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 service_last;
 151         long slice_resid;
 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         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 175 };
 176
 177 #define CFQ_CFQQ_FNS(name)                                              \
 178 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 179 {                                                                       \
 180         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 181 }                                                                       \
 182 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 183 {                                                                       \
 184         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 185 }                                                                       \
 186 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 187 {                                                                       \
 188         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 189 }
 190
 191 CFQ_CFQQ_FNS(on_rr);
 192 CFQ_CFQQ_FNS(wait_request);
 193 CFQ_CFQQ_FNS(must_alloc);
 194 CFQ_CFQQ_FNS(must_alloc_slice);
 195 CFQ_CFQQ_FNS(must_dispatch);
 196 CFQ_CFQQ_FNS(fifo_expire);
 197 CFQ_CFQQ_FNS(idle_window);
 198 CFQ_CFQQ_FNS(prio_changed);
 199 CFQ_CFQQ_FNS(queue_new);
 200 CFQ_CFQQ_FNS(slice_new);
 201 #undef CFQ_CFQQ_FNS
 202
 203 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 204 static void cfq_dispatch_insert(request_queue_t *, struct request *);
 205 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
 206
 207 /*
 208  * scheduler run of queue, if there are requests pending and no one in the
 209  * driver that will restart queueing
 210  */
 211 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 212 {
 213         if (cfqd->busy_queues)
 214                 kblockd_schedule_work(&cfqd->unplug_work);
 215 }
 216
 217 static int cfq_queue_empty(request_queue_t *q)
 218 {
 219         struct cfq_data *cfqd = q->elevator->elevator_data;
 220
 221         return !cfqd->busy_queues;
 222 }
 223
 224 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw, int is_sync)
 225 {
 226         /*
 227          * Use the per-process queue, for read requests and syncronous writes
 228          */
 229         if (!(rw & REQ_RW) || is_sync)
 230                 return task->pid;
 231
 232         return CFQ_KEY_ASYNC;
 233 }
 234
 235 /*
 236  * Scale schedule slice based on io priority. Use the sync time slice only
 237  * if a queue is marked sync and has sync io queued. A sync queue with async
 238  * io only, should not get full sync slice length.
 239  */
 240 static inline int
 241 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 242 {
 243         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 244
 245         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 246
 247         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
 248 }
 249
 250 static inline void
 251 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 252 {
 253         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 254         cfqq->slice_end += cfqq->slice_resid;
 255
 256         /*
 257          * Don't carry over residual for more than one slice, we only want
 258          * to slightly correct the fairness. Carrying over forever would
 259          * easily introduce oscillations.
 260          */
 261         cfqq->slice_resid = 0;
 262 }
 263
 264 /*
 265  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
 266  * isn't valid until the first request from the dispatch is activated
 267  * and the slice time set.
 268  */
 269 static inline int cfq_slice_used(struct cfq_queue *cfqq)
 270 {
 271         if (cfq_cfqq_slice_new(cfqq))
 272                 return 0;
 273         if (time_before(jiffies, cfqq->slice_end))
 274                 return 0;
 275
 276         return 1;
 277 }
 278
 279 /*
 280  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 281  * We choose the request that is closest to the head right now. Distance
 282  * behind the head is penalized and only allowed to a certain extent.
 283  */
 284 static struct request *
 285 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
 286 {
 287         sector_t last, s1, s2, d1 = 0, d2 = 0;
 288         unsigned long back_max;
 289 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 290 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 291         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 292
 293         if (rq1 == NULL || rq1 == rq2)
 294                 return rq2;
 295         if (rq2 == NULL)
 296                 return rq1;
 297
 298         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 299                 return rq1;
 300         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 301                 return rq2;
 302         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
 303                 return rq1;
 304         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
 305                 return rq2;
 306
 307         s1 = rq1->sector;
 308         s2 = rq2->sector;
 309
 310         last = cfqd->last_sector;
 311
 312         /*
 313          * by definition, 1KiB is 2 sectors
 314          */
 315         back_max = cfqd->cfq_back_max * 2;
 316
 317         /*
 318          * Strict one way elevator _except_ in the case where we allow
 319          * short backward seeks which are biased as twice the cost of a
 320          * similar forward seek.
 321          */
 322         if (s1 >= last)
 323                 d1 = s1 - last;
 324         else if (s1 + back_max >= last)
 325                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 326         else
 327                 wrap |= CFQ_RQ1_WRAP;
 328
 329         if (s2 >= last)
 330                 d2 = s2 - last;
 331         else if (s2 + back_max >= last)
 332                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 333         else
 334                 wrap |= CFQ_RQ2_WRAP;
 335
 336         /* Found required data */
 337
 338         /*
 339          * By doing switch() on the bit mask "wrap" we avoid having to
 340          * check two variables for all permutations: --> faster!
 341          */
 342         switch (wrap) {
 343         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 344                 if (d1 < d2)
 345                         return rq1;
 346                 else if (d2 < d1)
 347                         return rq2;
 348                 else {
 349                         if (s1 >= s2)
 350                                 return rq1;
 351                         else
 352                                 return rq2;
 353                 }
 354
 355         case CFQ_RQ2_WRAP:
 356                 return rq1;
 357         case CFQ_RQ1_WRAP:
 358                 return rq2;
 359         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
 360         default:
 361                 /*
 362                  * Since both rqs are wrapped,
 363                  * start with the one that's further behind head
 364                  * (--> only *one* back seek required),
 365                  * since back seek takes more time than forward.
 366                  */
 367                 if (s1 <= s2)
 368                         return rq1;
 369                 else
 370                         return rq2;
 371         }
 372 }
 373
 374 /*
 375  * would be nice to take fifo expire time into account as well
 376  */
 377 static struct request *
 378 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 379                   struct request *last)
 380 {
 381         struct rb_node *rbnext = rb_next(&last->rb_node);
 382         struct rb_node *rbprev = rb_prev(&last->rb_node);
 383         struct request *next = NULL, *prev = NULL;
 384
 385         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
 386
 387         if (rbprev)
 388                 prev = rb_entry_rq(rbprev);
 389
 390         if (rbnext)
 391                 next = rb_entry_rq(rbnext);
 392         else {
 393                 rbnext = rb_first(&cfqq->sort_list);
 394                 if (rbnext && rbnext != &last->rb_node)
 395                         next = rb_entry_rq(rbnext);
 396         }
 397
 398         return cfq_choose_req(cfqd, next, prev);
 399 }
 400
 401 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 402 {
 403         struct cfq_data *cfqd = cfqq->cfqd;
 404         struct list_head *list, *n;
 405         struct cfq_queue *__cfqq;
 406
 407         /*
 408          * Resorting requires the cfqq to be on the RR list already.
 409          */
 410         if (!cfq_cfqq_on_rr(cfqq))
 411                 return;
 412
 413         list_del(&cfqq->cfq_list);
 414
 415         if (cfq_class_rt(cfqq))
 416                 list = &cfqd->cur_rr;
 417         else if (cfq_class_idle(cfqq))
 418                 list = &cfqd->idle_rr;
 419         else {
 420                 /*
 421                  * if cfqq has requests in flight, don't allow it to be
 422                  * found in cfq_set_active_queue before it has finished them.
 423                  * this is done to increase fairness between a process that
 424                  * has lots of io pending vs one that only generates one
 425                  * sporadically or synchronously
 426                  */
 427                 if (cfq_cfqq_dispatched(cfqq))
 428                         list = &cfqd->busy_rr;
 429                 else
 430                         list = &cfqd->rr_list[cfqq->ioprio];
 431         }
 432
 433         if (preempted || cfq_cfqq_queue_new(cfqq)) {
 434                 /*
 435                  * If this queue was preempted or is new (never been serviced),
 436                  * let it be added first for fairness but beind other new
 437                  * queues.
