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

   1 /*
   2  *  CFQ, or complete fairness queueing, disk scheduler.
   3  *
   4  *  Based on ideas from a previously unfinished io
   5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   6  *
   7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
   8  */
   9 #include <linux/module.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/elevator.h>
  12 #include <linux/hash.h>
  13 #include <linux/rbtree.h>
  14 #include <linux/ioprio.h>
  15
  16 /*
  17  * tunables
  18  */
  19 static const int cfq_quantum = 4;               /* max queue in one round of service */
  20 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  21 static const int cfq_back_max = 16 * 1024;      /* maximum backwards seek, in KiB */
  22 static const int cfq_back_penalty = 2;          /* penalty of a backwards seek */
  23
  24 static const int cfq_slice_sync = HZ / 10;
  25 static int cfq_slice_async = HZ / 25;
  26 static const int cfq_slice_async_rq = 2;
  27 static int cfq_slice_idle = HZ / 125;
  28
  29 #define CFQ_IDLE_GRACE          (HZ / 10)
  30 #define CFQ_SLICE_SCALE         (5)
  31
  32 #define CFQ_KEY_ASYNC           (0)
  33
  34 /*
  35  * for the hash of cfqq inside the cfqd
  36  */
  37 #define CFQ_QHASH_SHIFT         6
  38 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
  39 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
  40
  41 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
  42
  43 #define RQ_CIC(rq)              ((struct cfq_io_context*)(rq)->elevator_private)
  44 #define RQ_CFQQ(rq)             ((rq)->elevator_private2)
  45
  46 static struct kmem_cache *cfq_pool;
  47 static struct kmem_cache *cfq_ioc_pool;
  48
  49 static DEFINE_PER_CPU(unsigned long, ioc_count);
  50 static struct completion *ioc_gone;
  51
  52 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  53 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  54 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  55
  56 #define ASYNC                   (0)
  57 #define SYNC                    (1)
  58
  59 #define cfq_cfqq_dispatched(cfqq)       \
  60         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
  61
  62 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
  63
  64 #define cfq_cfqq_sync(cfqq)             \
  65         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
  66
  67 #define sample_valid(samples)   ((samples) > 80)
  68
  69 /*
  70  * Per block device queue structure
  71  */
  72 struct cfq_data {
  73         request_queue_t *queue;
  74
  75         /*
  76          * rr list of queues with requests and the count of them
  77          */
  78         struct list_head rr_list[CFQ_PRIO_LISTS];
  79         struct list_head busy_rr;
  80         struct list_head cur_rr;
  81         struct list_head idle_rr;
  82         unsigned int busy_queues;
  83
  84         /*
  85          * cfqq lookup hash
  86          */
  87         struct hlist_head *cfq_hash;
  88
  89         int rq_in_driver;
  90         int hw_tag;
  91
  92         /*
  93          * idle window management
  94          */
  95         struct timer_list idle_slice_timer;
  96         struct work_struct unplug_work;
  97
  98         struct cfq_queue *active_queue;
  99         struct cfq_io_context *active_cic;
 100         int cur_prio, cur_end_prio;
 101         unsigned int dispatch_slice;
 102
 103         struct timer_list idle_class_timer;
 104
 105         sector_t last_sector;
 106         unsigned long last_end_request;
 107
 108         /*
 109          * tunables, see top of file
 110          */
 111         unsigned int cfq_quantum;
 112         unsigned int cfq_fifo_expire[2];
 113         unsigned int cfq_back_penalty;
 114         unsigned int cfq_back_max;
 115         unsigned int cfq_slice[2];
 116         unsigned int cfq_slice_async_rq;
 117         unsigned int cfq_slice_idle;
 118
 119         struct list_head cic_list;
 120 };
 121
 122 /*
 123  * Per process-grouping structure
 124  */
 125 struct cfq_queue {
 126         /* reference count */
 127         atomic_t ref;
 128         /* parent cfq_data */
 129         struct cfq_data *cfqd;
 130         /* cfqq lookup hash */
 131         struct hlist_node cfq_hash;
 132         /* hash key */
 133         unsigned int key;
 134         /* member of the rr/busy/cur/idle cfqd list */
 135         struct list_head cfq_list;
 136         /* sorted list of pending requests */
 137         struct rb_root sort_list;
 138         /* if fifo isn't expired, next request to serve */
 139         struct request *next_rq;
 140         /* requests queued in sort_list */
 141         int queued[2];
 142         /* currently allocated requests */
 143         int allocated[2];
 144         /* pending metadata requests */
 145         int meta_pending;
 146         /* fifo list of requests in sort_list */
 147         struct list_head fifo;
 148
 149         unsigned long slice_end;
 150         unsigned long slice_left;
 151         unsigned long service_last;
 152
 153         /* number of requests that are on the dispatch list */
 154         int on_dispatch[2];
 155
 156         /* io prio of this group */
 157         unsigned short ioprio, org_ioprio;
 158         unsigned short ioprio_class, org_ioprio_class;
 159
 160         /* various state flags, see below */
 161         unsigned int flags;
 162 };
 163
 164 enum cfqq_state_flags {
 165         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 166         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 167         CFQ_CFQQ_FLAG_must_alloc,       /* must be allowed rq alloc */
 168         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 169         CFQ_CFQQ_FLAG_must_dispatch,    /* must dispatch, even if expired */
 170         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 171         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 172         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 173         CFQ_CFQQ_FLAG_queue_new,        /* queue never been serviced */
 174 };
 175
 176 #define CFQ_CFQQ_FNS(name)                                              \
 177 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 178 {                                                                       \
 179         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 180 }                                                                       \
 181 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 182 {                                                                       \
 183         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 184 }                                                                       \
 185 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 186 {                                                                       \
 187         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 188 }
 189
 190 CFQ_CFQQ_FNS(on_rr);
 191 CFQ_CFQQ_FNS(wait_request);
 192 CFQ_CFQQ_FNS(must_alloc);
 193 CFQ_CFQQ_FNS(must_alloc_slice);
 194 CFQ_CFQQ_FNS(must_dispatch);
 195 CFQ_CFQQ_FNS(fifo_expire);
 196 CFQ_CFQQ_FNS(idle_window);
 197 CFQ_CFQQ_FNS(prio_changed);
 198 CFQ_CFQQ_FNS(queue_new);
 199 #undef CFQ_CFQQ_FNS
 200
 201 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 202 static void cfq_dispatch_insert(request_queue_t *, struct request *);
 203 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
 204
 205 /*
 206  * scheduler run of queue, if there are requests pending and no one in the
 207  * driver that will restart queueing
 208  */
 209 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 210 {
 211         if (cfqd->busy_queues)
 212                 kblockd_schedule_work(&cfqd->unplug_work);
 213 }
 214
 215 static int cfq_queue_empty(request_queue_t *q)
 216 {
 217         struct cfq_data *cfqd = q->elevator->elevator_data;
 218
 219         return !cfqd->busy_queues;
 220 }
 221
 222 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw, int is_sync)
 223 {
 224         /*
 225          * Use the per-process queue, for read requests and syncronous writes
 226          */
 227         if (!(rw & REQ_RW) || is_sync)
 228                 return task->pid;
 229
 230         return CFQ_KEY_ASYNC;
 231 }
 232
 233 /*
 234  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 235  * We choose the request that is closest to the head right now. Distance
 236  * behind the head is penalized and only allowed to a certain extent.
 237  */
 238 static struct request *
 239 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
 240 {
 241         sector_t last, s1, s2, d1 = 0, d2 = 0;
 242         unsigned long back_max;
 243 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 244 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 245         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 246
 247         if (rq1 == NULL || rq1 == rq2)
 248                 return rq2;
 249         if (rq2 == NULL)
 250                 return rq1;
 251
 252         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 253                 return rq1;
 254         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 255                 return rq2;
 256         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
 257                 return rq1;
 258         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
 259                 return rq2;
 260
 261         s1 = rq1->sector;
 262         s2 = rq2->sector;
 263
 264         last = cfqd->last_sector;
 265
 266         /*
 267          * by definition, 1KiB is 2 sectors
 268          */
 269         back_max = cfqd->cfq_back_max * 2;
 270
 271         /*
 272          * Strict one way elevator _except_ in the case where we allow
 273          * short backward seeks which are biased as twice the cost of a
 274          * similar forward seek.
