Root/block/as-iosched.c

1/*
2 * Anticipatory & deadline i/o scheduler.
3 *
4 * Copyright (C) 2002 Jens Axboe <axboe@kernel.dk>
5 * Nick Piggin <nickpiggin@yahoo.com.au>
6 *
7 */
8#include <linux/kernel.h>
9#include <linux/fs.h>
10#include <linux/blkdev.h>
11#include <linux/elevator.h>
12#include <linux/bio.h>
13#include <linux/module.h>
14#include <linux/slab.h>
15#include <linux/init.h>
16#include <linux/compiler.h>
17#include <linux/rbtree.h>
18#include <linux/interrupt.h>
19
20/*
21 * See Documentation/block/as-iosched.txt
22 */
23
24/*
25 * max time before a read is submitted.
26 */
27#define default_read_expire (HZ / 8)
28
29/*
30 * ditto for writes, these limits are not hard, even
31 * if the disk is capable of satisfying them.
32 */
33#define default_write_expire (HZ / 4)
34
35/*
36 * read_batch_expire describes how long we will allow a stream of reads to
37 * persist before looking to see whether it is time to switch over to writes.
38 */
39#define default_read_batch_expire (HZ / 2)
40
41/*
42 * write_batch_expire describes how long we want a stream of writes to run for.
43 * This is not a hard limit, but a target we set for the auto-tuning thingy.
44 * See, the problem is: we can send a lot of writes to disk cache / TCQ in
45 * a short amount of time...
46 */
47#define default_write_batch_expire (HZ / 8)
48
49/*
50 * max time we may wait to anticipate a read (default around 6ms)
51 */
52#define default_antic_expire ((HZ / 150) ? HZ / 150 : 1)
53
54/*
55 * Keep track of up to 20ms thinktimes. We can go as big as we like here,
56 * however huge values tend to interfere and not decay fast enough. A program
57 * might be in a non-io phase of operation. Waiting on user input for example,
58 * or doing a lengthy computation. A small penalty can be justified there, and
59 * will still catch out those processes that constantly have large thinktimes.
60 */
61#define MAX_THINKTIME (HZ/50UL)
62
63/* Bits in as_io_context.state */
64enum as_io_states {
65    AS_TASK_RUNNING=0, /* Process has not exited */
66    AS_TASK_IOSTARTED, /* Process has started some IO */
67    AS_TASK_IORUNNING, /* Process has completed some IO */
68};
69
70enum anticipation_status {
71    ANTIC_OFF=0, /* Not anticipating (normal operation) */
72    ANTIC_WAIT_REQ, /* The last read has not yet completed */
73    ANTIC_WAIT_NEXT, /* Currently anticipating a request vs
74                   last read (which has completed) */
75    ANTIC_FINISHED, /* Anticipating but have found a candidate
76                 * or timed out */
77};
78
79struct as_data {
80    /*
81     * run time data
82     */
83
84    struct request_queue *q; /* the "owner" queue */
85
86    /*
87     * requests (as_rq s) are present on both sort_list and fifo_list
88     */
89    struct rb_root sort_list[2];
90    struct list_head fifo_list[2];
91
92    struct request *next_rq[2]; /* next in sort order */
93    sector_t last_sector[2]; /* last SYNC & ASYNC sectors */
94
95    unsigned long exit_prob; /* probability a task will exit while
96                       being waited on */
97    unsigned long exit_no_coop; /* probablility an exited task will
98                       not be part of a later cooperating
99                       request */
100    unsigned long new_ttime_total; /* mean thinktime on new proc */
101    unsigned long new_ttime_mean;
102    u64 new_seek_total; /* mean seek on new proc */
103    sector_t new_seek_mean;
104
105    unsigned long current_batch_expires;
106    unsigned long last_check_fifo[2];
107    int changed_batch; /* 1: waiting for old batch to end */
108    int new_batch; /* 1: waiting on first read complete */
109    int batch_data_dir; /* current batch SYNC / ASYNC */
110    int write_batch_count; /* max # of reqs in a write batch */
111    int current_write_count; /* how many requests left this batch */
112    int write_batch_idled; /* has the write batch gone idle? */
113
114    enum anticipation_status antic_status;
115    unsigned long antic_start; /* jiffies: when it started */
116    struct timer_list antic_timer; /* anticipatory scheduling timer */
117    struct work_struct antic_work; /* Deferred unplugging */
118    struct io_context *io_context; /* Identify the expected process */
119    int ioc_finished; /* IO associated with io_context is finished */
120    int nr_dispatched;
121
122    /*
123     * settings that change how the i/o scheduler behaves
124     */
125    unsigned long fifo_expire[2];
126    unsigned long batch_expire[2];
127    unsigned long antic_expire;
128};
129
130/*
131 * per-request data.
132 */
133enum arq_state {
134    AS_RQ_NEW=0, /* New - not referenced and not on any lists */
135    AS_RQ_QUEUED, /* In the request queue. It belongs to the
136                   scheduler */
137    AS_RQ_DISPATCHED, /* On the dispatch list. It belongs to the
138                   driver now */
139    AS_RQ_PRESCHED, /* Debug poisoning for requests being used */
140    AS_RQ_REMOVED,
141    AS_RQ_MERGED,
142    AS_RQ_POSTSCHED, /* when they shouldn't be */
143};
144
145#define RQ_IOC(rq) ((struct io_context *) (rq)->elevator_private)
146#define RQ_STATE(rq) ((enum arq_state)(rq)->elevator_private2)
147#define RQ_SET_STATE(rq, state) ((rq)->elevator_private2 = (void *) state)
148
149static DEFINE_PER_CPU(unsigned long, as_ioc_count);
150static struct completion *ioc_gone;
151static DEFINE_SPINLOCK(ioc_gone_lock);
152
153static void as_move_to_dispatch(struct as_data *ad, struct request *rq);
154static void as_antic_stop(struct as_data *ad);
155
156/*
157 * IO Context helper functions
158 */
159
160/* Called to deallocate the as_io_context */
161static void free_as_io_context(struct as_io_context *aic)
162{
163    kfree(aic);
164    elv_ioc_count_dec(as_ioc_count);
165    if (ioc_gone) {
166        /*
167         * AS scheduler is exiting, grab exit lock and check
168         * the pending io context count. If it hits zero,
169         * complete ioc_gone and set it back to NULL.
