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 ON_RB(node)             ((node)->rb_color != RB_NONE)
  88 #define rb_entry_crq(node)      rb_entry((node), struct cfq_rq, rb_node)
  89 #define rq_rb_key(rq)           (rq)->sector
  90
  91 static kmem_cache_t *crq_pool;
  92 static kmem_cache_t *cfq_pool;
  93 static kmem_cache_t *cfq_ioc_pool;
  94
  95 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  96 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  97 #define cfq_class_be(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
  98 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  99
 100 #define ASYNC                   (0)
 101 #define SYNC                    (1)
 102
 103 #define cfq_cfqq_dispatched(cfqq)       \
 104         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
 105
 106 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
 107
 108 #define cfq_cfqq_sync(cfqq)             \
 109         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
 110
 111 /*
 112  * Per block device queue structure
 113  */
 114 struct cfq_data {
 115         atomic_t ref;
 116         request_queue_t *queue;
 117
 118         /*
 119          * rr list of queues with requests and the count of them
 120          */
 121         struct list_head rr_list[CFQ_PRIO_LISTS];
 122         struct list_head busy_rr;
 123         struct list_head cur_rr;
 124         struct list_head idle_rr;
 125         unsigned int busy_queues;
 126
 127         /*
 128          * non-ordered list of empty cfqq's
 129          */
 130         struct list_head empty_list;
 131
 132         /*
 133          * cfqq lookup hash
 134          */
 135         struct hlist_head *cfq_hash;
 136
 137         /*
 138          * global crq hash for all queues
 139          */
 140         struct hlist_head *crq_hash;
 141
 142         unsigned int max_queued;
 143
 144         mempool_t *crq_pool;
 145
 146         int rq_in_driver;
 147
 148         /*
 149          * schedule slice state info
 150          */
 151         /*
 152          * idle window management
 153          */
 154         struct timer_list idle_slice_timer;
 155         struct work_struct unplug_work;
 156
 157         struct cfq_queue *active_queue;
 158         struct cfq_io_context *active_cic;
 159         int cur_prio, cur_end_prio;
 160         unsigned int dispatch_slice;
 161
 162         struct timer_list idle_class_timer;
 163
 164         sector_t last_sector;
 165         unsigned long last_end_request;
 166
 167         unsigned int rq_starved;
 168
 169         /*
 170          * tunables, see top of file
 171          */
 172         unsigned int cfq_quantum;
 173         unsigned int cfq_queued;
 174         unsigned int cfq_fifo_expire[2];
 175         unsigned int cfq_back_penalty;
 176         unsigned int cfq_back_max;
 177         unsigned int cfq_slice[2];
 178         unsigned int cfq_slice_async_rq;
 179         unsigned int cfq_slice_idle;
 180         unsigned int cfq_max_depth;
 181 };
 182
 183 /*
 184  * Per process-grouping structure
 185  */
 186 struct cfq_queue {
 187         /* reference count */
 188         atomic_t ref;
 189         /* parent cfq_data */
 190         struct cfq_data *cfqd;
 191         /* cfqq lookup hash */
 192         struct hlist_node cfq_hash;
 193         /* hash key */
 194         unsigned int key;
 195         /* on either rr or empty list of cfqd */
 196         struct list_head cfq_list;
 197         /* sorted list of pending requests */
 198         struct rb_root sort_list;
 199         /* if fifo isn't expired, next request to serve */
 200         struct cfq_rq *next_crq;
 201         /* requests queued in sort_list */
 202         int queued[2];
 203         /* currently allocated requests */
 204         int allocated[2];
 205         /* fifo list of requests in sort_list */
 206         struct list_head fifo;
 207
 208         unsigned long slice_start;
 209         unsigned long slice_end;
 210         unsigned long slice_left;
 211         unsigned long service_last;
 212
 213         /* number of requests that are on the dispatch list */
 214         int on_dispatch[2];
 215
 216         /* io prio of this group */
 217         unsigned short ioprio, org_ioprio;
 218         unsigned short ioprio_class, org_ioprio_class;
 219
 220         /* various state flags, see below */
 221         unsigned int flags;
 222 };
 223
 224 struct cfq_rq {
 225         struct rb_node rb_node;
 226         sector_t rb_key;
 227         struct request *request;
 228         struct hlist_node hash;
 229
 230         struct cfq_queue *cfq_queue;
 231         struct cfq_io_context *io_context;
 232
 233         unsigned int crq_flags;
 234 };
 235
 236 enum cfqq_state_flags {
 237         CFQ_CFQQ_FLAG_on_rr = 0,
 238         CFQ_CFQQ_FLAG_wait_request,
 239         CFQ_CFQQ_FLAG_must_alloc,
 240         CFQ_CFQQ_FLAG_must_alloc_slice,
 241         CFQ_CFQQ_FLAG_must_dispatch,
 242         CFQ_CFQQ_FLAG_fifo_expire,
 243         CFQ_CFQQ_FLAG_idle_window,
 244         CFQ_CFQQ_FLAG_prio_changed,
 245         CFQ_CFQQ_FLAG_expired,
 246 };
 247
 248 #define CFQ_CFQQ_FNS(name)                                              \
 249 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 250 {                                                                       \
 251         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 252 }                                                                       \
 253 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 254 {                                                                       \
 255         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 256 }                                                                       \
 257 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 258 {                                                                       \
 259         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 260 }
 261
 262 CFQ_CFQQ_FNS(on_rr);
 263 CFQ_CFQQ_FNS(wait_request);
 264 CFQ_CFQQ_FNS(must_alloc);
 265 CFQ_CFQQ_FNS(must_alloc_slice);
 266 CFQ_CFQQ_FNS(must_dispatch);
 267 CFQ_CFQQ_FNS(fifo_expire);
 268 CFQ_CFQQ_FNS(idle_window);
 269 CFQ_CFQQ_FNS(prio_changed);
 270 CFQ_CFQQ_FNS(expired);
 271 #undef CFQ_CFQQ_FNS
 272
 273 enum cfq_rq_state_flags {
 274         CFQ_CRQ_FLAG_in_flight = 0,
 275         CFQ_CRQ_FLAG_in_driver,
 276         CFQ_CRQ_FLAG_is_sync,
 277         CFQ_CRQ_FLAG_requeued,
 278 };
 279
 280 #define CFQ_CRQ_FNS(name)                                               \
 281 static inline void cfq_mark_crq_##name(struct cfq_rq *crq)              \
 282 {                                                                       \
 283         crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name);                   \
 284 }                                                                       \
 285 static inline void cfq_clear_crq_##name(struct cfq_rq *crq)             \
 286 {                                                                       \
 287         crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name);                  \
 288 }                                                                       \
 289 static inline int cfq_crq_##name(const struct cfq_rq *crq)              \
 290 {                                                                       \
 291         return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0;      \
 292 }
 293
 294 CFQ_CRQ_FNS(in_flight);
 295 CFQ_CRQ_FNS(in_driver);
 296 CFQ_CRQ_FNS(is_sync);
 297 CFQ_CRQ_FNS(requeued);
 298 #undef CFQ_CRQ_FNS
 299
 300 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 301 static void cfq_dispatch_sort(request_queue_t *, struct cfq_rq *);
 302 static void cfq_put_cfqd(struct cfq_data *cfqd);
 303
 304 #define process_sync(tsk)       ((tsk)->flags & PF_SYNCWRITE)
 305
 306 /*
 307  * lots of deadline iosched dupes, can be abstracted later...
 308  */
 309 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
 310 {
 311         hlist_del_init(&crq->hash);
 312 }
 313
 314 static void cfq_remove_merge_hints(request_queue_t *q, struct cfq_rq *crq)
 315 {
 316         cfq_del_crq_hash(crq);
 317
 318         if (q->last_merge == crq->request)
 319                 q->last_merge = NULL;
 320 }
 321
 322 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
 323 {
 324         const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
 325
 326         hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
 327 }
 328
 329 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
 330 {
 331         struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
 332         struct hlist_node *entry, *next;
 333
 334         hlist_for_each_safe(entry, next, hash_list) {
 335                 struct cfq_rq *crq = list_entry_hash(entry);
 336                 struct request *__rq = crq->request;
 337
 338                 if (!rq_mergeable(__rq)) {
 339                         cfq_del_crq_hash(crq);
 340                         continue;
 341                 }
 342
 343                 if (rq_hash_key(__rq) == offset)
 344                         return __rq;
 345         }
 346
 347         return NULL;
 348 }
 349
 350 static inline int cfq_pending_requests(struct cfq_data *cfqd)
 351 {
 352         return !list_empty(&cfqd->queue->queue_head) || cfqd->busy_queues;
 353 }
 354
 355 /*
 356  * scheduler run of queue, if there are requests pending and no one in the
 357  * driver that will restart queueing
 358  */
 359 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 360 {
 361         if (!cfqd->rq_in_driver && cfq_pending_requests(cfqd))
 362                 kblockd_schedule_work(&cfqd->unplug_work);
 363 }
 364
 365 static int cfq_queue_empty(request_queue_t *q)
 366 {
 367         struct cfq_data *cfqd = q->elevator->elevator_data;
 368
 369         return !cfq_pending_requests(cfqd);
 370 }
 371
 372 /*
 373  * Lifted from AS - choose which of crq1 and crq2 that is best served now.
 374  * We choose the request that is closest to the head right now. Distance
 375  * behind the head are penalized and only allowed to a certain extent.
 376  */
 377 static struct cfq_rq *
 378 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
 379 {
 380         sector_t last, s1, s2, d1 = 0, d2 = 0;
 381         int r1_wrap = 0, r2_wrap = 0;   /* requests are behind the disk head */
 382         unsigned long back_max;
 383
 384         if (crq1 == NULL || crq1 == crq2)
 385                 return crq2;
 386         if (crq2 == NULL)
 387                 return crq1;
 388
 389         if (cfq_crq_requeued(crq1) && !cfq_crq_requeued(crq2))
 390                 return crq1;
 391         else if (cfq_crq_requeued(crq2) && !cfq_crq_requeued(crq1))
 392                 return crq2;
 393
 394         if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
 395                 return crq1;
 396         else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
 397                 return crq2;
 398
 399         s1 = crq1->request->sector;
 400         s2 = crq2->request->sector;
 401
 402         last = cfqd->last_sector;
 403
 404         /*
 405          * by definition, 1KiB is 2 sectors
 406          */
 407         back_max = cfqd->cfq_back_max * 2;
 408
 409         /*
 410          * Strict one way elevator _except_ in the case where we allow
 411          * short backward seeks which are biased as twice the cost of a
 412          * similar forward seek.
