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 / 70;
  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_NONE                 (2)
  64 #define RB_EMPTY(node)          ((node)->rb_node == NULL)
  65 #define RB_CLEAR_COLOR(node)    (node)->rb_color = RB_NONE
  66 #define RB_CLEAR(node)          do {    \
  67         (node)->rb_parent = NULL;       \
  68         RB_CLEAR_COLOR((node));         \
  69         (node)->rb_right = NULL;        \
  70         (node)->rb_left = NULL;         \
  71 } while (0)
  72 #define RB_CLEAR_ROOT(root)     ((root)->rb_node = NULL)
  73 #define rb_entry_crq(node)      rb_entry((node), struct cfq_rq, rb_node)
  74 #define rq_rb_key(rq)           (rq)->sector
  75
  76 static kmem_cache_t *crq_pool;
  77 static kmem_cache_t *cfq_pool;
  78 static kmem_cache_t *cfq_ioc_pool;
  79
  80 static atomic_t ioc_count = ATOMIC_INIT(0);
  81 static struct completion *ioc_gone;
  82
  83 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  84 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  85 #define cfq_class_be(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
  86 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  87
  88 #define ASYNC                   (0)
  89 #define SYNC                    (1)
  90
  91 #define cfq_cfqq_dispatched(cfqq)       \
  92         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
  93
  94 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
  95
  96 #define cfq_cfqq_sync(cfqq)             \
  97         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
  98
  99 #define sample_valid(samples)   ((samples) > 80)
 100
 101 /*
 102  * Per block device queue structure
 103  */
 104 struct cfq_data {
 105         request_queue_t *queue;
 106
 107         /*
 108          * rr list of queues with requests and the count of them
 109          */
 110         struct list_head rr_list[CFQ_PRIO_LISTS];
 111         struct list_head busy_rr;
 112         struct list_head cur_rr;
 113         struct list_head idle_rr;
 114         unsigned int busy_queues;
 115
 116         /*
 117          * non-ordered list of empty cfqq's
 118          */
 119         struct list_head empty_list;
 120
 121         /*
 122          * cfqq lookup hash
 123          */
 124         struct hlist_head *cfq_hash;
 125
 126         /*
 127          * global crq hash for all queues
 128          */
 129         struct hlist_head *crq_hash;
 130
 131         unsigned int max_queued;
 132
 133         mempool_t *crq_pool;
 134
 135         int rq_in_driver;
 136         int hw_tag;
 137
 138         /*
 139          * schedule slice state info
 140          */
 141         /*
 142          * idle window management
 143          */
 144         struct timer_list idle_slice_timer;
 145         struct work_struct unplug_work;
 146
 147         struct cfq_queue *active_queue;
 148         struct cfq_io_context *active_cic;
 149         int cur_prio, cur_end_prio;
 150         unsigned int dispatch_slice;
 151
 152         struct timer_list idle_class_timer;
 153
 154         sector_t last_sector;
 155         unsigned long last_end_request;
 156
 157         unsigned int rq_starved;
 158
 159         /*
 160          * tunables, see top of file
 161          */
 162         unsigned int cfq_quantum;
 163         unsigned int cfq_queued;
 164         unsigned int cfq_fifo_expire[2];
 165         unsigned int cfq_back_penalty;
 166         unsigned int cfq_back_max;
 167         unsigned int cfq_slice[2];
 168         unsigned int cfq_slice_async_rq;
 169         unsigned int cfq_slice_idle;
 170
 171         struct list_head cic_list;
 172 };
 173
 174 /*
 175  * Per process-grouping structure
 176  */
 177 struct cfq_queue {
 178         /* reference count */
 179         atomic_t ref;
 180         /* parent cfq_data */
 181         struct cfq_data *cfqd;
 182         /* cfqq lookup hash */
 183         struct hlist_node cfq_hash;
 184         /* hash key */
 185         unsigned int key;
 186         /* on either rr or empty list of cfqd */
 187         struct list_head cfq_list;
 188         /* sorted list of pending requests */
 189         struct rb_root sort_list;
 190         /* if fifo isn't expired, next request to serve */
 191         struct cfq_rq *next_crq;
 192         /* requests queued in sort_list */
 193         int queued[2];
 194         /* currently allocated requests */
 195         int allocated[2];
 196         /* fifo list of requests in sort_list */
 197         struct list_head fifo;
 198
 199         unsigned long slice_start;
 200         unsigned long slice_end;
 201         unsigned long slice_left;
 202         unsigned long service_last;
 203
 204         /* number of requests that are on the dispatch list */
 205         int on_dispatch[2];
 206
 207         /* io prio of this group */
 208         unsigned short ioprio, org_ioprio;
 209         unsigned short ioprio_class, org_ioprio_class;
 210
 211         /* various state flags, see below */
 212         unsigned int flags;
 213 };
 214
 215 struct cfq_rq {
 216         struct rb_node rb_node;
 217         sector_t rb_key;
 218         struct request *request;
 219         struct hlist_node hash;
 220
 221         struct cfq_queue *cfq_queue;
 222         struct cfq_io_context *io_context;
 223
 224         unsigned int crq_flags;
 225 };
 226
 227 enum cfqq_state_flags {
 228         CFQ_CFQQ_FLAG_on_rr = 0,
 229         CFQ_CFQQ_FLAG_wait_request,
 230         CFQ_CFQQ_FLAG_must_alloc,
 231         CFQ_CFQQ_FLAG_must_alloc_slice,
 232         CFQ_CFQQ_FLAG_must_dispatch,
 233         CFQ_CFQQ_FLAG_fifo_expire,
 234         CFQ_CFQQ_FLAG_idle_window,
 235         CFQ_CFQQ_FLAG_prio_changed,
 236 };
 237
 238 #define CFQ_CFQQ_FNS(name)                                              \
 239 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 240 {                                                                       \
 241         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 242 }                                                                       \
 243 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 244 {                                                                       \
 245         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 246 }                                                                       \
 247 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 248 {                                                                       \
 249         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 250 }
 251
 252 CFQ_CFQQ_FNS(on_rr);
 253 CFQ_CFQQ_FNS(wait_request);
 254 CFQ_CFQQ_FNS(must_alloc);
 255 CFQ_CFQQ_FNS(must_alloc_slice);
 256 CFQ_CFQQ_FNS(must_dispatch);
 257 CFQ_CFQQ_FNS(fifo_expire);
 258 CFQ_CFQQ_FNS(idle_window);
 259 CFQ_CFQQ_FNS(prio_changed);
 260 #undef CFQ_CFQQ_FNS
 261
 262 enum cfq_rq_state_flags {
 263         CFQ_CRQ_FLAG_is_sync = 0,
 264 };
 265
 266 #define CFQ_CRQ_FNS(name)                                               \
 267 static inline void cfq_mark_crq_##name(struct cfq_rq *crq)              \
 268 {                                                                       \
 269         crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name);                   \
 270 }                                                                       \
 271 static inline void cfq_clear_crq_##name(struct cfq_rq *crq)             \
 272 {                                                                       \
 273         crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name);                  \
 274 }                                                                       \
 275 static inline int cfq_crq_##name(const struct cfq_rq *crq)              \
 276 {                                                                       \
 277         return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0;      \
 278 }
 279
 280 CFQ_CRQ_FNS(is_sync);
 281 #undef CFQ_CRQ_FNS
 282
 283 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 284 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
 285 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
 286
 287 #define process_sync(tsk)       ((tsk)->flags & PF_SYNCWRITE)
 288
 289 /*
 290  * lots of deadline iosched dupes, can be abstracted later...
 291  */
 292 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
 293 {
 294         hlist_del_init(&crq->hash);
 295 }
 296
 297 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
 298 {
 299         const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
 300
 301         hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
 302 }
 303
 304 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
 305 {
 306         struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
 307         struct hlist_node *entry, *next;
 308
 309         hlist_for_each_safe(entry, next, hash_list) {
 310                 struct cfq_rq *crq = list_entry_hash(entry);
 311                 struct request *__rq = crq->request;
 312
 313                 if (!rq_mergeable(__rq)) {
 314                         cfq_del_crq_hash(crq);
 315                         continue;
 316                 }
 317
 318                 if (rq_hash_key(__rq) == offset)
 319                         return __rq;
 320         }
 321
 322         return NULL;
 323 }
 324
 325 /*
 326  * scheduler run of queue, if there are requests pending and no one in the
 327  * driver that will restart queueing
 328  */
 329 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 330 {
 331         if (cfqd->busy_queues)
 332                 kblockd_schedule_work(&cfqd->unplug_work);
 333 }
 334
 335 static int cfq_queue_empty(request_queue_t *q)
 336 {
 337         struct cfq_data *cfqd = q->elevator->elevator_data;
 338
 339         return !cfqd->busy_queues;
 340 }
 341
 342 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
 343 {
 344         if (rw == READ || process_sync(task))
 345                 return task->pid;
 346
 347         return CFQ_KEY_ASYNC;
 348 }
 349
 350 /*
 351  * Lifted from AS - choose which of crq1 and crq2 that is best served now.
 352  * We choose the request that is closest to the head right now. Distance
 353  * behind the head is penalized and only allowed to a certain extent.
