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

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