ia64/linux-2.6.18-xen.hg

view block/as-iosched.c @ 890:2e94884f0e8d

Dom0 PCI: fix a regression introduced by the SR-IOV change

The device class may be changed during the early fixup. So need to
re-read the device class from pci_dev after the fixup.

The patch "PCI: centralize device setup code" (c/s 825) wrongly
cleaned up the device class re-read. This patch reverts that change.

Signed-off-by: Yu Zhao <yu.zhao@intel.com>
author Keir Fraser <keir.fraser@citrix.com>
date Wed Jun 03 11:21:52 2009 +0100 (2009-06-03)
parents 831230e53067
children
line source
1 /*
2 * Anticipatory & deadline i/o scheduler.
3 *
4 * Copyright (C) 2002 Jens Axboe <axboe@suse.de>
5 * Nick Piggin <nickpiggin@yahoo.com.au>
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/module.h>
14 #include <linux/slab.h>
15 #include <linux/init.h>
16 #include <linux/compiler.h>
17 #include <linux/hash.h>
18 #include <linux/rbtree.h>
19 #include <linux/interrupt.h>
21 #define REQ_SYNC 1
22 #define REQ_ASYNC 0
24 /*
25 * See Documentation/block/as-iosched.txt
26 */
28 /*
29 * max time before a read is submitted.
30 */
31 #define default_read_expire (HZ / 8)
33 /*
34 * ditto for writes, these limits are not hard, even
35 * if the disk is capable of satisfying them.
36 */
37 #define default_write_expire (HZ / 4)
39 /*
40 * read_batch_expire describes how long we will allow a stream of reads to
41 * persist before looking to see whether it is time to switch over to writes.
42 */
43 #define default_read_batch_expire (HZ / 2)
45 /*
46 * write_batch_expire describes how long we want a stream of writes to run for.
47 * This is not a hard limit, but a target we set for the auto-tuning thingy.
48 * See, the problem is: we can send a lot of writes to disk cache / TCQ in
49 * a short amount of time...
50 */
51 #define default_write_batch_expire (HZ / 8)
53 /*
54 * max time we may wait to anticipate a read (default around 6ms)
55 */
56 #define default_antic_expire ((HZ / 150) ? HZ / 150 : 1)
58 /*
59 * Keep track of up to 20ms thinktimes. We can go as big as we like here,
60 * however huge values tend to interfere and not decay fast enough. A program
61 * might be in a non-io phase of operation. Waiting on user input for example,
62 * or doing a lengthy computation. A small penalty can be justified there, and
63 * will still catch out those processes that constantly have large thinktimes.
64 */
65 #define MAX_THINKTIME (HZ/50UL)
67 /* Bits in as_io_context.state */
68 enum as_io_states {
69 AS_TASK_RUNNING=0, /* Process has not exited */
70 AS_TASK_IOSTARTED, /* Process has started some IO */
71 AS_TASK_IORUNNING, /* Process has completed some IO */
72 };
74 enum anticipation_status {
75 ANTIC_OFF=0, /* Not anticipating (normal operation) */
76 ANTIC_WAIT_REQ, /* The last read has not yet completed */
77 ANTIC_WAIT_NEXT, /* Currently anticipating a request vs
78 last read (which has completed) */
79 ANTIC_FINISHED, /* Anticipating but have found a candidate
80 * or timed out */
81 };
83 struct as_data {
84 /*
85 * run time data
86 */
88 struct request_queue *q; /* the "owner" queue */
90 /*
91 * requests (as_rq s) are present on both sort_list and fifo_list
92 */
93 struct rb_root sort_list[2];
94 struct list_head fifo_list[2];
96 struct as_rq *next_arq[2]; /* next in sort order */
97 sector_t last_sector[2]; /* last REQ_SYNC & REQ_ASYNC sectors */
98 struct hlist_head *hash; /* request hash */
100 unsigned long exit_prob; /* probability a task will exit while
101 being waited on */
102 unsigned long exit_no_coop; /* probablility an exited task will
103 not be part of a later cooperating
104 request */
105 unsigned long new_ttime_total; /* mean thinktime on new proc */
106 unsigned long new_ttime_mean;
107 u64 new_seek_total; /* mean seek on new proc */
108 sector_t new_seek_mean;
110 unsigned long current_batch_expires;
111 unsigned long last_check_fifo[2];
112 int changed_batch; /* 1: waiting for old batch to end */
113 int new_batch; /* 1: waiting on first read complete */
114 int batch_data_dir; /* current batch REQ_SYNC / REQ_ASYNC */
115 int write_batch_count; /* max # of reqs in a write batch */
116 int current_write_count; /* how many requests left this batch */
117 int write_batch_idled; /* has the write batch gone idle? */
118 mempool_t *arq_pool;
120 enum anticipation_status antic_status;
121 unsigned long antic_start; /* jiffies: when it started */
122 struct timer_list antic_timer; /* anticipatory scheduling timer */
123 struct work_struct antic_work; /* Deferred unplugging */
124 struct io_context *io_context; /* Identify the expected process */
125 int ioc_finished; /* IO associated with io_context is finished */
126 int nr_dispatched;
128 /*
129 * settings that change how the i/o scheduler behaves
130 */
131 unsigned long fifo_expire[2];
132 unsigned long batch_expire[2];
133 unsigned long antic_expire;
134 };
136 #define list_entry_fifo(ptr) list_entry((ptr), struct as_rq, fifo)
138 /*
139 * per-request data.
140 */
141 enum arq_state {
142 AS_RQ_NEW=0, /* New - not referenced and not on any lists */
143 AS_RQ_QUEUED, /* In the request queue. It belongs to the
144 scheduler */
145 AS_RQ_DISPATCHED, /* On the dispatch list. It belongs to the
146 driver now */
147 AS_RQ_PRESCHED, /* Debug poisoning for requests being used */
148 AS_RQ_REMOVED,
149 AS_RQ_MERGED,
150 AS_RQ_POSTSCHED, /* when they shouldn't be */
151 };
153 struct as_rq {
154 /*
155 * rbtree index, key is the starting offset
156 */
157 struct rb_node rb_node;
158 sector_t rb_key;
160 struct request *request;
162 struct io_context *io_context; /* The submitting task */
164 /*
165 * request hash, key is the ending offset (for back merge lookup)
166 */
167 struct hlist_node hash;
169 /*
170 * expire fifo
171 */
172 struct list_head fifo;
173 unsigned long expires;
175 unsigned int is_sync;
176 enum arq_state state;
177 };
179 #define RQ_DATA(rq) ((struct as_rq *) (rq)->elevator_private)
181 static kmem_cache_t *arq_pool;
183 static atomic_t ioc_count = ATOMIC_INIT(0);
184 static struct completion *ioc_gone;
186 static void as_move_to_dispatch(struct as_data *ad, struct as_rq *arq);
187 static void as_antic_stop(struct as_data *ad);
189 /*
190 * IO Context helper functions
191 */
193 /* Called to deallocate the as_io_context */
194 static void free_as_io_context(struct as_io_context *aic)
195 {
196 kfree(aic);
197 if (atomic_dec_and_test(&ioc_count) && ioc_gone)
198 complete(ioc_gone);
199 }
201 static void as_trim(struct io_context *ioc)
202 {
203 if (ioc->aic)
204 free_as_io_context(ioc->aic);
205 ioc->aic = NULL;
206 }
208 /* Called when the task exits */
209 static void exit_as_io_context(struct as_io_context *aic)
210 {
211 WARN_ON(!test_bit(AS_TASK_RUNNING, &aic->state));
212 clear_bit(AS_TASK_RUNNING, &aic->state);
213 }
215 static struct as_io_context *alloc_as_io_context(void)
216 {
217 struct as_io_context *ret;
219 ret = kmalloc(sizeof(*ret), GFP_ATOMIC);
220 if (ret) {
221 ret->dtor = free_as_io_context;
222 ret->exit = exit_as_io_context;
223 ret->state = 1 << AS_TASK_RUNNING;
224 atomic_set(&ret->nr_queued, 0);
225 atomic_set(&ret->nr_dispatched, 0);
226 spin_lock_init(&ret->lock);
227 ret->ttime_total = 0;
228 ret->ttime_samples = 0;
229 ret->ttime_mean = 0;
230 ret->seek_total = 0;
231 ret->seek_samples = 0;
232 ret->seek_mean = 0;
233 atomic_inc(&ioc_count);
234 }
236 return ret;
237 }
239 /*
240 * If the current task has no AS IO context then create one and initialise it.
