linux/block/as-iosched.c
<<
>>
Prefs
   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#define REQ_SYNC        1
  21#define REQ_ASYNC       0
  22
  23/*
  24 * See Documentation/block/as-iosched.txt
  25 */
  26
  27/*
  28 * max time before a read is submitted.
  29 */
  30#define default_read_expire (HZ / 8)
  31
  32/*
  33 * ditto for writes, these limits are not hard, even
  34 * if the disk is capable of satisfying them.
  35 */
  36#define default_write_expire (HZ / 4)
  37
  38/*
  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.
  41 */
  42#define default_read_batch_expire (HZ / 2)
  43
  44/*
  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...
  49 */
  50#define default_write_batch_expire (HZ / 8)
  51
  52/*
  53 * max time we may wait to anticipate a read (default around 6ms)
  54 */
  55#define default_antic_expire ((HZ / 150) ? HZ / 150 : 1)
  56
  57/*
  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.
  63 */
  64#define MAX_THINKTIME (HZ/50UL)
  65
  66/* Bits in as_io_context.state */
  67enum as_io_states {
  68        AS_TASK_RUNNING=0,      /* Process has not exited */
  69        AS_TASK_IOSTARTED,      /* Process has started some IO */
  70        AS_TASK_IORUNNING,      /* Process has completed some IO */
  71};
  72
  73enum anticipation_status {
  74        ANTIC_OFF=0,            /* Not anticipating (normal operation)  */
  75        ANTIC_WAIT_REQ,         /* The last read has not yet completed  */
  76        ANTIC_WAIT_NEXT,        /* Currently anticipating a request vs
  77                                   last read (which has completed) */
  78        ANTIC_FINISHED,         /* Anticipating but have found a candidate
  79                                 * or timed out */
  80};
  81
  82struct as_data {
  83        /*
  84         * run time data
  85         */
  86
  87        struct request_queue *q;        /* the "owner" queue */
  88
  89        /*
  90         * requests (as_rq s) are present on both sort_list and fifo_list
  91         */
  92        struct rb_root sort_list[2];
  93        struct list_head fifo_list[2];
  94
  95        struct request *next_rq[2];     /* next in sort order */
  96        sector_t last_sector[2];        /* last REQ_SYNC & REQ_ASYNC sectors */
  97
  98        unsigned long exit_prob;        /* probability a task will exit while
  99                                           being waited on */
 100        unsigned long exit_no_coop;     /* probablility an exited task will
 101                                           not be part of a later cooperating
 102                                           request */
 103        unsigned long new_ttime_total;  /* mean thinktime on new proc */
 104        unsigned long new_ttime_mean;
 105        u64 new_seek_total;             /* mean seek on new proc */
 106        sector_t new_seek_mean;
 107
 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? */
 116
 117        enum anticipation_status antic_status;
 118        unsigned long antic_start;      /* jiffies: when it started */
 119        struct timer_list antic_timer;  /* anticipatory scheduling timer */
 120        struct work_struct antic_work;  /* Deferred unplugging */
 121        struct io_context *io_context;  /* Identify the expected process */
 122        int ioc_finished; /* IO associated with io_context is finished */
 123        int nr_dispatched;
 124
 125        /*
 126         * settings that change how the i/o scheduler behaves
 127         */
 128        unsigned long fifo_expire[2];
 129        unsigned long batch_expire[2];
 130        unsigned long antic_expire;
 131};
 132
 133/*
 134 * per-request data.
 135 */
 136enum arq_state {
 137        AS_RQ_NEW=0,            /* New - not referenced and not on any lists */
 138        AS_RQ_QUEUED,           /* In the request queue. It belongs to the
 139                                   scheduler */
 140        AS_RQ_DISPATCHED,       /* On the dispatch list. It belongs to the
 141                                   driver now */
 142        AS_RQ_PRESCHED,         /* Debug poisoning for requests being used */
 143        AS_RQ_REMOVED,
 144        AS_RQ_MERGED,
 145        AS_RQ_POSTSCHED,        /* when they shouldn't be */
 146};
 147
 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)
 151
 152static DEFINE_PER_CPU(unsigned long, ioc_count);
 153static struct completion *ioc_gone;
 154static DEFINE_SPINLOCK(ioc_gone_lock);
 155
 156static void as_move_to_dispatch(struct as_data *ad, struct request *rq);
 157static void as_antic_stop(struct as_data *ad);
 158
 159/*
 160 * IO Context helper functions
 161 */
 162
 163/* Called to deallocate the as_io_context */
 164static void free_as_io_context(struct as_io_context *aic)
 165{
 166        kfree(aic);
 167        elv_ioc_count_dec(ioc_count);
 168        if (ioc_gone) {
 169                /*
 170                 * AS scheduler is exiting, grab exit lock and check
 171                 * the pending io context count. If it hits zero,
 172                 * complete ioc_gone and set it back to NULL.
 173                 */
 174                spin_lock(&ioc_gone_lock);
 175                if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
 176                        complete(ioc_gone);
 177                        ioc_gone = NULL;
 178                }
 179                spin_unlock(&ioc_gone_lock);
 180        }
 181}
 182
 183static void as_trim(struct io_context *ioc)
 184{
 185        spin_lock_irq(&ioc->lock);
 186        if (ioc->aic)
 187                free_as_io_context(ioc->aic);
 188        ioc->aic = NULL;
 189        spin_unlock_irq(&ioc->lock);
 190}
 191
 192/* Called when the task exits */
 193static void exit_as_io_context(struct as_io_context *aic)
 194{
 195        WARN_ON(!test_bit(AS_TASK_RUNNING, &aic->state));
 196        clear_bit(AS_TASK_RUNNING, &aic->state);
 197}
 198
 199static struct as_io_context *alloc_as_io_context(void)
 200{
 201        struct as_io_context *ret;
 202
 203        ret = kmalloc(sizeof(*ret), GFP_ATOMIC);
 204        if (ret) {
 205                ret->dtor = free_as_io_context;
 206                ret->exit = exit_as_io_context;
 207                ret->state = 1 << AS_TASK_RUNNING;
 208                atomic_set(&ret->nr_queued, 0);
 209                atomic_set(&ret->nr_dispatched, 0);
 210                spin_lock_init(&ret->lock);
 211                ret->ttime_total = 0;
 212                ret->ttime_samples = 0;
 213                ret->ttime_mean = 0;
 214                ret->seek_total = 0;
 215                ret->seek_samples = 0;
 216                ret->seek_mean = 0;
 217                elv_ioc_count_inc(ioc_count);
 218        }
 219
 220        return ret;
 221}
 222
 223/*
 224 * If the current task has no AS IO context then create one and initialise it.
