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

   1 /*
   2  *  CFQ, or complete fairness queueing, disk scheduler.
   3  *
   4  *  Based on ideas from a previously unfinished io
   5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   6  *
   7  *  Copyright (C) 2003 Jens Axboe <axboe@suse.de>
   8  */
   9 #include <linux/config.h>
  10 #include <linux/module.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/elevator.h>
  13 #include <linux/hash.h>
  14 #include <linux/rbtree.h>
  15 #include <linux/ioprio.h>
  16
  17 /*
  18  * tunables
  19  */
  20 static const int cfq_quantum = 4;               /* max queue in one round of service */
  21 static const int cfq_queued = 8;                /* minimum rq allocate limit per-queue*/
  22 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  23 static const int cfq_back_max = 16 * 1024;      /* maximum backwards seek, in KiB */
  24 static const int cfq_back_penalty = 2;          /* penalty of a backwards seek */
  25
  26 static const int cfq_slice_sync = HZ / 10;
  27 static int cfq_slice_async = HZ / 25;
  28 static const int cfq_slice_async_rq = 2;
  29 static int cfq_slice_idle = HZ / 125;
  30
  31 #define CFQ_IDLE_GRACE          (HZ / 10)
  32 #define CFQ_SLICE_SCALE         (5)
  33
  34 #define CFQ_KEY_ASYNC           (0)
  35
  36 static DEFINE_SPINLOCK(cfq_exit_lock);
  37
  38 /*
  39  * for the hash of cfqq inside the cfqd
  40  */
  41 #define CFQ_QHASH_SHIFT         6
  42 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
  43 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
  44
  45 /*
  46  * for the hash of crq inside the cfqq
  47  */
  48 #define CFQ_MHASH_SHIFT         6
  49 #define CFQ_MHASH_BLOCK(sec)    ((sec) >> 3)
  50 #define CFQ_MHASH_ENTRIES       (1 << CFQ_MHASH_SHIFT)
  51 #define CFQ_MHASH_FN(sec)       hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
  52 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
  53 #define list_entry_hash(ptr)    hlist_entry((ptr), struct cfq_rq, hash)
  54
  55 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
  56 #define list_entry_fifo(ptr)    list_entry((ptr), struct request, queuelist)
  57
  58 #define RQ_DATA(rq)             (rq)->elevator_private
  59
  60 /*
  61  * rb-tree defines
  62  */
  63 #define rb_entry_crq(node)      rb_entry((node), struct cfq_rq, rb_node)
  64 #define rq_rb_key(rq)           (rq)->sector
  65
  66 static kmem_cache_t *crq_pool;
  67 static kmem_cache_t *cfq_pool;
  68 static kmem_cache_t *cfq_ioc_pool;
  69
  70 static atomic_t ioc_count = ATOMIC_INIT(0);
  71 static struct completion *ioc_gone;
  72
  73 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  74 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  75 #define cfq_class_be(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
  76 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  77
  78 #define ASYNC                   (0)
  79 #define SYNC                    (1)
  80
  81 #define cfq_cfqq_dispatched(cfqq)       \
  82         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
  83
  84 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
  85
  86 #define cfq_cfqq_sync(cfqq)             \
  87         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
  88
  89 #define sample_valid(samples)   ((samples) > 80)
  90
  91 /*
  92  * Per block device queue structure
  93  */
  94 struct cfq_data {
  95         request_queue_t *queue;
  96
  97         /*
  98          * rr list of queues with requests and the count of them
  99          */
 100         struct list_head rr_list[CFQ_PRIO_LISTS];
 101         struct list_head busy_rr;
 102         struct list_head cur_rr;
 103         struct list_head idle_rr;
 104         unsigned int busy_queues;
 105
 106         /*
 107          * non-ordered list of empty cfqq's
 108          */
 109         struct list_head empty_list;
 110
 111         /*
 112          * cfqq lookup hash
 113          */
 114         struct hlist_head *cfq_hash;
 115
 116         /*
 117          * global crq hash for all queues
 118          */
 119         struct hlist_head *crq_hash;
 120
 121         mempool_t *crq_pool;
 122
 123         int rq_in_driver;
 124         int hw_tag;
 125
 126         /*
 127          * schedule slice state info
 128          */
 129         /*
 130          * idle window management
 131          */
 132         struct timer_list idle_slice_timer;
 133         struct work_struct unplug_work;
 134
 135         struct cfq_queue *active_queue;
 136         struct cfq_io_context *active_cic;
 137         int cur_prio, cur_end_prio;
 138         unsigned int dispatch_slice;
 139
 140         struct timer_list idle_class_timer;
 141
 142         sector_t last_sector;
 143         unsigned long last_end_request;
 144
 145         unsigned int rq_starved;
 146
 147         /*
 148          * tunables, see top of file
 149          */
 150         unsigned int cfq_quantum;
 151         unsigned int cfq_queued;
 152         unsigned int cfq_fifo_expire[2];
 153         unsigned int cfq_back_penalty;
 154         unsigned int cfq_back_max;
 155         unsigned int cfq_slice[2];
 156         unsigned int cfq_slice_async_rq;
 157         unsigned int cfq_slice_idle;
 158
 159         struct list_head cic_list;
 160 };
 161
 162 /*
 163  * Per process-grouping structure
 164  */
 165 struct cfq_queue {
 166         /* reference count */
 167         atomic_t ref;
 168         /* parent cfq_data */
 169         struct cfq_data *cfqd;
 170         /* cfqq lookup hash */
 171         struct hlist_node cfq_hash;
 172         /* hash key */
 173         unsigned int key;
 174         /* on either rr or empty list of cfqd */
 175         struct list_head cfq_list;
 176         /* sorted list of pending requests */
 177         struct rb_root sort_list;
 178         /* if fifo isn't expired, next request to serve */
 179         struct cfq_rq *next_crq;
 180         /* requests queued in sort_list */
 181         int queued[2];
 182         /* currently allocated requests */
 183         int allocated[2];
 184         /* fifo list of requests in sort_list */
 185         struct list_head fifo;
 186
 187         unsigned long slice_start;
 188         unsigned long slice_end;
 189         unsigned long slice_left;
 190         unsigned long service_last;
 191
 192         /* number of requests that are on the dispatch list */
 193         int on_dispatch[2];
 194
 195         /* io prio of this group */
 196         unsigned short ioprio, org_ioprio;
 197         unsigned short ioprio_class, org_ioprio_class;
 198
 199         /* various state flags, see below */
 200         unsigned int flags;
 201 };
 202
 203 struct cfq_rq {
 204         struct rb_node rb_node;
 205         sector_t rb_key;
 206         struct request *request;
 207         struct hlist_node hash;
 208
 209         struct cfq_queue *cfq_queue;
 210         struct cfq_io_context *io_context;
 211
 212         unsigned int crq_flags;
 213 };
 214
 215 enum cfqq_state_flags {
 216         CFQ_CFQQ_FLAG_on_rr = 0,
 217         CFQ_CFQQ_FLAG_wait_request,
 218         CFQ_CFQQ_FLAG_must_alloc,
 219         CFQ_CFQQ_FLAG_must_alloc_slice,
 220         CFQ_CFQQ_FLAG_must_dispatch,
 221         CFQ_CFQQ_FLAG_fifo_expire,
 222         CFQ_CFQQ_FLAG_idle_window,
 223         CFQ_CFQQ_FLAG_prio_changed,
 224 };
 225
 226 #define CFQ_CFQQ_FNS(name)                                              \
 227 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 228 {                                                                       \
 229         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 230 }                                                                       \
 231 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 232 {                                                                       \
 233         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 234 }                                                                       \
 235 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 236 {                                                                       \
 237         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 238 }
 239
 240 CFQ_CFQQ_FNS(on_rr);
 241 CFQ_CFQQ_FNS(wait_request);
 242 CFQ_CFQQ_FNS(must_alloc);
 243 CFQ_CFQQ_FNS(must_alloc_slice);
 244 CFQ_CFQQ_FNS(must_dispatch);
 245 CFQ_CFQQ_FNS(fifo_expire);
 246 CFQ_CFQQ_FNS(idle_window);
 247 CFQ_CFQQ_FNS(prio_changed);
 248 #undef CFQ_CFQQ_FNS
 249
 250 enum cfq_rq_state_flags {
 251         CFQ_CRQ_FLAG_is_sync = 0,
 252 };
 253
 254 #define CFQ_CRQ_FNS(name)                                               \
 255 static inline void cfq_mark_crq_##name(struct cfq_rq *crq)              \
 256 {                                                                       \
 257         crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name);                   \
 258 }                                                                       \
 259 static inline void cfq_clear_crq_##name(struct cfq_rq *crq)             \
 260 {                                                                       \
 261         crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name);                  \
 262 }                                                                       \
 263 static inline int cfq_crq_##name(const struct cfq_rq *crq)              \
 264 {                                                                       \
 265         return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0;      \
 266 }
 267
 268 CFQ_CRQ_FNS(is_sync);
 269 #undef CFQ_CRQ_FNS
 270
 271 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 272 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
 273 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
 274
 275 /*
 276  * lots of deadline iosched dupes, can be abstracted later...