 354  */
 355 static struct cfq_rq *
 356 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
 357 {
 358         sector_t last, s1, s2, d1 = 0, d2 = 0;
 359         unsigned long back_max;
 360 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 361 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 362         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 363
 364         if (crq1 == NULL || crq1 == crq2)
 365                 return crq2;
 366         if (crq2 == NULL)
 367                 return crq1;
 368
 369         if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
 370                 return crq1;
 371         else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
 372                 return crq2;
 373
 374         s1 = crq1->request->sector;
 375         s2 = crq2->request->sector;
 376
 377         last = cfqd->last_sector;
 378
 379         /*
 380          * by definition, 1KiB is 2 sectors
 381          */
 382         back_max = cfqd->cfq_back_max * 2;
 383
 384         /*
 385          * Strict one way elevator _except_ in the case where we allow
 386          * short backward seeks which are biased as twice the cost of a
 387          * similar forward seek.
 388          */
 389         if (s1 >= last)
 390                 d1 = s1 - last;
 391         else if (s1 + back_max >= last)
 392                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 393         else
 394                 wrap |= CFQ_RQ1_WRAP;
 395
 396         if (s2 >= last)
 397                 d2 = s2 - last;
 398         else if (s2 + back_max >= last)
 399                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 400         else
 401                 wrap |= CFQ_RQ2_WRAP;
 402
 403         /* Found required data */
 404
 405         /*
 406          * By doing switch() on the bit mask "wrap" we avoid having to
 407          * check two variables for all permutations: --> faster!
 408          */
 409         switch (wrap) {
 410         case 0: /* common case for CFQ: crq1 and crq2 not wrapped */
 411                 if (d1 < d2)
 412                         return crq1;
 413                 else if (d2 < d1)
 414                         return crq2;
 415                 else {
 416                         if (s1 >= s2)
 417                                 return crq1;
 418                         else
 419                                 return crq2;
 420                 }
 421
 422         case CFQ_RQ2_WRAP:
 423                 return crq1;
 424         case CFQ_RQ1_WRAP:
 425                 return crq2;
 426         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both crqs wrapped */
 427         default:
 428                 /*
 429                  * Since both rqs are wrapped,
 430                  * start with the one that's further behind head
 431                  * (--> only *one* back seek required),
 432                  * since back seek takes more time than forward.
 433                  */
 434                 if (s1 <= s2)
 435                         return crq1;
 436                 else
 437                         return crq2;
 438         }
 439 }
 440
 441 /*
 442  * would be nice to take fifo expire time into account as well
 443  */
 444 static struct cfq_rq *
 445 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 446                   struct cfq_rq *last)
 447 {
 448         struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
 449         struct rb_node *rbnext, *rbprev;
 450
 451         if (!(rbnext = rb_next(&last->rb_node))) {
 452                 rbnext = rb_first(&cfqq->sort_list);
 453                 if (rbnext == &last->rb_node)
 454                         rbnext = NULL;
 455         }
 456
 457         rbprev = rb_prev(&last->rb_node);
 458
 459         if (rbprev)
 460                 crq_prev = rb_entry_crq(rbprev);
 461         if (rbnext)
 462                 crq_next = rb_entry_crq(rbnext);
 463
 464         return cfq_choose_req(cfqd, crq_next, crq_prev);
 465 }
 466
 467 static void cfq_update_next_crq(struct cfq_rq *crq)
 468 {
 469         struct cfq_queue *cfqq = crq->cfq_queue;
 470
 471         if (cfqq->next_crq == crq)
 472                 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
 473 }
 474
 475 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 476 {
 477         struct cfq_data *cfqd = cfqq->cfqd;
 478         struct list_head *list, *entry;
 479
 480         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 481
 482         list_del(&cfqq->cfq_list);
 483
 484         if (cfq_class_rt(cfqq))
 485                 list = &cfqd->cur_rr;
 486         else if (cfq_class_idle(cfqq))
 487                 list = &cfqd->idle_rr;
 488         else {
 489                 /*
 490                  * if cfqq has requests in flight, don't allow it to be
 491                  * found in cfq_set_active_queue before it has finished them.
 492                  * this is done to increase fairness between a process that
 493                  * has lots of io pending vs one that only generates one
 494                  * sporadically or synchronously
 495                  */
 496                 if (cfq_cfqq_dispatched(cfqq))
 497                         list = &cfqd->busy_rr;
 498                 else
 499                         list = &cfqd->rr_list[cfqq->ioprio];
 500         }
 501
 502         /*
 503          * if queue was preempted, just add to front to be fair. busy_rr
 504          * isn't sorted.
 505          */
 506         if (preempted || list == &cfqd->busy_rr) {
 507                 list_add(&cfqq->cfq_list, list);
 508                 return;
 509         }
 510
 511         /*
 512          * sort by when queue was last serviced
 513          */
 514         entry = list;
 515         while ((entry = entry->prev) != list) {
 516                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
 517
 518                 if (!__cfqq->service_last)
 519                         break;
 520                 if (time_before(__cfqq->service_last, cfqq->service_last))
 521                         break;
 522         }
 523
 524         list_add(&cfqq->cfq_list, entry);
 525 }
 526
 527 /*
 528  * add to busy list of queues for service, trying to be fair in ordering
 529  * the pending list according to last request service
 530  */
 531 static inline void
 532 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 533 {
 534         BUG_ON(cfq_cfqq_on_rr(cfqq));
 535         cfq_mark_cfqq_on_rr(cfqq);
 536         cfqd->busy_queues++;
 537
 538         cfq_resort_rr_list(cfqq, 0);
 539 }
 540
 541 static inline void
 542 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 543 {
 544         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 545         cfq_clear_cfqq_on_rr(cfqq);
 546         list_move(&cfqq->cfq_list, &cfqd->empty_list);
 547
 548         BUG_ON(!cfqd->busy_queues);
 549         cfqd->busy_queues--;
 550 }
 551
 552 /*
 553  * rb tree support functions
 554  */
 555 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
 556 {
 557         struct cfq_queue *cfqq = crq->cfq_queue;
 558         struct cfq_data *cfqd = cfqq->cfqd;
 559         const int sync = cfq_crq_is_sync(crq);
 560
 561         BUG_ON(!cfqq->queued[sync]);
 562         cfqq->queued[sync]--;
 563
 564         cfq_update_next_crq(crq);
 565
 566         rb_erase(&crq->rb_node, &cfqq->sort_list);
 567         RB_CLEAR_COLOR(&crq->rb_node);
 568
 569         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
 570                 cfq_del_cfqq_rr(cfqd, cfqq);
 571 }
 572
 573 static struct cfq_rq *
 574 __cfq_add_crq_rb(struct cfq_rq *crq)
 575 {
 576         struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
 577         struct rb_node *parent = NULL;
 578         struct cfq_rq *__crq;
 579
 580         while (*p) {
 581                 parent = *p;
 582                 __crq = rb_entry_crq(parent);
 583
 584                 if (crq->rb_key < __crq->rb_key)
 585                         p = &(*p)->rb_left;
 586                 else if (crq->rb_key > __crq->rb_key)
 587                         p = &(*p)->rb_right;
 588                 else
 589                         return __crq;
 590         }
 591
 592         rb_link_node(&crq->rb_node, parent, p);
 593         return NULL;
 594 }
 595
 596 static void cfq_add_crq_rb(struct cfq_rq *crq)
 597 {
 598         struct cfq_queue *cfqq = crq->cfq_queue;
 599         struct cfq_data *cfqd = cfqq->cfqd;
 600         struct request *rq = crq->request;
 601         struct cfq_rq *__alias;
 602
 603         crq->rb_key = rq_rb_key(rq);
 604         cfqq->queued[cfq_crq_is_sync(crq)]++;
 605
 606         /*
 607          * looks a little odd, but the first insert might return an alias.
 608          * if that happens, put the alias on the dispatch list
 609          */
 610         while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
 611                 cfq_dispatch_insert(cfqd->queue, __alias);
 612
 613         rb_insert_color(&crq->rb_node, &cfqq->sort_list);
 614
 615         if (!cfq_cfqq_on_rr(cfqq))
 616                 cfq_add_cfqq_rr(cfqd, cfqq);
 617
 618         /*
 619          * check if this request is a better next-serve candidate
 620          */
 621         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
 622 }
 623
 624 static inline void
 625 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
 626 {
 627         rb_erase(&crq->rb_node, &cfqq->sort_list);
 628         cfqq->queued[cfq_crq_is_sync(crq)]--;
 629
 630         cfq_add_crq_rb(crq);
 631 }
 632
 633 static struct request *
 634 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 635 {
 636         struct task_struct *tsk = current;
 637         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
 638         struct cfq_queue *cfqq;
 639         struct rb_node *n;
 640         sector_t sector;
 641
 642         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
 643         if (!cfqq)
 644                 goto out;
 645
 646         sector = bio->bi_sector + bio_sectors(bio);
 647         n = cfqq->sort_list.rb_node;
 648         while (n) {
 649                 struct cfq_rq *crq = rb_entry_crq(n);
 650
 651                 if (sector < crq->rb_key)
 652                         n = n->rb_left;
 653                 else if (sector > crq->rb_key)
 654                         n = n->rb_right;
 655                 else
 656                         return crq->request;
 657         }
 658
 659 out:
 660         return NULL;
 661 }
 662
 663 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 664 {
 665         struct cfq_data *cfqd = q->elevator->elevator_data;
 666
 667         cfqd->rq_in_driver++;
 668
 669         /*
 670          * If the depth is larger 1, it really could be queueing. But lets
 671          * make the mark a little higher - idling could still be good for
 672          * low queueing, and a low queueing number could also just indicate
 673          * a SCSI mid layer like behaviour where limit+1 is often seen.
