block/cfq-iosched.c

   1 /*
   2  *  CFQ, or complete fairness queueing, disk scheduler.
   3  *
   4  *  Based on ideas from a previously unfinished io
   5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   6  *
   7  *  Copyright (C) 2003 Jens Axboe <axboe@suse.de>
   8  */
   9 #include <linux/config.h>
  10 #include <linux/module.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/elevator.h>
  13 #include <linux/hash.h>
  14 #include <linux/rbtree.h>
  15 #include <linux/ioprio.h>
  16
  17 /*
  18  * tunables
  19  */
  20 static const int cfq_quantum = 4;               /* max queue in one round of service */
  21 static const int cfq_queued = 8;                /* minimum rq allocate limit per-queue*/
  22 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  23 static const int cfq_back_max = 16 * 1024;      /* maximum backwards seek, in KiB */
  24 static const int cfq_back_penalty = 2;          /* penalty of a backwards seek */
  25
  26 static const int cfq_slice_sync = HZ / 10;
  27 static int cfq_slice_async = HZ / 25;
  28 static const int cfq_slice_async_rq = 2;
  29 static int cfq_slice_idle = HZ / 70;
  30
  31 #define CFQ_IDLE_GRACE          (HZ / 10)
  32 #define CFQ_SLICE_SCALE         (5)
  33
  34 #define CFQ_KEY_ASYNC           (0)
  35
  36 static DEFINE_SPINLOCK(cfq_exit_lock);
  37
  38 /*
  39  * for the hash of cfqq inside the cfqd
  40  */
  41 #define CFQ_QHASH_SHIFT         6
  42 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
  43 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
  44
  45 /*
  46  * for the hash of crq inside the cfqq
  47  */
  48 #define CFQ_MHASH_SHIFT         6
  49 #define CFQ_MHASH_BLOCK(sec)    ((sec) >> 3)
  50 #define CFQ_MHASH_ENTRIES       (1 << CFQ_MHASH_SHIFT)
  51 #define CFQ_MHASH_FN(sec)       hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
  52 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
  53 #define list_entry_hash(ptr)    hlist_entry((ptr), struct cfq_rq, hash)
  54
  55 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
  56 #define list_entry_fifo(ptr)    list_entry((ptr), struct request, queuelist)
  57
  58 #define RQ_DATA(rq)             (rq)->elevator_private
  59
  60 /*
  61  * rb-tree defines
  62  */
  63 #define RB_EMPTY(node)          ((node)->rb_node == NULL)
  64 #define RB_CLEAR(node)          do {    \
  65                 memset(node, 0, sizeof(*node)); \
  66 } while (0)
  67 #define RB_CLEAR_ROOT(root)     ((root)->rb_node = NULL)
  68 #define rb_entry_crq(node)      rb_entry((node), struct cfq_rq, rb_node)
  69 #define rq_rb_key(rq)           (rq)->sector
  70
  71 static kmem_cache_t *crq_pool;
  72 static kmem_cache_t *cfq_pool;
  73 static kmem_cache_t *cfq_ioc_pool;
  74
  75 static atomic_t ioc_count = ATOMIC_INIT(0);
  76 static struct completion *ioc_gone;
  77
  78 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  79 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  80 #define cfq_class_be(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
  81 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  82
  83 #define ASYNC                   (0)
  84 #define SYNC                    (1)
  85
  86 #define cfq_cfqq_dispatched(cfqq)       \
  87         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
  88
  89 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
  90
  91 #define cfq_cfqq_sync(cfqq)             \
  92         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
  93
  94 #define sample_valid(samples)   ((samples) > 80)
  95
  96 /*
  97  * Per block device queue structure
  98  */
  99 struct cfq_data {
 100         request_queue_t *queue;
 101
 102         /*
 103          * rr list of queues with requests and the count of them
 104          */
 105         struct list_head rr_list[CFQ_PRIO_LISTS];
 106         struct list_head busy_rr;
 107         struct list_head cur_rr;
 108         struct list_head idle_rr;
 109         unsigned int busy_queues;
 110
 111         /*
 112          * non-ordered list of empty cfqq's
 113          */
 114         struct list_head empty_list;
 115
 116         /*
 117          * cfqq lookup hash
 118          */
 119         struct hlist_head *cfq_hash;
 120
 121         /*
 122          * global crq hash for all queues
 123          */
 124         struct hlist_head *crq_hash;
 125
 126         mempool_t *crq_pool;
 127
 128         int rq_in_driver;
 129         int hw_tag;
 130
 131         /*
 132          * schedule slice state info
 133          */
 134         /*
 135          * idle window management
 136          */
 137         struct timer_list idle_slice_timer;
 138         struct work_struct unplug_work;
 139
 140         struct cfq_queue *active_queue;
 141         struct cfq_io_context *active_cic;
 142         int cur_prio, cur_end_prio;
 143         unsigned int dispatch_slice;
 144
 145         struct timer_list idle_class_timer;
 146
 147         sector_t last_sector;
 148         unsigned long last_end_request;
 149
 150         unsigned int rq_starved;
 151
 152         /*
 153          * tunables, see top of file
 154          */
 155         unsigned int cfq_quantum;
 156         unsigned int cfq_queued;
 157         unsigned int cfq_fifo_expire[2];
 158         unsigned int cfq_back_penalty;
 159         unsigned int cfq_back_max;
 160         unsigned int cfq_slice[2];
 161         unsigned int cfq_slice_async_rq;
 162         unsigned int cfq_slice_idle;
 163
 164         struct list_head cic_list;
 165 };
 166
 167 /*
 168  * Per process-grouping structure
 169  */
 170 struct cfq_queue {
 171         /* reference count */
 172         atomic_t ref;
 173         /* parent cfq_data */
 174         struct cfq_data *cfqd;
 175         /* cfqq lookup hash */
 176         struct hlist_node cfq_hash;
 177         /* hash key */
 178         unsigned int key;
 179         /* on either rr or empty list of cfqd */
 180         struct list_head cfq_list;
 181         /* sorted list of pending requests */
 182         struct rb_root sort_list;
 183         /* if fifo isn't expired, next request to serve */
 184         struct cfq_rq *next_crq;
 185         /* requests queued in sort_list */
 186         int queued[2];
 187         /* currently allocated requests */
 188         int allocated[2];
 189         /* fifo list of requests in sort_list */
 190         struct list_head fifo;
 191
 192         unsigned long slice_start;
 193         unsigned long slice_end;
 194         unsigned long slice_left;
 195         unsigned long service_last;
 196
 197         /* number of requests that are on the dispatch list */
 198         int on_dispatch[2];
 199
 200         /* io prio of this group */
 201         unsigned short ioprio, org_ioprio;
 202         unsigned short ioprio_class, org_ioprio_class;
 203
 204         /* various state flags, see below */
 205         unsigned int flags;
 206 };
 207
 208 struct cfq_rq {
 209         struct rb_node rb_node;
 210         sector_t rb_key;
 211         struct request *request;
 212         struct hlist_node hash;
 213
 214         struct cfq_queue *cfq_queue;
 215         struct cfq_io_context *io_context;
 216
 217         unsigned int crq_flags;
 218 };
 219
 220 enum cfqq_state_flags {
 221         CFQ_CFQQ_FLAG_on_rr = 0,
 222         CFQ_CFQQ_FLAG_wait_request,
 223         CFQ_CFQQ_FLAG_must_alloc,
 224         CFQ_CFQQ_FLAG_must_alloc_slice,
 225         CFQ_CFQQ_FLAG_must_dispatch,
 226         CFQ_CFQQ_FLAG_fifo_expire,
 227         CFQ_CFQQ_FLAG_idle_window,
 228         CFQ_CFQQ_FLAG_prio_changed,
 229 };
 230
 231 #define CFQ_CFQQ_FNS(name)                                              \
 232 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 233 {                                                                       \
 234         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 235 }                                                                       \
 236 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 237 {                                                                       \
 238         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 239 }                                                                       \
 240 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 241 {                                                                       \
 242         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 243 }
 244
 245 CFQ_CFQQ_FNS(on_rr);
 246 CFQ_CFQQ_FNS(wait_request);
 247 CFQ_CFQQ_FNS(must_alloc);
 248 CFQ_CFQQ_FNS(must_alloc_slice);
 249 CFQ_CFQQ_FNS(must_dispatch);
 250 CFQ_CFQQ_FNS(fifo_expire);
 251 CFQ_CFQQ_FNS(idle_window);
 252 CFQ_CFQQ_FNS(prio_changed);
 253 #undef CFQ_CFQQ_FNS
 254
 255 enum cfq_rq_state_flags {
 256         CFQ_CRQ_FLAG_is_sync = 0,
 257 };
 258
 259 #define CFQ_CRQ_FNS(name)                                               \
 260 static inline void cfq_mark_crq_##name(struct cfq_rq *crq)              \
 261 {                                                                       \
 262         crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name);                   \
 263 }                                                                       \
 264 static inline void cfq_clear_crq_##name(struct cfq_rq *crq)             \
 265 {                                                                       \
 266         crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name);                  \
 267 }                                                                       \
 268 static inline int cfq_crq_##name(const struct cfq_rq *crq)              \
 269 {                                                                       \
 270         return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0;      \
 271 }
 272
 273 CFQ_CRQ_FNS(is_sync);
 274 #undef CFQ_CRQ_FNS
 275
 276 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 277 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
 278 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
 279
 280 /*
 281  * lots of deadline iosched dupes, can be abstracted later...
