block/cfq-iosched.c

   1 /*
   2  *  CFQ, or complete fairness queueing, disk scheduler.
   3  *
   4  *  Based on ideas from a previously unfinished io
   5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   6  *
   7  *  Copyright (C) 2003 Jens Axboe <axboe@suse.de>
   8  */
   9 #include <linux/kernel.h>
  10 #include <linux/fs.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/elevator.h>
  13 #include <linux/bio.h>
  14 #include <linux/config.h>
  15 #include <linux/module.h>
  16 #include <linux/slab.h>
  17 #include <linux/init.h>
  18 #include <linux/compiler.h>
  19 #include <linux/hash.h>
  20 #include <linux/rbtree.h>
  21 #include <linux/mempool.h>
  22 #include <linux/ioprio.h>
  23 #include <linux/writeback.h>
  24
  25 /*
  26  * tunables
  27  */
  28 static const int cfq_quantum = 4;               /* max queue in one round of service */
  29 static const int cfq_queued = 8;                /* minimum rq allocate limit per-queue*/
  30 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  31 static const int cfq_back_max = 16 * 1024;      /* maximum backwards seek, in KiB */
  32 static const int cfq_back_penalty = 2;          /* penalty of a backwards seek */
  33
  34 static const int cfq_slice_sync = HZ / 10;
  35 static int cfq_slice_async = HZ / 25;
  36 static const int cfq_slice_async_rq = 2;
  37 static int cfq_slice_idle = HZ / 100;
  38
  39 #define CFQ_IDLE_GRACE          (HZ / 10)
  40 #define CFQ_SLICE_SCALE         (5)
  41
  42 #define CFQ_KEY_ASYNC           (0)
  43 #define CFQ_KEY_ANY             (0xffff)
  44
  45 /*
  46  * disable queueing at the driver/hardware level
  47  */
  48 static const int cfq_max_depth = 2;
  49
  50 /*
  51  * for the hash of cfqq inside the cfqd
  52  */
  53 #define CFQ_QHASH_SHIFT         6
  54 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
  55 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
  56
  57 /*
  58  * for the hash of crq inside the cfqq
  59  */
  60 #define CFQ_MHASH_SHIFT         6
  61 #define CFQ_MHASH_BLOCK(sec)    ((sec) >> 3)
  62 #define CFQ_MHASH_ENTRIES       (1 << CFQ_MHASH_SHIFT)
  63 #define CFQ_MHASH_FN(sec)       hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
  64 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
  65 #define list_entry_hash(ptr)    hlist_entry((ptr), struct cfq_rq, hash)
  66
  67 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
  68 #define list_entry_fifo(ptr)    list_entry((ptr), struct request, queuelist)
  69
  70 #define RQ_DATA(rq)             (rq)->elevator_private
  71
  72 /*
  73  * rb-tree defines
  74  */
  75 #define RB_NONE                 (2)
  76 #define RB_EMPTY(node)          ((node)->rb_node == NULL)
  77 #define RB_CLEAR_COLOR(node)    (node)->rb_color = RB_NONE
  78 #define RB_CLEAR(node)          do {    \
  79         (node)->rb_parent = NULL;       \
  80         RB_CLEAR_COLOR((node));         \
  81         (node)->rb_right = NULL;        \
  82         (node)->rb_left = NULL;         \
  83 } while (0)
  84 #define RB_CLEAR_ROOT(root)     ((root)->rb_node = NULL)
  85 #define rb_entry_crq(node)      rb_entry((node), struct cfq_rq, rb_node)
  86 #define rq_rb_key(rq)           (rq)->sector
  87
  88 static kmem_cache_t *crq_pool;
  89 static kmem_cache_t *cfq_pool;
  90 static kmem_cache_t *cfq_ioc_pool;
  91
  92 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  93 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  94 #define cfq_class_be(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
  95 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  96
  97 #define ASYNC                   (0)
  98 #define SYNC                    (1)
  99
 100 #define cfq_cfqq_dispatched(cfqq)       \
 101         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
 102
 103 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
 104
 105 #define cfq_cfqq_sync(cfqq)             \
 106         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
 107
 108 /*
 109  * Per block device queue structure
 110  */
 111 struct cfq_data {
 112         atomic_t ref;
 113         request_queue_t *queue;
 114
 115         /*
 116          * rr list of queues with requests and the count of them
 117          */
 118         struct list_head rr_list[CFQ_PRIO_LISTS];
 119         struct list_head busy_rr;
 120         struct list_head cur_rr;
 121         struct list_head idle_rr;
 122         unsigned int busy_queues;
 123
 124         /*
 125          * non-ordered list of empty cfqq's
 126          */
 127         struct list_head empty_list;
 128
 129         /*
 130          * cfqq lookup hash
 131          */
 132         struct hlist_head *cfq_hash;
 133
 134         /*
 135          * global crq hash for all queues
 136          */
 137         struct hlist_head *crq_hash;
 138
 139         unsigned int max_queued;
 140
 141         mempool_t *crq_pool;
 142
 143         int rq_in_driver;
 144
 145         /*
 146          * schedule slice state info
 147          */
 148         /*
 149          * idle window management
 150          */
 151         struct timer_list idle_slice_timer;
 152         struct work_struct unplug_work;
 153
 154         struct cfq_queue *active_queue;
 155         struct cfq_io_context *active_cic;
 156         int cur_prio, cur_end_prio;
 157         unsigned int dispatch_slice;
 158
 159         struct timer_list idle_class_timer;
 160
 161         sector_t last_sector;
 162         unsigned long last_end_request;
 163
 164         unsigned int rq_starved;
 165
 166         /*
 167          * tunables, see top of file
 168          */
 169         unsigned int cfq_quantum;
 170         unsigned int cfq_queued;
 171         unsigned int cfq_fifo_expire[2];
 172         unsigned int cfq_back_penalty;
 173         unsigned int cfq_back_max;
 174         unsigned int cfq_slice[2];
 175         unsigned int cfq_slice_async_rq;
 176         unsigned int cfq_slice_idle;
 177         unsigned int cfq_max_depth;
 178 };
 179
 180 /*
 181  * Per process-grouping structure
 182  */
 183 struct cfq_queue {
 184         /* reference count */
 185         atomic_t ref;
 186         /* parent cfq_data */
 187         struct cfq_data *cfqd;
 188         /* cfqq lookup hash */
 189         struct hlist_node cfq_hash;
 190         /* hash key */
 191         unsigned int key;
 192         /* on either rr or empty list of cfqd */
 193         struct list_head cfq_list;
 194         /* sorted list of pending requests */
 195         struct rb_root sort_list;
 196         /* if fifo isn't expired, next request to serve */
 197         struct cfq_rq *next_crq;
 198         /* requests queued in sort_list */
 199         int queued[2];
 200         /* currently allocated requests */
 201         int allocated[2];
 202         /* fifo list of requests in sort_list */
 203         struct list_head fifo;
 204
 205         unsigned long slice_start;
 206         unsigned long slice_end;
 207         unsigned long slice_left;
 208         unsigned long service_last;
 209
 210         /* number of requests that are on the dispatch list */
 211         int on_dispatch[2];
 212
 213         /* io prio of this group */
 214         unsigned short ioprio, org_ioprio;
 215         unsigned short ioprio_class, org_ioprio_class;
 216
 217         /* various state flags, see below */
 218         unsigned int flags;
 219 };
 220
 221 struct cfq_rq {
 222         struct rb_node rb_node;
 223         sector_t rb_key;
 224         struct request *request;
 225         struct hlist_node hash;
 226
 227         struct cfq_queue *cfq_queue;
 228         struct cfq_io_context *io_context;
 229
 230         unsigned int crq_flags;
 231 };
 232
 233 enum cfqq_state_flags {
 234         CFQ_CFQQ_FLAG_on_rr = 0,
 235         CFQ_CFQQ_FLAG_wait_request,
 236         CFQ_CFQQ_FLAG_must_alloc,
 237         CFQ_CFQQ_FLAG_must_alloc_slice,
 238         CFQ_CFQQ_FLAG_must_dispatch,
 239         CFQ_CFQQ_FLAG_fifo_expire,
 240         CFQ_CFQQ_FLAG_idle_window,
 241         CFQ_CFQQ_FLAG_prio_changed,
 242         CFQ_CFQQ_FLAG_expired,
 243 };
 244
 245 #define CFQ_CFQQ_FNS(name)                                              \
 246 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 247 {                                                                       \
 248         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 249 }                                                                       \
 250 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 251 {                                                                       \
 252         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 253 }                                                                       \
 254 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 255 {                                                                       \
 256         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 257 }
 258
 259 CFQ_CFQQ_FNS(on_rr);
 260 CFQ_CFQQ_FNS(wait_request);
 261 CFQ_CFQQ_FNS(must_alloc);
 262 CFQ_CFQQ_FNS(must_alloc_slice);
 263 CFQ_CFQQ_FNS(must_dispatch);
 264 CFQ_CFQQ_FNS(fifo_expire);
 265 CFQ_CFQQ_FNS(idle_window);
 266 CFQ_CFQQ_FNS(prio_changed);
 267 CFQ_CFQQ_FNS(expired);
 268 #undef CFQ_CFQQ_FNS
 269
 270 enum cfq_rq_state_flags {
 271         CFQ_CRQ_FLAG_is_sync = 0,
 272 };
 273
 274 #define CFQ_CRQ_FNS(name)                                               \
 275 static inline void cfq_mark_crq_##name(struct cfq_rq *crq)              \
 276 {                                                                       \
 277         crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name);                   \
 278 }                                                                       \
 279 static inline void cfq_clear_crq_##name(struct cfq_rq *crq)             \
 280 {                                                                       \
 281         crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name);                  \
 282 }                                                                       \
 283 static inline int cfq_crq_##name(const struct cfq_rq *crq)              \
 284 {                                                                       \
 285         return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0;      \
 286 }
 287
 288 CFQ_CRQ_FNS(is_sync);
 289 #undef CFQ_CRQ_FNS
 290
 291 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 292 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
 293 static void cfq_put_cfqd(struct cfq_data *cfqd);
 294
 295 #define process_sync(tsk)       ((tsk)->flags & PF_SYNCWRITE)
 296
 297 /*
 298  * lots of deadline iosched dupes, can be abstracted later...
