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_sync(cfqq)     ((cfqq)->key != CFQ_KEY_ASYNC)
  60
  61 #define sample_valid(samples)   ((samples) > 80)
  62
  63 /*
  64  * Per block device queue structure
  65  */
  66 struct cfq_data {
  67         request_queue_t *queue;
  68
  69         /*
  70          * rr list of queues with requests and the count of them
  71          */
  72         struct list_head rr_list[CFQ_PRIO_LISTS];
  73         struct list_head busy_rr;
  74         struct list_head cur_rr;
  75         struct list_head idle_rr;
  76         unsigned long cur_rr_tick;
  77         unsigned int busy_queues;
  78
  79         /*
  80          * cfqq lookup hash
  81          */
  82         struct hlist_head *cfq_hash;
  83
  84         int rq_in_driver;
  85         int hw_tag;
  86
  87         /*
  88          * idle window management
  89          */
  90         struct timer_list idle_slice_timer;
  91         struct work_struct unplug_work;
  92
  93         struct cfq_queue *active_queue;
  94         struct cfq_io_context *active_cic;
  95         int cur_prio, cur_end_prio;
  96         unsigned long prio_time;
  97         unsigned int dispatch_slice;
  98
  99         struct timer_list idle_class_timer;
 100
 101         sector_t last_position;
 102         unsigned long last_end_request;
 103
 104         /*
 105          * tunables, see top of file
 106          */
 107         unsigned int cfq_quantum;
 108         unsigned int cfq_fifo_expire[2];
 109         unsigned int cfq_back_penalty;
 110         unsigned int cfq_back_max;
 111         unsigned int cfq_slice[2];
 112         unsigned int cfq_slice_async_rq;
 113         unsigned int cfq_slice_idle;
 114
 115         struct list_head cic_list;
 116
 117         sector_t new_seek_mean;
 118         u64 new_seek_total;
 119 };
 120
 121 /*
 122  * Per process-grouping structure
 123  */
 124 struct cfq_queue {
 125         /* reference count */
 126         atomic_t ref;
 127         /* parent cfq_data */
 128         struct cfq_data *cfqd;
 129         /* cfqq lookup hash */
 130         struct hlist_node cfq_hash;
 131         /* hash key */
 132         unsigned int key;
 133         /* member of the rr/busy/cur/idle cfqd list */
 134         struct list_head cfq_list;
 135         /* in what tick we were last serviced */
 136         unsigned long rr_tick;
 137         /* sorted list of pending requests */
 138         struct rb_root sort_list;
 139         /* if fifo isn't expired, next request to serve */
 140         struct request *next_rq;
 141         /* requests queued in sort_list */
 142         int queued[2];
 143         /* currently allocated requests */
 144         int allocated[2];
 145         /* pending metadata requests */
 146         int meta_pending;
 147         /* fifo list of requests in sort_list */
 148         struct list_head fifo;
 149
 150         unsigned long slice_end;
 151         unsigned long service_last;
 152         unsigned long slice_start;
 153         long slice_resid;
 154
 155         /* number of requests that are on the dispatch list or inside driver */
 156         int dispatched;
 157
 158         /* io prio of this group */
 159         unsigned short ioprio, org_ioprio;
 160         unsigned short ioprio_class, org_ioprio_class;
 161
 162         /* various state flags, see below */
 163         unsigned int flags;
 164
 165         sector_t last_request_pos;
 166 };
 167
 168 enum cfqq_state_flags {
 169         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 170         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 171         CFQ_CFQQ_FLAG_must_alloc,       /* must be allowed rq alloc */
 172         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 173         CFQ_CFQQ_FLAG_must_dispatch,    /* must dispatch, even if expired */
 174         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 175         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 176         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 177         CFQ_CFQQ_FLAG_queue_new,        /* queue never been serviced */
 178         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 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 CFQ_CFQQ_FNS(queue_new);
 204 CFQ_CFQQ_FNS(slice_new);
 205 #undef CFQ_CFQQ_FNS
 206
 207 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 208 static void cfq_dispatch_insert(request_queue_t *, struct request *);
 209 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
 210
 211 /*
 212  * scheduler run of queue, if there are requests pending and no one in the
 213  * driver that will restart queueing
 214  */
 215 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 216 {
 217         if (cfqd->busy_queues)
 218                 kblockd_schedule_work(&cfqd->unplug_work);
 219 }
 220
 221 static int cfq_queue_empty(request_queue_t *q)
 222 {
 223         struct cfq_data *cfqd = q->elevator->elevator_data;
 224
 225         return !cfqd->busy_queues;
 226 }
 227
 228 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw, int is_sync)
 229 {
 230         /*
 231          * Use the per-process queue, for read requests and syncronous writes
 232          */
 233         if (!(rw & REQ_RW) || is_sync)
 234                 return task->pid;
 235
 236         return CFQ_KEY_ASYNC;
 237 }
 238
 239 /*
 240  * Scale schedule slice based on io priority. Use the sync time slice only
 241  * if a queue is marked sync and has sync io queued. A sync queue with async
 242  * io only, should not get full sync slice length.
 243  */
 244 static inline int
 245 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 246 {
 247         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 248
 249         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 250
 251         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
 252 }
 253
 254 static inline void
 255 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 256 {
 257         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 258         cfqq->slice_end += cfqq->slice_resid;
 259
 260         /*
 261          * Don't carry over residual for more than one slice, we only want
 262          * to slightly correct the fairness. Carrying over forever would
 263          * easily introduce oscillations.
 264          */
 265         cfqq->slice_resid = 0;
 266
 267         cfqq->slice_start = jiffies;
 268 }
 269
 270 /*
 271  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
 272  * isn't valid until the first request from the dispatch is activated
 273  * and the slice time set.
 274  */
 275 static inline int cfq_slice_used(struct cfq_queue *cfqq)
 276 {
 277         if (cfq_cfqq_slice_new(cfqq))
 278                 return 0;
 279         if (time_before(jiffies, cfqq->slice_end))
 280                 return 0;
 281
 282         return 1;
 283 }
 284
 285 /*
 286  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 287  * We choose the request that is closest to the head right now. Distance
 288  * behind the head is penalized and only allowed to a certain extent.
 289  */
 290 static struct request *
 291 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
 292 {
 293         sector_t last, s1, s2, d1 = 0, d2 = 0;
 294         unsigned long back_max;
 295 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 296 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 297         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 298
 299         if (rq1 == NULL || rq1 == rq2)
 300                 return rq2;
 301         if (rq2 == NULL)
 302                 return rq1;
 303
 304         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 305                 return rq1;
 306         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 307                 return rq2;
 308         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
 309                 return rq1;
 310         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
 311                 return rq2;
 312
 313         s1 = rq1->sector;
 314         s2 = rq2->sector;
 315
 316         last = cfqd->last_position;
 317
 318         /*
 319          * by definition, 1KiB is 2 sectors
 320          */
 321         back_max = cfqd->cfq_back_max * 2;
 322
 323         /*
 324          * Strict one way elevator _except_ in the case where we allow
 325          * short backward seeks which are biased as twice the cost of a
 326          * similar forward seek.
 327          */
 328         if (s1 >= last)
 329                 d1 = s1 - last;
 330         else if (s1 + back_max >= last)
 331                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 332         else
 333                 wrap |= CFQ_RQ1_WRAP;
 334
 335         if (s2 >= last)
 336                 d2 = s2 - last;
 337         else if (s2 + back_max >= last)
 338                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 339         else
 340                 wrap |= CFQ_RQ2_WRAP;
 341
 342         /* Found required data */
 343
 344         /*
 345          * By doing switch() on the bit mask "wrap" we avoid having to
 346          * check two variables for all permutations: --> faster!
 347          */
 348         switch (wrap) {
 349         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 350                 if (d1 < d2)
 351                         return rq1;
 352                 else if (d2 < d1)
 353                         return rq2;
 354                 else {
 355                         if (s1 >= s2)
 356                                 return rq1;
 357                         else
 358                                 return rq2;
 359                 }
 360
 361         case CFQ_RQ2_WRAP:
 362                 return rq1;
 363         case CFQ_RQ1_WRAP:
 364                 return rq2;
 365         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
 366         default:
 367                 /*
 368                  * Since both rqs are wrapped,
 369                  * start with the one that's further behind head
 370                  * (--> only *one* back seek required),
 371                  * since back seek takes more time than forward.
