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 | ||
8a8e674c | 95 | struct request *next_rq[2]; /* next in sort order */ |
1da177e4 | 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? */ | |
1da177e4 LT |
116 | |
117 | enum anticipation_status antic_status; | |
118 | unsigned long antic_start; /* jiffies: when it started */ | |
119 | struct timer_list antic_timer; /* anticipatory scheduling timer */ | |
120 | struct work_struct antic_work; /* Deferred unplugging */ | |
121 | struct io_context *io_context; /* Identify the expected process */ | |
122 | int ioc_finished; /* IO associated with io_context is finished */ | |
123 | int nr_dispatched; | |
124 | ||
125 | /* | |
126 | * settings that change how the i/o scheduler behaves | |
127 | */ | |
128 | unsigned long fifo_expire[2]; | |
129 | unsigned long batch_expire[2]; | |
130 | unsigned long antic_expire; | |
131 | }; | |
132 | ||
1da177e4 LT |
133 | /* |
134 | * per-request data. | |
135 | */ | |
136 | enum arq_state { | |
137 | AS_RQ_NEW=0, /* New - not referenced and not on any lists */ | |
138 | AS_RQ_QUEUED, /* In the request queue. It belongs to the | |
139 | scheduler */ | |
140 | AS_RQ_DISPATCHED, /* On the dispatch list. It belongs to the | |
141 | driver now */ | |
142 | AS_RQ_PRESCHED, /* Debug poisoning for requests being used */ | |
143 | AS_RQ_REMOVED, | |
144 | AS_RQ_MERGED, | |
145 | AS_RQ_POSTSCHED, /* when they shouldn't be */ | |
146 | }; | |
147 | ||
8a8e674c JA |
148 | #define RQ_IOC(rq) ((struct io_context *) (rq)->elevator_private) |
149 | #define RQ_STATE(rq) ((enum arq_state)(rq)->elevator_private2) | |
150 | #define RQ_SET_STATE(rq, state) ((rq)->elevator_private2 = (void *) state) | |
1da177e4 | 151 | |
334e94de AV |
152 | static atomic_t ioc_count = ATOMIC_INIT(0); |
153 | static struct completion *ioc_gone; | |
154 | ||
8a8e674c | 155 | static void as_move_to_dispatch(struct as_data *ad, struct request *rq); |
ef9be1d3 TH |
156 | static void as_antic_stop(struct as_data *ad); |
157 | ||
1da177e4 LT |
158 | /* |
159 | * IO Context helper functions | |
160 | */ | |
161 | ||
162 | /* Called to deallocate the as_io_context */ | |
163 | static void free_as_io_context(struct as_io_context *aic) | |
164 | { | |
165 | kfree(aic); | |
334e94de AV |
166 | if (atomic_dec_and_test(&ioc_count) && ioc_gone) |
167 | complete(ioc_gone); | |
1da177e4 LT |
168 | } |
169 | ||
e17a9489 AV |
170 | static void as_trim(struct io_context *ioc) |
171 | { | |
334e94de AV |
172 | if (ioc->aic) |
173 | free_as_io_context(ioc->aic); | |
e17a9489 AV |
174 | ioc->aic = NULL; |
175 | } | |
176 | ||
1da177e4 LT |
177 | /* Called when the task exits */ |
178 | static void exit_as_io_context(struct as_io_context *aic) | |
179 | { | |
180 | WARN_ON(!test_bit(AS_TASK_RUNNING, &aic->state)); | |
181 | clear_bit(AS_TASK_RUNNING, &aic->state); | |
182 | } | |
183 | ||
184 | static struct as_io_context *alloc_as_io_context(void) | |
185 | { | |
186 | struct as_io_context *ret; | |
187 | ||
188 | ret = kmalloc(sizeof(*ret), GFP_ATOMIC); | |
189 | if (ret) { | |
190 | ret->dtor = free_as_io_context; | |
191 | ret->exit = exit_as_io_context; | |
192 | ret->state = 1 << AS_TASK_RUNNING; | |
193 | atomic_set(&ret->nr_queued, 0); | |
194 | atomic_set(&ret->nr_dispatched, 0); | |
195 | spin_lock_init(&ret->lock); | |
196 | ret->ttime_total = 0; | |
197 | ret->ttime_samples = 0; | |
198 | ret->ttime_mean = 0; | |
199 | ret->seek_total = 0; | |
200 | ret->seek_samples = 0; | |
201 | ret->seek_mean = 0; | |
334e94de | 202 | atomic_inc(&ioc_count); |
1da177e4 LT |
203 | } |
204 | ||
205 | return ret; | |
206 | } | |
207 | ||
208 | /* | |
209 | * If the current task has no AS IO context then create one and initialise it. | |
210 | * Then take a ref on the task's io context and return it. | |
211 | */ | |
212 | static struct io_context *as_get_io_context(void) | |
213 | { | |
214 | struct io_context *ioc = get_io_context(GFP_ATOMIC); | |
215 | if (ioc && !ioc->aic) { | |
216 | ioc->aic = alloc_as_io_context(); | |
217 | if (!ioc->aic) { | |
218 | put_io_context(ioc); | |
219 | ioc = NULL; | |
220 | } | |
221 | } | |
222 | return ioc; | |
223 | } | |
224 | ||
8a8e674c | 225 | static void as_put_io_context(struct request *rq) |
b4878f24 JA |
226 | { |
227 | struct as_io_context *aic; | |
228 | ||
8a8e674c | 229 | if (unlikely(!RQ_IOC(rq))) |
b4878f24 JA |
230 | return; |
231 | ||
8a8e674c | 232 | aic = RQ_IOC(rq)->aic; |
b4878f24 | 233 | |
8a8e674c | 234 | if (rq_is_sync(rq) && aic) { |
b4878f24 JA |
235 | spin_lock(&aic->lock); |
236 | set_bit(AS_TASK_IORUNNING, &aic->state); | |
237 | aic->last_end_request = jiffies; | |
238 | spin_unlock(&aic->lock); | |
239 | } | |
240 | ||
8a8e674c | 241 | put_io_context(RQ_IOC(rq)); |
b4878f24 JA |
242 | } |
243 | ||
1da177e4 LT |
244 | /* |
245 | * rb tree support functions | |
246 | */ | |
9e2585a8 | 247 | #define RQ_RB_ROOT(ad, rq) (&(ad)->sort_list[rq_is_sync((rq))]) |
1da177e4 | 248 | |
8a8e674c | 249 | static void as_add_rq_rb(struct as_data *ad, struct request *rq) |
ef9be1d3 | 250 | { |
e37f346e | 251 | struct request *alias; |
ef9be1d3 | 252 | |
9e2585a8 | 253 | while ((unlikely(alias = elv_rb_add(RQ_RB_ROOT(ad, rq), rq)))) { |
8a8e674c | 254 | as_move_to_dispatch(ad, alias); |
ef9be1d3 TH |
255 | as_antic_stop(ad); |
256 | } | |
257 | } | |
258 | ||
8a8e674c | 259 | static inline void as_del_rq_rb(struct as_data *ad, struct request *rq) |
1da177e4 | 260 | { |
9e2585a8 | 261 | elv_rb_del(RQ_RB_ROOT(ad, rq), rq); |
1da177e4 LT |
262 | } |
263 | ||
264 | /* | |
265 | * IO Scheduler proper | |
266 | */ | |
267 | ||
268 | #define MAXBACK (1024 * 1024) /* | |
269 | * Maximum distance the disk will go backward | |
270 | * for a request. | |
271 | */ | |
272 | ||
273 | #define BACK_PENALTY 2 | |
274 | ||
275 | /* | |
276 | * as_choose_req selects the preferred one of two requests of the same data_dir | |
277 | * ignoring time - eg. timeouts, which is the job of as_dispatch_request | |
278 | */ | |
8a8e674c JA |
279 | static struct request * |
280 | as_choose_req(struct as_data *ad, struct request *rq1, struct request *rq2) | |
1da177e4 LT |
281 | { |
282 | int data_dir; | |
283 | sector_t last, s1, s2, d1, d2; | |
284 | int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */ | |
285 | const sector_t maxback = MAXBACK; | |
286 | ||
8a8e674c JA |
287 | if (rq1 == NULL || rq1 == rq2) |
288 | return rq2; | |
289 | if (rq2 == NULL) | |