 277  */
 278 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
 279 {
 280         hlist_del_init(&crq->hash);
 281 }
 282
 283 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
 284 {
 285         const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
 286
 287         hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
 288 }
 289
 290 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
 291 {
 292         struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
 293         struct hlist_node *entry, *next;
 294
 295         hlist_for_each_safe(entry, next, hash_list) {
 296                 struct cfq_rq *crq = list_entry_hash(entry);
 297                 struct request *__rq = crq->request;
 298
 299                 if (!rq_mergeable(__rq)) {
 300                         cfq_del_crq_hash(crq);
 301                         continue;
 302                 }
 303
 304                 if (rq_hash_key(__rq) == offset)
 305                         return __rq;
 306         }
 307
 308         return NULL;
 309 }
 310
 311 /*
 312  * scheduler run of queue, if there are requests pending and no one in the
 313  * driver that will restart queueing
 314  */
 315 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 316 {
 317         if (cfqd->busy_queues)
 318                 kblockd_schedule_work(&cfqd->unplug_work);
 319 }
 320
 321 static int cfq_queue_empty(request_queue_t *q)
 322 {
 323         struct cfq_data *cfqd = q->elevator->elevator_data;
 324
 325         return !cfqd->busy_queues;
 326 }
 327
 328 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
 329 {
 330         if (rw == READ || rw == WRITE_SYNC)
 331                 return task->pid;
 332
 333         return CFQ_KEY_ASYNC;
 334 }
 335
 336 /*
 337  * Lifted from AS - choose which of crq1 and crq2 that is best served now.
 338  * We choose the request that is closest to the head right now. Distance
 339  * behind the head is penalized and only allowed to a certain extent.
 340  */
 341 static struct cfq_rq *
 342 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
 343 {
 344         sector_t last, s1, s2, d1 = 0, d2 = 0;
 345         unsigned long back_max;
 346 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 347 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 348         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 349
 350         if (crq1 == NULL || crq1 == crq2)
 351                 return crq2;
 352         if (crq2 == NULL)
 353                 return crq1;
 354
 355         if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
 356                 return crq1;
 357         else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
 358                 return crq2;
 359
 360         s1 = crq1->request->sector;
 361         s2 = crq2->request->sector;
 362
 363         last = cfqd->last_sector;
 364
 365         /*
 366          * by definition, 1KiB is 2 sectors
 367          */
 368         back_max = cfqd->cfq_back_max * 2;
 369
 370         /*
 371          * Strict one way elevator _except_ in the case where we allow
 372          * short backward seeks which are biased as twice the cost of a
 373          * similar forward seek.
 374          */
 375         if (s1 >= last)
 376                 d1 = s1 - last;
 377         else if (s1 + back_max >= last)
 378                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 379         else
 380                 wrap |= CFQ_RQ1_WRAP;
 381
 382         if (s2 >= last)
 383                 d2 = s2 - last;
 384         else if (s2 + back_max >= last)
 385                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 386         else
 387                 wrap |= CFQ_RQ2_WRAP;
 388
 389         /* Found required data */
 390
 391         /*
 392          * By doing switch() on the bit mask "wrap" we avoid having to
 393          * check two variables for all permutations: --> faster!
 394          */
 395         switch (wrap) {
 396         case 0: /* common case for CFQ: crq1 and crq2 not wrapped */
 397                 if (d1 < d2)
 398                         return crq1;
 399                 else if (d2 < d1)
 400                         return crq2;
 401                 else {
 402                         if (s1 >= s2)
 403                                 return crq1;
 404                         else
 405                                 return crq2;
 406                 }
 407
 408         case CFQ_RQ2_WRAP:
 409                 return crq1;
 410         case CFQ_RQ1_WRAP:
 411                 return crq2;
 412         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both crqs wrapped */
 413         default:
 414                 /*
 415                  * Since both rqs are wrapped,
 416                  * start with the one that's further behind head
 417                  * (--> only *one* back seek required),
 418                  * since back seek takes more time than forward.
 419                  */
 420                 if (s1 <= s2)
 421                         return crq1;
 422                 else
 423                         return crq2;
 424         }
 425 }
 426
 427 /*
 428  * would be nice to take fifo expire time into account as well
 429  */
 430 static struct cfq_rq *
 431 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 432                   struct cfq_rq *last)
 433 {
 434         struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
 435         struct rb_node *rbnext, *rbprev;
 436
 437         if (!(rbnext = rb_next(&last->rb_node))) {
 438                 rbnext = rb_first(&cfqq->sort_list);
 439                 if (rbnext == &last->rb_node)
 440                         rbnext = NULL;
 441         }
 442
 443         rbprev = rb_prev(&last->rb_node);
 444
 445         if (rbprev)
 446                 crq_prev = rb_entry_crq(rbprev);
 447         if (rbnext)
 448                 crq_next = rb_entry_crq(rbnext);
 449
 450         return cfq_choose_req(cfqd, crq_next, crq_prev);
 451 }
 452
 453 static void cfq_update_next_crq(struct cfq_rq *crq)
 454 {
 455         struct cfq_queue *cfqq = crq->cfq_queue;
 456
 457         if (cfqq->next_crq == crq)
 458                 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
 459 }
 460
 461 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 462 {
 463         struct cfq_data *cfqd = cfqq->cfqd;
 464         struct list_head *list, *entry;
 465
 466         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 467
 468         list_del(&cfqq->cfq_list);
 469
 470         if (cfq_class_rt(cfqq))
 471                 list = &cfqd->cur_rr;
 472         else if (cfq_class_idle(cfqq))
 473                 list = &cfqd->idle_rr;
 474         else {
 475                 /*
 476                  * if cfqq has requests in flight, don't allow it to be
 477                  * found in cfq_set_active_queue before it has finished them.
 478                  * this is done to increase fairness between a process that
 479                  * has lots of io pending vs one that only generates one
 480                  * sporadically or synchronously
 481                  */
 482                 if (cfq_cfqq_dispatched(cfqq))
 483                         list = &cfqd->busy_rr;
 484                 else
 485                         list = &cfqd->rr_list[cfqq->ioprio];
 486         }
 487
 488         /*
 489          * if queue was preempted, just add to front to be fair. busy_rr
 490          * isn't sorted, but insert at the back for fairness.
 491          */
 492         if (preempted || list == &cfqd->busy_rr) {
 493                 if (preempted)
 494                         list = list->prev;
 495
 496                 list_add_tail(&cfqq->cfq_list, list);
 497                 return;
 498         }
 499
 500         /*
 501          * sort by when queue was last serviced
 502          */
 503         entry = list;
 504         while ((entry = entry->prev) != list) {
 505                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
 506
 507                 if (!__cfqq->service_last)
 508                         break;
 509                 if (time_before(__cfqq->service_last, cfqq->service_last))
 510                         break;
 511         }
 512
 513         list_add(&cfqq->cfq_list, entry);
 514 }
 515
 516 /*
 517  * add to busy list of queues for service, trying to be fair in ordering
 518  * the pending list according to last request service
 519  */
 520 static inline void
 521 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 522 {
 523         BUG_ON(cfq_cfqq_on_rr(cfqq));
 524         cfq_mark_cfqq_on_rr(cfqq);
 525         cfqd->busy_queues++;
 526
 527         cfq_resort_rr_list(cfqq, 0);
 528 }
 529
 530 static inline void
 531 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 532 {
 533         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 534         cfq_clear_cfqq_on_rr(cfqq);
 535         list_move(&cfqq->cfq_list, &cfqd->empty_list);
 536
 537         BUG_ON(!cfqd->busy_queues);
 538         cfqd->busy_queues--;
 539 }
 540
 541 /*
 542  * rb tree support functions
 543  */
 544 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
 545 {
 546         struct cfq_queue *cfqq = crq->cfq_queue;
 547         struct cfq_data *cfqd = cfqq->cfqd;
 548         const int sync = cfq_crq_is_sync(crq);
 549
 550         BUG_ON(!cfqq->queued[sync]);
 551         cfqq->queued[sync]--;
 552
 553         cfq_update_next_crq(crq);
 554
 555         rb_erase(&crq->rb_node, &cfqq->sort_list);
 556
 557         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
 558                 cfq_del_cfqq_rr(cfqd, cfqq);
 559 }
 560
 561 static struct cfq_rq *
 562 __cfq_add_crq_rb(struct cfq_rq *crq)
 563 {
 564         struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
 565         struct rb_node *parent = NULL;
 566         struct cfq_rq *__crq;
 567
 568         while (*p) {
 569                 parent = *p;
 570                 __crq = rb_entry_crq(parent);
 571
 572                 if (crq->rb_key < __crq->rb_key)
 573                         p = &(*p)->rb_left;
 574                 else if (crq->rb_key > __crq->rb_key)
 575                         p = &(*p)->rb_right;
 576                 else
 577                         return __crq;
 578         }
 579
 580         rb_link_node(&crq->rb_node, parent, p);
 581         return NULL;
 582 }
 583
 584 static void cfq_add_crq_rb(struct cfq_rq *crq)
 585 {
 586         struct cfq_queue *cfqq = crq->cfq_queue;
 587         struct cfq_data *cfqd = cfqq->cfqd;
 588         struct request *rq = crq->request;
 589         struct cfq_rq *__alias;
 590
 591         crq->rb_key = rq_rb_key(rq);
 592         cfqq->queued[cfq_crq_is_sync(crq)]++;
 593
 594         /*
 595          * looks a little odd, but the first insert might return an alias.
