Pileus Git - ~andy/linux/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 / 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_RWLOCK(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_NONE                 (2)
  64 #define RB_EMPTY(node)          ((node)->rb_node == NULL)
  65 #define RB_CLEAR_COLOR(node)    (node)->rb_color = RB_NONE
  66 #define RB_CLEAR(node)          do {    \
  67         (node)->rb_parent = NULL;       \
  68         RB_CLEAR_COLOR((node));         \
  69         (node)->rb_right = NULL;        \
  70         (node)->rb_left = NULL;         \
  71 } while (0)
  72 #define RB_CLEAR_ROOT(root)     ((root)->rb_node = NULL)
  73 #define rb_entry_crq(node)      rb_entry((node), struct cfq_rq, rb_node)
  74 #define rq_rb_key(rq)           (rq)->sector
  75
  76 static kmem_cache_t *crq_pool;
  77 static kmem_cache_t *cfq_pool;
  78 static kmem_cache_t *cfq_ioc_pool;
  79
  80 static atomic_t ioc_count = ATOMIC_INIT(0);
  81 static struct completion *ioc_gone;
  82
  83 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  84 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  85 #define cfq_class_be(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
  86 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  87
  88 #define ASYNC                   (0)
  89 #define SYNC                    (1)
  90
  91 #define cfq_cfqq_dispatched(cfqq)       \
  92         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
  93
  94 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
  95
  96 #define cfq_cfqq_sync(cfqq)             \
  97         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
  98
  99 #define sample_valid(samples)   ((samples) > 80)
 100
 101 /*
 102  * Per block device queue structure
 103  */
 104 struct cfq_data {
 105         request_queue_t *queue;
 106
 107         /*
 108          * rr list of queues with requests and the count of them
 109          */
 110         struct list_head rr_list[CFQ_PRIO_LISTS];
 111         struct list_head busy_rr;
 112         struct list_head cur_rr;
 113         struct list_head idle_rr;
 114         unsigned int busy_queues;
 115
 116         /*
 117          * non-ordered list of empty cfqq's
 118          */
 119         struct list_head empty_list;
 120
 121         /*
 122          * cfqq lookup hash
 123          */
 124         struct hlist_head *cfq_hash;
 125
 126         /*
 127          * global crq hash for all queues
 128          */
 129         struct hlist_head *crq_hash;
 130
 131         unsigned int max_queued;
 132
 133         mempool_t *crq_pool;
 134
 135         int rq_in_driver;
 136
 137         /*
 138          * schedule slice state info
 139          */
 140         /*
 141          * idle window management
 142          */
 143         struct timer_list idle_slice_timer;
 144         struct work_struct unplug_work;
 145
 146         struct cfq_queue *active_queue;
 147         struct cfq_io_context *active_cic;
 148         int cur_prio, cur_end_prio;
 149         unsigned int dispatch_slice;
 150
 151         struct timer_list idle_class_timer;
 152
 153         sector_t last_sector;
 154         unsigned long last_end_request;
 155
 156         unsigned int rq_starved;
 157
 158         /*
 159          * tunables, see top of file
 160          */
 161         unsigned int cfq_quantum;
 162         unsigned int cfq_queued;
 163         unsigned int cfq_fifo_expire[2];
 164         unsigned int cfq_back_penalty;
 165         unsigned int cfq_back_max;
 166         unsigned int cfq_slice[2];
 167         unsigned int cfq_slice_async_rq;
 168         unsigned int cfq_slice_idle;
 169
 170         struct list_head cic_list;
 171 };
 172
 173 /*
 174  * Per process-grouping structure
 175  */
 176 struct cfq_queue {
 177         /* reference count */
 178         atomic_t ref;
 179         /* parent cfq_data */
 180         struct cfq_data *cfqd;
 181         /* cfqq lookup hash */
 182         struct hlist_node cfq_hash;
 183         /* hash key */
 184         unsigned int key;
 185         /* on either rr or empty list of cfqd */
 186         struct list_head cfq_list;
 187         /* sorted list of pending requests */
 188         struct rb_root sort_list;
 189         /* if fifo isn't expired, next request to serve */
 190         struct cfq_rq *next_crq;
 191         /* requests queued in sort_list */
 192         int queued[2];
 193         /* currently allocated requests */
 194         int allocated[2];
 195         /* fifo list of requests in sort_list */
 196         struct list_head fifo;
 197
 198         unsigned long slice_start;
 199         unsigned long slice_end;
 200         unsigned long slice_left;
 201         unsigned long service_last;
 202
 203         /* number of requests that are on the dispatch list */
 204         int on_dispatch[2];
 205
 206         /* io prio of this group */
 207         unsigned short ioprio, org_ioprio;
 208         unsigned short ioprio_class, org_ioprio_class;
 209
 210         /* various state flags, see below */
 211         unsigned int flags;
 212 };
 213
 214 struct cfq_rq {
 215         struct rb_node rb_node;
 216         sector_t rb_key;