290 | return rq1; | |
1da177e4 | 291 | |
8a8e674c | 292 | data_dir = rq_is_sync(rq1); |
1da177e4 LT |
293 | |
294 | last = ad->last_sector[data_dir]; | |
8a8e674c JA |
295 | s1 = rq1->sector; |
296 | s2 = rq2->sector; | |
1da177e4 | 297 | |
8a8e674c | 298 | BUG_ON(data_dir != rq_is_sync(rq2)); |
1da177e4 LT |
299 | |
300 | /* | |
301 | * Strict one way elevator _except_ in the case where we allow | |
302 | * short backward seeks which are biased as twice the cost of a | |
303 | * similar forward seek. | |
304 | */ | |
305 | if (s1 >= last) | |
306 | d1 = s1 - last; | |
307 | else if (s1+maxback >= last) | |
308 | d1 = (last - s1)*BACK_PENALTY; | |
309 | else { | |
310 | r1_wrap = 1; | |
311 | d1 = 0; /* shut up, gcc */ | |
312 | } | |
313 | ||
314 | if (s2 >= last) | |
315 | d2 = s2 - last; | |
316 | else if (s2+maxback >= last) | |
317 | d2 = (last - s2)*BACK_PENALTY; | |
318 | else { | |
319 | r2_wrap = 1; | |
320 | d2 = 0; | |
321 | } | |
322 | ||
323 | /* Found required data */ | |
324 | if (!r1_wrap && r2_wrap) | |
8a8e674c | 325 | return rq1; |
1da177e4 | 326 | else if (!r2_wrap && r1_wrap) |
8a8e674c | 327 | return rq2; |
1da177e4 LT |
328 | else if (r1_wrap && r2_wrap) { |
329 | /* both behind the head */ | |
330 | if (s1 <= s2) | |
8a8e674c | 331 | return rq1; |
1da177e4 | 332 | else |
8a8e674c | 333 | return rq2; |
1da177e4 LT |
334 | } |
335 | ||
336 | /* Both requests in front of the head */ | |
337 | if (d1 < d2) | |
8a8e674c | 338 | return rq1; |
1da177e4 | 339 | else if (d2 < d1) |
8a8e674c | 340 | return rq2; |
1da177e4 LT |
341 | else { |
342 | if (s1 >= s2) | |
8a8e674c | 343 | return rq1; |
1da177e4 | 344 | else |
8a8e674c | 345 | return rq2; |
1da177e4 LT |
346 | } |
347 | } | |
348 | ||
349 | /* | |
8a8e674c | 350 | * as_find_next_rq finds the next request after @prev in elevator order. |
1da177e4 LT |
351 | * this with as_choose_req form the basis for how the scheduler chooses |
352 | * what request to process next. Anticipation works on top of this. | |
353 | */ | |
8a8e674c JA |
354 | static struct request * |
355 | as_find_next_rq(struct as_data *ad, struct request *last) | |
1da177e4 | 356 | { |
1da177e4 LT |
357 | struct rb_node *rbnext = rb_next(&last->rb_node); |
358 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
8a8e674c | 359 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 360 | |
e37f346e | 361 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
362 | |
363 | if (rbprev) | |
8a8e674c | 364 | prev = rb_entry_rq(rbprev); |
1da177e4 LT |
365 | |
366 | if (rbnext) | |
8a8e674c | 367 | next = rb_entry_rq(rbnext); |
1da177e4 | 368 | else { |
9e2585a8 | 369 | const int data_dir = rq_is_sync(last); |
1da177e4 | 370 | |
e37f346e JA |
371 | rbnext = rb_first(&ad->sort_list[data_dir]); |
372 | if (rbnext && rbnext != &last->rb_node) | |
8a8e674c | 373 | next = rb_entry_rq(rbnext); |
e37f346e | 374 | } |
1da177e4 | 375 | |
e37f346e | 376 | return as_choose_req(ad, next, prev); |
1da177e4 LT |
377 | } |
378 | ||
379 | /* | |
380 | * anticipatory scheduling functions follow | |
381 | */ | |
382 | ||
383 | /* | |
384 | * as_antic_expired tells us when we have anticipated too long. | |
385 | * The funny "absolute difference" math on the elapsed time is to handle | |
386 | * jiffy wraps, and disks which have been idle for 0x80000000 jiffies. | |
387 | */ | |
388 | static int as_antic_expired(struct as_data *ad) | |
389 | { | |
390 | long delta_jif; | |
391 | ||
392 | delta_jif = jiffies - ad->antic_start; | |
393 | if (unlikely(delta_jif < 0)) | |
394 | delta_jif = -delta_jif; | |
395 | if (delta_jif < ad->antic_expire) | |
396 | return 0; | |
397 | ||
398 | return 1; | |
399 | } | |
400 | ||
401 | /* | |
402 | * as_antic_waitnext starts anticipating that a nice request will soon be | |
403 | * submitted. See also as_antic_waitreq | |
404 | */ | |
405 | static void as_antic_waitnext(struct as_data *ad) | |
406 | { | |
407 | unsigned long timeout; | |
408 | ||
409 | BUG_ON(ad->antic_status != ANTIC_OFF | |
410 | && ad->antic_status != ANTIC_WAIT_REQ); | |
411 | ||
412 | timeout = ad->antic_start + ad->antic_expire; | |
413 | ||
414 | mod_timer(&ad->antic_timer, timeout); | |
415 | ||
416 | ad->antic_status = ANTIC_WAIT_NEXT; | |
417 | } | |
418 | ||
419 | /* | |
420 | * as_antic_waitreq starts anticipating. We don't start timing the anticipation | |
421 | * until the request that we're anticipating on has finished. This means we | |
422 | * are timing from when the candidate process wakes up hopefully. | |
423 | */ | |
424 | static void as_antic_waitreq(struct as_data *ad) | |
425 | { | |
426 | BUG_ON(ad->antic_status == ANTIC_FINISHED); | |
427 | if (ad->antic_status == ANTIC_OFF) { | |
428 | if (!ad->io_context || ad->ioc_finished) | |
429 | as_antic_waitnext(ad); | |
430 | else | |
431 | ad->antic_status = ANTIC_WAIT_REQ; | |
432 | } | |
433 | } | |
434 | ||
435 | /* | |
436 | * This is called directly by the functions in this file to stop anticipation. | |
437 | * We kill the timer and schedule a call to the request_fn asap. | |
438 | */ | |
439 | static void as_antic_stop(struct as_data *ad) | |
440 | { | |
441 | int status = ad->antic_status; | |
442 | ||
443 | if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) { | |
444 | if (status == ANTIC_WAIT_NEXT) | |
445 | del_timer(&ad->antic_timer); | |
446 | ad->antic_status = ANTIC_FINISHED; | |
447 | /* see as_work_handler */ | |
448 | kblockd_schedule_work(&ad->antic_work); | |
449 | } | |
450 | } | |
451 | ||
452 | /* | |
453 | * as_antic_timeout is the timer function set by as_antic_waitnext. | |
454 | */ | |
455 | static void as_antic_timeout(unsigned long data) | |
456 | { | |
457 | struct request_queue *q = (struct request_queue *)data; | |
458 | struct as_data *ad = q->elevator->elevator_data; | |
459 | unsigned long flags; | |
460 | ||
461 | spin_lock_irqsave(q->queue_lock, flags); | |
462 | if (ad->antic_status == ANTIC_WAIT_REQ | |
463 | || ad->antic_status == ANTIC_WAIT_NEXT) { | |
464 | struct as_io_context *aic = ad->io_context->aic; | |
465 | ||
466 | ad->antic_status = ANTIC_FINISHED; | |
467 | kblockd_schedule_work(&ad->antic_work); | |
468 | ||
469 | if (aic->ttime_samples == 0) { | |
f5b3db00 | 470 | /* process anticipated on has exited or timed out*/ |
1da177e4 LT |
471 | ad->exit_prob = (7*ad->exit_prob + 256)/8; |
472 | } | |
f5b3db00 NP |
473 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
