git.oblomov.eu Git - linux-2.6/blob - block/cfq-iosched.c

   1 /*
   2  *  CFQ, or complete fairness queueing, disk scheduler.
   3  *
   4  *  Based on ideas from a previously unfinished io
   5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   6  *
   7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
   8  */
   9 #include <linux/module.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/elevator.h>
  12 #include <linux/hash.h>
  13 #include <linux/rbtree.h>
  14 #include <linux/ioprio.h>
  15
  16 /*
  17  * tunables
  18  */
  19 static const int cfq_quantum = 4;               /* max queue in one round of service */
  20 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  21 static const int cfq_back_max = 16 * 1024;      /* maximum backwards seek, in KiB */
  22 static const int cfq_back_penalty = 2;          /* penalty of a backwards seek */
  23
  24 static const int cfq_slice_sync = HZ / 10;
  25 static int cfq_slice_async = HZ / 25;
  26 static const int cfq_slice_async_rq = 2;
  27 static int cfq_slice_idle = HZ / 125;
  28
  29 #define CFQ_IDLE_GRACE          (HZ / 10)
  30 #define CFQ_SLICE_SCALE         (5)
  31
  32 #define CFQ_KEY_ASYNC           (0)
  33
  34 /*
  35  * for the hash of cfqq inside the cfqd
  36  */
  37 #define CFQ_QHASH_SHIFT         6
  38 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
  39 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
  40
  41 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
  42
  43 #define RQ_CIC(rq)              ((struct cfq_io_context*)(rq)->elevator_private)
  44 #define RQ_CFQQ(rq)             ((rq)->elevator_private2)
  45
  46 static kmem_cache_t *cfq_pool;
  47 static kmem_cache_t *cfq_ioc_pool;
  48
  49 static DEFINE_PER_CPU(unsigned long, ioc_count);
  50 static struct completion *ioc_gone;
  51
  52 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  53 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  54 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  55
  56 #define ASYNC                   (0)
  57 #define SYNC                    (1)
  58
  59 #define cfq_cfqq_dispatched(cfqq)       \
  60         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
  61
  62 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
  63
  64 #define cfq_cfqq_sync(cfqq)             \
  65         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
  66
  67 #define sample_valid(samples)   ((samples) > 80)
  68
  69 /*
  70  * Per block device queue structure
  71  */
  72 struct cfq_data {
  73         request_queue_t *queue;
  74
  75         /*
  76          * rr list of queues with requests and the count of them
  77          */
  78         struct list_head rr_list[CFQ_PRIO_LISTS];
  79         struct list_head busy_rr;
  80         struct list_head cur_rr;
  81         struct list_head idle_rr;
  82         unsigned int busy_queues;
  83
  84         /*
  85          * cfqq lookup hash
  86          */
  87         struct hlist_head *cfq_hash;
  88
  89         int rq_in_driver;
  90         int hw_tag;
  91
  92         /*
  93          * idle window management
  94          */
  95         struct timer_list idle_slice_timer;
  96         struct work_struct unplug_work;
  97
  98         struct cfq_queue *active_queue;
  99         struct cfq_io_context *active_cic;
 100         int cur_prio, cur_end_prio;
 101         unsigned int dispatch_slice;
 102
 103         struct timer_list idle_class_timer;
 104
 105         sector_t last_sector;
 106         unsigned long last_end_request;
 107
 108         /*
 109          * tunables, see top of file
 110          */
 111         unsigned int cfq_quantum;
 112         unsigned int cfq_fifo_expire[2];
 113         unsigned int cfq_back_penalty;
 114         unsigned int cfq_back_max;
 115         unsigned int cfq_slice[2];
 116         unsigned int cfq_slice_async_rq;
 117         unsigned int cfq_slice_idle;
 118
 119         struct list_head cic_list;
 120 };
 121
 122 /*
 123  * Per process-grouping structure
 124  */
 125 struct cfq_queue {
 126         /* reference count */
 127         atomic_t ref;
 128         /* parent cfq_data */
 129         struct cfq_data *cfqd;
 130         /* cfqq lookup hash */
 131         struct hlist_node cfq_hash;
 132         /* hash key */
 133         unsigned int key;
 134         /* member of the rr/busy/cur/idle cfqd list */
 135         struct list_head cfq_list;
 136         /* sorted list of pending requests */
 137         struct rb_root sort_list;
 138         /* if fifo isn't expired, next request to serve */
 139         struct request *next_rq;
 140         /* requests queued in sort_list */
 141         int queued[2];
 142         /* currently allocated requests */
 143         int allocated[2];
 144         /* pending metadata requests */
 145         int meta_pending;
 146         /* fifo list of requests in sort_list */
 147         struct list_head fifo;
 148
 149         unsigned long slice_start;
 150         unsigned long slice_end;
 151         unsigned long slice_left;
 152
 153         /* number of requests that are on the dispatch list */
 154         int on_dispatch[2];
 155
 156         /* io prio of this group */
 157         unsigned short ioprio, org_ioprio;
 158         unsigned short ioprio_class, org_ioprio_class;
 159
 160         /* various state flags, see below */
 161         unsigned int flags;
 162 };
 163
 164 enum cfqq_state_flags {
 165         CFQ_CFQQ_FLAG_on_rr = 0,
 166         CFQ_CFQQ_FLAG_wait_request,
 167         CFQ_CFQQ_FLAG_must_alloc,
 168         CFQ_CFQQ_FLAG_must_alloc_slice,
 169         CFQ_CFQQ_FLAG_must_dispatch,
 170         CFQ_CFQQ_FLAG_fifo_expire,
 171         CFQ_CFQQ_FLAG_idle_window,
 172         CFQ_CFQQ_FLAG_prio_changed,
 173         CFQ_CFQQ_FLAG_queue_new,
 174 };
 175
 176 #define CFQ_CFQQ_FNS(name)                                              \
 177 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 178 {                                                                       \
 179         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 180 }                                                                       \
 181 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 182 {                                                                       \
 183         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 184 }                                                                       \
 185 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 186 {                                                                       \
 187         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 188 }
 189
 190 CFQ_CFQQ_FNS(on_rr);
 191 CFQ_CFQQ_FNS(wait_request);
 192 CFQ_CFQQ_FNS(must_alloc);
 193 CFQ_CFQQ_FNS(must_alloc_slice);
 194 CFQ_CFQQ_FNS(must_dispatch);
 195 CFQ_CFQQ_FNS(fifo_expire);
 196 CFQ_CFQQ_FNS(idle_window);
 197 CFQ_CFQQ_FNS(prio_changed);
 198 CFQ_CFQQ_FNS(queue_new);
 199 #undef CFQ_CFQQ_FNS
 200
 201 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 202 static void cfq_dispatch_insert(request_queue_t *, struct request *);
 203 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
 204
 205 /*
 206  * scheduler run of queue, if there are requests pending and no one in the
 207  * driver that will restart queueing
 208  */
 209 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 210 {
 211         if (cfqd->busy_queues)
 212                 kblockd_schedule_work(&cfqd->unplug_work);
 213 }
 214
 215 static int cfq_queue_empty(request_queue_t *q)
 216 {
 217         struct cfq_data *cfqd = q->elevator->elevator_data;
 218
 219         return !cfqd->busy_queues;
 220 }
 221
 222 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
 223 {
 224         if (rw == READ || rw == WRITE_SYNC)
 225                 return task->pid;
 226
 227         return CFQ_KEY_ASYNC;
 228 }
 229
 230 /*
 231  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 232  * We choose the request that is closest to the head right now. Distance
 233  * behind the head is penalized and only allowed to a certain extent.
 234  */
 235 static struct request *
 236 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
 237 {
 238         sector_t last, s1, s2, d1 = 0, d2 = 0;
 239         unsigned long back_max;
 240 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 241 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 242         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 243
 244         if (rq1 == NULL || rq1 == rq2)
 245                 return rq2;
 246         if (rq2 == NULL)
 247                 return rq1;
 248
 249         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 250                 return rq1;
 251         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 252                 return rq2;
 253         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
 254                 return rq1;
 255         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
 256                 return rq2;
 257
 258         s1 = rq1->sector;
 259         s2 = rq2->sector;
 260
 261         last = cfqd->last_sector;
 262
 263         /*
 264          * by definition, 1KiB is 2 sectors
 265          */
 266         back_max = cfqd->cfq_back_max * 2;
 267
 268         /*
 269          * Strict one way elevator _except_ in the case where we allow
 270          * short backward seeks which are biased as twice the cost of a
 271          * similar forward seek.