 299  */
 300 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
 301 {
 302         hlist_del_init(&crq->hash);
 303 }
 304
 305 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
 306 {
 307         const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
 308
 309         hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
 310 }
 311
 312 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
 313 {
 314         struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
 315         struct hlist_node *entry, *next;
 316
 317         hlist_for_each_safe(entry, next, hash_list) {
 318                 struct cfq_rq *crq = list_entry_hash(entry);
 319                 struct request *__rq = crq->request;
 320
 321                 if (!rq_mergeable(__rq)) {
 322                         cfq_del_crq_hash(crq);
 323                         continue;
 324                 }
 325
 326                 if (rq_hash_key(__rq) == offset)
 327                         return __rq;
 328         }
 329
 330         return NULL;
 331 }
 332
 333 /*
 334  * scheduler run of queue, if there are requests pending and no one in the
 335  * driver that will restart queueing
 336  */
 337 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 338 {
 339         if (!cfqd->rq_in_driver && cfqd->busy_queues)
 340                 kblockd_schedule_work(&cfqd->unplug_work);
 341 }
 342
 343 static int cfq_queue_empty(request_queue_t *q)
 344 {
 345         struct cfq_data *cfqd = q->elevator->elevator_data;
 346
 347         return !cfqd->busy_queues;
 348 }
 349
 350 /*
 351  * Lifted from AS - choose which of crq1 and crq2 that is best served now.
 352  * We choose the request that is closest to the head right now. Distance
 353  * behind the head are penalized and only allowed to a certain extent.
 354  */
 355 static struct cfq_rq *
 356 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
 357 {
 358         sector_t last, s1, s2, d1 = 0, d2 = 0;
 359         int r1_wrap = 0, r2_wrap = 0;   /* requests are behind the disk head */
 360         unsigned long back_max;
 361
 362         if (crq1 == NULL || crq1 == crq2)
 363                 return crq2;
 364         if (crq2 == NULL)
 365                 return crq1;
 366
 367         if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
 368                 return crq1;
 369         else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
 370                 return crq2;
 371
 372         s1 = crq1->request->sector;
 373         s2 = crq2->request->sector;
 374
 375         last = cfqd->last_sector;
 376
 377         /*
 378          * by definition, 1KiB is 2 sectors
 379          */
 380         back_max = cfqd->cfq_back_max * 2;
 381
 382         /*
 383          * Strict one way elevator _except_ in the case where we allow
 384          * short backward seeks which are biased as twice the cost of a
 385          * similar forward seek.
 386          */
 387         if (s1 >= last)
 388                 d1 = s1 - last;
 389         else if (s1 + back_max >= last)
 390                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 391         else
 392                 r1_wrap = 1;
 393
 394         if (s2 >= last)
 395                 d2 = s2 - last;
 396         else if (s2 + back_max >= last)
 397                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 398         else
 399                 r2_wrap = 1;
 400
 401         /* Found required data */
 402         if (!r1_wrap && r2_wrap)
 403                 return crq1;
 404         else if (!r2_wrap && r1_wrap)
 405                 return crq2;
 406         else if (r1_wrap && r2_wrap) {
 407                 /* both behind the head */
 408                 if (s1 <= s2)
 409                         return crq1;
 410                 else
 411                         return crq2;
 412         }
 413
 414         /* Both requests in front of the head */
 415         if (d1 < d2)
 416                 return crq1;
 417         else if (d2 < d1)
 418                 return crq2;
 419         else {
 420                 if (s1 >= s2)
 421                         return crq1;
 422                 else
 423                         return crq2;
 424         }
 425 }
 426
 427 /*
 428  * would be nice to take fifo expire time into account as well
 429  */
 430 static struct cfq_rq *
 431 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 432                   struct cfq_rq *last)
 433 {
 434         struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
 435         struct rb_node *rbnext, *rbprev;
 436
 437         if (!(rbnext = rb_next(&last->rb_node))) {
 438                 rbnext = rb_first(&cfqq->sort_list);
 439                 if (rbnext == &last->rb_node)
 440                         rbnext = NULL;
 441         }
 442
 443         rbprev = rb_prev(&last->rb_node);
 444
 445         if (rbprev)
 446                 crq_prev = rb_entry_crq(rbprev);
 447         if (rbnext)
 448                 crq_next = rb_entry_crq(rbnext);
 449
 450         return cfq_choose_req(cfqd, crq_next, crq_prev);
 451 }
 452
 453 static void cfq_update_next_crq(struct cfq_rq *crq)
 454 {
 455         struct cfq_queue *cfqq = crq->cfq_queue;
 456
 457         if (cfqq->next_crq == crq)
 458                 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
 459 }
 460
 461 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 462 {
 463         struct cfq_data *cfqd = cfqq->cfqd;
 464         struct list_head *list, *entry;
 465
 466         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 467
 468         list_del(&cfqq->cfq_list);
 469
 470         if (cfq_class_rt(cfqq))
 471                 list = &cfqd->cur_rr;
 472         else if (cfq_class_idle(cfqq))
 473                 list = &cfqd->idle_rr;
 474         else {
 475                 /*
 476                  * if cfqq has requests in flight, don't allow it to be
 477                  * found in cfq_set_active_queue before it has finished them.
 478                  * this is done to increase fairness between a process that
 479                  * has lots of io pending vs one that only generates one
 480                  * sporadically or synchronously
 481                  */
 482                 if (cfq_cfqq_dispatched(cfqq))
 483                         list = &cfqd->busy_rr;
 484                 else
 485                         list = &cfqd->rr_list[cfqq->ioprio];
 486         }
 487
 488         /*
 489          * if queue was preempted, just add to front to be fair. busy_rr
 490          * isn't sorted.
 491          */
 492         if (preempted || list == &cfqd->busy_rr) {
 493                 list_add(&cfqq->cfq_list, list);
 494                 return;
 495         }
 496
 497         /*
 498          * sort by when queue was last serviced
 499          */
 500         entry = list;
 501         while ((entry = entry->prev) != list) {
 502                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
 503
 504                 if (!__cfqq->service_last)
 505                         break;
 506                 if (time_before(__cfqq->service_last, cfqq->service_last))
 507                         break;
 508         }
 509
 510         list_add(&cfqq->cfq_list, entry);
 511 }
 512
 513 /*
 514  * add to busy list of queues for service, trying to be fair in ordering
 515  * the pending list according to last request service
 516  */
 517 static inline void
 518 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 519 {
 520         BUG_ON(cfq_cfqq_on_rr(cfqq));
 521         cfq_mark_cfqq_on_rr(cfqq);
 522         cfqd->busy_queues++;
 523
 524         cfq_resort_rr_list(cfqq, 0);
 525 }
 526
 527 static inline void
 528 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 529 {
 530         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 531         cfq_clear_cfqq_on_rr(cfqq);
 532         list_move(&cfqq->cfq_list, &cfqd->empty_list);
 533
 534         BUG_ON(!cfqd->busy_queues);
 535         cfqd->busy_queues--;
 536 }
 537
 538 /*
 539  * rb tree support functions
 540  */
 541 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
 542 {
 543         struct cfq_queue *cfqq = crq->cfq_queue;
 544         struct cfq_data *cfqd = cfqq->cfqd;
 545         const int sync = cfq_crq_is_sync(crq);
 546
 547         BUG_ON(!cfqq->queued[sync]);
 548         cfqq->queued[sync]--;
 549
 550         cfq_update_next_crq(crq);
 551
 552         rb_erase(&crq->rb_node, &cfqq->sort_list);
 553         RB_CLEAR_COLOR(&crq->rb_node);
 554
 555         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
 556                 cfq_del_cfqq_rr(cfqd, cfqq);
 557 }
 558
 559 static struct cfq_rq *
 560 __cfq_add_crq_rb(struct cfq_rq *crq)
 561 {
 562         struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
 563         struct rb_node *parent = NULL;
 564         struct cfq_rq *__crq;
 565
 566         while (*p) {
 567                 parent = *p;
 568                 __crq = rb_entry_crq(parent);
 569
 570                 if (crq->rb_key < __crq->rb_key)
 571                         p = &(*p)->rb_left;
 572                 else if (crq->rb_key > __crq->rb_key)
 573                         p = &(*p)->rb_right;
 574                 else
 575                         return __crq;
 576         }
 577
 578         rb_link_node(&crq->rb_node, parent, p);
 579         return NULL;
 580 }
 581
 582 static void cfq_add_crq_rb(struct cfq_rq *crq)
 583 {
 584         struct cfq_queue *cfqq = crq->cfq_queue;
 585         struct cfq_data *cfqd = cfqq->cfqd;
 586         struct request *rq = crq->request;
 587         struct cfq_rq *__alias;
 588
 589         crq->rb_key = rq_rb_key(rq);
 590         cfqq->queued[cfq_crq_is_sync(crq)]++;
 591
 592         /*
 593          * looks a little odd, but the first insert might return an alias.