 596          * if that happens, put the alias on the dispatch list
 597          */
 598         while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
 599                 cfq_dispatch_insert(cfqd->queue, __alias);
 600
 601         rb_insert_color(&crq->rb_node, &cfqq->sort_list);
 602
 603         if (!cfq_cfqq_on_rr(cfqq))
 604                 cfq_add_cfqq_rr(cfqd, cfqq);
 605
 606         /*
 607          * check if this request is a better next-serve candidate
 608          */
 609         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
 610 }
 611
 612 static inline void
 613 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
 614 {
 615         rb_erase(&crq->rb_node, &cfqq->sort_list);
 616         cfqq->queued[cfq_crq_is_sync(crq)]--;
 617
 618         cfq_add_crq_rb(crq);
 619 }
 620
 621 static struct request *
 622 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 623 {
 624         struct task_struct *tsk = current;
 625         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
 626         struct cfq_queue *cfqq;
 627         struct rb_node *n;
 628         sector_t sector;
 629
 630         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
 631         if (!cfqq)
 632                 goto out;
 633
 634         sector = bio->bi_sector + bio_sectors(bio);
 635         n = cfqq->sort_list.rb_node;
 636         while (n) {
 637                 struct cfq_rq *crq = rb_entry_crq(n);
 638
 639                 if (sector < crq->rb_key)
 640                         n = n->rb_left;
 641                 else if (sector > crq->rb_key)
 642                         n = n->rb_right;
 643                 else
 644                         return crq->request;
 645         }
 646
 647 out:
 648         return NULL;
 649 }
 650
 651 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 652 {
 653         struct cfq_data *cfqd = q->elevator->elevator_data;
 654
 655         cfqd->rq_in_driver++;
 656
 657         /*
 658          * If the depth is larger 1, it really could be queueing. But lets
 659          * make the mark a little higher - idling could still be good for
 660          * low queueing, and a low queueing number could also just indicate
 661          * a SCSI mid layer like behaviour where limit+1 is often seen.
 662          */
 663         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
 664                 cfqd->hw_tag = 1;
 665 }
 666
 667 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 668 {
 669         struct cfq_data *cfqd = q->elevator->elevator_data;
 670
 671         WARN_ON(!cfqd->rq_in_driver);
 672         cfqd->rq_in_driver--;
 673 }
 674
 675 static void cfq_remove_request(struct request *rq)
 676 {
 677         struct cfq_rq *crq = RQ_DATA(rq);
 678
 679         list_del_init(&rq->queuelist);
 680         cfq_del_crq_rb(crq);
 681         cfq_del_crq_hash(crq);
 682 }
 683
 684 static int
 685 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 686 {
 687         struct cfq_data *cfqd = q->elevator->elevator_data;
 688         struct request *__rq;
 689         int ret;
 690
 691         __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
 692         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 693                 ret = ELEVATOR_BACK_MERGE;
 694                 goto out;
 695         }
 696
 697         __rq = cfq_find_rq_fmerge(cfqd, bio);
 698         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 699                 ret = ELEVATOR_FRONT_MERGE;
 700                 goto out;
 701         }
 702
 703         return ELEVATOR_NO_MERGE;
 704 out:
 705         *req = __rq;
 706         return ret;
 707 }
 708
 709 static void cfq_merged_request(request_queue_t *q, struct request *req)
 710 {
 711         struct cfq_data *cfqd = q->elevator->elevator_data;
 712         struct cfq_rq *crq = RQ_DATA(req);
 713
 714         cfq_del_crq_hash(crq);
 715         cfq_add_crq_hash(cfqd, crq);
 716
 717         if (rq_rb_key(req) != crq->rb_key) {
 718                 struct cfq_queue *cfqq = crq->cfq_queue;
 719
 720                 cfq_update_next_crq(crq);
 721                 cfq_reposition_crq_rb(cfqq, crq);
 722         }
 723 }
 724
 725 static void
 726 cfq_merged_requests(request_queue_t *q, struct request *rq,
 727                     struct request *next)
 728 {
 729         cfq_merged_request(q, rq);
 730
 731         /*
 732          * reposition in fifo if next is older than rq
 733          */
 734         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 735             time_before(next->start_time, rq->start_time))
 736                 list_move(&rq->queuelist, &next->queuelist);
 737
 738         cfq_remove_request(next);
 739 }
 740
 741 static inline void
 742 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 743 {
 744         if (cfqq) {
 745                 /*
 746                  * stop potential idle class queues waiting service
 747                  */
 748                 del_timer(&cfqd->idle_class_timer);
 749
 750                 cfqq->slice_start = jiffies;
 751                 cfqq->slice_end = 0;
 752                 cfqq->slice_left = 0;
 753                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 754                 cfq_clear_cfqq_fifo_expire(cfqq);
 755         }
 756
 757         cfqd->active_queue = cfqq;
 758 }
 759
 760 /*
 761  * current cfqq expired its slice (or was too idle), select new one
 762  */
 763 static void
 764 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 765                     int preempted)
 766 {
 767         unsigned long now = jiffies;
 768
 769         if (cfq_cfqq_wait_request(cfqq))
 770                 del_timer(&cfqd->idle_slice_timer);
 771
 772         if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
 773                 cfqq->service_last = now;
 774                 cfq_schedule_dispatch(cfqd);
 775         }
 776
 777         cfq_clear_cfqq_must_dispatch(cfqq);
 778         cfq_clear_cfqq_wait_request(cfqq);
 779
 780         /*
 781          * store what was left of this slice, if the queue idled out
 782          * or was preempted
 783          */
 784         if (time_after(cfqq->slice_end, now))
 785                 cfqq->slice_left = cfqq->slice_end - now;
 786         else
 787                 cfqq->slice_left = 0;
 788
 789         if (cfq_cfqq_on_rr(cfqq))
 790                 cfq_resort_rr_list(cfqq, preempted);
 791
 792         if (cfqq == cfqd->active_queue)
 793                 cfqd->active_queue = NULL;
 794
 795         if (cfqd->active_cic) {
 796                 put_io_context(cfqd->active_cic->ioc);
 797                 cfqd->active_cic = NULL;
 798         }
 799
 800         cfqd->dispatch_slice = 0;
 801 }
 802
 803 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 804 {
 805         struct cfq_queue *cfqq = cfqd->active_queue;
 806
 807         if (cfqq)
 808                 __cfq_slice_expired(cfqd, cfqq, preempted);
 809 }
 810
 811 /*
 812  * 0
 813  * 0,1
 814  * 0,1,2
 815  * 0,1,2,3
 816  * 0,1,2,3,4
 817  * 0,1,2,3,4,5
 818  * 0,1,2,3,4,5,6
 819  * 0,1,2,3,4,5,6,7
 820  */
 821 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 822 {
 823         int prio, wrap;
 824
 825         prio = -1;
 826         wrap = 0;
 827         do {
 828                 int p;
 829
 830                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 831                         if (!list_empty(&cfqd->rr_list[p])) {
 832                                 prio = p;
 833                                 break;
 834                         }
 835                 }
 836
 837                 if (prio != -1)
 838                         break;
 839                 cfqd->cur_prio = 0;
 840                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 841                         cfqd->cur_end_prio = 0;
 842                         if (wrap)
 843                                 break;
 844                         wrap = 1;
 845                 }
 846         } while (1);
 847
 848         if (unlikely(prio == -1))
 849                 return -1;
 850
 851         BUG_ON(prio >= CFQ_PRIO_LISTS);
 852
 853         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 854
 855         cfqd->cur_prio = prio + 1;
 856         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 857                 cfqd->cur_end_prio = cfqd->cur_prio;
 858                 cfqd->cur_prio = 0;
 859         }
 860         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 861                 cfqd->cur_prio = 0;
 862                 cfqd->cur_end_prio = 0;
 863         }
 864
 865         return prio;
 866 }
 867
 868 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 869 {
 870         struct cfq_queue *cfqq = NULL;
 871
 872         /*
 873          * if current list is non-empty, grab first entry. if it is empty,
 874          * get next prio level and grab first entry then if any are spliced
 875          */
 876         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
 877                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 878
 879         /*
 880          * If no new queues are available, check if the busy list has some
 881          * before falling back to idle io.