 674          */
 675         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
 676                 cfqd->hw_tag = 1;
 677 }
 678
 679 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 680 {
 681         struct cfq_data *cfqd = q->elevator->elevator_data;
 682
 683         WARN_ON(!cfqd->rq_in_driver);
 684         cfqd->rq_in_driver--;
 685 }
 686
 687 static void cfq_remove_request(struct request *rq)
 688 {
 689         struct cfq_rq *crq = RQ_DATA(rq);
 690
 691         list_del_init(&rq->queuelist);
 692         cfq_del_crq_rb(crq);
 693         cfq_del_crq_hash(crq);
 694 }
 695
 696 static int
 697 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 698 {
 699         struct cfq_data *cfqd = q->elevator->elevator_data;
 700         struct request *__rq;
 701         int ret;
 702
 703         __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
 704         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 705                 ret = ELEVATOR_BACK_MERGE;
 706                 goto out;
 707         }
 708
 709         __rq = cfq_find_rq_fmerge(cfqd, bio);
 710         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 711                 ret = ELEVATOR_FRONT_MERGE;
 712                 goto out;
 713         }
 714
 715         return ELEVATOR_NO_MERGE;
 716 out:
 717         *req = __rq;
 718         return ret;
 719 }
 720
 721 static void cfq_merged_request(request_queue_t *q, struct request *req)
 722 {
 723         struct cfq_data *cfqd = q->elevator->elevator_data;
 724         struct cfq_rq *crq = RQ_DATA(req);
 725
 726         cfq_del_crq_hash(crq);
 727         cfq_add_crq_hash(cfqd, crq);
 728
 729         if (rq_rb_key(req) != crq->rb_key) {
 730                 struct cfq_queue *cfqq = crq->cfq_queue;
 731
 732                 cfq_update_next_crq(crq);
 733                 cfq_reposition_crq_rb(cfqq, crq);
 734         }
 735 }
 736
 737 static void
 738 cfq_merged_requests(request_queue_t *q, struct request *rq,
 739                     struct request *next)
 740 {
 741         cfq_merged_request(q, rq);
 742
 743         /*
 744          * reposition in fifo if next is older than rq
 745          */
 746         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 747             time_before(next->start_time, rq->start_time))
 748                 list_move(&rq->queuelist, &next->queuelist);
 749
 750         cfq_remove_request(next);
 751 }
 752
 753 static inline void
 754 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 755 {
 756         if (cfqq) {
 757                 /*
 758                  * stop potential idle class queues waiting service
 759                  */
 760                 del_timer(&cfqd->idle_class_timer);
 761
 762                 cfqq->slice_start = jiffies;
 763                 cfqq->slice_end = 0;
 764                 cfqq->slice_left = 0;
 765                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 766                 cfq_clear_cfqq_fifo_expire(cfqq);
 767         }
 768
 769         cfqd->active_queue = cfqq;
 770 }
 771
 772 /*
 773  * current cfqq expired its slice (or was too idle), select new one
 774  */
 775 static void
 776 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 777                     int preempted)
 778 {
 779         unsigned long now = jiffies;
 780
 781         if (cfq_cfqq_wait_request(cfqq))
 782                 del_timer(&cfqd->idle_slice_timer);
 783
 784         if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
 785                 cfqq->service_last = now;
 786                 cfq_schedule_dispatch(cfqd);
 787         }
 788
 789         cfq_clear_cfqq_must_dispatch(cfqq);
 790         cfq_clear_cfqq_wait_request(cfqq);
 791
 792         /*
 793          * store what was left of this slice, if the queue idled out
 794          * or was preempted
 795          */
 796         if (time_after(cfqq->slice_end, now))
 797                 cfqq->slice_left = cfqq->slice_end - now;
 798         else
 799                 cfqq->slice_left = 0;
 800
 801         if (cfq_cfqq_on_rr(cfqq))
 802                 cfq_resort_rr_list(cfqq, preempted);
 803
 804         if (cfqq == cfqd->active_queue)
 805                 cfqd->active_queue = NULL;
 806
 807         if (cfqd->active_cic) {
 808                 put_io_context(cfqd->active_cic->ioc);
 809                 cfqd->active_cic = NULL;
 810         }
 811
 812         cfqd->dispatch_slice = 0;
 813 }
 814
 815 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 816 {
 817         struct cfq_queue *cfqq = cfqd->active_queue;
 818
 819         if (cfqq)
 820                 __cfq_slice_expired(cfqd, cfqq, preempted);
 821 }
 822
 823 /*
 824  * 0
 825  * 0,1
 826  * 0,1,2
 827  * 0,1,2,3
 828  * 0,1,2,3,4
 829  * 0,1,2,3,4,5
 830  * 0,1,2,3,4,5,6
 831  * 0,1,2,3,4,5,6,7
 832  */
 833 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 834 {
 835         int prio, wrap;
 836
 837         prio = -1;
 838         wrap = 0;
 839         do {
 840                 int p;
 841
 842                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 843                         if (!list_empty(&cfqd->rr_list[p])) {
 844                                 prio = p;
 845                                 break;
 846                         }
 847                 }
 848
 849                 if (prio != -1)
 850                         break;
 851                 cfqd->cur_prio = 0;
 852                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 853                         cfqd->cur_end_prio = 0;
 854                         if (wrap)
 855                                 break;
 856                         wrap = 1;
 857                 }
 858         } while (1);
 859
 860         if (unlikely(prio == -1))
 861                 return -1;
 862
 863         BUG_ON(prio >= CFQ_PRIO_LISTS);
 864
 865         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 866
 867         cfqd->cur_prio = prio + 1;
 868         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 869                 cfqd->cur_end_prio = cfqd->cur_prio;
 870                 cfqd->cur_prio = 0;
 871         }
 872         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 873                 cfqd->cur_prio = 0;
 874                 cfqd->cur_end_prio = 0;
 875         }
 876
 877         return prio;
 878 }
 879
 880 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 881 {
 882         struct cfq_queue *cfqq = NULL;
 883
 884         /*
 885          * if current list is non-empty, grab first entry. if it is empty,
 886          * get next prio level and grab first entry then if any are spliced
 887          */
 888         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
 889                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 890
 891         /*
 892          * If no new queues are available, check if the busy list has some
 893          * before falling back to idle io.
 894          */
 895         if (!cfqq && !list_empty(&cfqd->busy_rr))
 896                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
 897
 898         /*
 899          * if we have idle queues and no rt or be queues had pending
 900          * requests, either allow immediate service if the grace period
 901          * has passed or arm the idle grace timer
 902          */
 903         if (!cfqq && !list_empty(&cfqd->idle_rr)) {
 904                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 905
 906                 if (time_after_eq(jiffies, end))
 907                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 908                 else
 909                         mod_timer(&cfqd->idle_class_timer, end);
 910         }
 911
 912         __cfq_set_active_queue(cfqd, cfqq);
 913         return cfqq;
 914 }
 915
 916 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 917
 918 {
 919         struct cfq_io_context *cic;
 920         unsigned long sl;
 921
 922         WARN_ON(!RB_EMPTY(&cfqq->sort_list));
 923         WARN_ON(cfqq != cfqd->active_queue);
 924
 925         /*
 926          * idle is disabled, either manually or by past process history
 927          */
 928         if (!cfqd->cfq_slice_idle)
 929                 return 0;
 930         if (!cfq_cfqq_idle_window(cfqq))
 931                 return 0;
 932         /*
 933          * task has exited, don't wait
 934          */
 935         cic = cfqd->active_cic;
 936         if (!cic || !cic->ioc->task)
 937                 return 0;
 938
 939         cfq_mark_cfqq_must_dispatch(cfqq);
 940         cfq_mark_cfqq_wait_request(cfqq);
 941
 942         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 943
 944         /*
 945          * we don't want to idle for seeks, but we do want to allow
 946          * fair distribution of slice time for a process doing back-to-back
 947          * seeks. so allow a little bit of time for him to submit a new rq
 948          */
 949         if (sample_valid(cic->seek_samples) && cic->seek_mean > 131072)
 950                 sl = 2;
 951
 952         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 953         return 1;
 954 }
 955
 956 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
 957 {
 958         struct cfq_data *cfqd = q->elevator->elevator_data;
 959         struct cfq_queue *cfqq = crq->cfq_queue;
 960
 961         cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
 962         cfq_remove_request(crq->request);
 963         cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
 964         elv_dispatch_sort(q, crq->request);
 965 }
 966
 967 /*
 968  * return expired entry, or NULL to just start from scratch in rbtree
 969  */
 970 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
 971 {
 972         struct cfq_data *cfqd = cfqq->cfqd;
 973         struct request *rq;
 974         struct cfq_rq *crq;
 975
 976         if (cfq_cfqq_fifo_expire(cfqq))
 977                 return NULL;
 978
 979         if (!list_empty(&cfqq->fifo)) {
 980                 int fifo = cfq_cfqq_class_sync(cfqq);
 981
 982                 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
 983                 rq = crq->request;
 984                 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
 985                         cfq_mark_cfqq_fifo_expire(cfqq);
 986                         return crq;
 987                 }
 988         }
 989
 990         return NULL;
 991 }
 992
 993 /*
 994  * Scale schedule slice based on io priority. Use the sync time slice only
 995  * if a queue is marked sync and has sync io queued. A sync queue with async
 996  * io only, should not get full sync slice length.