 413          */
 414         if (s1 >= last)
 415                 d1 = s1 - last;
 416         else if (s1 + back_max >= last)
 417                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 418         else
 419                 r1_wrap = 1;
 420
 421         if (s2 >= last)
 422                 d2 = s2 - last;
 423         else if (s2 + back_max >= last)
 424                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 425         else
 426                 r2_wrap = 1;
 427
 428         /* Found required data */
 429         if (!r1_wrap && r2_wrap)
 430                 return crq1;
 431         else if (!r2_wrap && r1_wrap)
 432                 return crq2;
 433         else if (r1_wrap && r2_wrap) {
 434                 /* both behind the head */
 435                 if (s1 <= s2)
 436                         return crq1;
 437                 else
 438                         return crq2;
 439         }
 440
 441         /* Both requests in front of the head */
 442         if (d1 < d2)
 443                 return crq1;
 444         else if (d2 < d1)
 445                 return crq2;
 446         else {
 447                 if (s1 >= s2)
 448                         return crq1;
 449                 else
 450                         return crq2;
 451         }
 452 }
 453
 454 /*
 455  * would be nice to take fifo expire time into account as well
 456  */
 457 static struct cfq_rq *
 458 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 459                   struct cfq_rq *last)
 460 {
 461         struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
 462         struct rb_node *rbnext, *rbprev;
 463
 464         rbnext = NULL;
 465         if (ON_RB(&last->rb_node))
 466                 rbnext = rb_next(&last->rb_node);
 467         if (!rbnext) {
 468                 rbnext = rb_first(&cfqq->sort_list);
 469                 if (rbnext == &last->rb_node)
 470                         rbnext = NULL;
 471         }
 472
 473         rbprev = rb_prev(&last->rb_node);
 474
 475         if (rbprev)
 476                 crq_prev = rb_entry_crq(rbprev);
 477         if (rbnext)
 478                 crq_next = rb_entry_crq(rbnext);
 479
 480         return cfq_choose_req(cfqd, crq_next, crq_prev);
 481 }
 482
 483 static void cfq_update_next_crq(struct cfq_rq *crq)
 484 {
 485         struct cfq_queue *cfqq = crq->cfq_queue;
 486
 487         if (cfqq->next_crq == crq)
 488                 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
 489 }
 490
 491 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 492 {
 493         struct cfq_data *cfqd = cfqq->cfqd;
 494         struct list_head *list, *entry;
 495
 496         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 497
 498         list_del(&cfqq->cfq_list);
 499
 500         if (cfq_class_rt(cfqq))
 501                 list = &cfqd->cur_rr;
 502         else if (cfq_class_idle(cfqq))
 503                 list = &cfqd->idle_rr;
 504         else {
 505                 /*
 506                  * if cfqq has requests in flight, don't allow it to be
 507                  * found in cfq_set_active_queue before it has finished them.
 508                  * this is done to increase fairness between a process that
 509                  * has lots of io pending vs one that only generates one
 510                  * sporadically or synchronously
 511                  */
 512                 if (cfq_cfqq_dispatched(cfqq))
 513                         list = &cfqd->busy_rr;
 514                 else
 515                         list = &cfqd->rr_list[cfqq->ioprio];
 516         }
 517
 518         /*
 519          * if queue was preempted, just add to front to be fair. busy_rr
 520          * isn't sorted.
 521          */
 522         if (preempted || list == &cfqd->busy_rr) {
 523                 list_add(&cfqq->cfq_list, list);
 524                 return;
 525         }
 526
 527         /*
 528          * sort by when queue was last serviced
 529          */
 530         entry = list;
 531         while ((entry = entry->prev) != list) {
 532                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
 533
 534                 if (!__cfqq->service_last)
 535                         break;
 536                 if (time_before(__cfqq->service_last, cfqq->service_last))
 537                         break;
 538         }
 539
 540         list_add(&cfqq->cfq_list, entry);
 541 }
 542
 543 /*
 544  * add to busy list of queues for service, trying to be fair in ordering
 545  * the pending list according to last request service
 546  */
 547 static inline void
 548 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq, int requeue)
 549 {
 550         BUG_ON(cfq_cfqq_on_rr(cfqq));
 551         cfq_mark_cfqq_on_rr(cfqq);
 552         cfqd->busy_queues++;
 553
 554         cfq_resort_rr_list(cfqq, requeue);
 555 }
 556
 557 static inline void
 558 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 559 {
 560         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 561         cfq_clear_cfqq_on_rr(cfqq);
 562         list_move(&cfqq->cfq_list, &cfqd->empty_list);
 563
 564         BUG_ON(!cfqd->busy_queues);
 565         cfqd->busy_queues--;
 566 }
 567
 568 /*
 569  * rb tree support functions
 570  */
 571 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
 572 {
 573         struct cfq_queue *cfqq = crq->cfq_queue;
 574
 575         if (ON_RB(&crq->rb_node)) {
 576                 struct cfq_data *cfqd = cfqq->cfqd;
 577                 const int sync = cfq_crq_is_sync(crq);
 578
 579                 BUG_ON(!cfqq->queued[sync]);
 580                 cfqq->queued[sync]--;
 581
 582                 cfq_update_next_crq(crq);
 583
 584                 rb_erase(&crq->rb_node, &cfqq->sort_list);
 585                 RB_CLEAR_COLOR(&crq->rb_node);
 586
 587                 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
 588                         cfq_del_cfqq_rr(cfqd, cfqq);
 589         }
 590 }
 591
 592 static struct cfq_rq *
 593 __cfq_add_crq_rb(struct cfq_rq *crq)
 594 {
 595         struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
 596         struct rb_node *parent = NULL;
 597         struct cfq_rq *__crq;
 598
 599         while (*p) {
 600                 parent = *p;
 601                 __crq = rb_entry_crq(parent);
 602
 603                 if (crq->rb_key < __crq->rb_key)
 604                         p = &(*p)->rb_left;
 605                 else if (crq->rb_key > __crq->rb_key)
 606                         p = &(*p)->rb_right;
 607                 else
 608                         return __crq;
 609         }
 610
 611         rb_link_node(&crq->rb_node, parent, p);
 612         return NULL;
 613 }
 614
 615 static void cfq_add_crq_rb(struct cfq_rq *crq)
 616 {
 617         struct cfq_queue *cfqq = crq->cfq_queue;
 618         struct cfq_data *cfqd = cfqq->cfqd;
 619         struct request *rq = crq->request;
 620         struct cfq_rq *__alias;
 621
 622         crq->rb_key = rq_rb_key(rq);
 623         cfqq->queued[cfq_crq_is_sync(crq)]++;
 624
 625         /*
 626          * looks a little odd, but the first insert might return an alias.
 627          * if that happens, put the alias on the dispatch list
 628          */
 629         while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
 630                 cfq_dispatch_sort(cfqd->queue, __alias);
 631
 632         rb_insert_color(&crq->rb_node, &cfqq->sort_list);
 633
 634         if (!cfq_cfqq_on_rr(cfqq))
 635                 cfq_add_cfqq_rr(cfqd, cfqq, cfq_crq_requeued(crq));
 636
 637         /*
 638          * check if this request is a better next-serve candidate
 639          */
 640         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
 641 }
 642
 643 static inline void
 644 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
 645 {
 646         if (ON_RB(&crq->rb_node)) {
 647                 rb_erase(&crq->rb_node, &cfqq->sort_list);
 648                 cfqq->queued[cfq_crq_is_sync(crq)]--;
 649         }
 650
 651         cfq_add_crq_rb(crq);
 652 }
 653
 654 static struct request *cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector)
 655
 656 {
 657         struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, current->pid, CFQ_KEY_ANY);
 658         struct rb_node *n;
 659
 660         if (!cfqq)
 661                 goto out;
 662
 663         n = cfqq->sort_list.rb_node;
 664         while (n) {
 665                 struct cfq_rq *crq = rb_entry_crq(n);
 666
 667                 if (sector < crq->rb_key)
 668                         n = n->rb_left;
 669                 else if (sector > crq->rb_key)
 670                         n = n->rb_right;
 671                 else
 672                         return crq->request;
 673         }
 674
 675 out:
 676         return NULL;
 677 }
 678
 679 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 680 {
 681         struct cfq_data *cfqd = q->elevator->elevator_data;
 682         struct cfq_rq *crq = RQ_DATA(rq);
 683
 684         if (crq) {
 685                 struct cfq_queue *cfqq = crq->cfq_queue;
 686
 687                 if (cfq_crq_in_driver(crq)) {
 688                         cfq_clear_crq_in_driver(crq);
 689                         WARN_ON(!cfqd->rq_in_driver);
 690                         cfqd->rq_in_driver--;
 691                 }
 692                 if (cfq_crq_in_flight(crq)) {
 693                         const int sync = cfq_crq_is_sync(crq);
 694
 695                         cfq_clear_crq_in_flight(crq);
 696                         WARN_ON(!cfqq->on_dispatch[sync]);
 697                         cfqq->on_dispatch[sync]--;
 698                 }
 699                 cfq_mark_crq_requeued(crq);
 700         }
 701 }
 702
 703 /*
 704  * make sure the service time gets corrected on reissue of this request
 705  */
 706 static void cfq_requeue_request(request_queue_t *q, struct request *rq)
 707 {
 708         cfq_deactivate_request(q, rq);
 709         list_add(&rq->queuelist, &q->queue_head);
 710 }
 711
 712 static void cfq_remove_request(request_queue_t *q, struct request *rq)
 713 {
 714         struct cfq_rq *crq = RQ_DATA(rq);
 715
 716         if (crq) {
 717                 list_del_init(&rq->queuelist);
 718                 cfq_del_crq_rb(crq);
 719                 cfq_remove_merge_hints(q, crq);
 720
 721         }
 722 }
 723
 724 static int
 725 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 726 {
 727         struct cfq_data *cfqd = q->elevator->elevator_data;
 728         struct request *__rq;
 729         int ret;
 730
 731         ret = elv_try_last_merge(q, bio);
 732         if (ret != ELEVATOR_NO_MERGE) {
 733                 __rq = q->last_merge;
 734                 goto out_insert;
 735         }
 736
 737         __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
 738         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 739                 ret = ELEVATOR_BACK_MERGE;
 740                 goto out;
 741         }
 742
 743         __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio));
