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_irq(&ioc->lock);
175 free_as_io_context(ioc->aic);
177 spin_unlock_irq(&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) {
240 spin_lock_irqsave(&aic->lock, flags);
241 set_bit(AS_TASK_IORUNNING, &aic->state);
242 aic->last_end_request = jiffies;
243 spin_unlock_irqrestore(&aic->lock, flags);
246 put_io_context(RQ_IOC(rq));
250 * rb tree support functions
252 #define RQ_RB_ROOT(ad, rq) (&(ad)->sort_list[rq_is_sync((rq))])
254 static void as_add_rq_rb(struct as_data *ad, struct request *rq)
256 struct request *alias;
258 while ((unlikely(alias = elv_rb_add(RQ_RB_ROOT(ad, rq), rq)))) {
259 as_move_to_dispatch(ad, alias);
264 static inline void as_del_rq_rb(struct as_data *ad, struct request *rq)
266 elv_rb_del(RQ_RB_ROOT(ad, rq), rq);
270 * IO Scheduler proper
273 #define MAXBACK (1024 * 1024) /*
274 * Maximum distance the disk will go backward
278 #define BACK_PENALTY 2
281 * as_choose_req selects the preferred one of two requests of the same data_dir
282 * ignoring time - eg. timeouts, which is the job of as_dispatch_request
284 static struct request *
285 as_choose_req(struct as_data *ad, struct request *rq1, struct request *rq2)
288 sector_t last, s1, s2, d1, d2;
289 int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */
290 const sector_t maxback = MAXBACK;
292 if (rq1 == NULL || rq1 == rq2)
297 data_dir = rq_is_sync(rq1);
299 last = ad->last_sector[data_dir];
303 BUG_ON(data_dir != rq_is_sync(rq2));
306 * Strict one way elevator _except_ in the case where we allow
307 * short backward seeks which are biased as twice the cost of a
308 * similar forward seek.
312 else if (s1+maxback >= last)
313 d1 = (last - s1)*BACK_PENALTY;
316 d1 = 0; /* shut up, gcc */
321 else if (s2+maxback >= last)
322 d2 = (last - s2)*BACK_PENALTY;
328 /* Found required data */
329 if (!r1_wrap && r2_wrap)
331 else if (!r2_wrap && r1_wrap)
333 else if (r1_wrap && r2_wrap) {
334 /* both behind the head */
341 /* Both requests in front of the head */
355 * as_find_next_rq finds the next request after @prev in elevator order.
356 * this with as_choose_req form the basis for how the scheduler chooses
357 * what request to process next. Anticipation works on top of this.
359 static struct request *
360 as_find_next_rq(struct as_data *ad, struct request *last)
362 struct rb_node *rbnext = rb_next(&last->rb_node);
363 struct rb_node *rbprev = rb_prev(&last->rb_node);
364 struct request *next = NULL, *prev = NULL;
366 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
369 prev = rb_entry_rq(rbprev);
372 next = rb_entry_rq(rbnext);
374 const int data_dir = rq_is_sync(last);
376 rbnext = rb_first(&ad->sort_list[data_dir]);
377 if (rbnext && rbnext != &last->rb_node)
378 next = rb_entry_rq(rbnext);
381 return as_choose_req(ad, next, prev);
385 * anticipatory scheduling functions follow
389 * as_antic_expired tells us when we have anticipated too long.
390 * The funny "absolute difference" math on the elapsed time is to handle
391 * jiffy wraps, and disks which have been idle for 0x80000000 jiffies.
393 static int as_antic_expired(struct as_data *ad)
397 delta_jif = jiffies - ad->antic_start;
398 if (unlikely(delta_jif < 0))
399 delta_jif = -delta_jif;
400 if (delta_jif < ad->antic_expire)
407 * as_antic_waitnext starts anticipating that a nice request will soon be
408 * submitted. See also as_antic_waitreq
410 static void as_antic_waitnext(struct as_data *ad)
412 unsigned long timeout;
414 BUG_ON(ad->antic_status != ANTIC_OFF
415 && ad->antic_status != ANTIC_WAIT_REQ);
417 timeout = ad->antic_start + ad->antic_expire;
419 mod_timer(&ad->antic_timer, timeout);
421 ad->antic_status = ANTIC_WAIT_NEXT;
425 * as_antic_waitreq starts anticipating. We don't start timing the anticipation
426 * until the request that we're anticipating on has finished. This means we
427 * are timing from when the candidate process wakes up hopefully.