241 * Then take a ref on the task's io context and return it.
242 */
243 static struct io_context *as_get_io_context(void)
244 {
245 struct io_context *ioc = get_io_context(GFP_ATOMIC);
246 if (ioc && !ioc->aic) {
247 ioc->aic = alloc_as_io_context();
248 if (!ioc->aic) {
249 put_io_context(ioc);
250 ioc = NULL;
251 }
252 }
253 return ioc;
254 }
256 static void as_put_io_context(struct as_rq *arq)
257 {
258 struct as_io_context *aic;
260 if (unlikely(!arq->io_context))
261 return;
263 aic = arq->io_context->aic;
265 if (arq->is_sync == REQ_SYNC && aic) {
266 spin_lock(&aic->lock);
267 set_bit(AS_TASK_IORUNNING, &aic->state);
268 aic->last_end_request = jiffies;
269 spin_unlock(&aic->lock);
270 }
272 put_io_context(arq->io_context);
273 }
275 /*
276 * the back merge hash support functions
277 */
278 static const int as_hash_shift = 6;
279 #define AS_HASH_BLOCK(sec) ((sec) >> 3)
280 #define AS_HASH_FN(sec) (hash_long(AS_HASH_BLOCK((sec)), as_hash_shift))
281 #define AS_HASH_ENTRIES (1 << as_hash_shift)
282 #define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors)
284 static inline void __as_del_arq_hash(struct as_rq *arq)
285 {
286 hlist_del_init(&arq->hash);
287 }
289 static inline void as_del_arq_hash(struct as_rq *arq)
290 {
291 if (!hlist_unhashed(&arq->hash))
292 __as_del_arq_hash(arq);
293 }
295 static void as_add_arq_hash(struct as_data *ad, struct as_rq *arq)
296 {
297 struct request *rq = arq->request;
299 BUG_ON(!hlist_unhashed(&arq->hash));
301 hlist_add_head(&arq->hash, &ad->hash[AS_HASH_FN(rq_hash_key(rq))]);
302 }
304 /*
305 * move hot entry to front of chain
306 */
307 static inline void as_hot_arq_hash(struct as_data *ad, struct as_rq *arq)
308 {
309 struct request *rq = arq->request;
310 struct hlist_head *head = &ad->hash[AS_HASH_FN(rq_hash_key(rq))];
312 if (hlist_unhashed(&arq->hash)) {
313 WARN_ON(1);
314 return;
315 }
317 if (&arq->hash != head->first) {
318 hlist_del(&arq->hash);
319 hlist_add_head(&arq->hash, head);
320 }
321 }
323 static struct request *as_find_arq_hash(struct as_data *ad, sector_t offset)
324 {
325 struct hlist_head *hash_list = &ad->hash[AS_HASH_FN(offset)];
326 struct hlist_node *entry, *next;
327 struct as_rq *arq;
329 hlist_for_each_entry_safe(arq, entry, next, hash_list, hash) {
330 struct request *__rq = arq->request;
332 BUG_ON(hlist_unhashed(&arq->hash));
334 if (!rq_mergeable(__rq)) {
335 as_del_arq_hash(arq);
336 continue;
337 }
339 if (rq_hash_key(__rq) == offset)
340 return __rq;
341 }
343 return NULL;
344 }
346 /*
347 * rb tree support functions
348 */
349 #define rb_entry_arq(node) rb_entry((node), struct as_rq, rb_node)
350 #define ARQ_RB_ROOT(ad, arq) (&(ad)->sort_list[(arq)->is_sync])
351 #define rq_rb_key(rq) (rq)->sector
353 /*
354 * as_find_first_arq finds the first (lowest sector numbered) request
355 * for the specified data_dir. Used to sweep back to the start of the disk
356 * (1-way elevator) after we process the last (highest sector) request.
357 */
358 static struct as_rq *as_find_first_arq(struct as_data *ad, int data_dir)
359 {
360 struct rb_node *n = ad->sort_list[data_dir].rb_node;
362 if (n == NULL)
363 return NULL;
365 for (;;) {
366 if (n->rb_left == NULL)
367 return rb_entry_arq(n);
369 n = n->rb_left;
370 }
371 }
373 /*
374 * Add the request to the rb tree if it is unique. If there is an alias (an
375 * existing request against the same sector), which can happen when using
376 * direct IO, then return the alias.
377 */
378 static struct as_rq *__as_add_arq_rb(struct as_data *ad, struct as_rq *arq)
379 {
380 struct rb_node **p = &ARQ_RB_ROOT(ad, arq)->rb_node;
381 struct rb_node *parent = NULL;
382 struct as_rq *__arq;
383 struct request *rq = arq->request;
385 arq->rb_key = rq_rb_key(rq);
387 while (*p) {
388 parent = *p;
389 __arq = rb_entry_arq(parent);
391 if (arq->rb_key < __arq->rb_key)
392 p = &(*p)->rb_left;
393 else if (arq->rb_key > __arq->rb_key)
394 p = &(*p)->rb_right;
395 else
396 return __arq;
397 }
399 rb_link_node(&arq->rb_node, parent, p);
400 rb_insert_color(&arq->rb_node, ARQ_RB_ROOT(ad, arq));
402 return NULL;
403 }
405 static void as_add_arq_rb(struct as_data *ad, struct as_rq *arq)
406 {
407 struct as_rq *alias;
409 while ((unlikely(alias = __as_add_arq_rb(ad, arq)))) {
410 as_move_to_dispatch(ad, alias);
411 as_antic_stop(ad);
412 }
413 }
415 static inline void as_del_arq_rb(struct as_data *ad, struct as_rq *arq)
416 {
417 if (!RB_EMPTY_NODE(&arq->rb_node)) {
418 WARN_ON(1);
419 return;
420 }
422 rb_erase(&arq->rb_node, ARQ_RB_ROOT(ad, arq));
423 RB_CLEAR_NODE(&arq->rb_node);
424 }
426 static struct request *
427 as_find_arq_rb(struct as_data *ad, sector_t sector, int data_dir)
428 {
429 struct rb_node *n = ad->sort_list[data_dir].rb_node;
430 struct as_rq *arq;
432 while (n) {
433 arq = rb_entry_arq(n);
435 if (sector < arq->rb_key)
436 n = n->rb_left;
437 else if (sector > arq->rb_key)
438 n = n->rb_right;
439 else
440 return arq->request;
441 }
443 return NULL;
444 }
446 /*
447 * IO Scheduler proper
448 */
450 #define MAXBACK (1024 * 1024) /*
451 * Maximum distance the disk will go backward
452 * for a request.