 744         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 745                 ret = ELEVATOR_FRONT_MERGE;
 746                 goto out;
 747         }
 748
 749         return ELEVATOR_NO_MERGE;
 750 out:
 751         q->last_merge = __rq;
 752 out_insert:
 753         *req = __rq;
 754         return ret;
 755 }
 756
 757 static void cfq_merged_request(request_queue_t *q, struct request *req)
 758 {
 759         struct cfq_data *cfqd = q->elevator->elevator_data;
 760         struct cfq_rq *crq = RQ_DATA(req);
 761
 762         cfq_del_crq_hash(crq);
 763         cfq_add_crq_hash(cfqd, crq);
 764
 765         if (ON_RB(&crq->rb_node) && (rq_rb_key(req) != crq->rb_key)) {
 766                 struct cfq_queue *cfqq = crq->cfq_queue;
 767
 768                 cfq_update_next_crq(crq);
 769                 cfq_reposition_crq_rb(cfqq, crq);
 770         }
 771
 772         q->last_merge = req;
 773 }
 774
 775 static void
 776 cfq_merged_requests(request_queue_t *q, struct request *rq,
 777                     struct request *next)
 778 {
 779         cfq_merged_request(q, rq);
 780
 781         /*
 782          * reposition in fifo if next is older than rq
 783          */
 784         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 785             time_before(next->start_time, rq->start_time))
 786                 list_move(&rq->queuelist, &next->queuelist);
 787
 788         cfq_remove_request(q, next);
 789 }
 790
 791 static inline void
 792 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 793 {
 794         if (cfqq) {
 795                 /*
 796                  * stop potential idle class queues waiting service
 797                  */
 798                 del_timer(&cfqd->idle_class_timer);
 799
 800                 cfqq->slice_start = jiffies;
 801                 cfqq->slice_end = 0;
 802                 cfqq->slice_left = 0;
 803                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 804                 cfq_clear_cfqq_fifo_expire(cfqq);
 805                 cfq_clear_cfqq_expired(cfqq);
 806         }
 807
 808         cfqd->active_queue = cfqq;
 809 }
 810
 811 /*
 812  * 0
 813  * 0,1
 814  * 0,1,2
 815  * 0,1,2,3
 816  * 0,1,2,3,4
 817  * 0,1,2,3,4,5
 818  * 0,1,2,3,4,5,6
 819  * 0,1,2,3,4,5,6,7
 820  */
 821 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 822 {
 823         int prio, wrap;
 824
 825         prio = -1;
 826         wrap = 0;
 827         do {
 828                 int p;
 829
 830                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 831                         if (!list_empty(&cfqd->rr_list[p])) {
 832                                 prio = p;
 833                                 break;
 834                         }
 835                 }
 836
 837                 if (prio != -1)
 838                         break;
 839                 cfqd->cur_prio = 0;
 840                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 841                         cfqd->cur_end_prio = 0;
 842                         if (wrap)
 843                                 break;
 844                         wrap = 1;
 845                 }
 846         } while (1);
 847
 848         if (unlikely(prio == -1))
 849                 return -1;
 850
 851         BUG_ON(prio >= CFQ_PRIO_LISTS);
 852
 853         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 854
 855         cfqd->cur_prio = prio + 1;
 856         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 857                 cfqd->cur_end_prio = cfqd->cur_prio;
 858                 cfqd->cur_prio = 0;
 859         }
 860         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 861                 cfqd->cur_prio = 0;
 862                 cfqd->cur_end_prio = 0;
 863         }
 864
 865         return prio;
 866 }
 867
 868 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 869 {
 870         struct cfq_queue *cfqq;
 871
 872         /*
 873          * if current queue is expired but not done with its requests yet,
 874          * wait for that to happen
 875          */
 876         if ((cfqq = cfqd->active_queue) != NULL) {
 877                 if (cfq_cfqq_expired(cfqq) && cfq_cfqq_dispatched(cfqq))
 878                         return NULL;
 879         }
 880
 881         /*
 882          * if current list is non-empty, grab first entry. if it is empty,
 883          * get next prio level and grab first entry then if any are spliced
 884          */
 885         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
 886                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 887
 888         /*
 889          * if we have idle queues and no rt or be queues had pending
 890          * requests, either allow immediate service if the grace period
 891          * has passed or arm the idle grace timer
 892          */
 893         if (!cfqq && !list_empty(&cfqd->idle_rr)) {
 894                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 895
 896                 if (time_after_eq(jiffies, end))
 897                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 898                 else
 899                         mod_timer(&cfqd->idle_class_timer, end);
 900         }
 901
 902         __cfq_set_active_queue(cfqd, cfqq);
 903         return cfqq;
 904 }
 905
 906 /*
 907  * current cfqq expired its slice (or was too idle), select new one
 908  */
 909 static void
 910 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 911                     int preempted)
 912 {
 913         unsigned long now = jiffies;
 914
 915         if (cfq_cfqq_wait_request(cfqq))
 916                 del_timer(&cfqd->idle_slice_timer);
 917
 918         if (!preempted && !cfq_cfqq_dispatched(cfqq))
 919                 cfqq->service_last = now;
 920
 921         cfq_clear_cfqq_must_dispatch(cfqq);
 922         cfq_clear_cfqq_wait_request(cfqq);
 923
 924         /*
 925          * store what was left of this slice, if the queue idled out
 926          * or was preempted
 927          */
 928         if (time_after(now, cfqq->slice_end))
 929                 cfqq->slice_left = now - cfqq->slice_end;
 930         else
 931                 cfqq->slice_left = 0;
 932
 933         if (cfq_cfqq_on_rr(cfqq))
 934                 cfq_resort_rr_list(cfqq, preempted);
 935
 936         if (cfqq == cfqd->active_queue)
 937                 cfqd->active_queue = NULL;
 938
 939         if (cfqd->active_cic) {
 940                 put_io_context(cfqd->active_cic->ioc);
 941                 cfqd->active_cic = NULL;
 942         }
 943
 944         cfqd->dispatch_slice = 0;
 945 }
 946
 947 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 948 {
 949         struct cfq_queue *cfqq = cfqd->active_queue;
 950
 951         if (cfqq) {
 952                 /*
 953                  * use deferred expiry, if there are requests in progress as
 954                  * not to disturb the slice of the next queue
 955                  */
 956                 if (cfq_cfqq_dispatched(cfqq))
 957                         cfq_mark_cfqq_expired(cfqq);
 958                 else
 959                         __cfq_slice_expired(cfqd, cfqq, preempted);
 960         }
 961 }
 962
 963 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 964
 965 {
 966         WARN_ON(!RB_EMPTY(&cfqq->sort_list));
 967         WARN_ON(cfqq != cfqd->active_queue);
 968
 969         /*
 970          * idle is disabled, either manually or by past process history
 971          */
 972         if (!cfqd->cfq_slice_idle)
 973                 return 0;
 974         if (!cfq_cfqq_idle_window(cfqq))
 975                 return 0;
 976         /*
 977          * task has exited, don't wait
 978          */
 979         if (cfqd->active_cic && !cfqd->active_cic->ioc->task)
 980                 return 0;
 981
 982         cfq_mark_cfqq_must_dispatch(cfqq);
 983         cfq_mark_cfqq_wait_request(cfqq);
 984
 985         if (!timer_pending(&cfqd->idle_slice_timer)) {
 986                 unsigned long slice_left = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 987
 988                 cfqd->idle_slice_timer.expires = jiffies + slice_left;
 989                 add_timer(&cfqd->idle_slice_timer);
 990         }
 991
 992         return 1;
 993 }
 994
 995 /*
 996  * we dispatch cfqd->cfq_quantum requests in total from the rr_list queues,
 997  * this function sector sorts the selected request to minimize seeks. we start
 998  * at cfqd->last_sector, not 0.
 999  */
1000 static void cfq_dispatch_sort(request_queue_t *q, struct cfq_rq *crq)
1001 {
1002         struct cfq_data *cfqd = q->elevator->elevator_data;
1003         struct cfq_queue *cfqq = crq->cfq_queue;
1004         struct list_head *head = &q->queue_head, *entry = head;
1005         struct request *__rq;
1006         sector_t last;
1007
1008         list_del(&crq->request->queuelist);
1009
1010         last = cfqd->last_sector;
1011         list_for_each_entry_reverse(__rq, head, queuelist) {
1012                 struct cfq_rq *__crq = RQ_DATA(__rq);
1013
1014                 if (blk_barrier_rq(__rq))
1015                         break;
1016                 if (!blk_fs_request(__rq))
1017                         break;
1018                 if (cfq_crq_requeued(__crq))
1019                         break;
1020
1021                 if (__rq->sector <= crq->request->sector)
1022                         break;
1023                 if (__rq->sector > last && crq->request->sector < last) {
1024                         last = crq->request->sector + crq->request->nr_sectors;
1025                         break;
1026                 }
1027                 entry = &__rq->queuelist;
1028         }
1029
1030         cfqd->last_sector = last;
1031
1032         cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
1033
1034         cfq_del_crq_rb(crq);
1035         cfq_remove_merge_hints(q, crq);
1036
1037         cfq_mark_crq_in_flight(crq);
1038         cfq_clear_crq_requeued(crq);
1039
1040         cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
1041         list_add_tail(&crq->request->queuelist, entry);
1042 }
1043
1044 /*
1045  * return expired entry, or NULL to just start from scratch in rbtree
1046  */
1047 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
1048 {
1049         struct cfq_data *cfqd = cfqq->cfqd;
1050         struct request *rq;
1051         struct cfq_rq *crq;
1052
1053         if (cfq_cfqq_fifo_expire(cfqq))
1054                 return NULL;
1055
1056         if (!list_empty(&cfqq->fifo)) {
1057                 int fifo = cfq_cfqq_class_sync(cfqq);
1058
1059                 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
1060                 rq = crq->request;
1061                 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
1062                         cfq_mark_cfqq_fifo_expire(cfqq);
1063                         return crq;
1064                 }
1065         }
1066
1067         return NULL;
1068 }
1069
1070 /*
1071  * Scale schedule slice based on io priority. Use the sync time slice only
1072  * if a queue is marked sync and has sync io queued. A sync queue with async
1073  * io only, should not get full sync slice length.
