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 static DEFINE_SPINLOCK(cfq_exit_lock);
  35
  36 /*
  37  * for the hash of cfqq inside the cfqd
  38  */
  39 #define CFQ_QHASH_SHIFT         6
  40 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
  41 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
  42
  43 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
  44
  45 #define RQ_CIC(rq)              ((struct cfq_io_context*)(rq)->elevator_private)
  46 #define RQ_CFQQ(rq)             ((rq)->elevator_private2)
  47
  48 static kmem_cache_t *cfq_pool;
  49 static kmem_cache_t *cfq_ioc_pool;
  50
  51 static atomic_t ioc_count = ATOMIC_INIT(0);
  52 static struct completion *ioc_gone;
  53
  54 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  55 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  56 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  57
  58 #define ASYNC                   (0)
  59 #define SYNC                    (1)
  60
  61 #define cfq_cfqq_dispatched(cfqq)       \
  62         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
  63
  64 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
  65
  66 #define cfq_cfqq_sync(cfqq)             \
  67         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
  68
  69 #define sample_valid(samples)   ((samples) > 80)
  70
  71 /*
  72  * Per block device queue structure
  73  */
  74 struct cfq_data {
  75         request_queue_t *queue;
  76
  77         /*
  78          * rr list of queues with requests and the count of them
  79          */
  80         struct list_head rr_list[CFQ_PRIO_LISTS];
  81         struct list_head busy_rr;
  82         struct list_head cur_rr;
  83         struct list_head idle_rr;
  84         unsigned int busy_queues;
  85
  86         /*
  87          * non-ordered list of empty cfqq's
  88          */
  89         struct list_head empty_list;
  90
  91         /*
  92          * cfqq lookup hash
  93          */
  94         struct hlist_head *cfq_hash;
  95
  96         int rq_in_driver;
  97         int hw_tag;
  98
  99         /*
 100          * idle window management
 101          */
 102         struct timer_list idle_slice_timer;
 103         struct work_struct unplug_work;
 104
 105         struct cfq_queue *active_queue;
 106         struct cfq_io_context *active_cic;
 107         int cur_prio, cur_end_prio;
 108         unsigned int dispatch_slice;
 109
 110         struct timer_list idle_class_timer;
 111
 112         sector_t last_sector;
 113         unsigned long last_end_request;
 114
 115         /*
 116          * tunables, see top of file
 117          */
 118         unsigned int cfq_quantum;
 119         unsigned int cfq_fifo_expire[2];
 120         unsigned int cfq_back_penalty;
 121         unsigned int cfq_back_max;
 122         unsigned int cfq_slice[2];
 123         unsigned int cfq_slice_async_rq;
 124         unsigned int cfq_slice_idle;
 125
 126         struct list_head cic_list;
 127 };
 128
 129 /*
 130  * Per process-grouping structure
 131  */
 132 struct cfq_queue {
 133         /* reference count */
 134         atomic_t ref;
 135         /* parent cfq_data */
 136         struct cfq_data *cfqd;
 137         /* cfqq lookup hash */
 138         struct hlist_node cfq_hash;
 139         /* hash key */
 140         unsigned int key;
 141         /* on either rr or empty list of cfqd */
 142         struct list_head cfq_list;
 143         /* sorted list of pending requests */
 144         struct rb_root sort_list;
 145         /* if fifo isn't expired, next request to serve */
 146         struct request *next_rq;
 147         /* requests queued in sort_list */
 148         int queued[2];
 149         /* currently allocated requests */
 150         int allocated[2];
 151         /* fifo list of requests in sort_list */
 152         struct list_head fifo;
 153
 154         unsigned long slice_start;
 155         unsigned long slice_end;
 156         unsigned long slice_left;
 157         unsigned long service_last;
 158
 159         /* number of requests that are on the dispatch list */
 160         int on_dispatch[2];
 161
 162         /* io prio of this group */
 163         unsigned short ioprio, org_ioprio;
 164         unsigned short ioprio_class, org_ioprio_class;
 165
 166         /* various state flags, see below */
 167         unsigned int flags;
 168 };
 169
 170 enum cfqq_state_flags {
 171         CFQ_CFQQ_FLAG_on_rr = 0,
 172         CFQ_CFQQ_FLAG_wait_request,
 173         CFQ_CFQQ_FLAG_must_alloc,
 174         CFQ_CFQQ_FLAG_must_alloc_slice,
 175         CFQ_CFQQ_FLAG_must_dispatch,
 176         CFQ_CFQQ_FLAG_fifo_expire,
 177         CFQ_CFQQ_FLAG_idle_window,
 178         CFQ_CFQQ_FLAG_prio_changed,
 179 };
 180
 181 #define CFQ_CFQQ_FNS(name)                                              \
 182 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 183 {                                                                       \
 184         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 185 }                                                                       \
 186 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 187 {                                                                       \
 188         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 189 }                                                                       \
 190 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 191 {                                                                       \
 192         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 193 }
 194
 195 CFQ_CFQQ_FNS(on_rr);
 196 CFQ_CFQQ_FNS(wait_request);
 197 CFQ_CFQQ_FNS(must_alloc);
 198 CFQ_CFQQ_FNS(must_alloc_slice);
 199 CFQ_CFQQ_FNS(must_dispatch);
 200 CFQ_CFQQ_FNS(fifo_expire);
 201 CFQ_CFQQ_FNS(idle_window);
 202 CFQ_CFQQ_FNS(prio_changed);
 203 #undef CFQ_CFQQ_FNS
 204
 205 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 206 static void cfq_dispatch_insert(request_queue_t *, struct request *);
 207 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
 208
 209 /*
 210  * scheduler run of queue, if there are requests pending and no one in the
 211  * driver that will restart queueing
 212  */
 213 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 214 {
 215         if (cfqd->busy_queues)
 216                 kblockd_schedule_work(&cfqd->unplug_work);
 217 }
 218
 219 static int cfq_queue_empty(request_queue_t *q)
 220 {
 221         struct cfq_data *cfqd = q->elevator->elevator_data;
 222
 223         return !cfqd->busy_queues;
 224 }
 225
 226 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
 227 {
 228         if (rw == READ || rw == WRITE_SYNC)
 229                 return task->pid;
 230
 231         return CFQ_KEY_ASYNC;
 232 }
 233
 234 /*
 235  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 236  * We choose the request that is closest to the head right now. Distance
 237  * behind the head is penalized and only allowed to a certain extent.
 238  */
 239 static struct request *
 240 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
 241 {
 242         sector_t last, s1, s2, d1 = 0, d2 = 0;
 243         unsigned long back_max;
 244 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 245 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 246         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 247
 248         if (rq1 == NULL || rq1 == rq2)
 249                 return rq2;
 250         if (rq2 == NULL)
 251                 return rq1;
 252
 253         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 254                 return rq1;
 255         else if (rq_is_sync(rq2) && !rq_is_sync(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, *entry;
 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 queue was preempted, just add to front to be fair. busy_rr
 381          * isn't sorted, but insert at the back for fairness.