 997  */
 998 static inline int
 999 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1000 {
1001         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
1002
1003         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1004
1005         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
1006 }
1007
1008 static inline void
1009 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1010 {
1011         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
1012 }
1013
1014 static inline int
1015 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1016 {
1017         const int base_rq = cfqd->cfq_slice_async_rq;
1018
1019         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1020
1021         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1022 }
1023
1024 /*
1025  * get next queue for service
1026  */
1027 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1028 {
1029         unsigned long now = jiffies;
1030         struct cfq_queue *cfqq;
1031
1032         cfqq = cfqd->active_queue;
1033         if (!cfqq)
1034                 goto new_queue;
1035
1036         /*
1037          * slice has expired
1038          */
1039         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
1040                 goto expire;
1041
1042         /*
1043          * if queue has requests, dispatch one. if not, check if
1044          * enough slice is left to wait for one
1045          */
1046         if (!RB_EMPTY(&cfqq->sort_list))
1047                 goto keep_queue;
1048         else if (cfq_cfqq_class_sync(cfqq) &&
1049                  time_before(now, cfqq->slice_end)) {
1050                 if (cfq_arm_slice_timer(cfqd, cfqq))
1051                         return NULL;
1052         }
1053
1054 expire:
1055         cfq_slice_expired(cfqd, 0);
1056 new_queue:
1057         cfqq = cfq_set_active_queue(cfqd);
1058 keep_queue:
1059         return cfqq;
1060 }
1061
1062 static int
1063 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1064                         int max_dispatch)
1065 {
1066         int dispatched = 0;
1067
1068         BUG_ON(RB_EMPTY(&cfqq->sort_list));
1069
1070         do {
1071                 struct cfq_rq *crq;
1072
1073                 /*
1074                  * follow expired path, else get first next available
1075                  */
1076                 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1077                         crq = cfqq->next_crq;
1078
1079                 /*
1080                  * finally, insert request into driver dispatch list
1081                  */
1082                 cfq_dispatch_insert(cfqd->queue, crq);
1083
1084                 cfqd->dispatch_slice++;
1085                 dispatched++;
1086
1087                 if (!cfqd->active_cic) {
1088                         atomic_inc(&crq->io_context->ioc->refcount);
1089                         cfqd->active_cic = crq->io_context;
1090                 }
1091
1092                 if (RB_EMPTY(&cfqq->sort_list))
1093                         break;
1094
1095         } while (dispatched < max_dispatch);
1096
1097         /*
1098          * if slice end isn't set yet, set it. if at least one request was
1099          * sync, use the sync time slice value
1100          */
1101         if (!cfqq->slice_end)
1102                 cfq_set_prio_slice(cfqd, cfqq);
1103
1104         /*
1105          * expire an async queue immediately if it has used up its slice. idle
1106          * queue always expire after 1 dispatch round.
1107          */
1108         if ((!cfq_cfqq_sync(cfqq) &&
1109             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1110             cfq_class_idle(cfqq))
1111                 cfq_slice_expired(cfqd, 0);
1112
1113         return dispatched;
1114 }
1115
1116 static int
1117 cfq_forced_dispatch_cfqqs(struct list_head *list)
1118 {
1119         int dispatched = 0;
1120         struct cfq_queue *cfqq, *next;
1121         struct cfq_rq *crq;
1122
1123         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1124                 while ((crq = cfqq->next_crq)) {
1125                         cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1126                         dispatched++;
1127                 }
1128                 BUG_ON(!list_empty(&cfqq->fifo));
1129         }
1130         return dispatched;
1131 }
1132
1133 static int
1134 cfq_forced_dispatch(struct cfq_data *cfqd)
1135 {
1136         int i, dispatched = 0;
1137
1138         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1139                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1140
1141         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1142         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1143         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1144
1145         cfq_slice_expired(cfqd, 0);
1146
1147         BUG_ON(cfqd->busy_queues);
1148
1149         return dispatched;
1150 }
1151
1152 static int
1153 cfq_dispatch_requests(request_queue_t *q, int force)
1154 {
1155         struct cfq_data *cfqd = q->elevator->elevator_data;
1156         struct cfq_queue *cfqq;
1157
1158         if (!cfqd->busy_queues)
1159                 return 0;
1160
1161         if (unlikely(force))
1162                 return cfq_forced_dispatch(cfqd);
1163
1164         cfqq = cfq_select_queue(cfqd);
1165         if (cfqq) {
1166                 int max_dispatch;
1167
1168                 cfq_clear_cfqq_must_dispatch(cfqq);
1169                 cfq_clear_cfqq_wait_request(cfqq);
1170                 del_timer(&cfqd->idle_slice_timer);
1171
1172                 max_dispatch = cfqd->cfq_quantum;
1173                 if (cfq_class_idle(cfqq))
1174                         max_dispatch = 1;
1175
1176                 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1177         }
1178
1179         return 0;
1180 }
1181
1182 /*
1183  * task holds one reference to the queue, dropped when task exits. each crq
1184  * in-flight on this queue also holds a reference, dropped when crq is freed.
1185  *
1186  * queue lock must be held here.
1187  */
1188 static void cfq_put_queue(struct cfq_queue *cfqq)
1189 {
1190         struct cfq_data *cfqd = cfqq->cfqd;
1191
1192         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1193
1194         if (!atomic_dec_and_test(&cfqq->ref))
1195                 return;
1196
1197         BUG_ON(rb_first(&cfqq->sort_list));
1198         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1199         BUG_ON(cfq_cfqq_on_rr(cfqq));
1200
1201         if (unlikely(cfqd->active_queue == cfqq))
1202                 __cfq_slice_expired(cfqd, cfqq, 0);
1203
1204         /*
1205          * it's on the empty list and still hashed
1206          */
1207         list_del(&cfqq->cfq_list);
1208         hlist_del(&cfqq->cfq_hash);
1209         kmem_cache_free(cfq_pool, cfqq);
1210 }
1211
1212 static inline struct cfq_queue *
1213 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1214                     const int hashval)
1215 {
1216         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1217         struct hlist_node *entry;
1218         struct cfq_queue *__cfqq;
1219
1220         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1221                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1222
1223                 if (__cfqq->key == key && (__p == prio || !prio))
1224                         return __cfqq;
1225         }
1226
1227         return NULL;
1228 }
1229
1230 static struct cfq_queue *
1231 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1232 {
1233         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1234 }
1235
1236 static void cfq_free_io_context(struct io_context *ioc)
1237 {
1238         struct cfq_io_context *__cic;
1239         struct rb_node *n;
1240         int freed = 0;
1241
1242         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1243                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1244                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1245                 kmem_cache_free(cfq_ioc_pool, __cic);
1246                 freed++;
1247         }
1248
1249         if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1250                 complete(ioc_gone);
1251 }
1252
1253 static void cfq_trim(struct io_context *ioc)
1254 {
1255         ioc->set_ioprio = NULL;
1256         cfq_free_io_context(ioc);
1257 }
1258
1259 /*
1260  * Called with interrupts disabled
1261  */
1262 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1263 {
1264         struct cfq_data *cfqd = cic->key;
1265         request_queue_t *q;
1266
1267         if (!cfqd)
1268                 return;
1269
1270         q = cfqd->queue;
1271
1272         WARN_ON(!irqs_disabled());
1273
1274         spin_lock(q->queue_lock);
1275
1276         if (cic->cfqq[ASYNC]) {
1277                 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
1278                         __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
1279                 cfq_put_queue(cic->cfqq[ASYNC]);
1280                 cic->cfqq[ASYNC] = NULL;
1281         }
1282
1283         if (cic->cfqq[SYNC]) {
1284                 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
1285                         __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
1286                 cfq_put_queue(cic->cfqq[SYNC]);
1287                 cic->cfqq[SYNC] = NULL;
1288         }
1289
1290         cic->key = NULL;
1291         list_del_init(&cic->queue_list);
1292         spin_unlock(q->queue_lock);
1293 }
1294
1295 static void cfq_exit_io_context(struct io_context *ioc)
1296 {
1297         struct cfq_io_context *__cic;
1298         unsigned long flags;
1299         struct rb_node *n;
1300
1301         /*
1302          * put the reference this task is holding to the various queues
1303          */
1304         spin_lock_irqsave(&cfq_exit_lock, flags);
1305
1306         n = rb_first(&ioc->cic_root);
1307         while (n != NULL) {
1308                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1309
1310                 cfq_exit_single_io_context(__cic);
1311                 n = rb_next(n);
1312         }
1313
1314         spin_unlock_irqrestore(&cfq_exit_lock, flags);
1315 }
1316
1317 static struct cfq_io_context *
1318 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1319 {
1320         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1321
1322         if (cic) {
1323                 RB_CLEAR(&cic->rb_node);
1324                 cic->key = NULL;
1325                 cic->cfqq[ASYNC] = NULL;
1326                 cic->cfqq[SYNC] = NULL;
1327                 cic->last_end_request = jiffies;
1328                 cic->ttime_total = 0;
1329                 cic->ttime_samples = 0;
1330                 cic->ttime_mean = 0;
1331                 cic->dtor = cfq_free_io_context;
1332                 cic->exit = cfq_exit_io_context;
1333                 INIT_LIST_HEAD(&cic->queue_list);
1334                 atomic_inc(&ioc_count);
1335         }
1336
1337         return cic;
1338 }
1339
1340 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1341 {
1342         struct task_struct *tsk = current;
1343         int ioprio_class;
1344
1345         if (!cfq_cfqq_prio_changed(cfqq))
1346                 return;
1347
1348         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1349         switch (ioprio_class) {
1350                 default:
1351                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1352                 case IOPRIO_CLASS_NONE:
1353                         /*
1354                          * no prio set, place us in the middle of the BE classes
1355                          */
1356                         cfqq->ioprio = task_nice_ioprio(tsk);
1357                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1358                         break;