1074  */
1075 static inline int
1076 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1077 {
1078         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
1079
1080         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1081
1082         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
1083 }
1084
1085 static inline void
1086 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1087 {
1088         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
1089 }
1090
1091 static inline int
1092 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1093 {
1094         const int base_rq = cfqd->cfq_slice_async_rq;
1095
1096         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1097
1098         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1099 }
1100
1101 /*
1102  * get next queue for service
1103  */
1104 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd, int force)
1105 {
1106         unsigned long now = jiffies;
1107         struct cfq_queue *cfqq;
1108
1109         cfqq = cfqd->active_queue;
1110         if (!cfqq)
1111                 goto new_queue;
1112
1113         if (cfq_cfqq_expired(cfqq))
1114                 goto new_queue;
1115
1116         /*
1117          * slice has expired
1118          */
1119         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
1120                 goto expire;
1121
1122         /*
1123          * if queue has requests, dispatch one. if not, check if
1124          * enough slice is left to wait for one
1125          */
1126         if (!RB_EMPTY(&cfqq->sort_list))
1127                 goto keep_queue;
1128         else if (!force && cfq_cfqq_class_sync(cfqq) &&
1129                  time_before(now, cfqq->slice_end)) {
1130                 if (cfq_arm_slice_timer(cfqd, cfqq))
1131                         return NULL;
1132         }
1133
1134 expire:
1135         cfq_slice_expired(cfqd, 0);
1136 new_queue:
1137         cfqq = cfq_set_active_queue(cfqd);
1138 keep_queue:
1139         return cfqq;
1140 }
1141
1142 static int
1143 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1144                         int max_dispatch)
1145 {
1146         int dispatched = 0;
1147
1148         BUG_ON(RB_EMPTY(&cfqq->sort_list));
1149
1150         do {
1151                 struct cfq_rq *crq;
1152
1153                 /*
1154                  * follow expired path, else get first next available
1155                  */
1156                 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1157                         crq = cfqq->next_crq;
1158
1159                 /*
1160                  * finally, insert request into driver dispatch list
1161                  */
1162                 cfq_dispatch_sort(cfqd->queue, crq);
1163
1164                 cfqd->dispatch_slice++;
1165                 dispatched++;
1166
1167                 if (!cfqd->active_cic) {
1168                         atomic_inc(&crq->io_context->ioc->refcount);
1169                         cfqd->active_cic = crq->io_context;
1170                 }
1171
1172                 if (RB_EMPTY(&cfqq->sort_list))
1173                         break;
1174
1175         } while (dispatched < max_dispatch);
1176
1177         /*
1178          * if slice end isn't set yet, set it. if at least one request was
1179          * sync, use the sync time slice value
1180          */
1181         if (!cfqq->slice_end)
1182                 cfq_set_prio_slice(cfqd, cfqq);
1183
1184         /*
1185          * expire an async queue immediately if it has used up its slice. idle
1186          * queue always expire after 1 dispatch round.
1187          */
1188         if ((!cfq_cfqq_sync(cfqq) &&
1189             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1190             cfq_class_idle(cfqq))
1191                 cfq_slice_expired(cfqd, 0);
1192
1193         return dispatched;
1194 }
1195
1196 static int
1197 cfq_dispatch_requests(request_queue_t *q, int max_dispatch, int force)
1198 {
1199         struct cfq_data *cfqd = q->elevator->elevator_data;
1200         struct cfq_queue *cfqq;
1201
1202         if (!cfqd->busy_queues)
1203                 return 0;
1204
1205         cfqq = cfq_select_queue(cfqd, force);
1206         if (cfqq) {
1207                 cfq_clear_cfqq_must_dispatch(cfqq);
1208                 cfq_clear_cfqq_wait_request(cfqq);
1209                 del_timer(&cfqd->idle_slice_timer);
1210
1211                 if (cfq_class_idle(cfqq))
1212                         max_dispatch = 1;
1213
1214                 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1215         }
1216
1217         return 0;
1218 }
1219
1220 static inline void cfq_account_dispatch(struct cfq_rq *crq)
1221 {
1222         struct cfq_queue *cfqq = crq->cfq_queue;
1223         struct cfq_data *cfqd = cfqq->cfqd;
1224
1225         if (unlikely(!blk_fs_request(crq->request)))
1226                 return;
1227
1228         /*
1229          * accounted bit is necessary since some drivers will call
1230          * elv_next_request() many times for the same request (eg ide)
1231          */
1232         if (cfq_crq_in_driver(crq))
1233                 return;
1234
1235         cfq_mark_crq_in_driver(crq);
1236         cfqd->rq_in_driver++;
1237 }
1238
1239 static inline void
1240 cfq_account_completion(struct cfq_queue *cfqq, struct cfq_rq *crq)
1241 {
1242         struct cfq_data *cfqd = cfqq->cfqd;
1243         unsigned long now;
1244
1245         if (!cfq_crq_in_driver(crq))
1246                 return;
1247
1248         now = jiffies;
1249
1250         WARN_ON(!cfqd->rq_in_driver);
1251         cfqd->rq_in_driver--;
1252
1253         if (!cfq_class_idle(cfqq))
1254                 cfqd->last_end_request = now;
1255
1256         if (!cfq_cfqq_dispatched(cfqq)) {
1257                 if (cfq_cfqq_on_rr(cfqq)) {
1258                         cfqq->service_last = now;
1259                         cfq_resort_rr_list(cfqq, 0);
1260                 }
1261                 if (cfq_cfqq_expired(cfqq)) {
1262                         __cfq_slice_expired(cfqd, cfqq, 0);
1263                         cfq_schedule_dispatch(cfqd);
1264                 }
1265         }
1266
1267         if (cfq_crq_is_sync(crq))
1268                 crq->io_context->last_end_request = now;
1269 }
1270
1271 static struct request *cfq_next_request(request_queue_t *q)
1272 {
1273         struct cfq_data *cfqd = q->elevator->elevator_data;
1274         struct request *rq;
1275
1276         if (!list_empty(&q->queue_head)) {
1277                 struct cfq_rq *crq;
1278 dispatch:
1279                 rq = list_entry_rq(q->queue_head.next);
1280
1281                 crq = RQ_DATA(rq);
1282                 if (crq) {
1283                         struct cfq_queue *cfqq = crq->cfq_queue;
1284
1285                         /*
1286                          * if idle window is disabled, allow queue buildup
1287                          */
1288                         if (!cfq_crq_in_driver(crq) &&
1289                             !cfq_cfqq_idle_window(cfqq) &&
1290                             !blk_barrier_rq(rq) &&
1291                             cfqd->rq_in_driver >= cfqd->cfq_max_depth)
1292                                 return NULL;
1293
1294                         cfq_remove_merge_hints(q, crq);
1295                         cfq_account_dispatch(crq);
1296                 }
1297
1298                 return rq;
1299         }
1300
1301         if (cfq_dispatch_requests(q, cfqd->cfq_quantum, 0))
1302                 goto dispatch;
1303
1304         return NULL;
1305 }
1306
1307 /*
1308  * task holds one reference to the queue, dropped when task exits. each crq
1309  * in-flight on this queue also holds a reference, dropped when crq is freed.
1310  *
1311  * queue lock must be held here.
1312  */
1313 static void cfq_put_queue(struct cfq_queue *cfqq)
1314 {
1315         struct cfq_data *cfqd = cfqq->cfqd;
1316
1317         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1318
1319         if (!atomic_dec_and_test(&cfqq->ref))
1320                 return;
1321
1322         BUG_ON(rb_first(&cfqq->sort_list));
1323         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1324         BUG_ON(cfq_cfqq_on_rr(cfqq));
1325
1326         if (unlikely(cfqd->active_queue == cfqq)) {
1327                 __cfq_slice_expired(cfqd, cfqq, 0);
1328                 cfq_schedule_dispatch(cfqd);
1329         }
1330
1331         cfq_put_cfqd(cfqq->cfqd);
1332
1333         /*
1334          * it's on the empty list and still hashed
1335          */
1336         list_del(&cfqq->cfq_list);
1337         hlist_del(&cfqq->cfq_hash);
1338         kmem_cache_free(cfq_pool, cfqq);
1339 }
1340
1341 static inline struct cfq_queue *
1342 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1343                     const int hashval)
1344 {
1345         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1346         struct hlist_node *entry, *next;
1347
1348         hlist_for_each_safe(entry, next, hash_list) {
1349                 struct cfq_queue *__cfqq = list_entry_qhash(entry);
1350                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->ioprio_class, __cfqq->ioprio);
1351
1352                 if (__cfqq->key == key && (__p == prio || prio == CFQ_KEY_ANY))
1353                         return __cfqq;
1354         }
1355
1356         return NULL;
1357 }
1358
1359 static struct cfq_queue *
1360 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1361 {
1362         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1363 }
1364
1365 static void cfq_free_io_context(struct cfq_io_context *cic)
1366 {
1367         struct cfq_io_context *__cic;
1368         struct list_head *entry, *next;
1369
1370         list_for_each_safe(entry, next, &cic->list) {
1371                 __cic = list_entry(entry, struct cfq_io_context, list);
1372                 kmem_cache_free(cfq_ioc_pool, __cic);
1373         }
1374
1375         kmem_cache_free(cfq_ioc_pool, cic);
1376 }
1377
1378 /*
1379  * Called with interrupts disabled
1380  */
1381 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1382 {
1383         struct cfq_data *cfqd = cic->cfqq->cfqd;
1384         request_queue_t *q = cfqd->queue;
1385
1386         WARN_ON(!irqs_disabled());
1387
1388         spin_lock(q->queue_lock);
1389
1390         if (unlikely(cic->cfqq == cfqd->active_queue)) {
1391                 __cfq_slice_expired(cfqd, cic->cfqq, 0);
1392                 cfq_schedule_dispatch(cfqd);
1393         }
1394
1395         cfq_put_queue(cic->cfqq);
1396         cic->cfqq = NULL;
1397         spin_unlock(q->queue_lock);
1398 }
1399
1400 /*
1401  * Another task may update the task cic list, if it is doing a queue lookup
1402  * on its behalf. cfq_cic_lock excludes such concurrent updates
1403  */
1404 static void cfq_exit_io_context(struct cfq_io_context *cic)
1405 {
1406         struct cfq_io_context *__cic;
1407         struct list_head *entry;
1408         unsigned long flags;
1409
1410         local_irq_save(flags);
1411
1412         /*
1413          * put the reference this task is holding to the various queues
1414          */
1415         list_for_each(entry, &cic->list) {
1416                 __cic = list_entry(entry, struct cfq_io_context, list);
1417                 cfq_exit_single_io_context(__cic);
1418         }
1419
1420         cfq_exit_single_io_context(cic);
1421         local_irq_restore(flags);
1422 }
1423
1424 static struct cfq_io_context *
1425 cfq_alloc_io_context(struct cfq_data *cfqd, int gfp_mask)
