Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
1da177e4 LT |
2 | * Anticipatory & deadline i/o scheduler. |
3 | * | |
0fe23479 | 4 | * Copyright (C) 2002 Jens Axboe <axboe@kernel.dk> |
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 | |
e4313dd4 | 152 | static DEFINE_PER_CPU(unsigned long, ioc_count); |
334e94de AV |
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); | |
e4313dd4 JA |
166 | elv_ioc_count_dec(ioc_count); |
167 | if (ioc_gone && !elv_ioc_count_read(ioc_count)) | |
334e94de | 168 | complete(ioc_gone); |
1da177e4 LT |
169 | } |
170 | ||
e17a9489 AV |
171 | static void as_trim(struct io_context *ioc) |
172 | { | |
8bdd3f8a | 173 | spin_lock_irq(&ioc->lock); |
334e94de AV |
174 | if (ioc->aic) |
175 | free_as_io_context(ioc->aic); | |
e17a9489 | 176 | ioc->aic = NULL; |
8bdd3f8a | 177 | spin_unlock_irq(&ioc->lock); |
e17a9489 AV |
178 | } |
179 | ||
1da177e4 LT |
180 | /* Called when the task exits */ |
181 | static void exit_as_io_context(struct as_io_context *aic) | |
182 | { | |
183 | WARN_ON(!test_bit(AS_TASK_RUNNING, &aic->state)); | |
184 | clear_bit(AS_TASK_RUNNING, &aic->state); | |
185 | } | |
186 | ||
187 | static struct as_io_context *alloc_as_io_context(void) | |
188 | { | |
189 | struct as_io_context *ret; | |
190 | ||
191 | ret = kmalloc(sizeof(*ret), GFP_ATOMIC); | |
192 | if (ret) { | |
193 | ret->dtor = free_as_io_context; | |
194 | ret->exit = exit_as_io_context; | |
195 | ret->state = 1 << AS_TASK_RUNNING; | |
196 | atomic_set(&ret->nr_queued, 0); | |
197 | atomic_set(&ret->nr_dispatched, 0); | |
198 | spin_lock_init(&ret->lock); | |
199 | ret->ttime_total = 0; | |
200 | ret->ttime_samples = 0; | |
201 | ret->ttime_mean = 0; | |
202 | ret->seek_total = 0; | |
203 | ret->seek_samples = 0; | |
204 | ret->seek_mean = 0; | |
e4313dd4 | 205 | elv_ioc_count_inc(ioc_count); |
1da177e4 LT |
206 | } |
207 | ||
208 | return ret; | |
209 | } | |
210 | ||
211 | /* | |
212 | * If the current task has no AS IO context then create one and initialise it. | |
213 | * Then take a ref on the task's io context and return it. | |
214 | */ | |
b5deef90 | 215 | static struct io_context *as_get_io_context(int node) |
1da177e4 | 216 | { |
b5deef90 | 217 | struct io_context *ioc = get_io_context(GFP_ATOMIC, node); |
1da177e4 LT |
218 | if (ioc && !ioc->aic) { |
219 | ioc->aic = alloc_as_io_context(); | |
220 | if (!ioc->aic) { | |
221 | put_io_context(ioc); | |
222 | ioc = NULL; | |
223 | } | |
224 | } | |
225 | return ioc; | |
226 | } | |
227 | ||
8a8e674c | 228 | static void as_put_io_context(struct request *rq) |
b4878f24 JA |
229 | { |
230 | struct as_io_context *aic; | |
231 | ||
8a8e674c | 232 | if (unlikely(!RQ_IOC(rq))) |
b4878f24 JA |
233 | return; |
234 | ||
8a8e674c | 235 | aic = RQ_IOC(rq)->aic; |
b4878f24 | 236 | |
8a8e674c | 237 | if (rq_is_sync(rq) && aic) { |
8bdd3f8a JA |
238 | unsigned long flags; |
239 | ||
240 | spin_lock_irqsave(&aic->lock, flags); | |
b4878f24 JA |
241 | set_bit(AS_TASK_IORUNNING, &aic->state); |
242 | aic->last_end_request = jiffies; | |
8bdd3f8a | 243 | spin_unlock_irqrestore(&aic->lock, flags); |
b4878f24 JA |
244 | } |
245 | ||
8a8e674c | 246 | put_io_context(RQ_IOC(rq)); |
b4878f24 JA |
247 | } |
248 | ||
1da177e4 LT |
249 | /* |
250 | * rb tree support functions | |
251 | */ | |
9e2585a8 | 252 | #define RQ_RB_ROOT(ad, rq) (&(ad)->sort_list[rq_is_sync((rq))]) |
1da177e4 | 253 | |
8a8e674c | 254 | static void as_add_rq_rb(struct as_data *ad, struct request *rq) |
ef9be1d3 | 255 | { |
e37f346e | 256 | struct request *alias; |
ef9be1d3 | 257 | |
9e2585a8 | 258 | while ((unlikely(alias = elv_rb_add(RQ_RB_ROOT(ad, rq), rq)))) { |
8a8e674c | 259 | as_move_to_dispatch(ad, alias); |
ef9be1d3 TH |
260 | as_antic_stop(ad); |
261 | } | |
262 | } | |
263 | ||
8a8e674c | 264 | static inline void as_del_rq_rb(struct as_data *ad, struct request *rq) |
1da177e4 | 265 | { |
9e2585a8 | 266 | elv_rb_del(RQ_RB_ROOT(ad, rq), rq); |
1da177e4 LT |
267 | } |
268 | ||
269 | /* | |
270 | * IO Scheduler proper | |
271 | */ | |
272 | ||
273 | #define MAXBACK (1024 * 1024) /* | |
274 | * Maximum distance the disk will go backward | |
275 | * for a request. | |
276 | */ | |
277 | ||
278 | #define BACK_PENALTY 2 | |
279 | ||
280 | /* | |
281 | * as_choose_req selects the preferred one of two requests of the same data_dir | |
282 | * ignoring time - eg. timeouts, which is the job of as_dispatch_request | |
283 | */ | |
8a8e674c JA |
284 | static struct request * |
285 | as_choose_req(struct as_data *ad, struct request *rq1, struct request *rq2) | |
1da177e4 LT |
286 | { |
287 | int data_dir; | |
288 | sector_t last, s1, s2, d1, d2; | |
289 | int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */ | |
290 | const sector_t maxback = MAXBACK; | |
291 | ||
8a8e674c JA |
292 | if (rq1 == NULL || rq1 == rq2) |
293 | return rq2; | |
294 | if (rq2 == NULL) | |
295 | return rq1; | |
1da177e4 | 296 | |
8a8e674c | 297 | data_dir = rq_is_sync(rq1); |
1da177e4 LT |
298 | |
299 | last = ad->last_sector[data_dir]; | |
8a8e674c JA |
300 | s1 = rq1->sector; |
301 | s2 = rq2->sector; | |
1da177e4 | 302 | |
8a8e674c | 303 | BUG_ON(data_dir != rq_is_sync(rq2)); |
1da177e4 LT |
304 | |
305 | /* | |
306 | * Strict one way elevator _except_ in the case where we allow | |
307 | * short backward seeks which are biased as twice the cost of a | |
308 | * similar forward seek. | |
309 | */ | |
310 | if (s1 >= last) | |
311 | d1 = s1 - last; | |
312 | else if (s1+maxback >= last) | |
313 | d1 = (last - s1)*BACK_PENALTY; | |
314 | else { | |
315 | r1_wrap = 1; | |
316 | d1 = 0; /* shut up, gcc */ | |
317 | } | |
318 | ||
319 | if (s2 >= last) | |
320 | d2 = s2 - last; | |
321 | else if (s2+maxback >= last) | |
322 | d2 = (last - s2)*BACK_PENALTY; | |
323 | else { | |
324 | r2_wrap = 1; | |
325 | d2 = 0; | |
326 | } | |
327 | ||
328 | /* Found required data */ | |
329 | if (!r1_wrap && r2_wrap) | |
8a8e674c | 330 | return rq1; |
1da177e4 | 331 | else if (!r2_wrap && r1_wrap) |
8a8e674c | 332 | return rq2; |
1da177e4 LT |
333 | else if (r1_wrap && r2_wrap) { |
334 | /* both behind the head */ | |
335 | if (s1 <= s2) | |
8a8e674c | 336 | return rq1; |
1da177e4 | 337 | else |
8a8e674c | 338 | return rq2; |
1da177e4 LT |
339 | } |
340 | ||
341 | /* Both requests in front of the head */ | |
