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 | { | |
521f3bbd | 173 | spin_lock(&ioc->lock); |
334e94de AV |
174 | if (ioc->aic) |
175 | free_as_io_context(ioc->aic); | |
e17a9489 | 176 | ioc->aic = NULL; |
521f3bbd | 177 | spin_unlock(&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) { |
b4878f24 JA |
238 | spin_lock(&aic->lock); |
239 | set_bit(AS_TASK_IORUNNING, &aic->state); | |
240 | aic->last_end_request = jiffies; | |
241 | spin_unlock(&aic->lock); | |
242 | } | |
243 | ||
8a8e674c | 244 | put_io_context(RQ_IOC(rq)); |
b4878f24 JA |
245 | } |
246 | ||
1da177e4 LT |
247 | /* |
248 | * rb tree support functions | |
249 | */ | |
9e2585a8 | 250 | #define RQ_RB_ROOT(ad, rq) (&(ad)->sort_list[rq_is_sync((rq))]) |
1da177e4 | 251 | |
8a8e674c | 252 | static void as_add_rq_rb(struct as_data *ad, struct request *rq) |
ef9be1d3 | 253 | { |
e37f346e | 254 | struct request *alias; |
ef9be1d3 | 255 | |
9e2585a8 | 256 | while ((unlikely(alias = elv_rb_add(RQ_RB_ROOT(ad, rq), rq)))) { |
8a8e674c | 257 | as_move_to_dispatch(ad, alias); |
ef9be1d3 TH |
258 | as_antic_stop(ad); |
259 | } | |
260 | } | |
261 | ||
8a8e674c | 262 | static inline void as_del_rq_rb(struct as_data *ad, struct request *rq) |
1da177e4 | 263 | { |
9e2585a8 | 264 | elv_rb_del(RQ_RB_ROOT(ad, rq), rq); |
1da177e4 LT |
265 | } |
266 | ||
267 | /* | |
268 | * IO Scheduler proper | |
269 | */ | |
270 | ||
271 | #define MAXBACK (1024 * 1024) /* | |
272 | * Maximum distance the disk will go backward | |
273 | * for a request. | |
274 | */ | |
275 | ||
276 | #define BACK_PENALTY 2 | |
277 | ||
278 | /* | |
279 | * as_choose_req selects the preferred one of two requests of the same data_dir | |
280 | * ignoring time - eg. timeouts, which is the job of as_dispatch_request | |
281 | */ | |
8a8e674c JA |
282 | static struct request * |
283 | as_choose_req(struct as_data *ad, struct request *rq1, struct request *rq2) | |
1da177e4 LT |
284 | { |
285 | int data_dir; | |
286 | sector_t last, s1, s2, d1, d2; | |
287 | int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */ | |
288 | const sector_t maxback = MAXBACK; | |
289 | ||
8a8e674c JA |
290 | if (rq1 == NULL || rq1 == rq2) |
291 | return rq2; | |
292 | if (rq2 == NULL) | |
293 | return rq1; | |
1da177e4 | 294 | |
8a8e674c | 295 | data_dir = rq_is_sync(rq1); |
1da177e4 LT |
296 | |
297 | last = ad->last_sector[data_dir]; | |
8a8e674c JA |
298 | s1 = rq1->sector; |
299 | s2 = rq2->sector; | |
1da177e4 | 300 | |
8a8e674c | 301 | BUG_ON(data_dir != rq_is_sync(rq2)); |
1da177e4 LT |
302 | |
303 | /* | |
304 | * Strict one way elevator _except_ in the case where we allow | |
305 | * short backward seeks which are biased as twice the cost of a | |
306 | * similar forward seek. | |
307 | */ | |
308 | if (s1 >= last) | |
309 | d1 = s1 - last; | |
310 | else if (s1+maxback >= last) | |
311 | d1 = (last - s1)*BACK_PENALTY; | |
312 | else { | |
313 | r1_wrap = 1; | |
314 | d1 = 0; /* shut up, gcc */ | |
315 | } | |
316 | ||
317 | if (s2 >= last) | |
318 | d2 = s2 - last; | |
319 | else if (s2+maxback >= last) | |
320 | d2 = (last - s2)*BACK_PENALTY; | |
321 | else { | |
322 | r2_wrap = 1; | |
323 | d2 = 0; | |
324 | } | |
325 | ||
326 | /* Found required data */ | |
327 | if (!r1_wrap && r2_wrap) | |
8a8e674c | 328 | return rq1; |
1da177e4 | 329 | else if (!r2_wrap && r1_wrap) |
8a8e674c | 330 | return rq2; |
1da177e4 LT |
331 | else if (r1_wrap && r2_wrap) { |
332 | /* both behind the head */ | |
333 | if (s1 <= s2) | |
8a8e674c | 334 | return rq1; |
1da177e4 | 335 | else |
8a8e674c | 336 | return rq2; |
1da177e4 LT |
337 | } |
338 | ||
339 | /* Both requests in front of the head */ | |
340 | if (d1 < d2) | |
8a8e674c | 341 | return rq1; |
1da177e4 | 342 | else if (d2 < d1) |
8a8e674c | 343 | return rq2; |
1da177e4 LT |
344 | else { |
345 | if (s1 >= s2) | |
8a8e674c | 346 | return rq1; |
1da177e4 | 347 | else |
8a8e674c | 348 | return rq2; |
1da177e4 LT |
349 | } |
350 | } | |
351 | ||
352 | /* | |
8a8e674c | 353 | * as_find_next_rq finds the next request after @prev in elevator order. |
1da177e4 LT |
354 | * this with as_choose_req form the basis for how the scheduler chooses |
355 | * what request to process next. Anticipation works on top of this. | |
356 | */ | |
8a8e674c JA |
357 | static struct request * |
358 | as_find_next_rq(struct as_data *ad, struct request *last) | |
1da177e4 | 359 | { |
1da177e4 LT |
360 | struct rb_node *rbnext = rb_next(&last->rb_node); |
361 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
8a8e674c | 362 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 363 | |
e37f346e | 364 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
365 | |
366 | if (rbprev) | |
8a8e674c | 367 | prev = rb_entry_rq(rbprev); |
1da177e4 LT |
368 | |
369 | if (rbnext) | |
8a8e674c | 370 | next = rb_entry_rq(rbnext); |
1da177e4 | 371 | else { |
9e2585a8 | 372 | const int data_dir = rq_is_sync(last); |
1da177e4 | 373 | |
e37f346e JA |
374 | rbnext = rb_first(&ad->sort_list[data_dir]); |
375 | if (rbnext && rbnext != &last->rb_node) | |
8a8e674c | 376 | next = rb_entry_rq(rbnext); |
e37f346e | 377 | } |
1da177e4 | 378 | |
e37f346e | 379 | return as_choose_req(ad, next, prev); |
1da177e4 LT |
380 | } |
381 | ||
382 | /* | |
383 | * anticipatory scheduling functions follow | |
384 | */ | |
385 | ||
386 | /* | |
387 | * as_antic_expired tells us when we have anticipated too long. | |
388 | * The funny "absolute difference" math on the elapsed time is to handle | |
389 | * jiffy wraps, and disks which have been idle for 0x80000000 jiffies. | |
390 | */ | |
391 | static int as_antic_expired(struct as_data *ad) | |
392 | { | |
393 | long delta_jif; | |
394 | ||
395 | delta_jif = jiffies - ad->antic_start; | |
396 | if (unlikely(delta_jif < 0)) | |
397 | delta_jif = -delta_jif; | |
398 | if (delta_jif < ad->antic_expire) | |
399 | return 0; | |
400 | ||
401 | return 1; | |
402 | } | |
403 | ||
404 | /* | |
405 | * as_antic_waitnext starts anticipating that a nice request will soon be | |
406 | * submitted. See also as_antic_waitreq | |