170         */
171        spin_lock(&ioc_gone_lock);
172        if (ioc_gone && !elv_ioc_count_read(as_ioc_count)) {
173            complete(ioc_gone);
174            ioc_gone = NULL;
175        }
176        spin_unlock(&ioc_gone_lock);
177    }
178}
179
180static void as_trim(struct io_context *ioc)
181{
182    spin_lock_irq(&ioc->lock);
183    if (ioc->aic)
184        free_as_io_context(ioc->aic);
185    ioc->aic = NULL;
186    spin_unlock_irq(&ioc->lock);
187}
188
189/* Called when the task exits */
190static void exit_as_io_context(struct as_io_context *aic)
191{
192    WARN_ON(!test_bit(AS_TASK_RUNNING, &aic->state));
193    clear_bit(AS_TASK_RUNNING, &aic->state);
194}
195
196static struct as_io_context *alloc_as_io_context(void)
197{
198    struct as_io_context *ret;
199
200    ret = kmalloc(sizeof(*ret), GFP_ATOMIC);
201    if (ret) {
202        ret->dtor = free_as_io_context;
203        ret->exit = exit_as_io_context;
204        ret->state = 1 << AS_TASK_RUNNING;
205        atomic_set(&ret->nr_queued, 0);
206        atomic_set(&ret->nr_dispatched, 0);
207        spin_lock_init(&ret->lock);
208        ret->ttime_total = 0;
209        ret->ttime_samples = 0;
210        ret->ttime_mean = 0;
211        ret->seek_total = 0;
212        ret->seek_samples = 0;
213        ret->seek_mean = 0;
214        elv_ioc_count_inc(as_ioc_count);
215    }
216
217    return ret;
218}
219
220/*
221 * If the current task has no AS IO context then create one and initialise it.
222 * Then take a ref on the task's io context and return it.
223 */
224static struct io_context *as_get_io_context(int node)
225{
226    struct io_context *ioc = get_io_context(GFP_ATOMIC, node);
227    if (ioc && !ioc->aic) {
228        ioc->aic = alloc_as_io_context();
229        if (!ioc->aic) {
230            put_io_context(ioc);
231            ioc = NULL;
232        }
233    }
234    return ioc;
235}
236
237static void as_put_io_context(struct request *rq)
238{
239    struct as_io_context *aic;
240
241    if (unlikely(!RQ_IOC(rq)))
242        return;
243
244    aic = RQ_IOC(rq)->aic;
245
246    if (rq_is_sync(rq) && aic) {
247        unsigned long flags;
248
249        spin_lock_irqsave(&aic->lock, flags);
250        set_bit(AS_TASK_IORUNNING, &aic->state);
251        aic->last_end_request = jiffies;
252        spin_unlock_irqrestore(&aic->lock, flags);
253    }
254
255    put_io_context(RQ_IOC(rq));
256}
257
258/*
259 * rb tree support functions
260 */
261#define RQ_RB_ROOT(ad, rq) (&(ad)->sort_list[rq_is_sync((rq))])
262
263static void as_add_rq_rb(struct as_data *ad, struct request *rq)
264{
265    struct request *alias;
266
267    while ((unlikely(alias = elv_rb_add(RQ_RB_ROOT(ad, rq), rq)))) {
268        as_move_to_dispatch(ad, alias);
269        as_antic_stop(ad);
270    }
271}
272
273static inline void as_del_rq_rb(struct as_data *ad, struct request *rq)
274{
275    elv_rb_del(RQ_RB_ROOT(ad, rq), rq);
276}
277
278/*
279 * IO Scheduler proper
280 */
281
282#define MAXBACK (1024 * 1024) /*
283                 * Maximum distance the disk will go backward
284                 * for a request.
285                 */
286
287#define BACK_PENALTY 2
288
289/*
290 * as_choose_req selects the preferred one of two requests of the same data_dir
291 * ignoring time - eg. timeouts, which is the job of as_dispatch_request
292 */
293static struct request *
294as_choose_req(struct as_data *ad, struct request *rq1, struct request *rq2)
295{
296    int data_dir;
297    sector_t last, s1, s2, d1, d2;
298    int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */
299    const sector_t maxback = MAXBACK;
300
301    if (rq1 == NULL || rq1 == rq2)
302        return rq2;
303    if (rq2 == NULL)
304        return rq1;
305
306    data_dir = rq_is_sync(rq1);
307
308    last = ad->last_sector[data_dir];
309    s1 = blk_rq_pos(rq1);
310    s2 = blk_rq_pos(rq2);
311
312    BUG_ON(data_dir != rq_is_sync(rq2));
313
314    /*
315     * Strict one way elevator _except_ in the case where we allow
316     * short backward seeks which are biased as twice the cost of a
317     * similar forward seek.
318     */
319    if (s1 >= last)
320        d1 = s1 - last;
321    else if (s1+maxback >= last)
322        d1 = (last - s1)*BACK_PENALTY;
323    else {
324        r1_wrap = 1;
325        d1 = 0; /* shut up, gcc */
326    }
327
328    if (s2 >= last)
329        d2 = s2 - last;
330    else if (s2+maxback >= last)
331        d2 = (last - s2)*BACK_PENALTY;
332    else {
333        r2_wrap = 1;
334        d2 = 0;
335    }
336
337    /* Found required data */
338    if (!r1_wrap && r2_wrap)
339        return rq1;
340    else if (!r2_wrap && r1_wrap)
341        return rq2;
342    else if (r1_wrap && r2_wrap) {
343        /* both behind the head */
344        if (s1 <= s2)
345            return rq1;
346        else
347            return rq2;
348    }
349
350    /* Both requests in front of the head */
351    if (d1 < d2)
352        return rq1;
353    else if (d2 < d1)
354        return rq2;
355    else {
356        if (s1 >= s2)
357            return rq1;
358        else
359            return rq2;
360    }
361}
362
363/*
364 * as_find_next_rq finds the next request after @prev in elevator order.
365 * this with as_choose_req form the basis for how the scheduler chooses
366 * what request to process next. Anticipation works on top of this.
367 */
368static struct request *
369as_find_next_rq(struct as_data *ad, struct request *last)
370{
371    struct rb_node *rbnext = rb_next(&last->rb_node);
372    struct rb_node *rbprev = rb_prev(&last->rb_node);
373    struct request *next = NULL, *prev = NULL;
374
375    BUG_ON(RB_EMPTY_NODE(&last->rb_node));
376
377    if (rbprev)
378        prev = rb_entry_rq(rbprev);
379
380    if (rbnext)
381        next = rb_entry_rq(rbnext);
382    else {
383        const int data_dir = rq_is_sync(last);
384
385        rbnext = rb_first(&ad->sort_list[data_dir]);
386        if (rbnext && rbnext != &last->rb_node)
387            next = rb_entry_rq(rbnext);
388    }
389
390    return as_choose_req(ad, next, prev);
391}
392
393/*
394 * anticipatory scheduling functions follow
395 */
396
397/*
398 * as_antic_expired tells us when we have anticipated too long.