1426 {
1427         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1428
1429         if (cic) {
1430                 INIT_LIST_HEAD(&cic->list);
1431                 cic->cfqq = NULL;
1432                 cic->key = NULL;
1433                 cic->last_end_request = jiffies;
1434                 cic->ttime_total = 0;
1435                 cic->ttime_samples = 0;
1436                 cic->ttime_mean = 0;
1437                 cic->dtor = cfq_free_io_context;
1438                 cic->exit = cfq_exit_io_context;
1439         }
1440
1441         return cic;
1442 }
1443
1444 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1445 {
1446         struct task_struct *tsk = current;
1447         int ioprio_class;
1448
1449         if (!cfq_cfqq_prio_changed(cfqq))
1450                 return;
1451
1452         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1453         switch (ioprio_class) {
1454                 default:
1455                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1456                 case IOPRIO_CLASS_NONE:
1457                         /*
1458                          * no prio set, place us in the middle of the BE classes
1459                          */
1460                         cfqq->ioprio = task_nice_ioprio(tsk);
1461                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1462                         break;
1463                 case IOPRIO_CLASS_RT:
1464                         cfqq->ioprio = task_ioprio(tsk);
1465                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1466                         break;
1467                 case IOPRIO_CLASS_BE:
1468                         cfqq->ioprio = task_ioprio(tsk);
1469                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1470                         break;
1471                 case IOPRIO_CLASS_IDLE:
1472                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1473                         cfqq->ioprio = 7;
1474                         cfq_clear_cfqq_idle_window(cfqq);
1475                         break;
1476         }
1477
1478         /*
1479          * keep track of original prio settings in case we have to temporarily
1480          * elevate the priority of this queue
1481          */
1482         cfqq->org_ioprio = cfqq->ioprio;
1483         cfqq->org_ioprio_class = cfqq->ioprio_class;
1484
1485         if (cfq_cfqq_on_rr(cfqq))
1486                 cfq_resort_rr_list(cfqq, 0);
1487
1488         cfq_clear_cfqq_prio_changed(cfqq);
1489 }
1490
1491 static inline void changed_ioprio(struct cfq_queue *cfqq)
1492 {
1493         if (cfqq) {
1494                 struct cfq_data *cfqd = cfqq->cfqd;
1495
1496                 spin_lock(cfqd->queue->queue_lock);
1497                 cfq_mark_cfqq_prio_changed(cfqq);
1498                 cfq_init_prio_data(cfqq);
1499                 spin_unlock(cfqd->queue->queue_lock);
1500         }
1501 }
1502
1503 /*
1504  * callback from sys_ioprio_set, irqs are disabled
1505  */
1506 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1507 {
1508         struct cfq_io_context *cic = ioc->cic;
1509
1510         changed_ioprio(cic->cfqq);
1511
1512         list_for_each_entry(cic, &cic->list, list)
1513                 changed_ioprio(cic->cfqq);
1514
1515         return 0;
1516 }
1517
1518 static struct cfq_queue *
1519 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, unsigned short ioprio,
1520               int gfp_mask)
1521 {
1522         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1523         struct cfq_queue *cfqq, *new_cfqq = NULL;
1524
1525 retry:
1526         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1527
1528         if (!cfqq) {
1529                 if (new_cfqq) {
1530                         cfqq = new_cfqq;
1531                         new_cfqq = NULL;
1532                 } else if (gfp_mask & __GFP_WAIT) {
1533                         spin_unlock_irq(cfqd->queue->queue_lock);
1534                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1535                         spin_lock_irq(cfqd->queue->queue_lock);
1536                         goto retry;
1537                 } else {
1538                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1539                         if (!cfqq)
1540                                 goto out;
1541                 }
1542
1543                 memset(cfqq, 0, sizeof(*cfqq));
1544
1545                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1546                 INIT_LIST_HEAD(&cfqq->cfq_list);
1547                 RB_CLEAR_ROOT(&cfqq->sort_list);
1548                 INIT_LIST_HEAD(&cfqq->fifo);
1549
1550                 cfqq->key = key;
1551                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1552                 atomic_set(&cfqq->ref, 0);
1553                 cfqq->cfqd = cfqd;
1554                 atomic_inc(&cfqd->ref);
1555                 cfqq->service_last = 0;
1556                 /*
1557                  * set ->slice_left to allow preemption for a new process
1558                  */
1559                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1560                 cfq_mark_cfqq_idle_window(cfqq);
1561                 cfq_mark_cfqq_prio_changed(cfqq);
1562                 cfq_init_prio_data(cfqq);
1563         }
1564
1565         if (new_cfqq)
1566                 kmem_cache_free(cfq_pool, new_cfqq);
1567
1568         atomic_inc(&cfqq->ref);
1569 out:
1570         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1571         return cfqq;
1572 }
1573
1574 /*
1575  * Setup general io context and cfq io context. There can be several cfq
1576  * io contexts per general io context, if this process is doing io to more
1577  * than one device managed by cfq. Note that caller is holding a reference to
1578  * cfqq, so we don't need to worry about it disappearing
1579  */
1580 static struct cfq_io_context *
1581 cfq_get_io_context(struct cfq_data *cfqd, pid_t pid, int gfp_mask)
1582 {
1583         struct io_context *ioc = NULL;
1584         struct cfq_io_context *cic;
1585
1586         might_sleep_if(gfp_mask & __GFP_WAIT);
1587
1588         ioc = get_io_context(gfp_mask);
1589         if (!ioc)
1590                 return NULL;
1591
1592         if ((cic = ioc->cic) == NULL) {
1593                 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1594
1595                 if (cic == NULL)
1596                         goto err;
1597
1598                 /*
1599                  * manually increment generic io_context usage count, it
1600                  * cannot go away since we are already holding one ref to it
1601                  */
1602                 ioc->cic = cic;
1603                 ioc->set_ioprio = cfq_ioc_set_ioprio;
1604                 cic->ioc = ioc;
1605                 cic->key = cfqd;
1606                 atomic_inc(&cfqd->ref);
1607         } else {
1608                 struct cfq_io_context *__cic;
1609
1610                 /*
1611                  * the first cic on the list is actually the head itself
1612                  */
1613                 if (cic->key == cfqd)
1614                         goto out;
1615
1616                 /*
1617                  * cic exists, check if we already are there. linear search
1618                  * should be ok here, the list will usually not be more than
1619                  * 1 or a few entries long
1620                  */
1621                 list_for_each_entry(__cic, &cic->list, list) {
1622                         /*
1623                          * this process is already holding a reference to
1624                          * this queue, so no need to get one more
1625                          */
1626                         if (__cic->key == cfqd) {
1627                                 cic = __cic;
1628                                 goto out;
1629                         }
1630                 }
1631
1632                 /*
1633                  * nope, process doesn't have a cic assoicated with this
1634                  * cfqq yet. get a new one and add to list
1635                  */
1636                 __cic = cfq_alloc_io_context(cfqd, gfp_mask);
1637                 if (__cic == NULL)
1638                         goto err;
1639
1640                 __cic->ioc = ioc;
1641                 __cic->key = cfqd;
1642                 atomic_inc(&cfqd->ref);
1643                 list_add(&__cic->list, &cic->list);
1644                 cic = __cic;
1645         }
1646
1647 out:
1648         return cic;
1649 err:
1650         put_io_context(ioc);
1651         return NULL;
1652 }
1653
1654 static void
1655 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1656 {
1657         unsigned long elapsed, ttime;
1658
1659         /*
1660          * if this context already has stuff queued, thinktime is from
1661          * last queue not last end
1662          */
1663 #if 0
1664         if (time_after(cic->last_end_request, cic->last_queue))
1665                 elapsed = jiffies - cic->last_end_request;
1666         else
1667                 elapsed = jiffies - cic->last_queue;
1668 #else
1669                 elapsed = jiffies - cic->last_end_request;
1670 #endif
1671
1672         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1673
1674         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1675         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1676         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1677 }
1678
1679 #define sample_valid(samples)   ((samples) > 80)
1680
1681 /*
1682  * Disable idle window if the process thinks too long or seeks so much that
1683  * it doesn't matter
1684  */
1685 static void
1686 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1687                        struct cfq_io_context *cic)
1688 {
1689         int enable_idle = cfq_cfqq_idle_window(cfqq);
1690
1691         if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1692                 enable_idle = 0;
1693         else if (sample_valid(cic->ttime_samples)) {
1694                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1695                         enable_idle = 0;
1696                 else
1697                         enable_idle = 1;
1698         }
1699
1700         if (enable_idle)
1701                 cfq_mark_cfqq_idle_window(cfqq);
1702         else
1703                 cfq_clear_cfqq_idle_window(cfqq);
1704 }
1705
1706
1707 /*
1708  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1709  * no or if we aren't sure, a 1 will cause a preempt.
1710  */
1711 static int
1712 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1713                    struct cfq_rq *crq)
1714 {
1715         struct cfq_queue *cfqq = cfqd->active_queue;
1716
1717         if (cfq_class_idle(new_cfqq))
1718                 return 0;
1719
1720         if (!cfqq)
1721                 return 1;
1722
1723         if (cfq_class_idle(cfqq))
1724                 return 1;
1725         if (!cfq_cfqq_wait_request(new_cfqq))
1726                 return 0;
1727         /*
1728          * if it doesn't have slice left, forget it
1729          */
1730         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1731                 return 0;
1732         if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1733                 return 1;
1734
1735         return 0;
1736 }
1737
1738 /*
1739  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1740  * let it have half of its nominal slice.