 282  */
 283 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
 284 {
 285         hlist_del_init(&crq->hash);
 286 }
 287
 288 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
 289 {
 290         const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
 291
 292         hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
 293 }
 294
 295 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
 296 {
 297         struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
 298         struct hlist_node *entry, *next;
 299
 300         hlist_for_each_safe(entry, next, hash_list) {
 301                 struct cfq_rq *crq = list_entry_hash(entry);
 302                 struct request *__rq = crq->request;
 303
 304                 if (!rq_mergeable(__rq)) {
 305                         cfq_del_crq_hash(crq);
 306                         continue;
 307                 }
 308
 309                 if (rq_hash_key(__rq) == offset)
 310                         return __rq;
 311         }
 312
 313         return NULL;
 314 }
 315
 316 /*
 317  * scheduler run of queue, if there are requests pending and no one in the
 318  * driver that will restart queueing
 319  */
 320 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 321 {
 322         if (cfqd->busy_queues)
 323                 kblockd_schedule_work(&cfqd->unplug_work);
 324 }
 325
 326 static int cfq_queue_empty(request_queue_t *q)
 327 {
 328         struct cfq_data *cfqd = q->elevator->elevator_data;
 329
 330         return !cfqd->busy_queues;
 331 }
 332
 333 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
 334 {
 335         if (rw == READ || rw == WRITE_SYNC)
 336                 return task->pid;
 337
 338         return CFQ_KEY_ASYNC;
 339 }
 340
 341 /*
 342  * Lifted from AS - choose which of crq1 and crq2 that is best served now.
 343  * We choose the request that is closest to the head right now. Distance
 344  * behind the head is penalized and only allowed to a certain extent.
 345  */
 346 static struct cfq_rq *
 347 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
 348 {
 349         sector_t last, s1, s2, d1 = 0, d2 = 0;
 350         unsigned long back_max;
 351 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 352 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 353         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 354
 355         if (crq1 == NULL || crq1 == crq2)
 356                 return crq2;
 357         if (crq2 == NULL)
 358                 return crq1;
 359
 360         if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
 361                 return crq1;
 362         else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
 363                 return crq2;
 364
 365         s1 = crq1->request->sector;
 366         s2 = crq2->request->sector;
 367
 368         last = cfqd->last_sector;
 369
 370         /*
 371          * by definition, 1KiB is 2 sectors
 372          */
 373         back_max = cfqd->cfq_back_max * 2;
 374
 375         /*
 376          * Strict one way elevator _except_ in the case where we allow
 377          * short backward seeks which are biased as twice the cost of a
 378          * similar forward seek.
 379          */
 380         if (s1 >= last)
 381                 d1 = s1 - last;
 382         else if (s1 + back_max >= last)
 383                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 384         else
 385                 wrap |= CFQ_RQ1_WRAP;
 386
 387         if (s2 >= last)
 388                 d2 = s2 - last;
 389         else if (s2 + back_max >= last)
 390                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 391         else
 392                 wrap |= CFQ_RQ2_WRAP;
 393
 394         /* Found required data */
 395
 396         /*
 397          * By doing switch() on the bit mask "wrap" we avoid having to
 398          * check two variables for all permutations: --> faster!
 399          */
 400         switch (wrap) {
 401         case 0: /* common case for CFQ: crq1 and crq2 not wrapped */
 402                 if (d1 < d2)
 403                         return crq1;
 404                 else if (d2 < d1)
 405                         return crq2;
 406                 else {
 407                         if (s1 >= s2)
 408                                 return crq1;
 409                         else
 410                                 return crq2;
 411                 }
 412
 413         case CFQ_RQ2_WRAP:
 414                 return crq1;
 415         case CFQ_RQ1_WRAP:
 416                 return crq2;
 417         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both crqs wrapped */
 418         default:
 419                 /*
 420                  * Since both rqs are wrapped,
 421                  * start with the one that's further behind head
 422                  * (--> only *one* back seek required),
 423                  * since back seek takes more time than forward.
 424                  */
 425                 if (s1 <= s2)
 426                         return crq1;
 427                 else
 428                         return crq2;
 429         }
 430 }
 431
 432 /*
 433  * would be nice to take fifo expire time into account as well
 434  */
 435 static struct cfq_rq *
 436 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 437                   struct cfq_rq *last)
 438 {
 439         struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
 440         struct rb_node *rbnext, *rbprev;
 441
 442         if (!(rbnext = rb_next(&last->rb_node))) {
 443                 rbnext = rb_first(&cfqq->sort_list);
 444                 if (rbnext == &last->rb_node)
 445                         rbnext = NULL;
 446         }
 447
 448         rbprev = rb_prev(&last->rb_node);
 449
 450         if (rbprev)
 451                 crq_prev = rb_entry_crq(rbprev);
 452         if (rbnext)
 453                 crq_next = rb_entry_crq(rbnext);
 454
 455         return cfq_choose_req(cfqd, crq_next, crq_prev);
 456 }
 457
 458 static void cfq_update_next_crq(struct cfq_rq *crq)
 459 {
 460         struct cfq_queue *cfqq = crq->cfq_queue;
 461
 462         if (cfqq->next_crq == crq)
 463                 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
 464 }
 465
 466 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 467 {
 468         struct cfq_data *cfqd = cfqq->cfqd;
 469         struct list_head *list, *entry;
 470
 471         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 472
 473         list_del(&cfqq->cfq_list);
 474
 475         if (cfq_class_rt(cfqq))
 476                 list = &cfqd->cur_rr;
 477         else if (cfq_class_idle(cfqq))
 478                 list = &cfqd->idle_rr;
 479         else {
 480                 /*
 481                  * if cfqq has requests in flight, don't allow it to be
 482                  * found in cfq_set_active_queue before it has finished them.
 483                  * this is done to increase fairness between a process that
 484                  * has lots of io pending vs one that only generates one
 485                  * sporadically or synchronously
 486                  */
 487                 if (cfq_cfqq_dispatched(cfqq))
 488                         list = &cfqd->busy_rr;
 489                 else
 490                         list = &cfqd->rr_list[cfqq->ioprio];
 491         }
 492
 493         /*
 494          * if queue was preempted, just add to front to be fair. busy_rr
 495          * isn't sorted, but insert at the back for fairness.
 496          */
 497         if (preempted || list == &cfqd->busy_rr) {
 498                 if (preempted)
 499                         list = list->prev;
 500
 501                 list_add_tail(&cfqq->cfq_list, list);
 502                 return;
 503         }
 504
 505         /*
 506          * sort by when queue was last serviced
 507          */
 508         entry = list;
 509         while ((entry = entry->prev) != list) {
 510                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
 511
 512                 if (!__cfqq->service_last)
 513                         break;
 514                 if (time_before(__cfqq->service_last, cfqq->service_last))
 515                         break;
 516         }
 517
 518         list_add(&cfqq->cfq_list, entry);
 519 }
 520
 521 /*
 522  * add to busy list of queues for service, trying to be fair in ordering
 523  * the pending list according to last request service
 524  */
 525 static inline void
 526 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 527 {
 528         BUG_ON(cfq_cfqq_on_rr(cfqq));
 529         cfq_mark_cfqq_on_rr(cfqq);
 530         cfqd->busy_queues++;
 531
 532         cfq_resort_rr_list(cfqq, 0);
 533 }
 534
 535 static inline void
 536 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 537 {
 538         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 539         cfq_clear_cfqq_on_rr(cfqq);
 540         list_move(&cfqq->cfq_list, &cfqd->empty_list);
 541
 542         BUG_ON(!cfqd->busy_queues);
 543         cfqd->busy_queues--;
 544 }
 545
 546 /*
 547  * rb tree support functions
 548  */
 549 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
 550 {
 551         struct cfq_queue *cfqq = crq->cfq_queue;
 552         struct cfq_data *cfqd = cfqq->cfqd;
 553         const int sync = cfq_crq_is_sync(crq);
 554
 555         BUG_ON(!cfqq->queued[sync]);
 556         cfqq->queued[sync]--;
 557
 558         cfq_update_next_crq(crq);
 559
 560         rb_erase(&crq->rb_node, &cfqq->sort_list);
 561
 562         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
 563                 cfq_del_cfqq_rr(cfqd, cfqq);
 564 }
 565
 566 static struct cfq_rq *
 567 __cfq_add_crq_rb(struct cfq_rq *crq)
 568 {
 569         struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
 570         struct rb_node *parent = NULL;
 571         struct cfq_rq *__crq;
 572
 573         while (*p) {
 574                 parent = *p;
 575                 __crq = rb_entry_crq(parent);
 576
 577                 if (crq->rb_key < __crq->rb_key)
 578                         p = &(*p)->rb_left;
 579                 else if (crq->rb_key > __crq->rb_key)
 580                         p = &(*p)->rb_right;
 581                 else
 582                         return __crq;
 583         }
 584
 585         rb_link_node(&crq->rb_node, parent, p);
 586         return NULL;
 587 }
 588
 589 static void cfq_add_crq_rb(struct cfq_rq *crq)
 590 {
 591         struct cfq_queue *cfqq = crq->cfq_queue;
 592         struct cfq_data *cfqd = cfqq->cfqd;
 593         struct request *rq = crq->request;
 594         struct cfq_rq *__alias;
 595
 596         crq->rb_key = rq_rb_key(rq);
 597         cfqq->queued[cfq_crq_is_sync(crq)]++;
 598
 599         /*
 600          * looks a little odd, but the first insert might return an alias.