 882          */
 883         if (!cfqq && !list_empty(&cfqd->busy_rr))
 884                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
 885
 886         /*
 887          * if we have idle queues and no rt or be queues had pending
 888          * requests, either allow immediate service if the grace period
 889          * has passed or arm the idle grace timer
 890          */
 891         if (!cfqq && !list_empty(&cfqd->idle_rr)) {
 892                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 893
 894                 if (time_after_eq(jiffies, end))
 895                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 896                 else
 897                         mod_timer(&cfqd->idle_class_timer, end);
 898         }
 899
 900         __cfq_set_active_queue(cfqd, cfqq);
 901         return cfqq;
 902 }
 903
 904 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
 905
 906 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 907
 908 {
 909         struct cfq_io_context *cic;
 910         unsigned long sl;
 911
 912         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
 913         WARN_ON(cfqq != cfqd->active_queue);
 914
 915         /*
 916          * idle is disabled, either manually or by past process history
 917          */
 918         if (!cfqd->cfq_slice_idle)
 919                 return 0;
 920         if (!cfq_cfqq_idle_window(cfqq))
 921                 return 0;
 922         /*
 923          * task has exited, don't wait
 924          */
 925         cic = cfqd->active_cic;
 926         if (!cic || !cic->ioc->task)
 927                 return 0;
 928
 929         cfq_mark_cfqq_must_dispatch(cfqq);
 930         cfq_mark_cfqq_wait_request(cfqq);
 931
 932         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 933
 934         /*
 935          * we don't want to idle for seeks, but we do want to allow
 936          * fair distribution of slice time for a process doing back-to-back
 937          * seeks. so allow a little bit of time for him to submit a new rq
 938          */
 939         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
 940                 sl = 2;
 941
 942         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 943         return 1;
 944 }
 945
 946 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
 947 {
 948         struct cfq_data *cfqd = q->elevator->elevator_data;
 949         struct cfq_queue *cfqq = crq->cfq_queue;
 950         struct request *rq;
 951
 952         cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
 953         cfq_remove_request(crq->request);
 954         cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
 955         elv_dispatch_sort(q, crq->request);
 956
 957         rq = list_entry(q->queue_head.prev, struct request, queuelist);
 958         cfqd->last_sector = rq->sector + rq->nr_sectors;
 959 }
 960
 961 /*
 962  * return expired entry, or NULL to just start from scratch in rbtree
 963  */
 964 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
 965 {
 966         struct cfq_data *cfqd = cfqq->cfqd;
 967         struct request *rq;
 968         struct cfq_rq *crq;
 969
 970         if (cfq_cfqq_fifo_expire(cfqq))
 971                 return NULL;
 972
 973         if (!list_empty(&cfqq->fifo)) {
 974                 int fifo = cfq_cfqq_class_sync(cfqq);
 975
 976                 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
 977                 rq = crq->request;
 978                 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
 979                         cfq_mark_cfqq_fifo_expire(cfqq);
 980                         return crq;
 981                 }
 982         }
 983
 984         return NULL;
 985 }
 986
 987 /*
 988  * Scale schedule slice based on io priority. Use the sync time slice only
 989  * if a queue is marked sync and has sync io queued. A sync queue with async
 990  * io only, should not get full sync slice length.
 991  */
 992 static inline int
 993 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 994 {
 995         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 996
 997         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 998
 999         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
1000 }
1001
1002 static inline void
1003 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1004 {
1005         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
1006 }
1007
1008 static inline int
1009 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1010 {
1011         const int base_rq = cfqd->cfq_slice_async_rq;
1012
1013         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1014
1015         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1016 }
1017
1018 /*
1019  * get next queue for service
1020  */
1021 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1022 {
1023         unsigned long now = jiffies;
1024         struct cfq_queue *cfqq;
1025
1026         cfqq = cfqd->active_queue;
1027         if (!cfqq)
1028                 goto new_queue;
1029
1030         /*
1031          * slice has expired
1032          */
1033         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
1034                 goto expire;
1035
1036         /*
1037          * if queue has requests, dispatch one. if not, check if
1038          * enough slice is left to wait for one
1039          */
1040         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1041                 goto keep_queue;
1042         else if (cfq_cfqq_dispatched(cfqq)) {
1043                 cfqq = NULL;
1044                 goto keep_queue;
1045         } else if (cfq_cfqq_class_sync(cfqq)) {
1046                 if (cfq_arm_slice_timer(cfqd, cfqq))
1047                         return NULL;
1048         }
1049
1050 expire:
1051         cfq_slice_expired(cfqd, 0);
1052 new_queue:
1053         cfqq = cfq_set_active_queue(cfqd);
1054 keep_queue:
1055         return cfqq;
1056 }
1057
1058 static int
1059 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1060                         int max_dispatch)
1061 {
1062         int dispatched = 0;
1063
1064         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1065
1066         do {
1067                 struct cfq_rq *crq;
1068
1069                 /*
1070                  * follow expired path, else get first next available
1071                  */
1072                 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1073                         crq = cfqq->next_crq;
1074
1075                 /*
1076                  * finally, insert request into driver dispatch list
1077                  */
1078                 cfq_dispatch_insert(cfqd->queue, crq);
1079
1080                 cfqd->dispatch_slice++;
1081                 dispatched++;
1082
1083                 if (!cfqd->active_cic) {
1084                         atomic_inc(&crq->io_context->ioc->refcount);
1085                         cfqd->active_cic = crq->io_context;
1086                 }
1087
1088                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1089                         break;
1090
1091         } while (dispatched < max_dispatch);
1092
1093         /*
1094          * if slice end isn't set yet, set it.
1095          */
1096         if (!cfqq->slice_end)
1097                 cfq_set_prio_slice(cfqd, cfqq);
1098
1099         /*
1100          * expire an async queue immediately if it has used up its slice. idle
1101          * queue always expire after 1 dispatch round.
1102          */
1103         if ((!cfq_cfqq_sync(cfqq) &&
1104             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1105             cfq_class_idle(cfqq) ||
1106             !cfq_cfqq_idle_window(cfqq))
1107                 cfq_slice_expired(cfqd, 0);
1108
1109         return dispatched;
1110 }
1111
1112 static int
1113 cfq_forced_dispatch_cfqqs(struct list_head *list)
1114 {
1115         struct cfq_queue *cfqq, *next;
1116         struct cfq_rq *crq;
1117         int dispatched;
1118
1119         dispatched = 0;
1120         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1121                 while ((crq = cfqq->next_crq)) {
1122                         cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1123                         dispatched++;
1124                 }
1125                 BUG_ON(!list_empty(&cfqq->fifo));
1126         }
1127
1128         return dispatched;
1129 }
1130
1131 static int
1132 cfq_forced_dispatch(struct cfq_data *cfqd)
1133 {
1134         int i, dispatched = 0;
1135
1136         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1137                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1138
1139         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1140         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1141         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1142
1143         cfq_slice_expired(cfqd, 0);
1144
1145         BUG_ON(cfqd->busy_queues);
1146
1147         return dispatched;
1148 }
1149
1150 static int
1151 cfq_dispatch_requests(request_queue_t *q, int force)
1152 {
1153         struct cfq_data *cfqd = q->elevator->elevator_data;
1154         struct cfq_queue *cfqq, *prev_cfqq;
1155         int dispatched;
1156
1157         if (!cfqd->busy_queues)
1158                 return 0;
1159
1160         if (unlikely(force))
1161                 return cfq_forced_dispatch(cfqd);
1162
1163         dispatched = 0;
1164         prev_cfqq = NULL;
1165         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1166                 int max_dispatch;
1167
1168                 /*
1169                  * Don't repeat dispatch from the previous queue.
1170                  */
1171                 if (prev_cfqq == cfqq)
1172                         break;
1173
1174                 cfq_clear_cfqq_must_dispatch(cfqq);
1175                 cfq_clear_cfqq_wait_request(cfqq);
1176                 del_timer(&cfqd->idle_slice_timer);
1177
1178                 max_dispatch = cfqd->cfq_quantum;
1179                 if (cfq_class_idle(cfqq))
1180                         max_dispatch = 1;
1181
1182                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1183
1184                 /*
1185                  * If the dispatch cfqq has idling enabled and is still
1186                  * the active queue, break out.
1187                  */
1188                 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1189                         break;
1190
1191                 prev_cfqq = cfqq;
1192         }
1193
1194         return dispatched;
1195 }
1196
1197 /*
1198  * task holds one reference to the queue, dropped when task exits. each crq
1199  * in-flight on this queue also holds a reference, dropped when crq is freed.
1200  *
1201  * queue lock must be held here.
