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