 438                  */
 439                 n = list;
 440                 while (n->next != list) {
 441                         __cfqq = list_entry_cfqq(n->next);
 442                         if (!cfq_cfqq_queue_new(__cfqq))
 443                                 break;
 444
 445                         n = n->next;
 446                 }
 447                 list_add_tail(&cfqq->cfq_list, n);
 448         } else if (!cfq_cfqq_class_sync(cfqq)) {
 449                 /*
 450                  * async queue always goes to the end. this wont be overly
 451                  * unfair to writes, as the sort of the sync queue wont be
 452                  * allowed to pass the async queue again.
 453                  */
 454                 list_add_tail(&cfqq->cfq_list, list);
 455         } else {
 456                 /*
 457                  * sort by last service, but don't cross a new or async
 458                  * queue. we don't cross a new queue because it hasn't been
 459                  * service before, and we don't cross an async queue because
 460                  * it gets added to the end on expire.
 461                  */
 462                 n = list;
 463                 while ((n = n->prev) != list) {
 464                         struct cfq_queue *__cfqq = list_entry_cfqq(n);
 465
 466                         if (!cfq_cfqq_class_sync(cfqq) || !__cfqq->service_last)
 467                                 break;
 468                         if (time_before(__cfqq->service_last, cfqq->service_last))
 469                                 break;
 470                 }
 471                 list_add(&cfqq->cfq_list, n);
 472         }
 473 }
 474
 475 /*
 476  * add to busy list of queues for service, trying to be fair in ordering
 477  * the pending list according to last request service
 478  */
 479 static inline void
 480 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 481 {
 482         BUG_ON(cfq_cfqq_on_rr(cfqq));
 483         cfq_mark_cfqq_on_rr(cfqq);
 484         cfqd->busy_queues++;
 485
 486         cfq_resort_rr_list(cfqq, 0);
 487 }
 488
 489 static inline void
 490 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 491 {
 492         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 493         cfq_clear_cfqq_on_rr(cfqq);
 494         list_del_init(&cfqq->cfq_list);
 495
 496         BUG_ON(!cfqd->busy_queues);
 497         cfqd->busy_queues--;
 498 }
 499
 500 /*
 501  * rb tree support functions
 502  */
 503 static inline void cfq_del_rq_rb(struct request *rq)
 504 {
 505         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 506         struct cfq_data *cfqd = cfqq->cfqd;
 507         const int sync = rq_is_sync(rq);
 508
 509         BUG_ON(!cfqq->queued[sync]);
 510         cfqq->queued[sync]--;
 511
 512         elv_rb_del(&cfqq->sort_list, rq);
 513
 514         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
 515                 cfq_del_cfqq_rr(cfqd, cfqq);
 516 }
 517
 518 static void cfq_add_rq_rb(struct request *rq)
 519 {
 520         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 521         struct cfq_data *cfqd = cfqq->cfqd;
 522         struct request *__alias;
 523
 524         cfqq->queued[rq_is_sync(rq)]++;
 525
 526         /*
 527          * looks a little odd, but the first insert might return an alias.
 528          * if that happens, put the alias on the dispatch list
 529          */
 530         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
 531                 cfq_dispatch_insert(cfqd->queue, __alias);
 532
 533         if (!cfq_cfqq_on_rr(cfqq))
 534                 cfq_add_cfqq_rr(cfqd, cfqq);
 535 }
 536
 537 static inline void
 538 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
 539 {
 540         elv_rb_del(&cfqq->sort_list, rq);
 541         cfqq->queued[rq_is_sync(rq)]--;
 542         cfq_add_rq_rb(rq);
 543 }
 544
 545 static struct request *
 546 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 547 {
 548         struct task_struct *tsk = current;
 549         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
 550         struct cfq_queue *cfqq;
 551
 552         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
 553         if (cfqq) {
 554                 sector_t sector = bio->bi_sector + bio_sectors(bio);
 555
 556                 return elv_rb_find(&cfqq->sort_list, sector);
 557         }
 558
 559         return NULL;
 560 }
 561
 562 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 563 {
 564         struct cfq_data *cfqd = q->elevator->elevator_data;
 565
 566         cfqd->rq_in_driver++;
 567
 568         /*
 569          * If the depth is larger 1, it really could be queueing. But lets
 570          * make the mark a little higher - idling could still be good for
 571          * low queueing, and a low queueing number could also just indicate
 572          * a SCSI mid layer like behaviour where limit+1 is often seen.
 573          */
 574         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
 575                 cfqd->hw_tag = 1;
 576 }
 577
 578 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 579 {
 580         struct cfq_data *cfqd = q->elevator->elevator_data;
 581
 582         WARN_ON(!cfqd->rq_in_driver);
 583         cfqd->rq_in_driver--;
 584 }
 585
 586 static void cfq_remove_request(struct request *rq)
 587 {
 588         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 589
 590         if (cfqq->next_rq == rq)
 591                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
 592
 593         list_del_init(&rq->queuelist);
 594         cfq_del_rq_rb(rq);
 595
 596         if (rq_is_meta(rq)) {
 597                 WARN_ON(!cfqq->meta_pending);
 598                 cfqq->meta_pending--;
 599         }
 600 }
 601
 602 static int
 603 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 604 {
 605         struct cfq_data *cfqd = q->elevator->elevator_data;
 606         struct request *__rq;
 607
 608         __rq = cfq_find_rq_fmerge(cfqd, bio);
 609         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 610                 *req = __rq;
 611                 return ELEVATOR_FRONT_MERGE;
 612         }
 613
 614         return ELEVATOR_NO_MERGE;
 615 }
 616
 617 static void cfq_merged_request(request_queue_t *q, struct request *req,
 618                                int type)
 619 {
 620         if (type == ELEVATOR_FRONT_MERGE) {
 621                 struct cfq_queue *cfqq = RQ_CFQQ(req);
 622
 623                 cfq_reposition_rq_rb(cfqq, req);
 624         }
 625 }
 626
 627 static void
 628 cfq_merged_requests(request_queue_t *q, struct request *rq,
 629                     struct request *next)
 630 {
 631         /*
 632          * reposition in fifo if next is older than rq
 633          */
 634         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 635             time_before(next->start_time, rq->start_time))
 636                 list_move(&rq->queuelist, &next->queuelist);
 637
 638         cfq_remove_request(next);
 639 }
 640
 641 static int cfq_allow_merge(request_queue_t *q, struct request *rq,
 642                            struct bio *bio)
 643 {
 644         struct cfq_data *cfqd = q->elevator->elevator_data;
 645         const int rw = bio_data_dir(bio);
 646         struct cfq_queue *cfqq;
 647         pid_t key;
 648
 649         /*
 650          * Disallow merge of a sync bio into an async request.
 651          */
 652         if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq))
 653                 return 0;
 654
 655         /*
 656          * Lookup the cfqq that this bio will be queued with. Allow
 657          * merge only if rq is queued there.