 594          * if that happens, put the alias on the dispatch list
 595          */
 596         while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
 597                 cfq_dispatch_insert(cfqd->queue, __alias);
 598
 599         rb_insert_color(&crq->rb_node, &cfqq->sort_list);
 600
 601         if (!cfq_cfqq_on_rr(cfqq))
 602                 cfq_add_cfqq_rr(cfqd, cfqq);
 603
 604         /*
 605          * check if this request is a better next-serve candidate
 606          */
 607         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
 608 }
 609
 610 static inline void
 611 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
 612 {
 613         rb_erase(&crq->rb_node, &cfqq->sort_list);
 614         cfqq->queued[cfq_crq_is_sync(crq)]--;
 615
 616         cfq_add_crq_rb(crq);
 617 }
 618
 619 static struct request *cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector)
 620
 621 {
 622         struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, current->pid, CFQ_KEY_ANY);
 623         struct rb_node *n;
 624
 625         if (!cfqq)
 626                 goto out;
 627
 628         n = cfqq->sort_list.rb_node;
 629         while (n) {
 630                 struct cfq_rq *crq = rb_entry_crq(n);
 631
 632                 if (sector < crq->rb_key)
 633                         n = n->rb_left;
 634                 else if (sector > crq->rb_key)
 635                         n = n->rb_right;
 636                 else
 637                         return crq->request;
 638         }
 639
 640 out:
 641         return NULL;
 642 }
 643
 644 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 645 {
 646         struct cfq_data *cfqd = q->elevator->elevator_data;
 647
 648         cfqd->rq_in_driver++;
 649 }
 650
 651 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 652 {
 653         struct cfq_data *cfqd = q->elevator->elevator_data;
 654
 655         WARN_ON(!cfqd->rq_in_driver);
 656         cfqd->rq_in_driver--;
 657 }
 658
 659 static void cfq_remove_request(struct request *rq)
 660 {
 661         struct cfq_rq *crq = RQ_DATA(rq);
 662
 663         list_del_init(&rq->queuelist);
 664         cfq_del_crq_rb(crq);
 665         cfq_del_crq_hash(crq);
 666 }
 667
 668 static int
 669 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 670 {
 671         struct cfq_data *cfqd = q->elevator->elevator_data;
 672         struct request *__rq;
 673         int ret;
 674
 675         __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
 676         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 677                 ret = ELEVATOR_BACK_MERGE;
 678                 goto out;
 679         }
 680
 681         __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio));
 682         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 683                 ret = ELEVATOR_FRONT_MERGE;
 684                 goto out;
 685         }
 686
 687         return ELEVATOR_NO_MERGE;
 688 out:
 689         *req = __rq;
 690         return ret;
 691 }
 692
 693 static void cfq_merged_request(request_queue_t *q, struct request *req)
 694 {
 695         struct cfq_data *cfqd = q->elevator->elevator_data;
 696         struct cfq_rq *crq = RQ_DATA(req);
 697
 698         cfq_del_crq_hash(crq);
 699         cfq_add_crq_hash(cfqd, crq);
 700
 701         if (rq_rb_key(req) != crq->rb_key) {
 702                 struct cfq_queue *cfqq = crq->cfq_queue;
 703
 704                 cfq_update_next_crq(crq);
 705                 cfq_reposition_crq_rb(cfqq, crq);
 706         }
 707 }
 708
 709 static void
 710 cfq_merged_requests(request_queue_t *q, struct request *rq,
 711                     struct request *next)
 712 {
 713         cfq_merged_request(q, rq);
 714
 715         /*
 716          * reposition in fifo if next is older than rq
 717          */
 718         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 719             time_before(next->start_time, rq->start_time))
 720                 list_move(&rq->queuelist, &next->queuelist);
 721
 722         cfq_remove_request(next);
 723 }
 724
 725 static inline void
 726 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 727 {
 728         if (cfqq) {
 729                 /*
 730                  * stop potential idle class queues waiting service
 731                  */
 732                 del_timer(&cfqd->idle_class_timer);
 733
 734                 cfqq->slice_start = jiffies;
 735                 cfqq->slice_end = 0;
 736                 cfqq->slice_left = 0;
 737                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 738                 cfq_clear_cfqq_fifo_expire(cfqq);
 739                 cfq_clear_cfqq_expired(cfqq);
 740         }
 741
 742         cfqd->active_queue = cfqq;
 743 }
 744
 745 /*
 746  * 0
 747  * 0,1
 748  * 0,1,2
 749  * 0,1,2,3
 750  * 0,1,2,3,4
 751  * 0,1,2,3,4,5
 752  * 0,1,2,3,4,5,6
 753  * 0,1,2,3,4,5,6,7
 754  */
 755 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 756 {
 757         int prio, wrap;
 758
 759         prio = -1;
 760         wrap = 0;
 761         do {
 762                 int p;
 763
 764                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 765                         if (!list_empty(&cfqd->rr_list[p])) {
 766                                 prio = p;
 767                                 break;
 768                         }
 769                 }
 770
 771                 if (prio != -1)
 772                         break;
 773                 cfqd->cur_prio = 0;
 774                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 775                         cfqd->cur_end_prio = 0;
 776                         if (wrap)
 777                                 break;
 778                         wrap = 1;
 779                 }
 780         } while (1);
 781
 782         if (unlikely(prio == -1))
 783                 return -1;
 784
 785         BUG_ON(prio >= CFQ_PRIO_LISTS);
 786
 787         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 788
 789         cfqd->cur_prio = prio + 1;
 790         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 791                 cfqd->cur_end_prio = cfqd->cur_prio;
 792                 cfqd->cur_prio = 0;
 793         }
 794         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 795                 cfqd->cur_prio = 0;
 796                 cfqd->cur_end_prio = 0;
 797         }
 798
 799         return prio;
 800 }
 801
 802 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 803 {
 804         struct cfq_queue *cfqq;
 805
 806         /*
 807          * if current queue is expired but not done with its requests yet,
 808          * wait for that to happen
 809          */
 810         if ((cfqq = cfqd->active_queue) != NULL) {
 811                 if (cfq_cfqq_expired(cfqq) && cfq_cfqq_dispatched(cfqq))
 812                         return NULL;
 813         }
 814
 815         /*
 816          * if current list is non-empty, grab first entry. if it is empty,
 817          * get next prio level and grab first entry then if any are spliced
 818          */
 819         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
 820                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 821
 822         /*
 823          * if we have idle queues and no rt or be queues had pending
 824          * requests, either allow immediate service if the grace period
 825          * has passed or arm the idle grace timer
 826          */
 827         if (!cfqq && !list_empty(&cfqd->idle_rr)) {
 828                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 829
 830                 if (time_after_eq(jiffies, end))
 831                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 832                 else
 833                         mod_timer(&cfqd->idle_class_timer, end);
 834         }
 835
 836         __cfq_set_active_queue(cfqd, cfqq);
 837         return cfqq;
 838 }
 839
 840 /*
 841  * current cfqq expired its slice (or was too idle), select new one
 842  */
 843 static void
 844 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 845                     int preempted)
 846 {
 847         unsigned long now = jiffies;
 848
 849         if (cfq_cfqq_wait_request(cfqq))
 850                 del_timer(&cfqd->idle_slice_timer);
 851
 852         if (!preempted && !cfq_cfqq_dispatched(cfqq))
 853                 cfqq->service_last = now;
 854
 855         cfq_clear_cfqq_must_dispatch(cfqq);
 856         cfq_clear_cfqq_wait_request(cfqq);
 857
 858         /*
 859          * store what was left of this slice, if the queue idled out
 860          * or was preempted
 861          */
 862         if (time_after(cfqq->slice_end, now))
 863                 cfqq->slice_left = cfqq->slice_end - now;
 864         else
 865                 cfqq->slice_left = 0;
 866
 867         if (cfq_cfqq_on_rr(cfqq))
 868                 cfq_resort_rr_list(cfqq, preempted);
 869
 870         if (cfqq == cfqd->active_queue)
 871                 cfqd->active_queue = NULL;
 872
 873         if (cfqd->active_cic) {
 874                 put_io_context(cfqd->active_cic->ioc);
 875                 cfqd->active_cic = NULL;
 876         }
 877
 878         cfqd->dispatch_slice = 0;
 879 }
 880
 881 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 882 {
 883         struct cfq_queue *cfqq = cfqd->active_queue;
 884
 885         if (cfqq) {
 886                 /*
 887                  * use deferred expiry, if there are requests in progress as
 888                  * not to disturb the slice of the next queue
 889                  */
 890                 if (cfq_cfqq_dispatched(cfqq))
 891                         cfq_mark_cfqq_expired(cfqq);
 892                 else
 893                         __cfq_slice_expired(cfqd, cfqq, preempted);
 894         }
 895 }
 896
 897 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 898
 899 {
 900         WARN_ON(!RB_EMPTY(&cfqq->sort_list));
 901         WARN_ON(cfqq != cfqd->active_queue);
 902
 903         /*
 904          * idle is disabled, either manually or by past process history
 905          */
 906         if (!cfqd->cfq_slice_idle)
 907                 return 0;
 908         if (!cfq_cfqq_idle_window(cfqq))
 909                 return 0;
 910         /*
 911          * task has exited, don't wait
 912          */
 913         if (cfqd->active_cic && !cfqd->active_cic->ioc->task)
 914                 return 0;
 915
 916         cfq_mark_cfqq_must_dispatch(cfqq);
 917         cfq_mark_cfqq_wait_request(cfqq);
 918
 919         if (!timer_pending(&cfqd->idle_slice_timer)) {
 920                 unsigned long slice_left = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 921
 922                 cfqd->idle_slice_timer.expires = jiffies + slice_left;
 923                 add_timer(&cfqd->idle_slice_timer);
 924         }
 925
 926         return 1;
 927 }
 928
 929 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
 930 {
 931         struct cfq_data *cfqd = q->elevator->elevator_data;
 932         struct cfq_queue *cfqq = crq->cfq_queue;
 933
 934         cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
 935         cfq_remove_request(crq->request);
 936         cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
 937         elv_dispatch_sort(q, crq->request);
 938 }
 939
 940 /*
 941  * return expired entry, or NULL to just start from scratch in rbtree
 942  */
 943 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
 944 {
 945         struct cfq_data *cfqd = cfqq->cfqd;
 946         struct request *rq;
 947         struct cfq_rq *crq;
 948
 949         if (cfq_cfqq_fifo_expire(cfqq))
 950                 return NULL;
 951
 952         if (!list_empty(&cfqq->fifo)) {
 953                 int fifo = cfq_cfqq_class_sync(cfqq);
 954
 955                 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
 956                 rq = crq->request;
 957                 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
 958                         cfq_mark_cfqq_fifo_expire(cfqq);
 959                         return crq;
 960                 }
 961         }
 962
 963         return NULL;
 964 }
 965
 966 /*
 967  * Scale schedule slice based on io priority. Use the sync time slice only
 968  * if a queue is marked sync and has sync io queued. A sync queue with async
 969  * io only, should not get full sync slice length.