 382          */
 383         if (preempted || list == &cfqd->busy_rr) {
 384                 if (preempted)
 385                         list = list->prev;
 386
 387                 list_add_tail(&cfqq->cfq_list, list);
 388                 return;
 389         }
 390
 391         /*
 392          * sort by when queue was last serviced
 393          */
 394         entry = list;
 395         while ((entry = entry->prev) != list) {
 396                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
 397
 398                 if (!__cfqq->service_last)
 399                         break;
 400                 if (time_before(__cfqq->service_last, cfqq->service_last))
 401                         break;
 402         }
 403
 404         list_add(&cfqq->cfq_list, entry);
 405 }
 406
 407 /*
 408  * add to busy list of queues for service, trying to be fair in ordering
 409  * the pending list according to last request service
 410  */
 411 static inline void
 412 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 413 {
 414         BUG_ON(cfq_cfqq_on_rr(cfqq));
 415         cfq_mark_cfqq_on_rr(cfqq);
 416         cfqd->busy_queues++;
 417
 418         cfq_resort_rr_list(cfqq, 0);
 419 }
 420
 421 static inline void
 422 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 423 {
 424         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 425         cfq_clear_cfqq_on_rr(cfqq);
 426         list_move(&cfqq->cfq_list, &cfqd->empty_list);
 427
 428         BUG_ON(!cfqd->busy_queues);
 429         cfqd->busy_queues--;
 430 }
 431
 432 /*
 433  * rb tree support functions
 434  */
 435 static inline void cfq_del_rq_rb(struct request *rq)
 436 {
 437         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 438         struct cfq_data *cfqd = cfqq->cfqd;
 439         const int sync = rq_is_sync(rq);
 440
 441         BUG_ON(!cfqq->queued[sync]);
 442         cfqq->queued[sync]--;
 443
 444         elv_rb_del(&cfqq->sort_list, rq);
 445
 446         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
 447                 cfq_del_cfqq_rr(cfqd, cfqq);
 448 }
 449
 450 static void cfq_add_rq_rb(struct request *rq)
 451 {
 452         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 453         struct cfq_data *cfqd = cfqq->cfqd;
 454         struct request *__alias;
 455
 456         cfqq->queued[rq_is_sync(rq)]++;
 457
 458         /*
 459          * looks a little odd, but the first insert might return an alias.
 460          * if that happens, put the alias on the dispatch list
 461          */
 462         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
 463                 cfq_dispatch_insert(cfqd->queue, __alias);
 464 }
 465
 466 static inline void
 467 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
 468 {
 469         elv_rb_del(&cfqq->sort_list, rq);
 470         cfqq->queued[rq_is_sync(rq)]--;
 471         cfq_add_rq_rb(rq);
 472 }
 473
 474 static struct request *
 475 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 476 {
 477         struct task_struct *tsk = current;
 478         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
 479         struct cfq_queue *cfqq;
 480
 481         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
 482         if (cfqq) {
 483                 sector_t sector = bio->bi_sector + bio_sectors(bio);
 484
 485                 return elv_rb_find(&cfqq->sort_list, sector);
 486         }
 487
 488         return NULL;
 489 }
 490
 491 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 492 {
 493         struct cfq_data *cfqd = q->elevator->elevator_data;
 494
 495         cfqd->rq_in_driver++;
 496
 497         /*
 498          * If the depth is larger 1, it really could be queueing. But lets
 499          * make the mark a little higher - idling could still be good for
 500          * low queueing, and a low queueing number could also just indicate
 501          * a SCSI mid layer like behaviour where limit+1 is often seen.
 502          */
 503         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
 504                 cfqd->hw_tag = 1;
 505 }
 506
 507 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 508 {
 509         struct cfq_data *cfqd = q->elevator->elevator_data;
 510
 511         WARN_ON(!cfqd->rq_in_driver);
 512         cfqd->rq_in_driver--;
 513 }
 514
 515 static void cfq_remove_request(struct request *rq)
 516 {
 517         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 518
 519         if (cfqq->next_rq == rq)
 520                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
 521
 522         list_del_init(&rq->queuelist);
 523         cfq_del_rq_rb(rq);
 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                 cfqq->service_last = now;
 598                 cfq_schedule_dispatch(cfqd);
 599         }
 600
 601         cfq_clear_cfqq_must_dispatch(cfqq);
 602         cfq_clear_cfqq_wait_request(cfqq);
 603
 604         /*
 605          * store what was left of this slice, if the queue idled out
 606          * or was preempted
 607          */
 608         if (time_after(cfqq->slice_end, now))
 609                 cfqq->slice_left = cfqq->slice_end - now;
 610         else
 611                 cfqq->slice_left = 0;
 612
 613         if (cfq_cfqq_on_rr(cfqq))
 614                 cfq_resort_rr_list(cfqq, preempted);
 615
 616         if (cfqq == cfqd->active_queue)
 617                 cfqd->active_queue = NULL;
 618
 619         if (cfqd->active_cic) {
 620                 put_io_context(cfqd->active_cic->ioc);
 621                 cfqd->active_cic = NULL;
 622         }
 623
 624         cfqd->dispatch_slice = 0;
 625 }
 626
 627 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 628 {
 629         struct cfq_queue *cfqq = cfqd->active_queue;
 630
 631         if (cfqq)
 632                 __cfq_slice_expired(cfqd, cfqq, preempted);
 633 }
 634
 635 /*
 636  * 0
 637  * 0,1
 638  * 0,1,2
 639  * 0,1,2,3
 640  * 0,1,2,3,4
 641  * 0,1,2,3,4,5
 642  * 0,1,2,3,4,5,6
 643  * 0,1,2,3,4,5,6,7
 644  */
 645 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 646 {
 647         int prio, wrap;
 648
 649         prio = -1;
 650         wrap = 0;
 651         do {
 652                 int p;
 653
 654                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 655                         if (!list_empty(&cfqd->rr_list[p])) {
 656                                 prio = p;
 657                                 break;
 658                         }
 659                 }
 660
 661                 if (prio != -1)
 662                         break;
 663                 cfqd->cur_prio = 0;
 664                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 665                         cfqd->cur_end_prio = 0;
 666                         if (wrap)
 667                                 break;
 668                         wrap = 1;
 669                 }
 670         } while (1);
 671
 672         if (unlikely(prio == -1))
 673                 return -1;
 674
 675         BUG_ON(prio >= CFQ_PRIO_LISTS);
 676
 677         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 678
 679         cfqd->cur_prio = prio + 1;
 680         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 681                 cfqd->cur_end_prio = cfqd->cur_prio;
 682                 cfqd->cur_prio = 0;
 683         }
 684         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 685                 cfqd->cur_prio = 0;
 686                 cfqd->cur_end_prio = 0;
 687         }
 688
 689         return prio;
 690 }
 691
 692 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 693 {
 694         struct cfq_queue *cfqq = NULL;
 695
 696         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
 697                 /*
 698                  * if current list is non-empty, grab first entry. if it is
 699                  * empty, get next prio level and grab first entry then if any
 700                  * are spliced
 701                  */
 702                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 703         } else if (!list_empty(&cfqd->busy_rr)) {
 704                 /*
 705                  * If no new queues are available, check if the busy list has
 706                  * some before falling back to idle io.
