&arq->fifo); |
1476 | rq->elevator_private = arq; | |
1477 | return 0; | |
1478 | } | |
1479 | ||
1480 | return 1; | |
1481 | } | |
1482 | ||
22e2c507 | 1483 | static int as_may_queue(request_queue_t *q, int rw, struct bio *bio) |
1da177e4 LT |
1484 | { |
1485 | int ret = ELV_MQUEUE_MAY; | |
1486 | struct as_data *ad = q->elevator->elevator_data; | |
1487 | struct io_context *ioc; | |
1488 | if (ad->antic_status == ANTIC_WAIT_REQ || | |
1489 | ad->antic_status == ANTIC_WAIT_NEXT) { | |
1490 | ioc = as_get_io_context(); | |
1491 | if (ad->io_context == ioc) | |
1492 | ret = ELV_MQUEUE_MUST; | |
1493 | put_io_context(ioc); | |
1494 | } | |
1495 | ||
1496 | return ret; | |
1497 | } | |
1498 | ||
1499 | static void as_exit_queue(elevator_t *e) | |
1500 | { | |
1501 | struct as_data *ad = e->elevator_data; | |
1502 | ||
1503 | del_timer_sync(&ad->antic_timer); | |
1504 | kblockd_flush(); | |
1505 | ||
1506 | BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC])); | |
1507 | BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC])); | |
1508 | ||
1509 | mempool_destroy(ad->arq_pool); | |
1510 | put_io_context(ad->io_context); | |
1da177e4 LT |
1511 | kfree(ad); |
1512 | } | |
1513 | ||
1514 | /* | |
1515 | * initialize elevator private data (as_data), and alloc a arq for | |
1516 | * each request on the free lists | |
1517 | */ | |
bc1c1169 | 1518 | static void *as_init_queue(request_queue_t *q, elevator_t *e) |
1da177e4 LT |
1519 | { |
1520 | struct as_data *ad; | |
1da177e4 LT |
1521 | |
1522 | if (!arq_pool) | |
bc1c1169 | 1523 | return NULL; |
1da177e4 | 1524 | |
1946089a | 1525 | ad = kmalloc_node(sizeof(*ad), GFP_KERNEL, q->node); |
1da177e4 | 1526 | if (!ad) |
bc1c1169 | 1527 | return NULL; |
1da177e4 LT |
1528 | memset(ad, 0, sizeof(*ad)); |
1529 | ||
1530 | ad->q = q; /* Identify what queue the data belongs to */ | |
1531 | ||
1946089a CL |
1532 | ad->arq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab, |
1533 | mempool_free_slab, arq_pool, q->node); | |
1da177e4 | 1534 | if (!ad->arq_pool) { |
1da177e4 | 1535 | kfree(ad); |
bc1c1169 | 1536 | return NULL; |
1da177e4 LT |
1537 | } |
1538 | ||
1539 | /* anticipatory scheduling helpers */ | |
1540 | ad->antic_timer.function = as_antic_timeout; | |
1541 | ad->antic_timer.data = (unsigned long)q; | |
1542 | init_timer(&ad->antic_timer); | |
1543 | INIT_WORK(&ad->antic_work, as_work_handler, q); | |
1544 | ||
1da177e4 LT |
1545 | INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]); |
1546 | INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]); | |
1547 | ad->sort_list[REQ_SYNC] = RB_ROOT; | |
1548 | ad->sort_list[REQ_ASYNC] = RB_ROOT; | |
1da177e4 LT |
1549 | ad->fifo_expire[REQ_SYNC] = default_read_expire; |
1550 | ad->fifo_expire[REQ_ASYNC] = default_write_expire; | |
1551 | ad->antic_expire = default_antic_expire; | |
1552 | ad->batch_expire[REQ_SYNC] = default_read_batch_expire; | |
1553 | ad->batch_expire[REQ_ASYNC] = default_write_batch_expire; | |
1da177e4 LT |
1554 | |
1555 | ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC]; | |
1556 | ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10; | |
1557 | if (ad->write_batch_count < 2) | |
