Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
1da177e4 LT |
2 | * Anticipatory & deadline i/o scheduler. |
3 | * | |
4 | * Copyright (C) 2002 Jens Axboe <axboe@suse.de> | |
f5b3db00 | 5 | * Nick Piggin <nickpiggin@yahoo.com.au> |
1da177e4 LT |
6 | * |
7 | */ | |
8 | #include <linux/kernel.h> | |
9 | #include <linux/fs.h> | |
10 | #include <linux/blkdev.h> | |
11 | #include <linux/elevator.h> | |
12 | #include <linux/bio.h> | |
13 | #include <linux/config.h> | |
14 | #include <linux/module.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/init.h> | |
17 | #include <linux/compiler.h> | |
18 | #include <linux/hash.h> | |
19 | #include <linux/rbtree.h> | |
20 | #include <linux/interrupt.h> | |
21 | ||
22 | #define REQ_SYNC 1 | |
23 | #define REQ_ASYNC 0 | |
24 | ||
25 | /* | |
26 | * See Documentation/block/as-iosched.txt | |
27 | */ | |
28 | ||
29 | /* | |
30 | * max time before a read is submitted. | |
31 | */ | |
32 | #define default_read_expire (HZ / 8) | |
33 | ||
34 | /* | |
35 | * ditto for writes, these limits are not hard, even | |
36 | * if the disk is capable of satisfying them. | |
37 | */ | |
38 | #define default_write_expire (HZ / 4) | |
39 | ||
40 | /* | |
41 | * read_batch_expire describes how long we will allow a stream of reads to | |
42 | * persist before looking to see whether it is time to switch over to writes. | |
43 | */ | |
44 | #define default_read_batch_expire (HZ / 2) | |
45 | ||
46 | /* | |
47 | * write_batch_expire describes how long we want a stream of writes to run for. | |
48 | * This is not a hard limit, but a target we set for the auto-tuning thingy. | |
49 | * See, the problem is: we can send a lot of writes to disk cache / TCQ in | |
50 | * a short amount of time... | |
51 | */ | |
52 | #define default_write_batch_expire (HZ / 8) | |
53 | ||
54 | /* | |
55 | * max time we may wait to anticipate a read (default around 6ms) | |
56 | */ | |
57 | #define default_antic_expire ((HZ / 150) ? HZ / 150 : 1) | |
58 | ||
59 | /* | |
60 | * Keep track of up to 20ms thinktimes. We can go as big as we like here, | |
61 | * however huge values tend to interfere and not decay fast enough. A program | |
62 | * might be in a non-io phase of operation. Waiting on user input for example, | |
63 | * or doing a lengthy computation. A small penalty can be justified there, and | |
64 | * will still catch out those processes that constantly have large thinktimes. | |
65 | */ | |
66 | #define MAX_THINKTIME (HZ/50UL) | |
67 | ||
68 | /* Bits in as_io_context.state */ | |
69 | enum as_io_states { | |
f5b3db00 | 70 | AS_TASK_RUNNING=0, /* Process has not exited */ |
1da177e4 LT |
71 | AS_TASK_IOSTARTED, /* Process has started some IO */ |
72 | AS_TASK_IORUNNING, /* Process has completed some IO */ | |
73 | }; | |
74 | ||
75 | enum anticipation_status { | |
76 | ANTIC_OFF=0, /* Not anticipating (normal operation) */ | |
77 | ANTIC_WAIT_REQ, /* The last read has not yet completed */ | |
78 | ANTIC_WAIT_NEXT, /* Currently anticipating a request vs | |
79 | last read (which has completed) */ | |
80 | ANTIC_FINISHED, /* Anticipating but have found a candidate | |
81 | * or timed out */ | |
82 | }; | |
83 | ||
84 | struct as_data { | |
85 | /* | |
86 | * run time data | |
87 | */ | |
88 | ||
89 | struct request_queue *q; /* the "owner" queue */ | |
90 | ||
91 | /* | |
92 | * requests (as_rq s) are present on both sort_list and fifo_list | |
93 | */ | |
94 | struct rb_root sort_list[2]; | |
95 | struct list_head fifo_list[2]; | |
96 | ||
97 | struct as_rq *next_arq[2]; /* next in sort order */ | |
98 | sector_t last_sector[2]; /* last REQ_SYNC & REQ_ASYNC sectors */ | |
bae386f7 | 99 | struct hlist_head *hash; /* request hash */ |
1da177e4 LT |
100 | |
101 | unsigned long exit_prob; /* probability a task will exit while | |
102 | being waited on */ | |
f5b3db00 NP |
103 | unsigned long exit_no_coop; /* probablility an exited task will |
104 | not be part of a later cooperating | |
105 | request */ | |
1da177e4 LT |
106 | unsigned long new_ttime_total; /* mean thinktime on new proc */ |
107 | unsigned long new_ttime_mean; | |
108 | u64 new_seek_total; /* mean seek on new proc */ | |
109 | sector_t new_seek_mean; | |
110 | ||
111 | unsigned long current_batch_expires; | |
112 | unsigned long last_check_fifo[2]; | |
113 | int changed_batch; /* 1: waiting for old batch to end */ | |
114 | int new_batch; /* 1: waiting on first read complete */ | |
115 | int batch_data_dir; /* current batch REQ_SYNC / REQ_ASYNC */ | |
116 | int write_batch_count; /* max # of reqs in a write batch */ | |
117 | int current_write_count; /* how many requests left this batch */ | |
118 | int write_batch_idled; /* has the write batch gone idle? */ | |
119 | mempool_t *arq_pool; | |
120 | ||
121 | enum anticipation_status antic_status; | |
122 | unsigned long antic_start; /* jiffies: when it started */ | |
123 | struct timer_list antic_timer; /* anticipatory scheduling timer */ | |
124 | struct work_struct antic_work; /* Deferred unplugging */ | |
125 | struct io_context *io_context; /* Identify the expected process */ | |
126 | int ioc_finished; /* IO associated with io_context is finished */ | |
127 | int nr_dispatched; | |
128 | ||
129 | /* | |
130 | * settings that change how the i/o scheduler behaves | |
131 | */ | |
132 | unsigned long fifo_expire[2]; | |
133 | unsigned long batch_expire[2]; | |
134 | unsigned long antic_expire; | |
135 | }; | |
136 | ||
137 | #define list_entry_fifo(ptr) list_entry((ptr), struct as_rq, fifo) | |
138 | ||
139 | /* | |
140 | * per-request data. | |
141 | */ | |
142 | enum arq_state { | |
143 | AS_RQ_NEW=0, /* New - not referenced and not on any lists */ | |
144 | AS_RQ_QUEUED, /* In the request queue. It belongs to the | |
145 | scheduler */ | |
146 | AS_RQ_DISPATCHED, /* On the dispatch list. It belongs to the | |
147 | driver now */ | |
148 | AS_RQ_PRESCHED, /* Debug poisoning for requests being used */ | |
149 | AS_RQ_REMOVED, | |
150 | AS_RQ_MERGED, | |
151 | AS_RQ_POSTSCHED, /* when they shouldn't be */ | |
152 | }; | |
153 | ||
154 | struct as_rq { | |
155 | /* | |
156 | * rbtree index, key is the starting offset | |
157 | */ | |
158 | struct rb_node rb_node; | |
159 | sector_t rb_key; | |
160 | ||
161 | struct request *request; | |
162 | ||
163 | struct io_context *io_context; /* The submitting task */ | |
164 | ||
165 | /* | |
166 | * request hash, key is the ending offset (for back merge lookup) | |
167 | */ | |
bae386f7 | 168 | struct hlist_node hash; |
1da177e4 LT |
169 | |
170 | /* | |
171 | * expire fifo | |
172 | */ | |
173 | struct list_head fifo; | |
174 | unsigned long expires; | |
175 | ||
176 | unsigned int is_sync; | |
177 | enum arq_state state; | |
178 | }; | |
179 | ||
180 | #define RQ_DATA(rq) ((struct as_rq *) (rq)->elevator_private) | |
181 | ||
182 | static kmem_cache_t *arq_pool; | |
183 | ||
334e94de AV |
184 | static atomic_t ioc_count = ATOMIC_INIT(0); |
185 | static struct completion *ioc_gone; | |
186 | ||
ef9be1d3 TH |
187 | static void as_move_to_dispatch(struct as_data *ad, struct as_rq *arq); |
188 | static void as_antic_stop(struct as_data *ad); | |
189 | ||
1da177e4 LT |
190 | /* |
191 | * IO Context helper functions | |
192 | */ | |
193 | ||
194 | /* Called to deallocate the as_io_context */ | |
195 | static void free_as_io_context(struct as_io_context *aic) | |
196 | { | |
197 | kfree(aic); | |
334e94de AV |
198 | if (atomic_dec_and_test(&ioc_count) && ioc_gone) |
199 | complete(ioc_gone); | |
1da177e4 LT |
200 | } |
201 | ||
e17a9489 AV |
202 | static void as_trim(struct io_context *ioc) |
203 | { | |
334e94de AV |
204 | if (ioc->aic) |
205 | free_as_io_context(ioc->aic); | |
e17a9489 AV |
206 | ioc->aic = NULL; |
207 | } | |
208 | ||
1da177e4 LT |
209 | /* Called when the task exits */ |
210 | static void exit_as_io_context(struct as_io_context *aic) | |
211 | { | |
212 | WARN_ON(!test_bit(AS_TASK_RUNNING, &aic->state)); | |
213 | clear_bit(AS_TASK_RUNNING, &aic->state); | |
214 | } | |
215 | ||
216 | static struct as_io_context *alloc_as_io_context(void) | |
217 | { | |
218 | struct as_io_context *ret; | |
219 | ||
220 | ret = kmalloc(sizeof(*ret), GFP_ATOMIC); | |
221 | if (ret) { | |