407 | */ | |
408 | static void as_antic_waitnext(struct as_data *ad) | |
409 | { | |
410 | unsigned long timeout; | |
411 | ||
412 | BUG_ON(ad->antic_status != ANTIC_OFF | |
413 | && ad->antic_status != ANTIC_WAIT_REQ); | |
414 | ||
415 | timeout = ad->antic_start + ad->antic_expire; | |
416 | ||
417 | mod_timer(&ad->antic_timer, timeout); | |
418 | ||
419 | ad->antic_status = ANTIC_WAIT_NEXT; | |
420 | } | |
421 | ||
422 | /* | |
423 | * as_antic_waitreq starts anticipating. We don't start timing the anticipation | |
424 | * until the request that we're anticipating on has finished. This means we | |
425 | * are timing from when the candidate process wakes up hopefully. | |
426 | */ | |
427 | static void as_antic_waitreq(struct as_data *ad) | |
428 | { | |
429 | BUG_ON(ad->antic_status == ANTIC_FINISHED); | |
430 | if (ad->antic_status == ANTIC_OFF) { | |
431 | if (!ad->io_context || ad->ioc_finished) | |
432 | as_antic_waitnext(ad); | |
433 | else | |
434 | ad->antic_status = ANTIC_WAIT_REQ; | |
435 | } | |
436 | } | |
437 | ||
438 | /* | |
439 | * This is called directly by the functions in this file to stop anticipation. | |
440 | * We kill the timer and schedule a call to the request_fn asap. | |
441 | */ | |
442 | static void as_antic_stop(struct as_data *ad) | |
443 | { | |
444 | int status = ad->antic_status; | |
445 | ||
446 | if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) { | |
447 | if (status == ANTIC_WAIT_NEXT) | |
448 | del_timer(&ad->antic_timer); | |
449 | ad->antic_status = ANTIC_FINISHED; | |
450 | /* see as_work_handler */ | |
451 | kblockd_schedule_work(&ad->antic_work); | |
452 | } | |
453 | } | |
454 | ||
455 | /* | |
456 | * as_antic_timeout is the timer function set by as_antic_waitnext. | |
457 | */ | |
458 | static void as_antic_timeout(unsigned long data) | |
459 | { | |
460 | struct request_queue *q = (struct request_queue *)data; | |
461 | struct as_data *ad = q->elevator->elevator_data; | |
462 | unsigned long flags; | |
463 | ||
464 | spin_lock_irqsave(q->queue_lock, flags); | |
465 | if (ad->antic_status == ANTIC_WAIT_REQ | |
466 | || ad->antic_status == ANTIC_WAIT_NEXT) { | |
521f3bbd JA |
467 | struct as_io_context *aic; |
468 | spin_lock(&ad->io_context->lock); | |
469 | aic = ad->io_context->aic; | |
1da177e4 LT |
470 | |
471 | ad->antic_status = ANTIC_FINISHED; | |
472 | kblockd_schedule_work(&ad->antic_work); | |
473 | ||
474 | if (aic->ttime_samples == 0) { | |
f5b3db00 | 475 | /* process anticipated on has exited or timed out*/ |
1da177e4 LT |
476 | ad->exit_prob = (7*ad->exit_prob + 256)/8; |
477 | } | |
f5b3db00 NP |
478 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
479 | /* process not "saved" by a cooperating request */ | |
480 | ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8; | |
481 | } | |
521f3bbd | 482 | spin_unlock(&ad->io_context->lock); |
1da177e4 LT |
483 | } |
484 | spin_unlock_irqrestore(q->queue_lock, flags); | |
485 | } | |
486 | ||
f5b3db00 NP |
487 | static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic, |
488 | unsigned long ttime) | |
489 | { | |
490 | /* fixed point: 1.0 == 1<<8 */ | |
491 | if (aic->ttime_samples == 0) { | |
492 | ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8; | |
493 | ad->new_ttime_mean = ad->new_ttime_total / 256; | |
494 | ||
495 | ad->exit_prob = (7*ad->exit_prob)/8; | |
496 | } | |
497 | aic->ttime_samples = (7*aic->ttime_samples + 256) / 8; | |
498 | aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8; | |
499 | aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples; | |
500 | } | |
501 | ||
502 | static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic, | |
503 | sector_t sdist) | |
504 | { | |
505 | u64 total; | |
506 | ||
507 | if (aic->seek_samples == 0) { | |
508 | ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8; | |
509 | ad->new_seek_mean = ad->new_seek_total / 256; | |
510 | } | |
511 | ||
512 | /* | |
513 | * Don't allow the seek distance to get too large from the | |
514 | * odd fragment, pagein, etc | |
515 | */ | |
516 | if (aic->seek_samples <= 60) /* second&third seek */ | |
517 | sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024); | |
518 | else | |
519 | sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64); | |
520 | ||
521 | aic->seek_samples = (7*aic->seek_samples + 256) / 8; | |
522 | aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8; | |
523 | total = aic->seek_total + (aic->seek_samples/2); | |
524 | do_div(total, aic->seek_samples); | |
525 | aic->seek_mean = (sector_t)total; | |
526 | } | |
527 | ||
528 | /* | |
529 | * as_update_iohist keeps a decaying histogram of IO thinktimes, and | |
530 | * updates @aic->ttime_mean based on that. It is called when a new | |
531 | * request is queued. | |
532 | */ | |
533 | static void as_update_iohist(struct as_data *ad, struct as_io_context *aic, | |
534 | struct request *rq) | |
535 | { | |
9e2585a8 | 536 | int data_dir = rq_is_sync(rq); |
f5b3db00 NP |
537 | unsigned long thinktime = 0; |
538 | sector_t seek_dist; | |
539 | ||
540 | if (aic == NULL) | |
541 | return; | |
542 | ||
543 | if (data_dir == REQ_SYNC) { | |
544 | unsigned long in_flight = atomic_read(&aic->nr_queued) | |
545 | + atomic_read(&aic->nr_dispatched); | |
546 | spin_lock(&aic->lock); | |
547 | if (test_bit(AS_TASK_IORUNNING, &aic->state) || | |
548 | test_bit(AS_TASK_IOSTARTED, &aic->state)) { | |
549 | /* Calculate read -> read thinktime */ | |
550 | if (test_bit(AS_TASK_IORUNNING, &aic->state) | |
551 | && in_flight == 0) { | |
552 | thinktime = jiffies - aic->last_end_request; | |
553 | thinktime = min(thinktime, MAX_THINKTIME-1); | |
554 | } | |
555 | as_update_thinktime(ad, aic, thinktime); | |
556 | ||
557 | /* Calculate read -> read seek distance */ | |
558 | if (aic->last_request_pos < rq->sector) | |
559 | seek_dist = rq->sector - aic->last_request_pos; | |
560 | else | |
561 | seek_dist = aic->last_request_pos - rq->sector; | |
562 | as_update_seekdist(ad, aic, seek_dist); | |
563 | } | |