342 | if (d1 < d2) | |
8a8e674c | 343 | return rq1; |
1da177e4 | 344 | else if (d2 < d1) |
8a8e674c | 345 | return rq2; |
1da177e4 LT |
346 | else { |
347 | if (s1 >= s2) | |
8a8e674c | 348 | return rq1; |
1da177e4 | 349 | else |
8a8e674c | 350 | return rq2; |
1da177e4 LT |
351 | } |
352 | } | |
353 | ||
354 | /* | |
8a8e674c | 355 | * as_find_next_rq finds the next request after @prev in elevator order. |
1da177e4 LT |
356 | * this with as_choose_req form the basis for how the scheduler chooses |
357 | * what request to process next. Anticipation works on top of this. | |
358 | */ | |
8a8e674c JA |
359 | static struct request * |
360 | as_find_next_rq(struct as_data *ad, struct request *last) | |
1da177e4 | 361 | { |
1da177e4 LT |
362 | struct rb_node *rbnext = rb_next(&last->rb_node); |
363 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
8a8e674c | 364 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 365 | |
e37f346e | 366 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
367 | |
368 | if (rbprev) | |
8a8e674c | 369 | prev = rb_entry_rq(rbprev); |
1da177e4 LT |
370 | |
371 | if (rbnext) | |
8a8e674c | 372 | next = rb_entry_rq(rbnext); |
1da177e4 | 373 | else { |
9e2585a8 | 374 | const int data_dir = rq_is_sync(last); |
1da177e4 | 375 | |
e37f346e JA |
376 | rbnext = rb_first(&ad->sort_list[data_dir]); |
377 | if (rbnext && rbnext != &last->rb_node) | |
8a8e674c | 378 | next = rb_entry_rq(rbnext); |
e37f346e | 379 | } |
1da177e4 | 380 | |
e37f346e | 381 | return as_choose_req(ad, next, prev); |
1da177e4 LT |
382 | } |
383 | ||
384 | /* | |
385 | * anticipatory scheduling functions follow | |
386 | */ | |
387 | ||
388 | /* | |
389 | * as_antic_expired tells us when we have anticipated too long. | |
390 | * The funny "absolute difference" math on the elapsed time is to handle | |
391 | * jiffy wraps, and disks which have been idle for 0x80000000 jiffies. | |
392 | */ | |
393 | static int as_antic_expired(struct as_data *ad) | |
394 | { | |
395 | long delta_jif; | |
396 | ||
397 | delta_jif = jiffies - ad->antic_start; | |
398 | if (unlikely(delta_jif < 0)) | |
399 | delta_jif = -delta_jif; | |
400 | if (delta_jif < ad->antic_expire) | |
401 | return 0; | |
402 | ||
403 | return 1; | |
404 | } | |
405 | ||
406 | /* | |
407 | * as_antic_waitnext starts anticipating that a nice request will soon be | |
408 | * submitted. See also as_antic_waitreq | |
409 | */ | |
410 | static void as_antic_waitnext(struct as_data *ad) | |
411 | { | |
412 | unsigned long timeout; | |
413 | ||
414 | BUG_ON(ad->antic_status != ANTIC_OFF | |
415 | && ad->antic_status != ANTIC_WAIT_REQ); | |
416 | ||
417 | timeout = ad->antic_start + ad->antic_expire; | |
418 | ||
419 | mod_timer(&ad->antic_timer, timeout); | |
420 | ||
421 | ad->antic_status = ANTIC_WAIT_NEXT; | |
422 | } | |
423 | ||
424 | /* | |
425 | * as_antic_waitreq starts anticipating. We don't start timing the anticipation | |
426 | * until the request that we're anticipating on has finished. This means we | |
427 | * are timing from when the candidate process wakes up hopefully. | |
428 | */ | |
429 | static void as_antic_waitreq(struct as_data *ad) | |
430 | { | |
431 | BUG_ON(ad->antic_status == ANTIC_FINISHED); | |
432 | if (ad->antic_status == ANTIC_OFF) { | |
433 | if (!ad->io_context || ad->ioc_finished) | |
434 | as_antic_waitnext(ad); | |
435 | else | |
436 | ad->antic_status = ANTIC_WAIT_REQ; | |
437 | } | |
438 | } | |
439 | ||
440 | /* | |
441 | * This is called directly by the functions in this file to stop anticipation. | |
442 | * We kill the timer and schedule a call to the request_fn asap. | |
443 | */ | |
444 | static void as_antic_stop(struct as_data *ad) | |
445 | { | |
446 | int status = ad->antic_status; | |
447 | ||
448 | if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) { | |
449 | if (status == ANTIC_WAIT_NEXT) | |
450 | del_timer(&ad->antic_timer); | |
451 | ad->antic_status = ANTIC_FINISHED; | |
452 | /* see as_work_handler */ | |
453 | kblockd_schedule_work(&ad->antic_work); | |
454 | } | |
455 | } | |
456 | ||
457 | /* | |
458 | * as_antic_timeout is the timer function set by as_antic_waitnext. | |
459 | */ | |
460 | static void as_antic_timeout(unsigned long data) | |
461 | { | |
462 | struct request_queue *q = (struct request_queue *)data; | |
463 | struct as_data *ad = q->elevator->elevator_data; | |
464 | unsigned long flags; | |
465 | ||
466 | spin_lock_irqsave(q->queue_lock, flags); | |
467 | if (ad->antic_status == ANTIC_WAIT_REQ | |
468 | || ad->antic_status == ANTIC_WAIT_NEXT) { | |
521f3bbd JA |
469 | struct as_io_context *aic; |
470 | spin_lock(&ad->io_context->lock); | |
471 | aic = ad->io_context->aic; | |
1da177e4 LT |
472 | |
473 | ad->antic_status = ANTIC_FINISHED; | |
474 | kblockd_schedule_work(&ad->antic_work); | |
475 | ||
476 | if (aic->ttime_samples == 0) { | |
f5b3db00 | 477 | /* process anticipated on has exited or timed out*/ |
1da177e4 LT |
478 | ad->exit_prob = (7*ad->exit_prob + 256)/8; |
479 | } | |
f5b3db00 NP |
480 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
481 | /* process not "saved" by a cooperating request */ | |
482 | ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8; | |
483 | } | |
521f3bbd | 484 | spin_unlock(&ad->io_context->lock); |
1da177e4 LT |
485 | } |
486 | spin_unlock_irqrestore(q->queue_lock, flags); | |
487 | } | |
488 | ||
f5b3db00 NP |
489 | static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic, |
490 | unsigned long ttime) | |
491 | { | |
492 | /* fixed point: 1.0 == 1<<8 */ | |
493 | if (aic->ttime_samples == 0) { | |
494 | ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8; | |
495 | ad->new_ttime_mean = ad->new_ttime_total / 256; | |
496 | ||
497 | ad->exit_prob = (7*ad->exit_prob)/8; | |
498 | } | |
499 | aic->ttime_samples = (7*aic->ttime_samples + 256) / 8; | |
500 | aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8; | |
501 | aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples; | |
502 | } | |
503 | ||
504 | static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic, | |
505 | sector_t sdist) | |
506 | { | |
507 | u64 total; | |
508 | ||
509 | if (aic->seek_samples == 0) { | |
510 | ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8; | |
511 | ad->new_seek_mean = ad->new_seek_total / 256; | |
512 | } | |
513 | ||