474 | /* process not "saved" by a cooperating request */ | |
475 | ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8; | |
476 | } | |
1da177e4 LT |
477 | } |
478 | spin_unlock_irqrestore(q->queue_lock, flags); | |
479 | } | |
480 | ||
f5b3db00 NP |
481 | static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic, |
482 | unsigned long ttime) | |
483 | { | |
484 | /* fixed point: 1.0 == 1<<8 */ | |
485 | if (aic->ttime_samples == 0) { | |
486 | ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8; | |
487 | ad->new_ttime_mean = ad->new_ttime_total / 256; | |
488 | ||
489 | ad->exit_prob = (7*ad->exit_prob)/8; | |
490 | } | |
491 | aic->ttime_samples = (7*aic->ttime_samples + 256) / 8; | |
492 | aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8; | |
493 | aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples; | |
494 | } | |
495 | ||
496 | static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic, | |
497 | sector_t sdist) | |
498 | { | |
499 | u64 total; | |
500 | ||
501 | if (aic->seek_samples == 0) { | |
502 | ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8; | |
503 | ad->new_seek_mean = ad->new_seek_total / 256; | |
504 | } | |
505 | ||
506 | /* | |
507 | * Don't allow the seek distance to get too large from the | |
508 | * odd fragment, pagein, etc | |
509 | */ | |
510 | if (aic->seek_samples <= 60) /* second&third seek */ | |
511 | sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024); | |
512 | else | |
513 | sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64); | |
514 | ||
515 | aic->seek_samples = (7*aic->seek_samples + 256) / 8; | |
516 | aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8; | |
517 | total = aic->seek_total + (aic->seek_samples/2); | |
518 | do_div(total, aic->seek_samples); | |
519 | aic->seek_mean = (sector_t)total; | |
520 | } | |
521 | ||
522 | /* | |
523 | * as_update_iohist keeps a decaying histogram of IO thinktimes, and | |
524 | * updates @aic->ttime_mean based on that. It is called when a new | |
525 | * request is queued. | |
526 | */ | |
527 | static void as_update_iohist(struct as_data *ad, struct as_io_context *aic, | |
528 | struct request *rq) | |
529 | { | |
9e2585a8 | 530 | int data_dir = rq_is_sync(rq); |
f5b3db00 NP |
531 | unsigned long thinktime = 0; |
532 | sector_t seek_dist; | |
533 | ||
534 | if (aic == NULL) | |
535 | return; | |
536 | ||
537 | if (data_dir == REQ_SYNC) { | |
538 | unsigned long in_flight = atomic_read(&aic->nr_queued) | |
539 | + atomic_read(&aic->nr_dispatched); | |
540 | spin_lock(&aic->lock); | |
541 | if (test_bit(AS_TASK_IORUNNING, &aic->state) || | |
542 | test_bit(AS_TASK_IOSTARTED, &aic->state)) { | |
543 | /* Calculate read -> read thinktime */ | |
544 | if (test_bit(AS_TASK_IORUNNING, &aic->state) | |
545 | && in_flight == 0) { | |
546 | thinktime = jiffies - aic->last_end_request; | |
547 | thinktime = min(thinktime, MAX_THINKTIME-1); | |
548 | } | |
549 | as_update_thinktime(ad, aic, thinktime); | |
550 | ||
551 | /* Calculate read -> read seek distance */ | |
552 | if (aic->last_request_pos < rq->sector) | |
553 | seek_dist = rq->sector - aic->last_request_pos; | |
554 | else | |
555 | seek_dist = aic->last_request_pos - rq->sector; | |
556 | as_update_seekdist(ad, aic, seek_dist); | |
557 | } | |
558 | aic->last_request_pos = rq->sector + rq->nr_sectors; | |
559 | set_bit(AS_TASK_IOSTARTED, &aic->state); | |
560 | spin_unlock(&aic->lock); | |
561 | } | |
562 | } | |
563 | ||
1da177e4 LT |
564 | /* |
565 | * as_close_req decides if one request is considered "close" to the | |
566 | * previous one issued. | |
567 | */ | |
f5b3db00 | 568 | static int as_close_req(struct as_data *ad, struct as_io_context *aic, |
8a8e674c | 569 | struct request *rq) |
1da177e4 LT |
570 | { |
571 | unsigned long delay; /* milliseconds */ | |
572 | sector_t last = ad->last_sector[ad->batch_data_dir]; | |
8a8e674c | 573 | sector_t next = rq->sector; |
1da177e4 | 574 | sector_t delta; /* acceptable close offset (in sectors) */ |
f5b3db00 | 575 | sector_t s; |
1da177e4 LT |
576 | |
577 | if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished) | |
578 | delay = 0; | |
579 | else | |
580 | delay = ((jiffies - ad->antic_start) * 1000) / HZ; | |
581 | ||
f5b3db00 NP |
582 | if (delay == 0) |
583 | delta = 8192; | |
1da177e4 | 584 | else if (delay <= 20 && delay <= ad->antic_expire) |
f5b3db00 | 585 | delta = 8192 << delay; |
1da177e4 LT |
586 | else |
587 | return 1; | |
588 | ||
f5b3db00 NP |
589 | if ((last <= next + (delta>>1)) && (next <= last + delta)) |
590 | return 1; | |
591 | ||
592 | if (last < next) | |
593 | s = next - last; | |
594 | else | |
595 | s = last - next; | |
596 | ||
597 | if (aic->seek_samples == 0) { | |
598 | /* | |
599 | * Process has just started IO. Use past statistics to | |
600 | * gauge success possibility | |
601 | */ | |
602 | if (ad->new_seek_mean > s) { | |
603 | /* this request is better than what we're expecting */ | |
604 | return 1; | |
605 | } | |
606 | ||
607 | } else { | |
608 | if (aic->seek_mean > s) { | |
609 | /* this request is better than what we're expecting */ | |
610 | return 1; | |
611 | } | |
612 | } | |
613 | ||
614 | return 0; | |
1da177e4 LT |
615 | } |
616 | ||
617 | /* | |
618 | * as_can_break_anticipation returns true if we have been anticipating this | |
619 | * request. | |
620 | * | |
621 | * It also returns true if the process against which we are anticipating | |
622 | * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to | |
623 | * dispatch it ASAP, because we know that application will not be submitting | |
624 | * any new reads. | |
625 | * | |
f5b3db00 | 626 | * If the task which has submitted the request has exited, break anticipation. |
1da177e4 LT |
627 | * |
628 | * If this task has queued some other IO, do not enter enticipation. | |
629 | */ | |
8a8e674c | 630 | static int as_can_break_anticipation(struct as_data *ad, struct request *rq) |
1da177e4 LT |
631 | { |
632 | struct io_context *ioc; | |
633 | struct as_io_context *aic; | |
1da177e4 LT |
634 | |
635 | ioc = ad->io_context; | |
636 | BUG_ON(!ioc); | |
637 | ||
8a8e674c | 638 | if (rq && ioc == RQ_IOC(rq)) { |
1da177e4 LT |
639 | /* request from same process */ |
640 | return 1; | |
641 | } | |
642 | ||
643 | if (ad->ioc_finished && as_antic_expired(ad)) { | |
644 | /* | |
645 | * In this situation status should really be FINISHED, | |
646 | * however the timer hasn't had the chance to run yet. | |