 658          */
 659         key = cfq_queue_pid(current, rw, bio_sync(bio));
 660         cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio);
 661
 662         if (cfqq == RQ_CFQQ(rq))
 663                 return 1;
 664
 665         return 0;
 666 }
 667
 668 static inline void
 669 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 670 {
 671         if (cfqq) {
 672                 /*
 673                  * stop potential idle class queues waiting service
 674                  */
 675                 del_timer(&cfqd->idle_class_timer);
 676
 677                 cfqq->slice_end = 0;
 678                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 679                 cfq_clear_cfqq_fifo_expire(cfqq);
 680                 cfq_mark_cfqq_slice_new(cfqq);
 681         }
 682
 683         cfqd->active_queue = cfqq;
 684 }
 685
 686 /*
 687  * current cfqq expired its slice (or was too idle), select new one
 688  */
 689 static void
 690 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 691                     int preempted, int timed_out)
 692 {
 693         if (cfq_cfqq_wait_request(cfqq))
 694                 del_timer(&cfqd->idle_slice_timer);
 695
 696         cfq_clear_cfqq_must_dispatch(cfqq);
 697         cfq_clear_cfqq_wait_request(cfqq);
 698         cfq_clear_cfqq_queue_new(cfqq);
 699
 700         /*
 701          * store what was left of this slice, if the queue idled out
 702          * or was preempted
 703          */
 704         if (timed_out && !cfq_cfqq_slice_new(cfqq))
 705                 cfqq->slice_resid = cfqq->slice_end - jiffies;
 706
 707         cfq_resort_rr_list(cfqq, preempted);
 708
 709         if (cfqq == cfqd->active_queue)
 710                 cfqd->active_queue = NULL;
 711
 712         if (cfqd->active_cic) {
 713                 put_io_context(cfqd->active_cic->ioc);
 714                 cfqd->active_cic = NULL;
 715         }
 716
 717         cfqd->dispatch_slice = 0;
 718 }
 719
 720 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted,
 721                                      int timed_out)
 722 {
 723         struct cfq_queue *cfqq = cfqd->active_queue;
 724
 725         if (cfqq)
 726                 __cfq_slice_expired(cfqd, cfqq, preempted, timed_out);
 727 }
 728
 729 /*
 730  * 0
 731  * 0,1
 732  * 0,1,2
 733  * 0,1,2,3
 734  * 0,1,2,3,4
 735  * 0,1,2,3,4,5
 736  * 0,1,2,3,4,5,6
 737  * 0,1,2,3,4,5,6,7
 738  */
 739 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 740 {
 741         int prio, wrap;
 742
 743         prio = -1;
 744         wrap = 0;
 745         do {
 746                 int p;
 747
 748                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 749                         if (!list_empty(&cfqd->rr_list[p])) {
 750                                 prio = p;
 751                                 break;
 752                         }
 753                 }
 754
 755                 if (prio != -1)
 756                         break;
 757                 cfqd->cur_prio = 0;
 758                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 759                         cfqd->cur_end_prio = 0;
 760                         if (wrap)
 761                                 break;
 762                         wrap = 1;
 763                 }
 764         } while (1);
 765
 766         if (unlikely(prio == -1))
 767                 return -1;
 768
 769         BUG_ON(prio >= CFQ_PRIO_LISTS);
 770
 771         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 772
 773         cfqd->cur_prio = prio + 1;
 774         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 775                 cfqd->cur_end_prio = cfqd->cur_prio;
 776                 cfqd->cur_prio = 0;
 777         }
 778         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 779                 cfqd->cur_prio = 0;
 780                 cfqd->cur_end_prio = 0;
 781         }
 782
 783         return prio;
 784 }
 785
 786 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 787 {
 788         struct cfq_queue *cfqq = NULL;
 789
 790         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
 791                 /*
 792                  * if current list is non-empty, grab first entry. if it is
 793                  * empty, get next prio level and grab first entry then if any
 794                  * are spliced
 795                  */
 796                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 797         } else if (!list_empty(&cfqd->busy_rr)) {
 798                 /*
 799                  * If no new queues are available, check if the busy list has
 800                  * some before falling back to idle io.
 801                  */
 802                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
 803         } else if (!list_empty(&cfqd->idle_rr)) {
 804                 /*
 805                  * if we have idle queues and no rt or be queues had pending
 806                  * requests, either allow immediate service if the grace period
 807                  * has passed or arm the idle grace timer
 808                  */
 809                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 810
 811                 if (time_after_eq(jiffies, end))
 812                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 813                 else
 814                         mod_timer(&cfqd->idle_class_timer, end);
 815         }
 816
 817         __cfq_set_active_queue(cfqd, cfqq);
 818         return cfqq;
 819 }
 820
 821 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
 822
 823 static int cfq_arm_slice_timer(struct cfq_data *cfqd)
 824 {
 825         struct cfq_queue *cfqq = cfqd->active_queue;
 826         struct cfq_io_context *cic;
 827         unsigned long sl;
 828
 829         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
 830
 831         /*
 832          * idle is disabled, either manually or by past process history
 833          */
 834         if (!cfqd->cfq_slice_idle)
 835                 return 0;
 836         if (!cfq_cfqq_idle_window(cfqq))
 837                 return 0;
 838         /*
 839          * task has exited, don't wait
 840          */
 841         cic = cfqd->active_cic;
 842         if (!cic || !cic->ioc->task)
 843                 return 0;
 844
 845         cfq_mark_cfqq_must_dispatch(cfqq);
 846         cfq_mark_cfqq_wait_request(cfqq);
 847
 848         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 849
 850         /*
 851          * we don't want to idle for seeks, but we do want to allow
 852          * fair distribution of slice time for a process doing back-to-back
 853          * seeks. so allow a little bit of time for him to submit a new rq
 854          */
 855         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
 856                 sl = min(sl, msecs_to_jiffies(2));
 857
 858         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 859         return 1;
 860 }
 861
 862 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
 863 {
 864         struct cfq_data *cfqd = q->elevator->elevator_data;
 865         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 866
 867         cfq_remove_request(rq);
 868         cfqq->on_dispatch[rq_is_sync(rq)]++;
 869         elv_dispatch_sort(q, rq);
 870
 871         rq = list_entry(q->queue_head.prev, struct request, queuelist);
 872         cfqd->last_sector = rq->sector + rq->nr_sectors;
 873 }
 874
 875 /*
 876  * return expired entry, or NULL to just start from scratch in rbtree
 877  */
 878 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
 879 {
 880         struct cfq_data *cfqd = cfqq->cfqd;
 881         struct request *rq;
 882         int fifo;
 883
 884         if (cfq_cfqq_fifo_expire(cfqq))
 885                 return NULL;
 886
 887         cfq_mark_cfqq_fifo_expire(cfqq);
 888
 889         if (list_empty(&cfqq->fifo))
 890                 return NULL;
 891
 892         fifo = cfq_cfqq_class_sync(cfqq);
 893         rq = rq_entry_fifo(cfqq->fifo.next);
 894
 895         if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
 896                 return rq;
 897
 898         return NULL;
 899 }
 900
 901 static inline int
 902 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 903 {
 904         const int base_rq = cfqd->cfq_slice_async_rq;
 905
 906         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 907
 908         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
 909 }
 910
 911 /*
 912  * get next queue for service
 913  */
 914 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
 915 {
 916         struct cfq_queue *cfqq;
 917
 918         cfqq = cfqd->active_queue;
 919         if (!cfqq)
 920                 goto new_queue;
 921
 922         /*
 923          * slice has expired
 924          */
 925         if (!cfq_cfqq_must_dispatch(cfqq) && cfq_slice_used(cfqq))
 926                 goto expire;
 927
 928         /*
 929          * if queue has requests, dispatch one. if not, check if
 930          * enough slice is left to wait for one
 931          */
 932         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
 933                 goto keep_queue;
 934         else if (cfq_cfqq_slice_new(cfqq) || cfq_cfqq_dispatched(cfqq)) {
 935                 cfqq = NULL;
 936                 goto keep_queue;
 937         } else if (cfq_cfqq_class_sync(cfqq)) {
 938                 if (cfq_arm_slice_timer(cfqd))
 939                         return NULL;
 940         }
 941
 942 expire:
 943         cfq_slice_expired(cfqd, 0, 0);
 944 new_queue:
 945         cfqq = cfq_set_active_queue(cfqd);
 946 keep_queue:
 947         return cfqq;
 948 }
 949
 950 static int
 951 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 952                         int max_dispatch)
 953 {
 954         int dispatched = 0;
 955
 956         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
 957
 958         do {
 959                 struct request *rq;
 960
 961                 /*
 962                  * follow expired path, else get first next available
 963                  */
 964                 if ((rq = cfq_check_fifo(cfqq)) == NULL)
 965                         rq = cfqq->next_rq;
 966
 967                 /*
 968                  * finally, insert request into driver dispatch list
 969                  */
 970                 cfq_dispatch_insert(cfqd->queue, rq);
 971
 972                 cfqd->dispatch_slice++;
 973                 dispatched++;
 974
 975                 if (!cfqd->active_cic) {
 976                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
 977                         cfqd->active_cic = RQ_CIC(rq);
 978                 }
 979
 980                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
 981                         break;
 982
 983         } while (dispatched < max_dispatch);
 984
 985         /*
 986          * expire an async queue immediately if it has used up its slice. idle
 987          * queue always expire after 1 dispatch round.
