smaps: only define clear_refs for CONFIG_MMU
[linux-2.6/mini2440.git] / block / as-iosched.c
blob640aa839d63fe45ff7bcd6ca04f81b47bbdad409
1 /*
2 * Anticipatory & deadline i/o scheduler.
4 * Copyright (C) 2002 Jens Axboe <axboe@kernel.dk>
5 * Nick Piggin <nickpiggin@yahoo.com.au>
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/rbtree.h>
18 #include <linux/interrupt.h>
20 #define REQ_SYNC 1
21 #define REQ_ASYNC 0
24 * See Documentation/block/as-iosched.txt
28 * max time before a read is submitted.
30 #define default_read_expire (HZ / 8)
33 * ditto for writes, these limits are not hard, even
34 * if the disk is capable of satisfying them.
36 #define default_write_expire (HZ / 4)
39 * read_batch_expire describes how long we will allow a stream of reads to
40 * persist before looking to see whether it is time to switch over to writes.
42 #define default_read_batch_expire (HZ / 2)
45 * write_batch_expire describes how long we want a stream of writes to run for.
46 * This is not a hard limit, but a target we set for the auto-tuning thingy.
47 * See, the problem is: we can send a lot of writes to disk cache / TCQ in
48 * a short amount of time...
50 #define default_write_batch_expire (HZ / 8)
53 * max time we may wait to anticipate a read (default around 6ms)
55 #define default_antic_expire ((HZ / 150) ? HZ / 150 : 1)
58 * Keep track of up to 20ms thinktimes. We can go as big as we like here,
59 * however huge values tend to interfere and not decay fast enough. A program
60 * might be in a non-io phase of operation. Waiting on user input for example,
61 * or doing a lengthy computation. A small penalty can be justified there, and
62 * will still catch out those processes that constantly have large thinktimes.
64 #define MAX_THINKTIME (HZ/50UL)
66 /* Bits in as_io_context.state */
67 enum as_io_states {
68 AS_TASK_RUNNING=0, /* Process has not exited */
69 AS_TASK_IOSTARTED, /* Process has started some IO */
70 AS_TASK_IORUNNING, /* Process has completed some IO */
73 enum anticipation_status {
74 ANTIC_OFF=0, /* Not anticipating (normal operation) */
75 ANTIC_WAIT_REQ, /* The last read has not yet completed */
76 ANTIC_WAIT_NEXT, /* Currently anticipating a request vs
77 last read (which has completed) */
78 ANTIC_FINISHED, /* Anticipating but have found a candidate
79 * or timed out */
82 struct as_data {
84 * run time data
87 struct request_queue *q; /* the "owner" queue */
90 * requests (as_rq s) are present on both sort_list and fifo_list
92 struct rb_root sort_list[2];
93 struct list_head fifo_list[2];
95 struct request *next_rq[2]; /* next in sort order */
96 sector_t last_sector[2]; /* last REQ_SYNC & REQ_ASYNC sectors */
98 unsigned long exit_prob; /* probability a task will exit while
99 being waited on */
100 unsigned long exit_no_coop; /* probablility an exited task will
101 not be part of a later cooperating
102 request */
103 unsigned long new_ttime_total; /* mean thinktime on new proc */
104 unsigned long new_ttime_mean;
105 u64 new_seek_total; /* mean seek on new proc */
106 sector_t new_seek_mean;
108 unsigned long current_batch_expires;
109 unsigned long last_check_fifo[2];
110 int changed_batch; /* 1: waiting for old batch to end */
111 int new_batch; /* 1: waiting on first read complete */
112 int batch_data_dir; /* current batch REQ_SYNC / REQ_ASYNC */
113 int write_batch_count; /* max # of reqs in a write batch */
114 int current_write_count; /* how many requests left this batch */
115 int write_batch_idled; /* has the write batch gone idle? */
117 enum anticipation_status antic_status;
118 unsigned long antic_start; /* jiffies: when it started */
119 struct timer_list antic_timer; /* anticipatory scheduling timer */
120 struct work_struct antic_work; /* Deferred unplugging */
121 struct io_context *io_context; /* Identify the expected process */
122 int ioc_finished; /* IO associated with io_context is finished */
123 int nr_dispatched;
126 * settings that change how the i/o scheduler behaves
128 unsigned long fifo_expire[2];
129 unsigned long batch_expire[2];
130 unsigned long antic_expire;
134 * per-request data.
136 enum arq_state {
137 AS_RQ_NEW=0, /* New - not referenced and not on any lists */
138 AS_RQ_QUEUED, /* In the request queue. It belongs to the
139 scheduler */
140 AS_RQ_DISPATCHED, /* On the dispatch list. It belongs to the
141 driver now */
142 AS_RQ_PRESCHED, /* Debug poisoning for requests being used */
143 AS_RQ_REMOVED,
144 AS_RQ_MERGED,
145 AS_RQ_POSTSCHED, /* when they shouldn't be */
148 #define RQ_IOC(rq) ((struct io_context *) (rq)->elevator_private)
149 #define RQ_STATE(rq) ((enum arq_state)(rq)->elevator_private2)
150 #define RQ_SET_STATE(rq, state) ((rq)->elevator_private2 = (void *) state)
152 static DEFINE_PER_CPU(unsigned long, ioc_count);
153 static struct completion *ioc_gone;
155 static void as_move_to_dispatch(struct as_data *ad, struct request *rq);
156 static void as_antic_stop(struct as_data *ad);
159 * IO Context helper functions
162 /* Called to deallocate the as_io_context */
163 static void free_as_io_context(struct as_io_context *aic)
165 kfree(aic);
166 elv_ioc_count_dec(ioc_count);
167 if (ioc_gone && !elv_ioc_count_read(ioc_count))
168 complete(ioc_gone);
171 static void as_trim(struct io_context *ioc)
173 if (ioc->aic)
174 free_as_io_context(ioc->aic);
175 ioc->aic = NULL;
178 /* Called when the task exits */
179 static void exit_as_io_context(struct as_io_context *aic)
181 WARN_ON(!test_bit(AS_TASK_RUNNING, &aic->state));
182 clear_bit(AS_TASK_RUNNING, &aic->state);
185 static struct as_io_context *alloc_as_io_context(void)
187 struct as_io_context *ret;
189 ret = kmalloc(sizeof(*ret), GFP_ATOMIC);
190 if (ret) {
191 ret->dtor = free_as_io_context;
192 ret->exit = exit_as_io_context;
193 ret->state = 1 << AS_TASK_RUNNING;
194 atomic_set(&ret->nr_queued, 0);
195 atomic_set(&ret->nr_dispatched, 0);
196 spin_lock_init(&ret->lock);
197 ret->ttime_total = 0;
198 ret->ttime_samples = 0;
199 ret->ttime_mean = 0;
200 ret->seek_total = 0;
201 ret->seek_samples = 0;
202 ret->seek_mean = 0;
203 elv_ioc_count_inc(ioc_count);
206 return ret;
210 * If the current task has no AS IO context then create one and initialise it.