 372                  */
 373                 if (s1 <= s2)
 374                         return rq1;
 375                 else
 376                         return rq2;
 377         }
 378 }
 379
 380 /*
 381  * would be nice to take fifo expire time into account as well
 382  */
 383 static struct request *
 384 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 385                   struct request *last)
 386 {
 387         struct rb_node *rbnext = rb_next(&last->rb_node);
 388         struct rb_node *rbprev = rb_prev(&last->rb_node);
 389         struct request *next = NULL, *prev = NULL;
 390
 391         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
 392
 393         if (rbprev)
 394                 prev = rb_entry_rq(rbprev);
 395
 396         if (rbnext)
 397                 next = rb_entry_rq(rbnext);
 398         else {
 399                 rbnext = rb_first(&cfqq->sort_list);
 400                 if (rbnext && rbnext != &last->rb_node)
 401                         next = rb_entry_rq(rbnext);
 402         }
 403
 404         return cfq_choose_req(cfqd, next, prev);
 405 }
 406
 407 /*
 408  * This function finds out where to insert a BE queue in the service hierarchy
 409  */
 410 static void cfq_resort_be_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 411                                 int preempted)
 412 {
 413         struct list_head *list, *n;
 414         struct cfq_queue *__cfqq;
 415         int add_tail = 0;
 416
 417         /*
 418          * if cfqq has requests in flight, don't allow it to be
 419          * found in cfq_set_active_queue before it has finished them.
 420          * this is done to increase fairness between a process that
 421          * has lots of io pending vs one that only generates one
 422          * sporadically or synchronously
 423          */
 424         if (cfqq->dispatched)
 425                 list = &cfqd->busy_rr;
 426         else if (cfqq->ioprio == (cfqd->cur_prio + 1) &&
 427                  cfq_cfqq_sync(cfqq) &&
 428                  (time_before(cfqd->prio_time, cfqq->service_last) ||
 429                   cfq_cfqq_queue_new(cfqq) || preempted)) {
 430                 list = &cfqd->cur_rr;
 431                 add_tail = 1;
 432         } else
 433                 list = &cfqd->rr_list[cfqq->ioprio];
 434
 435         if (!cfq_cfqq_sync(cfqq) || add_tail) {
 436                 /*
 437                  * async queue always goes to the end. this wont be overly
 438                  * unfair to writes, as the sort of the sync queue wont be
 439                  * allowed to pass the async queue again.
 440                  */
 441                 list_add_tail(&cfqq->cfq_list, list);
 442         } else if (preempted || cfq_cfqq_queue_new(cfqq)) {
 443                 /*
 444                  * If this queue was preempted or is new (never been serviced),
 445                  * let it be added first for fairness but beind other new
 446                  * queues.
 447                  */
 448                 n = list;
 449                 while (n->next != list) {
 450                         __cfqq = list_entry_cfqq(n->next);
 451                         if (!cfq_cfqq_queue_new(__cfqq))
 452                                 break;
 453
 454                         n = n->next;
 455                 }
 456                 list_add(&cfqq->cfq_list, n);
 457         } else {
 458                 /*
 459                  * sort by last service, but don't cross a new or async
 460                  * queue. we don't cross a new queue because it hasn't been
 461                  * service before, and we don't cross an async queue because
 462                  * it gets added to the end on expire.
 463                  */
 464                 n = list;
 465                 while ((n = n->prev) != list) {
 466                         struct cfq_queue *__c = list_entry_cfqq(n);
 467
 468                         if (!cfq_cfqq_sync(__c) || !__c->service_last)
 469                                 break;
 470                         if (time_before(__c->service_last, cfqq->service_last))
 471                                 break;
 472                 }
 473                 list_add(&cfqq->cfq_list, n);
 474         }
 475 }
 476
 477 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 478 {
 479         struct cfq_data *cfqd = cfqq->cfqd;
 480         struct list_head *n;
 481
 482         /*
 483          * Resorting requires the cfqq to be on the RR list already.
 484          */
 485         if (!cfq_cfqq_on_rr(cfqq))
 486                 return;
 487
 488         list_del(&cfqq->cfq_list);
 489
 490         if (cfq_class_rt(cfqq)) {
 491                 /*
 492                  * At to the front of the current list, but behind other
 493                  * RT queues.
 494                  */
 495                 n = &cfqd->cur_rr;
 496                 while (n->next != &cfqd->cur_rr)
 497                         if (!cfq_class_rt(cfqq))
 498                                 break;
 499
 500                 list_add(&cfqq->cfq_list, n);
 501         } else if (cfq_class_idle(cfqq)) {
 502                 /*
 503                  * IDLE goes to the tail of the idle list
 504                  */
 505                 list_add_tail(&cfqq->cfq_list, &cfqd->idle_rr);
 506         } else {
 507                 /*
 508                  * So we get here, ergo the queue is a regular best-effort queue
 509                  */
 510                 cfq_resort_be_queue(cfqd, cfqq, preempted);
 511         }
 512 }
 513
 514 /*
 515  * add to busy list of queues for service, trying to be fair in ordering
 516  * the pending list according to last request service
 517  */
 518 static inline void
 519 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 520 {
 521         BUG_ON(cfq_cfqq_on_rr(cfqq));
 522         cfq_mark_cfqq_on_rr(cfqq);
 523         cfqd->busy_queues++;
 524
 525         cfq_resort_rr_list(cfqq, 0);
 526 }
 527
 528 static inline void
 529 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 530 {
 531         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 532         cfq_clear_cfqq_on_rr(cfqq);
 533         list_del_init(&cfqq->cfq_list);
 534
 535         BUG_ON(!cfqd->busy_queues);
 536         cfqd->busy_queues--;
 537 }
 538
 539 /*
 540  * rb tree support functions
 541  */
 542 static inline void cfq_del_rq_rb(struct request *rq)
 543 {
 544         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 545         struct cfq_data *cfqd = cfqq->cfqd;
 546         const int sync = rq_is_sync(rq);
 547
 548         BUG_ON(!cfqq->queued[sync]);
 549         cfqq->queued[sync]--;
 550
 551         elv_rb_del(&cfqq->sort_list, rq);
 552
 553         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
 554                 cfq_del_cfqq_rr(cfqd, cfqq);
 555 }
 556
 557 static void cfq_add_rq_rb(struct request *rq)
 558 {
 559         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 560         struct cfq_data *cfqd = cfqq->cfqd;
 561         struct request *__alias;
 562
 563         cfqq->queued[rq_is_sync(rq)]++;
 564
 565         /*
 566          * looks a little odd, but the first insert might return an alias.
 567          * if that happens, put the alias on the dispatch list
 568          */
 569         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
 570                 cfq_dispatch_insert(cfqd->queue, __alias);
 571
 572         if (!cfq_cfqq_on_rr(cfqq))
 573                 cfq_add_cfqq_rr(cfqd, cfqq);
 574
 575         /*
 576          * check if this request is a better next-serve candidate
 577          */
 578         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
 579         BUG_ON(!cfqq->next_rq);
 580 }
 581
 582 static inline void
 583 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
 584 {
 585         elv_rb_del(&cfqq->sort_list, rq);
 586         cfqq->queued[rq_is_sync(rq)]--;
 587         cfq_add_rq_rb(rq);
 588 }
 589
 590 static struct request *
 591 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 592 {
 593         struct task_struct *tsk = current;
 594         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
 595         struct cfq_queue *cfqq;
 596
 597         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
 598         if (cfqq) {
 599                 sector_t sector = bio->bi_sector + bio_sectors(bio);
 600
 601                 return elv_rb_find(&cfqq->sort_list, sector);
 602         }
 603
 604         return NULL;
 605 }
 606
 607 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 608 {
 609         struct cfq_data *cfqd = q->elevator->elevator_data;
 610
 611         cfqd->rq_in_driver++;
 612
 613         /*
 614          * If the depth is larger 1, it really could be queueing. But lets
 615          * make the mark a little higher - idling could still be good for
 616          * low queueing, and a low queueing number could also just indicate
 617          * a SCSI mid layer like behaviour where limit+1 is often seen.