 217         struct request *request;
 218         struct hlist_node hash;
 219
 220         struct cfq_queue *cfq_queue;
 221         struct cfq_io_context *io_context;
 222
 223         unsigned int crq_flags;
 224 };
 225
 226 enum cfqq_state_flags {
 227         CFQ_CFQQ_FLAG_on_rr = 0,
 228         CFQ_CFQQ_FLAG_wait_request,
 229         CFQ_CFQQ_FLAG_must_alloc,
 230         CFQ_CFQQ_FLAG_must_alloc_slice,
 231         CFQ_CFQQ_FLAG_must_dispatch,
 232         CFQ_CFQQ_FLAG_fifo_expire,
 233         CFQ_CFQQ_FLAG_idle_window,
 234         CFQ_CFQQ_FLAG_prio_changed,
 235 };
 236
 237 #define CFQ_CFQQ_FNS(name)                                              \
 238 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 239 {                                                                       \
 240         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 241 }                                                                       \
 242 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 243 {                                                                       \
 244         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 245 }                                                                       \
 246 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 247 {                                                                       \
 248         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 249 }
 250
 251 CFQ_CFQQ_FNS(on_rr);
 252 CFQ_CFQQ_FNS(wait_request);
 253 CFQ_CFQQ_FNS(must_alloc);
 254 CFQ_CFQQ_FNS(must_alloc_slice);
 255 CFQ_CFQQ_FNS(must_dispatch);
 256 CFQ_CFQQ_FNS(fifo_expire);
 257 CFQ_CFQQ_FNS(idle_window);
 258 CFQ_CFQQ_FNS(prio_changed);
 259 #undef CFQ_CFQQ_FNS
 260
 261 enum cfq_rq_state_flags {
 262         CFQ_CRQ_FLAG_is_sync = 0,
 263 };
 264
 265 #define CFQ_CRQ_FNS(name)                                               \
 266 static inline void cfq_mark_crq_##name(struct cfq_rq *crq)              \
 267 {                                                                       \
 268         crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name);                   \
 269 }                                                                       \
 270 static inline void cfq_clear_crq_##name(struct cfq_rq *crq)             \
 271 {                                                                       \
 272         crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name);                  \
 273 }                                                                       \
 274 static inline int cfq_crq_##name(const struct cfq_rq *crq)              \
 275 {                                                                       \
 276         return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0;      \
 277 }
 278
 279 CFQ_CRQ_FNS(is_sync);
 280 #undef CFQ_CRQ_FNS
 281
 282 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 283 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
 284 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
 285
 286 #define process_sync(tsk)       ((tsk)->flags & PF_SYNCWRITE)
 287
 288 /*
 289  * lots of deadline iosched dupes, can be abstracted later...
 290  */
 291 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
 292 {
 293         hlist_del_init(&crq->hash);
 294 }
 295
 296 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
 297 {
 298         const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
 299
 300         hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
 301 }
 302
 303 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
 304 {
 305         struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
 306         struct hlist_node *entry, *next;
 307
 308         hlist_for_each_safe(entry, next, hash_list) {
 309                 struct cfq_rq *crq = list_entry_hash(entry);
 310                 struct request *__rq = crq->request;
 311
 312                 if (!rq_mergeable(__rq)) {
 313                         cfq_del_crq_hash(crq);
 314                         continue;
 315                 }
 316
 317                 if (rq_hash_key(__rq) == offset)
 318                         return __rq;
 319         }
 320
 321         return NULL;
 322 }
 323
 324 /*
 325  * scheduler run of queue, if there are requests pending and no one in the
 326  * driver that will restart queueing
 327  */
 328 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 329 {
 330         if (cfqd->busy_queues)
 331                 kblockd_schedule_work(&cfqd->unplug_work);
 332 }
 333
 334 static int cfq_queue_empty(request_queue_t *q)
 335 {
 336         struct cfq_data *cfqd = q->elevator->elevator_data;
 337
 338         return !cfqd->busy_queues;
 339 }
 340
 341 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
 342 {
 343         if (rw == READ || process_sync(task))
 344                 return task->pid;
 345
 346         return CFQ_KEY_ASYNC;
 347 }
 348
 349 /*
 350  * Lifted from AS - choose which of crq1 and crq2 that is best served now.
 351  * We choose the request that is closest to the head right now. Distance
 352  * behind the head is penalized and only allowed to a certain extent.