222 | ret->dtor = free_as_io_context; | |
223 | ret->exit = exit_as_io_context; | |
224 | ret->state = 1 << AS_TASK_RUNNING; | |
225 | atomic_set(&ret->nr_queued, 0); | |
226 | atomic_set(&ret->nr_dispatched, 0); | |
227 | spin_lock_init(&ret->lock); | |
228 | ret->ttime_total = 0; | |
229 | ret->ttime_samples = 0; | |
230 | ret->ttime_mean = 0; | |
231 | ret->seek_total = 0; | |
232 | ret->seek_samples = 0; | |
233 | ret->seek_mean = 0; | |
334e94de | 234 | atomic_inc(&ioc_count); |
1da177e4 LT |
235 | } |
236 | ||
237 | return ret; | |
238 | } | |
239 | ||
240 | /* | |
241 | * If the current task has no AS IO context then create one and initialise it. | |
242 | * Then take a ref on the task's io context and return it. | |
243 | */ | |
244 | static struct io_context *as_get_io_context(void) | |
245 | { | |
246 | struct io_context *ioc = get_io_context(GFP_ATOMIC); | |
247 | if (ioc && !ioc->aic) { | |
248 | ioc->aic = alloc_as_io_context(); | |
249 | if (!ioc->aic) { | |
250 | put_io_context(ioc); | |
251 | ioc = NULL; | |
252 | } | |
253 | } | |
254 | return ioc; | |
255 | } | |
256 | ||
b4878f24 JA |
257 | static void as_put_io_context(struct as_rq *arq) |
258 | { | |
259 | struct as_io_context *aic; | |
260 | ||
261 | if (unlikely(!arq->io_context)) | |
262 | return; | |
263 | ||
264 | aic = arq->io_context->aic; | |
265 | ||
266 | if (arq->is_sync == REQ_SYNC && aic) { | |
267 | spin_lock(&aic->lock); | |
268 | set_bit(AS_TASK_IORUNNING, &aic->state); | |
269 | aic->last_end_request = jiffies; | |
270 | spin_unlock(&aic->lock); | |
271 | } | |
272 | ||
273 | put_io_context(arq->io_context); | |
274 | } | |
275 | ||
1da177e4 LT |
276 | /* |
277 | * the back merge hash support functions | |
278 | */ | |
279 | static const int as_hash_shift = 6; | |
280 | #define AS_HASH_BLOCK(sec) ((sec) >> 3) | |
281 | #define AS_HASH_FN(sec) (hash_long(AS_HASH_BLOCK((sec)), as_hash_shift)) | |
282 | #define AS_HASH_ENTRIES (1 << as_hash_shift) | |
283 | #define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors) | |
1da177e4 LT |
284 | |
285 | static inline void __as_del_arq_hash(struct as_rq *arq) | |
286 | { | |
bae386f7 | 287 | hlist_del_init(&arq->hash); |
1da177e4 LT |
288 | } |
289 | ||
290 | static inline void as_del_arq_hash(struct as_rq *arq) | |
291 | { | |
bae386f7 | 292 | if (!hlist_unhashed(&arq->hash)) |
1da177e4 LT |
293 | __as_del_arq_hash(arq); |
294 | } | |
295 | ||
1da177e4 LT |
296 | static void as_add_arq_hash(struct as_data *ad, struct as_rq *arq) |
297 | { | |
298 | struct request *rq = arq->request; | |
299 | ||
bae386f7 | 300 | BUG_ON(!hlist_unhashed(&arq->hash)); |
1da177e4 | 301 | |
bae386f7 | 302 | hlist_add_head(&arq->hash, &ad->hash[AS_HASH_FN(rq_hash_key(rq))]); |
1da177e4 LT |
303 | } |
304 | ||
305 | /* | |
306 | * move hot entry to front of chain | |
307 | */ | |
308 | static inline void as_hot_arq_hash(struct as_data *ad, struct as_rq *arq) | |
309 | { | |
310 | struct request *rq = arq->request; | |
bae386f7 | 311 | struct hlist_head *head = &ad->hash[AS_HASH_FN(rq_hash_key(rq))]; |
1da177e4 | 312 | |
bae386f7 | 313 | if (hlist_unhashed(&arq->hash)) { |
1da177e4 LT |
314 | WARN_ON(1); |
315 | return; | |
316 | } | |
317 | ||
bae386f7 AM |
318 | if (&arq->hash != head->first) { |
319 | hlist_del(&arq->hash); | |
320 | hlist_add_head(&arq->hash, head); | |
1da177e4 LT |
321 | } |
322 | } | |
323 | ||
324 | static struct request *as_find_arq_hash(struct as_data *ad, sector_t offset) | |
325 | { | |
bae386f7 AM |
326 | struct hlist_head *hash_list = &ad->hash[AS_HASH_FN(offset)]; |
327 | struct hlist_node *entry, *next; | |
328 | struct as_rq *arq; | |
1da177e4 | 329 | |
bae386f7 | 330 | hlist_for_each_entry_safe(arq, entry, next, hash_list, hash) { |
1da177e4 LT |
331 | struct request *__rq = arq->request; |
332 | ||
bae386f7 | 333 | BUG_ON(hlist_unhashed(&arq->hash)); |
1da177e4 LT |
334 | |
335 | if (!rq_mergeable(__rq)) { | |
98b11471 | 336 | as_del_arq_hash(arq); |
1da177e4 LT |
337 | continue; |
338 | } | |
339 | ||
340 | if (rq_hash_key(__rq) == offset) | |
341 | return __rq; | |
342 | } | |
343 | ||
344 | return NULL; | |
345 | } | |
346 | ||
347 | /* | |
348 | * rb tree support functions | |
349 | */ | |
1da177e4 | 350 | #define RB_EMPTY(root) ((root)->rb_node == NULL) |
3db3a445 DW |
351 | #define ON_RB(node) (rb_parent(node) != node) |
352 | #define RB_CLEAR(node) (rb_set_parent(node, node)) | |
1da177e4 LT |
353 | #define rb_entry_arq(node) rb_entry((node), struct as_rq, rb_node) |
354 | #define ARQ_RB_ROOT(ad, arq) (&(ad)->sort_list[(arq)->is_sync]) | |
355 | #define rq_rb_key(rq) (rq)->sector | |
356 | ||
357 | /* | |
358 | * as_find_first_arq finds the first (lowest sector numbered) request | |
359 | * for the specified data_dir. Used to sweep back to the start of the disk | |
360 | * (1-way elevator) after we process the last (highest sector) request. | |
361 | */ | |
362 | static struct as_rq *as_find_first_arq(struct as_data *ad, int data_dir) | |
363 | { | |
364 | struct rb_node *n = ad->sort_list[data_dir].rb_node; | |
365 | ||
366 | if (n == NULL) | |
367 | return NULL; | |
368 | ||
369 | for (;;) { | |
370 | if (n->rb_left == NULL) | |
371 | return rb_entry_arq(n); | |
372 | ||
373 | n = n->rb_left; | |
374 | } | |
375 | } | |
376 | ||
377 | /* | |
378 | * Add the request to the rb tree if it is unique. If there is an alias (an | |
379 | * existing request against the same sector), which can happen when using | |
380 | * direct IO, then return the alias. | |
381 | */ | |
ef9be1d3 | 382 | static struct as_rq *__as_add_arq_rb(struct as_data *ad, struct as_rq *arq) |
1da177e4 LT |
383 | { |
384 | struct rb_node **p = &ARQ_RB_ROOT(ad, arq)->rb_node; | |
385 | struct rb_node *parent = NULL; | |
386 | struct as_rq *__arq; | |
387 | struct request *rq = arq->request; | |
388 | ||
389 | arq->rb_key = rq_rb_key(rq); | |
390 | ||
391 | while (*p) { | |
392 | parent = *p; | |
393 | __arq = rb_entry_arq(parent); | |
394 | ||
395 | if (arq->rb_key < __arq->rb_key) | |
396 | p = &(*p)->rb_left; | |
397 | else if (arq->rb_key > __arq->rb_key) | |
398 | p = &(*p)->rb_right; | |
399 | else | |
400 | return __arq; | |
401 | } | |
402 | ||
403 | rb_link_node(&arq->rb_node, parent, p); | |
404 | rb_insert_color(&arq->rb_node, ARQ_RB_ROOT(ad, arq)); | |
405 | ||
406 | return NULL; | |
407 | } | |
408 | ||
ef9be1d3 TH |
409 | static void as_add_arq_rb(struct as_data *ad, struct as_rq *arq) |
410 | { | |
411 | struct as_rq *alias; | |
412 | ||
413 | while ((unlikely(alias = __as_add_arq_rb(ad, arq)))) { | |
414 | as_move_to_dispatch(ad, alias); | |
415 | as_antic_stop(ad); | |
416 | } | |
417 | } | |
418 | ||
1da177e4 LT |
419 | static inline void as_del_arq_rb(struct as_data *ad, struct as_rq *arq) |
420 | { | |
421 | if (!ON_RB(&arq->rb_node)) { | |
422 | WARN_ON(1); | |
423 | return; | |
424 | } | |
425 | ||
426 | rb_erase(&arq->rb_node, ARQ_RB_ROOT(ad, arq)); | |
427 | RB_CLEAR(&arq->rb_node); | |
428 | } | |
429 | ||
430 | static struct request * | |
431 | as_find_arq_rb(struct as_data *ad, sector_t sector, int data_dir) | |
432 | { | |
433 | struct rb_node *n = ad->sort_list[data_dir].rb_node; | |
434 | struct as_rq *arq; | |
435 | ||
436 | while (n) { | |
437 | arq = rb_entry_arq(n); | |
438 | ||
439 | if (sector < arq->rb_key) | |
440 | n = n->rb_left; | |
441 | else if (sector > arq->rb_key) | |
442 | n = n->rb_right; | |
443 | else | |
444 | return arq->request; | |
445 | } | |
446 | ||
447 | return NULL; | |
448 | } | |
449 | ||
450 | /* | |
451 | * IO Scheduler proper | |
452 | */ | |
453 | ||
454 | #define MAXBACK (1024 * 1024) /* | |
455 | * Maximum distance the disk will go backward | |
456 | * for a request. | |
457 | */ | |
458 | ||
459 | #define BACK_PENALTY 2 | |
460 | ||
461 | /* | |
462 | * as_choose_req selects the preferred one of two requests of the same data_dir | |
463 | * ignoring time - eg. timeouts, which is the job of as_dispatch_request | |
464 | */ | |
465 | static struct as_rq * | |
466 | as_choose_req(struct as_data *ad, struct as_rq *arq1, struct as_rq *arq2) | |
467 | { | |
468 | int data_dir; | |
469 | sector_t last, s1, s2, d1, d2; | |
470 | int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */ | |
471 | const sector_t maxback = MAXBACK; | |
472 | ||
473 | if (arq1 == NULL || arq1 == arq2) | |
474 | return arq2; | |
475 | if (arq2 == NULL) | |
476 | return arq1; | |
477 | ||
478 | data_dir = arq1->is_sync; | |
479 | ||
480 | last = ad->last_sector[data_dir]; | |
481 | s1 = arq1->request->sector; | |
482 | s2 = arq2->request->sector; | |
483 | ||
484 | BUG_ON(data_dir != arq2->is_sync); | |
485 | ||
486 | /* | |
487 | * Strict one way elevator _except_ in the case where we allow | |
488 | * short backward seeks which are biased as twice the cost of a | |
489 | * similar forward seek. | |
490 | */ | |
491 | if (s1 >= last) | |
492 | d1 = s1 - last; | |
493 | else if (s1+maxback >= last) | |
494 | d1 = (last - s1)*BACK_PENALTY; | |
495 | else { | |
496 | r1_wrap = 1; | |
497 | d1 = 0; /* shut up, gcc */ | |
498 | } | |
499 | ||
500 | if (s2 >= last) | |
501 | d2 = s2 - last; | |
502 | else if (s2+maxback >= last) | |
503 | d2 = (last - s2)*BACK_PENALTY; | |
504 | else { | |
505 | r2_wrap = 1; | |
506 | d2 = 0; | |
507 | } | |
508 | ||
509 | /* Found required data */ | |
510 | if (!r1_wrap && r2_wrap) | |
511 | return arq1; | |
512 | else if (!r2_wrap && r1_wrap) | |
513 | return arq2; | |
514 | else if (r1_wrap && r2_wrap) { | |
515 | /* both behind the head */ | |
516 | if (s1 <= s2) | |
517 | return arq1; | |
518 | else | |
519 | return arq2; | |
520 | } | |
521 | ||
522 | /* Both requests in front of the head */ | |
523 | if (d1 < d2) | |
524 | return arq1; | |
525 | else if (d2 < d1) | |
526 | return arq2; | |
527 | else { | |
528 | if (s1 >= s2) | |
529 | return arq1; | |
530 | else | |
531 | return arq2; | |
532 | } | |
533 | } | |
534 | ||
535 | /* | |
536 | * as_find_next_arq finds the next request after @prev in elevator order. | |
537 | * this with as_choose_req form the basis for how the scheduler chooses | |
538 | * what request to process next. Anticipation works on top of this. | |
539 | */ | |
540 | static struct as_rq *as_find_next_arq(struct as_data *ad, struct as_rq *last) | |
541 | { | |
542 | const int data_dir = last->is_sync; | |
543 | struct as_rq *ret; | |
544 | struct rb_node *rbnext = rb_next(&last->rb_node); | |
545 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
546 | struct as_rq *arq_next, *arq_prev; | |
547 | ||
548 | BUG_ON(!ON_RB(&last->rb_node)); | |
549 | ||
550 | if (rbprev) | |
551 | arq_prev = rb_entry_arq(rbprev); | |
552 | else | |
553 | arq_prev = NULL; | |
554 | ||
555 | if (rbnext) | |
556 | arq_next = rb_entry_arq(rbnext); | |
557 | else { | |
558 | arq_next = as_find_first_arq(ad, data_dir); | |
559 | if (arq_next == last) | |
560 | arq_next = NULL; | |
561 | } | |
562 | ||
563 | ret = as_choose_req(ad, arq_next, arq_prev); | |
564 | ||
565 | return ret; | |
566 | } | |
567 | ||
568 | /* | |
569 | * anticipatory scheduling functions follow | |
570 | */ | |
571 | ||
572 | /* | |
573 | * as_antic_expired tells us when we have anticipated too long. | |
574 | * The funny "absolute difference" math on the elapsed time is to handle | |
575 | * jiffy wraps, and disks which have been idle for 0x80000000 jiffies. | |
576 | */ | |
577 | static int as_antic_expired(struct as_data *ad) | |
578 | { | |
579 | long delta_jif; | |
580 | ||
581 | delta_jif = jiffies - ad->antic_start; | |
582 | if (unlikely(delta_jif < 0)) | |
583 | delta_jif = -delta_jif; | |
584 | if (delta_jif < ad->antic_expire) | |
585 | return 0; | |
586 | ||
587 | return 1; | |
588 | } | |
589 | ||
590 | /* | |
591 | * as_antic_waitnext starts anticipating that a nice request will soon be | |
592 | * submitted. See also as_antic_waitreq | |
593 | */ | |
594 | static void as_antic_waitnext(struct as_data *ad) | |
595 | { | |
596 | unsigned long timeout; | |
597 | ||
598 | BUG_ON(ad->antic_status != ANTIC_OFF | |
599 | && ad->antic_status != ANTIC_WAIT_REQ); | |
600 | ||
601 | timeout = ad->antic_start + ad->antic_expire; | |
602 | ||
603 | mod_timer(&ad->antic_timer, timeout); | |
604 | ||
605 | ad->antic_status = ANTIC_WAIT_NEXT; | |
606 | } | |
607 | ||
608 | /* | |
609 | * as_antic_waitreq starts anticipating. We don't start timing the anticipation | |
610 | * until the request that we're anticipating on has finished. This means we | |
611 | * are timing from when the candidate process wakes up hopefully. | |
612 | */ | |
613 | static void as_antic_waitreq(struct as_data *ad) | |
614 | { | |
615 | BUG_ON(ad->antic_status == ANTIC_FINISHED); | |
616 | if (ad->antic_status == ANTIC_OFF) { | |
617 | if (!ad->io_context || ad->ioc_finished) | |
618 | as_antic_waitnext(ad); | |
619 | else | |
620 | ad->antic_status = ANTIC_WAIT_REQ; | |
621 | } | |
622 | } | |
623 | ||
624 | /* | |
625 | * This is called directly by the functions in this file to stop anticipation. | |
626 | * We kill the timer and schedule a call to the request_fn asap. | |
627 | */ | |
628 | static void as_antic_stop(struct as_data *ad) | |
629 | { | |
630 | int status = ad->antic_status; | |
631 | ||
632 | if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) { | |
633 | if (status == ANTIC_WAIT_NEXT) | |
634 | del_timer(&ad->antic_timer); | |
635 | ad->antic_status = ANTIC_FINISHED; | |
636 | /* see as_work_handler */ | |
637 | kblockd_schedule_work(&ad->antic_work); | |
638 | } | |
639 | } | |
640 | ||
641 | /* | |
642 | * as_antic_timeout is the timer function set by as_antic_waitnext. | |
643 | */ | |
644 | static void as_antic_timeout(unsigned long data) | |
645 | { | |
646 | struct request_queue *q = (struct request_queue *)data; | |
647 | struct as_data *ad = q->elevator->elevator_data; | |
648 | unsigned long flags; | |
649 | ||
650 | spin_lock_irqsave(q->queue_lock, flags); | |
651 | if (ad->antic_status == ANTIC_WAIT_REQ | |
652 | || ad->antic_status == ANTIC_WAIT_NEXT) { | |
653 | struct as_io_context *aic = ad->io_context->aic; | |
654 | ||
655 | ad->antic_status = ANTIC_FINISHED; | |
656 | kblockd_schedule_work(&ad->antic_work); | |
657 | ||
658 | if (aic->ttime_samples == 0) { | |
f5b3db00 | 659 | /* process anticipated on has exited or timed out*/ |
1da177e4 LT |
660 | ad->exit_prob = (7*ad->exit_prob + 256)/8; |
661 | } | |
f5b3db00 NP |
662 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
663 | /* process not "saved" by a cooperating request */ | |
664 | ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8; | |
665 | } | |
1da177e4 LT |
666 | } |
667 | spin_unlock_irqrestore(q->queue_lock, flags); | |
668 | } | |
669 | ||
f5b3db00 NP |
670 | static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic, |
671 | unsigned long ttime) | |
672 | { | |
673 | /* fixed point: 1.0 == 1<<8 */ | |
674 | if (aic->ttime_samples == 0) { | |
675 | ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8; | |
676 | ad->new_ttime_mean = ad->new_ttime_total / 256; | |
677 | ||
678 | ad->exit_prob = (7*ad->exit_prob)/8; | |
679 | } | |
680 | aic->ttime_samples = (7*aic->ttime_samples + 256) / 8; | |
681 | aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8; | |
682 | aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples; | |
683 | } | |
684 | ||
685 | static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic, | |
686 | sector_t sdist) | |
687 | { | |
688 | u64 total; | |
689 | ||
690 | if (aic->seek_samples == 0) { | |
691 | ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8; | |
692 | ad->new_seek_mean = ad->new_seek_total / 256; | |
693 | } | |
694 | ||
695 | /* | |
696 | * Don't allow the seek distance to get too large from the | |
697 | * odd fragment, pagein, etc | |
698 | */ | |
699 | if (aic->seek_samples <= 60) /* second&third seek */ | |
700 | sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024); | |
701 | else | |
702 | sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64); | |
703 | ||
704 | aic->seek_samples = (7*aic->seek_samples + 256) / 8; | |
705 | aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8; | |
706 | total = aic->seek_total + (aic->seek_samples/2); | |
707 | do_div(total, aic->seek_samples); | |
708 | aic->seek_mean = (sector_t)total; | |
709 | } | |
710 | ||
711 | /* | |
712 | * as_update_iohist keeps a decaying histogram of IO thinktimes, and | |