429 static void as_antic_waitreq(struct as_data *ad)
431 BUG_ON(ad->antic_status == ANTIC_FINISHED);
432 if (ad->antic_status == ANTIC_OFF) {
433 if (!ad->io_context || ad->ioc_finished)
434 as_antic_waitnext(ad);
436 ad->antic_status = ANTIC_WAIT_REQ;
441 * This is called directly by the functions in this file to stop anticipation.
442 * We kill the timer and schedule a call to the request_fn asap.
444 static void as_antic_stop(struct as_data *ad)
446 int status = ad->antic_status;
448 if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) {
449 if (status == ANTIC_WAIT_NEXT)
450 del_timer(&ad->antic_timer);
451 ad->antic_status = ANTIC_FINISHED;
452 /* see as_work_handler */
453 kblockd_schedule_work(&ad->antic_work);
458 * as_antic_timeout is the timer function set by as_antic_waitnext.
460 static void as_antic_timeout(unsigned long data)
462 struct request_queue *q = (struct request_queue *)data;
463 struct as_data *ad = q->elevator->elevator_data;
466 spin_lock_irqsave(q->queue_lock, flags);
467 if (ad->antic_status == ANTIC_WAIT_REQ
468 || ad->antic_status == ANTIC_WAIT_NEXT) {
469 struct as_io_context *aic;
470 spin_lock(&ad->io_context->lock);
471 aic = ad->io_context->aic;
473 ad->antic_status = ANTIC_FINISHED;
474 kblockd_schedule_work(&ad->antic_work);
476 if (aic->ttime_samples == 0) {
477 /* process anticipated on has exited or timed out*/
478 ad->exit_prob = (7*ad->exit_prob + 256)/8;
480 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
481 /* process not "saved" by a cooperating request */
482 ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8;
484 spin_unlock(&ad->io_context->lock);
486 spin_unlock_irqrestore(q->queue_lock, flags);
489 static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic,
492 /* fixed point: 1.0 == 1<<8 */
493 if (aic->ttime_samples == 0) {
494 ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8;
495 ad->new_ttime_mean = ad->new_ttime_total / 256;
497 ad->exit_prob = (7*ad->exit_prob)/8;
499 aic->ttime_samples = (7*aic->ttime_samples + 256) / 8;
500 aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8;
501 aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples;
504 static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic,
509 if (aic->seek_samples == 0) {
510 ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8;
511 ad->new_seek_mean = ad->new_seek_total / 256;
515 * Don't allow the seek distance to get too large from the
516 * odd fragment, pagein, etc
518 if (aic->seek_samples <= 60) /* second&third seek */
519 sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024);
521 sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64);
523 aic->seek_samples = (7*aic->seek_samples + 256) / 8;
524 aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8;
525 total = aic->seek_total + (aic->seek_samples/2);
526 do_div(total, aic->seek_samples);
527 aic->seek_mean = (sector_t)total;
531 * as_update_iohist keeps a decaying histogram of IO thinktimes, and
532 * updates @aic->ttime_mean based on that. It is called when a new
535 static void as_update_iohist(struct as_data *ad, struct as_io_context *aic,
538 int data_dir = rq_is_sync(rq);
539 unsigned long thinktime = 0;
545 if (data_dir == REQ_SYNC) {
546 unsigned long in_flight = atomic_read(&aic->nr_queued)
547 + atomic_read(&aic->nr_dispatched);
548 spin_lock(&aic->lock);
549 if (test_bit(AS_TASK_IORUNNING, &aic->state) ||
550 test_bit(AS_TASK_IOSTARTED, &aic->state)) {
551 /* Calculate read -> read thinktime */
552 if (test_bit(AS_TASK_IORUNNING, &aic->state)
554 thinktime = jiffies - aic->last_end_request;
555 thinktime = min(thinktime, MAX_THINKTIME-1);
557 as_update_thinktime(ad, aic, thinktime);
559 /* Calculate read -> read seek distance */
560 if (aic->last_request_pos < rq->sector)
561 seek_dist = rq->sector - aic->last_request_pos;
563 seek_dist = aic->last_request_pos - rq->sector;
564 as_update_seekdist(ad, aic, seek_dist);
566 aic->last_request_pos = rq->sector + rq->nr_sectors;
567 set_bit(AS_TASK_IOSTARTED, &aic->state);
568 spin_unlock(&aic->lock);
573 * as_close_req decides if one request is considered "close" to the
574 * previous one issued.