 225 * Then take a ref on the task's io context and return it.
 226 */
 227static struct io_context *as_get_io_context(int node)
 228{
 229        struct io_context *ioc = get_io_context(GFP_ATOMIC, node);
 230        if (ioc && !ioc->aic) {
 231                ioc->aic = alloc_as_io_context();
 232                if (!ioc->aic) {
 233                        put_io_context(ioc);
 234                        ioc = NULL;
 235                }
 236        }
 237        return ioc;
 238}
 239
 240static void as_put_io_context(struct request *rq)
 241{
 242        struct as_io_context *aic;
 243
 244        if (unlikely(!RQ_IOC(rq)))
 245                return;
 246
 247        aic = RQ_IOC(rq)->aic;
 248
 249        if (rq_is_sync(rq) && aic) {
 250                unsigned long flags;
 251
 252                spin_lock_irqsave(&aic->lock, flags);
 253                set_bit(AS_TASK_IORUNNING, &aic->state);
 254                aic->last_end_request = jiffies;
 255                spin_unlock_irqrestore(&aic->lock, flags);
 256        }
 257
 258        put_io_context(RQ_IOC(rq));
 259}
 260
 261/*
 262 * rb tree support functions
 263 */
 264#define RQ_RB_ROOT(ad, rq)      (&(ad)->sort_list[rq_is_sync((rq))])
 265
 266static void as_add_rq_rb(struct as_data *ad, struct request *rq)
 267{
 268        struct request *alias;
 269
 270        while ((unlikely(alias = elv_rb_add(RQ_RB_ROOT(ad, rq), rq)))) {
 271                as_move_to_dispatch(ad, alias);
 272                as_antic_stop(ad);
 273        }
 274}
 275
 276static inline void as_del_rq_rb(struct as_data *ad, struct request *rq)
 277{
 278        elv_rb_del(RQ_RB_ROOT(ad, rq), rq);
 279}
 280
 281/*
 282 * IO Scheduler proper
 283 */
 284
 285#define MAXBACK (1024 * 1024)   /*
 286                                 * Maximum distance the disk will go backward
 287                                 * for a request.
 288                                 */
 289
 290#define BACK_PENALTY    2
 291
 292/*
 293 * as_choose_req selects the preferred one of two requests of the same data_dir
 294 * ignoring time - eg. timeouts, which is the job of as_dispatch_request
 295 */
 296static struct request *
 297as_choose_req(struct as_data *ad, struct request *rq1, struct request *rq2)
 298{
 299        int data_dir;
 300        sector_t last, s1, s2, d1, d2;
 301        int r1_wrap=0, r2_wrap=0;       /* requests are behind the disk head */
 302        const sector_t maxback = MAXBACK;
 303
 304        if (rq1 == NULL || rq1 == rq2)
 305                return rq2;
 306        if (rq2 == NULL)
 307                return rq1;
 308
 309        data_dir = rq_is_sync(rq1);
 310
 311        last = ad->last_sector[data_dir];
 312        s1 = rq1->sector;
 313        s2 = rq2->sector;
 314
 315        BUG_ON(data_dir != rq_is_sync(rq2));
 316
 317        /*
 318         * Strict one way elevator _except_ in the case where we allow
 319         * short backward seeks which are biased as twice the cost of a
 320         * similar forward seek.
 321         */
 322        if (s1 >= last)
 323                d1 = s1 - last;
 324        else if (s1+maxback >= last)
 325                d1 = (last - s1)*BACK_PENALTY;
 326        else {
 327                r1_wrap = 1;
 328                d1 = 0; /* shut up, gcc */
 329        }
 330
 331        if (s2 >= last)
 332                d2 = s2 - last;
 333        else if (s2+maxback >= last)
 334                d2 = (last - s2)*BACK_PENALTY;
 335        else {
 336                r2_wrap = 1;
 337                d2 = 0;
 338        }
 339
 340        /* Found required data */
 341        if (!r1_wrap && r2_wrap)
 342                return rq1;
 343        else if (!r2_wrap && r1_wrap)
 344                return rq2;
 345        else if (r1_wrap && r2_wrap) {
 346                /* both behind the head */
 347                if (s1 <= s2)
 348                        return rq1;
 349                else
 350                        return rq2;
 351        }
 352
 353        /* Both requests in front of the head */
 354        if (d1 < d2)
 355                return rq1;
 356        else if (d2 < d1)
 357                return rq2;
 358        else {
 359                if (s1 >= s2)
 360                        return rq1;
 361                else
 362                        return rq2;
 363        }
 364}
 365
 366/*
 367 * as_find_next_rq finds the next request after @prev in elevator order.
 368 * this with as_choose_req form the basis for how the scheduler chooses
 369 * what request to process next. Anticipation works on top of this.
 370 */
 371static struct request *
 372as_find_next_rq(struct as_data *ad, struct request *last)
 373{
 374        struct rb_node *rbnext = rb_next(&last->rb_node);
 375        struct rb_node *rbprev = rb_prev(&last->rb_node);
 376        struct request *next = NULL, *prev = NULL;
 377
 378        BUG_ON(RB_EMPTY_NODE(&last->rb_node));
 379
 380        if (rbprev)
 381                prev = rb_entry_rq(rbprev);
 382
 383        if (rbnext)
 384                next = rb_entry_rq(rbnext);
 385        else {
 386                const int data_dir = rq_is_sync(last);
 387
 388                rbnext = rb_first(&ad->sort_list[data_dir]);
 389                if (rbnext && rbnext != &last->rb_node)
 390                        next = rb_entry_rq(rbnext);
 391        }
 392
 393        return as_choose_req(ad, next, prev);
 394}
 395
 396/*
 397 * anticipatory scheduling functions follow
 398 */
 399
 400/*
 401 * as_antic_expired tells us when we have anticipated too long.
 402 * The funny "absolute difference" math on the elapsed time is to handle
 403 * jiffy wraps, and disks which have been idle for 0x80000000 jiffies.