713 | * updates @aic->ttime_mean based on that. It is called when a new | |
714 | * request is queued. | |
715 | */ | |
716 | static void as_update_iohist(struct as_data *ad, struct as_io_context *aic, | |
717 | struct request *rq) | |
718 | { | |
719 | struct as_rq *arq = RQ_DATA(rq); | |
720 | int data_dir = arq->is_sync; | |
721 | unsigned long thinktime = 0; | |
722 | sector_t seek_dist; | |
723 | ||
724 | if (aic == NULL) | |
725 | return; | |
726 | ||
727 | if (data_dir == REQ_SYNC) { | |
728 | unsigned long in_flight = atomic_read(&aic->nr_queued) | |
729 | + atomic_read(&aic->nr_dispatched); | |
730 | spin_lock(&aic->lock); | |
731 | if (test_bit(AS_TASK_IORUNNING, &aic->state) || | |
732 | test_bit(AS_TASK_IOSTARTED, &aic->state)) { | |
733 | /* Calculate read -> read thinktime */ | |
734 | if (test_bit(AS_TASK_IORUNNING, &aic->state) | |
735 | && in_flight == 0) { | |
736 | thinktime = jiffies - aic->last_end_request; | |
737 | thinktime = min(thinktime, MAX_THINKTIME-1); | |
738 | } | |
739 | as_update_thinktime(ad, aic, thinktime); | |
740 | ||
741 | /* Calculate read -> read seek distance */ | |
742 | if (aic->last_request_pos < rq->sector) | |
743 | seek_dist = rq->sector - aic->last_request_pos; | |
744 | else | |
745 | seek_dist = aic->last_request_pos - rq->sector; | |
746 | as_update_seekdist(ad, aic, seek_dist); | |
747 | } | |
748 | aic->last_request_pos = rq->sector + rq->nr_sectors; | |
749 | set_bit(AS_TASK_IOSTARTED, &aic->state); | |
750 | spin_unlock(&aic->lock); | |
751 | } | |
752 | } | |
753 | ||
1da177e4 LT |
754 | /* |
755 | * as_close_req decides if one request is considered "close" to the | |
756 | * previous one issued. | |
757 | */ | |
f5b3db00 NP |
758 | static int as_close_req(struct as_data *ad, struct as_io_context *aic, |
759 | struct as_rq *arq) | |
1da177e4 LT |
760 | { |
761 | unsigned long delay; /* milliseconds */ | |
762 | sector_t last = ad->last_sector[ad->batch_data_dir]; | |
763 | sector_t next = arq->request->sector; | |
764 | sector_t delta; /* acceptable close offset (in sectors) */ | |
f5b3db00 | 765 | sector_t s; |
1da177e4 LT |
766 | |
767 | if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished) | |
768 | delay = 0; | |
769 | else | |
770 | delay = ((jiffies - ad->antic_start) * 1000) / HZ; | |
771 | ||
f5b3db00 NP |
772 | if (delay == 0) |
773 | delta = 8192; | |
1da177e4 | 774 | else if (delay <= 20 && delay <= ad->antic_expire) |
f5b3db00 | 775 | delta = 8192 << delay; |
1da177e4 LT |
776 | else |
777 | return 1; | |
778 | ||
f5b3db00 NP |
779 | if ((last <= next + (delta>>1)) && (next <= last + delta)) |
780 | return 1; | |
781 | ||
782 | if (last < next) | |
783 | s = next - last; | |
784 | else | |
785 | s = last - next; | |
786 | ||
787 | if (aic->seek_samples == 0) { | |
788 | /* | |
789 | * Process has just started IO. Use past statistics to | |
790 | * gauge success possibility | |
791 | */ | |
792 | if (ad->new_seek_mean > s) { | |
793 | /* this request is better than what we're expecting */ | |
794 | return 1; | |
795 | } | |
796 | ||
797 | } else { | |
798 | if (aic->seek_mean > s) { | |
799 | /* this request is better than what we're expecting */ | |
800 | return 1; | |
801 | } | |
802 | } | |
803 | ||
804 | return 0; | |
1da177e4 LT |
805 | } |
806 | ||
807 | /* | |
808 | * as_can_break_anticipation returns true if we have been anticipating this | |
809 | * request. | |
810 | * | |
811 | * It also returns true if the process against which we are anticipating | |
812 | * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to | |
813 | * dispatch it ASAP, because we know that application will not be submitting | |
814 | * any new reads. | |
815 | * | |
f5b3db00 | 816 | * If the task which has submitted the request has exited, break anticipation. |
1da177e4 LT |
817 | * |
818 | * If this task has queued some other IO, do not enter enticipation. | |
819 | */ | |
820 | static int as_can_break_anticipation(struct as_data *ad, struct as_rq *arq) | |
821 | { | |
822 | struct io_context *ioc; | |
823 | struct as_io_context *aic; | |
1da177e4 LT |
824 | |
825 | ioc = ad->io_context; | |
826 | BUG_ON(!ioc); | |
827 | ||
828 | if (arq && ioc == arq->io_context) { | |
829 | /* request from same process */ | |
830 | return 1; | |
831 | } | |
832 | ||
833 | if (ad->ioc_finished && as_antic_expired(ad)) { | |
834 | /* | |
835 | * In this situation status should really be FINISHED, | |
836 | * however the timer hasn't had the chance to run yet. | |
837 | */ | |
838 | return 1; | |
839 | } | |
840 | ||
841 | aic = ioc->aic; | |
842 | if (!aic) | |
843 | return 0; | |
844 | ||
1da177e4 LT |
845 | if (atomic_read(&aic->nr_queued) > 0) { |
846 | /* process has more requests queued */ | |
847 | return 1; | |
848 | } | |
849 | ||
850 | if (atomic_read(&aic->nr_dispatched) > 0) { | |
851 | /* process has more requests dispatched */ | |
852 | return 1; | |
853 | } | |
854 | ||
f5b3db00 | 855 | if (arq && arq->is_sync == REQ_SYNC && as_close_req(ad, aic, arq)) { |
1da177e4 LT |
856 | /* |
857 | * Found a close request that is not one of ours. | |
858 | * | |
f5b3db00 NP |
859 | * This makes close requests from another process update |
860 | * our IO history. Is generally useful when there are | |
1da177e4 LT |
861 | * two or more cooperating processes working in the same |
862 | * area. | |
863 | */ | |
f5b3db00 NP |
864 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
865 | if (aic->ttime_samples == 0) | |
866 | ad->exit_prob = (7*ad->exit_prob + 256)/8; | |
867 | ||
868 | ad->exit_no_coop = (7*ad->exit_no_coop)/8; | |
869 | } | |
870 | ||
871 | as_update_iohist(ad, aic, arq->request); | |
1da177e4 LT |
872 | return 1; |
873 | } | |
874 | ||
f5b3db00 NP |
875 | if (!test_bit(AS_TASK_RUNNING, &aic->state)) { |
876 | /* process anticipated on has exited */ | |
877 | if (aic->ttime_samples == 0) | |
878 | ad->exit_prob = (7*ad->exit_prob + 256)/8; | |
879 | ||
880 | if (ad->exit_no_coop > 128) | |
881 | return 1; | |
882 | } | |
1da177e4 LT |
883 | |
884 | if (aic->ttime_samples == 0) { | |
885 | if (ad->new_ttime_mean > ad->antic_expire) | |
886 | return 1; | |
f5b3db00 | 887 | if (ad->exit_prob * ad->exit_no_coop > 128*256) |
1da177e4 LT |
888 | return 1; |
889 | } else if (aic->ttime_mean > ad->antic_expire) { | |
890 | /* the process thinks too much between requests */ | |
891 | return 1; | |
892 | } | |
893 | ||
1da177e4 LT |
894 | return 0; |
895 | } | |
896 | ||
897 | /* | |
898 | * as_can_anticipate indicates weather we should either run arq | |
899 | * or keep anticipating a better request. | |
900 | */ | |
901 | static int as_can_anticipate(struct as_data *ad, struct as_rq *arq) | |
902 | { | |
903 | if (!ad->io_context) | |
904 | /* | |
905 | * Last request submitted was a write | |
906 | */ | |
907 | return 0; | |
908 | ||
909 | if (ad->antic_status == ANTIC_FINISHED) | |
910 | /* | |
911 | * Don't restart if we have just finished. Run the next request | |
912 | */ | |
913 | return 0; | |
914 | ||
915 | if (as_can_break_anticipation(ad, arq)) | |
916 | /* | |
917 | * This request is a good candidate. Don't keep anticipating, | |
918 | * run it. | |
919 | */ | |
920 | return 0; | |
921 | ||
922 | /* | |
923 | * OK from here, we haven't finished, and don't have a decent request! | |
924 | * Status is either ANTIC_OFF so start waiting, | |
925 | * ANTIC_WAIT_REQ so continue waiting for request to finish | |
926 | * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request. | |
1da177e4 LT |
927 | */ |
928 | ||
929 | return 1; | |
930 | } | |
931 | ||
1da177e4 LT |
932 | /* |
933 | * as_update_arq must be called whenever a request (arq) is added to | |
934 | * the sort_list. This function keeps caches up to date, and checks if the | |
935 | * request might be one we are "anticipating" | |
936 | */ | |
937 | static void as_update_arq(struct as_data *ad, struct as_rq *arq) | |
938 | { | |
939 | const int data_dir = arq->is_sync; | |
940 | ||
941 | /* keep the next_arq cache up to date */ | |
942 | ad->next_arq[data_dir] = as_choose_req(ad, arq, ad->next_arq[data_dir]); | |
943 | ||
944 | /* | |
945 | * have we been anticipating this request? | |
946 | * or does it come from the same process as the one we are anticipating | |