576 static int as_close_req(struct as_data *ad, struct as_io_context *aic,
579 unsigned long delay; /* jiffies */
580 sector_t last = ad->last_sector[ad->batch_data_dir];
581 sector_t next = rq->sector;
582 sector_t delta; /* acceptable close offset (in sectors) */
585 if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished)
588 delay = jiffies - ad->antic_start;
592 else if (delay <= (20 * HZ / 1000) && delay <= ad->antic_expire)
593 delta = 8192 << delay;
597 if ((last <= next + (delta>>1)) && (next <= last + delta))
605 if (aic->seek_samples == 0) {
607 * Process has just started IO. Use past statistics to
608 * gauge success possibility
610 if (ad->new_seek_mean > s) {
611 /* this request is better than what we're expecting */
616 if (aic->seek_mean > s) {
617 /* this request is better than what we're expecting */
626 * as_can_break_anticipation returns true if we have been anticipating this
629 * It also returns true if the process against which we are anticipating
630 * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to
631 * dispatch it ASAP, because we know that application will not be submitting
634 * If the task which has submitted the request has exited, break anticipation.
636 * If this task has queued some other IO, do not enter enticipation.
638 static int as_can_break_anticipation(struct as_data *ad, struct request *rq)
640 struct io_context *ioc;
641 struct as_io_context *aic;
643 ioc = ad->io_context;
645 spin_lock(&ioc->lock);
647 if (rq && ioc == RQ_IOC(rq)) {
648 /* request from same process */
649 spin_unlock(&ioc->lock);
653 if (ad->ioc_finished && as_antic_expired(ad)) {
655 * In this situation status should really be FINISHED,
656 * however the timer hasn't had the chance to run yet.
658 spin_unlock(&ioc->lock);
664 spin_unlock(&ioc->lock);
668 if (atomic_read(&aic->nr_queued) > 0) {
669 /* process has more requests queued */
670 spin_unlock(&ioc->lock);
674 if (atomic_read(&aic->nr_dispatched) > 0) {
675 /* process has more requests dispatched */
676 spin_unlock(&ioc->lock);
680 if (rq && rq_is_sync(rq) && as_close_req(ad, aic, rq)) {
682 * Found a close request that is not one of ours.
684 * This makes close requests from another process update
685 * our IO history. Is generally useful when there are
686 * two or more cooperating processes working in the same
689 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
690 if (aic->ttime_samples == 0)
691 ad->exit_prob = (7*ad->exit_prob + 256)/8;
693 ad->exit_no_coop = (7*ad->exit_no_coop)/8;
696 as_update_iohist(ad, aic, rq);
697 spin_unlock(&ioc->lock);
701 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
702 /* process anticipated on has exited */
703 if (aic->ttime_samples == 0)
704 ad->exit_prob = (7*ad->exit_prob + 256)/8;
706 if (ad->exit_no_coop > 128) {
707 spin_unlock(&ioc->lock);
712 if (aic->ttime_samples == 0) {
713 if (ad->new_ttime_mean > ad->antic_expire) {
714 spin_unlock(&ioc->lock);
717 if (ad->exit_prob * ad->exit_no_coop > 128*256) {
718 spin_unlock(&ioc->lock);
721 } else if (aic->ttime_mean > ad->antic_expire) {
722 /* the process thinks too much between requests */
723 spin_unlock(&ioc->lock);
726 spin_unlock(&ioc->lock);
731 * as_can_anticipate indicates whether we should either run rq
732 * or keep anticipating a better request.
734 static int as_can_anticipate(struct as_data *ad, struct request *rq)
738 * Last request submitted was a write
742 if (ad->antic_status == ANTIC_FINISHED)
744 * Don't restart if we have just finished. Run the next request
748 if (as_can_break_anticipation(ad, rq))
750 * This request is a good candidate. Don't keep anticipating,
756 * OK from here, we haven't finished, and don't have a decent request!
757 * Status is either ANTIC_OFF so start waiting,
758 * ANTIC_WAIT_REQ so continue waiting for request to finish
759 * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request.