 988          */
 989         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
 990             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
 991             cfq_class_idle(cfqq))) {
 992                 cfqq->slice_end = jiffies + 1;
 993                 cfq_slice_expired(cfqd, 0, 0);
 994         }
 995
 996         return dispatched;
 997 }
 998
 999 static int
1000 cfq_forced_dispatch_cfqqs(struct list_head *list)
1001 {
1002         struct cfq_queue *cfqq, *next;
1003         int dispatched;
1004
1005         dispatched = 0;
1006         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1007                 while (cfqq->next_rq) {
1008                         cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1009                         dispatched++;
1010                 }
1011                 BUG_ON(!list_empty(&cfqq->fifo));
1012         }
1013
1014         return dispatched;
1015 }
1016
1017 static int
1018 cfq_forced_dispatch(struct cfq_data *cfqd)
1019 {
1020         int i, dispatched = 0;
1021
1022         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1023                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1024
1025         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1026         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1027         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1028
1029         cfq_slice_expired(cfqd, 0, 0);
1030
1031         BUG_ON(cfqd->busy_queues);
1032
1033         return dispatched;
1034 }
1035
1036 static int
1037 cfq_dispatch_requests(request_queue_t *q, int force)
1038 {
1039         struct cfq_data *cfqd = q->elevator->elevator_data;
1040         struct cfq_queue *cfqq, *prev_cfqq;
1041         int dispatched;
1042
1043         if (!cfqd->busy_queues)
1044                 return 0;
1045
1046         if (unlikely(force))
1047                 return cfq_forced_dispatch(cfqd);
1048
1049         dispatched = 0;
1050         prev_cfqq = NULL;
1051         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1052                 int max_dispatch;
1053
1054                 if (cfqd->busy_queues > 1) {
1055                         /*
1056                          * Don't repeat dispatch from the previous queue.
1057                          */
1058                         if (prev_cfqq == cfqq)
1059                                 break;
1060
1061                         /*
1062                          * So we have dispatched before in this round, if the
1063                          * next queue has idling enabled (must be sync), don't
1064                          * allow it service until the previous have continued.
1065                          */
1066                         if (cfqd->rq_in_driver && cfq_cfqq_idle_window(cfqq))
1067                                 break;
1068                 }
1069
1070                 cfq_clear_cfqq_must_dispatch(cfqq);
1071                 cfq_clear_cfqq_wait_request(cfqq);
1072                 del_timer(&cfqd->idle_slice_timer);
1073
1074                 max_dispatch = cfqd->cfq_quantum;
1075                 if (cfq_class_idle(cfqq))
1076                         max_dispatch = 1;
1077
1078                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1079                 prev_cfqq = cfqq;
1080         }
1081
1082         return dispatched;
1083 }
1084
1085 /*
1086  * task holds one reference to the queue, dropped when task exits. each rq
1087  * in-flight on this queue also holds a reference, dropped when rq is freed.
1088  *
1089  * queue lock must be held here.
1090  */
1091 static void cfq_put_queue(struct cfq_queue *cfqq)
1092 {
1093         struct cfq_data *cfqd = cfqq->cfqd;
1094
1095         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1096
1097         if (!atomic_dec_and_test(&cfqq->ref))
1098                 return;
1099
1100         BUG_ON(rb_first(&cfqq->sort_list));
1101         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1102         BUG_ON(cfq_cfqq_on_rr(cfqq));
1103
1104         if (unlikely(cfqd->active_queue == cfqq)) {
1105                 __cfq_slice_expired(cfqd, cfqq, 0, 0);
1106                 cfq_schedule_dispatch(cfqd);
1107         }
1108
1109         /*
1110          * it's on the empty list and still hashed
1111          */
1112         list_del(&cfqq->cfq_list);
1113         hlist_del(&cfqq->cfq_hash);
1114         kmem_cache_free(cfq_pool, cfqq);
1115 }
1116
1117 static struct cfq_queue *
1118 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1119                     const int hashval)
1120 {
1121         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1122         struct hlist_node *entry;
1123         struct cfq_queue *__cfqq;
1124
1125         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1126                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1127
1128                 if (__cfqq->key == key && (__p == prio || !prio))
1129                         return __cfqq;
1130         }
1131
1132         return NULL;
1133 }
1134
1135 static struct cfq_queue *
1136 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1137 {
1138         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1139 }
1140
1141 static void cfq_free_io_context(struct io_context *ioc)
1142 {
1143         struct cfq_io_context *__cic;
1144         struct rb_node *n;
1145         int freed = 0;
1146
1147         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1148                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1149                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1150                 kmem_cache_free(cfq_ioc_pool, __cic);
1151                 freed++;
1152         }
1153
1154         elv_ioc_count_mod(ioc_count, -freed);
1155
1156         if (ioc_gone && !elv_ioc_count_read(ioc_count))
1157                 complete(ioc_gone);
1158 }
1159
1160 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1161 {
1162         if (unlikely(cfqq == cfqd->active_queue)) {
1163                 __cfq_slice_expired(cfqd, cfqq, 0, 0);
1164                 cfq_schedule_dispatch(cfqd);
1165         }
1166
1167         cfq_put_queue(cfqq);
1168 }
1169
1170 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1171                                          struct cfq_io_context *cic)
1172 {
1173         list_del_init(&cic->queue_list);
1174         smp_wmb();
1175         cic->key = NULL;
1176
1177         if (cic->cfqq[ASYNC]) {
1178                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1179                 cic->cfqq[ASYNC] = NULL;
1180         }
1181
1182         if (cic->cfqq[SYNC]) {
1183                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1184                 cic->cfqq[SYNC] = NULL;
1185         }
1186 }
1187
1188
1189 /*
1190  * Called with interrupts disabled
1191  */
1192 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1193 {
1194         struct cfq_data *cfqd = cic->key;
1195
1196         if (cfqd) {
1197                 request_queue_t *q = cfqd->queue;
1198
1199                 spin_lock_irq(q->queue_lock);
1200                 __cfq_exit_single_io_context(cfqd, cic);
1201                 spin_unlock_irq(q->queue_lock);
1202         }
1203 }
1204
1205 static void cfq_exit_io_context(struct io_context *ioc)
1206 {
1207         struct cfq_io_context *__cic;
1208         struct rb_node *n;
1209
1210         /*
1211          * put the reference this task is holding to the various queues
1212          */
1213
1214         n = rb_first(&ioc->cic_root);
1215         while (n != NULL) {
1216                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1217
1218                 cfq_exit_single_io_context(__cic);
1219                 n = rb_next(n);
1220         }
1221 }
1222
1223 static struct cfq_io_context *
1224 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1225 {
1226         struct cfq_io_context *cic;
1227
1228         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
1229         if (cic) {
1230                 memset(cic, 0, sizeof(*cic));
1231                 cic->last_end_request = jiffies;
1232                 INIT_LIST_HEAD(&cic->queue_list);
1233                 cic->dtor = cfq_free_io_context;
1234                 cic->exit = cfq_exit_io_context;
1235                 elv_ioc_count_inc(ioc_count);
1236         }
1237
1238         return cic;
1239 }
1240
1241 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1242 {
1243         struct task_struct *tsk = current;
1244         int ioprio_class;