 601          * if that happens, put the alias on the dispatch list
 602          */
 603         while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
 604                 cfq_dispatch_insert(cfqd->queue, __alias);
 605
 606         rb_insert_color(&crq->rb_node, &cfqq->sort_list);
 607
 608         if (!cfq_cfqq_on_rr(cfqq))
 609                 cfq_add_cfqq_rr(cfqd, cfqq);
 610
 611         /*
 612          * check if this request is a better next-serve candidate
 613          */
 614         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
 615 }
 616
 617 static inline void
 618 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
 619 {
 620         rb_erase(&crq->rb_node, &cfqq->sort_list);
 621         cfqq->queued[cfq_crq_is_sync(crq)]--;
 622
 623         cfq_add_crq_rb(crq);
 624 }
 625
 626 static struct request *
 627 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 628 {
 629         struct task_struct *tsk = current;
 630         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
 631         struct cfq_queue *cfqq;
 632         struct rb_node *n;
 633         sector_t sector;
 634
 635         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
 636         if (!cfqq)
 637                 goto out;
 638
 639         sector = bio->bi_sector + bio_sectors(bio);
 640         n = cfqq->sort_list.rb_node;
 641         while (n) {
 642                 struct cfq_rq *crq = rb_entry_crq(n);
 643
 644                 if (sector < crq->rb_key)
 645                         n = n->rb_left;
 646                 else if (sector > crq->rb_key)
 647                         n = n->rb_right;
 648                 else
 649                         return crq->request;
 650         }
 651
 652 out:
 653         return NULL;
 654 }
 655
 656 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 657 {
 658         struct cfq_data *cfqd = q->elevator->elevator_data;
 659
 660         cfqd->rq_in_driver++;
 661
 662         /*
 663          * If the depth is larger 1, it really could be queueing. But lets
 664          * make the mark a little higher - idling could still be good for
 665          * low queueing, and a low queueing number could also just indicate
 666          * a SCSI mid layer like behaviour where limit+1 is often seen.
 667          */
 668         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
 669                 cfqd->hw_tag = 1;
 670 }
 671
 672 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 673 {
 674         struct cfq_data *cfqd = q->elevator->elevator_data;
 675
 676         WARN_ON(!cfqd->rq_in_driver);
 677         cfqd->rq_in_driver--;
 678 }
 679
 680 static void cfq_remove_request(struct request *rq)
 681 {
 682         struct cfq_rq *crq = RQ_DATA(rq);
 683
 684         list_del_init(&rq->queuelist);
 685         cfq_del_crq_rb(crq);
 686         cfq_del_crq_hash(crq);
 687 }
 688
 689 static int
 690 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 691 {
 692         struct cfq_data *cfqd = q->elevator->elevator_data;
 693         struct request *__rq;
 694         int ret;
 695
 696         __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
 697         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 698                 ret = ELEVATOR_BACK_MERGE;
 699                 goto out;
 700         }
 701
 702         __rq = cfq_find_rq_fmerge(cfqd, bio);
 703         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 704                 ret = ELEVATOR_FRONT_MERGE;
 705                 goto out;
 706         }
 707
 708         return ELEVATOR_NO_MERGE;
 709 out:
 710         *req = __rq;
 711         return ret;
 712 }
 713
 714 static void cfq_merged_request(request_queue_t *q, struct request *req)
 715 {
 716         struct cfq_data *cfqd = q->elevator->elevator_data;
 717         struct cfq_rq *crq = RQ_DATA(req);
 718
 719         cfq_del_crq_hash(crq);
 720         cfq_add_crq_hash(cfqd, crq);
 721
 722         if (rq_rb_key(req) != crq->rb_key) {
 723                 struct cfq_queue *cfqq = crq->cfq_queue;
 724
 725                 cfq_update_next_crq(crq);
 726                 cfq_reposition_crq_rb(cfqq, crq);
 727         }
 728 }
 729
 730 static void
 731 cfq_merged_requests(request_queue_t *q, struct request *rq,
 732                     struct request *next)
 733 {
 734         cfq_merged_request(q, rq);
 735
 736         /*
 737          * reposition in fifo if next is older than rq
 738          */
 739         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 740             time_before(next->start_time, rq->start_time))
 741                 list_move(&rq->queuelist, &next->queuelist);
 742
 743         cfq_remove_request(next);
 744 }
 745
 746 static inline void
 747 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 748 {
 749         if (cfqq) {
 750                 /*
 751                  * stop potential idle class queues waiting service
 752                  */
 753                 del_timer(&cfqd->idle_class_timer);
 754
 755                 cfqq->slice_start = jiffies;
 756                 cfqq->slice_end = 0;
 757                 cfqq->slice_left = 0;
 758                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 759                 cfq_clear_cfqq_fifo_expire(cfqq);
 760         }
 761
 762         cfqd->active_queue = cfqq;
 763 }
 764
 765 /*
 766  * current cfqq expired its slice (or was too idle), select new one
 767  */
 768 static void
 769 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 770                     int preempted)
 771 {
 772         unsigned long now = jiffies;
 773
 774         if (cfq_cfqq_wait_request(cfqq))
 775                 del_timer(&cfqd->idle_slice_timer);
 776
 777         if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
 778                 cfqq->service_last = now;
 779                 cfq_schedule_dispatch(cfqd);
 780         }
 781
 782         cfq_clear_cfqq_must_dispatch(cfqq);
 783         cfq_clear_cfqq_wait_request(cfqq);
 784
 785         /*
 786          * store what was left of this slice, if the queue idled out
 787          * or was preempted
 788          */
 789         if (time_after(cfqq->slice_end, now))
 790                 cfqq->slice_left = cfqq->slice_end - now;
 791         else
 792                 cfqq->slice_left = 0;
 793
 794         if (cfq_cfqq_on_rr(cfqq))
 795                 cfq_resort_rr_list(cfqq, preempted);
 796
 797         if (cfqq == cfqd->active_queue)
 798                 cfqd->active_queue = NULL;
 799
 800         if (cfqd->active_cic) {
 801                 put_io_context(cfqd->active_cic->ioc);
 802                 cfqd->active_cic = NULL;
 803         }
 804
 805         cfqd->dispatch_slice = 0;
 806 }
 807
 808 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 809 {
 810         struct cfq_queue *cfqq = cfqd->active_queue;
 811
 812         if (cfqq)
 813                 __cfq_slice_expired(cfqd, cfqq, preempted);
 814 }
 815
 816 /*
 817  * 0
 818  * 0,1
 819  * 0,1,2
 820  * 0,1,2,3
 821  * 0,1,2,3,4
 822  * 0,1,2,3,4,5
 823  * 0,1,2,3,4,5,6
 824  * 0,1,2,3,4,5,6,7
 825  */
 826 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 827 {
 828         int prio, wrap;
 829
 830         prio = -1;
 831         wrap = 0;
 832         do {
 833                 int p;
 834
 835                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 836                         if (!list_empty(&cfqd->rr_list[p])) {
 837                                 prio = p;
 838                                 break;
 839                         }
 840                 }
 841
 842                 if (prio != -1)
 843                         break;
 844                 cfqd->cur_prio = 0;
 845                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 846                         cfqd->cur_end_prio = 0;
 847                         if (wrap)
 848                                 break;
 849                         wrap = 1;
 850                 }
 851         } while (1);
 852
 853         if (unlikely(prio == -1))
 854                 return -1;
 855
 856         BUG_ON(prio >= CFQ_PRIO_LISTS);
 857
 858         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 859
 860         cfqd->cur_prio = prio + 1;
 861         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 862                 cfqd->cur_end_prio = cfqd->cur_prio;
 863                 cfqd->cur_prio = 0;
 864         }
 865         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 866                 cfqd->cur_prio = 0;
 867                 cfqd->cur_end_prio = 0;
 868         }
 869
 870         return prio;
 871 }
 872
 873 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 874 {
 875         struct cfq_queue *cfqq = NULL;
 876
 877         /*
 878          * if current list is non-empty, grab first entry. if it is empty,
 879          * get next prio level and grab first entry then if any are spliced
 880          */
 881         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
 882                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 883
 884         /*
 885          * If no new queues are available, check if the busy list has some
 886          * before falling back to idle io.
