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 | ||
1da177e4 LT |
134 | /* |
135 | * per-request data. | |
136 | */ | |
137 | enum arq_state { | |
138 | AS_RQ_NEW=0, /* New - not referenced and not on any lists */ | |
139 | AS_RQ_QUEUED, /* In the request queue. It belongs to the | |
140 | scheduler */ | |
141 | AS_RQ_DISPATCHED, /* On the dispatch list. It belongs to the | |
142 | driver now */ | |
143 | AS_RQ_PRESCHED, /* Debug poisoning for requests being used */ | |
144 | AS_RQ_REMOVED, | |
145 | AS_RQ_MERGED, | |
146 | AS_RQ_POSTSCHED, /* when they shouldn't be */ | |
147 | }; | |
148 | ||
149 | struct as_rq { | |
1da177e4 LT |
150 | struct request *request; |
151 | ||
152 | struct io_context *io_context; /* The submitting task */ | |
153 | ||
1da177e4 LT |
154 | enum arq_state state; |
155 | }; | |
156 | ||
157 | #define RQ_DATA(rq) ((struct as_rq *) (rq)->elevator_private) | |
158 | ||
159 | static kmem_cache_t *arq_pool; | |
160 | ||
334e94de AV |
161 | static atomic_t ioc_count = ATOMIC_INIT(0); |
162 | static struct completion *ioc_gone; | |
163 | ||
ef9be1d3 TH |
164 | static void as_move_to_dispatch(struct as_data *ad, struct as_rq *arq); |
165 | static void as_antic_stop(struct as_data *ad); | |
166 | ||
1da177e4 LT |
167 | /* |
168 | * IO Context helper functions | |
169 | */ | |
170 | ||
171 | /* Called to deallocate the as_io_context */ | |
172 | static void free_as_io_context(struct as_io_context *aic) | |
173 | { | |
174 | kfree(aic); | |
334e94de AV |
175 | if (atomic_dec_and_test(&ioc_count) && ioc_gone) |
176 | complete(ioc_gone); | |
1da177e4 LT |
177 | } |
178 | ||
e17a9489 AV |
179 | static void as_trim(struct io_context *ioc) |
180 | { | |
334e94de AV |
181 | if (ioc->aic) |
182 | free_as_io_context(ioc->aic); | |
e17a9489 AV |
183 | ioc->aic = NULL; |
184 | } | |
185 | ||
1da177e4 LT |
186 | /* Called when the task exits */ |
187 | static void exit_as_io_context(struct as_io_context *aic) | |
188 | { | |
189 | WARN_ON(!test_bit(AS_TASK_RUNNING, &aic->state)); | |
190 | clear_bit(AS_TASK_RUNNING, &aic->state); | |
191 | } | |
192 | ||
193 | static struct as_io_context *alloc_as_io_context(void) | |
194 | { | |
195 | struct as_io_context *ret; | |
196 | ||
197 | ret = kmalloc(sizeof(*ret), GFP_ATOMIC); | |
198 | if (ret) { | |
199 | ret->dtor = free_as_io_context; | |
200 | ret->exit = exit_as_io_context; | |
201 | ret->state = 1 << AS_TASK_RUNNING; | |
202 | atomic_set(&ret->nr_queued, 0); | |
203 | atomic_set(&ret->nr_dispatched, 0); | |
204 | spin_lock_init(&ret->lock); | |
205 | ret->ttime_total = 0; | |
206 | ret->ttime_samples = 0; | |
207 | ret->ttime_mean = 0; | |
208 | ret->seek_total = 0; | |
209 | ret->seek_samples = 0; | |
210 | ret->seek_mean = 0; | |
334e94de | 211 | atomic_inc(&ioc_count); |
1da177e4 LT |
212 | } |
213 | ||
214 | return ret; | |
215 | } | |
216 | ||
217 | /* | |
218 | * If the current task has no AS IO context then create one and initialise it. | |
219 | * Then take a ref on the task's io context and return it. | |
220 | */ | |
221 | static struct io_context *as_get_io_context(void) | |
222 | { | |
223 | struct io_context *ioc = get_io_context(GFP_ATOMIC); | |
224 | if (ioc && !ioc->aic) { | |
225 | ioc->aic = alloc_as_io_context(); | |
226 | if (!ioc->aic) { | |
227 | put_io_context(ioc); | |
228 | ioc = NULL; | |
229 | } | |
230 | } | |
231 | return ioc; | |
232 | } | |
233 | ||
b4878f24 JA |
234 | static void as_put_io_context(struct as_rq *arq) |
235 | { | |
236 | struct as_io_context *aic; | |
237 | ||
238 | if (unlikely(!arq->io_context)) | |
239 | return; | |
240 | ||
241 | aic = arq->io_context->aic; | |
242 | ||
9e2585a8 | 243 | if (rq_is_sync(arq->request) && aic) { |
b4878f24 JA |
244 | spin_lock(&aic->lock); |
245 | set_bit(AS_TASK_IORUNNING, &aic->state); | |
246 | aic->last_end_request = jiffies; | |
247 | spin_unlock(&aic->lock); | |
248 | } | |
249 | ||
250 | put_io_context(arq->io_context); | |
251 | } | |
252 | ||
1da177e4 LT |
253 | /* |
254 | * rb tree support functions | |
255 | */ | |
9e2585a8 | 256 | #define RQ_RB_ROOT(ad, rq) (&(ad)->sort_list[rq_is_sync((rq))]) |
1da177e4 | 257 | |
e37f346e | 258 | static void as_add_arq_rb(struct as_data *ad, struct request *rq) |
ef9be1d3 | 259 | { |
e37f346e | 260 | struct request *alias; |
ef9be1d3 | 261 | |
9e2585a8 | 262 | while ((unlikely(alias = elv_rb_add(RQ_RB_ROOT(ad, rq), rq)))) { |
e37f346e | 263 | as_move_to_dispatch(ad, RQ_DATA(alias)); |
ef9be1d3 TH |
264 | as_antic_stop(ad); |
265 | } | |
266 | } | |
267 | ||
e37f346e | 268 | static inline void as_del_arq_rb(struct as_data *ad, struct request *rq) |
1da177e4 | 269 | { |
9e2585a8 | 270 | elv_rb_del(RQ_RB_ROOT(ad, rq), rq); |
1da177e4 LT |
271 | } |
272 | ||
273 | /* | |
274 | * IO Scheduler proper | |
275 | */ | |
276 | ||
277 | #define MAXBACK (1024 * 1024) /* | |
278 | * Maximum distance the disk will go backward | |
279 | * for a request. | |
280 | */ | |
281 | ||
282 | #define BACK_PENALTY 2 | |
283 | ||
284 | /* | |
285 | * as_choose_req selects the preferred one of two requests of the same data_dir | |
286 | * ignoring time - eg. timeouts, which is the job of as_dispatch_request | |
287 | */ | |
288 | static struct as_rq * | |
289 | as_choose_req(struct as_data *ad, struct as_rq *arq1, struct as_rq *arq2) | |
290 | { | |
291 | int data_dir; | |
292 | sector_t last, s1, s2, d1, d2; | |
293 | int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */ | |
294 | const sector_t maxback = MAXBACK; | |
295 | ||
296 | if (arq1 == NULL || arq1 == arq2) | |
297 | return arq2; | |
298 | if (arq2 == NULL) | |
299 | return arq1; | |
300 | ||
9e2585a8 | 301 | data_dir = rq_is_sync(arq1->request); |
1da177e4 LT |
302 | |
303 | last = ad->last_sector[data_dir]; | |
304 | s1 = arq1->request->sector; | |
305 | s2 = arq2->request->sector; | |
306 | ||
9e2585a8 | 307 | BUG_ON(data_dir != rq_is_sync(arq2->request)); |
1da177e4 LT |
308 | |
309 | /* | |
310 | * Strict one way elevator _except_ in the case where we allow | |
311 | * short backward seeks which are biased as twice the cost of a | |
312 | * similar forward seek. | |
313 | */ | |
314 | if (s1 >= last) | |
315 | d1 = s1 - last; | |
316 | else if (s1+maxback >= last) | |
317 | d1 = (last - s1)*BACK_PENALTY; | |
318 | else { | |
319 | r1_wrap = 1; | |
320 | d1 = 0; /* shut up, gcc */ | |
321 | } | |
322 | ||
323 | if (s2 >= last) | |
324 | d2 = s2 - last; | |
325 | else if (s2+maxback >= last) | |
326 | d2 = (last - s2)*BACK_PENALTY; | |
327 | else { | |
328 | r2_wrap = 1; | |