 275          */
 276         if (s1 >= last)
 277                 d1 = s1 - last;
 278         else if (s1 + back_max >= last)
 279                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 280         else
 281                 wrap |= CFQ_RQ1_WRAP;
 282
 283         if (s2 >= last)
 284                 d2 = s2 - last;
 285         else if (s2 + back_max >= last)
 286                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 287         else
 288                 wrap |= CFQ_RQ2_WRAP;
 289
 290         /* Found required data */
 291
 292         /*
 293          * By doing switch() on the bit mask "wrap" we avoid having to
 294          * check two variables for all permutations: --> faster!
 295          */
 296         switch (wrap) {
 297         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 298                 if (d1 < d2)
 299                         return rq1;
 300                 else if (d2 < d1)
 301                         return rq2;
 302                 else {
 303                         if (s1 >= s2)
 304                                 return rq1;
 305                         else
 306                                 return rq2;
 307                 }
 308
 309         case CFQ_RQ2_WRAP:
 310                 return rq1;
 311         case CFQ_RQ1_WRAP:
 312                 return rq2;
 313         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
 314         default:
 315                 /*
 316                  * Since both rqs are wrapped,
 317                  * start with the one that's further behind head
 318                  * (--> only *one* back seek required),
 319                  * since back seek takes more time than forward.
 320                  */
 321                 if (s1 <= s2)
 322                         return rq1;
 323                 else
 324                         return rq2;
 325         }
 326 }
 327
 328 /*
 329  * would be nice to take fifo expire time into account as well
 330  */
 331 static struct request *
 332 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 333                   struct request *last)
 334 {
 335         struct rb_node *rbnext = rb_next(&last->rb_node);
 336         struct rb_node *rbprev = rb_prev(&last->rb_node);
 337         struct request *next = NULL, *prev = NULL;
 338
 339         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
 340
 341         if (rbprev)
 342                 prev = rb_entry_rq(rbprev);
 343
 344         if (rbnext)
 345                 next = rb_entry_rq(rbnext);
 346         else {
 347                 rbnext = rb_first(&cfqq->sort_list);
 348                 if (rbnext && rbnext != &last->rb_node)
 349                         next = rb_entry_rq(rbnext);
 350         }
 351
 352         return cfq_choose_req(cfqd, next, prev);
 353 }
 354
 355 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 356 {
 357         struct cfq_data *cfqd = cfqq->cfqd;
 358         struct list_head *list, *n;
 359         struct cfq_queue *__cfqq;
 360
 361         /*
 362          * Resorting requires the cfqq to be on the RR list already.
 363          */
 364         if (!cfq_cfqq_on_rr(cfqq))
 365                 return;
 366
 367         list_del(&cfqq->cfq_list);
 368
 369         if (cfq_class_rt(cfqq))
 370                 list = &cfqd->cur_rr;
 371         else if (cfq_class_idle(cfqq))
 372                 list = &cfqd->idle_rr;
 373         else {
 374                 /*
 375                  * if cfqq has requests in flight, don't allow it to be
 376                  * found in cfq_set_active_queue before it has finished them.
 377                  * this is done to increase fairness between a process that
 378                  * has lots of io pending vs one that only generates one
 379                  * sporadically or synchronously
 380                  */
 381                 if (cfq_cfqq_dispatched(cfqq))
 382                         list = &cfqd->busy_rr;
 383                 else
 384                         list = &cfqd->rr_list[cfqq->ioprio];
 385         }
 386
 387         if (preempted || cfq_cfqq_queue_new(cfqq)) {
 388                 /*
 389                  * If this queue was preempted or is new (never been serviced),
 390                  * let it be added first for fairness but beind other new
 391                  * queues.
 392                  */
 393                 n = list;
 394                 while (n->next != list) {
 395                         __cfqq = list_entry_cfqq(n->next);
 396                         if (!cfq_cfqq_queue_new(__cfqq))
 397                                 break;
 398
 399                         n = n->next;
 400                 }
 401                 list_add_tail(&cfqq->cfq_list, n);
 402         } else if (!cfq_cfqq_class_sync(cfqq)) {
 403                 /*
 404                  * async queue always goes to the end. this wont be overly
 405                  * unfair to writes, as the sort of the sync queue wont be
 406                  * allowed to pass the async queue again.
 407                  */
 408                 list_add_tail(&cfqq->cfq_list, list);
 409         } else {
 410                 /*
 411                  * sort by last service, but don't cross a new or async
 412                  * queue. we don't cross a new queue because it hasn't been
 413                  * service before, and we don't cross an async queue because
 414                  * it gets added to the end on expire.
 415                  */
 416                 n = list;
 417                 while ((n = n->prev) != list) {
 418                         struct cfq_queue *__cfqq = list_entry_cfqq(n);
 419
 420                         if (!cfq_cfqq_class_sync(cfqq) || !__cfqq->service_last)
 421                                 break;
 422                         if (time_before(__cfqq->service_last, cfqq->service_last))
 423                                 break;
 424                 }
 425                 list_add(&cfqq->cfq_list, n);
 426         }
 427 }
 428
 429 /*
 430  * add to busy list of queues for service, trying to be fair in ordering
 431  * the pending list according to last request service
 432  */
 433 static inline void
 434 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 435 {
 436         BUG_ON(cfq_cfqq_on_rr(cfqq));
 437         cfq_mark_cfqq_on_rr(cfqq);
 438         cfqd->busy_queues++;
 439
 440         cfq_resort_rr_list(cfqq, 0);
 441 }
 442
 443 static inline void
 444 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 445 {
 446         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 447         cfq_clear_cfqq_on_rr(cfqq);
 448         list_del_init(&cfqq->cfq_list);
 449
 450         BUG_ON(!cfqd->busy_queues);
 451         cfqd->busy_queues--;
 452 }
 453
 454 /*
 455  * rb tree support functions
 456  */
 457 static inline void cfq_del_rq_rb(struct request *rq)
 458 {
 459         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 460         struct cfq_data *cfqd = cfqq->cfqd;
 461         const int sync = rq_is_sync(rq);
 462
 463         BUG_ON(!cfqq->queued[sync]);
 464         cfqq->queued[sync]--;
 465
 466         elv_rb_del(&cfqq->sort_list, rq);
 467
 468         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
 469                 cfq_del_cfqq_rr(cfqd, cfqq);
 470 }
 471
 472 static void cfq_add_rq_rb(struct request *rq)
 473 {
 474         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 475         struct cfq_data *cfqd = cfqq->cfqd;
 476         struct request *__alias;
 477
 478         cfqq->queued[rq_is_sync(rq)]++;
 479
 480         /*
 481          * looks a little odd, but the first insert might return an alias.
 482          * if that happens, put the alias on the dispatch list
 483          */
 484         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
 485                 cfq_dispatch_insert(cfqd->queue, __alias);
 486
 487         if (!cfq_cfqq_on_rr(cfqq))
 488                 cfq_add_cfqq_rr(cfqd, cfqq);
 489 }
 490
 491 static inline void
 492 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
 493 {
 494         elv_rb_del(&cfqq->sort_list, rq);
 495         cfqq->queued[rq_is_sync(rq)]--;
 496         cfq_add_rq_rb(rq);
 497 }
 498
 499 static struct request *
 500 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 501 {
 502         struct task_struct *tsk = current;
 503         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
 504         struct cfq_queue *cfqq;
 505
 506         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
 507         if (cfqq) {
 508                 sector_t sector = bio->bi_sector + bio_sectors(bio);
 509
 510                 return elv_rb_find(&cfqq->sort_list, sector);
 511         }
 512
 513         return NULL;
 514 }
 515
 516 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 517 {
 518         struct cfq_data *cfqd = q->elevator->elevator_data;
 519
 520         cfqd->rq_in_driver++;
 521
 522         /*
 523          * If the depth is larger 1, it really could be queueing. But lets
 524          * make the mark a little higher - idling could still be good for
 525          * low queueing, and a low queueing number could also just indicate
 526          * a SCSI mid layer like behaviour where limit+1 is often seen.