 887          */
 888         if (!cfqq && !list_empty(&cfqd->busy_rr))
 889                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
 890
 891         /*
 892          * if we have idle queues and no rt or be queues had pending
 893          * requests, either allow immediate service if the grace period
 894          * has passed or arm the idle grace timer
 895          */
 896         if (!cfqq && !list_empty(&cfqd->idle_rr)) {
 897                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 898
 899                 if (time_after_eq(jiffies, end))
 900                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 901                 else
 902                         mod_timer(&cfqd->idle_class_timer, end);
 903         }
 904
 905         __cfq_set_active_queue(cfqd, cfqq);
 906         return cfqq;
 907 }
 908
 909 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 910
 911 {
 912         struct cfq_io_context *cic;
 913         unsigned long sl;
 914
 915         WARN_ON(!RB_EMPTY(&cfqq->sort_list));
 916         WARN_ON(cfqq != cfqd->active_queue);
 917
 918         /*
 919          * idle is disabled, either manually or by past process history
 920          */
 921         if (!cfqd->cfq_slice_idle)
 922                 return 0;
 923         if (!cfq_cfqq_idle_window(cfqq))
 924                 return 0;
 925         /*
 926          * task has exited, don't wait
 927          */
 928         cic = cfqd->active_cic;
 929         if (!cic || !cic->ioc->task)
 930                 return 0;
 931
 932         cfq_mark_cfqq_must_dispatch(cfqq);
 933         cfq_mark_cfqq_wait_request(cfqq);
 934
 935         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 936
 937         /*
 938          * we don't want to idle for seeks, but we do want to allow
 939          * fair distribution of slice time for a process doing back-to-back
 940          * seeks. so allow a little bit of time for him to submit a new rq
 941          */
 942         if (sample_valid(cic->seek_samples) && cic->seek_mean > 131072)
 943                 sl = 2;
 944
 945         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 946         return 1;
 947 }
 948
 949 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
 950 {
 951         struct cfq_data *cfqd = q->elevator->elevator_data;
 952         struct cfq_queue *cfqq = crq->cfq_queue;
 953
 954         cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
 955         cfq_remove_request(crq->request);
 956         cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
 957         elv_dispatch_sort(q, crq->request);
 958 }
 959
 960 /*
 961  * return expired entry, or NULL to just start from scratch in rbtree
 962  */
 963 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
 964 {
 965         struct cfq_data *cfqd = cfqq->cfqd;
 966         struct request *rq;
 967         struct cfq_rq *crq;
 968
 969         if (cfq_cfqq_fifo_expire(cfqq))
 970                 return NULL;
 971
 972         if (!list_empty(&cfqq->fifo)) {
 973                 int fifo = cfq_cfqq_class_sync(cfqq);
 974
 975                 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
 976                 rq = crq->request;
 977                 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
 978                         cfq_mark_cfqq_fifo_expire(cfqq);
 979                         return crq;
 980                 }
 981         }
 982
 983         return NULL;
 984 }
 985
 986 /*
 987  * Scale schedule slice based on io priority. Use the sync time slice only
 988  * if a queue is marked sync and has sync io queued. A sync queue with async
 989  * io only, should not get full sync slice length.
 990  */
 991 static inline int
 992 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 993 {
 994         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 995
 996         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 997
 998         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
 999 }
1000
1001 static inline void
1002 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1003 {
1004         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
1005 }
1006
1007 static inline int
1008 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1009 {
1010         const int base_rq = cfqd->cfq_slice_async_rq;
1011
1012         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1013
1014         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1015 }
1016
1017 /*
1018  * get next queue for service
1019  */
1020 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1021 {
1022         unsigned long now = jiffies;
1023         struct cfq_queue *cfqq;
1024
1025         cfqq = cfqd->active_queue;
1026         if (!cfqq)
1027                 goto new_queue;
1028
1029         /*
1030          * slice has expired
1031          */
1032         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
1033                 goto expire;
1034
1035         /*
1036          * if queue has requests, dispatch one. if not, check if
1037          * enough slice is left to wait for one
1038          */
1039         if (!RB_EMPTY(&cfqq->sort_list))
1040                 goto keep_queue;
1041         else if (cfq_cfqq_class_sync(cfqq) &&
1042                  time_before(now, cfqq->slice_end)) {
1043                 if (cfq_arm_slice_timer(cfqd, cfqq))
1044                         return NULL;
1045         }
1046
1047 expire:
1048         cfq_slice_expired(cfqd, 0);
1049 new_queue:
1050         cfqq = cfq_set_active_queue(cfqd);
1051 keep_queue:
1052         return cfqq;
1053 }
1054
1055 static int
1056 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1057                         int max_dispatch)
1058 {
1059         int dispatched = 0;
1060
1061         BUG_ON(RB_EMPTY(&cfqq->sort_list));
1062
1063         do {
1064                 struct cfq_rq *crq;
1065
1066                 /*
1067                  * follow expired path, else get first next available
1068                  */
1069                 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1070                         crq = cfqq->next_crq;
1071
1072                 /*
1073                  * finally, insert request into driver dispatch list
1074                  */
1075                 cfq_dispatch_insert(cfqd->queue, crq);
1076
1077                 cfqd->dispatch_slice++;
1078                 dispatched++;
1079
1080                 if (!cfqd->active_cic) {
1081                         atomic_inc(&crq->io_context->ioc->refcount);
1082                         cfqd->active_cic = crq->io_context;
1083                 }
1084
1085                 if (RB_EMPTY(&cfqq->sort_list))
1086                         break;
1087
1088         } while (dispatched < max_dispatch);
1089
1090         /*
1091          * if slice end isn't set yet, set it. if at least one request was
1092          * sync, use the sync time slice value
1093          */
1094         if (!cfqq->slice_end)
1095                 cfq_set_prio_slice(cfqd, cfqq);
1096
1097         /*
1098          * expire an async queue immediately if it has used up its slice. idle
1099          * queue always expire after 1 dispatch round.
1100          */
1101         if ((!cfq_cfqq_sync(cfqq) &&
1102             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1103             cfq_class_idle(cfqq))
1104                 cfq_slice_expired(cfqd, 0);
1105
1106         return dispatched;
1107 }
1108
1109 static int
1110 cfq_forced_dispatch_cfqqs(struct list_head *list)
1111 {
1112         int dispatched = 0;
1113         struct cfq_queue *cfqq, *next;
1114         struct cfq_rq *crq;
1115
1116         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1117                 while ((crq = cfqq->next_crq)) {
1118                         cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1119                         dispatched++;
1120                 }
1121                 BUG_ON(!list_empty(&cfqq->fifo));
1122         }
1123         return dispatched;
1124 }
1125
1126 static int
1127 cfq_forced_dispatch(struct cfq_data *cfqd)
1128 {
1129         int i, dispatched = 0;
1130
1131         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1132                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1133
1134         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1135         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1136         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1137
1138         cfq_slice_expired(cfqd, 0);
1139
1140         BUG_ON(cfqd->busy_queues);
1141
1142         return dispatched;
1143 }
1144
1145 static int
1146 cfq_dispatch_requests(request_queue_t *q, int force)
1147 {
1148         struct cfq_data *cfqd = q->elevator->elevator_data;
1149         struct cfq_queue *cfqq;
1150
1151         if (!cfqd->busy_queues)
1152                 return 0;
1153
1154         if (unlikely(force))
1155                 return cfq_forced_dispatch(cfqd);
1156
1157         cfqq = cfq_select_queue(cfqd);
1158         if (cfqq) {
1159                 int max_dispatch;
1160
1161                 cfq_clear_cfqq_must_dispatch(cfqq);
1162                 cfq_clear_cfqq_wait_request(cfqq);
1163                 del_timer(&cfqd->idle_slice_timer);
1164
1165                 max_dispatch = cfqd->cfq_quantum;
1166                 if (cfq_class_idle(cfqq))
1167                         max_dispatch = 1;
1168
1169                 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1170         }
1171
1172         return 0;
1173 }
1174
1175 /*
1176  * task holds one reference to the queue, dropped when task exits. each crq
1177  * in-flight on this queue also holds a reference, dropped when crq is freed.
1178  *
1179  * queue lock must be held here.
