2 * Anticipatory & deadline i/o scheduler.
4 * Copyright (C) 2002 Jens Axboe <axboe@kernel.dk>
5 * Nick Piggin <nickpiggin@yahoo.com.au>
8 #include <linux/kernel.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>
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 */
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
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
100 unsigned long exit_no_coop; /* probablility an exited task will
101 not be part of a later cooperating
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 */
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;
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
140 AS_RQ_DISPATCHED, /* On the dispatch list. It belongs to the
142 AS_RQ_PRESCHED, /* Debug poisoning for requests being used */
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)
166 elv_ioc_count_dec(ioc_count);
167 if (ioc_gone && !elv_ioc_count_read(ioc_count))
171 static void as_trim(struct io_context *ioc)
173 spin_lock(&ioc->lock);
175 free_as_io_context(ioc->aic);
177 spin_unlock(&ioc->lock);
180 /* Called when the task exits */
181 static void exit_as_io_context(struct as_io_context *aic)
183 WARN_ON(!test_bit(AS_TASK_RUNNING, &aic->state));
184 clear_bit(AS_TASK_RUNNING, &aic->state);
187 static struct as_io_context *alloc_as_io_context(void)
189 struct as_io_context *ret;
191 ret = kmalloc(sizeof(*ret), GFP_ATOMIC);
193 ret->dtor = free_as_io_context;
194 ret->exit = exit_as_io_context;
195 ret->state = 1 << AS_TASK_RUNNING;
196 atomic_set(&ret->nr_queued, 0);
197 atomic_set(&ret->nr_dispatched, 0);
198 spin_lock_init(&ret->lock);
199 ret->ttime_total = 0;
200 ret->ttime_samples = 0;
203 ret->seek_samples = 0;
205 elv_ioc_count_inc(ioc_count);
212 * If the current task has no AS IO context then create one and initialise it.
213 * Then take a ref on the task's io context and return it.
215 static struct io_context *as_get_io_context(int node)
217 struct io_context *ioc = get_io_context(GFP_ATOMIC, node);
218 if (ioc && !ioc->aic) {
219 ioc->aic = alloc_as_io_context();
228 static void as_put_io_context(struct request *rq)
230 struct as_io_context *aic;
232 if (unlikely(!RQ_IOC(rq)))
235 aic = RQ_IOC(rq)->aic;
237 if (rq_is_sync(rq) && aic) {
238 spin_lock(&aic->lock);
239 set_bit(AS_TASK_IORUNNING, &aic->state);
240 aic->last_end_request = jiffies;
241 spin_unlock(&aic->lock);
244 put_io_context(RQ_IOC(rq));
248 * rb tree support functions
250 #define RQ_RB_ROOT(ad, rq) (&(ad)->sort_list[rq_is_sync((rq))])
252 static void as_add_rq_rb(struct as_data *ad, struct request *rq)
254 struct request *alias;
256 while ((unlikely(alias = elv_rb_add(RQ_RB_ROOT(ad, rq), rq)))) {
257 as_move_to_dispatch(ad, alias);
262 static inline void as_del_rq_rb(struct as_data *ad, struct request *rq)
264 elv_rb_del(RQ_RB_ROOT(ad, rq), rq);
268 * IO Scheduler proper
271 #define MAXBACK (1024 * 1024) /*
272 * Maximum distance the disk will go backward
276 #define BACK_PENALTY 2
279 * as_choose_req selects the preferred one of two requests of the same data_dir
280 * ignoring time - eg. timeouts, which is the job of as_dispatch_request
282 static struct request *
283 as_choose_req(struct as_data *ad, struct request *rq1, struct request *rq2)
286 sector_t last, s1, s2, d1, d2;
287 int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */
288 const sector_t maxback = MAXBACK;
290 if (rq1 == NULL || rq1 == rq2)
295 data_dir = rq_is_sync(rq1);
297 last = ad->last_sector[data_dir];
301 BUG_ON(data_dir != rq_is_sync(rq2));
304 * Strict one way elevator _except_ in the case where we allow
305 * short backward seeks which are biased as twice the cost of a
306 * similar forward seek.
310 else if (s1+maxback >= last)
311 d1 = (last - s1)*BACK_PENALTY;
314 d1 = 0; /* shut up, gcc */
319 else if (s2+maxback >= last)
320 d2 = (last - s2)*BACK_PENALTY;
326 /* Found required data */
327 if (!r1_wrap && r2_wrap)
329 else if (!r2_wrap && r1_wrap)
331 else if (r1_wrap && r2_wrap) {
332 /* both behind the head */
339 /* Both requests in front of the head */
353 * as_find_next_rq finds the next request after @prev in elevator order.
