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

   1 /*
   2  *  linux/drivers/block/cfq-iosched.c
   3  *
   4  *  CFQ, or complete fairness queueing, disk scheduler.
   5  *
   6  *  Based on ideas from a previously unfinished io
   7  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   8  *
   9  *  Copyright (C) 2003 Jens Axboe <axboe@suse.de>
  10  */
  11 #include <linux/kernel.h>
  12 #include <linux/fs.h>
  13 #include <linux/blkdev.h>
  14 #include <linux/elevator.h>
  15 #include <linux/bio.h>
  16 #include <linux/config.h>
  17 #include <linux/module.h>
  18 #include <linux/slab.h>
  19 #include <linux/init.h>
  20 #include <linux/compiler.h>
  21 #include <linux/hash.h>
  22 #include <linux/rbtree.h>
  23 #include <linux/mempool.h>
  24 #include <linux/ioprio.h>
  25 #include <linux/writeback.h>
  26
  27 /*
  28  * tunables
  29  */
  30 static int cfq_quantum = 4;             /* max queue in one round of service */
  31 static int cfq_queued = 8;              /* minimum rq allocate limit per-queue*/
  32 static int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  33 static int cfq_back_max = 16 * 1024;    /* maximum backwards seek, in KiB */
  34 static int cfq_back_penalty = 2;        /* penalty of a backwards seek */
  35
  36 static int cfq_slice_sync = HZ / 10;
  37 static int cfq_slice_async = HZ / 25;
  38 static int cfq_slice_async_rq = 2;
  39 static int cfq_slice_idle = HZ / 100;
  40
  41 #define CFQ_IDLE_GRACE          (HZ / 10)
  42 #define CFQ_SLICE_SCALE         (5)
  43
  44 #define CFQ_KEY_ASYNC           (0)
  45 #define CFQ_KEY_ANY             (0xffff)
  46
  47 /*
  48  * disable queueing at the driver/hardware level
  49  */
  50 static int cfq_max_depth = 2;
  51
  52 /*
  53  * for the hash of cfqq inside the cfqd
  54  */
  55 #define CFQ_QHASH_SHIFT         6
  56 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
  57 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
  58
  59 /*
  60  * for the hash of crq inside the cfqq
  61  */
  62 #define CFQ_MHASH_SHIFT         6
  63 #define CFQ_MHASH_BLOCK(sec)    ((sec) >> 3)
  64 #define CFQ_MHASH_ENTRIES       (1 << CFQ_MHASH_SHIFT)
  65 #define CFQ_MHASH_FN(sec)       hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
  66 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
  67 #define list_entry_hash(ptr)    hlist_entry((ptr), struct cfq_rq, hash)
  68
  69 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
  70 #define list_entry_fifo(ptr)    list_entry((ptr), struct request, queuelist)
  71
  72 #define RQ_DATA(rq)             (rq)->elevator_private
  73
  74 /*
  75  * rb-tree defines
  76  */
  77 #define RB_NONE                 (2)
  78 #define RB_EMPTY(node)          ((node)->rb_node == NULL)
  79 #define RB_CLEAR_COLOR(node)    (node)->rb_color = RB_NONE
  80 #define RB_CLEAR(node)          do {    \
  81         (node)->rb_parent = NULL;       \
  82         RB_CLEAR_COLOR((node));         \
  83         (node)->rb_right = NULL;        \
  84         (node)->rb_left = NULL;         \
  85 } while (0)
  86 #define RB_CLEAR_ROOT(root)     ((root)->rb_node = NULL)
  87 #define rb_entry_crq(node)      rb_entry((node), struct cfq_rq, rb_node)
  88 #define rq_rb_key(rq)           (rq)->sector
  89
  90 static kmem_cache_t *crq_pool;
  91 static kmem_cache_t *cfq_pool;
  92 static kmem_cache_t *cfq_ioc_pool;
  93
  94 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  95 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  96 #define cfq_class_be(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
  97 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  98
  99 #define ASYNC                   (0)
 100 #define SYNC                    (1)
 101
 102 #define cfq_cfqq_dispatched(cfqq)       \
 103         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
 104
 105 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
 106
 107 #define cfq_cfqq_sync(cfqq)             \
 108         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
 109
 110 /*
 111  * Per block device queue structure
 112  */
 113 struct cfq_data {
 114         atomic_t ref;
 115         request_queue_t *queue;
 116
 117         /*
 118          * rr list of queues with requests and the count of them
 119          */
 120         struct list_head rr_list[CFQ_PRIO_LISTS];
 121         struct list_head busy_rr;
 122         struct list_head cur_rr;
 123         struct list_head idle_rr;
 124         unsigned int busy_queues;
 125
 126         /*
 127          * non-ordered list of empty cfqq's
 128          */
 129         struct list_head empty_list;
 130
 131         /*
 132          * cfqq lookup hash
 133          */
 134         struct hlist_head *cfq_hash;
 135
 136         /*
 137          * global crq hash for all queues
 138          */
 139         struct hlist_head *crq_hash;
 140
 141         unsigned int max_queued;
 142
 143         mempool_t *crq_pool;
 144
 145         int rq_in_driver;
 146
 147         /*
 148          * schedule slice state info
 149          */
 150         /*
 151          * idle window management
 152          */
 153         struct timer_list idle_slice_timer;
 154         struct work_struct unplug_work;
 155
 156         struct cfq_queue *active_queue;
 157         struct cfq_io_context *active_cic;
 158         int cur_prio, cur_end_prio;
 159         unsigned int dispatch_slice;
 160
 161         struct timer_list idle_class_timer;
 162
 163         sector_t last_sector;
 164         unsigned long last_end_request;
 165
 166         unsigned int rq_starved;
 167
 168         /*
 169          * tunables, see top of file
 170          */
 171         unsigned int cfq_quantum;
 172         unsigned int cfq_queued;
 173         unsigned int cfq_fifo_expire[2];
 174         unsigned int cfq_back_penalty;
 175         unsigned int cfq_back_max;
 176         unsigned int cfq_slice[2];
 177         unsigned int cfq_slice_async_rq;
 178         unsigned int cfq_slice_idle;
 179         unsigned int cfq_max_depth;
 180 };
 181
 182 /*
 183  * Per process-grouping structure
 184  */
 185 struct cfq_queue {
 186         /* reference count */
 187         atomic_t ref;
 188         /* parent cfq_data */
 189         struct cfq_data *cfqd;
 190         /* cfqq lookup hash */
 191         struct hlist_node cfq_hash;
 192         /* hash key */
 193         unsigned int key;
 194         /* on either rr or empty list of cfqd */
 195         struct list_head cfq_list;
 196         /* sorted list of pending requests */
 197         struct rb_root sort_list;
 198         /* if fifo isn't expired, next request to serve */
 199         struct cfq_rq *next_crq;
 200         /* requests queued in sort_list */
 201         int queued[2];
 202         /* currently allocated requests */
 203         int allocated[2];
 204         /* fifo list of requests in sort_list */
 205         struct list_head fifo;
 206
 207         unsigned long slice_start;
 208         unsigned long slice_end;
 209         unsigned long slice_left;
 210         unsigned long service_last;
 211
 212         /* number of requests that are on the dispatch list */
 213         int on_dispatch[2];
 214
 215         /* io prio of this group */
 216         unsigned short ioprio, org_ioprio;
 217         unsigned short ioprio_class, org_ioprio_class;
 218
 219         /* various state flags, see below */
 220         unsigned int flags;
 221 };
 222
 223 struct cfq_rq {
 224         struct rb_node rb_node;
 225         sector_t rb_key;
 226         struct request *request;
 227         struct hlist_node hash;
 228
 229         struct cfq_queue *cfq_queue;
 230         struct cfq_io_context *io_context;
 231
 232         unsigned int crq_flags;
 233 };
 234
 235 enum cfqq_state_flags {
 236         CFQ_CFQQ_FLAG_on_rr = 0,
 237         CFQ_CFQQ_FLAG_wait_request,
 238         CFQ_CFQQ_FLAG_must_alloc,
 239         CFQ_CFQQ_FLAG_must_alloc_slice,
 240         CFQ_CFQQ_FLAG_must_dispatch,
 241         CFQ_CFQQ_FLAG_fifo_expire,
 242         CFQ_CFQQ_FLAG_idle_window,
 243         CFQ_CFQQ_FLAG_prio_changed,
 244         CFQ_CFQQ_FLAG_expired,
 245 };
 246
 247 #define CFQ_CFQQ_FNS(name)                                              \
 248 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 249 {                                                                       \
 250         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 251 }                                                                       \
 252 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 253 {                                                                       \
 254         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 255 }                                                                       \
 256 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 257 {                                                                       \
 258         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 259 }
 260
 261 CFQ_CFQQ_FNS(on_rr);
 262 CFQ_CFQQ_FNS(wait_request);
 263 CFQ_CFQQ_FNS(must_alloc);
 264 CFQ_CFQQ_FNS(must_alloc_slice);
 265 CFQ_CFQQ_FNS(must_dispatch);
 266 CFQ_CFQQ_FNS(fifo_expire);
 267 CFQ_CFQQ_FNS(idle_window);
 268 CFQ_CFQQ_FNS(prio_changed);
 269 CFQ_CFQQ_FNS(expired);
 270 #undef CFQ_CFQQ_FNS
 271
 272 enum cfq_rq_state_flags {
 273         CFQ_CRQ_FLAG_is_sync = 0,
 274 };
 275
 276 #define CFQ_CRQ_FNS(name)                                               \
 277 static inline void cfq_mark_crq_##name(struct cfq_rq *crq)              \
 278 {                                                                       \
 279         crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name);                   \
 280 }                                                                       \
 281 static inline void cfq_clear_crq_##name(struct cfq_rq *crq)             \
 282 {                                                                       \
 283         crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name);                  \
 284 }                                                                       \
 285 static inline int cfq_crq_##name(const struct cfq_rq *crq)              \
 286 {                                                                       \
 287         return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0;      \
 288 }
 289
 290 CFQ_CRQ_FNS(is_sync);
 291 #undef CFQ_CRQ_FNS
 292
 293 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 294 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
 295 static void cfq_put_cfqd(struct cfq_data *cfqd);
 296
 297 #define process_sync(tsk)       ((tsk)->flags & PF_SYNCWRITE)
 298
 299 /*
 300  * lots of deadline iosched dupes, can be abstracted later...