514 | /* | |
515 | * Don't allow the seek distance to get too large from the | |
516 | * odd fragment, pagein, etc | |
517 | */ | |
518 | if (aic->seek_samples <= 60) /* second&third seek */ | |
519 | sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024); | |
520 | else | |
521 | sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64); | |
522 | ||
523 | aic->seek_samples = (7*aic->seek_samples + 256) / 8; | |
524 | aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8; | |
525 | total = aic->seek_total + (aic->seek_samples/2); | |
526 | do_div(total, aic->seek_samples); | |
527 | aic->seek_mean = (sector_t)total; | |
528 | } | |
529 | ||
530 | /* | |
531 | * as_update_iohist keeps a decaying histogram of IO thinktimes, and | |
532 | * updates @aic->ttime_mean based on that. It is called when a new | |
533 | * request is queued. | |
534 | */ | |
535 | static void as_update_iohist(struct as_data *ad, struct as_io_context *aic, | |
536 | struct request *rq) | |
537 | { | |
9e2585a8 | 538 | int data_dir = rq_is_sync(rq); |
f5b3db00 NP |
539 | unsigned long thinktime = 0; |
540 | sector_t seek_dist; | |
541 | ||
542 | if (aic == NULL) | |
543 | return; | |
544 | ||
545 | if (data_dir == REQ_SYNC) { | |
546 | unsigned long in_flight = atomic_read(&aic->nr_queued) | |
547 | + atomic_read(&aic->nr_dispatched); | |
548 | spin_lock(&aic->lock); | |
549 | if (test_bit(AS_TASK_IORUNNING, &aic->state) || | |
550 | test_bit(AS_TASK_IOSTARTED, &aic->state)) { | |
551 | /* Calculate read -> read thinktime */ | |
552 | if (test_bit(AS_TASK_IORUNNING, &aic->state) | |
553 | && in_flight == 0) { | |
554 | thinktime = jiffies - aic->last_end_request; | |
555 | thinktime = min(thinktime, MAX_THINKTIME-1); | |
556 | } | |
557 | as_update_thinktime(ad, aic, thinktime); | |
558 | ||
559 | /* Calculate read -> read seek distance */ | |
560 | if (aic->last_request_pos < rq->sector) | |
561 | seek_dist = rq->sector - aic->last_request_pos; | |
562 | else | |
563 | seek_dist = aic->last_request_pos - rq->sector; | |
564 | as_update_seekdist(ad, aic, seek_dist); | |
565 | } | |
566 | aic->last_request_pos = rq->sector + rq->nr_sectors; | |
567 | set_bit(AS_TASK_IOSTARTED, &aic->state); | |
568 | spin_unlock(&aic->lock); | |
569 | } | |
570 | } | |
571 | ||
1da177e4 LT |
572 | /* |
573 | * as_close_req decides if one request is considered "close" to the | |
574 | * previous one issued. | |
575 | */ | |
f5b3db00 | 576 | static int as_close_req(struct as_data *ad, struct as_io_context *aic, |
8a8e674c | 577 | struct request *rq) |
1da177e4 | 578 | { |
c6a632a2 | 579 | unsigned long delay; /* jiffies */ |
1da177e4 | 580 | sector_t last = ad->last_sector[ad->batch_data_dir]; |
8a8e674c | 581 | sector_t next = rq->sector; |
1da177e4 | 582 | sector_t delta; /* acceptable close offset (in sectors) */ |
f5b3db00 | 583 | sector_t s; |
1da177e4 LT |
584 | |
585 | if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished) | |
586 | delay = 0; | |
587 | else | |
c6a632a2 | 588 | delay = jiffies - ad->antic_start; |
1da177e4 | 589 | |
f5b3db00 NP |
590 | if (delay == 0) |
591 | delta = 8192; | |
c6a632a2 | 592 | else if (delay <= (20 * HZ / 1000) && delay <= ad->antic_expire) |
f5b3db00 | 593 | delta = 8192 << delay; |
1da177e4 LT |
594 | else |
595 | return 1; | |
596 | ||
f5b3db00 NP |
597 | if ((last <= next + (delta>>1)) && (next <= last + delta)) |
598 | return 1; | |
599 | ||
600 | if (last < next) | |
601 | s = next - last; | |
602 | else | |
603 | s = last - next; | |
604 | ||
605 | if (aic->seek_samples == 0) { | |
606 | /* | |
607 | * Process has just started IO. Use past statistics to | |
608 | * gauge success possibility | |
609 | */ | |
610 | if (ad->new_seek_mean > s) { | |
611 | /* this request is better than what we're expecting */ | |
612 | return 1; | |
613 | } | |
614 | ||
615 | } else { | |
616 | if (aic->seek_mean > s) { | |
617 | /* this request is better than what we're expecting */ | |
618 | return 1; | |
619 | } | |
620 | } | |
621 | ||
622 | return 0; | |
1da177e4 LT |
623 | } |
624 | ||
625 | /* | |
626 | * as_can_break_anticipation returns true if we have been anticipating this | |
627 | * request. | |
628 | * | |
629 | * It also returns true if the process against which we are anticipating | |
630 | * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to | |
631 | * dispatch it ASAP, because we know that application will not be submitting | |
632 | * any new reads. | |
633 | * | |
f5b3db00 | 634 | * If the task which has submitted the request has exited, break anticipation. |
1da177e4 LT |
635 | * |
636 | * If this task has queued some other IO, do not enter enticipation. | |
637 | */ | |
8a8e674c | 638 | static int as_can_break_anticipation(struct as_data *ad, struct request *rq) |
1da177e4 LT |
639 | { |
640 | struct io_context *ioc; | |
641 | struct as_io_context *aic; | |
1da177e4 LT |
642 | |
643 | ioc = ad->io_context; | |
644 | BUG_ON(!ioc); | |
521f3bbd | 645 | spin_lock(&ioc->lock); |
1da177e4 | 646 | |
8a8e674c | 647 | if (rq && ioc == RQ_IOC(rq)) { |
1da177e4 | 648 | /* request from same process */ |
521f3bbd | 649 | spin_unlock(&ioc->lock); |
1da177e4 LT |
650 | return 1; |
651 | } | |
652 | ||
653 | if (ad->ioc_finished && as_antic_expired(ad)) { | |
654 | /* | |
655 | * In this situation status should really be FINISHED, | |
656 | * however the timer hasn't had the chance to run yet. | |
657 | */ | |
521f3bbd | 658 | spin_unlock(&ioc->lock); |
1da177e4 LT |
659 | return 1; |
660 | } | |
661 | ||
662 | aic = ioc->aic; | |
521f3bbd JA |
663 | if (!aic) { |
664 | spin_unlock(&ioc->lock); | |
1da177e4 | 665 | return 0; |
521f3bbd | 666 | } |
1da177e4 | 667 | |
1da177e4 LT |
668 | if (atomic_read(&aic->nr_queued) > 0) { |
669 | /* process has more requests queued */ | |
521f3bbd | 670 | spin_unlock(&ioc->lock); |
1da177e4 LT |
671 | return 1; |
672 | } | |
673 | ||
674 | if (atomic_read(&aic->nr_dispatched) > 0) { | |
675 | /* process has more requests dispatched */ | |
521f3bbd | 676 | spin_unlock(&ioc->lock); |
1da177e4 LT |
677 | return 1; |
678 | } | |
679 | ||
8a8e674c | 680 | if (rq && rq_is_sync(rq) && as_close_req(ad, aic, rq)) { |
1da177e4 LT |
681 | /* |
682 | * Found a close request that is not one of ours. | |
683 | * | |
f5b3db00 NP |
684 | * This makes close requests from another process update |