947 | * for? | |
948 | */ | |
949 | if (ad->antic_status == ANTIC_WAIT_REQ | |
950 | || ad->antic_status == ANTIC_WAIT_NEXT) { | |
951 | if (as_can_break_anticipation(ad, arq)) | |
952 | as_antic_stop(ad); | |
953 | } | |
954 | } | |
955 | ||
956 | /* | |
957 | * Gathers timings and resizes the write batch automatically | |
958 | */ | |
959 | static void update_write_batch(struct as_data *ad) | |
960 | { | |
961 | unsigned long batch = ad->batch_expire[REQ_ASYNC]; | |
962 | long write_time; | |
963 | ||
964 | write_time = (jiffies - ad->current_batch_expires) + batch; | |
965 | if (write_time < 0) | |
966 | write_time = 0; | |
967 | ||
968 | if (write_time > batch && !ad->write_batch_idled) { | |
969 | if (write_time > batch * 3) | |
970 | ad->write_batch_count /= 2; | |
971 | else | |
972 | ad->write_batch_count--; | |
973 | } else if (write_time < batch && ad->current_write_count == 0) { | |
974 | if (batch > write_time * 3) | |
975 | ad->write_batch_count *= 2; | |
976 | else | |
977 | ad->write_batch_count++; | |
978 | } | |
979 | ||
980 | if (ad->write_batch_count < 1) | |
981 | ad->write_batch_count = 1; | |
982 | } | |
983 | ||
984 | /* | |
985 | * as_completed_request is to be called when a request has completed and | |
986 | * returned something to the requesting process, be it an error or data. | |
987 | */ | |
988 | static void as_completed_request(request_queue_t *q, struct request *rq) | |
989 | { | |
990 | struct as_data *ad = q->elevator->elevator_data; | |
991 | struct as_rq *arq = RQ_DATA(rq); | |
992 | ||
993 | WARN_ON(!list_empty(&rq->queuelist)); | |
994 | ||
1da177e4 LT |
995 | if (arq->state != AS_RQ_REMOVED) { |
996 | printk("arq->state %d\n", arq->state); | |
997 | WARN_ON(1); | |
998 | goto out; | |
999 | } | |
1000 | ||
1da177e4 LT |
1001 | if (ad->changed_batch && ad->nr_dispatched == 1) { |
1002 | kblockd_schedule_work(&ad->antic_work); | |
1003 | ad->changed_batch = 0; | |
1004 | ||
1005 | if (ad->batch_data_dir == REQ_SYNC) | |
1006 | ad->new_batch = 1; | |
1007 | } | |
1008 | WARN_ON(ad->nr_dispatched == 0); | |
1009 | ad->nr_dispatched--; | |
1010 | ||
1011 | /* | |
1012 | * Start counting the batch from when a request of that direction is | |
1013 | * actually serviced. This should help devices with big TCQ windows | |
1014 | * and writeback caches | |
1015 | */ | |
1016 | if (ad->new_batch && ad->batch_data_dir == arq->is_sync) { | |
1017 | update_write_batch(ad); | |
1018 | ad->current_batch_expires = jiffies + | |
1019 | ad->batch_expire[REQ_SYNC]; | |
1020 | ad->new_batch = 0; | |
1021 | } | |
1022 | ||
1023 | if (ad->io_context == arq->io_context && ad->io_context) { | |
1024 | ad->antic_start = jiffies; | |
1025 | ad->ioc_finished = 1; | |
1026 | if (ad->antic_status == ANTIC_WAIT_REQ) { | |
1027 | /* | |
1028 | * We were waiting on this request, now anticipate | |
1029 | * the next one | |
1030 | */ | |
1031 | as_antic_waitnext(ad); | |
1032 | } | |
1033 | } | |
1034 | ||
b4878f24 | 1035 | as_put_io_context(arq); |
1da177e4 LT |
1036 | out: |
1037 | arq->state = AS_RQ_POSTSCHED; | |
1038 | } | |
1039 | ||
1040 | /* | |
1041 | * as_remove_queued_request removes a request from the pre dispatch queue | |
1042 | * without updating refcounts. It is expected the caller will drop the | |
1043 | * reference unless it replaces the request at somepart of the elevator | |
1044 | * (ie. the dispatch queue) | |
1045 | */ | |
1046 | static void as_remove_queued_request(request_queue_t *q, struct request *rq) | |
1047 | { | |
1048 | struct as_rq *arq = RQ_DATA(rq); | |
1049 | const int data_dir = arq->is_sync; | |
1050 | struct as_data *ad = q->elevator->elevator_data; | |
1051 | ||
1052 | WARN_ON(arq->state != AS_RQ_QUEUED); | |
1053 | ||
1054 | if (arq->io_context && arq->io_context->aic) { | |
1055 | BUG_ON(!atomic_read(&arq->io_context->aic->nr_queued)); | |
1056 | atomic_dec(&arq->io_context->aic->nr_queued); | |
1057 | } | |
1058 | ||
1059 | /* | |
1060 | * Update the "next_arq" cache if we are about to remove its | |
1061 | * entry | |
1062 | */ | |
1063 | if (ad->next_arq[data_dir] == arq) | |
1064 | ad->next_arq[data_dir] = as_find_next_arq(ad, arq); | |
1065 | ||
1066 | list_del_init(&arq->fifo); | |
98b11471 | 1067 | as_del_arq_hash(arq); |
1da177e4 LT |
1068 | as_del_arq_rb(ad, arq); |
1069 | } | |
1070 | ||
1da177e4 LT |
1071 | /* |
1072 | * as_fifo_expired returns 0 if there are no expired reads on the fifo, | |
1073 | * 1 otherwise. It is ratelimited so that we only perform the check once per | |
1074 | * `fifo_expire' interval. Otherwise a large number of expired requests | |
1075 | * would create a hopeless seekstorm. | |
1076 | * | |
1077 | * See as_antic_expired comment. | |
1078 | */ | |
1079 | static int as_fifo_expired(struct as_data *ad, int adir) | |
1080 | { | |
1081 | struct as_rq *arq; | |
1082 | long delta_jif; | |
1083 | ||
1084 | delta_jif = jiffies - ad->last_check_fifo[adir]; | |
1085 | if (unlikely(delta_jif < 0)) | |
1086 | delta_jif = -delta_jif; | |
1087 | if (delta_jif < ad->fifo_expire[adir]) | |
1088 | return 0; | |
1089 | ||
1090 | ad->last_check_fifo[adir] = jiffies; | |
1091 | ||
1092 | if (list_empty(&ad->fifo_list[adir])) | |
1093 | return 0; | |
1094 | ||
1095 | arq = list_entry_fifo(ad->fifo_list[adir].next); | |
1096 | ||
1097 | return time_after(jiffies, arq->expires); | |
1098 | } | |
1099 | ||
1100 | /* | |
1101 | * as_batch_expired returns true if the current batch has expired. A batch | |
1102 | * is a set of reads or a set of writes. | |
1103 | */ | |
1104 | static inline int as_batch_expired(struct as_data *ad) | |
1105 | { | |
1106 | if (ad->changed_batch || ad->new_batch) | |
1107 | return 0; | |
1108 | ||
1109 | if (ad->batch_data_dir == REQ_SYNC) | |
1110 | /* TODO! add a check so a complete fifo gets written? */ | |
1111 | return time_after(jiffies, ad->current_batch_expires); | |
1112 | ||
1113 | return time_after(jiffies, ad->current_batch_expires) | |
1114 | || ad->current_write_count == 0; | |
1115 | } | |
1116 | ||
1117 | /* | |
1118 | * move an entry to dispatch queue | |
1119 | */ | |
1120 | static void as_move_to_dispatch(struct as_data *ad, struct as_rq *arq) | |
1121 | { | |
1122 | struct request *rq = arq->request; | |
1da177e4 LT |
1123 | const int data_dir = arq->is_sync; |
1124 | ||
1125 | BUG_ON(!ON_RB(&arq->rb_node)); | |
1126 | ||
1127 | as_antic_stop(ad); | |
1128 | ad->antic_status = ANTIC_OFF; | |
1129 | ||
1130 | /* | |
1131 | * This has to be set in order to be correctly updated by | |
1132 | * as_find_next_arq | |
1133 | */ | |
1134 | ad->last_sector[data_dir] = rq->sector + rq->nr_sectors; | |
1135 | ||
1136 | if (data_dir == REQ_SYNC) { | |
1137 | /* In case we have to anticipate after this */ | |
1138 | copy_io_context(&ad->io_context, &arq->io_context); | |
1139 | } else { | |
1140 | if (ad->io_context) { | |
1141 | put_io_context(ad->io_context); | |
1142 | ad->io_context = NULL; | |
1143 | } | |
1144 | ||
1145 | if (ad->current_write_count != 0) | |
1146 | ad->current_write_count--; | |
1147 | } | |
1148 | ad->ioc_finished = 0; | |
1149 | ||
1150 | ad->next_arq[data_dir] = as_find_next_arq(ad, arq); | |
1151 | ||
1152 | /* | |
1153 | * take it off the sort and fifo list, add to dispatch queue | |
1154 | */ | |
1da177e4 LT |
1155 | as_remove_queued_request(ad->q, rq); |
1156 | WARN_ON(arq->state != AS_RQ_QUEUED); | |
1157 | ||
b4878f24 JA |
1158 | elv_dispatch_sort(ad->q, rq); |
1159 | ||
1da177e4 LT |
1160 | arq->state = AS_RQ_DISPATCHED; |
1161 | if (arq->io_context && arq->io_context->aic) | |
1162 | atomic_inc(&arq->io_context->aic->nr_dispatched); | |
1163 | ad->nr_dispatched++; | |
1164 | } | |
1165 | ||
1166 | /* | |
1167 | * as_dispatch_request selects the best request according to | |
1168 | * read/write expire, batch expire, etc, and moves it to the dispatch | |
1169 | * queue. Returns 1 if a request was found, 0 otherwise. | |
1170 | */ | |
b4878f24 | 1171 | static int as_dispatch_request(request_queue_t *q, int force) |
1da177e4 | 1172 | { |
b4878f24 | 1173 | struct as_data *ad = q->elevator->elevator_data; |
1da177e4 LT |
1174 | struct as_rq *arq; |
1175 | const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]); | |
