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@suse.de>
   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          * non-ordered list of empty cfqq's
  86          */
  87         struct list_head empty_list;
  88
  89         /*
  90          * cfqq lookup hash
  91          */
  92         struct hlist_head *cfq_hash;
  93
  94         int rq_in_driver;
  95         int hw_tag;
  96
  97         /*
  98          * idle window management
  99          */
 100         struct timer_list idle_slice_timer;
 101         struct work_struct unplug_work;
 102
 103         struct cfq_queue *active_queue;
 104         struct cfq_io_context *active_cic;
 105         int cur_prio, cur_end_prio;
 106         unsigned int dispatch_slice;
 107
 108         struct timer_list idle_class_timer;
 109
 110         sector_t last_sector;
 111         unsigned long last_end_request;
 112
 113         /*
 114          * tunables, see top of file
 115          */
 116         unsigned int cfq_quantum;
 117         unsigned int cfq_fifo_expire[2];
 118         unsigned int cfq_back_penalty;
 119         unsigned int cfq_back_max;
 120         unsigned int cfq_slice[2];
 121         unsigned int cfq_slice_async_rq;
 122         unsigned int cfq_slice_idle;
 123
 124         struct list_head cic_list;
 125 };
 126
 127 /*
 128  * Per process-grouping structure
 129  */
 130 struct cfq_queue {
 131         /* reference count */
 132         atomic_t ref;
 133         /* parent cfq_data */
 134         struct cfq_data *cfqd;
 135         /* cfqq lookup hash */
 136         struct hlist_node cfq_hash;
 137         /* hash key */
 138         unsigned int key;
 139         /* on either rr or empty list of cfqd */
 140         struct list_head cfq_list;
 141         /* sorted list of pending requests */
 142         struct rb_root sort_list;
 143         /* if fifo isn't expired, next request to serve */
 144         struct request *next_rq;
 145         /* requests queued in sort_list */
 146         int queued[2];
 147         /* currently allocated requests */
 148         int allocated[2];
 149         /* fifo list of requests in sort_list */
 150         struct list_head fifo;
 151
 152         unsigned long slice_start;
 153         unsigned long slice_end;
 154         unsigned long slice_left;
 155         unsigned long service_last;
 156
 157         /* number of requests that are on the dispatch list */
 158         int on_dispatch[2];
 159
 160         /* io prio of this group */
 161         unsigned short ioprio, org_ioprio;
 162         unsigned short ioprio_class, org_ioprio_class;
 163
 164         /* various state flags, see below */
 165         unsigned int flags;
 166 };
 167
 168 enum cfqq_state_flags {
 169         CFQ_CFQQ_FLAG_on_rr = 0,
 170         CFQ_CFQQ_FLAG_wait_request,
 171         CFQ_CFQQ_FLAG_must_alloc,
 172         CFQ_CFQQ_FLAG_must_alloc_slice,
 173         CFQ_CFQQ_FLAG_must_dispatch,
 174         CFQ_CFQQ_FLAG_fifo_expire,
 175         CFQ_CFQQ_FLAG_idle_window,
 176         CFQ_CFQQ_FLAG_prio_changed,
 177 };
 178
 179 #define CFQ_CFQQ_FNS(name)                                              \
 180 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 181 {                                                                       \
 182         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 183 }                                                                       \
 184 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 185 {                                                                       \
 186         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 187 }                                                                       \
 188 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 189 {                                                                       \
 190         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 191 }
 192
 193 CFQ_CFQQ_FNS(on_rr);
 194 CFQ_CFQQ_FNS(wait_request);
 195 CFQ_CFQQ_FNS(must_alloc);
 196 CFQ_CFQQ_FNS(must_alloc_slice);
 197 CFQ_CFQQ_FNS(must_dispatch);
 198 CFQ_CFQQ_FNS(fifo_expire);
 199 CFQ_CFQQ_FNS(idle_window);
 200 CFQ_CFQQ_FNS(prio_changed);
 201 #undef CFQ_CFQQ_FNS
 202
 203 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 204 static void cfq_dispatch_insert(request_queue_t *, struct request *);
 205 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
 206
 207 /*
 208  * scheduler run of queue, if there are requests pending and no one in the
 209  * driver that will restart queueing
 210  */
 211 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 212 {
 213         if (cfqd->busy_queues)
 214                 kblockd_schedule_work(&cfqd->unplug_work);
 215 }
 216
 217 static int cfq_queue_empty(request_queue_t *q)
 218 {
 219         struct cfq_data *cfqd = q->elevator->elevator_data;
 220
 221         return !cfqd->busy_queues;
 222 }
 223
 224 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
 225 {
 226         if (rw == READ || rw == WRITE_SYNC)
 227                 return task->pid;
 228
 229         return CFQ_KEY_ASYNC;
 230 }
 231
 232 /*
 233  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 234  * We choose the request that is closest to the head right now. Distance
 235  * behind the head is penalized and only allowed to a certain extent.
 236  */
 237 static struct request *
 238 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
 239 {
 240         sector_t last, s1, s2, d1 = 0, d2 = 0;
 241         unsigned long back_max;
 242 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 243 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 244         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 245
 246         if (rq1 == NULL || rq1 == rq2)
 247                 return rq2;
 248         if (rq2 == NULL)
 249                 return rq1;
 250
 251         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 252                 return rq1;
 253         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 254                 return rq2;
 255
 256         s1 = rq1->sector;
 257         s2 = rq2->sector;
 258
 259         last = cfqd->last_sector;
 260
 261         /*
 262          * by definition, 1KiB is 2 sectors
 263          */
 264         back_max = cfqd->cfq_back_max * 2;
 265
 266         /*
 267          * Strict one way elevator _except_ in the case where we allow
 268          * short backward seeks which are biased as twice the cost of a
 269          * similar forward seek.
 270          */
 271         if (s1 >= last)
 272                 d1 = s1 - last;
 273         else if (s1 + back_max >= last)
 274                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 275         else
 276                 wrap |= CFQ_RQ1_WRAP;
 277
 278         if (s2 >= last)
 279                 d2 = s2 - last;
 280         else if (s2 + back_max >= last)
 281                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 282         else
 283                 wrap |= CFQ_RQ2_WRAP;
 284
 285         /* Found required data */
 286
 287         /*
 288          * By doing switch() on the bit mask "wrap" we avoid having to
 289          * check two variables for all permutations: --> faster!
 290          */
 291         switch (wrap) {
 292         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 293                 if (d1 < d2)
 294                         return rq1;
 295                 else if (d2 < d1)
 296                         return rq2;
 297                 else {
 298                         if (s1 >= s2)
 299                                 return rq1;
 300                         else
 301                                 return rq2;
 302                 }
 303
 304         case CFQ_RQ2_WRAP:
 305                 return rq1;
 306         case CFQ_RQ1_WRAP:
 307                 return rq2;
 308         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
 309         default:
 310                 /*
 311                  * Since both rqs are wrapped,
 312                  * start with the one that's further behind head
 313                  * (--> only *one* back seek required),
 314                  * since back seek takes more time than forward.