453 */
455 #define BACK_PENALTY 2
457 /*
458 * as_choose_req selects the preferred one of two requests of the same data_dir
459 * ignoring time - eg. timeouts, which is the job of as_dispatch_request
460 */
461 static struct as_rq *
462 as_choose_req(struct as_data *ad, struct as_rq *arq1, struct as_rq *arq2)
463 {
464 int data_dir;
465 sector_t last, s1, s2, d1, d2;
466 int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */
467 const sector_t maxback = MAXBACK;
469 if (arq1 == NULL || arq1 == arq2)
470 return arq2;
471 if (arq2 == NULL)
472 return arq1;
474 data_dir = arq1->is_sync;
476 last = ad->last_sector[data_dir];
477 s1 = arq1->request->sector;
478 s2 = arq2->request->sector;
480 BUG_ON(data_dir != arq2->is_sync);
482 /*
483 * Strict one way elevator _except_ in the case where we allow
484 * short backward seeks which are biased as twice the cost of a
485 * similar forward seek.
486 */
487 if (s1 >= last)
488 d1 = s1 - last;
489 else if (s1+maxback >= last)
490 d1 = (last - s1)*BACK_PENALTY;
491 else {
492 r1_wrap = 1;
493 d1 = 0; /* shut up, gcc */
494 }
496 if (s2 >= last)
497 d2 = s2 - last;
498 else if (s2+maxback >= last)
499 d2 = (last - s2)*BACK_PENALTY;
500 else {
501 r2_wrap = 1;
502 d2 = 0;
503 }
505 /* Found required data */
506 if (!r1_wrap && r2_wrap)
507 return arq1;
508 else if (!r2_wrap && r1_wrap)
509 return arq2;
510 else if (r1_wrap && r2_wrap) {
511 /* both behind the head */
512 if (s1 <= s2)
513 return arq1;
514 else
515 return arq2;
516 }
518 /* Both requests in front of the head */
519 if (d1 < d2)
520 return arq1;
521 else if (d2 < d1)
522 return arq2;
523 else {
524 if (s1 >= s2)
525 return arq1;
526 else
527 return arq2;
528 }
529 }
531 /*
532 * as_find_next_arq finds the next request after @prev in elevator order.
533 * this with as_choose_req form the basis for how the scheduler chooses
534 * what request to process next. Anticipation works on top of this.
535 */
536 static struct as_rq *as_find_next_arq(struct as_data *ad, struct as_rq *last)
537 {
538 const int data_dir = last->is_sync;
539 struct as_rq *ret;
540 struct rb_node *rbnext = rb_next(&last->rb_node);
541 struct rb_node *rbprev = rb_prev(&last->rb_node);
542 struct as_rq *arq_next, *arq_prev;
544 BUG_ON(!RB_EMPTY_NODE(&last->rb_node));
546 if (rbprev)
547 arq_prev = rb_entry_arq(rbprev);
548 else
549 arq_prev = NULL;
551 if (rbnext)
552 arq_next = rb_entry_arq(rbnext);
553 else {
554 arq_next = as_find_first_arq(ad, data_dir);
555 if (arq_next == last)
556 arq_next = NULL;
557 }
559 ret = as_choose_req(ad, arq_next, arq_prev);
561 return ret;
562 }
564 /*
565 * anticipatory scheduling functions follow
566 */
568 /*
569 * as_antic_expired tells us when we have anticipated too long.
570 * The funny "absolute difference" math on the elapsed time is to handle
571 * jiffy wraps, and disks which have been idle for 0x80000000 jiffies.
572 */
573 static int as_antic_expired(struct as_data *ad)
574 {
575 long delta_jif;
577 delta_jif = jiffies - ad->antic_start;
578 if (unlikely(delta_jif < 0))
579 delta_jif = -delta_jif;
580 if (delta_jif < ad->antic_expire)
581 return 0;
583 return 1;
584 }
586 /*
587 * as_antic_waitnext starts anticipating that a nice request will soon be
588 * submitted. See also as_antic_waitreq
589 */
590 static void as_antic_waitnext(struct as_data *ad)
591 {
592 unsigned long timeout;
594 BUG_ON(ad->antic_status != ANTIC_OFF
595 && ad->antic_status != ANTIC_WAIT_REQ);
597 timeout = ad->antic_start + ad->antic_expire;
599 mod_timer(&ad->antic_timer, timeout);
601 ad->antic_status = ANTIC_WAIT_NEXT;
602 }
604 /*
605 * as_antic_waitreq starts anticipating. We don't start timing the anticipation
606 * until the request that we're anticipating on has finished. This means we
607 * are timing from when the candidate process wakes up hopefully.
608 */
609 static void as_antic_waitreq(struct as_data *ad)
610 {
611 BUG_ON(ad->antic_status == ANTIC_FINISHED);
612 if (ad->antic_status == ANTIC_OFF) {
613 if (!ad->io_context || ad->ioc_finished)
614 as_antic_waitnext(ad);
615 else
616 ad->antic_status = ANTIC_WAIT_REQ;
617 }
618 }
620 /*
621 * This is called directly by the functions in this file to stop anticipation.
622 * We kill the timer and schedule a call to the request_fn asap.
623 */
624 static void as_antic_stop(struct as_data *ad)
625 {
626 int status = ad->antic_status;
628 if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) {
629 if (status == ANTIC_WAIT_NEXT)
630 del_timer(&ad->antic_timer);
631 ad->antic_status = ANTIC_FINISHED;
632 /* see as_work_handler */
633 kblockd_schedule_work(&ad->antic_work);
634 }
635 }
637 /*
638 * as_antic_timeout is the timer function set by as_antic_waitnext.
639 */
640 static void as_antic_timeout(unsigned long data)
641 {
642 struct request_queue *q = (struct request_queue *)data;
643 struct as_data *ad = q->elevator->elevator_data;
644 unsigned long flags;
646 spin_lock_irqsave(q->queue_lock, flags);
647 if (ad->antic_status == ANTIC_WAIT_REQ
648 || ad->antic_status == ANTIC_WAIT_NEXT) {
649 struct as_io_context *aic = ad->io_context->aic;
651 ad->antic_status = ANTIC_FINISHED;
652 kblockd_schedule_work(&ad->antic_work);
654 if (aic->ttime_samples == 0) {
655 /* process anticipated on has exited or timed out*/
656 ad->exit_prob = (7*ad->exit_prob + 256)/8;
657 }
658 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
659 /* process not "saved" by a cooperating request */
660 ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8;
661 }
662 }
663 spin_unlock_irqrestore(q->queue_lock, flags);
664 }
666 static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic,
667 unsigned long ttime)
668 {
669 /* fixed point: 1.0 == 1<<8 */
670 if (aic->ttime_samples == 0) {
671 ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8;
672 ad->new_ttime_mean = ad->new_ttime_total / 256;
674 ad->exit_prob = (7*ad->exit_prob)/8;
675 }
676 aic->ttime_samples = (7*aic->ttime_samples + 256) / 8;
677 aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8;
678 aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples;
679 }
681 static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic,
682 sector_t sdist)
683 {
684 u64 total;
686 if (aic->seek_samples == 0) {
687 ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8;
688 ad->new_seek_mean = ad->new_seek_total / 256;
689 }
691 /*
692 * Don't allow the seek distance to get too large from the
693 * odd fragment, pagein, etc
694 */
695 if (aic->seek_samples <= 60) /* second&third seek */
696 sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024);
697 else
698 sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64);
700 aic->seek_samples = (7*aic->seek_samples + 256) / 8;
701 aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8;
702 total = aic->seek_total + (aic->seek_samples/2);
703 do_div(total, aic->seek_samples);
704 aic->seek_mean = (sector_t)total;
705 }
707 /*
708 * as_update_iohist keeps a decaying histogram of IO thinktimes, and
709 * updates @aic->ttime_mean based on that. It is called when a new
710 * request is queued.