685 | * our IO history. Is generally useful when there are | |
1da177e4 LT |
686 | * two or more cooperating processes working in the same |
687 | * area. | |
688 | */ | |
f5b3db00 NP |
689 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
690 | if (aic->ttime_samples == 0) | |
691 | ad->exit_prob = (7*ad->exit_prob + 256)/8; | |
692 | ||
693 | ad->exit_no_coop = (7*ad->exit_no_coop)/8; | |
694 | } | |
695 | ||
8a8e674c | 696 | as_update_iohist(ad, aic, rq); |
521f3bbd | 697 | spin_unlock(&ioc->lock); |
1da177e4 LT |
698 | return 1; |
699 | } | |
700 | ||
f5b3db00 NP |
701 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
702 | /* process anticipated on has exited */ | |
703 | if (aic->ttime_samples == 0) | |
704 | ad->exit_prob = (7*ad->exit_prob + 256)/8; | |
705 | ||
521f3bbd JA |
706 | if (ad->exit_no_coop > 128) { |
707 | spin_unlock(&ioc->lock); | |
f5b3db00 | 708 | return 1; |
521f3bbd | 709 | } |
f5b3db00 | 710 | } |
1da177e4 LT |
711 | |
712 | if (aic->ttime_samples == 0) { | |
521f3bbd JA |
713 | if (ad->new_ttime_mean > ad->antic_expire) { |
714 | spin_unlock(&ioc->lock); | |
1da177e4 | 715 | return 1; |
521f3bbd JA |
716 | } |
717 | if (ad->exit_prob * ad->exit_no_coop > 128*256) { | |
718 | spin_unlock(&ioc->lock); | |
1da177e4 | 719 | return 1; |
521f3bbd | 720 | } |
1da177e4 LT |
721 | } else if (aic->ttime_mean > ad->antic_expire) { |
722 | /* the process thinks too much between requests */ | |
521f3bbd | 723 | spin_unlock(&ioc->lock); |
1da177e4 LT |
724 | return 1; |
725 | } | |
521f3bbd | 726 | spin_unlock(&ioc->lock); |
1da177e4 LT |
727 | return 0; |
728 | } | |
729 | ||
730 | /* | |
8a8e674c | 731 | * as_can_anticipate indicates whether we should either run rq |
1da177e4 LT |
732 | * or keep anticipating a better request. |
733 | */ | |
8a8e674c | 734 | static int as_can_anticipate(struct as_data *ad, struct request *rq) |
1da177e4 LT |
735 | { |
736 | if (!ad->io_context) | |
737 | /* | |
738 | * Last request submitted was a write | |
739 | */ | |
740 | return 0; | |
741 | ||
742 | if (ad->antic_status == ANTIC_FINISHED) | |
743 | /* | |
744 | * Don't restart if we have just finished. Run the next request | |
745 | */ | |
746 | return 0; | |
747 | ||
8a8e674c | 748 | if (as_can_break_anticipation(ad, rq)) |
1da177e4 LT |
749 | /* |
750 | * This request is a good candidate. Don't keep anticipating, | |
751 | * run it. | |
752 | */ | |
753 | return 0; | |
754 | ||
755 | /* | |
756 | * OK from here, we haven't finished, and don't have a decent request! | |
757 | * Status is either ANTIC_OFF so start waiting, | |
758 | * ANTIC_WAIT_REQ so continue waiting for request to finish | |
759 | * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request. | |
1da177e4 LT |
760 | */ |
761 | ||
762 | return 1; | |
763 | } | |
764 | ||
1da177e4 | 765 | /* |
8a8e674c | 766 | * as_update_rq must be called whenever a request (rq) is added to |
1da177e4 LT |
767 | * the sort_list. This function keeps caches up to date, and checks if the |
768 | * request might be one we are "anticipating" | |
769 | */ | |
8a8e674c | 770 | static void as_update_rq(struct as_data *ad, struct request *rq) |
1da177e4 | 771 | { |
8a8e674c | 772 | const int data_dir = rq_is_sync(rq); |
1da177e4 | 773 | |
8a8e674c JA |
774 | /* keep the next_rq cache up to date */ |
775 | ad->next_rq[data_dir] = as_choose_req(ad, rq, ad->next_rq[data_dir]); | |
1da177e4 LT |
776 | |
777 | /* | |
778 | * have we been anticipating this request? | |
779 | * or does it come from the same process as the one we are anticipating | |
780 | * for? | |
781 | */ | |
782 | if (ad->antic_status == ANTIC_WAIT_REQ | |
783 | || ad->antic_status == ANTIC_WAIT_NEXT) { | |
8a8e674c | 784 | if (as_can_break_anticipation(ad, rq)) |
1da177e4 LT |
785 | as_antic_stop(ad); |
786 | } | |
787 | } | |
788 | ||
789 | /* | |
790 | * Gathers timings and resizes the write batch automatically | |
791 | */ | |
792 | static void update_write_batch(struct as_data *ad) | |
793 | { | |
794 | unsigned long batch = ad->batch_expire[REQ_ASYNC]; | |
795 | long write_time; | |
796 | ||
797 | write_time = (jiffies - ad->current_batch_expires) + batch; | |
798 | if (write_time < 0) | |
799 | write_time = 0; | |
800 | ||
801 | if (write_time > batch && !ad->write_batch_idled) { | |
802 | if (write_time > batch * 3) | |
803 | ad->write_batch_count /= 2; | |
804 | else | |
805 | ad->write_batch_count--; | |
806 | } else if (write_time < batch && ad->current_write_count == 0) { | |
807 | if (batch > write_time * 3) | |
808 | ad->write_batch_count *= 2; | |
809 | else | |
810 | ad->write_batch_count++; | |
811 | } | |
812 | ||
813 | if (ad->write_batch_count < 1) | |
814 | ad->write_batch_count = 1; | |
815 | } | |
816 | ||
817 | /* | |
818 | * as_completed_request is to be called when a request has completed and | |
819 | * returned something to the requesting process, be it an error or data. | |
820 | */ | |
165125e1 | 821 | static void as_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 LT |
822 | { |
823 | struct as_data *ad = q->elevator->elevator_data; | |
1da177e4 LT |
824 | |
825 | WARN_ON(!list_empty(&rq->queuelist)); | |
826 | ||
8a8e674c JA |
827 | if (RQ_STATE(rq) != AS_RQ_REMOVED) { |
828 | printk("rq->state %d\n", RQ_STATE(rq)); | |
1da177e4 LT |
829 | WARN_ON(1); |
830 | goto out; | |
831 | } | |
832 | ||
1da177e4 LT |
833 | if (ad->changed_batch && ad->nr_dispatched == 1) { |
834 | kblockd_schedule_work(&ad->antic_work); | |
835 | ad->changed_batch = 0; | |
836 | ||
837 | if (ad->batch_data_dir == REQ_SYNC) | |
838 | ad->new_batch = 1; | |
839 | } | |
840 | WARN_ON(ad->nr_dispatched == 0); | |
841 | ad->nr_dispatched--; | |
842 | ||
843 | /* | |
844 | * Start counting the batch from when a request of that direction is | |
845 | * actually serviced. This should help devices with big TCQ windows | |
846 | * and writeback caches | |
847 | */ | |
9e2585a8 | 848 | if (ad->new_batch && ad->batch_data_dir == rq_is_sync(rq)) { |
1da177e4 LT |
849 | update_write_batch(ad); |
850 | ad->current_batch_expires = jiffies + | |
851 | ad->batch_expire[REQ_SYNC]; | |
852 | ad->new_batch = 0; | |
853 | } | |
854 | ||
8a8e674c | 855 | if (ad->io_context == RQ_IOC(rq) && ad->io_context) { |
1da177e4 LT |
856 | ad->antic_start = jiffies; |