 527          */
 528         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
 529                 cfqd->hw_tag = 1;
 530 }
 531
 532 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 533 {
 534         struct cfq_data *cfqd = q->elevator->elevator_data;
 535
 536         WARN_ON(!cfqd->rq_in_driver);
 537         cfqd->rq_in_driver--;
 538 }
 539
 540 static void cfq_remove_request(struct request *rq)
 541 {
 542         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 543
 544         if (cfqq->next_rq == rq)
 545                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
 546
 547         list_del_init(&rq->queuelist);
 548         cfq_del_rq_rb(rq);
 549
 550         if (rq_is_meta(rq)) {
 551                 WARN_ON(!cfqq->meta_pending);
 552                 cfqq->meta_pending--;
 553         }
 554 }
 555
 556 static int
 557 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 558 {
 559         struct cfq_data *cfqd = q->elevator->elevator_data;
 560         struct request *__rq;
 561
 562         __rq = cfq_find_rq_fmerge(cfqd, bio);
 563         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 564                 *req = __rq;
 565                 return ELEVATOR_FRONT_MERGE;
 566         }
 567
 568         return ELEVATOR_NO_MERGE;
 569 }
 570
 571 static void cfq_merged_request(request_queue_t *q, struct request *req,
 572                                int type)
 573 {
 574         if (type == ELEVATOR_FRONT_MERGE) {
 575                 struct cfq_queue *cfqq = RQ_CFQQ(req);
 576
 577                 cfq_reposition_rq_rb(cfqq, req);
 578         }
 579 }
 580
 581 static void
 582 cfq_merged_requests(request_queue_t *q, struct request *rq,
 583                     struct request *next)
 584 {
 585         /*
 586          * reposition in fifo if next is older than rq
 587          */
 588         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 589             time_before(next->start_time, rq->start_time))
 590                 list_move(&rq->queuelist, &next->queuelist);
 591
 592         cfq_remove_request(next);
 593 }
 594
 595 static int cfq_allow_merge(request_queue_t *q, struct request *rq,
 596                            struct bio *bio)
 597 {
 598         struct cfq_data *cfqd = q->elevator->elevator_data;
 599         const int rw = bio_data_dir(bio);
 600         struct cfq_queue *cfqq;
 601         pid_t key;
 602
 603         /*
 604          * Disallow merge of a sync bio into an async request.
 605          */
 606         if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq))
 607                 return 0;
 608
 609         /*
 610          * Lookup the cfqq that this bio will be queued with. Allow
 611          * merge only if rq is queued there.
 612          */
 613         key = cfq_queue_pid(current, rw, bio_sync(bio));
 614         cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio);
 615
 616         if (cfqq == RQ_CFQQ(rq))
 617                 return 1;
 618
 619         return 0;
 620 }
 621
 622 static inline void
 623 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 624 {
 625         if (cfqq) {
 626                 /*
 627                  * stop potential idle class queues waiting service
 628                  */
 629                 del_timer(&cfqd->idle_class_timer);
 630
 631                 cfqq->slice_end = 0;
 632                 cfqq->slice_left = 0;
 633                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 634                 cfq_clear_cfqq_fifo_expire(cfqq);
 635         }
 636
 637         cfqd->active_queue = cfqq;
 638 }
 639
 640 /*
 641  * current cfqq expired its slice (or was too idle), select new one
 642  */
 643 static void
 644 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 645                     int preempted)
 646 {
 647         unsigned long now = jiffies;
 648
 649         if (cfq_cfqq_wait_request(cfqq))
 650                 del_timer(&cfqd->idle_slice_timer);
 651
 652         if (!preempted && !cfq_cfqq_dispatched(cfqq))
 653                 cfq_schedule_dispatch(cfqd);
 654
 655         cfq_clear_cfqq_must_dispatch(cfqq);
 656         cfq_clear_cfqq_wait_request(cfqq);
 657         cfq_clear_cfqq_queue_new(cfqq);
 658
 659         /*
 660          * store what was left of this slice, if the queue idled out
 661          * or was preempted
 662          */
 663         if (time_after(cfqq->slice_end, now))
 664                 cfqq->slice_left = cfqq->slice_end - now;
 665         else
 666                 cfqq->slice_left = 0;
 667
 668         cfq_resort_rr_list(cfqq, preempted);
 669
 670         if (cfqq == cfqd->active_queue)
 671                 cfqd->active_queue = NULL;
 672
 673         if (cfqd->active_cic) {
 674                 put_io_context(cfqd->active_cic->ioc);
 675                 cfqd->active_cic = NULL;
 676         }
 677
 678         cfqd->dispatch_slice = 0;
 679 }
 680
 681 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 682 {
 683         struct cfq_queue *cfqq = cfqd->active_queue;
 684
 685         if (cfqq)
 686                 __cfq_slice_expired(cfqd, cfqq, preempted);
 687 }
 688
 689 /*
 690  * 0
 691  * 0,1
 692  * 0,1,2
 693  * 0,1,2,3
 694  * 0,1,2,3,4
 695  * 0,1,2,3,4,5
 696  * 0,1,2,3,4,5,6
 697  * 0,1,2,3,4,5,6,7
 698  */
 699 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 700 {
 701         int prio, wrap;
 702
 703         prio = -1;
 704         wrap = 0;
 705         do {
 706                 int p;
 707
 708                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 709                         if (!list_empty(&cfqd->rr_list[p])) {
 710                                 prio = p;
 711                                 break;
 712                         }
 713                 }
 714
 715                 if (prio != -1)
 716                         break;
 717                 cfqd->cur_prio = 0;
 718                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 719                         cfqd->cur_end_prio = 0;
 720                         if (wrap)
 721                                 break;
 722                         wrap = 1;
 723                 }
 724         } while (1);
 725
 726         if (unlikely(prio == -1))
 727                 return -1;
 728
 729         BUG_ON(prio >= CFQ_PRIO_LISTS);
 730
 731         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 732
 733         cfqd->cur_prio = prio + 1;
 734         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 735                 cfqd->cur_end_prio = cfqd->cur_prio;
 736                 cfqd->cur_prio = 0;
 737         }
 738         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 739                 cfqd->cur_prio = 0;
 740                 cfqd->cur_end_prio = 0;
 741         }
 742
 743         return prio;
 744 }
 745
 746 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 747 {
 748         struct cfq_queue *cfqq = NULL;
 749
 750         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
 751                 /*
 752                  * if current list is non-empty, grab first entry. if it is
 753                  * empty, get next prio level and grab first entry then if any
 754                  * are spliced
 755                  */
 756                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 757         } else if (!list_empty(&cfqd->busy_rr)) {
 758                 /*
 759                  * If no new queues are available, check if the busy list has
 760                  * some before falling back to idle io.
 761                  */
 762                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
 763         } else if (!list_empty(&cfqd->idle_rr)) {
 764                 /*
 765                  * if we have idle queues and no rt or be queues had pending
 766                  * requests, either allow immediate service if the grace period
 767                  * has passed or arm the idle grace timer
 768                  */
 769                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 770
 771                 if (time_after_eq(jiffies, end))
 772                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 773                 else
 774                         mod_timer(&cfqd->idle_class_timer, end);
 775         }
 776
 777         __cfq_set_active_queue(cfqd, cfqq);
 778         return cfqq;
 779 }
 780
 781 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
 782
 783 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 784
 785 {
 786         struct cfq_io_context *cic;
 787         unsigned long sl;
 788
 789         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
 790         WARN_ON(cfqq != cfqd->active_queue);
 791
 792         /*
 793          * idle is disabled, either manually or by past process history
 794          */
 795         if (!cfqd->cfq_slice_idle)
 796                 return 0;
 797         if (!cfq_cfqq_idle_window(cfqq))
 798                 return 0;
 799         /*
 800          * task has exited, don't wait
 801          */
 802         cic = cfqd->active_cic;
 803         if (!cic || !cic->ioc->task)
 804                 return 0;
 805
 806         cfq_mark_cfqq_must_dispatch(cfqq);
 807         cfq_mark_cfqq_wait_request(cfqq);
 808
 809         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 810
 811         /*
 812          * we don't want to idle for seeks, but we do want to allow
 813          * fair distribution of slice time for a process doing back-to-back
 814          * seeks. so allow a little bit of time for him to submit a new rq
 815          */
 816         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
 817                 sl = min(sl, msecs_to_jiffies(2));
 818
 819         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 820         return 1;
 821 }
 822
 823 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
 824 {
 825         struct cfq_data *cfqd = q->elevator->elevator_data;
 826         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 827
 828         cfq_remove_request(rq);
 829         cfqq->on_dispatch[rq_is_sync(rq)]++;
 830         elv_dispatch_sort(q, rq);
 831
 832         rq = list_entry(q->queue_head.prev, struct request, queuelist);
 833         cfqd->last_sector = rq->sector + rq->nr_sectors;
 834 }
 835
 836 /*
 837  * return expired entry, or NULL to just start from scratch in rbtree
 838  */
 839 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
 840 {
 841         struct cfq_data *cfqd = cfqq->cfqd;
 842         struct request *rq;
 843         int fifo;
 844
 845         if (cfq_cfqq_fifo_expire(cfqq))
 846                 return NULL;
 847         if (list_empty(&cfqq->fifo))
 848                 return NULL;
 849
 850         fifo = cfq_cfqq_class_sync(cfqq);
 851         rq = rq_entry_fifo(cfqq->fifo.next);
 852
 853         if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
 854                 cfq_mark_cfqq_fifo_expire(cfqq);
 855                 return rq;
 856         }
 857
 858         return NULL;
 859 }
 860
 861 /*
 862  * Scale schedule slice based on io priority. Use the sync time slice only
 863  * if a queue is marked sync and has sync io queued. A sync queue with async
 864  * io only, should not get full sync slice length.