329 | d2 = 0; | |
330 | } | |
331 | ||
332 | /* Found required data */ | |
333 | if (!r1_wrap && r2_wrap) | |
334 | return arq1; | |
335 | else if (!r2_wrap && r1_wrap) | |
336 | return arq2; | |
337 | else if (r1_wrap && r2_wrap) { | |
338 | /* both behind the head */ | |
339 | if (s1 <= s2) | |
340 | return arq1; | |
341 | else | |
342 | return arq2; | |
343 | } | |
344 | ||
345 | /* Both requests in front of the head */ | |
346 | if (d1 < d2) | |
347 | return arq1; | |
348 | else if (d2 < d1) | |
349 | return arq2; | |
350 | else { | |
351 | if (s1 >= s2) | |
352 | return arq1; | |
353 | else | |
354 | return arq2; | |
355 | } | |
356 | } | |
357 | ||
358 | /* | |
359 | * as_find_next_arq finds the next request after @prev in elevator order. | |
360 | * this with as_choose_req form the basis for how the scheduler chooses | |
361 | * what request to process next. Anticipation works on top of this. | |
362 | */ | |
e37f346e | 363 | static struct as_rq *as_find_next_arq(struct as_data *ad, struct as_rq *arq) |
1da177e4 | 364 | { |
e37f346e | 365 | struct request *last = arq->request; |
1da177e4 LT |
366 | struct rb_node *rbnext = rb_next(&last->rb_node); |
367 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
e37f346e | 368 | struct as_rq *next = NULL, *prev = NULL; |
1da177e4 | 369 | |
e37f346e | 370 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
371 | |
372 | if (rbprev) | |
e37f346e | 373 | prev = RQ_DATA(rb_entry_rq(rbprev)); |
1da177e4 LT |
374 | |
375 | if (rbnext) | |
e37f346e | 376 | next = RQ_DATA(rb_entry_rq(rbnext)); |
1da177e4 | 377 | else { |
9e2585a8 | 378 | const int data_dir = rq_is_sync(last); |
1da177e4 | 379 | |
e37f346e JA |
380 | rbnext = rb_first(&ad->sort_list[data_dir]); |
381 | if (rbnext && rbnext != &last->rb_node) | |
382 | next = RQ_DATA(rb_entry_rq(rbnext)); | |
383 | } | |
1da177e4 | 384 | |
e37f346e | 385 | return as_choose_req(ad, next, prev); |
1da177e4 LT |
386 | } |
387 | ||
388 | /* | |
389 | * anticipatory scheduling functions follow | |
390 | */ | |
391 | ||
392 | /* | |
393 | * as_antic_expired tells us when we have anticipated too long. | |
394 | * The funny "absolute difference" math on the elapsed time is to handle | |
395 | * jiffy wraps, and disks which have been idle for 0x80000000 jiffies. | |
396 | */ | |
397 | static int as_antic_expired(struct as_data *ad) | |
398 | { | |
399 | long delta_jif; | |
400 | ||
401 | delta_jif = jiffies - ad->antic_start; | |
402 | if (unlikely(delta_jif < 0)) | |
403 | delta_jif = -delta_jif; | |
404 | if (delta_jif < ad->antic_expire) | |
405 | return 0; | |
406 | ||
407 | return 1; | |
408 | } | |
409 | ||
410 | /* | |
411 | * as_antic_waitnext starts anticipating that a nice request will soon be | |
412 | * submitted. See also as_antic_waitreq | |
413 | */ | |
414 | static void as_antic_waitnext(struct as_data *ad) | |
415 | { | |
416 | unsigned long timeout; | |
417 | ||
418 | BUG_ON(ad->antic_status != ANTIC_OFF | |
419 | && ad->antic_status != ANTIC_WAIT_REQ); | |
420 | ||
421 | timeout = ad->antic_start + ad->antic_expire; | |
422 | ||
423 | mod_timer(&ad->antic_timer, timeout); | |
424 | ||
425 | ad->antic_status = ANTIC_WAIT_NEXT; | |
426 | } | |
427 | ||
428 | /* | |
429 | * as_antic_waitreq starts anticipating. We don't start timing the anticipation | |
430 | * until the request that we're anticipating on has finished. This means we | |
431 | * are timing from when the candidate process wakes up hopefully. | |
432 | */ | |
433 | static void as_antic_waitreq(struct as_data *ad) | |
434 | { | |
435 | BUG_ON(ad->antic_status == ANTIC_FINISHED); | |
436 | if (ad->antic_status == ANTIC_OFF) { | |
437 | if (!ad->io_context || ad->ioc_finished) | |
438 | as_antic_waitnext(ad); | |
439 | else | |
440 | ad->antic_status = ANTIC_WAIT_REQ; | |
441 | } | |
442 | } | |
443 | ||
444 | /* | |
445 | * This is called directly by the functions in this file to stop anticipation. | |
446 | * We kill the timer and schedule a call to the request_fn asap. | |
447 | */ | |
448 | static void as_antic_stop(struct as_data *ad) | |
449 | { | |
450 | int status = ad->antic_status; | |
451 | ||
452 | if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) { | |
453 | if (status == ANTIC_WAIT_NEXT) | |
454 | del_timer(&ad->antic_timer); | |
455 | ad->antic_status = ANTIC_FINISHED; | |
456 | /* see as_work_handler */ | |
457 | kblockd_schedule_work(&ad->antic_work); | |
458 | } | |
459 | } | |
460 | ||
461 | /* | |
462 | * as_antic_timeout is the timer function set by as_antic_waitnext. | |
463 | */ | |
464 | static void as_antic_timeout(unsigned long data) | |
465 | { | |
466 | struct request_queue *q = (struct request_queue *)data; | |
467 | struct as_data *ad = q->elevator->elevator_data; | |
468 | unsigned long flags; | |
469 | ||
470 | spin_lock_irqsave(q->queue_lock, flags); | |
471 | if (ad->antic_status == ANTIC_WAIT_REQ | |
472 | || ad->antic_status == ANTIC_WAIT_NEXT) { | |
473 | struct as_io_context *aic = ad->io_context->aic; | |
474 | ||
475 | ad->antic_status = ANTIC_FINISHED; | |
476 | kblockd_schedule_work(&ad->antic_work); | |
477 | ||
478 | if (aic->ttime_samples == 0) { | |
f5b3db00 | 479 | /* process anticipated on has exited or timed out*/ |
1da177e4 LT |
480 | ad->exit_prob = (7*ad->exit_prob + 256)/8; |
481 | } | |
f5b3db00 NP |
482 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
483 | /* process not "saved" by a cooperating request */ | |
484 | ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8; | |
485 | } | |
1da177e4 LT |
486 | } |
487 | spin_unlock_irqrestore(q->queue_lock, flags); | |
488 | } | |
489 | ||
f5b3db00 NP |
490 | static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic, |
491 | unsigned long ttime) | |
492 | { | |
493 | /* fixed point: 1.0 == 1<<8 */ | |
494 | if (aic->ttime_samples == 0) { | |
495 | ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8; | |
496 | ad->new_ttime_mean = ad->new_ttime_total / 256; | |
497 | ||
498 | ad->exit_prob = (7*ad->exit_prob)/8; | |
499 | } | |
500 | aic->ttime_samples = (7*aic->ttime_samples + 256) / 8; | |
501 | aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8; | |
502 | aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples; | |
503 | } | |
504 | ||
505 | static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic, | |
506 | sector_t sdist) | |
507 | { | |
508 | u64 total; | |
509 | ||
510 | if (aic->seek_samples == 0) { | |
511 | ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8; | |
512 | ad->new_seek_mean = ad->new_seek_total / 256; | |