857 | ad->ioc_finished = 1; | |
858 | if (ad->antic_status == ANTIC_WAIT_REQ) { | |
859 | /* | |
860 | * We were waiting on this request, now anticipate | |
861 | * the next one | |
862 | */ | |
863 | as_antic_waitnext(ad); | |
864 | } | |
865 | } | |
866 | ||
8a8e674c | 867 | as_put_io_context(rq); |
1da177e4 | 868 | out: |
8a8e674c | 869 | RQ_SET_STATE(rq, AS_RQ_POSTSCHED); |
1da177e4 LT |
870 | } |
871 | ||
872 | /* | |
873 | * as_remove_queued_request removes a request from the pre dispatch queue | |
874 | * without updating refcounts. It is expected the caller will drop the | |
875 | * reference unless it replaces the request at somepart of the elevator | |
876 | * (ie. the dispatch queue) | |
877 | */ | |
165125e1 JA |
878 | static void as_remove_queued_request(struct request_queue *q, |
879 | struct request *rq) | |
1da177e4 | 880 | { |
9e2585a8 | 881 | const int data_dir = rq_is_sync(rq); |
1da177e4 | 882 | struct as_data *ad = q->elevator->elevator_data; |
8a8e674c | 883 | struct io_context *ioc; |
1da177e4 | 884 | |
8a8e674c | 885 | WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED); |
1da177e4 | 886 | |
8a8e674c JA |
887 | ioc = RQ_IOC(rq); |
888 | if (ioc && ioc->aic) { | |
889 | BUG_ON(!atomic_read(&ioc->aic->nr_queued)); | |
890 | atomic_dec(&ioc->aic->nr_queued); | |
1da177e4 LT |
891 | } |
892 | ||
893 | /* | |
8a8e674c | 894 | * Update the "next_rq" cache if we are about to remove its |
1da177e4 LT |
895 | * entry |
896 | */ | |
8a8e674c JA |
897 | if (ad->next_rq[data_dir] == rq) |
898 | ad->next_rq[data_dir] = as_find_next_rq(ad, rq); | |
1da177e4 | 899 | |
d4f2f462 | 900 | rq_fifo_clear(rq); |
8a8e674c | 901 | as_del_rq_rb(ad, rq); |
1da177e4 LT |
902 | } |
903 | ||
1da177e4 | 904 | /* |
8896f3c0 | 905 | * as_fifo_expired returns 0 if there are no expired requests on the fifo, |
1da177e4 LT |
906 | * 1 otherwise. It is ratelimited so that we only perform the check once per |
907 | * `fifo_expire' interval. Otherwise a large number of expired requests | |
908 | * would create a hopeless seekstorm. | |
909 | * | |
910 | * See as_antic_expired comment. | |
911 | */ | |
912 | static int as_fifo_expired(struct as_data *ad, int adir) | |
913 | { | |
d4f2f462 | 914 | struct request *rq; |
1da177e4 LT |
915 | long delta_jif; |
916 | ||
917 | delta_jif = jiffies - ad->last_check_fifo[adir]; | |
918 | if (unlikely(delta_jif < 0)) | |
919 | delta_jif = -delta_jif; | |
920 | if (delta_jif < ad->fifo_expire[adir]) | |
921 | return 0; | |
922 | ||
923 | ad->last_check_fifo[adir] = jiffies; | |
924 | ||
925 | if (list_empty(&ad->fifo_list[adir])) | |
926 | return 0; | |
927 | ||
d4f2f462 | 928 | rq = rq_entry_fifo(ad->fifo_list[adir].next); |
1da177e4 | 929 | |
d4f2f462 | 930 | return time_after(jiffies, rq_fifo_time(rq)); |
1da177e4 LT |
931 | } |
932 | ||
933 | /* | |
934 | * as_batch_expired returns true if the current batch has expired. A batch | |
935 | * is a set of reads or a set of writes. | |
936 | */ | |
937 | static inline int as_batch_expired(struct as_data *ad) | |
938 | { | |
939 | if (ad->changed_batch || ad->new_batch) | |
940 | return 0; | |
941 | ||
942 | if (ad->batch_data_dir == REQ_SYNC) | |
943 | /* TODO! add a check so a complete fifo gets written? */ | |
944 | return time_after(jiffies, ad->current_batch_expires); | |
945 | ||
946 | return time_after(jiffies, ad->current_batch_expires) | |
947 | || ad->current_write_count == 0; | |
948 | } | |
949 | ||
950 | /* | |
951 | * move an entry to dispatch queue | |
952 | */ | |
8a8e674c | 953 | static void as_move_to_dispatch(struct as_data *ad, struct request *rq) |
1da177e4 | 954 | { |
9e2585a8 | 955 | const int data_dir = rq_is_sync(rq); |
1da177e4 | 956 | |
e37f346e | 957 | BUG_ON(RB_EMPTY_NODE(&rq->rb_node)); |
1da177e4 LT |
958 | |
959 | as_antic_stop(ad); | |
960 | ad->antic_status = ANTIC_OFF; | |
961 | ||
962 | /* | |
963 | * This has to be set in order to be correctly updated by | |
8a8e674c | 964 | * as_find_next_rq |
1da177e4 LT |
965 | */ |
966 | ad->last_sector[data_dir] = rq->sector + rq->nr_sectors; | |
967 | ||
968 | if (data_dir == REQ_SYNC) { | |
8a8e674c | 969 | struct io_context *ioc = RQ_IOC(rq); |
1da177e4 | 970 | /* In case we have to anticipate after this */ |
8a8e674c | 971 | copy_io_context(&ad->io_context, &ioc); |
1da177e4 LT |
972 | } else { |
973 | if (ad->io_context) { | |
974 | put_io_context(ad->io_context); | |
975 | ad->io_context = NULL; | |
976 | } | |
977 | ||
978 | if (ad->current_write_count != 0) | |
979 | ad->current_write_count--; | |
980 | } | |
981 | ad->ioc_finished = 0; | |
982 | ||
8a8e674c | 983 | ad->next_rq[data_dir] = as_find_next_rq(ad, rq); |
1da177e4 LT |
984 | |
985 | /* | |
986 | * take it off the sort and fifo list, add to dispatch queue | |
987 | */ | |
1da177e4 | 988 | as_remove_queued_request(ad->q, rq); |
8a8e674c | 989 | WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED); |
1da177e4 | 990 | |
b4878f24 JA |
991 | elv_dispatch_sort(ad->q, rq); |
992 | ||
8a8e674c JA |
993 | RQ_SET_STATE(rq, AS_RQ_DISPATCHED); |
994 | if (RQ_IOC(rq) && RQ_IOC(rq)->aic) | |
995 | atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched); | |
1da177e4 LT |
996 | ad->nr_dispatched++; |
997 | } | |
998 | ||
999 | /* | |
1000 | * as_dispatch_request selects the best request according to | |
1001 | * read/write expire, batch expire, etc, and moves it to the dispatch | |
1002 | * queue. Returns 1 if a request was found, 0 otherwise. | |
1003 | */ | |
165125e1 | 1004 | static int as_dispatch_request(struct request_queue *q, int force) |
1da177e4 | 1005 | { |
b4878f24 | 1006 | struct as_data *ad = q->elevator->elevator_data; |
1da177e4 LT |
1007 | const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]); |
1008 | const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]); | |
8a8e674c | 1009 | struct request *rq; |
1da177e4 | 1010 | |
b4878f24 JA |
1011 | if (unlikely(force)) { |
1012 | /* | |
1013 | * Forced dispatch, accounting is useless. Reset | |
1014 | * accounting states and dump fifo_lists. Note that | |
1015 | * batch_data_dir is reset to REQ_SYNC to avoid | |
1016 | * screwing write batch accounting as write batch | |
1017 | * accounting occurs on W->R transition. | |
1018 | */ | |
1019 | int dispatched = 0; | |
1020 | ||
1021 | ad->batch_data_dir = REQ_SYNC; | |