 353  */
 354 static struct cfq_rq *
 355 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
 356 {
 357         sector_t last, s1, s2, d1 = 0, d2 = 0;
 358         unsigned long back_max;
 359 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 360 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 361         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 362
 363         if (crq1 == NULL || crq1 == crq2)
 364                 return crq2;
 365         if (crq2 == NULL)
 366                 return crq1;
 367
 368         if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
 369                 return crq1;
 370         else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
 371                 return crq2;
 372
 373         s1 = crq1->request->sector;
 374         s2 = crq2->request->sector;
 375
 376         last = cfqd->last_sector;
 377
 378         /*
 379          * by definition, 1KiB is 2 sectors
 380          */
 381         back_max = cfqd->cfq_back_max * 2;
 382
 383         /*
 384          * Strict one way elevator _except_ in the case where we allow
 385          * short backward seeks which are biased as twice the cost of a
 386          * similar forward seek.
 387          */
 388         if (s1 >= last)
 389                 d1 = s1 - last;
 390         else if (s1 + back_max >= last)
 391                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 392         else
 393                 wrap |= CFQ_RQ1_WRAP;
 394
 395         if (s2 >= last)
 396                 d2 = s2 - last;
 397         else if (s2 + back_max >= last)
 398                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 399         else
 400                 wrap |= CFQ_RQ2_WRAP;
 401
 402         /* Found required data */
 403
 404         /*
 405          * By doing switch() on the bit mask "wrap" we avoid having to
 406          * check two variables for all permutations: --> faster!
 407          */
 408         switch (wrap) {
 409         case 0: /* common case for CFQ: crq1 and crq2 not wrapped */
 410                 if (d1 < d2)
 411                         return crq1;
 412                 else if (d2 < d1)
 413                         return crq2;
 414                 else {
 415                         if (s1 >= s2)
 416                                 return crq1;
 417                         else
 418                                 return crq2;
 419                 }
 420
 421         case CFQ_RQ2_WRAP:
 422                 return crq1;
 423         case CFQ_RQ1_WRAP:
 424                 return crq2;
 425         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both crqs wrapped */
 426         default:
 427                 /*
 428                  * Since both rqs are wrapped,
 429                  * start with the one that's further behind head
 430                  * (--> only *one* back seek required),
 431                  * since back seek takes more time than forward.
 432                  */
 433                 if (s1 <= s2)
 434                         return crq1;
 435                 else
 436                         return crq2;
 437         }
 438 }
 439
 440 /*
 441  * would be nice to take fifo expire time into account as well
 442  */
 443 static struct cfq_rq *
 444 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 445                   struct cfq_rq *last)
 446 {
 447         struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
 448         struct rb_node *rbnext, *rbprev;
 449
 450         if (!(rbnext = rb_next(&last->rb_node))) {
 451                 rbnext = rb_first(&cfqq->sort_list);
 452                 if (rbnext == &last->rb_node)
 453                         rbnext = NULL;
 454         }
 455
 456         rbprev = rb_prev(&last->rb_node);
 457
 458         if (rbprev)
 459                 crq_prev = rb_entry_crq(rbprev);
 460         if (rbnext)
 461                 crq_next = rb_entry_crq(rbnext);
 462
 463         return cfq_choose_req(cfqd, crq_next, crq_prev);
 464 }
 465
 466 static void cfq_update_next_crq(struct cfq_rq *crq)
 467 {
 468         struct cfq_queue *cfqq = crq->cfq_queue;
 469
 470         if (cfqq->next_crq == crq)
 471                 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
 472 }
 473
 474 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 475 {
 476         struct cfq_data *cfqd = cfqq->cfqd;
 477         struct list_head *list, *entry;
 478
 479         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 480
 481         list_del(&cfqq->cfq_list);
 482
 483         if (cfq_class_rt(cfqq))
 484                 list = &cfqd->cur_rr;
 485         else if (cfq_class_idle(cfqq))
 486                 list = &cfqd->idle_rr;
 487         else {
 488                 /*
 489                  * if cfqq has requests in flight, don't allow it to be
 490                  * found in cfq_set_active_queue before it has finished them.
 491                  * this is done to increase fairness between a process that
 492                  * has lots of io pending vs one that only generates one
 493                  * sporadically or synchronously
 494                  */
 495                 if (cfq_cfqq_dispatched(cfqq))
 496                         list = &cfqd->busy_rr;
 497                 else
 498                         list = &cfqd->rr_list[cfqq->ioprio];
 499         }
 500
 501         /*
 502          * if queue was preempted, just add to front to be fair. busy_rr
 503          * isn't sorted.