211 * Then take a ref on the task's io context and return it.
213 static struct io_context *as_get_io_context(int node)
215 struct io_context *ioc = get_io_context(GFP_ATOMIC, node);
216 if (ioc && !ioc->aic) {
217 ioc->aic = alloc_as_io_context();
218 if (!ioc->aic) {
219 put_io_context(ioc);
220 ioc = NULL;
223 return ioc;
226 static void as_put_io_context(struct request *rq)
228 struct as_io_context *aic;
230 if (unlikely(!RQ_IOC(rq)))
231 return;
233 aic = RQ_IOC(rq)->aic;
235 if (rq_is_sync(rq) && aic) {
236 spin_lock(&aic->lock);
237 set_bit(AS_TASK_IORUNNING, &aic->state);
238 aic->last_end_request = jiffies;
239 spin_unlock(&aic->lock);
242 put_io_context(RQ_IOC(rq));
246 * rb tree support functions
248 #define RQ_RB_ROOT(ad, rq) (&(ad)->sort_list[rq_is_sync((rq))])
250 static void as_add_rq_rb(struct as_data *ad, struct request *rq)
252 struct request *alias;
254 while ((unlikely(alias = elv_rb_add(RQ_RB_ROOT(ad, rq), rq)))) {
255 as_move_to_dispatch(ad, alias);
256 as_antic_stop(ad);
260 static inline void as_del_rq_rb(struct as_data *ad, struct request *rq)
262 elv_rb_del(RQ_RB_ROOT(ad, rq), rq);
266 * IO Scheduler proper
269 #define MAXBACK (1024 * 1024) /*
270 * Maximum distance the disk will go backward
271 * for a request.
274 #define BACK_PENALTY 2
277 * as_choose_req selects the preferred one of two requests of the same data_dir
278 * ignoring time - eg. timeouts, which is the job of as_dispatch_request
280 static struct request *
281 as_choose_req(struct as_data *ad, struct request *rq1, struct request *rq2)
283 int data_dir;
284 sector_t last, s1, s2, d1, d2;
285 int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */
286 const sector_t maxback = MAXBACK;
288 if (rq1 == NULL || rq1 == rq2)
289 return rq2;
290 if (rq2 == NULL)
291 return rq1;
293 data_dir = rq_is_sync(rq1);
295 last = ad->last_sector[data_dir];
296 s1 = rq1->sector;
297 s2 = rq2->sector;
299 BUG_ON(data_dir != rq_is_sync(rq2));
302 * Strict one way elevator _except_ in the case where we allow
303 * short backward seeks which are biased as twice the cost of a
304 * similar forward seek.
306 if (s1 >= last)
307 d1 = s1 - last;
308 else if (s1+maxback >= last)
309 d1 = (last - s1)*BACK_PENALTY;
310 else {
311 r1_wrap = 1;
312 d1 = 0; /* shut up, gcc */
315 if (s2 >= last)
316 d2 = s2 - last;
317 else if (s2+maxback >= last)
318 d2 = (last - s2)*BACK_PENALTY;
319 else {
320 r2_wrap = 1;
321 d2 = 0;
324 /* Found required data */
325 if (!r1_wrap && r2_wrap)
326 return rq1;
327 else if (!r2_wrap && r1_wrap)
328 return rq2;
329 else if (r1_wrap && r2_wrap) {
330 /* both behind the head */
331 if (s1 <= s2)
332 return rq1;
333 else
334 return rq2;
337 /* Both requests in front of the head */
338 if (d1 < d2)
339 return rq1;
340 else if (d2 < d1)
341 return rq2;
342 else {
343 if (s1 >= s2)
344 return rq1;
345 else
346 return rq2;
351 * as_find_next_rq finds the next request after @prev in elevator order.
352 * this with as_choose_req form the basis for how the scheduler chooses
353 * what request to process next. Anticipation works on top of this.
355 static struct request *
356 as_find_next_rq(struct as_data *ad, struct request *last)
358 struct rb_node *rbnext = rb_next(&last->rb_node);
359 struct rb_node *rbprev = rb_prev(&last->rb_node);
360 struct request *next = NULL, *prev = NULL;
362 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
364 if (rbprev)
365 prev = rb_entry_rq(rbprev);
367 if (rbnext)
368 next = rb_entry_rq(rbnext);
369 else {
370 const int data_dir = rq_is_sync(last);
372 rbnext = rb_first(&ad->sort_list[data_dir]);
373 if (rbnext && rbnext != &last->rb_node)
374 next = rb_entry_rq(rbnext);
377 return as_choose_req(ad, next, prev);
381 * anticipatory scheduling functions follow
385 * as_antic_expired tells us when we have anticipated too long.