354 * this with as_choose_req form the basis for how the scheduler chooses
355 * what request to process next. Anticipation works on top of this.
357 static struct request *
358 as_find_next_rq(struct as_data *ad, struct request *last)
360 struct rb_node *rbnext = rb_next(&last->rb_node);
361 struct rb_node *rbprev = rb_prev(&last->rb_node);
362 struct request *next = NULL, *prev = NULL;
364 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
367 prev = rb_entry_rq(rbprev);
370 next = rb_entry_rq(rbnext);
372 const int data_dir = rq_is_sync(last);
374 rbnext = rb_first(&ad->sort_list[data_dir]);
375 if (rbnext && rbnext != &last->rb_node)
376 next = rb_entry_rq(rbnext);
379 return as_choose_req(ad, next, prev);
383 * anticipatory scheduling functions follow
387 * as_antic_expired tells us when we have anticipated too long.
388 * The funny "absolute difference" math on the elapsed time is to handle
389 * jiffy wraps, and disks which have been idle for 0x80000000 jiffies.
391 static int as_antic_expired(struct as_data *ad)
395 delta_jif = jiffies - ad->antic_start;
396 if (unlikely(delta_jif < 0))
397 delta_jif = -delta_jif;
398 if (delta_jif < ad->antic_expire)
405 * as_antic_waitnext starts anticipating that a nice request will soon be
406 * submitted. See also as_antic_waitreq
408 static void as_antic_waitnext(struct as_data *ad)
410 unsigned long timeout;
412 BUG_ON(ad->antic_status != ANTIC_OFF
413 && ad->antic_status != ANTIC_WAIT_REQ);
415 timeout = ad->antic_start + ad->antic_expire;
417 mod_timer(&ad->antic_timer, timeout);
419 ad->antic_status = ANTIC_WAIT_NEXT;
423 * as_antic_waitreq starts anticipating. We don't start timing the anticipation
424 * until the request that we're anticipating on has finished. This means we
425 * are timing from when the candidate process wakes up hopefully.
427 static void as_antic_waitreq(struct as_data *ad)
429 BUG_ON(ad->antic_status == ANTIC_FINISHED);
430 if (ad->antic_status == ANTIC_OFF) {
431 if (!ad->io_context || ad->ioc_finished)
432 as_antic_waitnext(ad);
434 ad->antic_status = ANTIC_WAIT_REQ;
439 * This is called directly by the functions in this file to stop anticipation.
440 * We kill the timer and schedule a call to the request_fn asap.
442 static void as_antic_stop(struct as_data *ad)
444 int status = ad->antic_status;
446 if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) {
447 if (status == ANTIC_WAIT_NEXT)
448 del_timer(&ad->antic_timer);
449 ad->antic_status = ANTIC_FINISHED;
450 /* see as_work_handler */
451 kblockd_schedule_work(&ad->antic_work);
456 * as_antic_timeout is the timer function set by as_antic_waitnext.
458 static void as_antic_timeout(unsigned long data)
460 struct request_queue *q = (struct request_queue *)data;
461 struct as_data *ad = q->elevator->elevator_data;
464 spin_lock_irqsave(q->queue_lock, flags);
465 if (ad->antic_status == ANTIC_WAIT_REQ
466 || ad->antic_status == ANTIC_WAIT_NEXT) {
467 struct as_io_context *aic;
468 spin_lock(&ad->io_context->lock);
469 aic = ad->io_context->aic;
471 ad->antic_status = ANTIC_FINISHED;
472 kblockd_schedule_work(&ad->antic_work);
474 if (aic->ttime_samples == 0) {
475 /* process anticipated on has exited or timed out*/
476 ad->exit_prob = (7*ad->exit_prob + 256)/8;
478 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
479 /* process not "saved" by a cooperating request */
480 ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8;
482 spin_unlock(&ad->io_context->lock);
484 spin_unlock_irqrestore(q->queue_lock, flags);
487 static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic,
490 /* fixed point: 1.0 == 1<<8 */
491 if (aic->ttime_samples == 0) {
492 ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8;
493 ad->new_ttime_mean = ad->new_ttime_total / 256;
495 ad->exit_prob = (7*ad->exit_prob)/8;
497 aic->ttime_samples = (7*aic->ttime_samples + 256) / 8;
498 aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8;
499 aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples;
502 static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic,
507 if (aic->seek_samples == 0) {
508 ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8;
509 ad->new_seek_mean = ad->new_seek_total / 256;
513 * Don't allow the seek distance to get too large from the
514 * odd fragment, pagein, etc
516 if (aic->seek_samples <= 60) /* second&third seek */
517 sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024);
519 sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64);
521 aic->seek_samples = (7*aic->seek_samples + 256) / 8;
522 aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8;
523 total = aic->seek_total + (aic->seek_samples/2);