 301  */
 302 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
 303 {
 304         hlist_del_init(&crq->hash);
 305 }
 306
 307 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
 308 {
 309         const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
 310
 311         hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
 312 }
 313
 314 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
 315 {
 316         struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
 317         struct hlist_node *entry, *next;
 318
 319         hlist_for_each_safe(entry, next, hash_list) {
 320                 struct cfq_rq *crq = list_entry_hash(entry);
 321                 struct request *__rq = crq->request;
 322
 323                 if (!rq_mergeable(__rq)) {
 324                         cfq_del_crq_hash(crq);
 325                         continue;
 326                 }
 327
 328                 if (rq_hash_key(__rq) == offset)
 329                         return __rq;
 330         }
 331
 332         return NULL;
 333 }
 334
 335 /*
 336  * scheduler run of queue, if there are requests pending and no one in the
 337  * driver that will restart queueing
 338  */
 339 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 340 {
 341         if (!cfqd->rq_in_driver && cfqd->busy_queues)
 342                 kblockd_schedule_work(&cfqd->unplug_work);
 343 }
 344
 345 static int cfq_queue_empty(request_queue_t *q)
 346 {
 347         struct cfq_data *cfqd = q->elevator->elevator_data;
 348
 349         return !cfqd->busy_queues;
 350 }
 351
 352 /*
 353  * Lifted from AS - choose which of crq1 and crq2 that is best served now.
 354  * We choose the request that is closest to the head right now. Distance
 355  * behind the head are penalized and only allowed to a certain extent.
 356  */
 357 static struct cfq_rq *
 358 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
 359 {
 360         sector_t last, s1, s2, d1 = 0, d2 = 0;
 361         int r1_wrap = 0, r2_wrap = 0;   /* requests are behind the disk head */
 362         unsigned long back_max;
 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                 r1_wrap = 1;
 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                 r2_wrap = 1;
 402
 403         /* Found required data */
 404         if (!r1_wrap && r2_wrap)
 405                 return crq1;
 406         else if (!r2_wrap && r1_wrap)
 407                 return crq2;
 408         else if (r1_wrap && r2_wrap) {
 409                 /* both behind the head */
 410                 if (s1 <= s2)
 411                         return crq1;
 412                 else
 413                         return crq2;
 414         }
 415
 416         /* Both requests in front of the head */
 417         if (d1 < d2)
 418                 return crq1;
 419         else if (d2 < d1)
 420                 return crq2;
 421         else {
 422                 if (s1 >= s2)
 423                         return crq1;
 424                 else
 425                         return crq2;
 426         }
 427 }
 428
 429 /*
 430  * would be nice to take fifo expire time into account as well
 431  */
 432 static struct cfq_rq *
 433 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 434                   struct cfq_rq *last)
 435 {
 436         struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
 437         struct rb_node *rbnext, *rbprev;
 438
 439         if (!(rbnext = rb_next(&last->rb_node))) {
 440                 rbnext = rb_first(&cfqq->sort_list);
 441                 if (rbnext == &last->rb_node)
 442                         rbnext = NULL;
 443         }
 444
 445         rbprev = rb_prev(&last->rb_node);
 446
 447         if (rbprev)
 448                 crq_prev = rb_entry_crq(rbprev);
 449         if (rbnext)
 450                 crq_next = rb_entry_crq(rbnext);
 451
 452         return cfq_choose_req(cfqd, crq_next, crq_prev);
 453 }
 454
 455 static void cfq_update_next_crq(struct cfq_rq *crq)
 456 {
 457         struct cfq_queue *cfqq = crq->cfq_queue;
 458
 459         if (cfqq->next_crq == crq)
 460                 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
 461 }
 462
 463 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 464 {
 465         struct cfq_data *cfqd = cfqq->cfqd;
 466         struct list_head *list, *entry;
 467
 468         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 469
 470         list_del(&cfqq->cfq_list);
 471
 472         if (cfq_class_rt(cfqq))
 473                 list = &cfqd->cur_rr;
 474         else if (cfq_class_idle(cfqq))
 475                 list = &cfqd->idle_rr;
 476         else {
 477                 /*
 478                  * if cfqq has requests in flight, don't allow it to be
 479                  * found in cfq_set_active_queue before it has finished them.
 480                  * this is done to increase fairness between a process that
 481                  * has lots of io pending vs one that only generates one
 482                  * sporadically or synchronously
 483                  */
 484                 if (cfq_cfqq_dispatched(cfqq))
 485                         list = &cfqd->busy_rr;
 486                 else
 487                         list = &cfqd->rr_list[cfqq->ioprio];
 488         }
 489
 490         /*
 491          * if queue was preempted, just add to front to be fair. busy_rr
 492          * isn't sorted.
 493          */
 494         if (preempted || list == &cfqd->busy_rr) {
 495                 list_add(&cfqq->cfq_list, list);
 496                 return;
 497         }
 498
 499         /*
 500          * sort by when queue was last serviced
 501          */
 502         entry = list;
 503         while ((entry = entry->prev) != list) {
 504                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
 505
 506                 if (!__cfqq->service_last)
 507                         break;
 508                 if (time_before(__cfqq->service_last, cfqq->service_last))
 509                         break;
 510         }
 511
 512         list_add(&cfqq->cfq_list, entry);
 513 }
 514
 515 /*
 516  * add to busy list of queues for service, trying to be fair in ordering
 517  * the pending list according to last request service
 518  */
 519 static inline void
 520 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 521 {
 522         BUG_ON(cfq_cfqq_on_rr(cfqq));
 523         cfq_mark_cfqq_on_rr(cfqq);
 524         cfqd->busy_queues++;
 525
 526         cfq_resort_rr_list(cfqq, 0);
 527 }
 528
 529 static inline void
 530 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 531 {
 532         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 533         cfq_clear_cfqq_on_rr(cfqq);
 534         list_move(&cfqq->cfq_list, &cfqd->empty_list);
 535
 536         BUG_ON(!cfqd->busy_queues);
 537         cfqd->busy_queues--;
 538 }
 539
 540 /*
 541  * rb tree support functions
 542  */
 543 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
 544 {
 545         struct cfq_queue *cfqq = crq->cfq_queue;
 546         struct cfq_data *cfqd = cfqq->cfqd;
 547         const int sync = cfq_crq_is_sync(crq);
 548
 549         BUG_ON(!cfqq->queued[sync]);
 550         cfqq->queued[sync]--;
 551
 552         cfq_update_next_crq(crq);
 553
 554         rb_erase(&crq->rb_node, &cfqq->sort_list);
 555         RB_CLEAR_COLOR(&crq->rb_node);
 556
 557         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
 558                 cfq_del_cfqq_rr(cfqd, cfqq);
 559 }
 560
 561 static struct cfq_rq *
 562 __cfq_add_crq_rb(struct cfq_rq *crq)
 563 {
 564         struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
 565         struct rb_node *parent = NULL;
 566         struct cfq_rq *__crq;
 567
 568         while (*p) {
 569                 parent = *p;
 570                 __crq = rb_entry_crq(parent);
 571
 572                 if (crq->rb_key < __crq->rb_key)
 573                         p = &(*p)->rb_left;
 574                 else if (crq->rb_key > __crq->rb_key)
 575                         p = &(*p)->rb_right;
 576                 else
 577                         return __crq;
 578         }
 579
 580         rb_link_node(&crq->rb_node, parent, p);
 581         return NULL;
 582 }
 583
 584 static void cfq_add_crq_rb(struct cfq_rq *crq)
 585 {
 586         struct cfq_queue *cfqq = crq->cfq_queue;
 587         struct cfq_data *cfqd = cfqq->cfqd;
 588         struct request *rq = crq->request;
 589         struct cfq_rq *__alias;
 590
 591         crq->rb_key = rq_rb_key(rq);
 592         cfqq->queued[cfq_crq_is_sync(crq)]++;
 593
 594         /*
 595          * looks a little odd, but the first insert might return an alias.