 970  */
 971 static inline int
 972 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 973 {
 974         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 975
 976         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 977
 978         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
 979 }
 980
 981 static inline void
 982 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 983 {
 984         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 985 }
 986
 987 static inline int
 988 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 989 {
 990         const int base_rq = cfqd->cfq_slice_async_rq;
 991
 992         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 993
 994         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
 995 }
 996
 997 /*
 998  * get next queue for service
 999  */
1000 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1001 {
1002         unsigned long now = jiffies;
1003         struct cfq_queue *cfqq;
1004
1005         cfqq = cfqd->active_queue;
1006         if (!cfqq)
1007                 goto new_queue;
1008
1009         if (cfq_cfqq_expired(cfqq))
1010                 goto new_queue;
1011
1012         /*
1013          * slice has expired
1014          */
1015         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
1016                 goto expire;
1017
1018         /*
1019          * if queue has requests, dispatch one. if not, check if
1020          * enough slice is left to wait for one
1021          */
1022         if (!RB_EMPTY(&cfqq->sort_list))
1023                 goto keep_queue;
1024         else if (cfq_cfqq_class_sync(cfqq) &&
1025                  time_before(now, cfqq->slice_end)) {
1026                 if (cfq_arm_slice_timer(cfqd, cfqq))
1027                         return NULL;
1028         }
1029
1030 expire:
1031         cfq_slice_expired(cfqd, 0);
1032 new_queue:
1033         cfqq = cfq_set_active_queue(cfqd);
1034 keep_queue:
1035         return cfqq;
1036 }
1037
1038 static int
1039 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1040                         int max_dispatch)
1041 {
1042         int dispatched = 0;
1043
1044         BUG_ON(RB_EMPTY(&cfqq->sort_list));
1045
1046         do {
1047                 struct cfq_rq *crq;
1048
1049                 /*
1050                  * follow expired path, else get first next available
1051                  */
1052                 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1053                         crq = cfqq->next_crq;
1054
1055                 /*
1056                  * finally, insert request into driver dispatch list
1057                  */
1058                 cfq_dispatch_insert(cfqd->queue, crq);
1059
1060                 cfqd->dispatch_slice++;
1061                 dispatched++;
1062
1063                 if (!cfqd->active_cic) {
1064                         atomic_inc(&crq->io_context->ioc->refcount);
1065                         cfqd->active_cic = crq->io_context;
1066                 }
1067
1068                 if (RB_EMPTY(&cfqq->sort_list))
1069                         break;
1070
1071         } while (dispatched < max_dispatch);
1072
1073         /*
1074          * if slice end isn't set yet, set it. if at least one request was
1075          * sync, use the sync time slice value
1076          */
1077         if (!cfqq->slice_end)
1078                 cfq_set_prio_slice(cfqd, cfqq);
1079
1080         /*
1081          * expire an async queue immediately if it has used up its slice. idle
1082          * queue always expire after 1 dispatch round.
1083          */
1084         if ((!cfq_cfqq_sync(cfqq) &&
1085             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1086             cfq_class_idle(cfqq))
1087                 cfq_slice_expired(cfqd, 0);
1088
1089         return dispatched;
1090 }
1091
1092 static int
1093 cfq_forced_dispatch_cfqqs(struct list_head *list)
1094 {
1095         int dispatched = 0;
1096         struct cfq_queue *cfqq, *next;
1097         struct cfq_rq *crq;
1098
1099         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1100                 while ((crq = cfqq->next_crq)) {
1101                         cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1102                         dispatched++;
1103                 }
1104                 BUG_ON(!list_empty(&cfqq->fifo));
1105         }
1106         return dispatched;
1107 }
1108
1109 static int
1110 cfq_forced_dispatch(struct cfq_data *cfqd)
1111 {
1112         int i, dispatched = 0;
1113
1114         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1115                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1116
1117         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1118         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1119         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1120
1121         cfq_slice_expired(cfqd, 0);
1122
1123         BUG_ON(cfqd->busy_queues);
1124
1125         return dispatched;
1126 }
1127
1128 static int
1129 cfq_dispatch_requests(request_queue_t *q, int force)
1130 {
1131         struct cfq_data *cfqd = q->elevator->elevator_data;
1132         struct cfq_queue *cfqq;
1133
1134         if (!cfqd->busy_queues)
1135                 return 0;
1136
1137         if (unlikely(force))
1138                 return cfq_forced_dispatch(cfqd);
1139
1140         cfqq = cfq_select_queue(cfqd);
1141         if (cfqq) {
1142                 int max_dispatch;
1143
1144                 /*
1145                  * if idle window is disabled, allow queue buildup
1146                  */
1147                 if (!cfq_cfqq_idle_window(cfqq) &&
1148                     cfqd->rq_in_driver >= cfqd->cfq_max_depth)
1149                         return 0;
1150
1151                 cfq_clear_cfqq_must_dispatch(cfqq);
1152                 cfq_clear_cfqq_wait_request(cfqq);
1153                 del_timer(&cfqd->idle_slice_timer);
1154
1155                 max_dispatch = cfqd->cfq_quantum;
1156                 if (cfq_class_idle(cfqq))
1157                         max_dispatch = 1;
1158
1159                 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1160         }
1161
1162         return 0;
1163 }
1164
1165 /*
1166  * task holds one reference to the queue, dropped when task exits. each crq
1167  * in-flight on this queue also holds a reference, dropped when crq is freed.
1168  *
1169  * queue lock must be held here.
1170  */
1171 static void cfq_put_queue(struct cfq_queue *cfqq)
1172 {
1173         struct cfq_data *cfqd = cfqq->cfqd;
1174
1175         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1176
1177         if (!atomic_dec_and_test(&cfqq->ref))
1178                 return;
1179
1180         BUG_ON(rb_first(&cfqq->sort_list));
1181         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1182         BUG_ON(cfq_cfqq_on_rr(cfqq));
1183
1184         if (unlikely(cfqd->active_queue == cfqq)) {
1185                 __cfq_slice_expired(cfqd, cfqq, 0);
1186                 cfq_schedule_dispatch(cfqd);
1187         }
1188
1189         cfq_put_cfqd(cfqq->cfqd);
1190
1191         /*
1192          * it's on the empty list and still hashed
1193          */
1194         list_del(&cfqq->cfq_list);
1195         hlist_del(&cfqq->cfq_hash);
1196         kmem_cache_free(cfq_pool, cfqq);
1197 }
1198
1199 static inline struct cfq_queue *
1200 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1201                     const int hashval)
1202 {
1203         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1204         struct hlist_node *entry, *next;
1205
1206         hlist_for_each_safe(entry, next, hash_list) {
1207                 struct cfq_queue *__cfqq = list_entry_qhash(entry);
1208                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->ioprio_class, __cfqq->ioprio);
1209
1210                 if (__cfqq->key == key && (__p == prio || prio == CFQ_KEY_ANY))
1211                         return __cfqq;
1212         }
1213
1214         return NULL;
1215 }
1216
1217 static struct cfq_queue *
1218 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1219 {
1220         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1221 }
1222
1223 static void cfq_free_io_context(struct cfq_io_context *cic)
1224 {
1225         struct cfq_io_context *__cic;
1226         struct list_head *entry, *next;
1227
1228         list_for_each_safe(entry, next, &cic->list) {
1229                 __cic = list_entry(entry, struct cfq_io_context, list);
1230                 kmem_cache_free(cfq_ioc_pool, __cic);
1231         }
1232
1233         kmem_cache_free(cfq_ioc_pool, cic);
1234 }
1235
1236 /*
1237  * Called with interrupts disabled
1238  */
1239 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1240 {
1241         struct cfq_data *cfqd = cic->cfqq->cfqd;
1242         request_queue_t *q = cfqd->queue;
1243
1244         WARN_ON(!irqs_disabled());
1245
1246         spin_lock(q->queue_lock);
1247
1248         if (unlikely(cic->cfqq == cfqd->active_queue)) {
1249                 __cfq_slice_expired(cfqd, cic->cfqq, 0);
1250                 cfq_schedule_dispatch(cfqd);
1251         }
1252
1253         cfq_put_queue(cic->cfqq);
1254         cic->cfqq = NULL;
1255         spin_unlock(q->queue_lock);
1256 }
1257
1258 /*
1259  * Another task may update the task cic list, if it is doing a queue lookup
1260  * on its behalf. cfq_cic_lock excludes such concurrent updates
1261  */
1262 static void cfq_exit_io_context(struct cfq_io_context *cic)
1263 {
1264         struct cfq_io_context *__cic;
1265         struct list_head *entry;
1266         unsigned long flags;
1267
1268         local_irq_save(flags);
1269
1270         /*
1271          * put the reference this task is holding to the various queues
1272          */
1273         list_for_each(entry, &cic->list) {
1274                 __cic = list_entry(entry, struct cfq_io_context, list);
1275                 cfq_exit_single_io_context(__cic);
1276         }
1277
1278         cfq_exit_single_io_context(cic);
1279         local_irq_restore(flags);
1280 }
1281
1282 static struct cfq_io_context *
1283 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1284 {
1285         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1286
1287         if (cic) {
1288                 INIT_LIST_HEAD(&cic->list);
1289                 cic->cfqq = NULL;
1290                 cic->key = NULL;
1291                 cic->last_end_request = jiffies;
1292                 cic->ttime_total = 0;
1293                 cic->ttime_samples = 0;
1294                 cic->ttime_mean = 0;
1295                 cic->dtor = cfq_free_io_context;
1296                 cic->exit = cfq_exit_io_context;
1297         }
1298
1299         return cic;
1300 }
1301
1302 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1303 {