711 */
712 static void as_update_iohist(struct as_data *ad, struct as_io_context *aic,
713 struct request *rq)
714 {
715 struct as_rq *arq = RQ_DATA(rq);
716 int data_dir = arq->is_sync;
717 unsigned long thinktime = 0;
718 sector_t seek_dist;
720 if (aic == NULL)
721 return;
723 if (data_dir == REQ_SYNC) {
724 unsigned long in_flight = atomic_read(&aic->nr_queued)
725 + atomic_read(&aic->nr_dispatched);
726 spin_lock(&aic->lock);
727 if (test_bit(AS_TASK_IORUNNING, &aic->state) ||
728 test_bit(AS_TASK_IOSTARTED, &aic->state)) {
729 /* Calculate read -> read thinktime */
730 if (test_bit(AS_TASK_IORUNNING, &aic->state)
731 && in_flight == 0) {
732 thinktime = jiffies - aic->last_end_request;
733 thinktime = min(thinktime, MAX_THINKTIME-1);
734 }
735 as_update_thinktime(ad, aic, thinktime);
737 /* Calculate read -> read seek distance */
738 if (aic->last_request_pos < rq->sector)
739 seek_dist = rq->sector - aic->last_request_pos;
740 else
741 seek_dist = aic->last_request_pos - rq->sector;
742 as_update_seekdist(ad, aic, seek_dist);
743 }
744 aic->last_request_pos = rq->sector + rq->nr_sectors;
745 set_bit(AS_TASK_IOSTARTED, &aic->state);
746 spin_unlock(&aic->lock);
747 }
748 }
750 /*
751 * as_close_req decides if one request is considered "close" to the
752 * previous one issued.
753 */
754 static int as_close_req(struct as_data *ad, struct as_io_context *aic,
755 struct as_rq *arq)
756 {
757 unsigned long delay; /* milliseconds */
758 sector_t last = ad->last_sector[ad->batch_data_dir];
759 sector_t next = arq->request->sector;
760 sector_t delta; /* acceptable close offset (in sectors) */
761 sector_t s;
763 if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished)
764 delay = 0;
765 else
766 delay = ((jiffies - ad->antic_start) * 1000) / HZ;
768 if (delay == 0)
769 delta = 8192;
770 else if (delay <= 20 && delay <= ad->antic_expire)
771 delta = 8192 << delay;
772 else
773 return 1;
775 if ((last <= next + (delta>>1)) && (next <= last + delta))
776 return 1;
778 if (last < next)
779 s = next - last;
780 else
781 s = last - next;
783 if (aic->seek_samples == 0) {
784 /*
785 * Process has just started IO. Use past statistics to
786 * gauge success possibility
787 */
788 if (ad->new_seek_mean > s) {
789 /* this request is better than what we're expecting */
790 return 1;
791 }
793 } else {
794 if (aic->seek_mean > s) {
795 /* this request is better than what we're expecting */
796 return 1;
797 }
798 }
800 return 0;
801 }
803 /*
804 * as_can_break_anticipation returns true if we have been anticipating this
805 * request.
806 *
807 * It also returns true if the process against which we are anticipating
808 * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to
809 * dispatch it ASAP, because we know that application will not be submitting
810 * any new reads.
811 *
812 * If the task which has submitted the request has exited, break anticipation.
813 *
814 * If this task has queued some other IO, do not enter enticipation.
815 */
816 static int as_can_break_anticipation(struct as_data *ad, struct as_rq *arq)
817 {
818 struct io_context *ioc;
819 struct as_io_context *aic;
821 ioc = ad->io_context;
822 BUG_ON(!ioc);
824 if (arq && ioc == arq->io_context) {
825 /* request from same process */
826 return 1;
827 }
829 if (ad->ioc_finished && as_antic_expired(ad)) {
830 /*
831 * In this situation status should really be FINISHED,
832 * however the timer hasn't had the chance to run yet.
833 */
834 return 1;
835 }
837 aic = ioc->aic;
838 if (!aic)
839 return 0;
841 if (atomic_read(&aic->nr_queued) > 0) {
842 /* process has more requests queued */
843 return 1;
844 }
846 if (atomic_read(&aic->nr_dispatched) > 0) {
847 /* process has more requests dispatched */
848 return 1;
849 }
851 if (arq && arq->is_sync == REQ_SYNC && as_close_req(ad, aic, arq)) {
852 /*
853 * Found a close request that is not one of ours.
854 *
855 * This makes close requests from another process update
856 * our IO history. Is generally useful when there are
857 * two or more cooperating processes working in the same
858 * area.
859 */
860 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
861 if (aic->ttime_samples == 0)
862 ad->exit_prob = (7*ad->exit_prob + 256)/8;
864 ad->exit_no_coop = (7*ad->exit_no_coop)/8;
865 }
867 as_update_iohist(ad, aic, arq->request);
868 return 1;
869 }
871 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
872 /* process anticipated on has exited */
873 if (aic->ttime_samples == 0)
874 ad->exit_prob = (7*ad->exit_prob + 256)/8;
876 if (ad->exit_no_coop > 128)
877 return 1;
878 }
880 if (aic->ttime_samples == 0) {
881 if (ad->new_ttime_mean > ad->antic_expire)
882 return 1;
883 if (ad->exit_prob * ad->exit_no_coop > 128*256)
884 return 1;
885 } else if (aic->ttime_mean > ad->antic_expire) {
886 /* the process thinks too much between requests */
887 return 1;
888 }
890 return 0;
891 }
893 /*
894 * as_can_anticipate indicates whether we should either run arq
895 * or keep anticipating a better request.
896 */
897 static int as_can_anticipate(struct as_data *ad, struct as_rq *arq)
898 {
899 if (!ad->io_context)
900 /*
901 * Last request submitted was a write
902 */
903 return 0;
905 if (ad->antic_status == ANTIC_FINISHED)
906 /*
907 * Don't restart if we have just finished. Run the next request
908 */
909 return 0;
911 if (as_can_break_anticipation(ad, arq))
912 /*
913 * This request is a good candidate. Don't keep anticipating,
914 * run it.
915 */
916 return 0;
918 /*
919 * OK from here, we haven't finished, and don't have a decent request!
920 * Status is either ANTIC_OFF so start waiting,
921 * ANTIC_WAIT_REQ so continue waiting for request to finish
922 * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request.
923 */
925 return 1;
926 }
928 /*
929 * as_update_arq must be called whenever a request (arq) is added to
930 * the sort_list. This function keeps caches up to date, and checks if the
931 * request might be one we are "anticipating"
932 */
933 static void as_update_arq(struct as_data *ad, struct as_rq *arq)
934 {
935 const int data_dir = arq->is_sync;
937 /* keep the next_arq cache up to date */
938 ad->next_arq[data_dir] = as_choose_req(ad, arq, ad->next_arq[data_dir]);
940 /*
941 * have we been anticipating this request?
942 * or does it come from the same process as the one we are anticipating
943 * for?