 618          */
 619         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
 620                 cfqd->hw_tag = 1;
 621
 622         cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
 623 }
 624
 625 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 626 {
 627         struct cfq_data *cfqd = q->elevator->elevator_data;
 628
 629         WARN_ON(!cfqd->rq_in_driver);
 630         cfqd->rq_in_driver--;
 631 }
 632
 633 static void cfq_remove_request(struct request *rq)
 634 {
 635         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 636
 637         if (cfqq->next_rq == rq)
 638                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
 639
 640         list_del_init(&rq->queuelist);
 641         cfq_del_rq_rb(rq);
 642
 643         if (rq_is_meta(rq)) {
 644                 WARN_ON(!cfqq->meta_pending);
 645                 cfqq->meta_pending--;
 646         }
 647 }
 648
 649 static int
 650 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 651 {
 652         struct cfq_data *cfqd = q->elevator->elevator_data;
 653         struct request *__rq;
 654
 655         __rq = cfq_find_rq_fmerge(cfqd, bio);
 656         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 657                 *req = __rq;
 658                 return ELEVATOR_FRONT_MERGE;
 659         }
 660
 661         return ELEVATOR_NO_MERGE;
 662 }
 663
 664 static void cfq_merged_request(request_queue_t *q, struct request *req,
 665                                int type)
 666 {
 667         if (type == ELEVATOR_FRONT_MERGE) {
 668                 struct cfq_queue *cfqq = RQ_CFQQ(req);
 669
 670                 cfq_reposition_rq_rb(cfqq, req);
 671         }
 672 }
 673
 674 static void
 675 cfq_merged_requests(request_queue_t *q, struct request *rq,
 676                     struct request *next)
 677 {
 678         /*
 679          * reposition in fifo if next is older than rq
 680          */
 681         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 682             time_before(next->start_time, rq->start_time))
 683                 list_move(&rq->queuelist, &next->queuelist);
 684
 685         cfq_remove_request(next);
 686 }
 687
 688 static int cfq_allow_merge(request_queue_t *q, struct request *rq,
 689                            struct bio *bio)
 690 {
 691         struct cfq_data *cfqd = q->elevator->elevator_data;
 692         const int rw = bio_data_dir(bio);
 693         struct cfq_queue *cfqq;
 694         pid_t key;
 695
 696         /*
 697          * Disallow merge of a sync bio into an async request.
 698          */
 699         if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq))
 700                 return 0;
 701
 702         /*
 703          * Lookup the cfqq that this bio will be queued with. Allow
 704          * merge only if rq is queued there.
 705          */
 706         key = cfq_queue_pid(current, rw, bio_sync(bio));
 707         cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio);
 708
 709         if (cfqq == RQ_CFQQ(rq))
 710                 return 1;
 711
 712         return 0;
 713 }
 714
 715 static inline void
 716 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 717 {
 718         if (cfqq) {
 719                 /*
 720                  * stop potential idle class queues waiting service
 721                  */
 722                 del_timer(&cfqd->idle_class_timer);
 723
 724                 cfqq->slice_end = 0;
 725                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 726                 cfq_clear_cfqq_fifo_expire(cfqq);
 727                 cfq_mark_cfqq_slice_new(cfqq);
 728                 cfqq->rr_tick = cfqd->cur_rr_tick;
 729         }
 730
 731         cfqd->active_queue = cfqq;
 732 }
 733
 734 /*
 735  * current cfqq expired its slice (or was too idle), select new one
 736  */
 737 static void
 738 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 739                     int preempted, int timed_out)
 740 {
 741         if (cfq_cfqq_wait_request(cfqq))
 742                 del_timer(&cfqd->idle_slice_timer);
 743
 744         cfq_clear_cfqq_must_dispatch(cfqq);
 745         cfq_clear_cfqq_wait_request(cfqq);
 746         cfq_clear_cfqq_queue_new(cfqq);
 747
 748         /*
 749          * store what was left of this slice, if the queue idled out
 750          * or was preempted
 751          */
 752         if (timed_out && !cfq_cfqq_slice_new(cfqq))
 753                 cfqq->slice_resid = cfqq->slice_end - jiffies;
 754
 755         cfq_resort_rr_list(cfqq, preempted);
 756
 757         if (cfqq == cfqd->active_queue)
 758                 cfqd->active_queue = NULL;
 759
 760         if (cfqd->active_cic) {
 761                 put_io_context(cfqd->active_cic->ioc);
 762                 cfqd->active_cic = NULL;
 763         }
 764
 765         cfqd->dispatch_slice = 0;
 766 }
 767
 768 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted,
 769                                      int timed_out)
 770 {
 771         struct cfq_queue *cfqq = cfqd->active_queue;
 772
 773         if (cfqq)
 774                 __cfq_slice_expired(cfqd, cfqq, preempted, timed_out);
 775 }
 776
 777 /*
 778  * 0
 779  * 0,1
 780  * 0,1,2
 781  * 0,1,2,3
 782  * 0,1,2,3,4
 783  * 0,1,2,3,4,5
 784  * 0,1,2,3,4,5,6
 785  * 0,1,2,3,4,5,6,7
 786  */
 787 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 788 {
 789         int prio, wrap;
 790
 791         prio = -1;
 792         wrap = 0;
 793         do {
 794                 int p;
 795
 796                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 797                         if (!list_empty(&cfqd->rr_list[p])) {
 798                                 prio = p;
 799                                 break;
 800                         }
 801                 }
 802
 803                 if (prio != -1)
 804                         break;
 805                 cfqd->cur_prio = 0;
 806                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 807                         cfqd->cur_end_prio = 0;
 808                         if (wrap)
 809                                 break;
 810                         wrap = 1;
 811                 }
 812         } while (1);
 813
 814         if (unlikely(prio == -1))
 815                 return -1;
 816
 817         BUG_ON(prio >= CFQ_PRIO_LISTS);
 818
 819         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 820
 821         cfqd->cur_prio = prio + 1;
 822         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 823                 cfqd->cur_end_prio = cfqd->cur_prio;
 824                 cfqd->cur_prio = 0;
 825         }
 826         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 827                 cfqd->cur_prio = 0;
 828                 cfqd->cur_end_prio = 0;
 829         }
 830
 831         cfqd->cur_rr_tick++;
 832         cfqd->prio_time = jiffies;
 833         return prio;
 834 }
 835
 836 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
 837                                           struct request *rq)
 838 {
 839         if (rq->sector >= cfqd->last_position)
 840                 return rq->sector - cfqd->last_position;
 841         else
 842                 return cfqd->last_position - rq->sector;
 843 }
 844
 845 static struct cfq_queue *cfq_get_best_queue(struct cfq_data *cfqd)
 846 {
 847         struct cfq_queue *cfqq = NULL, *__cfqq;
 848         sector_t best = -1, dist;
 849
 850         list_for_each_entry(__cfqq, &cfqd->cur_rr, cfq_list) {
 851                 if (!__cfqq->next_rq || !cfq_cfqq_sync(__cfqq))
 852                         continue;
 853
 854                 dist = cfq_dist_from_last(cfqd, __cfqq->next_rq);
 855                 if (dist < best) {
 856                         best = dist;
 857                         cfqq = __cfqq;
 858                 }
 859         }
 860
 861         /*
 862          * Only async queue(s) available, grab first entry
 863          */
 864         if (!cfqq)
 865                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 866
 867         return cfqq;
 868 }
 869
 870 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
 871 {
 872         struct cfq_queue *cfqq = NULL;
 873
 874         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
 875                 /*
 876                  * if current list is non-empty, grab first entry. if it is
 877                  * empty, get next prio level and grab first entry then if any
 878                  * are spliced
 879                  */
 880                 cfqq = cfq_get_best_queue(cfqd);
 881         } else if (!list_empty(&cfqd->busy_rr)) {
 882                 /*
 883                  * If no new queues are available, check if the busy list has
 884                  * some before falling back to idle io.