564 | aic->last_request_pos = rq->sector + rq->nr_sectors; | |
565 | set_bit(AS_TASK_IOSTARTED, &aic->state); | |
566 | spin_unlock(&aic->lock); | |
567 | } | |
568 | } | |
569 | ||
1da177e4 LT |
570 | /* |
571 | * as_close_req decides if one request is considered "close" to the | |
572 | * previous one issued. | |
573 | */ | |
f5b3db00 | 574 | static int as_close_req(struct as_data *ad, struct as_io_context *aic, |
8a8e674c | 575 | struct request *rq) |
1da177e4 | 576 | { |
c6a632a2 | 577 | unsigned long delay; /* jiffies */ |
1da177e4 | 578 | sector_t last = ad->last_sector[ad->batch_data_dir]; |
8a8e674c | 579 | sector_t next = rq->sector; |
1da177e4 | 580 | sector_t delta; /* acceptable close offset (in sectors) */ |
f5b3db00 | 581 | sector_t s; |
1da177e4 LT |
582 | |
583 | if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished) | |
584 | delay = 0; | |
585 | else | |
c6a632a2 | 586 | delay = jiffies - ad->antic_start; |
1da177e4 | 587 | |
f5b3db00 NP |
588 | if (delay == 0) |
589 | delta = 8192; | |
c6a632a2 | 590 | else if (delay <= (20 * HZ / 1000) && delay <= ad->antic_expire) |
f5b3db00 | 591 | delta = 8192 << delay; |
1da177e4 LT |
592 | else |
593 | return 1; | |
594 | ||
f5b3db00 NP |
595 | if ((last <= next + (delta>>1)) && (next <= last + delta)) |
596 | return 1; | |
597 | ||
598 | if (last < next) | |
599 | s = next - last; | |
600 | else | |
601 | s = last - next; | |
602 | ||
603 | if (aic->seek_samples == 0) { | |
604 | /* | |
605 | * Process has just started IO. Use past statistics to | |
606 | * gauge success possibility | |
607 | */ | |
608 | if (ad->new_seek_mean > s) { | |
609 | /* this request is better than what we're expecting */ | |
610 | return 1; | |
611 | } | |
612 | ||
613 | } else { | |
614 | if (aic->seek_mean > s) { | |
615 | /* this request is better than what we're expecting */ | |
616 | return 1; | |
617 | } | |
618 | } | |
619 | ||
620 | return 0; | |
1da177e4 LT |
621 | } |
622 | ||
623 | /* | |
624 | * as_can_break_anticipation returns true if we have been anticipating this | |
625 | * request. | |
626 | * | |
627 | * It also returns true if the process against which we are anticipating | |
628 | * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to | |
629 | * dispatch it ASAP, because we know that application will not be submitting | |
630 | * any new reads. | |
631 | * | |
f5b3db00 | 632 | * If the task which has submitted the request has exited, break anticipation. |
1da177e4 LT |
633 | * |
634 | * If this task has queued some other IO, do not enter enticipation. | |
635 | */ | |
8a8e674c | 636 | static int as_can_break_anticipation(struct as_data *ad, struct request *rq) |
1da177e4 LT |
637 | { |
638 | struct io_context *ioc; | |
639 | struct as_io_context *aic; | |
1da177e4 LT |
640 | |
641 | ioc = ad->io_context; | |
642 | BUG_ON(!ioc); | |
521f3bbd | 643 | spin_lock(&ioc->lock); |
1da177e4 | 644 | |
8a8e674c | 645 | if (rq && ioc == RQ_IOC(rq)) { |
1da177e4 | 646 | /* request from same process */ |
521f3bbd | 647 | spin_unlock(&ioc->lock); |
1da177e4 LT |
648 | return 1; |
649 | } | |
650 | ||
651 | if (ad->ioc_finished && as_antic_expired(ad)) { | |
652 | /* | |
653 | * In this situation status should really be FINISHED, | |
654 | * however the timer hasn't had the chance to run yet. | |
655 | */ | |
521f3bbd | 656 | spin_unlock(&ioc->lock); |
1da177e4 LT |
657 | return 1; |
658 | } | |
659 | ||
660 | aic = ioc->aic; | |
521f3bbd JA |
661 | if (!aic) { |
662 | spin_unlock(&ioc->lock); | |
1da177e4 | 663 | return 0; |
521f3bbd | 664 | } |
1da177e4 | 665 | |
1da177e4 LT |
666 | if (atomic_read(&aic->nr_queued) > 0) { |
667 | /* process has more requests queued */ | |
521f3bbd | 668 | spin_unlock(&ioc->lock); |
1da177e4 LT |
669 | return 1; |
670 | } | |
671 | ||
672 | if (atomic_read(&aic->nr_dispatched) > 0) { | |
673 | /* process has more requests dispatched */ | |
521f3bbd | 674 | spin_unlock(&ioc->lock); |
1da177e4 LT |
675 | return 1; |
676 | } | |
677 | ||
8a8e674c | 678 | if (rq && rq_is_sync(rq) && as_close_req(ad, aic, rq)) { |
1da177e4 LT |
679 | /* |
680 | * Found a close request that is not one of ours. | |
681 | * | |
f5b3db00 NP |
682 | * This makes close requests from another process update |
683 | * our IO history. Is generally useful when there are | |
1da177e4 LT |
684 | * two or more cooperating processes working in the same |
685 | * area. | |
686 | */ | |
f5b3db00 NP |
687 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
688 | if (aic->ttime_samples == 0) | |
689 | ad->exit_prob = (7*ad->exit_prob + 256)/8; | |
690 | ||
691 | ad->exit_no_coop = (7*ad->exit_no_coop)/8; | |
692 | } | |
693 | ||
8a8e674c | 694 | as_update_iohist(ad, aic, rq); |
521f3bbd | 695 | spin_unlock(&ioc->lock); |
1da177e4 LT |
696 | return 1; |
697 | } | |
698 | ||
f5b3db00 NP |
699 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
700 | /* process anticipated on has exited */ | |
701 | if (aic->ttime_samples == 0) | |
702 | ad->exit_prob = (7*ad->exit_prob + 256)/8; | |
703 | ||
521f3bbd JA |
704 | if (ad->exit_no_coop > 128) { |
705 | spin_unlock(&ioc->lock); | |
f5b3db00 | 706 | return 1; |
521f3bbd | 707 | } |
f5b3db00 | 708 | } |
1da177e4 LT |
709 | |
710 | if (aic->ttime_samples == 0) { | |
521f3bbd JA |
711 | if (ad->new_ttime_mean > ad->antic_expire) { |
712 | spin_unlock(&ioc->lock); | |
1da177e4 | 713 | return 1; |
521f3bbd JA |
714 | } |
715 | if (ad->exit_prob * ad->exit_no_coop > 128*256) { | |
716 | spin_unlock(&ioc->lock); | |
1da177e4 | 717 | return 1; |
521f3bbd | 718 | } |
1da177e4 LT |
719 | } else if (aic->ttime_mean > ad->antic_expire) { |
720 | /* the process thinks too much between requests */ | |
521f3bbd | 721 | spin_unlock(&ioc->lock); |
1da177e4 LT |
722 | return 1; |
723 | } | |
521f3bbd | 724 | spin_unlock(&ioc->lock); |
1da177e4 LT |
725 | return 0; |
726 | } | |
727 | ||
728 | /* | |
8a8e674c | 729 | * as_can_anticipate indicates whether we should either run rq |
1da177e4 LT |
730 | * or keep anticipating a better request. |
731 | */ | |