 272          */
 273         if (s1 >= last)
 274                 d1 = s1 - last;
 275         else if (s1 + back_max >= last)
 276                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 277         else
 278                 wrap |= CFQ_RQ1_WRAP;
 279
 280         if (s2 >= last)
 281                 d2 = s2 - last;
 282         else if (s2 + back_max >= last)
 283                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 284         else
 285                 wrap |= CFQ_RQ2_WRAP;
 286
 287         /* Found required data */
 288
 289         /*
 290          * By doing switch() on the bit mask "wrap" we avoid having to
 291          * check two variables for all permutations: --> faster!
 292          */
 293         switch (wrap) {
 294         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 295                 if (d1 < d2)
 296                         return rq1;
 297                 else if (d2 < d1)
 298                         return rq2;
 299                 else {
 300                         if (s1 >= s2)
 301                                 return rq1;
 302                         else
 303                                 return rq2;
 304                 }
 305
 306         case CFQ_RQ2_WRAP:
 307                 return rq1;
 308         case CFQ_RQ1_WRAP:
 309                 return rq2;
 310         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
 311         default:
 312                 /*
 313                  * Since both rqs are wrapped,
 314                  * start with the one that's further behind head
 315                  * (--> only *one* back seek required),
 316                  * since back seek takes more time than forward.
 317                  */
 318                 if (s1 <= s2)
 319                         return rq1;
 320                 else
 321                         return rq2;
 322         }
 323 }
 324
 325 /*
 326  * would be nice to take fifo expire time into account as well
 327  */
 328 static struct request *
 329 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 330                   struct request *last)
 331 {
 332         struct rb_node *rbnext = rb_next(&last->rb_node);
 333         struct rb_node *rbprev = rb_prev(&last->rb_node);
 334         struct request *next = NULL, *prev = NULL;
 335
 336         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
 337
 338         if (rbprev)
 339                 prev = rb_entry_rq(rbprev);
 340
 341         if (rbnext)
 342                 next = rb_entry_rq(rbnext);
 343         else {
 344                 rbnext = rb_first(&cfqq->sort_list);
 345                 if (rbnext && rbnext != &last->rb_node)
 346                         next = rb_entry_rq(rbnext);
 347         }
 348
 349         return cfq_choose_req(cfqd, next, prev);
 350 }
 351
 352 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 353 {
 354         struct cfq_data *cfqd = cfqq->cfqd;
 355         struct list_head *list;
 356
 357         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 358
 359         list_del(&cfqq->cfq_list);
 360
 361         if (cfq_class_rt(cfqq))
 362                 list = &cfqd->cur_rr;
 363         else if (cfq_class_idle(cfqq))
 364                 list = &cfqd->idle_rr;
 365         else {
 366                 /*
 367                  * if cfqq has requests in flight, don't allow it to be
 368                  * found in cfq_set_active_queue before it has finished them.
 369                  * this is done to increase fairness between a process that
 370                  * has lots of io pending vs one that only generates one
 371                  * sporadically or synchronously
 372                  */
 373                 if (cfq_cfqq_dispatched(cfqq))
 374                         list = &cfqd->busy_rr;
 375                 else
 376                         list = &cfqd->rr_list[cfqq->ioprio];
 377         }
 378
 379         /*
 380          * If this queue was preempted or is new (never been serviced), let
 381          * it be added first for fairness but beind other new queues.
 382          * Otherwise, just add to the back  of the list.
 383          */
 384         if (preempted || cfq_cfqq_queue_new(cfqq)) {
 385                 struct list_head *n = list;
 386                 struct cfq_queue *__cfqq;
 387
 388                 while (n->next != list) {
 389                         __cfqq = list_entry_cfqq(n->next);
 390                         if (!cfq_cfqq_queue_new(__cfqq))
 391                                 break;
 392
 393                         n = n->next;
 394                 }
 395
 396                 list = n;
 397         }
 398
 399         list_add_tail(&cfqq->cfq_list, list);
 400 }
 401
 402 /*
 403  * add to busy list of queues for service, trying to be fair in ordering
 404  * the pending list according to last request service
 405  */
 406 static inline void
 407 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 408 {
 409         BUG_ON(cfq_cfqq_on_rr(cfqq));
 410         cfq_mark_cfqq_on_rr(cfqq);
 411         cfqd->busy_queues++;
 412
 413         cfq_resort_rr_list(cfqq, 0);
 414 }
 415
 416 static inline void
 417 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 418 {
 419         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 420         cfq_clear_cfqq_on_rr(cfqq);
 421         list_del_init(&cfqq->cfq_list);
 422
 423         BUG_ON(!cfqd->busy_queues);
 424         cfqd->busy_queues--;
 425 }
 426
 427 /*
 428  * rb tree support functions
 429  */
 430 static inline void cfq_del_rq_rb(struct request *rq)
 431 {
 432         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 433         struct cfq_data *cfqd = cfqq->cfqd;
 434         const int sync = rq_is_sync(rq);
 435
 436         BUG_ON(!cfqq->queued[sync]);
 437         cfqq->queued[sync]--;
 438
 439         elv_rb_del(&cfqq->sort_list, rq);
 440
 441         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
 442                 cfq_del_cfqq_rr(cfqd, cfqq);
 443 }
 444
 445 static void cfq_add_rq_rb(struct request *rq)
 446 {
 447         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 448         struct cfq_data *cfqd = cfqq->cfqd;
 449         struct request *__alias;
 450
 451         cfqq->queued[rq_is_sync(rq)]++;
 452
 453         /*
 454          * looks a little odd, but the first insert might return an alias.
 455          * if that happens, put the alias on the dispatch list
 456          */
 457         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
 458                 cfq_dispatch_insert(cfqd->queue, __alias);
 459 }
 460
 461 static inline void
 462 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
 463 {
 464         elv_rb_del(&cfqq->sort_list, rq);
 465         cfqq->queued[rq_is_sync(rq)]--;
 466         cfq_add_rq_rb(rq);
 467 }
 468
 469 static struct request *
 470 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 471 {
 472         struct task_struct *tsk = current;
 473         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
 474         struct cfq_queue *cfqq;
 475
 476         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
 477         if (cfqq) {
 478                 sector_t sector = bio->bi_sector + bio_sectors(bio);
 479
 480                 return elv_rb_find(&cfqq->sort_list, sector);
 481         }
 482
 483         return NULL;
 484 }
 485
 486 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 487 {
 488         struct cfq_data *cfqd = q->elevator->elevator_data;
 489
 490         cfqd->rq_in_driver++;
 491
 492         /*
 493          * If the depth is larger 1, it really could be queueing. But lets
 494          * make the mark a little higher - idling could still be good for
 495          * low queueing, and a low queueing number could also just indicate
 496          * a SCSI mid layer like behaviour where limit+1 is often seen.