513 | } | |
514 | ||
515 | /* | |
516 | * Don't allow the seek distance to get too large from the | |
517 | * odd fragment, pagein, etc | |
518 | */ | |
519 | if (aic->seek_samples <= 60) /* second&third seek */ | |
520 | sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024); | |
521 | else | |
522 | sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64); | |
523 | ||
524 | aic->seek_samples = (7*aic->seek_samples + 256) / 8; | |
525 | aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8; | |
526 | total = aic->seek_total + (aic->seek_samples/2); | |
527 | do_div(total, aic->seek_samples); | |
528 | aic->seek_mean = (sector_t)total; | |
529 | } | |
530 | ||
531 | /* | |
532 | * as_update_iohist keeps a decaying histogram of IO thinktimes, and | |
533 | * updates @aic->ttime_mean based on that. It is called when a new | |
534 | * request is queued. | |
535 | */ | |
536 | static void as_update_iohist(struct as_data *ad, struct as_io_context *aic, | |
537 | struct request *rq) | |
538 | { | |
9e2585a8 | 539 | int data_dir = rq_is_sync(rq); |
f5b3db00 NP |
540 | unsigned long thinktime = 0; |
541 | sector_t seek_dist; | |
542 | ||
543 | if (aic == NULL) | |
544 | return; | |
545 | ||
546 | if (data_dir == REQ_SYNC) { | |
547 | unsigned long in_flight = atomic_read(&aic->nr_queued) | |
548 | + atomic_read(&aic->nr_dispatched); | |
549 | spin_lock(&aic->lock); | |
550 | if (test_bit(AS_TASK_IORUNNING, &aic->state) || | |
551 | test_bit(AS_TASK_IOSTARTED, &aic->state)) { | |
552 | /* Calculate read -> read thinktime */ | |
553 | if (test_bit(AS_TASK_IORUNNING, &aic->state) | |
554 | && in_flight == 0) { | |
555 | thinktime = jiffies - aic->last_end_request; | |
556 | thinktime = min(thinktime, MAX_THINKTIME-1); | |
557 | } | |
558 | as_update_thinktime(ad, aic, thinktime); | |
559 | ||
560 | /* Calculate read -> read seek distance */ | |
561 | if (aic->last_request_pos < rq->sector) | |
562 | seek_dist = rq->sector - aic->last_request_pos; | |
563 | else | |
564 | seek_dist = aic->last_request_pos - rq->sector; | |
565 | as_update_seekdist(ad, aic, seek_dist); | |
566 | } | |
567 | aic->last_request_pos = rq->sector + rq->nr_sectors; | |
568 | set_bit(AS_TASK_IOSTARTED, &aic->state); | |
569 | spin_unlock(&aic->lock); | |
570 | } | |
571 | } | |
572 | ||
1da177e4 LT |
573 | /* |
574 | * as_close_req decides if one request is considered "close" to the | |
575 | * previous one issued. | |
576 | */ | |
f5b3db00 NP |
577 | static int as_close_req(struct as_data *ad, struct as_io_context *aic, |
578 | struct as_rq *arq) | |
1da177e4 LT |
579 | { |
580 | unsigned long delay; /* milliseconds */ | |
581 | sector_t last = ad->last_sector[ad->batch_data_dir]; | |
582 | sector_t next = arq->request->sector; | |
583 | sector_t delta; /* acceptable close offset (in sectors) */ | |
f5b3db00 | 584 | sector_t s; |
1da177e4 LT |
585 | |
586 | if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished) | |
587 | delay = 0; | |
588 | else | |
589 | delay = ((jiffies - ad->antic_start) * 1000) / HZ; | |
590 | ||
f5b3db00 NP |
591 | if (delay == 0) |
592 | delta = 8192; | |
1da177e4 | 593 | else if (delay <= 20 && delay <= ad->antic_expire) |
f5b3db00 | 594 | delta = 8192 << delay; |
1da177e4 LT |
595 | else |
596 | return 1; | |
597 | ||
f5b3db00 NP |
598 | if ((last <= next + (delta>>1)) && (next <= last + delta)) |
599 | return 1; | |
600 | ||
601 | if (last < next) | |
602 | s = next - last; | |
603 | else | |
604 | s = last - next; | |
605 | ||
606 | if (aic->seek_samples == 0) { | |
607 | /* | |
608 | * Process has just started IO. Use past statistics to | |
609 | * gauge success possibility | |
610 | */ | |
611 | if (ad->new_seek_mean > s) { | |
612 | /* this request is better than what we're expecting */ | |
613 | return 1; | |
614 | } | |
615 | ||
616 | } else { | |
617 | if (aic->seek_mean > s) { | |
618 | /* this request is better than what we're expecting */ | |
619 | return 1; | |
620 | } | |
621 | } | |
622 | ||
623 | return 0; | |
1da177e4 LT |
624 | } |
625 | ||
626 | /* | |
627 | * as_can_break_anticipation returns true if we have been anticipating this | |
628 | * request. | |
629 | * | |
630 | * It also returns true if the process against which we are anticipating | |
631 | * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to | |
632 | * dispatch it ASAP, because we know that application will not be submitting | |
633 | * any new reads. | |
634 | * | |
f5b3db00 | 635 | * If the task which has submitted the request has exited, break anticipation. |
1da177e4 LT |
636 | * |
637 | * If this task has queued some other IO, do not enter enticipation. | |
638 | */ | |
639 | static int as_can_break_anticipation(struct as_data *ad, struct as_rq *arq) | |
640 | { | |
641 | struct io_context *ioc; | |
642 | struct as_io_context *aic; | |
1da177e4 LT |
643 | |
644 | ioc = ad->io_context; | |
645 | BUG_ON(!ioc); | |
646 | ||
647 | if (arq && ioc == arq->io_context) { | |
648 | /* request from same process */ | |
649 | return 1; | |
650 | } | |
651 | ||
652 | if (ad->ioc_finished && as_antic_expired(ad)) { | |
653 | /* | |
654 | * In this situation status should really be FINISHED, | |
655 | * however the timer hasn't had the chance to run yet. | |
656 | */ | |
657 | return 1; | |
658 | } | |
659 | ||
660 | aic = ioc->aic; | |
661 | if (!aic) | |
662 | return 0; | |
663 | ||
1da177e4 LT |
664 | if (atomic_read(&aic->nr_queued) > 0) { |
665 | /* process has more requests queued */ | |
666 | return 1; | |
667 | } | |
668 | ||
669 | if (atomic_read(&aic->nr_dispatched) > 0) { | |
670 | /* process has more requests dispatched */ | |
671 | return 1; | |
672 | } | |
673 | ||
9e2585a8 | 674 | if (arq && rq_is_sync(arq->request) && as_close_req(ad, aic, arq)) { |
1da177e4 LT |
675 | /* |
676 | * Found a close request that is not one of ours. | |
677 | * | |
f5b3db00 NP |
678 | * This makes close requests from another process update |
679 | * our IO history. Is generally useful when there are | |
1da177e4 LT |
680 | * two or more cooperating processes working in the same |
681 | * area. | |
682 | */ | |
f5b3db00 NP |
683 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
684 | if (aic->ttime_samples == 0) | |
685 | ad->exit_prob = (7*ad->exit_prob + 256)/8; | |
686 | ||
687 | ad->exit_no_coop = (7*ad->exit_no_coop)/8; | |
688 | } | |
689 | ||
690 | as_update_iohist(ad, aic, arq->request); | |
1da177e4 LT |
691 | return 1; |
692 | } | |
693 | ||
f5b3db00 NP |
694 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