 596          * if that happens, put the alias on the dispatch list
 597          */
 598         while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
 599                 cfq_dispatch_insert(cfqd->queue, __alias);
 600
 601         rb_insert_color(&crq->rb_node, &cfqq->sort_list);
 602
 603         if (!cfq_cfqq_on_rr(cfqq))
 604                 cfq_add_cfqq_rr(cfqd, cfqq);
 605
 606         /*
 607          * check if this request is a better next-serve candidate
 608          */
 609         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
 610 }
 611
 612 static inline void
 613 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
 614 {
 615         rb_erase(&crq->rb_node, &cfqq->sort_list);
 616         cfqq->queued[cfq_crq_is_sync(crq)]--;
 617
 618         cfq_add_crq_rb(crq);
 619 }
 620
 621 static struct request *cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector)
 622
 623 {
 624         struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, current->pid, CFQ_KEY_ANY);
 625         struct rb_node *n;
 626
 627         if (!cfqq)
 628                 goto out;
 629
 630         n = cfqq->sort_list.rb_node;
 631         while (n) {
 632                 struct cfq_rq *crq = rb_entry_crq(n);
 633
 634                 if (sector < crq->rb_key)
 635                         n = n->rb_left;
 636                 else if (sector > crq->rb_key)
 637                         n = n->rb_right;
 638                 else
 639                         return crq->request;
 640         }
 641
 642 out:
 643         return NULL;
 644 }
 645
 646 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 647 {
 648         struct cfq_data *cfqd = q->elevator->elevator_data;
 649
 650         cfqd->rq_in_driver++;
 651 }
 652
 653 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 654 {
 655         struct cfq_data *cfqd = q->elevator->elevator_data;
 656
 657         WARN_ON(!cfqd->rq_in_driver);
 658         cfqd->rq_in_driver--;
 659 }
 660
 661 static void cfq_remove_request(struct request *rq)
 662 {
 663         struct cfq_rq *crq = RQ_DATA(rq);
 664
 665         list_del_init(&rq->queuelist);
 666         cfq_del_crq_rb(crq);
 667         cfq_del_crq_hash(crq);
 668 }
 669
 670 static int
 671 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 672 {
 673         struct cfq_data *cfqd = q->elevator->elevator_data;
 674         struct request *__rq;
 675         int ret;
 676
 677         __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
 678         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 679                 ret = ELEVATOR_BACK_MERGE;
 680                 goto out;
 681         }
 682
 683         __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio));
 684         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 685                 ret = ELEVATOR_FRONT_MERGE;
 686                 goto out;
 687         }
 688
 689         return ELEVATOR_NO_MERGE;
 690 out:
 691         *req = __rq;
 692         return ret;
 693 }
 694
 695 static void cfq_merged_request(request_queue_t *q, struct request *req)
 696 {
 697         struct cfq_data *cfqd = q->elevator->elevator_data;
 698         struct cfq_rq *crq = RQ_DATA(req);
 699
 700         cfq_del_crq_hash(crq);
 701         cfq_add_crq_hash(cfqd, crq);
 702
 703         if (rq_rb_key(req) != crq->rb_key) {
 704                 struct cfq_queue *cfqq = crq->cfq_queue;
 705
 706                 cfq_update_next_crq(crq);
 707                 cfq_reposition_crq_rb(cfqq, crq);
 708         }
 709 }
 710
 711 static void
 712 cfq_merged_requests(request_queue_t *q, struct request *rq,
 713                     struct request *next)
 714 {
 715         cfq_merged_request(q, rq);
 716
 717         /*
 718          * reposition in fifo if next is older than rq
 719          */
 720         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 721             time_before(next->start_time, rq->start_time))
 722                 list_move(&rq->queuelist, &next->queuelist);
 723
 724         cfq_remove_request(next);
 725 }
 726
 727 static inline void
 728 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 729 {
 730         if (cfqq) {
 731                 /*
 732                  * stop potential idle class queues waiting service
 733                  */
 734                 del_timer(&cfqd->idle_class_timer);
 735
 736                 cfqq->slice_start = jiffies;
 737                 cfqq->slice_end = 0;
 738                 cfqq->slice_left = 0;
 739                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 740                 cfq_clear_cfqq_fifo_expire(cfqq);
 741                 cfq_clear_cfqq_expired(cfqq);
 742         }
 743
 744         cfqd->active_queue = cfqq;
 745 }
 746
 747 /*
 748  * 0
 749  * 0,1
 750  * 0,1,2
 751  * 0,1,2,3
 752  * 0,1,2,3,4
 753  * 0,1,2,3,4,5
 754  * 0,1,2,3,4,5,6
 755  * 0,1,2,3,4,5,6,7
 756  */
 757 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 758 {
 759         int prio, wrap;
 760
 761         prio = -1;
 762         wrap = 0;
 763         do {
 764                 int p;
 765
 766                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 767                         if (!list_empty(&cfqd->rr_list[p])) {
 768                                 prio = p;
 769                                 break;
 770                         }
 771                 }
 772
 773                 if (prio != -1)
 774                         break;
 775                 cfqd->cur_prio = 0;
 776                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 777                         cfqd->cur_end_prio = 0;
 778                         if (wrap)
 779                                 break;
 780                         wrap = 1;
 781                 }
 782         } while (1);
 783
 784         if (unlikely(prio == -1))
 785                 return -1;
 786
 787         BUG_ON(prio >= CFQ_PRIO_LISTS);
 788
 789         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 790
 791         cfqd->cur_prio = prio + 1;
 792         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 793                 cfqd->cur_end_prio = cfqd->cur_prio;
 794                 cfqd->cur_prio = 0;
 795         }
 796         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 797                 cfqd->cur_prio = 0;
 798                 cfqd->cur_end_prio = 0;
 799         }
 800
 801         return prio;
 802 }
 803
 804 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 805 {
 806         struct cfq_queue *cfqq;
 807
 808         /*
 809          * if current queue is expired but not done with its requests yet,
 810          * wait for that to happen
 811          */
 812         if ((cfqq = cfqd->active_queue) != NULL) {
 813                 if (cfq_cfqq_expired(cfqq) && cfq_cfqq_dispatched(cfqq))
 814                         return NULL;
 815         }
 816
 817         /*
 818          * if current list is non-empty, grab first entry. if it is empty,
 819          * get next prio level and grab first entry then if any are spliced
 820          */
 821         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
 822                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 823
 824         /*
 825          * if we have idle queues and no rt or be queues had pending
 826          * requests, either allow immediate service if the grace period
 827          * has passed or arm the idle grace timer
 828          */
 829         if (!cfqq && !list_empty(&cfqd->idle_rr)) {
 830                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 831
 832                 if (time_after_eq(jiffies, end))
 833                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 834                 else
 835                         mod_timer(&cfqd->idle_class_timer, end);
 836         }
 837
 838         __cfq_set_active_queue(cfqd, cfqq);
 839         return cfqq;
 840 }
 841
 842 /*
 843  * current cfqq expired its slice (or was too idle), select new one
 844  */
 845 static void
 846 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 847                     int preempted)
 848 {
 849         unsigned long now = jiffies;
 850
 851         if (cfq_cfqq_wait_request(cfqq))
 852                 del_timer(&cfqd->idle_slice_timer);
 853
 854         if (!preempted && !cfq_cfqq_dispatched(cfqq))
 855                 cfqq->service_last = now;
 856
 857         cfq_clear_cfqq_must_dispatch(cfqq);
 858         cfq_clear_cfqq_wait_request(cfqq);
 859
 860         /*
 861          * store what was left of this slice, if the queue idled out
 862          * or was preempted
 863          */
 864         if (time_after(now, cfqq->slice_end))
 865                 cfqq->slice_left = now - cfqq->slice_end;
 866         else
 867                 cfqq->slice_left = 0;
 868
 869         if (cfq_cfqq_on_rr(cfqq))
 870                 cfq_resort_rr_list(cfqq, preempted);
 871
 872         if (cfqq == cfqd->active_queue)
 873                 cfqd->active_queue = NULL;
 874
 875         if (cfqd->active_cic) {
 876                 put_io_context(cfqd->active_cic->ioc);
 877                 cfqd->active_cic = NULL;
 878         }
 879
 880         cfqd->dispatch_slice = 0;
 881 }
 882
 883 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 884 {
 885         struct cfq_queue *cfqq = cfqd->active_queue;
 886
 887         if (cfqq) {
 888                 /*
 889                  * use deferred expiry, if there are requests in progress as
 890                  * not to disturb the slice of the next queue
 891                  */
 892                 if (cfq_cfqq_dispatched(cfqq))
 893                         cfq_mark_cfqq_expired(cfqq);
 894                 else
 895                         __cfq_slice_expired(cfqd, cfqq, preempted);
 896         }
 897 }
 898
 899 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 900
 901 {
 902         WARN_ON(!RB_EMPTY(&cfqq->sort_list));
 903         WARN_ON(cfqq != cfqd->active_queue);
 904
 905         /*
 906          * idle is disabled, either manually or by past process history
 907          */
 908         if (!cfqd->cfq_slice_idle)
 909                 return 0;
 910         if (!cfq_cfqq_idle_window(cfqq))
 911                 return 0;
 912         /*
 913          * task has exited, don't wait
 914          */
 915         if (cfqd->active_cic && !cfqd->active_cic->ioc->task)
 916                 return 0;
 917
 918         cfq_mark_cfqq_must_dispatch(cfqq);
 919         cfq_mark_cfqq_wait_request(cfqq);
 920
 921         if (!timer_pending(&cfqd->idle_slice_timer)) {
 922                 unsigned long slice_left = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 923
 924                 cfqd->idle_slice_timer.expires = jiffies + slice_left;
 925                 add_timer(&cfqd->idle_slice_timer);
 926         }
 927
 928         return 1;
 929 }
 930
 931 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
 932 {
 933         struct cfq_data *cfqd = q->elevator->elevator_data;
 934         struct cfq_queue *cfqq = crq->cfq_queue;
 935
 936         cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
 937         cfq_remove_request(crq->request);
 938         cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
 939         elv_dispatch_sort(q, crq->request);
 940 }
 941
 942 /*
 943  * return expired entry, or NULL to just start from scratch in rbtree
 944  */
 945 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
 946 {
 947         struct cfq_data *cfqd = cfqq->cfqd;
 948         struct request *rq;
 949         struct cfq_rq *crq;
 950
 951         if (cfq_cfqq_fifo_expire(cfqq))
 952                 return NULL;
 953
 954         if (!list_empty(&cfqq->fifo)) {
 955                 int fifo = cfq_cfqq_class_sync(cfqq);
 956
 957                 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
 958                 rq = crq->request;
 959                 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
 960                         cfq_mark_cfqq_fifo_expire(cfqq);
 961                         return crq;
 962                 }
 963         }
 964
 965         return NULL;
 966 }
 967
 968 /*
 969  * Scale schedule slice based on io priority. Use the sync time slice only
 970  * if a queue is marked sync and has sync io queued. A sync queue with async
 971  * io only, should not get full sync slice length.