 404 */
 405static int as_antic_expired(struct as_data *ad)
 406{
 407        long delta_jif;
 408
 409        delta_jif = jiffies - ad->antic_start;
 410        if (unlikely(delta_jif < 0))
 411                delta_jif = -delta_jif;
 412        if (delta_jif < ad->antic_expire)
 413                return 0;
 414
 415        return 1;
 416}
 417
 418/*
 419 * as_antic_waitnext starts anticipating that a nice request will soon be
 420 * submitted. See also as_antic_waitreq
 421 */
 422static void as_antic_waitnext(struct as_data *ad)
 423{
 424        unsigned long timeout;
 425
 426        BUG_ON(ad->antic_status != ANTIC_OFF
 427                        && ad->antic_status != ANTIC_WAIT_REQ);
 428
 429        timeout = ad->antic_start + ad->antic_expire;
 430
 431        mod_timer(&ad->antic_timer, timeout);
 432
 433        ad->antic_status = ANTIC_WAIT_NEXT;
 434}
 435
 436/*
 437 * as_antic_waitreq starts anticipating. We don't start timing the anticipation
 438 * until the request that we're anticipating on has finished. This means we
 439 * are timing from when the candidate process wakes up hopefully.
 440 */
 441static void as_antic_waitreq(struct as_data *ad)
 442{
 443        BUG_ON(ad->antic_status == ANTIC_FINISHED);
 444        if (ad->antic_status == ANTIC_OFF) {
 445                if (!ad->io_context || ad->ioc_finished)
 446                        as_antic_waitnext(ad);
 447                else
 448                        ad->antic_status = ANTIC_WAIT_REQ;
 449        }
 450}
 451
 452/*
 453 * This is called directly by the functions in this file to stop anticipation.
 454 * We kill the timer and schedule a call to the request_fn asap.
 455 */
 456static void as_antic_stop(struct as_data *ad)
 457{
 458        int status = ad->antic_status;
 459
 460        if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) {
 461                if (status == ANTIC_WAIT_NEXT)
 462                        del_timer(&ad->antic_timer);
 463                ad->antic_status = ANTIC_FINISHED;
 464                /* see as_work_handler */
 465                kblockd_schedule_work(ad->q, &ad->antic_work);
 466        }
 467}
 468
 469/*
 470 * as_antic_timeout is the timer function set by as_antic_waitnext.
 471 */
 472static void as_antic_timeout(unsigned long data)
 473{
 474        struct request_queue *q = (struct request_queue *)data;
 475        struct as_data *ad = q->elevator->elevator_data;
 476        unsigned long flags;
 477
 478        spin_lock_irqsave(q->queue_lock, flags);
 479        if (ad->antic_status == ANTIC_WAIT_REQ
 480                        || ad->antic_status == ANTIC_WAIT_NEXT) {
 481                struct as_io_context *aic;
 482                spin_lock(&ad->io_context->lock);
 483                aic = ad->io_context->aic;
 484
 485                ad->antic_status = ANTIC_FINISHED;
 486                kblockd_schedule_work(q, &ad->antic_work);
 487
 488                if (aic->ttime_samples == 0) {
 489                        /* process anticipated on has exited or timed out*/
 490                        ad->exit_prob = (7*ad->exit_prob + 256)/8;
 491                }
 492                if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
 493                        /* process not "saved" by a cooperating request */
 494                        ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8;
 495                }
 496                spin_unlock(&ad->io_context->lock);
 497        }
 498        spin_unlock_irqrestore(q->queue_lock, flags);
 499}
 500
 501static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic,
 502                                unsigned long ttime)
 503{
 504        /* fixed point: 1.0 == 1<<8 */
 505        if (aic->ttime_samples == 0) {
 506                ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8;
 507                ad->new_ttime_mean = ad->new_ttime_total / 256;
 508
 509                ad->exit_prob = (7*ad->exit_prob)/8;
 510        }
 511        aic->ttime_samples = (7*aic->ttime_samples + 256) / 8;
 512        aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8;
 513        aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples;
 514}
 515
 516static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic,
 517                                sector_t sdist)
 518{
 519        u64 total;
 520
 521        if (aic->seek_samples == 0) {
 522                ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8;
 523                ad->new_seek_mean = ad->new_seek_total / 256;
 524        }
 525
 526        /*
 527         * Don't allow the seek distance to get too large from the
 528         * odd fragment, pagein, etc
 529         */
 530        if (aic->seek_samples <= 60) /* second&third seek */
 531                sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024);
 532        else
 533                sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64);
 534
 535        aic->seek_samples = (7*aic->seek_samples + 256) / 8;
 536        aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8;
 537        total = aic->seek_total + (aic->seek_samples/2);
 538        do_div(total, aic->seek_samples);
 539        aic->seek_mean = (sector_t)total;
 540}
 541
 542/*
 543 * as_update_iohist keeps a decaying histogram of IO thinktimes, and
 544 * updates @aic->ttime_mean based on that. It is called when a new
 545 * request is queued.
 546 */
 547static void as_update_iohist(struct as_data *ad, struct as_io_context *aic,
 548                                struct request *rq)
 549{
 550        int data_dir = rq_is_sync(rq);
 551        unsigned long thinktime = 0;
 552        sector_t seek_dist;
 553
 554        if (aic == NULL)
 555                return;
 556
 557        if (data_dir == REQ_SYNC) {
 558                unsigned long in_flight = atomic_read(&aic->nr_queued)
 559                                        + atomic_read(&aic->nr_dispatched);
 560                spin_lock(&aic->lock);
 561                if (test_bit(AS_TASK_IORUNNING, &aic->state) ||
 562                        test_bit(AS_TASK_IOSTARTED, &aic->state)) {
 563                        /* Calculate read -> read thinktime */
 564                        if (test_bit(AS_TASK_IORUNNING, &aic->state)
 565                                                        && in_flight == 0) {
 566                                thinktime = jiffies - aic->last_end_request;
 567                                thinktime = min(thinktime, MAX_THINKTIME-1);
 568                        }
 569                        as_update_thinktime(ad, aic, thinktime);
 570
 571                        /* Calculate read -> read seek distance */
 572                        if (aic->last_request_pos < rq->sector)
 573                                seek_dist = rq->sector - aic->last_request_pos;
 574                        else
 575                                seek_dist = aic->last_request_pos - rq->sector;
 576                        as_update_seekdist(ad, aic, seek_dist);
 577                }
 578                aic->last_request_pos = rq->sector + rq->nr_sectors;
 579                set_bit(AS_TASK_IOSTARTED, &aic->state);
 580                spin_unlock(&aic->lock);
 581        }
 582}
 583
 584/*
 585 * as_close_req decides if one request is considered "close" to the
 586 * previous one issued.