1359                 case IOPRIO_CLASS_RT:
1360                         cfqq->ioprio = task_ioprio(tsk);
1361                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1362                         break;
1363                 case IOPRIO_CLASS_BE:
1364                         cfqq->ioprio = task_ioprio(tsk);
1365                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1366                         break;
1367                 case IOPRIO_CLASS_IDLE:
1368                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1369                         cfqq->ioprio = 7;
1370                         cfq_clear_cfqq_idle_window(cfqq);
1371                         break;
1372         }
1373
1374         /*
1375          * keep track of original prio settings in case we have to temporarily
1376          * elevate the priority of this queue
1377          */
1378         cfqq->org_ioprio = cfqq->ioprio;
1379         cfqq->org_ioprio_class = cfqq->ioprio_class;
1380
1381         if (cfq_cfqq_on_rr(cfqq))
1382                 cfq_resort_rr_list(cfqq, 0);
1383
1384         cfq_clear_cfqq_prio_changed(cfqq);
1385 }
1386
1387 static inline void changed_ioprio(struct cfq_io_context *cic)
1388 {
1389         struct cfq_data *cfqd = cic->key;
1390         struct cfq_queue *cfqq;
1391         if (cfqd) {
1392                 spin_lock(cfqd->queue->queue_lock);
1393                 cfqq = cic->cfqq[ASYNC];
1394                 if (cfqq) {
1395                         struct cfq_queue *new_cfqq;
1396                         new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC,
1397                                                 cic->ioc->task, GFP_ATOMIC);
1398                         if (new_cfqq) {
1399                                 cic->cfqq[ASYNC] = new_cfqq;
1400                                 cfq_put_queue(cfqq);
1401                         }
1402                 }
1403                 cfqq = cic->cfqq[SYNC];
1404                 if (cfqq) {
1405                         cfq_mark_cfqq_prio_changed(cfqq);
1406                         cfq_init_prio_data(cfqq);
1407                 }
1408                 spin_unlock(cfqd->queue->queue_lock);
1409         }
1410 }
1411
1412 /*
1413  * callback from sys_ioprio_set, irqs are disabled
1414  */
1415 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1416 {
1417         struct cfq_io_context *cic;
1418         struct rb_node *n;
1419
1420         spin_lock(&cfq_exit_lock);
1421
1422         n = rb_first(&ioc->cic_root);
1423         while (n != NULL) {
1424                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1425
1426                 changed_ioprio(cic);
1427                 n = rb_next(n);
1428         }
1429
1430         spin_unlock(&cfq_exit_lock);
1431
1432         return 0;
1433 }
1434
1435 static struct cfq_queue *
1436 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1437               gfp_t gfp_mask)
1438 {
1439         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1440         struct cfq_queue *cfqq, *new_cfqq = NULL;
1441         unsigned short ioprio;
1442
1443 retry:
1444         ioprio = tsk->ioprio;
1445         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1446
1447         if (!cfqq) {
1448                 if (new_cfqq) {
1449                         cfqq = new_cfqq;
1450                         new_cfqq = NULL;
1451                 } else if (gfp_mask & __GFP_WAIT) {
1452                         spin_unlock_irq(cfqd->queue->queue_lock);
1453                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1454                         spin_lock_irq(cfqd->queue->queue_lock);
1455                         goto retry;
1456                 } else {
1457                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1458                         if (!cfqq)
1459                                 goto out;
1460                 }
1461
1462                 memset(cfqq, 0, sizeof(*cfqq));
1463
1464                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1465                 INIT_LIST_HEAD(&cfqq->cfq_list);
1466                 RB_CLEAR_ROOT(&cfqq->sort_list);
1467                 INIT_LIST_HEAD(&cfqq->fifo);
1468
1469                 cfqq->key = key;
1470                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1471                 atomic_set(&cfqq->ref, 0);
1472                 cfqq->cfqd = cfqd;
1473                 cfqq->service_last = 0;
1474                 /*
1475                  * set ->slice_left to allow preemption for a new process
1476                  */
1477                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1478                 if (!cfqd->hw_tag)
1479                         cfq_mark_cfqq_idle_window(cfqq);
1480                 cfq_mark_cfqq_prio_changed(cfqq);
1481                 cfq_init_prio_data(cfqq);
1482         }
1483
1484         if (new_cfqq)
1485                 kmem_cache_free(cfq_pool, new_cfqq);
1486
1487         atomic_inc(&cfqq->ref);
1488 out:
1489         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1490         return cfqq;
1491 }
1492
1493 static void
1494 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1495 {
1496         spin_lock(&cfq_exit_lock);
1497         rb_erase(&cic->rb_node, &ioc->cic_root);
1498         list_del_init(&cic->queue_list);
1499         spin_unlock(&cfq_exit_lock);
1500         kmem_cache_free(cfq_ioc_pool, cic);
1501         atomic_dec(&ioc_count);
1502 }
1503
1504 static struct cfq_io_context *
1505 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1506 {
1507         struct rb_node *n;
1508         struct cfq_io_context *cic;
1509         void *k, *key = cfqd;
1510
1511 restart:
1512         n = ioc->cic_root.rb_node;
1513         while (n) {
1514                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1515                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1516                 k = cic->key;
1517                 if (unlikely(!k)) {
1518                         cfq_drop_dead_cic(ioc, cic);
1519                         goto restart;
1520                 }
1521
1522                 if (key < k)
1523                         n = n->rb_left;
1524                 else if (key > k)
1525                         n = n->rb_right;
1526                 else
1527                         return cic;
1528         }
1529
1530         return NULL;
1531 }
1532
1533 static inline void
1534 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1535              struct cfq_io_context *cic)
1536 {
1537         struct rb_node **p;
1538         struct rb_node *parent;
1539         struct cfq_io_context *__cic;
1540         void *k;
1541
1542         cic->ioc = ioc;
1543         cic->key = cfqd;
1544
1545         ioc->set_ioprio = cfq_ioc_set_ioprio;
1546 restart:
1547         parent = NULL;
1548         p = &ioc->cic_root.rb_node;
1549         while (*p) {
1550                 parent = *p;
1551                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1552                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1553                 k = __cic->key;
1554                 if (unlikely(!k)) {
1555                         cfq_drop_dead_cic(ioc, cic);
1556                         goto restart;
1557                 }
1558
1559                 if (cic->key < k)
1560                         p = &(*p)->rb_left;
1561                 else if (cic->key > k)
1562                         p = &(*p)->rb_right;
1563                 else
1564                         BUG();
1565         }
1566
1567         spin_lock(&cfq_exit_lock);
1568         rb_link_node(&cic->rb_node, parent, p);
1569         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1570         list_add(&cic->queue_list, &cfqd->cic_list);
1571         spin_unlock(&cfq_exit_lock);
1572 }
1573
1574 /*
1575  * Setup general io context and cfq io context. There can be several cfq
1576  * io contexts per general io context, if this process is doing io to more
1577  * than one device managed by cfq.
1578  */
1579 static struct cfq_io_context *
1580 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1581 {
1582         struct io_context *ioc = NULL;
1583         struct cfq_io_context *cic;
1584
1585         might_sleep_if(gfp_mask & __GFP_WAIT);
1586
1587         ioc = get_io_context(gfp_mask);
1588         if (!ioc)
1589                 return NULL;
1590
1591         cic = cfq_cic_rb_lookup(cfqd, ioc);
1592         if (cic)
1593                 goto out;
1594
1595         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1596         if (cic == NULL)
1597                 goto err;
1598
1599         cfq_cic_link(cfqd, ioc, cic);
1600 out:
1601         return cic;
1602 err:
1603         put_io_context(ioc);
1604         return NULL;
1605 }
1606
1607 static void
1608 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1609 {
1610         unsigned long elapsed, ttime;
1611
1612         /*
1613          * if this context already has stuff queued, thinktime is from
1614          * last queue not last end
1615          */
1616 #if 0
1617         if (time_after(cic->last_end_request, cic->last_queue))
1618                 elapsed = jiffies - cic->last_end_request;
1619         else
1620                 elapsed = jiffies - cic->last_queue;
1621 #else
1622                 elapsed = jiffies - cic->last_end_request;
1623 #endif
1624
1625         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1626
1627         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1628         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1629         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1630 }
1631
1632 static void
1633 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1634                        struct cfq_rq *crq)
1635 {
1636         sector_t sdist;
1637         u64 total;
1638
1639         if (cic->last_request_pos < crq->request->sector)
1640                 sdist = crq->request->sector - cic->last_request_pos;
1641         else
1642                 sdist = cic->last_request_pos - crq->request->sector;
1643
1644         /*
1645          * Don't allow the seek distance to get too large from the
1646          * odd fragment, pagein, etc
1647          */
1648         if (cic->seek_samples <= 60) /* second&third seek */
1649                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1650         else
1651                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1652
1653         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1654         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1655         total = cic->seek_total + (cic->seek_samples/2);
1656         do_div(total, cic->seek_samples);
1657         cic->seek_mean = (sector_t)total;
1658 }
1659
1660 /*
1661  * Disable idle window if the process thinks too long or seeks so much that
1662  * it doesn't matter
1663  */
1664 static void
1665 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1666                        struct cfq_io_context *cic)
1667 {
1668         int enable_idle = cfq_cfqq_idle_window(cfqq);
1669
1670         if (!cic->ioc->task || !cfqd->cfq_slice_idle || cfqd->hw_tag)
1671                 enable_idle = 0;
1672         else if (sample_valid(cic->ttime_samples)) {
1673                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1674                         enable_idle = 0;
1675                 else
1676                         enable_idle = 1;
1677         }
1678
1679         if (enable_idle)
1680                 cfq_mark_cfqq_idle_window(cfqq);
1681         else
1682                 cfq_clear_cfqq_idle_window(cfqq);
1683 }
1684
1685
1686 /*
1687  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1688  * no or if we aren't sure, a 1 will cause a preempt.