695 | /* process anticipated on has exited */ | |
696 | if (aic->ttime_samples == 0) | |
697 | ad->exit_prob = (7*ad->exit_prob + 256)/8; | |
698 | ||
699 | if (ad->exit_no_coop > 128) | |
700 | return 1; | |
701 | } | |
1da177e4 LT |
702 | |
703 | if (aic->ttime_samples == 0) { | |
704 | if (ad->new_ttime_mean > ad->antic_expire) | |
705 | return 1; | |
f5b3db00 | 706 | if (ad->exit_prob * ad->exit_no_coop > 128*256) |
1da177e4 LT |
707 | return 1; |
708 | } else if (aic->ttime_mean > ad->antic_expire) { | |
709 | /* the process thinks too much between requests */ | |
710 | return 1; | |
711 | } | |
712 | ||
1da177e4 LT |
713 | return 0; |
714 | } | |
715 | ||
716 | /* | |
d6e05edc | 717 | * as_can_anticipate indicates whether we should either run arq |
1da177e4 LT |
718 | * or keep anticipating a better request. |
719 | */ | |
720 | static int as_can_anticipate(struct as_data *ad, struct as_rq *arq) | |
721 | { | |
722 | if (!ad->io_context) | |
723 | /* | |
724 | * Last request submitted was a write | |
725 | */ | |
726 | return 0; | |
727 | ||
728 | if (ad->antic_status == ANTIC_FINISHED) | |
729 | /* | |
730 | * Don't restart if we have just finished. Run the next request | |
731 | */ | |
732 | return 0; | |
733 | ||
734 | if (as_can_break_anticipation(ad, arq)) | |
735 | /* | |
736 | * This request is a good candidate. Don't keep anticipating, | |
737 | * run it. | |
738 | */ | |
739 | return 0; | |
740 | ||
741 | /* | |
742 | * OK from here, we haven't finished, and don't have a decent request! | |
743 | * Status is either ANTIC_OFF so start waiting, | |
744 | * ANTIC_WAIT_REQ so continue waiting for request to finish | |
745 | * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request. | |
1da177e4 LT |
746 | */ |
747 | ||
748 | return 1; | |
749 | } | |
750 | ||
1da177e4 LT |
751 | /* |
752 | * as_update_arq must be called whenever a request (arq) is added to | |
753 | * the sort_list. This function keeps caches up to date, and checks if the | |
754 | * request might be one we are "anticipating" | |
755 | */ | |
756 | static void as_update_arq(struct as_data *ad, struct as_rq *arq) | |
757 | { | |
9e2585a8 | 758 | const int data_dir = rq_is_sync(arq->request); |
1da177e4 LT |
759 | |
760 | /* keep the next_arq cache up to date */ | |
761 | ad->next_arq[data_dir] = as_choose_req(ad, arq, ad->next_arq[data_dir]); | |
762 | ||
763 | /* | |
764 | * have we been anticipating this request? | |
765 | * or does it come from the same process as the one we are anticipating | |
766 | * for? | |
767 | */ | |
768 | if (ad->antic_status == ANTIC_WAIT_REQ | |
769 | || ad->antic_status == ANTIC_WAIT_NEXT) { | |
770 | if (as_can_break_anticipation(ad, arq)) | |
771 | as_antic_stop(ad); | |
772 | } | |
773 | } | |
774 | ||
775 | /* | |
776 | * Gathers timings and resizes the write batch automatically | |
777 | */ | |
778 | static void update_write_batch(struct as_data *ad) | |
779 | { | |
780 | unsigned long batch = ad->batch_expire[REQ_ASYNC]; | |
781 | long write_time; | |
782 | ||
783 | write_time = (jiffies - ad->current_batch_expires) + batch; | |
784 | if (write_time < 0) | |
785 | write_time = 0; | |
786 | ||
787 | if (write_time > batch && !ad->write_batch_idled) { | |
788 | if (write_time > batch * 3) | |
789 | ad->write_batch_count /= 2; | |
790 | else | |
791 | ad->write_batch_count--; | |
792 | } else if (write_time < batch && ad->current_write_count == 0) { | |
793 | if (batch > write_time * 3) | |
794 | ad->write_batch_count *= 2; | |
795 | else | |
796 | ad->write_batch_count++; | |
797 | } | |
798 | ||
799 | if (ad->write_batch_count < 1) | |
800 | ad->write_batch_count = 1; | |
801 | } | |
802 | ||
803 | /* | |
804 | * as_completed_request is to be called when a request has completed and | |
805 | * returned something to the requesting process, be it an error or data. | |
806 | */ | |
807 | static void as_completed_request(request_queue_t *q, struct request *rq) | |
808 | { | |
809 | struct as_data *ad = q->elevator->elevator_data; | |
810 | struct as_rq *arq = RQ_DATA(rq); | |
811 | ||
812 | WARN_ON(!list_empty(&rq->queuelist)); | |
813 | ||
1da177e4 LT |
814 | if (arq->state != AS_RQ_REMOVED) { |
815 | printk("arq->state %d\n", arq->state); | |
816 | WARN_ON(1); | |
817 | goto out; | |
818 | } | |
819 | ||
1da177e4 LT |
820 | if (ad->changed_batch && ad->nr_dispatched == 1) { |
821 | kblockd_schedule_work(&ad->antic_work); | |
822 | ad->changed_batch = 0; | |
823 | ||
824 | if (ad->batch_data_dir == REQ_SYNC) | |
825 | ad->new_batch = 1; | |
826 | } | |
827 | WARN_ON(ad->nr_dispatched == 0); | |
828 | ad->nr_dispatched--; | |
829 | ||
830 | /* | |
831 | * Start counting the batch from when a request of that direction is | |
832 | * actually serviced. This should help devices with big TCQ windows | |
833 | * and writeback caches | |
834 | */ | |
9e2585a8 | 835 | if (ad->new_batch && ad->batch_data_dir == rq_is_sync(rq)) { |
1da177e4 LT |
836 | update_write_batch(ad); |
837 | ad->current_batch_expires = jiffies + | |
838 | ad->batch_expire[REQ_SYNC]; | |
839 | ad->new_batch = 0; | |
840 | } | |
841 | ||
842 | if (ad->io_context == arq->io_context && ad->io_context) { | |
843 | ad->antic_start = jiffies; | |
844 | ad->ioc_finished = 1; | |
845 | if (ad->antic_status == ANTIC_WAIT_REQ) { | |
846 | /* | |
847 | * We were waiting on this request, now anticipate | |
848 | * the next one | |
849 | */ | |
850 | as_antic_waitnext(ad); | |
851 | } | |
852 | } | |
853 | ||
b4878f24 | 854 | as_put_io_context(arq); |
1da177e4 LT |
855 | out: |
856 | arq->state = AS_RQ_POSTSCHED; | |
857 | } | |
858 | ||
859 | /* | |
860 | * as_remove_queued_request removes a request from the pre dispatch queue | |
861 | * without updating refcounts. It is expected the caller will drop the | |
862 | * reference unless it replaces the request at somepart of the elevator | |
863 | * (ie. the dispatch queue) | |
864 | */ | |
865 | static void as_remove_queued_request(request_queue_t *q, struct request *rq) | |
866 | { | |
867 | struct as_rq *arq = RQ_DATA(rq); | |
9e2585a8 | 868 | const int data_dir = rq_is_sync(rq); |
1da177e4 LT |
869 | struct as_data *ad = q->elevator->elevator_data; |
870 | ||
871 | WARN_ON(arq->state != AS_RQ_QUEUED); | |
872 | ||
873 | if (arq->io_context && arq->io_context->aic) { | |
874 | BUG_ON(!atomic_read(&arq->io_context->aic->nr_queued)); | |
875 | atomic_dec(&arq->io_context->aic->nr_queued); | |
876 | } | |
877 | ||
878 | /* | |