944 */
945 if (ad->antic_status == ANTIC_WAIT_REQ
946 || ad->antic_status == ANTIC_WAIT_NEXT) {
947 if (as_can_break_anticipation(ad, arq))
948 as_antic_stop(ad);
949 }
950 }
952 /*
953 * Gathers timings and resizes the write batch automatically
954 */
955 static void update_write_batch(struct as_data *ad)
956 {
957 unsigned long batch = ad->batch_expire[REQ_ASYNC];
958 long write_time;
960 write_time = (jiffies - ad->current_batch_expires) + batch;
961 if (write_time < 0)
962 write_time = 0;
964 if (write_time > batch && !ad->write_batch_idled) {
965 if (write_time > batch * 3)
966 ad->write_batch_count /= 2;
967 else
968 ad->write_batch_count--;
969 } else if (write_time < batch && ad->current_write_count == 0) {
970 if (batch > write_time * 3)
971 ad->write_batch_count *= 2;
972 else
973 ad->write_batch_count++;
974 }
976 if (ad->write_batch_count < 1)
977 ad->write_batch_count = 1;
978 }
980 /*
981 * as_completed_request is to be called when a request has completed and
982 * returned something to the requesting process, be it an error or data.
983 */
984 static void as_completed_request(request_queue_t *q, struct request *rq)
985 {
986 struct as_data *ad = q->elevator->elevator_data;
987 struct as_rq *arq = RQ_DATA(rq);
989 WARN_ON(!list_empty(&rq->queuelist));
991 if (arq->state != AS_RQ_REMOVED) {
992 printk("arq->state %d\n", arq->state);
993 WARN_ON(1);
994 goto out;
995 }
997 if (ad->changed_batch && ad->nr_dispatched == 1) {
998 kblockd_schedule_work(&ad->antic_work);
999 ad->changed_batch = 0;
1001 if (ad->batch_data_dir == REQ_SYNC)
1002 ad->new_batch = 1;
1004 WARN_ON(ad->nr_dispatched == 0);
1005 ad->nr_dispatched--;
1007 /*
1008 * Start counting the batch from when a request of that direction is
1009 * actually serviced. This should help devices with big TCQ windows
1010 * and writeback caches
1011 */
1012 if (ad->new_batch && ad->batch_data_dir == arq->is_sync) {
1013 update_write_batch(ad);
1014 ad->current_batch_expires = jiffies +
1015 ad->batch_expire[REQ_SYNC];
1016 ad->new_batch = 0;
1019 if (ad->io_context == arq->io_context && ad->io_context) {
1020 ad->antic_start = jiffies;
1021 ad->ioc_finished = 1;
1022 if (ad->antic_status == ANTIC_WAIT_REQ) {
1023 /*
1024 * We were waiting on this request, now anticipate
1025 * the next one
1026 */
1027 as_antic_waitnext(ad);
1031 as_put_io_context(arq);
1032 out:
1033 arq->state = AS_RQ_POSTSCHED;
1036 /*
1037 * as_remove_queued_request removes a request from the pre dispatch queue
1038 * without updating refcounts. It is expected the caller will drop the
1039 * reference unless it replaces the request at somepart of the elevator
1040 * (ie. the dispatch queue)
1041 */
1042 static void as_remove_queued_request(request_queue_t *q, struct request *rq)
1044 struct as_rq *arq = RQ_DATA(rq);
1045 const int data_dir = arq->is_sync;
1046 struct as_data *ad = q->elevator->elevator_data;
1048 WARN_ON(arq->state != AS_RQ_QUEUED);
1050 if (arq->io_context && arq->io_context->aic) {
1051 BUG_ON(!atomic_read(&arq->io_context->aic->nr_queued));
1052 atomic_dec(&arq->io_context->aic->nr_queued);
1055 /*
1056 * Update the "next_arq" cache if we are about to remove its
1057 * entry
1058 */
1059 if (ad->next_arq[data_dir] == arq)
1060 ad->next_arq[data_dir] = as_find_next_arq(ad, arq);
1062 list_del_init(&arq->fifo);
1063 as_del_arq_hash(arq);
1064 as_del_arq_rb(ad, arq);
1067 /*
1068 * as_fifo_expired returns 0 if there are no expired reads on the fifo,
1069 * 1 otherwise. It is ratelimited so that we only perform the check once per
1070 * `fifo_expire' interval. Otherwise a large number of expired requests
1071 * would create a hopeless seekstorm.
1073 * See as_antic_expired comment.
1074 */
1075 static int as_fifo_expired(struct as_data *ad, int adir)
1077 struct as_rq *arq;
1078 long delta_jif;
1080 delta_jif = jiffies - ad->last_check_fifo[adir];
1081 if (unlikely(delta_jif < 0))
1082 delta_jif = -delta_jif;
1083 if (delta_jif < ad->fifo_expire[adir])
1084 return 0;
1086 ad->last_check_fifo[adir] = jiffies;
1088 if (list_empty(&ad->fifo_list[adir]))
1089 return 0;
1091 arq = list_entry_fifo(ad->fifo_list[adir].next);
1093 return time_after(jiffies, arq->expires);
1096 /*
1097 * as_batch_expired returns true if the current batch has expired. A batch
1098 * is a set of reads or a set of writes.
1099 */
1100 static inline int as_batch_expired(struct as_data *ad)
1102 if (ad->changed_batch || ad->new_batch)
1103 return 0;
1105 if (ad->batch_data_dir == REQ_SYNC)
1106 /* TODO! add a check so a complete fifo gets written? */
1107 return time_after(jiffies, ad->current_batch_expires);
1109 return time_after(jiffies, ad->current_batch_expires)
1110 || ad->current_write_count == 0;
1113 /*
1114 * move an entry to dispatch queue
1115 */
1116 static void as_move_to_dispatch(struct as_data *ad, struct as_rq *arq)
1118 struct request *rq = arq->request;
1119 const int data_dir = arq->is_sync;
1121 BUG_ON(!RB_EMPTY_NODE(&arq->rb_node));
1123 as_antic_stop(ad);
1124 ad->antic_status = ANTIC_OFF;
1126 /*
1127 * This has to be set in order to be correctly updated by
1128 * as_find_next_arq
1129 */
1130 ad->last_sector[data_dir] = rq->sector + rq->nr_sectors;
1132 if (data_dir == REQ_SYNC) {
1133 /* In case we have to anticipate after this */
1134 copy_io_context(&ad->io_context, &arq->io_context);
1135 } else {
1136 if (ad->io_context) {
1137 put_io_context(ad->io_context);
1138 ad->io_context = NULL;
1141 if (ad->current_write_count != 0)
1142 ad->current_write_count--;
1144 ad->ioc_finished = 0;
1146 ad->next_arq[data_dir] = as_find_next_arq(ad, arq);
1148 /*
1149 * take it off the sort and fifo list, add to dispatch queue
1150 */
1151 as_remove_queued_request(ad->q, rq);
1152 WARN_ON(arq->state != AS_RQ_QUEUED);
1154 elv_dispatch_sort(ad->q, rq);
1156 arq->state = AS_RQ_DISPATCHED;
1157 if (arq->io_context && arq->io_context->aic)
1158 atomic_inc(&arq->io_context->aic->nr_dispatched);
1159 ad->nr_dispatched++;
1162 /*
1163 * as_dispatch_request selects the best request according to
1164 * read/write expire, batch expire, etc, and moves it to the dispatch
1165 * queue. Returns 1 if a request was found, 0 otherwise.
1166 */
1167 static int as_dispatch_request(request_queue_t *q, int force)
1169 struct as_data *ad = q->elevator->elevator_data;
1170 struct as_rq *arq;
1171 const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]);
1172 const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]);
1174 if (unlikely(force)) {
1175 /*
1176 * Forced dispatch, accounting is useless. Reset
1177 * accounting states and dump fifo_lists. Note that
1178 * batch_data_dir is reset to REQ_SYNC to avoid
1179 * screwing write batch accounting as write batch
1180 * accounting occurs on W->R transition.