1689  */
1690 static int
1691 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1692                    struct cfq_rq *crq)
1693 {
1694         struct cfq_queue *cfqq = cfqd->active_queue;
1695
1696         if (cfq_class_idle(new_cfqq))
1697                 return 0;
1698
1699         if (!cfqq)
1700                 return 1;
1701
1702         if (cfq_class_idle(cfqq))
1703                 return 1;
1704         if (!cfq_cfqq_wait_request(new_cfqq))
1705                 return 0;
1706         /*
1707          * if it doesn't have slice left, forget it
1708          */
1709         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1710                 return 0;
1711         if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1712                 return 1;
1713
1714         return 0;
1715 }
1716
1717 /*
1718  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1719  * let it have half of its nominal slice.
1720  */
1721 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1722 {
1723         struct cfq_queue *__cfqq, *next;
1724
1725         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1726                 cfq_resort_rr_list(__cfqq, 1);
1727
1728         if (!cfqq->slice_left)
1729                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1730
1731         cfqq->slice_end = cfqq->slice_left + jiffies;
1732         __cfq_slice_expired(cfqd, cfqq, 1);
1733         __cfq_set_active_queue(cfqd, cfqq);
1734 }
1735
1736 /*
1737  * should really be a ll_rw_blk.c helper
1738  */
1739 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1740 {
1741         request_queue_t *q = cfqd->queue;
1742
1743         if (!blk_queue_plugged(q))
1744                 q->request_fn(q);
1745         else
1746                 __generic_unplug_device(q);
1747 }
1748
1749 /*
1750  * Called when a new fs request (crq) is added (to cfqq). Check if there's
1751  * something we should do about it
1752  */
1753 static void
1754 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1755                  struct cfq_rq *crq)
1756 {
1757         struct cfq_io_context *cic;
1758
1759         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1760
1761         cic = crq->io_context;
1762
1763         /*
1764          * we never wait for an async request and we don't allow preemption
1765          * of an async request. so just return early
1766          */
1767         if (!cfq_crq_is_sync(crq)) {
1768                 /*
1769                  * sync process issued an async request, if it's waiting
1770                  * then expire it and kick rq handling.
1771                  */
1772                 if (cic == cfqd->active_cic &&
1773                     del_timer(&cfqd->idle_slice_timer)) {
1774                         cfq_slice_expired(cfqd, 0);
1775                         cfq_start_queueing(cfqd, cfqq);
1776                 }
1777                 return;
1778         }
1779
1780         cfq_update_io_thinktime(cfqd, cic);
1781         cfq_update_io_seektime(cfqd, cic, crq);
1782         cfq_update_idle_window(cfqd, cfqq, cic);
1783
1784         cic->last_queue = jiffies;
1785         cic->last_request_pos = crq->request->sector + crq->request->nr_sectors;
1786
1787         if (cfqq == cfqd->active_queue) {
1788                 /*
1789                  * if we are waiting for a request for this queue, let it rip
1790                  * immediately and flag that we must not expire this queue
1791                  * just now
1792                  */
1793                 if (cfq_cfqq_wait_request(cfqq)) {
1794                         cfq_mark_cfqq_must_dispatch(cfqq);
1795                         del_timer(&cfqd->idle_slice_timer);
1796                         cfq_start_queueing(cfqd, cfqq);
1797                 }
1798         } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1799                 /*
1800                  * not the active queue - expire current slice if it is
1801                  * idle and has expired it's mean thinktime or this new queue
1802                  * has some old slice time left and is of higher priority
1803                  */
1804                 cfq_preempt_queue(cfqd, cfqq);
1805                 cfq_mark_cfqq_must_dispatch(cfqq);
1806                 cfq_start_queueing(cfqd, cfqq);
1807         }
1808 }
1809
1810 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1811 {
1812         struct cfq_data *cfqd = q->elevator->elevator_data;
1813         struct cfq_rq *crq = RQ_DATA(rq);
1814         struct cfq_queue *cfqq = crq->cfq_queue;
1815
1816         cfq_init_prio_data(cfqq);
1817
1818         cfq_add_crq_rb(crq);
1819
1820         list_add_tail(&rq->queuelist, &cfqq->fifo);
1821
1822         if (rq_mergeable(rq))
1823                 cfq_add_crq_hash(cfqd, crq);
1824
1825         cfq_crq_enqueued(cfqd, cfqq, crq);
1826 }
1827
1828 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1829 {
1830         struct cfq_rq *crq = RQ_DATA(rq);
1831         struct cfq_queue *cfqq = crq->cfq_queue;
1832         struct cfq_data *cfqd = cfqq->cfqd;
1833         const int sync = cfq_crq_is_sync(crq);
1834         unsigned long now;
1835
1836         now = jiffies;
1837
1838         WARN_ON(!cfqd->rq_in_driver);
1839         WARN_ON(!cfqq->on_dispatch[sync]);
1840         cfqd->rq_in_driver--;
1841         cfqq->on_dispatch[sync]--;
1842
1843         if (!cfq_class_idle(cfqq))
1844                 cfqd->last_end_request = now;
1845
1846         if (!cfq_cfqq_dispatched(cfqq)) {
1847                 if (cfq_cfqq_on_rr(cfqq)) {
1848                         cfqq->service_last = now;
1849                         cfq_resort_rr_list(cfqq, 0);
1850                 }
1851                 cfq_schedule_dispatch(cfqd);
1852         }
1853
1854         if (cfq_crq_is_sync(crq))
1855                 crq->io_context->last_end_request = now;
1856 }
1857
1858 static struct request *
1859 cfq_former_request(request_queue_t *q, struct request *rq)
1860 {
1861         struct cfq_rq *crq = RQ_DATA(rq);
1862         struct rb_node *rbprev = rb_prev(&crq->rb_node);
1863
1864         if (rbprev)
1865                 return rb_entry_crq(rbprev)->request;
1866
1867         return NULL;
1868 }
1869
1870 static struct request *
1871 cfq_latter_request(request_queue_t *q, struct request *rq)
1872 {
1873         struct cfq_rq *crq = RQ_DATA(rq);
1874         struct rb_node *rbnext = rb_next(&crq->rb_node);
1875
1876         if (rbnext)
1877                 return rb_entry_crq(rbnext)->request;
1878
1879         return NULL;
1880 }
1881
1882 /*
1883  * we temporarily boost lower priority queues if they are holding fs exclusive
1884  * resources. they are boosted to normal prio (CLASS_BE/4)
1885  */
1886 static void cfq_prio_boost(struct cfq_queue *cfqq)
1887 {
1888         const int ioprio_class = cfqq->ioprio_class;
1889         const int ioprio = cfqq->ioprio;
1890
1891         if (has_fs_excl()) {
1892                 /*
1893                  * boost idle prio on transactions that would lock out other
1894                  * users of the filesystem
1895                  */
1896                 if (cfq_class_idle(cfqq))
1897                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1898                 if (cfqq->ioprio > IOPRIO_NORM)
1899                         cfqq->ioprio = IOPRIO_NORM;
1900         } else {
1901                 /*
1902                  * check if we need to unboost the queue
1903                  */
1904                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1905                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1906                 if (cfqq->ioprio != cfqq->org_ioprio)
1907                         cfqq->ioprio = cfqq->org_ioprio;
1908         }
1909
1910         /*
1911          * refile between round-robin lists if we moved the priority class
1912          */
1913         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1914             cfq_cfqq_on_rr(cfqq))
1915                 cfq_resort_rr_list(cfqq, 0);
1916 }
1917
1918 static inline int
1919 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1920                 struct task_struct *task, int rw)
1921 {
1922 #if 1
1923         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1924             !cfq_cfqq_must_alloc_slice(cfqq)) {
1925                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1926                 return ELV_MQUEUE_MUST;
1927         }
1928
1929         return ELV_MQUEUE_MAY;
1930 #else
1931         if (!cfqq || task->flags & PF_MEMALLOC)
1932                 return ELV_MQUEUE_MAY;
1933         if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1934                 if (cfq_cfqq_wait_request(cfqq))
1935                         return ELV_MQUEUE_MUST;
1936
1937                 /*
1938                  * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1939                  * can quickly flood the queue with writes from a single task
1940                  */
1941                 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
1942                         cfq_mark_cfqq_must_alloc_slice(cfqq);
1943                         return ELV_MQUEUE_MUST;
1944                 }
1945
1946                 return ELV_MQUEUE_MAY;
1947         }
1948         if (cfq_class_idle(cfqq))
1949                 return ELV_MQUEUE_NO;
1950         if (cfqq->allocated[rw] >= cfqd->max_queued) {
1951                 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1952                 int ret = ELV_MQUEUE_NO;
1953
1954                 if (ioc && ioc->nr_batch_requests)
1955                         ret = ELV_MQUEUE_MAY;
1956
1957                 put_io_context(ioc);
1958                 return ret;
1959         }
1960
1961         return ELV_MQUEUE_MAY;
1962 #endif
1963 }
1964
1965 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1966 {
1967         struct cfq_data *cfqd = q->elevator->elevator_data;
1968         struct task_struct *tsk = current;
1969         struct cfq_queue *cfqq;
1970
1971         /*
1972          * don't force setup of a queue from here, as a call to may_queue
1973          * does not necessarily imply that a request actually will be queued.