766 * as_update_rq must be called whenever a request (rq) is added to
767 * the sort_list. This function keeps caches up to date, and checks if the
768 * request might be one we are "anticipating"
770 static void as_update_rq(struct as_data *ad, struct request *rq)
772 const int data_dir = rq_is_sync(rq);
774 /* keep the next_rq cache up to date */
775 ad->next_rq[data_dir] = as_choose_req(ad, rq, ad->next_rq[data_dir]);
778 * have we been anticipating this request?
779 * or does it come from the same process as the one we are anticipating
782 if (ad->antic_status == ANTIC_WAIT_REQ
783 || ad->antic_status == ANTIC_WAIT_NEXT) {
784 if (as_can_break_anticipation(ad, rq))
790 * Gathers timings and resizes the write batch automatically
792 static void update_write_batch(struct as_data *ad)
794 unsigned long batch = ad->batch_expire[REQ_ASYNC];
797 write_time = (jiffies - ad->current_batch_expires) + batch;
801 if (write_time > batch && !ad->write_batch_idled) {
802 if (write_time > batch * 3)
803 ad->write_batch_count /= 2;
805 ad->write_batch_count--;
806 } else if (write_time < batch && ad->current_write_count == 0) {
807 if (batch > write_time * 3)
808 ad->write_batch_count *= 2;
810 ad->write_batch_count++;
813 if (ad->write_batch_count < 1)
814 ad->write_batch_count = 1;
818 * as_completed_request is to be called when a request has completed and
819 * returned something to the requesting process, be it an error or data.
821 static void as_completed_request(struct request_queue *q, struct request *rq)
823 struct as_data *ad = q->elevator->elevator_data;
825 WARN_ON(!list_empty(&rq->queuelist));
827 if (RQ_STATE(rq) != AS_RQ_REMOVED) {
828 printk("rq->state %d\n", RQ_STATE(rq));
833 if (ad->changed_batch && ad->nr_dispatched == 1) {
834 kblockd_schedule_work(&ad->antic_work);
835 ad->changed_batch = 0;
837 if (ad->batch_data_dir == REQ_SYNC)
840 WARN_ON(ad->nr_dispatched == 0);
844 * Start counting the batch from when a request of that direction is
845 * actually serviced. This should help devices with big TCQ windows
846 * and writeback caches
848 if (ad->new_batch && ad->batch_data_dir == rq_is_sync(rq)) {
849 update_write_batch(ad);
850 ad->current_batch_expires = jiffies +
851 ad->batch_expire[REQ_SYNC];
855 if (ad->io_context == RQ_IOC(rq) && ad->io_context) {
856 ad->antic_start = jiffies;
857 ad->ioc_finished = 1;
858 if (ad->antic_status == ANTIC_WAIT_REQ) {
860 * We were waiting on this request, now anticipate
863 as_antic_waitnext(ad);
867 as_put_io_context(rq);
869 RQ_SET_STATE(rq, AS_RQ_POSTSCHED);
873 * as_remove_queued_request removes a request from the pre dispatch queue
874 * without updating refcounts. It is expected the caller will drop the
875 * reference unless it replaces the request at somepart of the elevator
876 * (ie. the dispatch queue)
878 static void as_remove_queued_request(struct request_queue *q,
881 const int data_dir = rq_is_sync(rq);
882 struct as_data *ad = q->elevator->elevator_data;
883 struct io_context *ioc;
885 WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED);
888 if (ioc && ioc->aic) {
889 BUG_ON(!atomic_read(&ioc->aic->nr_queued));
890 atomic_dec(&ioc->aic->nr_queued);
894 * Update the "next_rq" cache if we are about to remove its
897 if (ad->next_rq[data_dir] == rq)
898 ad->next_rq[data_dir] = as_find_next_rq(ad, rq);
901 as_del_rq_rb(ad, rq);
905 * as_fifo_expired returns 0 if there are no expired requests on the fifo,
906 * 1 otherwise. It is ratelimited so that we only perform the check once per
907 * `fifo_expire' interval. Otherwise a large number of expired requests
908 * would create a hopeless seekstorm.
910 * See as_antic_expired comment.
912 static int as_fifo_expired(struct as_data *ad, int adir)
917 delta_jif = jiffies - ad->last_check_fifo[adir];
918 if (unlikely(delta_jif < 0))
919 delta_jif = -delta_jif;
920 if (delta_jif < ad->fifo_expire[adir])
923 ad->last_check_fifo[adir] = jiffies;
925 if (list_empty(&ad->fifo_list[adir]))
928 rq = rq_entry_fifo(ad->fifo_list[adir].next);
930 return time_after(jiffies, rq_fifo_time(rq));
934 * as_batch_expired returns true if the current batch has expired. A batch
935 * is a set of reads or a set of writes.