 504          */
 505         if (preempted || list == &cfqd->busy_rr) {
 506                 list_add(&cfqq->cfq_list, list);
 507                 return;
 508         }
 509
 510         /*
 511          * sort by when queue was last serviced
 512          */
 513         entry = list;
 514         while ((entry = entry->prev) != list) {
 515                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
 516
 517                 if (!__cfqq->service_last)
 518                         break;
 519                 if (time_before(__cfqq->service_last, cfqq->service_last))
 520                         break;
 521         }
 522
 523         list_add(&cfqq->cfq_list, entry);
 524 }
 525
 526 /*
 527  * add to busy list of queues for service, trying to be fair in ordering
 528  * the pending list according to last request service
 529  */
 530 static inline void
 531 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 532 {
 533         BUG_ON(cfq_cfqq_on_rr(cfqq));
 534         cfq_mark_cfqq_on_rr(cfqq);
 535         cfqd->busy_queues++;
 536
 537         cfq_resort_rr_list(cfqq, 0);
 538 }
 539
 540 static inline void
 541 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 542 {
 543         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 544         cfq_clear_cfqq_on_rr(cfqq);
 545         list_move(&cfqq->cfq_list, &cfqd->empty_list);
 546
 547         BUG_ON(!cfqd->busy_queues);
 548         cfqd->busy_queues--;
 549 }
 550
 551 /*
 552  * rb tree support functions
 553  */
 554 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
 555 {
 556         struct cfq_queue *cfqq = crq->cfq_queue;
 557         struct cfq_data *cfqd = cfqq->cfqd;
 558         const int sync = cfq_crq_is_sync(crq);
 559
 560         BUG_ON(!cfqq->queued[sync]);
 561         cfqq->queued[sync]--;
 562
 563         cfq_update_next_crq(crq);
 564
 565         rb_erase(&crq->rb_node, &cfqq->sort_list);
 566         RB_CLEAR_COLOR(&crq->rb_node);
 567
 568         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
 569                 cfq_del_cfqq_rr(cfqd, cfqq);
 570 }
 571
 572 static struct cfq_rq *
 573 __cfq_add_crq_rb(struct cfq_rq *crq)
 574 {
 575         struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
 576         struct rb_node *parent = NULL;
 577         struct cfq_rq *__crq;
 578
 579         while (*p) {
 580                 parent = *p;
 581                 __crq = rb_entry_crq(parent);
 582
 583                 if (crq->rb_key < __crq->rb_key)
 584                         p = &(*p)->rb_left;
 585                 else if (crq->rb_key > __crq->rb_key)
 586                         p = &(*p)->rb_right;
 587                 else
 588                         return __crq;
 589         }
 590
 591         rb_link_node(&crq->rb_node, parent, p);
 592         return NULL;
 593 }
 594
 595 static void cfq_add_crq_rb(struct cfq_rq *crq)
 596 {
 597         struct cfq_queue *cfqq = crq->cfq_queue;
 598         struct cfq_data *cfqd = cfqq->cfqd;
 599         struct request *rq = crq->request;
 600         struct cfq_rq *__alias;
 601
 602         crq->rb_key = rq_rb_key(rq);
 603         cfqq->queued[cfq_crq_is_sync(crq)]++;
 604
 605         /*
 606          * looks a little odd, but the first insert might return an alias.
 607          * if that happens, put the alias on the dispatch list
 608          */
 609         while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
 610                 cfq_dispatch_insert(cfqd->queue, __alias);
 611
 612         rb_insert_color(&crq->rb_node, &cfqq->sort_list);
 613
 614         if (!cfq_cfqq_on_rr(cfqq))
 615                 cfq_add_cfqq_rr(cfqd, cfqq);
 616
 617         /*
 618          * check if this request is a better next-serve candidate
 619          */
 620         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
 621 }
 622
 623 static inline void
 624 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
 625 {
 626         rb_erase(&crq->rb_node, &cfqq->sort_list);
 627         cfqq->queued[cfq_crq_is_sync(crq)]--;
 628
 629         cfq_add_crq_rb(crq);
 630 }
 631
 632 static struct request *
 633 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 634 {
 635         struct task_struct *tsk = current;
 636         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
 637         struct cfq_queue *cfqq;
 638         struct rb_node *n;
 639         sector_t sector;
 640
 641         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
 642         if (!cfqq)
 643                 goto out;
 644
 645         sector = bio->bi_sector + bio_sectors(bio);
 646         n = cfqq->sort_list.rb_node;
 647         while (n) {
 648                 struct cfq_rq *crq = rb_entry_crq(n);
 649
 650                 if (sector < crq->rb_key)
 651                         n = n->rb_left;
 652                 else if (sector > crq->rb_key)
 653                         n = n->rb_right;
 654                 else
 655                         return crq->request;
 656         }
 657
 658 out:
 659         return NULL;
 660 }
 661
 662 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 663 {