1180  */
1181 static void cfq_put_queue(struct cfq_queue *cfqq)
1182 {
1183         struct cfq_data *cfqd = cfqq->cfqd;
1184
1185         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1186
1187         if (!atomic_dec_and_test(&cfqq->ref))
1188                 return;
1189
1190         BUG_ON(rb_first(&cfqq->sort_list));
1191         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1192         BUG_ON(cfq_cfqq_on_rr(cfqq));
1193
1194         if (unlikely(cfqd->active_queue == cfqq))
1195                 __cfq_slice_expired(cfqd, cfqq, 0);
1196
1197         /*
1198          * it's on the empty list and still hashed
1199          */
1200         list_del(&cfqq->cfq_list);
1201         hlist_del(&cfqq->cfq_hash);
1202         kmem_cache_free(cfq_pool, cfqq);
1203 }
1204
1205 static inline struct cfq_queue *
1206 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1207                     const int hashval)
1208 {
1209         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1210         struct hlist_node *entry;
1211         struct cfq_queue *__cfqq;
1212
1213         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1214                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1215
1216                 if (__cfqq->key == key && (__p == prio || !prio))
1217                         return __cfqq;
1218         }
1219
1220         return NULL;
1221 }
1222
1223 static struct cfq_queue *
1224 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1225 {
1226         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1227 }
1228
1229 static void cfq_free_io_context(struct io_context *ioc)
1230 {
1231         struct cfq_io_context *__cic;
1232         struct rb_node *n;
1233         int freed = 0;
1234
1235         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1236                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1237                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1238                 kmem_cache_free(cfq_ioc_pool, __cic);
1239                 freed++;
1240         }
1241
1242         if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1243                 complete(ioc_gone);
1244 }
1245
1246 static void cfq_trim(struct io_context *ioc)
1247 {
1248         ioc->set_ioprio = NULL;
1249         cfq_free_io_context(ioc);
1250 }
1251
1252 /*
1253  * Called with interrupts disabled
1254  */
1255 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1256 {
1257         struct cfq_data *cfqd = cic->key;
1258         request_queue_t *q;
1259
1260         if (!cfqd)
1261                 return;
1262
1263         q = cfqd->queue;
1264
1265         WARN_ON(!irqs_disabled());
1266
1267         spin_lock(q->queue_lock);
1268
1269         if (cic->cfqq[ASYNC]) {
1270                 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
1271                         __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
1272                 cfq_put_queue(cic->cfqq[ASYNC]);
1273                 cic->cfqq[ASYNC] = NULL;
1274         }
1275
1276         if (cic->cfqq[SYNC]) {
1277                 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
1278                         __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
1279                 cfq_put_queue(cic->cfqq[SYNC]);
1280                 cic->cfqq[SYNC] = NULL;
1281         }
1282
1283         cic->key = NULL;
1284         list_del_init(&cic->queue_list);
1285         spin_unlock(q->queue_lock);
1286 }
1287
1288 static void cfq_exit_io_context(struct io_context *ioc)
1289 {
1290         struct cfq_io_context *__cic;
1291         unsigned long flags;
1292         struct rb_node *n;
1293
1294         /*
1295          * put the reference this task is holding to the various queues
1296          */
1297         spin_lock_irqsave(&cfq_exit_lock, flags);
1298
1299         n = rb_first(&ioc->cic_root);
1300         while (n != NULL) {
1301                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1302
1303                 cfq_exit_single_io_context(__cic);
1304                 n = rb_next(n);
1305         }
1306
1307         spin_unlock_irqrestore(&cfq_exit_lock, flags);
1308 }
1309
1310 static struct cfq_io_context *
1311 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1312 {
1313         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1314
1315         if (cic) {
1316                 memset(cic, 0, sizeof(*cic));
1317                 cic->last_end_request = jiffies;
1318                 INIT_LIST_HEAD(&cic->queue_list);
1319                 cic->dtor = cfq_free_io_context;
1320                 cic->exit = cfq_exit_io_context;
1321                 atomic_inc(&ioc_count);
1322         }
1323
1324         return cic;
1325 }
1326
1327 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1328 {
1329         struct task_struct *tsk = current;
1330         int ioprio_class;
1331
1332         if (!cfq_cfqq_prio_changed(cfqq))
1333                 return;
1334
1335         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1336         switch (ioprio_class) {
1337                 default:
1338                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1339                 case IOPRIO_CLASS_NONE:
1340                         /*
1341                          * no prio set, place us in the middle of the BE classes
1342                          */
1343                         cfqq->ioprio = task_nice_ioprio(tsk);
1344                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1345                         break;
1346                 case IOPRIO_CLASS_RT:
1347                         cfqq->ioprio = task_ioprio(tsk);
1348                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1349                         break;
1350                 case IOPRIO_CLASS_BE:
1351                         cfqq->ioprio = task_ioprio(tsk);
1352                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1353                         break;
1354                 case IOPRIO_CLASS_IDLE:
1355                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1356                         cfqq->ioprio = 7;
1357                         cfq_clear_cfqq_idle_window(cfqq);
1358                         break;
1359         }
1360
1361         /*
1362          * keep track of original prio settings in case we have to temporarily
1363          * elevate the priority of this queue
1364          */
1365         cfqq->org_ioprio = cfqq->ioprio;
1366         cfqq->org_ioprio_class = cfqq->ioprio_class;
1367
1368         if (cfq_cfqq_on_rr(cfqq))
1369                 cfq_resort_rr_list(cfqq, 0);
1370
1371         cfq_clear_cfqq_prio_changed(cfqq);
1372 }
1373
1374 static inline void changed_ioprio(struct cfq_io_context *cic)
1375 {
1376         struct cfq_data *cfqd = cic->key;
1377         struct cfq_queue *cfqq;
1378         if (cfqd) {
1379                 spin_lock(cfqd->queue->queue_lock);
1380                 cfqq = cic->cfqq[ASYNC];
1381                 if (cfqq) {
1382                         struct cfq_queue *new_cfqq;
1383                         new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC,
1384                                                 cic->ioc->task, GFP_ATOMIC);
1385                         if (new_cfqq) {
1386                                 cic->cfqq[ASYNC] = new_cfqq;
1387                                 cfq_put_queue(cfqq);
1388                         }
1389                 }
1390                 cfqq = cic->cfqq[SYNC];
1391                 if (cfqq)
1392                         cfq_mark_cfqq_prio_changed(cfqq);
1393
1394                 spin_unlock(cfqd->queue->queue_lock);
1395         }
1396 }
1397
1398 /*
1399  * callback from sys_ioprio_set, irqs are disabled
1400  */
1401 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1402 {
1403         struct cfq_io_context *cic;
1404         struct rb_node *n;
1405
1406         spin_lock(&cfq_exit_lock);
1407
1408         n = rb_first(&ioc->cic_root);
1409         while (n != NULL) {
1410                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1411
1412                 changed_ioprio(cic);
1413                 n = rb_next(n);
1414         }
1415
1416         spin_unlock(&cfq_exit_lock);
1417
1418         return 0;
1419 }
1420
1421 static struct cfq_queue *
1422 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1423               gfp_t gfp_mask)
1424 {
1425         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1426         struct cfq_queue *cfqq, *new_cfqq = NULL;
1427         unsigned short ioprio;
1428
1429 retry:
1430         ioprio = tsk->ioprio;
1431         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1432
1433         if (!cfqq) {
1434                 if (new_cfqq) {
1435                         cfqq = new_cfqq;
1436                         new_cfqq = NULL;
1437                 } else if (gfp_mask & __GFP_WAIT) {
1438                         spin_unlock_irq(cfqd->queue->queue_lock);
1439                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1440                         spin_lock_irq(cfqd->queue->queue_lock);
1441                         goto retry;
1442                 } else {
1443                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1444                         if (!cfqq)
1445                                 goto out;
1446                 }
1447
1448                 memset(cfqq, 0, sizeof(*cfqq));
1449
1450                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1451                 INIT_LIST_HEAD(&cfqq->cfq_list);
1452                 RB_CLEAR_ROOT(&cfqq->sort_list);
1453                 INIT_LIST_HEAD(&cfqq->fifo);
1454
1455                 cfqq->key = key;
1456                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1457                 atomic_set(&cfqq->ref, 0);
1458                 cfqq->cfqd = cfqd;
1459                 cfqq->service_last = 0;
1460                 /*
1461                  * set ->slice_left to allow preemption for a new process
1462                  */
1463                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1464                 if (!cfqd->hw_tag)
1465                         cfq_mark_cfqq_idle_window(cfqq);
1466                 cfq_mark_cfqq_prio_changed(cfqq);
1467                 cfq_init_prio_data(cfqq);
1468         }
1469
1470         if (new_cfqq)
1471                 kmem_cache_free(cfq_pool, new_cfqq);
1472
1473         atomic_inc(&cfqq->ref);
1474 out:
1475         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1476         return cfqq;
1477 }
1478
1479 static void
1480 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1481 {
1482         spin_lock(&cfq_exit_lock);
1483         rb_erase(&cic->rb_node, &ioc->cic_root);
1484         list_del_init(&cic->queue_list);
1485         spin_unlock(&cfq_exit_lock);
1486         kmem_cache_free(cfq_ioc_pool, cic);
1487         atomic_dec(&ioc_count);
1488 }
1489
1490 static struct cfq_io_context *
1491 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1492 {
1493         struct rb_node *n;
1494         struct cfq_io_context *cic;
1495         void *k, *key = cfqd;
1496
1497 restart:
1498         n = ioc->cic_root.rb_node;
1499         while (n) {
1500                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1501                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1502                 k = cic->key;
1503                 if (unlikely(!k)) {
1504                         cfq_drop_dead_cic(ioc, cic);
1505                         goto restart;
1506                 }
1507
1508                 if (key < k)
1509                         n = n->rb_left;
1510                 else if (key > k)
1511                         n = n->rb_right;
1512                 else
1513                         return cic;
1514         }
1515
1516         return NULL;
1517 }
1518
1519 static inline void
1520 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1521              struct cfq_io_context *cic)
1522 {
1523         struct rb_node **p;
1524         struct rb_node *parent;
1525         struct cfq_io_context *__cic;
1526         void *k;
1527
1528         cic->ioc = ioc;
1529         cic->key = cfqd;
1530
1531         ioc->set_ioprio = cfq_ioc_set_ioprio;
1532 restart:
1533         parent = NULL;
1534         p = &ioc->cic_root.rb_node;
1535         while (*p) {
1536                 parent = *p;
1537                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1538                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1539                 k = __cic->key;
1540                 if (unlikely(!k)) {
1541                         cfq_drop_dead_cic(ioc, cic);
1542                         goto restart;
1543                 }
1544
1545                 if (cic->key < k)
1546                         p = &(*p)->rb_left;
1547                 else if (cic->key > k)
1548                         p = &(*p)->rb_right;
1549                 else
1550                         BUG();
1551         }
1552
1553         spin_lock(&cfq_exit_lock);
1554         rb_link_node(&cic->rb_node, parent, p);
1555         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1556         list_add(&cic->queue_list, &cfqd->cic_list);
1557         spin_unlock(&cfq_exit_lock);
1558 }
1559
1560 /*
1561  * Setup general io context and cfq io context. There can be several cfq
1562  * io contexts per general io context, if this process is doing io to more
1563  * than one device managed by cfq.