1558 | ad->write_batch_count = 2; | |
1559 | ||
bc1c1169 | 1560 | return ad; |
1da177e4 LT |
1561 | } |
1562 | ||
1563 | /* | |
1564 | * sysfs parts below | |
1565 | */ | |
1da177e4 LT |
1566 | |
1567 | static ssize_t | |
1568 | as_var_show(unsigned int var, char *page) | |
1569 | { | |
1da177e4 LT |
1570 | return sprintf(page, "%d\n", var); |
1571 | } | |
1572 | ||
1573 | static ssize_t | |
1574 | as_var_store(unsigned long *var, const char *page, size_t count) | |
1575 | { | |
1da177e4 LT |
1576 | char *p = (char *) page; |
1577 | ||
c9b3ad67 | 1578 | *var = simple_strtoul(p, &p, 10); |
1da177e4 LT |
1579 | return count; |
1580 | } | |
1581 | ||
e572ec7e | 1582 | static ssize_t est_time_show(elevator_t *e, char *page) |
1da177e4 | 1583 | { |
3d1ab40f | 1584 | struct as_data *ad = e->elevator_data; |
1da177e4 LT |
1585 | int pos = 0; |
1586 | ||
f5b3db00 NP |
1587 | pos += sprintf(page+pos, "%lu %% exit probability\n", |
1588 | 100*ad->exit_prob/256); | |
1589 | pos += sprintf(page+pos, "%lu %% probability of exiting without a " | |
1590 | "cooperating process submitting IO\n", | |
1591 | 100*ad->exit_no_coop/256); | |
1da177e4 | 1592 | pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean); |
f5b3db00 NP |
1593 | pos += sprintf(page+pos, "%llu sectors new seek distance\n", |
1594 | (unsigned long long)ad->new_seek_mean); | |
1da177e4 LT |
1595 | |
1596 | return pos; | |
1597 | } | |
1598 | ||
1599 | #define SHOW_FUNCTION(__FUNC, __VAR) \ | |
3d1ab40f | 1600 | static ssize_t __FUNC(elevator_t *e, char *page) \ |
1da177e4 | 1601 | { \ |
3d1ab40f | 1602 | struct as_data *ad = e->elevator_data; \ |
1da177e4 LT |
1603 | return as_var_show(jiffies_to_msecs((__VAR)), (page)); \ |
1604 | } | |
e572ec7e AV |
1605 | SHOW_FUNCTION(as_read_expire_show, ad->fifo_expire[REQ_SYNC]); |
1606 | SHOW_FUNCTION(as_write_expire_show, ad->fifo_expire[REQ_ASYNC]); | |
1607 | SHOW_FUNCTION(as_antic_expire_show, ad->antic_expire); | |
1608 | SHOW_FUNCTION(as_read_batch_expire_show, ad->batch_expire[REQ_SYNC]); | |
1609 | SHOW_FUNCTION(as_write_batch_expire_show, ad->batch_expire[REQ_ASYNC]); | |
1da177e4 LT |
1610 | #undef SHOW_FUNCTION |
1611 | ||
1612 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \ | |
3d1ab40f | 1613 | static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \ |
1da177e4 | 1614 | { \ |
3d1ab40f AV |
1615 | struct as_data *ad = e->elevator_data; \ |
1616 | int ret = as_var_store(__PTR, (page), count); \ | |
1da177e4 LT |
1617 | if (*(__PTR) < (MIN)) \ |
1618 | *(__PTR) = (MIN); \ | |
1619 | else if (*(__PTR) > (MAX)) \ | |
1620 | *(__PTR) = (MAX); \ | |
1621 | *(__PTR) = msecs_to_jiffies(*(__PTR)); \ | |
1622 | return ret; \ | |
1623 | } | |
e572ec7e AV |
1624 | STORE_FUNCTION(as_read_expire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX); |
1625 | STORE_FUNCTION(as_write_expire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX); | |
1626 | STORE_FUNCTION(as_antic_expire_store, &ad->antic_expire, 0, INT_MAX); | |
1627 | STORE_FUNCTION(as_read_batch_expire_store, | |
1da177e4 | 1628 | &ad->batch_expire[REQ_SYNC], 0, INT_MAX); |
e572ec7e | 1629 | STORE_FUNCTION(as_write_batch_expire_store, |
1da177e4 LT |