524 do_div(total, aic->seek_samples);
525 aic->seek_mean = (sector_t)total;
529 * as_update_iohist keeps a decaying histogram of IO thinktimes, and
530 * updates @aic->ttime_mean based on that. It is called when a new
533 static void as_update_iohist(struct as_data *ad, struct as_io_context *aic,
536 int data_dir = rq_is_sync(rq);
537 unsigned long thinktime = 0;
543 if (data_dir == REQ_SYNC) {
544 unsigned long in_flight = atomic_read(&aic->nr_queued)
545 + atomic_read(&aic->nr_dispatched);
546 spin_lock(&aic->lock);
547 if (test_bit(AS_TASK_IORUNNING, &aic->state) ||
548 test_bit(AS_TASK_IOSTARTED, &aic->state)) {
549 /* Calculate read -> read thinktime */
550 if (test_bit(AS_TASK_IORUNNING, &aic->state)
552 thinktime = jiffies - aic->last_end_request;
553 thinktime = min(thinktime, MAX_THINKTIME-1);
555 as_update_thinktime(ad, aic, thinktime);
557 /* Calculate read -> read seek distance */
558 if (aic->last_request_pos < rq->sector)
559 seek_dist = rq->sector - aic->last_request_pos;
561 seek_dist = aic->last_request_pos - rq->sector;
562 as_update_seekdist(ad, aic, seek_dist);
564 aic->last_request_pos = rq->sector + rq->nr_sectors;
565 set_bit(AS_TASK_IOSTARTED, &aic->state);
566 spin_unlock(&aic->lock);
571 * as_close_req decides if one request is considered "close" to the
572 * previous one issued.
574 static int as_close_req(struct as_data *ad, struct as_io_context *aic,
577 unsigned long delay; /* jiffies */
578 sector_t last = ad->last_sector[ad->batch_data_dir];
579 sector_t next = rq->sector;
580 sector_t delta; /* acceptable close offset (in sectors) */
583 if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished)
586 delay = jiffies - ad->antic_start;
590 else if (delay <= (20 * HZ / 1000) && delay <= ad->antic_expire)
591 delta = 8192 << delay;
595 if ((last <= next + (delta>>1)) && (next <= last + delta))
603 if (aic->seek_samples == 0) {
605 * Process has just started IO. Use past statistics to
606 * gauge success possibility
608 if (ad->new_seek_mean > s) {
609 /* this request is better than what we're expecting */
614 if (aic->seek_mean > s) {
615 /* this request is better than what we're expecting */
624 * as_can_break_anticipation returns true if we have been anticipating this
627 * It also returns true if the process against which we are anticipating
628 * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to
629 * dispatch it ASAP, because we know that application will not be submitting
632 * If the task which has submitted the request has exited, break anticipation.
634 * If this task has queued some other IO, do not enter enticipation.
636 static int as_can_break_anticipation(struct as_data *ad, struct request *rq)
638 struct io_context *ioc;
639 struct as_io_context *aic;
641 ioc = ad->io_context;
643 spin_lock(&ioc->lock);
645 if (rq && ioc == RQ_IOC(rq)) {
646 /* request from same process */
647 spin_unlock(&ioc->lock);
651 if (ad->ioc_finished && as_antic_expired(ad)) {
653 * In this situation status should really be FINISHED,
654 * however the timer hasn't had the chance to run yet.
656 spin_unlock(&ioc->lock);
662 spin_unlock(&ioc->lock);
666 if (atomic_read(&aic->nr_queued) > 0) {
667 /* process has more requests queued */
668 spin_unlock(&ioc->lock);
672 if (atomic_read(&aic->nr_dispatched) > 0) {
673 /* process has more requests dispatched */
674 spin_unlock(&ioc->lock);
678 if (rq && rq_is_sync(rq) && as_close_req(ad, aic, rq)) {
680 * Found a close request that is not one of ours.
682 * This makes close requests from another process update
683 * our IO history. Is generally useful when there are
684 * two or more cooperating processes working in the same
687 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
688 if (aic->ttime_samples == 0)
689 ad->exit_prob = (7*ad->exit_prob + 256)/8;
691 ad->exit_no_coop = (7*ad->exit_no_coop)/8;
694 as_update_iohist(ad, aic, rq);
695 spin_unlock(&ioc->lock);
699 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
700 /* process anticipated on has exited */
701 if (aic->ttime_samples == 0)
702 ad->exit_prob = (7*ad->exit_prob + 256)/8;
704 if (ad->exit_no_coop > 128) {
705 spin_unlock(&ioc->lock);
710 if (aic->ttime_samples == 0) {
711 if (ad->new_ttime_mean > ad->antic_expire) {
712 spin_unlock(&ioc->lock);
715 if (ad->exit_prob * ad->exit_no_coop > 128*256) {
716 spin_unlock(&ioc->lock);
719 } else if (aic->ttime_mean > ad->antic_expire) {
720 /* the process thinks too much between requests */
721 spin_unlock(&ioc->lock);
724 spin_unlock(&ioc->lock);
729 * as_can_anticipate indicates whether we should either run rq
730 * or keep anticipating a better request.