1181 */
1182 int dispatched = 0;
1184 ad->batch_data_dir = REQ_SYNC;
1185 ad->changed_batch = 0;
1186 ad->new_batch = 0;
1188 while (ad->next_arq[REQ_SYNC]) {
1189 as_move_to_dispatch(ad, ad->next_arq[REQ_SYNC]);
1190 dispatched++;
1192 ad->last_check_fifo[REQ_SYNC] = jiffies;
1194 while (ad->next_arq[REQ_ASYNC]) {
1195 as_move_to_dispatch(ad, ad->next_arq[REQ_ASYNC]);
1196 dispatched++;
1198 ad->last_check_fifo[REQ_ASYNC] = jiffies;
1200 return dispatched;
1203 /* Signal that the write batch was uncontended, so we can't time it */
1204 if (ad->batch_data_dir == REQ_ASYNC && !reads) {
1205 if (ad->current_write_count == 0 || !writes)
1206 ad->write_batch_idled = 1;
1209 if (!(reads || writes)
1210 || ad->antic_status == ANTIC_WAIT_REQ
1211 || ad->antic_status == ANTIC_WAIT_NEXT
1212 || ad->changed_batch)
1213 return 0;
1215 if (!(reads && writes && as_batch_expired(ad))) {
1216 /*
1217 * batch is still running or no reads or no writes
1218 */
1219 arq = ad->next_arq[ad->batch_data_dir];
1221 if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) {
1222 if (as_fifo_expired(ad, REQ_SYNC))
1223 goto fifo_expired;
1225 if (as_can_anticipate(ad, arq)) {
1226 as_antic_waitreq(ad);
1227 return 0;
1231 if (arq) {
1232 /* we have a "next request" */
1233 if (reads && !writes)
1234 ad->current_batch_expires =
1235 jiffies + ad->batch_expire[REQ_SYNC];
1236 goto dispatch_request;
1240 /*
1241 * at this point we are not running a batch. select the appropriate
1242 * data direction (read / write)
1243 */
1245 if (reads) {
1246 BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_SYNC]));
1248 if (writes && ad->batch_data_dir == REQ_SYNC)
1249 /*
1250 * Last batch was a read, switch to writes
1251 */
1252 goto dispatch_writes;
1254 if (ad->batch_data_dir == REQ_ASYNC) {
1255 WARN_ON(ad->new_batch);
1256 ad->changed_batch = 1;
1258 ad->batch_data_dir = REQ_SYNC;
1259 arq = list_entry_fifo(ad->fifo_list[ad->batch_data_dir].next);
1260 ad->last_check_fifo[ad->batch_data_dir] = jiffies;
1261 goto dispatch_request;
1264 /*
1265 * the last batch was a read
1266 */
1268 if (writes) {
1269 dispatch_writes:
1270 BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_ASYNC]));
1272 if (ad->batch_data_dir == REQ_SYNC) {
1273 ad->changed_batch = 1;
1275 /*
1276 * new_batch might be 1 when the queue runs out of
1277 * reads. A subsequent submission of a write might
1278 * cause a change of batch before the read is finished.
1279 */
1280 ad->new_batch = 0;
1282 ad->batch_data_dir = REQ_ASYNC;
1283 ad->current_write_count = ad->write_batch_count;
1284 ad->write_batch_idled = 0;
1285 arq = ad->next_arq[ad->batch_data_dir];
1286 goto dispatch_request;
1289 BUG();
1290 return 0;
1292 dispatch_request:
1293 /*
1294 * If a request has expired, service it.
1295 */
1297 if (as_fifo_expired(ad, ad->batch_data_dir)) {
1298 fifo_expired:
1299 arq = list_entry_fifo(ad->fifo_list[ad->batch_data_dir].next);
1300 BUG_ON(arq == NULL);
1303 if (ad->changed_batch) {
1304 WARN_ON(ad->new_batch);
1306 if (ad->nr_dispatched)
1307 return 0;
1309 if (ad->batch_data_dir == REQ_ASYNC)
1310 ad->current_batch_expires = jiffies +
1311 ad->batch_expire[REQ_ASYNC];
1312 else
1313 ad->new_batch = 1;
1315 ad->changed_batch = 0;
1318 /*
1319 * arq is the selected appropriate request.
1320 */
1321 as_move_to_dispatch(ad, arq);
1323 return 1;
1326 /*
1327 * add arq to rbtree and fifo
1328 */
1329 static void as_add_request(request_queue_t *q, struct request *rq)
1331 struct as_data *ad = q->elevator->elevator_data;
1332 struct as_rq *arq = RQ_DATA(rq);
1333 int data_dir;
1335 arq->state = AS_RQ_NEW;
1337 if (rq_data_dir(arq->request) == READ
1338 || (arq->request->flags & REQ_RW_SYNC))
1339 arq->is_sync = 1;
1340 else
1341 arq->is_sync = 0;
1342 data_dir = arq->is_sync;
1344 arq->io_context = as_get_io_context();
1346 if (arq->io_context) {
1347 as_update_iohist(ad, arq->io_context->aic, arq->request);
1348 atomic_inc(&arq->io_context->aic->nr_queued);
1351 as_add_arq_rb(ad, arq);
1352 if (rq_mergeable(arq->request))
1353 as_add_arq_hash(ad, arq);
1355 /*
1356 * set expire time (only used for reads) and add to fifo list
1357 */
1358 arq->expires = jiffies + ad->fifo_expire[data_dir];
1359 list_add_tail(&arq->fifo, &ad->fifo_list[data_dir]);
1361 as_update_arq(ad, arq); /* keep state machine up to date */
1362 arq->state = AS_RQ_QUEUED;
1365 static void as_activate_request(request_queue_t *q, struct request *rq)
1367 struct as_rq *arq = RQ_DATA(rq);
1369 WARN_ON(arq->state != AS_RQ_DISPATCHED);
1370 arq->state = AS_RQ_REMOVED;
1371 if (arq->io_context && arq->io_context->aic)
1372 atomic_dec(&arq->io_context->aic->nr_dispatched);
1375 static void as_deactivate_request(request_queue_t *q, struct request *rq)
1377 struct as_rq *arq = RQ_DATA(rq);
1379 WARN_ON(arq->state != AS_RQ_REMOVED);
1380 arq->state = AS_RQ_DISPATCHED;
1381 if (arq->io_context && arq->io_context->aic)
1382 atomic_inc(&arq->io_context->aic->nr_dispatched);
1385 /*
1386 * as_queue_empty tells us if there are requests left in the device. It may
1387 * not be the case that a driver can get the next request even if the queue
1388 * is not empty - it is used in the block layer to check for plugging and
1389 * merging opportunities
1390 */
1391 static int as_queue_empty(request_queue_t *q)
1393 struct as_data *ad = q->elevator->elevator_data;
1395 return list_empty(&ad->fifo_list[REQ_ASYNC])
1396 && list_empty(&ad->fifo_list[REQ_SYNC]);
1399 static struct request *as_former_request(request_queue_t *q,
1400 struct request *rq)
1402 struct as_rq *arq = RQ_DATA(rq);
1403 struct rb_node *rbprev = rb_prev(&arq->rb_node);
1404 struct request *ret = NULL;
1406 if (rbprev)
1407 ret = rb_entry_arq(rbprev)->request;
1409 return ret;
1412 static struct request *as_latter_request(request_queue_t *q,
1413 struct request *rq)
1415 struct as_rq *arq = RQ_DATA(rq);
1416 struct rb_node *rbnext = rb_next(&arq->rb_node);
1417 struct request *ret = NULL;