386 * The funny "absolute difference" math on the elapsed time is to handle
387 * jiffy wraps, and disks which have been idle for 0x80000000 jiffies.
389 static int as_antic_expired(struct as_data *ad)
391 long delta_jif;
393 delta_jif = jiffies - ad->antic_start;
394 if (unlikely(delta_jif < 0))
395 delta_jif = -delta_jif;
396 if (delta_jif < ad->antic_expire)
397 return 0;
399 return 1;
403 * as_antic_waitnext starts anticipating that a nice request will soon be
404 * submitted. See also as_antic_waitreq
406 static void as_antic_waitnext(struct as_data *ad)
408 unsigned long timeout;
410 BUG_ON(ad->antic_status != ANTIC_OFF
411 && ad->antic_status != ANTIC_WAIT_REQ);
413 timeout = ad->antic_start + ad->antic_expire;
415 mod_timer(&ad->antic_timer, timeout);
417 ad->antic_status = ANTIC_WAIT_NEXT;
421 * as_antic_waitreq starts anticipating. We don't start timing the anticipation
422 * until the request that we're anticipating on has finished. This means we
423 * are timing from when the candidate process wakes up hopefully.
425 static void as_antic_waitreq(struct as_data *ad)
427 BUG_ON(ad->antic_status == ANTIC_FINISHED);
428 if (ad->antic_status == ANTIC_OFF) {
429 if (!ad->io_context || ad->ioc_finished)
430 as_antic_waitnext(ad);
431 else
432 ad->antic_status = ANTIC_WAIT_REQ;
437 * This is called directly by the functions in this file to stop anticipation.
438 * We kill the timer and schedule a call to the request_fn asap.
440 static void as_antic_stop(struct as_data *ad)
442 int status = ad->antic_status;
444 if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) {
445 if (status == ANTIC_WAIT_NEXT)
446 del_timer(&ad->antic_timer);
447 ad->antic_status = ANTIC_FINISHED;
448 /* see as_work_handler */
449 kblockd_schedule_work(&ad->antic_work);
454 * as_antic_timeout is the timer function set by as_antic_waitnext.
456 static void as_antic_timeout(unsigned long data)
458 struct request_queue *q = (struct request_queue *)data;
459 struct as_data *ad = q->elevator->elevator_data;
460 unsigned long flags;
462 spin_lock_irqsave(q->queue_lock, flags);
463 if (ad->antic_status == ANTIC_WAIT_REQ
464 || ad->antic_status == ANTIC_WAIT_NEXT) {
465 struct as_io_context *aic = ad->io_context->aic;
467 ad->antic_status = ANTIC_FINISHED;
468 kblockd_schedule_work(&ad->antic_work);
470 if (aic->ttime_samples == 0) {
471 /* process anticipated on has exited or timed out*/
472 ad->exit_prob = (7*ad->exit_prob + 256)/8;
474 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
475 /* process not "saved" by a cooperating request */
476 ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8;
479 spin_unlock_irqrestore(q->queue_lock, flags);
482 static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic,
483 unsigned long ttime)
485 /* fixed point: 1.0 == 1<<8 */
486 if (aic->ttime_samples == 0) {
487 ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8;
488 ad->new_ttime_mean = ad->new_ttime_total / 256;
490 ad->exit_prob = (7*ad->exit_prob)/8;
492 aic->ttime_samples = (7*aic->ttime_samples + 256) / 8;
493 aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8;
494 aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples;
497 static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic,
498 sector_t sdist)
500 u64 total;
502 if (aic->seek_samples == 0) {
503 ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8;
504 ad->new_seek_mean = ad->new_seek_total / 256;
508 * Don't allow the seek distance to get too large from the
509 * odd fragment, pagein, etc
511 if (aic->seek_samples <= 60) /* second&third seek */
512 sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024);
513 else
514 sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64);
516 aic->seek_samples = (7*aic->seek_samples + 256) / 8;
517 aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8;
518 total = aic->seek_total + (aic->seek_samples/2);
519 do_div(total, aic->seek_samples);
520 aic->seek_mean = (sector_t)total;
524 * as_update_iohist keeps a decaying histogram of IO thinktimes, and
525 * updates @aic->ttime_mean based on that. It is called when a new
526 * request is queued.
528 static void as_update_iohist(struct as_data *ad, struct as_io_context *aic,
529 struct request *rq)
531 int data_dir = rq_is_sync(rq);
532 unsigned long thinktime = 0;
533 sector_t seek_dist;
535 if (aic == NULL)
536 return;
538 if (data_dir == REQ_SYNC) {
539 unsigned long in_flight = atomic_read(&aic->nr_queued)
540 + atomic_read(&aic->nr_dispatched);
541 spin_lock(&aic->lock);
542 if (test_bit(AS_TASK_IORUNNING, &aic->state) ||
543 test_bit(AS_TASK_IOSTARTED, &aic->state)) {
544 /* Calculate read -> read thinktime */
545 if (test_bit(AS_TASK_IORUNNING, &aic->state)
546 && in_flight == 0) {
547 thinktime = jiffies - aic->last_end_request;
548 thinktime = min(thinktime, MAX_THINKTIME-1);
550 as_update_thinktime(ad, aic, thinktime);
552 /* Calculate read -> read seek distance */
553 if (aic->last_request_pos < rq->sector)
554 seek_dist = rq->sector - aic->last_request_pos;
555 else
556 seek_dist = aic->last_request_pos - rq->sector;
557 as_update_seekdist(ad, aic, seek_dist);
559 aic->last_request_pos = rq->sector + rq->nr_sectors;
560 set_bit(AS_TASK_IOSTARTED, &aic->state);
561 spin_unlock(&aic->lock);
566 * as_close_req decides if one request is considered "close" to the
567 * previous one issued.