 587 */
 588static int as_close_req(struct as_data *ad, struct as_io_context *aic,
 589                        struct request *rq)
 590{
 591        unsigned long delay;    /* jiffies */
 592        sector_t last = ad->last_sector[ad->batch_data_dir];
 593        sector_t next = rq->sector;
 594        sector_t delta; /* acceptable close offset (in sectors) */
 595        sector_t s;
 596
 597        if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished)
 598                delay = 0;
 599        else
 600                delay = jiffies - ad->antic_start;
 601
 602        if (delay == 0)
 603                delta = 8192;
 604        else if (delay <= (20 * HZ / 1000) && delay <= ad->antic_expire)
 605                delta = 8192 << delay;
 606        else
 607                return 1;
 608
 609        if ((last <= next + (delta>>1)) && (next <= last + delta))
 610                return 1;
 611
 612        if (last < next)
 613                s = next - last;
 614        else
 615                s = last - next;
 616
 617        if (aic->seek_samples == 0) {
 618                /*
 619                 * Process has just started IO. Use past statistics to
 620                 * gauge success possibility
 621                 */
 622                if (ad->new_seek_mean > s) {
 623                        /* this request is better than what we're expecting */
 624                        return 1;
 625                }
 626
 627        } else {
 628                if (aic->seek_mean > s) {
 629                        /* this request is better than what we're expecting */
 630                        return 1;
 631                }
 632        }
 633
 634        return 0;
 635}
 636
 637/*
 638 * as_can_break_anticipation returns true if we have been anticipating this
 639 * request.
 640 *
 641 * It also returns true if the process against which we are anticipating
 642 * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to
 643 * dispatch it ASAP, because we know that application will not be submitting
 644 * any new reads.
 645 *
 646 * If the task which has submitted the request has exited, break anticipation.
 647 *
 648 * If this task has queued some other IO, do not enter enticipation.
 649 */
 650static int as_can_break_anticipation(struct as_data *ad, struct request *rq)
 651{
 652        struct io_context *ioc;
 653        struct as_io_context *aic;
 654
 655        ioc = ad->io_context;
 656        BUG_ON(!ioc);
 657        spin_lock(&ioc->lock);
 658
 659        if (rq && ioc == RQ_IOC(rq)) {
 660                /* request from same process */
 661                spin_unlock(&ioc->lock);
 662                return 1;
 663        }
 664
 665        if (ad->ioc_finished && as_antic_expired(ad)) {
 666                /*
 667                 * In this situation status should really be FINISHED,
 668                 * however the timer hasn't had the chance to run yet.
 669                 */
 670                spin_unlock(&ioc->lock);
 671                return 1;
 672        }
 673
 674        aic = ioc->aic;
 675        if (!aic) {
 676                spin_unlock(&ioc->lock);
 677                return 0;
 678        }
 679
 680        if (atomic_read(&aic->nr_queued) > 0) {
 681                /* process has more requests queued */
 682                spin_unlock(&ioc->lock);
 683                return 1;
 684        }
 685
 686        if (atomic_read(&aic->nr_dispatched) > 0) {
 687                /* process has more requests dispatched */
 688                spin_unlock(&ioc->lock);
 689                return 1;
 690        }
 691
 692        if (rq && rq_is_sync(rq) && as_close_req(ad, aic, rq)) {
 693                /*
 694                 * Found a close request that is not one of ours.
 695                 *
 696                 * This makes close requests from another process update
 697                 * our IO history. Is generally useful when there are
 698                 * two or more cooperating processes working in the same
 699                 * area.
 700                 */
 701                if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
 702                        if (aic->ttime_samples == 0)
 703                                ad->exit_prob = (7*ad->exit_prob + 256)/8;
 704
 705                        ad->exit_no_coop = (7*ad->exit_no_coop)/8;
 706                }
 707
 708                as_update_iohist(ad, aic, rq);
 709                spin_unlock(&ioc->lock);
 710                return 1;
 711        }
 712
 713        if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
 714                /* process anticipated on has exited */
 715                if (aic->ttime_samples == 0)
 716                        ad->exit_prob = (7*ad->exit_prob + 256)/8;
 717
 718                if (ad->exit_no_coop > 128) {
 719                        spin_unlock(&ioc->lock);
 720                        return 1;
 721                }
 722        }
 723
 724        if (aic->ttime_samples == 0) {
 725                if (ad->new_ttime_mean > ad->antic_expire) {
 726                        spin_unlock(&ioc->lock);
 727                        return 1;
 728                }
 729                if (ad->exit_prob * ad->exit_no_coop > 128*256) {
 730                        spin_unlock(&ioc->lock);
 731                        return 1;
 732                }
 733        } else if (aic->ttime_mean > ad->antic_expire) {
 734                /* the process thinks too much between requests */
 735                spin_unlock(&ioc->lock);
 736                return 1;
 737        }
 738        spin_unlock(&ioc->lock);
 739        return 0;
 740}
 741
 742/*
 743 * as_can_anticipate indicates whether we should either run rq
 744 * or keep anticipating a better request.
 745 */
 746static int as_can_anticipate(struct as_data *ad, struct request *rq)
 747{
 748#if 0 /* disable for now, we need to check tag level as well */
 749        /*
 750         * SSD device without seek penalty, disable idling
 751         */
 752        if (blk_queue_nonrot(ad->q)) axman
 753                return 0;
 754#endif
 755
 756        if (!ad->io_context)
 757                /*
 758                 * Last request submitted was a write
 759                 */
 760                return 0;
 761
 762        if (ad->antic_status == ANTIC_FINISHED)
 763                /*
 764                 * Don't restart if we have just finished. Run the next request
 765                 */
 766                return 0;
 767
 768        if (as_can_break_anticipation(ad, rq))
 769                /*
 770                 * This request is a good candidate. Don't keep anticipating,
 771                 * run it.
 772                 */
 773                return 0;
 774
 775        /*
 776         * OK from here, we haven't finished, and don't have a decent request!
 777         * Status is either ANTIC_OFF so start waiting,
 778         * ANTIC_WAIT_REQ so continue waiting for request to finish
 779         * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request.
 780         */
 781
 782        return 1;
 783}
 784
 785/*
 786 * as_update_rq must be called whenever a request (rq) is added to
 787 * the sort_list. This function keeps caches up to date, and checks if the
 788 * request might be one we are "anticipating"
 789 */
 790static void as_update_rq(struct as_data *ad, struct request *rq)
 791{
 792        const int data_dir = rq_is_sync(rq);
 793
 794        /* keep the next_rq cache up to date */
 795        ad->next_rq[data_dir] = as_choose_req(ad, rq, ad->next_rq[data_dir]);
 796
 797        /*
 798         * have we been anticipating this request?
 799         * or does it come from the same process as the one we are anticipating
 800         * for?