 664         struct cfq_data *cfqd = q->elevator->elevator_data;
 665
 666         cfqd->rq_in_driver++;
 667 }
 668
 669 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 670 {
 671         struct cfq_data *cfqd = q->elevator->elevator_data;
 672
 673         WARN_ON(!cfqd->rq_in_driver);
 674         cfqd->rq_in_driver--;
 675 }
 676
 677 static void cfq_remove_request(struct request *rq)
 678 {
 679         struct cfq_rq *crq = RQ_DATA(rq);
 680
 681         list_del_init(&rq->queuelist);
 682         cfq_del_crq_rb(crq);
 683         cfq_del_crq_hash(crq);
 684 }
 685
 686 static int
 687 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 688 {
 689         struct cfq_data *cfqd = q->elevator->elevator_data;
 690         struct request *__rq;
 691         int ret;
 692
 693         __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
 694         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 695                 ret = ELEVATOR_BACK_MERGE;
 696                 goto out;
 697         }
 698
 699         __rq = cfq_find_rq_fmerge(cfqd, bio);
 700         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 701                 ret = ELEVATOR_FRONT_MERGE;
 702                 goto out;
 703         }
 704
 705         return ELEVATOR_NO_MERGE;
 706 out:
 707         *req = __rq;
 708         return ret;
 709 }
 710
 711 static void cfq_merged_request(request_queue_t *q, struct request *req)
 712 {
 713         struct cfq_data *cfqd = q->elevator->elevator_data;
 714         struct cfq_rq *crq = RQ_DATA(req);
 715
 716         cfq_del_crq_hash(crq);
 717         cfq_add_crq_hash(cfqd, crq);
 718
 719         if (rq_rb_key(req) != crq->rb_key) {
 720                 struct cfq_queue *cfqq = crq->cfq_queue;
 721
 722                 cfq_update_next_crq(crq);
 723                 cfq_reposition_crq_rb(cfqq, crq);
 724         }
 725 }
 726
 727 static void
 728 cfq_merged_requests(request_queue_t *q, struct request *rq,
 729                     struct request *next)
 730 {
 731         cfq_merged_request(q, rq);
 732
 733         /*
 734          * reposition in fifo if next is older than rq
 735          */
 736         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 737             time_before(next->start_time, rq->start_time))
 738                 list_move(&rq->queuelist, &next->queuelist);
 739
 740         cfq_remove_request(next);
 741 }
 742
 743 static inline void
 744 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 745 {
 746         if (cfqq) {
 747                 /*
 748                  * stop potential idle class queues waiting service
 749                  */
 750                 del_timer(&cfqd->idle_class_timer);
 751
 752                 cfqq->slice_start = jiffies;
 753                 cfqq->slice_end = 0;
 754                 cfqq->slice_left = 0;
 755                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 756                 cfq_clear_cfqq_fifo_expire(cfqq);
 757         }
 758
 759         cfqd->active_queue = cfqq;
 760 }
 761
 762 /*
 763  * current cfqq expired its slice (or was too idle), select new one
 764  */
 765 static void
 766 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 767                     int preempted)
 768 {
 769         unsigned long now = jiffies;
 770
 771         if (cfq_cfqq_wait_request(cfqq))
 772                 del_timer(&cfqd->idle_slice_timer);
 773
 774         if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
 775                 cfqq->service_last = now;
 776                 cfq_schedule_dispatch(cfqd);
 777         }
 778
 779         cfq_clear_cfqq_must_dispatch(cfqq);
 780         cfq_clear_cfqq_wait_request(cfqq);
 781
 782         /*
 783          * store what was left of this slice, if the queue idled out
 784          * or was preempted
 785          */
 786         if (time_after(cfqq->slice_end, now))
 787                 cfqq->slice_left = cfqq->slice_end - now;
 788         else
 789                 cfqq->slice_left = 0;
 790
 791         if (cfq_cfqq_on_rr(cfqq))
 792                 cfq_resort_rr_list(cfqq, preempted);
 793
 794         if (cfqq == cfqd->active_queue)
 795                 cfqd->active_queue = NULL;
 796
 797         if (cfqd->active_cic) {
 798                 put_io_context(cfqd->active_cic->ioc);
 799                 cfqd->active_cic = NULL;
 800         }
 801
 802         cfqd->dispatch_slice = 0;
 803 }
 804
 805 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 806 {
 807         struct cfq_queue *cfqq = cfqd->active_queue;
 808
 809         if (cfqq)
 810                 __cfq_slice_expired(cfqd, cfqq, preempted);
 811 }
 812
 813 /*
 814  * 0
 815  * 0,1
 816  * 0,1,2
 817  * 0,1,2,3
 818  * 0,1,2,3,4
 819  * 0,1,2,3,4,5
 820  * 0,1,2,3,4,5,6
 821  * 0,1,2,3,4,5,6,7
 822  */
 823 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 824 {
 825         int prio, wrap;
 826
 827         prio = -1;
 828         wrap = 0;
 829         do {
 830                 int p;
 831
 832                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 833                         if (!list_empty(&cfqd->rr_list[p])) {
 834                                 prio = p;