1245
1246         if (!cfq_cfqq_prio_changed(cfqq))
1247                 return;
1248
1249         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1250         switch (ioprio_class) {
1251                 default:
1252                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1253                 case IOPRIO_CLASS_NONE:
1254                         /*
1255                          * no prio set, place us in the middle of the BE classes
1256                          */
1257                         cfqq->ioprio = task_nice_ioprio(tsk);
1258                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1259                         break;
1260                 case IOPRIO_CLASS_RT:
1261                         cfqq->ioprio = task_ioprio(tsk);
1262                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1263                         break;
1264                 case IOPRIO_CLASS_BE:
1265                         cfqq->ioprio = task_ioprio(tsk);
1266                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1267                         break;
1268                 case IOPRIO_CLASS_IDLE:
1269                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1270                         cfqq->ioprio = 7;
1271                         cfq_clear_cfqq_idle_window(cfqq);
1272                         break;
1273         }
1274
1275         /*
1276          * keep track of original prio settings in case we have to temporarily
1277          * elevate the priority of this queue
1278          */
1279         cfqq->org_ioprio = cfqq->ioprio;
1280         cfqq->org_ioprio_class = cfqq->ioprio_class;
1281
1282         cfq_resort_rr_list(cfqq, 0);
1283         cfq_clear_cfqq_prio_changed(cfqq);
1284 }
1285
1286 static inline void changed_ioprio(struct cfq_io_context *cic)
1287 {
1288         struct cfq_data *cfqd = cic->key;
1289         struct cfq_queue *cfqq;
1290         unsigned long flags;
1291
1292         if (unlikely(!cfqd))
1293                 return;
1294
1295         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1296
1297         cfqq = cic->cfqq[ASYNC];
1298         if (cfqq) {
1299                 struct cfq_queue *new_cfqq;
1300                 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1301                                          GFP_ATOMIC);
1302                 if (new_cfqq) {
1303                         cic->cfqq[ASYNC] = new_cfqq;
1304                         cfq_put_queue(cfqq);
1305                 }
1306         }
1307
1308         cfqq = cic->cfqq[SYNC];
1309         if (cfqq)
1310                 cfq_mark_cfqq_prio_changed(cfqq);
1311
1312         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1313 }
1314
1315 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1316 {
1317         struct cfq_io_context *cic;
1318         struct rb_node *n;
1319
1320         ioc->ioprio_changed = 0;
1321
1322         n = rb_first(&ioc->cic_root);
1323         while (n != NULL) {
1324                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1325
1326                 changed_ioprio(cic);
1327                 n = rb_next(n);
1328         }
1329 }
1330
1331 static struct cfq_queue *
1332 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1333               gfp_t gfp_mask)
1334 {
1335         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1336         struct cfq_queue *cfqq, *new_cfqq = NULL;
1337         unsigned short ioprio;
1338
1339 retry:
1340         ioprio = tsk->ioprio;
1341         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1342
1343         if (!cfqq) {
1344                 if (new_cfqq) {
1345                         cfqq = new_cfqq;
1346                         new_cfqq = NULL;
1347                 } else if (gfp_mask & __GFP_WAIT) {
1348                         /*
1349                          * Inform the allocator of the fact that we will
1350                          * just repeat this allocation if it fails, to allow
1351                          * the allocator to do whatever it needs to attempt to
1352                          * free memory.
1353                          */
1354                         spin_unlock_irq(cfqd->queue->queue_lock);
1355                         new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
1356                         spin_lock_irq(cfqd->queue->queue_lock);
1357                         goto retry;
1358                 } else {
1359                         cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
1360                         if (!cfqq)
1361                                 goto out;
1362                 }
1363
1364                 memset(cfqq, 0, sizeof(*cfqq));
1365
1366                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1367                 INIT_LIST_HEAD(&cfqq->cfq_list);
1368                 INIT_LIST_HEAD(&cfqq->fifo);
1369
1370                 cfqq->key = key;
1371                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1372                 atomic_set(&cfqq->ref, 0);
1373                 cfqq->cfqd = cfqd;
1374
1375                 if (key != CFQ_KEY_ASYNC)
1376                         cfq_mark_cfqq_idle_window(cfqq);
1377
1378                 cfq_mark_cfqq_prio_changed(cfqq);
1379                 cfq_mark_cfqq_queue_new(cfqq);
1380                 cfq_init_prio_data(cfqq);
1381         }
1382
1383         if (new_cfqq)
1384                 kmem_cache_free(cfq_pool, new_cfqq);
1385
1386         atomic_inc(&cfqq->ref);
1387 out:
1388         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1389         return cfqq;
1390 }
1391
1392 static void
1393 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1394 {
1395         WARN_ON(!list_empty(&cic->queue_list));
1396         rb_erase(&cic->rb_node, &ioc->cic_root);
1397         kmem_cache_free(cfq_ioc_pool, cic);
1398         elv_ioc_count_dec(ioc_count);
1399 }
1400
1401 static struct cfq_io_context *
1402 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1403 {
1404         struct rb_node *n;
1405         struct cfq_io_context *cic;
1406         void *k, *key = cfqd;
1407
1408 restart:
1409         n = ioc->cic_root.rb_node;
1410         while (n) {
1411                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1412                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1413                 k = cic->key;
1414                 if (unlikely(!k)) {
1415                         cfq_drop_dead_cic(ioc, cic);
1416                         goto restart;
1417                 }
1418
1419                 if (key < k)
1420                         n = n->rb_left;
1421                 else if (key > k)
1422                         n = n->rb_right;
1423                 else
1424                         return cic;
1425         }
1426
1427         return NULL;
1428 }
1429
1430 static inline void
1431 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1432              struct cfq_io_context *cic)
1433 {
1434         struct rb_node **p;
1435         struct rb_node *parent;
1436         struct cfq_io_context *__cic;
1437         unsigned long flags;
1438         void *k;
1439
1440         cic->ioc = ioc;
1441         cic->key = cfqd;
1442
1443 restart:
1444         parent = NULL;
1445         p = &ioc->cic_root.rb_node;
1446         while (*p) {
1447                 parent = *p;
1448                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1449                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1450                 k = __cic->key;
1451                 if (unlikely(!k)) {
1452                         cfq_drop_dead_cic(ioc, __cic);
1453                         goto restart;
1454                 }
1455
1456                 if (cic->key < k)
1457                         p = &(*p)->rb_left;
1458                 else if (cic->key > k)
1459                         p = &(*p)->rb_right;
1460                 else
1461                         BUG();
1462         }
1463
1464         rb_link_node(&cic->rb_node, parent, p);
1465         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1466
1467         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1468         list_add(&cic->queue_list, &cfqd->cic_list);
1469         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1470 }
1471
1472 /*
1473  * Setup general io context and cfq io context. There can be several cfq
1474  * io contexts per general io context, if this process is doing io to more
1475  * than one device managed by cfq.