879 | * Update the "next_arq" cache if we are about to remove its | |
880 | * entry | |
881 | */ | |
882 | if (ad->next_arq[data_dir] == arq) | |
883 | ad->next_arq[data_dir] = as_find_next_arq(ad, arq); | |
884 | ||
d4f2f462 | 885 | rq_fifo_clear(rq); |
e37f346e | 886 | as_del_arq_rb(ad, rq); |
1da177e4 LT |
887 | } |
888 | ||
1da177e4 LT |
889 | /* |
890 | * as_fifo_expired returns 0 if there are no expired reads on the fifo, | |
891 | * 1 otherwise. It is ratelimited so that we only perform the check once per | |
892 | * `fifo_expire' interval. Otherwise a large number of expired requests | |
893 | * would create a hopeless seekstorm. | |
894 | * | |
895 | * See as_antic_expired comment. | |
896 | */ | |
897 | static int as_fifo_expired(struct as_data *ad, int adir) | |
898 | { | |
d4f2f462 | 899 | struct request *rq; |
1da177e4 LT |
900 | long delta_jif; |
901 | ||
902 | delta_jif = jiffies - ad->last_check_fifo[adir]; | |
903 | if (unlikely(delta_jif < 0)) | |
904 | delta_jif = -delta_jif; | |
905 | if (delta_jif < ad->fifo_expire[adir]) | |
906 | return 0; | |
907 | ||
908 | ad->last_check_fifo[adir] = jiffies; | |
909 | ||
910 | if (list_empty(&ad->fifo_list[adir])) | |
911 | return 0; | |
912 | ||
d4f2f462 | 913 | rq = rq_entry_fifo(ad->fifo_list[adir].next); |
1da177e4 | 914 | |
d4f2f462 | 915 | return time_after(jiffies, rq_fifo_time(rq)); |
1da177e4 LT |
916 | } |
917 | ||
918 | /* | |
919 | * as_batch_expired returns true if the current batch has expired. A batch | |
920 | * is a set of reads or a set of writes. | |
921 | */ | |
922 | static inline int as_batch_expired(struct as_data *ad) | |
923 | { | |
924 | if (ad->changed_batch || ad->new_batch) | |
925 | return 0; | |
926 | ||
927 | if (ad->batch_data_dir == REQ_SYNC) | |
928 | /* TODO! add a check so a complete fifo gets written? */ | |
929 | return time_after(jiffies, ad->current_batch_expires); | |
930 | ||
931 | return time_after(jiffies, ad->current_batch_expires) | |
932 | || ad->current_write_count == 0; | |
933 | } | |
934 | ||
935 | /* | |
936 | * move an entry to dispatch queue | |
937 | */ | |
938 | static void as_move_to_dispatch(struct as_data *ad, struct as_rq *arq) | |
939 | { | |
940 | struct request *rq = arq->request; | |
9e2585a8 | 941 | const int data_dir = rq_is_sync(rq); |
1da177e4 | 942 | |
e37f346e | 943 | BUG_ON(RB_EMPTY_NODE(&rq->rb_node)); |
1da177e4 LT |
944 | |
945 | as_antic_stop(ad); | |
946 | ad->antic_status = ANTIC_OFF; | |
947 | ||
948 | /* | |
949 | * This has to be set in order to be correctly updated by | |
950 | * as_find_next_arq | |
951 | */ | |
952 | ad->last_sector[data_dir] = rq->sector + rq->nr_sectors; | |
953 | ||
954 | if (data_dir == REQ_SYNC) { | |
955 | /* In case we have to anticipate after this */ | |
956 | copy_io_context(&ad->io_context, &arq->io_context); | |
957 | } else { | |
958 | if (ad->io_context) { | |
959 | put_io_context(ad->io_context); | |
960 | ad->io_context = NULL; | |
961 | } | |
962 | ||
963 | if (ad->current_write_count != 0) | |
964 | ad->current_write_count--; | |
965 | } | |
966 | ad->ioc_finished = 0; | |
967 | ||
968 | ad->next_arq[data_dir] = as_find_next_arq(ad, arq); | |
969 | ||
970 | /* | |
971 | * take it off the sort and fifo list, add to dispatch queue | |
972 | */ | |
1da177e4 LT |
973 | as_remove_queued_request(ad->q, rq); |
974 | WARN_ON(arq->state != AS_RQ_QUEUED); | |
975 | ||
b4878f24 JA |
976 | elv_dispatch_sort(ad->q, rq); |
977 | ||
1da177e4 LT |
978 | arq->state = AS_RQ_DISPATCHED; |
979 | if (arq->io_context && arq->io_context->aic) | |
980 | atomic_inc(&arq->io_context->aic->nr_dispatched); | |
981 | ad->nr_dispatched++; | |
982 | } | |
983 | ||
984 | /* | |
985 | * as_dispatch_request selects the best request according to | |
986 | * read/write expire, batch expire, etc, and moves it to the dispatch | |
987 | * queue. Returns 1 if a request was found, 0 otherwise. | |
988 | */ | |
b4878f24 | 989 | static int as_dispatch_request(request_queue_t *q, int force) |
1da177e4 | 990 | { |
b4878f24 | 991 | struct as_data *ad = q->elevator->elevator_data; |
1da177e4 LT |
992 | struct as_rq *arq; |
993 | const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]); | |
994 | const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]); | |
995 | ||
b4878f24 JA |
996 | if (unlikely(force)) { |
997 | /* | |
998 | * Forced dispatch, accounting is useless. Reset | |
999 | * accounting states and dump fifo_lists. Note that | |
1000 | * batch_data_dir is reset to REQ_SYNC to avoid | |
1001 | * screwing write batch accounting as write batch | |
1002 | * accounting occurs on W->R transition. | |
1003 | */ | |
1004 | int dispatched = 0; | |
1005 | ||
1006 | ad->batch_data_dir = REQ_SYNC; | |
1007 | ad->changed_batch = 0; | |
1008 | ad->new_batch = 0; | |
1009 | ||
1010 | while (ad->next_arq[REQ_SYNC]) { | |
1011 | as_move_to_dispatch(ad, ad->next_arq[REQ_SYNC]); | |
1012 | dispatched++; | |
1013 | } | |
1014 | ad->last_check_fifo[REQ_SYNC] = jiffies; | |
1015 | ||
1016 | while (ad->next_arq[REQ_ASYNC]) { | |
1017 | as_move_to_dispatch(ad, ad->next_arq[REQ_ASYNC]); | |
1018 | dispatched++; | |
1019 | } | |
1020 | ad->last_check_fifo[REQ_ASYNC] = jiffies; | |
1021 | ||
1022 | return dispatched; | |
1023 | } | |
1024 | ||
1da177e4 LT |
1025 | /* Signal that the write batch was uncontended, so we can't time it */ |
1026 | if (ad->batch_data_dir == REQ_ASYNC && !reads) { | |
1027 | if (ad->current_write_count == 0 || !writes) | |
1028 | ad->write_batch_idled = 1; | |
1029 | } | |
1030 | ||
1031 | if (!(reads || writes) | |
1032 | || ad->antic_status == ANTIC_WAIT_REQ | |
1033 | || ad->antic_status == ANTIC_WAIT_NEXT | |
1034 | || ad->changed_batch) | |
1035 | return 0; | |
1036 | ||
f5b3db00 | 1037 | if (!(reads && writes && as_batch_expired(ad))) { |
1da177e4 LT |
1038 | /* |
1039 | * batch is still running or no reads or no writes | |
1040 | */ | |
1041 | arq = ad->next_arq[ad->batch_data_dir]; | |
1042 | ||
1043 | if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) { | |
1044 | if (as_fifo_expired(ad, REQ_SYNC)) | |
1045 | goto fifo_expired; | |
1046 | ||
1047 | if (as_can_anticipate(ad, arq)) { | |
1048 | as_antic_waitreq(ad); | |
1049 | return 0; | |
1050 | } | |
1051 | } | |
1052 | ||
1053 | if (arq) { | |
1054 | /* we have a "next request" */ | |
1055 | if (reads && !writes) | |
1056 | ad->current_batch_expires = | |
1057 | jiffies + ad->batch_expire[REQ_SYNC]; | |
1058 | goto dispatch_request; | |
1059 | } | |
1060 | } | |
1061 | ||
1062 | /* | |