 315                  */
 316                 if (s1 <= s2)
 317                         return rq1;
 318                 else
 319                         return rq2;
 320         }
 321 }
 322
 323 /*
 324  * would be nice to take fifo expire time into account as well
 325  */
 326 static struct request *
 327 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 328                   struct request *last)
 329 {
 330         struct rb_node *rbnext = rb_next(&last->rb_node);
 331         struct rb_node *rbprev = rb_prev(&last->rb_node);
 332         struct request *next = NULL, *prev = NULL;
 333
 334         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
 335
 336         if (rbprev)
 337                 prev = rb_entry_rq(rbprev);
 338
 339         if (rbnext)
 340                 next = rb_entry_rq(rbnext);
 341         else {
 342                 rbnext = rb_first(&cfqq->sort_list);
 343                 if (rbnext && rbnext != &last->rb_node)
 344                         next = rb_entry_rq(rbnext);
 345         }
 346
 347         return cfq_choose_req(cfqd, next, prev);
 348 }
 349
 350 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 351 {
 352         struct cfq_data *cfqd = cfqq->cfqd;
 353         struct list_head *list, *entry;
 354
 355         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 356
 357         list_del(&cfqq->cfq_list);
 358
 359         if (cfq_class_rt(cfqq))
 360                 list = &cfqd->cur_rr;
 361         else if (cfq_class_idle(cfqq))
 362                 list = &cfqd->idle_rr;
 363         else {
 364                 /*
 365                  * if cfqq has requests in flight, don't allow it to be
 366                  * found in cfq_set_active_queue before it has finished them.
 367                  * this is done to increase fairness between a process that
 368                  * has lots of io pending vs one that only generates one
 369                  * sporadically or synchronously
 370                  */
 371                 if (cfq_cfqq_dispatched(cfqq))
 372                         list = &cfqd->busy_rr;
 373                 else
 374                         list = &cfqd->rr_list[cfqq->ioprio];
 375         }
 376
 377         /*
 378          * if queue was preempted, just add to front to be fair. busy_rr
 379          * isn't sorted, but insert at the back for fairness.
 380          */
 381         if (preempted || list == &cfqd->busy_rr) {
 382                 if (preempted)
 383                         list = list->prev;
 384
 385                 list_add_tail(&cfqq->cfq_list, list);
 386                 return;
 387         }
 388
 389         /*
 390          * sort by when queue was last serviced
 391          */
 392         entry = list;
 393         while ((entry = entry->prev) != list) {
 394                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
 395
 396                 if (!__cfqq->service_last)
 397                         break;
 398                 if (time_before(__cfqq->service_last, cfqq->service_last))
 399                         break;
 400         }
 401
 402         list_add(&cfqq->cfq_list, entry);
 403 }
 404
 405 /*
 406  * add to busy list of queues for service, trying to be fair in ordering
 407  * the pending list according to last request service
 408  */
 409 static inline void
 410 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 411 {
 412         BUG_ON(cfq_cfqq_on_rr(cfqq));
 413         cfq_mark_cfqq_on_rr(cfqq);
 414         cfqd->busy_queues++;
 415
 416         cfq_resort_rr_list(cfqq, 0);
 417 }
 418
 419 static inline void
 420 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 421 {
 422         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 423         cfq_clear_cfqq_on_rr(cfqq);
 424         list_move(&cfqq->cfq_list, &cfqd->empty_list);
 425
 426         BUG_ON(!cfqd->busy_queues);
 427         cfqd->busy_queues--;
 428 }
 429
 430 /*
 431  * rb tree support functions
 432  */
 433 static inline void cfq_del_rq_rb(struct request *rq)
 434 {
 435         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 436         struct cfq_data *cfqd = cfqq->cfqd;
 437         const int sync = rq_is_sync(rq);
 438
 439         BUG_ON(!cfqq->queued[sync]);
 440         cfqq->queued[sync]--;
 441
 442         elv_rb_del(&cfqq->sort_list, rq);
 443
 444         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
 445                 cfq_del_cfqq_rr(cfqd, cfqq);
 446 }
 447
 448 static void cfq_add_rq_rb(struct request *rq)
 449 {
 450         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 451         struct cfq_data *cfqd = cfqq->cfqd;
 452         struct request *__alias;
 453
 454         cfqq->queued[rq_is_sync(rq)]++;
 455
 456         /*
 457          * looks a little odd, but the first insert might return an alias.
 458          * if that happens, put the alias on the dispatch list
 459          */
 460         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
 461                 cfq_dispatch_insert(cfqd->queue, __alias);
 462 }
 463
 464 static inline void
 465 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
 466 {
 467         elv_rb_del(&cfqq->sort_list, rq);
 468         cfqq->queued[rq_is_sync(rq)]--;
 469         cfq_add_rq_rb(rq);
 470 }
 471
 472 static struct request *
 473 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 474 {
 475         struct task_struct *tsk = current;
 476         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
 477         struct cfq_queue *cfqq;
 478
 479         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
 480         if (cfqq) {
 481                 sector_t sector = bio->bi_sector + bio_sectors(bio);
 482
 483                 return elv_rb_find(&cfqq->sort_list, sector);
 484         }
 485
 486         return NULL;
 487 }
 488
 489 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 490 {
 491         struct cfq_data *cfqd = q->elevator->elevator_data;
 492
 493         cfqd->rq_in_driver++;
 494
 495         /*
 496          * If the depth is larger 1, it really could be queueing. But lets
 497          * make the mark a little higher - idling could still be good for
 498          * low queueing, and a low queueing number could also just indicate
 499          * a SCSI mid layer like behaviour where limit+1 is often seen.