 801         */
 802        if (ad->antic_status == ANTIC_WAIT_REQ
 803                        || ad->antic_status == ANTIC_WAIT_NEXT) {
 804                if (as_can_break_anticipation(ad, rq))
 805                        as_antic_stop(ad);
 806        }
 807}
 808
 809/*
 810 * Gathers timings and resizes the write batch automatically
 811 */
 812static void update_write_batch(struct as_data *ad)
 813{
 814        unsigned long batch = ad->batch_expire[REQ_ASYNC];
 815        long write_time;
 816
 817        write_time = (jiffies - ad->current_batch_expires) + batch;
 818        if (write_time < 0)
 819                write_time = 0;
 820
 821        if (write_time > batch && !ad->write_batch_idled) {
 822                if (write_time > batch * 3)
 823                        ad->write_batch_count /= 2;
 824                else
 825                        ad->write_batch_count--;
 826        } else if (write_time < batch && ad->current_write_count == 0) {
 827                if (batch > write_time * 3)
 828                        ad->write_batch_count *= 2;
 829                else
 830                        ad->write_batch_count++;
 831        }
 832
 833        if (ad->write_batch_count < 1)
 834                ad->write_batch_count = 1;
 835}
 836
 837/*
 838 * as_completed_request is to be called when a request has completed and
 839 * returned something to the requesting process, be it an error or data.
 840 */
 841static void as_completed_request(struct request_queue *q, struct request *rq)
 842{
 843        struct as_data *ad = q->elevator->elevator_data;
 844
 845        WARN_ON(!list_empty(&rq->queuelist));
 846
 847        if (RQ_STATE(rq) != AS_RQ_REMOVED) {
 848                WARN(1, "rq->state %d\n", RQ_STATE(rq));
 849                goto out;
 850        }
 851
 852        if (ad->changed_batch && ad->nr_dispatched == 1) {
 853                ad->current_batch_expires = jiffies +
 854                                        ad->batch_expire[ad->batch_data_dir];
 855                kblockd_schedule_work(q, &ad->antic_work);
 856                ad->changed_batch = 0;
 857
 858                if (ad->batch_data_dir == REQ_SYNC)
 859                        ad->new_batch = 1;
 860        }
 861        WARN_ON(ad->nr_dispatched == 0);
 862        ad->nr_dispatched--;
 863
 864        /*
 865         * Start counting the batch from when a request of that direction is
 866         * actually serviced. This should help devices with big TCQ windows
 867         * and writeback caches
 868         */
 869        if (ad->new_batch && ad->batch_data_dir == rq_is_sync(rq)) {
 870                update_write_batch(ad);
 871                ad->current_batch_expires = jiffies +
 872                                ad->batch_expire[REQ_SYNC];
 873                ad->new_batch = 0;
 874        }
 875
 876        if (ad->io_context == RQ_IOC(rq) && ad->io_context) {
 877                ad->antic_start = jiffies;
 878                ad->ioc_finished = 1;
 879                if (ad->antic_status == ANTIC_WAIT_REQ) {
 880                        /*
 881                         * We were waiting on this request, now anticipate
 882                         * the next one
 883                         */
 884                        as_antic_waitnext(ad);
 885                }
 886        }
 887
 888        as_put_io_context(rq);
 889out:
 890        RQ_SET_STATE(rq, AS_RQ_POSTSCHED);
 891}
 892
 893/*
 894 * as_remove_queued_request removes a request from the pre dispatch queue
 895 * without updating refcounts. It is expected the caller will drop the
 896 * reference unless it replaces the request at somepart of the elevator
 897 * (ie. the dispatch queue)
 898 */
 899static void as_remove_queued_request(struct request_queue *q,
 900                                     struct request *rq)
 901{
 902        const int data_dir = rq_is_sync(rq);
 903        struct as_data *ad = q->elevator->elevator_data;
 904        struct io_context *ioc;
 905
 906        WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED);
 907
 908        ioc = RQ_IOC(rq);
 909        if (ioc && ioc->aic) {
 910                BUG_ON(!atomic_read(&ioc->aic->nr_queued));
 911                atomic_dec(&ioc->aic->nr_queued);
 912        }
 913
 914        /*
 915         * Update the "next_rq" cache if we are about to remove its
 916         * entry
 917         */
 918        if (ad->next_rq[data_dir] == rq)
 919                ad->next_rq[data_dir] = as_find_next_rq(ad, rq);
 920
 921        rq_fifo_clear(rq);
 922        as_del_rq_rb(ad, rq);
 923}
 924
 925/*
 926 * as_fifo_expired returns 0 if there are no expired requests on the fifo,
 927 * 1 otherwise.  It is ratelimited so that we only perform the check once per
 928 * `fifo_expire' interval.  Otherwise a large number of expired requests
 929 * would create a hopeless seekstorm.
 930 *
 931 * See as_antic_expired comment.
 932 */
 933static int as_fifo_expired(struct as_data *ad, int adir)
 934{
 935        struct request *rq;
 936        long delta_jif;
 937
 938        delta_jif = jiffies - ad->last_check_fifo[adir];
 939        if (unlikely(delta_jif < 0))
 940                delta_jif = -delta_jif;
 941        if (delta_jif < ad->fifo_expire[adir])
 942                return 0;
 943
 944        ad->last_check_fifo[adir] = jiffies;
 945
 946        if (list_empty(&ad->fifo_list[adir]))
 947                return 0;
 948
 949        rq = rq_entry_fifo(ad->fifo_list[adir].next);
 950
 951        return time_after(jiffies, rq_fifo_time(rq));
 952}
 953
 954/*
 955 * as_batch_expired returns true if the current batch has expired. A batch
 956 * is a set of reads or a set of writes.
 957 */
 958static inline int as_batch_expired(struct as_data *ad)
 959{
 960        if (ad->changed_batch || ad->new_batch)
 961                return 0;
 962
 963        if (ad->batch_data_dir == REQ_SYNC)
 964                /* TODO! add a check so a complete fifo gets written? */
 965                return time_after(jiffies, ad->current_batch_expires);
 966
 967        return time_after(jiffies, ad->current_batch_expires)
 968                || ad->current_write_count == 0;
 969}
 970
 971/*
 972 * move an entry to dispatch queue
 973 */
 974static void as_move_to_dispatch(struct as_data *ad, struct request *rq)
 975{
 976        const int data_dir = rq_is_sync(rq);
 977
 978        BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 979
 980        as_antic_stop(ad);
 981        ad->antic_status = ANTIC_OFF;
 982
 983        /*
 984         * This has to be set in order to be correctly updated by
 985         * as_find_next_rq
 986         */
 987        ad->last_sector[data_dir] = rq->sector + rq->nr_sectors;
 988
 989        if (data_dir == REQ_SYNC) {
 990                struct io_context *ioc = RQ_IOC(rq);
 991                /* In case we have to anticipate after this */
 992                copy_io_context(&ad->io_context, &ioc);
 993        } else {
 994                if (ad->io_context) {
 995                        put_io_context(ad->io_context);
 996                        ad->io_context = NULL;
 997                }
 998
 999                if (ad->current_write_count != 0)
1000                        ad->current_write_count--;
1001        }
1002        ad->ioc_finished = 0;
1003
1004        ad->next_rq[data_dir] = as_find_next_rq(ad, rq);
1005
1006        /*
1007         * take it off the sort and fifo list, add to dispatch queue
1008         */
1009        as_remove_queued_request(ad->q, rq);
1010        WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED);
1011
1012        elv_dispatch_sort(ad->q, rq);
1013
1014        RQ_SET_STATE(rq, AS_RQ_DISPATCHED);
1015        if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
1016                atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched);
1017        ad->nr_dispatched++;
1018}
1019
1020/*
1021 * as_dispatch_request selects the best request according to
1022 * read/write expire, batch expire, etc, and moves it to the dispatch
1023 * queue. Returns 1 if a request was found, 0 otherwise.