399 * The funny "absolute difference" math on the elapsed time is to handle
400 * jiffy wraps, and disks which have been idle for 0x80000000 jiffies.
401 */
402static int as_antic_expired(struct as_data *ad)
403{
404    long delta_jif;
405
406    delta_jif = jiffies - ad->antic_start;
407    if (unlikely(delta_jif < 0))
408        delta_jif = -delta_jif;
409    if (delta_jif < ad->antic_expire)
410        return 0;
411
412    return 1;
413}
414
415/*
416 * as_antic_waitnext starts anticipating that a nice request will soon be
417 * submitted. See also as_antic_waitreq
418 */
419static void as_antic_waitnext(struct as_data *ad)
420{
421    unsigned long timeout;
422
423    BUG_ON(ad->antic_status != ANTIC_OFF
424            && ad->antic_status != ANTIC_WAIT_REQ);
425
426    timeout = ad->antic_start + ad->antic_expire;
427
428    mod_timer(&ad->antic_timer, timeout);
429
430    ad->antic_status = ANTIC_WAIT_NEXT;
431}
432
433/*
434 * as_antic_waitreq starts anticipating. We don't start timing the anticipation
435 * until the request that we're anticipating on has finished. This means we
436 * are timing from when the candidate process wakes up hopefully.
437 */
438static void as_antic_waitreq(struct as_data *ad)
439{
440    BUG_ON(ad->antic_status == ANTIC_FINISHED);
441    if (ad->antic_status == ANTIC_OFF) {
442        if (!ad->io_context || ad->ioc_finished)
443            as_antic_waitnext(ad);
444        else
445            ad->antic_status = ANTIC_WAIT_REQ;
446    }
447}
448
449/*
450 * This is called directly by the functions in this file to stop anticipation.
451 * We kill the timer and schedule a call to the request_fn asap.
452 */
453static void as_antic_stop(struct as_data *ad)
454{
455    int status = ad->antic_status;
456
457    if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) {
458        if (status == ANTIC_WAIT_NEXT)
459            del_timer(&ad->antic_timer);
460        ad->antic_status = ANTIC_FINISHED;
461        /* see as_work_handler */
462        kblockd_schedule_work(ad->q, &ad->antic_work);
463    }
464}
465
466/*
467 * as_antic_timeout is the timer function set by as_antic_waitnext.
468 */
469static void as_antic_timeout(unsigned long data)
470{
471    struct request_queue *q = (struct request_queue *)data;
472    struct as_data *ad = q->elevator->elevator_data;
473    unsigned long flags;
474
475    spin_lock_irqsave(q->queue_lock, flags);
476    if (ad->antic_status == ANTIC_WAIT_REQ
477            || ad->antic_status == ANTIC_WAIT_NEXT) {
478        struct as_io_context *aic;
479        spin_lock(&ad->io_context->lock);
480        aic = ad->io_context->aic;
481
482        ad->antic_status = ANTIC_FINISHED;
483        kblockd_schedule_work(q, &ad->antic_work);
484
485        if (aic->ttime_samples == 0) {
486            /* process anticipated on has exited or timed out*/
487            ad->exit_prob = (7*ad->exit_prob + 256)/8;
488        }
489        if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
490            /* process not "saved" by a cooperating request */
491            ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8;
492        }
493        spin_unlock(&ad->io_context->lock);
494    }
495    spin_unlock_irqrestore(q->queue_lock, flags);
496}
497
498static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic,
499                unsigned long ttime)
500{
501    /* fixed point: 1.0 == 1<<8 */
502    if (aic->ttime_samples == 0) {
503        ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8;
504        ad->new_ttime_mean = ad->new_ttime_total / 256;
505
506        ad->exit_prob = (7*ad->exit_prob)/8;
507    }
508    aic->ttime_samples = (7*aic->ttime_samples + 256) / 8;
509    aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8;
510    aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples;
511}
512
513static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic,
514                sector_t sdist)
515{
516    u64 total;
517
518    if (aic->seek_samples == 0) {
519        ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8;
520        ad->new_seek_mean = ad->new_seek_total / 256;
521    }
522
523    /*
524     * Don't allow the seek distance to get too large from the
525     * odd fragment, pagein, etc
526     */
527    if (aic->seek_samples <= 60) /* second&third seek */
528        sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024);
529    else
530        sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64);
531
532    aic->seek_samples = (7*aic->seek_samples + 256) / 8;
533    aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8;
534    total = aic->seek_total + (aic->seek_samples/2);
535    do_div(total, aic->seek_samples);
536    aic->seek_mean = (sector_t)total;
537}
538
539/*
540 * as_update_iohist keeps a decaying histogram of IO thinktimes, and
541 * updates @aic->ttime_mean based on that. It is called when a new
542 * request is queued.
543 */
544static void as_update_iohist(struct as_data *ad, struct as_io_context *aic,
545                struct request *rq)
546{
547    int data_dir = rq_is_sync(rq);
548    unsigned long thinktime = 0;
549    sector_t seek_dist;
550
551    if (aic == NULL)
552        return;
553
554    if (data_dir == BLK_RW_SYNC) {
555        unsigned long in_flight = atomic_read(&aic->nr_queued)
556                    + atomic_read(&aic->nr_dispatched);
557        spin_lock(&aic->lock);
558        if (test_bit(AS_TASK_IORUNNING, &aic->state) ||
559            test_bit(AS_TASK_IOSTARTED, &aic->state)) {
560            /* Calculate read -> read thinktime */
561            if (test_bit(AS_TASK_IORUNNING, &aic->state)
562                            && in_flight == 0) {
563                thinktime = jiffies - aic->last_end_request;
564                thinktime = min(thinktime, MAX_THINKTIME-1);
565            }
566            as_update_thinktime(ad, aic, thinktime);
567
568            /* Calculate read -> read seek distance */
569            if (aic->last_request_pos < blk_rq_pos(rq))
570                seek_dist = blk_rq_pos(rq) -
571                        aic->last_request_pos;
572            else
573                seek_dist = aic->last_request_pos -
574                        blk_rq_pos(rq);
575            as_update_seekdist(ad, aic, seek_dist);
576        }
577        aic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
578        set_bit(AS_TASK_IOSTARTED, &aic->state);
579        spin_unlock(&aic->lock);
580    }
581}
582
583/*
584 * as_close_req decides if one request is considered "close" to the
585 * previous one issued.