569 static int as_close_req(struct as_data *ad, struct as_io_context *aic,
570 struct request *rq)
572 unsigned long delay; /* jiffies */
573 sector_t last = ad->last_sector[ad->batch_data_dir];
574 sector_t next = rq->sector;
575 sector_t delta; /* acceptable close offset (in sectors) */
576 sector_t s;
578 if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished)
579 delay = 0;
580 else
581 delay = jiffies - ad->antic_start;
583 if (delay == 0)
584 delta = 8192;
585 else if (delay <= (20 * HZ / 1000) && delay <= ad->antic_expire)
586 delta = 8192 << delay;
587 else
588 return 1;
590 if ((last <= next + (delta>>1)) && (next <= last + delta))
591 return 1;
593 if (last < next)
594 s = next - last;
595 else
596 s = last - next;
598 if (aic->seek_samples == 0) {
600 * Process has just started IO. Use past statistics to
601 * gauge success possibility
603 if (ad->new_seek_mean > s) {
604 /* this request is better than what we're expecting */
605 return 1;
608 } else {
609 if (aic->seek_mean > s) {
610 /* this request is better than what we're expecting */
611 return 1;
615 return 0;
619 * as_can_break_anticipation returns true if we have been anticipating this
620 * request.
622 * It also returns true if the process against which we are anticipating
623 * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to
624 * dispatch it ASAP, because we know that application will not be submitting
625 * any new reads.
627 * If the task which has submitted the request has exited, break anticipation.
629 * If this task has queued some other IO, do not enter enticipation.
631 static int as_can_break_anticipation(struct as_data *ad, struct request *rq)
633 struct io_context *ioc;
634 struct as_io_context *aic;
636 ioc = ad->io_context;
637 BUG_ON(!ioc);
639 if (rq && ioc == RQ_IOC(rq)) {
640 /* request from same process */
641 return 1;
644 if (ad->ioc_finished && as_antic_expired(ad)) {
646 * In this situation status should really be FINISHED,
647 * however the timer hasn't had the chance to run yet.
649 return 1;
652 aic = ioc->aic;
653 if (!aic)
654 return 0;
656 if (atomic_read(&aic->nr_queued) > 0) {
657 /* process has more requests queued */
658 return 1;
661 if (atomic_read(&aic->nr_dispatched) > 0) {
662 /* process has more requests dispatched */
663 return 1;
666 if (rq && rq_is_sync(rq) && as_close_req(ad, aic, rq)) {
668 * Found a close request that is not one of ours.
670 * This makes close requests from another process update
671 * our IO history. Is generally useful when there are
672 * two or more cooperating processes working in the same
673 * area.
675 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
676 if (aic->ttime_samples == 0)
677 ad->exit_prob = (7*ad->exit_prob + 256)/8;
679 ad->exit_no_coop = (7*ad->exit_no_coop)/8;
682 as_update_iohist(ad, aic, rq);
683 return 1;
686 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
687 /* process anticipated on has exited */
688 if (aic->ttime_samples == 0)
689 ad->exit_prob = (7*ad->exit_prob + 256)/8;
691 if (ad->exit_no_coop > 128)
692 return 1;
695 if (aic->ttime_samples == 0) {
696 if (ad->new_ttime_mean > ad->antic_expire)
697 return 1;
698 if (ad->exit_prob * ad->exit_no_coop > 128*256)
699 return 1;
700 } else if (aic->ttime_mean > ad->antic_expire) {
701 /* the process thinks too much between requests */
702 return 1;
705 return 0;
709 * as_can_anticipate indicates whether we should either run rq
710 * or keep anticipating a better request.
712 static int as_can_anticipate(struct as_data *ad, struct request *rq)
714 if (!ad->io_context)
716 * Last request submitted was a write
718 return 0;
720 if (ad->antic_status == ANTIC_FINISHED)
722 * Don't restart if we have just finished. Run the next request
724 return 0;
726 if (as_can_break_anticipation(ad, rq))
728 * This request is a good candidate. Don't keep anticipating,
729 * run it.
731 return 0;
734 * OK from here, we haven't finished, and don't have a decent request!
735 * Status is either ANTIC_OFF so start waiting,
736 * ANTIC_WAIT_REQ so continue waiting for request to finish
737 * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request.
740 return 1;
744 * as_update_rq must be called whenever a request (rq) is added to
745 * the sort_list. This function keeps caches up to date, and checks if the
746 * request might be one we are "anticipating"
748 static void as_update_rq(struct as_data *ad, struct request *rq)
750 const int data_dir = rq_is_sync(rq);
752 /* keep the next_rq cache up to date */
753 ad->next_rq[data_dir] = as_choose_req(ad, rq, ad->next_rq[data_dir]);
756 * have we been anticipating this request?
757 * or does it come from the same process as the one we are anticipating
758 * for?
760 if (ad->antic_status == ANTIC_WAIT_REQ
761 || ad->antic_status == ANTIC_WAIT_NEXT) {
762 if (as_can_break_anticipation(ad, rq))
763 as_antic_stop(ad);
768 * Gathers timings and resizes the write batch automatically
770 static void update_write_batch(struct as_data *ad)
772 unsigned long batch = ad->batch_expire[REQ_ASYNC];
773 long write_time;
775 write_time = (jiffies - ad->current_batch_expires) + batch;
776 if (write_time < 0)
777 write_time = 0;
779 if (write_time > batch && !ad->write_batch_idled) {
780 if (write_time > batch * 3)
781 ad->write_batch_count /= 2;
782 else
783 ad->write_batch_count--;
784 } else if (write_time < batch && ad->current_write_count == 0) {
785 if (batch > write_time * 3)
786 ad->write_batch_count *= 2;
787 else
788 ad->write_batch_count++;
791 if (ad->write_batch_count < 1)
792 ad->write_batch_count = 1;
796 * as_completed_request is to be called when a request has completed and
797 * returned something to the requesting process, be it an error or data.