1022 | ad->changed_batch = 0; | |
1023 | ad->new_batch = 0; | |
1024 | ||
8a8e674c JA |
1025 | while (ad->next_rq[REQ_SYNC]) { |
1026 | as_move_to_dispatch(ad, ad->next_rq[REQ_SYNC]); | |
b4878f24 JA |
1027 | dispatched++; |
1028 | } | |
1029 | ad->last_check_fifo[REQ_SYNC] = jiffies; | |
1030 | ||
8a8e674c JA |
1031 | while (ad->next_rq[REQ_ASYNC]) { |
1032 | as_move_to_dispatch(ad, ad->next_rq[REQ_ASYNC]); | |
b4878f24 JA |
1033 | dispatched++; |
1034 | } | |
1035 | ad->last_check_fifo[REQ_ASYNC] = jiffies; | |
1036 | ||
1037 | return dispatched; | |
1038 | } | |
1039 | ||
1da177e4 LT |
1040 | /* Signal that the write batch was uncontended, so we can't time it */ |
1041 | if (ad->batch_data_dir == REQ_ASYNC && !reads) { | |
1042 | if (ad->current_write_count == 0 || !writes) | |
1043 | ad->write_batch_idled = 1; | |
1044 | } | |
1045 | ||
1046 | if (!(reads || writes) | |
1047 | || ad->antic_status == ANTIC_WAIT_REQ | |
1048 | || ad->antic_status == ANTIC_WAIT_NEXT | |
1049 | || ad->changed_batch) | |
1050 | return 0; | |
1051 | ||
f5b3db00 | 1052 | if (!(reads && writes && as_batch_expired(ad))) { |
1da177e4 LT |
1053 | /* |
1054 | * batch is still running or no reads or no writes | |
1055 | */ | |
8a8e674c | 1056 | rq = ad->next_rq[ad->batch_data_dir]; |
1da177e4 LT |
1057 | |
1058 | if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) { | |
1059 | if (as_fifo_expired(ad, REQ_SYNC)) | |
1060 | goto fifo_expired; | |
1061 | ||
8a8e674c | 1062 | if (as_can_anticipate(ad, rq)) { |
1da177e4 LT |
1063 | as_antic_waitreq(ad); |
1064 | return 0; | |
1065 | } | |
1066 | } | |
1067 | ||
8a8e674c | 1068 | if (rq) { |
1da177e4 LT |
1069 | /* we have a "next request" */ |
1070 | if (reads && !writes) | |
1071 | ad->current_batch_expires = | |
1072 | jiffies + ad->batch_expire[REQ_SYNC]; | |
1073 | goto dispatch_request; | |
1074 | } | |
1075 | } | |
1076 | ||
1077 | /* | |
1078 | * at this point we are not running a batch. select the appropriate | |
1079 | * data direction (read / write) | |
1080 | */ | |
1081 | ||
1082 | if (reads) { | |
dd67d051 | 1083 | BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_SYNC])); |
1da177e4 LT |
1084 | |
1085 | if (writes && ad->batch_data_dir == REQ_SYNC) | |
1086 | /* | |
1087 | * Last batch was a read, switch to writes | |
1088 | */ | |
1089 | goto dispatch_writes; | |
1090 | ||
1091 | if (ad->batch_data_dir == REQ_ASYNC) { | |
1092 | WARN_ON(ad->new_batch); | |
1093 | ad->changed_batch = 1; | |
1094 | } | |
1095 | ad->batch_data_dir = REQ_SYNC; | |
8a8e674c | 1096 | rq = rq_entry_fifo(ad->fifo_list[REQ_SYNC].next); |
1da177e4 LT |
1097 | ad->last_check_fifo[ad->batch_data_dir] = jiffies; |
1098 | goto dispatch_request; | |
1099 | } | |
1100 | ||
1101 | /* | |
1102 | * the last batch was a read | |
1103 | */ | |
1104 | ||
1105 | if (writes) { | |
1106 | dispatch_writes: | |
dd67d051 | 1107 | BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_ASYNC])); |
1da177e4 LT |
1108 | |
1109 | if (ad->batch_data_dir == REQ_SYNC) { | |
1110 | ad->changed_batch = 1; | |
1111 | ||
1112 | /* | |
1113 | * new_batch might be 1 when the queue runs out of | |
1114 | * reads. A subsequent submission of a write might | |
1115 | * cause a change of batch before the read is finished. | |
1116 | */ | |
1117 | ad->new_batch = 0; | |
1118 | } | |
1119 | ad->batch_data_dir = REQ_ASYNC; | |
1120 | ad->current_write_count = ad->write_batch_count; | |
1121 | ad->write_batch_idled = 0; | |
49565124 AC |
1122 | rq = rq_entry_fifo(ad->fifo_list[REQ_ASYNC].next); |
1123 | ad->last_check_fifo[REQ_ASYNC] = jiffies; | |
1da177e4 LT |
1124 | goto dispatch_request; |
1125 | } | |
1126 | ||
1127 | BUG(); | |
1128 | return 0; | |
1129 | ||
1130 | dispatch_request: | |
1131 | /* | |
1132 | * If a request has expired, service it. | |
1133 | */ | |
1134 | ||
1135 | if (as_fifo_expired(ad, ad->batch_data_dir)) { | |
1136 | fifo_expired: | |
8a8e674c | 1137 | rq = rq_entry_fifo(ad->fifo_list[ad->batch_data_dir].next); |
1da177e4 LT |
1138 | } |
1139 | ||
1140 | if (ad->changed_batch) { | |
1141 | WARN_ON(ad->new_batch); | |
1142 | ||
1143 | if (ad->nr_dispatched) | |
1144 | return 0; | |
1145 | ||
1146 | if (ad->batch_data_dir == REQ_ASYNC) | |
1147 | ad->current_batch_expires = jiffies + | |
1148 | ad->batch_expire[REQ_ASYNC]; | |
1149 | else | |
1150 | ad->new_batch = 1; | |
1151 | ||
1152 | ad->changed_batch = 0; | |
1153 | } | |
1154 | ||
1155 | /* | |
8a8e674c | 1156 | * rq is the selected appropriate request. |
1da177e4 | 1157 | */ |
8a8e674c | 1158 | as_move_to_dispatch(ad, rq); |
1da177e4 LT |
1159 | |
1160 | return 1; | |
1161 | } | |
1162 | ||
1da177e4 | 1163 | /* |
8a8e674c | 1164 | * add rq to rbtree and fifo |
1da177e4 | 1165 | */ |
165125e1 | 1166 | static void as_add_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1167 | { |
b4878f24 | 1168 | struct as_data *ad = q->elevator->elevator_data; |
1da177e4 LT |
1169 | int data_dir; |
1170 | ||
8a8e674c | 1171 | RQ_SET_STATE(rq, AS_RQ_NEW); |
b4878f24 | 1172 | |
9e2585a8 | 1173 | data_dir = rq_is_sync(rq); |
1da177e4 | 1174 | |
b5deef90 | 1175 | rq->elevator_private = as_get_io_context(q->node); |
1da177e4 | 1176 | |
8a8e674c JA |
1177 | if (RQ_IOC(rq)) { |
1178 | as_update_iohist(ad, RQ_IOC(rq)->aic, rq); | |
1179 | atomic_inc(&RQ_IOC(rq)->aic->nr_queued); | |
1da177e4 LT |
1180 | } |
1181 | ||
8a8e674c | 1182 | as_add_rq_rb(ad, rq); |
1da177e4 | 1183 | |
ef9be1d3 | 1184 | /* |
8896f3c0 | 1185 | * set expire time and add to fifo list |
ef9be1d3 | 1186 | */ |
d4f2f462 JA |
1187 | rq_set_fifo_time(rq, jiffies + ad->fifo_expire[data_dir]); |
1188 | list_add_tail(&rq->queuelist, &ad->fifo_list[data_dir]); | |
1da177e4 | 1189 | |
8a8e674c JA |
1190 | as_update_rq(ad, rq); /* keep state machine up to date */ |
1191 | RQ_SET_STATE(rq, AS_RQ_QUEUED); | |
1da177e4 LT |
1192 | } |
1193 | ||
165125e1 | 1194 | static void as_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1195 | { |
8a8e674c JA |
1196 | WARN_ON(RQ_STATE(rq) != AS_RQ_DISPATCHED); |
1197 | RQ_SET_STATE(rq, AS_RQ_REMOVED); | |
1198 | if (RQ_IOC(rq) && RQ_IOC(rq)->aic) | |
1199 | atomic_dec(&RQ_IOC(rq)->aic->nr_dispatched); | |
1da177e4 LT |
1200 | } |
1201 | ||
165125e1 | 1202 | static void as_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1203 | { |
8a8e674c JA |