732 static int as_can_anticipate(struct as_data *ad, struct request *rq)
736 * Last request submitted was a write
740 if (ad->antic_status == ANTIC_FINISHED)
742 * Don't restart if we have just finished. Run the next request
746 if (as_can_break_anticipation(ad, rq))
748 * This request is a good candidate. Don't keep anticipating,
754 * OK from here, we haven't finished, and don't have a decent request!
755 * Status is either ANTIC_OFF so start waiting,
756 * ANTIC_WAIT_REQ so continue waiting for request to finish
757 * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request.
764 * as_update_rq must be called whenever a request (rq) is added to
765 * the sort_list. This function keeps caches up to date, and checks if the
766 * request might be one we are "anticipating"
768 static void as_update_rq(struct as_data *ad, struct request *rq)
770 const int data_dir = rq_is_sync(rq);
772 /* keep the next_rq cache up to date */
773 ad->next_rq[data_dir] = as_choose_req(ad, rq, ad->next_rq[data_dir]);
776 * have we been anticipating this request?
777 * or does it come from the same process as the one we are anticipating
780 if (ad->antic_status == ANTIC_WAIT_REQ
781 || ad->antic_status == ANTIC_WAIT_NEXT) {
782 if (as_can_break_anticipation(ad, rq))
788 * Gathers timings and resizes the write batch automatically
790 static void update_write_batch(struct as_data *ad)
792 unsigned long batch = ad->batch_expire[REQ_ASYNC];
795 write_time = (jiffies - ad->current_batch_expires) + batch;
799 if (write_time > batch && !ad->write_batch_idled) {
800 if (write_time > batch * 3)
801 ad->write_batch_count /= 2;
803 ad->write_batch_count--;
804 } else if (write_time < batch && ad->current_write_count == 0) {
805 if (batch > write_time * 3)
806 ad->write_batch_count *= 2;
808 ad->write_batch_count++;
811 if (ad->write_batch_count < 1)
812 ad->write_batch_count = 1;
816 * as_completed_request is to be called when a request has completed and
817 * returned something to the requesting process, be it an error or data.
819 static void as_completed_request(struct request_queue *q, struct request *rq)
821 struct as_data *ad = q->elevator->elevator_data;
823 WARN_ON(!list_empty(&rq->queuelist));
825 if (RQ_STATE(rq) != AS_RQ_REMOVED) {
826 printk("rq->state %d\n", RQ_STATE(rq));
831 if (ad->changed_batch && ad->nr_dispatched == 1) {
832 kblockd_schedule_work(&ad->antic_work);
833 ad->changed_batch = 0;
835 if (ad->batch_data_dir == REQ_SYNC)
838 WARN_ON(ad->nr_dispatched == 0);
842 * Start counting the batch from when a request of that direction is
843 * actually serviced. This should help devices with big TCQ windows
844 * and writeback caches
846 if (ad->new_batch && ad->batch_data_dir == rq_is_sync(rq)) {
847 update_write_batch(ad);
848 ad->current_batch_expires = jiffies +
849 ad->batch_expire[REQ_SYNC];
853 if (ad->io_context == RQ_IOC(rq) && ad->io_context) {
854 ad->antic_start = jiffies;
855 ad->ioc_finished = 1;
856 if (ad->antic_status == ANTIC_WAIT_REQ) {
858 * We were waiting on this request, now anticipate
861 as_antic_waitnext(ad);
865 as_put_io_context(rq);
867 RQ_SET_STATE(rq, AS_RQ_POSTSCHED);
871 * as_remove_queued_request removes a request from the pre dispatch queue
872 * without updating refcounts. It is expected the caller will drop the
873 * reference unless it replaces the request at somepart of the elevator
874 * (ie. the dispatch queue)
876 static void as_remove_queued_request(struct request_queue *q,
879 const int data_dir = rq_is_sync(rq);
880 struct as_data *ad = q->elevator->elevator_data;
881 struct io_context *ioc;
883 WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED);
886 if (ioc && ioc->aic) {
887 BUG_ON(!atomic_read(&ioc->aic->nr_queued));
888 atomic_dec(&ioc->aic->nr_queued);
892 * Update the "next_rq" cache if we are about to remove its
895 if (ad->next_rq[data_dir] == rq)
896 ad->next_rq[data_dir] = as_find_next_rq(ad, rq);
899 as_del_rq_rb(ad, rq);
903 * as_fifo_expired returns 0 if there are no expired requests on the fifo,
904 * 1 otherwise. It is ratelimited so that we only perform the check once per
905 * `fifo_expire' interval. Otherwise a large number of expired requests
906 * would create a hopeless seekstorm.