586 */
587static int as_close_req(struct as_data *ad, struct as_io_context *aic,
588            struct request *rq)
589{
590    unsigned long delay; /* jiffies */
591    sector_t last = ad->last_sector[ad->batch_data_dir];
592    sector_t next = blk_rq_pos(rq);
593    sector_t delta; /* acceptable close offset (in sectors) */
594    sector_t s;
595
596    if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished)
597        delay = 0;
598    else
599        delay = jiffies - ad->antic_start;
600
601    if (delay == 0)
602        delta = 8192;
603    else if (delay <= (20 * HZ / 1000) && delay <= ad->antic_expire)
604        delta = 8192 << delay;
605    else
606        return 1;
607
608    if ((last <= next + (delta>>1)) && (next <= last + delta))
609        return 1;
610
611    if (last < next)
612        s = next - last;
613    else
614        s = last - next;
615
616    if (aic->seek_samples == 0) {
617        /*
618         * Process has just started IO. Use past statistics to
619         * gauge success possibility
620         */
621        if (ad->new_seek_mean > s) {
622            /* this request is better than what we're expecting */
623            return 1;
624        }
625
626    } else {
627        if (aic->seek_mean > s) {
628            /* this request is better than what we're expecting */
629            return 1;
630        }
631    }
632
633    return 0;
634}
635
636/*
637 * as_can_break_anticipation returns true if we have been anticipating this
638 * request.
639 *
640 * It also returns true if the process against which we are anticipating
641 * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to
642 * dispatch it ASAP, because we know that application will not be submitting
643 * any new reads.
644 *
645 * If the task which has submitted the request has exited, break anticipation.
646 *
647 * If this task has queued some other IO, do not enter enticipation.
648 */
649static int as_can_break_anticipation(struct as_data *ad, struct request *rq)
650{
651    struct io_context *ioc;
652    struct as_io_context *aic;
653
654    ioc = ad->io_context;
655    BUG_ON(!ioc);
656    spin_lock(&ioc->lock);
657
658    if (rq && ioc == RQ_IOC(rq)) {
659        /* request from same process */
660        spin_unlock(&ioc->lock);
661        return 1;
662    }
663
664    if (ad->ioc_finished && as_antic_expired(ad)) {
665        /*
666         * In this situation status should really be FINISHED,
667         * however the timer hasn't had the chance to run yet.
668         */
669        spin_unlock(&ioc->lock);
670        return 1;
671    }
672
673    aic = ioc->aic;
674    if (!aic) {
675        spin_unlock(&ioc->lock);
676        return 0;
677    }
678
679    if (atomic_read(&aic->nr_queued) > 0) {
680        /* process has more requests queued */
681        spin_unlock(&ioc->lock);
682        return 1;
683    }
684
685    if (atomic_read(&aic->nr_dispatched) > 0) {
686        /* process has more requests dispatched */
687        spin_unlock(&ioc->lock);
688        return 1;
689    }
690
691    if (rq && rq_is_sync(rq) && as_close_req(ad, aic, rq)) {
692        /*
693         * Found a close request that is not one of ours.
694         *
695         * This makes close requests from another process update
696         * our IO history. Is generally useful when there are
697         * two or more cooperating processes working in the same
698         * area.
699         */
700        if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
701            if (aic->ttime_samples == 0)
702                ad->exit_prob = (7*ad->exit_prob + 256)/8;
703
704            ad->exit_no_coop = (7*ad->exit_no_coop)/8;
705        }
706
707        as_update_iohist(ad, aic, rq);
708        spin_unlock(&ioc->lock);
709        return 1;
710    }
711
712    if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
713        /* process anticipated on has exited */
714        if (aic->ttime_samples == 0)
715            ad->exit_prob = (7*ad->exit_prob + 256)/8;
716
717        if (ad->exit_no_coop > 128) {
718            spin_unlock(&ioc->lock);
719            return 1;
720        }
721    }
722
723    if (aic->ttime_samples == 0) {
724        if (ad->new_ttime_mean > ad->antic_expire) {
725            spin_unlock(&ioc->lock);
726            return 1;
727        }
728        if (ad->exit_prob * ad->exit_no_coop > 128*256) {
729            spin_unlock(&ioc->lock);
730            return 1;
731        }
732    } else if (aic->ttime_mean > ad->antic_expire) {
733        /* the process thinks too much between requests */
734        spin_unlock(&ioc->lock);
735        return 1;
736    }
737    spin_unlock(&ioc->lock);
738    return 0;
739}
740
741/*
742 * as_can_anticipate indicates whether we should either run rq
743 * or keep anticipating a better request.
744 */
745static int as_can_anticipate(struct as_data *ad, struct request *rq)
746{
747#if 0 /* disable for now, we need to check tag level as well */
748    /*
749     * SSD device without seek penalty, disable idling
750     */
751    if (blk_queue_nonrot(ad->q)) axman
752        return 0;
753#endif
754
755    if (!ad->io_context)
756        /*
757         * Last request submitted was a write
758         */
759        return 0;
760
761    if (ad->antic_status == ANTIC_FINISHED)
762        /*
763         * Don't restart if we have just finished. Run the next request
764         */
765        return 0;
766
767    if (as_can_break_anticipation(ad, rq))
768        /*
769         * This request is a good candidate. Don't keep anticipating,
770         * run it.
771         */
772        return 0;
773
774    /*
775     * OK from here, we haven't finished, and don't have a decent request!
776     * Status is either ANTIC_OFF so start waiting,
777     * ANTIC_WAIT_REQ so continue waiting for request to finish
778     * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request.
779     */
780
781    return 1;
782}
783
784/*
785 * as_update_rq must be called whenever a request (rq) is added to
786 * the sort_list. This function keeps caches up to date, and checks if the
787 * request might be one we are "anticipating"
788 */
789static void as_update_rq(struct as_data *ad, struct request *rq)
790{
791    const int data_dir = rq_is_sync(rq);
792
793    /* keep the next_rq cache up to date */
794    ad->next_rq[data_dir] = as_choose_req(ad, rq, ad->next_rq[data_dir]);
795
796    /*
797     * have we been anticipating this request?
798     * or does it come from the same process as the one we are anticipating
799     * for?