1202  */
1203 static void cfq_put_queue(struct cfq_queue *cfqq)
1204 {
1205         struct cfq_data *cfqd = cfqq->cfqd;
1206
1207         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1208
1209         if (!atomic_dec_and_test(&cfqq->ref))
1210                 return;
1211
1212         BUG_ON(rb_first(&cfqq->sort_list));
1213         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1214         BUG_ON(cfq_cfqq_on_rr(cfqq));
1215
1216         if (unlikely(cfqd->active_queue == cfqq))
1217                 __cfq_slice_expired(cfqd, cfqq, 0);
1218
1219         /*
1220          * it's on the empty list and still hashed
1221          */
1222         list_del(&cfqq->cfq_list);
1223         hlist_del(&cfqq->cfq_hash);
1224         kmem_cache_free(cfq_pool, cfqq);
1225 }
1226
1227 static inline struct cfq_queue *
1228 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1229                     const int hashval)
1230 {
1231         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1232         struct hlist_node *entry;
1233         struct cfq_queue *__cfqq;
1234
1235         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1236                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1237
1238                 if (__cfqq->key == key && (__p == prio || !prio))
1239                         return __cfqq;
1240         }
1241
1242         return NULL;
1243 }
1244
1245 static struct cfq_queue *
1246 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1247 {
1248         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1249 }
1250
1251 static void cfq_free_io_context(struct io_context *ioc)
1252 {
1253         struct cfq_io_context *__cic;
1254         struct rb_node *n;
1255         int freed = 0;
1256
1257         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1258                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1259                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1260                 kmem_cache_free(cfq_ioc_pool, __cic);
1261                 freed++;
1262         }
1263
1264         if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1265                 complete(ioc_gone);
1266 }
1267
1268 static void cfq_trim(struct io_context *ioc)
1269 {
1270         ioc->set_ioprio = NULL;
1271         cfq_free_io_context(ioc);
1272 }
1273
1274 /*
1275  * Called with interrupts disabled
1276  */
1277 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1278 {
1279         struct cfq_data *cfqd = cic->key;
1280         request_queue_t *q;
1281
1282         if (!cfqd)
1283                 return;
1284
1285         q = cfqd->queue;
1286
1287         WARN_ON(!irqs_disabled());
1288
1289         spin_lock(q->queue_lock);
1290
1291         if (cic->cfqq[ASYNC]) {
1292                 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
1293                         __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
1294                 cfq_put_queue(cic->cfqq[ASYNC]);
1295                 cic->cfqq[ASYNC] = NULL;
1296         }
1297
1298         if (cic->cfqq[SYNC]) {
1299                 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
1300                         __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
1301                 cfq_put_queue(cic->cfqq[SYNC]);
1302                 cic->cfqq[SYNC] = NULL;
1303         }
1304
1305         cic->key = NULL;
1306         list_del_init(&cic->queue_list);
1307         spin_unlock(q->queue_lock);
1308 }
1309
1310 static void cfq_exit_io_context(struct io_context *ioc)
1311 {
1312         struct cfq_io_context *__cic;
1313         unsigned long flags;
1314         struct rb_node *n;
1315
1316         /*
1317          * put the reference this task is holding to the various queues
1318          */
1319         spin_lock_irqsave(&cfq_exit_lock, flags);
1320
1321         n = rb_first(&ioc->cic_root);
1322         while (n != NULL) {
1323                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1324
1325                 cfq_exit_single_io_context(__cic);
1326                 n = rb_next(n);
1327         }
1328
1329         spin_unlock_irqrestore(&cfq_exit_lock, flags);
1330 }
1331
1332 static struct cfq_io_context *
1333 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1334 {
1335         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1336
1337         if (cic) {
1338                 memset(cic, 0, sizeof(*cic));
1339                 cic->last_end_request = jiffies;
1340                 INIT_LIST_HEAD(&cic->queue_list);
1341                 cic->dtor = cfq_free_io_context;
1342                 cic->exit = cfq_exit_io_context;
1343                 atomic_inc(&ioc_count);
1344         }
1345
1346         return cic;
1347 }
1348
1349 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1350 {
1351         struct task_struct *tsk = current;
1352         int ioprio_class;
1353
1354         if (!cfq_cfqq_prio_changed(cfqq))
1355                 return;
1356
1357         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1358         switch (ioprio_class) {
1359                 default:
1360                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1361                 case IOPRIO_CLASS_NONE:
1362                         /*
1363                          * no prio set, place us in the middle of the BE classes
1364                          */
1365                         cfqq->ioprio = task_nice_ioprio(tsk);
1366                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1367                         break;
1368                 case IOPRIO_CLASS_RT:
1369                         cfqq->ioprio = task_ioprio(tsk);
1370                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1371                         break;
1372                 case IOPRIO_CLASS_BE:
1373                         cfqq->ioprio = task_ioprio(tsk);
1374                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1375                         break;
1376                 case IOPRIO_CLASS_IDLE:
1377                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1378                         cfqq->ioprio = 7;
1379                         cfq_clear_cfqq_idle_window(cfqq);
1380                         break;
1381         }
1382
1383         /*
1384          * keep track of original prio settings in case we have to temporarily
1385          * elevate the priority of this queue
1386          */
1387         cfqq->org_ioprio = cfqq->ioprio;
1388         cfqq->org_ioprio_class = cfqq->ioprio_class;
1389
1390         if (cfq_cfqq_on_rr(cfqq))
1391                 cfq_resort_rr_list(cfqq, 0);
1392
1393         cfq_clear_cfqq_prio_changed(cfqq);
1394 }
1395
1396 static inline void changed_ioprio(struct cfq_io_context *cic)
1397 {
1398         struct cfq_data *cfqd = cic->key;
1399         struct cfq_queue *cfqq;
1400
1401         if (unlikely(!cfqd))
1402                 return;
1403
1404         spin_lock(cfqd->queue->queue_lock);
1405
1406         cfqq = cic->cfqq[ASYNC];
1407         if (cfqq) {
1408                 struct cfq_queue *new_cfqq;
1409                 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1410                                          GFP_ATOMIC);
1411                 if (new_cfqq) {
1412                         cic->cfqq[ASYNC] = new_cfqq;
1413                         cfq_put_queue(cfqq);
1414                 }
1415         }
1416
1417         cfqq = cic->cfqq[SYNC];
1418         if (cfqq)
1419                 cfq_mark_cfqq_prio_changed(cfqq);
1420
1421         spin_unlock(cfqd->queue->queue_lock);
1422 }
1423
1424 /*
1425  * callback from sys_ioprio_set, irqs are disabled
1426  */
1427 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1428 {
1429         struct cfq_io_context *cic;
1430         struct rb_node *n;
1431
1432         spin_lock(&cfq_exit_lock);
1433
1434         n = rb_first(&ioc->cic_root);
1435         while (n != NULL) {
1436                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1437
1438                 changed_ioprio(cic);
1439                 n = rb_next(n);
1440         }
1441
1442         spin_unlock(&cfq_exit_lock);
1443
1444         return 0;
1445 }
1446
1447 static struct cfq_queue *
1448 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1449               gfp_t gfp_mask)
1450 {
1451         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1452         struct cfq_queue *cfqq, *new_cfqq = NULL;
1453         unsigned short ioprio;
1454
1455 retry:
1456         ioprio = tsk->ioprio;
1457         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1458
1459         if (!cfqq) {
1460                 if (new_cfqq) {
1461                         cfqq = new_cfqq;
1462                         new_cfqq = NULL;
1463                 } else if (gfp_mask & __GFP_WAIT) {
1464                         spin_unlock_irq(cfqd->queue->queue_lock);
1465                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1466                         spin_lock_irq(cfqd->queue->queue_lock);
1467                         goto retry;
1468                 } else {
1469                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1470                         if (!cfqq)
1471                                 goto out;
1472                 }
1473
1474                 memset(cfqq, 0, sizeof(*cfqq));
1475
1476                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1477                 INIT_LIST_HEAD(&cfqq->cfq_list);
1478                 INIT_LIST_HEAD(&cfqq->fifo);
1479
1480                 cfqq->key = key;
1481                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1482                 atomic_set(&cfqq->ref, 0);
1483                 cfqq->cfqd = cfqd;
1484                 cfqq->service_last = 0;
1485                 /*
1486                  * set ->slice_left to allow preemption for a new process
1487                  */
1488                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1489                 cfq_mark_cfqq_idle_window(cfqq);
1490                 cfq_mark_cfqq_prio_changed(cfqq);
1491                 cfq_init_prio_data(cfqq);
1492         }
1493
1494         if (new_cfqq)
1495                 kmem_cache_free(cfq_pool, new_cfqq);
1496
1497         atomic_inc(&cfqq->ref);
1498 out:
1499         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1500         return cfqq;
1501 }
1502
1503 static void
1504 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1505 {
1506         spin_lock(&cfq_exit_lock);
1507         rb_erase(&cic->rb_node, &ioc->cic_root);
1508         list_del_init(&cic->queue_list);
1509         spin_unlock(&cfq_exit_lock);
1510         kmem_cache_free(cfq_ioc_pool, cic);
1511         atomic_dec(&ioc_count);
1512 }
1513
1514 static struct cfq_io_context *
1515 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1516 {
1517         struct rb_node *n;
1518         struct cfq_io_context *cic;
1519         void *k, *key = cfqd;
1520
1521 restart:
1522         n = ioc->cic_root.rb_node;
1523         while (n) {
1524                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1525                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1526                 k = cic->key;
1527                 if (unlikely(!k)) {
1528                         cfq_drop_dead_cic(ioc, cic);
1529                         goto restart;
1530                 }
1531
1532                 if (key < k)
1533                         n = n->rb_left;
1534                 else if (key > k)
1535                         n = n->rb_right;
1536                 else
1537                         return cic;
1538         }
1539
1540         return NULL;
1541 }
1542
1543 static inline void
1544 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1545              struct cfq_io_context *cic)
1546 {
1547         struct rb_node **p;
1548         struct rb_node *parent;
1549         struct cfq_io_context *__cic;
1550         void *k;
1551
1552         cic->ioc = ioc;
1553         cic->key = cfqd;
1554
1555         ioc->set_ioprio = cfq_ioc_set_ioprio;
1556 restart:
1557         parent = NULL;
1558         p = &ioc->cic_root.rb_node;
1559         while (*p) {
1560                 parent = *p;
1561                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1562                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1563                 k = __cic->key;
1564                 if (unlikely(!k)) {
1565                         cfq_drop_dead_cic(ioc, cic);
1566                         goto restart;
1567                 }
1568
1569                 if (cic->key < k)
1570                         p = &(*p)->rb_left;
1571                 else if (cic->key > k)
1572                         p = &(*p)->rb_right;
1573                 else
1574                         BUG();
1575         }
1576
1577         spin_lock(&cfq_exit_lock);
1578         rb_link_node(&cic->rb_node, parent, p);
1579         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1580         list_add(&cic->queue_list, &cfqd->cic_list);
1581         spin_unlock(&cfq_exit_lock);
1582 }
1583
1584 /*
1585  * Setup general io context and cfq io context. There can be several cfq
1586  * io contexts per general io context, if this process is doing io to more
1587  * than one device managed by cfq.