 865  */
 866 static inline int
 867 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 868 {
 869         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 870
 871         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 872
 873         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
 874 }
 875
 876 static inline void
 877 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 878 {
 879         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 880 }
 881
 882 static inline int
 883 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 884 {
 885         const int base_rq = cfqd->cfq_slice_async_rq;
 886
 887         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 888
 889         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
 890 }
 891
 892 /*
 893  * get next queue for service
 894  */
 895 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
 896 {
 897         unsigned long now = jiffies;
 898         struct cfq_queue *cfqq;
 899
 900         cfqq = cfqd->active_queue;
 901         if (!cfqq)
 902                 goto new_queue;
 903
 904         /*
 905          * slice has expired
 906          */
 907         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
 908                 goto expire;
 909
 910         /*
 911          * if queue has requests, dispatch one. if not, check if
 912          * enough slice is left to wait for one
 913          */
 914         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
 915                 goto keep_queue;
 916         else if (cfq_cfqq_dispatched(cfqq)) {
 917                 cfqq = NULL;
 918                 goto keep_queue;
 919         } else if (cfq_cfqq_class_sync(cfqq)) {
 920                 if (cfq_arm_slice_timer(cfqd, cfqq))
 921                         return NULL;
 922         }
 923
 924 expire:
 925         cfq_slice_expired(cfqd, 0);
 926 new_queue:
 927         cfqq = cfq_set_active_queue(cfqd);
 928 keep_queue:
 929         return cfqq;
 930 }
 931
 932 static int
 933 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 934                         int max_dispatch)
 935 {
 936         int dispatched = 0;
 937
 938         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
 939
 940         do {
 941                 struct request *rq;
 942
 943                 /*
 944                  * follow expired path, else get first next available
 945                  */
 946                 if ((rq = cfq_check_fifo(cfqq)) == NULL)
 947                         rq = cfqq->next_rq;
 948
 949                 /*
 950                  * finally, insert request into driver dispatch list
 951                  */
 952                 cfq_dispatch_insert(cfqd->queue, rq);
 953
 954                 cfqd->dispatch_slice++;
 955                 dispatched++;
 956
 957                 if (!cfqd->active_cic) {
 958                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
 959                         cfqd->active_cic = RQ_CIC(rq);
 960                 }
 961
 962                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
 963                         break;
 964
 965         } while (dispatched < max_dispatch);
 966
 967         /*
 968          * if slice end isn't set yet, set it.
 969          */
 970         if (!cfqq->slice_end)
 971                 cfq_set_prio_slice(cfqd, cfqq);
 972
 973         /*
 974          * expire an async queue immediately if it has used up its slice. idle
 975          * queue always expire after 1 dispatch round.
 976          */
 977         if ((!cfq_cfqq_sync(cfqq) &&
 978             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
 979             cfq_class_idle(cfqq) ||
 980             !cfq_cfqq_idle_window(cfqq))
 981                 cfq_slice_expired(cfqd, 0);
 982
 983         return dispatched;
 984 }
 985
 986 static int
 987 cfq_forced_dispatch_cfqqs(struct list_head *list)
 988 {
 989         struct cfq_queue *cfqq, *next;
 990         int dispatched;
 991
 992         dispatched = 0;
 993         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
 994                 while (cfqq->next_rq) {
 995                         cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
 996                         dispatched++;
 997                 }
 998                 BUG_ON(!list_empty(&cfqq->fifo));
 999         }
1000
1001         return dispatched;
1002 }
1003
1004 static int
1005 cfq_forced_dispatch(struct cfq_data *cfqd)
1006 {
1007         int i, dispatched = 0;
1008
1009         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1010                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1011
1012         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1013         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1014         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1015
1016         cfq_slice_expired(cfqd, 0);
1017
1018         BUG_ON(cfqd->busy_queues);
1019
1020         return dispatched;
1021 }
1022
1023 static int
1024 cfq_dispatch_requests(request_queue_t *q, int force)
1025 {
1026         struct cfq_data *cfqd = q->elevator->elevator_data;
1027         struct cfq_queue *cfqq, *prev_cfqq;
1028         int dispatched;
1029
1030         if (!cfqd->busy_queues)
1031                 return 0;
1032
1033         if (unlikely(force))
1034                 return cfq_forced_dispatch(cfqd);
1035
1036         dispatched = 0;
1037         prev_cfqq = NULL;
1038         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1039                 int max_dispatch;
1040
1041                 /*
1042                  * Don't repeat dispatch from the previous queue.
1043                  */
1044                 if (prev_cfqq == cfqq)
1045                         break;
1046
1047                 cfq_clear_cfqq_must_dispatch(cfqq);
1048                 cfq_clear_cfqq_wait_request(cfqq);
1049                 del_timer(&cfqd->idle_slice_timer);
1050
1051                 max_dispatch = cfqd->cfq_quantum;
1052                 if (cfq_class_idle(cfqq))
1053                         max_dispatch = 1;
1054
1055                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1056
1057                 /*
1058                  * If the dispatch cfqq has idling enabled and is still
1059                  * the active queue, break out.
1060                  */
1061                 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1062                         break;
1063
1064                 prev_cfqq = cfqq;
1065         }
1066
1067         return dispatched;
1068 }
1069
1070 /*
1071  * task holds one reference to the queue, dropped when task exits. each rq
1072  * in-flight on this queue also holds a reference, dropped when rq is freed.
1073  *
1074  * queue lock must be held here.