8a8e674c | 732 | static int as_can_anticipate(struct as_data *ad, struct request *rq) |
1da177e4 LT |
733 | { |
734 | if (!ad->io_context) | |
735 | /* | |
736 | * Last request submitted was a write | |
737 | */ | |
738 | return 0; | |
739 | ||
740 | if (ad->antic_status == ANTIC_FINISHED) | |
741 | /* | |
742 | * Don't restart if we have just finished. Run the next request | |
743 | */ | |
744 | return 0; | |
745 | ||
8a8e674c | 746 | if (as_can_break_anticipation(ad, rq)) |
1da177e4 LT |
747 | /* |
748 | * This request is a good candidate. Don't keep anticipating, | |
749 | * run it. | |
750 | */ | |
751 | return 0; | |
752 | ||
753 | /* | |
754 | * OK from here, we haven't finished, and don't have a decent request! | |
755 | * Status is either ANTIC_OFF so start waiting, | |
756 | * ANTIC_WAIT_REQ so continue waiting for request to finish | |
757 | * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request. | |
1da177e4 LT |
758 | */ |
759 | ||
760 | return 1; | |
761 | } | |
762 | ||
1da177e4 | 763 | /* |
8a8e674c | 764 | * as_update_rq must be called whenever a request (rq) is added to |
1da177e4 LT |
765 | * the sort_list. This function keeps caches up to date, and checks if the |
766 | * request might be one we are "anticipating" | |
767 | */ | |
8a8e674c | 768 | static void as_update_rq(struct as_data *ad, struct request *rq) |
1da177e4 | 769 | { |
8a8e674c | 770 | const int data_dir = rq_is_sync(rq); |
1da177e4 | 771 | |
8a8e674c JA |
772 | /* keep the next_rq cache up to date */ |
773 | ad->next_rq[data_dir] = as_choose_req(ad, rq, ad->next_rq[data_dir]); | |
1da177e4 LT |
774 | |
775 | /* | |
776 | * have we been anticipating this request? | |
777 | * or does it come from the same process as the one we are anticipating | |
778 | * for? | |
779 | */ | |
780 | if (ad->antic_status == ANTIC_WAIT_REQ | |
781 | || ad->antic_status == ANTIC_WAIT_NEXT) { | |
8a8e674c | 782 | if (as_can_break_anticipation(ad, rq)) |
1da177e4 LT |
783 | as_antic_stop(ad); |
784 | } | |
785 | } | |
786 | ||
787 | /* | |
788 | * Gathers timings and resizes the write batch automatically | |
789 | */ | |
790 | static void update_write_batch(struct as_data *ad) | |
791 | { | |
792 | unsigned long batch = ad->batch_expire[REQ_ASYNC]; | |
793 | long write_time; | |
794 | ||
795 | write_time = (jiffies - ad->current_batch_expires) + batch; | |
796 | if (write_time < 0) | |
797 | write_time = 0; | |
798 | ||
799 | if (write_time > batch && !ad->write_batch_idled) { | |
800 | if (write_time > batch * 3) | |
801 | ad->write_batch_count /= 2; | |
802 | else | |
803 | ad->write_batch_count--; | |
804 | } else if (write_time < batch && ad->current_write_count == 0) { | |
805 | if (batch > write_time * 3) | |
806 | ad->write_batch_count *= 2; | |
807 | else | |
808 | ad->write_batch_count++; | |
809 | } | |
810 | ||
811 | if (ad->write_batch_count < 1) | |
812 | ad->write_batch_count = 1; | |
813 | } | |
814 | ||
815 | /* | |
816 | * as_completed_request is to be called when a request has completed and | |
817 | * returned something to the requesting process, be it an error or data. | |
818 | */ | |
165125e1 | 819 | static void as_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 LT |
820 | { |
821 | struct as_data *ad = q->elevator->elevator_data; | |
1da177e4 LT |
822 | |
823 | WARN_ON(!list_empty(&rq->queuelist)); | |
824 | ||
8a8e674c JA |
825 | if (RQ_STATE(rq) != AS_RQ_REMOVED) { |
826 | printk("rq->state %d\n", RQ_STATE(rq)); | |
1da177e4 LT |
827 | WARN_ON(1); |
828 | goto out; | |
829 | } | |
830 | ||
1da177e4 LT |
831 | if (ad->changed_batch && ad->nr_dispatched == 1) { |
832 | kblockd_schedule_work(&ad->antic_work); | |
833 | ad->changed_batch = 0; | |
834 | ||
835 | if (ad->batch_data_dir == REQ_SYNC) | |
836 | ad->new_batch = 1; | |
837 | } | |
838 | WARN_ON(ad->nr_dispatched == 0); | |
839 | ad->nr_dispatched--; | |
840 | ||
841 | /* | |
842 | * Start counting the batch from when a request of that direction is | |
843 | * actually serviced. This should help devices with big TCQ windows | |
844 | * and writeback caches | |
845 | */ | |
9e2585a8 | 846 | if (ad->new_batch && ad->batch_data_dir == rq_is_sync(rq)) { |
1da177e4 LT |
847 | update_write_batch(ad); |
848 | ad->current_batch_expires = jiffies + | |
849 | ad->batch_expire[REQ_SYNC]; | |
850 | ad->new_batch = 0; | |
851 | } | |
852 | ||
8a8e674c | 853 | if (ad->io_context == RQ_IOC(rq) && ad->io_context) { |
1da177e4 LT |
854 | ad->antic_start = jiffies; |
855 | ad->ioc_finished = 1; | |
856 | if (ad->antic_status == ANTIC_WAIT_REQ) { | |
857 | /* | |
858 | * We were waiting on this request, now anticipate | |
859 | * the next one | |
860 | */ | |
861 | as_antic_waitnext(ad); | |
862 | } | |
863 | } | |
864 | ||
8a8e674c | 865 | as_put_io_context(rq); |
1da177e4 | 866 | out: |
8a8e674c | 867 | RQ_SET_STATE(rq, AS_RQ_POSTSCHED); |
1da177e4 LT |
868 | } |
869 | ||
870 | /* | |
871 | * as_remove_queued_request removes a request from the pre dispatch queue | |
872 | * without updating refcounts. It is expected the caller will drop the | |
873 | * reference unless it replaces the request at somepart of the elevator | |
874 | * (ie. the dispatch queue) | |
875 | */ | |
165125e1 JA |
876 | static void as_remove_queued_request(struct request_queue *q, |
877 | struct request *rq) | |
1da177e4 | 878 | { |
9e2585a8 | 879 | const int data_dir = rq_is_sync(rq); |
1da177e4 | 880 | struct as_data *ad = q->elevator->elevator_data; |
8a8e674c | 881 | struct io_context *ioc; |
1da177e4 | 882 | |
8a8e674c | 883 | WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED); |
1da177e4 | 884 | |
8a8e674c JA |
885 | ioc = RQ_IOC(rq); |
886 | if (ioc && ioc->aic) { | |
887 | BUG_ON(!atomic_read(&ioc->aic->nr_queued)); | |
888 | atomic_dec(&ioc->aic->nr_queued); | |
1da177e4 LT |
889 | } |
890 | ||
891 | /* | |
8a8e674c | 892 | * Update the "next_rq" cache if we are about to remove its |
1da177e4 LT |
893 | * entry |
894 | */ | |
8a8e674c JA |
895 | if (ad->next_rq[data_dir] == rq) |
896 | ad->next_rq[data_dir] = as_find_next_rq(ad, rq); | |