800     */
801    if (ad->antic_status == ANTIC_WAIT_REQ
802            || ad->antic_status == ANTIC_WAIT_NEXT) {
803        if (as_can_break_anticipation(ad, rq))
804            as_antic_stop(ad);
805    }
806}
807
808/*
809 * Gathers timings and resizes the write batch automatically
810 */
811static void update_write_batch(struct as_data *ad)
812{
813    unsigned long batch = ad->batch_expire[BLK_RW_ASYNC];
814    long write_time;
815
816    write_time = (jiffies - ad->current_batch_expires) + batch;
817    if (write_time < 0)
818        write_time = 0;
819
820    if (write_time > batch && !ad->write_batch_idled) {
821        if (write_time > batch * 3)
822            ad->write_batch_count /= 2;
823        else
824            ad->write_batch_count--;
825    } else if (write_time < batch && ad->current_write_count == 0) {
826        if (batch > write_time * 3)
827            ad->write_batch_count *= 2;
828        else
829            ad->write_batch_count++;
830    }
831
832    if (ad->write_batch_count < 1)
833        ad->write_batch_count = 1;
834}
835
836/*
837 * as_completed_request is to be called when a request has completed and
838 * returned something to the requesting process, be it an error or data.
839 */
840static void as_completed_request(struct request_queue *q, struct request *rq)
841{
842    struct as_data *ad = q->elevator->elevator_data;
843
844    WARN_ON(!list_empty(&rq->queuelist));
845
846    if (RQ_STATE(rq) != AS_RQ_REMOVED) {
847        WARN(1, "rq->state %d\n", RQ_STATE(rq));
848        goto out;
849    }
850
851    if (ad->changed_batch && ad->nr_dispatched == 1) {
852        ad->current_batch_expires = jiffies +
853                    ad->batch_expire[ad->batch_data_dir];
854        kblockd_schedule_work(q, &ad->antic_work);
855        ad->changed_batch = 0;
856
857        if (ad->batch_data_dir == BLK_RW_SYNC)
858            ad->new_batch = 1;
859    }
860    WARN_ON(ad->nr_dispatched == 0);
861    ad->nr_dispatched--;
862
863    /*
864     * Start counting the batch from when a request of that direction is
865     * actually serviced. This should help devices with big TCQ windows
866     * and writeback caches
867     */
868    if (ad->new_batch && ad->batch_data_dir == rq_is_sync(rq)) {
869        update_write_batch(ad);
870        ad->current_batch_expires = jiffies +
871                ad->batch_expire[BLK_RW_SYNC];
872        ad->new_batch = 0;
873    }
874
875    if (ad->io_context == RQ_IOC(rq) && ad->io_context) {
876        ad->antic_start = jiffies;
877        ad->ioc_finished = 1;
878        if (ad->antic_status == ANTIC_WAIT_REQ) {
879            /*
880             * We were waiting on this request, now anticipate
881             * the next one
882             */
883            as_antic_waitnext(ad);
884        }
885    }
886
887    as_put_io_context(rq);
888out:
889    RQ_SET_STATE(rq, AS_RQ_POSTSCHED);
890}
891
892/*
893 * as_remove_queued_request removes a request from the pre dispatch queue
894 * without updating refcounts. It is expected the caller will drop the
895 * reference unless it replaces the request at somepart of the elevator
896 * (ie. the dispatch queue)
897 */
898static void as_remove_queued_request(struct request_queue *q,
899                     struct request *rq)
900{
901    const int data_dir = rq_is_sync(rq);
902    struct as_data *ad = q->elevator->elevator_data;
903    struct io_context *ioc;
904
905    WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED);
906
907    ioc = RQ_IOC(rq);
908    if (ioc && ioc->aic) {
909        BUG_ON(!atomic_read(&ioc->aic->nr_queued));
910        atomic_dec(&ioc->aic->nr_queued);
911    }
912
913    /*
914     * Update the "next_rq" cache if we are about to remove its
915     * entry
916     */
917    if (ad->next_rq[data_dir] == rq)
918        ad->next_rq[data_dir] = as_find_next_rq(ad, rq);
919
920    rq_fifo_clear(rq);
921    as_del_rq_rb(ad, rq);
922}
923
924/*
925 * as_fifo_expired returns 0 if there are no expired requests on the fifo,
926 * 1 otherwise. It is ratelimited so that we only perform the check once per
927 * `fifo_expire' interval. Otherwise a large number of expired requests
928 * would create a hopeless seekstorm.
929 *
930 * See as_antic_expired comment.
931 */
932static int as_fifo_expired(struct as_data *ad, int adir)
933{
934    struct request *rq;
935    long delta_jif;
936
937    delta_jif = jiffies - ad->last_check_fifo[adir];
938    if (unlikely(delta_jif < 0))
939        delta_jif = -delta_jif;
940    if (delta_jif < ad->fifo_expire[adir])
941        return 0;
942
943    ad->last_check_fifo[adir] = jiffies;
944
945    if (list_empty(&ad->fifo_list[adir]))
946        return 0;
947
948    rq = rq_entry_fifo(ad->fifo_list[adir].next);
949
950    return time_after(jiffies, rq_fifo_time(rq));
951}
952
953/*
954 * as_batch_expired returns true if the current batch has expired. A batch
955 * is a set of reads or a set of writes.
956 */
957static inline int as_batch_expired(struct as_data *ad)
958{
959    if (ad->changed_batch || ad->new_batch)
960        return 0;
961
962    if (ad->batch_data_dir == BLK_RW_SYNC)
963        /* TODO! add a check so a complete fifo gets written? */
964        return time_after(jiffies, ad->current_batch_expires);
965
966    return time_after(jiffies, ad->current_batch_expires)
967        || ad->current_write_count == 0;
968}
969
970/*
971 * move an entry to dispatch queue
972 */
973static void as_move_to_dispatch(struct as_data *ad, struct request *rq)
974{
975    const int data_dir = rq_is_sync(rq);
976
977    BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
978
979    as_antic_stop(ad);
980    ad->antic_status = ANTIC_OFF;
981
982    /*
983     * This has to be set in order to be correctly updated by
984     * as_find_next_rq
985     */
986    ad->last_sector[data_dir] = blk_rq_pos(rq) + blk_rq_sectors(rq);
987
988    if (data_dir == BLK_RW_SYNC) {
989        struct io_context *ioc = RQ_IOC(rq);
990        /* In case we have to anticipate after this */
991        copy_io_context(&ad->io_context, &ioc);
992    } else {
993        if (ad->io_context) {
994            put_io_context(ad->io_context);
995            ad->io_context = NULL;
996        }
997
998        if (ad->current_write_count != 0)
999            ad->current_write_count--;
1000    }
1001    ad->ioc_finished = 0;
1002
1003    ad->next_rq[data_dir] = as_find_next_rq(ad, rq);
1004
1005    /*
1006     * take it off the sort and fifo list, add to dispatch queue
1007     */
1008    as_remove_queued_request(ad->q, rq);
1009    WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED);
1010
1011    elv_dispatch_sort(ad->q, rq);
1012
1013    RQ_SET_STATE(rq, AS_RQ_DISPATCHED);
1014    if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
1015        atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched);
1016    ad->nr_dispatched++;
1017}
1018
1019/*
1020 * as_dispatch_request selects the best request according to
1021 * read/write expire, batch expire, etc, and moves it to the dispatch
1022 * queue. Returns 1 if a request was found, 0 otherwise.