1974          * so just lookup a possibly existing queue, or return 'may queue'
1975          * if that fails
1976          */
1977         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1978         if (cfqq) {
1979                 cfq_init_prio_data(cfqq);
1980                 cfq_prio_boost(cfqq);
1981
1982                 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1983         }
1984
1985         return ELV_MQUEUE_MAY;
1986 }
1987
1988 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1989 {
1990         struct cfq_data *cfqd = q->elevator->elevator_data;
1991         struct request_list *rl = &q->rq;
1992
1993         if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1994                 smp_mb();
1995                 if (waitqueue_active(&rl->wait[READ]))
1996                         wake_up(&rl->wait[READ]);
1997         }
1998
1999         if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
2000                 smp_mb();
2001                 if (waitqueue_active(&rl->wait[WRITE]))
2002                         wake_up(&rl->wait[WRITE]);
2003         }
2004 }
2005
2006 /*
2007  * queue lock held here
2008  */
2009 static void cfq_put_request(request_queue_t *q, struct request *rq)
2010 {
2011         struct cfq_data *cfqd = q->elevator->elevator_data;
2012         struct cfq_rq *crq = RQ_DATA(rq);
2013
2014         if (crq) {
2015                 struct cfq_queue *cfqq = crq->cfq_queue;
2016                 const int rw = rq_data_dir(rq);
2017
2018                 BUG_ON(!cfqq->allocated[rw]);
2019                 cfqq->allocated[rw]--;
2020
2021                 put_io_context(crq->io_context->ioc);
2022
2023                 mempool_free(crq, cfqd->crq_pool);
2024                 rq->elevator_private = NULL;
2025
2026                 cfq_check_waiters(q, cfqq);
2027                 cfq_put_queue(cfqq);
2028         }
2029 }
2030
2031 /*
2032  * Allocate cfq data structures associated with this request.
2033  */
2034 static int
2035 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
2036                 gfp_t gfp_mask)
2037 {
2038         struct cfq_data *cfqd = q->elevator->elevator_data;
2039         struct task_struct *tsk = current;
2040         struct cfq_io_context *cic;
2041         const int rw = rq_data_dir(rq);
2042         pid_t key = cfq_queue_pid(tsk, rw);
2043         struct cfq_queue *cfqq;
2044         struct cfq_rq *crq;
2045         unsigned long flags;
2046         int is_sync = key != CFQ_KEY_ASYNC;
2047
2048         might_sleep_if(gfp_mask & __GFP_WAIT);
2049
2050         cic = cfq_get_io_context(cfqd, gfp_mask);
2051
2052         spin_lock_irqsave(q->queue_lock, flags);
2053
2054         if (!cic)
2055                 goto queue_fail;
2056
2057         if (!cic->cfqq[is_sync]) {
2058                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
2059                 if (!cfqq)
2060                         goto queue_fail;
2061
2062                 cic->cfqq[is_sync] = cfqq;
2063         } else
2064                 cfqq = cic->cfqq[is_sync];
2065
2066         cfqq->allocated[rw]++;
2067         cfq_clear_cfqq_must_alloc(cfqq);
2068         cfqd->rq_starved = 0;
2069         atomic_inc(&cfqq->ref);
2070         spin_unlock_irqrestore(q->queue_lock, flags);
2071
2072         crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
2073         if (crq) {
2074                 RB_CLEAR(&crq->rb_node);
2075                 crq->rb_key = 0;
2076                 crq->request = rq;
2077                 INIT_HLIST_NODE(&crq->hash);
2078                 crq->cfq_queue = cfqq;
2079                 crq->io_context = cic;
2080
2081                 if (is_sync)
2082                         cfq_mark_crq_is_sync(crq);
2083                 else
2084                         cfq_clear_crq_is_sync(crq);
2085
2086                 rq->elevator_private = crq;
2087                 return 0;
2088         }
2089
2090         spin_lock_irqsave(q->queue_lock, flags);
2091         cfqq->allocated[rw]--;
2092         if (!(cfqq->allocated[0] + cfqq->allocated[1]))
2093                 cfq_mark_cfqq_must_alloc(cfqq);
2094         cfq_put_queue(cfqq);
2095 queue_fail:
2096         if (cic)
2097                 put_io_context(cic->ioc);
2098         /*
2099          * mark us rq allocation starved. we need to kickstart the process
2100          * ourselves if there are no pending requests that can do it for us.
2101          * that would be an extremely rare OOM situation
2102          */
2103         cfqd->rq_starved = 1;
2104         cfq_schedule_dispatch(cfqd);
2105         spin_unlock_irqrestore(q->queue_lock, flags);
2106         return 1;
2107 }
2108
2109 static void cfq_kick_queue(void *data)
2110 {
2111         request_queue_t *q = data;
2112         struct cfq_data *cfqd = q->elevator->elevator_data;
2113         unsigned long flags;
2114
2115         spin_lock_irqsave(q->queue_lock, flags);
2116
2117         if (cfqd->rq_starved) {
2118                 struct request_list *rl = &q->rq;
2119
2120                 /*
2121                  * we aren't guaranteed to get a request after this, but we
2122                  * have to be opportunistic
2123                  */
2124                 smp_mb();
2125                 if (waitqueue_active(&rl->wait[READ]))
2126                         wake_up(&rl->wait[READ]);
2127                 if (waitqueue_active(&rl->wait[WRITE]))
2128                         wake_up(&rl->wait[WRITE]);
2129         }
2130
2131         blk_remove_plug(q);
2132         q->request_fn(q);
2133         spin_unlock_irqrestore(q->queue_lock, flags);
2134 }
2135
2136 /*
2137  * Timer running if the active_queue is currently idling inside its time slice
2138  */
2139 static void cfq_idle_slice_timer(unsigned long data)
2140 {
2141         struct cfq_data *cfqd = (struct cfq_data *) data;
2142         struct cfq_queue *cfqq;
2143         unsigned long flags;
2144
2145         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2146
2147         if ((cfqq = cfqd->active_queue) != NULL) {
2148                 unsigned long now = jiffies;
2149
2150                 /*
2151                  * expired
2152                  */
2153                 if (time_after(now, cfqq->slice_end))
2154                         goto expire;
2155
2156                 /*
2157                  * only expire and reinvoke request handler, if there are
2158                  * other queues with pending requests
2159                  */
2160                 if (!cfqd->busy_queues) {
2161                         cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2162                         add_timer(&cfqd->idle_slice_timer);
2163                         goto out_cont;
2164                 }
2165
2166                 /*
2167                  * not expired and it has a request pending, let it dispatch
2168                  */
2169                 if (!RB_EMPTY(&cfqq->sort_list)) {
2170                         cfq_mark_cfqq_must_dispatch(cfqq);
2171                         goto out_kick;
2172                 }
2173         }
2174 expire:
2175         cfq_slice_expired(cfqd, 0);
2176 out_kick:
2177         cfq_schedule_dispatch(cfqd);
2178 out_cont:
2179         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2180 }
2181
2182 /*
2183  * Timer running if an idle class queue is waiting for service
2184  */
2185 static void cfq_idle_class_timer(unsigned long data)
2186 {
2187         struct cfq_data *cfqd = (struct cfq_data *) data;
2188         unsigned long flags, end;
2189
2190         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2191
2192         /*
2193          * race with a non-idle queue, reset timer
2194          */
2195         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2196         if (!time_after_eq(jiffies, end)) {
2197                 cfqd->idle_class_timer.expires = end;
2198                 add_timer(&cfqd->idle_class_timer);
2199         } else
2200                 cfq_schedule_dispatch(cfqd);
2201
2202         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2203 }
2204
2205 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2206 {
2207         del_timer_sync(&cfqd->idle_slice_timer);
2208         del_timer_sync(&cfqd->idle_class_timer);
2209         blk_sync_queue(cfqd->queue);
2210 }
2211
2212 static void cfq_exit_queue(elevator_t *e)
2213 {
2214         struct cfq_data *cfqd = e->elevator_data;
2215         request_queue_t *q = cfqd->queue;
2216
2217         cfq_shutdown_timer_wq(cfqd);
2218
2219         spin_lock(&cfq_exit_lock);
2220         spin_lock_irq(q->queue_lock);
2221
2222         if (cfqd->active_queue)
2223                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2224
2225         while (!list_empty(&cfqd->cic_list)) {
2226                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2227                                                         struct cfq_io_context,
2228                                                         queue_list);
2229                 if (cic->cfqq[ASYNC]) {
2230                         cfq_put_queue(cic->cfqq[ASYNC]);
2231                         cic->cfqq[ASYNC] = NULL;
2232                 }
2233                 if (cic->cfqq[SYNC]) {
2234                         cfq_put_queue(cic->cfqq[SYNC]);
2235                         cic->cfqq[SYNC] = NULL;
2236                 }
2237                 cic->key = NULL;
2238                 list_del_init(&cic->queue_list);
2239         }
2240
2241         spin_unlock_irq(q->queue_lock);
2242         spin_unlock(&cfq_exit_lock);
2243
2244         cfq_shutdown_timer_wq(cfqd);
2245
2246         mempool_destroy(cfqd->crq_pool);
2247         kfree(cfqd->crq_hash);
2248         kfree(cfqd->cfq_hash);
2249         kfree(cfqd);
2250 }
2251
2252 static int cfq_init_queue(request_queue_t *q, elevator_t *e)
2253 {
2254         struct cfq_data *cfqd;
2255         int i;
2256
2257         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2258         if (!cfqd)
2259                 return -ENOMEM;
2260
2261         memset(cfqd, 0, sizeof(*cfqd));