647 | */ | |
648 | return 1; | |
649 | } | |
650 | ||
651 | aic = ioc->aic; | |
652 | if (!aic) | |
653 | return 0; | |
654 | ||
1da177e4 LT |
655 | if (atomic_read(&aic->nr_queued) > 0) { |
656 | /* process has more requests queued */ | |
657 | return 1; | |
658 | } | |
659 | ||
660 | if (atomic_read(&aic->nr_dispatched) > 0) { | |
661 | /* process has more requests dispatched */ | |
662 | return 1; | |
663 | } | |
664 | ||
8a8e674c | 665 | if (rq && rq_is_sync(rq) && as_close_req(ad, aic, rq)) { |
1da177e4 LT |
666 | /* |
667 | * Found a close request that is not one of ours. | |
668 | * | |
f5b3db00 NP |
669 | * This makes close requests from another process update |
670 | * our IO history. Is generally useful when there are | |
1da177e4 LT |
671 | * two or more cooperating processes working in the same |
672 | * area. | |
673 | */ | |
f5b3db00 NP |
674 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
675 | if (aic->ttime_samples == 0) | |
676 | ad->exit_prob = (7*ad->exit_prob + 256)/8; | |
677 | ||
678 | ad->exit_no_coop = (7*ad->exit_no_coop)/8; | |
679 | } | |
680 | ||
8a8e674c | 681 | as_update_iohist(ad, aic, rq); |
1da177e4 LT |
682 | return 1; |
683 | } | |
684 | ||
f5b3db00 NP |
685 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
686 | /* process anticipated on has exited */ | |
687 | if (aic->ttime_samples == 0) | |
688 | ad->exit_prob = (7*ad->exit_prob + 256)/8; | |
689 | ||
690 | if (ad->exit_no_coop > 128) | |
691 | return 1; | |
692 | } | |
1da177e4 LT |
693 | |
694 | if (aic->ttime_samples == 0) { | |
695 | if (ad->new_ttime_mean > ad->antic_expire) | |
696 | return 1; | |
f5b3db00 | 697 | if (ad->exit_prob * ad->exit_no_coop > 128*256) |
1da177e4 LT |
698 | return 1; |
699 | } else if (aic->ttime_mean > ad->antic_expire) { | |
700 | /* the process thinks too much between requests */ | |
701 | return 1; | |
702 | } | |
703 | ||
1da177e4 LT |
704 | return 0; |
705 | } | |
706 | ||
707 | /* | |
8a8e674c | 708 | * as_can_anticipate indicates whether we should either run rq |
1da177e4 LT |
709 | * or keep anticipating a better request. |
710 | */ | |
8a8e674c | 711 | static int as_can_anticipate(struct as_data *ad, struct request *rq) |
1da177e4 LT |
712 | { |
713 | if (!ad->io_context) | |
714 | /* | |
715 | * Last request submitted was a write | |
716 | */ | |
717 | return 0; | |
718 | ||
719 | if (ad->antic_status == ANTIC_FINISHED) | |
720 | /* | |
721 | * Don't restart if we have just finished. Run the next request | |
722 | */ | |
723 | return 0; | |
724 | ||
8a8e674c | 725 | if (as_can_break_anticipation(ad, rq)) |
1da177e4 LT |
726 | /* |
727 | * This request is a good candidate. Don't keep anticipating, | |
728 | * run it. | |
729 | */ | |
730 | return 0; | |
731 | ||
732 | /* | |
733 | * OK from here, we haven't finished, and don't have a decent request! | |
734 | * Status is either ANTIC_OFF so start waiting, | |
735 | * ANTIC_WAIT_REQ so continue waiting for request to finish | |
736 | * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request. | |
1da177e4 LT |
737 | */ |
738 | ||
739 | return 1; | |
740 | } | |
741 | ||
1da177e4 | 742 | /* |
8a8e674c | 743 | * as_update_rq must be called whenever a request (rq) is added to |
1da177e4 LT |
744 | * the sort_list. This function keeps caches up to date, and checks if the |
745 | * request might be one we are "anticipating" | |
746 | */ | |
8a8e674c | 747 | static void as_update_rq(struct as_data *ad, struct request *rq) |
1da177e4 | 748 | { |
8a8e674c | 749 | const int data_dir = rq_is_sync(rq); |
1da177e4 | 750 | |
8a8e674c JA |
751 | /* keep the next_rq cache up to date */ |
752 | ad->next_rq[data_dir] = as_choose_req(ad, rq, ad->next_rq[data_dir]); | |
1da177e4 LT |
753 | |
754 | /* | |
755 | * have we been anticipating this request? | |
756 | * or does it come from the same process as the one we are anticipating | |
757 | * for? | |
758 | */ | |
759 | if (ad->antic_status == ANTIC_WAIT_REQ | |
760 | || ad->antic_status == ANTIC_WAIT_NEXT) { | |
8a8e674c | 761 | if (as_can_break_anticipation(ad, rq)) |
1da177e4 LT |
762 | as_antic_stop(ad); |
763 | } | |
764 | } | |
765 | ||
766 | /* | |
767 | * Gathers timings and resizes the write batch automatically | |
768 | */ | |
769 | static void update_write_batch(struct as_data *ad) | |
770 | { | |
771 | unsigned long batch = ad->batch_expire[REQ_ASYNC]; | |
772 | long write_time; | |
773 | ||
774 | write_time = (jiffies - ad->current_batch_expires) + batch; | |
775 | if (write_time < 0) | |
776 | write_time = 0; | |
777 | ||
778 | if (write_time > batch && !ad->write_batch_idled) { | |
779 | if (write_time > batch * 3) | |
780 | ad->write_batch_count /= 2; | |
781 | else | |
782 | ad->write_batch_count--; | |
783 | } else if (write_time < batch && ad->current_write_count == 0) { | |
784 | if (batch > write_time * 3) | |
785 | ad->write_batch_count *= 2; | |
786 | else | |
787 | ad->write_batch_count++; | |
788 | } | |
789 | ||
790 | if (ad->write_batch_count < 1) | |
791 | ad->write_batch_count = 1; | |
792 | } | |
793 | ||
794 | /* | |
795 | * as_completed_request is to be called when a request has completed and | |
796 | * returned something to the requesting process, be it an error or data. | |
797 | */ | |
798 | static void as_completed_request(request_queue_t *q, struct request *rq) | |
799 | { | |
800 | struct as_data *ad = q->elevator->elevator_data; | |
1da177e4 LT |
801 | |
802 | WARN_ON(!list_empty(&rq->queuelist)); | |
803 | ||
8a8e674c JA |
804 | if (RQ_STATE(rq) != AS_RQ_REMOVED) { |
805 | printk("rq->state %d\n", RQ_STATE(rq)); | |
1da177e4 LT |
806 | WARN_ON(1); |
807 | goto out; | |
808 | } | |
809 | ||
1da177e4 LT |
810 | if (ad->changed_batch && ad->nr_dispatched == 1) { |
811 | kblockd_schedule_work(&ad->antic_work); | |
812 | ad->changed_batch = 0; | |
813 | ||
814 | if (ad->batch_data_dir == REQ_SYNC) | |
815 | ad->new_batch = 1; | |
816 | } | |
817 | WARN_ON(ad->nr_dispatched == 0); | |
818 | ad->nr_dispatched--; | |
819 | ||
820 | /* | |
821 | * Start counting the batch from when a request of that direction is | |
822 | * actually serviced. This should help devices with big TCQ windows | |
823 | * and writeback caches | |
824 | */ | |
9e2585a8 | 825 | if (ad->new_batch && ad->batch_data_dir == rq_is_sync(rq)) { |
1da177e4 LT |
826 | update_write_batch(ad); |
827 | ad->current_batch_expires = jiffies + | |
828 | ad->batch_expire[REQ_SYNC]; | |