1588  */
1589 static struct cfq_io_context *
1590 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1591 {
1592         struct io_context *ioc = NULL;
1593         struct cfq_io_context *cic;
1594
1595         might_sleep_if(gfp_mask & __GFP_WAIT);
1596
1597         ioc = get_io_context(gfp_mask);
1598         if (!ioc)
1599                 return NULL;
1600
1601         cic = cfq_cic_rb_lookup(cfqd, ioc);
1602         if (cic)
1603                 goto out;
1604
1605         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1606         if (cic == NULL)
1607                 goto err;
1608
1609         cfq_cic_link(cfqd, ioc, cic);
1610 out:
1611         return cic;
1612 err:
1613         put_io_context(ioc);
1614         return NULL;
1615 }
1616
1617 static void
1618 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1619 {
1620         unsigned long elapsed, ttime;
1621
1622         /*
1623          * if this context already has stuff queued, thinktime is from
1624          * last queue not last end
1625          */
1626 #if 0
1627         if (time_after(cic->last_end_request, cic->last_queue))
1628                 elapsed = jiffies - cic->last_end_request;
1629         else
1630                 elapsed = jiffies - cic->last_queue;
1631 #else
1632                 elapsed = jiffies - cic->last_end_request;
1633 #endif
1634
1635         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1636
1637         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1638         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1639         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1640 }
1641
1642 static void
1643 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1644                        struct cfq_rq *crq)
1645 {
1646         sector_t sdist;
1647         u64 total;
1648
1649         if (cic->last_request_pos < crq->request->sector)
1650                 sdist = crq->request->sector - cic->last_request_pos;
1651         else
1652                 sdist = cic->last_request_pos - crq->request->sector;
1653
1654         /*
1655          * Don't allow the seek distance to get too large from the
1656          * odd fragment, pagein, etc
1657          */
1658         if (cic->seek_samples <= 60) /* second&third seek */
1659                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1660         else
1661                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1662
1663         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1664         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1665         total = cic->seek_total + (cic->seek_samples/2);
1666         do_div(total, cic->seek_samples);
1667         cic->seek_mean = (sector_t)total;
1668 }
1669
1670 /*
1671  * Disable idle window if the process thinks too long or seeks so much that
1672  * it doesn't matter
1673  */
1674 static void
1675 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1676                        struct cfq_io_context *cic)
1677 {
1678         int enable_idle = cfq_cfqq_idle_window(cfqq);
1679
1680         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1681             (cfqd->hw_tag && CIC_SEEKY(cic)))
1682                 enable_idle = 0;
1683         else if (sample_valid(cic->ttime_samples)) {
1684                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1685                         enable_idle = 0;
1686                 else
1687                         enable_idle = 1;
1688         }
1689
1690         if (enable_idle)
1691                 cfq_mark_cfqq_idle_window(cfqq);
1692         else
1693                 cfq_clear_cfqq_idle_window(cfqq);
1694 }
1695
1696
1697 /*
1698  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1699  * no or if we aren't sure, a 1 will cause a preempt.
1700  */
1701 static int
1702 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1703                    struct cfq_rq *crq)
1704 {
1705         struct cfq_queue *cfqq = cfqd->active_queue;
1706
1707         if (cfq_class_idle(new_cfqq))
1708                 return 0;
1709
1710         if (!cfqq)
1711                 return 0;
1712
1713         if (cfq_class_idle(cfqq))
1714                 return 1;
1715         if (!cfq_cfqq_wait_request(new_cfqq))
1716                 return 0;
1717         /*
1718          * if it doesn't have slice left, forget it
1719          */
1720         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1721                 return 0;
1722         if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1723                 return 1;
1724
1725         return 0;
1726 }
1727
1728 /*
1729  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1730  * let it have half of its nominal slice.
1731  */
1732 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1733 {
1734         struct cfq_queue *__cfqq, *next;
1735
1736         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1737                 cfq_resort_rr_list(__cfqq, 1);
1738
1739         if (!cfqq->slice_left)
1740                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1741
1742         cfqq->slice_end = cfqq->slice_left + jiffies;
1743         cfq_slice_expired(cfqd, 1);
1744         __cfq_set_active_queue(cfqd, cfqq);
1745 }
1746
1747 /*
1748  * should really be a ll_rw_blk.c helper
1749  */
1750 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1751 {
1752         request_queue_t *q = cfqd->queue;
1753
1754         if (!blk_queue_plugged(q))
1755                 q->request_fn(q);
1756         else
1757                 __generic_unplug_device(q);
1758 }
1759
1760 /*
1761  * Called when a new fs request (crq) is added (to cfqq). Check if there's
1762  * something we should do about it
1763  */
1764 static void
1765 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1766                  struct cfq_rq *crq)
1767 {
1768         struct cfq_io_context *cic = crq->io_context;
1769
1770         /*
1771          * we never wait for an async request and we don't allow preemption
1772          * of an async request. so just return early
1773          */
1774         if (!cfq_crq_is_sync(crq)) {
1775                 /*
1776                  * sync process issued an async request, if it's waiting
1777                  * then expire it and kick rq handling.
1778                  */
1779                 if (cic == cfqd->active_cic &&
1780                     del_timer(&cfqd->idle_slice_timer)) {
1781                         cfq_slice_expired(cfqd, 0);
1782                         cfq_start_queueing(cfqd, cfqq);
1783                 }
1784                 return;
1785         }
1786
1787         cfq_update_io_thinktime(cfqd, cic);
1788         cfq_update_io_seektime(cfqd, cic, crq);
1789         cfq_update_idle_window(cfqd, cfqq, cic);
1790
1791         cic->last_queue = jiffies;
1792         cic->last_request_pos = crq->request->sector + crq->request->nr_sectors;
1793
1794         if (cfqq == cfqd->active_queue) {
1795                 /*
1796                  * if we are waiting for a request for this queue, let it rip
1797                  * immediately and flag that we must not expire this queue
1798                  * just now
1799                  */
1800                 if (cfq_cfqq_wait_request(cfqq)) {
1801                         cfq_mark_cfqq_must_dispatch(cfqq);
1802                         del_timer(&cfqd->idle_slice_timer);
1803                         cfq_start_queueing(cfqd, cfqq);
1804                 }
1805         } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1806                 /*
1807                  * not the active queue - expire current slice if it is
1808                  * idle and has expired it's mean thinktime or this new queue
1809                  * has some old slice time left and is of higher priority
1810                  */
1811                 cfq_preempt_queue(cfqd, cfqq);
1812                 cfq_mark_cfqq_must_dispatch(cfqq);
1813                 cfq_start_queueing(cfqd, cfqq);
1814         }
1815 }
1816
1817 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1818 {
1819         struct cfq_data *cfqd = q->elevator->elevator_data;
1820         struct cfq_rq *crq = RQ_DATA(rq);
1821         struct cfq_queue *cfqq = crq->cfq_queue;
1822
1823         cfq_init_prio_data(cfqq);
1824
1825         cfq_add_crq_rb(crq);
1826
1827         list_add_tail(&rq->queuelist, &cfqq->fifo);
1828
1829         if (rq_mergeable(rq))
1830                 cfq_add_crq_hash(cfqd, crq);
1831
1832         cfq_crq_enqueued(cfqd, cfqq, crq);
1833 }
1834
1835 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1836 {
1837         struct cfq_rq *crq = RQ_DATA(rq);
1838         struct cfq_queue *cfqq = crq->cfq_queue;
1839         struct cfq_data *cfqd = cfqq->cfqd;
1840         const int sync = cfq_crq_is_sync(crq);
1841         unsigned long now;
1842
1843         now = jiffies;
1844
1845         WARN_ON(!cfqd->rq_in_driver);
1846         WARN_ON(!cfqq->on_dispatch[sync]);
1847         cfqd->rq_in_driver--;
1848         cfqq->on_dispatch[sync]--;
1849
1850         if (!cfq_class_idle(cfqq))
1851                 cfqd->last_end_request = now;
1852
1853         if (!cfq_cfqq_dispatched(cfqq)) {
1854                 if (cfq_cfqq_on_rr(cfqq)) {
1855                         cfqq->service_last = now;
1856                         cfq_resort_rr_list(cfqq, 0);
1857                 }
1858         }
1859
1860         if (sync)
1861                 crq->io_context->last_end_request = now;
1862
1863         /*
1864          * If this is the active queue, check if it needs to be expired,
1865          * or if we want to idle in case it has no pending requests.