 500          */
 501         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
 502                 cfqd->hw_tag = 1;
 503 }
 504
 505 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 506 {
 507         struct cfq_data *cfqd = q->elevator->elevator_data;
 508
 509         WARN_ON(!cfqd->rq_in_driver);
 510         cfqd->rq_in_driver--;
 511 }
 512
 513 static void cfq_remove_request(struct request *rq)
 514 {
 515         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 516
 517         if (cfqq->next_rq == rq)
 518                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
 519
 520         list_del_init(&rq->queuelist);
 521         cfq_del_rq_rb(rq);
 522 }
 523
 524 static int
 525 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 526 {
 527         struct cfq_data *cfqd = q->elevator->elevator_data;
 528         struct request *__rq;
 529
 530         __rq = cfq_find_rq_fmerge(cfqd, bio);
 531         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 532                 *req = __rq;
 533                 return ELEVATOR_FRONT_MERGE;
 534         }
 535
 536         return ELEVATOR_NO_MERGE;
 537 }
 538
 539 static void cfq_merged_request(request_queue_t *q, struct request *req,
 540                                int type)
 541 {
 542         if (type == ELEVATOR_FRONT_MERGE) {
 543                 struct cfq_queue *cfqq = RQ_CFQQ(req);
 544
 545                 cfq_reposition_rq_rb(cfqq, req);
 546         }
 547 }
 548
 549 static void
 550 cfq_merged_requests(request_queue_t *q, struct request *rq,
 551                     struct request *next)
 552 {
 553         /*
 554          * reposition in fifo if next is older than rq
 555          */
 556         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 557             time_before(next->start_time, rq->start_time))
 558                 list_move(&rq->queuelist, &next->queuelist);
 559
 560         cfq_remove_request(next);
 561 }
 562
 563 static inline void
 564 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 565 {
 566         if (cfqq) {
 567                 /*
 568                  * stop potential idle class queues waiting service
 569                  */
 570                 del_timer(&cfqd->idle_class_timer);
 571
 572                 cfqq->slice_start = jiffies;
 573                 cfqq->slice_end = 0;
 574                 cfqq->slice_left = 0;
 575                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 576                 cfq_clear_cfqq_fifo_expire(cfqq);
 577         }
 578
 579         cfqd->active_queue = cfqq;
 580 }
 581
 582 /*
 583  * current cfqq expired its slice (or was too idle), select new one
 584  */
 585 static void
 586 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 587                     int preempted)
 588 {
 589         unsigned long now = jiffies;
 590
 591         if (cfq_cfqq_wait_request(cfqq))
 592                 del_timer(&cfqd->idle_slice_timer);
 593
 594         if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
 595                 cfqq->service_last = now;
 596                 cfq_schedule_dispatch(cfqd);
 597         }
 598
 599         cfq_clear_cfqq_must_dispatch(cfqq);
 600         cfq_clear_cfqq_wait_request(cfqq);
 601
 602         /*
 603          * store what was left of this slice, if the queue idled out
 604          * or was preempted
 605          */
 606         if (time_after(cfqq->slice_end, now))
 607                 cfqq->slice_left = cfqq->slice_end - now;
 608         else
 609                 cfqq->slice_left = 0;
 610
 611         if (cfq_cfqq_on_rr(cfqq))
 612                 cfq_resort_rr_list(cfqq, preempted);
 613
 614         if (cfqq == cfqd->active_queue)
 615                 cfqd->active_queue = NULL;
 616
 617         if (cfqd->active_cic) {
 618                 put_io_context(cfqd->active_cic->ioc);
 619                 cfqd->active_cic = NULL;
 620         }
 621
 622         cfqd->dispatch_slice = 0;
 623 }
 624
 625 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 626 {
 627         struct cfq_queue *cfqq = cfqd->active_queue;
 628
 629         if (cfqq)
 630                 __cfq_slice_expired(cfqd, cfqq, preempted);
 631 }
 632
 633 /*
 634  * 0
 635  * 0,1
 636  * 0,1,2
 637  * 0,1,2,3
 638  * 0,1,2,3,4
 639  * 0,1,2,3,4,5
 640  * 0,1,2,3,4,5,6
 641  * 0,1,2,3,4,5,6,7
 642  */
 643 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 644 {
 645         int prio, wrap;
 646
 647         prio = -1;
 648         wrap = 0;
 649         do {
 650                 int p;
 651
 652                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 653                         if (!list_empty(&cfqd->rr_list[p])) {
 654                                 prio = p;
 655                                 break;
 656                         }
 657                 }
 658
 659                 if (prio != -1)
 660                         break;
 661                 cfqd->cur_prio = 0;
 662                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 663                         cfqd->cur_end_prio = 0;
 664                         if (wrap)
 665                                 break;
 666                         wrap = 1;
 667                 }
 668         } while (1);
 669
 670         if (unlikely(prio == -1))
 671                 return -1;
 672
 673         BUG_ON(prio >= CFQ_PRIO_LISTS);
 674
 675         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 676
 677         cfqd->cur_prio = prio + 1;
 678         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 679                 cfqd->cur_end_prio = cfqd->cur_prio;
 680                 cfqd->cur_prio = 0;
 681         }
 682         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 683                 cfqd->cur_prio = 0;
 684                 cfqd->cur_end_prio = 0;
 685         }
 686
 687         return prio;
 688 }
 689
 690 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 691 {
 692         struct cfq_queue *cfqq = NULL;
 693
 694         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
 695                 /*
 696                  * if current list is non-empty, grab first entry. if it is
 697                  * empty, get next prio level and grab first entry then if any
 698                  * are spliced
 699                  */
 700                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 701         } else if (!list_empty(&cfqd->busy_rr)) {
 702                 /*
 703                  * If no new queues are available, check if the busy list has
 704                  * some before falling back to idle io.
 705                  */
 706                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
 707         } else if (!list_empty(&cfqd->idle_rr)) {
 708                 /*
 709                  * if we have idle queues and no rt or be queues had pending
 710                  * requests, either allow immediate service if the grace period
 711                  * has passed or arm the idle grace timer
 712                  */
 713                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 714
 715                 if (time_after_eq(jiffies, end))
 716                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 717                 else
 718                         mod_timer(&cfqd->idle_class_timer, end);
 719         }
 720
 721         __cfq_set_active_queue(cfqd, cfqq);
 722         return cfqq;
 723 }
 724
 725 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
 726
 727 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 728
 729 {
 730         struct cfq_io_context *cic;
 731         unsigned long sl;
 732
 733         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
 734         WARN_ON(cfqq != cfqd->active_queue);
 735
 736         /*
 737          * idle is disabled, either manually or by past process history
 738          */
 739         if (!cfqd->cfq_slice_idle)
 740                 return 0;
 741         if (!cfq_cfqq_idle_window(cfqq))
 742                 return 0;
 743         /*
 744          * task has exited, don't wait
 745          */
 746         cic = cfqd->active_cic;
 747         if (!cic || !cic->ioc->task)
 748                 return 0;
 749
 750         cfq_mark_cfqq_must_dispatch(cfqq);
 751         cfq_mark_cfqq_wait_request(cfqq);
 752
 753         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 754
 755         /*
 756          * we don't want to idle for seeks, but we do want to allow
 757          * fair distribution of slice time for a process doing back-to-back
 758          * seeks. so allow a little bit of time for him to submit a new rq
 759          */
 760         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
 761                 sl = min(sl, msecs_to_jiffies(2));
 762
 763         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 764         return 1;
 765 }
 766
 767 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
 768 {
 769         struct cfq_data *cfqd = q->elevator->elevator_data;
 770         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 771
 772         cfq_remove_request(rq);
 773         cfqq->on_dispatch[rq_is_sync(rq)]++;
 774         elv_dispatch_sort(q, rq);
 775
 776         rq = list_entry(q->queue_head.prev, struct request, queuelist);
 777         cfqd->last_sector = rq->sector + rq->nr_sectors;
 778 }
 779
 780 /*
 781  * return expired entry, or NULL to just start from scratch in rbtree
 782  */
 783 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
 784 {
 785         struct cfq_data *cfqd = cfqq->cfqd;
 786         struct request *rq;
 787         int fifo;
 788
 789         if (cfq_cfqq_fifo_expire(cfqq))
 790                 return NULL;
 791         if (list_empty(&cfqq->fifo))
 792                 return NULL;
 793
 794         fifo = cfq_cfqq_class_sync(cfqq);
 795         rq = rq_entry_fifo(cfqq->fifo.next);
 796
 797         if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
 798                 cfq_mark_cfqq_fifo_expire(cfqq);
 799                 return rq;
 800         }
 801
 802         return NULL;
 803 }
 804
 805 /*
 806  * Scale schedule slice based on io priority. Use the sync time slice only
 807  * if a queue is marked sync and has sync io queued. A sync queue with async
 808  * io only, should not get full sync slice length.