 835                                 break;
 836                         }
 837                 }
 838
 839                 if (prio != -1)
 840                         break;
 841                 cfqd->cur_prio = 0;
 842                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 843                         cfqd->cur_end_prio = 0;
 844                         if (wrap)
 845                                 break;
 846                         wrap = 1;
 847                 }
 848         } while (1);
 849
 850         if (unlikely(prio == -1))
 851                 return -1;
 852
 853         BUG_ON(prio >= CFQ_PRIO_LISTS);
 854
 855         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 856
 857         cfqd->cur_prio = prio + 1;
 858         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 859                 cfqd->cur_end_prio = cfqd->cur_prio;
 860                 cfqd->cur_prio = 0;
 861         }
 862         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 863                 cfqd->cur_prio = 0;
 864                 cfqd->cur_end_prio = 0;
 865         }
 866
 867         return prio;
 868 }
 869
 870 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 871 {
 872         struct cfq_queue *cfqq = NULL;
 873
 874         /*
 875          * if current list is non-empty, grab first entry. if it is empty,
 876          * get next prio level and grab first entry then if any are spliced
 877          */
 878         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
 879                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 880
 881         /*
 882          * if we have idle queues and no rt or be queues had pending
 883          * requests, either allow immediate service if the grace period
 884          * has passed or arm the idle grace timer
 885          */
 886         if (!cfqq && !list_empty(&cfqd->idle_rr)) {
 887                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 888
 889                 if (time_after_eq(jiffies, end))
 890                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 891                 else
 892                         mod_timer(&cfqd->idle_class_timer, end);
 893         }
 894
 895         __cfq_set_active_queue(cfqd, cfqq);
 896         return cfqq;
 897 }
 898
 899 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 900
 901 {
 902         struct cfq_io_context *cic;
 903         unsigned long sl;
 904
 905         WARN_ON(!RB_EMPTY(&cfqq->sort_list));
 906         WARN_ON(cfqq != cfqd->active_queue);
 907
 908         /*
 909          * idle is disabled, either manually or by past process history
 910          */
 911         if (!cfqd->cfq_slice_idle)
 912                 return 0;
 913         if (!cfq_cfqq_idle_window(cfqq))
 914                 return 0;
 915         /*
 916          * task has exited, don't wait
 917          */
 918         cic = cfqd->active_cic;
 919         if (!cic || !cic->ioc->task)
 920                 return 0;
 921
 922         cfq_mark_cfqq_must_dispatch(cfqq);
 923         cfq_mark_cfqq_wait_request(cfqq);
 924
 925         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 926
 927         /*
 928          * we don't want to idle for seeks, but we do want to allow
 929          * fair distribution of slice time for a process doing back-to-back
 930          * seeks. so allow a little bit of time for him to submit a new rq
 931          */
 932         if (sample_valid(cic->seek_samples) && cic->seek_mean > 131072)
 933                 sl = 2;
 934
 935         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 936         return 1;
 937 }
 938
 939 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
 940 {
 941         struct cfq_data *cfqd = q->elevator->elevator_data;
 942         struct cfq_queue *cfqq = crq->cfq_queue;
 943
 944         cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
 945         cfq_remove_request(crq->request);
 946         cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
 947         elv_dispatch_sort(q, crq->request);
 948 }
 949
 950 /*
 951  * return expired entry, or NULL to just start from scratch in rbtree
 952  */
 953 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
 954 {
 955         struct cfq_data *cfqd = cfqq->cfqd;
 956         struct request *rq;
 957         struct cfq_rq *crq;
 958
 959         if (cfq_cfqq_fifo_expire(cfqq))
 960                 return NULL;
 961
 962         if (!list_empty(&cfqq->fifo)) {
 963                 int fifo = cfq_cfqq_class_sync(cfqq);
 964
 965                 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
 966                 rq = crq->request;
 967                 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
 968                         cfq_mark_cfqq_fifo_expire(cfqq);
 969                         return crq;
 970                 }
 971         }
 972
 973         return NULL;
 974 }
 975
 976 /*
 977  * Scale schedule slice based on io priority. Use the sync time slice only
 978  * if a queue is marked sync and has sync io queued. A sync queue with async
 979  * io only, should not get full sync slice length.