1419 if (rbnext)
1420 ret = rb_entry_arq(rbnext)->request;
1422 return ret;
1425 static int
1426 as_merge(request_queue_t *q, struct request **req, struct bio *bio)
1428 struct as_data *ad = q->elevator->elevator_data;
1429 sector_t rb_key = bio->bi_sector + bio_sectors(bio);
1430 struct request *__rq;
1431 int ret;
1433 /*
1434 * see if the merge hash can satisfy a back merge
1435 */
1436 __rq = as_find_arq_hash(ad, bio->bi_sector);
1437 if (__rq) {
1438 BUG_ON(__rq->sector + __rq->nr_sectors != bio->bi_sector);
1440 if (elv_rq_merge_ok(__rq, bio)) {
1441 ret = ELEVATOR_BACK_MERGE;
1442 goto out;
1446 /*
1447 * check for front merge
1448 */
1449 __rq = as_find_arq_rb(ad, rb_key, bio_data_dir(bio));
1450 if (__rq) {
1451 BUG_ON(rb_key != rq_rb_key(__rq));
1453 if (elv_rq_merge_ok(__rq, bio)) {
1454 ret = ELEVATOR_FRONT_MERGE;
1455 goto out;
1459 return ELEVATOR_NO_MERGE;
1460 out:
1461 if (ret) {
1462 if (rq_mergeable(__rq))
1463 as_hot_arq_hash(ad, RQ_DATA(__rq));
1465 *req = __rq;
1466 return ret;
1469 static void as_merged_request(request_queue_t *q, struct request *req)
1471 struct as_data *ad = q->elevator->elevator_data;
1472 struct as_rq *arq = RQ_DATA(req);
1474 /*
1475 * hash always needs to be repositioned, key is end sector
1476 */
1477 as_del_arq_hash(arq);
1478 as_add_arq_hash(ad, arq);
1480 /*
1481 * if the merge was a front merge, we need to reposition request
1482 */
1483 if (rq_rb_key(req) != arq->rb_key) {
1484 as_del_arq_rb(ad, arq);
1485 as_add_arq_rb(ad, arq);
1486 /*
1487 * Note! At this stage of this and the next function, our next
1488 * request may not be optimal - eg the request may have "grown"
1489 * behind the disk head. We currently don't bother adjusting.
1490 */
1494 static void as_merged_requests(request_queue_t *q, struct request *req,
1495 struct request *next)
1497 struct as_data *ad = q->elevator->elevator_data;
1498 struct as_rq *arq = RQ_DATA(req);
1499 struct as_rq *anext = RQ_DATA(next);
1501 BUG_ON(!arq);
1502 BUG_ON(!anext);
1504 /*
1505 * reposition arq (this is the merged request) in hash, and in rbtree
1506 * in case of a front merge
1507 */
1508 as_del_arq_hash(arq);
1509 as_add_arq_hash(ad, arq);
1511 if (rq_rb_key(req) != arq->rb_key) {
1512 as_del_arq_rb(ad, arq);
1513 as_add_arq_rb(ad, arq);
1516 /*
1517 * if anext expires before arq, assign its expire time to arq
1518 * and move into anext position (anext will be deleted) in fifo
1519 */
1520 if (!list_empty(&arq->fifo) && !list_empty(&anext->fifo)) {
1521 if (time_before(anext->expires, arq->expires)) {
1522 list_move(&arq->fifo, &anext->fifo);
1523 arq->expires = anext->expires;
1524 /*
1525 * Don't copy here but swap, because when anext is
1526 * removed below, it must contain the unused context
1527 */
1528 swap_io_context(&arq->io_context, &anext->io_context);
1532 /*
1533 * kill knowledge of next, this one is a goner
1534 */
1535 as_remove_queued_request(q, next);
1536 as_put_io_context(anext);
1538 anext->state = AS_RQ_MERGED;
1541 /*
1542 * This is executed in a "deferred" process context, by kblockd. It calls the
1543 * driver's request_fn so the driver can submit that request.
1545 * IMPORTANT! This guy will reenter the elevator, so set up all queue global
1546 * state before calling, and don't rely on any state over calls.
1548 * FIXME! dispatch queue is not a queue at all!
1549 */
1550 static void as_work_handler(void *data)
1552 struct request_queue *q = data;
1553 unsigned long flags;
1555 spin_lock_irqsave(q->queue_lock, flags);
1556 if (!as_queue_empty(q))
1557 q->request_fn(q);
1558 spin_unlock_irqrestore(q->queue_lock, flags);
1561 static void as_put_request(request_queue_t *q, struct request *rq)
1563 struct as_data *ad = q->elevator->elevator_data;
1564 struct as_rq *arq = RQ_DATA(rq);
1566 if (!arq) {
1567 WARN_ON(1);
1568 return;
1571 if (unlikely(arq->state != AS_RQ_POSTSCHED &&
1572 arq->state != AS_RQ_PRESCHED &&
1573 arq->state != AS_RQ_MERGED)) {
1574 printk("arq->state %d\n", arq->state);
1575 WARN_ON(1);
1578 mempool_free(arq, ad->arq_pool);
1579 rq->elevator_private = NULL;
1582 static int as_set_request(request_queue_t *q, struct request *rq,
1583 struct bio *bio, gfp_t gfp_mask)
1585 struct as_data *ad = q->elevator->elevator_data;
1586 struct as_rq *arq = mempool_alloc(ad->arq_pool, gfp_mask);
1588 if (arq) {
1589 memset(arq, 0, sizeof(*arq));
1590 RB_CLEAR_NODE(&arq->rb_node);
1591 arq->request = rq;
1592 arq->state = AS_RQ_PRESCHED;
1593 arq->io_context = NULL;
1594 INIT_HLIST_NODE(&arq->hash);
1595 INIT_LIST_HEAD(&arq->fifo);
1596 rq->elevator_private = arq;
1597 return 0;
1600 return 1;
1603 static int as_may_queue(request_queue_t *q, int rw, struct bio *bio)
1605 int ret = ELV_MQUEUE_MAY;
1606 struct as_data *ad = q->elevator->elevator_data;
1607 struct io_context *ioc;
1608 if (ad->antic_status == ANTIC_WAIT_REQ ||
1609 ad->antic_status == ANTIC_WAIT_NEXT) {
1610 ioc = as_get_io_context();
1611 if (ad->io_context == ioc)
1612 ret = ELV_MQUEUE_MUST;
1613 put_io_context(ioc);
1616 return ret;
1619 static void as_exit_queue(elevator_t *e)
1621 struct as_data *ad = e->elevator_data;
1623 del_timer_sync(&ad->antic_timer);
1624 kblockd_flush();
1626 BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC]));
1627 BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC]));
1629 mempool_destroy(ad->arq_pool);
1630 put_io_context(ad->io_context);
1631 kfree(ad->hash);
1632 kfree(ad);
1635 /*
1636 * initialize elevator private data (as_data), and alloc a arq for
1637 * each request on the free lists
1638 */
1639 static void *as_init_queue(request_queue_t *q, elevator_t *e)
1641 struct as_data *ad;
1642 int i;
1644 if (!arq_pool)
1645 return NULL;
1647 ad = kmalloc_node(sizeof(*ad), GFP_KERNEL, q->node);
1648 if (!ad)
1649 return NULL;
1650 memset(ad, 0, sizeof(*ad));
1652 ad->q = q; /* Identify what queue the data belongs to */
1654 ad->hash = kmalloc_node(sizeof(struct hlist_head)*AS_HASH_ENTRIES,