1176 | const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]); | |
1177 | ||
b4878f24 JA |
1178 | if (unlikely(force)) { |
1179 | /* | |
1180 | * Forced dispatch, accounting is useless. Reset | |
1181 | * accounting states and dump fifo_lists. Note that | |
1182 | * batch_data_dir is reset to REQ_SYNC to avoid | |
1183 | * screwing write batch accounting as write batch | |
1184 | * accounting occurs on W->R transition. | |
1185 | */ | |
1186 | int dispatched = 0; | |
1187 | ||
1188 | ad->batch_data_dir = REQ_SYNC; | |
1189 | ad->changed_batch = 0; | |
1190 | ad->new_batch = 0; | |
1191 | ||
1192 | while (ad->next_arq[REQ_SYNC]) { | |
1193 | as_move_to_dispatch(ad, ad->next_arq[REQ_SYNC]); | |
1194 | dispatched++; | |
1195 | } | |
1196 | ad->last_check_fifo[REQ_SYNC] = jiffies; | |
1197 | ||
1198 | while (ad->next_arq[REQ_ASYNC]) { | |
1199 | as_move_to_dispatch(ad, ad->next_arq[REQ_ASYNC]); | |
1200 | dispatched++; | |
1201 | } | |
1202 | ad->last_check_fifo[REQ_ASYNC] = jiffies; | |
1203 | ||
1204 | return dispatched; | |
1205 | } | |
1206 | ||
1da177e4 LT |
1207 | /* Signal that the write batch was uncontended, so we can't time it */ |
1208 | if (ad->batch_data_dir == REQ_ASYNC && !reads) { | |
1209 | if (ad->current_write_count == 0 || !writes) | |
1210 | ad->write_batch_idled = 1; | |
1211 | } | |
1212 | ||
1213 | if (!(reads || writes) | |
1214 | || ad->antic_status == ANTIC_WAIT_REQ | |
1215 | || ad->antic_status == ANTIC_WAIT_NEXT | |
1216 | || ad->changed_batch) | |
1217 | return 0; | |
1218 | ||
f5b3db00 | 1219 | if (!(reads && writes && as_batch_expired(ad))) { |
1da177e4 LT |
1220 | /* |
1221 | * batch is still running or no reads or no writes | |
1222 | */ | |
1223 | arq = ad->next_arq[ad->batch_data_dir]; | |
1224 | ||
1225 | if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) { | |
1226 | if (as_fifo_expired(ad, REQ_SYNC)) | |
1227 | goto fifo_expired; | |
1228 | ||
1229 | if (as_can_anticipate(ad, arq)) { | |
1230 | as_antic_waitreq(ad); | |
1231 | return 0; | |
1232 | } | |
1233 | } | |
1234 | ||
1235 | if (arq) { | |
1236 | /* we have a "next request" */ | |
1237 | if (reads && !writes) | |
1238 | ad->current_batch_expires = | |
1239 | jiffies + ad->batch_expire[REQ_SYNC]; | |
1240 | goto dispatch_request; | |
1241 | } | |
1242 | } | |
1243 | ||
1244 | /* | |
1245 | * at this point we are not running a batch. select the appropriate | |
1246 | * data direction (read / write) | |
1247 | */ | |
1248 | ||
1249 | if (reads) { | |
1250 | BUG_ON(RB_EMPTY(&ad->sort_list[REQ_SYNC])); | |
1251 | ||
1252 | if (writes && ad->batch_data_dir == REQ_SYNC) | |
1253 | /* | |
1254 | * Last batch was a read, switch to writes | |
1255 | */ | |
1256 | goto dispatch_writes; | |
1257 | ||
1258 | if (ad->batch_data_dir == REQ_ASYNC) { | |
1259 | WARN_ON(ad->new_batch); | |
1260 | ad->changed_batch = 1; | |
1261 | } | |
1262 | ad->batch_data_dir = REQ_SYNC; | |
1263 | arq = list_entry_fifo(ad->fifo_list[ad->batch_data_dir].next); | |
1264 | ad->last_check_fifo[ad->batch_data_dir] = jiffies; | |
1265 | goto dispatch_request; | |
1266 | } | |
1267 | ||
1268 | /* | |
1269 | * the last batch was a read | |
1270 | */ | |
1271 | ||
1272 | if (writes) { | |
1273 | dispatch_writes: | |
1274 | BUG_ON(RB_EMPTY(&ad->sort_list[REQ_ASYNC])); | |
1275 | ||
1276 | if (ad->batch_data_dir == REQ_SYNC) { | |
1277 | ad->changed_batch = 1; | |
1278 | ||
1279 | /* | |
1280 | * new_batch might be 1 when the queue runs out of | |
1281 | * reads. A subsequent submission of a write might | |
1282 | * cause a change of batch before the read is finished. | |
1283 | */ | |
1284 | ad->new_batch = 0; | |
1285 | } | |
1286 | ad->batch_data_dir = REQ_ASYNC; | |
1287 | ad->current_write_count = ad->write_batch_count; | |
1288 | ad->write_batch_idled = 0; | |
1289 | arq = ad->next_arq[ad->batch_data_dir]; | |
1290 | goto dispatch_request; | |
1291 | } | |
1292 | ||
1293 | BUG(); | |
1294 | return 0; | |
1295 | ||
1296 | dispatch_request: | |
1297 | /* | |
1298 | * If a request has expired, service it. | |
1299 | */ | |
1300 | ||
1301 | if (as_fifo_expired(ad, ad->batch_data_dir)) { | |
1302 | fifo_expired: | |
1303 | arq = list_entry_fifo(ad->fifo_list[ad->batch_data_dir].next); | |
1304 | BUG_ON(arq == NULL); | |
1305 | } | |
1306 | ||
1307 | if (ad->changed_batch) { | |
1308 | WARN_ON(ad->new_batch); | |
1309 | ||
1310 | if (ad->nr_dispatched) | |
1311 | return 0; | |
1312 | ||
1313 | if (ad->batch_data_dir == REQ_ASYNC) | |
1314 | ad->current_batch_expires = jiffies + | |
1315 | ad->batch_expire[REQ_ASYNC]; | |
1316 | else | |
1317 | ad->new_batch = 1; | |
1318 | ||
1319 | ad->changed_batch = 0; | |
1320 | } | |
1321 | ||
1322 | /* | |
1323 | * arq is the selected appropriate request. | |
1324 | */ | |
1325 | as_move_to_dispatch(ad, arq); | |
1326 | ||
1327 | return 1; | |
1328 | } | |
1329 | ||
1da177e4 LT |
1330 | /* |
1331 | * add arq to rbtree and fifo | |
1332 | */ | |
b4878f24 | 1333 | static void as_add_request(request_queue_t *q, struct request *rq) |
1da177e4 | 1334 | { |
b4878f24 JA |
1335 | struct as_data *ad = q->elevator->elevator_data; |
1336 | struct as_rq *arq = RQ_DATA(rq); | |
1da177e4 LT |
1337 | int data_dir; |
1338 | ||
b4878f24 JA |
1339 | arq->state = AS_RQ_NEW; |
1340 | ||
1da177e4 | 1341 | if (rq_data_dir(arq->request) == READ |
b31dc66a | 1342 | || (arq->request->flags & REQ_RW_SYNC)) |
1da177e4 LT |
1343 | arq->is_sync = 1; |
1344 | else | |
1345 | arq->is_sync = 0; | |
1346 | data_dir = arq->is_sync; | |
1347 | ||
1348 | arq->io_context = as_get_io_context(); | |
1349 | ||
1350 | if (arq->io_context) { | |
1351 | as_update_iohist(ad, arq->io_context->aic, arq->request); | |
1352 | atomic_inc(&arq->io_context->aic->nr_queued); | |
1353 | } | |
1354 | ||
ef9be1d3 TH |
1355 | as_add_arq_rb(ad, arq); |
1356 | if (rq_mergeable(arq->request)) | |
1357 | as_add_arq_hash(ad, arq); | |
1da177e4 | 1358 | |
ef9be1d3 TH |
1359 | /* |
1360 | * set expire time (only used for reads) and add to fifo list | |
1361 | */ | |
1362 | arq->expires = jiffies + ad->fifo_expire[data_dir]; | |
1363 | list_add_tail(&arq->fifo, &ad->fifo_list[data_dir]); | |
1da177e4 | 1364 | |
ef9be1d3 | 1365 | as_update_arq(ad, arq); /* keep state machine up to date */ |
1da177e4 LT |
1366 | arq->state = AS_RQ_QUEUED; |
1367 | } | |
1368 | ||
b4878f24 | 1369 | static void as_activate_request(request_queue_t *q, struct request *rq) |
1da177e4 | 1370 | { |
1da177e4 LT |
1371 | struct as_rq *arq = RQ_DATA(rq); |
1372 | ||
b4878f24 JA |
1373 | WARN_ON(arq->state != AS_RQ_DISPATCHED); |
1374 | arq->state = AS_RQ_REMOVED; | |
1375 | if (arq->io_context && arq->io_context->aic) | |
1376 | atomic_dec(&arq->io_context->aic->nr_dispatched); | |
1da177e4 LT |
1377 | } |
1378 | ||
b4878f24 | 1379 | static void as_deactivate_request(request_queue_t *q, struct request *rq) |
1da177e4 | 1380 | { |
1da177e4 LT |
1381 | struct as_rq *arq = RQ_DATA(rq); |
1382 | ||
b4878f24 JA |
1383 | WARN_ON(arq->state != AS_RQ_REMOVED); |
1384 | arq->state = AS_RQ_DISPATCHED; | |
1385 | if (arq->io_context && arq->io_context->aic) | |
1386 | atomic_inc(&arq->io_context->aic->nr_dispatched); | |
1da177e4 LT |
1387 | } |
1388 | ||
1389 | /* | |
1390 | * as_queue_empty tells us if there are requests left in the device. It may | |
1391 | * not be the case that a driver can get the next request even if the queue | |
1392 | * is not empty - it is used in the block layer to check for plugging and | |
1393 | * merging opportunities | |
1394 | */ | |
1395 | static int as_queue_empty(request_queue_t *q) | |
1396 | { | |
1397 | struct as_data *ad = q->elevator->elevator_data; | |
1398 | ||
b4878f24 JA |
1399 | return list_empty(&ad->fifo_list[REQ_ASYNC]) |
1400 | && list_empty(&ad->fifo_list[REQ_SYNC]); | |
1da177e4 LT |
1401 | } |
1402 | ||
f5b3db00 NP |
1403 | static struct request *as_former_request(request_queue_t *q, |
1404 | struct request *rq) | |
1da177e4 LT |
1405 | { |
1406 | struct as_rq *arq = RQ_DATA(rq); | |
1407 | struct rb_node *rbprev = rb_prev(&arq->rb_node); | |
1408 | struct request *ret = NULL; | |
1409 | ||
1410 | if (rbprev) | |
1411 | ret = rb_entry_arq(rbprev)->request; | |
1412 | ||