1866          */
1867         if (cfqd->active_queue == cfqq) {
1868                 if (time_after(now, cfqq->slice_end))
1869                         cfq_slice_expired(cfqd, 0);
1870                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1871                         if (!cfq_arm_slice_timer(cfqd, cfqq))
1872                                 cfq_schedule_dispatch(cfqd);
1873                 }
1874         }
1875 }
1876
1877 static struct request *
1878 cfq_former_request(request_queue_t *q, struct request *rq)
1879 {
1880         struct cfq_rq *crq = RQ_DATA(rq);
1881         struct rb_node *rbprev = rb_prev(&crq->rb_node);
1882
1883         if (rbprev)
1884                 return rb_entry_crq(rbprev)->request;
1885
1886         return NULL;
1887 }
1888
1889 static struct request *
1890 cfq_latter_request(request_queue_t *q, struct request *rq)
1891 {
1892         struct cfq_rq *crq = RQ_DATA(rq);
1893         struct rb_node *rbnext = rb_next(&crq->rb_node);
1894
1895         if (rbnext)
1896                 return rb_entry_crq(rbnext)->request;
1897
1898         return NULL;
1899 }
1900
1901 /*
1902  * we temporarily boost lower priority queues if they are holding fs exclusive
1903  * resources. they are boosted to normal prio (CLASS_BE/4)
1904  */
1905 static void cfq_prio_boost(struct cfq_queue *cfqq)
1906 {
1907         const int ioprio_class = cfqq->ioprio_class;
1908         const int ioprio = cfqq->ioprio;
1909
1910         if (has_fs_excl()) {
1911                 /*
1912                  * boost idle prio on transactions that would lock out other
1913                  * users of the filesystem
1914                  */
1915                 if (cfq_class_idle(cfqq))
1916                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1917                 if (cfqq->ioprio > IOPRIO_NORM)
1918                         cfqq->ioprio = IOPRIO_NORM;
1919         } else {
1920                 /*
1921                  * check if we need to unboost the queue
1922                  */
1923                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1924                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1925                 if (cfqq->ioprio != cfqq->org_ioprio)
1926                         cfqq->ioprio = cfqq->org_ioprio;
1927         }
1928
1929         /*
1930          * refile between round-robin lists if we moved the priority class
1931          */
1932         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1933             cfq_cfqq_on_rr(cfqq))
1934                 cfq_resort_rr_list(cfqq, 0);
1935 }
1936
1937 static inline int
1938 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1939                 struct task_struct *task, int rw)
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, struct bio *bio)
1951 {
1952         struct cfq_data *cfqd = q->elevator->elevator_data;
1953         struct task_struct *tsk = current;
1954         struct cfq_queue *cfqq;
1955
1956         /*
1957          * don't force setup of a queue from here, as a call to may_queue
1958          * does not necessarily imply that a request actually will be queued.
1959          * so just lookup a possibly existing queue, or return 'may queue'
1960          * if that fails
1961          */
1962         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1963         if (cfqq) {
1964                 cfq_init_prio_data(cfqq);
1965                 cfq_prio_boost(cfqq);
1966
1967                 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1968         }
1969
1970         return ELV_MQUEUE_MAY;
1971 }
1972
1973 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1974 {
1975         struct cfq_data *cfqd = q->elevator->elevator_data;
1976
1977         if (unlikely(cfqd->rq_starved)) {
1978                 struct request_list *rl = &q->rq;
1979
1980                 smp_mb();
1981                 if (waitqueue_active(&rl->wait[READ]))
1982                         wake_up(&rl->wait[READ]);
1983                 if (waitqueue_active(&rl->wait[WRITE]))
1984                         wake_up(&rl->wait[WRITE]);
1985         }
1986 }
1987
1988 /*
1989  * queue lock held here
1990  */
1991 static void cfq_put_request(request_queue_t *q, struct request *rq)
1992 {
1993         struct cfq_data *cfqd = q->elevator->elevator_data;
1994         struct cfq_rq *crq = RQ_DATA(rq);
1995
1996         if (crq) {
1997                 struct cfq_queue *cfqq = crq->cfq_queue;
1998                 const int rw = rq_data_dir(rq);
1999
2000                 BUG_ON(!cfqq->allocated[rw]);
2001                 cfqq->allocated[rw]--;
2002
2003                 put_io_context(crq->io_context->ioc);
2004
2005                 mempool_free(crq, cfqd->crq_pool);
2006                 rq->elevator_private = NULL;
2007
2008                 cfq_check_waiters(q, cfqq);
2009                 cfq_put_queue(cfqq);
2010         }
2011 }
2012
2013 /*
2014  * Allocate cfq data structures associated with this request.
2015  */
2016 static int
2017 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
2018                 gfp_t gfp_mask)
2019 {
2020         struct cfq_data *cfqd = q->elevator->elevator_data;
2021         struct task_struct *tsk = current;
2022         struct cfq_io_context *cic;
2023         const int rw = rq_data_dir(rq);
2024         pid_t key = cfq_queue_pid(tsk, rw);
2025         struct cfq_queue *cfqq;
2026         struct cfq_rq *crq;
2027         unsigned long flags;
2028         int is_sync = key != CFQ_KEY_ASYNC;
2029
2030         might_sleep_if(gfp_mask & __GFP_WAIT);
2031
2032         cic = cfq_get_io_context(cfqd, gfp_mask);
2033
2034         spin_lock_irqsave(q->queue_lock, flags);
2035
2036         if (!cic)
2037                 goto queue_fail;
2038
2039         if (!cic->cfqq[is_sync]) {
2040                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
2041                 if (!cfqq)
2042                         goto queue_fail;
2043
2044                 cic->cfqq[is_sync] = cfqq;
2045         } else
2046                 cfqq = cic->cfqq[is_sync];
2047
2048         cfqq->allocated[rw]++;
2049         cfq_clear_cfqq_must_alloc(cfqq);
2050         cfqd->rq_starved = 0;
2051         atomic_inc(&cfqq->ref);
2052         spin_unlock_irqrestore(q->queue_lock, flags);
2053
2054         crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
2055         if (crq) {
2056                 RB_CLEAR_NODE(&crq->rb_node);
2057                 crq->rb_key = 0;
2058                 crq->request = rq;
2059                 INIT_HLIST_NODE(&crq->hash);
2060                 crq->cfq_queue = cfqq;
2061                 crq->io_context = cic;
2062
2063                 if (is_sync)
2064                         cfq_mark_crq_is_sync(crq);
2065                 else
2066                         cfq_clear_crq_is_sync(crq);
2067
2068                 rq->elevator_private = crq;
2069                 return 0;
2070         }
2071
2072         spin_lock_irqsave(q->queue_lock, flags);
2073         cfqq->allocated[rw]--;
2074         if (!(cfqq->allocated[0] + cfqq->allocated[1]))
2075                 cfq_mark_cfqq_must_alloc(cfqq);
2076         cfq_put_queue(cfqq);
2077 queue_fail:
2078         if (cic)
2079                 put_io_context(cic->ioc);
2080         /*
2081          * mark us rq allocation starved. we need to kickstart the process
2082          * ourselves if there are no pending requests that can do it for us.
2083          * that would be an extremely rare OOM situation
2084          */
2085         cfqd->rq_starved = 1;
2086         cfq_schedule_dispatch(cfqd);
2087         spin_unlock_irqrestore(q->queue_lock, flags);
2088         return 1;
2089 }
2090
2091 static void cfq_kick_queue(void *data)
2092 {
2093         request_queue_t *q = data;
2094         struct cfq_data *cfqd = q->elevator->elevator_data;
2095         unsigned long flags;
2096
2097         spin_lock_irqsave(q->queue_lock, flags);
2098
2099         if (cfqd->rq_starved) {
2100                 struct request_list *rl = &q->rq;
2101
2102                 /*
2103                  * we aren't guaranteed to get a request after this, but we
2104                  * have to be opportunistic
2105                  */
2106                 smp_mb();
2107                 if (waitqueue_active(&rl->wait[READ]))
2108                         wake_up(&rl->wait[READ]);
2109                 if (waitqueue_active(&rl->wait[WRITE]))
2110                         wake_up(&rl->wait[WRITE]);
2111         }
2112
2113         blk_remove_plug(q);
2114         q->request_fn(q);
2115         spin_unlock_irqrestore(q->queue_lock, flags);
2116 }
2117
2118 /*
2119  * Timer running if the active_queue is currently idling inside its time slice
2120  */
2121 static void cfq_idle_slice_timer(unsigned long data)
2122 {
2123         struct cfq_data *cfqd = (struct cfq_data *) data;
2124         struct cfq_queue *cfqq;
2125         unsigned long flags;
2126
2127         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2128
2129         if ((cfqq = cfqd->active_queue) != NULL) {
2130                 unsigned long now = jiffies;
2131
2132                 /*
2133                  * expired
2134                  */
2135                 if (time_after(now, cfqq->slice_end))
2136                         goto expire;
2137
2138                 /*
2139                  * only expire and reinvoke request handler, if there are
2140                  * other queues with pending requests
2141                  */
2142                 if (!cfqd->busy_queues)
2143                         goto out_cont;
2144
2145                 /*
2146                  * not expired and it has a request pending, let it dispatch
2147                  */
2148                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
2149                         cfq_mark_cfqq_must_dispatch(cfqq);
2150                         goto out_kick;
2151                 }
2152         }
2153 expire:
2154         cfq_slice_expired(cfqd, 0);
2155 out_kick:
2156         cfq_schedule_dispatch(cfqd);
2157 out_cont:
2158         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2159 }
2160
2161 /*
2162  * Timer running if an idle class queue is waiting for service
2163  */
2164 static void cfq_idle_class_timer(unsigned long data)
2165 {
2166         struct cfq_data *cfqd = (struct cfq_data *) data;
2167         unsigned long flags, end;
2168
2169         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2170
2171         /*
2172          * race with a non-idle queue, reset timer
2173          */
2174         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2175         if (!time_after_eq(jiffies, end))
2176                 mod_timer(&cfqd->idle_class_timer, end);
2177         else
2178                 cfq_schedule_dispatch(cfqd);
2179
2180         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2181 }
2182
2183 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2184 {
2185         del_timer_sync(&cfqd->idle_slice_timer);
2186         del_timer_sync(&cfqd->idle_class_timer);
2187         blk_sync_queue(cfqd->queue);
2188 }
2189
2190 static void cfq_exit_queue(elevator_t *e)
2191 {
2192         struct cfq_data *cfqd = e->elevator_data;
2193         request_queue_t *q = cfqd->queue;
2194
2195         cfq_shutdown_timer_wq(cfqd);
2196
2197         spin_lock(&cfq_exit_lock);
2198         spin_lock_irq(q->queue_lock);
2199