 885                  */
 886                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
 887         } else if (!list_empty(&cfqd->idle_rr)) {
 888                 /*
 889                  * if we have idle queues and no rt or be queues had pending
 890                  * requests, either allow immediate service if the grace period
 891                  * has passed or arm the idle grace timer
 892                  */
 893                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 894
 895                 if (time_after_eq(jiffies, end))
 896                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 897                 else
 898                         mod_timer(&cfqd->idle_class_timer, end);
 899         }
 900
 901         return cfqq;
 902 }
 903
 904 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 905 {
 906         struct cfq_queue *cfqq;
 907
 908         do {
 909                 long prio;
 910
 911                 cfqq = cfq_get_next_queue(cfqd);
 912                 if (!cfqq)
 913                         break;
 914
 915                 prio = cfq_prio_to_slice(cfqd, cfqq);
 916                 if (cfqq->slice_resid > -prio)
 917                         break;
 918
 919                 cfqq->slice_resid += prio;
 920                 list_del_init(&cfqq->cfq_list);
 921                 list_add_tail(&cfqq->cfq_list, &cfqd->rr_list[cfqq->ioprio]);
 922                 cfqq = NULL;
 923         } while (1);
 924
 925         __cfq_set_active_queue(cfqd, cfqq);
 926         return cfqq;
 927 }
 928
 929 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
 930 {
 931         struct cfq_io_context *cic = cfqd->active_cic;
 932
 933         if (!sample_valid(cic->seek_samples))
 934                 return 0;
 935
 936         return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
 937 }
 938
 939 static struct cfq_queue *__cfq_close_cooperator(struct cfq_data *cfqd,
 940                                                 struct cfq_queue *cur_cfqq,
 941                                                 struct list_head *list)
 942 {
 943         struct cfq_queue *cfqq;
 944
 945         list_for_each_entry(cfqq, list, cfq_list) {
 946                 if (cfqq == cur_cfqq || !cfq_cfqq_sync(cfqq))
 947                         continue;
 948
 949                 BUG_ON(!cfqq->next_rq);
 950
 951                 if (cfq_rq_close(cfqd, cfqq->next_rq))
 952                         return cfqq;
 953         }
 954
 955         return NULL;
 956 }
 957
 958 static int cfq_close_cooperator(struct cfq_data *cfqd,
 959                                 struct cfq_queue *cur_cfqq)
 960 {
 961         struct cfq_queue *cfqq;
 962
 963         if (!cfqd->busy_queues)
 964                 return 0;
 965
 966         /*
 967          * check cur_rr and same-prio rr_list for candidates
 968          */
 969         cfqq = __cfq_close_cooperator(cfqd, cur_cfqq, &cfqd->cur_rr);
 970         if (cfqq)
 971                 return 1;
 972
 973         cfqq = __cfq_close_cooperator(cfqd, cur_cfqq, &cfqd->rr_list[cur_cfqq->ioprio]);
 974         if (cfqq && (cfqq->rr_tick == cfqd->cur_rr_tick))
 975                 cfqq = NULL;
 976
 977         return cfqq != NULL;
 978 }
 979
 980 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
 981
 982 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
 983 {
 984         struct cfq_queue *cfqq = cfqd->active_queue;
 985         struct cfq_io_context *cic;
 986         unsigned long sl;
 987
 988         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
 989         WARN_ON(cfq_cfqq_slice_new(cfqq));
 990
 991         /*
 992          * idle is disabled, either manually or by past process history
 993          */
 994         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
 995                 return;
 996
 997         /*
 998          * task has exited, don't wait
 999          */
1000         cic = cfqd->active_cic;
1001         if (!cic || !cic->ioc->task)
1002                 return;
1003
1004         /*
1005          * See if this prio level has a good candidate
1006          */
1007         if (cfq_close_cooperator(cfqd, cfqq))
1008                 return;
1009
1010         cfq_mark_cfqq_must_dispatch(cfqq);
1011         cfq_mark_cfqq_wait_request(cfqq);
1012
1013         /*
1014          * we don't want to idle for seeks, but we do want to allow
1015          * fair distribution of slice time for a process doing back-to-back
1016          * seeks. so allow a little bit of time for him to submit a new rq
1017          */
1018         sl = cfqd->cfq_slice_idle;
1019         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
1020                 sl = min(sl, msecs_to_jiffies(2));
1021
1022         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1023 }
1024
1025 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
1026 {
1027         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1028
1029         cfq_remove_request(rq);
1030         cfqq->dispatched++;
1031         elv_dispatch_sort(q, rq);
1032 }
1033
1034 /*
1035  * return expired entry, or NULL to just start from scratch in rbtree
1036  */
1037 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1038 {
1039         struct cfq_data *cfqd = cfqq->cfqd;
1040         struct request *rq;
1041         int fifo;
1042
1043         if (cfq_cfqq_fifo_expire(cfqq))
1044                 return NULL;
1045
1046         cfq_mark_cfqq_fifo_expire(cfqq);
1047
1048         if (list_empty(&cfqq->fifo))
1049                 return NULL;
1050
1051         fifo = cfq_cfqq_sync(cfqq);
1052         rq = rq_entry_fifo(cfqq->fifo.next);
1053
1054         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
1055                 return NULL;
1056
1057         return rq;
1058 }
1059
1060 static inline int
1061 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1062 {
1063         const int base_rq = cfqd->cfq_slice_async_rq;
1064
1065         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1066
1067         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1068 }
1069
1070 /*
1071  * get next queue for service
1072  */
1073 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1074 {
1075         struct cfq_queue *cfqq;
1076
1077         cfqq = cfqd->active_queue;
1078         if (!cfqq)
1079                 goto new_queue;
1080
1081         /*
1082          * The active queue has run out of time, expire it and select new.
1083          */
1084         if (cfq_slice_used(cfqq))
1085                 goto expire;
1086
1087         /*
1088          * The active queue has requests and isn't expired, allow it to
1089          * dispatch.
1090          */
1091         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1092                 goto keep_queue;
1093
1094         /*
1095          * No requests pending. If the active queue still has requests in
1096          * flight or is idling for a new request, allow either of these
1097          * conditions to happen (or time out) before selecting a new queue.
1098          */
1099         if (cfqq->dispatched || timer_pending(&cfqd->idle_slice_timer)) {
1100                 cfqq = NULL;
1101                 goto keep_queue;
1102         }
1103
1104 expire:
1105         cfq_slice_expired(cfqd, 0, 0);
1106 new_queue:
1107         cfqq = cfq_set_active_queue(cfqd);
1108 keep_queue:
1109         return cfqq;
1110 }
1111
1112 static int
1113 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1114                         int max_dispatch)
1115 {
1116         int dispatched = 0;
1117
1118         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1119
1120         do {
1121                 struct request *rq;
1122
1123                 /*
1124                  * follow expired path, else get first next available
1125                  */
1126                 if ((rq = cfq_check_fifo(cfqq)) == NULL)
1127                         rq = cfqq->next_rq;
1128
1129                 /*
1130                  * finally, insert request into driver dispatch list
1131                  */
1132                 cfq_dispatch_insert(cfqd->queue, rq);
1133
1134                 cfqd->dispatch_slice++;
1135                 dispatched++;
1136
1137                 if (!cfqd->active_cic) {
1138                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1139                         cfqd->active_cic = RQ_CIC(rq);
1140                 }
1141
1142                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1143                         break;
1144
1145         } while (dispatched < max_dispatch);
1146
1147         /*
1148          * expire an async queue immediately if it has used up its slice. idle
1149          * queue always expire after 1 dispatch round.
1150          */
1151         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1152             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1153             cfq_class_idle(cfqq))) {
1154                 cfqq->slice_end = jiffies + 1;
1155                 cfq_slice_expired(cfqd, 0, 0);
1156         }
1157
1158         return dispatched;
1159 }
1160
1161 static int
1162 cfq_forced_dispatch_cfqqs(struct list_head *list)
1163 {
1164         struct cfq_queue *cfqq, *next;
1165         int dispatched;
1166
1167         dispatched = 0;
1168         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1169                 while (cfqq->next_rq) {
1170                         cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1171                         dispatched++;
1172                 }
1173                 BUG_ON(!list_empty(&cfqq->fifo));
1174         }
1175
1176         return dispatched;
1177 }
1178
1179 static int
1180 cfq_forced_dispatch(struct cfq_data *cfqd)
1181 {
1182         int i, dispatched = 0;
1183
1184         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1185                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1186
1187         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1188         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1189         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1190
1191         cfq_slice_expired(cfqd, 0, 0);
1192
1193         BUG_ON(cfqd->busy_queues);
1194
1195         return dispatched;
1196 }
1197
1198 static int
1199 cfq_dispatch_requests(request_queue_t *q, int force)
1200 {
1201         struct cfq_data *cfqd = q->elevator->elevator_data;
1202         struct cfq_queue *cfqq;
1203         int dispatched;
1204
1205         if (!cfqd->busy_queues)
1206                 return 0;
1207
1208         if (unlikely(force))
1209                 return cfq_forced_dispatch(cfqd);
1210
1211         dispatched = 0;
1212         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1213                 int max_dispatch;
1214
1215                 if (cfqd->busy_queues > 1) {
1216                         /*
1217                          * So we have dispatched before in this round, if the
1218                          * next queue has idling enabled (must be sync), don't
1219                          * allow it service until the previous have completed.
1220                          */
1221                         if (cfqd->rq_in_driver && cfq_cfqq_idle_window(cfqq) &&
1222                             dispatched)
1223                                 break;
1224                         if (cfqq->dispatched >= cfqd->cfq_quantum)
1225                                 break;
1226                 }
1227
1228                 cfq_clear_cfqq_must_dispatch(cfqq);
1229                 cfq_clear_cfqq_wait_request(cfqq);
1230                 del_timer(&cfqd->idle_slice_timer);
1231
1232                 max_dispatch = cfqd->cfq_quantum;
1233                 if (cfq_class_idle(cfqq))
1234                         max_dispatch = 1;
1235
1236                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1237         }
1238
1239         return dispatched;
1240 }
1241
1242 /*
1243  * task holds one reference to the queue, dropped when task exits. each rq
1244  * in-flight on this queue also holds a reference, dropped when rq is freed.