 972  */
 973 static inline int
 974 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 975 {
 976         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 977
 978         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 979
 980         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
 981 }
 982
 983 static inline void
 984 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 985 {
 986         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 987 }
 988
 989 static inline int
 990 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 991 {
 992         const int base_rq = cfqd->cfq_slice_async_rq;
 993
 994         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 995
 996         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
 997 }
 998
 999 /*
1000  * get next queue for service
1001  */
1002 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd, int force)
1003 {
1004         unsigned long now = jiffies;
1005         struct cfq_queue *cfqq;
1006
1007         cfqq = cfqd->active_queue;
1008         if (!cfqq)
1009                 goto new_queue;
1010
1011         if (cfq_cfqq_expired(cfqq))
1012                 goto new_queue;
1013
1014         /*
1015          * slice has expired
1016          */
1017         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
1018                 goto expire;
1019
1020         /*
1021          * if queue has requests, dispatch one. if not, check if
1022          * enough slice is left to wait for one
1023          */
1024         if (!RB_EMPTY(&cfqq->sort_list))
1025                 goto keep_queue;
1026         else if (!force && cfq_cfqq_class_sync(cfqq) &&
1027                  time_before(now, cfqq->slice_end)) {
1028                 if (cfq_arm_slice_timer(cfqd, cfqq))
1029                         return NULL;
1030         }
1031
1032 expire:
1033         cfq_slice_expired(cfqd, 0);
1034 new_queue:
1035         cfqq = cfq_set_active_queue(cfqd);
1036 keep_queue:
1037         return cfqq;
1038 }
1039
1040 static int
1041 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1042                         int max_dispatch)
1043 {
1044         int dispatched = 0;
1045
1046         BUG_ON(RB_EMPTY(&cfqq->sort_list));
1047
1048         do {
1049                 struct cfq_rq *crq;
1050
1051                 /*
1052                  * follow expired path, else get first next available
1053                  */
1054                 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1055                         crq = cfqq->next_crq;
1056
1057                 /*
1058                  * finally, insert request into driver dispatch list
1059                  */
1060                 cfq_dispatch_insert(cfqd->queue, crq);
1061
1062                 cfqd->dispatch_slice++;
1063                 dispatched++;
1064
1065                 if (!cfqd->active_cic) {
1066                         atomic_inc(&crq->io_context->ioc->refcount);
1067                         cfqd->active_cic = crq->io_context;
1068                 }
1069
1070                 if (RB_EMPTY(&cfqq->sort_list))
1071                         break;
1072
1073         } while (dispatched < max_dispatch);
1074
1075         /*
1076          * if slice end isn't set yet, set it. if at least one request was
1077          * sync, use the sync time slice value
1078          */
1079         if (!cfqq->slice_end)
1080                 cfq_set_prio_slice(cfqd, cfqq);
1081
1082         /*
1083          * expire an async queue immediately if it has used up its slice. idle
1084          * queue always expire after 1 dispatch round.
1085          */
1086         if ((!cfq_cfqq_sync(cfqq) &&
1087             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1088             cfq_class_idle(cfqq))
1089                 cfq_slice_expired(cfqd, 0);
1090
1091         return dispatched;
1092 }
1093
1094 static int
1095 cfq_dispatch_requests(request_queue_t *q, int force)
1096 {
1097         struct cfq_data *cfqd = q->elevator->elevator_data;
1098         struct cfq_queue *cfqq;
1099
1100         if (!cfqd->busy_queues)
1101                 return 0;
1102
1103         cfqq = cfq_select_queue(cfqd, force);
1104         if (cfqq) {
1105                 int max_dispatch;
1106
1107                 /*
1108                  * if idle window is disabled, allow queue buildup
1109                  */
1110                 if (!cfq_cfqq_idle_window(cfqq) &&
1111                     cfqd->rq_in_driver >= cfqd->cfq_max_depth)
1112                         return 0;
1113
1114                 cfq_clear_cfqq_must_dispatch(cfqq);
1115                 cfq_clear_cfqq_wait_request(cfqq);
1116                 del_timer(&cfqd->idle_slice_timer);
1117
1118                 if (!force) {
1119                         max_dispatch = cfqd->cfq_quantum;
1120                         if (cfq_class_idle(cfqq))
1121                                 max_dispatch = 1;
1122                 } else
1123                         max_dispatch = INT_MAX;
1124
1125                 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1126         }
1127
1128         return 0;
1129 }
1130
1131 /*
1132  * task holds one reference to the queue, dropped when task exits. each crq
1133  * in-flight on this queue also holds a reference, dropped when crq is freed.
1134  *
1135  * queue lock must be held here.
1136  */
1137 static void cfq_put_queue(struct cfq_queue *cfqq)
1138 {
1139         struct cfq_data *cfqd = cfqq->cfqd;
1140
1141         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1142
1143         if (!atomic_dec_and_test(&cfqq->ref))
1144                 return;
1145
1146         BUG_ON(rb_first(&cfqq->sort_list));
1147         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1148         BUG_ON(cfq_cfqq_on_rr(cfqq));
1149
1150         if (unlikely(cfqd->active_queue == cfqq)) {
1151                 __cfq_slice_expired(cfqd, cfqq, 0);
1152                 cfq_schedule_dispatch(cfqd);
1153         }
1154
1155         cfq_put_cfqd(cfqq->cfqd);
1156
1157         /*
1158          * it's on the empty list and still hashed
1159          */
1160         list_del(&cfqq->cfq_list);
1161         hlist_del(&cfqq->cfq_hash);
1162         kmem_cache_free(cfq_pool, cfqq);
1163 }
1164
1165 static inline struct cfq_queue *
1166 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1167                     const int hashval)
1168 {
1169         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1170         struct hlist_node *entry, *next;
1171
1172         hlist_for_each_safe(entry, next, hash_list) {
1173                 struct cfq_queue *__cfqq = list_entry_qhash(entry);
1174                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->ioprio_class, __cfqq->ioprio);
1175
1176                 if (__cfqq->key == key && (__p == prio || prio == CFQ_KEY_ANY))
1177                         return __cfqq;
1178         }
1179
1180         return NULL;
1181 }
1182
1183 static struct cfq_queue *
1184 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1185 {
1186         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1187 }
1188
1189 static void cfq_free_io_context(struct cfq_io_context *cic)
1190 {
1191         struct cfq_io_context *__cic;
1192         struct list_head *entry, *next;
1193
1194         list_for_each_safe(entry, next, &cic->list) {
1195                 __cic = list_entry(entry, struct cfq_io_context, list);
1196                 kmem_cache_free(cfq_ioc_pool, __cic);
1197         }
1198
1199         kmem_cache_free(cfq_ioc_pool, cic);
1200 }
1201
1202 /*
1203  * Called with interrupts disabled
1204  */
1205 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1206 {
1207         struct cfq_data *cfqd = cic->cfqq->cfqd;
1208         request_queue_t *q = cfqd->queue;
1209
1210         WARN_ON(!irqs_disabled());
1211
1212         spin_lock(q->queue_lock);
1213
1214         if (unlikely(cic->cfqq == cfqd->active_queue)) {
1215                 __cfq_slice_expired(cfqd, cic->cfqq, 0);
1216                 cfq_schedule_dispatch(cfqd);
1217         }
1218
1219         cfq_put_queue(cic->cfqq);
1220         cic->cfqq = NULL;
1221         spin_unlock(q->queue_lock);
1222 }
1223
1224 /*
1225  * Another task may update the task cic list, if it is doing a queue lookup
1226  * on its behalf. cfq_cic_lock excludes such concurrent updates
1227  */
1228 static void cfq_exit_io_context(struct cfq_io_context *cic)
1229 {
1230         struct cfq_io_context *__cic;
1231         struct list_head *entry;
1232         unsigned long flags;
1233
1234         local_irq_save(flags);
1235
1236         /*
1237          * put the reference this task is holding to the various queues
1238          */
1239         list_for_each(entry, &cic->list) {
1240                 __cic = list_entry(entry, struct cfq_io_context, list);
1241                 cfq_exit_single_io_context(__cic);
1242         }
1243
1244         cfq_exit_single_io_context(cic);
1245         local_irq_restore(flags);
1246 }
1247
1248 static struct cfq_io_context *
1249 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1250 {
1251         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1252
1253         if (cic) {
1254                 INIT_LIST_HEAD(&cic->list);
1255                 cic->cfqq = NULL;
1256                 cic->key = NULL;
1257                 cic->last_end_request = jiffies;
1258                 cic->ttime_total = 0;
1259                 cic->ttime_samples = 0;
1260                 cic->ttime_mean = 0;
1261                 cic->dtor = cfq_free_io_context;
1262                 cic->exit = cfq_exit_io_context;
1263         }
1264
1265         return cic;
1266 }
1267
1268 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1269 {
1270         struct task_struct *tsk = current;
1271         int ioprio_class;
1272
1273         if (!cfq_cfqq_prio_changed(cfqq))
1274                 return;
1275
1276         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1277         switch (ioprio_class) {
1278                 default:
1279                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1280                 case IOPRIO_CLASS_NONE:
1281                         /*
1282                          * no prio set, place us in the middle of the BE classes