1024 */
1025static int as_dispatch_request(struct request_queue *q, int force)
1026{
1027        struct as_data *ad = q->elevator->elevator_data;
1028        const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]);
1029        const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]);
1030        struct request *rq;
1031
1032        if (unlikely(force)) {
1033                /*
1034                 * Forced dispatch, accounting is useless.  Reset
1035                 * accounting states and dump fifo_lists.  Note that
1036                 * batch_data_dir is reset to REQ_SYNC to avoid
1037                 * screwing write batch accounting as write batch
1038                 * accounting occurs on W->R transition.
1039                 */
1040                int dispatched = 0;
1041
1042                ad->batch_data_dir = REQ_SYNC;
1043                ad->changed_batch = 0;
1044                ad->new_batch = 0;
1045
1046                while (ad->next_rq[REQ_SYNC]) {
1047                        as_move_to_dispatch(ad, ad->next_rq[REQ_SYNC]);
1048                        dispatched++;
1049                }
1050                ad->last_check_fifo[REQ_SYNC] = jiffies;
1051
1052                while (ad->next_rq[REQ_ASYNC]) {
1053                        as_move_to_dispatch(ad, ad->next_rq[REQ_ASYNC]);
1054                        dispatched++;
1055                }
1056                ad->last_check_fifo[REQ_ASYNC] = jiffies;
1057
1058                return dispatched;
1059        }
1060
1061        /* Signal that the write batch was uncontended, so we can't time it */
1062        if (ad->batch_data_dir == REQ_ASYNC && !reads) {
1063                if (ad->current_write_count == 0 || !writes)
1064                        ad->write_batch_idled = 1;
1065        }
1066
1067        if (!(reads || writes)
1068                || ad->antic_status == ANTIC_WAIT_REQ
1069                || ad->antic_status == ANTIC_WAIT_NEXT
1070                || ad->changed_batch)
1071                return 0;
1072
1073        if (!(reads && writes && as_batch_expired(ad))) {
1074                /*
1075                 * batch is still running or no reads or no writes
1076                 */
1077                rq = ad->next_rq[ad->batch_data_dir];
1078
1079                if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) {
1080                        if (as_fifo_expired(ad, REQ_SYNC))
1081                                goto fifo_expired;
1082
1083                        if (as_can_anticipate(ad, rq)) {
1084                                as_antic_waitreq(ad);
1085                                return 0;
1086                        }
1087                }
1088
1089                if (rq) {
1090                        /* we have a "next request" */
1091                        if (reads && !writes)
1092                                ad->current_batch_expires =
1093                                        jiffies + ad->batch_expire[REQ_SYNC];
1094                        goto dispatch_request;
1095                }
1096        }
1097
1098        /*
1099         * at this point we are not running a batch. select the appropriate
1100         * data direction (read / write)
1101         */
1102
1103        if (reads) {
1104                BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_SYNC]));
1105
1106                if (writes && ad->batch_data_dir == REQ_SYNC)
1107                        /*
1108                         * Last batch was a read, switch to writes
1109                         */
1110                        goto dispatch_writes;
1111
1112                if (ad->batch_data_dir == REQ_ASYNC) {
1113                        WARN_ON(ad->new_batch);
1114                        ad->changed_batch = 1;
1115                }
1116                ad->batch_data_dir = REQ_SYNC;
1117                rq = rq_entry_fifo(ad->fifo_list[REQ_SYNC].next);
1118                ad->last_check_fifo[ad->batch_data_dir] = jiffies;
1119                goto dispatch_request;
1120        }
1121
1122        /*
1123         * the last batch was a read
1124         */
1125
1126        if (writes) {
1127dispatch_writes:
1128                BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_ASYNC]));
1129
1130                if (ad->batch_data_dir == REQ_SYNC) {
1131                        ad->changed_batch = 1;
1132
1133                        /*
1134                         * new_batch might be 1 when the queue runs out of
1135                         * reads. A subsequent submission of a write might
1136                         * cause a change of batch before the read is finished.
1137                         */
1138                        ad->new_batch = 0;
1139                }
1140                ad->batch_data_dir = REQ_ASYNC;
1141                ad->current_write_count = ad->write_batch_count;
1142                ad->write_batch_idled = 0;
1143                rq = rq_entry_fifo(ad->fifo_list[REQ_ASYNC].next);
1144                ad->last_check_fifo[REQ_ASYNC] = jiffies;
1145                goto dispatch_request;
1146        }
1147
1148        BUG();
1149        return 0;
1150
1151dispatch_request:
1152        /*
1153         * If a request has expired, service it.
1154         */
1155
1156        if (as_fifo_expired(ad, ad->batch_data_dir)) {
1157fifo_expired:
1158                rq = rq_entry_fifo(ad->fifo_list[ad->batch_data_dir].next);
1159        }
1160
1161        if (ad->changed_batch) {
1162                WARN_ON(ad->new_batch);
1163
1164                if (ad->nr_dispatched)
1165                        return 0;
1166
1167                if (ad->batch_data_dir == REQ_ASYNC)
1168                        ad->current_batch_expires = jiffies +
1169                                        ad->batch_expire[REQ_ASYNC];
1170                else
1171                        ad->new_batch = 1;
1172
1173                ad->changed_batch = 0;
1174        }
1175
1176        /*
1177         * rq is the selected appropriate request.