1413 | return ret; | |
1414 | } | |
1415 | ||
f5b3db00 NP |
1416 | static struct request *as_latter_request(request_queue_t *q, |
1417 | struct request *rq) | |
1da177e4 LT |
1418 | { |
1419 | struct as_rq *arq = RQ_DATA(rq); | |
1420 | struct rb_node *rbnext = rb_next(&arq->rb_node); | |
1421 | struct request *ret = NULL; | |
1422 | ||
1423 | if (rbnext) | |
1424 | ret = rb_entry_arq(rbnext)->request; | |
1425 | ||
1426 | return ret; | |
1427 | } | |
1428 | ||
1429 | static int | |
1430 | as_merge(request_queue_t *q, struct request **req, struct bio *bio) | |
1431 | { | |
1432 | struct as_data *ad = q->elevator->elevator_data; | |
1433 | sector_t rb_key = bio->bi_sector + bio_sectors(bio); | |
1434 | struct request *__rq; | |
1435 | int ret; | |
1436 | ||
1da177e4 LT |
1437 | /* |
1438 | * see if the merge hash can satisfy a back merge | |
1439 | */ | |
1440 | __rq = as_find_arq_hash(ad, bio->bi_sector); | |
1441 | if (__rq) { | |
1442 | BUG_ON(__rq->sector + __rq->nr_sectors != bio->bi_sector); | |
1443 | ||
1444 | if (elv_rq_merge_ok(__rq, bio)) { | |
1445 | ret = ELEVATOR_BACK_MERGE; | |
1446 | goto out; | |
1447 | } | |
1448 | } | |
1449 | ||
1450 | /* | |
1451 | * check for front merge | |
1452 | */ | |
1453 | __rq = as_find_arq_rb(ad, rb_key, bio_data_dir(bio)); | |
1454 | if (__rq) { | |
1455 | BUG_ON(rb_key != rq_rb_key(__rq)); | |
1456 | ||
1457 | if (elv_rq_merge_ok(__rq, bio)) { | |
1458 | ret = ELEVATOR_FRONT_MERGE; | |
1459 | goto out; | |
1460 | } | |
1461 | } | |
1462 | ||
1463 | return ELEVATOR_NO_MERGE; | |
1464 | out: | |
1da177e4 LT |
1465 | if (ret) { |
1466 | if (rq_mergeable(__rq)) | |
1467 | as_hot_arq_hash(ad, RQ_DATA(__rq)); | |
1468 | } | |
1469 | *req = __rq; | |
1470 | return ret; | |
1471 | } | |
1472 | ||
1473 | static void as_merged_request(request_queue_t *q, struct request *req) | |
1474 | { | |
1475 | struct as_data *ad = q->elevator->elevator_data; | |
1476 | struct as_rq *arq = RQ_DATA(req); | |
1477 | ||
1478 | /* | |
1479 | * hash always needs to be repositioned, key is end sector | |
1480 | */ | |
1481 | as_del_arq_hash(arq); | |
1482 | as_add_arq_hash(ad, arq); | |
1483 | ||
1484 | /* | |
1485 | * if the merge was a front merge, we need to reposition request | |
1486 | */ | |
1487 | if (rq_rb_key(req) != arq->rb_key) { | |
1da177e4 | 1488 | as_del_arq_rb(ad, arq); |
ef9be1d3 | 1489 | as_add_arq_rb(ad, arq); |
1da177e4 LT |
1490 | /* |
1491 | * Note! At this stage of this and the next function, our next | |
1492 | * request may not be optimal - eg the request may have "grown" | |
1493 | * behind the disk head. We currently don't bother adjusting. | |
1494 | */ | |
1495 | } | |
1da177e4 LT |
1496 | } |
1497 | ||
f5b3db00 NP |
1498 | static void as_merged_requests(request_queue_t *q, struct request *req, |
1499 | struct request *next) | |
1da177e4 LT |
1500 | { |
1501 | struct as_data *ad = q->elevator->elevator_data; | |
1502 | struct as_rq *arq = RQ_DATA(req); | |
1503 | struct as_rq *anext = RQ_DATA(next); | |
1504 | ||
1505 | BUG_ON(!arq); | |
1506 | BUG_ON(!anext); | |
1507 | ||
1508 | /* | |
1509 | * reposition arq (this is the merged request) in hash, and in rbtree | |
1510 | * in case of a front merge | |
1511 | */ | |
1512 | as_del_arq_hash(arq); | |
1513 | as_add_arq_hash(ad, arq); | |
1514 | ||
1515 | if (rq_rb_key(req) != arq->rb_key) { | |
1da177e4 | 1516 | as_del_arq_rb(ad, arq); |
ef9be1d3 | 1517 | as_add_arq_rb(ad, arq); |
1da177e4 LT |
1518 | } |
1519 | ||
1520 | /* | |
1521 | * if anext expires before arq, assign its expire time to arq | |
1522 | * and move into anext position (anext will be deleted) in fifo | |
1523 | */ | |
1524 | if (!list_empty(&arq->fifo) && !list_empty(&anext->fifo)) { | |
1525 | if (time_before(anext->expires, arq->expires)) { | |
1526 | list_move(&arq->fifo, &anext->fifo); | |
1527 | arq->expires = anext->expires; | |
1528 | /* | |
1529 | * Don't copy here but swap, because when anext is | |
1530 | * removed below, it must contain the unused context | |
1531 | */ | |
1532 | swap_io_context(&arq->io_context, &anext->io_context); | |
1533 | } | |
1534 | } | |
1535 | ||
1da177e4 LT |
1536 | /* |
1537 | * kill knowledge of next, this one is a goner | |
1538 | */ | |
1539 | as_remove_queued_request(q, next); | |
b4878f24 | 1540 | as_put_io_context(anext); |
1da177e4 LT |
1541 | |
1542 | anext->state = AS_RQ_MERGED; | |
1543 | } | |
1544 | ||
1545 | /* | |
1546 | * This is executed in a "deferred" process context, by kblockd. It calls the | |
1547 | * driver's request_fn so the driver can submit that request. | |
1548 | * | |
1549 | * IMPORTANT! This guy will reenter the elevator, so set up all queue global | |
1550 | * state before calling, and don't rely on any state over calls. | |
1551 | * | |
1552 | * FIXME! dispatch queue is not a queue at all! | |
1553 | */ | |
1554 | static void as_work_handler(void *data) | |
1555 | { | |
1556 | struct request_queue *q = data; | |
1557 | unsigned long flags; | |
1558 | ||
1559 | spin_lock_irqsave(q->queue_lock, flags); | |
b4878f24 | 1560 | if (!as_queue_empty(q)) |
1da177e4 LT |
1561 | q->request_fn(q); |
1562 | spin_unlock_irqrestore(q->queue_lock, flags); | |
1563 | } | |
1564 | ||
1565 | static void as_put_request(request_queue_t *q, struct request *rq) | |
1566 | { | |
1567 | struct as_data *ad = q->elevator->elevator_data; | |
1568 | struct as_rq *arq = RQ_DATA(rq); | |
1569 | ||
1570 | if (!arq) { | |
1571 | WARN_ON(1); | |
1572 | return; | |
1573 | } | |
1574 | ||
b4878f24 JA |
1575 | if (unlikely(arq->state != AS_RQ_POSTSCHED && |
1576 | arq->state != AS_RQ_PRESCHED && | |
1577 | arq->state != AS_RQ_MERGED)) { | |
1da177e4 LT |
1578 | printk("arq->state %d\n", arq->state); |
1579 | WARN_ON(1); | |
1580 | } | |
1581 | ||
1582 | mempool_free(arq, ad->arq_pool); | |
1583 | rq->elevator_private = NULL; | |
1584 | } | |
1585 | ||
22e2c507 | 1586 | static int as_set_request(request_queue_t *q, struct request *rq, |
8267e268 | 1587 | struct bio *bio, gfp_t gfp_mask) |
1da177e4 LT |
1588 | { |
1589 | struct as_data *ad = q->elevator->elevator_data; | |
1590 | struct as_rq *arq = mempool_alloc(ad->arq_pool, gfp_mask); | |
1591 | ||
1592 | if (arq) { | |
1593 | memset(arq, 0, sizeof(*arq)); | |
1594 | RB_CLEAR(&arq->rb_node); | |
1595 | arq->request = rq; | |
1596 | arq->state = AS_RQ_PRESCHED; | |
1597 | arq->io_context = NULL; | |
bae386f7 | 1598 | INIT_HLIST_NODE(&arq->hash); |
1da177e4 LT |
1599 | INIT_LIST_HEAD(&arq->fifo); |
1600 | rq->elevator_private = arq; | |
1601 | return 0; | |
1602 | } | |
1603 | ||
1604 | return 1; | |
1605 | } | |
1606 | ||
22e2c507 | 1607 | static int as_may_queue(request_queue_t *q, int rw, struct bio *bio) |
1da177e4 LT |
1608 | { |
1609 | int ret = ELV_MQUEUE_MAY; | |
1610 | struct as_data *ad = q->elevator->elevator_data; | |
1611 | struct io_context *ioc; | |
1612 | if (ad->antic_status == ANTIC_WAIT_REQ || | |
1613 | ad->antic_status == ANTIC_WAIT_NEXT) { | |
1614 | ioc = as_get_io_context(); | |
1615 | if (ad->io_context == ioc) | |
1616 | ret = ELV_MQUEUE_MUST; | |
1617 | put_io_context(ioc); | |
1618 | } | |
1619 | ||
1620 | return ret; | |
1621 | } | |
1622 | ||
1623 | static void as_exit_queue(elevator_t *e) | |
1624 | { | |
1625 | struct as_data *ad = e->elevator_data; | |
1626 | ||
1627 | del_timer_sync(&ad->antic_timer); | |
1628 | kblockd_flush(); | |
1629 | ||
1630 | BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC])); | |
1631 | BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC])); | |
1632 | ||
1633 | mempool_destroy(ad->arq_pool); | |
1634 | put_io_context(ad->io_context); | |
1635 | kfree(ad->hash); | |
1636 | kfree(ad); | |
1637 | } | |
1638 | ||
1639 | /* | |
1640 | * initialize elevator private data (as_data), and alloc a arq for | |
1641 | * each request on the free lists | |
1642 | */ | |
bc1c1169 | 1643 | static void *as_init_queue(request_queue_t *q, elevator_t *e) |
1da177e4 LT |
1644 | { |
1645 | struct as_data *ad; | |
1646 | int i; | |
1647 | ||
1648 | if (!arq_pool) | |
bc1c1169 | 1649 | return NULL; |
1da177e4 | 1650 | |
1946089a | 1651 | ad = kmalloc_node(sizeof(*ad), GFP_KERNEL, q->node); |
1da177e4 | 1652 | if (!ad) |
bc1c1169 | 1653 | return NULL; |