1245  *
1246  * queue lock must be held here.
1247  */
1248 static void cfq_put_queue(struct cfq_queue *cfqq)
1249 {
1250         struct cfq_data *cfqd = cfqq->cfqd;
1251
1252         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1253
1254         if (!atomic_dec_and_test(&cfqq->ref))
1255                 return;
1256
1257         BUG_ON(rb_first(&cfqq->sort_list));
1258         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1259         BUG_ON(cfq_cfqq_on_rr(cfqq));
1260
1261         if (unlikely(cfqd->active_queue == cfqq)) {
1262                 __cfq_slice_expired(cfqd, cfqq, 0, 0);
1263                 cfq_schedule_dispatch(cfqd);
1264         }
1265
1266         /*
1267          * it's on the empty list and still hashed
1268          */
1269         list_del(&cfqq->cfq_list);
1270         hlist_del(&cfqq->cfq_hash);
1271         kmem_cache_free(cfq_pool, cfqq);
1272 }
1273
1274 static struct cfq_queue *
1275 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1276                     const int hashval)
1277 {
1278         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1279         struct hlist_node *entry;
1280         struct cfq_queue *__cfqq;
1281
1282         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1283                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1284
1285                 if (__cfqq->key == key && (__p == prio || !prio))
1286                         return __cfqq;
1287         }
1288
1289         return NULL;
1290 }
1291
1292 static struct cfq_queue *
1293 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1294 {
1295         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1296 }
1297
1298 static void cfq_free_io_context(struct io_context *ioc)
1299 {
1300         struct cfq_io_context *__cic;
1301         struct rb_node *n;
1302         int freed = 0;
1303
1304         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1305                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1306                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1307                 kmem_cache_free(cfq_ioc_pool, __cic);
1308                 freed++;
1309         }
1310
1311         elv_ioc_count_mod(ioc_count, -freed);
1312
1313         if (ioc_gone && !elv_ioc_count_read(ioc_count))
1314                 complete(ioc_gone);
1315 }
1316
1317 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1318 {
1319         if (unlikely(cfqq == cfqd->active_queue)) {
1320                 __cfq_slice_expired(cfqd, cfqq, 0, 0);
1321                 cfq_schedule_dispatch(cfqd);
1322         }
1323
1324         cfq_put_queue(cfqq);
1325 }
1326
1327 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1328                                          struct cfq_io_context *cic)
1329 {
1330         list_del_init(&cic->queue_list);
1331         smp_wmb();
1332         cic->key = NULL;
1333
1334         if (cic->cfqq[ASYNC]) {
1335                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1336                 cic->cfqq[ASYNC] = NULL;
1337         }
1338
1339         if (cic->cfqq[SYNC]) {
1340                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1341                 cic->cfqq[SYNC] = NULL;
1342         }
1343 }
1344
1345
1346 /*
1347  * Called with interrupts disabled
1348  */
1349 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1350 {
1351         struct cfq_data *cfqd = cic->key;
1352
1353         if (cfqd) {
1354                 request_queue_t *q = cfqd->queue;
1355
1356                 spin_lock_irq(q->queue_lock);
1357                 __cfq_exit_single_io_context(cfqd, cic);
1358                 spin_unlock_irq(q->queue_lock);
1359         }
1360 }
1361
1362 static void cfq_exit_io_context(struct io_context *ioc)
1363 {
1364         struct cfq_io_context *__cic;
1365         struct rb_node *n;
1366
1367         /*
1368          * put the reference this task is holding to the various queues
1369          */
1370
1371         n = rb_first(&ioc->cic_root);
1372         while (n != NULL) {
1373                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1374
1375                 cfq_exit_single_io_context(__cic);
1376                 n = rb_next(n);
1377         }
1378 }
1379
1380 static struct cfq_io_context *
1381 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1382 {
1383         struct cfq_io_context *cic;
1384
1385         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
1386         if (cic) {
1387                 memset(cic, 0, sizeof(*cic));
1388                 cic->last_end_request = jiffies;
1389                 INIT_LIST_HEAD(&cic->queue_list);
1390                 cic->dtor = cfq_free_io_context;
1391                 cic->exit = cfq_exit_io_context;
1392                 elv_ioc_count_inc(ioc_count);
1393         }
1394
1395         return cic;
1396 }
1397
1398 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1399 {
1400         struct task_struct *tsk = current;
1401         int ioprio_class;
1402
1403         if (!cfq_cfqq_prio_changed(cfqq))
1404                 return;
1405
1406         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1407         switch (ioprio_class) {
1408                 default:
1409                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1410                 case IOPRIO_CLASS_NONE:
1411                         /*
1412                          * no prio set, place us in the middle of the BE classes
1413                          */
1414                         cfqq->ioprio = task_nice_ioprio(tsk);
1415                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1416                         break;
1417                 case IOPRIO_CLASS_RT:
1418                         cfqq->ioprio = task_ioprio(tsk);
1419                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1420                         break;
1421                 case IOPRIO_CLASS_BE:
1422                         cfqq->ioprio = task_ioprio(tsk);
1423                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1424                         break;
1425                 case IOPRIO_CLASS_IDLE:
1426                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1427                         cfqq->ioprio = 7;
1428                         cfq_clear_cfqq_idle_window(cfqq);
1429                         break;
1430         }
1431
1432         /*
1433          * keep track of original prio settings in case we have to temporarily
1434          * elevate the priority of this queue
1435          */
1436         cfqq->org_ioprio = cfqq->ioprio;
1437         cfqq->org_ioprio_class = cfqq->ioprio_class;
1438
1439         cfq_resort_rr_list(cfqq, 0);
1440         cfq_clear_cfqq_prio_changed(cfqq);
1441 }
1442
1443 static inline void changed_ioprio(struct cfq_io_context *cic)
1444 {
1445         struct cfq_data *cfqd = cic->key;
1446         struct cfq_queue *cfqq;
1447         unsigned long flags;
1448
1449         if (unlikely(!cfqd))
1450                 return;
1451
1452         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1453
1454         cfqq = cic->cfqq[ASYNC];
1455         if (cfqq) {
1456                 struct cfq_queue *new_cfqq;
1457                 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1458                                          GFP_ATOMIC);
1459                 if (new_cfqq) {
1460                         cic->cfqq[ASYNC] = new_cfqq;
1461                         cfq_put_queue(cfqq);
1462                 }
1463         }
1464
1465         cfqq = cic->cfqq[SYNC];
1466         if (cfqq)
1467                 cfq_mark_cfqq_prio_changed(cfqq);
1468
1469         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1470 }
1471
1472 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1473 {
1474         struct cfq_io_context *cic;
1475         struct rb_node *n;
1476
1477         ioc->ioprio_changed = 0;
1478
1479         n = rb_first(&ioc->cic_root);
1480         while (n != NULL) {
1481                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1482
1483                 changed_ioprio(cic);
1484                 n = rb_next(n);
1485         }
1486 }
1487
1488 static struct cfq_queue *
1489 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1490               gfp_t gfp_mask)
1491 {
1492         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1493         struct cfq_queue *cfqq, *new_cfqq = NULL;
1494         unsigned short ioprio;
1495
1496 retry:
1497         ioprio = tsk->ioprio;
1498         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1499
1500         if (!cfqq) {
1501                 if (new_cfqq) {
1502                         cfqq = new_cfqq;
1503                         new_cfqq = NULL;
1504                 } else if (gfp_mask & __GFP_WAIT) {
1505                         /*
1506                          * Inform the allocator of the fact that we will
1507                          * just repeat this allocation if it fails, to allow
1508                          * the allocator to do whatever it needs to attempt to
1509                          * free memory.