1283                          */
1284                         cfqq->ioprio = task_nice_ioprio(tsk);
1285                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1286                         break;
1287                 case IOPRIO_CLASS_RT:
1288                         cfqq->ioprio = task_ioprio(tsk);
1289                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1290                         break;
1291                 case IOPRIO_CLASS_BE:
1292                         cfqq->ioprio = task_ioprio(tsk);
1293                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1294                         break;
1295                 case IOPRIO_CLASS_IDLE:
1296                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1297                         cfqq->ioprio = 7;
1298                         cfq_clear_cfqq_idle_window(cfqq);
1299                         break;
1300         }
1301
1302         /*
1303          * keep track of original prio settings in case we have to temporarily
1304          * elevate the priority of this queue
1305          */
1306         cfqq->org_ioprio = cfqq->ioprio;
1307         cfqq->org_ioprio_class = cfqq->ioprio_class;
1308
1309         if (cfq_cfqq_on_rr(cfqq))
1310                 cfq_resort_rr_list(cfqq, 0);
1311
1312         cfq_clear_cfqq_prio_changed(cfqq);
1313 }
1314
1315 static inline void changed_ioprio(struct cfq_queue *cfqq)
1316 {
1317         if (cfqq) {
1318                 struct cfq_data *cfqd = cfqq->cfqd;
1319
1320                 spin_lock(cfqd->queue->queue_lock);
1321                 cfq_mark_cfqq_prio_changed(cfqq);
1322                 cfq_init_prio_data(cfqq);
1323                 spin_unlock(cfqd->queue->queue_lock);
1324         }
1325 }
1326
1327 /*
1328  * callback from sys_ioprio_set, irqs are disabled
1329  */
1330 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1331 {
1332         struct cfq_io_context *cic = ioc->cic;
1333
1334         changed_ioprio(cic->cfqq);
1335
1336         list_for_each_entry(cic, &cic->list, list)
1337                 changed_ioprio(cic->cfqq);
1338
1339         return 0;
1340 }
1341
1342 static struct cfq_queue *
1343 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, unsigned short ioprio,
1344               gfp_t gfp_mask)
1345 {
1346         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1347         struct cfq_queue *cfqq, *new_cfqq = NULL;
1348
1349 retry:
1350         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1351
1352         if (!cfqq) {
1353                 if (new_cfqq) {
1354                         cfqq = new_cfqq;
1355                         new_cfqq = NULL;
1356                 } else if (gfp_mask & __GFP_WAIT) {
1357                         spin_unlock_irq(cfqd->queue->queue_lock);
1358                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1359                         spin_lock_irq(cfqd->queue->queue_lock);
1360                         goto retry;
1361                 } else {
1362                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1363                         if (!cfqq)
1364                                 goto out;
1365                 }
1366
1367                 memset(cfqq, 0, sizeof(*cfqq));
1368
1369                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1370                 INIT_LIST_HEAD(&cfqq->cfq_list);
1371                 RB_CLEAR_ROOT(&cfqq->sort_list);
1372                 INIT_LIST_HEAD(&cfqq->fifo);
1373
1374                 cfqq->key = key;
1375                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1376                 atomic_set(&cfqq->ref, 0);
1377                 cfqq->cfqd = cfqd;
1378                 atomic_inc(&cfqd->ref);
1379                 cfqq->service_last = 0;
1380                 /*
1381                  * set ->slice_left to allow preemption for a new process
1382                  */
1383                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1384                 cfq_mark_cfqq_idle_window(cfqq);
1385                 cfq_mark_cfqq_prio_changed(cfqq);
1386                 cfq_init_prio_data(cfqq);
1387         }
1388
1389         if (new_cfqq)
1390                 kmem_cache_free(cfq_pool, new_cfqq);
1391
1392         atomic_inc(&cfqq->ref);
1393 out:
1394         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1395         return cfqq;
1396 }
1397
1398 /*
1399  * Setup general io context and cfq io context. There can be several cfq
1400  * io contexts per general io context, if this process is doing io to more
1401  * than one device managed by cfq. Note that caller is holding a reference to
1402  * cfqq, so we don't need to worry about it disappearing
1403  */
1404 static struct cfq_io_context *
1405 cfq_get_io_context(struct cfq_data *cfqd, pid_t pid, gfp_t gfp_mask)
1406 {
1407         struct io_context *ioc = NULL;
1408         struct cfq_io_context *cic;
1409
1410         might_sleep_if(gfp_mask & __GFP_WAIT);
1411
1412         ioc = get_io_context(gfp_mask);
1413         if (!ioc)
1414                 return NULL;
1415
1416         if ((cic = ioc->cic) == NULL) {
1417                 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1418
1419                 if (cic == NULL)
1420                         goto err;
1421
1422                 /*
1423                  * manually increment generic io_context usage count, it
1424                  * cannot go away since we are already holding one ref to it
1425                  */
1426                 ioc->cic = cic;
1427                 ioc->set_ioprio = cfq_ioc_set_ioprio;
1428                 cic->ioc = ioc;
1429                 cic->key = cfqd;
1430                 atomic_inc(&cfqd->ref);
1431         } else {
1432                 struct cfq_io_context *__cic;
1433
1434                 /*
1435                  * the first cic on the list is actually the head itself
1436                  */
1437                 if (cic->key == cfqd)
1438                         goto out;
1439
1440                 /*
1441                  * cic exists, check if we already are there. linear search
1442                  * should be ok here, the list will usually not be more than
1443                  * 1 or a few entries long
1444                  */
1445                 list_for_each_entry(__cic, &cic->list, list) {
1446                         /*
1447                          * this process is already holding a reference to
1448                          * this queue, so no need to get one more
1449                          */
1450                         if (__cic->key == cfqd) {
1451                                 cic = __cic;
1452                                 goto out;
1453                         }
1454                 }
1455
1456                 /*
1457                  * nope, process doesn't have a cic assoicated with this
1458                  * cfqq yet. get a new one and add to list
1459                  */
1460                 __cic = cfq_alloc_io_context(cfqd, gfp_mask);
1461                 if (__cic == NULL)
1462                         goto err;
1463
1464                 __cic->ioc = ioc;
1465                 __cic->key = cfqd;
1466                 atomic_inc(&cfqd->ref);
1467                 list_add(&__cic->list, &cic->list);
1468                 cic = __cic;
1469         }
1470
1471 out:
1472         return cic;
1473 err:
1474         put_io_context(ioc);
1475         return NULL;
1476 }
1477
1478 static void
1479 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1480 {
1481         unsigned long elapsed, ttime;
1482
1483         /*
1484          * if this context already has stuff queued, thinktime is from
1485          * last queue not last end
1486          */
1487 #if 0
1488         if (time_after(cic->last_end_request, cic->last_queue))
1489                 elapsed = jiffies - cic->last_end_request;
1490         else
1491                 elapsed = jiffies - cic->last_queue;
1492 #else
1493                 elapsed = jiffies - cic->last_end_request;
1494 #endif
1495
1496         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1497
1498         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1499         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1500         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1501 }
1502
1503 #define sample_valid(samples)   ((samples) > 80)
1504
1505 /*
1506  * Disable idle window if the process thinks too long or seeks so much that
1507  * it doesn't matter
1508  */
1509 static void
1510 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1511                        struct cfq_io_context *cic)
1512 {
1513         int enable_idle = cfq_cfqq_idle_window(cfqq);
1514
1515         if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1516                 enable_idle = 0;
1517         else if (sample_valid(cic->ttime_samples)) {
1518                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1519                         enable_idle = 0;
1520                 else
1521                         enable_idle = 1;
1522         }
1523
1524         if (enable_idle)
1525                 cfq_mark_cfqq_idle_window(cfqq);
1526         else
1527                 cfq_clear_cfqq_idle_window(cfqq);
1528 }
1529
1530
1531 /*
1532  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1533  * no or if we aren't sure, a 1 will cause a preempt.
1534  */
1535 static int
1536 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1537                    struct cfq_rq *crq)
1538 {
1539         struct cfq_queue *cfqq = cfqd->active_queue;
1540
1541         if (cfq_class_idle(new_cfqq))
1542                 return 0;
1543
1544         if (!cfqq)
1545                 return 1;
1546
1547         if (cfq_class_idle(cfqq))
1548                 return 1;
1549         if (!cfq_cfqq_wait_request(new_cfqq))
1550                 return 0;
1551         /*
1552          * if it doesn't have slice left, forget it
1553          */
1554         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1555                 return 0;
1556         if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1557                 return 1;
1558
1559         return 0;
1560 }
1561
1562 /*
1563  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1564  * let it have half of its nominal slice.