 809  */
 810 static inline int
 811 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 812 {
 813         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 814
 815         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 816
 817         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
 818 }
 819
 820 static inline void
 821 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 822 {
 823         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 824 }
 825
 826 static inline int
 827 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 828 {
 829         const int base_rq = cfqd->cfq_slice_async_rq;
 830
 831         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 832
 833         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
 834 }
 835
 836 /*
 837  * get next queue for service
 838  */
 839 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
 840 {
 841         unsigned long now = jiffies;
 842         struct cfq_queue *cfqq;
 843
 844         cfqq = cfqd->active_queue;
 845         if (!cfqq)
 846                 goto new_queue;
 847
 848         /*
 849          * slice has expired
 850          */
 851         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
 852                 goto expire;
 853
 854         /*
 855          * if queue has requests, dispatch one. if not, check if
 856          * enough slice is left to wait for one
 857          */
 858         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
 859                 goto keep_queue;
 860         else if (cfq_cfqq_dispatched(cfqq)) {
 861                 cfqq = NULL;
 862                 goto keep_queue;
 863         } else if (cfq_cfqq_class_sync(cfqq)) {
 864                 if (cfq_arm_slice_timer(cfqd, cfqq))
 865                         return NULL;
 866         }
 867
 868 expire:
 869         cfq_slice_expired(cfqd, 0);
 870 new_queue:
 871         cfqq = cfq_set_active_queue(cfqd);
 872 keep_queue:
 873         return cfqq;
 874 }
 875
 876 static int
 877 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 878                         int max_dispatch)
 879 {
 880         int dispatched = 0;
 881
 882         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
 883
 884         do {
 885                 struct request *rq;
 886
 887                 /*
 888                  * follow expired path, else get first next available
 889                  */
 890                 if ((rq = cfq_check_fifo(cfqq)) == NULL)
 891                         rq = cfqq->next_rq;
 892
 893                 /*
 894                  * finally, insert request into driver dispatch list
 895                  */
 896                 cfq_dispatch_insert(cfqd->queue, rq);
 897
 898                 cfqd->dispatch_slice++;
 899                 dispatched++;
 900
 901                 if (!cfqd->active_cic) {
 902                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
 903                         cfqd->active_cic = RQ_CIC(rq);
 904                 }
 905
 906                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
 907                         break;
 908
 909         } while (dispatched < max_dispatch);
 910
 911         /*
 912          * if slice end isn't set yet, set it.
 913          */
 914         if (!cfqq->slice_end)
 915                 cfq_set_prio_slice(cfqd, cfqq);
 916
 917         /*
 918          * expire an async queue immediately if it has used up its slice. idle
 919          * queue always expire after 1 dispatch round.
 920          */
 921         if ((!cfq_cfqq_sync(cfqq) &&
 922             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
 923             cfq_class_idle(cfqq) ||
 924             !cfq_cfqq_idle_window(cfqq))
 925                 cfq_slice_expired(cfqd, 0);
 926
 927         return dispatched;
 928 }
 929
 930 static int
 931 cfq_forced_dispatch_cfqqs(struct list_head *list)
 932 {
 933         struct cfq_queue *cfqq, *next;
 934         int dispatched;
 935
 936         dispatched = 0;
 937         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
 938                 while (cfqq->next_rq) {
 939                         cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
 940                         dispatched++;
 941                 }
 942                 BUG_ON(!list_empty(&cfqq->fifo));
 943         }
 944
 945         return dispatched;
 946 }
 947
 948 static int
 949 cfq_forced_dispatch(struct cfq_data *cfqd)
 950 {
 951         int i, dispatched = 0;
 952
 953         for (i = 0; i < CFQ_PRIO_LISTS; i++)
 954                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
 955
 956         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
 957         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
 958         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
 959
 960         cfq_slice_expired(cfqd, 0);
 961
 962         BUG_ON(cfqd->busy_queues);
 963
 964         return dispatched;
 965 }
 966
 967 static int
 968 cfq_dispatch_requests(request_queue_t *q, int force)
 969 {
 970         struct cfq_data *cfqd = q->elevator->elevator_data;
 971         struct cfq_queue *cfqq, *prev_cfqq;
 972         int dispatched;
 973
 974         if (!cfqd->busy_queues)
 975                 return 0;
 976
 977         if (unlikely(force))
 978                 return cfq_forced_dispatch(cfqd);
 979
 980         dispatched = 0;
 981         prev_cfqq = NULL;
 982         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
 983                 int max_dispatch;
 984
 985                 /*
 986                  * Don't repeat dispatch from the previous queue.
 987                  */
 988                 if (prev_cfqq == cfqq)
 989                         break;
 990
 991                 cfq_clear_cfqq_must_dispatch(cfqq);
 992                 cfq_clear_cfqq_wait_request(cfqq);
 993                 del_timer(&cfqd->idle_slice_timer);
 994
 995                 max_dispatch = cfqd->cfq_quantum;
 996                 if (cfq_class_idle(cfqq))
 997                         max_dispatch = 1;
 998
 999                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1000
1001                 /*
1002                  * If the dispatch cfqq has idling enabled and is still
1003                  * the active queue, break out.
1004                  */
1005                 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1006                         break;
1007
1008                 prev_cfqq = cfqq;
1009         }
1010
1011         return dispatched;
1012 }
1013
1014 /*
1015  * task holds one reference to the queue, dropped when task exits. each rq
1016  * in-flight on this queue also holds a reference, dropped when rq is freed.
1017  *
1018  * queue lock must be held here.