 980  */
 981 static inline int
 982 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 983 {
 984         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 985
 986         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 987
 988         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
 989 }
 990
 991 static inline void
 992 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 993 {
 994         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 995 }
 996
 997 static inline int
 998 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 999 {
1000         const int base_rq = cfqd->cfq_slice_async_rq;
1001
1002         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1003
1004         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1005 }
1006
1007 /*
1008  * get next queue for service
1009  */
1010 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1011 {
1012         unsigned long now = jiffies;
1013         struct cfq_queue *cfqq;
1014
1015         cfqq = cfqd->active_queue;
1016         if (!cfqq)
1017                 goto new_queue;
1018
1019         /*
1020          * slice has expired
1021          */
1022         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
1023                 goto expire;
1024
1025         /*
1026          * if queue has requests, dispatch one. if not, check if
1027          * enough slice is left to wait for one
1028          */
1029         if (!RB_EMPTY(&cfqq->sort_list))
1030                 goto keep_queue;
1031         else if (cfq_cfqq_class_sync(cfqq) &&
1032                  time_before(now, cfqq->slice_end)) {
1033                 if (cfq_arm_slice_timer(cfqd, cfqq))
1034                         return NULL;
1035         }
1036
1037 expire:
1038         cfq_slice_expired(cfqd, 0);
1039 new_queue:
1040         cfqq = cfq_set_active_queue(cfqd);
1041 keep_queue:
1042         return cfqq;
1043 }
1044
1045 static int
1046 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1047                         int max_dispatch)
1048 {
1049         int dispatched = 0;
1050
1051         BUG_ON(RB_EMPTY(&cfqq->sort_list));
1052
1053         do {
1054                 struct cfq_rq *crq;
1055
1056                 /*
1057                  * follow expired path, else get first next available
1058                  */
1059                 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1060                         crq = cfqq->next_crq;
1061
1062                 /*
1063                  * finally, insert request into driver dispatch list
1064                  */
1065                 cfq_dispatch_insert(cfqd->queue, crq);
1066
1067                 cfqd->dispatch_slice++;
1068                 dispatched++;
1069
1070                 if (!cfqd->active_cic) {
1071                         atomic_inc(&crq->io_context->ioc->refcount);
1072                         cfqd->active_cic = crq->io_context;
1073                 }
1074
1075                 if (RB_EMPTY(&cfqq->sort_list))
1076                         break;
1077
1078         } while (dispatched < max_dispatch);
1079
1080         /*
1081          * if slice end isn't set yet, set it. if at least one request was
1082          * sync, use the sync time slice value
1083          */
1084         if (!cfqq->slice_end)
1085                 cfq_set_prio_slice(cfqd, cfqq);
1086
1087         /*
1088          * expire an async queue immediately if it has used up its slice. idle
1089          * queue always expire after 1 dispatch round.
1090          */
1091         if ((!cfq_cfqq_sync(cfqq) &&
1092             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1093             cfq_class_idle(cfqq))
1094                 cfq_slice_expired(cfqd, 0);
1095
1096         return dispatched;
1097 }
1098
1099 static int
1100 cfq_forced_dispatch_cfqqs(struct list_head *list)
1101 {
1102         int dispatched = 0;
1103         struct cfq_queue *cfqq, *next;
1104         struct cfq_rq *crq;
1105
1106         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1107                 while ((crq = cfqq->next_crq)) {
1108                         cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1109                         dispatched++;
1110                 }
1111                 BUG_ON(!list_empty(&cfqq->fifo));
1112         }
1113         return dispatched;
1114 }
1115
1116 static int
1117 cfq_forced_dispatch(struct cfq_data *cfqd)
1118 {
1119         int i, dispatched = 0;
1120
1121         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1122                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1123
1124         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1125         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1126         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1127
1128         cfq_slice_expired(cfqd, 0);
1129
1130         BUG_ON(cfqd->busy_queues);
1131
1132         return dispatched;
1133 }
1134
1135 static int
1136 cfq_dispatch_requests(request_queue_t *q, int force)
1137 {
1138         struct cfq_data *cfqd = q->elevator->elevator_data;
1139         struct cfq_queue *cfqq;
1140
1141         if (!cfqd->busy_queues)
1142                 return 0;
1143
1144         if (unlikely(force))
1145                 return cfq_forced_dispatch(cfqd);
1146
1147         cfqq = cfq_select_queue(cfqd);
1148         if (cfqq) {
1149                 int max_dispatch;
1150
1151                 cfq_clear_cfqq_must_dispatch(cfqq);
1152                 cfq_clear_cfqq_wait_request(cfqq);
1153                 del_timer(&cfqd->idle_slice_timer);
1154
1155                 max_dispatch = cfqd->cfq_quantum;
1156                 if (cfq_class_idle(cfqq))
1157                         max_dispatch = 1;
1158
1159                 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1160         }
1161
1162         return 0;
1163 }
1164
1165 /*
1166  * task holds one reference to the queue, dropped when task exits. each crq
1167  * in-flight on this queue also holds a reference, dropped when crq is freed.
1168  *
1169  * queue lock must be held here.