799 static void as_completed_request(request_queue_t *q, struct request *rq)
801 struct as_data *ad = q->elevator->elevator_data;
803 WARN_ON(!list_empty(&rq->queuelist));
805 if (RQ_STATE(rq) != AS_RQ_REMOVED) {
806 printk("rq->state %d\n", RQ_STATE(rq));
807 WARN_ON(1);
808 goto out;
811 if (ad->changed_batch && ad->nr_dispatched == 1) {
812 kblockd_schedule_work(&ad->antic_work);
813 ad->changed_batch = 0;
815 if (ad->batch_data_dir == REQ_SYNC)
816 ad->new_batch = 1;
818 WARN_ON(ad->nr_dispatched == 0);
819 ad->nr_dispatched--;
822 * Start counting the batch from when a request of that direction is
823 * actually serviced. This should help devices with big TCQ windows
824 * and writeback caches
826 if (ad->new_batch && ad->batch_data_dir == rq_is_sync(rq)) {
827 update_write_batch(ad);
828 ad->current_batch_expires = jiffies +
829 ad->batch_expire[REQ_SYNC];
830 ad->new_batch = 0;
833 if (ad->io_context == RQ_IOC(rq) && ad->io_context) {
834 ad->antic_start = jiffies;
835 ad->ioc_finished = 1;
836 if (ad->antic_status == ANTIC_WAIT_REQ) {
838 * We were waiting on this request, now anticipate
839 * the next one
841 as_antic_waitnext(ad);
845 as_put_io_context(rq);
846 out:
847 RQ_SET_STATE(rq, AS_RQ_POSTSCHED);
851 * as_remove_queued_request removes a request from the pre dispatch queue
852 * without updating refcounts. It is expected the caller will drop the
853 * reference unless it replaces the request at somepart of the elevator
854 * (ie. the dispatch queue)
856 static void as_remove_queued_request(request_queue_t *q, struct request *rq)
858 const int data_dir = rq_is_sync(rq);
859 struct as_data *ad = q->elevator->elevator_data;
860 struct io_context *ioc;
862 WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED);
864 ioc = RQ_IOC(rq);
865 if (ioc && ioc->aic) {
866 BUG_ON(!atomic_read(&ioc->aic->nr_queued));
867 atomic_dec(&ioc->aic->nr_queued);
871 * Update the "next_rq" cache if we are about to remove its
872 * entry
874 if (ad->next_rq[data_dir] == rq)
875 ad->next_rq[data_dir] = as_find_next_rq(ad, rq);
877 rq_fifo_clear(rq);
878 as_del_rq_rb(ad, rq);
882 * as_fifo_expired returns 0 if there are no expired reads on the fifo,
883 * 1 otherwise. It is ratelimited so that we only perform the check once per
884 * `fifo_expire' interval. Otherwise a large number of expired requests
885 * would create a hopeless seekstorm.
887 * See as_antic_expired comment.
889 static int as_fifo_expired(struct as_data *ad, int adir)
891 struct request *rq;
892 long delta_jif;
894 delta_jif = jiffies - ad->last_check_fifo[adir];
895 if (unlikely(delta_jif < 0))
896 delta_jif = -delta_jif;
897 if (delta_jif < ad->fifo_expire[adir])
898 return 0;
900 ad->last_check_fifo[adir] = jiffies;
902 if (list_empty(&ad->fifo_list[adir]))
903 return 0;
905 rq = rq_entry_fifo(ad->fifo_list[adir].next);
907 return time_after(jiffies, rq_fifo_time(rq));
911 * as_batch_expired returns true if the current batch has expired. A batch
912 * is a set of reads or a set of writes.
914 static inline int as_batch_expired(struct as_data *ad)
916 if (ad->changed_batch || ad->new_batch)
917 return 0;
919 if (ad->batch_data_dir == REQ_SYNC)
920 /* TODO! add a check so a complete fifo gets written? */
921 return time_after(jiffies, ad->current_batch_expires);
923 return time_after(jiffies, ad->current_batch_expires)
924 || ad->current_write_count == 0;
928 * move an entry to dispatch queue
930 static void as_move_to_dispatch(struct as_data *ad, struct request *rq)
932 const int data_dir = rq_is_sync(rq);
934 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
936 as_antic_stop(ad);
937 ad->antic_status = ANTIC_OFF;
940 * This has to be set in order to be correctly updated by
941 * as_find_next_rq
943 ad->last_sector[data_dir] = rq->sector + rq->nr_sectors;
945 if (data_dir == REQ_SYNC) {
946 struct io_context *ioc = RQ_IOC(rq);
947 /* In case we have to anticipate after this */
948 copy_io_context(&ad->io_context, &ioc);
949 } else {
950 if (ad->io_context) {
951 put_io_context(ad->io_context);
952 ad->io_context = NULL;
955 if (ad->current_write_count != 0)
956 ad->current_write_count--;
958 ad->ioc_finished = 0;
960 ad->next_rq[data_dir] = as_find_next_rq(ad, rq);
963 * take it off the sort and fifo list, add to dispatch queue
965 as_remove_queued_request(ad->q, rq);
966 WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED);
968 elv_dispatch_sort(ad->q, rq);
970 RQ_SET_STATE(rq, AS_RQ_DISPATCHED);
971 if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
972 atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched);
973 ad->nr_dispatched++;
977 * as_dispatch_request selects the best request according to
978 * read/write expire, batch expire, etc, and moves it to the dispatch
979 * queue. Returns 1 if a request was found, 0 otherwise.