829 | ad->new_batch = 0; | |
830 | } | |
831 | ||
8a8e674c | 832 | if (ad->io_context == RQ_IOC(rq) && ad->io_context) { |
1da177e4 LT |
833 | ad->antic_start = jiffies; |
834 | ad->ioc_finished = 1; | |
835 | if (ad->antic_status == ANTIC_WAIT_REQ) { | |
836 | /* | |
837 | * We were waiting on this request, now anticipate | |
838 | * the next one | |
839 | */ | |
840 | as_antic_waitnext(ad); | |
841 | } | |
842 | } | |
843 | ||
8a8e674c | 844 | as_put_io_context(rq); |
1da177e4 | 845 | out: |
8a8e674c | 846 | RQ_SET_STATE(rq, AS_RQ_POSTSCHED); |
1da177e4 LT |
847 | } |
848 | ||
849 | /* | |
850 | * as_remove_queued_request removes a request from the pre dispatch queue | |
851 | * without updating refcounts. It is expected the caller will drop the | |
852 | * reference unless it replaces the request at somepart of the elevator | |
853 | * (ie. the dispatch queue) | |
854 | */ | |
855 | static void as_remove_queued_request(request_queue_t *q, struct request *rq) | |
856 | { | |
9e2585a8 | 857 | const int data_dir = rq_is_sync(rq); |
1da177e4 | 858 | struct as_data *ad = q->elevator->elevator_data; |
8a8e674c | 859 | struct io_context *ioc; |
1da177e4 | 860 | |
8a8e674c | 861 | WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED); |
1da177e4 | 862 | |
8a8e674c JA |
863 | ioc = RQ_IOC(rq); |
864 | if (ioc && ioc->aic) { | |
865 | BUG_ON(!atomic_read(&ioc->aic->nr_queued)); | |
866 | atomic_dec(&ioc->aic->nr_queued); | |
1da177e4 LT |
867 | } |
868 | ||
869 | /* | |
8a8e674c | 870 | * Update the "next_rq" cache if we are about to remove its |
1da177e4 LT |
871 | * entry |
872 | */ | |
8a8e674c JA |
873 | if (ad->next_rq[data_dir] == rq) |
874 | ad->next_rq[data_dir] = as_find_next_rq(ad, rq); | |
1da177e4 | 875 | |
d4f2f462 | 876 | rq_fifo_clear(rq); |
8a8e674c | 877 | as_del_rq_rb(ad, rq); |
1da177e4 LT |
878 | } |
879 | ||
1da177e4 LT |
880 | /* |
881 | * as_fifo_expired returns 0 if there are no expired reads on the fifo, | |
882 | * 1 otherwise. It is ratelimited so that we only perform the check once per | |
883 | * `fifo_expire' interval. Otherwise a large number of expired requests | |
884 | * would create a hopeless seekstorm. | |
885 | * | |
886 | * See as_antic_expired comment. | |
887 | */ | |
888 | static int as_fifo_expired(struct as_data *ad, int adir) | |
889 | { | |
d4f2f462 | 890 | struct request *rq; |
1da177e4 LT |
891 | long delta_jif; |
892 | ||
893 | delta_jif = jiffies - ad->last_check_fifo[adir]; | |
894 | if (unlikely(delta_jif < 0)) | |
895 | delta_jif = -delta_jif; | |
896 | if (delta_jif < ad->fifo_expire[adir]) | |
897 | return 0; | |
898 | ||
899 | ad->last_check_fifo[adir] = jiffies; | |
900 | ||
901 | if (list_empty(&ad->fifo_list[adir])) | |
902 | return 0; | |
903 | ||
d4f2f462 | 904 | rq = rq_entry_fifo(ad->fifo_list[adir].next); |
1da177e4 | 905 | |
d4f2f462 | 906 | return time_after(jiffies, rq_fifo_time(rq)); |
1da177e4 LT |
907 | } |
908 | ||
909 | /* | |
910 | * as_batch_expired returns true if the current batch has expired. A batch | |
911 | * is a set of reads or a set of writes. | |
912 | */ | |
913 | static inline int as_batch_expired(struct as_data *ad) | |
914 | { | |
915 | if (ad->changed_batch || ad->new_batch) | |
916 | return 0; | |
917 | ||
918 | if (ad->batch_data_dir == REQ_SYNC) | |
919 | /* TODO! add a check so a complete fifo gets written? */ | |
920 | return time_after(jiffies, ad->current_batch_expires); | |
921 | ||
922 | return time_after(jiffies, ad->current_batch_expires) | |
923 | || ad->current_write_count == 0; | |
924 | } | |
925 | ||
926 | /* | |
927 | * move an entry to dispatch queue | |
928 | */ | |
8a8e674c | 929 | static void as_move_to_dispatch(struct as_data *ad, struct request *rq) |
1da177e4 | 930 | { |
9e2585a8 | 931 | const int data_dir = rq_is_sync(rq); |
1da177e4 | 932 | |
e37f346e | 933 | BUG_ON(RB_EMPTY_NODE(&rq->rb_node)); |
1da177e4 LT |
934 | |
935 | as_antic_stop(ad); | |
936 | ad->antic_status = ANTIC_OFF; | |
937 | ||
938 | /* | |
939 | * This has to be set in order to be correctly updated by | |
8a8e674c | 940 | * as_find_next_rq |
1da177e4 LT |
941 | */ |
942 | ad->last_sector[data_dir] = rq->sector + rq->nr_sectors; | |
943 | ||
944 | if (data_dir == REQ_SYNC) { | |
8a8e674c | 945 | struct io_context *ioc = RQ_IOC(rq); |
1da177e4 | 946 | /* In case we have to anticipate after this */ |
8a8e674c | 947 | copy_io_context(&ad->io_context, &ioc); |
1da177e4 LT |
948 | } else { |
949 | if (ad->io_context) { | |
950 | put_io_context(ad->io_context); | |
951 | ad->io_context = NULL; | |
952 | } | |
953 | ||
954 | if (ad->current_write_count != 0) | |
955 | ad->current_write_count--; | |
956 | } | |
957 | ad->ioc_finished = 0; | |
958 | ||
8a8e674c | 959 | ad->next_rq[data_dir] = as_find_next_rq(ad, rq); |
1da177e4 LT |
960 | |
961 | /* | |
962 | * take it off the sort and fifo list, add to dispatch queue | |
963 | */ | |
1da177e4 | 964 | as_remove_queued_request(ad->q, rq); |
8a8e674c | 965 | WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED); |
1da177e4 | 966 | |
b4878f24 JA |
967 | elv_dispatch_sort(ad->q, rq); |
968 | ||
8a8e674c JA |
969 | RQ_SET_STATE(rq, AS_RQ_DISPATCHED); |
970 | if (RQ_IOC(rq) && RQ_IOC(rq)->aic) | |
971 | atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched); | |
1da177e4 LT |
972 | ad->nr_dispatched++; |
973 | } | |
974 | ||
975 | /* | |
976 | * as_dispatch_request selects the best request according to | |
977 | * read/write expire, batch expire, etc, and moves it to the dispatch | |
978 | * queue. Returns 1 if a request was found, 0 otherwise. | |
979 | */ | |
b4878f24 | 980 | static int as_dispatch_request(request_queue_t *q, int force) |
1da177e4 | 981 | { |
b4878f24 | 982 | struct as_data *ad = q->elevator->elevator_data; |
1da177e4 LT |
983 | const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]); |
984 | const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]); | |
8a8e674c | 985 | struct request *rq; |
1da177e4 | 986 | |
b4878f24 JA |
987 | if (unlikely(force)) { |
988 | /* | |
989 | * Forced dispatch, accounting is useless. Reset | |
990 | * accounting states and dump fifo_lists. Note that | |
991 | * batch_data_dir is reset to REQ_SYNC to avoid | |
992 | * screwing write batch accounting as write batch | |
993 | * accounting occurs on W->R transition. | |
994 | */ | |
995 | int dispatched = 0; | |
996 | ||
997 | ad->batch_data_dir = REQ_SYNC; | |
998 | ad->changed_batch = 0; | |