1063 | * at this point we are not running a batch. select the appropriate | |
1064 | * data direction (read / write) | |
1065 | */ | |
1066 | ||
1067 | if (reads) { | |
dd67d051 | 1068 | BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_SYNC])); |
1da177e4 LT |
1069 | |
1070 | if (writes && ad->batch_data_dir == REQ_SYNC) | |
1071 | /* | |
1072 | * Last batch was a read, switch to writes | |
1073 | */ | |
1074 | goto dispatch_writes; | |
1075 | ||
1076 | if (ad->batch_data_dir == REQ_ASYNC) { | |
1077 | WARN_ON(ad->new_batch); | |
1078 | ad->changed_batch = 1; | |
1079 | } | |
1080 | ad->batch_data_dir = REQ_SYNC; | |
d4f2f462 | 1081 | arq = RQ_DATA(rq_entry_fifo(ad->fifo_list[REQ_SYNC].next)); |
1da177e4 LT |
1082 | ad->last_check_fifo[ad->batch_data_dir] = jiffies; |
1083 | goto dispatch_request; | |
1084 | } | |
1085 | ||
1086 | /* | |
1087 | * the last batch was a read | |
1088 | */ | |
1089 | ||
1090 | if (writes) { | |
1091 | dispatch_writes: | |
dd67d051 | 1092 | BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_ASYNC])); |
1da177e4 LT |
1093 | |
1094 | if (ad->batch_data_dir == REQ_SYNC) { | |
1095 | ad->changed_batch = 1; | |
1096 | ||
1097 | /* | |
1098 | * new_batch might be 1 when the queue runs out of | |
1099 | * reads. A subsequent submission of a write might | |
1100 | * cause a change of batch before the read is finished. | |
1101 | */ | |
1102 | ad->new_batch = 0; | |
1103 | } | |
1104 | ad->batch_data_dir = REQ_ASYNC; | |
1105 | ad->current_write_count = ad->write_batch_count; | |
1106 | ad->write_batch_idled = 0; | |
1107 | arq = ad->next_arq[ad->batch_data_dir]; | |
1108 | goto dispatch_request; | |
1109 | } | |
1110 | ||
1111 | BUG(); | |
1112 | return 0; | |
1113 | ||
1114 | dispatch_request: | |
1115 | /* | |
1116 | * If a request has expired, service it. | |
1117 | */ | |
1118 | ||
1119 | if (as_fifo_expired(ad, ad->batch_data_dir)) { | |
1120 | fifo_expired: | |
d4f2f462 | 1121 | arq = RQ_DATA(rq_entry_fifo(ad->fifo_list[ad->batch_data_dir].next)); |
1da177e4 LT |
1122 | } |
1123 | ||
1124 | if (ad->changed_batch) { | |
1125 | WARN_ON(ad->new_batch); | |
1126 | ||
1127 | if (ad->nr_dispatched) | |
1128 | return 0; | |
1129 | ||
1130 | if (ad->batch_data_dir == REQ_ASYNC) | |
1131 | ad->current_batch_expires = jiffies + | |
1132 | ad->batch_expire[REQ_ASYNC]; | |
1133 | else | |
1134 | ad->new_batch = 1; | |
1135 | ||
1136 | ad->changed_batch = 0; | |
1137 | } | |
1138 | ||
1139 | /* | |
1140 | * arq is the selected appropriate request. | |
1141 | */ | |
1142 | as_move_to_dispatch(ad, arq); | |
1143 | ||
1144 | return 1; | |
1145 | } | |
1146 | ||
1da177e4 LT |
1147 | /* |
1148 | * add arq to rbtree and fifo | |
1149 | */ | |
b4878f24 | 1150 | static void as_add_request(request_queue_t *q, struct request *rq) |
1da177e4 | 1151 | { |
b4878f24 JA |
1152 | struct as_data *ad = q->elevator->elevator_data; |
1153 | struct as_rq *arq = RQ_DATA(rq); | |
1da177e4 LT |
1154 | int data_dir; |
1155 | ||
b4878f24 JA |
1156 | arq->state = AS_RQ_NEW; |
1157 | ||
9e2585a8 | 1158 | data_dir = rq_is_sync(rq); |
1da177e4 LT |
1159 | |
1160 | arq->io_context = as_get_io_context(); | |
1161 | ||
1162 | if (arq->io_context) { | |
1163 | as_update_iohist(ad, arq->io_context->aic, arq->request); | |
1164 | atomic_inc(&arq->io_context->aic->nr_queued); | |
1165 | } | |
1166 | ||
e37f346e | 1167 | as_add_arq_rb(ad, rq); |
1da177e4 | 1168 | |
ef9be1d3 TH |
1169 | /* |
1170 | * set expire time (only used for reads) and add to fifo list | |
1171 | */ | |
d4f2f462 JA |
1172 | rq_set_fifo_time(rq, jiffies + ad->fifo_expire[data_dir]); |
1173 | list_add_tail(&rq->queuelist, &ad->fifo_list[data_dir]); | |
1da177e4 | 1174 | |
ef9be1d3 | 1175 | as_update_arq(ad, arq); /* keep state machine up to date */ |
1da177e4 LT |
1176 | arq->state = AS_RQ_QUEUED; |
1177 | } | |
1178 | ||
b4878f24 | 1179 | static void as_activate_request(request_queue_t *q, struct request *rq) |
1da177e4 | 1180 | { |
1da177e4 LT |
1181 | struct as_rq *arq = RQ_DATA(rq); |
1182 | ||
b4878f24 JA |
1183 | WARN_ON(arq->state != AS_RQ_DISPATCHED); |
1184 | arq->state = AS_RQ_REMOVED; | |
1185 | if (arq->io_context && arq->io_context->aic) | |
1186 | atomic_dec(&arq->io_context->aic->nr_dispatched); | |
1da177e4 LT |
1187 | } |
1188 | ||
b4878f24 | 1189 | static void as_deactivate_request(request_queue_t *q, struct request *rq) |
1da177e4 | 1190 | { |
1da177e4 LT |
1191 | struct as_rq *arq = RQ_DATA(rq); |
1192 | ||
b4878f24 JA |
1193 | WARN_ON(arq->state != AS_RQ_REMOVED); |
1194 | arq->state = AS_RQ_DISPATCHED; | |
1195 | if (arq->io_context && arq->io_context->aic) | |
1196 | atomic_inc(&arq->io_context->aic->nr_dispatched); | |
1da177e4 LT |
1197 | } |
1198 | ||
1199 | /* | |
1200 | * as_queue_empty tells us if there are requests left in the device. It may | |
1201 | * not be the case that a driver can get the next request even if the queue | |
1202 | * is not empty - it is used in the block layer to check for plugging and | |
1203 | * merging opportunities | |
1204 | */ | |
1205 | static int as_queue_empty(request_queue_t *q) | |
1206 | { | |
1207 | struct as_data *ad = q->elevator->elevator_data; | |
1208 | ||
b4878f24 JA |
1209 | return list_empty(&ad->fifo_list[REQ_ASYNC]) |
1210 | && list_empty(&ad->fifo_list[REQ_SYNC]); | |
1da177e4 LT |
1211 | } |
1212 | ||
1da177e4 LT |
1213 | static int |
1214 | as_merge(request_queue_t *q, struct request **req, struct bio *bio) | |
1215 | { | |
1216 | struct as_data *ad = q->elevator->elevator_data; | |
1217 | sector_t rb_key = bio->bi_sector + bio_sectors(bio); | |
1218 | struct request *__rq; | |
1da177e4 LT |
1219 | |
1220 | /* | |
1221 | * check for front merge | |
1222 | */ | |
e37f346e | 1223 | __rq = elv_rb_find(&ad->sort_list[bio_data_dir(bio)], rb_key); |
9817064b JA |
1224 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
1225 | *req = __rq; | |
1226 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
1227 | } |
1228 | ||
1229 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
1230 | } |
1231 | ||
e37f346e | 1232 | static void as_merged_request(request_queue_t *q, struct request *req, int type) |
1da177e4 LT |
1233 | { |
1234 | struct as_data *ad = q->elevator->elevator_data; | |
1da177e4 | 1235 | |
1da177e4 LT |
1236 | /* |
1237 | * if the merge was a front merge, we need to reposition request | |
1238 | */ | |
e37f346e JA |
1239 | if (type == ELEVATOR_FRONT_MERGE) { |
1240 | as_del_arq_rb(ad, req); | |
1241 | as_add_arq_rb(ad, req); | |
1da177e4 LT |
1242 | /* |
1243 | * Note! At this stage of this and the next function, our next | |