1630 | &ad->batch_expire[REQ_ASYNC], 0, INT_MAX); |
1631 | #undef STORE_FUNCTION | |
1632 | ||
e572ec7e AV |
1633 | #define AS_ATTR(name) \ |
1634 | __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store) | |
1635 | ||
1636 | static struct elv_fs_entry as_attrs[] = { | |
1637 | __ATTR_RO(est_time), | |
1638 | AS_ATTR(read_expire), | |
1639 | AS_ATTR(write_expire), | |
1640 | AS_ATTR(antic_expire), | |
1641 | AS_ATTR(read_batch_expire), | |
1642 | AS_ATTR(write_batch_expire), | |
1643 | __ATTR_NULL | |
1da177e4 LT |
1644 | }; |
1645 | ||
1da177e4 LT |
1646 | static struct elevator_type iosched_as = { |
1647 | .ops = { | |
1648 | .elevator_merge_fn = as_merge, | |
1649 | .elevator_merged_fn = as_merged_request, | |
1650 | .elevator_merge_req_fn = as_merged_requests, | |
b4878f24 JA |
1651 | .elevator_dispatch_fn = as_dispatch_request, |
1652 | .elevator_add_req_fn = as_add_request, | |
1653 | .elevator_activate_req_fn = as_activate_request, | |
1da177e4 LT |
1654 | .elevator_deactivate_req_fn = as_deactivate_request, |
1655 | .elevator_queue_empty_fn = as_queue_empty, | |
1656 | .elevator_completed_req_fn = as_completed_request, | |
1657 | .elevator_former_req_fn = as_former_request, | |
1658 | .elevator_latter_req_fn = as_latter_request, | |
1659 | .elevator_set_req_fn = as_set_request, | |
1660 | .elevator_put_req_fn = as_put_request, | |
1661 | .elevator_may_queue_fn = as_may_queue, | |
1662 | .elevator_init_fn = as_init_queue, | |
1663 | .elevator_exit_fn = as_exit_queue, | |
e17a9489 | 1664 | .trim = as_trim, |
1da177e4 LT |
1665 | }, |
1666 | ||
3d1ab40f | 1667 | .elevator_attrs = as_attrs, |
1da177e4 LT |
1668 | .elevator_name = "anticipatory", |
1669 | .elevator_owner = THIS_MODULE, | |
1670 | }; | |
1671 | ||
1672 | static int __init as_init(void) | |
1673 | { | |
1674 | int ret; | |
1675 | ||
1676 | arq_pool = kmem_cache_create("as_arq", sizeof(struct as_rq), | |
1677 | 0, 0, NULL, NULL); | |
1678 | if (!arq_pool) | |
1679 | return -ENOMEM; | |
1680 | ||
1681 | ret = elv_register(&iosched_as); | |
1682 | if (!ret) { | |
1683 | /* | |
1684 | * don't allow AS to get unregistered, since we would have | |
1685 | * to browse all tasks in the system and release their | |
1686 | * as_io_context first | |
1687 | */ | |
1688 | __module_get(THIS_MODULE); | |
1689 | return 0; | |
1690 | } | |
1691 | ||
1692 | kmem_cache_destroy(arq_pool); | |
1693 | return ret; | |
1694 | } | |
1695 | ||
1696 | static void __exit as_exit(void) | |
1697 | { | |
334e94de | 1698 | DECLARE_COMPLETION(all_gone); |
1da177e4 | 1699 | elv_unregister(&iosched_as); |
334e94de | 1700 | ioc_gone = &all_gone; |
fba82272 OH |
1701 | /* ioc_gone's update must be visible before reading ioc_count */ |
1702 | smp_wmb(); | |
334e94de | 1703 | if (atomic_read(&ioc_count)) |
fba82272 | 1704 | wait_for_completion(ioc_gone); |
334e94de | 1705 | synchronize_rcu(); |
83521d3e | 1706 | kmem_cache_destroy(arq_pool); |
1da177e4 LT |
1707 | } |
1708 | ||
1709 | module_init(as_init); | |
1710 | module_exit(as_exit); | |
1711 | ||
1712 | MODULE_AUTHOR("Nick Piggin"); | |
1713 | MODULE_LICENSE("GPL"); | |
1714 | MODULE_DESCRIPTION("anticipatory IO scheduler"); |