981 static int as_dispatch_request(request_queue_t *q, int force)
983 struct as_data *ad = q->elevator->elevator_data;
984 const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]);
985 const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]);
986 struct request *rq;
988 if (unlikely(force)) {
990 * Forced dispatch, accounting is useless. Reset
991 * accounting states and dump fifo_lists. Note that
992 * batch_data_dir is reset to REQ_SYNC to avoid
993 * screwing write batch accounting as write batch
994 * accounting occurs on W->R transition.
996 int dispatched = 0;
998 ad->batch_data_dir = REQ_SYNC;
999 ad->changed_batch = 0;
1000 ad->new_batch = 0;
1002 while (ad->next_rq[REQ_SYNC]) {
1003 as_move_to_dispatch(ad, ad->next_rq[REQ_SYNC]);
1004 dispatched++;
1006 ad->last_check_fifo[REQ_SYNC] = jiffies;
1008 while (ad->next_rq[REQ_ASYNC]) {
1009 as_move_to_dispatch(ad, ad->next_rq[REQ_ASYNC]);
1010 dispatched++;
1012 ad->last_check_fifo[REQ_ASYNC] = jiffies;
1014 return dispatched;
1017 /* Signal that the write batch was uncontended, so we can't time it */
1018 if (ad->batch_data_dir == REQ_ASYNC && !reads) {
1019 if (ad->current_write_count == 0 || !writes)
1020 ad->write_batch_idled = 1;
1023 if (!(reads || writes)
1024 || ad->antic_status == ANTIC_WAIT_REQ
1025 || ad->antic_status == ANTIC_WAIT_NEXT
1026 || ad->changed_batch)
1027 return 0;
1029 if (!(reads && writes && as_batch_expired(ad))) {
1031 * batch is still running or no reads or no writes
1033 rq = ad->next_rq[ad->batch_data_dir];
1035 if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) {
1036 if (as_fifo_expired(ad, REQ_SYNC))
1037 goto fifo_expired;
1039 if (as_can_anticipate(ad, rq)) {
1040 as_antic_waitreq(ad);
1041 return 0;
1045 if (rq) {
1046 /* we have a "next request" */
1047 if (reads && !writes)
1048 ad->current_batch_expires =
1049 jiffies + ad->batch_expire[REQ_SYNC];
1050 goto dispatch_request;
1055 * at this point we are not running a batch. select the appropriate
1056 * data direction (read / write)
1059 if (reads) {
1060 BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_SYNC]));
1062 if (writes && ad->batch_data_dir == REQ_SYNC)
1064 * Last batch was a read, switch to writes
1066 goto dispatch_writes;
1068 if (ad->batch_data_dir == REQ_ASYNC) {
1069 WARN_ON(ad->new_batch);
1070 ad->changed_batch = 1;
1072 ad->batch_data_dir = REQ_SYNC;
1073 rq = rq_entry_fifo(ad->fifo_list[REQ_SYNC].next);
1074 ad->last_check_fifo[ad->batch_data_dir] = jiffies;
1075 goto dispatch_request;
1079 * the last batch was a read
1082 if (writes) {
1083 dispatch_writes:
1084 BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_ASYNC]));
1086 if (ad->batch_data_dir == REQ_SYNC) {
1087 ad->changed_batch = 1;
1090 * new_batch might be 1 when the queue runs out of
1091 * reads. A subsequent submission of a write might
1092 * cause a change of batch before the read is finished.
1094 ad->new_batch = 0;
1096 ad->batch_data_dir = REQ_ASYNC;
1097 ad->current_write_count = ad->write_batch_count;
1098 ad->write_batch_idled = 0;
1099 rq = ad->next_rq[ad->batch_data_dir];
1100 goto dispatch_request;
1103 BUG();
1104 return 0;
1106 dispatch_request:
1108 * If a request has expired, service it.
1111 if (as_fifo_expired(ad, ad->batch_data_dir)) {
1112 fifo_expired:
1113 rq = rq_entry_fifo(ad->fifo_list[ad->batch_data_dir].next);
1116 if (ad->changed_batch) {
1117 WARN_ON(ad->new_batch);
1119 if (ad->nr_dispatched)
1120 return 0;
1122 if (ad->batch_data_dir == REQ_ASYNC)
1123 ad->current_batch_expires = jiffies +
1124 ad->batch_expire[REQ_ASYNC];
1125 else
1126 ad->new_batch = 1;
1128 ad->changed_batch = 0;
1132 * rq is the selected appropriate request.