1510                          */
1511                         spin_unlock_irq(cfqd->queue->queue_lock);
1512                         new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
1513                         spin_lock_irq(cfqd->queue->queue_lock);
1514                         goto retry;
1515                 } else {
1516                         cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
1517                         if (!cfqq)
1518                                 goto out;
1519                 }
1520
1521                 memset(cfqq, 0, sizeof(*cfqq));
1522
1523                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1524                 INIT_LIST_HEAD(&cfqq->cfq_list);
1525                 INIT_LIST_HEAD(&cfqq->fifo);
1526
1527                 cfqq->key = key;
1528                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1529                 atomic_set(&cfqq->ref, 0);
1530                 cfqq->cfqd = cfqd;
1531
1532                 if (key != CFQ_KEY_ASYNC)
1533                         cfq_mark_cfqq_idle_window(cfqq);
1534
1535                 cfq_mark_cfqq_prio_changed(cfqq);
1536                 cfq_mark_cfqq_queue_new(cfqq);
1537                 cfq_init_prio_data(cfqq);
1538         }
1539
1540         if (new_cfqq)
1541                 kmem_cache_free(cfq_pool, new_cfqq);
1542
1543         atomic_inc(&cfqq->ref);
1544 out:
1545         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1546         return cfqq;
1547 }
1548
1549 static void
1550 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1551 {
1552         WARN_ON(!list_empty(&cic->queue_list));
1553         rb_erase(&cic->rb_node, &ioc->cic_root);
1554         kmem_cache_free(cfq_ioc_pool, cic);
1555         elv_ioc_count_dec(ioc_count);
1556 }
1557
1558 static struct cfq_io_context *
1559 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1560 {
1561         struct rb_node *n;
1562         struct cfq_io_context *cic;
1563         void *k, *key = cfqd;
1564
1565 restart:
1566         n = ioc->cic_root.rb_node;
1567         while (n) {
1568                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1569                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1570                 k = cic->key;
1571                 if (unlikely(!k)) {
1572                         cfq_drop_dead_cic(ioc, cic);
1573                         goto restart;
1574                 }
1575
1576                 if (key < k)
1577                         n = n->rb_left;
1578                 else if (key > k)
1579                         n = n->rb_right;
1580                 else
1581                         return cic;
1582         }
1583
1584         return NULL;
1585 }
1586
1587 static inline void
1588 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1589              struct cfq_io_context *cic)
1590 {
1591         struct rb_node **p;
1592         struct rb_node *parent;
1593         struct cfq_io_context *__cic;
1594         unsigned long flags;
1595         void *k;
1596
1597         cic->ioc = ioc;
1598         cic->key = cfqd;
1599
1600 restart:
1601         parent = NULL;
1602         p = &ioc->cic_root.rb_node;
1603         while (*p) {
1604                 parent = *p;
1605                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1606                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1607                 k = __cic->key;
1608                 if (unlikely(!k)) {
1609                         cfq_drop_dead_cic(ioc, __cic);
1610                         goto restart;
1611                 }
1612
1613                 if (cic->key < k)
1614                         p = &(*p)->rb_left;
1615                 else if (cic->key > k)
1616                         p = &(*p)->rb_right;
1617                 else
1618                         BUG();
1619         }
1620
1621         rb_link_node(&cic->rb_node, parent, p);
1622         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1623
1624         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1625         list_add(&cic->queue_list, &cfqd->cic_list);
1626         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1627 }
1628
1629 /*
1630  * Setup general io context and cfq io context. There can be several cfq
1631  * io contexts per general io context, if this process is doing io to more
1632  * than one device managed by cfq.
1633  */
1634 static struct cfq_io_context *
1635 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1636 {
1637         struct io_context *ioc = NULL;
1638         struct cfq_io_context *cic;
1639
1640         might_sleep_if(gfp_mask & __GFP_WAIT);
1641
1642         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1643         if (!ioc)
1644                 return NULL;
1645
1646         cic = cfq_cic_rb_lookup(cfqd, ioc);
1647         if (cic)
1648                 goto out;
1649
1650         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1651         if (cic == NULL)
1652                 goto err;
1653
1654         cfq_cic_link(cfqd, ioc, cic);
1655 out:
1656         smp_read_barrier_depends();
1657         if (unlikely(ioc->ioprio_changed))
1658                 cfq_ioc_set_ioprio(ioc);
1659
1660         return cic;
1661 err:
1662         put_io_context(ioc);
1663         return NULL;
1664 }
1665
1666 static void
1667 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1668 {
1669         unsigned long elapsed = jiffies - cic->last_end_request;
1670         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1671
1672         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1673         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1674         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1675 }
1676
1677 static void
1678 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1679                        struct request *rq)
1680 {
1681         sector_t sdist;
1682         u64 total;
1683
1684         if (cic->last_request_pos < rq->sector)
1685                 sdist = rq->sector - cic->last_request_pos;
1686         else
1687                 sdist = cic->last_request_pos - rq->sector;
1688
1689         if (!cic->seek_samples) {
1690                 cfqd->new_seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1691                 cfqd->new_seek_mean = cfqd->new_seek_total / 256;
1692         }
1693
1694         /*
1695          * Don't allow the seek distance to get too large from the
1696          * odd fragment, pagein, etc
1697          */
1698         if (cic->seek_samples <= 60) /* second&third seek */
1699                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1700         else
1701                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1702
1703         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1704         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1705         total = cic->seek_total + (cic->seek_samples/2);
1706         do_div(total, cic->seek_samples);
1707         cic->seek_mean = (sector_t)total;
1708 }
1709
1710 /*
1711  * Disable idle window if the process thinks too long or seeks so much that
1712  * it doesn't matter
1713  */
1714 static void
1715 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1716                        struct cfq_io_context *cic)
1717 {
1718         int enable_idle = cfq_cfqq_idle_window(cfqq);
1719
1720         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1721             (cfqd->hw_tag && CIC_SEEKY(cic)))
1722                 enable_idle = 0;
1723         else if (sample_valid(cic->ttime_samples)) {
1724                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1725                         enable_idle = 0;
1726                 else
1727                         enable_idle = 1;
1728         }
1729
1730         if (enable_idle)
1731                 cfq_mark_cfqq_idle_window(cfqq);
1732         else
1733                 cfq_clear_cfqq_idle_window(cfqq);
1734 }
1735
1736 /*
1737  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1738  * no or if we aren't sure, a 1 will cause a preempt.
1739  */
1740 static int
1741 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1742                    struct request *rq)
1743 {
1744         struct cfq_queue *cfqq;
1745
1746         cfqq = cfqd->active_queue;
1747         if (!cfqq)
1748                 return 0;
1749
1750         if (cfq_slice_used(cfqq))
1751                 return 1;
1752
1753         if (cfq_class_idle(new_cfqq))
1754                 return 0;
1755
1756         if (cfq_class_idle(cfqq))
1757                 return 1;
1758
1759         /*
1760          * if the new request is sync, but the currently running queue is
1761          * not, let the sync request have priority.
1762          */
1763         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1764                 return 1;
1765
1766         /*
1767          * So both queues are sync. Let the new request get disk time if
1768          * it's a metadata request and the current queue is doing regular IO.
1769          */
1770         if (rq_is_meta(rq) && !cfqq->meta_pending)
1771                 return 1;
1772
1773         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1774                 return 0;
1775
1776         /*
1777          * if this request is as-good as one we would expect from the
1778          * current cfqq, let it preempt
1779          */
1780         if (cfq_rq_close(cfqd, rq))
1781                 return 1;
1782
1783         return 0;
1784 }
1785
1786 /*
1787  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1788  * let it have half of its nominal slice.
1789  */
1790 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1791 {
1792         cfq_slice_expired(cfqd, 1, 1);
1793
1794         /*
1795          * Put the new queue at the front of the of the current list,
1796          * so we know that it will be selected next.
1797          */
1798         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1799         list_move(&cfqq->cfq_list, &cfqd->cur_rr);
1800
1801         cfqq->slice_end = 0;
1802         cfq_mark_cfqq_slice_new(cfqq);
1803 }
1804
1805 /*
1806  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1807  * something we should do about it
1808  */
1809 static void
1810 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1811                 struct request *rq)
1812 {
1813         struct cfq_io_context *cic = RQ_CIC(rq);
1814
1815         if (rq_is_meta(rq))
1816                 cfqq->meta_pending++;
1817
1818         cfq_update_io_thinktime(cfqd, cic);
1819         cfq_update_io_seektime(cfqd, cic, rq);
1820         cfq_update_idle_window(cfqd, cfqq, cic);
1821
1822         cic->last_request_pos = rq->sector + rq->nr_sectors;
1823         cfqq->last_request_pos = cic->last_request_pos;
1824
1825         if (cfqq == cfqd->active_queue) {
1826                 /*
1827                  * if we are waiting for a request for this queue, let it rip
1828                  * immediately and flag that we must not expire this queue
1829                  * just now
1830                  */
1831                 if (cfq_cfqq_wait_request(cfqq)) {
1832                         cfq_mark_cfqq_must_dispatch(cfqq);
1833                         del_timer(&cfqd->idle_slice_timer);
1834                         blk_start_queueing(cfqd->queue);
1835                 }
1836         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1837                 /*
1838                  * not the active queue - expire current slice if it is
1839                  * idle and has expired it's mean thinktime or this new queue
1840                  * has some old slice time left and is of higher priority
1841                  */
1842                 cfq_preempt_queue(cfqd, cfqq);
1843                 cfq_mark_cfqq_must_dispatch(cfqq);
1844                 blk_start_queueing(cfqd->queue);
1845         }
1846 }
1847
1848 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1849 {
1850         struct cfq_data *cfqd = q->elevator->elevator_data;
1851         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1852
1853         cfq_init_prio_data(cfqq);
1854
1855         cfq_add_rq_rb(rq);
1856
1857         list_add_tail(&rq->queuelist, &cfqq->fifo);
1858
1859         cfq_rq_enqueued(cfqd, cfqq, rq);
1860 }
1861
1862 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1863 {
1864         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1865         struct cfq_data *cfqd = cfqq->cfqd;
1866         const int sync = rq_is_sync(rq);
1867         unsigned long now;
1868
1869         now = jiffies;
1870
1871         WARN_ON(!cfqd->rq_in_driver);
1872         WARN_ON(!cfqq->dispatched);
1873         cfqd->rq_in_driver--;
1874         cfqq->dispatched--;
1875         cfqq->service_last = now;
1876
1877         if (!cfq_class_idle(cfqq))
1878                 cfqd->last_end_request = now;
1879
1880         cfq_resort_rr_list(cfqq, 0);
1881
1882         if (sync)
1883                 RQ_CIC(rq)->last_end_request = now;
1884
1885         /*
1886          * If this is the active queue, check if it needs to be expired,
1887          * or if we want to idle in case it has no pending requests.