1da177e4 | 897 | |
d4f2f462 | 898 | rq_fifo_clear(rq); |
8a8e674c | 899 | as_del_rq_rb(ad, rq); |
1da177e4 LT |
900 | } |
901 | ||
1da177e4 | 902 | /* |
8896f3c0 | 903 | * as_fifo_expired returns 0 if there are no expired requests on the fifo, |
1da177e4 LT |
904 | * 1 otherwise. It is ratelimited so that we only perform the check once per |
905 | * `fifo_expire' interval. Otherwise a large number of expired requests | |
906 | * would create a hopeless seekstorm. | |
907 | * | |
908 | * See as_antic_expired comment. | |
909 | */ | |
910 | static int as_fifo_expired(struct as_data *ad, int adir) | |
911 | { | |
d4f2f462 | 912 | struct request *rq; |
1da177e4 LT |
913 | long delta_jif; |
914 | ||
915 | delta_jif = jiffies - ad->last_check_fifo[adir]; | |
916 | if (unlikely(delta_jif < 0)) | |
917 | delta_jif = -delta_jif; | |
918 | if (delta_jif < ad->fifo_expire[adir]) | |
919 | return 0; | |
920 | ||
921 | ad->last_check_fifo[adir] = jiffies; | |
922 | ||
923 | if (list_empty(&ad->fifo_list[adir])) | |
924 | return 0; | |
925 | ||
d4f2f462 | 926 | rq = rq_entry_fifo(ad->fifo_list[adir].next); |
1da177e4 | 927 | |
d4f2f462 | 928 | return time_after(jiffies, rq_fifo_time(rq)); |
1da177e4 LT |
929 | } |
930 | ||
931 | /* | |
932 | * as_batch_expired returns true if the current batch has expired. A batch | |
933 | * is a set of reads or a set of writes. | |
934 | */ | |
935 | static inline int as_batch_expired(struct as_data *ad) | |
936 | { | |
937 | if (ad->changed_batch || ad->new_batch) | |
938 | return 0; | |
939 | ||
940 | if (ad->batch_data_dir == REQ_SYNC) | |
941 | /* TODO! add a check so a complete fifo gets written? */ | |
942 | return time_after(jiffies, ad->current_batch_expires); | |
943 | ||
944 | return time_after(jiffies, ad->current_batch_expires) | |
945 | || ad->current_write_count == 0; | |
946 | } | |
947 | ||
948 | /* | |
949 | * move an entry to dispatch queue | |
950 | */ | |
8a8e674c | 951 | static void as_move_to_dispatch(struct as_data *ad, struct request *rq) |
1da177e4 | 952 | { |
9e2585a8 | 953 | const int data_dir = rq_is_sync(rq); |
1da177e4 | 954 | |
e37f346e | 955 | BUG_ON(RB_EMPTY_NODE(&rq->rb_node)); |
1da177e4 LT |
956 | |
957 | as_antic_stop(ad); | |
958 | ad->antic_status = ANTIC_OFF; | |
959 | ||
960 | /* | |
961 | * This has to be set in order to be correctly updated by | |
8a8e674c | 962 | * as_find_next_rq |
1da177e4 LT |
963 | */ |
964 | ad->last_sector[data_dir] = rq->sector + rq->nr_sectors; | |
965 | ||
966 | if (data_dir == REQ_SYNC) { | |
8a8e674c | 967 | struct io_context *ioc = RQ_IOC(rq); |
1da177e4 | 968 | /* In case we have to anticipate after this */ |
8a8e674c | 969 | copy_io_context(&ad->io_context, &ioc); |
1da177e4 LT |
970 | } else { |
971 | if (ad->io_context) { | |
972 | put_io_context(ad->io_context); | |
973 | ad->io_context = NULL; | |
974 | } | |
975 | ||
976 | if (ad->current_write_count != 0) | |
977 | ad->current_write_count--; | |
978 | } | |
979 | ad->ioc_finished = 0; | |
980 | ||
8a8e674c | 981 | ad->next_rq[data_dir] = as_find_next_rq(ad, rq); |
1da177e4 LT |
982 | |
983 | /* | |
984 | * take it off the sort and fifo list, add to dispatch queue | |
985 | */ | |
1da177e4 | 986 | as_remove_queued_request(ad->q, rq); |
8a8e674c | 987 | WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED); |
1da177e4 | 988 | |
b4878f24 JA |
989 | elv_dispatch_sort(ad->q, rq); |
990 | ||
8a8e674c JA |
991 | RQ_SET_STATE(rq, AS_RQ_DISPATCHED); |
992 | if (RQ_IOC(rq) && RQ_IOC(rq)->aic) | |
993 | atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched); | |
1da177e4 LT |
994 | ad->nr_dispatched++; |
995 | } | |
996 | ||
997 | /* | |
998 | * as_dispatch_request selects the best request according to | |
999 | * read/write expire, batch expire, etc, and moves it to the dispatch | |
1000 | * queue. Returns 1 if a request was found, 0 otherwise. | |
1001 | */ | |
165125e1 | 1002 | static int as_dispatch_request(struct request_queue *q, int force) |
1da177e4 | 1003 | { |
b4878f24 | 1004 | struct as_data *ad = q->elevator->elevator_data; |
1da177e4 LT |
1005 | const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]); |
1006 | const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]); | |
8a8e674c | 1007 | struct request *rq; |
1da177e4 | 1008 | |
b4878f24 JA |
1009 | if (unlikely(force)) { |
1010 | /* | |
1011 | * Forced dispatch, accounting is useless. Reset | |
1012 | * accounting states and dump fifo_lists. Note that | |
1013 | * batch_data_dir is reset to REQ_SYNC to avoid | |
1014 | * screwing write batch accounting as write batch | |
1015 | * accounting occurs on W->R transition. | |
1016 | */ | |
1017 | int dispatched = 0; | |
1018 | ||
1019 | ad->batch_data_dir = REQ_SYNC; | |
1020 | ad->changed_batch = 0; | |
1021 | ad->new_batch = 0; | |
1022 | ||
8a8e674c JA |
1023 | while (ad->next_rq[REQ_SYNC]) { |
1024 | as_move_to_dispatch(ad, ad->next_rq[REQ_SYNC]); | |
b4878f24 JA |
1025 | dispatched++; |
1026 | } | |
1027 | ad->last_check_fifo[REQ_SYNC] = jiffies; | |
1028 | ||
8a8e674c JA |
1029 | while (ad->next_rq[REQ_ASYNC]) { |
1030 | as_move_to_dispatch(ad, ad->next_rq[REQ_ASYNC]); | |
b4878f24 JA |
1031 | dispatched++; |
1032 | } | |
1033 | ad->last_check_fifo[REQ_ASYNC] = jiffies; | |
1034 | ||
1035 | return dispatched; | |
1036 | } | |
1037 | ||
1da177e4 LT |
1038 | /* Signal that the write batch was uncontended, so we can't time it */ |
1039 | if (ad->batch_data_dir == REQ_ASYNC && !reads) { | |
1040 | if (ad->current_write_count == 0 || !writes) | |
1041 | ad->write_batch_idled = 1; | |
1042 | } | |
1043 | ||
1044 | if (!(reads || writes) | |
1045 | || ad->antic_status == ANTIC_WAIT_REQ | |
1046 | || ad->antic_status == ANTIC_WAIT_NEXT | |
1047 | || ad->changed_batch) | |
1048 | return 0; | |
1049 | ||
f5b3db00 | 1050 | if (!(reads && writes && as_batch_expired(ad))) { |
1da177e4 LT |
1051 | /* |
1052 | * batch is still running or no reads or no writes | |
1053 | */ | |
8a8e674c | 1054 | rq = ad->next_rq[ad->batch_data_dir]; |
1da177e4 LT |
1055 | |