1565  */
1566 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1567 {
1568         struct cfq_queue *__cfqq, *next;
1569
1570         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1571                 cfq_resort_rr_list(__cfqq, 1);
1572
1573         if (!cfqq->slice_left)
1574                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1575
1576         cfqq->slice_end = cfqq->slice_left + jiffies;
1577         __cfq_slice_expired(cfqd, cfqq, 1);
1578         __cfq_set_active_queue(cfqd, cfqq);
1579 }
1580
1581 /*
1582  * should really be a ll_rw_blk.c helper
1583  */
1584 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1585 {
1586         request_queue_t *q = cfqd->queue;
1587
1588         if (!blk_queue_plugged(q))
1589                 q->request_fn(q);
1590         else
1591                 __generic_unplug_device(q);
1592 }
1593
1594 /*
1595  * Called when a new fs request (crq) is added (to cfqq). Check if there's
1596  * something we should do about it
1597  */
1598 static void
1599 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1600                  struct cfq_rq *crq)
1601 {
1602         struct cfq_io_context *cic;
1603
1604         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1605
1606         /*
1607          * we never wait for an async request and we don't allow preemption
1608          * of an async request. so just return early
1609          */
1610         if (!cfq_crq_is_sync(crq))
1611                 return;
1612
1613         cic = crq->io_context;
1614
1615         cfq_update_io_thinktime(cfqd, cic);
1616         cfq_update_idle_window(cfqd, cfqq, cic);
1617
1618         cic->last_queue = jiffies;
1619
1620         if (cfqq == cfqd->active_queue) {
1621                 /*
1622                  * if we are waiting for a request for this queue, let it rip
1623                  * immediately and flag that we must not expire this queue
1624                  * just now
1625                  */
1626                 if (cfq_cfqq_wait_request(cfqq)) {
1627                         cfq_mark_cfqq_must_dispatch(cfqq);
1628                         del_timer(&cfqd->idle_slice_timer);
1629                         cfq_start_queueing(cfqd, cfqq);
1630                 }
1631         } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1632                 /*
1633                  * not the active queue - expire current slice if it is
1634                  * idle and has expired it's mean thinktime or this new queue
1635                  * has some old slice time left and is of higher priority
1636                  */
1637                 cfq_preempt_queue(cfqd, cfqq);
1638                 cfq_mark_cfqq_must_dispatch(cfqq);
1639                 cfq_start_queueing(cfqd, cfqq);
1640         }
1641 }
1642
1643 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1644 {
1645         struct cfq_data *cfqd = q->elevator->elevator_data;
1646         struct cfq_rq *crq = RQ_DATA(rq);
1647         struct cfq_queue *cfqq = crq->cfq_queue;
1648
1649         cfq_init_prio_data(cfqq);
1650
1651         cfq_add_crq_rb(crq);
1652
1653         list_add_tail(&rq->queuelist, &cfqq->fifo);
1654
1655         if (rq_mergeable(rq))
1656                 cfq_add_crq_hash(cfqd, crq);
1657
1658         cfq_crq_enqueued(cfqd, cfqq, crq);
1659 }
1660
1661 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1662 {
1663         struct cfq_rq *crq = RQ_DATA(rq);
1664         struct cfq_queue *cfqq = crq->cfq_queue;
1665         struct cfq_data *cfqd = cfqq->cfqd;
1666         const int sync = cfq_crq_is_sync(crq);
1667         unsigned long now;
1668
1669         now = jiffies;
1670
1671         WARN_ON(!cfqd->rq_in_driver);
1672         WARN_ON(!cfqq->on_dispatch[sync]);
1673         cfqd->rq_in_driver--;
1674         cfqq->on_dispatch[sync]--;
1675
1676         if (!cfq_class_idle(cfqq))
1677                 cfqd->last_end_request = now;
1678
1679         if (!cfq_cfqq_dispatched(cfqq)) {
1680                 if (cfq_cfqq_on_rr(cfqq)) {
1681                         cfqq->service_last = now;
1682                         cfq_resort_rr_list(cfqq, 0);
1683                 }
1684                 if (cfq_cfqq_expired(cfqq)) {
1685                         __cfq_slice_expired(cfqd, cfqq, 0);
1686                         cfq_schedule_dispatch(cfqd);
1687                 }
1688         }
1689
1690         if (cfq_crq_is_sync(crq))
1691                 crq->io_context->last_end_request = now;
1692 }
1693
1694 static struct request *
1695 cfq_former_request(request_queue_t *q, struct request *rq)
1696 {
1697         struct cfq_rq *crq = RQ_DATA(rq);
1698         struct rb_node *rbprev = rb_prev(&crq->rb_node);
1699
1700         if (rbprev)
1701                 return rb_entry_crq(rbprev)->request;
1702
1703         return NULL;
1704 }
1705
1706 static struct request *
1707 cfq_latter_request(request_queue_t *q, struct request *rq)
1708 {
1709         struct cfq_rq *crq = RQ_DATA(rq);
1710         struct rb_node *rbnext = rb_next(&crq->rb_node);
1711
1712         if (rbnext)
1713                 return rb_entry_crq(rbnext)->request;
1714
1715         return NULL;
1716 }
1717
1718 /*
1719  * we temporarily boost lower priority queues if they are holding fs exclusive
1720  * resources. they are boosted to normal prio (CLASS_BE/4)
1721  */
1722 static void cfq_prio_boost(struct cfq_queue *cfqq)
1723 {
1724         const int ioprio_class = cfqq->ioprio_class;
1725         const int ioprio = cfqq->ioprio;
1726
1727         if (has_fs_excl()) {
1728                 /*
1729                  * boost idle prio on transactions that would lock out other
1730                  * users of the filesystem
1731                  */
1732                 if (cfq_class_idle(cfqq))
1733                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1734                 if (cfqq->ioprio > IOPRIO_NORM)
1735                         cfqq->ioprio = IOPRIO_NORM;
1736         } else {
1737                 /*
1738                  * check if we need to unboost the queue
1739                  */
1740                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1741                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1742                 if (cfqq->ioprio != cfqq->org_ioprio)
1743                         cfqq->ioprio = cfqq->org_ioprio;
1744         }
1745
1746         /*
1747          * refile between round-robin lists if we moved the priority class
1748          */
1749         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1750             cfq_cfqq_on_rr(cfqq))
1751                 cfq_resort_rr_list(cfqq, 0);
1752 }
1753
1754 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
1755 {
1756         if (rw == READ || process_sync(task))
1757                 return task->pid;
1758
1759         return CFQ_KEY_ASYNC;
1760 }
1761
1762 static inline int
1763 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1764                 struct task_struct *task, int rw)
1765 {
1766 #if 1
1767         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1768             !cfq_cfqq_must_alloc_slice(cfqq)) {
1769                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1770                 return ELV_MQUEUE_MUST;
1771         }
1772
1773         return ELV_MQUEUE_MAY;
1774 #else
1775         if (!cfqq || task->flags & PF_MEMALLOC)
1776                 return ELV_MQUEUE_MAY;
1777         if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1778                 if (cfq_cfqq_wait_request(cfqq))
1779                         return ELV_MQUEUE_MUST;
1780
1781                 /*
1782                  * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1783                  * can quickly flood the queue with writes from a single task
1784                  */
1785                 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
1786                         cfq_mark_cfqq_must_alloc_slice(cfqq);
1787                         return ELV_MQUEUE_MUST;
1788                 }
1789
1790                 return ELV_MQUEUE_MAY;
1791         }
1792         if (cfq_class_idle(cfqq))
1793                 return ELV_MQUEUE_NO;
1794         if (cfqq->allocated[rw] >= cfqd->max_queued) {
1795                 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1796                 int ret = ELV_MQUEUE_NO;
1797
1798                 if (ioc && ioc->nr_batch_requests)
1799                         ret = ELV_MQUEUE_MAY;
1800
1801                 put_io_context(ioc);
1802                 return ret;
1803         }
1804
1805         return ELV_MQUEUE_MAY;
1806 #endif
1807 }
1808
1809 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1810 {
1811         struct cfq_data *cfqd = q->elevator->elevator_data;
1812         struct task_struct *tsk = current;
1813         struct cfq_queue *cfqq;
1814
1815         /*
1816          * don't force setup of a queue from here, as a call to may_queue
1817          * does not necessarily imply that a request actually will be queued.
1818          * so just lookup a possibly existing queue, or return 'may queue'
1819          * if that fails
1820          */
1821         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1822         if (cfqq) {
1823                 cfq_init_prio_data(cfqq);
1824                 cfq_prio_boost(cfqq);
1825
1826                 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1827         }
1828
1829         return ELV_MQUEUE_MAY;
1830 }
1831
1832 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1833 {
1834         struct cfq_data *cfqd = q->elevator->elevator_data;
1835         struct request_list *rl = &q->rq;
1836
1837         if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1838                 smp_mb();
1839                 if (waitqueue_active(&rl->wait[READ]))
1840                         wake_up(&rl->wait[READ]);
1841         }
1842
1843         if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
1844                 smp_mb();
1845                 if (waitqueue_active(&rl->wait[WRITE]))
1846                         wake_up(&rl->wait[WRITE]);
1847         }
1848 }
1849
1850 /*
1851  * queue lock held here
1852  */
1853 static void cfq_put_request(request_queue_t *q, struct request *rq)
1854 {
1855         struct cfq_data *cfqd = q->elevator->elevator_data;
1856         struct cfq_rq *crq = RQ_DATA(rq);
1857
1858         if (crq) {
1859                 struct cfq_queue *cfqq = crq->cfq_queue;
1860                 const int rw = rq_data_dir(rq);
1861
1862                 BUG_ON(!cfqq->allocated[rw]);
1863                 cfqq->allocated[rw]--;
1864
1865                 put_io_context(crq->io_context->ioc);
1866
1867                 mempool_free(crq, cfqd->crq_pool);
1868                 rq->elevator_private = NULL;
1869
1870                 cfq_check_waiters(q, cfqq);
1871                 cfq_put_queue(cfqq);
1872         }
1873 }
1874
1875 /*
1876  * Allocate cfq data structures associated with this request.