1023 */
1024static int as_dispatch_request(struct request_queue *q, int force)
1025{
1026    struct as_data *ad = q->elevator->elevator_data;
1027    const int reads = !list_empty(&ad->fifo_list[BLK_RW_SYNC]);
1028    const int writes = !list_empty(&ad->fifo_list[BLK_RW_ASYNC]);
1029    struct request *rq;
1030
1031    if (unlikely(force)) {
1032        /*
1033         * Forced dispatch, accounting is useless. Reset
1034         * accounting states and dump fifo_lists. Note that
1035         * batch_data_dir is reset to BLK_RW_SYNC to avoid
1036         * screwing write batch accounting as write batch
1037         * accounting occurs on W->R transition.
1038         */
1039        int dispatched = 0;
1040
1041        ad->batch_data_dir = BLK_RW_SYNC;
1042        ad->changed_batch = 0;
1043        ad->new_batch = 0;
1044
1045        while (ad->next_rq[BLK_RW_SYNC]) {
1046            as_move_to_dispatch(ad, ad->next_rq[BLK_RW_SYNC]);
1047            dispatched++;
1048        }
1049        ad->last_check_fifo[BLK_RW_SYNC] = jiffies;
1050
1051        while (ad->next_rq[BLK_RW_ASYNC]) {
1052            as_move_to_dispatch(ad, ad->next_rq[BLK_RW_ASYNC]);
1053            dispatched++;
1054        }
1055        ad->last_check_fifo[BLK_RW_ASYNC] = jiffies;
1056
1057        return dispatched;
1058    }
1059
1060    /* Signal that the write batch was uncontended, so we can't time it */
1061    if (ad->batch_data_dir == BLK_RW_ASYNC && !reads) {
1062        if (ad->current_write_count == 0 || !writes)
1063            ad->write_batch_idled = 1;
1064    }
1065
1066    if (!(reads || writes)
1067        || ad->antic_status == ANTIC_WAIT_REQ
1068        || ad->antic_status == ANTIC_WAIT_NEXT
1069        || ad->changed_batch)
1070        return 0;
1071
1072    if (!(reads && writes && as_batch_expired(ad))) {
1073        /*
1074         * batch is still running or no reads or no writes
1075         */
1076        rq = ad->next_rq[ad->batch_data_dir];
1077
1078        if (ad->batch_data_dir == BLK_RW_SYNC && ad->antic_expire) {
1079            if (as_fifo_expired(ad, BLK_RW_SYNC))
1080                goto fifo_expired;
1081
1082            if (as_can_anticipate(ad, rq)) {
1083                as_antic_waitreq(ad);
1084                return 0;
1085            }
1086        }
1087
1088        if (rq) {
1089            /* we have a "next request" */
1090            if (reads && !writes)
1091                ad->current_batch_expires =
1092                    jiffies + ad->batch_expire[BLK_RW_SYNC];
1093            goto dispatch_request;
1094        }
1095    }
1096
1097    /*
1098     * at this point we are not running a batch. select the appropriate
1099     * data direction (read / write)
1100     */
1101
1102    if (reads) {
1103        BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[BLK_RW_SYNC]));
1104
1105        if (writes && ad->batch_data_dir == BLK_RW_SYNC)
1106            /*
1107             * Last batch was a read, switch to writes
1108             */
1109            goto dispatch_writes;
1110
1111        if (ad->batch_data_dir == BLK_RW_ASYNC) {
1112            WARN_ON(ad->new_batch);
1113            ad->changed_batch = 1;
1114        }
1115        ad->batch_data_dir = BLK_RW_SYNC;
1116        rq = rq_entry_fifo(ad->fifo_list[BLK_RW_SYNC].next);
1117        ad->last_check_fifo[ad->batch_data_dir] = jiffies;
1118        goto dispatch_request;
1119    }
1120
1121    /*
1122     * the last batch was a read
1123     */
1124
1125    if (writes) {
1126dispatch_writes:
1127        BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[BLK_RW_ASYNC]));
1128
1129        if (ad->batch_data_dir == BLK_RW_SYNC) {
1130            ad->changed_batch = 1;
1131
1132            /*
1133             * new_batch might be 1 when the queue runs out of
1134             * reads. A subsequent submission of a write might
1135             * cause a change of batch before the read is finished.
1136             */
1137            ad->new_batch = 0;
1138        }
1139        ad->batch_data_dir = BLK_RW_ASYNC;
1140        ad->current_write_count = ad->write_batch_count;
1141        ad->write_batch_idled = 0;
1142        rq = rq_entry_fifo(ad->fifo_list[BLK_RW_ASYNC].next);
1143        ad->last_check_fifo[BLK_RW_ASYNC] = jiffies;
1144        goto dispatch_request;
1145    }
1146
1147    BUG();
1148    return 0;
1149
1150dispatch_request:
1151    /*
1152     * If a request has expired, service it.
1153     */
1154
1155    if (as_fifo_expired(ad, ad->batch_data_dir)) {
1156fifo_expired:
1157        rq = rq_entry_fifo(ad->fifo_list[ad->batch_data_dir].next);
1158    }
1159
1160    if (ad->changed_batch) {
1161        WARN_ON(ad->new_batch);
1162
1163        if (ad->nr_dispatched)
1164            return 0;
1165
1166        if (ad->batch_data_dir == BLK_RW_ASYNC)
1167            ad->current_batch_expires = jiffies +
1168                    ad->batch_expire[BLK_RW_ASYNC];
1169        else
1170            ad->new_batch = 1;
1171
1172        ad->changed_batch = 0;
1173    }
1174
1175    /*
1176     * rq is the selected appropriate request.