908 * See as_antic_expired comment.
910 static int as_fifo_expired(struct as_data *ad, int adir)
915 delta_jif = jiffies - ad->last_check_fifo[adir];
916 if (unlikely(delta_jif < 0))
917 delta_jif = -delta_jif;
918 if (delta_jif < ad->fifo_expire[adir])
921 ad->last_check_fifo[adir] = jiffies;
923 if (list_empty(&ad->fifo_list[adir]))
926 rq = rq_entry_fifo(ad->fifo_list[adir].next);
928 return time_after(jiffies, rq_fifo_time(rq));
932 * as_batch_expired returns true if the current batch has expired. A batch
933 * is a set of reads or a set of writes.
935 static inline int as_batch_expired(struct as_data *ad)
937 if (ad->changed_batch || ad->new_batch)
940 if (ad->batch_data_dir == REQ_SYNC)
941 /* TODO! add a check so a complete fifo gets written? */
942 return time_after(jiffies, ad->current_batch_expires);
944 return time_after(jiffies, ad->current_batch_expires)
945 || ad->current_write_count == 0;
949 * move an entry to dispatch queue
951 static void as_move_to_dispatch(struct as_data *ad, struct request *rq)
953 const int data_dir = rq_is_sync(rq);
955 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
958 ad->antic_status = ANTIC_OFF;
961 * This has to be set in order to be correctly updated by
964 ad->last_sector[data_dir] = rq->sector + rq->nr_sectors;
966 if (data_dir == REQ_SYNC) {
967 struct io_context *ioc = RQ_IOC(rq);
968 /* In case we have to anticipate after this */
969 copy_io_context(&ad->io_context, &ioc);
971 if (ad->io_context) {
972 put_io_context(ad->io_context);
973 ad->io_context = NULL;
976 if (ad->current_write_count != 0)
977 ad->current_write_count--;
979 ad->ioc_finished = 0;
981 ad->next_rq[data_dir] = as_find_next_rq(ad, rq);
984 * take it off the sort and fifo list, add to dispatch queue
986 as_remove_queued_request(ad->q, rq);
987 WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED);
989 elv_dispatch_sort(ad->q, rq);
991 RQ_SET_STATE(rq, AS_RQ_DISPATCHED);
992 if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
993 atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched);
998 * as_dispatch_request selects the best request according to
999 * read/write expire, batch expire, etc, and moves it to the dispatch
1000 * queue. Returns 1 if a request was found, 0 otherwise.
1002 static int as_dispatch_request(struct request_queue *q, int force)
1004 struct as_data *ad = q->elevator->elevator_data;
1005 const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]);
1006 const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]);
1009 if (unlikely(force)) {
1011 * Forced dispatch, accounting is useless. Reset
1012 * accounting states and dump fifo_lists. Note that
1013 * batch_data_dir is reset to REQ_SYNC to avoid
1014 * screwing write batch accounting as write batch
1015 * accounting occurs on W->R transition.
1019 ad->batch_data_dir = REQ_SYNC;
1020 ad->changed_batch = 0;
1023 while (ad->next_rq[REQ_SYNC]) {
1024 as_move_to_dispatch(ad, ad->next_rq[REQ_SYNC]);
1027 ad->last_check_fifo[REQ_SYNC] = jiffies;
1029 while (ad->next_rq[REQ_ASYNC]) {
1030 as_move_to_dispatch(ad, ad->next_rq[REQ_ASYNC]);
1033 ad->last_check_fifo[REQ_ASYNC] = jiffies;
1038 /* Signal that the write batch was uncontended, so we can't time it */
1039 if (ad->batch_data_dir == REQ_ASYNC && !reads) {
1040 if (ad->current_write_count == 0 || !writes)
1041 ad->write_batch_idled = 1;
1044 if (!(reads || writes)
1045 || ad->antic_status == ANTIC_WAIT_REQ
1046 || ad->antic_status == ANTIC_WAIT_NEXT
1047 || ad->changed_batch)
1050 if (!(reads && writes && as_batch_expired(ad))) {
1052 * batch is still running or no reads or no writes
1054 rq = ad->next_rq[ad->batch_data_dir];
1056 if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) {
1057 if (as_fifo_expired(ad, REQ_SYNC))
1060 if (as_can_anticipate(ad, rq)) {
1061 as_antic_waitreq(ad);
1067 /* we have a "next request" */
1068 if (reads && !writes)
1069 ad->current_batch_expires =
1070 jiffies + ad->batch_expire[REQ_SYNC];
1071 goto dispatch_request;
1076 * at this point we are not running a batch. select the appropriate
1077 * data direction (read / write)
1081 BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_SYNC]));
1083 if (writes && ad->batch_data_dir == REQ_SYNC)
1085 * Last batch was a read, switch to writes
1087 goto dispatch_writes;
1089 if (ad->batch_data_dir == REQ_ASYNC) {
1090 WARN_ON(ad->new_batch);
1091 ad->changed_batch = 1;
1093 ad->batch_data_dir = REQ_SYNC;
1094 rq = rq_entry_fifo(ad->fifo_list[REQ_SYNC].next);
1095 ad->last_check_fifo[ad->batch_data_dir] = jiffies;
1096 goto dispatch_request;
1100 * the last batch was a read
1105 BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_ASYNC]));
1107 if (ad->batch_data_dir == REQ_SYNC) {
1108 ad->changed_batch = 1;
1111 * new_batch might be 1 when the queue runs out of
1112 * reads. A subsequent submission of a write might
1113 * cause a change of batch before the read is finished.