1056 | if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) { | |
1057 | if (as_fifo_expired(ad, REQ_SYNC)) | |
1058 | goto fifo_expired; | |
1059 | ||
8a8e674c | 1060 | if (as_can_anticipate(ad, rq)) { |
1da177e4 LT |
1061 | as_antic_waitreq(ad); |
1062 | return 0; | |
1063 | } | |
1064 | } | |
1065 | ||
8a8e674c | 1066 | if (rq) { |
1da177e4 LT |
1067 | /* we have a "next request" */ |
1068 | if (reads && !writes) | |
1069 | ad->current_batch_expires = | |
1070 | jiffies + ad->batch_expire[REQ_SYNC]; | |
1071 | goto dispatch_request; | |
1072 | } | |
1073 | } | |
1074 | ||
1075 | /* | |
1076 | * at this point we are not running a batch. select the appropriate | |
1077 | * data direction (read / write) | |
1078 | */ | |
1079 | ||
1080 | if (reads) { | |
dd67d051 | 1081 | BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_SYNC])); |
1da177e4 LT |
1082 | |
1083 | if (writes && ad->batch_data_dir == REQ_SYNC) | |
1084 | /* | |
1085 | * Last batch was a read, switch to writes | |
1086 | */ | |
1087 | goto dispatch_writes; | |
1088 | ||
1089 | if (ad->batch_data_dir == REQ_ASYNC) { | |
1090 | WARN_ON(ad->new_batch); | |
1091 | ad->changed_batch = 1; | |
1092 | } | |
1093 | ad->batch_data_dir = REQ_SYNC; | |
8a8e674c | 1094 | rq = rq_entry_fifo(ad->fifo_list[REQ_SYNC].next); |
1da177e4 LT |
1095 | ad->last_check_fifo[ad->batch_data_dir] = jiffies; |
1096 | goto dispatch_request; | |
1097 | } | |
1098 | ||
1099 | /* | |
1100 | * the last batch was a read | |
1101 | */ | |
1102 | ||
1103 | if (writes) { | |
1104 | dispatch_writes: | |
dd67d051 | 1105 | BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_ASYNC])); |
1da177e4 LT |
1106 | |
1107 | if (ad->batch_data_dir == REQ_SYNC) { | |
1108 | ad->changed_batch = 1; | |
1109 | ||
1110 | /* | |
1111 | * new_batch might be 1 when the queue runs out of | |
1112 | * reads. A subsequent submission of a write might | |
1113 | * cause a change of batch before the read is finished. | |
1114 | */ | |
1115 | ad->new_batch = 0; | |
1116 | } | |
1117 | ad->batch_data_dir = REQ_ASYNC; | |
1118 | ad->current_write_count = ad->write_batch_count; | |
1119 | ad->write_batch_idled = 0; | |
49565124 AC |
1120 | rq = rq_entry_fifo(ad->fifo_list[REQ_ASYNC].next); |
1121 | ad->last_check_fifo[REQ_ASYNC] = jiffies; | |
1da177e4 LT |
1122 | goto dispatch_request; |
1123 | } | |
1124 | ||
1125 | BUG(); | |
1126 | return 0; | |
1127 | ||
1128 | dispatch_request: | |
1129 | /* | |
1130 | * If a request has expired, service it. | |
1131 | */ | |
1132 | ||
1133 | if (as_fifo_expired(ad, ad->batch_data_dir)) { | |
1134 | fifo_expired: | |
8a8e674c | 1135 | rq = rq_entry_fifo(ad->fifo_list[ad->batch_data_dir].next); |
1da177e4 LT |
1136 | } |
1137 | ||
1138 | if (ad->changed_batch) { | |
1139 | WARN_ON(ad->new_batch); | |
1140 | ||
1141 | if (ad->nr_dispatched) | |
1142 | return 0; | |
1143 | ||
1144 | if (ad->batch_data_dir == REQ_ASYNC) | |
1145 | ad->current_batch_expires = jiffies + | |
1146 | ad->batch_expire[REQ_ASYNC]; | |
1147 | else | |
1148 | ad->new_batch = 1; | |
1149 | ||
1150 | ad->changed_batch = 0; | |
1151 | } | |
1152 | ||
1153 | /* | |
8a8e674c | 1154 | * rq is the selected appropriate request. |
1da177e4 | 1155 | */ |
8a8e674c | 1156 | as_move_to_dispatch(ad, rq); |
1da177e4 LT |
1157 | |
1158 | return 1; | |
1159 | } | |
1160 | ||
1da177e4 | 1161 | /* |
8a8e674c | 1162 | * add rq to rbtree and fifo |
1da177e4 | 1163 | */ |
165125e1 | 1164 | static void as_add_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1165 | { |
b4878f24 | 1166 | struct as_data *ad = q->elevator->elevator_data; |
1da177e4 LT |
1167 | int data_dir; |
1168 | ||
8a8e674c | 1169 | RQ_SET_STATE(rq, AS_RQ_NEW); |
b4878f24 | 1170 | |
9e2585a8 | 1171 | data_dir = rq_is_sync(rq); |
1da177e4 | 1172 | |
b5deef90 | 1173 | rq->elevator_private = as_get_io_context(q->node); |
1da177e4 | 1174 | |
8a8e674c JA |
1175 | if (RQ_IOC(rq)) { |
1176 | as_update_iohist(ad, RQ_IOC(rq)->aic, rq); | |
1177 | atomic_inc(&RQ_IOC(rq)->aic->nr_queued); | |
1da177e4 LT |
1178 | } |
1179 | ||
8a8e674c | 1180 | as_add_rq_rb(ad, rq); |
1da177e4 | 1181 | |
ef9be1d3 | 1182 | /* |
8896f3c0 | 1183 | * set expire time and add to fifo list |
ef9be1d3 | 1184 | */ |
d4f2f462 JA |
1185 | rq_set_fifo_time(rq, jiffies + ad->fifo_expire[data_dir]); |
1186 | list_add_tail(&rq->queuelist, &ad->fifo_list[data_dir]); | |
1da177e4 | 1187 | |
8a8e674c JA |
1188 | as_update_rq(ad, rq); /* keep state machine up to date */ |
1189 | RQ_SET_STATE(rq, AS_RQ_QUEUED); | |
1da177e4 LT |
1190 | } |
1191 | ||
165125e1 | 1192 | static void as_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1193 | { |
8a8e674c JA |
1194 | WARN_ON(RQ_STATE(rq) != AS_RQ_DISPATCHED); |
1195 | RQ_SET_STATE(rq, AS_RQ_REMOVED); | |
1196 | if (RQ_IOC(rq) && RQ_IOC(rq)->aic) | |
1197 | atomic_dec(&RQ_IOC(rq)->aic->nr_dispatched); | |
1da177e4 LT |
1198 | } |
1199 | ||
165125e1 | 1200 | static void as_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1201 | { |
8a8e674c JA |
1202 | WARN_ON(RQ_STATE(rq) != AS_RQ_REMOVED); |
1203 | RQ_SET_STATE(rq, AS_RQ_DISPATCHED); | |
1204 | if (RQ_IOC(rq) && RQ_IOC(rq)->aic) | |
1205 | atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched); | |
1da177e4 LT |
1206 | } |
1207 | ||
1208 | /* | |
1209 | * as_queue_empty tells us if there are requests left in the device. It may | |
1210 | * not be the case that a driver can get the next request even if the queue | |
1211 | * is not empty - it is used in the block layer to check for plugging and | |
1212 | * merging opportunities | |
1213 | */ | |
165125e1 | 1214 | static int as_queue_empty(struct request_queue *q) |
1da177e4 LT |
1215 | { |
1216 | struct as_data *ad = q->elevator->elevator_data; | |
1217 | ||
b4878f24 JA |
1218 | return list_empty(&ad->fifo_list[REQ_ASYNC]) |
1219 | && list_empty(&ad->fifo_list[REQ_SYNC]); | |
1da177e4 LT |
1220 | } |
1221 | ||
1da177e4 | 1222 | static int |
165125e1 | 1223 | as_merge(struct request_queue *q, struct request **req, struct bio *bio) |
1da177e4 LT |
1224 | { |