1075  */
1076 static void cfq_put_queue(struct cfq_queue *cfqq)
1077 {
1078         struct cfq_data *cfqd = cfqq->cfqd;
1079
1080         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1081
1082         if (!atomic_dec_and_test(&cfqq->ref))
1083                 return;
1084
1085         BUG_ON(rb_first(&cfqq->sort_list));
1086         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1087         BUG_ON(cfq_cfqq_on_rr(cfqq));
1088
1089         if (unlikely(cfqd->active_queue == cfqq))
1090                 __cfq_slice_expired(cfqd, cfqq, 0);
1091
1092         /*
1093          * it's on the empty list and still hashed
1094          */
1095         list_del(&cfqq->cfq_list);
1096         hlist_del(&cfqq->cfq_hash);
1097         kmem_cache_free(cfq_pool, cfqq);
1098 }
1099
1100 static struct cfq_queue *
1101 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1102                     const int hashval)
1103 {
1104         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1105         struct hlist_node *entry;
1106         struct cfq_queue *__cfqq;
1107
1108         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1109                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1110
1111                 if (__cfqq->key == key && (__p == prio || !prio))
1112                         return __cfqq;
1113         }
1114
1115         return NULL;
1116 }
1117
1118 static struct cfq_queue *
1119 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1120 {
1121         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1122 }
1123
1124 static void cfq_free_io_context(struct io_context *ioc)
1125 {
1126         struct cfq_io_context *__cic;
1127         struct rb_node *n;
1128         int freed = 0;
1129
1130         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1131                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1132                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1133                 kmem_cache_free(cfq_ioc_pool, __cic);
1134                 freed++;
1135         }
1136
1137         elv_ioc_count_mod(ioc_count, -freed);
1138
1139         if (ioc_gone && !elv_ioc_count_read(ioc_count))
1140                 complete(ioc_gone);
1141 }
1142
1143 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1144 {
1145         if (unlikely(cfqq == cfqd->active_queue))
1146                 __cfq_slice_expired(cfqd, cfqq, 0);
1147
1148         cfq_put_queue(cfqq);
1149 }
1150
1151 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1152                                          struct cfq_io_context *cic)
1153 {
1154         list_del_init(&cic->queue_list);
1155         smp_wmb();
1156         cic->key = NULL;
1157
1158         if (cic->cfqq[ASYNC]) {
1159                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1160                 cic->cfqq[ASYNC] = NULL;
1161         }
1162
1163         if (cic->cfqq[SYNC]) {
1164                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1165                 cic->cfqq[SYNC] = NULL;
1166         }
1167 }
1168
1169
1170 /*
1171  * Called with interrupts disabled
1172  */
1173 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1174 {
1175         struct cfq_data *cfqd = cic->key;
1176
1177         if (cfqd) {
1178                 request_queue_t *q = cfqd->queue;
1179
1180                 spin_lock_irq(q->queue_lock);
1181                 __cfq_exit_single_io_context(cfqd, cic);
1182                 spin_unlock_irq(q->queue_lock);
1183         }
1184 }
1185
1186 static void cfq_exit_io_context(struct io_context *ioc)
1187 {
1188         struct cfq_io_context *__cic;
1189         struct rb_node *n;
1190
1191         /*
1192          * put the reference this task is holding to the various queues
1193          */
1194
1195         n = rb_first(&ioc->cic_root);
1196         while (n != NULL) {
1197                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1198
1199                 cfq_exit_single_io_context(__cic);
1200                 n = rb_next(n);
1201         }
1202 }
1203
1204 static struct cfq_io_context *
1205 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1206 {
1207         struct cfq_io_context *cic;
1208
1209         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
1210         if (cic) {
1211                 memset(cic, 0, sizeof(*cic));
1212                 cic->last_end_request = jiffies;
1213                 INIT_LIST_HEAD(&cic->queue_list);
1214                 cic->dtor = cfq_free_io_context;
1215                 cic->exit = cfq_exit_io_context;
1216                 elv_ioc_count_inc(ioc_count);
1217         }
1218
1219         return cic;
1220 }
1221
1222 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1223 {
1224         struct task_struct *tsk = current;
1225         int ioprio_class;
1226
1227         if (!cfq_cfqq_prio_changed(cfqq))
1228                 return;
1229
1230         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1231         switch (ioprio_class) {
1232                 default:
1233                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1234                 case IOPRIO_CLASS_NONE:
1235                         /*
1236                          * no prio set, place us in the middle of the BE classes
1237                          */
1238                         cfqq->ioprio = task_nice_ioprio(tsk);
1239                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1240                         break;
1241                 case IOPRIO_CLASS_RT:
1242                         cfqq->ioprio = task_ioprio(tsk);
1243                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1244                         break;
1245                 case IOPRIO_CLASS_BE:
1246                         cfqq->ioprio = task_ioprio(tsk);
1247                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1248                         break;
1249                 case IOPRIO_CLASS_IDLE:
1250                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1251                         cfqq->ioprio = 7;
1252                         cfq_clear_cfqq_idle_window(cfqq);
1253                         break;
1254         }
1255
1256         /*
1257          * keep track of original prio settings in case we have to temporarily
1258          * elevate the priority of this queue
1259          */
1260         cfqq->org_ioprio = cfqq->ioprio;
1261         cfqq->org_ioprio_class = cfqq->ioprio_class;
1262
1263         cfq_resort_rr_list(cfqq, 0);
1264         cfq_clear_cfqq_prio_changed(cfqq);
1265 }
1266
1267 static inline void changed_ioprio(struct cfq_io_context *cic)
1268 {
1269         struct cfq_data *cfqd = cic->key;
1270         struct cfq_queue *cfqq;
1271         unsigned long flags;
1272
1273         if (unlikely(!cfqd))
1274                 return;
1275
1276         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1277
1278         cfqq = cic->cfqq[ASYNC];
1279         if (cfqq) {
1280                 struct cfq_queue *new_cfqq;
1281                 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1282                                          GFP_ATOMIC);
1283                 if (new_cfqq) {
1284                         cic->cfqq[ASYNC] = new_cfqq;
1285                         cfq_put_queue(cfqq);
1286                 }
1287         }
1288
1289         cfqq = cic->cfqq[SYNC];
1290         if (cfqq)
1291                 cfq_mark_cfqq_prio_changed(cfqq);
1292
1293         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1294 }
1295
1296 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1297 {
1298         struct cfq_io_context *cic;
1299         struct rb_node *n;
1300
1301         ioc->ioprio_changed = 0;
1302
1303         n = rb_first(&ioc->cic_root);
1304         while (n != NULL) {
1305                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1306
1307                 changed_ioprio(cic);
1308                 n = rb_next(n);
1309         }
1310 }
1311
1312 static struct cfq_queue *
1313 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1314               gfp_t gfp_mask)
1315 {
1316         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1317         struct cfq_queue *cfqq, *new_cfqq = NULL;
1318         unsigned short ioprio;
1319
1320 retry:
1321         ioprio = tsk->ioprio;
1322         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1323
1324         if (!cfqq) {
1325                 if (new_cfqq) {
1326                         cfqq = new_cfqq;
1327                         new_cfqq = NULL;
1328                 } else if (gfp_mask & __GFP_WAIT) {
1329                         /*
1330                          * Inform the allocator of the fact that we will
1331                          * just repeat this allocation if it fails, to allow
1332                          * the allocator to do whatever it needs to attempt to
1333                          * free memory.
1334                          */
1335                         spin_unlock_irq(cfqd->queue->queue_lock);
1336                         new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
1337                         spin_lock_irq(cfqd->queue->queue_lock);
1338                         goto retry;
1339                 } else {
1340                         cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
1341                         if (!cfqq)
1342                                 goto out;
1343                 }
1344
1345                 memset(cfqq, 0, sizeof(*cfqq));
1346
1347                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1348                 INIT_LIST_HEAD(&cfqq->cfq_list);
1349                 INIT_LIST_HEAD(&cfqq->fifo);
1350
1351                 cfqq->key = key;
1352                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1353                 atomic_set(&cfqq->ref, 0);
1354                 cfqq->cfqd = cfqd;
1355                 /*
1356                  * set ->slice_left to allow preemption for a new process
1357                  */
1358                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1359                 cfq_mark_cfqq_idle_window(cfqq);
1360                 cfq_mark_cfqq_prio_changed(cfqq);
1361                 cfq_mark_cfqq_queue_new(cfqq);
1362                 cfq_init_prio_data(cfqq);
1363         }
1364
1365         if (new_cfqq)
1366                 kmem_cache_free(cfq_pool, new_cfqq);
1367
1368         atomic_inc(&cfqq->ref);
1369 out:
1370         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1371         return cfqq;
1372 }
1373
1374 static void
1375 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1376 {
1377         WARN_ON(!list_empty(&cic->queue_list));
1378         rb_erase(&cic->rb_node, &ioc->cic_root);
1379         kmem_cache_free(cfq_ioc_pool, cic);
1380         elv_ioc_count_dec(ioc_count);
1381 }
1382
1383 static struct cfq_io_context *
1384 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1385 {
1386         struct rb_node *n;
1387         struct cfq_io_context *cic;
1388         void *k, *key = cfqd;
1389
1390 restart:
1391         n = ioc->cic_root.rb_node;
1392         while (n) {
1393                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1394                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1395                 k = cic->key;
1396                 if (unlikely(!k)) {
1397                         cfq_drop_dead_cic(ioc, cic);
1398                         goto restart;
1399                 }
1400
1401                 if (key < k)
1402                         n = n->rb_left;
1403                 else if (key > k)
1404                         n = n->rb_right;
1405                 else
1406                         return cic;
1407         }
1408
1409         return NULL;
1410 }
1411
1412 static inline void
1413 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1414              struct cfq_io_context *cic)
1415 {
1416         struct rb_node **p;
1417         struct rb_node *parent;
1418         struct cfq_io_context *__cic;
1419         unsigned long flags;
1420         void *k;
1421
1422         cic->ioc = ioc;
1423         cic->key = cfqd;
1424
1425 restart:
1426         parent = NULL;
1427         p = &ioc->cic_root.rb_node;
1428         while (*p) {
1429                 parent = *p;
1430                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1431                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1432                 k = __cic->key;
1433                 if (unlikely(!k)) {
1434                         cfq_drop_dead_cic(ioc, __cic);
1435                         goto restart;
1436                 }
1437
1438                 if (cic->key < k)
1439                         p = &(*p)->rb_left;
1440                 else if (cic->key > k)
1441                         p = &(*p)->rb_right;
1442                 else
1443                         BUG();
1444         }
1445
1446         rb_link_node(&cic->rb_node, parent, p);
1447         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1448
1449         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1450         list_add(&cic->queue_list, &cfqd->cic_list);
1451         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1452 }
1453
1454 /*
1455  * Setup general io context and cfq io context. There can be several cfq
1456  * io contexts per general io context, if this process is doing io to more
1457  * than one device managed by cfq.