 497          */
 498         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
 499                 cfqd->hw_tag = 1;
 500 }
 501
 502 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 503 {
 504         struct cfq_data *cfqd = q->elevator->elevator_data;
 505
 506         WARN_ON(!cfqd->rq_in_driver);
 507         cfqd->rq_in_driver--;
 508 }
 509
 510 static void cfq_remove_request(struct request *rq)
 511 {
 512         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 513
 514         if (cfqq->next_rq == rq)
 515                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
 516
 517         list_del_init(&rq->queuelist);
 518         cfq_del_rq_rb(rq);
 519
 520         if (rq_is_meta(rq)) {
 521                 WARN_ON(!cfqq->meta_pending);
 522                 cfqq->meta_pending--;
 523         }
 524 }
 525
 526 static int
 527 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 528 {
 529         struct cfq_data *cfqd = q->elevator->elevator_data;
 530         struct request *__rq;
 531
 532         __rq = cfq_find_rq_fmerge(cfqd, bio);
 533         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 534                 *req = __rq;
 535                 return ELEVATOR_FRONT_MERGE;
 536         }
 537
 538         return ELEVATOR_NO_MERGE;
 539 }
 540
 541 static void cfq_merged_request(request_queue_t *q, struct request *req,
 542                                int type)
 543 {
 544         if (type == ELEVATOR_FRONT_MERGE) {
 545                 struct cfq_queue *cfqq = RQ_CFQQ(req);
 546
 547                 cfq_reposition_rq_rb(cfqq, req);
 548         }
 549 }
 550
 551 static void
 552 cfq_merged_requests(request_queue_t *q, struct request *rq,
 553                     struct request *next)
 554 {
 555         /*
 556          * reposition in fifo if next is older than rq
 557          */
 558         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 559             time_before(next->start_time, rq->start_time))
 560                 list_move(&rq->queuelist, &next->queuelist);
 561
 562         cfq_remove_request(next);
 563 }
 564
 565 static inline void
 566 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 567 {
 568         if (cfqq) {
 569                 /*
 570                  * stop potential idle class queues waiting service
 571                  */
 572                 del_timer(&cfqd->idle_class_timer);
 573
 574                 cfqq->slice_start = jiffies;
 575                 cfqq->slice_end = 0;
 576                 cfqq->slice_left = 0;
 577                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 578                 cfq_clear_cfqq_fifo_expire(cfqq);
 579         }
 580
 581         cfqd->active_queue = cfqq;
 582 }
 583
 584 /*
 585  * current cfqq expired its slice (or was too idle), select new one
 586  */
 587 static void
 588 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 589                     int preempted)
 590 {
 591         unsigned long now = jiffies;
 592
 593         if (cfq_cfqq_wait_request(cfqq))
 594                 del_timer(&cfqd->idle_slice_timer);
 595
 596         if (!preempted && !cfq_cfqq_dispatched(cfqq))
 597                 cfq_schedule_dispatch(cfqd);
 598
 599         cfq_clear_cfqq_must_dispatch(cfqq);
 600         cfq_clear_cfqq_wait_request(cfqq);
 601         cfq_clear_cfqq_queue_new(cfqq);
 602
 603         /*
 604          * store what was left of this slice, if the queue idled out
 605          * or was preempted
 606          */
 607         if (time_after(cfqq->slice_end, now))
 608                 cfqq->slice_left = cfqq->slice_end - now;
 609         else
 610                 cfqq->slice_left = 0;
 611
 612         if (cfq_cfqq_on_rr(cfqq))
 613                 cfq_resort_rr_list(cfqq, preempted);
 614
 615         if (cfqq == cfqd->active_queue)
 616                 cfqd->active_queue = NULL;
 617
 618         if (cfqd->active_cic) {
 619                 put_io_context(cfqd->active_cic->ioc);
 620                 cfqd->active_cic = NULL;
 621         }
 622
 623         cfqd->dispatch_slice = 0;
 624 }
 625
 626 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 627 {
 628         struct cfq_queue *cfqq = cfqd->active_queue;
 629
 630         if (cfqq)
 631                 __cfq_slice_expired(cfqd, cfqq, preempted);
 632 }
 633
 634 /*
 635  * 0
 636  * 0,1
 637  * 0,1,2
 638  * 0,1,2,3
 639  * 0,1,2,3,4
 640  * 0,1,2,3,4,5
 641  * 0,1,2,3,4,5,6
 642  * 0,1,2,3,4,5,6,7
 643  */
 644 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 645 {
 646         int prio, wrap;
 647
 648         prio = -1;
 649         wrap = 0;
 650         do {
 651                 int p;
 652
 653                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 654                         if (!list_empty(&cfqd->rr_list[p])) {
 655                                 prio = p;
 656                                 break;
 657                         }
 658                 }
 659
 660                 if (prio != -1)
 661                         break;
 662                 cfqd->cur_prio = 0;
 663                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 664                         cfqd->cur_end_prio = 0;
 665                         if (wrap)
 666                                 break;
 667                         wrap = 1;
 668                 }
 669         } while (1);
 670
 671         if (unlikely(prio == -1))
 672                 return -1;
 673
 674         BUG_ON(prio >= CFQ_PRIO_LISTS);
 675
 676         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 677
 678         cfqd->cur_prio = prio + 1;
 679         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 680                 cfqd->cur_end_prio = cfqd->cur_prio;
 681                 cfqd->cur_prio = 0;
 682         }
 683         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 684                 cfqd->cur_prio = 0;
 685                 cfqd->cur_end_prio = 0;
 686         }
 687
 688         return prio;
 689 }
 690
 691 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 692 {
 693         struct cfq_queue *cfqq = NULL;
 694
 695         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
 696                 /*
 697                  * if current list is non-empty, grab first entry. if it is
 698                  * empty, get next prio level and grab first entry then if any
 699                  * are spliced
 700                  */
 701                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 702         } else if (!list_empty(&cfqd->busy_rr)) {
 703                 /*
 704                  * If no new queues are available, check if the busy list has
 705                  * some before falling back to idle io.
 706                  */
 707                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
 708         } else if (!list_empty(&cfqd->idle_rr)) {
 709                 /*
 710                  * if we have idle queues and no rt or be queues had pending
 711                  * requests, either allow immediate service if the grace period
 712                  * has passed or arm the idle grace timer
 713                  */
 714                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 715
 716                 if (time_after_eq(jiffies, end))
 717                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 718                 else
 719                         mod_timer(&cfqd->idle_class_timer, end);
 720         }
 721
 722         __cfq_set_active_queue(cfqd, cfqq);
 723         return cfqq;
 724 }
 725
 726 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
 727
 728 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 729
 730 {
 731         struct cfq_io_context *cic;
 732         unsigned long sl;
 733
 734         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
 735         WARN_ON(cfqq != cfqd->active_queue);
 736
 737         /*
 738          * idle is disabled, either manually or by past process history
 739          */
 740         if (!cfqd->cfq_slice_idle)
 741                 return 0;
 742         if (!cfq_cfqq_idle_window(cfqq))
 743                 return 0;
 744         /*
 745          * task has exited, don't wait
 746          */
 747         cic = cfqd->active_cic;
 748         if (!cic || !cic->ioc->task)
 749                 return 0;
 750
 751         cfq_mark_cfqq_must_dispatch(cfqq);
 752         cfq_mark_cfqq_wait_request(cfqq);
 753
 754         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 755
 756         /*
 757          * we don't want to idle for seeks, but we do want to allow
 758          * fair distribution of slice time for a process doing back-to-back
 759          * seeks. so allow a little bit of time for him to submit a new rq
 760          */
 761         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
 762                 sl = min(sl, msecs_to_jiffies(2));
 763
 764         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 765         return 1;
 766 }
 767
 768 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
 769 {
 770         struct cfq_data *cfqd = q->elevator->elevator_data;
 771         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 772
 773         cfq_remove_request(rq);
 774         cfqq->on_dispatch[rq_is_sync(rq)]++;
 775         elv_dispatch_sort(q, rq);
 776
 777         rq = list_entry(q->queue_head.prev, struct request, queuelist);
 778         cfqd->last_sector = rq->sector + rq->nr_sectors;
 779 }
 780
 781 /*
 782  * return expired entry, or NULL to just start from scratch in rbtree
 783  */
 784 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
 785 {
 786         struct cfq_data *cfqd = cfqq->cfqd;
 787         struct request *rq;
 788         int fifo;
 789
 790         if (cfq_cfqq_fifo_expire(cfqq))
 791                 return NULL;
 792         if (list_empty(&cfqq->fifo))
 793                 return NULL;
 794
 795         fifo = cfq_cfqq_class_sync(cfqq);
 796         rq = rq_entry_fifo(cfqq->fifo.next);
 797
 798         if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
 799                 cfq_mark_cfqq_fifo_expire(cfqq);
 800                 return rq;
 801         }
 802
 803         return NULL;
 804 }
 805
 806 /*
 807  * Scale schedule slice based on io priority. Use the sync time slice only
 808  * if a queue is marked sync and has sync io queued. A sync queue with async
 809  * io only, should not get full sync slice length.