1225 | struct as_data *ad = q->elevator->elevator_data; | |
1226 | sector_t rb_key = bio->bi_sector + bio_sectors(bio); | |
1227 | struct request *__rq; | |
1da177e4 LT |
1228 | |
1229 | /* | |
1230 | * check for front merge | |
1231 | */ | |
e37f346e | 1232 | __rq = elv_rb_find(&ad->sort_list[bio_data_dir(bio)], rb_key); |
9817064b JA |
1233 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
1234 | *req = __rq; | |
1235 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
1236 | } |
1237 | ||
1238 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
1239 | } |
1240 | ||
165125e1 JA |
1241 | static void as_merged_request(struct request_queue *q, struct request *req, |
1242 | int type) | |
1da177e4 LT |
1243 | { |
1244 | struct as_data *ad = q->elevator->elevator_data; | |
1da177e4 | 1245 | |
1da177e4 LT |
1246 | /* |
1247 | * if the merge was a front merge, we need to reposition request | |
1248 | */ | |
e37f346e | 1249 | if (type == ELEVATOR_FRONT_MERGE) { |
8a8e674c JA |
1250 | as_del_rq_rb(ad, req); |
1251 | as_add_rq_rb(ad, req); | |
1da177e4 LT |
1252 | /* |
1253 | * Note! At this stage of this and the next function, our next | |
1254 | * request may not be optimal - eg the request may have "grown" | |
1255 | * behind the disk head. We currently don't bother adjusting. | |
1256 | */ | |
1257 | } | |
1da177e4 LT |
1258 | } |
1259 | ||
165125e1 | 1260 | static void as_merged_requests(struct request_queue *q, struct request *req, |
f5b3db00 | 1261 | struct request *next) |
1da177e4 | 1262 | { |
1da177e4 | 1263 | /* |
8a8e674c JA |
1264 | * if next expires before rq, assign its expire time to arq |
1265 | * and move into next position (next will be deleted) in fifo | |
1da177e4 | 1266 | */ |
d4f2f462 JA |
1267 | if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) { |
1268 | if (time_before(rq_fifo_time(next), rq_fifo_time(req))) { | |
8a8e674c JA |
1269 | struct io_context *rioc = RQ_IOC(req); |
1270 | struct io_context *nioc = RQ_IOC(next); | |
1271 | ||
d4f2f462 JA |
1272 | list_move(&req->queuelist, &next->queuelist); |
1273 | rq_set_fifo_time(req, rq_fifo_time(next)); | |
1da177e4 LT |
1274 | /* |
1275 | * Don't copy here but swap, because when anext is | |
1276 | * removed below, it must contain the unused context | |
1277 | */ | |
521f3bbd | 1278 | double_spin_lock(&rioc->lock, &nioc->lock, rioc < nioc); |
8a8e674c | 1279 | swap_io_context(&rioc, &nioc); |
521f3bbd | 1280 | double_spin_unlock(&rioc->lock, &nioc->lock, rioc < nioc); |
1da177e4 LT |
1281 | } |
1282 | } | |
1283 | ||
1da177e4 LT |
1284 | /* |
1285 | * kill knowledge of next, this one is a goner | |
1286 | */ | |
1287 | as_remove_queued_request(q, next); | |
8a8e674c | 1288 | as_put_io_context(next); |
1da177e4 | 1289 | |
8a8e674c | 1290 | RQ_SET_STATE(next, AS_RQ_MERGED); |
1da177e4 LT |
1291 | } |
1292 | ||
1293 | /* | |
1294 | * This is executed in a "deferred" process context, by kblockd. It calls the | |
1295 | * driver's request_fn so the driver can submit that request. | |
1296 | * | |
1297 | * IMPORTANT! This guy will reenter the elevator, so set up all queue global | |
1298 | * state before calling, and don't rely on any state over calls. | |
1299 | * | |
1300 | * FIXME! dispatch queue is not a queue at all! | |
1301 | */ | |
65f27f38 | 1302 | static void as_work_handler(struct work_struct *work) |
1da177e4 | 1303 | { |
65f27f38 DH |
1304 | struct as_data *ad = container_of(work, struct as_data, antic_work); |
1305 | struct request_queue *q = ad->q; | |
1da177e4 LT |
1306 | unsigned long flags; |
1307 | ||
1308 | spin_lock_irqsave(q->queue_lock, flags); | |
dc72ef4a | 1309 | blk_start_queueing(q); |
1da177e4 LT |
1310 | spin_unlock_irqrestore(q->queue_lock, flags); |
1311 | } | |
1312 | ||
165125e1 | 1313 | static int as_may_queue(struct request_queue *q, int rw) |
1da177e4 LT |
1314 | { |
1315 | int ret = ELV_MQUEUE_MAY; | |
1316 | struct as_data *ad = q->elevator->elevator_data; | |
1317 | struct io_context *ioc; | |
1318 | if (ad->antic_status == ANTIC_WAIT_REQ || | |
1319 | ad->antic_status == ANTIC_WAIT_NEXT) { | |
b5deef90 | 1320 | ioc = as_get_io_context(q->node); |
1da177e4 LT |
1321 | if (ad->io_context == ioc) |
1322 | ret = ELV_MQUEUE_MUST; | |
1323 | put_io_context(ioc); | |
1324 | } | |
1325 | ||
1326 | return ret; | |
1327 | } | |
1328 | ||
1329 | static void as_exit_queue(elevator_t *e) | |
1330 | { | |
1331 | struct as_data *ad = e->elevator_data; | |
1332 | ||
1333 | del_timer_sync(&ad->antic_timer); | |
19a75d83 | 1334 | kblockd_flush_work(&ad->antic_work); |
1da177e4 LT |
1335 | |
1336 | BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC])); | |
1337 | BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC])); | |
1338 | ||
1da177e4 | 1339 | put_io_context(ad->io_context); |
1da177e4 LT |
1340 | kfree(ad); |
1341 | } | |
1342 | ||
1343 | /* | |
8a8e674c | 1344 | * initialize elevator private data (as_data). |
1da177e4 | 1345 | */ |
165125e1 | 1346 | static void *as_init_queue(struct request_queue *q) |
1da177e4 LT |
1347 | { |
1348 | struct as_data *ad; | |
1da177e4 | 1349 | |
94f6030c | 1350 | ad = kmalloc_node(sizeof(*ad), GFP_KERNEL | __GFP_ZERO, q->node); |
1da177e4 | 1351 | if (!ad) |
bc1c1169 | 1352 | return NULL; |
1da177e4 LT |
1353 | |
1354 | ad->q = q; /* Identify what queue the data belongs to */ | |
1355 | ||
1da177e4 LT |
1356 | /* anticipatory scheduling helpers */ |
1357 | ad->antic_timer.function = as_antic_timeout; | |
1358 | ad->antic_timer.data = (unsigned long)q; | |
1359 | init_timer(&ad->antic_timer); | |
65f27f38 | 1360 | INIT_WORK(&ad->antic_work, as_work_handler); |
1da177e4 | 1361 | |
1da177e4 LT |
1362 | INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]); |
1363 | INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]); | |
1364 | ad->sort_list[REQ_SYNC] = RB_ROOT; | |
1365 | ad->sort_list[REQ_ASYNC] = RB_ROOT; | |
1da177e4 LT |
1366 | ad->fifo_expire[REQ_SYNC] = default_read_expire; |
1367 | ad->fifo_expire[REQ_ASYNC] = default_write_expire; | |
1368 | ad->antic_expire = default_antic_expire; | |