1476  */
1477 static struct cfq_io_context *
1478 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1479 {
1480         struct io_context *ioc = NULL;
1481         struct cfq_io_context *cic;
1482
1483         might_sleep_if(gfp_mask & __GFP_WAIT);
1484
1485         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1486         if (!ioc)
1487                 return NULL;
1488
1489         cic = cfq_cic_rb_lookup(cfqd, ioc);
1490         if (cic)
1491                 goto out;
1492
1493         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1494         if (cic == NULL)
1495                 goto err;
1496
1497         cfq_cic_link(cfqd, ioc, cic);
1498 out:
1499         smp_read_barrier_depends();
1500         if (unlikely(ioc->ioprio_changed))
1501                 cfq_ioc_set_ioprio(ioc);
1502
1503         return cic;
1504 err:
1505         put_io_context(ioc);
1506         return NULL;
1507 }
1508
1509 static void
1510 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1511 {
1512         unsigned long elapsed = jiffies - cic->last_end_request;
1513         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1514
1515         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1516         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1517         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1518 }
1519
1520 static void
1521 cfq_update_io_seektime(struct cfq_io_context *cic, struct request *rq)
1522 {
1523         sector_t sdist;
1524         u64 total;
1525
1526         if (cic->last_request_pos < rq->sector)
1527                 sdist = rq->sector - cic->last_request_pos;
1528         else
1529                 sdist = cic->last_request_pos - rq->sector;
1530
1531         /*
1532          * Don't allow the seek distance to get too large from the
1533          * odd fragment, pagein, etc
1534          */
1535         if (cic->seek_samples <= 60) /* second&third seek */
1536                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1537         else
1538                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1539
1540         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1541         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1542         total = cic->seek_total + (cic->seek_samples/2);
1543         do_div(total, cic->seek_samples);
1544         cic->seek_mean = (sector_t)total;
1545 }
1546
1547 /*
1548  * Disable idle window if the process thinks too long or seeks so much that
1549  * it doesn't matter
1550  */
1551 static void
1552 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1553                        struct cfq_io_context *cic)
1554 {
1555         int enable_idle = cfq_cfqq_idle_window(cfqq);
1556
1557         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1558             (cfqd->hw_tag && CIC_SEEKY(cic)))
1559                 enable_idle = 0;
1560         else if (sample_valid(cic->ttime_samples)) {
1561                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1562                         enable_idle = 0;
1563                 else
1564                         enable_idle = 1;
1565         }
1566
1567         if (enable_idle)
1568                 cfq_mark_cfqq_idle_window(cfqq);
1569         else
1570                 cfq_clear_cfqq_idle_window(cfqq);
1571 }
1572
1573 /*
1574  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1575  * no or if we aren't sure, a 1 will cause a preempt.
1576  */
1577 static int
1578 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1579                    struct request *rq)
1580 {
1581         struct cfq_queue *cfqq = cfqd->active_queue;
1582
1583         if (cfq_class_idle(new_cfqq))
1584                 return 0;
1585
1586         if (!cfqq)
1587                 return 0;
1588
1589         if (cfq_class_idle(cfqq))
1590                 return 1;
1591         if (!cfq_cfqq_wait_request(new_cfqq))
1592                 return 0;
1593         /*
1594          * if the new request is sync, but the currently running queue is
1595          * not, let the sync request have priority.
1596          */
1597         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1598                 return 1;
1599         /*
1600          * So both queues are sync. Let the new request get disk time if
1601          * it's a metadata request and the current queue is doing regular IO.
1602          */
1603         if (rq_is_meta(rq) && !cfqq->meta_pending)
1604                 return 1;
1605
1606         return 0;
1607 }
1608
1609 /*
1610  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1611  * let it have half of its nominal slice.
1612  */
1613 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1614 {
1615         cfq_slice_expired(cfqd, 1, 1);
1616
1617         /*
1618          * Put the new queue at the front of the of the current list,
1619          * so we know that it will be selected next.
1620          */
1621         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1622         list_move(&cfqq->cfq_list, &cfqd->cur_rr);
1623
1624         cfqq->slice_end = 0;
1625         cfq_mark_cfqq_slice_new(cfqq);
1626 }
1627
1628 /*
1629  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1630  * something we should do about it
1631  */
1632 static void
1633 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1634                 struct request *rq)
1635 {
1636         struct cfq_io_context *cic = RQ_CIC(rq);
1637
1638         if (rq_is_meta(rq))
1639                 cfqq->meta_pending++;
1640
1641         /*
1642          * check if this request is a better next-serve candidate)) {
1643          */
1644         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1645         BUG_ON(!cfqq->next_rq);
1646
1647         /*
1648          * we never wait for an async request and we don't allow preemption
1649          * of an async request. so just return early
1650          */
1651         if (!rq_is_sync(rq)) {
1652                 /*
1653                  * sync process issued an async request, if it's waiting
1654                  * then expire it and kick rq handling.
1655                  */
1656                 if (cic == cfqd->active_cic &&
1657                     del_timer(&cfqd->idle_slice_timer)) {
1658                         cfq_slice_expired(cfqd, 0, 0);
1659                         blk_start_queueing(cfqd->queue);
1660                 }
1661                 return;
1662         }
1663
1664         cfq_update_io_thinktime(cfqd, cic);
1665         cfq_update_io_seektime(cic, rq);
1666         cfq_update_idle_window(cfqd, cfqq, cic);
1667
1668         cic->last_request_pos = rq->sector + rq->nr_sectors;
1669
1670         if (cfqq == cfqd->active_queue) {
1671                 /*
1672                  * if we are waiting for a request for this queue, let it rip
1673                  * immediately and flag that we must not expire this queue
1674                  * just now
1675                  */
1676                 if (cfq_cfqq_wait_request(cfqq)) {
1677                         cfq_mark_cfqq_must_dispatch(cfqq);
1678                         del_timer(&cfqd->idle_slice_timer);
1679                         blk_start_queueing(cfqd->queue);
1680                 }
1681         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1682                 /*
1683                  * not the active queue - expire current slice if it is
1684                  * idle and has expired it's mean thinktime or this new queue
1685                  * has some old slice time left and is of higher priority
1686                  */
1687                 cfq_preempt_queue(cfqd, cfqq);
1688                 cfq_mark_cfqq_must_dispatch(cfqq);
1689                 blk_start_queueing(cfqd->queue);
1690         }
1691 }
1692
1693 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1694 {
1695         struct cfq_data *cfqd = q->elevator->elevator_data;
1696         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1697
1698         cfq_init_prio_data(cfqq);
1699
1700         cfq_add_rq_rb(rq);
1701
1702         list_add_tail(&rq->queuelist, &cfqq->fifo);
1703
1704         cfq_rq_enqueued(cfqd, cfqq, rq);
1705 }
1706
1707 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1708 {
1709         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1710         struct cfq_data *cfqd = cfqq->cfqd;
1711         const int sync = rq_is_sync(rq);
1712         unsigned long now;
1713
1714         now = jiffies;
1715
1716         WARN_ON(!cfqd->rq_in_driver);
1717         WARN_ON(!cfqq->on_dispatch[sync]);
1718         cfqd->rq_in_driver--;
1719         cfqq->on_dispatch[sync]--;
1720         cfqq->service_last = now;
1721
1722         if (!cfq_class_idle(cfqq))
1723                 cfqd->last_end_request = now;
1724
1725         cfq_resort_rr_list(cfqq, 0);
1726
1727         if (sync)
1728                 RQ_CIC(rq)->last_end_request = now;
1729
1730         /*
1731          * If this is the active queue, check if it needs to be expired,
1732          * or if we want to idle in case it has no pending requests.