937 static inline int as_batch_expired(struct as_data *ad)
939 if (ad->changed_batch || ad->new_batch)
942 if (ad->batch_data_dir == REQ_SYNC)
943 /* TODO! add a check so a complete fifo gets written? */
944 return time_after(jiffies, ad->current_batch_expires);
946 return time_after(jiffies, ad->current_batch_expires)
947 || ad->current_write_count == 0;
951 * move an entry to dispatch queue
953 static void as_move_to_dispatch(struct as_data *ad, struct request *rq)
955 const int data_dir = rq_is_sync(rq);
957 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
960 ad->antic_status = ANTIC_OFF;
963 * This has to be set in order to be correctly updated by
966 ad->last_sector[data_dir] = rq->sector + rq->nr_sectors;
968 if (data_dir == REQ_SYNC) {
969 struct io_context *ioc = RQ_IOC(rq);
970 /* In case we have to anticipate after this */
971 copy_io_context(&ad->io_context, &ioc);
973 if (ad->io_context) {
974 put_io_context(ad->io_context);
975 ad->io_context = NULL;
978 if (ad->current_write_count != 0)
979 ad->current_write_count--;
981 ad->ioc_finished = 0;
983 ad->next_rq[data_dir] = as_find_next_rq(ad, rq);
986 * take it off the sort and fifo list, add to dispatch queue
988 as_remove_queued_request(ad->q, rq);
989 WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED);
991 elv_dispatch_sort(ad->q, rq);
993 RQ_SET_STATE(rq, AS_RQ_DISPATCHED);
994 if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
995 atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched);
1000 * as_dispatch_request selects the best request according to
1001 * read/write expire, batch expire, etc, and moves it to the dispatch
1002 * queue. Returns 1 if a request was found, 0 otherwise.
1004 static int as_dispatch_request(struct request_queue *q, int force)
1006 struct as_data *ad = q->elevator->elevator_data;
1007 const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]);
1008 const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]);
1011 if (unlikely(force)) {
1013 * Forced dispatch, accounting is useless. Reset
1014 * accounting states and dump fifo_lists. Note that
1015 * batch_data_dir is reset to REQ_SYNC to avoid
1016 * screwing write batch accounting as write batch
1017 * accounting occurs on W->R transition.
1021 ad->batch_data_dir = REQ_SYNC;
1022 ad->changed_batch = 0;
1025 while (ad->next_rq[REQ_SYNC]) {
1026 as_move_to_dispatch(ad, ad->next_rq[REQ_SYNC]);
1029 ad->last_check_fifo[REQ_SYNC] = jiffies;
1031 while (ad->next_rq[REQ_ASYNC]) {
1032 as_move_to_dispatch(ad, ad->next_rq[REQ_ASYNC]);
1035 ad->last_check_fifo[REQ_ASYNC] = jiffies;
1040 /* Signal that the write batch was uncontended, so we can't time it */
1041 if (ad->batch_data_dir == REQ_ASYNC && !reads) {
1042 if (ad->current_write_count == 0 || !writes)
1043 ad->write_batch_idled = 1;
1046 if (!(reads || writes)
1047 || ad->antic_status == ANTIC_WAIT_REQ
1048 || ad->antic_status == ANTIC_WAIT_NEXT
1049 || ad->changed_batch)
1052 if (!(reads && writes && as_batch_expired(ad))) {
1054 * batch is still running or no reads or no writes
1056 rq = ad->next_rq[ad->batch_data_dir];
1058 if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) {
1059 if (as_fifo_expired(ad, REQ_SYNC))
1062 if (as_can_anticipate(ad, rq)) {
1063 as_antic_waitreq(ad);
1069 /* we have a "next request" */
1070 if (reads && !writes)
1071 ad->current_batch_expires =
1072 jiffies + ad->batch_expire[REQ_SYNC];
1073 goto dispatch_request;
1078 * at this point we are not running a batch. select the appropriate
1079 * data direction (read / write)
1083 BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_SYNC]));
1085 if (writes && ad->batch_data_dir == REQ_SYNC)
1087 * Last batch was a read, switch to writes
1089 goto dispatch_writes;
1091 if (ad->batch_data_dir == REQ_ASYNC) {
1092 WARN_ON(ad->new_batch);
1093 ad->changed_batch = 1;
1095 ad->batch_data_dir = REQ_SYNC;
1096 rq = rq_entry_fifo(ad->fifo_list[REQ_SYNC].next);
1097 ad->last_check_fifo[ad->batch_data_dir] = jiffies;
1098 goto dispatch_request;
1102 * the last batch was a read
1107 BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_ASYNC]));
1109 if (ad->batch_data_dir == REQ_SYNC) {
1110 ad->changed_batch = 1;
1113 * new_batch might be 1 when the queue runs out of
1114 * reads. A subsequent submission of a write might
1115 * cause a change of batch before the read is finished.