1117 ad->batch_data_dir = REQ_ASYNC;
1118 ad->current_write_count = ad->write_batch_count;
1119 ad->write_batch_idled = 0;
1120 rq = rq_entry_fifo(ad->fifo_list[REQ_ASYNC].next);
1121 ad->last_check_fifo[REQ_ASYNC] = jiffies;
1122 goto dispatch_request;
1130 * If a request has expired, service it.
1133 if (as_fifo_expired(ad, ad->batch_data_dir)) {
1135 rq = rq_entry_fifo(ad->fifo_list[ad->batch_data_dir].next);
1138 if (ad->changed_batch) {
1139 WARN_ON(ad->new_batch);
1141 if (ad->nr_dispatched)
1144 if (ad->batch_data_dir == REQ_ASYNC)
1145 ad->current_batch_expires = jiffies +
1146 ad->batch_expire[REQ_ASYNC];
1150 ad->changed_batch = 0;
1154 * rq is the selected appropriate request.
1156 as_move_to_dispatch(ad, rq);
1162 * add rq to rbtree and fifo
1164 static void as_add_request(struct request_queue *q, struct request *rq)
1166 struct as_data *ad = q->elevator->elevator_data;
1169 RQ_SET_STATE(rq, AS_RQ_NEW);
1171 data_dir = rq_is_sync(rq);
1173 rq->elevator_private = as_get_io_context(q->node);
1176 as_update_iohist(ad, RQ_IOC(rq)->aic, rq);
1177 atomic_inc(&RQ_IOC(rq)->aic->nr_queued);
1180 as_add_rq_rb(ad, rq);
1183 * set expire time and add to fifo list
1185 rq_set_fifo_time(rq, jiffies + ad->fifo_expire[data_dir]);
1186 list_add_tail(&rq->queuelist, &ad->fifo_list[data_dir]);
1188 as_update_rq(ad, rq); /* keep state machine up to date */
1189 RQ_SET_STATE(rq, AS_RQ_QUEUED);
1192 static void as_activate_request(struct request_queue *q, struct request *rq)
1194 WARN_ON(RQ_STATE(rq) != AS_RQ_DISPATCHED);
1195 RQ_SET_STATE(rq, AS_RQ_REMOVED);
1196 if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
1197 atomic_dec(&RQ_IOC(rq)->aic->nr_dispatched);
1200 static void as_deactivate_request(struct request_queue *q, struct request *rq)
1202 WARN_ON(RQ_STATE(rq) != AS_RQ_REMOVED);
1203 RQ_SET_STATE(rq, AS_RQ_DISPATCHED);
1204 if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
1205 atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched);
1209 * as_queue_empty tells us if there are requests left in the device. It may
1210 * not be the case that a driver can get the next request even if the queue
1211 * is not empty - it is used in the block layer to check for plugging and
1212 * merging opportunities
1214 static int as_queue_empty(struct request_queue *q)
1216 struct as_data *ad = q->elevator->elevator_data;
1218 return list_empty(&ad->fifo_list[REQ_ASYNC])
1219 && list_empty(&ad->fifo_list[REQ_SYNC]);
1223 as_merge(struct request_queue *q, struct request **req, struct bio *bio)
1225 struct as_data *ad = q->elevator->elevator_data;
1226 sector_t rb_key = bio->bi_sector + bio_sectors(bio);
1227 struct request *__rq;
1230 * check for front merge
1232 __rq = elv_rb_find(&ad->sort_list[bio_data_dir(bio)], rb_key);
1233 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1235 return ELEVATOR_FRONT_MERGE;
1238 return ELEVATOR_NO_MERGE;
1241 static void as_merged_request(struct request_queue *q, struct request *req,
1244 struct as_data *ad = q->elevator->elevator_data;
1247 * if the merge was a front merge, we need to reposition request
1249 if (type == ELEVATOR_FRONT_MERGE) {
1250 as_del_rq_rb(ad, req);
1251 as_add_rq_rb(ad, req);
1253 * Note! At this stage of this and the next function, our next
1254 * request may not be optimal - eg the request may have "grown"
1255 * behind the disk head. We currently don't bother adjusting.