 707                  */
 708                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
 709         } else if (!list_empty(&cfqd->idle_rr)) {
 710                 /*
 711                  * if we have idle queues and no rt or be queues had pending
 712                  * requests, either allow immediate service if the grace period
 713                  * has passed or arm the idle grace timer
 714                  */
 715                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 716
 717                 if (time_after_eq(jiffies, end))
 718                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 719                 else
 720                         mod_timer(&cfqd->idle_class_timer, end);
 721         }
 722
 723         __cfq_set_active_queue(cfqd, cfqq);
 724         return cfqq;
 725 }
 726
 727 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
 728
 729 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 730
 731 {
 732         struct cfq_io_context *cic;
 733         unsigned long sl;
 734
 735         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
 736         WARN_ON(cfqq != cfqd->active_queue);
 737
 738         /*
 739          * idle is disabled, either manually or by past process history
 740          */
 741         if (!cfqd->cfq_slice_idle)
 742                 return 0;
 743         if (!cfq_cfqq_idle_window(cfqq))
 744                 return 0;
 745         /*
 746          * task has exited, don't wait
 747          */
 748         cic = cfqd->active_cic;
 749         if (!cic || !cic->ioc->task)
 750                 return 0;
 751
 752         cfq_mark_cfqq_must_dispatch(cfqq);
 753         cfq_mark_cfqq_wait_request(cfqq);
 754
 755         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 756
 757         /*
 758          * we don't want to idle for seeks, but we do want to allow
 759          * fair distribution of slice time for a process doing back-to-back
 760          * seeks. so allow a little bit of time for him to submit a new rq
 761          */
 762         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
 763                 sl = min(sl, msecs_to_jiffies(2));
 764
 765         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 766         return 1;
 767 }
 768
 769 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
 770 {
 771         struct cfq_data *cfqd = q->elevator->elevator_data;
 772         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 773
 774         cfq_remove_request(rq);
 775         cfqq->on_dispatch[rq_is_sync(rq)]++;
 776         elv_dispatch_sort(q, rq);
 777
 778         rq = list_entry(q->queue_head.prev, struct request, queuelist);
 779         cfqd->last_sector = rq->sector + rq->nr_sectors;
 780 }
 781
 782 /*
 783  * return expired entry, or NULL to just start from scratch in rbtree
 784  */
 785 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
 786 {
 787         struct cfq_data *cfqd = cfqq->cfqd;
 788         struct request *rq;
 789         int fifo;
 790
 791         if (cfq_cfqq_fifo_expire(cfqq))
 792                 return NULL;
 793         if (list_empty(&cfqq->fifo))
 794                 return NULL;
 795
 796         fifo = cfq_cfqq_class_sync(cfqq);
 797         rq = rq_entry_fifo(cfqq->fifo.next);
 798
 799         if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
 800                 cfq_mark_cfqq_fifo_expire(cfqq);
 801                 return rq;
 802         }
 803
 804         return NULL;
 805 }
 806
 807 /*
 808  * Scale schedule slice based on io priority. Use the sync time slice only
 809  * if a queue is marked sync and has sync io queued. A sync queue with async
 810  * io only, should not get full sync slice length.
 811  */
 812 static inline int
 813 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 814 {
 815         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 816
 817         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 818
 819         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
 820 }
 821
 822 static inline void
 823 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 824 {
 825         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 826 }
 827
 828 static inline int
 829 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 830 {
 831         const int base_rq = cfqd->cfq_slice_async_rq;
 832
 833         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 834
 835         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
 836 }
 837
 838 /*
 839  * get next queue for service
 840  */
 841 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
 842 {
 843         unsigned long now = jiffies;
 844         struct cfq_queue *cfqq;
 845
 846         cfqq = cfqd->active_queue;
 847         if (!cfqq)
 848                 goto new_queue;
 849
 850         /*
 851          * slice has expired
 852          */
 853         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
 854                 goto expire;
 855
 856         /*
 857          * if queue has requests, dispatch one. if not, check if
 858          * enough slice is left to wait for one
 859          */
 860         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
 861                 goto keep_queue;
 862         else if (cfq_cfqq_dispatched(cfqq)) {
 863                 cfqq = NULL;
 864                 goto keep_queue;
 865         } else if (cfq_cfqq_class_sync(cfqq)) {
 866                 if (cfq_arm_slice_timer(cfqd, cfqq))
 867                         return NULL;
 868         }
 869
 870 expire:
 871         cfq_slice_expired(cfqd, 0);
 872 new_queue:
 873         cfqq = cfq_set_active_queue(cfqd);
 874 keep_queue:
 875         return cfqq;
 876 }
 877
 878 static int
 879 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 880                         int max_dispatch)
 881 {
 882         int dispatched = 0;
 883
 884         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
 885
 886         do {
 887                 struct request *rq;
 888
 889                 /*
 890                  * follow expired path, else get first next available
 891                  */
 892                 if ((rq = cfq_check_fifo(cfqq)) == NULL)
 893                         rq = cfqq->next_rq;
 894
 895                 /*
 896                  * finally, insert request into driver dispatch list
 897                  */
 898                 cfq_dispatch_insert(cfqd->queue, rq);
 899
 900                 cfqd->dispatch_slice++;
 901                 dispatched++;
 902
 903                 if (!cfqd->active_cic) {
 904                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
 905                         cfqd->active_cic = RQ_CIC(rq);
 906                 }
 907
 908                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
 909                         break;
 910
 911         } while (dispatched < max_dispatch);
 912
 913         /*
 914          * if slice end isn't set yet, set it.
 915          */
 916         if (!cfqq->slice_end)
 917                 cfq_set_prio_slice(cfqd, cfqq);
 918
 919         /*
 920          * expire an async queue immediately if it has used up its slice. idle
 921          * queue always expire after 1 dispatch round.
 922          */
 923         if ((!cfq_cfqq_sync(cfqq) &&
 924             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
 925             cfq_class_idle(cfqq) ||
 926             !cfq_cfqq_idle_window(cfqq))
 927                 cfq_slice_expired(cfqd, 0);
 928
 929         return dispatched;
 930 }
 931
 932 static int
 933 cfq_forced_dispatch_cfqqs(struct list_head *list)
 934 {
 935         struct cfq_queue *cfqq, *next;
 936         int dispatched;
 937
 938         dispatched = 0;
 939         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
 940                 while (cfqq->next_rq) {
 941                         cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
 942                         dispatched++;
 943                 }
 944                 BUG_ON(!list_empty(&cfqq->fifo));
 945         }
 946
 947         return dispatched;
 948 }
 949
 950 static int
 951 cfq_forced_dispatch(struct cfq_data *cfqd)
 952 {
 953         int i, dispatched = 0;
 954
 955         for (i = 0; i < CFQ_PRIO_LISTS; i++)
 956                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
 957
 958         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
 959         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
 960         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
 961
 962         cfq_slice_expired(cfqd, 0);
 963
 964         BUG_ON(cfqd->busy_queues);
 965
 966         return dispatched;
 967 }
 968
 969 static int
 970 cfq_dispatch_requests(request_queue_t *q, int force)
 971 {
 972         struct cfq_data *cfqd = q->elevator->elevator_data;
 973         struct cfq_queue *cfqq, *prev_cfqq;
 974         int dispatched;
 975
 976         if (!cfqd->busy_queues)
 977                 return 0;
 978
 979         if (unlikely(force))
 980                 return cfq_forced_dispatch(cfqd);
 981
 982         dispatched = 0;
 983         prev_cfqq = NULL;
 984         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
 985                 int max_dispatch;
 986
 987                 /*
 988                  * Don't repeat dispatch from the previous queue.