1304         struct task_struct *tsk = current;
1305         int ioprio_class;
1306
1307         if (!cfq_cfqq_prio_changed(cfqq))
1308                 return;
1309
1310         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1311         switch (ioprio_class) {
1312                 default:
1313                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1314                 case IOPRIO_CLASS_NONE:
1315                         /*
1316                          * no prio set, place us in the middle of the BE classes
1317                          */
1318                         cfqq->ioprio = task_nice_ioprio(tsk);
1319                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1320                         break;
1321                 case IOPRIO_CLASS_RT:
1322                         cfqq->ioprio = task_ioprio(tsk);
1323                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1324                         break;
1325                 case IOPRIO_CLASS_BE:
1326                         cfqq->ioprio = task_ioprio(tsk);
1327                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1328                         break;
1329                 case IOPRIO_CLASS_IDLE:
1330                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1331                         cfqq->ioprio = 7;
1332                         cfq_clear_cfqq_idle_window(cfqq);
1333                         break;
1334         }
1335
1336         /*
1337          * keep track of original prio settings in case we have to temporarily
1338          * elevate the priority of this queue
1339          */
1340         cfqq->org_ioprio = cfqq->ioprio;
1341         cfqq->org_ioprio_class = cfqq->ioprio_class;
1342
1343         if (cfq_cfqq_on_rr(cfqq))
1344                 cfq_resort_rr_list(cfqq, 0);
1345
1346         cfq_clear_cfqq_prio_changed(cfqq);
1347 }
1348
1349 static inline void changed_ioprio(struct cfq_queue *cfqq)
1350 {
1351         if (cfqq) {
1352                 struct cfq_data *cfqd = cfqq->cfqd;
1353
1354                 spin_lock(cfqd->queue->queue_lock);
1355                 cfq_mark_cfqq_prio_changed(cfqq);
1356                 cfq_init_prio_data(cfqq);
1357                 spin_unlock(cfqd->queue->queue_lock);
1358         }
1359 }
1360
1361 /*
1362  * callback from sys_ioprio_set, irqs are disabled
1363  */
1364 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1365 {
1366         struct cfq_io_context *cic = ioc->cic;
1367
1368         changed_ioprio(cic->cfqq);
1369
1370         list_for_each_entry(cic, &cic->list, list)
1371                 changed_ioprio(cic->cfqq);
1372
1373         return 0;
1374 }
1375
1376 static struct cfq_queue *
1377 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, unsigned short ioprio,
1378               gfp_t gfp_mask)
1379 {
1380         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1381         struct cfq_queue *cfqq, *new_cfqq = NULL;
1382
1383 retry:
1384         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1385
1386         if (!cfqq) {
1387                 if (new_cfqq) {
1388                         cfqq = new_cfqq;
1389                         new_cfqq = NULL;
1390                 } else if (gfp_mask & __GFP_WAIT) {
1391                         spin_unlock_irq(cfqd->queue->queue_lock);
1392                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1393                         spin_lock_irq(cfqd->queue->queue_lock);
1394                         goto retry;
1395                 } else {
1396                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1397                         if (!cfqq)
1398                                 goto out;
1399                 }
1400
1401                 memset(cfqq, 0, sizeof(*cfqq));
1402
1403                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1404                 INIT_LIST_HEAD(&cfqq->cfq_list);
1405                 RB_CLEAR_ROOT(&cfqq->sort_list);
1406                 INIT_LIST_HEAD(&cfqq->fifo);
1407
1408                 cfqq->key = key;
1409                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1410                 atomic_set(&cfqq->ref, 0);
1411                 cfqq->cfqd = cfqd;
1412                 atomic_inc(&cfqd->ref);
1413                 cfqq->service_last = 0;
1414                 /*
1415                  * set ->slice_left to allow preemption for a new process
1416                  */
1417                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1418                 cfq_mark_cfqq_idle_window(cfqq);
1419                 cfq_mark_cfqq_prio_changed(cfqq);
1420                 cfq_init_prio_data(cfqq);
1421         }
1422
1423         if (new_cfqq)
1424                 kmem_cache_free(cfq_pool, new_cfqq);
1425
1426         atomic_inc(&cfqq->ref);
1427 out:
1428         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1429         return cfqq;
1430 }
1431
1432 /*
1433  * Setup general io context and cfq io context. There can be several cfq
1434  * io contexts per general io context, if this process is doing io to more
1435  * than one device managed by cfq. Note that caller is holding a reference to
1436  * cfqq, so we don't need to worry about it disappearing
1437  */
1438 static struct cfq_io_context *
1439 cfq_get_io_context(struct cfq_data *cfqd, pid_t pid, gfp_t gfp_mask)
1440 {
1441         struct io_context *ioc = NULL;
1442         struct cfq_io_context *cic;
1443
1444         might_sleep_if(gfp_mask & __GFP_WAIT);
1445
1446         ioc = get_io_context(gfp_mask);
1447         if (!ioc)
1448                 return NULL;
1449
1450         if ((cic = ioc->cic) == NULL) {
1451                 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1452
1453                 if (cic == NULL)
1454                         goto err;
1455
1456                 /*
1457                  * manually increment generic io_context usage count, it
1458                  * cannot go away since we are already holding one ref to it
1459                  */
1460                 ioc->cic = cic;
1461                 ioc->set_ioprio = cfq_ioc_set_ioprio;
1462                 cic->ioc = ioc;
1463                 cic->key = cfqd;
1464                 atomic_inc(&cfqd->ref);
1465         } else {
1466                 struct cfq_io_context *__cic;
1467
1468                 /*
1469                  * the first cic on the list is actually the head itself
1470                  */
1471                 if (cic->key == cfqd)
1472                         goto out;
1473
1474                 /*
1475                  * cic exists, check if we already are there. linear search
1476                  * should be ok here, the list will usually not be more than
1477                  * 1 or a few entries long
1478                  */
1479                 list_for_each_entry(__cic, &cic->list, list) {
1480                         /*
1481                          * this process is already holding a reference to
1482                          * this queue, so no need to get one more
1483                          */
1484                         if (__cic->key == cfqd) {
1485                                 cic = __cic;
1486                                 goto out;
1487                         }
1488                 }
1489
1490                 /*
1491                  * nope, process doesn't have a cic assoicated with this
1492                  * cfqq yet. get a new one and add to list
1493                  */
1494                 __cic = cfq_alloc_io_context(cfqd, gfp_mask);
1495                 if (__cic == NULL)
1496                         goto err;
1497
1498                 __cic->ioc = ioc;
1499                 __cic->key = cfqd;
1500                 atomic_inc(&cfqd->ref);
1501                 list_add(&__cic->list, &cic->list);
1502                 cic = __cic;
1503         }
1504
1505 out:
1506         return cic;
1507 err:
1508         put_io_context(ioc);
1509         return NULL;
1510 }
1511
1512 static void
1513 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1514 {
1515         unsigned long elapsed, ttime;
1516
1517         /*
1518          * if this context already has stuff queued, thinktime is from
1519          * last queue not last end
1520          */
1521 #if 0
1522         if (time_after(cic->last_end_request, cic->last_queue))
1523                 elapsed = jiffies - cic->last_end_request;
1524         else
1525                 elapsed = jiffies - cic->last_queue;
1526 #else
1527                 elapsed = jiffies - cic->last_end_request;
1528 #endif
1529
1530         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1531
1532         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1533         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1534         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1535 }
1536
1537 #define sample_valid(samples)   ((samples) > 80)
1538
1539 /*
1540  * Disable idle window if the process thinks too long or seeks so much that
1541  * it doesn't matter
1542  */
1543 static void
1544 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1545                        struct cfq_io_context *cic)
1546 {
1547         int enable_idle = cfq_cfqq_idle_window(cfqq);
1548
1549         if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1550                 enable_idle = 0;
1551         else if (sample_valid(cic->ttime_samples)) {
1552                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1553                         enable_idle = 0;
1554                 else
1555                         enable_idle = 1;
1556         }
1557
1558         if (enable_idle)
1559                 cfq_mark_cfqq_idle_window(cfqq);
1560         else
1561                 cfq_clear_cfqq_idle_window(cfqq);
1562 }
1563
1564
1565 /*
1566  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1567  * no or if we aren't sure, a 1 will cause a preempt.
1568  */
1569 static int
1570 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1571                    struct cfq_rq *crq)
1572 {
1573         struct cfq_queue *cfqq = cfqd->active_queue;
1574
1575         if (cfq_class_idle(new_cfqq))
1576                 return 0;
1577
1578         if (!cfqq)
1579                 return 1;
1580
1581         if (cfq_class_idle(cfqq))
1582                 return 1;
1583         if (!cfq_cfqq_wait_request(new_cfqq))
1584                 return 0;
1585         /*
1586          * if it doesn't have slice left, forget it
1587          */
1588         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1589                 return 0;
1590         if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1591                 return 1;
1592
1593         return 0;
1594 }
1595
1596 /*
1597  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1598  * let it have half of its nominal slice.