1655 GFP_KERNEL, q->node);
1656 if (!ad->hash) {
1657 kfree(ad);
1658 return NULL;
1661 ad->arq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1662 mempool_free_slab, arq_pool, q->node);
1663 if (!ad->arq_pool) {
1664 kfree(ad->hash);
1665 kfree(ad);
1666 return NULL;
1669 /* anticipatory scheduling helpers */
1670 ad->antic_timer.function = as_antic_timeout;
1671 ad->antic_timer.data = (unsigned long)q;
1672 init_timer(&ad->antic_timer);
1673 INIT_WORK(&ad->antic_work, as_work_handler, q);
1675 for (i = 0; i < AS_HASH_ENTRIES; i++)
1676 INIT_HLIST_HEAD(&ad->hash[i]);
1678 INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]);
1679 INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]);
1680 ad->sort_list[REQ_SYNC] = RB_ROOT;
1681 ad->sort_list[REQ_ASYNC] = RB_ROOT;
1682 ad->fifo_expire[REQ_SYNC] = default_read_expire;
1683 ad->fifo_expire[REQ_ASYNC] = default_write_expire;
1684 ad->antic_expire = default_antic_expire;
1685 ad->batch_expire[REQ_SYNC] = default_read_batch_expire;
1686 ad->batch_expire[REQ_ASYNC] = default_write_batch_expire;
1688 ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC];
1689 ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10;
1690 if (ad->write_batch_count < 2)
1691 ad->write_batch_count = 2;
1693 return ad;
1696 /*
1697 * sysfs parts below
1698 */
1700 static ssize_t
1701 as_var_show(unsigned int var, char *page)
1703 return sprintf(page, "%d\n", var);
1706 static ssize_t
1707 as_var_store(unsigned long *var, const char *page, size_t count)
1709 char *p = (char *) page;
1711 *var = simple_strtoul(p, &p, 10);
1712 return count;
1715 static ssize_t est_time_show(elevator_t *e, char *page)
1717 struct as_data *ad = e->elevator_data;
1718 int pos = 0;
1720 pos += sprintf(page+pos, "%lu %% exit probability\n",
1721 100*ad->exit_prob/256);
1722 pos += sprintf(page+pos, "%lu %% probability of exiting without a "
1723 "cooperating process submitting IO\n",
1724 100*ad->exit_no_coop/256);
1725 pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean);
1726 pos += sprintf(page+pos, "%llu sectors new seek distance\n",
1727 (unsigned long long)ad->new_seek_mean);
1729 return pos;
1732 #define SHOW_FUNCTION(__FUNC, __VAR) \
1733 static ssize_t __FUNC(elevator_t *e, char *page) \
1734 { \
1735 struct as_data *ad = e->elevator_data; \
1736 return as_var_show(jiffies_to_msecs((__VAR)), (page)); \
1738 SHOW_FUNCTION(as_read_expire_show, ad->fifo_expire[REQ_SYNC]);
1739 SHOW_FUNCTION(as_write_expire_show, ad->fifo_expire[REQ_ASYNC]);
1740 SHOW_FUNCTION(as_antic_expire_show, ad->antic_expire);
1741 SHOW_FUNCTION(as_read_batch_expire_show, ad->batch_expire[REQ_SYNC]);
1742 SHOW_FUNCTION(as_write_batch_expire_show, ad->batch_expire[REQ_ASYNC]);
1743 #undef SHOW_FUNCTION
1745 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
1746 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
1747 { \
1748 struct as_data *ad = e->elevator_data; \
1749 int ret = as_var_store(__PTR, (page), count); \
1750 if (*(__PTR) < (MIN)) \
1751 *(__PTR) = (MIN); \
1752 else if (*(__PTR) > (MAX)) \
1753 *(__PTR) = (MAX); \
1754 *(__PTR) = msecs_to_jiffies(*(__PTR)); \
1755 return ret; \
1757 STORE_FUNCTION(as_read_expire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX);
1758 STORE_FUNCTION(as_write_expire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX);
1759 STORE_FUNCTION(as_antic_expire_store, &ad->antic_expire, 0, INT_MAX);
1760 STORE_FUNCTION(as_read_batch_expire_store,
1761 &ad->batch_expire[REQ_SYNC], 0, INT_MAX);
1762 STORE_FUNCTION(as_write_batch_expire_store,
1763 &ad->batch_expire[REQ_ASYNC], 0, INT_MAX);
1764 #undef STORE_FUNCTION
1766 #define AS_ATTR(name) \
1767 __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store)
1769 static struct elv_fs_entry as_attrs[] = {
1770 __ATTR_RO(est_time),
1771 AS_ATTR(read_expire),
1772 AS_ATTR(write_expire),
1773 AS_ATTR(antic_expire),
1774 AS_ATTR(read_batch_expire),
1775 AS_ATTR(write_batch_expire),
1776 __ATTR_NULL
1777 };
1779 static struct elevator_type iosched_as = {
1780 .ops = {
1781 .elevator_merge_fn = as_merge,
1782 .elevator_merged_fn = as_merged_request,
1783 .elevator_merge_req_fn = as_merged_requests,
1784 .elevator_dispatch_fn = as_dispatch_request,
1785 .elevator_add_req_fn = as_add_request,
1786 .elevator_activate_req_fn = as_activate_request,
1787 .elevator_deactivate_req_fn = as_deactivate_request,
1788 .elevator_queue_empty_fn = as_queue_empty,
1789 .elevator_completed_req_fn = as_completed_request,
1790 .elevator_former_req_fn = as_former_request,
1791 .elevator_latter_req_fn = as_latter_request,
1792 .elevator_set_req_fn = as_set_request,
1793 .elevator_put_req_fn = as_put_request,
1794 .elevator_may_queue_fn = as_may_queue,
1795 .elevator_init_fn = as_init_queue,
1796 .elevator_exit_fn = as_exit_queue,
1797 .trim = as_trim,
1798 },
1800 .elevator_attrs = as_attrs,
1801 .elevator_name = "anticipatory",
1802 .elevator_owner = THIS_MODULE,
1803 };
1805 static int __init as_init(void)
1807 int ret;
1809 arq_pool = kmem_cache_create("as_arq", sizeof(struct as_rq),
1810 0, 0, NULL, NULL);
1811 if (!arq_pool)
1812 return -ENOMEM;
1814 ret = elv_register(&iosched_as);
1815 if (!ret) {
1816 /*
1817 * don't allow AS to get unregistered, since we would have
1818 * to browse all tasks in the system and release their
1819 * as_io_context first
1820 */
1821 __module_get(THIS_MODULE);
1822 return 0;
1825 kmem_cache_destroy(arq_pool);
1826 return ret;
1829 static void __exit as_exit(void)
1831 DECLARE_COMPLETION(all_gone);
1832 elv_unregister(&iosched_as);
1833 ioc_gone = &all_gone;
1834 /* ioc_gone's update must be visible before reading ioc_count */
1835 smp_wmb();
1836 if (atomic_read(&ioc_count))
1837 wait_for_completion(ioc_gone);
1838 synchronize_rcu();
1839 kmem_cache_destroy(arq_pool);
1842 module_init(as_init);
1843 module_exit(as_exit);
1845 MODULE_AUTHOR("Nick Piggin");
1846 MODULE_LICENSE("GPL");
1847 MODULE_DESCRIPTION("anticipatory IO scheduler");