 989                  */
 990                 if (prev_cfqq == cfqq)
 991                         break;
 992
 993                 cfq_clear_cfqq_must_dispatch(cfqq);
 994                 cfq_clear_cfqq_wait_request(cfqq);
 995                 del_timer(&cfqd->idle_slice_timer);
 996
 997                 max_dispatch = cfqd->cfq_quantum;
 998                 if (cfq_class_idle(cfqq))
 999                         max_dispatch = 1;
1000
1001                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1002
1003                 /*
1004                  * If the dispatch cfqq has idling enabled and is still
1005                  * the active queue, break out.
1006                  */
1007                 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1008                         break;
1009
1010                 prev_cfqq = cfqq;
1011         }
1012
1013         return dispatched;
1014 }
1015
1016 /*
1017  * task holds one reference to the queue, dropped when task exits. each rq
1018  * in-flight on this queue also holds a reference, dropped when rq is freed.
1019  *
1020  * queue lock must be held here.
1021  */
1022 static void cfq_put_queue(struct cfq_queue *cfqq)
1023 {
1024         struct cfq_data *cfqd = cfqq->cfqd;
1025
1026         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1027
1028         if (!atomic_dec_and_test(&cfqq->ref))
1029                 return;
1030
1031         BUG_ON(rb_first(&cfqq->sort_list));
1032         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1033         BUG_ON(cfq_cfqq_on_rr(cfqq));
1034
1035         if (unlikely(cfqd->active_queue == cfqq))
1036                 __cfq_slice_expired(cfqd, cfqq, 0);
1037
1038         /*
1039          * it's on the empty list and still hashed
1040          */
1041         list_del(&cfqq->cfq_list);
1042         hlist_del(&cfqq->cfq_hash);
1043         kmem_cache_free(cfq_pool, cfqq);
1044 }
1045
1046 static inline struct cfq_queue *
1047 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1048                     const int hashval)
1049 {
1050         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1051         struct hlist_node *entry;
1052         struct cfq_queue *__cfqq;
1053
1054         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1055                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1056
1057                 if (__cfqq->key == key && (__p == prio || !prio))
1058                         return __cfqq;
1059         }
1060
1061         return NULL;
1062 }
1063
1064 static struct cfq_queue *
1065 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1066 {
1067         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1068 }
1069
1070 static void cfq_free_io_context(struct io_context *ioc)
1071 {
1072         struct cfq_io_context *__cic;
1073         struct rb_node *n;
1074         int freed = 0;
1075
1076         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1077                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1078                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1079                 kmem_cache_free(cfq_ioc_pool, __cic);
1080                 freed++;
1081         }
1082
1083         if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1084                 complete(ioc_gone);
1085 }
1086
1087 static void cfq_trim(struct io_context *ioc)
1088 {
1089         ioc->set_ioprio = NULL;
1090         cfq_free_io_context(ioc);
1091 }
1092
1093 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1094 {
1095         if (unlikely(cfqq == cfqd->active_queue))
1096                 __cfq_slice_expired(cfqd, cfqq, 0);
1097
1098         cfq_put_queue(cfqq);
1099 }
1100
1101 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1102                                          struct cfq_io_context *cic)
1103 {
1104         if (cic->cfqq[ASYNC]) {
1105                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1106                 cic->cfqq[ASYNC] = NULL;
1107         }
1108
1109         if (cic->cfqq[SYNC]) {
1110                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1111                 cic->cfqq[SYNC] = NULL;
1112         }
1113
1114         cic->key = NULL;
1115         list_del_init(&cic->queue_list);
1116 }
1117
1118
1119 /*
1120  * Called with interrupts disabled
1121  */
1122 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1123 {
1124         struct cfq_data *cfqd = cic->key;
1125
1126         WARN_ON(!irqs_disabled());
1127
1128         if (cfqd) {
1129                 request_queue_t *q = cfqd->queue;
1130
1131                 spin_lock(q->queue_lock);
1132                 __cfq_exit_single_io_context(cfqd, cic);
1133                 spin_unlock(q->queue_lock);
1134         }
1135 }
1136
1137 static void cfq_exit_io_context(struct io_context *ioc)
1138 {
1139         struct cfq_io_context *__cic;
1140         unsigned long flags;
1141         struct rb_node *n;
1142
1143         /*
1144          * put the reference this task is holding to the various queues
1145          */
1146         spin_lock_irqsave(&cfq_exit_lock, flags);
1147
1148         n = rb_first(&ioc->cic_root);
1149         while (n != NULL) {
1150                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1151
1152                 cfq_exit_single_io_context(__cic);
1153                 n = rb_next(n);
1154         }
1155
1156         spin_unlock_irqrestore(&cfq_exit_lock, flags);
1157 }
1158
1159 static struct cfq_io_context *
1160 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1161 {
1162         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1163
1164         if (cic) {
1165                 memset(cic, 0, sizeof(*cic));
1166                 cic->last_end_request = jiffies;
1167                 INIT_LIST_HEAD(&cic->queue_list);
1168                 cic->dtor = cfq_free_io_context;
1169                 cic->exit = cfq_exit_io_context;
1170                 atomic_inc(&ioc_count);
1171         }
1172
1173         return cic;
1174 }
1175
1176 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1177 {
1178         struct task_struct *tsk = current;
1179         int ioprio_class;
1180
1181         if (!cfq_cfqq_prio_changed(cfqq))
1182                 return;
1183
1184         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1185         switch (ioprio_class) {
1186                 default:
1187                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1188                 case IOPRIO_CLASS_NONE:
1189                         /*
1190                          * no prio set, place us in the middle of the BE classes
1191                          */
1192                         cfqq->ioprio = task_nice_ioprio(tsk);
1193                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1194                         break;
1195                 case IOPRIO_CLASS_RT:
1196                         cfqq->ioprio = task_ioprio(tsk);
1197                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1198                         break;
1199                 case IOPRIO_CLASS_BE:
1200                         cfqq->ioprio = task_ioprio(tsk);
1201                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1202                         break;
1203                 case IOPRIO_CLASS_IDLE:
1204                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1205                         cfqq->ioprio = 7;
1206                         cfq_clear_cfqq_idle_window(cfqq);
1207                         break;
1208         }
1209
1210         /*
1211          * keep track of original prio settings in case we have to temporarily
1212          * elevate the priority of this queue
1213          */
1214         cfqq->org_ioprio = cfqq->ioprio;
1215         cfqq->org_ioprio_class = cfqq->ioprio_class;
1216
1217         if (cfq_cfqq_on_rr(cfqq))
1218                 cfq_resort_rr_list(cfqq, 0);
1219
1220         cfq_clear_cfqq_prio_changed(cfqq);
1221 }
1222
1223 static inline void changed_ioprio(struct cfq_io_context *cic)
1224 {
1225         struct cfq_data *cfqd = cic->key;
1226         struct cfq_queue *cfqq;
1227
1228         if (unlikely(!cfqd))
1229                 return;
1230
1231         spin_lock(cfqd->queue->queue_lock);
1232
1233         cfqq = cic->cfqq[ASYNC];
1234         if (cfqq) {
1235                 struct cfq_queue *new_cfqq;
1236                 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1237                                          GFP_ATOMIC);
1238                 if (new_cfqq) {
1239                         cic->cfqq[ASYNC] = new_cfqq;
1240                         cfq_put_queue(cfqq);
1241                 }
1242         }
1243
1244         cfqq = cic->cfqq[SYNC];
1245         if (cfqq)
1246                 cfq_mark_cfqq_prio_changed(cfqq);
1247
1248         spin_unlock(cfqd->queue->queue_lock);
1249 }
1250
1251 /*
1252  * callback from sys_ioprio_set, irqs are disabled
1253  */
1254 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1255 {
1256         struct cfq_io_context *cic;
1257         struct rb_node *n;
1258
1259         spin_lock(&cfq_exit_lock);
1260
1261         n = rb_first(&ioc->cic_root);
1262         while (n != NULL) {
1263                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1264
1265                 changed_ioprio(cic);
1266                 n = rb_next(n);
1267         }
1268
1269         spin_unlock(&cfq_exit_lock);
1270
1271         return 0;
1272 }
1273
1274 static struct cfq_queue *
1275 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1276               gfp_t gfp_mask)
1277 {
1278         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1279         struct cfq_queue *cfqq, *new_cfqq = NULL;
1280         unsigned short ioprio;
1281
1282 retry:
1283         ioprio = tsk->ioprio;
1284         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1285
1286         if (!cfqq) {
1287                 if (new_cfqq) {
1288                         cfqq = new_cfqq;
1289                         new_cfqq = NULL;
1290                 } else if (gfp_mask & __GFP_WAIT) {
1291                         /*
1292                          * Inform the allocator of the fact that we will
1293                          * just repeat this allocation if it fails, to allow
1294                          * the allocator to do whatever it needs to attempt to
1295                          * free memory.