1599  */
1600 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1601 {
1602         struct cfq_queue *__cfqq, *next;
1603
1604         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1605                 cfq_resort_rr_list(__cfqq, 1);
1606
1607         if (!cfqq->slice_left)
1608                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1609
1610         cfqq->slice_end = cfqq->slice_left + jiffies;
1611         __cfq_slice_expired(cfqd, cfqq, 1);
1612         __cfq_set_active_queue(cfqd, cfqq);
1613 }
1614
1615 /*
1616  * should really be a ll_rw_blk.c helper
1617  */
1618 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1619 {
1620         request_queue_t *q = cfqd->queue;
1621
1622         if (!blk_queue_plugged(q))
1623                 q->request_fn(q);
1624         else
1625                 __generic_unplug_device(q);
1626 }
1627
1628 /*
1629  * Called when a new fs request (crq) is added (to cfqq). Check if there's
1630  * something we should do about it
1631  */
1632 static void
1633 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1634                  struct cfq_rq *crq)
1635 {
1636         struct cfq_io_context *cic;
1637
1638         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1639
1640         /*
1641          * we never wait for an async request and we don't allow preemption
1642          * of an async request. so just return early
1643          */
1644         if (!cfq_crq_is_sync(crq))
1645                 return;
1646
1647         cic = crq->io_context;
1648
1649         cfq_update_io_thinktime(cfqd, cic);
1650         cfq_update_idle_window(cfqd, cfqq, cic);
1651
1652         cic->last_queue = jiffies;
1653
1654         if (cfqq == cfqd->active_queue) {
1655                 /*
1656                  * if we are waiting for a request for this queue, let it rip
1657                  * immediately and flag that we must not expire this queue
1658                  * just now
1659                  */
1660                 if (cfq_cfqq_wait_request(cfqq)) {
1661                         cfq_mark_cfqq_must_dispatch(cfqq);
1662                         del_timer(&cfqd->idle_slice_timer);
1663                         cfq_start_queueing(cfqd, cfqq);
1664                 }
1665         } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1666                 /*
1667                  * not the active queue - expire current slice if it is
1668                  * idle and has expired it's mean thinktime or this new queue
1669                  * has some old slice time left and is of higher priority
1670                  */
1671                 cfq_preempt_queue(cfqd, cfqq);
1672                 cfq_mark_cfqq_must_dispatch(cfqq);
1673                 cfq_start_queueing(cfqd, cfqq);
1674         }
1675 }
1676
1677 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1678 {
1679         struct cfq_data *cfqd = q->elevator->elevator_data;
1680         struct cfq_rq *crq = RQ_DATA(rq);
1681         struct cfq_queue *cfqq = crq->cfq_queue;
1682
1683         cfq_init_prio_data(cfqq);
1684
1685         cfq_add_crq_rb(crq);
1686
1687         list_add_tail(&rq->queuelist, &cfqq->fifo);
1688
1689         if (rq_mergeable(rq))
1690                 cfq_add_crq_hash(cfqd, crq);
1691
1692         cfq_crq_enqueued(cfqd, cfqq, crq);
1693 }
1694
1695 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1696 {
1697         struct cfq_rq *crq = RQ_DATA(rq);
1698         struct cfq_queue *cfqq = crq->cfq_queue;
1699         struct cfq_data *cfqd = cfqq->cfqd;
1700         const int sync = cfq_crq_is_sync(crq);
1701         unsigned long now;
1702
1703         now = jiffies;
1704
1705         WARN_ON(!cfqd->rq_in_driver);
1706         WARN_ON(!cfqq->on_dispatch[sync]);
1707         cfqd->rq_in_driver--;
1708         cfqq->on_dispatch[sync]--;
1709
1710         if (!cfq_class_idle(cfqq))
1711                 cfqd->last_end_request = now;
1712
1713         if (!cfq_cfqq_dispatched(cfqq)) {
1714                 if (cfq_cfqq_on_rr(cfqq)) {
1715                         cfqq->service_last = now;
1716                         cfq_resort_rr_list(cfqq, 0);
1717                 }
1718                 if (cfq_cfqq_expired(cfqq)) {
1719                         __cfq_slice_expired(cfqd, cfqq, 0);
1720                         cfq_schedule_dispatch(cfqd);
1721                 }
1722         }
1723
1724         if (cfq_crq_is_sync(crq))
1725                 crq->io_context->last_end_request = now;
1726 }
1727
1728 static struct request *
1729 cfq_former_request(request_queue_t *q, struct request *rq)
1730 {
1731         struct cfq_rq *crq = RQ_DATA(rq);
1732         struct rb_node *rbprev = rb_prev(&crq->rb_node);
1733
1734         if (rbprev)
1735                 return rb_entry_crq(rbprev)->request;
1736
1737         return NULL;
1738 }
1739
1740 static struct request *
1741 cfq_latter_request(request_queue_t *q, struct request *rq)
1742 {
1743         struct cfq_rq *crq = RQ_DATA(rq);
1744         struct rb_node *rbnext = rb_next(&crq->rb_node);
1745
1746         if (rbnext)
1747                 return rb_entry_crq(rbnext)->request;
1748
1749         return NULL;
1750 }
1751
1752 /*
1753  * we temporarily boost lower priority queues if they are holding fs exclusive
1754  * resources. they are boosted to normal prio (CLASS_BE/4)
1755  */
1756 static void cfq_prio_boost(struct cfq_queue *cfqq)
1757 {
1758         const int ioprio_class = cfqq->ioprio_class;
1759         const int ioprio = cfqq->ioprio;
1760
1761         if (has_fs_excl()) {
1762                 /*
1763                  * boost idle prio on transactions that would lock out other
1764                  * users of the filesystem
1765                  */
1766                 if (cfq_class_idle(cfqq))
1767                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1768                 if (cfqq->ioprio > IOPRIO_NORM)
1769                         cfqq->ioprio = IOPRIO_NORM;
1770         } else {
1771                 /*
1772                  * check if we need to unboost the queue
1773                  */
1774                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1775                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1776                 if (cfqq->ioprio != cfqq->org_ioprio)
1777                         cfqq->ioprio = cfqq->org_ioprio;
1778         }
1779
1780         /*
1781          * refile between round-robin lists if we moved the priority class
1782          */
1783         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1784             cfq_cfqq_on_rr(cfqq))
1785                 cfq_resort_rr_list(cfqq, 0);
1786 }
1787
1788 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
1789 {
1790         if (rw == READ || process_sync(task))
1791                 return task->pid;
1792
1793         return CFQ_KEY_ASYNC;
1794 }
1795
1796 static inline int
1797 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1798                 struct task_struct *task, int rw)
1799 {
1800 #if 1
1801         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1802             !cfq_cfqq_must_alloc_slice(cfqq)) {
1803                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1804                 return ELV_MQUEUE_MUST;
1805         }
1806
1807         return ELV_MQUEUE_MAY;
1808 #else
1809         if (!cfqq || task->flags & PF_MEMALLOC)
1810                 return ELV_MQUEUE_MAY;
1811         if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1812                 if (cfq_cfqq_wait_request(cfqq))
1813                         return ELV_MQUEUE_MUST;
1814
1815                 /*
1816                  * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1817                  * can quickly flood the queue with writes from a single task
1818                  */
1819                 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
1820                         cfq_mark_cfqq_must_alloc_slice(cfqq);
1821                         return ELV_MQUEUE_MUST;
1822                 }
1823
1824                 return ELV_MQUEUE_MAY;
1825         }
1826         if (cfq_class_idle(cfqq))
1827                 return ELV_MQUEUE_NO;
1828         if (cfqq->allocated[rw] >= cfqd->max_queued) {
1829                 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1830                 int ret = ELV_MQUEUE_NO;
1831
1832                 if (ioc && ioc->nr_batch_requests)
1833                         ret = ELV_MQUEUE_MAY;
1834
1835                 put_io_context(ioc);
1836                 return ret;
1837         }
1838
1839         return ELV_MQUEUE_MAY;
1840 #endif
1841 }
1842
1843 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1844 {
1845         struct cfq_data *cfqd = q->elevator->elevator_data;
1846         struct task_struct *tsk = current;
1847         struct cfq_queue *cfqq;
1848
1849         /*
1850          * don't force setup of a queue from here, as a call to may_queue
1851          * does not necessarily imply that a request actually will be queued.
1852          * so just lookup a possibly existing queue, or return 'may queue'
1853          * if that fails
1854          */
1855         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1856         if (cfqq) {
1857                 cfq_init_prio_data(cfqq);
1858                 cfq_prio_boost(cfqq);
1859
1860                 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1861         }
1862
1863         return ELV_MQUEUE_MAY;
1864 }
1865
1866 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1867 {
1868         struct cfq_data *cfqd = q->elevator->elevator_data;
1869         struct request_list *rl = &q->rq;
1870
1871         if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1872                 smp_mb();
1873                 if (waitqueue_active(&rl->wait[READ]))
1874                         wake_up(&rl->wait[READ]);
1875         }
1876
1877         if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
1878                 smp_mb();
1879                 if (waitqueue_active(&rl->wait[WRITE]))
1880                         wake_up(&rl->wait[WRITE]);
1881         }
1882 }
1883
1884 /*
1885  * queue lock held here
1886  */
1887 static void cfq_put_request(request_queue_t *q, struct request *rq)
1888 {
1889         struct cfq_data *cfqd = q->elevator->elevator_data;
1890         struct cfq_rq *crq = RQ_DATA(rq);
1891
1892         if (crq) {
1893                 struct cfq_queue *cfqq = crq->cfq_queue;
1894                 const int rw = rq_data_dir(rq);
1895
1896                 BUG_ON(!cfqq->allocated[rw]);
1897                 cfqq->allocated[rw]--;
1898
1899                 put_io_context(crq->io_context->ioc);
1900
1901                 mempool_free(crq, cfqd->crq_pool);
1902                 rq->elevator_private = NULL;
1903
1904                 cfq_check_waiters(q, cfqq);
1905                 cfq_put_queue(cfqq);
1906         }
1907 }
1908
1909 /*
1910  * Allocate cfq data structures associated with this request.