1134 as_move_to_dispatch(ad, rq);
1136 return 1;
1140 * add rq to rbtree and fifo
1142 static void as_add_request(request_queue_t *q, struct request *rq)
1144 struct as_data *ad = q->elevator->elevator_data;
1145 int data_dir;
1147 RQ_SET_STATE(rq, AS_RQ_NEW);
1149 data_dir = rq_is_sync(rq);
1151 rq->elevator_private = as_get_io_context(q->node);
1153 if (RQ_IOC(rq)) {
1154 as_update_iohist(ad, RQ_IOC(rq)->aic, rq);
1155 atomic_inc(&RQ_IOC(rq)->aic->nr_queued);
1158 as_add_rq_rb(ad, rq);
1161 * set expire time (only used for reads) and add to fifo list
1163 rq_set_fifo_time(rq, jiffies + ad->fifo_expire[data_dir]);
1164 list_add_tail(&rq->queuelist, &ad->fifo_list[data_dir]);
1166 as_update_rq(ad, rq); /* keep state machine up to date */
1167 RQ_SET_STATE(rq, AS_RQ_QUEUED);
1170 static void as_activate_request(request_queue_t *q, struct request *rq)
1172 WARN_ON(RQ_STATE(rq) != AS_RQ_DISPATCHED);
1173 RQ_SET_STATE(rq, AS_RQ_REMOVED);
1174 if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
1175 atomic_dec(&RQ_IOC(rq)->aic->nr_dispatched);
1178 static void as_deactivate_request(request_queue_t *q, struct request *rq)
1180 WARN_ON(RQ_STATE(rq) != AS_RQ_REMOVED);
1181 RQ_SET_STATE(rq, AS_RQ_DISPATCHED);
1182 if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
1183 atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched);
1187 * as_queue_empty tells us if there are requests left in the device. It may
1188 * not be the case that a driver can get the next request even if the queue
1189 * is not empty - it is used in the block layer to check for plugging and
1190 * merging opportunities
1192 static int as_queue_empty(request_queue_t *q)
1194 struct as_data *ad = q->elevator->elevator_data;
1196 return list_empty(&ad->fifo_list[REQ_ASYNC])
1197 && list_empty(&ad->fifo_list[REQ_SYNC]);
1200 static int
1201 as_merge(request_queue_t *q, struct request **req, struct bio *bio)
1203 struct as_data *ad = q->elevator->elevator_data;
1204 sector_t rb_key = bio->bi_sector + bio_sectors(bio);
1205 struct request *__rq;
1208 * check for front merge
1210 __rq = elv_rb_find(&ad->sort_list[bio_data_dir(bio)], rb_key);
1211 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1212 *req = __rq;
1213 return ELEVATOR_FRONT_MERGE;
1216 return ELEVATOR_NO_MERGE;
1219 static void as_merged_request(request_queue_t *q, struct request *req, int type)
1221 struct as_data *ad = q->elevator->elevator_data;
1224 * if the merge was a front merge, we need to reposition request
1226 if (type == ELEVATOR_FRONT_MERGE) {
1227 as_del_rq_rb(ad, req);
1228 as_add_rq_rb(ad, req);
1230 * Note! At this stage of this and the next function, our next
1231 * request may not be optimal - eg the request may have "grown"
1232 * behind the disk head. We currently don't bother adjusting.
1237 static void as_merged_requests(request_queue_t *q, struct request *req,
1238 struct request *next)
1241 * if next expires before rq, assign its expire time to arq
1242 * and move into next position (next will be deleted) in fifo
1244 if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
1245 if (time_before(rq_fifo_time(next), rq_fifo_time(req))) {
1246 struct io_context *rioc = RQ_IOC(req);
1247 struct io_context *nioc = RQ_IOC(next);
1249 list_move(&req->queuelist, &next->queuelist);
1250 rq_set_fifo_time(req, rq_fifo_time(next));
1252 * Don't copy here but swap, because when anext is
1253 * removed below, it must contain the unused context
1255 swap_io_context(&rioc, &nioc);
1260 * kill knowledge of next, this one is a goner
1262 as_remove_queued_request(q, next);
1263 as_put_io_context(next);
1265 RQ_SET_STATE(next, AS_RQ_MERGED);
1269 * This is executed in a "deferred" process context, by kblockd. It calls the
1270 * driver's request_fn so the driver can submit that request.
1272 * IMPORTANT! This guy will reenter the elevator, so set up all queue global
1273 * state before calling, and don't rely on any state over calls.
1275 * FIXME! dispatch queue is not a queue at all!
1277 static void as_work_handler(struct work_struct *work)
1279 struct as_data *ad = container_of(work, struct as_data, antic_work);
1280 struct request_queue *q = ad->q;
1281 unsigned long flags;
1283 spin_lock_irqsave(q->queue_lock, flags);
1284 blk_start_queueing(q);
1285 spin_unlock_irqrestore(q->queue_lock, flags);
1288 static int as_may_queue(request_queue_t *q, int rw)
1290 int ret = ELV_MQUEUE_MAY;
1291 struct as_data *ad = q->elevator->elevator_data;
1292 struct io_context *ioc;
1293 if (ad->antic_status == ANTIC_WAIT_REQ ||
1294 ad->antic_status == ANTIC_WAIT_NEXT) {
1295 ioc = as_get_io_context(q->node);
1296 if (ad->io_context == ioc)
1297 ret = ELV_MQUEUE_MUST;
1298 put_io_context(ioc);
1301 return ret;
1304 static void as_exit_queue(elevator_t *e)
1306 struct as_data *ad = e->elevator_data;
1308 del_timer_sync(&ad->antic_timer);
1309 kblockd_flush();
1311 BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC]));
1312 BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC]));
1314 put_io_context(ad->io_context);
1315 kfree(ad);
1319 * initialize elevator private data (as_data).