1296                          */
1297                         spin_unlock_irq(cfqd->queue->queue_lock);
1298                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask|__GFP_NOFAIL);
1299                         spin_lock_irq(cfqd->queue->queue_lock);
1300                         goto retry;
1301                 } else {
1302                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1303                         if (!cfqq)
1304                                 goto out;
1305                 }
1306
1307                 memset(cfqq, 0, sizeof(*cfqq));
1308
1309                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1310                 INIT_LIST_HEAD(&cfqq->cfq_list);
1311                 INIT_LIST_HEAD(&cfqq->fifo);
1312
1313                 cfqq->key = key;
1314                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1315                 atomic_set(&cfqq->ref, 0);
1316                 cfqq->cfqd = cfqd;
1317                 cfqq->service_last = 0;
1318                 /*
1319                  * set ->slice_left to allow preemption for a new process
1320                  */
1321                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1322                 cfq_mark_cfqq_idle_window(cfqq);
1323                 cfq_mark_cfqq_prio_changed(cfqq);
1324                 cfq_init_prio_data(cfqq);
1325         }
1326
1327         if (new_cfqq)
1328                 kmem_cache_free(cfq_pool, new_cfqq);
1329
1330         atomic_inc(&cfqq->ref);
1331 out:
1332         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1333         return cfqq;
1334 }
1335
1336 static void
1337 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1338 {
1339         spin_lock(&cfq_exit_lock);
1340         rb_erase(&cic->rb_node, &ioc->cic_root);
1341         list_del_init(&cic->queue_list);
1342         spin_unlock(&cfq_exit_lock);
1343         kmem_cache_free(cfq_ioc_pool, cic);
1344         atomic_dec(&ioc_count);
1345 }
1346
1347 static struct cfq_io_context *
1348 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1349 {
1350         struct rb_node *n;
1351         struct cfq_io_context *cic;
1352         void *k, *key = cfqd;
1353
1354 restart:
1355         n = ioc->cic_root.rb_node;
1356         while (n) {
1357                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1358                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1359                 k = cic->key;
1360                 if (unlikely(!k)) {
1361                         cfq_drop_dead_cic(ioc, cic);
1362                         goto restart;
1363                 }
1364
1365                 if (key < k)
1366                         n = n->rb_left;
1367                 else if (key > k)
1368                         n = n->rb_right;
1369                 else
1370                         return cic;
1371         }
1372
1373         return NULL;
1374 }
1375
1376 static inline void
1377 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1378              struct cfq_io_context *cic)
1379 {
1380         struct rb_node **p;
1381         struct rb_node *parent;
1382         struct cfq_io_context *__cic;
1383         void *k;
1384
1385         cic->ioc = ioc;
1386         cic->key = cfqd;
1387
1388         ioc->set_ioprio = cfq_ioc_set_ioprio;
1389 restart:
1390         parent = NULL;
1391         p = &ioc->cic_root.rb_node;
1392         while (*p) {
1393                 parent = *p;
1394                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1395                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1396                 k = __cic->key;
1397                 if (unlikely(!k)) {
1398                         cfq_drop_dead_cic(ioc, __cic);
1399                         goto restart;
1400                 }
1401
1402                 if (cic->key < k)
1403                         p = &(*p)->rb_left;
1404                 else if (cic->key > k)
1405                         p = &(*p)->rb_right;
1406                 else
1407                         BUG();
1408         }
1409
1410         spin_lock(&cfq_exit_lock);
1411         rb_link_node(&cic->rb_node, parent, p);
1412         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1413         list_add(&cic->queue_list, &cfqd->cic_list);
1414         spin_unlock(&cfq_exit_lock);
1415 }
1416
1417 /*
1418  * Setup general io context and cfq io context. There can be several cfq
1419  * io contexts per general io context, if this process is doing io to more
1420  * than one device managed by cfq.
1421  */
1422 static struct cfq_io_context *
1423 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1424 {
1425         struct io_context *ioc = NULL;
1426         struct cfq_io_context *cic;
1427
1428         might_sleep_if(gfp_mask & __GFP_WAIT);
1429
1430         ioc = get_io_context(gfp_mask);
1431         if (!ioc)
1432                 return NULL;
1433
1434         cic = cfq_cic_rb_lookup(cfqd, ioc);
1435         if (cic)
1436                 goto out;
1437
1438         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1439         if (cic == NULL)
1440                 goto err;
1441
1442         cfq_cic_link(cfqd, ioc, cic);
1443 out:
1444         return cic;
1445 err:
1446         put_io_context(ioc);
1447         return NULL;
1448 }
1449
1450 static void
1451 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1452 {
1453         unsigned long elapsed, ttime;
1454
1455         /*
1456          * if this context already has stuff queued, thinktime is from
1457          * last queue not last end
1458          */
1459 #if 0
1460         if (time_after(cic->last_end_request, cic->last_queue))
1461                 elapsed = jiffies - cic->last_end_request;
1462         else
1463                 elapsed = jiffies - cic->last_queue;
1464 #else
1465                 elapsed = jiffies - cic->last_end_request;
1466 #endif
1467
1468         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1469
1470         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1471         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1472         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1473 }
1474
1475 static void
1476 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1477                        struct request *rq)
1478 {
1479         sector_t sdist;
1480         u64 total;
1481
1482         if (cic->last_request_pos < rq->sector)
1483                 sdist = rq->sector - cic->last_request_pos;
1484         else
1485                 sdist = cic->last_request_pos - rq->sector;
1486
1487         /*
1488          * Don't allow the seek distance to get too large from the
1489          * odd fragment, pagein, etc
1490          */
1491         if (cic->seek_samples <= 60) /* second&third seek */
1492                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1493         else
1494                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1495
1496         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1497         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1498         total = cic->seek_total + (cic->seek_samples/2);
1499         do_div(total, cic->seek_samples);
1500         cic->seek_mean = (sector_t)total;
1501 }
1502
1503 /*
1504  * Disable idle window if the process thinks too long or seeks so much that
1505  * it doesn't matter
1506  */
1507 static void
1508 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1509                        struct cfq_io_context *cic)
1510 {
1511         int enable_idle = cfq_cfqq_idle_window(cfqq);
1512
1513         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1514             (cfqd->hw_tag && CIC_SEEKY(cic)))
1515                 enable_idle = 0;
1516         else if (sample_valid(cic->ttime_samples)) {
1517                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1518                         enable_idle = 0;
1519                 else
1520                         enable_idle = 1;
1521         }
1522
1523         if (enable_idle)
1524                 cfq_mark_cfqq_idle_window(cfqq);
1525         else
1526                 cfq_clear_cfqq_idle_window(cfqq);
1527 }
1528
1529
1530 /*
1531  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1532  * no or if we aren't sure, a 1 will cause a preempt.