1733          */
1734         if (cfqd->active_queue == cfqq) {
1735                 if (cfq_cfqq_slice_new(cfqq)) {
1736                         cfq_set_prio_slice(cfqd, cfqq);
1737                         cfq_clear_cfqq_slice_new(cfqq);
1738                 }
1739                 if (cfq_slice_used(cfqq))
1740                         cfq_slice_expired(cfqd, 0, 1);
1741                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1742                         if (!cfq_arm_slice_timer(cfqd))
1743                                 cfq_schedule_dispatch(cfqd);
1744                 }
1745         }
1746 }
1747
1748 /*
1749  * we temporarily boost lower priority queues if they are holding fs exclusive
1750  * resources. they are boosted to normal prio (CLASS_BE/4)
1751  */
1752 static void cfq_prio_boost(struct cfq_queue *cfqq)
1753 {
1754         const int ioprio_class = cfqq->ioprio_class;
1755         const int ioprio = cfqq->ioprio;
1756
1757         if (has_fs_excl()) {
1758                 /*
1759                  * boost idle prio on transactions that would lock out other
1760                  * users of the filesystem
1761                  */
1762                 if (cfq_class_idle(cfqq))
1763                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1764                 if (cfqq->ioprio > IOPRIO_NORM)
1765                         cfqq->ioprio = IOPRIO_NORM;
1766         } else {
1767                 /*
1768                  * check if we need to unboost the queue
1769                  */
1770                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1771                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1772                 if (cfqq->ioprio != cfqq->org_ioprio)
1773                         cfqq->ioprio = cfqq->org_ioprio;
1774         }
1775
1776         /*
1777          * refile between round-robin lists if we moved the priority class
1778          */
1779         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio))
1780                 cfq_resort_rr_list(cfqq, 0);
1781 }
1782
1783 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1784 {
1785         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1786             !cfq_cfqq_must_alloc_slice(cfqq)) {
1787                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1788                 return ELV_MQUEUE_MUST;
1789         }
1790
1791         return ELV_MQUEUE_MAY;
1792 }
1793
1794 static int cfq_may_queue(request_queue_t *q, int rw)
1795 {
1796         struct cfq_data *cfqd = q->elevator->elevator_data;
1797         struct task_struct *tsk = current;
1798         struct cfq_queue *cfqq;
1799         unsigned int key;
1800
1801         key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
1802
1803         /*
1804          * don't force setup of a queue from here, as a call to may_queue
1805          * does not necessarily imply that a request actually will be queued.
1806          * so just lookup a possibly existing queue, or return 'may queue'
1807          * if that fails
1808          */
1809         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
1810         if (cfqq) {
1811                 cfq_init_prio_data(cfqq);
1812                 cfq_prio_boost(cfqq);
1813
1814                 return __cfq_may_queue(cfqq);
1815         }
1816
1817         return ELV_MQUEUE_MAY;
1818 }
1819
1820 /*
1821  * queue lock held here
1822  */
1823 static void cfq_put_request(struct request *rq)
1824 {
1825         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1826
1827         if (cfqq) {
1828                 const int rw = rq_data_dir(rq);
1829
1830                 BUG_ON(!cfqq->allocated[rw]);
1831                 cfqq->allocated[rw]--;
1832
1833                 put_io_context(RQ_CIC(rq)->ioc);
1834
1835                 rq->elevator_private = NULL;
1836                 rq->elevator_private2 = NULL;
1837
1838                 cfq_put_queue(cfqq);
1839         }
1840 }
1841
1842 /*
1843  * Allocate cfq data structures associated with this request.
1844  */
1845 static int
1846 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1847 {
1848         struct cfq_data *cfqd = q->elevator->elevator_data;
1849         struct task_struct *tsk = current;
1850         struct cfq_io_context *cic;
1851         const int rw = rq_data_dir(rq);
1852         const int is_sync = rq_is_sync(rq);
1853         pid_t key = cfq_queue_pid(tsk, rw, is_sync);
1854         struct cfq_queue *cfqq;
1855         unsigned long flags;
1856
1857         might_sleep_if(gfp_mask & __GFP_WAIT);
1858
1859         cic = cfq_get_io_context(cfqd, gfp_mask);
1860
1861         spin_lock_irqsave(q->queue_lock, flags);
1862
1863         if (!cic)
1864                 goto queue_fail;
1865
1866         if (!cic->cfqq[is_sync]) {
1867                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1868                 if (!cfqq)
1869                         goto queue_fail;
1870
1871                 cic->cfqq[is_sync] = cfqq;
1872         } else
1873                 cfqq = cic->cfqq[is_sync];
1874
1875         cfqq->allocated[rw]++;
1876         cfq_clear_cfqq_must_alloc(cfqq);
1877         atomic_inc(&cfqq->ref);
1878
1879         spin_unlock_irqrestore(q->queue_lock, flags);
1880
1881         rq->elevator_private = cic;
1882         rq->elevator_private2 = cfqq;
1883         return 0;
1884
1885 queue_fail:
1886         if (cic)
1887                 put_io_context(cic->ioc);
1888
1889         cfq_schedule_dispatch(cfqd);
1890         spin_unlock_irqrestore(q->queue_lock, flags);
1891         return 1;
1892 }
1893
1894 static void cfq_kick_queue(struct work_struct *work)
1895 {
1896         struct cfq_data *cfqd =
1897                 container_of(work, struct cfq_data, unplug_work);
1898         request_queue_t *q = cfqd->queue;
1899         unsigned long flags;
1900
1901         spin_lock_irqsave(q->queue_lock, flags);
1902         blk_start_queueing(q);
1903         spin_unlock_irqrestore(q->queue_lock, flags);
1904 }
1905
1906 /*
1907  * Timer running if the active_queue is currently idling inside its time slice
1908  */
1909 static void cfq_idle_slice_timer(unsigned long data)
1910 {
1911         struct cfq_data *cfqd = (struct cfq_data *) data;
1912         struct cfq_queue *cfqq;
1913         unsigned long flags;
1914         int timed_out = 1;
1915
1916         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1917
1918         if ((cfqq = cfqd->active_queue) != NULL) {
1919                 timed_out = 0;
1920
1921                 /*
1922                  * expired
1923                  */
1924                 if (cfq_slice_used(cfqq))
1925                         goto expire;
1926
1927                 /*
1928                  * only expire and reinvoke request handler, if there are
1929                  * other queues with pending requests
1930                  */
1931                 if (!cfqd->busy_queues)
1932                         goto out_cont;
1933
1934                 /*
1935                  * not expired and it has a request pending, let it dispatch
1936                  */
1937                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1938                         cfq_mark_cfqq_must_dispatch(cfqq);
1939                         goto out_kick;
1940                 }
1941         }
1942 expire:
1943         cfq_slice_expired(cfqd, 0, timed_out);
1944 out_kick:
1945         cfq_schedule_dispatch(cfqd);
1946 out_cont:
1947         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1948 }
1949
1950 /*
1951  * Timer running if an idle class queue is waiting for service
1952  */
1953 static void cfq_idle_class_timer(unsigned long data)
1954 {
1955         struct cfq_data *cfqd = (struct cfq_data *) data;
1956         unsigned long flags, end;
1957
1958         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1959
1960         /*
1961          * race with a non-idle queue, reset timer
1962          */
1963         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
1964         if (!time_after_eq(jiffies, end))
1965                 mod_timer(&cfqd->idle_class_timer, end);
1966         else
1967                 cfq_schedule_dispatch(cfqd);
1968
1969         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1970 }
1971
1972 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1973 {
1974         del_timer_sync(&cfqd->idle_slice_timer);
1975         del_timer_sync(&cfqd->idle_class_timer);
1976         blk_sync_queue(cfqd->queue);
1977 }
1978
1979 static void cfq_exit_queue(elevator_t *e)
1980 {
1981         struct cfq_data *cfqd = e->elevator_data;
1982         request_queue_t *q = cfqd->queue;
1983
1984         cfq_shutdown_timer_wq(cfqd);
1985
1986         spin_lock_irq(q->queue_lock);
1987
1988         if (cfqd->active_queue)
1989                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0, 0);
1990
1991         while (!list_empty(&cfqd->cic_list)) {
1992                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1993                                                         struct cfq_io_context,
1994                                                         queue_list);
1995
1996                 __cfq_exit_single_io_context(cfqd, cic);
1997         }
1998
1999         spin_unlock_irq(q->queue_lock);
2000
2001         cfq_shutdown_timer_wq(cfqd);