1119 ad->batch_data_dir = REQ_ASYNC;
1120 ad->current_write_count = ad->write_batch_count;
1121 ad->write_batch_idled = 0;
1122 rq = rq_entry_fifo(ad->fifo_list[REQ_ASYNC].next);
1123 ad->last_check_fifo[REQ_ASYNC] = jiffies;
1124 goto dispatch_request;
1132 * If a request has expired, service it.
1135 if (as_fifo_expired(ad, ad->batch_data_dir)) {
1137 rq = rq_entry_fifo(ad->fifo_list[ad->batch_data_dir].next);
1140 if (ad->changed_batch) {
1141 WARN_ON(ad->new_batch);
1143 if (ad->nr_dispatched)
1146 if (ad->batch_data_dir == REQ_ASYNC)
1147 ad->current_batch_expires = jiffies +
1148 ad->batch_expire[REQ_ASYNC];
1152 ad->changed_batch = 0;
1156 * rq is the selected appropriate request.
1158 as_move_to_dispatch(ad, rq);
1164 * add rq to rbtree and fifo
1166 static void as_add_request(struct request_queue *q, struct request *rq)
1168 struct as_data *ad = q->elevator->elevator_data;
1171 RQ_SET_STATE(rq, AS_RQ_NEW);
1173 data_dir = rq_is_sync(rq);
1175 rq->elevator_private = as_get_io_context(q->node);
1178 as_update_iohist(ad, RQ_IOC(rq)->aic, rq);
1179 atomic_inc(&RQ_IOC(rq)->aic->nr_queued);
1182 as_add_rq_rb(ad, rq);
1185 * set expire time and add to fifo list
1187 rq_set_fifo_time(rq, jiffies + ad->fifo_expire[data_dir]);
1188 list_add_tail(&rq->queuelist, &ad->fifo_list[data_dir]);
1190 as_update_rq(ad, rq); /* keep state machine up to date */
1191 RQ_SET_STATE(rq, AS_RQ_QUEUED);
1194 static void as_activate_request(struct request_queue *q, struct request *rq)
1196 WARN_ON(RQ_STATE(rq) != AS_RQ_DISPATCHED);
1197 RQ_SET_STATE(rq, AS_RQ_REMOVED);
1198 if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
1199 atomic_dec(&RQ_IOC(rq)->aic->nr_dispatched);
1202 static void as_deactivate_request(struct request_queue *q, struct request *rq)
1204 WARN_ON(RQ_STATE(rq) != AS_RQ_REMOVED);
1205 RQ_SET_STATE(rq, AS_RQ_DISPATCHED);
1206 if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
1207 atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched);
1211 * as_queue_empty tells us if there are requests left in the device. It may
1212 * not be the case that a driver can get the next request even if the queue
1213 * is not empty - it is used in the block layer to check for plugging and
1214 * merging opportunities
1216 static int as_queue_empty(struct request_queue *q)
1218 struct as_data *ad = q->elevator->elevator_data;
1220 return list_empty(&ad->fifo_list[REQ_ASYNC])
1221 && list_empty(&ad->fifo_list[REQ_SYNC]);
1225 as_merge(struct request_queue *q, struct request **req, struct bio *bio)
1227 struct as_data *ad = q->elevator->elevator_data;
1228 sector_t rb_key = bio->bi_sector + bio_sectors(bio);
1229 struct request *__rq;
1232 * check for front merge
1234 __rq = elv_rb_find(&ad->sort_list[bio_data_dir(bio)], rb_key);
1235 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1237 return ELEVATOR_FRONT_MERGE;
1240 return ELEVATOR_NO_MERGE;
1243 static void as_merged_request(struct request_queue *q, struct request *req,
1246 struct as_data *ad = q->elevator->elevator_data;
1249 * if the merge was a front merge, we need to reposition request
1251 if (type == ELEVATOR_FRONT_MERGE) {
1252 as_del_rq_rb(ad, req);
1253 as_add_rq_rb(ad, req);
1255 * Note! At this stage of this and the next function, our next
1256 * request may not be optimal - eg the request may have "grown"
1257 * behind the disk head. We currently don't bother adjusting.