1178         */
1179        as_move_to_dispatch(ad, rq);
1180
1181        return 1;
1182}
1183
1184/*
1185 * add rq to rbtree and fifo
1186 */
1187static void as_add_request(struct request_queue *q, struct request *rq)
1188{
1189        struct as_data *ad = q->elevator->elevator_data;
1190        int data_dir;
1191
1192        RQ_SET_STATE(rq, AS_RQ_NEW);
1193
1194        data_dir = rq_is_sync(rq);
1195
1196        rq->elevator_private = as_get_io_context(q->node);
1197
1198        if (RQ_IOC(rq)) {
1199                as_update_iohist(ad, RQ_IOC(rq)->aic, rq);
1200                atomic_inc(&RQ_IOC(rq)->aic->nr_queued);
1201        }
1202
1203        as_add_rq_rb(ad, rq);
1204
1205        /*
1206         * set expire time and add to fifo list
1207         */
1208        rq_set_fifo_time(rq, jiffies + ad->fifo_expire[data_dir]);
1209        list_add_tail(&rq->queuelist, &ad->fifo_list[data_dir]);
1210
1211        as_update_rq(ad, rq); /* keep state machine up to date */
1212        RQ_SET_STATE(rq, AS_RQ_QUEUED);
1213}
1214
1215static void as_activate_request(struct request_queue *q, struct request *rq)
1216{
1217        WARN_ON(RQ_STATE(rq) != AS_RQ_DISPATCHED);
1218        RQ_SET_STATE(rq, AS_RQ_REMOVED);
1219        if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
1220                atomic_dec(&RQ_IOC(rq)->aic->nr_dispatched);
1221}
1222
1223static void as_deactivate_request(struct request_queue *q, struct request *rq)
1224{
1225        WARN_ON(RQ_STATE(rq) != AS_RQ_REMOVED);
1226        RQ_SET_STATE(rq, AS_RQ_DISPATCHED);
1227        if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
1228                atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched);
1229}
1230
1231/*
1232 * as_queue_empty tells us if there are requests left in the device. It may
1233 * not be the case that a driver can get the next request even if the queue
1234 * is not empty - it is used in the block layer to check for plugging and
1235 * merging opportunities
1236 */
1237static int as_queue_empty(struct request_queue *q)
1238{
1239        struct as_data *ad = q->elevator->elevator_data;
1240
1241        return list_empty(&ad->fifo_list[REQ_ASYNC])
1242                && list_empty(&ad->fifo_list[REQ_SYNC]);
1243}
1244
1245static int
1246as_merge(struct request_queue *q, struct request **req, struct bio *bio)
1247{
1248        struct as_data *ad = q->elevator->elevator_data;
1249        sector_t rb_key = bio->bi_sector + bio_sectors(bio);
1250        struct request *__rq;
1251
1252        /*
1253         * check for front merge
1254         */
1255        __rq = elv_rb_find(&ad->sort_list[bio_data_dir(bio)], rb_key);
1256        if (__rq && elv_rq_merge_ok(__rq, bio)) {
1257                *req = __rq;
1258                return ELEVATOR_FRONT_MERGE;
1259        }
1260
1261        return ELEVATOR_NO_MERGE;
1262}
1263
1264static void as_merged_request(struct request_queue *q, struct request *req,
1265                              int type)
1266{
1267        struct as_data *ad = q->elevator->elevator_data;
1268
1269        /*
1270         * if the merge was a front merge, we need to reposition request
1271         */
1272        if (type == ELEVATOR_FRONT_MERGE) {
1273                as_del_rq_rb(ad, req);
1274                as_add_rq_rb(ad, req);
1275                /*
1276                 * Note! At this stage of this and the next function, our next
1277                 * request may not be optimal - eg the request may have "grown"
1278                 * behind the disk head. We currently don't bother adjusting.
1279                 */
1280        }
1281}
1282
1283static void as_merged_requests(struct request_queue *q, struct request *req,
1284                                struct request *next)
1285{
1286        /*
1287         * if next expires before rq, assign its expire time to arq
1288         * and move into next position (next will be deleted) in fifo
1289         */
1290        if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
1291                if (time_before(rq_fifo_time(next), rq_fifo_time(req))) {
1292                        list_move(&req->queuelist, &next->queuelist);
1293                        rq_set_fifo_time(req, rq_fifo_time(next));
1294                }
1295        }
1296
1297        /*
1298         * kill knowledge of next, this one is a goner
1299         */
1300        as_remove_queued_request(q, next);
1301        as_put_io_context(next);
1302
1303        RQ_SET_STATE(next, AS_RQ_MERGED);
1304}
1305
1306/*
1307 * This is executed in a "deferred" process context, by kblockd. It calls the
1308 * driver's request_fn so the driver can submit that request.
1309 *
1310 * IMPORTANT! This guy will reenter the elevator, so set up all queue global
1311 * state before calling, and don't rely on any state over calls.
1312 *
1313 * FIXME! dispatch queue is not a queue at all!
1314 */
1315static void as_work_handler(struct work_struct *work)
1316{
1317        struct as_data *ad = container_of(work, struct as_data, antic_work);
1318        struct request_queue *q = ad->q;
1319        unsigned long flags;
1320
1321        spin_lock_irqsave(q->queue_lock, flags);
1322        blk_start_queueing(q);
1323        spin_unlock_irqrestore(q->queue_lock, flags);
1324}
1325
1326static int as_may_queue(struct request_queue *q, int rw)
1327{
1328        int ret = ELV_MQUEUE_MAY;
1329        struct as_data *ad = q->elevator->elevator_data;
1330        struct io_context *ioc;
1331        if (ad->antic_status == ANTIC_WAIT_REQ ||
1332                        ad->antic_status == ANTIC_WAIT_NEXT) {
1333                ioc = as_get_io_context(q->node);
1334                if (ad->io_context == ioc)
1335                        ret = ELV_MQUEUE_MUST;
1336                put_io_context(ioc);
1337        }
1338
1339        return ret;
1340}
1341
1342static void as_exit_queue(elevator_t *e)
1343{
1344        struct as_data *ad = e->elevator_data;
1345
1346        del_timer_sync(&ad->antic_timer);
1347        kblockd_flush_work(&ad->antic_work);
1348
1349        BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC]));
1350        BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC]));
1351
1352        put_io_context(ad->io_context);
1353        kfree(ad);
1354}
1355
1356/*
1357 * initialize elevator private data (as_data).