1888          */
1889         if (cfqd->active_queue == cfqq) {
1890                 if (cfq_cfqq_slice_new(cfqq)) {
1891                         cfq_set_prio_slice(cfqd, cfqq);
1892                         cfq_clear_cfqq_slice_new(cfqq);
1893                 }
1894                 if (cfq_slice_used(cfqq))
1895                         cfq_slice_expired(cfqd, 0, 1);
1896                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1897                         cfq_arm_slice_timer(cfqd);
1898         }
1899
1900         if (!cfqd->rq_in_driver)
1901                 cfq_schedule_dispatch(cfqd);
1902 }
1903
1904 /*
1905  * we temporarily boost lower priority queues if they are holding fs exclusive
1906  * resources. they are boosted to normal prio (CLASS_BE/4)
1907  */
1908 static void cfq_prio_boost(struct cfq_queue *cfqq)
1909 {
1910         const int ioprio_class = cfqq->ioprio_class;
1911         const int ioprio = cfqq->ioprio;
1912
1913         if (has_fs_excl()) {
1914                 /*
1915                  * boost idle prio on transactions that would lock out other
1916                  * users of the filesystem
1917                  */
1918                 if (cfq_class_idle(cfqq))
1919                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1920                 if (cfqq->ioprio > IOPRIO_NORM)
1921                         cfqq->ioprio = IOPRIO_NORM;
1922         } else {
1923                 /*
1924                  * check if we need to unboost the queue
1925                  */
1926                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1927                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1928                 if (cfqq->ioprio != cfqq->org_ioprio)
1929                         cfqq->ioprio = cfqq->org_ioprio;
1930         }
1931
1932         /*
1933          * refile between round-robin lists if we moved the priority class
1934          */
1935         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio))
1936                 cfq_resort_rr_list(cfqq, 0);
1937 }
1938
1939 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1940 {
1941         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1942             !cfq_cfqq_must_alloc_slice(cfqq)) {
1943                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1944                 return ELV_MQUEUE_MUST;
1945         }
1946
1947         return ELV_MQUEUE_MAY;
1948 }
1949
1950 static int cfq_may_queue(request_queue_t *q, int rw)
1951 {
1952         struct cfq_data *cfqd = q->elevator->elevator_data;
1953         struct task_struct *tsk = current;
1954         struct cfq_queue *cfqq;
1955         unsigned int key;
1956
1957         key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
1958
1959         /*
1960          * don't force setup of a queue from here, as a call to may_queue
1961          * does not necessarily imply that a request actually will be queued.
1962          * so just lookup a possibly existing queue, or return 'may queue'
1963          * if that fails
1964          */
1965         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
1966         if (cfqq) {
1967                 cfq_init_prio_data(cfqq);
1968                 cfq_prio_boost(cfqq);
1969
1970                 return __cfq_may_queue(cfqq);
1971         }
1972
1973         return ELV_MQUEUE_MAY;
1974 }
1975
1976 /*
1977  * queue lock held here
1978  */
1979 static void cfq_put_request(struct request *rq)
1980 {
1981         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1982
1983         if (cfqq) {
1984                 const int rw = rq_data_dir(rq);
1985
1986                 BUG_ON(!cfqq->allocated[rw]);
1987                 cfqq->allocated[rw]--;
1988
1989                 put_io_context(RQ_CIC(rq)->ioc);
1990
1991                 rq->elevator_private = NULL;
1992                 rq->elevator_private2 = NULL;
1993
1994                 cfq_put_queue(cfqq);
1995         }
1996 }
1997
1998 /*
1999  * Allocate cfq data structures associated with this request.
2000  */
2001 static int
2002 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
2003 {
2004         struct cfq_data *cfqd = q->elevator->elevator_data;
2005         struct task_struct *tsk = current;
2006         struct cfq_io_context *cic;
2007         const int rw = rq_data_dir(rq);
2008         const int is_sync = rq_is_sync(rq);
2009         pid_t key = cfq_queue_pid(tsk, rw, is_sync);
2010         struct cfq_queue *cfqq;
2011         unsigned long flags;
2012
2013         might_sleep_if(gfp_mask & __GFP_WAIT);
2014
2015         cic = cfq_get_io_context(cfqd, gfp_mask);
2016
2017         spin_lock_irqsave(q->queue_lock, flags);
2018
2019         if (!cic)
2020                 goto queue_fail;
2021
2022         if (!cic->cfqq[is_sync]) {
2023                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
2024                 if (!cfqq)
2025                         goto queue_fail;
2026
2027                 cic->cfqq[is_sync] = cfqq;
2028         } else
2029                 cfqq = cic->cfqq[is_sync];
2030
2031         cfqq->allocated[rw]++;
2032         cfq_clear_cfqq_must_alloc(cfqq);
2033         atomic_inc(&cfqq->ref);
2034
2035         spin_unlock_irqrestore(q->queue_lock, flags);
2036
2037         rq->elevator_private = cic;
2038         rq->elevator_private2 = cfqq;
2039         return 0;
2040
2041 queue_fail:
2042         if (cic)
2043                 put_io_context(cic->ioc);
2044
2045         cfq_schedule_dispatch(cfqd);
2046         spin_unlock_irqrestore(q->queue_lock, flags);
2047         return 1;
2048 }
2049
2050 static void cfq_kick_queue(struct work_struct *work)
2051 {
2052         struct cfq_data *cfqd =
2053                 container_of(work, struct cfq_data, unplug_work);
2054         request_queue_t *q = cfqd->queue;
2055         unsigned long flags;
2056
2057         spin_lock_irqsave(q->queue_lock, flags);
2058         blk_start_queueing(q);
2059         spin_unlock_irqrestore(q->queue_lock, flags);
2060 }
2061
2062 /*
2063  * Timer running if the active_queue is currently idling inside its time slice
2064  */
2065 static void cfq_idle_slice_timer(unsigned long data)
2066 {
2067         struct cfq_data *cfqd = (struct cfq_data *) data;
2068         struct cfq_queue *cfqq;
2069         unsigned long flags;
2070         int timed_out = 1;
2071
2072         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2073
2074         if ((cfqq = cfqd->active_queue) != NULL) {
2075                 timed_out = 0;
2076
2077                 /*
2078                  * expired
2079                  */
2080                 if (cfq_slice_used(cfqq))
2081                         goto expire;
2082
2083                 /*
2084                  * only expire and reinvoke request handler, if there are
2085                  * other queues with pending requests
2086                  */
2087                 if (!cfqd->busy_queues)
2088                         goto out_cont;
2089
2090                 /*
2091                  * not expired and it has a request pending, let it dispatch
2092                  */
2093                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
2094                         cfq_mark_cfqq_must_dispatch(cfqq);
2095                         goto out_kick;
2096                 }
2097         }
2098 expire:
2099         cfq_slice_expired(cfqd, 0, timed_out);
2100 out_kick:
2101         cfq_schedule_dispatch(cfqd);
2102 out_cont:
2103         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2104 }
2105
2106 /*
2107  * Timer running if an idle class queue is waiting for service
2108  */
2109 static void cfq_idle_class_timer(unsigned long data)
2110 {
2111         struct cfq_data *cfqd = (struct cfq_data *) data;
2112         unsigned long flags, end;
2113
2114         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2115
2116         /*
2117          * race with a non-idle queue, reset timer
2118          */
2119         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2120         if (!time_after_eq(jiffies, end))
2121                 mod_timer(&cfqd->idle_class_timer, end);
2122         else
2123                 cfq_schedule_dispatch(cfqd);
2124
2125         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2126 }
2127
2128 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2129 {
2130         del_timer_sync(&cfqd->idle_slice_timer);
2131         del_timer_sync(&cfqd->idle_class_timer);
2132         blk_sync_queue(cfqd->queue);
2133 }
2134
2135 static void cfq_exit_queue(elevator_t *e)
2136 {
2137         struct cfq_data *cfqd = e->elevator_data;
2138         request_queue_t *q = cfqd->queue;
2139
2140         cfq_shutdown_timer_wq(cfqd);
2141
2142         spin_lock_irq(q->queue_lock);
2143
2144         if (cfqd->active_queue)
2145                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0, 0);
2146
2147         while (!list_empty(&cfqd->cic_list)) {
2148                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2149                                                         struct cfq_io_context,
2150                                                         queue_list);
2151
2152                 __cfq_exit_single_io_context(cfqd, cic);
2153         }
2154
2155         spin_unlock_irq(q->queue_lock);
2156
2157         cfq_shutdown_timer_wq(cfqd);