1533  */
1534 static int
1535 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1536                    struct request *rq)
1537 {
1538         struct cfq_queue *cfqq = cfqd->active_queue;
1539
1540         if (cfq_class_idle(new_cfqq))
1541                 return 0;
1542
1543         if (!cfqq)
1544                 return 0;
1545
1546         if (cfq_class_idle(cfqq))
1547                 return 1;
1548         if (!cfq_cfqq_wait_request(new_cfqq))
1549                 return 0;
1550         /*
1551          * if it doesn't have slice left, forget it
1552          */
1553         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1554                 return 0;
1555         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1556                 return 1;
1557
1558         return 0;
1559 }
1560
1561 /*
1562  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1563  * let it have half of its nominal slice.
1564  */
1565 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1566 {
1567         struct cfq_queue *__cfqq, *next;
1568
1569         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1570                 cfq_resort_rr_list(__cfqq, 1);
1571
1572         if (!cfqq->slice_left)
1573                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1574
1575         cfqq->slice_end = cfqq->slice_left + jiffies;
1576         cfq_slice_expired(cfqd, 1);
1577         __cfq_set_active_queue(cfqd, cfqq);
1578 }
1579
1580 /*
1581  * should really be a ll_rw_blk.c helper
1582  */
1583 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1584 {
1585         request_queue_t *q = cfqd->queue;
1586
1587         if (!blk_queue_plugged(q))
1588                 q->request_fn(q);
1589         else
1590                 __generic_unplug_device(q);
1591 }
1592
1593 /*
1594  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1595  * something we should do about it
1596  */
1597 static void
1598 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1599                 struct request *rq)
1600 {
1601         struct cfq_io_context *cic = RQ_CIC(rq);
1602
1603         /*
1604          * check if this request is a better next-serve candidate)) {
1605          */
1606         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1607         BUG_ON(!cfqq->next_rq);
1608
1609         /*
1610          * we never wait for an async request and we don't allow preemption
1611          * of an async request. so just return early
1612          */
1613         if (!rq_is_sync(rq)) {
1614                 /*
1615                  * sync process issued an async request, if it's waiting
1616                  * then expire it and kick rq handling.
1617                  */
1618                 if (cic == cfqd->active_cic &&
1619                     del_timer(&cfqd->idle_slice_timer)) {
1620                         cfq_slice_expired(cfqd, 0);
1621                         cfq_start_queueing(cfqd, cfqq);
1622                 }
1623                 return;
1624         }
1625
1626         cfq_update_io_thinktime(cfqd, cic);
1627         cfq_update_io_seektime(cfqd, cic, rq);
1628         cfq_update_idle_window(cfqd, cfqq, cic);
1629
1630         cic->last_queue = jiffies;
1631         cic->last_request_pos = rq->sector + rq->nr_sectors;
1632
1633         if (cfqq == cfqd->active_queue) {
1634                 /*
1635                  * if we are waiting for a request for this queue, let it rip
1636                  * immediately and flag that we must not expire this queue
1637                  * just now
1638                  */
1639                 if (cfq_cfqq_wait_request(cfqq)) {
1640                         cfq_mark_cfqq_must_dispatch(cfqq);
1641                         del_timer(&cfqd->idle_slice_timer);
1642                         cfq_start_queueing(cfqd, cfqq);
1643                 }
1644         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1645                 /*
1646                  * not the active queue - expire current slice if it is
1647                  * idle and has expired it's mean thinktime or this new queue
1648                  * has some old slice time left and is of higher priority
1649                  */
1650                 cfq_preempt_queue(cfqd, cfqq);
1651                 cfq_mark_cfqq_must_dispatch(cfqq);
1652                 cfq_start_queueing(cfqd, cfqq);
1653         }
1654 }
1655
1656 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1657 {
1658         struct cfq_data *cfqd = q->elevator->elevator_data;
1659         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1660
1661         cfq_init_prio_data(cfqq);
1662
1663         cfq_add_rq_rb(rq);
1664
1665         if (!cfq_cfqq_on_rr(cfqq))
1666                 cfq_add_cfqq_rr(cfqd, cfqq);
1667
1668         list_add_tail(&rq->queuelist, &cfqq->fifo);
1669
1670         cfq_rq_enqueued(cfqd, cfqq, rq);
1671 }
1672
1673 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1674 {
1675         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1676         struct cfq_data *cfqd = cfqq->cfqd;
1677         const int sync = rq_is_sync(rq);
1678         unsigned long now;
1679
1680         now = jiffies;
1681
1682         WARN_ON(!cfqd->rq_in_driver);
1683         WARN_ON(!cfqq->on_dispatch[sync]);
1684         cfqd->rq_in_driver--;
1685         cfqq->on_dispatch[sync]--;
1686
1687         if (!cfq_class_idle(cfqq))
1688                 cfqd->last_end_request = now;
1689
1690         if (!cfq_cfqq_dispatched(cfqq)) {
1691                 if (cfq_cfqq_on_rr(cfqq)) {
1692                         cfqq->service_last = now;
1693                         cfq_resort_rr_list(cfqq, 0);
1694                 }
1695         }
1696
1697         if (sync)
1698                 RQ_CIC(rq)->last_end_request = now;
1699
1700         /*
1701          * If this is the active queue, check if it needs to be expired,
1702          * or if we want to idle in case it has no pending requests.