1019  */
1020 static void cfq_put_queue(struct cfq_queue *cfqq)
1021 {
1022         struct cfq_data *cfqd = cfqq->cfqd;
1023
1024         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1025
1026         if (!atomic_dec_and_test(&cfqq->ref))
1027                 return;
1028
1029         BUG_ON(rb_first(&cfqq->sort_list));
1030         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1031         BUG_ON(cfq_cfqq_on_rr(cfqq));
1032
1033         if (unlikely(cfqd->active_queue == cfqq))
1034                 __cfq_slice_expired(cfqd, cfqq, 0);
1035
1036         /*
1037          * it's on the empty list and still hashed
1038          */
1039         list_del(&cfqq->cfq_list);
1040         hlist_del(&cfqq->cfq_hash);
1041         kmem_cache_free(cfq_pool, cfqq);
1042 }
1043
1044 static inline struct cfq_queue *
1045 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1046                     const int hashval)
1047 {
1048         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1049         struct hlist_node *entry;
1050         struct cfq_queue *__cfqq;
1051
1052         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1053                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1054
1055                 if (__cfqq->key == key && (__p == prio || !prio))
1056                         return __cfqq;
1057         }
1058
1059         return NULL;
1060 }
1061
1062 static struct cfq_queue *
1063 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1064 {
1065         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1066 }
1067
1068 static void cfq_free_io_context(struct io_context *ioc)
1069 {
1070         struct cfq_io_context *__cic;
1071         struct rb_node *n;
1072         int freed = 0;
1073
1074         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1075                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1076                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1077                 kmem_cache_free(cfq_ioc_pool, __cic);
1078                 freed++;
1079         }
1080
1081         elv_ioc_count_mod(ioc_count, -freed);
1082
1083         if (ioc_gone && !elv_ioc_count_read(ioc_count))
1084                 complete(ioc_gone);
1085 }
1086
1087 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1088 {
1089         if (unlikely(cfqq == cfqd->active_queue))
1090                 __cfq_slice_expired(cfqd, cfqq, 0);
1091
1092         cfq_put_queue(cfqq);
1093 }
1094
1095 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1096                                          struct cfq_io_context *cic)
1097 {
1098         list_del_init(&cic->queue_list);
1099         smp_wmb();
1100         cic->key = NULL;
1101
1102         if (cic->cfqq[ASYNC]) {
1103                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1104                 cic->cfqq[ASYNC] = NULL;
1105         }
1106
1107         if (cic->cfqq[SYNC]) {
1108                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1109                 cic->cfqq[SYNC] = NULL;
1110         }
1111 }
1112
1113
1114 /*
1115  * Called with interrupts disabled
1116  */
1117 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1118 {
1119         struct cfq_data *cfqd = cic->key;
1120
1121         if (cfqd) {
1122                 request_queue_t *q = cfqd->queue;
1123
1124                 spin_lock_irq(q->queue_lock);
1125                 __cfq_exit_single_io_context(cfqd, cic);
1126                 spin_unlock_irq(q->queue_lock);
1127         }
1128 }
1129
1130 static void cfq_exit_io_context(struct io_context *ioc)
1131 {
1132         struct cfq_io_context *__cic;
1133         struct rb_node *n;
1134
1135         /*
1136          * put the reference this task is holding to the various queues
1137          */
1138
1139         n = rb_first(&ioc->cic_root);
1140         while (n != NULL) {
1141                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1142
1143                 cfq_exit_single_io_context(__cic);
1144                 n = rb_next(n);
1145         }
1146 }
1147
1148 static struct cfq_io_context *
1149 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1150 {
1151         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1152
1153         if (cic) {
1154                 memset(cic, 0, sizeof(*cic));
1155                 cic->last_end_request = jiffies;
1156                 INIT_LIST_HEAD(&cic->queue_list);
1157                 cic->dtor = cfq_free_io_context;
1158                 cic->exit = cfq_exit_io_context;
1159                 elv_ioc_count_inc(ioc_count);
1160         }
1161
1162         return cic;
1163 }
1164
1165 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1166 {
1167         struct task_struct *tsk = current;
1168         int ioprio_class;
1169
1170         if (!cfq_cfqq_prio_changed(cfqq))
1171                 return;
1172
1173         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1174         switch (ioprio_class) {
1175                 default:
1176                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1177                 case IOPRIO_CLASS_NONE:
1178                         /*
1179                          * no prio set, place us in the middle of the BE classes
1180                          */
1181                         cfqq->ioprio = task_nice_ioprio(tsk);
1182                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1183                         break;
1184                 case IOPRIO_CLASS_RT:
1185                         cfqq->ioprio = task_ioprio(tsk);
1186                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1187                         break;
1188                 case IOPRIO_CLASS_BE:
1189                         cfqq->ioprio = task_ioprio(tsk);
1190                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1191                         break;
1192                 case IOPRIO_CLASS_IDLE:
1193                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1194                         cfqq->ioprio = 7;
1195                         cfq_clear_cfqq_idle_window(cfqq);
1196                         break;
1197         }
1198
1199         /*
1200          * keep track of original prio settings in case we have to temporarily
1201          * elevate the priority of this queue
1202          */
1203         cfqq->org_ioprio = cfqq->ioprio;
1204         cfqq->org_ioprio_class = cfqq->ioprio_class;
1205
1206         if (cfq_cfqq_on_rr(cfqq))
1207                 cfq_resort_rr_list(cfqq, 0);
1208
1209         cfq_clear_cfqq_prio_changed(cfqq);
1210 }
1211
1212 static inline void changed_ioprio(struct cfq_io_context *cic)
1213 {
1214         struct cfq_data *cfqd = cic->key;
1215         struct cfq_queue *cfqq;
1216
1217         if (unlikely(!cfqd))
1218                 return;
1219
1220         spin_lock(cfqd->queue->queue_lock);
1221
1222         cfqq = cic->cfqq[ASYNC];
1223         if (cfqq) {
1224                 struct cfq_queue *new_cfqq;
1225                 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1226                                          GFP_ATOMIC);
1227                 if (new_cfqq) {
1228                         cic->cfqq[ASYNC] = new_cfqq;
1229                         cfq_put_queue(cfqq);
1230                 }
1231         }
1232
1233         cfqq = cic->cfqq[SYNC];
1234         if (cfqq)
1235                 cfq_mark_cfqq_prio_changed(cfqq);
1236
1237         spin_unlock(cfqd->queue->queue_lock);
1238 }
1239
1240 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1241 {
1242         struct cfq_io_context *cic;
1243         struct rb_node *n;
1244
1245         ioc->ioprio_changed = 0;
1246
1247         n = rb_first(&ioc->cic_root);
1248         while (n != NULL) {
1249                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1250
1251                 changed_ioprio(cic);
1252                 n = rb_next(n);
1253         }
1254 }
1255
1256 static struct cfq_queue *
1257 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1258               gfp_t gfp_mask)
1259 {
1260         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1261         struct cfq_queue *cfqq, *new_cfqq = NULL;
1262         unsigned short ioprio;
1263
1264 retry:
1265         ioprio = tsk->ioprio;
1266         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1267
1268         if (!cfqq) {
1269                 if (new_cfqq) {
1270                         cfqq = new_cfqq;
1271                         new_cfqq = NULL;
1272                 } else if (gfp_mask & __GFP_WAIT) {
1273                         /*
1274                          * Inform the allocator of the fact that we will
1275                          * just repeat this allocation if it fails, to allow
1276                          * the allocator to do whatever it needs to attempt to
1277                          * free memory.