1458  */
1459 static struct cfq_io_context *
1460 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1461 {
1462         struct io_context *ioc = NULL;
1463         struct cfq_io_context *cic;
1464
1465         might_sleep_if(gfp_mask & __GFP_WAIT);
1466
1467         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1468         if (!ioc)
1469                 return NULL;
1470
1471         cic = cfq_cic_rb_lookup(cfqd, ioc);
1472         if (cic)
1473                 goto out;
1474
1475         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1476         if (cic == NULL)
1477                 goto err;
1478
1479         cfq_cic_link(cfqd, ioc, cic);
1480 out:
1481         smp_read_barrier_depends();
1482         if (unlikely(ioc->ioprio_changed))
1483                 cfq_ioc_set_ioprio(ioc);
1484
1485         return cic;
1486 err:
1487         put_io_context(ioc);
1488         return NULL;
1489 }
1490
1491 static void
1492 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1493 {
1494         unsigned long elapsed = jiffies - cic->last_end_request;
1495         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1496
1497         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1498         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1499         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1500 }
1501
1502 static void
1503 cfq_update_io_seektime(struct cfq_io_context *cic, struct request *rq)
1504 {
1505         sector_t sdist;
1506         u64 total;
1507
1508         if (cic->last_request_pos < rq->sector)
1509                 sdist = rq->sector - cic->last_request_pos;
1510         else
1511                 sdist = cic->last_request_pos - rq->sector;
1512
1513         /*
1514          * Don't allow the seek distance to get too large from the
1515          * odd fragment, pagein, etc
1516          */
1517         if (cic->seek_samples <= 60) /* second&third seek */
1518                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1519         else
1520                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1521
1522         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1523         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1524         total = cic->seek_total + (cic->seek_samples/2);
1525         do_div(total, cic->seek_samples);
1526         cic->seek_mean = (sector_t)total;
1527 }
1528
1529 /*
1530  * Disable idle window if the process thinks too long or seeks so much that
1531  * it doesn't matter
1532  */
1533 static void
1534 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1535                        struct cfq_io_context *cic)
1536 {
1537         int enable_idle = cfq_cfqq_idle_window(cfqq);
1538
1539         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1540             (cfqd->hw_tag && CIC_SEEKY(cic)))
1541                 enable_idle = 0;
1542         else if (sample_valid(cic->ttime_samples)) {
1543                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1544                         enable_idle = 0;
1545                 else
1546                         enable_idle = 1;
1547         }
1548
1549         if (enable_idle)
1550                 cfq_mark_cfqq_idle_window(cfqq);
1551         else
1552                 cfq_clear_cfqq_idle_window(cfqq);
1553 }
1554
1555
1556 /*
1557  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1558  * no or if we aren't sure, a 1 will cause a preempt.
1559  */
1560 static int
1561 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1562                    struct request *rq)
1563 {
1564         struct cfq_queue *cfqq = cfqd->active_queue;
1565
1566         if (cfq_class_idle(new_cfqq))
1567                 return 0;
1568
1569         if (!cfqq)
1570                 return 0;
1571
1572         if (cfq_class_idle(cfqq))
1573                 return 1;
1574         if (!cfq_cfqq_wait_request(new_cfqq))
1575                 return 0;
1576         /*
1577          * if it doesn't have slice left, forget it
1578          */
1579         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1580                 return 0;
1581         /*
1582          * if the new request is sync, but the currently running queue is
1583          * not, let the sync request have priority.
1584          */
1585         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1586                 return 1;
1587         /*
1588          * So both queues are sync. Let the new request get disk time if
1589          * it's a metadata request and the current queue is doing regular IO.
1590          */
1591         if (rq_is_meta(rq) && !cfqq->meta_pending)
1592                 return 1;
1593
1594         return 0;
1595 }
1596
1597 /*
1598  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1599  * let it have half of its nominal slice.
1600  */
1601 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1602 {
1603         cfq_slice_expired(cfqd, 1);
1604
1605         if (!cfqq->slice_left)
1606                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1607
1608         /*
1609          * Put the new queue at the front of the of the current list,
1610          * so we know that it will be selected next.
1611          */
1612         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1613         list_move(&cfqq->cfq_list, &cfqd->cur_rr);
1614
1615         cfqq->slice_end = cfqq->slice_left + jiffies;
1616 }
1617
1618 /*
1619  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1620  * something we should do about it
1621  */
1622 static void
1623 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1624                 struct request *rq)
1625 {
1626         struct cfq_io_context *cic = RQ_CIC(rq);
1627
1628         if (rq_is_meta(rq))
1629                 cfqq->meta_pending++;
1630
1631         /*
1632          * check if this request is a better next-serve candidate)) {
1633          */
1634         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1635         BUG_ON(!cfqq->next_rq);
1636
1637         /*
1638          * we never wait for an async request and we don't allow preemption
1639          * of an async request. so just return early
1640          */
1641         if (!rq_is_sync(rq)) {
1642                 /*
1643                  * sync process issued an async request, if it's waiting
1644                  * then expire it and kick rq handling.
1645                  */
1646                 if (cic == cfqd->active_cic &&
1647                     del_timer(&cfqd->idle_slice_timer)) {
1648                         cfq_slice_expired(cfqd, 0);
1649                         blk_start_queueing(cfqd->queue);
1650                 }
1651                 return;
1652         }
1653
1654         cfq_update_io_thinktime(cfqd, cic);
1655         cfq_update_io_seektime(cic, rq);
1656         cfq_update_idle_window(cfqd, cfqq, cic);
1657
1658         cic->last_request_pos = rq->sector + rq->nr_sectors;
1659
1660         if (cfqq == cfqd->active_queue) {
1661                 /*
1662                  * if we are waiting for a request for this queue, let it rip
1663                  * immediately and flag that we must not expire this queue
1664                  * just now
1665                  */
1666                 if (cfq_cfqq_wait_request(cfqq)) {
1667                         cfq_mark_cfqq_must_dispatch(cfqq);
1668                         del_timer(&cfqd->idle_slice_timer);
1669                         blk_start_queueing(cfqd->queue);
1670                 }
1671         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1672                 /*
1673                  * not the active queue - expire current slice if it is
1674                  * idle and has expired it's mean thinktime or this new queue
1675                  * has some old slice time left and is of higher priority
1676                  */
1677                 cfq_preempt_queue(cfqd, cfqq);
1678                 cfq_mark_cfqq_must_dispatch(cfqq);
1679                 blk_start_queueing(cfqd->queue);
1680         }
1681 }
1682
1683 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1684 {
1685         struct cfq_data *cfqd = q->elevator->elevator_data;
1686         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1687
1688         cfq_init_prio_data(cfqq);
1689
1690         cfq_add_rq_rb(rq);
1691
1692         list_add_tail(&rq->queuelist, &cfqq->fifo);
1693
1694         cfq_rq_enqueued(cfqd, cfqq, rq);
1695 }
1696
1697 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1698 {
1699         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1700         struct cfq_data *cfqd = cfqq->cfqd;
1701         const int sync = rq_is_sync(rq);
1702         unsigned long now;
1703
1704         now = jiffies;
1705
1706         WARN_ON(!cfqd->rq_in_driver);
1707         WARN_ON(!cfqq->on_dispatch[sync]);
1708         cfqd->rq_in_driver--;
1709         cfqq->on_dispatch[sync]--;
1710         cfqq->service_last = now;
1711
1712         if (!cfq_class_idle(cfqq))
1713                 cfqd->last_end_request = now;
1714
1715         cfq_resort_rr_list(cfqq, 0);
1716
1717         if (sync)
1718                 RQ_CIC(rq)->last_end_request = now;
1719
1720         /*
1721          * If this is the active queue, check if it needs to be expired,
1722          * or if we want to idle in case it has no pending requests.