1262 static void as_merged_requests(struct request_queue *q, struct request *req,
1263 struct request *next)
1266 * if next expires before rq, assign its expire time to arq
1267 * and move into next position (next will be deleted) in fifo
1269 if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
1270 if (time_before(rq_fifo_time(next), rq_fifo_time(req))) {
1271 list_move(&req->queuelist, &next->queuelist);
1272 rq_set_fifo_time(req, rq_fifo_time(next));
1277 * kill knowledge of next, this one is a goner
1279 as_remove_queued_request(q, next);
1280 as_put_io_context(next);
1282 RQ_SET_STATE(next, AS_RQ_MERGED);
1286 * This is executed in a "deferred" process context, by kblockd. It calls the
1287 * driver's request_fn so the driver can submit that request.
1289 * IMPORTANT! This guy will reenter the elevator, so set up all queue global
1290 * state before calling, and don't rely on any state over calls.
1292 * FIXME! dispatch queue is not a queue at all!
1294 static void as_work_handler(struct work_struct *work)
1296 struct as_data *ad = container_of(work, struct as_data, antic_work);
1297 struct request_queue *q = ad->q;
1298 unsigned long flags;
1300 spin_lock_irqsave(q->queue_lock, flags);
1301 blk_start_queueing(q);
1302 spin_unlock_irqrestore(q->queue_lock, flags);
1305 static int as_may_queue(struct request_queue *q, int rw)
1307 int ret = ELV_MQUEUE_MAY;
1308 struct as_data *ad = q->elevator->elevator_data;
1309 struct io_context *ioc;
1310 if (ad->antic_status == ANTIC_WAIT_REQ ||
1311 ad->antic_status == ANTIC_WAIT_NEXT) {
1312 ioc = as_get_io_context(q->node);
1313 if (ad->io_context == ioc)
1314 ret = ELV_MQUEUE_MUST;
1315 put_io_context(ioc);
1321 static void as_exit_queue(elevator_t *e)
1323 struct as_data *ad = e->elevator_data;
1325 del_timer_sync(&ad->antic_timer);
1326 kblockd_flush_work(&ad->antic_work);
1328 BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC]));
1329 BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC]));
1331 put_io_context(ad->io_context);
1336 * initialize elevator private data (as_data).
1338 static void *as_init_queue(struct request_queue *q)
1342 ad = kmalloc_node(sizeof(*ad), GFP_KERNEL | __GFP_ZERO, q->node);
1346 ad->q = q; /* Identify what queue the data belongs to */
1348 /* anticipatory scheduling helpers */
1349 ad->antic_timer.function = as_antic_timeout;
1350 ad->antic_timer.data = (unsigned long)q;
1351 init_timer(&ad->antic_timer);
1352 INIT_WORK(&ad->antic_work, as_work_handler);
1354 INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]);
1355 INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]);
1356 ad->sort_list[REQ_SYNC] = RB_ROOT;
1357 ad->sort_list[REQ_ASYNC] = RB_ROOT;
1358 ad->fifo_expire[REQ_SYNC] = default_read_expire;
1359 ad->fifo_expire[REQ_ASYNC] = default_write_expire;
1360 ad->antic_expire = default_antic_expire;
1361 ad->batch_expire[REQ_SYNC] = default_read_batch_expire;
1362 ad->batch_expire[REQ_ASYNC] = default_write_batch_expire;
1364 ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC];
1365 ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10;
1366 if (ad->write_batch_count < 2)
1367 ad->write_batch_count = 2;
1377 as_var_show(unsigned int var, char *page)
1379 return sprintf(page, "%d\n", var);
1383 as_var_store(unsigned long *var, const char *page, size_t count)
1385 char *p = (char *) page;
1387 *var = simple_strtoul(p, &p, 10);
1391 static ssize_t est_time_show(elevator_t *e, char *page)
1393 struct as_data *ad = e->elevator_data;