1278                          */
1279                         spin_unlock_irq(cfqd->queue->queue_lock);
1280                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask|__GFP_NOFAIL);
1281                         spin_lock_irq(cfqd->queue->queue_lock);
1282                         goto retry;
1283                 } else {
1284                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1285                         if (!cfqq)
1286                                 goto out;
1287                 }
1288
1289                 memset(cfqq, 0, sizeof(*cfqq));
1290
1291                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1292                 INIT_LIST_HEAD(&cfqq->cfq_list);
1293                 INIT_LIST_HEAD(&cfqq->fifo);
1294
1295                 cfqq->key = key;
1296                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1297                 atomic_set(&cfqq->ref, 0);
1298                 cfqq->cfqd = cfqd;
1299                 cfqq->service_last = 0;
1300                 /*
1301                  * set ->slice_left to allow preemption for a new process
1302                  */
1303                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1304                 cfq_mark_cfqq_idle_window(cfqq);
1305                 cfq_mark_cfqq_prio_changed(cfqq);
1306                 cfq_init_prio_data(cfqq);
1307         }
1308
1309         if (new_cfqq)
1310                 kmem_cache_free(cfq_pool, new_cfqq);
1311
1312         atomic_inc(&cfqq->ref);
1313 out:
1314         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1315         return cfqq;
1316 }
1317
1318 static void
1319 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1320 {
1321         WARN_ON(!list_empty(&cic->queue_list));
1322         rb_erase(&cic->rb_node, &ioc->cic_root);
1323         kmem_cache_free(cfq_ioc_pool, cic);
1324         elv_ioc_count_dec(ioc_count);
1325 }
1326
1327 static struct cfq_io_context *
1328 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1329 {
1330         struct rb_node *n;
1331         struct cfq_io_context *cic;
1332         void *k, *key = cfqd;
1333
1334 restart:
1335         n = ioc->cic_root.rb_node;
1336         while (n) {
1337                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1338                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1339                 k = cic->key;
1340                 if (unlikely(!k)) {
1341                         cfq_drop_dead_cic(ioc, cic);
1342                         goto restart;
1343                 }
1344
1345                 if (key < k)
1346                         n = n->rb_left;
1347                 else if (key > k)
1348                         n = n->rb_right;
1349                 else
1350                         return cic;
1351         }
1352
1353         return NULL;
1354 }
1355
1356 static inline void
1357 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1358              struct cfq_io_context *cic)
1359 {
1360         struct rb_node **p;
1361         struct rb_node *parent;
1362         struct cfq_io_context *__cic;
1363         void *k;
1364
1365         cic->ioc = ioc;
1366         cic->key = cfqd;
1367
1368 restart:
1369         parent = NULL;
1370         p = &ioc->cic_root.rb_node;
1371         while (*p) {
1372                 parent = *p;
1373                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1374                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1375                 k = __cic->key;
1376                 if (unlikely(!k)) {
1377                         cfq_drop_dead_cic(ioc, __cic);
1378                         goto restart;
1379                 }
1380
1381                 if (cic->key < k)
1382                         p = &(*p)->rb_left;
1383                 else if (cic->key > k)
1384                         p = &(*p)->rb_right;
1385                 else
1386                         BUG();
1387         }
1388
1389         rb_link_node(&cic->rb_node, parent, p);
1390         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1391
1392         spin_lock_irq(cfqd->queue->queue_lock);
1393         list_add(&cic->queue_list, &cfqd->cic_list);
1394         spin_unlock_irq(cfqd->queue->queue_lock);
1395 }
1396
1397 /*
1398  * Setup general io context and cfq io context. There can be several cfq
1399  * io contexts per general io context, if this process is doing io to more
1400  * than one device managed by cfq.
1401  */
1402 static struct cfq_io_context *
1403 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1404 {
1405         struct io_context *ioc = NULL;
1406         struct cfq_io_context *cic;
1407
1408         might_sleep_if(gfp_mask & __GFP_WAIT);
1409
1410         ioc = get_io_context(gfp_mask);
1411         if (!ioc)
1412                 return NULL;
1413
1414         cic = cfq_cic_rb_lookup(cfqd, ioc);
1415         if (cic)
1416                 goto out;
1417
1418         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1419         if (cic == NULL)
1420                 goto err;
1421
1422         cfq_cic_link(cfqd, ioc, cic);
1423 out:
1424         smp_read_barrier_depends();
1425         if (unlikely(ioc->ioprio_changed))
1426                 cfq_ioc_set_ioprio(ioc);
1427
1428         return cic;
1429 err:
1430         put_io_context(ioc);
1431         return NULL;
1432 }
1433
1434 static void
1435 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1436 {
1437         unsigned long elapsed, ttime;
1438
1439         /*
1440          * if this context already has stuff queued, thinktime is from
1441          * last queue not last end
1442          */
1443 #if 0
1444         if (time_after(cic->last_end_request, cic->last_queue))
1445                 elapsed = jiffies - cic->last_end_request;
1446         else
1447                 elapsed = jiffies - cic->last_queue;
1448 #else
1449                 elapsed = jiffies - cic->last_end_request;
1450 #endif
1451
1452         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1453
1454         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1455         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1456         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1457 }
1458
1459 static void
1460 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1461                        struct request *rq)
1462 {
1463         sector_t sdist;
1464         u64 total;
1465
1466         if (cic->last_request_pos < rq->sector)
1467                 sdist = rq->sector - cic->last_request_pos;
1468         else
1469                 sdist = cic->last_request_pos - rq->sector;
1470
1471         /*
1472          * Don't allow the seek distance to get too large from the
1473          * odd fragment, pagein, etc
1474          */
1475         if (cic->seek_samples <= 60) /* second&third seek */
1476                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1477         else
1478                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1479
1480         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1481         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1482         total = cic->seek_total + (cic->seek_samples/2);
1483         do_div(total, cic->seek_samples);
1484         cic->seek_mean = (sector_t)total;
1485 }
1486
1487 /*
1488  * Disable idle window if the process thinks too long or seeks so much that
1489  * it doesn't matter
1490  */
1491 static void
1492 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1493                        struct cfq_io_context *cic)
1494 {
1495         int enable_idle = cfq_cfqq_idle_window(cfqq);
1496
1497         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1498             (cfqd->hw_tag && CIC_SEEKY(cic)))
1499                 enable_idle = 0;
1500         else if (sample_valid(cic->ttime_samples)) {
1501                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1502                         enable_idle = 0;
1503                 else
1504                         enable_idle = 1;
1505         }
1506
1507         if (enable_idle)
1508                 cfq_mark_cfqq_idle_window(cfqq);
1509         else
1510                 cfq_clear_cfqq_idle_window(cfqq);
1511 }
1512
1513
1514 /*
1515  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1516  * no or if we aren't sure, a 1 will cause a preempt.
1517  */
1518 static int
1519 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1520                    struct request *rq)
1521 {
1522         struct cfq_queue *cfqq = cfqd->active_queue;
1523
1524         if (cfq_class_idle(new_cfqq))
1525                 return 0;
1526
1527         if (!cfqq)
1528                 return 0;
1529
1530         if (cfq_class_idle(cfqq))
1531                 return 1;
1532         if (!cfq_cfqq_wait_request(new_cfqq))
1533                 return 0;
1534         /*
1535          * if it doesn't have slice left, forget it
1536          */
1537         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1538                 return 0;
1539         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1540                 return 1;
1541
1542         return 0;
1543 }
1544
1545 /*
1546  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1547  * let it have half of its nominal slice.