1703          */
1704         if (cfqd->active_queue == cfqq) {
1705                 if (time_after(now, cfqq->slice_end))
1706                         cfq_slice_expired(cfqd, 0);
1707                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1708                         if (!cfq_arm_slice_timer(cfqd, cfqq))
1709                                 cfq_schedule_dispatch(cfqd);
1710                 }
1711         }
1712 }
1713
1714 /*
1715  * we temporarily boost lower priority queues if they are holding fs exclusive
1716  * resources. they are boosted to normal prio (CLASS_BE/4)
1717  */
1718 static void cfq_prio_boost(struct cfq_queue *cfqq)
1719 {
1720         const int ioprio_class = cfqq->ioprio_class;
1721         const int ioprio = cfqq->ioprio;
1722
1723         if (has_fs_excl()) {
1724                 /*
1725                  * boost idle prio on transactions that would lock out other
1726                  * users of the filesystem
1727                  */
1728                 if (cfq_class_idle(cfqq))
1729                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1730                 if (cfqq->ioprio > IOPRIO_NORM)
1731                         cfqq->ioprio = IOPRIO_NORM;
1732         } else {
1733                 /*
1734                  * check if we need to unboost the queue
1735                  */
1736                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1737                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1738                 if (cfqq->ioprio != cfqq->org_ioprio)
1739                         cfqq->ioprio = cfqq->org_ioprio;
1740         }
1741
1742         /*
1743          * refile between round-robin lists if we moved the priority class
1744          */
1745         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1746             cfq_cfqq_on_rr(cfqq))
1747                 cfq_resort_rr_list(cfqq, 0);
1748 }
1749
1750 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1751 {
1752         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1753             !cfq_cfqq_must_alloc_slice(cfqq)) {
1754                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1755                 return ELV_MQUEUE_MUST;
1756         }
1757
1758         return ELV_MQUEUE_MAY;
1759 }
1760
1761 static int cfq_may_queue(request_queue_t *q, int rw)
1762 {
1763         struct cfq_data *cfqd = q->elevator->elevator_data;
1764         struct task_struct *tsk = current;
1765         struct cfq_queue *cfqq;
1766
1767         /*
1768          * don't force setup of a queue from here, as a call to may_queue
1769          * does not necessarily imply that a request actually will be queued.
1770          * so just lookup a possibly existing queue, or return 'may queue'
1771          * if that fails
1772          */
1773         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1774         if (cfqq) {
1775                 cfq_init_prio_data(cfqq);
1776                 cfq_prio_boost(cfqq);
1777
1778                 return __cfq_may_queue(cfqq);
1779         }
1780
1781         return ELV_MQUEUE_MAY;
1782 }
1783
1784 /*
1785  * queue lock held here
1786  */
1787 static void cfq_put_request(request_queue_t *q, struct request *rq)
1788 {
1789         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1790
1791         if (cfqq) {
1792                 const int rw = rq_data_dir(rq);
1793
1794                 BUG_ON(!cfqq->allocated[rw]);
1795                 cfqq->allocated[rw]--;
1796
1797                 put_io_context(RQ_CIC(rq)->ioc);
1798
1799                 rq->elevator_private = NULL;
1800                 rq->elevator_private2 = NULL;
1801
1802                 cfq_put_queue(cfqq);
1803         }
1804 }
1805
1806 /*
1807  * Allocate cfq data structures associated with this request.
1808  */
1809 static int
1810 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1811 {
1812         struct cfq_data *cfqd = q->elevator->elevator_data;
1813         struct task_struct *tsk = current;
1814         struct cfq_io_context *cic;
1815         const int rw = rq_data_dir(rq);
1816         pid_t key = cfq_queue_pid(tsk, rw);
1817         struct cfq_queue *cfqq;
1818         unsigned long flags;
1819         int is_sync = key != CFQ_KEY_ASYNC;
1820
1821         might_sleep_if(gfp_mask & __GFP_WAIT);
1822
1823         cic = cfq_get_io_context(cfqd, gfp_mask);
1824
1825         spin_lock_irqsave(q->queue_lock, flags);
1826
1827         if (!cic)
1828                 goto queue_fail;
1829
1830         if (!cic->cfqq[is_sync]) {
1831                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1832                 if (!cfqq)
1833                         goto queue_fail;
1834
1835                 cic->cfqq[is_sync] = cfqq;
1836         } else
1837                 cfqq = cic->cfqq[is_sync];
1838
1839         cfqq->allocated[rw]++;
1840         cfq_clear_cfqq_must_alloc(cfqq);
1841         atomic_inc(&cfqq->ref);
1842
1843         spin_unlock_irqrestore(q->queue_lock, flags);
1844
1845         rq->elevator_private = cic;
1846         rq->elevator_private2 = cfqq;
1847         return 0;
1848
1849 queue_fail:
1850         if (cic)
1851                 put_io_context(cic->ioc);
1852
1853         cfq_schedule_dispatch(cfqd);
1854         spin_unlock_irqrestore(q->queue_lock, flags);
1855         return 1;
1856 }
1857
1858 static void cfq_kick_queue(void *data)
1859 {
1860         request_queue_t *q = data;
1861         unsigned long flags;
1862
1863         spin_lock_irqsave(q->queue_lock, flags);
1864         blk_remove_plug(q);
1865         q->request_fn(q);
1866         spin_unlock_irqrestore(q->queue_lock, flags);
1867 }
1868
1869 /*
1870  * Timer running if the active_queue is currently idling inside its time slice
1871  */
1872 static void cfq_idle_slice_timer(unsigned long data)
1873 {
1874         struct cfq_data *cfqd = (struct cfq_data *) data;
1875         struct cfq_queue *cfqq;
1876         unsigned long flags;
1877
1878         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1879
1880         if ((cfqq = cfqd->active_queue) != NULL) {
1881                 unsigned long now = jiffies;
1882
1883                 /*
1884                  * expired
1885                  */
1886                 if (time_after(now, cfqq->slice_end))
1887                         goto expire;
1888
1889                 /*
1890                  * only expire and reinvoke request handler, if there are
1891                  * other queues with pending requests
1892                  */
1893                 if (!cfqd->busy_queues)
1894                         goto out_cont;
1895
1896                 /*
1897                  * not expired and it has a request pending, let it dispatch
1898                  */
1899                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1900                         cfq_mark_cfqq_must_dispatch(cfqq);
1901                         goto out_kick;
1902                 }
1903         }
1904 expire:
1905         cfq_slice_expired(cfqd, 0);
1906 out_kick:
1907         cfq_schedule_dispatch(cfqd);
1908 out_cont:
1909         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1910 }
1911
1912 /*
1913  * Timer running if an idle class queue is waiting for service
1914  */
1915 static void cfq_idle_class_timer(unsigned long data)
1916 {
1917         struct cfq_data *cfqd = (struct cfq_data *) data;
1918         unsigned long flags, end;
1919
1920         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1921
1922         /*
1923          * race with a non-idle queue, reset timer
1924          */
1925         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
1926         if (!time_after_eq(jiffies, end))
1927                 mod_timer(&cfqd->idle_class_timer, end);
1928         else
1929                 cfq_schedule_dispatch(cfqd);
1930
1931         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1932 }
1933
1934 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1935 {
1936         del_timer_sync(&cfqd->idle_slice_timer);
1937         del_timer_sync(&cfqd->idle_class_timer);
1938         blk_sync_queue(cfqd->queue);
1939 }
1940
1941 static void cfq_exit_queue(elevator_t *e)
1942 {
1943         struct cfq_data *cfqd = e->elevator_data;
1944         request_queue_t *q = cfqd->queue;
1945
1946         cfq_shutdown_timer_wq(cfqd);
1947
1948         spin_lock(&cfq_exit_lock);
1949         spin_lock_irq(q->queue_lock);
1950
1951         if (cfqd->active_queue)
1952                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
1953
1954         while (!list_empty(&cfqd->cic_list)) {
1955                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1956                                                         struct cfq_io_context,
1957                                                         queue_list);
1958
1959                 __cfq_exit_single_io_context(cfqd, cic);
1960         }
1961
1962         spin_unlock_irq(q->queue_lock);
1963         spin_unlock(&cfq_exit_lock);
1964
1965         cfq_shutdown_timer_wq(cfqd);
1966
1967         kfree(cfqd->cfq_hash);
1968         kfree(cfqd);
1969 }
1970
1971 static void *cfq_init_queue(request_queue_t *q, elevator_t *e)