2262
2263         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2264                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2265
2266         INIT_LIST_HEAD(&cfqd->busy_rr);
2267         INIT_LIST_HEAD(&cfqd->cur_rr);
2268         INIT_LIST_HEAD(&cfqd->idle_rr);
2269         INIT_LIST_HEAD(&cfqd->empty_list);
2270         INIT_LIST_HEAD(&cfqd->cic_list);
2271
2272         cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2273         if (!cfqd->crq_hash)
2274                 goto out_crqhash;
2275
2276         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2277         if (!cfqd->cfq_hash)
2278                 goto out_cfqhash;
2279
2280         cfqd->crq_pool = mempool_create_slab_pool(BLKDEV_MIN_RQ, crq_pool);
2281         if (!cfqd->crq_pool)
2282                 goto out_crqpool;
2283
2284         for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2285                 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2286         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2287                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2288
2289         e->elevator_data = cfqd;
2290
2291         cfqd->queue = q;
2292
2293         cfqd->max_queued = q->nr_requests / 4;
2294         q->nr_batching = cfq_queued;
2295
2296         init_timer(&cfqd->idle_slice_timer);
2297         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2298         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2299
2300         init_timer(&cfqd->idle_class_timer);
2301         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2302         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2303
2304         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2305
2306         cfqd->cfq_queued = cfq_queued;
2307         cfqd->cfq_quantum = cfq_quantum;
2308         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2309         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2310         cfqd->cfq_back_max = cfq_back_max;
2311         cfqd->cfq_back_penalty = cfq_back_penalty;
2312         cfqd->cfq_slice[0] = cfq_slice_async;
2313         cfqd->cfq_slice[1] = cfq_slice_sync;
2314         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2315         cfqd->cfq_slice_idle = cfq_slice_idle;
2316
2317         return 0;
2318 out_crqpool:
2319         kfree(cfqd->cfq_hash);
2320 out_cfqhash:
2321         kfree(cfqd->crq_hash);
2322 out_crqhash:
2323         kfree(cfqd);
2324         return -ENOMEM;
2325 }
2326
2327 static void cfq_slab_kill(void)
2328 {
2329         if (crq_pool)
2330                 kmem_cache_destroy(crq_pool);
2331         if (cfq_pool)
2332                 kmem_cache_destroy(cfq_pool);
2333         if (cfq_ioc_pool)
2334                 kmem_cache_destroy(cfq_ioc_pool);
2335 }
2336
2337 static int __init cfq_slab_setup(void)
2338 {
2339         crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2340                                         NULL, NULL);
2341         if (!crq_pool)
2342                 goto fail;
2343
2344         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2345                                         NULL, NULL);
2346         if (!cfq_pool)
2347                 goto fail;
2348
2349         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2350                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2351         if (!cfq_ioc_pool)
2352                 goto fail;
2353
2354         return 0;
2355 fail:
2356         cfq_slab_kill();
2357         return -ENOMEM;
2358 }
2359
2360 /*
2361  * sysfs parts below -->
2362  */
2363
2364 static ssize_t
2365 cfq_var_show(unsigned int var, char *page)
2366 {
2367         return sprintf(page, "%d\n", var);
2368 }
2369
2370 static ssize_t
2371 cfq_var_store(unsigned int *var, const char *page, size_t count)
2372 {
2373         char *p = (char *) page;
2374
2375         *var = simple_strtoul(p, &p, 10);
2376         return count;
2377 }
2378
2379 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2380 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2381 {                                                                       \
2382         struct cfq_data *cfqd = e->elevator_data;                       \
2383         unsigned int __data = __VAR;                                    \
2384         if (__CONV)                                                     \
2385                 __data = jiffies_to_msecs(__data);                      \
2386         return cfq_var_show(__data, (page));                            \
2387 }
2388 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2389 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2390 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2391 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2392 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2393 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2394 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2395 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2396 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2397 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2398 #undef SHOW_FUNCTION
2399
2400 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2401 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2402 {                                                                       \
2403         struct cfq_data *cfqd = e->elevator_data;                       \
2404         unsigned int __data;                                            \
2405         int ret = cfq_var_store(&__data, (page), count);                \
2406         if (__data < (MIN))                                             \
2407                 __data = (MIN);                                         \
2408         else if (__data > (MAX))                                        \
2409                 __data = (MAX);                                         \
2410         if (__CONV)                                                     \
2411                 *(__PTR) = msecs_to_jiffies(__data);                    \
2412         else                                                            \
2413                 *(__PTR) = __data;                                      \
2414         return ret;                                                     \
2415 }
2416 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2417 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2418 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2419 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2420 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2421 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2422 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2423 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2424 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2425 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2426 #undef STORE_FUNCTION
2427
2428 #define CFQ_ATTR(name) \
2429         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2430
2431 static struct elv_fs_entry cfq_attrs[] = {
2432         CFQ_ATTR(quantum),
2433         CFQ_ATTR(queued),
2434         CFQ_ATTR(fifo_expire_sync),
2435         CFQ_ATTR(fifo_expire_async),
2436         CFQ_ATTR(back_seek_max),
2437         CFQ_ATTR(back_seek_penalty),
2438         CFQ_ATTR(slice_sync),
2439         CFQ_ATTR(slice_async),
2440         CFQ_ATTR(slice_async_rq),
2441         CFQ_ATTR(slice_idle),
2442         __ATTR_NULL
2443 };
2444
2445 static struct elevator_type iosched_cfq = {
2446         .ops = {
2447                 .elevator_merge_fn =            cfq_merge,
2448                 .elevator_merged_fn =           cfq_merged_request,
2449                 .elevator_merge_req_fn =        cfq_merged_requests,
2450                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2451                 .elevator_add_req_fn =          cfq_insert_request,
2452                 .elevator_activate_req_fn =     cfq_activate_request,
2453                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2454                 .elevator_queue_empty_fn =      cfq_queue_empty,
2455                 .elevator_completed_req_fn =    cfq_completed_request,
2456                 .elevator_former_req_fn =       cfq_former_request,
2457                 .elevator_latter_req_fn =       cfq_latter_request,
2458                 .elevator_set_req_fn =          cfq_set_request,
2459                 .elevator_put_req_fn =          cfq_put_request,
2460                 .elevator_may_queue_fn =        cfq_may_queue,
2461                 .elevator_init_fn =             cfq_init_queue,
2462                 .elevator_exit_fn =             cfq_exit_queue,
2463                 .trim =                         cfq_trim,
2464         },
2465         .elevator_attrs =       cfq_attrs,
2466         .elevator_name =        "cfq",
2467         .elevator_owner =       THIS_MODULE,
2468 };
2469
2470 static int __init cfq_init(void)
2471 {
2472         int ret;
2473
2474         /*
2475          * could be 0 on HZ < 1000 setups
2476          */
2477         if (!cfq_slice_async)
2478                 cfq_slice_async = 1;
2479         if (!cfq_slice_idle)
2480                 cfq_slice_idle = 1;
2481
2482         if (cfq_slab_setup())
2483                 return -ENOMEM;
2484
2485         ret = elv_register(&iosched_cfq);
2486         if (ret)
2487                 cfq_slab_kill();
2488
2489         return ret;
2490 }
2491
2492 static void __exit cfq_exit(void)
2493 {
2494         DECLARE_COMPLETION(all_gone);
2495         elv_unregister(&iosched_cfq);
2496         ioc_gone = &all_gone;
2497         /* ioc_gone's update must be visible before reading ioc_count */
2498         smp_wmb();
2499         if (atomic_read(&ioc_count))
2500                 wait_for_completion(ioc_gone);
2501         synchronize_rcu();
2502         cfq_slab_kill();
2503 }
2504
2505 module_init(cfq_init);
2506 module_exit(cfq_exit);
2507
2508 MODULE_AUTHOR("Jens Axboe");
2509 MODULE_LICENSE("GPL");
2510 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");