2158
2159         kfree(cfqd->cfq_hash);
2160         kfree(cfqd);
2161 }
2162
2163 static void *cfq_init_queue(request_queue_t *q)
2164 {
2165         struct cfq_data *cfqd;
2166         int i;
2167
2168         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
2169         if (!cfqd)
2170                 return NULL;
2171
2172         memset(cfqd, 0, sizeof(*cfqd));
2173
2174         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2175                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2176
2177         INIT_LIST_HEAD(&cfqd->busy_rr);
2178         INIT_LIST_HEAD(&cfqd->cur_rr);
2179         INIT_LIST_HEAD(&cfqd->idle_rr);
2180         INIT_LIST_HEAD(&cfqd->cic_list);
2181
2182         cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
2183         if (!cfqd->cfq_hash)
2184                 goto out_free;
2185
2186         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2187                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2188
2189         cfqd->queue = q;
2190
2191         init_timer(&cfqd->idle_slice_timer);
2192         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2193         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2194
2195         init_timer(&cfqd->idle_class_timer);
2196         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2197         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2198
2199         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2200
2201         cfqd->cfq_quantum = cfq_quantum;
2202         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2203         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2204         cfqd->cfq_back_max = cfq_back_max;
2205         cfqd->cfq_back_penalty = cfq_back_penalty;
2206         cfqd->cfq_slice[0] = cfq_slice_async;
2207         cfqd->cfq_slice[1] = cfq_slice_sync;
2208         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2209         cfqd->cfq_slice_idle = cfq_slice_idle;
2210
2211         return cfqd;
2212 out_free:
2213         kfree(cfqd);
2214         return NULL;
2215 }
2216
2217 static void cfq_slab_kill(void)
2218 {
2219         if (cfq_pool)
2220                 kmem_cache_destroy(cfq_pool);
2221         if (cfq_ioc_pool)
2222                 kmem_cache_destroy(cfq_ioc_pool);
2223 }
2224
2225 static int __init cfq_slab_setup(void)
2226 {
2227         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2228                                         NULL, NULL);
2229         if (!cfq_pool)
2230                 goto fail;
2231
2232         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2233                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2234         if (!cfq_ioc_pool)
2235                 goto fail;
2236
2237         return 0;
2238 fail:
2239         cfq_slab_kill();
2240         return -ENOMEM;
2241 }
2242
2243 /*
2244  * sysfs parts below -->
2245  */
2246 static ssize_t
2247 cfq_var_show(unsigned int var, char *page)
2248 {
2249         return sprintf(page, "%d\n", var);
2250 }
2251
2252 static ssize_t
2253 cfq_var_store(unsigned int *var, const char *page, size_t count)
2254 {
2255         char *p = (char *) page;
2256
2257         *var = simple_strtoul(p, &p, 10);
2258         return count;
2259 }
2260
2261 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2262 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2263 {                                                                       \
2264         struct cfq_data *cfqd = e->elevator_data;                       \
2265         unsigned int __data = __VAR;                                    \
2266         if (__CONV)                                                     \
2267                 __data = jiffies_to_msecs(__data);                      \
2268         return cfq_var_show(__data, (page));                            \
2269 }
2270 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2271 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2272 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2273 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2274 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2275 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2276 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2277 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2278 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2279 #undef SHOW_FUNCTION
2280
2281 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2282 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2283 {                                                                       \
2284         struct cfq_data *cfqd = e->elevator_data;                       \
2285         unsigned int __data;                                            \
2286         int ret = cfq_var_store(&__data, (page), count);                \
2287         if (__data < (MIN))                                             \
2288                 __data = (MIN);                                         \
2289         else if (__data > (MAX))                                        \
2290                 __data = (MAX);                                         \
2291         if (__CONV)                                                     \
2292                 *(__PTR) = msecs_to_jiffies(__data);                    \
2293         else                                                            \
2294                 *(__PTR) = __data;                                      \
2295         return ret;                                                     \
2296 }
2297 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2298 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2299 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2300 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2301 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2302 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2303 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2304 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2305 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2306 #undef STORE_FUNCTION
2307
2308 #define CFQ_ATTR(name) \
2309         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2310
2311 static struct elv_fs_entry cfq_attrs[] = {
2312         CFQ_ATTR(quantum),
2313         CFQ_ATTR(fifo_expire_sync),
2314         CFQ_ATTR(fifo_expire_async),
2315         CFQ_ATTR(back_seek_max),
2316         CFQ_ATTR(back_seek_penalty),
2317         CFQ_ATTR(slice_sync),
2318         CFQ_ATTR(slice_async),
2319         CFQ_ATTR(slice_async_rq),
2320         CFQ_ATTR(slice_idle),
2321         __ATTR_NULL
2322 };
2323
2324 static struct elevator_type iosched_cfq = {
2325         .ops = {
2326                 .elevator_merge_fn =            cfq_merge,
2327                 .elevator_merged_fn =           cfq_merged_request,
2328                 .elevator_merge_req_fn =        cfq_merged_requests,
2329                 .elevator_allow_merge_fn =      cfq_allow_merge,
2330                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2331                 .elevator_add_req_fn =          cfq_insert_request,
2332                 .elevator_activate_req_fn =     cfq_activate_request,
2333                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2334                 .elevator_queue_empty_fn =      cfq_queue_empty,
2335                 .elevator_completed_req_fn =    cfq_completed_request,
2336                 .elevator_former_req_fn =       elv_rb_former_request,
2337                 .elevator_latter_req_fn =       elv_rb_latter_request,
2338                 .elevator_set_req_fn =          cfq_set_request,
2339                 .elevator_put_req_fn =          cfq_put_request,
2340                 .elevator_may_queue_fn =        cfq_may_queue,
2341                 .elevator_init_fn =             cfq_init_queue,
2342                 .elevator_exit_fn =             cfq_exit_queue,
2343                 .trim =                         cfq_free_io_context,
2344         },
2345         .elevator_attrs =       cfq_attrs,
2346         .elevator_name =        "cfq",
2347         .elevator_owner =       THIS_MODULE,
2348 };
2349
2350 static int __init cfq_init(void)
2351 {
2352         int ret;
2353
2354         /*
2355          * could be 0 on HZ < 1000 setups
2356          */
2357         if (!cfq_slice_async)
2358                 cfq_slice_async = 1;
2359         if (!cfq_slice_idle)
2360                 cfq_slice_idle = 1;
2361
2362         if (cfq_slab_setup())
2363                 return -ENOMEM;
2364
2365         ret = elv_register(&iosched_cfq);
2366         if (ret)
2367                 cfq_slab_kill();
2368
2369         return ret;
2370 }
2371
2372 static void __exit cfq_exit(void)
2373 {
2374         DECLARE_COMPLETION_ONSTACK(all_gone);
2375         elv_unregister(&iosched_cfq);
2376         ioc_gone = &all_gone;
2377         /* ioc_gone's update must be visible before reading ioc_count */
2378         smp_wmb();
2379         if (elv_ioc_count_read(ioc_count))
2380                 wait_for_completion(ioc_gone);
2381         synchronize_rcu();
2382         cfq_slab_kill();
2383 }
2384
2385 module_init(cfq_init);
2386 module_exit(cfq_exit);
2387
2388 MODULE_AUTHOR("Jens Axboe");
2389 MODULE_LICENSE("GPL");
2390 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");