1911  */
1912 static int
1913 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
1914                 gfp_t gfp_mask)
1915 {
1916         struct cfq_data *cfqd = q->elevator->elevator_data;
1917         struct task_struct *tsk = current;
1918         struct cfq_io_context *cic;
1919         const int rw = rq_data_dir(rq);
1920         pid_t key = cfq_queue_pid(tsk, rw);
1921         struct cfq_queue *cfqq;
1922         struct cfq_rq *crq;
1923         unsigned long flags;
1924
1925         might_sleep_if(gfp_mask & __GFP_WAIT);
1926
1927         cic = cfq_get_io_context(cfqd, key, gfp_mask);
1928
1929         spin_lock_irqsave(q->queue_lock, flags);
1930
1931         if (!cic)
1932                 goto queue_fail;
1933
1934         if (!cic->cfqq) {
1935                 cfqq = cfq_get_queue(cfqd, key, tsk->ioprio, gfp_mask);
1936                 if (!cfqq)
1937                         goto queue_fail;
1938
1939                 cic->cfqq = cfqq;
1940         } else
1941                 cfqq = cic->cfqq;
1942
1943         cfqq->allocated[rw]++;
1944         cfq_clear_cfqq_must_alloc(cfqq);
1945         cfqd->rq_starved = 0;
1946         atomic_inc(&cfqq->ref);
1947         spin_unlock_irqrestore(q->queue_lock, flags);
1948
1949         crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
1950         if (crq) {
1951                 RB_CLEAR(&crq->rb_node);
1952                 crq->rb_key = 0;
1953                 crq->request = rq;
1954                 INIT_HLIST_NODE(&crq->hash);
1955                 crq->cfq_queue = cfqq;
1956                 crq->io_context = cic;
1957
1958                 if (rw == READ || process_sync(tsk))
1959                         cfq_mark_crq_is_sync(crq);
1960                 else
1961                         cfq_clear_crq_is_sync(crq);
1962
1963                 rq->elevator_private = crq;
1964                 return 0;
1965         }
1966
1967         spin_lock_irqsave(q->queue_lock, flags);
1968         cfqq->allocated[rw]--;
1969         if (!(cfqq->allocated[0] + cfqq->allocated[1]))
1970                 cfq_mark_cfqq_must_alloc(cfqq);
1971         cfq_put_queue(cfqq);
1972 queue_fail:
1973         if (cic)
1974                 put_io_context(cic->ioc);
1975         /*
1976          * mark us rq allocation starved. we need to kickstart the process
1977          * ourselves if there are no pending requests that can do it for us.
1978          * that would be an extremely rare OOM situation
1979          */
1980         cfqd->rq_starved = 1;
1981         cfq_schedule_dispatch(cfqd);
1982         spin_unlock_irqrestore(q->queue_lock, flags);
1983         return 1;
1984 }
1985
1986 static void cfq_kick_queue(void *data)
1987 {
1988         request_queue_t *q = data;
1989         struct cfq_data *cfqd = q->elevator->elevator_data;
1990         unsigned long flags;
1991
1992         spin_lock_irqsave(q->queue_lock, flags);
1993
1994         if (cfqd->rq_starved) {
1995                 struct request_list *rl = &q->rq;
1996
1997                 /*
1998                  * we aren't guaranteed to get a request after this, but we
1999                  * have to be opportunistic
2000                  */
2001                 smp_mb();
2002                 if (waitqueue_active(&rl->wait[READ]))
2003                         wake_up(&rl->wait[READ]);
2004                 if (waitqueue_active(&rl->wait[WRITE]))
2005                         wake_up(&rl->wait[WRITE]);
2006         }
2007
2008         blk_remove_plug(q);
2009         q->request_fn(q);
2010         spin_unlock_irqrestore(q->queue_lock, flags);
2011 }
2012
2013 /*
2014  * Timer running if the active_queue is currently idling inside its time slice
2015  */
2016 static void cfq_idle_slice_timer(unsigned long data)
2017 {
2018         struct cfq_data *cfqd = (struct cfq_data *) data;
2019         struct cfq_queue *cfqq;
2020         unsigned long flags;
2021
2022         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2023
2024         if ((cfqq = cfqd->active_queue) != NULL) {
2025                 unsigned long now = jiffies;
2026
2027                 /*
2028                  * expired
2029                  */
2030                 if (time_after(now, cfqq->slice_end))
2031                         goto expire;
2032
2033                 /*
2034                  * only expire and reinvoke request handler, if there are
2035                  * other queues with pending requests
2036                  */
2037                 if (!cfqd->busy_queues) {
2038                         cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2039                         add_timer(&cfqd->idle_slice_timer);
2040                         goto out_cont;
2041                 }
2042
2043                 /*
2044                  * not expired and it has a request pending, let it dispatch
2045                  */
2046                 if (!RB_EMPTY(&cfqq->sort_list)) {
2047                         cfq_mark_cfqq_must_dispatch(cfqq);
2048                         goto out_kick;
2049                 }
2050         }
2051 expire:
2052         cfq_slice_expired(cfqd, 0);
2053 out_kick:
2054         cfq_schedule_dispatch(cfqd);
2055 out_cont:
2056         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2057 }
2058
2059 /*
2060  * Timer running if an idle class queue is waiting for service
2061  */
2062 static void cfq_idle_class_timer(unsigned long data)
2063 {
2064         struct cfq_data *cfqd = (struct cfq_data *) data;
2065         unsigned long flags, end;
2066
2067         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2068
2069         /*
2070          * race with a non-idle queue, reset timer
2071          */
2072         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2073         if (!time_after_eq(jiffies, end)) {
2074                 cfqd->idle_class_timer.expires = end;
2075                 add_timer(&cfqd->idle_class_timer);
2076         } else
2077                 cfq_schedule_dispatch(cfqd);
2078
2079         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2080 }
2081
2082 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2083 {
2084         del_timer_sync(&cfqd->idle_slice_timer);
2085         del_timer_sync(&cfqd->idle_class_timer);
2086         blk_sync_queue(cfqd->queue);
2087 }
2088
2089 static void cfq_put_cfqd(struct cfq_data *cfqd)
2090 {
2091         request_queue_t *q = cfqd->queue;
2092
2093         if (!atomic_dec_and_test(&cfqd->ref))
2094                 return;
2095
2096         cfq_shutdown_timer_wq(cfqd);
2097         blk_put_queue(q);
2098
2099         mempool_destroy(cfqd->crq_pool);
2100         kfree(cfqd->crq_hash);
2101         kfree(cfqd->cfq_hash);
2102         kfree(cfqd);
2103 }
2104
2105 static void cfq_exit_queue(elevator_t *e)
2106 {
2107         struct cfq_data *cfqd = e->elevator_data;
2108
2109         cfq_shutdown_timer_wq(cfqd);
2110         cfq_put_cfqd(cfqd);
2111 }
2112
2113 static int cfq_init_queue(request_queue_t *q, elevator_t *e)
2114 {
2115         struct cfq_data *cfqd;
2116         int i;
2117
2118         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2119         if (!cfqd)
2120                 return -ENOMEM;
2121
2122         memset(cfqd, 0, sizeof(*cfqd));
2123
2124         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2125                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2126
2127         INIT_LIST_HEAD(&cfqd->busy_rr);
2128         INIT_LIST_HEAD(&cfqd->cur_rr);
2129         INIT_LIST_HEAD(&cfqd->idle_rr);
2130         INIT_LIST_HEAD(&cfqd->empty_list);
2131
2132         cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2133         if (!cfqd->crq_hash)
2134                 goto out_crqhash;
2135
2136         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2137         if (!cfqd->cfq_hash)
2138                 goto out_cfqhash;
2139
2140         cfqd->crq_pool = mempool_create(BLKDEV_MIN_RQ, mempool_alloc_slab, mempool_free_slab, crq_pool);
2141         if (!cfqd->crq_pool)
2142                 goto out_crqpool;
2143
2144         for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2145                 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2146         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2147                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2148
2149         e->elevator_data = cfqd;
2150
2151         cfqd->queue = q;
2152         atomic_inc(&q->refcnt);
2153
2154         cfqd->max_queued = q->nr_requests / 4;
2155         q->nr_batching = cfq_queued;
2156
2157         init_timer(&cfqd->idle_slice_timer);
2158         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2159         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2160
2161         init_timer(&cfqd->idle_class_timer);
2162         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2163         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2164
2165         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2166
2167         atomic_set(&cfqd->ref, 1);
2168
2169         cfqd->cfq_queued = cfq_queued;
2170         cfqd->cfq_quantum = cfq_quantum;
2171         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2172         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2173         cfqd->cfq_back_max = cfq_back_max;
2174         cfqd->cfq_back_penalty = cfq_back_penalty;
2175         cfqd->cfq_slice[0] = cfq_slice_async;
2176         cfqd->cfq_slice[1] = cfq_slice_sync;
2177         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2178         cfqd->cfq_slice_idle = cfq_slice_idle;
2179         cfqd->cfq_max_depth = cfq_max_depth;
2180
2181         return 0;
2182 out_crqpool:
2183         kfree(cfqd->cfq_hash);
2184 out_cfqhash:
2185         kfree(cfqd->crq_hash);
2186 out_crqhash:
2187         kfree(cfqd);
2188         return -ENOMEM;
2189 }
2190
2191 static void cfq_slab_kill(void)
2192 {
2193         if (crq_pool)
2194                 kmem_cache_destroy(crq_pool);
2195         if (cfq_pool)
2196                 kmem_cache_destroy(cfq_pool);
2197         if (cfq_ioc_pool)
2198                 kmem_cache_destroy(cfq_ioc_pool);
2199 }
2200
2201 static int __init cfq_slab_setup(void)
2202 {
2203         crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2204                                         NULL, NULL);
2205         if (!crq_pool)
2206                 goto fail;
2207
2208         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2209                                         NULL, NULL);