1da177e4 LT |
1654 | memset(ad, 0, sizeof(*ad)); |
1655 | ||
1656 | ad->q = q; /* Identify what queue the data belongs to */ | |
1657 | ||
bae386f7 | 1658 | ad->hash = kmalloc_node(sizeof(struct hlist_head)*AS_HASH_ENTRIES, |
1946089a | 1659 | GFP_KERNEL, q->node); |
1da177e4 LT |
1660 | if (!ad->hash) { |
1661 | kfree(ad); | |
bc1c1169 | 1662 | return NULL; |
1da177e4 LT |
1663 | } |
1664 | ||
1946089a CL |
1665 | ad->arq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab, |
1666 | mempool_free_slab, arq_pool, q->node); | |
1da177e4 LT |
1667 | if (!ad->arq_pool) { |
1668 | kfree(ad->hash); | |
1669 | kfree(ad); | |
bc1c1169 | 1670 | return NULL; |
1da177e4 LT |
1671 | } |
1672 | ||
1673 | /* anticipatory scheduling helpers */ | |
1674 | ad->antic_timer.function = as_antic_timeout; | |
1675 | ad->antic_timer.data = (unsigned long)q; | |
1676 | init_timer(&ad->antic_timer); | |
1677 | INIT_WORK(&ad->antic_work, as_work_handler, q); | |
1678 | ||
1679 | for (i = 0; i < AS_HASH_ENTRIES; i++) | |
bae386f7 | 1680 | INIT_HLIST_HEAD(&ad->hash[i]); |
1da177e4 LT |
1681 | |
1682 | INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]); | |
1683 | INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]); | |
1684 | ad->sort_list[REQ_SYNC] = RB_ROOT; | |
1685 | ad->sort_list[REQ_ASYNC] = RB_ROOT; | |
1da177e4 LT |
1686 | ad->fifo_expire[REQ_SYNC] = default_read_expire; |
1687 | ad->fifo_expire[REQ_ASYNC] = default_write_expire; | |
1688 | ad->antic_expire = default_antic_expire; | |
1689 | ad->batch_expire[REQ_SYNC] = default_read_batch_expire; | |
1690 | ad->batch_expire[REQ_ASYNC] = default_write_batch_expire; | |
1da177e4 LT |
1691 | |
1692 | ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC]; | |
1693 | ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10; | |
1694 | if (ad->write_batch_count < 2) | |
1695 | ad->write_batch_count = 2; | |
1696 | ||
bc1c1169 | 1697 | return ad; |
1da177e4 LT |
1698 | } |
1699 | ||
1700 | /* | |
1701 | * sysfs parts below | |
1702 | */ | |
1da177e4 LT |
1703 | |
1704 | static ssize_t | |
1705 | as_var_show(unsigned int var, char *page) | |
1706 | { | |
1da177e4 LT |
1707 | return sprintf(page, "%d\n", var); |
1708 | } | |
1709 | ||
1710 | static ssize_t | |
1711 | as_var_store(unsigned long *var, const char *page, size_t count) | |
1712 | { | |
1da177e4 LT |
1713 | char *p = (char *) page; |
1714 | ||
c9b3ad67 | 1715 | *var = simple_strtoul(p, &p, 10); |
1da177e4 LT |
1716 | return count; |
1717 | } | |
1718 | ||
e572ec7e | 1719 | static ssize_t est_time_show(elevator_t *e, char *page) |
1da177e4 | 1720 | { |
3d1ab40f | 1721 | struct as_data *ad = e->elevator_data; |
1da177e4 LT |
1722 | int pos = 0; |
1723 | ||
f5b3db00 NP |
1724 | pos += sprintf(page+pos, "%lu %% exit probability\n", |
1725 | 100*ad->exit_prob/256); | |
1726 | pos += sprintf(page+pos, "%lu %% probability of exiting without a " | |
1727 | "cooperating process submitting IO\n", | |
1728 | 100*ad->exit_no_coop/256); | |
1da177e4 | 1729 | pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean); |
f5b3db00 NP |
1730 | pos += sprintf(page+pos, "%llu sectors new seek distance\n", |
1731 | (unsigned long long)ad->new_seek_mean); | |
1da177e4 LT |
1732 | |
1733 | return pos; | |
1734 | } | |
1735 | ||
1736 | #define SHOW_FUNCTION(__FUNC, __VAR) \ | |
3d1ab40f | 1737 | static ssize_t __FUNC(elevator_t *e, char *page) \ |
1da177e4 | 1738 | { \ |
3d1ab40f | 1739 | struct as_data *ad = e->elevator_data; \ |
1da177e4 LT |
1740 | return as_var_show(jiffies_to_msecs((__VAR)), (page)); \ |
1741 | } | |
e572ec7e AV |
1742 | SHOW_FUNCTION(as_read_expire_show, ad->fifo_expire[REQ_SYNC]); |
1743 | SHOW_FUNCTION(as_write_expire_show, ad->fifo_expire[REQ_ASYNC]); | |
1744 | SHOW_FUNCTION(as_antic_expire_show, ad->antic_expire); | |
1745 | SHOW_FUNCTION(as_read_batch_expire_show, ad->batch_expire[REQ_SYNC]); | |
1746 | SHOW_FUNCTION(as_write_batch_expire_show, ad->batch_expire[REQ_ASYNC]); | |
1da177e4 LT |
1747 | #undef SHOW_FUNCTION |
1748 | ||
1749 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \ | |
3d1ab40f | 1750 | static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \ |
1da177e4 | 1751 | { \ |
3d1ab40f AV |
1752 | struct as_data *ad = e->elevator_data; \ |
1753 | int ret = as_var_store(__PTR, (page), count); \ | |
1da177e4 LT |
1754 | if (*(__PTR) < (MIN)) \ |
1755 | *(__PTR) = (MIN); \ | |
1756 | else if (*(__PTR) > (MAX)) \ | |
1757 | *(__PTR) = (MAX); \ | |
1758 | *(__PTR) = msecs_to_jiffies(*(__PTR)); \ | |
1759 | return ret; \ | |
1760 | } | |
e572ec7e AV |
1761 | STORE_FUNCTION(as_read_expire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX); |
1762 | STORE_FUNCTION(as_write_expire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX); | |
1763 | STORE_FUNCTION(as_antic_expire_store, &ad->antic_expire, 0, INT_MAX); | |
1764 | STORE_FUNCTION(as_read_batch_expire_store, | |
1da177e4 | 1765 | &ad->batch_expire[REQ_SYNC], 0, INT_MAX); |
e572ec7e | 1766 | STORE_FUNCTION(as_write_batch_expire_store, |
1da177e4 LT |
1767 | &ad->batch_expire[REQ_ASYNC], 0, INT_MAX); |
1768 | #undef STORE_FUNCTION | |
1769 | ||
e572ec7e AV |
1770 | #define AS_ATTR(name) \ |
1771 | __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store) | |
1772 | ||
1773 | static struct elv_fs_entry as_attrs[] = { | |
1774 | __ATTR_RO(est_time), | |
1775 | AS_ATTR(read_expire), | |
1776 | AS_ATTR(write_expire), | |
1777 | AS_ATTR(antic_expire), | |
1778 | AS_ATTR(read_batch_expire), | |
1779 | AS_ATTR(write_batch_expire), | |
1780 | __ATTR_NULL | |
1da177e4 LT |
1781 | }; |
1782 | ||
1da177e4 LT |
1783 | static struct elevator_type iosched_as = { |
1784 | .ops = { | |
1785 | .elevator_merge_fn = as_merge, | |
1786 | .elevator_merged_fn = as_merged_request, | |
1787 | .elevator_merge_req_fn = as_merged_requests, | |
b4878f24 JA |
1788 | .elevator_dispatch_fn = as_dispatch_request, |
1789 | .elevator_add_req_fn = as_add_request, | |
1790 | .elevator_activate_req_fn = as_activate_request, | |
1da177e4 LT |
1791 | .elevator_deactivate_req_fn = as_deactivate_request, |
1792 | .elevator_queue_empty_fn = as_queue_empty, | |
1793 | .elevator_completed_req_fn = as_completed_request, | |
1794 | .elevator_former_req_fn = as_former_request, | |
1795 | .elevator_latter_req_fn = as_latter_request, | |
1796 | .elevator_set_req_fn = as_set_request, | |
1797 | .elevator_put_req_fn = as_put_request, | |
1798 | .elevator_may_queue_fn = as_may_queue, | |
1799 | .elevator_init_fn = as_init_queue, | |
1800 | .elevator_exit_fn = as_exit_queue, | |
e17a9489 | 1801 | .trim = as_trim, |
1da177e4 LT |
1802 | }, |
1803 | ||
3d1ab40f | 1804 | .elevator_attrs = as_attrs, |
1da177e4 LT |
1805 | .elevator_name = "anticipatory", |
1806 | .elevator_owner = THIS_MODULE, | |
1807 | }; | |
1808 | ||
1809 | static int __init as_init(void) | |
1810 | { | |
1811 | int ret; | |
1812 | ||
1813 | arq_pool = kmem_cache_create("as_arq", sizeof(struct as_rq), | |
1814 | 0, 0, NULL, NULL); | |
1815 | if (!arq_pool) | |
1816 | return -ENOMEM; | |
1817 | ||
1818 | ret = elv_register(&iosched_as); | |
1819 | if (!ret) { | |
1820 | /* | |
1821 | * don't allow AS to get unregistered, since we would have | |
1822 | * to browse all tasks in the system and release their | |
1823 | * as_io_context first | |
1824 | */ | |
1825 | __module_get(THIS_MODULE); | |
1826 | return 0; | |
1827 | } | |
1828 | ||
1829 | kmem_cache_destroy(arq_pool); | |
1830 | return ret; | |
1831 | } | |
1832 | ||
1833 | static void __exit as_exit(void) | |
1834 | { | |
334e94de | 1835 | DECLARE_COMPLETION(all_gone); |
1da177e4 | 1836 | elv_unregister(&iosched_as); |
334e94de | 1837 | ioc_gone = &all_gone; |
fba82272 OH |
1838 | /* ioc_gone's update must be visible before reading ioc_count */ |
1839 | smp_wmb(); | |
334e94de | 1840 | if (atomic_read(&ioc_count)) |
fba82272 | 1841 | wait_for_completion(ioc_gone); |
334e94de | 1842 | synchronize_rcu(); |
83521d3e | 1843 | kmem_cache_destroy(arq_pool); |
1da177e4 LT |
1844 | } |
1845 | ||
1846 | module_init(as_init); | |
1847 | module_exit(as_exit); | |
1848 | ||
1849 | MODULE_AUTHOR("Nick Piggin"); | |
1850 | MODULE_LICENSE("GPL"); | |
1851 | MODULE_DESCRIPTION("anticipatory IO scheduler"); |