1564  */
1565 static struct cfq_io_context *
1566 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1567 {
1568         struct io_context *ioc = NULL;
1569         struct cfq_io_context *cic;
1570
1571         might_sleep_if(gfp_mask & __GFP_WAIT);
1572
1573         ioc = get_io_context(gfp_mask);
1574         if (!ioc)
1575                 return NULL;
1576
1577         cic = cfq_cic_rb_lookup(cfqd, ioc);
1578         if (cic)
1579                 goto out;
1580
1581         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1582         if (cic == NULL)
1583                 goto err;
1584
1585         cfq_cic_link(cfqd, ioc, cic);
1586 out:
1587         return cic;
1588 err:
1589         put_io_context(ioc);
1590         return NULL;
1591 }
1592
1593 static void
1594 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1595 {
1596         unsigned long elapsed, ttime;
1597
1598         /*
1599          * if this context already has stuff queued, thinktime is from
1600          * last queue not last end
1601          */
1602 #if 0
1603         if (time_after(cic->last_end_request, cic->last_queue))
1604                 elapsed = jiffies - cic->last_end_request;
1605         else
1606                 elapsed = jiffies - cic->last_queue;
1607 #else
1608                 elapsed = jiffies - cic->last_end_request;
1609 #endif
1610
1611         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1612
1613         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1614         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1615         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1616 }
1617
1618 static void
1619 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1620                        struct cfq_rq *crq)
1621 {
1622         sector_t sdist;
1623         u64 total;
1624
1625         if (cic->last_request_pos < crq->request->sector)
1626                 sdist = crq->request->sector - cic->last_request_pos;
1627         else
1628                 sdist = cic->last_request_pos - crq->request->sector;
1629
1630         /*
1631          * Don't allow the seek distance to get too large from the
1632          * odd fragment, pagein, etc
1633          */
1634         if (cic->seek_samples <= 60) /* second&third seek */
1635                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1636         else
1637                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1638
1639         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1640         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1641         total = cic->seek_total + (cic->seek_samples/2);
1642         do_div(total, cic->seek_samples);
1643         cic->seek_mean = (sector_t)total;
1644 }
1645
1646 /*
1647  * Disable idle window if the process thinks too long or seeks so much that
1648  * it doesn't matter
1649  */
1650 static void
1651 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1652                        struct cfq_io_context *cic)
1653 {
1654         int enable_idle = cfq_cfqq_idle_window(cfqq);
1655
1656         if (!cic->ioc->task || !cfqd->cfq_slice_idle || cfqd->hw_tag)
1657                 enable_idle = 0;
1658         else if (sample_valid(cic->ttime_samples)) {
1659                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1660                         enable_idle = 0;
1661                 else
1662                         enable_idle = 1;
1663         }
1664
1665         if (enable_idle)
1666                 cfq_mark_cfqq_idle_window(cfqq);
1667         else
1668                 cfq_clear_cfqq_idle_window(cfqq);
1669 }
1670
1671
1672 /*
1673  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1674  * no or if we aren't sure, a 1 will cause a preempt.
1675  */
1676 static int
1677 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1678                    struct cfq_rq *crq)
1679 {
1680         struct cfq_queue *cfqq = cfqd->active_queue;
1681
1682         if (cfq_class_idle(new_cfqq))
1683                 return 0;
1684
1685         if (!cfqq)
1686                 return 1;
1687
1688         if (cfq_class_idle(cfqq))
1689                 return 1;
1690         if (!cfq_cfqq_wait_request(new_cfqq))
1691                 return 0;
1692         /*
1693          * if it doesn't have slice left, forget it
1694          */
1695         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1696                 return 0;
1697         if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1698                 return 1;
1699
1700         return 0;
1701 }
1702
1703 /*
1704  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1705  * let it have half of its nominal slice.
1706  */
1707 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1708 {
1709         struct cfq_queue *__cfqq, *next;
1710
1711         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1712                 cfq_resort_rr_list(__cfqq, 1);
1713
1714         if (!cfqq->slice_left)
1715                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1716
1717         cfqq->slice_end = cfqq->slice_left + jiffies;
1718         __cfq_slice_expired(cfqd, cfqq, 1);
1719         __cfq_set_active_queue(cfqd, cfqq);
1720 }
1721
1722 /*
1723  * should really be a ll_rw_blk.c helper
1724  */
1725 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1726 {
1727         request_queue_t *q = cfqd->queue;
1728
1729         if (!blk_queue_plugged(q))
1730                 q->request_fn(q);
1731         else
1732                 __generic_unplug_device(q);
1733 }
1734
1735 /*
1736  * Called when a new fs request (crq) is added (to cfqq). Check if there's
1737  * something we should do about it
1738  */
1739 static void
1740 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1741                  struct cfq_rq *crq)
1742 {
1743         struct cfq_io_context *cic;
1744
1745         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1746
1747         cic = crq->io_context;
1748
1749         /*
1750          * we never wait for an async request and we don't allow preemption
1751          * of an async request. so just return early
1752          */
1753         if (!cfq_crq_is_sync(crq)) {
1754                 /*
1755                  * sync process issued an async request, if it's waiting
1756                  * then expire it and kick rq handling.
1757                  */
1758                 if (cic == cfqd->active_cic &&
1759                     del_timer(&cfqd->idle_slice_timer)) {
1760                         cfq_slice_expired(cfqd, 0);
1761                         cfq_start_queueing(cfqd, cfqq);
1762                 }
1763                 return;
1764         }
1765
1766         cfq_update_io_thinktime(cfqd, cic);
1767         cfq_update_io_seektime(cfqd, cic, crq);
1768         cfq_update_idle_window(cfqd, cfqq, cic);
1769
1770         cic->last_queue = jiffies;
1771         cic->last_request_pos = crq->request->sector + crq->request->nr_sectors;
1772
1773         if (cfqq == cfqd->active_queue) {
1774                 /*
1775                  * if we are waiting for a request for this queue, let it rip
1776                  * immediately and flag that we must not expire this queue
1777                  * just now
1778                  */
1779                 if (cfq_cfqq_wait_request(cfqq)) {
1780                         cfq_mark_cfqq_must_dispatch(cfqq);
1781                         del_timer(&cfqd->idle_slice_timer);
1782                         cfq_start_queueing(cfqd, cfqq);
1783                 }
1784         } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1785                 /*
1786                  * not the active queue - expire current slice if it is
1787                  * idle and has expired it's mean thinktime or this new queue
1788                  * has some old slice time left and is of higher priority
1789                  */
1790                 cfq_preempt_queue(cfqd, cfqq);
1791                 cfq_mark_cfqq_must_dispatch(cfqq);
1792                 cfq_start_queueing(cfqd, cfqq);
1793         }
1794 }
1795
1796 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1797 {
1798         struct cfq_data *cfqd = q->elevator->elevator_data;
1799         struct cfq_rq *crq = RQ_DATA(rq);
1800         struct cfq_queue *cfqq = crq->cfq_queue;
1801
1802         cfq_init_prio_data(cfqq);
1803
1804         cfq_add_crq_rb(crq);
1805
1806         list_add_tail(&rq->queuelist, &cfqq->fifo);
1807
1808         if (rq_mergeable(rq))
1809                 cfq_add_crq_hash(cfqd, crq);
1810
1811         cfq_crq_enqueued(cfqd, cfqq, crq);
1812 }
1813
1814 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1815 {
1816         struct cfq_rq *crq = RQ_DATA(rq);
1817         struct cfq_queue *cfqq = crq->cfq_queue;
1818         struct cfq_data *cfqd = cfqq->cfqd;
1819         const int sync = cfq_crq_is_sync(crq);
1820         unsigned long now;
1821
1822         now = jiffies;
1823
1824         WARN_ON(!cfqd->rq_in_driver);
1825         WARN_ON(!cfqq->on_dispatch[sync]);
1826         cfqd->rq_in_driver--;
1827         cfqq->on_dispatch[sync]--;
1828
1829         if (!cfq_class_idle(cfqq))
1830                 cfqd->last_end_request = now;
1831
1832         if (!cfq_cfqq_dispatched(cfqq)) {
1833                 if (cfq_cfqq_on_rr(cfqq)) {
1834                         cfqq->service_last = now;
1835                         cfq_resort_rr_list(cfqq, 0);
1836                 }
1837                 cfq_schedule_dispatch(cfqd);
1838         }
1839
1840         if (cfq_crq_is_sync(crq))
1841                 crq->io_context->last_end_request = now;
1842 }
1843
1844 static struct request *
1845 cfq_former_request(request_queue_t *q, struct request *rq)
1846 {
1847         struct cfq_rq *crq = RQ_DATA(rq);
1848         struct rb_node *rbprev = rb_prev(&crq->rb_node);
1849
1850         if (rbprev)
1851                 return rb_entry_crq(rbprev)->request;
1852
1853         return NULL;
1854 }
1855
1856 static struct request *
1857 cfq_latter_request(request_queue_t *q, struct request *rq)
1858 {
1859         struct cfq_rq *crq = RQ_DATA(rq);
1860         struct rb_node *rbnext = rb_next(&crq->rb_node);
1861
1862         if (rbnext)
1863                 return rb_entry_crq(rbnext)->request;
1864
1865         return NULL;
1866 }
1867
1868 /*
1869  * we temporarily boost lower priority queues if they are holding fs exclusive
1870  * resources. they are boosted to normal prio (CLASS_BE/4)
1871  */
1872 static void cfq_prio_boost(struct cfq_queue *cfqq)
1873 {
1874         const int ioprio_class = cfqq->ioprio_class;
1875         const int ioprio = cfqq->ioprio;
1876
1877         if (has_fs_excl()) {
1878                 /*
1879                  * boost idle prio on transactions that would lock out other
1880                  * users of the filesystem
1881                  */
1882                 if (cfq_class_idle(cfqq))
1883                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1884                 if (cfqq->ioprio > IOPRIO_NORM)
1885                         cfqq->ioprio = IOPRIO_NORM;
1886         } else {
1887                 /*
1888                  * check if we need to unboost the queue
1889                  */
1890                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1891                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1892                 if (cfqq->ioprio != cfqq->org_ioprio)
1893                         cfqq->ioprio = cfqq->org_ioprio;
1894         }
1895
1896         /*
1897          * refile between round-robin lists if we moved the priority class
1898          */
1899         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1900             cfq_cfqq_on_rr(cfqq))
1901                 cfq_resort_rr_list(cfqq, 0);
1902 }
1903
1904 static inline int
1905 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1906                 struct task_struct *task, int rw)
1907 {
1908         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1909             !cfq_cfqq_must_alloc_slice(cfqq)) {
1910                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1911                 return ELV_MQUEUE_MUST;
1912         }
1913
1914         return ELV_MQUEUE_MAY;
1915 }
1916
1917 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1918 {
1919         struct cfq_data *cfqd = q->elevator->elevator_data;
1920         struct task_struct *tsk = current;
1921         struct cfq_queue *cfqq;
1922
1923         /*
1924          * don't force setup of a queue from here, as a call to may_queue
1925          * does not necessarily imply that a request actually will be queued.