 810  */
 811 static inline int
 812 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 813 {
 814         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 815
 816         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 817
 818         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
 819 }
 820
 821 static inline void
 822 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 823 {
 824         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 825 }
 826
 827 static inline int
 828 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 829 {
 830         const int base_rq = cfqd->cfq_slice_async_rq;
 831
 832         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 833
 834         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
 835 }
 836
 837 /*
 838  * get next queue for service
 839  */
 840 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
 841 {
 842         unsigned long now = jiffies;
 843         struct cfq_queue *cfqq;
 844
 845         cfqq = cfqd->active_queue;
 846         if (!cfqq)
 847                 goto new_queue;
 848
 849         /*
 850          * slice has expired
 851          */
 852         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
 853                 goto expire;
 854
 855         /*
 856          * if queue has requests, dispatch one. if not, check if
 857          * enough slice is left to wait for one
 858          */
 859         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
 860                 goto keep_queue;
 861         else if (cfq_cfqq_dispatched(cfqq)) {
 862                 cfqq = NULL;
 863                 goto keep_queue;
 864         } else if (cfq_cfqq_class_sync(cfqq)) {
 865                 if (cfq_arm_slice_timer(cfqd, cfqq))
 866                         return NULL;
 867         }
 868
 869 expire:
 870         cfq_slice_expired(cfqd, 0);
 871 new_queue:
 872         cfqq = cfq_set_active_queue(cfqd);
 873 keep_queue:
 874         return cfqq;
 875 }
 876
 877 static int
 878 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 879                         int max_dispatch)
 880 {
 881         int dispatched = 0;
 882
 883         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
 884
 885         do {
 886                 struct request *rq;
 887
 888                 /*
 889                  * follow expired path, else get first next available
 890                  */
 891                 if ((rq = cfq_check_fifo(cfqq)) == NULL)
 892                         rq = cfqq->next_rq;
 893
 894                 /*
 895                  * finally, insert request into driver dispatch list
 896                  */
 897                 cfq_dispatch_insert(cfqd->queue, rq);
 898
 899                 cfqd->dispatch_slice++;
 900                 dispatched++;
 901
 902                 if (!cfqd->active_cic) {
 903                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
 904                         cfqd->active_cic = RQ_CIC(rq);
 905                 }
 906
 907                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
 908                         break;
 909
 910         } while (dispatched < max_dispatch);
 911
 912         /*
 913          * if slice end isn't set yet, set it.
 914          */
 915         if (!cfqq->slice_end)
 916                 cfq_set_prio_slice(cfqd, cfqq);
 917
 918         /*
 919          * expire an async queue immediately if it has used up its slice. idle
 920          * queue always expire after 1 dispatch round.
 921          */
 922         if ((!cfq_cfqq_sync(cfqq) &&
 923             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
 924             cfq_class_idle(cfqq) ||
 925             !cfq_cfqq_idle_window(cfqq))
 926                 cfq_slice_expired(cfqd, 0);
 927
 928         return dispatched;
 929 }
 930
 931 static int
 932 cfq_forced_dispatch_cfqqs(struct list_head *list)
 933 {
 934         struct cfq_queue *cfqq, *next;
 935         int dispatched;
 936
 937         dispatched = 0;
 938         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
 939                 while (cfqq->next_rq) {
 940                         cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
 941                         dispatched++;
 942                 }
 943                 BUG_ON(!list_empty(&cfqq->fifo));
 944         }
 945
 946         return dispatched;
 947 }
 948
 949 static int
 950 cfq_forced_dispatch(struct cfq_data *cfqd)
 951 {
 952         int i, dispatched = 0;
 953
 954         for (i = 0; i < CFQ_PRIO_LISTS; i++)
 955                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
 956
 957         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
 958         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
 959         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
 960
 961         cfq_slice_expired(cfqd, 0);
 962
 963         BUG_ON(cfqd->busy_queues);
 964
 965         return dispatched;
 966 }
 967
 968 static int
 969 cfq_dispatch_requests(request_queue_t *q, int force)
 970 {
 971         struct cfq_data *cfqd = q->elevator->elevator_data;
 972         struct cfq_queue *cfqq, *prev_cfqq;
 973         int dispatched;
 974
 975         if (!cfqd->busy_queues)
 976                 return 0;
 977
 978         if (unlikely(force))
 979                 return cfq_forced_dispatch(cfqd);
 980
 981         dispatched = 0;
 982         prev_cfqq = NULL;
 983         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
 984                 int max_dispatch;
 985
 986                 /*
 987                  * Don't repeat dispatch from the previous queue.
 988                  */
 989                 if (prev_cfqq == cfqq)
 990                         break;
 991
 992                 cfq_clear_cfqq_must_dispatch(cfqq);
 993                 cfq_clear_cfqq_wait_request(cfqq);
 994                 del_timer(&cfqd->idle_slice_timer);
 995
 996                 max_dispatch = cfqd->cfq_quantum;
 997                 if (cfq_class_idle(cfqq))
 998                         max_dispatch = 1;
 999
1000                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1001
1002                 /*
1003                  * If the dispatch cfqq has idling enabled and is still
1004                  * the active queue, break out.
1005                  */
1006                 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1007                         break;
1008
1009                 prev_cfqq = cfqq;
1010         }
1011
1012         return dispatched;
1013 }
1014
1015 /*
1016  * task holds one reference to the queue, dropped when task exits. each rq
1017  * in-flight on this queue also holds a reference, dropped when rq is freed.
1018  *
1019  * queue lock must be held here.
1020  */
1021 static void cfq_put_queue(struct cfq_queue *cfqq)
1022 {
1023         struct cfq_data *cfqd = cfqq->cfqd;
1024
1025         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1026
1027         if (!atomic_dec_and_test(&cfqq->ref))
1028                 return;
1029
1030         BUG_ON(rb_first(&cfqq->sort_list));
1031         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1032         BUG_ON(cfq_cfqq_on_rr(cfqq));
1033
1034         if (unlikely(cfqd->active_queue == cfqq))
1035                 __cfq_slice_expired(cfqd, cfqq, 0);
1036
1037         /*
1038          * it's on the empty list and still hashed
1039          */
1040         list_del(&cfqq->cfq_list);
1041         hlist_del(&cfqq->cfq_hash);
1042         kmem_cache_free(cfq_pool, cfqq);
1043 }
1044
1045 static struct cfq_queue *
1046 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1047                     const int hashval)
1048 {
1049         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1050         struct hlist_node *entry;
1051         struct cfq_queue *__cfqq;
1052
1053         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1054                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1055
1056                 if (__cfqq->key == key && (__p == prio || !prio))
1057                         return __cfqq;
1058         }
1059
1060         return NULL;
1061 }
1062
1063 static struct cfq_queue *
1064 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1065 {
1066         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1067 }
1068
1069 static void cfq_free_io_context(struct io_context *ioc)
1070 {
1071         struct cfq_io_context *__cic;
1072         struct rb_node *n;
1073         int freed = 0;
1074
1075         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1076                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1077                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1078                 kmem_cache_free(cfq_ioc_pool, __cic);
1079                 freed++;
1080         }
1081
1082         elv_ioc_count_mod(ioc_count, -freed);
1083
1084         if (ioc_gone && !elv_ioc_count_read(ioc_count))
1085                 complete(ioc_gone);
1086 }
1087
1088 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1089 {
1090         if (unlikely(cfqq == cfqd->active_queue))
1091                 __cfq_slice_expired(cfqd, cfqq, 0);
1092
1093         cfq_put_queue(cfqq);
1094 }
1095
1096 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1097                                          struct cfq_io_context *cic)
1098 {
1099         list_del_init(&cic->queue_list);
1100         smp_wmb();
1101         cic->key = NULL;
1102
1103         if (cic->cfqq[ASYNC]) {
1104                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1105                 cic->cfqq[ASYNC] = NULL;
1106         }
1107
1108         if (cic->cfqq[SYNC]) {
1109                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1110                 cic->cfqq[SYNC] = NULL;
1111         }
1112 }
1113
1114
1115 /*
1116  * Called with interrupts disabled
1117  */
1118 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1119 {
1120         struct cfq_data *cfqd = cic->key;
1121
1122         if (cfqd) {
1123                 request_queue_t *q = cfqd->queue;