1260 static void as_merged_requests(struct request_queue *q, struct request *req,
1261 struct request *next)
1264 * if next expires before rq, assign its expire time to arq
1265 * and move into next position (next will be deleted) in fifo
1267 if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
1268 if (time_before(rq_fifo_time(next), rq_fifo_time(req))) {
1269 struct io_context *rioc = RQ_IOC(req);
1270 struct io_context *nioc = RQ_IOC(next);
1272 list_move(&req->queuelist, &next->queuelist);
1273 rq_set_fifo_time(req, rq_fifo_time(next));
1275 * Don't copy here but swap, because when anext is
1276 * removed below, it must contain the unused context
1279 double_spin_lock(&rioc->lock, &nioc->lock,
1281 swap_io_context(&rioc, &nioc);
1282 double_spin_unlock(&rioc->lock, &nioc->lock,
1289 * kill knowledge of next, this one is a goner
1291 as_remove_queued_request(q, next);
1292 as_put_io_context(next);
1294 RQ_SET_STATE(next, AS_RQ_MERGED);
1298 * This is executed in a "deferred" process context, by kblockd. It calls the
1299 * driver's request_fn so the driver can submit that request.
1301 * IMPORTANT! This guy will reenter the elevator, so set up all queue global
1302 * state before calling, and don't rely on any state over calls.
1304 * FIXME! dispatch queue is not a queue at all!
1306 static void as_work_handler(struct work_struct *work)
1308 struct as_data *ad = container_of(work, struct as_data, antic_work);
1309 struct request_queue *q = ad->q;
1310 unsigned long flags;
1312 spin_lock_irqsave(q->queue_lock, flags);
1313 blk_start_queueing(q);
1314 spin_unlock_irqrestore(q->queue_lock, flags);
1317 static int as_may_queue(struct request_queue *q, int rw)
1319 int ret = ELV_MQUEUE_MAY;
1320 struct as_data *ad = q->elevator->elevator_data;
1321 struct io_context *ioc;
1322 if (ad->antic_status == ANTIC_WAIT_REQ ||
1323 ad->antic_status == ANTIC_WAIT_NEXT) {
1324 ioc = as_get_io_context(q->node);
1325 if (ad->io_context == ioc)
1326 ret = ELV_MQUEUE_MUST;
1327 put_io_context(ioc);
1333 static void as_exit_queue(elevator_t *e)
1335 struct as_data *ad = e->elevator_data;
1337 del_timer_sync(&ad->antic_timer);
1338 kblockd_flush_work(&ad->antic_work);
1340 BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC]));
1341 BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC]));
1343 put_io_context(ad->io_context);
1348 * initialize elevator private data (as_data).