1204 | WARN_ON(RQ_STATE(rq) != AS_RQ_REMOVED); |
1205 | RQ_SET_STATE(rq, AS_RQ_DISPATCHED); | |
1206 | if (RQ_IOC(rq) && RQ_IOC(rq)->aic) | |
1207 | atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched); | |
1da177e4 LT |
1208 | } |
1209 | ||
1210 | /* | |
1211 | * as_queue_empty tells us if there are requests left in the device. It may | |
1212 | * not be the case that a driver can get the next request even if the queue | |
1213 | * is not empty - it is used in the block layer to check for plugging and | |
1214 | * merging opportunities | |
1215 | */ | |
165125e1 | 1216 | static int as_queue_empty(struct request_queue *q) |
1da177e4 LT |
1217 | { |
1218 | struct as_data *ad = q->elevator->elevator_data; | |
1219 | ||
b4878f24 JA |
1220 | return list_empty(&ad->fifo_list[REQ_ASYNC]) |
1221 | && list_empty(&ad->fifo_list[REQ_SYNC]); | |
1da177e4 LT |
1222 | } |
1223 | ||
1da177e4 | 1224 | static int |
165125e1 | 1225 | as_merge(struct request_queue *q, struct request **req, struct bio *bio) |
1da177e4 LT |
1226 | { |
1227 | struct as_data *ad = q->elevator->elevator_data; | |
1228 | sector_t rb_key = bio->bi_sector + bio_sectors(bio); | |
1229 | struct request *__rq; | |
1da177e4 LT |
1230 | |
1231 | /* | |
1232 | * check for front merge | |
1233 | */ | |
e37f346e | 1234 | __rq = elv_rb_find(&ad->sort_list[bio_data_dir(bio)], rb_key); |
9817064b JA |
1235 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
1236 | *req = __rq; | |
1237 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
1238 | } |
1239 | ||
1240 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
1241 | } |
1242 | ||
165125e1 JA |
1243 | static void as_merged_request(struct request_queue *q, struct request *req, |
1244 | int type) | |
1da177e4 LT |
1245 | { |
1246 | struct as_data *ad = q->elevator->elevator_data; | |
1da177e4 | 1247 | |
1da177e4 LT |
1248 | /* |
1249 | * if the merge was a front merge, we need to reposition request | |
1250 | */ | |
e37f346e | 1251 | if (type == ELEVATOR_FRONT_MERGE) { |
8a8e674c JA |
1252 | as_del_rq_rb(ad, req); |
1253 | as_add_rq_rb(ad, req); | |
1da177e4 LT |
1254 | /* |
1255 | * Note! At this stage of this and the next function, our next | |
1256 | * request may not be optimal - eg the request may have "grown" | |
1257 | * behind the disk head. We currently don't bother adjusting. | |
1258 | */ | |
1259 | } | |
1da177e4 LT |
1260 | } |
1261 | ||
165125e1 | 1262 | static void as_merged_requests(struct request_queue *q, struct request *req, |
f5b3db00 | 1263 | struct request *next) |
1da177e4 | 1264 | { |
1da177e4 | 1265 | /* |
8a8e674c JA |
1266 | * if next expires before rq, assign its expire time to arq |
1267 | * and move into next position (next will be deleted) in fifo | |
1da177e4 | 1268 | */ |
d4f2f462 JA |
1269 | if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) { |
1270 | if (time_before(rq_fifo_time(next), rq_fifo_time(req))) { | |
1271 | list_move(&req->queuelist, &next->queuelist); | |
1272 | rq_set_fifo_time(req, rq_fifo_time(next)); | |
1da177e4 LT |
1273 | } |
1274 | } | |
1275 | ||
1da177e4 LT |
1276 | /* |
1277 | * kill knowledge of next, this one is a goner | |
1278 | */ | |
1279 | as_remove_queued_request(q, next); | |
8a8e674c | 1280 | as_put_io_context(next); |
1da177e4 | 1281 | |
8a8e674c | 1282 | RQ_SET_STATE(next, AS_RQ_MERGED); |
1da177e4 LT |
1283 | } |
1284 | ||
1285 | /* | |
1286 | * This is executed in a "deferred" process context, by kblockd. It calls the | |
1287 | * driver's request_fn so the driver can submit that request. | |
1288 | * | |
1289 | * IMPORTANT! This guy will reenter the elevator, so set up all queue global | |
1290 | * state before calling, and don't rely on any state over calls. | |
1291 | * | |
1292 | * FIXME! dispatch queue is not a queue at all! | |
1293 | */ | |
65f27f38 | 1294 | static void as_work_handler(struct work_struct *work) |
1da177e4 | 1295 | { |
65f27f38 DH |
1296 | struct as_data *ad = container_of(work, struct as_data, antic_work); |
1297 | struct request_queue *q = ad->q; | |
1da177e4 LT |
1298 | unsigned long flags; |
1299 | ||
1300 | spin_lock_irqsave(q->queue_lock, flags); | |
dc72ef4a | 1301 | blk_start_queueing(q); |
1da177e4 LT |
1302 | spin_unlock_irqrestore(q->queue_lock, flags); |
1303 | } | |
1304 | ||
165125e1 | 1305 | static int as_may_queue(struct request_queue *q, int rw) |
1da177e4 LT |
1306 | { |
1307 | int ret = ELV_MQUEUE_MAY; | |
1308 | struct as_data *ad = q->elevator->elevator_data; | |
1309 | struct io_context *ioc; | |
1310 | if (ad->antic_status == ANTIC_WAIT_REQ || | |
1311 | ad->antic_status == ANTIC_WAIT_NEXT) { | |
b5deef90 | 1312 | ioc = as_get_io_context(q->node); |
1da177e4 LT |
1313 | if (ad->io_context == ioc) |
1314 | ret = ELV_MQUEUE_MUST; | |
1315 | put_io_context(ioc); | |
1316 | } | |
1317 | ||
1318 | return ret; | |
1319 | } | |
1320 | ||
1321 | static void as_exit_queue(elevator_t *e) | |
1322 | { | |
1323 | struct as_data *ad = e->elevator_data; | |
1324 | ||
1325 | del_timer_sync(&ad->antic_timer); | |
19a75d83 | 1326 | kblockd_flush_work(&ad->antic_work); |
1da177e4 LT |
1327 | |
1328 | BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC])); | |
1329 | BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC])); | |
1330 | ||
1da177e4 | 1331 | put_io_context(ad->io_context); |
1da177e4 LT |
1332 | kfree(ad); |
1333 | } | |
1334 | ||
1335 | /* | |
8a8e674c | 1336 | * initialize elevator private data (as_data). |
1da177e4 | 1337 | */ |
165125e1 | 1338 | static void *as_init_queue(struct request_queue *q) |
1da177e4 LT |
1339 | { |
1340 | struct as_data *ad; | |
1da177e4 | 1341 | |
94f6030c | 1342 | ad = kmalloc_node(sizeof(*ad), GFP_KERNEL | __GFP_ZERO, q->node); |
1da177e4 | 1343 | if (!ad) |
bc1c1169 | 1344 | return NULL; |
1da177e4 LT |
1345 | |
1346 | ad->q = q; /* Identify what queue the data belongs to */ | |
1347 | ||
1da177e4 LT |
1348 | /* anticipatory scheduling helpers */ |
1349 | ad->antic_timer.function = as_antic_timeout; | |
1350 | ad->antic_timer.data = (unsigned long)q; | |
1351 | init_timer(&ad->antic_timer); | |
65f27f38 | 1352 | INIT_WORK(&ad->antic_work, as_work_handler); |
1da177e4 | 1353 | |
1da177e4 LT |
1354 | INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]); |
1355 | INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]); | |
1356 | ad->sort_list[REQ_SYNC] = RB_ROOT; | |