1124
1125                 spin_lock_irq(q->queue_lock);
1126                 __cfq_exit_single_io_context(cfqd, cic);
1127                 spin_unlock_irq(q->queue_lock);
1128         }
1129 }
1130
1131 static void cfq_exit_io_context(struct io_context *ioc)
1132 {
1133         struct cfq_io_context *__cic;
1134         struct rb_node *n;
1135
1136         /*
1137          * put the reference this task is holding to the various queues
1138          */
1139
1140         n = rb_first(&ioc->cic_root);
1141         while (n != NULL) {
1142                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1143
1144                 cfq_exit_single_io_context(__cic);
1145                 n = rb_next(n);
1146         }
1147 }
1148
1149 static struct cfq_io_context *
1150 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1151 {
1152         struct cfq_io_context *cic;
1153
1154         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
1155         if (cic) {
1156                 memset(cic, 0, sizeof(*cic));
1157                 cic->last_end_request = jiffies;
1158                 INIT_LIST_HEAD(&cic->queue_list);
1159                 cic->dtor = cfq_free_io_context;
1160                 cic->exit = cfq_exit_io_context;
1161                 elv_ioc_count_inc(ioc_count);
1162         }
1163
1164         return cic;
1165 }
1166
1167 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1168 {
1169         struct task_struct *tsk = current;
1170         int ioprio_class;
1171
1172         if (!cfq_cfqq_prio_changed(cfqq))
1173                 return;
1174
1175         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1176         switch (ioprio_class) {
1177                 default:
1178                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1179                 case IOPRIO_CLASS_NONE:
1180                         /*
1181                          * no prio set, place us in the middle of the BE classes
1182                          */
1183                         cfqq->ioprio = task_nice_ioprio(tsk);
1184                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1185                         break;
1186                 case IOPRIO_CLASS_RT:
1187                         cfqq->ioprio = task_ioprio(tsk);
1188                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1189                         break;
1190                 case IOPRIO_CLASS_BE:
1191                         cfqq->ioprio = task_ioprio(tsk);
1192                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1193                         break;
1194                 case IOPRIO_CLASS_IDLE:
1195                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1196                         cfqq->ioprio = 7;
1197                         cfq_clear_cfqq_idle_window(cfqq);
1198                         break;
1199         }
1200
1201         /*
1202          * keep track of original prio settings in case we have to temporarily
1203          * elevate the priority of this queue
1204          */
1205         cfqq->org_ioprio = cfqq->ioprio;
1206         cfqq->org_ioprio_class = cfqq->ioprio_class;
1207
1208         if (cfq_cfqq_on_rr(cfqq))
1209                 cfq_resort_rr_list(cfqq, 0);
1210
1211         cfq_clear_cfqq_prio_changed(cfqq);
1212 }
1213
1214 static inline void changed_ioprio(struct cfq_io_context *cic)
1215 {
1216         struct cfq_data *cfqd = cic->key;
1217         struct cfq_queue *cfqq;
1218
1219         if (unlikely(!cfqd))
1220                 return;
1221
1222         spin_lock(cfqd->queue->queue_lock);
1223
1224         cfqq = cic->cfqq[ASYNC];
1225         if (cfqq) {
1226                 struct cfq_queue *new_cfqq;
1227                 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1228                                          GFP_ATOMIC);
1229                 if (new_cfqq) {
1230                         cic->cfqq[ASYNC] = new_cfqq;
1231                         cfq_put_queue(cfqq);
1232                 }
1233         }
1234
1235         cfqq = cic->cfqq[SYNC];
1236         if (cfqq)
1237                 cfq_mark_cfqq_prio_changed(cfqq);
1238
1239         spin_unlock(cfqd->queue->queue_lock);
1240 }
1241
1242 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1243 {
1244         struct cfq_io_context *cic;
1245         struct rb_node *n;
1246
1247         ioc->ioprio_changed = 0;
1248
1249         n = rb_first(&ioc->cic_root);
1250         while (n != NULL) {
1251                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1252
1253                 changed_ioprio(cic);
1254                 n = rb_next(n);
1255         }
1256 }
1257
1258 static struct cfq_queue *
1259 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1260               gfp_t gfp_mask)
1261 {
1262         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1263         struct cfq_queue *cfqq, *new_cfqq = NULL;
1264         unsigned short ioprio;
1265
1266 retry:
1267         ioprio = tsk->ioprio;
1268         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1269
1270         if (!cfqq) {
1271                 if (new_cfqq) {
1272                         cfqq = new_cfqq;
1273                         new_cfqq = NULL;
1274                 } else if (gfp_mask & __GFP_WAIT) {
1275                         /*
1276                          * Inform the allocator of the fact that we will
1277                          * just repeat this allocation if it fails, to allow
1278                          * the allocator to do whatever it needs to attempt to
1279                          * free memory.
1280                          */
1281                         spin_unlock_irq(cfqd->queue->queue_lock);
1282                         new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
1283                         spin_lock_irq(cfqd->queue->queue_lock);
1284                         goto retry;
1285                 } else {
1286                         cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
1287                         if (!cfqq)
1288                                 goto out;
1289                 }
1290
1291                 memset(cfqq, 0, sizeof(*cfqq));
1292
1293                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1294                 INIT_LIST_HEAD(&cfqq->cfq_list);
1295                 INIT_LIST_HEAD(&cfqq->fifo);
1296
1297                 cfqq->key = key;
1298                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1299                 atomic_set(&cfqq->ref, 0);
1300                 cfqq->cfqd = cfqd;
1301                 /*
1302                  * set ->slice_left to allow preemption for a new process
1303                  */
1304                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1305                 cfq_mark_cfqq_idle_window(cfqq);
1306                 cfq_mark_cfqq_prio_changed(cfqq);
1307                 cfq_mark_cfqq_queue_new(cfqq);
1308                 cfq_init_prio_data(cfqq);
1309         }
1310
1311         if (new_cfqq)
1312                 kmem_cache_free(cfq_pool, new_cfqq);
1313
1314         atomic_inc(&cfqq->ref);
1315 out:
1316         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1317         return cfqq;
1318 }
1319
1320 static void
1321 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1322 {
1323         WARN_ON(!list_empty(&cic->queue_list));
1324         rb_erase(&cic->rb_node, &ioc->cic_root);
1325         kmem_cache_free(cfq_ioc_pool, cic);
1326         elv_ioc_count_dec(ioc_count);
1327 }
1328
1329 static struct cfq_io_context *
1330 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1331 {
1332         struct rb_node *n;
1333         struct cfq_io_context *cic;
1334         void *k, *key = cfqd;
1335
1336 restart:
1337         n = ioc->cic_root.rb_node;
1338         while (n) {
1339                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1340                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1341                 k = cic->key;
1342                 if (unlikely(!k)) {
1343                         cfq_drop_dead_cic(ioc, cic);
1344                         goto restart;
1345                 }
1346
1347                 if (key < k)
1348                         n = n->rb_left;
1349                 else if (key > k)
1350                         n = n->rb_right;
1351                 else
1352                         return cic;
1353         }
1354
1355         return NULL;
1356 }
1357
1358 static inline void
1359 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1360              struct cfq_io_context *cic)
1361 {
1362         struct rb_node **p;
1363         struct rb_node *parent;
1364         struct cfq_io_context *__cic;
1365         void *k;
1366
1367         cic->ioc = ioc;
1368         cic->key = cfqd;
1369
1370 restart:
1371         parent = NULL;
1372         p = &ioc->cic_root.rb_node;
1373         while (*p) {
1374                 parent = *p;
1375                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1376                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1377                 k = __cic->key;
1378                 if (unlikely(!k)) {
1379                         cfq_drop_dead_cic(ioc, __cic);
1380                         goto restart;
1381                 }
1382
1383                 if (cic->key < k)
1384                         p = &(*p)->rb_left;
1385                 else if (cic->key > k)
1386                         p = &(*p)->rb_right;
1387                 else
1388                         BUG();
1389         }
1390
1391         rb_link_node(&cic->rb_node, parent, p);
1392         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1393
1394         spin_lock_irq(cfqd->queue->queue_lock);
1395         list_add(&cic->queue_list, &cfqd->cic_list);
1396         spin_unlock_irq(cfqd->queue->queue_lock);
1397 }
1398
1399 /*
1400  * Setup general io context and cfq io context. There can be several cfq
1401  * io contexts per general io context, if this process is doing io to more
1402  * than one device managed by cfq.