1877  */
1878 static int
1879 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
1880                 gfp_t gfp_mask)
1881 {
1882         struct cfq_data *cfqd = q->elevator->elevator_data;
1883         struct task_struct *tsk = current;
1884         struct cfq_io_context *cic;
1885         const int rw = rq_data_dir(rq);
1886         pid_t key = cfq_queue_pid(tsk, rw);
1887         struct cfq_queue *cfqq;
1888         struct cfq_rq *crq;
1889         unsigned long flags;
1890
1891         might_sleep_if(gfp_mask & __GFP_WAIT);
1892
1893         cic = cfq_get_io_context(cfqd, key, gfp_mask);
1894
1895         spin_lock_irqsave(q->queue_lock, flags);
1896
1897         if (!cic)
1898                 goto queue_fail;
1899
1900         if (!cic->cfqq) {
1901                 cfqq = cfq_get_queue(cfqd, key, tsk->ioprio, gfp_mask);
1902                 if (!cfqq)
1903                         goto queue_fail;
1904
1905                 cic->cfqq = cfqq;
1906         } else
1907                 cfqq = cic->cfqq;
1908
1909         cfqq->allocated[rw]++;
1910         cfq_clear_cfqq_must_alloc(cfqq);
1911         cfqd->rq_starved = 0;
1912         atomic_inc(&cfqq->ref);
1913         spin_unlock_irqrestore(q->queue_lock, flags);
1914
1915         crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
1916         if (crq) {
1917                 RB_CLEAR(&crq->rb_node);
1918                 crq->rb_key = 0;
1919                 crq->request = rq;
1920                 INIT_HLIST_NODE(&crq->hash);
1921                 crq->cfq_queue = cfqq;
1922                 crq->io_context = cic;
1923
1924                 if (rw == READ || process_sync(tsk))
1925                         cfq_mark_crq_is_sync(crq);
1926                 else
1927                         cfq_clear_crq_is_sync(crq);
1928
1929                 rq->elevator_private = crq;
1930                 return 0;
1931         }
1932
1933         spin_lock_irqsave(q->queue_lock, flags);
1934         cfqq->allocated[rw]--;
1935         if (!(cfqq->allocated[0] + cfqq->allocated[1]))
1936                 cfq_mark_cfqq_must_alloc(cfqq);
1937         cfq_put_queue(cfqq);
1938 queue_fail:
1939         if (cic)
1940                 put_io_context(cic->ioc);
1941         /*
1942          * mark us rq allocation starved. we need to kickstart the process
1943          * ourselves if there are no pending requests that can do it for us.
1944          * that would be an extremely rare OOM situation
1945          */
1946         cfqd->rq_starved = 1;
1947         cfq_schedule_dispatch(cfqd);
1948         spin_unlock_irqrestore(q->queue_lock, flags);
1949         return 1;
1950 }
1951
1952 static void cfq_kick_queue(void *data)
1953 {
1954         request_queue_t *q = data;
1955         struct cfq_data *cfqd = q->elevator->elevator_data;
1956         unsigned long flags;
1957
1958         spin_lock_irqsave(q->queue_lock, flags);
1959
1960         if (cfqd->rq_starved) {
1961                 struct request_list *rl = &q->rq;
1962
1963                 /*
1964                  * we aren't guaranteed to get a request after this, but we
1965                  * have to be opportunistic
1966                  */
1967                 smp_mb();
1968                 if (waitqueue_active(&rl->wait[READ]))
1969                         wake_up(&rl->wait[READ]);
1970                 if (waitqueue_active(&rl->wait[WRITE]))
1971                         wake_up(&rl->wait[WRITE]);
1972         }
1973
1974         blk_remove_plug(q);
1975         q->request_fn(q);
1976         spin_unlock_irqrestore(q->queue_lock, flags);
1977 }
1978
1979 /*
1980  * Timer running if the active_queue is currently idling inside its time slice
1981  */
1982 static void cfq_idle_slice_timer(unsigned long data)
1983 {
1984         struct cfq_data *cfqd = (struct cfq_data *) data;
1985         struct cfq_queue *cfqq;
1986         unsigned long flags;
1987
1988         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1989
1990         if ((cfqq = cfqd->active_queue) != NULL) {
1991                 unsigned long now = jiffies;
1992
1993                 /*
1994                  * expired
1995                  */
1996                 if (time_after(now, cfqq->slice_end))
1997                         goto expire;
1998
1999                 /*
2000                  * only expire and reinvoke request handler, if there are
2001                  * other queues with pending requests
2002                  */
2003                 if (!cfqd->busy_queues) {
2004                         cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2005                         add_timer(&cfqd->idle_slice_timer);
2006                         goto out_cont;
2007                 }
2008
2009                 /*
2010                  * not expired and it has a request pending, let it dispatch
2011                  */
2012                 if (!RB_EMPTY(&cfqq->sort_list)) {
2013                         cfq_mark_cfqq_must_dispatch(cfqq);
2014                         goto out_kick;
2015                 }
2016         }
2017 expire:
2018         cfq_slice_expired(cfqd, 0);
2019 out_kick:
2020         cfq_schedule_dispatch(cfqd);
2021 out_cont:
2022         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2023 }
2024
2025 /*
2026  * Timer running if an idle class queue is waiting for service
2027  */
2028 static void cfq_idle_class_timer(unsigned long data)
2029 {
2030         struct cfq_data *cfqd = (struct cfq_data *) data;
2031         unsigned long flags, end;
2032
2033         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2034
2035         /*
2036          * race with a non-idle queue, reset timer
2037          */
2038         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2039         if (!time_after_eq(jiffies, end)) {
2040                 cfqd->idle_class_timer.expires = end;
2041                 add_timer(&cfqd->idle_class_timer);
2042         } else
2043                 cfq_schedule_dispatch(cfqd);
2044
2045         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2046 }
2047
2048 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2049 {
2050         del_timer_sync(&cfqd->idle_slice_timer);
2051         del_timer_sync(&cfqd->idle_class_timer);
2052         blk_sync_queue(cfqd->queue);
2053 }
2054
2055 static void cfq_put_cfqd(struct cfq_data *cfqd)
2056 {
2057         request_queue_t *q = cfqd->queue;
2058
2059         if (!atomic_dec_and_test(&cfqd->ref))
2060                 return;
2061
2062         blk_put_queue(q);
2063
2064         cfq_shutdown_timer_wq(cfqd);
2065         q->elevator->elevator_data = NULL;
2066
2067         mempool_destroy(cfqd->crq_pool);
2068         kfree(cfqd->crq_hash);
2069         kfree(cfqd->cfq_hash);
2070         kfree(cfqd);
2071 }
2072
2073 static void cfq_exit_queue(elevator_t *e)
2074 {
2075         struct cfq_data *cfqd = e->elevator_data;
2076
2077         cfq_shutdown_timer_wq(cfqd);
2078         cfq_put_cfqd(cfqd);
2079 }
2080
2081 static int cfq_init_queue(request_queue_t *q, elevator_t *e)
2082 {
2083         struct cfq_data *cfqd;
2084         int i;
2085
2086         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2087         if (!cfqd)
2088                 return -ENOMEM;
2089
2090         memset(cfqd, 0, sizeof(*cfqd));
2091
2092         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2093                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2094
2095         INIT_LIST_HEAD(&cfqd->busy_rr);
2096         INIT_LIST_HEAD(&cfqd->cur_rr);
2097         INIT_LIST_HEAD(&cfqd->idle_rr);
2098         INIT_LIST_HEAD(&cfqd->empty_list);
2099
2100         cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2101         if (!cfqd->crq_hash)
2102                 goto out_crqhash;
2103
2104         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2105         if (!cfqd->cfq_hash)
2106                 goto out_cfqhash;
2107
2108         cfqd->crq_pool = mempool_create(BLKDEV_MIN_RQ, mempool_alloc_slab, mempool_free_slab, crq_pool);
2109         if (!cfqd->crq_pool)
2110                 goto out_crqpool;
2111
2112         for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2113                 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2114         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2115                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2116
2117         e->elevator_data = cfqd;
2118
2119         cfqd->queue = q;
2120         atomic_inc(&q->refcnt);
2121
2122         cfqd->max_queued = q->nr_requests / 4;
2123         q->nr_batching = cfq_queued;
2124
2125         init_timer(&cfqd->idle_slice_timer);
2126         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2127         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2128
2129         init_timer(&cfqd->idle_class_timer);
2130         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2131         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2132
2133         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2134
2135         atomic_set(&cfqd->ref, 1);
2136
2137         cfqd->cfq_queued = cfq_queued;
2138         cfqd->cfq_quantum = cfq_quantum;
2139         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2140         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2141         cfqd->cfq_back_max = cfq_back_max;
2142         cfqd->cfq_back_penalty = cfq_back_penalty;
2143         cfqd->cfq_slice[0] = cfq_slice_async;
2144         cfqd->cfq_slice[1] = cfq_slice_sync;
2145         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2146         cfqd->cfq_slice_idle = cfq_slice_idle;
2147         cfqd->cfq_max_depth = cfq_max_depth;
2148
2149         return 0;
2150 out_crqpool:
2151         kfree(cfqd->cfq_hash);
2152 out_cfqhash:
2153         kfree(cfqd->crq_hash);
2154 out_crqhash:
2155         kfree(cfqd);
2156         return -ENOMEM;
2157 }
2158
2159 static void cfq_slab_kill(void)
2160 {
2161         if (crq_pool)
2162                 kmem_cache_destroy(crq_pool);
2163         if (cfq_pool)
2164                 kmem_cache_destroy(cfq_pool);
2165         if (cfq_ioc_pool)
2166                 kmem_cache_destroy(cfq_ioc_pool);
2167 }
2168
2169 static int __init cfq_slab_setup(void)
2170 {
2171         crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2172                                         NULL, NULL);
2173         if (!crq_pool)
2174                 goto fail;
2175
2176         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2177                                         NULL, NULL);
2178         if (!cfq_pool)
2179                 goto fail;
2180
2181         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2182                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2183         if (!cfq_ioc_pool)
2184                 goto fail;
2185
2186         return 0;
2187 fail:
2188         cfq_slab_kill();
2189         return -ENOMEM;
2190 }
2191
2192 /*
2193  * sysfs parts below -->
2194  */
2195 struct cfq_fs_entry {
2196         struct attribute attr;
2197         ssize_t (*show)(struct cfq_data *, char *);
2198         ssize_t (*store)(struct cfq_data *, const char *, size_t);
2199 };
2200
2201 static ssize_t
2202 cfq_var_show(unsigned int var, char *page)
2203 {
2204         return sprintf(page, "%d\n", var);
2205 }
2206
2207 static ssize_t
2208 cfq_var_store(unsigned int *var, const char *page, size_t count)
2209 {
2210         char *p = (char *) page;
2211
2212         *var = simple_strtoul(p, &p, 10);
2213         return count;
2214 }
2215
2216 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2217 static ssize_t __FUNC(struct cfq_data *cfqd, char *page)                \
2218 {                                                                       \
2219         unsigned int __data = __VAR;                                    \
2220         if (__CONV)                                                     \
2221                 __data = jiffies_to_msecs(__data);                      \
2222         return cfq_var_show(__data, (page));                            \
2223 }
2224 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2225 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2226 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2227 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2228 SHOW_FUNCTION(cfq_back_max_show, cfqd->cfq_back_max, 0);
2229 SHOW_FUNCTION(cfq_back_penalty_show, cfqd->cfq_back_penalty, 0);
2230 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2231 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2232 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2233 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2234 SHOW_FUNCTION(cfq_max_depth_show, cfqd->cfq_max_depth, 0);
2235 #undef SHOW_FUNCTION
2236
2237 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2238 static ssize_t __FUNC(struct cfq_data *cfqd, const char *page, size_t count)    \
2239 {                                                                       \
2240         unsigned int __data;                                            \
2241         int ret = cfq_var_store(&__data, (page), count);                \
2242         if (__data < (MIN))                                             \
2243                 __data = (MIN);                                         \
2244         else if (__data > (MAX))                                        \
2245                 __data = (MAX);                                         \
2246         if (__CONV)                                                     \
2247                 *(__PTR) = msecs_to_jiffies(__data);                    \
2248         else                                                            \
2249                 *(__PTR) = __data;                                      \
2250         return ret;                                                     \
2251 }
2252 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2253 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2254 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2255 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2256 STORE_FUNCTION(cfq_back_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2257 STORE_FUNCTION(cfq_back_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2258 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2259 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2260 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2261 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2262 STORE_FUNCTION(cfq_max_depth_store, &cfqd->cfq_max_depth, 1, UINT_MAX, 0);
2263 #undef STORE_FUNCTION
2264
2265 static struct cfq_fs_entry cfq_quantum_entry = {
2266         .attr = {.name = "quantum", .mode = S_IRUGO | S_IWUSR },
2267         .show = cfq_quantum_show,
2268         .store = cfq_quantum_store,
2269 };
2270 static struct cfq_fs_entry cfq_queued_entry = {
2271         .attr = {.name = "queued", .mode = S_IRUGO | S_IWUSR },
2272         .show = cfq_queued_show,
2273         .store = cfq_queued_store,
2274 };
2275 static struct cfq_fs_entry cfq_fifo_expire_sync_entry = {
2276         .attr = {.name = "fifo_expire_sync", .mode = S_IRUGO | S_IWUSR },
2277         .show = cfq_fifo_expire_sync_show,
2278         .store = cfq_fifo_expire_sync_store,
2279 };
2280 static struct cfq_fs_entry cfq_fifo_expire_async_entry = {
2281         .attr = {.name = "fifo_expire_async", .mode = S_IRUGO | S_IWUSR },
2282         .show = cfq_fifo_expire_async_show,
2283         .store = cfq_fifo_expire_async_store,
2284 };
2285 static struct cfq_fs_entry cfq_back_max_entry = {
2286         .attr = {.name = "back_seek_max", .mode = S_IRUGO | S_IWUSR },
2287         .show = cfq_back_max_show,
2288         .store = cfq_back_max_store,
2289 };
2290 static struct cfq_fs_entry cfq_back_penalty_entry = {
2291         .attr = {.name = "back_seek_penalty", .mode = S_IRUGO | S_IWUSR },
2292         .show = cfq_back_penalty_show,
2293         .store = cfq_back_penalty_store,
2294 };
2295 static struct cfq_fs_entry cfq_slice_sync_entry = {
2296         .attr = {.name = "slice_sync", .mode = S_IRUGO | S_IWUSR },
2297         .show = cfq_slice_sync_show,
2298         .store = cfq_slice_sync_store,
2299 };
2300 static struct cfq_fs_entry cfq_slice_async_entry = {
2301         .attr = {.name = "slice_async", .mode = S_IRUGO | S_IWUSR },
2302         .show = cfq_slice_async_show,
2303         .store = cfq_slice_async_store,
2304 };
2305 static struct cfq_fs_entry cfq_slice_async_rq_entry = {
2306         .attr = {.name = "slice_async_rq", .mode = S_IRUGO | S_IWUSR },
2307         .show = cfq_slice_async_rq_show,
2308         .store = cfq_slice_async_rq_store,
2309 };
2310 static struct cfq_fs_entry cfq_slice_idle_entry = {
2311         .attr = {.name = "slice_idle", .mode = S_IRUGO | S_IWUSR },
2312         .show = cfq_slice_idle_show,
2313         .store = cfq_slice_idle_store,
2314 };
2315 static struct cfq_fs_entry cfq_max_depth_entry = {
2316         .attr = {.name = "max_depth", .mode = S_IRUGO | S_IWUSR },
2317         .show = cfq_max_depth_show,
2318         .store = cfq_max_depth_store,
2319 };
2320
2321 static struct attribute *default_attrs[] = {
2322         &cfq_quantum_entry.attr,
2323         &cfq_queued_entry.attr,
2324         &cfq_fifo_expire_sync_entry.attr,
2325         &cfq_fifo_expire_async_entry.attr,
2326         &cfq_back_max_entry.attr,
2327         &cfq_back_penalty_entry.attr,
2328         &cfq_slice_sync_entry.attr,
2329         &cfq_slice_async_entry.attr,
2330         &cfq_slice_async_rq_entry.attr,
2331         &cfq_slice_idle_entry.attr,
2332         &cfq_max_depth_entry.attr,
2333         NULL,
2334 };
2335
2336 #define to_cfq(atr) container_of((atr), struct cfq_fs_entry, attr)
2337
2338 static ssize_t
2339 cfq_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
2340 {
2341         elevator_t *e = container_of(kobj, elevator_t, kobj);
2342         struct cfq_fs_entry *entry = to_cfq(attr);
2343
2344         if (!entry->show)
2345                 return -EIO;
2346
2347         return entry->show(e->elevator_data, page);
2348 }
2349
2350 static ssize_t
2351 cfq_attr_store(struct kobject *kobj, struct attribute *attr,
2352                const char *page, size_t length)
2353 {
2354         elevator_t *e = container_of(kobj, elevator_t, kobj);
2355         struct cfq_fs_entry *entry = to_cfq(attr);
2356
2357         if (!entry->store)
2358                 return -EIO;
2359
2360         return entry->store(e->elevator_data, page, length);
2361 }
2362
2363 static struct sysfs_ops cfq_sysfs_ops = {
2364         .show   = cfq_attr_show,
2365         .store  = cfq_attr_store,
2366 };
2367
2368 static struct kobj_type cfq_ktype = {
2369         .sysfs_ops      = &cfq_sysfs_ops,
2370         .default_attrs  = default_attrs,
2371 };
2372
2373 static struct elevator_type iosched_cfq = {
2374         .ops = {
2375                 .elevator_merge_fn =            cfq_merge,
2376                 .elevator_merged_fn =           cfq_merged_request,
2377                 .elevator_merge_req_fn =        cfq_merged_requests,
2378                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2379                 .elevator_add_req_fn =          cfq_insert_request,
2380                 .elevator_activate_req_fn =     cfq_activate_request,
2381                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2382                 .elevator_queue_empty_fn =      cfq_queue_empty,
2383                 .elevator_completed_req_fn =    cfq_completed_request,
2384                 .elevator_former_req_fn =       cfq_former_request,
2385                 .elevator_latter_req_fn =       cfq_latter_request,
2386                 .elevator_set_req_fn =          cfq_set_request,
2387                 .elevator_put_req_fn =          cfq_put_request,
2388                 .elevator_may_queue_fn =        cfq_may_queue,
2389                 .elevator_init_fn =             cfq_init_queue,
2390                 .elevator_exit_fn =             cfq_exit_queue,
2391         },
2392         .elevator_ktype =       &cfq_ktype,
2393         .elevator_name =        "cfq",
2394         .elevator_owner =       THIS_MODULE,
2395 };
2396
2397 static int __init cfq_init(void)
2398 {
2399         int ret;
2400
2401         /*
2402          * could be 0 on HZ < 1000 setups
2403          */
2404         if (!cfq_slice_async)
2405                 cfq_slice_async = 1;
2406         if (!cfq_slice_idle)
2407                 cfq_slice_idle = 1;
2408
2409         if (cfq_slab_setup())
2410                 return -ENOMEM;
2411
2412         ret = elv_register(&iosched_cfq);
2413         if (ret)
2414                 cfq_slab_kill();
2415
2416         return ret;
2417 }
2418
2419 static void __exit cfq_exit(void)
2420 {
2421         struct task_struct *g, *p;
2422         unsigned long flags;
2423
2424         read_lock_irqsave(&tasklist_lock, flags);
2425
2426         /*
2427          * iterate each process in the system, removing our io_context
2428          */
2429         do_each_thread(g, p) {
2430                 struct io_context *ioc = p->io_context;
2431
2432                 if (ioc && ioc->cic) {
2433                         ioc->cic->exit(ioc->cic);
2434                         cfq_free_io_context(ioc->cic);
2435                         ioc->cic = NULL;
2436                 }
2437         } while_each_thread(g, p);
2438
2439         read_unlock_irqrestore(&tasklist_lock, flags);
2440
2441         cfq_slab_kill();
2442         elv_unregister(&iosched_cfq);
2443 }
2444
2445 module_init(cfq_init);
2446 module_exit(cfq_exit);
2447
2448 MODULE_AUTHOR("Jens Axboe");
2449 MODULE_LICENSE("GPL");
2450 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");