1741  */
1742 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1743 {
1744         struct cfq_queue *__cfqq, *next;
1745
1746         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1747                 cfq_resort_rr_list(__cfqq, 1);
1748
1749         if (!cfqq->slice_left)
1750                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1751
1752         cfqq->slice_end = cfqq->slice_left + jiffies;
1753         __cfq_slice_expired(cfqd, cfqq, 1);
1754         __cfq_set_active_queue(cfqd, cfqq);
1755 }
1756
1757 /*
1758  * should really be a ll_rw_blk.c helper
1759  */
1760 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1761 {
1762         request_queue_t *q = cfqd->queue;
1763
1764         if (!blk_queue_plugged(q))
1765                 q->request_fn(q);
1766         else
1767                 __generic_unplug_device(q);
1768 }
1769
1770 /*
1771  * Called when a new fs request (crq) is added (to cfqq). Check if there's
1772  * something we should do about it
1773  */
1774 static void
1775 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1776                  struct cfq_rq *crq)
1777 {
1778         struct cfq_io_context *cic;
1779
1780         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1781
1782         /*
1783          * we never wait for an async request and we don't allow preemption
1784          * of an async request. so just return early
1785          */
1786         if (!cfq_crq_is_sync(crq))
1787                 return;
1788
1789         cic = crq->io_context;
1790
1791         cfq_update_io_thinktime(cfqd, cic);
1792         cfq_update_idle_window(cfqd, cfqq, cic);
1793
1794         cic->last_queue = jiffies;
1795
1796         if (cfqq == cfqd->active_queue) {
1797                 /*
1798                  * if we are waiting for a request for this queue, let it rip
1799                  * immediately and flag that we must not expire this queue
1800                  * just now
1801                  */
1802                 if (cfq_cfqq_wait_request(cfqq)) {
1803                         cfq_mark_cfqq_must_dispatch(cfqq);
1804                         del_timer(&cfqd->idle_slice_timer);
1805                         cfq_start_queueing(cfqd, cfqq);
1806                 }
1807         } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1808                 /*
1809                  * not the active queue - expire current slice if it is
1810                  * idle and has expired it's mean thinktime or this new queue
1811                  * has some old slice time left and is of higher priority
1812                  */
1813                 cfq_preempt_queue(cfqd, cfqq);
1814                 cfq_mark_cfqq_must_dispatch(cfqq);
1815                 cfq_start_queueing(cfqd, cfqq);
1816         }
1817 }
1818
1819 static void cfq_enqueue(struct cfq_data *cfqd, struct request *rq)
1820 {
1821         struct cfq_rq *crq = RQ_DATA(rq);
1822         struct cfq_queue *cfqq = crq->cfq_queue;
1823
1824         cfq_init_prio_data(cfqq);
1825
1826         cfq_add_crq_rb(crq);
1827
1828         list_add_tail(&rq->queuelist, &cfqq->fifo);
1829
1830         if (rq_mergeable(rq)) {
1831                 cfq_add_crq_hash(cfqd, crq);
1832
1833                 if (!cfqd->queue->last_merge)
1834                         cfqd->queue->last_merge = rq;
1835         }
1836
1837         cfq_crq_enqueued(cfqd, cfqq, crq);
1838 }
1839
1840 static void
1841 cfq_insert_request(request_queue_t *q, struct request *rq, int where)
1842 {
1843         struct cfq_data *cfqd = q->elevator->elevator_data;
1844
1845         switch (where) {
1846                 case ELEVATOR_INSERT_BACK:
1847                         while (cfq_dispatch_requests(q, INT_MAX, 1))
1848                                 ;
1849                         list_add_tail(&rq->queuelist, &q->queue_head);
1850                         /*
1851                          * If we were idling with pending requests on
1852                          * inactive cfqqs, force dispatching will
1853                          * remove the idle timer and the queue won't
1854                          * be kicked by __make_request() afterward.
1855                          * Kick it here.
1856                          */
1857                         cfq_schedule_dispatch(cfqd);
1858                         break;
1859                 case ELEVATOR_INSERT_FRONT:
1860                         list_add(&rq->queuelist, &q->queue_head);
1861                         break;
1862                 case ELEVATOR_INSERT_SORT:
1863                         BUG_ON(!blk_fs_request(rq));
1864                         cfq_enqueue(cfqd, rq);
1865                         break;
1866                 default:
1867                         printk("%s: bad insert point %d\n", __FUNCTION__,where);
1868                         return;
1869         }
1870 }
1871
1872 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1873 {
1874         struct cfq_rq *crq = RQ_DATA(rq);
1875         struct cfq_queue *cfqq;
1876
1877         if (unlikely(!blk_fs_request(rq)))
1878                 return;
1879
1880         cfqq = crq->cfq_queue;
1881
1882         if (cfq_crq_in_flight(crq)) {
1883                 const int sync = cfq_crq_is_sync(crq);
1884
1885                 WARN_ON(!cfqq->on_dispatch[sync]);
1886                 cfqq->on_dispatch[sync]--;
1887         }
1888
1889         cfq_account_completion(cfqq, crq);
1890 }
1891
1892 static struct request *
1893 cfq_former_request(request_queue_t *q, struct request *rq)
1894 {
1895         struct cfq_rq *crq = RQ_DATA(rq);
1896         struct rb_node *rbprev = rb_prev(&crq->rb_node);
1897
1898         if (rbprev)
1899                 return rb_entry_crq(rbprev)->request;
1900
1901         return NULL;
1902 }
1903
1904 static struct request *
1905 cfq_latter_request(request_queue_t *q, struct request *rq)
1906 {
1907         struct cfq_rq *crq = RQ_DATA(rq);
1908         struct rb_node *rbnext = rb_next(&crq->rb_node);
1909
1910         if (rbnext)
1911                 return rb_entry_crq(rbnext)->request;
1912
1913         return NULL;
1914 }
1915
1916 /*
1917  * we temporarily boost lower priority queues if they are holding fs exclusive
1918  * resources. they are boosted to normal prio (CLASS_BE/4)
1919  */
1920 static void cfq_prio_boost(struct cfq_queue *cfqq)
1921 {
1922         const int ioprio_class = cfqq->ioprio_class;
1923         const int ioprio = cfqq->ioprio;
1924
1925         if (has_fs_excl()) {
1926                 /*
1927                  * boost idle prio on transactions that would lock out other
1928                  * users of the filesystem
1929                  */
1930                 if (cfq_class_idle(cfqq))
1931                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1932                 if (cfqq->ioprio > IOPRIO_NORM)
1933                         cfqq->ioprio = IOPRIO_NORM;
1934         } else {
1935                 /*
1936                  * check if we need to unboost the queue
1937                  */
1938                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1939                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1940                 if (cfqq->ioprio != cfqq->org_ioprio)
1941                         cfqq->ioprio = cfqq->org_ioprio;
1942         }
1943
1944         /*
1945          * refile between round-robin lists if we moved the priority class
1946          */
1947         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1948             cfq_cfqq_on_rr(cfqq))
1949                 cfq_resort_rr_list(cfqq, 0);
1950 }
1951
1952 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
1953 {
1954         if (rw == READ || process_sync(task))
1955                 return task->pid;
1956
1957         return CFQ_KEY_ASYNC;
1958 }
1959
1960 static inline int
1961 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1962                 struct task_struct *task, int rw)
1963 {
1964 #if 1
1965         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1966             !cfq_cfqq_must_alloc_slice(cfqq)) {
1967                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1968                 return ELV_MQUEUE_MUST;
1969         }
1970
1971         return ELV_MQUEUE_MAY;
1972 #else
1973         if (!cfqq || task->flags & PF_MEMALLOC)
1974                 return ELV_MQUEUE_MAY;
1975         if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1976                 if (cfq_cfqq_wait_request(cfqq))
1977                         return ELV_MQUEUE_MUST;
1978
1979                 /*
1980                  * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1981                  * can quickly flood the queue with writes from a single task
1982                  */
1983                 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
1984                         cfq_mark_cfqq_must_alloc_slice(cfqq);
1985                         return ELV_MQUEUE_MUST;
1986                 }
1987
1988                 return ELV_MQUEUE_MAY;
1989         }
1990         if (cfq_class_idle(cfqq))
1991                 return ELV_MQUEUE_NO;
1992         if (cfqq->allocated[rw] >= cfqd->max_queued) {
1993                 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1994                 int ret = ELV_MQUEUE_NO;
1995
1996                 if (ioc && ioc->nr_batch_requests)
1997                         ret = ELV_MQUEUE_MAY;
1998
1999                 put_io_context(ioc);
2000                 return ret;
2001         }
2002
2003         return ELV_MQUEUE_MAY;
2004 #endif
2005 }
2006
2007 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
2008 {
2009         struct cfq_data *cfqd = q->elevator->elevator_data;
2010         struct task_struct *tsk = current;
2011         struct cfq_queue *cfqq;
2012
2013         /*
2014          * don't force setup of a queue from here, as a call to may_queue
2015          * does not necessarily imply that a request actually will be queued.
2016          * so just lookup a possibly existing queue, or return 'may queue'
2017          * if that fails
2018          */
2019         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
2020         if (cfqq) {
2021                 cfq_init_prio_data(cfqq);
2022                 cfq_prio_boost(cfqq);
2023
2024                 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
2025         }
2026
2027         return ELV_MQUEUE_MAY;
2028 }
2029
2030 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
2031 {
2032         struct cfq_data *cfqd = q->elevator->elevator_data;
2033         struct request_list *rl = &q->rq;
2034
2035         if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
2036                 smp_mb();
2037                 if (waitqueue_active(&rl->wait[READ]))
2038                         wake_up(&rl->wait[READ]);
2039         }
2040
2041         if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
2042                 smp_mb();
2043                 if (waitqueue_active(&rl->wait[WRITE]))
2044                         wake_up(&rl->wait[WRITE]);
2045         }
2046 }
2047
2048 /*
2049  * queue lock held here
2050  */
2051 static void cfq_put_request(request_queue_t *q, struct request *rq)
2052 {
2053         struct cfq_data *cfqd = q->elevator->elevator_data;
2054         struct cfq_rq *crq = RQ_DATA(rq);
2055
2056         if (crq) {
2057                 struct cfq_queue *cfqq = crq->cfq_queue;
2058                 const int rw = rq_data_dir(rq);
2059
2060                 BUG_ON(!cfqq->allocated[rw]);
2061                 cfqq->allocated[rw]--;
2062
2063                 put_io_context(crq->io_context->ioc);
2064
2065                 mempool_free(crq, cfqd->crq_pool);
2066                 rq->elevator_private = NULL;
2067
2068                 cfq_check_waiters(q, cfqq);
2069                 cfq_put_queue(cfqq);
2070         }
2071 }
2072
2073 /*
2074  * Allocate cfq data structures associated with this request.