2200         if (cfqd->active_queue)
2201                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2202
2203         while (!list_empty(&cfqd->cic_list)) {
2204                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2205                                                         struct cfq_io_context,
2206                                                         queue_list);
2207                 if (cic->cfqq[ASYNC]) {
2208                         cfq_put_queue(cic->cfqq[ASYNC]);
2209                         cic->cfqq[ASYNC] = NULL;
2210                 }
2211                 if (cic->cfqq[SYNC]) {
2212                         cfq_put_queue(cic->cfqq[SYNC]);
2213                         cic->cfqq[SYNC] = NULL;
2214                 }
2215                 cic->key = NULL;
2216                 list_del_init(&cic->queue_list);
2217         }
2218
2219         spin_unlock_irq(q->queue_lock);
2220         spin_unlock(&cfq_exit_lock);
2221
2222         cfq_shutdown_timer_wq(cfqd);
2223
2224         mempool_destroy(cfqd->crq_pool);
2225         kfree(cfqd->crq_hash);
2226         kfree(cfqd->cfq_hash);
2227         kfree(cfqd);
2228 }
2229
2230 static void *cfq_init_queue(request_queue_t *q, elevator_t *e)
2231 {
2232         struct cfq_data *cfqd;
2233         int i;
2234
2235         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2236         if (!cfqd)
2237                 return NULL;
2238
2239         memset(cfqd, 0, sizeof(*cfqd));
2240
2241         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2242                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2243
2244         INIT_LIST_HEAD(&cfqd->busy_rr);
2245         INIT_LIST_HEAD(&cfqd->cur_rr);
2246         INIT_LIST_HEAD(&cfqd->idle_rr);
2247         INIT_LIST_HEAD(&cfqd->empty_list);
2248         INIT_LIST_HEAD(&cfqd->cic_list);
2249
2250         cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2251         if (!cfqd->crq_hash)
2252                 goto out_crqhash;
2253
2254         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2255         if (!cfqd->cfq_hash)
2256                 goto out_cfqhash;
2257
2258         cfqd->crq_pool = mempool_create_slab_pool(BLKDEV_MIN_RQ, crq_pool);
2259         if (!cfqd->crq_pool)
2260                 goto out_crqpool;
2261
2262         for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2263                 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2264         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2265                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2266
2267         cfqd->queue = q;
2268
2269         init_timer(&cfqd->idle_slice_timer);
2270         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2271         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2272
2273         init_timer(&cfqd->idle_class_timer);
2274         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2275         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2276
2277         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2278
2279         cfqd->cfq_queued = cfq_queued;
2280         cfqd->cfq_quantum = cfq_quantum;
2281         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2282         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2283         cfqd->cfq_back_max = cfq_back_max;
2284         cfqd->cfq_back_penalty = cfq_back_penalty;
2285         cfqd->cfq_slice[0] = cfq_slice_async;
2286         cfqd->cfq_slice[1] = cfq_slice_sync;
2287         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2288         cfqd->cfq_slice_idle = cfq_slice_idle;
2289
2290         return cfqd;
2291 out_crqpool:
2292         kfree(cfqd->cfq_hash);
2293 out_cfqhash:
2294         kfree(cfqd->crq_hash);
2295 out_crqhash:
2296         kfree(cfqd);
2297         return NULL;
2298 }
2299
2300 static void cfq_slab_kill(void)
2301 {
2302         if (crq_pool)
2303                 kmem_cache_destroy(crq_pool);
2304         if (cfq_pool)
2305                 kmem_cache_destroy(cfq_pool);
2306         if (cfq_ioc_pool)
2307                 kmem_cache_destroy(cfq_ioc_pool);
2308 }
2309
2310 static int __init cfq_slab_setup(void)
2311 {
2312         crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2313                                         NULL, NULL);
2314         if (!crq_pool)
2315                 goto fail;
2316
2317         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2318                                         NULL, NULL);
2319         if (!cfq_pool)
2320                 goto fail;
2321
2322         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2323                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2324         if (!cfq_ioc_pool)
2325                 goto fail;
2326
2327         return 0;
2328 fail:
2329         cfq_slab_kill();
2330         return -ENOMEM;
2331 }
2332
2333 /*
2334  * sysfs parts below -->
2335  */
2336
2337 static ssize_t
2338 cfq_var_show(unsigned int var, char *page)
2339 {
2340         return sprintf(page, "%d\n", var);
2341 }
2342
2343 static ssize_t
2344 cfq_var_store(unsigned int *var, const char *page, size_t count)
2345 {
2346         char *p = (char *) page;
2347
2348         *var = simple_strtoul(p, &p, 10);
2349         return count;
2350 }
2351
2352 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2353 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2354 {                                                                       \
2355         struct cfq_data *cfqd = e->elevator_data;                       \
2356         unsigned int __data = __VAR;                                    \
2357         if (__CONV)                                                     \
2358                 __data = jiffies_to_msecs(__data);                      \
2359         return cfq_var_show(__data, (page));                            \
2360 }
2361 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2362 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2363 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2364 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2365 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2366 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2367 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2368 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2369 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2370 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2371 #undef SHOW_FUNCTION
2372
2373 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2374 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2375 {                                                                       \
2376         struct cfq_data *cfqd = e->elevator_data;                       \
2377         unsigned int __data;                                            \
2378         int ret = cfq_var_store(&__data, (page), count);                \
2379         if (__data < (MIN))                                             \
2380                 __data = (MIN);                                         \
2381         else if (__data > (MAX))                                        \
2382                 __data = (MAX);                                         \
2383         if (__CONV)                                                     \
2384                 *(__PTR) = msecs_to_jiffies(__data);                    \
2385         else                                                            \
2386                 *(__PTR) = __data;                                      \
2387         return ret;                                                     \
2388 }
2389 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2390 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2391 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2392 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2393 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2394 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2395 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2396 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2397 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2398 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2399 #undef STORE_FUNCTION
2400
2401 #define CFQ_ATTR(name) \
2402         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2403
2404 static struct elv_fs_entry cfq_attrs[] = {
2405         CFQ_ATTR(quantum),
2406         CFQ_ATTR(queued),
2407         CFQ_ATTR(fifo_expire_sync),
2408         CFQ_ATTR(fifo_expire_async),
2409         CFQ_ATTR(back_seek_max),
2410         CFQ_ATTR(back_seek_penalty),
2411         CFQ_ATTR(slice_sync),
2412         CFQ_ATTR(slice_async),
2413         CFQ_ATTR(slice_async_rq),
2414         CFQ_ATTR(slice_idle),
2415         __ATTR_NULL
2416 };
2417
2418 static struct elevator_type iosched_cfq = {
2419         .ops = {
2420                 .elevator_merge_fn =            cfq_merge,
2421                 .elevator_merged_fn =           cfq_merged_request,
2422                 .elevator_merge_req_fn =        cfq_merged_requests,
2423                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2424                 .elevator_add_req_fn =          cfq_insert_request,
2425                 .elevator_activate_req_fn =     cfq_activate_request,
2426                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2427                 .elevator_queue_empty_fn =      cfq_queue_empty,
2428                 .elevator_completed_req_fn =    cfq_completed_request,
2429                 .elevator_former_req_fn =       cfq_former_request,
2430                 .elevator_latter_req_fn =       cfq_latter_request,
2431                 .elevator_set_req_fn =          cfq_set_request,
2432                 .elevator_put_req_fn =          cfq_put_request,
2433                 .elevator_may_queue_fn =        cfq_may_queue,
2434                 .elevator_init_fn =             cfq_init_queue,
2435                 .elevator_exit_fn =             cfq_exit_queue,
2436                 .trim =                         cfq_trim,
2437         },
2438         .elevator_attrs =       cfq_attrs,
2439         .elevator_name =        "cfq",
2440         .elevator_owner =       THIS_MODULE,
2441 };
2442
2443 static int __init cfq_init(void)
2444 {
2445         int ret;
2446
2447         /*
2448          * could be 0 on HZ < 1000 setups
2449          */
2450         if (!cfq_slice_async)
2451                 cfq_slice_async = 1;
2452         if (!cfq_slice_idle)
2453                 cfq_slice_idle = 1;
2454
2455         if (cfq_slab_setup())
2456                 return -ENOMEM;
2457
2458         ret = elv_register(&iosched_cfq);
2459         if (ret)
2460                 cfq_slab_kill();
2461
2462         return ret;
2463 }
2464
2465 static void __exit cfq_exit(void)
2466 {
2467         DECLARE_COMPLETION(all_gone);
2468         elv_unregister(&iosched_cfq);
2469         ioc_gone = &all_gone;
2470         /* ioc_gone's update must be visible before reading ioc_count */
2471         smp_wmb();
2472         if (atomic_read(&ioc_count))
2473                 wait_for_completion(ioc_gone);
2474         synchronize_rcu();
2475         cfq_slab_kill();
2476 }
2477
2478 module_init(cfq_init);
2479 module_exit(cfq_exit);
2480
2481 MODULE_AUTHOR("Jens Axboe");
2482 MODULE_LICENSE("GPL");
2483 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");