1177     */
1178    as_move_to_dispatch(ad, rq);
1179
1180    return 1;
1181}
1182
1183/*
1184 * add rq to rbtree and fifo
1185 */
1186static void as_add_request(struct request_queue *q, struct request *rq)
1187{
1188    struct as_data *ad = q->elevator->elevator_data;
1189    int data_dir;
1190
1191    RQ_SET_STATE(rq, AS_RQ_NEW);
1192
1193    data_dir = rq_is_sync(rq);
1194
1195    rq->elevator_private = as_get_io_context(q->node);
1196
1197    if (RQ_IOC(rq)) {
1198        as_update_iohist(ad, RQ_IOC(rq)->aic, rq);
1199        atomic_inc(&RQ_IOC(rq)->aic->nr_queued);
1200    }
1201
1202    as_add_rq_rb(ad, rq);
1203
1204    /*
1205     * set expire time and add to fifo list
1206     */
1207    rq_set_fifo_time(rq, jiffies + ad->fifo_expire[data_dir]);
1208    list_add_tail(&rq->queuelist, &ad->fifo_list[data_dir]);
1209
1210    as_update_rq(ad, rq); /* keep state machine up to date */
1211    RQ_SET_STATE(rq, AS_RQ_QUEUED);
1212}
1213
1214static void as_activate_request(struct request_queue *q, struct request *rq)
1215{
1216    WARN_ON(RQ_STATE(rq) != AS_RQ_DISPATCHED);
1217    RQ_SET_STATE(rq, AS_RQ_REMOVED);
1218    if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
1219        atomic_dec(&RQ_IOC(rq)->aic->nr_dispatched);
1220}
1221
1222static void as_deactivate_request(struct request_queue *q, struct request *rq)
1223{
1224    WARN_ON(RQ_STATE(rq) != AS_RQ_REMOVED);
1225    RQ_SET_STATE(rq, AS_RQ_DISPATCHED);
1226    if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
1227        atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched);
1228}
1229
1230/*
1231 * as_queue_empty tells us if there are requests left in the device. It may
1232 * not be the case that a driver can get the next request even if the queue
1233 * is not empty - it is used in the block layer to check for plugging and
1234 * merging opportunities
1235 */
1236static int as_queue_empty(struct request_queue *q)
1237{
1238    struct as_data *ad = q->elevator->elevator_data;
1239
1240    return list_empty(&ad->fifo_list[BLK_RW_ASYNC])
1241        && list_empty(&ad->fifo_list[BLK_RW_SYNC]);
1242}
1243
1244static int
1245as_merge(struct request_queue *q, struct request **req, struct bio *bio)
1246{
1247    struct as_data *ad = q->elevator->elevator_data;
1248    sector_t rb_key = bio->bi_sector + bio_sectors(bio);
1249    struct request *__rq;
1250
1251    /*
1252     * check for front merge
1253     */
1254    __rq = elv_rb_find(&ad->sort_list[bio_data_dir(bio)], rb_key);
1255    if (__rq && elv_rq_merge_ok(__rq, bio)) {
1256        *req = __rq;
1257        return ELEVATOR_FRONT_MERGE;
1258    }
1259
1260    return ELEVATOR_NO_MERGE;
1261}
1262
1263static void as_merged_request(struct request_queue *q, struct request *req,
1264                  int type)
1265{
1266    struct as_data *ad = q->elevator->elevator_data;
1267
1268    /*
1269     * if the merge was a front merge, we need to reposition request
1270     */
1271    if (type == ELEVATOR_FRONT_MERGE) {
1272        as_del_rq_rb(ad, req);
1273        as_add_rq_rb(ad, req);
1274        /*
1275         * Note! At this stage of this and the next function, our next
1276         * request may not be optimal - eg the request may have "grown"
1277         * behind the disk head. We currently don't bother adjusting.
1278         */
1279    }
1280}
1281
1282static void as_merged_requests(struct request_queue *q, struct request *req,
1283                 struct request *next)
1284{
1285    /*
1286     * if next expires before rq, assign its expire time to arq
1287     * and move into next position (next will be deleted) in fifo
1288     */
1289    if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
1290        if (time_before(rq_fifo_time(next), rq_fifo_time(req))) {
1291            list_move(&req->queuelist, &next->queuelist);
1292            rq_set_fifo_time(req, rq_fifo_time(next));
1293        }
1294    }
1295
1296    /*
1297     * kill knowledge of next, this one is a goner
1298     */
1299    as_remove_queued_request(q, next);
1300    as_put_io_context(next);
1301
1302    RQ_SET_STATE(next, AS_RQ_MERGED);
1303}
1304
1305/*
1306 * This is executed in a "deferred" process context, by kblockd. It calls the
1307 * driver's request_fn so the driver can submit that request.
1308 *
1309 * IMPORTANT! This guy will reenter the elevator, so set up all queue global
1310 * state before calling, and don't rely on any state over calls.
1311 *
1312 * FIXME! dispatch queue is not a queue at all!
1313 */
1314static void as_work_handler(struct work_struct *work)
1315{
1316    struct as_data *ad = container_of(work, struct as_data, antic_work);
1317
1318    blk_run_queue(ad->q);
1319}
1320
1321static int as_may_queue(struct request_queue *q, int rw)
1322{
1323    int ret = ELV_MQUEUE_MAY;
1324    struct as_data *ad = q->elevator->elevator_data;
1325    struct io_context *ioc;
1326    if (ad->antic_status == ANTIC_WAIT_REQ ||
1327            ad->antic_status == ANTIC_WAIT_NEXT) {
1328        ioc = as_get_io_context(q->node);
1329        if (ad->io_context == ioc)
1330            ret = ELV_MQUEUE_MUST;
1331        put_io_context(ioc);
1332    }
1333
1334    return ret;
1335}
1336
1337static void as_exit_queue(struct elevator_queue *e)
1338{
1339    struct as_data *ad = e->elevator_data;
1340
1341    del_timer_sync(&ad->antic_timer);
1342    cancel_work_sync(&ad->antic_work);
1343
1344    BUG_ON(!list_empty(&ad->fifo_list[BLK_RW_SYNC]));
1345    BUG_ON(!list_empty(&ad->fifo_list[BLK_RW_ASYNC]));
1346
1347    put_io_context(ad->io_context);
1348    kfree(ad);
1349}
1350
1351/*
1352 * initialize elevator private data (as_data).