999 | ad->new_batch = 0; | |
1000 | ||
8a8e674c JA |
1001 | while (ad->next_rq[REQ_SYNC]) { |
1002 | as_move_to_dispatch(ad, ad->next_rq[REQ_SYNC]); | |
b4878f24 JA |
1003 | dispatched++; |
1004 | } | |
1005 | ad->last_check_fifo[REQ_SYNC] = jiffies; | |
1006 | ||
8a8e674c JA |
1007 | while (ad->next_rq[REQ_ASYNC]) { |
1008 | as_move_to_dispatch(ad, ad->next_rq[REQ_ASYNC]); | |
b4878f24 JA |
1009 | dispatched++; |
1010 | } | |
1011 | ad->last_check_fifo[REQ_ASYNC] = jiffies; | |
1012 | ||
1013 | return dispatched; | |
1014 | } | |
1015 | ||
1da177e4 LT |
1016 | /* Signal that the write batch was uncontended, so we can't time it */ |
1017 | if (ad->batch_data_dir == REQ_ASYNC && !reads) { | |
1018 | if (ad->current_write_count == 0 || !writes) | |
1019 | ad->write_batch_idled = 1; | |
1020 | } | |
1021 | ||
1022 | if (!(reads || writes) | |
1023 | || ad->antic_status == ANTIC_WAIT_REQ | |
1024 | || ad->antic_status == ANTIC_WAIT_NEXT | |
1025 | || ad->changed_batch) | |
1026 | return 0; | |
1027 | ||
f5b3db00 | 1028 | if (!(reads && writes && as_batch_expired(ad))) { |
1da177e4 LT |
1029 | /* |
1030 | * batch is still running or no reads or no writes | |
1031 | */ | |
8a8e674c | 1032 | rq = ad->next_rq[ad->batch_data_dir]; |
1da177e4 LT |
1033 | |
1034 | if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) { | |
1035 | if (as_fifo_expired(ad, REQ_SYNC)) | |
1036 | goto fifo_expired; | |
1037 | ||
8a8e674c | 1038 | if (as_can_anticipate(ad, rq)) { |
1da177e4 LT |
1039 | as_antic_waitreq(ad); |
1040 | return 0; | |
1041 | } | |
1042 | } | |
1043 | ||
8a8e674c | 1044 | if (rq) { |
1da177e4 LT |
1045 | /* we have a "next request" */ |
1046 | if (reads && !writes) | |
1047 | ad->current_batch_expires = | |
1048 | jiffies + ad->batch_expire[REQ_SYNC]; | |
1049 | goto dispatch_request; | |
1050 | } | |
1051 | } | |
1052 | ||
1053 | /* | |
1054 | * at this point we are not running a batch. select the appropriate | |
1055 | * data direction (read / write) | |
1056 | */ | |
1057 | ||
1058 | if (reads) { | |
dd67d051 | 1059 | BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_SYNC])); |
1da177e4 LT |
1060 | |
1061 | if (writes && ad->batch_data_dir == REQ_SYNC) | |
1062 | /* | |
1063 | * Last batch was a read, switch to writes | |
1064 | */ | |
1065 | goto dispatch_writes; | |
1066 | ||
1067 | if (ad->batch_data_dir == REQ_ASYNC) { | |
1068 | WARN_ON(ad->new_batch); | |
1069 | ad->changed_batch = 1; | |
1070 | } | |
1071 | ad->batch_data_dir = REQ_SYNC; | |
8a8e674c | 1072 | rq = rq_entry_fifo(ad->fifo_list[REQ_SYNC].next); |
1da177e4 LT |
1073 | ad->last_check_fifo[ad->batch_data_dir] = jiffies; |
1074 | goto dispatch_request; | |
1075 | } | |
1076 | ||
1077 | /* | |
1078 | * the last batch was a read | |
1079 | */ | |
1080 | ||
1081 | if (writes) { | |
1082 | dispatch_writes: | |
dd67d051 | 1083 | BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_ASYNC])); |
1da177e4 LT |
1084 | |
1085 | if (ad->batch_data_dir == REQ_SYNC) { | |
1086 | ad->changed_batch = 1; | |
1087 | ||
1088 | /* | |
1089 | * new_batch might be 1 when the queue runs out of | |
1090 | * reads. A subsequent submission of a write might | |
1091 | * cause a change of batch before the read is finished. | |
1092 | */ | |
1093 | ad->new_batch = 0; | |
1094 | } | |
1095 | ad->batch_data_dir = REQ_ASYNC; | |
1096 | ad->current_write_count = ad->write_batch_count; | |
1097 | ad->write_batch_idled = 0; | |
8a8e674c | 1098 | rq = ad->next_rq[ad->batch_data_dir]; |
1da177e4 LT |
1099 | goto dispatch_request; |
1100 | } | |
1101 | ||
1102 | BUG(); | |
1103 | return 0; | |
1104 | ||
1105 | dispatch_request: | |
1106 | /* | |
1107 | * If a request has expired, service it. | |
1108 | */ | |
1109 | ||
1110 | if (as_fifo_expired(ad, ad->batch_data_dir)) { | |
1111 | fifo_expired: | |
8a8e674c | 1112 | rq = rq_entry_fifo(ad->fifo_list[ad->batch_data_dir].next); |
1da177e4 LT |
1113 | } |
1114 | ||
1115 | if (ad->changed_batch) { | |
1116 | WARN_ON(ad->new_batch); | |
1117 | ||
1118 | if (ad->nr_dispatched) | |
1119 | return 0; | |
1120 | ||
1121 | if (ad->batch_data_dir == REQ_ASYNC) | |
1122 | ad->current_batch_expires = jiffies + | |
1123 | ad->batch_expire[REQ_ASYNC]; | |
1124 | else | |
1125 | ad->new_batch = 1; | |
1126 | ||
1127 | ad->changed_batch = 0; | |
1128 | } | |
1129 | ||
1130 | /* | |
8a8e674c | 1131 | * rq is the selected appropriate request. |
1da177e4 | 1132 | */ |
8a8e674c | 1133 | as_move_to_dispatch(ad, rq); |
1da177e4 LT |
1134 | |
1135 | return 1; | |
1136 | } | |
1137 | ||
1da177e4 | 1138 | /* |
8a8e674c | 1139 | * add rq to rbtree and fifo |
1da177e4 | 1140 | */ |
b4878f24 | 1141 | static void as_add_request(request_queue_t *q, struct request *rq) |
1da177e4 | 1142 | { |
b4878f24 | 1143 | struct as_data *ad = q->elevator->elevator_data; |
1da177e4 LT |
1144 | int data_dir; |
1145 | ||
8a8e674c | 1146 | RQ_SET_STATE(rq, AS_RQ_NEW); |
b4878f24 | 1147 | |
9e2585a8 | 1148 | data_dir = rq_is_sync(rq); |
1da177e4 | 1149 | |
8a8e674c | 1150 | rq->elevator_private = as_get_io_context(); |
1da177e4 | 1151 | |
8a8e674c JA |
1152 | if (RQ_IOC(rq)) { |
1153 | as_update_iohist(ad, RQ_IOC(rq)->aic, rq); | |
1154 | atomic_inc(&RQ_IOC(rq)->aic->nr_queued); | |
1da177e4 LT |
1155 | } |
1156 | ||
8a8e674c | 1157 | as_add_rq_rb(ad, rq); |
1da177e4 | 1158 | |
ef9be1d3 TH |
1159 | /* |
1160 | * set expire time (only used for reads) and add to fifo list | |
1161 | */ | |
d4f2f462 JA |
1162 | rq_set_fifo_time(rq, jiffies + ad->fifo_expire[data_dir]); |
1163 | list_add_tail(&rq->queuelist, &ad->fifo_list[data_dir]); | |
1da177e4 | 1164 | |
8a8e674c JA |
1165 | as_update_rq(ad, rq); /* keep state machine up to date */ |
1166 | RQ_SET_STATE(rq, AS_RQ_QUEUED); | |
1da177e4 LT |
1167 | } |
1168 | ||
b4878f24 | 1169 | static void as_activate_request(request_queue_t *q, struct request *rq) |
1da177e4 | 1170 | { |
8a8e674c JA |
1171 | WARN_ON(RQ_STATE(rq) != AS_RQ_DISPATCHED); |
1172 | RQ_SET_STATE(rq, AS_RQ_REMOVED); | |
1173 | if (RQ_IOC(rq) && RQ_IOC(rq)->aic) | |
1174 | atomic_dec(&RQ_IOC(rq)->aic->nr_dispatched); | |
1da177e4 LT |
1175 | } |
1176 | ||
b4878f24 | 1177 | static void as_deactivate_request(request_queue_t *q, struct request *rq) |
1da177e4 | 1178 | { |
8a8e674c JA |
1179 | WARN_ON(RQ_STATE(rq) != AS_RQ_REMOVED); |
1180 | RQ_SET_STATE(rq, AS_RQ_DISPATCHED); | |
1181 | if (RQ_IOC(rq) && RQ_IOC(rq)->aic) | |
1182 | atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched); | |