1244 | * request may not be optimal - eg the request may have "grown" | |
1245 | * behind the disk head. We currently don't bother adjusting. | |
1246 | */ | |
1247 | } | |
1da177e4 LT |
1248 | } |
1249 | ||
f5b3db00 NP |
1250 | static void as_merged_requests(request_queue_t *q, struct request *req, |
1251 | struct request *next) | |
1da177e4 | 1252 | { |
1da177e4 LT |
1253 | struct as_rq *arq = RQ_DATA(req); |
1254 | struct as_rq *anext = RQ_DATA(next); | |
1255 | ||
1256 | BUG_ON(!arq); | |
1257 | BUG_ON(!anext); | |
1258 | ||
1da177e4 LT |
1259 | /* |
1260 | * if anext expires before arq, assign its expire time to arq | |
1261 | * and move into anext position (anext will be deleted) in fifo | |
1262 | */ | |
d4f2f462 JA |
1263 | if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) { |
1264 | if (time_before(rq_fifo_time(next), rq_fifo_time(req))) { | |
1265 | list_move(&req->queuelist, &next->queuelist); | |
1266 | rq_set_fifo_time(req, rq_fifo_time(next)); | |
1da177e4 LT |
1267 | /* |
1268 | * Don't copy here but swap, because when anext is | |
1269 | * removed below, it must contain the unused context | |
1270 | */ | |
1271 | swap_io_context(&arq->io_context, &anext->io_context); | |
1272 | } | |
1273 | } | |
1274 | ||
1da177e4 LT |
1275 | /* |
1276 | * kill knowledge of next, this one is a goner | |
1277 | */ | |
1278 | as_remove_queued_request(q, next); | |
b4878f24 | 1279 | as_put_io_context(anext); |
1da177e4 LT |
1280 | |
1281 | anext->state = AS_RQ_MERGED; | |
1282 | } | |
1283 | ||
1284 | /* | |
1285 | * This is executed in a "deferred" process context, by kblockd. It calls the | |
1286 | * driver's request_fn so the driver can submit that request. | |
1287 | * | |
1288 | * IMPORTANT! This guy will reenter the elevator, so set up all queue global | |
1289 | * state before calling, and don't rely on any state over calls. | |
1290 | * | |
1291 | * FIXME! dispatch queue is not a queue at all! | |
1292 | */ | |
1293 | static void as_work_handler(void *data) | |
1294 | { | |
1295 | struct request_queue *q = data; | |
1296 | unsigned long flags; | |
1297 | ||
1298 | spin_lock_irqsave(q->queue_lock, flags); | |
b4878f24 | 1299 | if (!as_queue_empty(q)) |
1da177e4 LT |
1300 | q->request_fn(q); |
1301 | spin_unlock_irqrestore(q->queue_lock, flags); | |
1302 | } | |
1303 | ||
1304 | static void as_put_request(request_queue_t *q, struct request *rq) | |
1305 | { | |
1306 | struct as_data *ad = q->elevator->elevator_data; | |
1307 | struct as_rq *arq = RQ_DATA(rq); | |
1308 | ||
1309 | if (!arq) { | |
1310 | WARN_ON(1); | |
1311 | return; | |
1312 | } | |
1313 | ||
b4878f24 JA |
1314 | if (unlikely(arq->state != AS_RQ_POSTSCHED && |
1315 | arq->state != AS_RQ_PRESCHED && | |
1316 | arq->state != AS_RQ_MERGED)) { | |
1da177e4 LT |
1317 | printk("arq->state %d\n", arq->state); |
1318 | WARN_ON(1); | |
1319 | } | |
1320 | ||
1321 | mempool_free(arq, ad->arq_pool); | |
1322 | rq->elevator_private = NULL; | |
1323 | } | |
1324 | ||
22e2c507 | 1325 | static int as_set_request(request_queue_t *q, struct request *rq, |
8267e268 | 1326 | struct bio *bio, gfp_t gfp_mask) |
1da177e4 LT |
1327 | { |
1328 | struct as_data *ad = q->elevator->elevator_data; | |
1329 | struct as_rq *arq = mempool_alloc(ad->arq_pool, gfp_mask); | |
1330 | ||
1331 | if (arq) { | |
1332 | memset(arq, 0, sizeof(*arq)); | |
1da177e4 LT |
1333 | arq->request = rq; |
1334 | arq->state = AS_RQ_PRESCHED; | |
1335 | arq->io_context = NULL; | |
1da177e4 LT |
1336 | rq->elevator_private = arq; |
1337 | return 0; | |
1338 | } | |
1339 | ||
1340 | return 1; | |
1341 | } | |
1342 | ||
22e2c507 | 1343 | static int as_may_queue(request_queue_t *q, int rw, struct bio *bio) |
1da177e4 LT |
1344 | { |
1345 | int ret = ELV_MQUEUE_MAY; | |
1346 | struct as_data *ad = q->elevator->elevator_data; | |
1347 | struct io_context *ioc; | |
1348 | if (ad->antic_status == ANTIC_WAIT_REQ || | |
1349 | ad->antic_status == ANTIC_WAIT_NEXT) { | |
1350 | ioc = as_get_io_context(); | |
1351 | if (ad->io_context == ioc) | |
1352 | ret = ELV_MQUEUE_MUST; | |
1353 | put_io_context(ioc); | |
1354 | } | |
1355 | ||
1356 | return ret; | |
1357 | } | |
1358 | ||
1359 | static void as_exit_queue(elevator_t *e) | |
1360 | { | |
1361 | struct as_data *ad = e->elevator_data; | |
1362 | ||
1363 | del_timer_sync(&ad->antic_timer); | |
1364 | kblockd_flush(); | |
1365 | ||
1366 | BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC])); | |
1367 | BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC])); | |
1368 | ||
1369 | mempool_destroy(ad->arq_pool); | |
1370 | put_io_context(ad->io_context); | |
1da177e4 LT |
1371 | kfree(ad); |
1372 | } | |
1373 | ||
1374 | /* | |
1375 | * initialize elevator private data (as_data), and alloc a arq for | |
1376 | * each request on the free lists | |
1377 | */ | |
bc1c1169 | 1378 | static void *as_init_queue(request_queue_t *q, elevator_t *e) |
1da177e4 LT |
1379 | { |
1380 | struct as_data *ad; | |
1da177e4 LT |
1381 | |
1382 | if (!arq_pool) | |
bc1c1169 | 1383 | return NULL; |
1da177e4 | 1384 | |
1946089a | 1385 | ad = kmalloc_node(sizeof(*ad), GFP_KERNEL, q->node); |
1da177e4 | 1386 | if (!ad) |
bc1c1169 | 1387 | return NULL; |
1da177e4 LT |
1388 | memset(ad, 0, sizeof(*ad)); |
1389 | ||
1390 | ad->q = q; /* Identify what queue the data belongs to */ | |
1391 | ||
1946089a CL |
1392 | ad->arq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab, |
1393 | mempool_free_slab, arq_pool, q->node); | |
1da177e4 | 1394 | if (!ad->arq_pool) { |
1da177e4 | 1395 | kfree(ad); |
bc1c1169 | 1396 | return NULL; |
1da177e4 LT |
1397 | } |
1398 | ||
1399 | /* anticipatory scheduling helpers */ | |
1400 | ad->antic_timer.function = as_antic_timeout; | |
1401 | ad->antic_timer.data = (unsigned long)q; | |
1402 | init_timer(&ad->antic_timer); | |
1403 | INIT_WORK(&ad->antic_work, as_work_handler, q); | |
1404 | ||
1da177e4 LT |
1405 | INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]); |
1406 | INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]); | |
1407 | ad->sort_list[REQ_SYNC] = RB_ROOT; | |
1408 | ad->sort_list[REQ_ASYNC] = RB_ROOT; | |
1da177e4 LT |
1409 | ad->fifo_expire[REQ_SYNC] = default_read_expire; |
1410 | ad->fifo_expire[REQ_ASYNC] = default_write_expire; | |
1411 | ad->antic_expire = default_antic_expire; | |
1412 | ad->batch_expire[REQ_SYNC] = default_read_batch_expire; | |
1413 | ad->batch_expire[REQ_ASYNC] = default_write_batch_expire; | |
1da177e4 LT |
1414 | |