1321 static void *as_init_queue(request_queue_t *q)
1323 struct as_data *ad;
1325 ad = kmalloc_node(sizeof(*ad), GFP_KERNEL, q->node);
1326 if (!ad)
1327 return NULL;
1328 memset(ad, 0, sizeof(*ad));
1330 ad->q = q; /* Identify what queue the data belongs to */
1332 /* anticipatory scheduling helpers */
1333 ad->antic_timer.function = as_antic_timeout;
1334 ad->antic_timer.data = (unsigned long)q;
1335 init_timer(&ad->antic_timer);
1336 INIT_WORK(&ad->antic_work, as_work_handler);
1338 INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]);
1339 INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]);
1340 ad->sort_list[REQ_SYNC] = RB_ROOT;
1341 ad->sort_list[REQ_ASYNC] = RB_ROOT;
1342 ad->fifo_expire[REQ_SYNC] = default_read_expire;
1343 ad->fifo_expire[REQ_ASYNC] = default_write_expire;
1344 ad->antic_expire = default_antic_expire;
1345 ad->batch_expire[REQ_SYNC] = default_read_batch_expire;
1346 ad->batch_expire[REQ_ASYNC] = default_write_batch_expire;
1348 ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC];
1349 ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10;
1350 if (ad->write_batch_count < 2)
1351 ad->write_batch_count = 2;
1353 return ad;
1357 * sysfs parts below
1360 static ssize_t
1361 as_var_show(unsigned int var, char *page)
1363 return sprintf(page, "%d\n", var);
1366 static ssize_t
1367 as_var_store(unsigned long *var, const char *page, size_t count)
1369 char *p = (char *) page;
1371 *var = simple_strtoul(p, &p, 10);
1372 return count;
1375 static ssize_t est_time_show(elevator_t *e, char *page)
1377 struct as_data *ad = e->elevator_data;
1378 int pos = 0;
1380 pos += sprintf(page+pos, "%lu %% exit probability\n",
1381 100*ad->exit_prob/256);
1382 pos += sprintf(page+pos, "%lu %% probability of exiting without a "
1383 "cooperating process submitting IO\n",
1384 100*ad->exit_no_coop/256);
1385 pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean);
1386 pos += sprintf(page+pos, "%llu sectors new seek distance\n",
1387 (unsigned long long)ad->new_seek_mean);
1389 return pos;
1392 #define SHOW_FUNCTION(__FUNC, __VAR) \
1393 static ssize_t __FUNC(elevator_t *e, char *page) \
1395 struct as_data *ad = e->elevator_data; \
1396 return as_var_show(jiffies_to_msecs((__VAR)), (page)); \
1398 SHOW_FUNCTION(as_read_expire_show, ad->fifo_expire[REQ_SYNC]);
1399 SHOW_FUNCTION(as_write_expire_show, ad->fifo_expire[REQ_ASYNC]);
1400 SHOW_FUNCTION(as_antic_expire_show, ad->antic_expire);
1401 SHOW_FUNCTION(as_read_batch_expire_show, ad->batch_expire[REQ_SYNC]);
1402 SHOW_FUNCTION(as_write_batch_expire_show, ad->batch_expire[REQ_ASYNC]);
1403 #undef SHOW_FUNCTION
1405 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
1406 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
1408 struct as_data *ad = e->elevator_data; \
1409 int ret = as_var_store(__PTR, (page), count); \
1410 if (*(__PTR) < (MIN)) \
1411 *(__PTR) = (MIN); \
1412 else if (*(__PTR) > (MAX)) \
1413 *(__PTR) = (MAX); \
1414 *(__PTR) = msecs_to_jiffies(*(__PTR)); \
1415 return ret; \
1417 STORE_FUNCTION(as_read_expire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX);
1418 STORE_FUNCTION(as_write_expire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX);
1419 STORE_FUNCTION(as_antic_expire_store, &ad->antic_expire, 0, INT_MAX);
1420 STORE_FUNCTION(as_read_batch_expire_store,
1421 &ad->batch_expire[REQ_SYNC], 0, INT_MAX);
1422 STORE_FUNCTION(as_write_batch_expire_store,
1423 &ad->batch_expire[REQ_ASYNC], 0, INT_MAX);
1424 #undef STORE_FUNCTION
1426 #define AS_ATTR(name) \
1427 __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store)
1429 static struct elv_fs_entry as_attrs[] = {
1430 __ATTR_RO(est_time),
1431 AS_ATTR(read_expire),
1432 AS_ATTR(write_expire),
1433 AS_ATTR(antic_expire),
1434 AS_ATTR(read_batch_expire),
1435 AS_ATTR(write_batch_expire),
1436 __ATTR_NULL
1439 static struct elevator_type iosched_as = {
1440 .ops = {
1441 .elevator_merge_fn = as_merge,
1442 .elevator_merged_fn = as_merged_request,
1443 .elevator_merge_req_fn = as_merged_requests,
1444 .elevator_dispatch_fn = as_dispatch_request,
1445 .elevator_add_req_fn = as_add_request,
1446 .elevator_activate_req_fn = as_activate_request,
1447 .elevator_deactivate_req_fn = as_deactivate_request,
1448 .elevator_queue_empty_fn = as_queue_empty,
1449 .elevator_completed_req_fn = as_completed_request,
1450 .elevator_former_req_fn = elv_rb_former_request,
1451 .elevator_latter_req_fn = elv_rb_latter_request,
1452 .elevator_may_queue_fn = as_may_queue,
1453 .elevator_init_fn = as_init_queue,
1454 .elevator_exit_fn = as_exit_queue,
1455 .trim = as_trim,
1458 .elevator_attrs = as_attrs,
1459 .elevator_name = "anticipatory",
1460 .elevator_owner = THIS_MODULE,
1463 static int __init as_init(void)
1465 return elv_register(&iosched_as);
1468 static void __exit as_exit(void)
1470 DECLARE_COMPLETION_ONSTACK(all_gone);
1471 elv_unregister(&iosched_as);
1472 ioc_gone = &all_gone;
1473 /* ioc_gone's update must be visible before reading ioc_count */
1474 smp_wmb();
1475 if (elv_ioc_count_read(ioc_count))
1476 wait_for_completion(ioc_gone);
1477 synchronize_rcu();
1480 module_init(as_init);
1481 module_exit(as_exit);
1483 MODULE_AUTHOR("Nick Piggin");
1484 MODULE_LICENSE("GPL");
1485 MODULE_DESCRIPTION("anticipatory IO scheduler");