1396 pos += sprintf(page+pos, "%lu %% exit probability\n",
1397 100*ad->exit_prob/256);
1398 pos += sprintf(page+pos, "%lu %% probability of exiting without a "
1399 "cooperating process submitting IO\n",
1400 100*ad->exit_no_coop/256);
1401 pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean);
1402 pos += sprintf(page+pos, "%llu sectors new seek distance\n",
1403 (unsigned long long)ad->new_seek_mean);
1408 #define SHOW_FUNCTION(__FUNC, __VAR) \
1409 static ssize_t __FUNC(elevator_t *e, char *page) \
1411 struct as_data *ad = e->elevator_data; \
1412 return as_var_show(jiffies_to_msecs((__VAR)), (page)); \
1414 SHOW_FUNCTION(as_read_expire_show, ad->fifo_expire[REQ_SYNC]);
1415 SHOW_FUNCTION(as_write_expire_show, ad->fifo_expire[REQ_ASYNC]);
1416 SHOW_FUNCTION(as_antic_expire_show, ad->antic_expire);
1417 SHOW_FUNCTION(as_read_batch_expire_show, ad->batch_expire[REQ_SYNC]);
1418 SHOW_FUNCTION(as_write_batch_expire_show, ad->batch_expire[REQ_ASYNC]);
1419 #undef SHOW_FUNCTION
1421 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
1422 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
1424 struct as_data *ad = e->elevator_data; \
1425 int ret = as_var_store(__PTR, (page), count); \
1426 if (*(__PTR) < (MIN)) \
1428 else if (*(__PTR) > (MAX)) \
1430 *(__PTR) = msecs_to_jiffies(*(__PTR)); \
1433 STORE_FUNCTION(as_read_expire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX);
1434 STORE_FUNCTION(as_write_expire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX);
1435 STORE_FUNCTION(as_antic_expire_store, &ad->antic_expire, 0, INT_MAX);
1436 STORE_FUNCTION(as_read_batch_expire_store,
1437 &ad->batch_expire[REQ_SYNC], 0, INT_MAX);
1438 STORE_FUNCTION(as_write_batch_expire_store,
1439 &ad->batch_expire[REQ_ASYNC], 0, INT_MAX);
1440 #undef STORE_FUNCTION
1442 #define AS_ATTR(name) \
1443 __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store)
1445 static struct elv_fs_entry as_attrs[] = {
1446 __ATTR_RO(est_time),
1447 AS_ATTR(read_expire),
1448 AS_ATTR(write_expire),
1449 AS_ATTR(antic_expire),
1450 AS_ATTR(read_batch_expire),
1451 AS_ATTR(write_batch_expire),
1455 static struct elevator_type iosched_as = {
1457 .elevator_merge_fn = as_merge,
1458 .elevator_merged_fn = as_merged_request,
1459 .elevator_merge_req_fn = as_merged_requests,
1460 .elevator_dispatch_fn = as_dispatch_request,
1461 .elevator_add_req_fn = as_add_request,
1462 .elevator_activate_req_fn = as_activate_request,
1463 .elevator_deactivate_req_fn = as_deactivate_request,
1464 .elevator_queue_empty_fn = as_queue_empty,
1465 .elevator_completed_req_fn = as_completed_request,
1466 .elevator_former_req_fn = elv_rb_former_request,
1467 .elevator_latter_req_fn = elv_rb_latter_request,
1468 .elevator_may_queue_fn = as_may_queue,
1469 .elevator_init_fn = as_init_queue,
1470 .elevator_exit_fn = as_exit_queue,
1474 .elevator_attrs = as_attrs,
1475 .elevator_name = "anticipatory",
1476 .elevator_owner = THIS_MODULE,
1479 static int __init as_init(void)
1481 elv_register(&iosched_as);
1486 static void __exit as_exit(void)
1488 DECLARE_COMPLETION_ONSTACK(all_gone);
1489 elv_unregister(&iosched_as);
1490 ioc_gone = &all_gone;
1491 /* ioc_gone's update must be visible before reading ioc_count */
1493 if (elv_ioc_count_read(ioc_count))
1494 wait_for_completion(ioc_gone);
1498 module_init(as_init);
1499 module_exit(as_exit);
1501 MODULE_AUTHOR("Nick Piggin");
1502 MODULE_LICENSE("GPL");
1503 MODULE_DESCRIPTION("anticipatory IO scheduler");