1723          */
1724         if (cfqd->active_queue == cfqq) {
1725                 if (time_after(now, cfqq->slice_end))
1726                         cfq_slice_expired(cfqd, 0);
1727                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1728                         if (!cfq_arm_slice_timer(cfqd, cfqq))
1729                                 cfq_schedule_dispatch(cfqd);
1730                 }
1731         }
1732 }
1733
1734 /*
1735  * we temporarily boost lower priority queues if they are holding fs exclusive
1736  * resources. they are boosted to normal prio (CLASS_BE/4)
1737  */
1738 static void cfq_prio_boost(struct cfq_queue *cfqq)
1739 {
1740         const int ioprio_class = cfqq->ioprio_class;
1741         const int ioprio = cfqq->ioprio;
1742
1743         if (has_fs_excl()) {
1744                 /*
1745                  * boost idle prio on transactions that would lock out other
1746                  * users of the filesystem
1747                  */
1748                 if (cfq_class_idle(cfqq))
1749                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1750                 if (cfqq->ioprio > IOPRIO_NORM)
1751                         cfqq->ioprio = IOPRIO_NORM;
1752         } else {
1753                 /*
1754                  * check if we need to unboost the queue
1755                  */
1756                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1757                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1758                 if (cfqq->ioprio != cfqq->org_ioprio)
1759                         cfqq->ioprio = cfqq->org_ioprio;
1760         }
1761
1762         /*
1763          * refile between round-robin lists if we moved the priority class
1764          */
1765         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio))
1766                 cfq_resort_rr_list(cfqq, 0);
1767 }
1768
1769 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1770 {
1771         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1772             !cfq_cfqq_must_alloc_slice(cfqq)) {
1773                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1774                 return ELV_MQUEUE_MUST;
1775         }
1776
1777         return ELV_MQUEUE_MAY;
1778 }
1779
1780 static int cfq_may_queue(request_queue_t *q, int rw)
1781 {
1782         struct cfq_data *cfqd = q->elevator->elevator_data;
1783         struct task_struct *tsk = current;
1784         struct cfq_queue *cfqq;
1785         unsigned int key;
1786
1787         key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
1788
1789         /*
1790          * don't force setup of a queue from here, as a call to may_queue
1791          * does not necessarily imply that a request actually will be queued.
1792          * so just lookup a possibly existing queue, or return 'may queue'
1793          * if that fails
1794          */
1795         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
1796         if (cfqq) {
1797                 cfq_init_prio_data(cfqq);
1798                 cfq_prio_boost(cfqq);
1799
1800                 return __cfq_may_queue(cfqq);
1801         }
1802
1803         return ELV_MQUEUE_MAY;
1804 }
1805
1806 /*
1807  * queue lock held here
1808  */
1809 static void cfq_put_request(struct request *rq)
1810 {
1811         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1812
1813         if (cfqq) {
1814                 const int rw = rq_data_dir(rq);
1815
1816                 BUG_ON(!cfqq->allocated[rw]);
1817                 cfqq->allocated[rw]--;
1818
1819                 put_io_context(RQ_CIC(rq)->ioc);
1820
1821                 rq->elevator_private = NULL;
1822                 rq->elevator_private2 = NULL;
1823
1824                 cfq_put_queue(cfqq);
1825         }
1826 }
1827
1828 /*
1829  * Allocate cfq data structures associated with this request.
1830  */
1831 static int
1832 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1833 {
1834         struct cfq_data *cfqd = q->elevator->elevator_data;
1835         struct task_struct *tsk = current;
1836         struct cfq_io_context *cic;
1837         const int rw = rq_data_dir(rq);
1838         const int is_sync = rq_is_sync(rq);
1839         pid_t key = cfq_queue_pid(tsk, rw, is_sync);
1840         struct cfq_queue *cfqq;
1841         unsigned long flags;
1842
1843         might_sleep_if(gfp_mask & __GFP_WAIT);
1844
1845         cic = cfq_get_io_context(cfqd, gfp_mask);
1846
1847         spin_lock_irqsave(q->queue_lock, flags);
1848
1849         if (!cic)
1850                 goto queue_fail;
1851
1852         if (!cic->cfqq[is_sync]) {
1853                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1854                 if (!cfqq)
1855                         goto queue_fail;
1856
1857                 cic->cfqq[is_sync] = cfqq;
1858         } else
1859                 cfqq = cic->cfqq[is_sync];
1860
1861         cfqq->allocated[rw]++;
1862         cfq_clear_cfqq_must_alloc(cfqq);
1863         atomic_inc(&cfqq->ref);
1864
1865         spin_unlock_irqrestore(q->queue_lock, flags);
1866
1867         rq->elevator_private = cic;
1868         rq->elevator_private2 = cfqq;
1869         return 0;
1870
1871 queue_fail:
1872         if (cic)
1873                 put_io_context(cic->ioc);
1874
1875         cfq_schedule_dispatch(cfqd);
1876         spin_unlock_irqrestore(q->queue_lock, flags);
1877         return 1;
1878 }
1879
1880 static void cfq_kick_queue(struct work_struct *work)
1881 {
1882         struct cfq_data *cfqd =
1883                 container_of(work, struct cfq_data, unplug_work);
1884         request_queue_t *q = cfqd->queue;
1885         unsigned long flags;
1886
1887         spin_lock_irqsave(q->queue_lock, flags);
1888         blk_start_queueing(q);
1889         spin_unlock_irqrestore(q->queue_lock, flags);
1890 }
1891
1892 /*
1893  * Timer running if the active_queue is currently idling inside its time slice
1894  */
1895 static void cfq_idle_slice_timer(unsigned long data)
1896 {
1897         struct cfq_data *cfqd = (struct cfq_data *) data;
1898         struct cfq_queue *cfqq;
1899         unsigned long flags;
1900
1901         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1902
1903         if ((cfqq = cfqd->active_queue) != NULL) {
1904                 unsigned long now = jiffies;
1905
1906                 /*
1907                  * expired
1908                  */
1909                 if (time_after(now, cfqq->slice_end))
1910                         goto expire;
1911
1912                 /*
1913                  * only expire and reinvoke request handler, if there are
1914                  * other queues with pending requests
1915                  */
1916                 if (!cfqd->busy_queues)
1917                         goto out_cont;
1918
1919                 /*
1920                  * not expired and it has a request pending, let it dispatch
1921                  */
1922                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1923                         cfq_mark_cfqq_must_dispatch(cfqq);
1924                         goto out_kick;
1925                 }
1926         }
1927 expire:
1928         cfq_slice_expired(cfqd, 0);
1929 out_kick:
1930         cfq_schedule_dispatch(cfqd);
1931 out_cont:
1932         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1933 }
1934
1935 /*
1936  * Timer running if an idle class queue is waiting for service
1937  */
1938 static void cfq_idle_class_timer(unsigned long data)
1939 {
1940         struct cfq_data *cfqd = (struct cfq_data *) data;
1941         unsigned long flags, end;
1942
1943         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1944
1945         /*
1946          * race with a non-idle queue, reset timer
1947          */
1948         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
1949         if (!time_after_eq(jiffies, end))
1950                 mod_timer(&cfqd->idle_class_timer, end);
1951         else
1952                 cfq_schedule_dispatch(cfqd);
1953
1954         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1955 }
1956
1957 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1958 {
1959         del_timer_sync(&cfqd->idle_slice_timer);
1960         del_timer_sync(&cfqd->idle_class_timer);
1961         blk_sync_queue(cfqd->queue);
1962 }
1963
1964 static void cfq_exit_queue(elevator_t *e)
1965 {
1966         struct cfq_data *cfqd = e->elevator_data;
1967         request_queue_t *q = cfqd->queue;
1968
1969         cfq_shutdown_timer_wq(cfqd);
1970
1971         spin_lock_irq(q->queue_lock);
1972
1973         if (cfqd->active_queue)
1974                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
1975
1976         while (!list_empty(&cfqd->cic_list)) {
1977                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1978                                                         struct cfq_io_context,
1979                                                         queue_list);
1980
1981                 __cfq_exit_single_io_context(cfqd, cic);
1982         }
1983
1984         spin_unlock_irq(q->queue_lock);
1985
1986         cfq_shutdown_timer_wq(cfqd);
1987
1988         kfree(cfqd->cfq_hash);
1989         kfree(cfqd);
1990 }
1991