1350 static void *as_init_queue(struct request_queue *q)
1354 ad = kmalloc_node(sizeof(*ad), GFP_KERNEL | __GFP_ZERO, q->node);
1358 ad->q = q; /* Identify what queue the data belongs to */
1360 /* anticipatory scheduling helpers */
1361 ad->antic_timer.function = as_antic_timeout;
1362 ad->antic_timer.data = (unsigned long)q;
1363 init_timer(&ad->antic_timer);
1364 INIT_WORK(&ad->antic_work, as_work_handler);
1366 INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]);
1367 INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]);
1368 ad->sort_list[REQ_SYNC] = RB_ROOT;
1369 ad->sort_list[REQ_ASYNC] = RB_ROOT;
1370 ad->fifo_expire[REQ_SYNC] = default_read_expire;
1371 ad->fifo_expire[REQ_ASYNC] = default_write_expire;
1372 ad->antic_expire = default_antic_expire;
1373 ad->batch_expire[REQ_SYNC] = default_read_batch_expire;
1374 ad->batch_expire[REQ_ASYNC] = default_write_batch_expire;
1376 ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC];
1377 ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10;
1378 if (ad->write_batch_count < 2)
1379 ad->write_batch_count = 2;
1389 as_var_show(unsigned int var, char *page)
1391 return sprintf(page, "%d\n", var);
1395 as_var_store(unsigned long *var, const char *page, size_t count)
1397 char *p = (char *) page;
1399 *var = simple_strtoul(p, &p, 10);
1403 static ssize_t est_time_show(elevator_t *e, char *page)
1405 struct as_data *ad = e->elevator_data;
1408 pos += sprintf(page+pos, "%lu %% exit probability\n",
1409 100*ad->exit_prob/256);
1410 pos += sprintf(page+pos, "%lu %% probability of exiting without a "
1411 "cooperating process submitting IO\n",
1412 100*ad->exit_no_coop/256);
1413 pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean);
1414 pos += sprintf(page+pos, "%llu sectors new seek distance\n",
1415 (unsigned long long)ad->new_seek_mean);
1420 #define SHOW_FUNCTION(__FUNC, __VAR) \
1421 static ssize_t __FUNC(elevator_t *e, char *page) \
1423 struct as_data *ad = e->elevator_data; \
1424 return as_var_show(jiffies_to_msecs((__VAR)), (page)); \
1426 SHOW_FUNCTION(as_read_expire_show, ad->fifo_expire[REQ_SYNC]);
1427 SHOW_FUNCTION(as_write_expire_show, ad->fifo_expire[REQ_ASYNC]);
1428 SHOW_FUNCTION(as_antic_expire_show, ad->antic_expire);
1429 SHOW_FUNCTION(as_read_batch_expire_show, ad->batch_expire[REQ_SYNC]);
1430 SHOW_FUNCTION(as_write_batch_expire_show, ad->batch_expire[REQ_ASYNC]);
1431 #undef SHOW_FUNCTION
1433 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
1434 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
1436 struct as_data *ad = e->elevator_data; \
1437 int ret = as_var_store(__PTR, (page), count); \
1438 if (*(__PTR) < (MIN)) \
1440 else if (*(__PTR) > (MAX)) \
1442 *(__PTR) = msecs_to_jiffies(*(__PTR)); \
1445 STORE_FUNCTION(as_read_expire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX);
1446 STORE_FUNCTION(as_write_expire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX);
1447 STORE_FUNCTION(as_antic_expire_store, &ad->antic_expire, 0, INT_MAX);
1448 STORE_FUNCTION(as_read_batch_expire_store,
1449 &ad->batch_expire[REQ_SYNC], 0, INT_MAX);
1450 STORE_FUNCTION(as_write_batch_expire_store,
1451 &ad->batch_expire[REQ_ASYNC], 0, INT_MAX);
1452 #undef STORE_FUNCTION
1454 #define AS_ATTR(name) \
1455 __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store)
1457 static struct elv_fs_entry as_attrs[] = {
1458 __ATTR_RO(est_time),
1459 AS_ATTR(read_expire),
1460 AS_ATTR(write_expire),
1461 AS_ATTR(antic_expire),
1462 AS_ATTR(read_batch_expire),
1463 AS_ATTR(write_batch_expire),
1467 static struct elevator_type iosched_as = {
1469 .elevator_merge_fn = as_merge,
1470 .elevator_merged_fn = as_merged_request,
1471 .elevator_merge_req_fn = as_merged_requests,
1472 .elevator_dispatch_fn = as_dispatch_request,
1473 .elevator_add_req_fn = as_add_request,
1474 .elevator_activate_req_fn = as_activate_request,
1475 .elevator_deactivate_req_fn = as_deactivate_request,
1476 .elevator_queue_empty_fn = as_queue_empty,
1477 .elevator_completed_req_fn = as_completed_request,
1478 .elevator_former_req_fn = elv_rb_former_request,
1479 .elevator_latter_req_fn = elv_rb_latter_request,
1480 .elevator_may_queue_fn = as_may_queue,
1481 .elevator_init_fn = as_init_queue,
1482 .elevator_exit_fn = as_exit_queue,
1486 .elevator_attrs = as_attrs,
1487 .elevator_name = "anticipatory",
1488 .elevator_owner = THIS_MODULE,
1491 static int __init as_init(void)
1493 elv_register(&iosched_as);
1498 static void __exit as_exit(void)
1500 DECLARE_COMPLETION_ONSTACK(all_gone);
1501 elv_unregister(&iosched_as);
1502 ioc_gone = &all_gone;
1503 /* ioc_gone's update must be visible before reading ioc_count */
1505 if (elv_ioc_count_read(ioc_count))
1506 wait_for_completion(ioc_gone);
1510 module_init(as_init);
1511 module_exit(as_exit);
1513 MODULE_AUTHOR("Nick Piggin");
1514 MODULE_LICENSE("GPL");
1515 MODULE_DESCRIPTION("anticipatory IO scheduler");