1369 | ad->batch_expire[REQ_SYNC] = default_read_batch_expire; | |
1370 | ad->batch_expire[REQ_ASYNC] = default_write_batch_expire; | |
1da177e4 LT |
1371 | |
1372 | ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC]; | |
1373 | ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10; | |
1374 | if (ad->write_batch_count < 2) | |
1375 | ad->write_batch_count = 2; | |
1376 | ||
bc1c1169 | 1377 | return ad; |
1da177e4 LT |
1378 | } |
1379 | ||
1380 | /* | |
1381 | * sysfs parts below | |
1382 | */ | |
1da177e4 LT |
1383 | |
1384 | static ssize_t | |
1385 | as_var_show(unsigned int var, char *page) | |
1386 | { | |
1da177e4 LT |
1387 | return sprintf(page, "%d\n", var); |
1388 | } | |
1389 | ||
1390 | static ssize_t | |
1391 | as_var_store(unsigned long *var, const char *page, size_t count) | |
1392 | { | |
1da177e4 LT |
1393 | char *p = (char *) page; |
1394 | ||
c9b3ad67 | 1395 | *var = simple_strtoul(p, &p, 10); |
1da177e4 LT |
1396 | return count; |
1397 | } | |
1398 | ||
e572ec7e | 1399 | static ssize_t est_time_show(elevator_t *e, char *page) |
1da177e4 | 1400 | { |
3d1ab40f | 1401 | struct as_data *ad = e->elevator_data; |
1da177e4 LT |
1402 | int pos = 0; |
1403 | ||
f5b3db00 NP |
1404 | pos += sprintf(page+pos, "%lu %% exit probability\n", |
1405 | 100*ad->exit_prob/256); | |
1406 | pos += sprintf(page+pos, "%lu %% probability of exiting without a " | |
1407 | "cooperating process submitting IO\n", | |
1408 | 100*ad->exit_no_coop/256); | |
1da177e4 | 1409 | pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean); |
f5b3db00 NP |
1410 | pos += sprintf(page+pos, "%llu sectors new seek distance\n", |
1411 | (unsigned long long)ad->new_seek_mean); | |
1da177e4 LT |
1412 | |
1413 | return pos; | |
1414 | } | |
1415 | ||
1416 | #define SHOW_FUNCTION(__FUNC, __VAR) \ | |
3d1ab40f | 1417 | static ssize_t __FUNC(elevator_t *e, char *page) \ |
1da177e4 | 1418 | { \ |
3d1ab40f | 1419 | struct as_data *ad = e->elevator_data; \ |
1da177e4 LT |
1420 | return as_var_show(jiffies_to_msecs((__VAR)), (page)); \ |
1421 | } | |
e572ec7e AV |
1422 | SHOW_FUNCTION(as_read_expire_show, ad->fifo_expire[REQ_SYNC]); |
1423 | SHOW_FUNCTION(as_write_expire_show, ad->fifo_expire[REQ_ASYNC]); | |
1424 | SHOW_FUNCTION(as_antic_expire_show, ad->antic_expire); | |
1425 | SHOW_FUNCTION(as_read_batch_expire_show, ad->batch_expire[REQ_SYNC]); | |
1426 | SHOW_FUNCTION(as_write_batch_expire_show, ad->batch_expire[REQ_ASYNC]); | |
1da177e4 LT |
1427 | #undef SHOW_FUNCTION |
1428 | ||
1429 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \ | |
3d1ab40f | 1430 | static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \ |
1da177e4 | 1431 | { \ |
3d1ab40f AV |
1432 | struct as_data *ad = e->elevator_data; \ |
1433 | int ret = as_var_store(__PTR, (page), count); \ | |
1da177e4 LT |
1434 | if (*(__PTR) < (MIN)) \ |
1435 | *(__PTR) = (MIN); \ | |
1436 | else if (*(__PTR) > (MAX)) \ | |
1437 | *(__PTR) = (MAX); \ | |
1438 | *(__PTR) = msecs_to_jiffies(*(__PTR)); \ | |
1439 | return ret; \ | |
1440 | } | |
e572ec7e AV |
1441 | STORE_FUNCTION(as_read_expire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX); |
1442 | STORE_FUNCTION(as_write_expire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX); | |
1443 | STORE_FUNCTION(as_antic_expire_store, &ad->antic_expire, 0, INT_MAX); | |
1444 | STORE_FUNCTION(as_read_batch_expire_store, | |
1da177e4 | 1445 | &ad->batch_expire[REQ_SYNC], 0, INT_MAX); |
e572ec7e | 1446 | STORE_FUNCTION(as_write_batch_expire_store, |
1da177e4 LT |
1447 | &ad->batch_expire[REQ_ASYNC], 0, INT_MAX); |
1448 | #undef STORE_FUNCTION | |
1449 | ||
e572ec7e AV |
1450 | #define AS_ATTR(name) \ |
1451 | __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store) | |
1452 | ||
1453 | static struct elv_fs_entry as_attrs[] = { | |
1454 | __ATTR_RO(est_time), | |
1455 | AS_ATTR(read_expire), | |
1456 | AS_ATTR(write_expire), | |
1457 | AS_ATTR(antic_expire), | |
1458 | AS_ATTR(read_batch_expire), | |
1459 | AS_ATTR(write_batch_expire), | |
1460 | __ATTR_NULL | |
1da177e4 LT |
1461 | }; |
1462 | ||
1da177e4 LT |
1463 | static struct elevator_type iosched_as = { |
1464 | .ops = { | |
1465 | .elevator_merge_fn = as_merge, | |
1466 | .elevator_merged_fn = as_merged_request, | |
1467 | .elevator_merge_req_fn = as_merged_requests, | |
b4878f24 JA |
1468 | .elevator_dispatch_fn = as_dispatch_request, |
1469 | .elevator_add_req_fn = as_add_request, | |
1470 | .elevator_activate_req_fn = as_activate_request, | |
1da177e4 LT |
1471 | .elevator_deactivate_req_fn = as_deactivate_request, |
1472 | .elevator_queue_empty_fn = as_queue_empty, | |
1473 | .elevator_completed_req_fn = as_completed_request, | |
e37f346e JA |
1474 | .elevator_former_req_fn = elv_rb_former_request, |
1475 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
1476 | .elevator_may_queue_fn = as_may_queue, |
1477 | .elevator_init_fn = as_init_queue, | |
1478 | .elevator_exit_fn = as_exit_queue, | |
e17a9489 | 1479 | .trim = as_trim, |
1da177e4 LT |
1480 | }, |
1481 | ||
3d1ab40f | 1482 | .elevator_attrs = as_attrs, |
1da177e4 LT |
1483 | .elevator_name = "anticipatory", |
1484 | .elevator_owner = THIS_MODULE, | |
1485 | }; | |
1486 | ||
1487 | static int __init as_init(void) | |
1488 | { | |
2fdd82bd AB |
1489 | elv_register(&iosched_as); |
1490 | ||
1491 | return 0; | |
1da177e4 LT |
1492 | } |
1493 | ||
1494 | static void __exit as_exit(void) | |
1495 | { | |
6e9a4738 | 1496 | DECLARE_COMPLETION_ONSTACK(all_gone); |
1da177e4 | 1497 | elv_unregister(&iosched_as); |
334e94de | 1498 | ioc_gone = &all_gone; |
fba82272 OH |
1499 | /* ioc_gone's update must be visible before reading ioc_count */ |
1500 | smp_wmb(); | |
e4313dd4 | 1501 | if (elv_ioc_count_read(ioc_count)) |
fba82272 | 1502 | wait_for_completion(ioc_gone); |
334e94de | 1503 | synchronize_rcu(); |
1da177e4 LT |
1504 | } |
1505 | ||
1506 | module_init(as_init); | |
1507 | module_exit(as_exit); | |
1508 | ||
1509 | MODULE_AUTHOR("Nick Piggin"); | |
1510 | MODULE_LICENSE("GPL"); | |
1511 | MODULE_DESCRIPTION("anticipatory IO scheduler"); |