1548  */
1549 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1550 {
1551         struct cfq_queue *__cfqq, *next;
1552
1553         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1554                 cfq_resort_rr_list(__cfqq, 1);
1555
1556         if (!cfqq->slice_left)
1557                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1558
1559         cfqq->slice_end = cfqq->slice_left + jiffies;
1560         cfq_slice_expired(cfqd, 1);
1561         __cfq_set_active_queue(cfqd, cfqq);
1562 }
1563
1564 /*
1565  * should really be a ll_rw_blk.c helper
1566  */
1567 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1568 {
1569         request_queue_t *q = cfqd->queue;
1570
1571         if (!blk_queue_plugged(q))
1572                 q->request_fn(q);
1573         else
1574                 __generic_unplug_device(q);
1575 }
1576
1577 /*
1578  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1579  * something we should do about it
1580  */
1581 static void
1582 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1583                 struct request *rq)
1584 {
1585         struct cfq_io_context *cic = RQ_CIC(rq);
1586
1587         /*
1588          * check if this request is a better next-serve candidate)) {
1589          */
1590         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1591         BUG_ON(!cfqq->next_rq);
1592
1593         /*
1594          * we never wait for an async request and we don't allow preemption
1595          * of an async request. so just return early
1596          */
1597         if (!rq_is_sync(rq)) {
1598                 /*
1599                  * sync process issued an async request, if it's waiting
1600                  * then expire it and kick rq handling.
1601                  */
1602                 if (cic == cfqd->active_cic &&
1603                     del_timer(&cfqd->idle_slice_timer)) {
1604                         cfq_slice_expired(cfqd, 0);
1605                         cfq_start_queueing(cfqd, cfqq);
1606                 }
1607                 return;
1608         }
1609
1610         cfq_update_io_thinktime(cfqd, cic);
1611         cfq_update_io_seektime(cfqd, cic, rq);
1612         cfq_update_idle_window(cfqd, cfqq, cic);
1613
1614         cic->last_queue = jiffies;
1615         cic->last_request_pos = rq->sector + rq->nr_sectors;
1616
1617         if (cfqq == cfqd->active_queue) {
1618                 /*
1619                  * if we are waiting for a request for this queue, let it rip
1620                  * immediately and flag that we must not expire this queue
1621                  * just now
1622                  */
1623                 if (cfq_cfqq_wait_request(cfqq)) {
1624                         cfq_mark_cfqq_must_dispatch(cfqq);
1625                         del_timer(&cfqd->idle_slice_timer);
1626                         cfq_start_queueing(cfqd, cfqq);
1627                 }
1628         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1629                 /*
1630                  * not the active queue - expire current slice if it is
1631                  * idle and has expired it's mean thinktime or this new queue
1632                  * has some old slice time left and is of higher priority
1633                  */
1634                 cfq_preempt_queue(cfqd, cfqq);
1635                 cfq_mark_cfqq_must_dispatch(cfqq);
1636                 cfq_start_queueing(cfqd, cfqq);
1637         }
1638 }
1639
1640 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1641 {
1642         struct cfq_data *cfqd = q->elevator->elevator_data;
1643         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1644
1645         cfq_init_prio_data(cfqq);
1646
1647         cfq_add_rq_rb(rq);
1648
1649         if (!cfq_cfqq_on_rr(cfqq))
1650                 cfq_add_cfqq_rr(cfqd, cfqq);
1651
1652         list_add_tail(&rq->queuelist, &cfqq->fifo);
1653
1654         cfq_rq_enqueued(cfqd, cfqq, rq);
1655 }
1656
1657 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1658 {
1659         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1660         struct cfq_data *cfqd = cfqq->cfqd;
1661         const int sync = rq_is_sync(rq);
1662         unsigned long now;
1663
1664         now = jiffies;
1665
1666         WARN_ON(!cfqd->rq_in_driver);
1667         WARN_ON(!cfqq->on_dispatch[sync]);
1668         cfqd->rq_in_driver--;
1669         cfqq->on_dispatch[sync]--;
1670
1671         if (!cfq_class_idle(cfqq))
1672                 cfqd->last_end_request = now;
1673
1674         if (!cfq_cfqq_dispatched(cfqq)) {
1675                 if (cfq_cfqq_on_rr(cfqq)) {
1676                         cfqq->service_last = now;
1677                         cfq_resort_rr_list(cfqq, 0);
1678                 }
1679         }
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_remove_plug(q);
1849         q->request_fn(q);
1850         spin_unlock_irqrestore(q->queue_lock, flags);
1851 }
1852
1853 /*
1854  * Timer running if the active_queue is currently idling inside its time slice
1855  */
1856 static void cfq_idle_slice_timer(unsigned long data)
1857 {
1858         struct cfq_data *cfqd = (struct cfq_data *) data;
1859         struct cfq_queue *cfqq;
1860         unsigned long flags;
1861
1862         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1863
1864         if ((cfqq = cfqd->active_queue) != NULL) {
1865                 unsigned long now = jiffies;
1866
1867                 /*
1868                  * expired
1869                  */
1870                 if (time_after(now, cfqq->slice_end))
1871                         goto expire;
1872
1873                 /*
1874                  * only expire and reinvoke request handler, if there are
1875                  * other queues with pending requests
1876                  */
1877                 if (!cfqd->busy_queues)
1878                         goto out_cont;
1879
1880                 /*
1881                  * not expired and it has a request pending, let it dispatch
1882                  */
1883                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1884                         cfq_mark_cfqq_must_dispatch(cfqq);
1885                         goto out_kick;
1886                 }
1887         }
1888 expire:
1889         cfq_slice_expired(cfqd, 0);
1890 out_kick:
1891         cfq_schedule_dispatch(cfqd);
1892 out_cont:
1893         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1894 }
1895
1896 /*
1897  * Timer running if an idle class queue is waiting for service
1898  */
1899 static void cfq_idle_class_timer(unsigned long data)
1900 {
1901         struct cfq_data *cfqd = (struct cfq_data *) data;
1902         unsigned long flags, end;
1903
1904         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1905
1906         /*
1907          * race with a non-idle queue, reset timer
1908          */
1909         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
1910         if (!time_after_eq(jiffies, end))
1911                 mod_timer(&cfqd->idle_class_timer, end);
1912         else
1913                 cfq_schedule_dispatch(cfqd);
1914
1915         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1916 }
1917
1918 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1919 {
1920         del_timer_sync(&cfqd->idle_slice_timer);
1921         del_timer_sync(&cfqd->idle_class_timer);
1922         blk_sync_queue(cfqd->queue);
1923 }
1924
1925 static void cfq_exit_queue(elevator_t *e)
1926 {
1927         struct cfq_data *cfqd = e->elevator_data;
1928         request_queue_t *q = cfqd->queue;
1929
1930         cfq_shutdown_timer_wq(cfqd);
1931
1932         spin_lock_irq(q->queue_lock);
1933
1934         if (cfqd->active_queue)
1935                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
1936
1937         while (!list_empty(&cfqd->cic_list)) {
1938                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1939                                                         struct cfq_io_context,
1940                                                         queue_list);
1941
1942                 __cfq_exit_single_io_context(cfqd, cic);
1943         }
1944
1945         spin_unlock_irq(q->queue_lock);
1946
1947         cfq_shutdown_timer_wq(cfqd);
1948
1949         kfree(cfqd->cfq_hash);
1950         kfree(cfqd);
1951 }
1952
1953 static void *cfq_init_queue(request_queue_t *q, elevator_t *e)
1954 {
1955         struct cfq_data *cfqd;
1956         int i;
1957
1958         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
1959         if (!cfqd)
1960                 return NULL;
1961
1962         memset(cfqd, 0, sizeof(*cfqd));
1963
1964         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1965                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
1966