1da177e4 LT |
1183 | } |
1184 | ||
1185 | /* | |
1186 | * as_queue_empty tells us if there are requests left in the device. It may | |
1187 | * not be the case that a driver can get the next request even if the queue | |
1188 | * is not empty - it is used in the block layer to check for plugging and | |
1189 | * merging opportunities | |
1190 | */ | |
1191 | static int as_queue_empty(request_queue_t *q) | |
1192 | { | |
1193 | struct as_data *ad = q->elevator->elevator_data; | |
1194 | ||
b4878f24 JA |
1195 | return list_empty(&ad->fifo_list[REQ_ASYNC]) |
1196 | && list_empty(&ad->fifo_list[REQ_SYNC]); | |
1da177e4 LT |
1197 | } |
1198 | ||
1da177e4 LT |
1199 | static int |
1200 | as_merge(request_queue_t *q, struct request **req, struct bio *bio) | |
1201 | { | |
1202 | struct as_data *ad = q->elevator->elevator_data; | |
1203 | sector_t rb_key = bio->bi_sector + bio_sectors(bio); | |
1204 | struct request *__rq; | |
1da177e4 LT |
1205 | |
1206 | /* | |
1207 | * check for front merge | |
1208 | */ | |
e37f346e | 1209 | __rq = elv_rb_find(&ad->sort_list[bio_data_dir(bio)], rb_key); |
9817064b JA |
1210 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
1211 | *req = __rq; | |
1212 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
1213 | } |
1214 | ||
1215 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
1216 | } |
1217 | ||
e37f346e | 1218 | static void as_merged_request(request_queue_t *q, struct request *req, int type) |
1da177e4 LT |
1219 | { |
1220 | struct as_data *ad = q->elevator->elevator_data; | |
1da177e4 | 1221 | |
1da177e4 LT |
1222 | /* |
1223 | * if the merge was a front merge, we need to reposition request | |
1224 | */ | |
e37f346e | 1225 | if (type == ELEVATOR_FRONT_MERGE) { |
8a8e674c JA |
1226 | as_del_rq_rb(ad, req); |
1227 | as_add_rq_rb(ad, req); | |
1da177e4 LT |
1228 | /* |
1229 | * Note! At this stage of this and the next function, our next | |
1230 | * request may not be optimal - eg the request may have "grown" | |
1231 | * behind the disk head. We currently don't bother adjusting. | |
1232 | */ | |
1233 | } | |
1da177e4 LT |
1234 | } |
1235 | ||
f5b3db00 NP |
1236 | static void as_merged_requests(request_queue_t *q, struct request *req, |
1237 | struct request *next) | |
1da177e4 | 1238 | { |
1da177e4 | 1239 | /* |
8a8e674c JA |
1240 | * if next expires before rq, assign its expire time to arq |
1241 | * and move into next position (next will be deleted) in fifo | |
1da177e4 | 1242 | */ |
d4f2f462 JA |
1243 | if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) { |
1244 | if (time_before(rq_fifo_time(next), rq_fifo_time(req))) { | |
8a8e674c JA |
1245 | struct io_context *rioc = RQ_IOC(req); |
1246 | struct io_context *nioc = RQ_IOC(next); | |
1247 | ||
d4f2f462 JA |
1248 | list_move(&req->queuelist, &next->queuelist); |
1249 | rq_set_fifo_time(req, rq_fifo_time(next)); | |
1da177e4 LT |
1250 | /* |
1251 | * Don't copy here but swap, because when anext is | |
1252 | * removed below, it must contain the unused context | |
1253 | */ | |
8a8e674c | 1254 | swap_io_context(&rioc, &nioc); |
1da177e4 LT |
1255 | } |
1256 | } | |
1257 | ||
1da177e4 LT |
1258 | /* |
1259 | * kill knowledge of next, this one is a goner | |
1260 | */ | |
1261 | as_remove_queued_request(q, next); | |
8a8e674c | 1262 | as_put_io_context(next); |
1da177e4 | 1263 | |
8a8e674c | 1264 | RQ_SET_STATE(next, AS_RQ_MERGED); |
1da177e4 LT |
1265 | } |
1266 | ||
1267 | /* | |
1268 | * This is executed in a "deferred" process context, by kblockd. It calls the | |
1269 | * driver's request_fn so the driver can submit that request. | |
1270 | * | |
1271 | * IMPORTANT! This guy will reenter the elevator, so set up all queue global | |
1272 | * state before calling, and don't rely on any state over calls. | |
1273 | * | |
1274 | * FIXME! dispatch queue is not a queue at all! | |
1275 | */ | |
1276 | static void as_work_handler(void *data) | |
1277 | { | |
1278 | struct request_queue *q = data; | |
1279 | unsigned long flags; | |
1280 | ||
1281 | spin_lock_irqsave(q->queue_lock, flags); | |
b4878f24 | 1282 | if (!as_queue_empty(q)) |
1da177e4 LT |
1283 | q->request_fn(q); |
1284 | spin_unlock_irqrestore(q->queue_lock, flags); | |
1285 | } | |
1286 | ||
cb78b285 | 1287 | static int as_may_queue(request_queue_t *q, int rw) |
1da177e4 LT |
1288 | { |
1289 | int ret = ELV_MQUEUE_MAY; | |
1290 | struct as_data *ad = q->elevator->elevator_data; | |
1291 | struct io_context *ioc; | |
1292 | if (ad->antic_status == ANTIC_WAIT_REQ || | |
1293 | ad->antic_status == ANTIC_WAIT_NEXT) { | |
1294 | ioc = as_get_io_context(); | |
1295 | if (ad->io_context == ioc) | |
1296 | ret = ELV_MQUEUE_MUST; | |
1297 | put_io_context(ioc); | |
1298 | } | |
1299 | ||
1300 | return ret; | |
1301 | } | |
1302 | ||
1303 | static void as_exit_queue(elevator_t *e) | |
1304 | { | |
1305 | struct as_data *ad = e->elevator_data; | |
1306 | ||
1307 | del_timer_sync(&ad->antic_timer); | |
1308 | kblockd_flush(); | |
1309 | ||
1310 | BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC])); | |
1311 | BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC])); | |
1312 | ||
1da177e4 | 1313 | put_io_context(ad->io_context); |
1da177e4 LT |
1314 | kfree(ad); |
1315 | } | |
1316 | ||
1317 | /* | |
8a8e674c | 1318 | * initialize elevator private data (as_data). |
1da177e4 | 1319 | */ |
bc1c1169 | 1320 | static void *as_init_queue(request_queue_t *q, elevator_t *e) |
1da177e4 LT |
1321 | { |
1322 | struct as_data *ad; | |
1da177e4 | 1323 | |
1946089a | 1324 | ad = kmalloc_node(sizeof(*ad), GFP_KERNEL, q->node); |
1da177e4 | 1325 | if (!ad) |
bc1c1169 | 1326 | return NULL; |
1da177e4 LT |
1327 | memset(ad, 0, sizeof(*ad)); |
1328 | ||
1329 | ad->q = q; /* Identify what queue the data belongs to */ | |
1330 | ||
1da177e4 LT |
1331 | /* anticipatory scheduling helpers */ |
1332 | ad->antic_timer.function = as_antic_timeout; | |
1333 | ad->antic_timer.data = (unsigned long)q; | |
1334 | init_timer(&ad->antic_timer); | |
1335 | INIT_WORK(&ad->antic_work, as_work_handler, q); | |
1336 | ||
1da177e4 LT |
1337 | INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]); |
1338 | INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]); | |
1339 | ad->sort_list[REQ_SYNC] = RB_ROOT; | |
1340 | ad->sort_list[REQ_ASYNC] = RB_ROOT; | |
1da177e4 LT |
1341 | ad->fifo_expire[REQ_SYNC] = default_read_expire; |
1342 | ad->fifo_expire[REQ_ASYNC] = default_write_expire; | |