2075  */
2076 static int
2077 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
2078                 int gfp_mask)
2079 {
2080         struct cfq_data *cfqd = q->elevator->elevator_data;
2081         struct task_struct *tsk = current;
2082         struct cfq_io_context *cic;
2083         const int rw = rq_data_dir(rq);
2084         pid_t key = cfq_queue_pid(tsk, rw);
2085         struct cfq_queue *cfqq;
2086         struct cfq_rq *crq;
2087         unsigned long flags;
2088
2089         might_sleep_if(gfp_mask & __GFP_WAIT);
2090
2091         cic = cfq_get_io_context(cfqd, key, gfp_mask);
2092
2093         spin_lock_irqsave(q->queue_lock, flags);
2094
2095         if (!cic)
2096                 goto queue_fail;
2097
2098         if (!cic->cfqq) {
2099                 cfqq = cfq_get_queue(cfqd, key, tsk->ioprio, gfp_mask);
2100                 if (!cfqq)
2101                         goto queue_fail;
2102
2103                 cic->cfqq = cfqq;
2104         } else
2105                 cfqq = cic->cfqq;
2106
2107         cfqq->allocated[rw]++;
2108         cfq_clear_cfqq_must_alloc(cfqq);
2109         cfqd->rq_starved = 0;
2110         atomic_inc(&cfqq->ref);
2111         spin_unlock_irqrestore(q->queue_lock, flags);
2112
2113         crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
2114         if (crq) {
2115                 RB_CLEAR(&crq->rb_node);
2116                 crq->rb_key = 0;
2117                 crq->request = rq;
2118                 INIT_HLIST_NODE(&crq->hash);
2119                 crq->cfq_queue = cfqq;
2120                 crq->io_context = cic;
2121                 cfq_clear_crq_in_flight(crq);
2122                 cfq_clear_crq_in_driver(crq);
2123                 cfq_clear_crq_requeued(crq);
2124
2125                 if (rw == READ || process_sync(tsk))
2126                         cfq_mark_crq_is_sync(crq);
2127                 else
2128                         cfq_clear_crq_is_sync(crq);
2129
2130                 rq->elevator_private = crq;
2131                 return 0;
2132         }
2133
2134         spin_lock_irqsave(q->queue_lock, flags);
2135         cfqq->allocated[rw]--;
2136         if (!(cfqq->allocated[0] + cfqq->allocated[1]))
2137                 cfq_mark_cfqq_must_alloc(cfqq);
2138         cfq_put_queue(cfqq);
2139 queue_fail:
2140         if (cic)
2141                 put_io_context(cic->ioc);
2142         /*
2143          * mark us rq allocation starved. we need to kickstart the process
2144          * ourselves if there are no pending requests that can do it for us.
2145          * that would be an extremely rare OOM situation
2146          */
2147         cfqd->rq_starved = 1;
2148         cfq_schedule_dispatch(cfqd);
2149         spin_unlock_irqrestore(q->queue_lock, flags);
2150         return 1;
2151 }
2152
2153 static void cfq_kick_queue(void *data)
2154 {
2155         request_queue_t *q = data;
2156         struct cfq_data *cfqd = q->elevator->elevator_data;
2157         unsigned long flags;
2158
2159         spin_lock_irqsave(q->queue_lock, flags);
2160
2161         if (cfqd->rq_starved) {
2162                 struct request_list *rl = &q->rq;
2163
2164                 /*
2165                  * we aren't guaranteed to get a request after this, but we
2166                  * have to be opportunistic
2167                  */
2168                 smp_mb();
2169                 if (waitqueue_active(&rl->wait[READ]))
2170                         wake_up(&rl->wait[READ]);
2171                 if (waitqueue_active(&rl->wait[WRITE]))
2172                         wake_up(&rl->wait[WRITE]);
2173         }
2174
2175         blk_remove_plug(q);
2176         q->request_fn(q);
2177         spin_unlock_irqrestore(q->queue_lock, flags);
2178 }
2179
2180 /*
2181  * Timer running if the active_queue is currently idling inside its time slice
2182  */
2183 static void cfq_idle_slice_timer(unsigned long data)
2184 {
2185         struct cfq_data *cfqd = (struct cfq_data *) data;
2186         struct cfq_queue *cfqq;
2187         unsigned long flags;
2188
2189         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2190
2191         if ((cfqq = cfqd->active_queue) != NULL) {
2192                 unsigned long now = jiffies;
2193
2194                 /*
2195                  * expired
2196                  */
2197                 if (time_after(now, cfqq->slice_end))
2198                         goto expire;
2199
2200                 /*
2201                  * only expire and reinvoke request handler, if there are
2202                  * other queues with pending requests
2203                  */
2204                 if (!cfq_pending_requests(cfqd)) {
2205                         cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2206                         add_timer(&cfqd->idle_slice_timer);
2207                         goto out_cont;
2208                 }
2209
2210                 /*
2211                  * not expired and it has a request pending, let it dispatch
2212                  */
2213                 if (!RB_EMPTY(&cfqq->sort_list)) {
2214                         cfq_mark_cfqq_must_dispatch(cfqq);
2215                         goto out_kick;
2216                 }
2217         }
2218 expire:
2219         cfq_slice_expired(cfqd, 0);
2220 out_kick:
2221         cfq_schedule_dispatch(cfqd);
2222 out_cont:
2223         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2224 }
2225
2226 /*
2227  * Timer running if an idle class queue is waiting for service
2228  */
2229 static void cfq_idle_class_timer(unsigned long data)
2230 {
2231         struct cfq_data *cfqd = (struct cfq_data *) data;
2232         unsigned long flags, end;
2233
2234         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2235
2236         /*
2237          * race with a non-idle queue, reset timer
2238          */
2239         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2240         if (!time_after_eq(jiffies, end)) {
2241                 cfqd->idle_class_timer.expires = end;
2242                 add_timer(&cfqd->idle_class_timer);
2243         } else
2244                 cfq_schedule_dispatch(cfqd);
2245
2246         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2247 }
2248
2249 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2250 {
2251         del_timer_sync(&cfqd->idle_slice_timer);
2252         del_timer_sync(&cfqd->idle_class_timer);
2253         blk_sync_queue(cfqd->queue);
2254 }
2255
2256 static void cfq_put_cfqd(struct cfq_data *cfqd)
2257 {
2258         request_queue_t *q = cfqd->queue;
2259
2260         if (!atomic_dec_and_test(&cfqd->ref))
2261                 return;
2262
2263         cfq_shutdown_timer_wq(cfqd);
2264         q->elevator->elevator_data = NULL;
2265
2266         mempool_destroy(cfqd->crq_pool);
2267         kfree(cfqd->crq_hash);
2268         kfree(cfqd->cfq_hash);
2269         kfree(cfqd);
2270 }
2271
2272 static void cfq_exit_queue(elevator_t *e)
2273 {
2274         struct cfq_data *cfqd = e->elevator_data;
2275
2276         cfq_shutdown_timer_wq(cfqd);
2277         cfq_put_cfqd(cfqd);
2278 }
2279
2280 static int cfq_init_queue(request_queue_t *q, elevator_t *e)
2281 {
2282         struct cfq_data *cfqd;
2283         int i;
2284
2285         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2286         if (!cfqd)
2287                 return -ENOMEM;
2288
2289         memset(cfqd, 0, sizeof(*cfqd));
2290
2291         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2292                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2293
2294         INIT_LIST_HEAD(&cfqd->busy_rr);
2295         INIT_LIST_HEAD(&cfqd->cur_rr);
2296         INIT_LIST_HEAD(&cfqd->idle_rr);
2297         INIT_LIST_HEAD(&cfqd->empty_list);
2298
2299         cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2300         if (!cfqd->crq_hash)
2301                 goto out_crqhash;
2302
2303         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2304         if (!cfqd->cfq_hash)
2305                 goto out_cfqhash;
2306
2307         cfqd->crq_pool = mempool_create(BLKDEV_MIN_RQ, mempool_alloc_slab, mempool_free_slab, crq_pool);
2308         if (!cfqd->crq_pool)
2309                 goto out_crqpool;
2310
2311         for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2312                 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2313         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2314                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2315
2316         e->elevator_data = cfqd;
2317
2318         cfqd->queue = q;
2319
2320         cfqd->max_queued = q->nr_requests / 4;
2321         q->nr_batching = cfq_queued;
2322
2323         init_timer(&cfqd->idle_slice_timer);
2324         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2325         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2326
2327         init_timer(&cfqd->idle_class_timer);
2328         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2329         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2330
2331         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2332
2333         atomic_set(&cfqd->ref, 1);
2334
2335         cfqd->cfq_queued = cfq_queued;
2336         cfqd->cfq_quantum = cfq_quantum;
2337         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2338         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2339         cfqd->cfq_back_max = cfq_back_max;
2340         cfqd->cfq_back_penalty = cfq_back_penalty;
2341         cfqd->cfq_slice[0] = cfq_slice_async;
2342         cfqd->cfq_slice[1] = cfq_slice_sync;
2343         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2344         cfqd->cfq_slice_idle = cfq_slice_idle;
2345         cfqd->cfq_max_depth = cfq_max_depth;
2346
2347         return 0;
2348 out_crqpool:
2349         kfree(cfqd->cfq_hash);
2350 out_cfqhash:
2351         kfree(cfqd->crq_hash);
2352 out_crqhash:
2353         kfree(cfqd);
2354         return -ENOMEM;
2355 }
2356
2357 static void cfq_slab_kill(void)
2358 {
2359         if (crq_pool)
2360                 kmem_cache_destroy(crq_pool);
2361         if (cfq_pool)
2362                 kmem_cache_destroy(cfq_pool);
2363         if (cfq_ioc_pool)
2364                 kmem_cache_destroy(cfq_ioc_pool);
2365 }
2366
2367 static int __init cfq_slab_setup(void)
2368 {
2369         crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2370                                         NULL, NULL);
2371         if (!crq_pool)
2372                 goto fail;
2373
2374         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2375                                         NULL, NULL);
2376         if (!cfq_pool)
2377                 goto fail;
2378
2379         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2380                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2381         if (!cfq_ioc_pool)
2382                 goto fail;
2383
2384         return 0;