2002
2003         kfree(cfqd->cfq_hash);
2004         kfree(cfqd);
2005 }
2006
2007 static void *cfq_init_queue(request_queue_t *q)
2008 {
2009         struct cfq_data *cfqd;
2010         int i;
2011
2012         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
2013         if (!cfqd)
2014                 return NULL;
2015
2016         memset(cfqd, 0, sizeof(*cfqd));
2017
2018         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2019                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2020
2021         INIT_LIST_HEAD(&cfqd->busy_rr);
2022         INIT_LIST_HEAD(&cfqd->cur_rr);
2023         INIT_LIST_HEAD(&cfqd->idle_rr);
2024         INIT_LIST_HEAD(&cfqd->cic_list);
2025
2026         cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
2027         if (!cfqd->cfq_hash)
2028                 goto out_free;
2029
2030         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2031                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2032
2033         cfqd->queue = q;
2034
2035         init_timer(&cfqd->idle_slice_timer);
2036         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2037         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2038
2039         init_timer(&cfqd->idle_class_timer);
2040         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2041         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2042
2043         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2044
2045         cfqd->cfq_quantum = cfq_quantum;
2046         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2047         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2048         cfqd->cfq_back_max = cfq_back_max;
2049         cfqd->cfq_back_penalty = cfq_back_penalty;
2050         cfqd->cfq_slice[0] = cfq_slice_async;
2051         cfqd->cfq_slice[1] = cfq_slice_sync;
2052         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2053         cfqd->cfq_slice_idle = cfq_slice_idle;
2054
2055         return cfqd;
2056 out_free:
2057         kfree(cfqd);
2058         return NULL;
2059 }
2060
2061 static void cfq_slab_kill(void)
2062 {
2063         if (cfq_pool)
2064                 kmem_cache_destroy(cfq_pool);
2065         if (cfq_ioc_pool)
2066                 kmem_cache_destroy(cfq_ioc_pool);
2067 }
2068
2069 static int __init cfq_slab_setup(void)
2070 {
2071         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2072                                         NULL, NULL);
2073         if (!cfq_pool)
2074                 goto fail;
2075
2076         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2077                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2078         if (!cfq_ioc_pool)
2079                 goto fail;
2080
2081         return 0;
2082 fail:
2083         cfq_slab_kill();
2084         return -ENOMEM;
2085 }
2086
2087 /*
2088  * sysfs parts below -->
2089  */
2090
2091 static ssize_t
2092 cfq_var_show(unsigned int var, char *page)
2093 {
2094         return sprintf(page, "%d\n", var);
2095 }
2096
2097 static ssize_t
2098 cfq_var_store(unsigned int *var, const char *page, size_t count)
2099 {
2100         char *p = (char *) page;
2101
2102         *var = simple_strtoul(p, &p, 10);
2103         return count;
2104 }
2105
2106 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2107 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2108 {                                                                       \
2109         struct cfq_data *cfqd = e->elevator_data;                       \
2110         unsigned int __data = __VAR;                                    \
2111         if (__CONV)                                                     \
2112                 __data = jiffies_to_msecs(__data);                      \
2113         return cfq_var_show(__data, (page));                            \
2114 }
2115 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2116 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2117 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2118 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2119 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2120 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2121 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2122 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2123 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2124 #undef SHOW_FUNCTION
2125
2126 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2127 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2128 {                                                                       \
2129         struct cfq_data *cfqd = e->elevator_data;                       \
2130         unsigned int __data;                                            \
2131         int ret = cfq_var_store(&__data, (page), count);                \
2132         if (__data < (MIN))                                             \
2133                 __data = (MIN);                                         \
2134         else if (__data > (MAX))                                        \
2135                 __data = (MAX);                                         \
2136         if (__CONV)                                                     \
2137                 *(__PTR) = msecs_to_jiffies(__data);                    \
2138         else                                                            \
2139                 *(__PTR) = __data;                                      \
2140         return ret;                                                     \
2141 }
2142 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2143 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2144 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2145 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2146 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2147 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2148 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2149 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2150 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2151 #undef STORE_FUNCTION
2152
2153 #define CFQ_ATTR(name) \
2154         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2155
2156 static struct elv_fs_entry cfq_attrs[] = {
2157         CFQ_ATTR(quantum),
2158         CFQ_ATTR(fifo_expire_sync),
2159         CFQ_ATTR(fifo_expire_async),
2160         CFQ_ATTR(back_seek_max),
2161         CFQ_ATTR(back_seek_penalty),
2162         CFQ_ATTR(slice_sync),
2163         CFQ_ATTR(slice_async),
2164         CFQ_ATTR(slice_async_rq),
2165         CFQ_ATTR(slice_idle),
2166         __ATTR_NULL
2167 };
2168
2169 static struct elevator_type iosched_cfq = {
2170         .ops = {
2171                 .elevator_merge_fn =            cfq_merge,
2172                 .elevator_merged_fn =           cfq_merged_request,
2173                 .elevator_merge_req_fn =        cfq_merged_requests,
2174                 .elevator_allow_merge_fn =      cfq_allow_merge,
2175                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2176                 .elevator_add_req_fn =          cfq_insert_request,
2177                 .elevator_activate_req_fn =     cfq_activate_request,
2178                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2179                 .elevator_queue_empty_fn =      cfq_queue_empty,
2180                 .elevator_completed_req_fn =    cfq_completed_request,
2181                 .elevator_former_req_fn =       elv_rb_former_request,
2182                 .elevator_latter_req_fn =       elv_rb_latter_request,
2183                 .elevator_set_req_fn =          cfq_set_request,
2184                 .elevator_put_req_fn =          cfq_put_request,
2185                 .elevator_may_queue_fn =        cfq_may_queue,
2186                 .elevator_init_fn =             cfq_init_queue,
2187                 .elevator_exit_fn =             cfq_exit_queue,
2188                 .trim =                         cfq_free_io_context,
2189         },
2190         .elevator_attrs =       cfq_attrs,
2191         .elevator_name =        "cfq",
2192         .elevator_owner =       THIS_MODULE,
2193 };
2194
2195 static int __init cfq_init(void)
2196 {
2197         int ret;
2198
2199         /*
2200          * could be 0 on HZ < 1000 setups
2201          */
2202         if (!cfq_slice_async)
2203                 cfq_slice_async = 1;
2204         if (!cfq_slice_idle)
2205                 cfq_slice_idle = 1;
2206
2207         if (cfq_slab_setup())
2208                 return -ENOMEM;
2209
2210         ret = elv_register(&iosched_cfq);
2211         if (ret)
2212                 cfq_slab_kill();
2213
2214         return ret;
2215 }
2216
2217 static void __exit cfq_exit(void)
2218 {
2219         DECLARE_COMPLETION_ONSTACK(all_gone);
2220         elv_unregister(&iosched_cfq);
2221         ioc_gone = &all_gone;
2222         /* ioc_gone's update must be visible before reading ioc_count */
2223         smp_wmb();
2224         if (elv_ioc_count_read(ioc_count))
2225                 wait_for_completion(ioc_gone);
2226         synchronize_rcu();
2227         cfq_slab_kill();
2228 }
2229
2230 module_init(cfq_init);
2231 module_exit(cfq_exit);
2232
2233 MODULE_AUTHOR("Jens Axboe");
2234 MODULE_LICENSE("GPL");
2235 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");