1926          * so just lookup a possibly existing queue, or return 'may queue'
1927          * if that fails
1928          */
1929         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1930         if (cfqq) {
1931                 cfq_init_prio_data(cfqq);
1932                 cfq_prio_boost(cfqq);
1933
1934                 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1935         }
1936
1937         return ELV_MQUEUE_MAY;
1938 }
1939
1940 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1941 {
1942         struct cfq_data *cfqd = q->elevator->elevator_data;
1943
1944         if (unlikely(cfqd->rq_starved)) {
1945                 struct request_list *rl = &q->rq;
1946
1947                 smp_mb();
1948                 if (waitqueue_active(&rl->wait[READ]))
1949                         wake_up(&rl->wait[READ]);
1950                 if (waitqueue_active(&rl->wait[WRITE]))
1951                         wake_up(&rl->wait[WRITE]);
1952         }
1953 }
1954
1955 /*
1956  * queue lock held here
1957  */
1958 static void cfq_put_request(request_queue_t *q, struct request *rq)
1959 {
1960         struct cfq_data *cfqd = q->elevator->elevator_data;
1961         struct cfq_rq *crq = RQ_DATA(rq);
1962
1963         if (crq) {
1964                 struct cfq_queue *cfqq = crq->cfq_queue;
1965                 const int rw = rq_data_dir(rq);
1966
1967                 BUG_ON(!cfqq->allocated[rw]);
1968                 cfqq->allocated[rw]--;
1969
1970                 put_io_context(crq->io_context->ioc);
1971
1972                 mempool_free(crq, cfqd->crq_pool);
1973                 rq->elevator_private = NULL;
1974
1975                 cfq_check_waiters(q, cfqq);
1976                 cfq_put_queue(cfqq);
1977         }
1978 }
1979
1980 /*
1981  * Allocate cfq data structures associated with this request.
1982  */
1983 static int
1984 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
1985                 gfp_t gfp_mask)
1986 {
1987         struct cfq_data *cfqd = q->elevator->elevator_data;
1988         struct task_struct *tsk = current;
1989         struct cfq_io_context *cic;
1990         const int rw = rq_data_dir(rq);
1991         pid_t key = cfq_queue_pid(tsk, rw);
1992         struct cfq_queue *cfqq;
1993         struct cfq_rq *crq;
1994         unsigned long flags;
1995         int is_sync = key != CFQ_KEY_ASYNC;
1996
1997         might_sleep_if(gfp_mask & __GFP_WAIT);
1998
1999         cic = cfq_get_io_context(cfqd, gfp_mask);
2000
2001         spin_lock_irqsave(q->queue_lock, flags);
2002
2003         if (!cic)
2004                 goto queue_fail;
2005
2006         if (!cic->cfqq[is_sync]) {
2007                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
2008                 if (!cfqq)
2009                         goto queue_fail;
2010
2011                 cic->cfqq[is_sync] = cfqq;
2012         } else
2013                 cfqq = cic->cfqq[is_sync];
2014
2015         cfqq->allocated[rw]++;
2016         cfq_clear_cfqq_must_alloc(cfqq);
2017         cfqd->rq_starved = 0;
2018         atomic_inc(&cfqq->ref);
2019         spin_unlock_irqrestore(q->queue_lock, flags);
2020
2021         crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
2022         if (crq) {
2023                 RB_CLEAR(&crq->rb_node);
2024                 crq->rb_key = 0;
2025                 crq->request = rq;
2026                 INIT_HLIST_NODE(&crq->hash);
2027                 crq->cfq_queue = cfqq;
2028                 crq->io_context = cic;
2029
2030                 if (is_sync)
2031                         cfq_mark_crq_is_sync(crq);
2032                 else
2033                         cfq_clear_crq_is_sync(crq);
2034
2035                 rq->elevator_private = crq;
2036                 return 0;
2037         }
2038
2039         spin_lock_irqsave(q->queue_lock, flags);
2040         cfqq->allocated[rw]--;
2041         if (!(cfqq->allocated[0] + cfqq->allocated[1]))
2042                 cfq_mark_cfqq_must_alloc(cfqq);
2043         cfq_put_queue(cfqq);
2044 queue_fail:
2045         if (cic)
2046                 put_io_context(cic->ioc);
2047         /*
2048          * mark us rq allocation starved. we need to kickstart the process
2049          * ourselves if there are no pending requests that can do it for us.
2050          * that would be an extremely rare OOM situation
2051          */
2052         cfqd->rq_starved = 1;
2053         cfq_schedule_dispatch(cfqd);
2054         spin_unlock_irqrestore(q->queue_lock, flags);
2055         return 1;
2056 }
2057
2058 static void cfq_kick_queue(void *data)
2059 {
2060         request_queue_t *q = data;
2061         struct cfq_data *cfqd = q->elevator->elevator_data;
2062         unsigned long flags;
2063
2064         spin_lock_irqsave(q->queue_lock, flags);
2065
2066         if (cfqd->rq_starved) {
2067                 struct request_list *rl = &q->rq;
2068
2069                 /*
2070                  * we aren't guaranteed to get a request after this, but we
2071                  * have to be opportunistic
2072                  */
2073                 smp_mb();
2074                 if (waitqueue_active(&rl->wait[READ]))
2075                         wake_up(&rl->wait[READ]);
2076                 if (waitqueue_active(&rl->wait[WRITE]))
2077                         wake_up(&rl->wait[WRITE]);
2078         }
2079
2080         blk_remove_plug(q);
2081         q->request_fn(q);
2082         spin_unlock_irqrestore(q->queue_lock, flags);
2083 }
2084
2085 /*
2086  * Timer running if the active_queue is currently idling inside its time slice
2087  */
2088 static void cfq_idle_slice_timer(unsigned long data)
2089 {
2090         struct cfq_data *cfqd = (struct cfq_data *) data;
2091         struct cfq_queue *cfqq;
2092         unsigned long flags;
2093
2094         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2095
2096         if ((cfqq = cfqd->active_queue) != NULL) {
2097                 unsigned long now = jiffies;
2098
2099                 /*
2100                  * expired
2101                  */
2102                 if (time_after(now, cfqq->slice_end))
2103                         goto expire;
2104
2105                 /*
2106                  * only expire and reinvoke request handler, if there are
2107                  * other queues with pending requests
2108                  */
2109                 if (!cfqd->busy_queues) {
2110                         cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2111                         add_timer(&cfqd->idle_slice_timer);
2112                         goto out_cont;
2113                 }
2114
2115                 /*
2116                  * not expired and it has a request pending, let it dispatch
2117                  */
2118                 if (!RB_EMPTY(&cfqq->sort_list)) {
2119                         cfq_mark_cfqq_must_dispatch(cfqq);
2120                         goto out_kick;
2121                 }
2122         }
2123 expire:
2124         cfq_slice_expired(cfqd, 0);
2125 out_kick:
2126         cfq_schedule_dispatch(cfqd);
2127 out_cont:
2128         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2129 }
2130
2131 /*
2132  * Timer running if an idle class queue is waiting for service
2133  */
2134 static void cfq_idle_class_timer(unsigned long data)
2135 {
2136         struct cfq_data *cfqd = (struct cfq_data *) data;
2137         unsigned long flags, end;
2138
2139         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2140
2141         /*
2142          * race with a non-idle queue, reset timer
2143          */
2144         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2145         if (!time_after_eq(jiffies, end))
2146                 mod_timer(&cfqd->idle_class_timer, end);
2147         else
2148                 cfq_schedule_dispatch(cfqd);
2149
2150         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2151 }
2152
2153 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2154 {
2155         del_timer_sync(&cfqd->idle_slice_timer);
2156         del_timer_sync(&cfqd->idle_class_timer);
2157         blk_sync_queue(cfqd->queue);
2158 }
2159
2160 static void cfq_exit_queue(elevator_t *e)
2161 {
2162         struct cfq_data *cfqd = e->elevator_data;
2163         request_queue_t *q = cfqd->queue;
2164
2165         cfq_shutdown_timer_wq(cfqd);
2166
2167         spin_lock(&cfq_exit_lock);
2168         spin_lock_irq(q->queue_lock);
2169
2170         if (cfqd->active_queue)
2171                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2172
2173         while (!list_empty(&cfqd->cic_list)) {
2174                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2175                                                         struct cfq_io_context,
2176                                                         queue_list);
2177                 if (cic->cfqq[ASYNC]) {
2178                         cfq_put_queue(cic->cfqq[ASYNC]);
2179                         cic->cfqq[ASYNC] = NULL;
2180                 }
2181                 if (cic->cfqq[SYNC]) {
2182                         cfq_put_queue(cic->cfqq[SYNC]);
2183                         cic->cfqq[SYNC] = NULL;
2184                 }
2185                 cic->key = NULL;
2186                 list_del_init(&cic->queue_list);
2187         }
2188
2189         spin_unlock_irq(q->queue_lock);
2190         spin_unlock(&cfq_exit_lock);
2191
2192         cfq_shutdown_timer_wq(cfqd);
2193
2194         mempool_destroy(cfqd->crq_pool);
2195         kfree(cfqd->crq_hash);
2196         kfree(cfqd->cfq_hash);
2197         kfree(cfqd);
2198 }
2199
2200 static void *cfq_init_queue(request_queue_t *q, elevator_t *e)
2201 {
2202         struct cfq_data *cfqd;
2203         int i;
2204
2205         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2206         if (!cfqd)
2207                 return NULL;
2208
2209         memset(cfqd, 0, sizeof(*cfqd));
2210