1357 | ad->sort_list[REQ_ASYNC] = RB_ROOT; | |
1da177e4 LT |
1358 | ad->fifo_expire[REQ_SYNC] = default_read_expire; |
1359 | ad->fifo_expire[REQ_ASYNC] = default_write_expire; | |
1360 | ad->antic_expire = default_antic_expire; | |
1361 | ad->batch_expire[REQ_SYNC] = default_read_batch_expire; | |
1362 | ad->batch_expire[REQ_ASYNC] = default_write_batch_expire; | |
1da177e4 LT |
1363 | |
1364 | ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC]; | |
1365 | ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10; | |
1366 | if (ad->write_batch_count < 2) | |
1367 | ad->write_batch_count = 2; | |
1368 | ||
bc1c1169 | 1369 | return ad; |
1da177e4 LT |
1370 | } |
1371 | ||
1372 | /* | |
1373 | * sysfs parts below | |
1374 | */ | |
1da177e4 LT |
1375 | |
1376 | static ssize_t | |
1377 | as_var_show(unsigned int var, char *page) | |
1378 | { | |
1da177e4 LT |
1379 | return sprintf(page, "%d\n", var); |
1380 | } | |
1381 | ||
1382 | static ssize_t | |
1383 | as_var_store(unsigned long *var, const char *page, size_t count) | |
1384 | { | |
1da177e4 LT |
1385 | char *p = (char *) page; |
1386 | ||
c9b3ad67 | 1387 | *var = simple_strtoul(p, &p, 10); |
1da177e4 LT |
1388 | return count; |
1389 | } | |
1390 | ||
e572ec7e | 1391 | static ssize_t est_time_show(elevator_t *e, char *page) |
1da177e4 | 1392 | { |
3d1ab40f | 1393 | struct as_data *ad = e->elevator_data; |
1da177e4 LT |
1394 | int pos = 0; |
1395 | ||
f5b3db00 NP |
1396 | pos += sprintf(page+pos, "%lu %% exit probability\n", |
1397 | 100*ad->exit_prob/256); | |
1398 | pos += sprintf(page+pos, "%lu %% probability of exiting without a " | |
1399 | "cooperating process submitting IO\n", | |
1400 | 100*ad->exit_no_coop/256); | |
1da177e4 | 1401 | pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean); |
f5b3db00 NP |
1402 | pos += sprintf(page+pos, "%llu sectors new seek distance\n", |
1403 | (unsigned long long)ad->new_seek_mean); | |
1da177e4 LT |
1404 | |
1405 | return pos; | |
1406 | } | |
1407 | ||
1408 | #define SHOW_FUNCTION(__FUNC, __VAR) \ | |
3d1ab40f | 1409 | static ssize_t __FUNC(elevator_t *e, char *page) \ |
1da177e4 | 1410 | { \ |
3d1ab40f | 1411 | struct as_data *ad = e->elevator_data; \ |
1da177e4 LT |
1412 | return as_var_show(jiffies_to_msecs((__VAR)), (page)); \ |
1413 | } | |
e572ec7e AV |
1414 | SHOW_FUNCTION(as_read_expire_show, ad->fifo_expire[REQ_SYNC]); |
1415 | SHOW_FUNCTION(as_write_expire_show, ad->fifo_expire[REQ_ASYNC]); | |
1416 | SHOW_FUNCTION(as_antic_expire_show, ad->antic_expire); | |
1417 | SHOW_FUNCTION(as_read_batch_expire_show, ad->batch_expire[REQ_SYNC]); | |
1418 | SHOW_FUNCTION(as_write_batch_expire_show, ad->batch_expire[REQ_ASYNC]); | |
1da177e4 LT |
1419 | #undef SHOW_FUNCTION |
1420 | ||
1421 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \ | |
3d1ab40f | 1422 | static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \ |
1da177e4 | 1423 | { \ |
3d1ab40f AV |
1424 | struct as_data *ad = e->elevator_data; \ |
1425 | int ret = as_var_store(__PTR, (page), count); \ | |
1da177e4 LT |
1426 | if (*(__PTR) < (MIN)) \ |
1427 | *(__PTR) = (MIN); \ | |
1428 | else if (*(__PTR) > (MAX)) \ | |
1429 | *(__PTR) = (MAX); \ | |
1430 | *(__PTR) = msecs_to_jiffies(*(__PTR)); \ | |
1431 | return ret; \ | |
1432 | } | |
e572ec7e AV |
1433 | STORE_FUNCTION(as_read_expire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX); |
1434 | STORE_FUNCTION(as_write_expire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX); | |
1435 | STORE_FUNCTION(as_antic_expire_store, &ad->antic_expire, 0, INT_MAX); | |
1436 | STORE_FUNCTION(as_read_batch_expire_store, | |
1da177e4 | 1437 | &ad->batch_expire[REQ_SYNC], 0, INT_MAX); |
e572ec7e | 1438 | STORE_FUNCTION(as_write_batch_expire_store, |
1da177e4 LT |
1439 | &ad->batch_expire[REQ_ASYNC], 0, INT_MAX); |
1440 | #undef STORE_FUNCTION | |
1441 | ||
e572ec7e AV |
1442 | #define AS_ATTR(name) \ |
1443 | __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store) | |
1444 | ||
1445 | static struct elv_fs_entry as_attrs[] = { | |
1446 | __ATTR_RO(est_time), | |
1447 | AS_ATTR(read_expire), | |
1448 | AS_ATTR(write_expire), | |
1449 | AS_ATTR(antic_expire), | |
1450 | AS_ATTR(read_batch_expire), | |
1451 | AS_ATTR(write_batch_expire), | |
1452 | __ATTR_NULL | |
1da177e4 LT |
1453 | }; |
1454 | ||
1da177e4 LT |
1455 | static struct elevator_type iosched_as = { |
1456 | .ops = { | |
1457 | .elevator_merge_fn = as_merge, | |
1458 | .elevator_merged_fn = as_merged_request, | |
1459 | .elevator_merge_req_fn = as_merged_requests, | |
b4878f24 JA |
1460 | .elevator_dispatch_fn = as_dispatch_request, |
1461 | .elevator_add_req_fn = as_add_request, | |
1462 | .elevator_activate_req_fn = as_activate_request, | |
1da177e4 LT |
1463 | .elevator_deactivate_req_fn = as_deactivate_request, |
1464 | .elevator_queue_empty_fn = as_queue_empty, | |
1465 | .elevator_completed_req_fn = as_completed_request, | |
e37f346e JA |
1466 | .elevator_former_req_fn = elv_rb_former_request, |
1467 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
1468 | .elevator_may_queue_fn = as_may_queue, |
1469 | .elevator_init_fn = as_init_queue, | |
1470 | .elevator_exit_fn = as_exit_queue, | |
e17a9489 | 1471 | .trim = as_trim, |
1da177e4 LT |
1472 | }, |
1473 | ||
3d1ab40f | 1474 | .elevator_attrs = as_attrs, |
1da177e4 LT |
1475 | .elevator_name = "anticipatory", |
1476 | .elevator_owner = THIS_MODULE, | |
1477 | }; | |
1478 | ||
1479 | static int __init as_init(void) | |
1480 | { | |
2fdd82bd AB |
1481 | elv_register(&iosched_as); |
1482 | ||
1483 | return 0; | |
1da177e4 LT |
1484 | } |
1485 | ||
1486 | static void __exit as_exit(void) | |
1487 | { | |
6e9a4738 | 1488 | DECLARE_COMPLETION_ONSTACK(all_gone); |
1da177e4 | 1489 | elv_unregister(&iosched_as); |
334e94de | 1490 | ioc_gone = &all_gone; |
fba82272 OH |
1491 | /* ioc_gone's update must be visible before reading ioc_count */ |
1492 | smp_wmb(); | |
e4313dd4 | 1493 | if (elv_ioc_count_read(ioc_count)) |
fba82272 | 1494 | wait_for_completion(ioc_gone); |
334e94de | 1495 | synchronize_rcu(); |
1da177e4 LT |
1496 | } |
1497 | ||
1498 | module_init(as_init); | |
1499 | module_exit(as_exit); | |
1500 | ||
1501 | MODULE_AUTHOR("Nick Piggin"); | |
1502 | MODULE_LICENSE("GPL"); | |
1503 | MODULE_DESCRIPTION("anticipatory IO scheduler"); |