2210         if (!cfq_pool)
2211                 goto fail;
2212
2213         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2214                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2215         if (!cfq_ioc_pool)
2216                 goto fail;
2217
2218         return 0;
2219 fail:
2220         cfq_slab_kill();
2221         return -ENOMEM;
2222 }
2223
2224 /*
2225  * sysfs parts below -->
2226  */
2227 struct cfq_fs_entry {
2228         struct attribute attr;
2229         ssize_t (*show)(struct cfq_data *, char *);
2230         ssize_t (*store)(struct cfq_data *, const char *, size_t);
2231 };
2232
2233 static ssize_t
2234 cfq_var_show(unsigned int var, char *page)
2235 {
2236         return sprintf(page, "%d\n", var);
2237 }
2238
2239 static ssize_t
2240 cfq_var_store(unsigned int *var, const char *page, size_t count)
2241 {
2242         char *p = (char *) page;
2243
2244         *var = simple_strtoul(p, &p, 10);
2245         return count;
2246 }
2247
2248 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2249 static ssize_t __FUNC(struct cfq_data *cfqd, char *page)                \
2250 {                                                                       \
2251         unsigned int __data = __VAR;                                    \
2252         if (__CONV)                                                     \
2253                 __data = jiffies_to_msecs(__data);                      \
2254         return cfq_var_show(__data, (page));                            \
2255 }
2256 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2257 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2258 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2259 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2260 SHOW_FUNCTION(cfq_back_max_show, cfqd->cfq_back_max, 0);
2261 SHOW_FUNCTION(cfq_back_penalty_show, cfqd->cfq_back_penalty, 0);
2262 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2263 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2264 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2265 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2266 SHOW_FUNCTION(cfq_max_depth_show, cfqd->cfq_max_depth, 0);
2267 #undef SHOW_FUNCTION
2268
2269 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2270 static ssize_t __FUNC(struct cfq_data *cfqd, const char *page, size_t count)    \
2271 {                                                                       \
2272         unsigned int __data;                                            \
2273         int ret = cfq_var_store(&__data, (page), count);                \
2274         if (__data < (MIN))                                             \
2275                 __data = (MIN);                                         \
2276         else if (__data > (MAX))                                        \
2277                 __data = (MAX);                                         \
2278         if (__CONV)                                                     \
2279                 *(__PTR) = msecs_to_jiffies(__data);                    \
2280         else                                                            \
2281                 *(__PTR) = __data;                                      \
2282         return ret;                                                     \
2283 }
2284 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2285 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2286 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2287 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2288 STORE_FUNCTION(cfq_back_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2289 STORE_FUNCTION(cfq_back_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2290 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2291 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2292 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2293 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2294 STORE_FUNCTION(cfq_max_depth_store, &cfqd->cfq_max_depth, 1, UINT_MAX, 0);
2295 #undef STORE_FUNCTION
2296
2297 static struct cfq_fs_entry cfq_quantum_entry = {
2298         .attr = {.name = "quantum", .mode = S_IRUGO | S_IWUSR },
2299         .show = cfq_quantum_show,
2300         .store = cfq_quantum_store,
2301 };
2302 static struct cfq_fs_entry cfq_queued_entry = {
2303         .attr = {.name = "queued", .mode = S_IRUGO | S_IWUSR },
2304         .show = cfq_queued_show,
2305         .store = cfq_queued_store,
2306 };
2307 static struct cfq_fs_entry cfq_fifo_expire_sync_entry = {
2308         .attr = {.name = "fifo_expire_sync", .mode = S_IRUGO | S_IWUSR },
2309         .show = cfq_fifo_expire_sync_show,
2310         .store = cfq_fifo_expire_sync_store,
2311 };
2312 static struct cfq_fs_entry cfq_fifo_expire_async_entry = {
2313         .attr = {.name = "fifo_expire_async", .mode = S_IRUGO | S_IWUSR },
2314         .show = cfq_fifo_expire_async_show,
2315         .store = cfq_fifo_expire_async_store,
2316 };
2317 static struct cfq_fs_entry cfq_back_max_entry = {
2318         .attr = {.name = "back_seek_max", .mode = S_IRUGO | S_IWUSR },
2319         .show = cfq_back_max_show,
2320         .store = cfq_back_max_store,
2321 };
2322 static struct cfq_fs_entry cfq_back_penalty_entry = {
2323         .attr = {.name = "back_seek_penalty", .mode = S_IRUGO | S_IWUSR },
2324         .show = cfq_back_penalty_show,
2325         .store = cfq_back_penalty_store,
2326 };
2327 static struct cfq_fs_entry cfq_slice_sync_entry = {
2328         .attr = {.name = "slice_sync", .mode = S_IRUGO | S_IWUSR },
2329         .show = cfq_slice_sync_show,
2330         .store = cfq_slice_sync_store,
2331 };
2332 static struct cfq_fs_entry cfq_slice_async_entry = {
2333         .attr = {.name = "slice_async", .mode = S_IRUGO | S_IWUSR },
2334         .show = cfq_slice_async_show,
2335         .store = cfq_slice_async_store,
2336 };
2337 static struct cfq_fs_entry cfq_slice_async_rq_entry = {
2338         .attr = {.name = "slice_async_rq", .mode = S_IRUGO | S_IWUSR },
2339         .show = cfq_slice_async_rq_show,
2340         .store = cfq_slice_async_rq_store,
2341 };
2342 static struct cfq_fs_entry cfq_slice_idle_entry = {
2343         .attr = {.name = "slice_idle", .mode = S_IRUGO | S_IWUSR },
2344         .show = cfq_slice_idle_show,
2345         .store = cfq_slice_idle_store,
2346 };
2347 static struct cfq_fs_entry cfq_max_depth_entry = {
2348         .attr = {.name = "max_depth", .mode = S_IRUGO | S_IWUSR },
2349         .show = cfq_max_depth_show,
2350         .store = cfq_max_depth_store,
2351 };
2352
2353 static struct attribute *default_attrs[] = {
2354         &cfq_quantum_entry.attr,
2355         &cfq_queued_entry.attr,
2356         &cfq_fifo_expire_sync_entry.attr,
2357         &cfq_fifo_expire_async_entry.attr,
2358         &cfq_back_max_entry.attr,
2359         &cfq_back_penalty_entry.attr,
2360         &cfq_slice_sync_entry.attr,
2361         &cfq_slice_async_entry.attr,
2362         &cfq_slice_async_rq_entry.attr,
2363         &cfq_slice_idle_entry.attr,
2364         &cfq_max_depth_entry.attr,
2365         NULL,
2366 };
2367
2368 #define to_cfq(atr) container_of((atr), struct cfq_fs_entry, attr)
2369
2370 static ssize_t
2371 cfq_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
2372 {
2373         elevator_t *e = container_of(kobj, elevator_t, kobj);
2374         struct cfq_fs_entry *entry = to_cfq(attr);
2375
2376         if (!entry->show)
2377                 return -EIO;
2378
2379         return entry->show(e->elevator_data, page);
2380 }
2381
2382 static ssize_t
2383 cfq_attr_store(struct kobject *kobj, struct attribute *attr,
2384                const char *page, size_t length)
2385 {
2386         elevator_t *e = container_of(kobj, elevator_t, kobj);
2387         struct cfq_fs_entry *entry = to_cfq(attr);
2388
2389         if (!entry->store)
2390                 return -EIO;
2391
2392         return entry->store(e->elevator_data, page, length);
2393 }
2394
2395 static struct sysfs_ops cfq_sysfs_ops = {
2396         .show   = cfq_attr_show,
2397         .store  = cfq_attr_store,
2398 };
2399
2400 static struct kobj_type cfq_ktype = {
2401         .sysfs_ops      = &cfq_sysfs_ops,
2402         .default_attrs  = default_attrs,
2403 };
2404
2405 static struct elevator_type iosched_cfq = {
2406         .ops = {
2407                 .elevator_merge_fn =            cfq_merge,
2408                 .elevator_merged_fn =           cfq_merged_request,
2409                 .elevator_merge_req_fn =        cfq_merged_requests,
2410                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2411                 .elevator_add_req_fn =          cfq_insert_request,
2412                 .elevator_activate_req_fn =     cfq_activate_request,
2413                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2414                 .elevator_queue_empty_fn =      cfq_queue_empty,
2415                 .elevator_completed_req_fn =    cfq_completed_request,
2416                 .elevator_former_req_fn =       cfq_former_request,
2417                 .elevator_latter_req_fn =       cfq_latter_request,
2418                 .elevator_set_req_fn =          cfq_set_request,
2419                 .elevator_put_req_fn =          cfq_put_request,
2420                 .elevator_may_queue_fn =        cfq_may_queue,
2421                 .elevator_init_fn =             cfq_init_queue,
2422                 .elevator_exit_fn =             cfq_exit_queue,
2423         },
2424         .elevator_ktype =       &cfq_ktype,
2425         .elevator_name =        "cfq",
2426         .elevator_owner =       THIS_MODULE,
2427 };
2428
2429 static int __init cfq_init(void)
2430 {
2431         int ret;
2432
2433         /*
2434          * could be 0 on HZ < 1000 setups
2435          */
2436         if (!cfq_slice_async)
2437                 cfq_slice_async = 1;
2438         if (!cfq_slice_idle)
2439                 cfq_slice_idle = 1;
2440
2441         if (cfq_slab_setup())
2442                 return -ENOMEM;
2443
2444         ret = elv_register(&iosched_cfq);
2445         if (ret)
2446                 cfq_slab_kill();
2447
2448         return ret;
2449 }
2450
2451 static void __exit cfq_exit(void)
2452 {
2453         elv_unregister(&iosched_cfq);
2454         cfq_slab_kill();
2455 }
2456
2457 module_init(cfq_init);
2458 module_exit(cfq_exit);
2459
2460 MODULE_AUTHOR("Jens Axboe");
2461 MODULE_LICENSE("GPL");
2462 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");