1403  */
1404 static struct cfq_io_context *
1405 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1406 {
1407         struct io_context *ioc = NULL;
1408         struct cfq_io_context *cic;
1409
1410         might_sleep_if(gfp_mask & __GFP_WAIT);
1411
1412         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1413         if (!ioc)
1414                 return NULL;
1415
1416         cic = cfq_cic_rb_lookup(cfqd, ioc);
1417         if (cic)
1418                 goto out;
1419
1420         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1421         if (cic == NULL)
1422                 goto err;
1423
1424         cfq_cic_link(cfqd, ioc, cic);
1425 out:
1426         smp_read_barrier_depends();
1427         if (unlikely(ioc->ioprio_changed))
1428                 cfq_ioc_set_ioprio(ioc);
1429
1430         return cic;
1431 err:
1432         put_io_context(ioc);
1433         return NULL;
1434 }
1435
1436 static void
1437 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1438 {
1439         unsigned long elapsed, ttime;
1440
1441         /*
1442          * if this context already has stuff queued, thinktime is from
1443          * last queue not last end
1444          */
1445 #if 0
1446         if (time_after(cic->last_end_request, cic->last_queue))
1447                 elapsed = jiffies - cic->last_end_request;
1448         else
1449                 elapsed = jiffies - cic->last_queue;
1450 #else
1451                 elapsed = jiffies - cic->last_end_request;
1452 #endif
1453
1454         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1455
1456         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1457         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1458         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1459 }
1460
1461 static void
1462 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1463                        struct request *rq)
1464 {
1465         sector_t sdist;
1466         u64 total;
1467
1468         if (cic->last_request_pos < rq->sector)
1469                 sdist = rq->sector - cic->last_request_pos;
1470         else
1471                 sdist = cic->last_request_pos - rq->sector;
1472
1473         /*
1474          * Don't allow the seek distance to get too large from the
1475          * odd fragment, pagein, etc
1476          */
1477         if (cic->seek_samples <= 60) /* second&third seek */
1478                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1479         else
1480                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1481
1482         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1483         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1484         total = cic->seek_total + (cic->seek_samples/2);
1485         do_div(total, cic->seek_samples);
1486         cic->seek_mean = (sector_t)total;
1487 }
1488
1489 /*
1490  * Disable idle window if the process thinks too long or seeks so much that
1491  * it doesn't matter
1492  */
1493 static void
1494 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1495                        struct cfq_io_context *cic)
1496 {
1497         int enable_idle = cfq_cfqq_idle_window(cfqq);
1498
1499         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1500             (cfqd->hw_tag && CIC_SEEKY(cic)))
1501                 enable_idle = 0;
1502         else if (sample_valid(cic->ttime_samples)) {
1503                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1504                         enable_idle = 0;
1505                 else
1506                         enable_idle = 1;
1507         }
1508
1509         if (enable_idle)
1510                 cfq_mark_cfqq_idle_window(cfqq);
1511         else
1512                 cfq_clear_cfqq_idle_window(cfqq);
1513 }
1514
1515
1516 /*
1517  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1518  * no or if we aren't sure, a 1 will cause a preempt.
1519  */
1520 static int
1521 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1522                    struct request *rq)
1523 {
1524         struct cfq_queue *cfqq = cfqd->active_queue;
1525
1526         if (cfq_class_idle(new_cfqq))
1527                 return 0;
1528
1529         if (!cfqq)
1530                 return 0;
1531
1532         if (cfq_class_idle(cfqq))
1533                 return 1;
1534         if (!cfq_cfqq_wait_request(new_cfqq))
1535                 return 0;
1536         /*
1537          * if it doesn't have slice left, forget it
1538          */
1539         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1540                 return 0;
1541         /*
1542          * if the new request is sync, but the currently running queue is
1543          * not, let the sync request have priority.
1544          */
1545         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1546                 return 1;
1547         /*
1548          * So both queues are sync. Let the new request get disk time if
1549          * it's a metadata request and the current queue is doing regular IO.
1550          */
1551         if (rq_is_meta(rq) && !cfqq->meta_pending)
1552                 return 1;
1553
1554         return 0;
1555 }
1556
1557 /*
1558  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1559  * let it have half of its nominal slice.
1560  */
1561 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1562 {
1563         cfq_slice_expired(cfqd, 1);
1564
1565         if (!cfqq->slice_left)
1566                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1567
1568         /*
1569          * Put the new queue at the front of the of the current list,
1570          * so we know that it will be selected next.
1571          */
1572         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1573         list_move(&cfqq->cfq_list, &cfqd->cur_rr);
1574
1575         cfqq->slice_end = cfqq->slice_left + jiffies;
1576 }
1577
1578 /*
1579  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1580  * something we should do about it
1581  */
1582 static void
1583 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1584                 struct request *rq)
1585 {
1586         struct cfq_io_context *cic = RQ_CIC(rq);
1587
1588         if (rq_is_meta(rq))
1589                 cfqq->meta_pending++;
1590
1591         /*
1592          * check if this request is a better next-serve candidate)) {
1593          */
1594         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1595         BUG_ON(!cfqq->next_rq);
1596
1597         /*
1598          * we never wait for an async request and we don't allow preemption
1599          * of an async request. so just return early
1600          */
1601         if (!rq_is_sync(rq)) {
1602                 /*
1603                  * sync process issued an async request, if it's waiting
1604                  * then expire it and kick rq handling.
1605                  */
1606                 if (cic == cfqd->active_cic &&
1607                     del_timer(&cfqd->idle_slice_timer)) {
1608                         cfq_slice_expired(cfqd, 0);
1609                         blk_start_queueing(cfqd->queue);
1610                 }
1611                 return;
1612         }
1613
1614         cfq_update_io_thinktime(cfqd, cic);
1615         cfq_update_io_seektime(cfqd, cic, rq);
1616         cfq_update_idle_window(cfqd, cfqq, cic);
1617
1618         cic->last_queue = jiffies;
1619         cic->last_request_pos = rq->sector + rq->nr_sectors;
1620
1621         if (cfqq == cfqd->active_queue) {
1622                 /*
1623                  * if we are waiting for a request for this queue, let it rip
1624                  * immediately and flag that we must not expire this queue
1625                  * just now
1626                  */
1627                 if (cfq_cfqq_wait_request(cfqq)) {
1628                         cfq_mark_cfqq_must_dispatch(cfqq);
1629                         del_timer(&cfqd->idle_slice_timer);
1630                         blk_start_queueing(cfqd->queue);
1631                 }
1632         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1633                 /*
1634                  * not the active queue - expire current slice if it is
1635                  * idle and has expired it's mean thinktime or this new queue
1636                  * has some old slice time left and is of higher priority
1637                  */
1638                 cfq_preempt_queue(cfqd, cfqq);
1639                 cfq_mark_cfqq_must_dispatch(cfqq);
1640                 blk_start_queueing(cfqd->queue);
1641         }
1642 }
1643
1644 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1645 {
1646         struct cfq_data *cfqd = q->elevator->elevator_data;
1647         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1648
1649         cfq_init_prio_data(cfqq);
1650
1651         cfq_add_rq_rb(rq);
1652
1653         if (!cfq_cfqq_on_rr(cfqq))
1654                 cfq_add_cfqq_rr(cfqd, cfqq);
1655
1656         list_add_tail(&rq->queuelist, &cfqq->fifo);
1657
1658         cfq_rq_enqueued(cfqd, cfqq, rq);
1659 }
1660
1661 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1662 {
1663         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1664         struct cfq_data *cfqd = cfqq->cfqd;
1665         const int sync = rq_is_sync(rq);
1666         unsigned long now;
1667
1668         now = jiffies;
1669
1670         WARN_ON(!cfqd->rq_in_driver);
1671         WARN_ON(!cfqq->on_dispatch[sync]);
1672         cfqd->rq_in_driver--;
1673         cfqq->on_dispatch[sync]--;
1674
1675         if (!cfq_class_idle(cfqq))
1676                 cfqd->last_end_request = now;
1677
1678         if (!cfq_cfqq_dispatched(cfqq) && cfq_cfqq_on_rr(cfqq))
1679                 cfq_resort_rr_list(cfqq, 0);
1680
1681         if (sync)
1682                 RQ_CIC(rq)->last_end_request = now;
1683
1684         /*
1685          * If this is the active queue, check if it needs to be expired,
1686          * or if we want to idle in case it has no pending requests.