2385 fail:
2386         cfq_slab_kill();
2387         return -ENOMEM;
2388 }
2389
2390 /*
2391  * sysfs parts below -->
2392  */
2393 struct cfq_fs_entry {
2394         struct attribute attr;
2395         ssize_t (*show)(struct cfq_data *, char *);
2396         ssize_t (*store)(struct cfq_data *, const char *, size_t);
2397 };
2398
2399 static ssize_t
2400 cfq_var_show(unsigned int var, char *page)
2401 {
2402         return sprintf(page, "%d\n", var);
2403 }
2404
2405 static ssize_t
2406 cfq_var_store(unsigned int *var, const char *page, size_t count)
2407 {
2408         char *p = (char *) page;
2409
2410         *var = simple_strtoul(p, &p, 10);
2411         return count;
2412 }
2413
2414 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2415 static ssize_t __FUNC(struct cfq_data *cfqd, char *page)                \
2416 {                                                                       \
2417         unsigned int __data = __VAR;                                    \
2418         if (__CONV)                                                     \
2419                 __data = jiffies_to_msecs(__data);                      \
2420         return cfq_var_show(__data, (page));                            \
2421 }
2422 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2423 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2424 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2425 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2426 SHOW_FUNCTION(cfq_back_max_show, cfqd->cfq_back_max, 0);
2427 SHOW_FUNCTION(cfq_back_penalty_show, cfqd->cfq_back_penalty, 0);
2428 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2429 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2430 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2431 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2432 SHOW_FUNCTION(cfq_max_depth_show, cfqd->cfq_max_depth, 0);
2433 #undef SHOW_FUNCTION
2434
2435 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2436 static ssize_t __FUNC(struct cfq_data *cfqd, const char *page, size_t count)    \
2437 {                                                                       \
2438         unsigned int __data;                                            \
2439         int ret = cfq_var_store(&__data, (page), count);                \
2440         if (__data < (MIN))                                             \
2441                 __data = (MIN);                                         \
2442         else if (__data > (MAX))                                        \
2443                 __data = (MAX);                                         \
2444         if (__CONV)                                                     \
2445                 *(__PTR) = msecs_to_jiffies(__data);                    \
2446         else                                                            \
2447                 *(__PTR) = __data;                                      \
2448         return ret;                                                     \
2449 }
2450 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2451 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2452 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2453 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2454 STORE_FUNCTION(cfq_back_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2455 STORE_FUNCTION(cfq_back_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2456 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2457 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2458 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2459 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2460 STORE_FUNCTION(cfq_max_depth_store, &cfqd->cfq_max_depth, 1, UINT_MAX, 0);
2461 #undef STORE_FUNCTION
2462
2463 static struct cfq_fs_entry cfq_quantum_entry = {
2464         .attr = {.name = "quantum", .mode = S_IRUGO | S_IWUSR },
2465         .show = cfq_quantum_show,
2466         .store = cfq_quantum_store,
2467 };
2468 static struct cfq_fs_entry cfq_queued_entry = {
2469         .attr = {.name = "queued", .mode = S_IRUGO | S_IWUSR },
2470         .show = cfq_queued_show,
2471         .store = cfq_queued_store,
2472 };
2473 static struct cfq_fs_entry cfq_fifo_expire_sync_entry = {
2474         .attr = {.name = "fifo_expire_sync", .mode = S_IRUGO | S_IWUSR },
2475         .show = cfq_fifo_expire_sync_show,
2476         .store = cfq_fifo_expire_sync_store,
2477 };
2478 static struct cfq_fs_entry cfq_fifo_expire_async_entry = {
2479         .attr = {.name = "fifo_expire_async", .mode = S_IRUGO | S_IWUSR },
2480         .show = cfq_fifo_expire_async_show,
2481         .store = cfq_fifo_expire_async_store,
2482 };
2483 static struct cfq_fs_entry cfq_back_max_entry = {
2484         .attr = {.name = "back_seek_max", .mode = S_IRUGO | S_IWUSR },
2485         .show = cfq_back_max_show,
2486         .store = cfq_back_max_store,
2487 };
2488 static struct cfq_fs_entry cfq_back_penalty_entry = {
2489         .attr = {.name = "back_seek_penalty", .mode = S_IRUGO | S_IWUSR },
2490         .show = cfq_back_penalty_show,
2491         .store = cfq_back_penalty_store,
2492 };
2493 static struct cfq_fs_entry cfq_slice_sync_entry = {
2494         .attr = {.name = "slice_sync", .mode = S_IRUGO | S_IWUSR },
2495         .show = cfq_slice_sync_show,
2496         .store = cfq_slice_sync_store,
2497 };
2498 static struct cfq_fs_entry cfq_slice_async_entry = {
2499         .attr = {.name = "slice_async", .mode = S_IRUGO | S_IWUSR },
2500         .show = cfq_slice_async_show,
2501         .store = cfq_slice_async_store,
2502 };
2503 static struct cfq_fs_entry cfq_slice_async_rq_entry = {
2504         .attr = {.name = "slice_async_rq", .mode = S_IRUGO | S_IWUSR },
2505         .show = cfq_slice_async_rq_show,
2506         .store = cfq_slice_async_rq_store,
2507 };
2508 static struct cfq_fs_entry cfq_slice_idle_entry = {
2509         .attr = {.name = "slice_idle", .mode = S_IRUGO | S_IWUSR },
2510         .show = cfq_slice_idle_show,
2511         .store = cfq_slice_idle_store,
2512 };
2513 static struct cfq_fs_entry cfq_max_depth_entry = {
2514         .attr = {.name = "max_depth", .mode = S_IRUGO | S_IWUSR },
2515         .show = cfq_max_depth_show,
2516         .store = cfq_max_depth_store,
2517 };
2518
2519 static struct attribute *default_attrs[] = {
2520         &cfq_quantum_entry.attr,
2521         &cfq_queued_entry.attr,
2522         &cfq_fifo_expire_sync_entry.attr,
2523         &cfq_fifo_expire_async_entry.attr,
2524         &cfq_back_max_entry.attr,
2525         &cfq_back_penalty_entry.attr,
2526         &cfq_slice_sync_entry.attr,
2527         &cfq_slice_async_entry.attr,
2528         &cfq_slice_async_rq_entry.attr,
2529         &cfq_slice_idle_entry.attr,
2530         &cfq_max_depth_entry.attr,
2531         NULL,
2532 };
2533
2534 #define to_cfq(atr) container_of((atr), struct cfq_fs_entry, attr)
2535
2536 static ssize_t
2537 cfq_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
2538 {
2539         elevator_t *e = container_of(kobj, elevator_t, kobj);
2540         struct cfq_fs_entry *entry = to_cfq(attr);
2541
2542         if (!entry->show)
2543                 return -EIO;
2544
2545         return entry->show(e->elevator_data, page);
2546 }
2547
2548 static ssize_t
2549 cfq_attr_store(struct kobject *kobj, struct attribute *attr,
2550                const char *page, size_t length)
2551 {
2552         elevator_t *e = container_of(kobj, elevator_t, kobj);
2553         struct cfq_fs_entry *entry = to_cfq(attr);
2554
2555         if (!entry->store)
2556                 return -EIO;
2557
2558         return entry->store(e->elevator_data, page, length);
2559 }
2560
2561 static struct sysfs_ops cfq_sysfs_ops = {
2562         .show   = cfq_attr_show,
2563         .store  = cfq_attr_store,
2564 };
2565
2566 static struct kobj_type cfq_ktype = {
2567         .sysfs_ops      = &cfq_sysfs_ops,
2568         .default_attrs  = default_attrs,
2569 };
2570
2571 static struct elevator_type iosched_cfq = {
2572         .ops = {
2573                 .elevator_merge_fn =            cfq_merge,
2574                 .elevator_merged_fn =           cfq_merged_request,
2575                 .elevator_merge_req_fn =        cfq_merged_requests,
2576                 .elevator_next_req_fn =         cfq_next_request,
2577                 .elevator_add_req_fn =          cfq_insert_request,
2578                 .elevator_remove_req_fn =       cfq_remove_request,
2579                 .elevator_requeue_req_fn =      cfq_requeue_request,
2580                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2581                 .elevator_queue_empty_fn =      cfq_queue_empty,
2582                 .elevator_completed_req_fn =    cfq_completed_request,
2583                 .elevator_former_req_fn =       cfq_former_request,
2584                 .elevator_latter_req_fn =       cfq_latter_request,
2585                 .elevator_set_req_fn =          cfq_set_request,
2586                 .elevator_put_req_fn =          cfq_put_request,
2587                 .elevator_may_queue_fn =        cfq_may_queue,
2588                 .elevator_init_fn =             cfq_init_queue,
2589                 .elevator_exit_fn =             cfq_exit_queue,
2590         },
2591         .elevator_ktype =       &cfq_ktype,
2592         .elevator_name =        "cfq",
2593         .elevator_owner =       THIS_MODULE,
2594 };
2595
2596 static int __init cfq_init(void)
2597 {
2598         int ret;
2599
2600         /*
2601          * could be 0 on HZ < 1000 setups
2602          */
2603         if (!cfq_slice_async)
2604                 cfq_slice_async = 1;
2605         if (!cfq_slice_idle)
2606                 cfq_slice_idle = 1;
2607
2608         if (cfq_slab_setup())
2609                 return -ENOMEM;
2610
2611         ret = elv_register(&iosched_cfq);
2612         if (ret)
2613                 cfq_slab_kill();
2614
2615         return ret;
2616 }
2617
2618 static void __exit cfq_exit(void)
2619 {
2620         struct task_struct *g, *p;
2621         unsigned long flags;
2622
2623         read_lock_irqsave(&tasklist_lock, flags);
2624
2625         /*
2626          * iterate each process in the system, removing our io_context
2627          */
2628         do_each_thread(g, p) {
2629                 struct io_context *ioc = p->io_context;
2630
2631                 if (ioc && ioc->cic) {
2632                         ioc->cic->exit(ioc->cic);
2633                         cfq_free_io_context(ioc->cic);
2634                         ioc->cic = NULL;
2635                 }
2636         } while_each_thread(g, p);
2637
2638         read_unlock_irqrestore(&tasklist_lock, flags);
2639
2640         cfq_slab_kill();
2641         elv_unregister(&iosched_cfq);
2642 }
2643
2644 module_init(cfq_init);
2645 module_exit(cfq_exit);
2646
2647 MODULE_AUTHOR("Jens Axboe");
2648 MODULE_LICENSE("GPL");
2649 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");