1415 | ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC]; | |
1416 | ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10; | |
1417 | if (ad->write_batch_count < 2) | |
1418 | ad->write_batch_count = 2; | |
1419 | ||
bc1c1169 | 1420 | return ad; |
1da177e4 LT |
1421 | } |
1422 | ||
1423 | /* | |
1424 | * sysfs parts below | |
1425 | */ | |
1da177e4 LT |
1426 | |
1427 | static ssize_t | |
1428 | as_var_show(unsigned int var, char *page) | |
1429 | { | |
1da177e4 LT |
1430 | return sprintf(page, "%d\n", var); |
1431 | } | |
1432 | ||
1433 | static ssize_t | |
1434 | as_var_store(unsigned long *var, const char *page, size_t count) | |
1435 | { | |
1da177e4 LT |
1436 | char *p = (char *) page; |
1437 | ||
c9b3ad67 | 1438 | *var = simple_strtoul(p, &p, 10); |
1da177e4 LT |
1439 | return count; |
1440 | } | |
1441 | ||
e572ec7e | 1442 | static ssize_t est_time_show(elevator_t *e, char *page) |
1da177e4 | 1443 | { |
3d1ab40f | 1444 | struct as_data *ad = e->elevator_data; |
1da177e4 LT |
1445 | int pos = 0; |
1446 | ||
f5b3db00 NP |
1447 | pos += sprintf(page+pos, "%lu %% exit probability\n", |
1448 | 100*ad->exit_prob/256); | |
1449 | pos += sprintf(page+pos, "%lu %% probability of exiting without a " | |
1450 | "cooperating process submitting IO\n", | |
1451 | 100*ad->exit_no_coop/256); | |
1da177e4 | 1452 | pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean); |
f5b3db00 NP |
1453 | pos += sprintf(page+pos, "%llu sectors new seek distance\n", |
1454 | (unsigned long long)ad->new_seek_mean); | |
1da177e4 LT |
1455 | |
1456 | return pos; | |
1457 | } | |
1458 | ||
1459 | #define SHOW_FUNCTION(__FUNC, __VAR) \ | |
3d1ab40f | 1460 | static ssize_t __FUNC(elevator_t *e, char *page) \ |
1da177e4 | 1461 | { \ |
3d1ab40f | 1462 | struct as_data *ad = e->elevator_data; \ |
1da177e4 LT |
1463 | return as_var_show(jiffies_to_msecs((__VAR)), (page)); \ |
1464 | } | |
e572ec7e AV |
1465 | SHOW_FUNCTION(as_read_expire_show, ad->fifo_expire[REQ_SYNC]); |
1466 | SHOW_FUNCTION(as_write_expire_show, ad->fifo_expire[REQ_ASYNC]); | |
1467 | SHOW_FUNCTION(as_antic_expire_show, ad->antic_expire); | |
1468 | SHOW_FUNCTION(as_read_batch_expire_show, ad->batch_expire[REQ_SYNC]); | |
1469 | SHOW_FUNCTION(as_write_batch_expire_show, ad->batch_expire[REQ_ASYNC]); | |
1da177e4 LT |
1470 | #undef SHOW_FUNCTION |
1471 | ||
1472 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \ | |
3d1ab40f | 1473 | static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \ |
1da177e4 | 1474 | { \ |
3d1ab40f AV |
1475 | struct as_data *ad = e->elevator_data; \ |
1476 | int ret = as_var_store(__PTR, (page), count); \ | |
1da177e4 LT |
1477 | if (*(__PTR) < (MIN)) \ |
1478 | *(__PTR) = (MIN); \ | |
1479 | else if (*(__PTR) > (MAX)) \ | |
1480 | *(__PTR) = (MAX); \ | |
1481 | *(__PTR) = msecs_to_jiffies(*(__PTR)); \ | |
1482 | return ret; \ | |
1483 | } | |
e572ec7e AV |
1484 | STORE_FUNCTION(as_read_expire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX); |
1485 | STORE_FUNCTION(as_write_expire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX); | |
1486 | STORE_FUNCTION(as_antic_expire_store, &ad->antic_expire, 0, INT_MAX); | |
1487 | STORE_FUNCTION(as_read_batch_expire_store, | |
1da177e4 | 1488 | &ad->batch_expire[REQ_SYNC], 0, INT_MAX); |
e572ec7e | 1489 | STORE_FUNCTION(as_write_batch_expire_store, |
1da177e4 LT |
1490 | &ad->batch_expire[REQ_ASYNC], 0, INT_MAX); |
1491 | #undef STORE_FUNCTION | |
1492 | ||
e572ec7e AV |
1493 | #define AS_ATTR(name) \ |
1494 | __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store) | |
1495 | ||
1496 | static struct elv_fs_entry as_attrs[] = { | |
1497 | __ATTR_RO(est_time), | |
1498 | AS_ATTR(read_expire), | |
1499 | AS_ATTR(write_expire), | |
1500 | AS_ATTR(antic_expire), | |
1501 | AS_ATTR(read_batch_expire), | |
1502 | AS_ATTR(write_batch_expire), | |
1503 | __ATTR_NULL | |
1da177e4 LT |
1504 | }; |
1505 | ||
1da177e4 LT |
1506 | static struct elevator_type iosched_as = { |
1507 | .ops = { | |
1508 | .elevator_merge_fn = as_merge, | |
1509 | .elevator_merged_fn = as_merged_request, | |
1510 | .elevator_merge_req_fn = as_merged_requests, | |
b4878f24 JA |
1511 | .elevator_dispatch_fn = as_dispatch_request, |
1512 | .elevator_add_req_fn = as_add_request, | |
1513 | .elevator_activate_req_fn = as_activate_request, | |
1da177e4 LT |
1514 | .elevator_deactivate_req_fn = as_deactivate_request, |
1515 | .elevator_queue_empty_fn = as_queue_empty, | |
1516 | .elevator_completed_req_fn = as_completed_request, | |
e37f346e JA |
1517 | .elevator_former_req_fn = elv_rb_former_request, |
1518 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
1519 | .elevator_set_req_fn = as_set_request, |
1520 | .elevator_put_req_fn = as_put_request, | |
1521 | .elevator_may_queue_fn = as_may_queue, | |
1522 | .elevator_init_fn = as_init_queue, | |
1523 | .elevator_exit_fn = as_exit_queue, | |
e17a9489 | 1524 | .trim = as_trim, |
1da177e4 LT |
1525 | }, |
1526 | ||
3d1ab40f | 1527 | .elevator_attrs = as_attrs, |
1da177e4 LT |
1528 | .elevator_name = "anticipatory", |
1529 | .elevator_owner = THIS_MODULE, | |
1530 | }; | |
1531 | ||
1532 | static int __init as_init(void) | |
1533 | { | |
1534 | int ret; | |
1535 | ||
1536 | arq_pool = kmem_cache_create("as_arq", sizeof(struct as_rq), | |
1537 | 0, 0, NULL, NULL); | |
1538 | if (!arq_pool) | |
1539 | return -ENOMEM; | |
1540 | ||
1541 | ret = elv_register(&iosched_as); | |
1542 | if (!ret) { | |
1543 | /* | |
1544 | * don't allow AS to get unregistered, since we would have | |
1545 | * to browse all tasks in the system and release their | |
1546 | * as_io_context first | |
1547 | */ | |
1548 | __module_get(THIS_MODULE); | |
1549 | return 0; | |
1550 | } | |
1551 | ||
1552 | kmem_cache_destroy(arq_pool); | |
1553 | return ret; | |
1554 | } | |
1555 | ||
1556 | static void __exit as_exit(void) | |
1557 | { | |
334e94de | 1558 | DECLARE_COMPLETION(all_gone); |
1da177e4 | 1559 | elv_unregister(&iosched_as); |
334e94de | 1560 | ioc_gone = &all_gone; |
fba82272 OH |
1561 | /* ioc_gone's update must be visible before reading ioc_count */ |
1562 | smp_wmb(); | |
334e94de | 1563 | if (atomic_read(&ioc_count)) |
fba82272 | 1564 | wait_for_completion(ioc_gone); |
334e94de | 1565 | synchronize_rcu(); |
83521d3e | 1566 | kmem_cache_destroy(arq_pool); |
1da177e4 LT |
1567 | } |
1568 | ||
1569 | module_init(as_init); | |
1570 | module_exit(as_exit); | |
1571 | ||
1572 | MODULE_AUTHOR("Nick Piggin"); | |
1573 | MODULE_LICENSE("GPL"); | |
1574 | MODULE_DESCRIPTION("anticipatory IO scheduler"); |