2211         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2212                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2213
2214         INIT_LIST_HEAD(&cfqd->busy_rr);
2215         INIT_LIST_HEAD(&cfqd->cur_rr);
2216         INIT_LIST_HEAD(&cfqd->idle_rr);
2217         INIT_LIST_HEAD(&cfqd->empty_list);
2218         INIT_LIST_HEAD(&cfqd->cic_list);
2219
2220         cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2221         if (!cfqd->crq_hash)
2222                 goto out_crqhash;
2223
2224         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2225         if (!cfqd->cfq_hash)
2226                 goto out_cfqhash;
2227
2228         cfqd->crq_pool = mempool_create_slab_pool(BLKDEV_MIN_RQ, crq_pool);
2229         if (!cfqd->crq_pool)
2230                 goto out_crqpool;
2231
2232         for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2233                 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2234         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2235                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2236
2237         cfqd->queue = q;
2238
2239         init_timer(&cfqd->idle_slice_timer);
2240         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2241         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2242
2243         init_timer(&cfqd->idle_class_timer);
2244         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2245         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2246
2247         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2248
2249         cfqd->cfq_queued = cfq_queued;
2250         cfqd->cfq_quantum = cfq_quantum;
2251         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2252         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2253         cfqd->cfq_back_max = cfq_back_max;
2254         cfqd->cfq_back_penalty = cfq_back_penalty;
2255         cfqd->cfq_slice[0] = cfq_slice_async;
2256         cfqd->cfq_slice[1] = cfq_slice_sync;
2257         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2258         cfqd->cfq_slice_idle = cfq_slice_idle;
2259
2260         return cfqd;
2261 out_crqpool:
2262         kfree(cfqd->cfq_hash);
2263 out_cfqhash:
2264         kfree(cfqd->crq_hash);
2265 out_crqhash:
2266         kfree(cfqd);
2267         return NULL;
2268 }
2269
2270 static void cfq_slab_kill(void)
2271 {
2272         if (crq_pool)
2273                 kmem_cache_destroy(crq_pool);
2274         if (cfq_pool)
2275                 kmem_cache_destroy(cfq_pool);
2276         if (cfq_ioc_pool)
2277                 kmem_cache_destroy(cfq_ioc_pool);
2278 }
2279
2280 static int __init cfq_slab_setup(void)
2281 {
2282         crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2283                                         NULL, NULL);
2284         if (!crq_pool)
2285                 goto fail;
2286
2287         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2288                                         NULL, NULL);
2289         if (!cfq_pool)
2290                 goto fail;
2291
2292         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2293                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2294         if (!cfq_ioc_pool)
2295                 goto fail;
2296
2297         return 0;
2298 fail:
2299         cfq_slab_kill();
2300         return -ENOMEM;
2301 }
2302
2303 /*
2304  * sysfs parts below -->
2305  */
2306
2307 static ssize_t
2308 cfq_var_show(unsigned int var, char *page)
2309 {
2310         return sprintf(page, "%d\n", var);
2311 }
2312
2313 static ssize_t
2314 cfq_var_store(unsigned int *var, const char *page, size_t count)
2315 {
2316         char *p = (char *) page;
2317
2318         *var = simple_strtoul(p, &p, 10);
2319         return count;
2320 }
2321
2322 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2323 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2324 {                                                                       \
2325         struct cfq_data *cfqd = e->elevator_data;                       \
2326         unsigned int __data = __VAR;                                    \
2327         if (__CONV)                                                     \
2328                 __data = jiffies_to_msecs(__data);                      \
2329         return cfq_var_show(__data, (page));                            \
2330 }
2331 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2332 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2333 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2334 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2335 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2336 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2337 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2338 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2339 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2340 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2341 #undef SHOW_FUNCTION
2342
2343 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2344 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2345 {                                                                       \
2346         struct cfq_data *cfqd = e->elevator_data;                       \
2347         unsigned int __data;                                            \
2348         int ret = cfq_var_store(&__data, (page), count);                \
2349         if (__data < (MIN))                                             \
2350                 __data = (MIN);                                         \
2351         else if (__data > (MAX))                                        \
2352                 __data = (MAX);                                         \
2353         if (__CONV)                                                     \
2354                 *(__PTR) = msecs_to_jiffies(__data);                    \
2355         else                                                            \
2356                 *(__PTR) = __data;                                      \
2357         return ret;                                                     \
2358 }
2359 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2360 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2361 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2362 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2363 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2364 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2365 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2366 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2367 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2368 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2369 #undef STORE_FUNCTION
2370
2371 #define CFQ_ATTR(name) \
2372         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2373
2374 static struct elv_fs_entry cfq_attrs[] = {
2375         CFQ_ATTR(quantum),
2376         CFQ_ATTR(queued),
2377         CFQ_ATTR(fifo_expire_sync),
2378         CFQ_ATTR(fifo_expire_async),
2379         CFQ_ATTR(back_seek_max),
2380         CFQ_ATTR(back_seek_penalty),
2381         CFQ_ATTR(slice_sync),
2382         CFQ_ATTR(slice_async),
2383         CFQ_ATTR(slice_async_rq),
2384         CFQ_ATTR(slice_idle),
2385         __ATTR_NULL
2386 };
2387
2388 static struct elevator_type iosched_cfq = {
2389         .ops = {
2390                 .elevator_merge_fn =            cfq_merge,
2391                 .elevator_merged_fn =           cfq_merged_request,
2392                 .elevator_merge_req_fn =        cfq_merged_requests,
2393                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2394                 .elevator_add_req_fn =          cfq_insert_request,
2395                 .elevator_activate_req_fn =     cfq_activate_request,
2396                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2397                 .elevator_queue_empty_fn =      cfq_queue_empty,
2398                 .elevator_completed_req_fn =    cfq_completed_request,
2399                 .elevator_former_req_fn =       cfq_former_request,
2400                 .elevator_latter_req_fn =       cfq_latter_request,
2401                 .elevator_set_req_fn =          cfq_set_request,
2402                 .elevator_put_req_fn =          cfq_put_request,
2403                 .elevator_may_queue_fn =        cfq_may_queue,
2404                 .elevator_init_fn =             cfq_init_queue,
2405                 .elevator_exit_fn =             cfq_exit_queue,
2406                 .trim =                         cfq_trim,
2407         },
2408         .elevator_attrs =       cfq_attrs,
2409         .elevator_name =        "cfq",
2410         .elevator_owner =       THIS_MODULE,
2411 };
2412
2413 static int __init cfq_init(void)
2414 {
2415         int ret;
2416
2417         /*
2418          * could be 0 on HZ < 1000 setups
2419          */
2420         if (!cfq_slice_async)
2421                 cfq_slice_async = 1;
2422         if (!cfq_slice_idle)
2423                 cfq_slice_idle = 1;
2424
2425         if (cfq_slab_setup())
2426                 return -ENOMEM;
2427
2428         ret = elv_register(&iosched_cfq);
2429         if (ret)
2430                 cfq_slab_kill();
2431
2432         return ret;
2433 }
2434
2435 static void __exit cfq_exit(void)
2436 {
2437         DECLARE_COMPLETION(all_gone);
2438         elv_unregister(&iosched_cfq);
2439         ioc_gone = &all_gone;
2440         /* ioc_gone's update must be visible before reading ioc_count */
2441         smp_wmb();
2442         if (atomic_read(&ioc_count))
2443                 wait_for_completion(ioc_gone);
2444         synchronize_rcu();
2445         cfq_slab_kill();
2446 }
2447
2448 module_init(cfq_init);
2449 module_exit(cfq_exit);
2450
2451 MODULE_AUTHOR("Jens Axboe");
2452 MODULE_LICENSE("GPL");
2453 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");