1687          */
1688         if (cfqd->active_queue == cfqq) {
1689                 if (time_after(now, cfqq->slice_end))
1690                         cfq_slice_expired(cfqd, 0);
1691                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1692                         if (!cfq_arm_slice_timer(cfqd, cfqq))
1693                                 cfq_schedule_dispatch(cfqd);
1694                 }
1695         }
1696 }
1697
1698 /*
1699  * we temporarily boost lower priority queues if they are holding fs exclusive
1700  * resources. they are boosted to normal prio (CLASS_BE/4)
1701  */
1702 static void cfq_prio_boost(struct cfq_queue *cfqq)
1703 {
1704         const int ioprio_class = cfqq->ioprio_class;
1705         const int ioprio = cfqq->ioprio;
1706
1707         if (has_fs_excl()) {
1708                 /*
1709                  * boost idle prio on transactions that would lock out other
1710                  * users of the filesystem
1711                  */
1712                 if (cfq_class_idle(cfqq))
1713                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1714                 if (cfqq->ioprio > IOPRIO_NORM)
1715                         cfqq->ioprio = IOPRIO_NORM;
1716         } else {
1717                 /*
1718                  * check if we need to unboost the queue
1719                  */
1720                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1721                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1722                 if (cfqq->ioprio != cfqq->org_ioprio)
1723                         cfqq->ioprio = cfqq->org_ioprio;
1724         }
1725
1726         /*
1727          * refile between round-robin lists if we moved the priority class
1728          */
1729         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1730             cfq_cfqq_on_rr(cfqq))
1731                 cfq_resort_rr_list(cfqq, 0);
1732 }
1733
1734 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1735 {
1736         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1737             !cfq_cfqq_must_alloc_slice(cfqq)) {
1738                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1739                 return ELV_MQUEUE_MUST;
1740         }
1741
1742         return ELV_MQUEUE_MAY;
1743 }
1744
1745 static int cfq_may_queue(request_queue_t *q, int rw)
1746 {
1747         struct cfq_data *cfqd = q->elevator->elevator_data;
1748         struct task_struct *tsk = current;
1749         struct cfq_queue *cfqq;
1750
1751         /*
1752          * don't force setup of a queue from here, as a call to may_queue
1753          * does not necessarily imply that a request actually will be queued.
1754          * so just lookup a possibly existing queue, or return 'may queue'
1755          * if that fails
1756          */
1757         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1758         if (cfqq) {
1759                 cfq_init_prio_data(cfqq);
1760                 cfq_prio_boost(cfqq);
1761
1762                 return __cfq_may_queue(cfqq);
1763         }
1764
1765         return ELV_MQUEUE_MAY;
1766 }
1767
1768 /*
1769  * queue lock held here
1770  */
1771 static void cfq_put_request(request_queue_t *q, struct request *rq)
1772 {
1773         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1774
1775         if (cfqq) {
1776                 const int rw = rq_data_dir(rq);
1777
1778                 BUG_ON(!cfqq->allocated[rw]);
1779                 cfqq->allocated[rw]--;
1780
1781                 put_io_context(RQ_CIC(rq)->ioc);
1782
1783                 rq->elevator_private = NULL;
1784                 rq->elevator_private2 = NULL;
1785
1786                 cfq_put_queue(cfqq);
1787         }
1788 }
1789
1790 /*
1791  * Allocate cfq data structures associated with this request.
1792  */
1793 static int
1794 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1795 {
1796         struct cfq_data *cfqd = q->elevator->elevator_data;
1797         struct task_struct *tsk = current;
1798         struct cfq_io_context *cic;
1799         const int rw = rq_data_dir(rq);
1800         pid_t key = cfq_queue_pid(tsk, rw);
1801         struct cfq_queue *cfqq;
1802         unsigned long flags;
1803         int is_sync = key != CFQ_KEY_ASYNC;
1804
1805         might_sleep_if(gfp_mask & __GFP_WAIT);
1806
1807         cic = cfq_get_io_context(cfqd, gfp_mask);
1808
1809         spin_lock_irqsave(q->queue_lock, flags);
1810
1811         if (!cic)
1812                 goto queue_fail;
1813
1814         if (!cic->cfqq[is_sync]) {
1815                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1816                 if (!cfqq)
1817                         goto queue_fail;
1818
1819                 cic->cfqq[is_sync] = cfqq;
1820         } else
1821                 cfqq = cic->cfqq[is_sync];
1822
1823         cfqq->allocated[rw]++;
1824         cfq_clear_cfqq_must_alloc(cfqq);
1825         atomic_inc(&cfqq->ref);
1826
1827         spin_unlock_irqrestore(q->queue_lock, flags);
1828
1829         rq->elevator_private = cic;
1830         rq->elevator_private2 = cfqq;
1831         return 0;
1832
1833 queue_fail:
1834         if (cic)
1835                 put_io_context(cic->ioc);
1836
1837         cfq_schedule_dispatch(cfqd);
1838         spin_unlock_irqrestore(q->queue_lock, flags);
1839         return 1;
1840 }
1841
1842 static void cfq_kick_queue(void *data)
1843 {
1844         request_queue_t *q = data;
1845         unsigned long flags;
1846
1847         spin_lock_irqsave(q->queue_lock, flags);
1848         blk_start_queueing(q);
1849         spin_unlock_irqrestore(q->queue_lock, flags);
1850 }
1851
1852 /*
1853  * Timer running if the active_queue is currently idling inside its time slice
1854  */
1855 static void cfq_idle_slice_timer(unsigned long data)
1856 {
1857         struct cfq_data *cfqd = (struct cfq_data *) data;
1858         struct cfq_queue *cfqq;
1859         unsigned long flags;
1860
1861         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1862
1863         if ((cfqq = cfqd->active_queue) != NULL) {
1864                 unsigned long now = jiffies;
1865
1866                 /*
1867                  * expired
1868                  */
1869                 if (time_after(now, cfqq->slice_end))
1870                         goto expire;
1871
1872                 /*
1873                  * only expire and reinvoke request handler, if there are
1874                  * other queues with pending requests
1875                  */
1876                 if (!cfqd->busy_queues)
1877                         goto out_cont;
1878
1879                 /*
1880                  * not expired and it has a request pending, let it dispatch
1881                  */
1882                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1883                         cfq_mark_cfqq_must_dispatch(cfqq);
1884                         goto out_kick;
1885                 }
1886         }
1887 expire:
1888         cfq_slice_expired(cfqd, 0);
1889 out_kick:
1890         cfq_schedule_dispatch(cfqd);
1891 out_cont:
1892         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1893 }
1894
1895 /*
1896  * Timer running if an idle class queue is waiting for service
1897  */
1898 static void cfq_idle_class_timer(unsigned long data)
1899 {
1900         struct cfq_data *cfqd = (struct cfq_data *) data;
1901         unsigned long flags, end;
1902
1903         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1904
1905         /*
1906          * race with a non-idle queue, reset timer
1907          */
1908         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
1909         if (!time_after_eq(jiffies, end))
1910                 mod_timer(&cfqd->idle_class_timer, end);
1911         else
1912                 cfq_schedule_dispatch(cfqd);
1913
1914         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1915 }
1916
1917 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1918 {
1919         del_timer_sync(&cfqd->idle_slice_timer);
1920         del_timer_sync(&cfqd->idle_class_timer);
1921         blk_sync_queue(cfqd->queue);
1922 }
1923
1924 static void cfq_exit_queue(elevator_t *e)
1925 {
1926         struct cfq_data *cfqd = e->elevator_data;
1927         request_queue_t *q = cfqd->queue;
1928
1929         cfq_shutdown_timer_wq(cfqd);
1930
1931         spin_lock_irq(q->queue_lock);
1932
1933         if (cfqd->active_queue)
1934                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
1935
1936         while (!list_empty(&cfqd->cic_list)) {
1937                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1938                                                         struct cfq_io_context,
1939                                                         queue_list);
1940
1941                 __cfq_exit_single_io_context(cfqd, cic);
1942         }
1943
1944         spin_unlock_irq(q->queue_lock);
1945
1946         cfq_shutdown_timer_wq(cfqd);
1947
1948         kfree(cfqd->cfq_hash);
1949         kfree(cfqd);
1950 }
1951
1952 static void *cfq_init_queue(request_queue_t *q, elevator_t *e)
1953 {
1954         struct cfq_data *cfqd;