1343 | ad->antic_expire = default_antic_expire; | |
1344 | ad->batch_expire[REQ_SYNC] = default_read_batch_expire; | |
1345 | ad->batch_expire[REQ_ASYNC] = default_write_batch_expire; | |
1da177e4 LT |
1346 | |
1347 | ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC]; | |
1348 | ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10; | |
1349 | if (ad->write_batch_count < 2) | |
1350 | ad->write_batch_count = 2; | |
1351 | ||
bc1c1169 | 1352 | return ad; |
1da177e4 LT |
1353 | } |
1354 | ||
1355 | /* | |
1356 | * sysfs parts below | |
1357 | */ | |
1da177e4 LT |
1358 | |
1359 | static ssize_t | |
1360 | as_var_show(unsigned int var, char *page) | |
1361 | { | |
1da177e4 LT |
1362 | return sprintf(page, "%d\n", var); |
1363 | } | |
1364 | ||
1365 | static ssize_t | |
1366 | as_var_store(unsigned long *var, const char *page, size_t count) | |
1367 | { | |
1da177e4 LT |
1368 | char *p = (char *) page; |
1369 | ||
c9b3ad67 | 1370 | *var = simple_strtoul(p, &p, 10); |
1da177e4 LT |
1371 | return count; |
1372 | } | |
1373 | ||
e572ec7e | 1374 | static ssize_t est_time_show(elevator_t *e, char *page) |
1da177e4 | 1375 | { |
3d1ab40f | 1376 | struct as_data *ad = e->elevator_data; |
1da177e4 LT |
1377 | int pos = 0; |
1378 | ||
f5b3db00 NP |
1379 | pos += sprintf(page+pos, "%lu %% exit probability\n", |
1380 | 100*ad->exit_prob/256); | |
1381 | pos += sprintf(page+pos, "%lu %% probability of exiting without a " | |
1382 | "cooperating process submitting IO\n", | |
1383 | 100*ad->exit_no_coop/256); | |
1da177e4 | 1384 | pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean); |
f5b3db00 NP |
1385 | pos += sprintf(page+pos, "%llu sectors new seek distance\n", |
1386 | (unsigned long long)ad->new_seek_mean); | |
1da177e4 LT |
1387 | |
1388 | return pos; | |
1389 | } | |
1390 | ||
1391 | #define SHOW_FUNCTION(__FUNC, __VAR) \ | |
3d1ab40f | 1392 | static ssize_t __FUNC(elevator_t *e, char *page) \ |
1da177e4 | 1393 | { \ |
3d1ab40f | 1394 | struct as_data *ad = e->elevator_data; \ |
1da177e4 LT |
1395 | return as_var_show(jiffies_to_msecs((__VAR)), (page)); \ |
1396 | } | |
e572ec7e AV |
1397 | SHOW_FUNCTION(as_read_expire_show, ad->fifo_expire[REQ_SYNC]); |
1398 | SHOW_FUNCTION(as_write_expire_show, ad->fifo_expire[REQ_ASYNC]); | |
1399 | SHOW_FUNCTION(as_antic_expire_show, ad->antic_expire); | |
1400 | SHOW_FUNCTION(as_read_batch_expire_show, ad->batch_expire[REQ_SYNC]); | |
1401 | SHOW_FUNCTION(as_write_batch_expire_show, ad->batch_expire[REQ_ASYNC]); | |
1da177e4 LT |
1402 | #undef SHOW_FUNCTION |
1403 | ||
1404 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \ | |
3d1ab40f | 1405 | static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \ |
1da177e4 | 1406 | { \ |
3d1ab40f AV |
1407 | struct as_data *ad = e->elevator_data; \ |
1408 | int ret = as_var_store(__PTR, (page), count); \ | |
1da177e4 LT |
1409 | if (*(__PTR) < (MIN)) \ |
1410 | *(__PTR) = (MIN); \ | |
1411 | else if (*(__PTR) > (MAX)) \ | |
1412 | *(__PTR) = (MAX); \ | |
1413 | *(__PTR) = msecs_to_jiffies(*(__PTR)); \ | |
1414 | return ret; \ | |
1415 | } | |
e572ec7e AV |
1416 | STORE_FUNCTION(as_read_expire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX); |
1417 | STORE_FUNCTION(as_write_expire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX); | |
1418 | STORE_FUNCTION(as_antic_expire_store, &ad->antic_expire, 0, INT_MAX); | |
1419 | STORE_FUNCTION(as_read_batch_expire_store, | |
1da177e4 | 1420 | &ad->batch_expire[REQ_SYNC], 0, INT_MAX); |
e572ec7e | 1421 | STORE_FUNCTION(as_write_batch_expire_store, |
1da177e4 LT |
1422 | &ad->batch_expire[REQ_ASYNC], 0, INT_MAX); |
1423 | #undef STORE_FUNCTION | |
1424 | ||
e572ec7e AV |
1425 | #define AS_ATTR(name) \ |
1426 | __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store) | |
1427 | ||
1428 | static struct elv_fs_entry as_attrs[] = { | |
1429 | __ATTR_RO(est_time), | |
1430 | AS_ATTR(read_expire), | |
1431 | AS_ATTR(write_expire), | |
1432 | AS_ATTR(antic_expire), | |
1433 | AS_ATTR(read_batch_expire), | |
1434 | AS_ATTR(write_batch_expire), | |
1435 | __ATTR_NULL | |
1da177e4 LT |
1436 | }; |
1437 | ||
1da177e4 LT |
1438 | static struct elevator_type iosched_as = { |
1439 | .ops = { | |
1440 | .elevator_merge_fn = as_merge, | |
1441 | .elevator_merged_fn = as_merged_request, | |
1442 | .elevator_merge_req_fn = as_merged_requests, | |
b4878f24 JA |
1443 | .elevator_dispatch_fn = as_dispatch_request, |
1444 | .elevator_add_req_fn = as_add_request, | |
1445 | .elevator_activate_req_fn = as_activate_request, | |
1da177e4 LT |
1446 | .elevator_deactivate_req_fn = as_deactivate_request, |
1447 | .elevator_queue_empty_fn = as_queue_empty, | |
1448 | .elevator_completed_req_fn = as_completed_request, | |
e37f346e JA |
1449 | .elevator_former_req_fn = elv_rb_former_request, |
1450 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
1451 | .elevator_may_queue_fn = as_may_queue, |
1452 | .elevator_init_fn = as_init_queue, | |
1453 | .elevator_exit_fn = as_exit_queue, | |
e17a9489 | 1454 | .trim = as_trim, |
1da177e4 LT |
1455 | }, |
1456 | ||
3d1ab40f | 1457 | .elevator_attrs = as_attrs, |
1da177e4 LT |
1458 | .elevator_name = "anticipatory", |
1459 | .elevator_owner = THIS_MODULE, | |
1460 | }; | |
1461 | ||
1462 | static int __init as_init(void) | |
1463 | { | |
1464 | int ret; | |
1465 | ||
1da177e4 LT |
1466 | ret = elv_register(&iosched_as); |
1467 | if (!ret) { | |
1468 | /* | |
1469 | * don't allow AS to get unregistered, since we would have | |
1470 | * to browse all tasks in the system and release their | |
1471 | * as_io_context first | |
1472 | */ | |
1473 | __module_get(THIS_MODULE); | |
1474 | return 0; | |
1475 | } | |
1476 | ||
1da177e4 LT |
1477 | return ret; |
1478 | } | |
1479 | ||
1480 | static void __exit as_exit(void) | |
1481 | { | |
334e94de | 1482 | DECLARE_COMPLETION(all_gone); |
1da177e4 | 1483 | elv_unregister(&iosched_as); |
334e94de | 1484 | ioc_gone = &all_gone; |
fba82272 OH |
1485 | /* ioc_gone's update must be visible before reading ioc_count */ |
1486 | smp_wmb(); | |
334e94de | 1487 | if (atomic_read(&ioc_count)) |
fba82272 | 1488 | wait_for_completion(ioc_gone); |
334e94de | 1489 | synchronize_rcu(); |
1da177e4 LT |
1490 | } |
1491 | ||
1492 | module_init(as_init); | |
1493 | module_exit(as_exit); | |
1494 | ||
1495 | MODULE_AUTHOR("Nick Piggin"); | |
1496 | MODULE_LICENSE("GPL"); | |
1497 | MODULE_DESCRIPTION("anticipatory IO scheduler"); |