1358 */
1359static void *as_init_queue(struct request_queue *q)
1360{
1361        struct as_data *ad;
1362
1363        ad = kmalloc_node(sizeof(*ad), GFP_KERNEL | __GFP_ZERO, q->node);
1364        if (!ad)
1365                return NULL;
1366
1367        ad->q = q; /* Identify what queue the data belongs to */
1368
1369        /* anticipatory scheduling helpers */
1370        ad->antic_timer.function = as_antic_timeout;
1371        ad->antic_timer.data = (unsigned long)q;
1372        init_timer(&ad->antic_timer);
1373        INIT_WORK(&ad->antic_work, as_work_handler);
1374
1375        INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]);
1376        INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]);
1377        ad->sort_list[REQ_SYNC] = RB_ROOT;
1378        ad->sort_list[REQ_ASYNC] = RB_ROOT;
1379        ad->fifo_expire[REQ_SYNC] = default_read_expire;
1380        ad->fifo_expire[REQ_ASYNC] = default_write_expire;
1381        ad->antic_expire = default_antic_expire;
1382        ad->batch_expire[REQ_SYNC] = default_read_batch_expire;
1383        ad->batch_expire[REQ_ASYNC] = default_write_batch_expire;
1384
1385        ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC];
1386        ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10;
1387        if (ad->write_batch_count < 2)
1388                ad->write_batch_count = 2;
1389
1390        return ad;
1391}
1392
1393/*
1394 * sysfs parts below
1395 */
1396
1397static ssize_t
1398as_var_show(unsigned int var, char *page)
1399{
1400        return sprintf(page, "%d\n", var);
1401}
1402
1403static ssize_t
1404as_var_store(unsigned long *var, const char *page, size_t count)
1405{
1406        char *p = (char *) page;
1407
1408        *var = simple_strtoul(p, &p, 10);
1409        return count;
1410}
1411
1412static ssize_t est_time_show(elevator_t *e, char *page)
1413{
1414        struct as_data *ad = e->elevator_data;
1415        int pos = 0;
1416
1417        pos += sprintf(page+pos, "%lu %% exit probability\n",
1418                                100*ad->exit_prob/256);
1419        pos += sprintf(page+pos, "%lu %% probability of exiting without a "
1420                                "cooperating process submitting IO\n",
1421                                100*ad->exit_no_coop/256);
1422        pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean);
1423        pos += sprintf(page+pos, "%llu sectors new seek distance\n",
1424                                (unsigned long long)ad->new_seek_mean);
1425
1426        return pos;
1427}
1428
1429#define SHOW_FUNCTION(__FUNC, __VAR)                            \
1430static ssize_t __FUNC(elevator_t *e, char *page)                \
1431{                                                               \
1432        struct as_data *ad = e->elevator_data;                  \
1433        return as_var_show(jiffies_to_msecs((__VAR)), (page));  \
1434}
1435SHOW_FUNCTION(as_read_expire_show, ad->fifo_expire[REQ_SYNC]);
1436SHOW_FUNCTION(as_write_expire_show, ad->fifo_expire[REQ_ASYNC]);
1437SHOW_FUNCTION(as_antic_expire_show, ad->antic_expire);
1438SHOW_FUNCTION(as_read_batch_expire_show, ad->batch_expire[REQ_SYNC]);
1439SHOW_FUNCTION(as_write_batch_expire_show, ad->batch_expire[REQ_ASYNC]);
1440#undef SHOW_FUNCTION
1441
1442#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX)                         \
1443static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
1444{                                                                       \
1445        struct as_data *ad = e->elevator_data;                          \
1446        int ret = as_var_store(__PTR, (page), count);                   \
1447        if (*(__PTR) < (MIN))                                           \
1448                *(__PTR) = (MIN);                                       \
1449        else if (*(__PTR) > (MAX))                                      \
1450                *(__PTR) = (MAX);                                       \
1451        *(__PTR) = msecs_to_jiffies(*(__PTR));                          \
1452        return ret;                                                     \
1453}
1454STORE_FUNCTION(as_read_expire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX);
1455STORE_FUNCTION(as_write_expire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX);
1456STORE_FUNCTION(as_antic_expire_store, &ad->antic_expire, 0, INT_MAX);
1457STORE_FUNCTION(as_read_batch_expire_store,
1458                        &ad->batch_expire[REQ_SYNC], 0, INT_MAX);
1459STORE_FUNCTION(as_write_batch_expire_store,
1460                        &ad->batch_expire[REQ_ASYNC], 0, INT_MAX);
1461#undef STORE_FUNCTION
1462
1463#define AS_ATTR(name) \
1464        __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store)
1465
1466static struct elv_fs_entry as_attrs[] = {
1467        __ATTR_RO(est_time),
1468        AS_ATTR(read_expire),
1469        AS_ATTR(write_expire),
1470        AS_ATTR(antic_expire),
1471        AS_ATTR(read_batch_expire),
1472        AS_ATTR(write_batch_expire),
1473        __ATTR_NULL
1474};
1475
1476static struct elevator_type iosched_as = {
1477        .ops = {
1478                .elevator_merge_fn =            as_merge,
1479                .elevator_merged_fn =           as_merged_request,
1480                .elevator_merge_req_fn =        as_merged_requests,
1481                .elevator_dispatch_fn =         as_dispatch_request,
1482                .elevator_add_req_fn =          as_add_request,
1483                .elevator_activate_req_fn =     as_activate_request,
1484                .elevator_deactivate_req_fn =   as_deactivate_request,
1485                .elevator_queue_empty_fn =      as_queue_empty,
1486                .elevator_completed_req_fn =    as_completed_request,
1487                .elevator_former_req_fn =       elv_rb_former_request,
1488                .elevator_latter_req_fn =       elv_rb_latter_request,
1489                .elevator_may_queue_fn =        as_may_queue,
1490                .elevator_init_fn =             as_init_queue,
1491                .elevator_exit_fn =             as_exit_queue,
1492                .trim =                         as_trim,
1493        },
1494
1495        .elevator_attrs = as_attrs,
1496        .elevator_name = "anticipatory",
1497        .elevator_owner = THIS_MODULE,
1498};
1499
1500static int __init as_init(void)
1501{
1502        elv_register(&iosched_as);
1503
1504        return 0;
1505}
1506
1507static void __exit as_exit(void)
1508{
1509        DECLARE_COMPLETION_ONSTACK(all_gone);
1510        elv_unregister(&iosched_as);
1511        ioc_gone = &all_gone;
1512        /* ioc_gone's update must be visible before reading ioc_count */
1513        smp_wmb();
1514        if (elv_ioc_count_read(ioc_count))
1515                wait_for_completion(&all_gone);
1516        synchronize_rcu();
1517}
1518
1519module_init(as_init);
1520module_exit(as_exit);
1521
1522MODULE_AUTHOR("Nick Piggin");
1523MODULE_LICENSE("GPL");
1524MODULE_DESCRIPTION("anticipatory IO scheduler");
1525