1992 static void *cfq_init_queue(request_queue_t *q)
1993 {
1994         struct cfq_data *cfqd;
1995         int i;
1996
1997         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
1998         if (!cfqd)
1999                 return NULL;
2000
2001         memset(cfqd, 0, sizeof(*cfqd));
2002
2003         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2004                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2005
2006         INIT_LIST_HEAD(&cfqd->busy_rr);
2007         INIT_LIST_HEAD(&cfqd->cur_rr);
2008         INIT_LIST_HEAD(&cfqd->idle_rr);
2009         INIT_LIST_HEAD(&cfqd->cic_list);
2010
2011         cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
2012         if (!cfqd->cfq_hash)
2013                 goto out_free;
2014
2015         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2016                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2017
2018         cfqd->queue = q;
2019
2020         init_timer(&cfqd->idle_slice_timer);
2021         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2022         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2023
2024         init_timer(&cfqd->idle_class_timer);
2025         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2026         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2027
2028         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2029
2030         cfqd->cfq_quantum = cfq_quantum;
2031         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2032         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2033         cfqd->cfq_back_max = cfq_back_max;
2034         cfqd->cfq_back_penalty = cfq_back_penalty;
2035         cfqd->cfq_slice[0] = cfq_slice_async;
2036         cfqd->cfq_slice[1] = cfq_slice_sync;
2037         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2038         cfqd->cfq_slice_idle = cfq_slice_idle;
2039
2040         return cfqd;
2041 out_free:
2042         kfree(cfqd);
2043         return NULL;
2044 }
2045
2046 static void cfq_slab_kill(void)
2047 {
2048         if (cfq_pool)
2049                 kmem_cache_destroy(cfq_pool);
2050         if (cfq_ioc_pool)
2051                 kmem_cache_destroy(cfq_ioc_pool);
2052 }
2053
2054 static int __init cfq_slab_setup(void)
2055 {
2056         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2057                                         NULL, NULL);
2058         if (!cfq_pool)
2059                 goto fail;
2060
2061         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2062                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2063         if (!cfq_ioc_pool)
2064                 goto fail;
2065
2066         return 0;
2067 fail:
2068         cfq_slab_kill();
2069         return -ENOMEM;
2070 }
2071
2072 /*
2073  * sysfs parts below -->
2074  */
2075
2076 static ssize_t
2077 cfq_var_show(unsigned int var, char *page)
2078 {
2079         return sprintf(page, "%d\n", var);
2080 }
2081
2082 static ssize_t
2083 cfq_var_store(unsigned int *var, const char *page, size_t count)
2084 {
2085         char *p = (char *) page;
2086
2087         *var = simple_strtoul(p, &p, 10);
2088         return count;
2089 }
2090
2091 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2092 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2093 {                                                                       \
2094         struct cfq_data *cfqd = e->elevator_data;                       \
2095         unsigned int __data = __VAR;                                    \
2096         if (__CONV)                                                     \
2097                 __data = jiffies_to_msecs(__data);                      \
2098         return cfq_var_show(__data, (page));                            \
2099 }
2100 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2101 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2102 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2103 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2104 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2105 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2106 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2107 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2108 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2109 #undef SHOW_FUNCTION
2110
2111 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2112 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2113 {                                                                       \
2114         struct cfq_data *cfqd = e->elevator_data;                       \
2115         unsigned int __data;                                            \
2116         int ret = cfq_var_store(&__data, (page), count);                \
2117         if (__data < (MIN))                                             \
2118                 __data = (MIN);                                         \
2119         else if (__data > (MAX))                                        \
2120                 __data = (MAX);                                         \
2121         if (__CONV)                                                     \
2122                 *(__PTR) = msecs_to_jiffies(__data);                    \
2123         else                                                            \
2124                 *(__PTR) = __data;                                      \
2125         return ret;                                                     \
2126 }
2127 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2128 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2129 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2130 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2131 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2132 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2133 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2134 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2135 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2136 #undef STORE_FUNCTION
2137
2138 #define CFQ_ATTR(name) \
2139         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2140
2141 static struct elv_fs_entry cfq_attrs[] = {
2142         CFQ_ATTR(quantum),
2143         CFQ_ATTR(fifo_expire_sync),
2144         CFQ_ATTR(fifo_expire_async),
2145         CFQ_ATTR(back_seek_max),
2146         CFQ_ATTR(back_seek_penalty),
2147         CFQ_ATTR(slice_sync),
2148         CFQ_ATTR(slice_async),
2149         CFQ_ATTR(slice_async_rq),
2150         CFQ_ATTR(slice_idle),
2151         __ATTR_NULL
2152 };
2153
2154 static struct elevator_type iosched_cfq = {
2155         .ops = {
2156                 .elevator_merge_fn =            cfq_merge,
2157                 .elevator_merged_fn =           cfq_merged_request,
2158                 .elevator_merge_req_fn =        cfq_merged_requests,
2159                 .elevator_allow_merge_fn =      cfq_allow_merge,
2160                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2161                 .elevator_add_req_fn =          cfq_insert_request,
2162                 .elevator_activate_req_fn =     cfq_activate_request,
2163                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2164                 .elevator_queue_empty_fn =      cfq_queue_empty,
2165                 .elevator_completed_req_fn =    cfq_completed_request,
2166                 .elevator_former_req_fn =       elv_rb_former_request,
2167                 .elevator_latter_req_fn =       elv_rb_latter_request,
2168                 .elevator_set_req_fn =          cfq_set_request,
2169                 .elevator_put_req_fn =          cfq_put_request,
2170                 .elevator_may_queue_fn =        cfq_may_queue,
2171                 .elevator_init_fn =             cfq_init_queue,
2172                 .elevator_exit_fn =             cfq_exit_queue,
2173                 .trim =                         cfq_free_io_context,
2174         },
2175         .elevator_attrs =       cfq_attrs,
2176         .elevator_name =        "cfq",
2177         .elevator_owner =       THIS_MODULE,
2178 };
2179
2180 static int __init cfq_init(void)
2181 {
2182         int ret;
2183
2184         /*
2185          * could be 0 on HZ < 1000 setups
2186          */
2187         if (!cfq_slice_async)
2188                 cfq_slice_async = 1;
2189         if (!cfq_slice_idle)
2190                 cfq_slice_idle = 1;
2191
2192         if (cfq_slab_setup())
2193                 return -ENOMEM;
2194
2195         ret = elv_register(&iosched_cfq);
2196         if (ret)
2197                 cfq_slab_kill();
2198
2199         return ret;
2200 }
2201
2202 static void __exit cfq_exit(void)
2203 {
2204         DECLARE_COMPLETION_ONSTACK(all_gone);
2205         elv_unregister(&iosched_cfq);
2206         ioc_gone = &all_gone;
2207         /* ioc_gone's update must be visible before reading ioc_count */
2208         smp_wmb();
2209         if (elv_ioc_count_read(ioc_count))
2210                 wait_for_completion(ioc_gone);
2211         synchronize_rcu();
2212         cfq_slab_kill();
2213 }
2214
2215 module_init(cfq_init);
2216 module_exit(cfq_exit);
2217
2218 MODULE_AUTHOR("Jens Axboe");
2219 MODULE_LICENSE("GPL");
2220 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");