1967         INIT_LIST_HEAD(&cfqd->busy_rr);
1968         INIT_LIST_HEAD(&cfqd->cur_rr);
1969         INIT_LIST_HEAD(&cfqd->idle_rr);
1970         INIT_LIST_HEAD(&cfqd->empty_list);
1971         INIT_LIST_HEAD(&cfqd->cic_list);
1972
1973         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
1974         if (!cfqd->cfq_hash)
1975                 goto out_free;
1976
1977         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
1978                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
1979
1980         cfqd->queue = q;
1981
1982         init_timer(&cfqd->idle_slice_timer);
1983         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
1984         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
1985
1986         init_timer(&cfqd->idle_class_timer);
1987         cfqd->idle_class_timer.function = cfq_idle_class_timer;
1988         cfqd->idle_class_timer.data = (unsigned long) cfqd;
1989
1990         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
1991
1992         cfqd->cfq_quantum = cfq_quantum;
1993         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
1994         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
1995         cfqd->cfq_back_max = cfq_back_max;
1996         cfqd->cfq_back_penalty = cfq_back_penalty;
1997         cfqd->cfq_slice[0] = cfq_slice_async;
1998         cfqd->cfq_slice[1] = cfq_slice_sync;
1999         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2000         cfqd->cfq_slice_idle = cfq_slice_idle;
2001
2002         return cfqd;
2003 out_free:
2004         kfree(cfqd);
2005         return NULL;
2006 }
2007
2008 static void cfq_slab_kill(void)
2009 {
2010         if (cfq_pool)
2011                 kmem_cache_destroy(cfq_pool);
2012         if (cfq_ioc_pool)
2013                 kmem_cache_destroy(cfq_ioc_pool);
2014 }
2015
2016 static int __init cfq_slab_setup(void)
2017 {
2018         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2019                                         NULL, NULL);
2020         if (!cfq_pool)
2021                 goto fail;
2022
2023         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2024                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2025         if (!cfq_ioc_pool)
2026                 goto fail;
2027
2028         return 0;
2029 fail:
2030         cfq_slab_kill();
2031         return -ENOMEM;
2032 }
2033
2034 /*
2035  * sysfs parts below -->
2036  */
2037
2038 static ssize_t
2039 cfq_var_show(unsigned int var, char *page)
2040 {
2041         return sprintf(page, "%d\n", var);
2042 }
2043
2044 static ssize_t
2045 cfq_var_store(unsigned int *var, const char *page, size_t count)
2046 {
2047         char *p = (char *) page;
2048
2049         *var = simple_strtoul(p, &p, 10);
2050         return count;
2051 }
2052
2053 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2054 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2055 {                                                                       \
2056         struct cfq_data *cfqd = e->elevator_data;                       \
2057         unsigned int __data = __VAR;                                    \
2058         if (__CONV)                                                     \
2059                 __data = jiffies_to_msecs(__data);                      \
2060         return cfq_var_show(__data, (page));                            \
2061 }
2062 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2063 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2064 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2065 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2066 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2067 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2068 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2069 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2070 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2071 #undef SHOW_FUNCTION
2072
2073 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2074 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2075 {                                                                       \
2076         struct cfq_data *cfqd = e->elevator_data;                       \
2077         unsigned int __data;                                            \
2078         int ret = cfq_var_store(&__data, (page), count);                \
2079         if (__data < (MIN))                                             \
2080                 __data = (MIN);                                         \
2081         else if (__data > (MAX))                                        \
2082                 __data = (MAX);                                         \
2083         if (__CONV)                                                     \
2084                 *(__PTR) = msecs_to_jiffies(__data);                    \
2085         else                                                            \
2086                 *(__PTR) = __data;                                      \
2087         return ret;                                                     \
2088 }
2089 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2090 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2091 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2092 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2093 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2094 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2095 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2096 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2097 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2098 #undef STORE_FUNCTION
2099
2100 #define CFQ_ATTR(name) \
2101         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2102
2103 static struct elv_fs_entry cfq_attrs[] = {
2104         CFQ_ATTR(quantum),
2105         CFQ_ATTR(fifo_expire_sync),
2106         CFQ_ATTR(fifo_expire_async),
2107         CFQ_ATTR(back_seek_max),
2108         CFQ_ATTR(back_seek_penalty),
2109         CFQ_ATTR(slice_sync),
2110         CFQ_ATTR(slice_async),
2111         CFQ_ATTR(slice_async_rq),
2112         CFQ_ATTR(slice_idle),
2113         __ATTR_NULL
2114 };
2115
2116 static struct elevator_type iosched_cfq = {
2117         .ops = {
2118                 .elevator_merge_fn =            cfq_merge,
2119                 .elevator_merged_fn =           cfq_merged_request,
2120                 .elevator_merge_req_fn =        cfq_merged_requests,
2121                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2122                 .elevator_add_req_fn =          cfq_insert_request,
2123                 .elevator_activate_req_fn =     cfq_activate_request,
2124                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2125                 .elevator_queue_empty_fn =      cfq_queue_empty,
2126                 .elevator_completed_req_fn =    cfq_completed_request,
2127                 .elevator_former_req_fn =       elv_rb_former_request,
2128                 .elevator_latter_req_fn =       elv_rb_latter_request,
2129                 .elevator_set_req_fn =          cfq_set_request,
2130                 .elevator_put_req_fn =          cfq_put_request,
2131                 .elevator_may_queue_fn =        cfq_may_queue,
2132                 .elevator_init_fn =             cfq_init_queue,
2133                 .elevator_exit_fn =             cfq_exit_queue,
2134                 .trim =                         cfq_free_io_context,
2135         },
2136         .elevator_attrs =       cfq_attrs,
2137         .elevator_name =        "cfq",
2138         .elevator_owner =       THIS_MODULE,
2139 };
2140
2141 static int __init cfq_init(void)
2142 {
2143         int ret;
2144
2145         /*
2146          * could be 0 on HZ < 1000 setups
2147          */
2148         if (!cfq_slice_async)
2149                 cfq_slice_async = 1;
2150         if (!cfq_slice_idle)
2151                 cfq_slice_idle = 1;
2152
2153         if (cfq_slab_setup())
2154                 return -ENOMEM;
2155
2156         ret = elv_register(&iosched_cfq);
2157         if (ret)
2158                 cfq_slab_kill();
2159
2160         return ret;
2161 }
2162
2163 static void __exit cfq_exit(void)
2164 {
2165         DECLARE_COMPLETION(all_gone);
2166         elv_unregister(&iosched_cfq);
2167         ioc_gone = &all_gone;
2168         /* ioc_gone's update must be visible before reading ioc_count */
2169         smp_wmb();
2170         if (elv_ioc_count_read(ioc_count))
2171                 wait_for_completion(ioc_gone);
2172         synchronize_rcu();
2173         cfq_slab_kill();
2174 }
2175
2176 module_init(cfq_init);
2177 module_exit(cfq_exit);
2178
2179 MODULE_AUTHOR("Jens Axboe");
2180 MODULE_LICENSE("GPL");
2181 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");