1955         int i;
1956
1957         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
1958         if (!cfqd)
1959                 return NULL;
1960
1961         memset(cfqd, 0, sizeof(*cfqd));
1962
1963         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1964                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
1965
1966         INIT_LIST_HEAD(&cfqd->busy_rr);
1967         INIT_LIST_HEAD(&cfqd->cur_rr);
1968         INIT_LIST_HEAD(&cfqd->idle_rr);
1969         INIT_LIST_HEAD(&cfqd->cic_list);
1970
1971         cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
1972         if (!cfqd->cfq_hash)
1973                 goto out_free;
1974
1975         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
1976                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
1977
1978         cfqd->queue = q;
1979
1980         init_timer(&cfqd->idle_slice_timer);
1981         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
1982         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
1983
1984         init_timer(&cfqd->idle_class_timer);
1985         cfqd->idle_class_timer.function = cfq_idle_class_timer;
1986         cfqd->idle_class_timer.data = (unsigned long) cfqd;
1987
1988         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
1989
1990         cfqd->cfq_quantum = cfq_quantum;
1991         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
1992         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
1993         cfqd->cfq_back_max = cfq_back_max;
1994         cfqd->cfq_back_penalty = cfq_back_penalty;
1995         cfqd->cfq_slice[0] = cfq_slice_async;
1996         cfqd->cfq_slice[1] = cfq_slice_sync;
1997         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
1998         cfqd->cfq_slice_idle = cfq_slice_idle;
1999
2000         return cfqd;
2001 out_free:
2002         kfree(cfqd);
2003         return NULL;
2004 }
2005
2006 static void cfq_slab_kill(void)
2007 {
2008         if (cfq_pool)
2009                 kmem_cache_destroy(cfq_pool);
2010         if (cfq_ioc_pool)
2011                 kmem_cache_destroy(cfq_ioc_pool);
2012 }
2013
2014 static int __init cfq_slab_setup(void)
2015 {
2016         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2017                                         NULL, NULL);
2018         if (!cfq_pool)
2019                 goto fail;
2020
2021         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2022                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2023         if (!cfq_ioc_pool)
2024                 goto fail;
2025
2026         return 0;
2027 fail:
2028         cfq_slab_kill();
2029         return -ENOMEM;
2030 }
2031
2032 /*
2033  * sysfs parts below -->
2034  */
2035
2036 static ssize_t
2037 cfq_var_show(unsigned int var, char *page)
2038 {
2039         return sprintf(page, "%d\n", var);
2040 }
2041
2042 static ssize_t
2043 cfq_var_store(unsigned int *var, const char *page, size_t count)
2044 {
2045         char *p = (char *) page;
2046
2047         *var = simple_strtoul(p, &p, 10);
2048         return count;
2049 }
2050
2051 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2052 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2053 {                                                                       \
2054         struct cfq_data *cfqd = e->elevator_data;                       \
2055         unsigned int __data = __VAR;                                    \
2056         if (__CONV)                                                     \
2057                 __data = jiffies_to_msecs(__data);                      \
2058         return cfq_var_show(__data, (page));                            \
2059 }
2060 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2061 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2062 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2063 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2064 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2065 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2066 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2067 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2068 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2069 #undef SHOW_FUNCTION
2070
2071 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2072 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2073 {                                                                       \
2074         struct cfq_data *cfqd = e->elevator_data;                       \
2075         unsigned int __data;                                            \
2076         int ret = cfq_var_store(&__data, (page), count);                \
2077         if (__data < (MIN))                                             \
2078                 __data = (MIN);                                         \
2079         else if (__data > (MAX))                                        \
2080                 __data = (MAX);                                         \
2081         if (__CONV)                                                     \
2082                 *(__PTR) = msecs_to_jiffies(__data);                    \
2083         else                                                            \
2084                 *(__PTR) = __data;                                      \
2085         return ret;                                                     \
2086 }
2087 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2088 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2089 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2090 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2091 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2092 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2093 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2094 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2095 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2096 #undef STORE_FUNCTION
2097
2098 #define CFQ_ATTR(name) \
2099         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2100
2101 static struct elv_fs_entry cfq_attrs[] = {
2102         CFQ_ATTR(quantum),
2103         CFQ_ATTR(fifo_expire_sync),
2104         CFQ_ATTR(fifo_expire_async),
2105         CFQ_ATTR(back_seek_max),
2106         CFQ_ATTR(back_seek_penalty),
2107         CFQ_ATTR(slice_sync),
2108         CFQ_ATTR(slice_async),
2109         CFQ_ATTR(slice_async_rq),
2110         CFQ_ATTR(slice_idle),
2111         __ATTR_NULL
2112 };
2113
2114 static struct elevator_type iosched_cfq = {
2115         .ops = {
2116                 .elevator_merge_fn =            cfq_merge,
2117                 .elevator_merged_fn =           cfq_merged_request,
2118                 .elevator_merge_req_fn =        cfq_merged_requests,
2119                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2120                 .elevator_add_req_fn =          cfq_insert_request,
2121                 .elevator_activate_req_fn =     cfq_activate_request,
2122                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2123                 .elevator_queue_empty_fn =      cfq_queue_empty,
2124                 .elevator_completed_req_fn =    cfq_completed_request,
2125                 .elevator_former_req_fn =       elv_rb_former_request,
2126                 .elevator_latter_req_fn =       elv_rb_latter_request,
2127                 .elevator_set_req_fn =          cfq_set_request,
2128                 .elevator_put_req_fn =          cfq_put_request,
2129                 .elevator_may_queue_fn =        cfq_may_queue,
2130                 .elevator_init_fn =             cfq_init_queue,
2131                 .elevator_exit_fn =             cfq_exit_queue,
2132                 .trim =                         cfq_free_io_context,
2133         },
2134         .elevator_attrs =       cfq_attrs,
2135         .elevator_name =        "cfq",
2136         .elevator_owner =       THIS_MODULE,
2137 };
2138
2139 static int __init cfq_init(void)
2140 {
2141         int ret;
2142
2143         /*
2144          * could be 0 on HZ < 1000 setups
2145          */
2146         if (!cfq_slice_async)
2147                 cfq_slice_async = 1;
2148         if (!cfq_slice_idle)
2149                 cfq_slice_idle = 1;
2150
2151         if (cfq_slab_setup())
2152                 return -ENOMEM;
2153
2154         ret = elv_register(&iosched_cfq);
2155         if (ret)
2156                 cfq_slab_kill();
2157
2158         return ret;
2159 }
2160
2161 static void __exit cfq_exit(void)
2162 {
2163         DECLARE_COMPLETION_ONSTACK(all_gone);
2164         elv_unregister(&iosched_cfq);
2165         ioc_gone = &all_gone;
2166         /* ioc_gone's update must be visible before reading ioc_count */
2167         smp_wmb();
2168         if (elv_ioc_count_read(ioc_count))
2169                 wait_for_completion(ioc_gone);
2170         synchronize_rcu();
2171         cfq_slab_kill();
2172 }
2173
2174 module_init(cfq_init);
2175 module_exit(cfq_exit);
2176
2177 MODULE_AUTHOR("Jens Axboe");
2178 MODULE_LICENSE("GPL");
2179 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");