1353 */
1354static void *as_init_queue(struct request_queue *q)
1355{
1356    struct as_data *ad;
1357
1358    ad = kmalloc_node(sizeof(*ad), GFP_KERNEL | __GFP_ZERO, q->node);
1359    if (!ad)
1360        return NULL;
1361
1362    ad->q = q; /* Identify what queue the data belongs to */
1363
1364    /* anticipatory scheduling helpers */
1365    ad->antic_timer.function = as_antic_timeout;
1366    ad->antic_timer.data = (unsigned long)q;
1367    init_timer(&ad->antic_timer);
1368    INIT_WORK(&ad->antic_work, as_work_handler);
1369
1370    INIT_LIST_HEAD(&ad->fifo_list[BLK_RW_SYNC]);
1371    INIT_LIST_HEAD(&ad->fifo_list[BLK_RW_ASYNC]);
1372    ad->sort_list[BLK_RW_SYNC] = RB_ROOT;
1373    ad->sort_list[BLK_RW_ASYNC] = RB_ROOT;
1374    ad->fifo_expire[BLK_RW_SYNC] = default_read_expire;
1375    ad->fifo_expire[BLK_RW_ASYNC] = default_write_expire;
1376    ad->antic_expire = default_antic_expire;
1377    ad->batch_expire[BLK_RW_SYNC] = default_read_batch_expire;
1378    ad->batch_expire[BLK_RW_ASYNC] = default_write_batch_expire;
1379
1380    ad->current_batch_expires = jiffies + ad->batch_expire[BLK_RW_SYNC];
1381    ad->write_batch_count = ad->batch_expire[BLK_RW_ASYNC] / 10;
1382    if (ad->write_batch_count < 2)
1383        ad->write_batch_count = 2;
1384
1385    return ad;
1386}
1387
1388/*
1389 * sysfs parts below
1390 */
1391
1392static ssize_t
1393as_var_show(unsigned int var, char *page)
1394{
1395    return sprintf(page, "%d\n", var);
1396}
1397
1398static ssize_t
1399as_var_store(unsigned long *var, const char *page, size_t count)
1400{
1401    char *p = (char *) page;
1402
1403    *var = simple_strtoul(p, &p, 10);
1404    return count;
1405}
1406
1407static ssize_t est_time_show(struct elevator_queue *e, char *page)
1408{
1409    struct as_data *ad = e->elevator_data;
1410    int pos = 0;
1411
1412    pos += sprintf(page+pos, "%lu %% exit probability\n",
1413                100*ad->exit_prob/256);
1414    pos += sprintf(page+pos, "%lu %% probability of exiting without a "
1415                "cooperating process submitting IO\n",
1416                100*ad->exit_no_coop/256);
1417    pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean);
1418    pos += sprintf(page+pos, "%llu sectors new seek distance\n",
1419                (unsigned long long)ad->new_seek_mean);
1420
1421    return pos;
1422}
1423
1424#define SHOW_FUNCTION(__FUNC, __VAR) \
1425static ssize_t __FUNC(struct elevator_queue *e, char *page) \
1426{ \
1427    struct as_data *ad = e->elevator_data; \
1428    return as_var_show(jiffies_to_msecs((__VAR)), (page)); \
1429}
1430SHOW_FUNCTION(as_read_expire_show, ad->fifo_expire[BLK_RW_SYNC]);
1431SHOW_FUNCTION(as_write_expire_show, ad->fifo_expire[BLK_RW_ASYNC]);
1432SHOW_FUNCTION(as_antic_expire_show, ad->antic_expire);
1433SHOW_FUNCTION(as_read_batch_expire_show, ad->batch_expire[BLK_RW_SYNC]);
1434SHOW_FUNCTION(as_write_batch_expire_show, ad->batch_expire[BLK_RW_ASYNC]);
1435#undef SHOW_FUNCTION
1436
1437#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
1438static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
1439{ \
1440    struct as_data *ad = e->elevator_data; \
1441    int ret = as_var_store(__PTR, (page), count); \
1442    if (*(__PTR) < (MIN)) \
1443        *(__PTR) = (MIN); \
1444    else if (*(__PTR) > (MAX)) \
1445        *(__PTR) = (MAX); \
1446    *(__PTR) = msecs_to_jiffies(*(__PTR)); \
1447    return ret; \
1448}
1449STORE_FUNCTION(as_read_expire_store, &ad->fifo_expire[BLK_RW_SYNC], 0, INT_MAX);
1450STORE_FUNCTION(as_write_expire_store,
1451            &ad->fifo_expire[BLK_RW_ASYNC], 0, INT_MAX);
1452STORE_FUNCTION(as_antic_expire_store, &ad->antic_expire, 0, INT_MAX);
1453STORE_FUNCTION(as_read_batch_expire_store,
1454            &ad->batch_expire[BLK_RW_SYNC], 0, INT_MAX);
1455STORE_FUNCTION(as_write_batch_expire_store,
1456            &ad->batch_expire[BLK_RW_ASYNC], 0, INT_MAX);
1457#undef STORE_FUNCTION
1458
1459#define AS_ATTR(name) \
1460    __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store)
1461
1462static struct elv_fs_entry as_attrs[] = {
1463    __ATTR_RO(est_time),
1464    AS_ATTR(read_expire),
1465    AS_ATTR(write_expire),
1466    AS_ATTR(antic_expire),
1467    AS_ATTR(read_batch_expire),
1468    AS_ATTR(write_batch_expire),
1469    __ATTR_NULL
1470};
1471
1472static struct elevator_type iosched_as = {
1473    .ops = {
1474        .elevator_merge_fn = as_merge,
1475        .elevator_merged_fn = as_merged_request,
1476        .elevator_merge_req_fn = as_merged_requests,
1477        .elevator_dispatch_fn = as_dispatch_request,
1478        .elevator_add_req_fn = as_add_request,
1479        .elevator_activate_req_fn = as_activate_request,
1480        .elevator_deactivate_req_fn = as_deactivate_request,
1481        .elevator_queue_empty_fn = as_queue_empty,
1482        .elevator_completed_req_fn = as_completed_request,
1483        .elevator_former_req_fn = elv_rb_former_request,
1484        .elevator_latter_req_fn = elv_rb_latter_request,
1485        .elevator_may_queue_fn = as_may_queue,
1486        .elevator_init_fn = as_init_queue,
1487        .elevator_exit_fn = as_exit_queue,
1488        .trim = as_trim,
1489    },
1490
1491    .elevator_attrs = as_attrs,
1492    .elevator_name = "anticipatory",
1493    .elevator_owner = THIS_MODULE,
1494};
1495
1496static int __init as_init(void)
1497{
1498    elv_register(&iosched_as);
1499
1500    return 0;
1501}
1502
1503static void __exit as_exit(void)
1504{
1505    DECLARE_COMPLETION_ONSTACK(all_gone);
1506    elv_unregister(&iosched_as);
1507    ioc_gone = &all_gone;
1508    /* ioc_gone's update must be visible before reading ioc_count */
1509    smp_wmb();
1510    if (elv_ioc_count_read(as_ioc_count))
1511        wait_for_completion(&all_gone);
1512    synchronize_rcu();
1513}
1514
1515module_init(as_init);
1516module_exit(as_exit);
1517
1518MODULE_AUTHOR("Nick Piggin");
1519MODULE_LICENSE("GPL");
1520MODULE_DESCRIPTION("anticipatory IO scheduler");
1521

Archive Download this file



interactive