1972 {
1973         struct cfq_data *cfqd;
1974         int i;
1975
1976         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
1977         if (!cfqd)
1978                 return NULL;
1979
1980         memset(cfqd, 0, sizeof(*cfqd));
1981
1982         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1983                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
1984
1985         INIT_LIST_HEAD(&cfqd->busy_rr);
1986         INIT_LIST_HEAD(&cfqd->cur_rr);
1987         INIT_LIST_HEAD(&cfqd->idle_rr);
1988         INIT_LIST_HEAD(&cfqd->empty_list);
1989         INIT_LIST_HEAD(&cfqd->cic_list);
1990
1991         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
1992         if (!cfqd->cfq_hash)
1993                 goto out_free;
1994
1995         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
1996                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
1997
1998         cfqd->queue = q;
1999
2000         init_timer(&cfqd->idle_slice_timer);
2001         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2002         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2003
2004         init_timer(&cfqd->idle_class_timer);
2005         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2006         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2007
2008         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2009
2010         cfqd->cfq_quantum = cfq_quantum;
2011         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2012         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2013         cfqd->cfq_back_max = cfq_back_max;
2014         cfqd->cfq_back_penalty = cfq_back_penalty;
2015         cfqd->cfq_slice[0] = cfq_slice_async;
2016         cfqd->cfq_slice[1] = cfq_slice_sync;
2017         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2018         cfqd->cfq_slice_idle = cfq_slice_idle;
2019
2020         return cfqd;
2021 out_free:
2022         kfree(cfqd);
2023         return NULL;
2024 }
2025
2026 static void cfq_slab_kill(void)
2027 {
2028         if (cfq_pool)
2029                 kmem_cache_destroy(cfq_pool);
2030         if (cfq_ioc_pool)
2031                 kmem_cache_destroy(cfq_ioc_pool);
2032 }
2033
2034 static int __init cfq_slab_setup(void)
2035 {
2036         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2037                                         NULL, NULL);
2038         if (!cfq_pool)
2039                 goto fail;
2040
2041         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2042                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2043         if (!cfq_ioc_pool)
2044                 goto fail;
2045
2046         return 0;
2047 fail:
2048         cfq_slab_kill();
2049         return -ENOMEM;
2050 }
2051
2052 /*
2053  * sysfs parts below -->
2054  */
2055
2056 static ssize_t
2057 cfq_var_show(unsigned int var, char *page)
2058 {
2059         return sprintf(page, "%d\n", var);
2060 }
2061
2062 static ssize_t
2063 cfq_var_store(unsigned int *var, const char *page, size_t count)
2064 {
2065         char *p = (char *) page;
2066
2067         *var = simple_strtoul(p, &p, 10);
2068         return count;
2069 }
2070
2071 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2072 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2073 {                                                                       \
2074         struct cfq_data *cfqd = e->elevator_data;                       \
2075         unsigned int __data = __VAR;                                    \
2076         if (__CONV)                                                     \
2077                 __data = jiffies_to_msecs(__data);                      \
2078         return cfq_var_show(__data, (page));                            \
2079 }
2080 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2081 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2082 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2083 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2084 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2085 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2086 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2087 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2088 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2089 #undef SHOW_FUNCTION
2090
2091 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2092 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2093 {                                                                       \
2094         struct cfq_data *cfqd = e->elevator_data;                       \
2095         unsigned int __data;                                            \
2096         int ret = cfq_var_store(&__data, (page), count);                \
2097         if (__data < (MIN))                                             \
2098                 __data = (MIN);                                         \
2099         else if (__data > (MAX))                                        \
2100                 __data = (MAX);                                         \
2101         if (__CONV)                                                     \
2102                 *(__PTR) = msecs_to_jiffies(__data);                    \
2103         else                                                            \
2104                 *(__PTR) = __data;                                      \
2105         return ret;                                                     \
2106 }
2107 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2108 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2109 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2110 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2111 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2112 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2113 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2114 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2115 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2116 #undef STORE_FUNCTION
2117
2118 #define CFQ_ATTR(name) \
2119         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2120
2121 static struct elv_fs_entry cfq_attrs[] = {
2122         CFQ_ATTR(quantum),
2123         CFQ_ATTR(fifo_expire_sync),
2124         CFQ_ATTR(fifo_expire_async),
2125         CFQ_ATTR(back_seek_max),
2126         CFQ_ATTR(back_seek_penalty),
2127         CFQ_ATTR(slice_sync),
2128         CFQ_ATTR(slice_async),
2129         CFQ_ATTR(slice_async_rq),
2130         CFQ_ATTR(slice_idle),
2131         __ATTR_NULL
2132 };
2133
2134 static struct elevator_type iosched_cfq = {
2135         .ops = {
2136                 .elevator_merge_fn =            cfq_merge,
2137                 .elevator_merged_fn =           cfq_merged_request,
2138                 .elevator_merge_req_fn =        cfq_merged_requests,
2139                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2140                 .elevator_add_req_fn =          cfq_insert_request,
2141                 .elevator_activate_req_fn =     cfq_activate_request,
2142                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2143                 .elevator_queue_empty_fn =      cfq_queue_empty,
2144                 .elevator_completed_req_fn =    cfq_completed_request,
2145                 .elevator_former_req_fn =       elv_rb_former_request,
2146                 .elevator_latter_req_fn =       elv_rb_latter_request,
2147                 .elevator_set_req_fn =          cfq_set_request,
2148                 .elevator_put_req_fn =          cfq_put_request,
2149                 .elevator_may_queue_fn =        cfq_may_queue,
2150                 .elevator_init_fn =             cfq_init_queue,
2151                 .elevator_exit_fn =             cfq_exit_queue,
2152                 .trim =                         cfq_trim,
2153         },
2154         .elevator_attrs =       cfq_attrs,
2155         .elevator_name =        "cfq",
2156         .elevator_owner =       THIS_MODULE,
2157 };
2158
2159 static int __init cfq_init(void)
2160 {
2161         int ret;
2162
2163         /*
2164          * could be 0 on HZ < 1000 setups
2165          */
2166         if (!cfq_slice_async)
2167                 cfq_slice_async = 1;
2168         if (!cfq_slice_idle)
2169                 cfq_slice_idle = 1;
2170
2171         if (cfq_slab_setup())
2172                 return -ENOMEM;
2173
2174         ret = elv_register(&iosched_cfq);
2175         if (ret)
2176                 cfq_slab_kill();
2177
2178         return ret;
2179 }
2180
2181 static void __exit cfq_exit(void)
2182 {
2183         DECLARE_COMPLETION(all_gone);
2184         elv_unregister(&iosched_cfq);
2185         ioc_gone = &all_gone;
2186         /* ioc_gone's update must be visible before reading ioc_count */
2187         smp_wmb();
2188         if (atomic_read(&ioc_count))
2189                 wait_for_completion(ioc_gone);
2190         synchronize_rcu();
2191         cfq_slab_kill();
2192 }
2193
2194 module_init(cfq_init);
2195 module_exit(cfq_exit);
2196
2197 MODULE_AUTHOR("Jens Axboe");
2198 MODULE_LICENSE("GPL");
2199 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");