1170  */
1171 static void cfq_put_queue(struct cfq_queue *cfqq)
1172 {
1173         struct cfq_data *cfqd = cfqq->cfqd;
1174
1175         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1176
1177         if (!atomic_dec_and_test(&cfqq->ref))
1178                 return;
1179
1180         BUG_ON(rb_first(&cfqq->sort_list));
1181         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1182         BUG_ON(cfq_cfqq_on_rr(cfqq));
1183
1184         if (unlikely(cfqd->active_queue == cfqq))
1185                 __cfq_slice_expired(cfqd, cfqq, 0);
1186
1187         /*
1188          * it's on the empty list and still hashed
1189          */
1190         list_del(&cfqq->cfq_list);
1191         hlist_del(&cfqq->cfq_hash);
1192         kmem_cache_free(cfq_pool, cfqq);
1193 }
1194
1195 static inline struct cfq_queue *
1196 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1197                     const int hashval)
1198 {
1199         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1200         struct hlist_node *entry;
1201         struct cfq_queue *__cfqq;
1202
1203         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1204                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1205
1206                 if (__cfqq->key == key && (__p == prio || !prio))
1207                         return __cfqq;
1208         }
1209
1210         return NULL;
1211 }
1212
1213 static struct cfq_queue *
1214 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1215 {
1216         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1217 }
1218
1219 static void cfq_free_io_context(struct io_context *ioc)
1220 {
1221         struct cfq_io_context *__cic;
1222         struct rb_node *n;
1223         int freed = 0;
1224
1225         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1226                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1227                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1228                 kmem_cache_free(cfq_ioc_pool, __cic);
1229                 freed++;
1230         }
1231
1232         if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1233                 complete(ioc_gone);
1234 }
1235
1236 static void cfq_trim(struct io_context *ioc)
1237 {
1238         ioc->set_ioprio = NULL;
1239         cfq_free_io_context(ioc);
1240 }
1241
1242 /*
1243  * Called with interrupts disabled
1244  */
1245 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1246 {
1247         struct cfq_data *cfqd = cic->key;
1248         request_queue_t *q;
1249
1250         if (!cfqd)
1251                 return;
1252
1253         q = cfqd->queue;
1254
1255         WARN_ON(!irqs_disabled());
1256
1257         spin_lock(q->queue_lock);
1258
1259         if (cic->cfqq[ASYNC]) {
1260                 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
1261                         __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
1262                 cfq_put_queue(cic->cfqq[ASYNC]);
1263                 cic->cfqq[ASYNC] = NULL;
1264         }
1265
1266         if (cic->cfqq[SYNC]) {
1267                 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
1268                         __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
1269                 cfq_put_queue(cic->cfqq[SYNC]);
1270                 cic->cfqq[SYNC] = NULL;
1271         }
1272
1273         cic->key = NULL;
1274         list_del_init(&cic->queue_list);
1275         spin_unlock(q->queue_lock);
1276 }
1277
1278 static void cfq_exit_io_context(struct io_context *ioc)
1279 {
1280         struct cfq_io_context *__cic;
1281         unsigned long flags;
1282         struct rb_node *n;
1283
1284         /*
1285          * put the reference this task is holding to the various queues
1286          */
1287         read_lock_irqsave(&cfq_exit_lock, flags);
1288
1289         n = rb_first(&ioc->cic_root);
1290         while (n != NULL) {
1291                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1292
1293                 cfq_exit_single_io_context(__cic);
1294                 n = rb_next(n);
1295         }
1296
1297         read_unlock_irqrestore(&cfq_exit_lock, flags);
1298 }
1299
1300 static struct cfq_io_context *
1301 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1302 {
1303         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1304
1305         if (cic) {
1306                 RB_CLEAR(&cic->rb_node);
1307                 cic->key = NULL;
1308                 cic->cfqq[ASYNC] = NULL;
1309                 cic->cfqq[SYNC] = NULL;
1310                 cic->last_end_request = jiffies;
1311                 cic->ttime_total = 0;
1312                 cic->ttime_samples = 0;
1313                 cic->ttime_mean = 0;
1314                 cic->dtor = cfq_free_io_context;
1315                 cic->exit = cfq_exit_io_context;
1316                 INIT_LIST_HEAD(&cic->queue_list);
1317                 atomic_inc(&ioc_count);
1318         }
1319
1320         return cic;
1321 }
1322
1323 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1324 {
1325         struct task_struct *tsk = current;
1326         int ioprio_class;
1327
1328         if (!cfq_cfqq_prio_changed(cfqq))
1329                 return;
1330
1331         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1332         switch (ioprio_class) {
1333                 default:
1334                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1335                 case IOPRIO_CLASS_NONE:
1336                         /*
1337                          * no prio set, place us in the middle of the BE classes
1338                          */
1339                         cfqq->ioprio = task_nice_ioprio(tsk);
1340                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1341                         break;
1342                 case IOPRIO_CLASS_RT:
1343                         cfqq->ioprio = task_ioprio(tsk);
1344                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1345                         break;
1346                 case IOPRIO_CLASS_BE:
1347                         cfqq->ioprio = task_ioprio(tsk);
1348                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1349                         break;
1350                 case IOPRIO_CLASS_IDLE:
1351                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1352                         cfqq->ioprio = 7;
1353                         cfq_clear_cfqq_idle_window(cfqq);
1354                         break;
1355         }
1356
1357         /*
1358          * keep track of original prio settings in case we have to temporarily
1359          * elevate the priority of this queue
1360          */
1361         cfqq->org_ioprio = cfqq->ioprio;
1362         cfqq->org_ioprio_class = cfqq->ioprio_class;
1363
1364         if (cfq_cfqq_on_rr(cfqq))
1365                 cfq_resort_rr_list(cfqq, 0);
1366
1367         cfq_clear_cfqq_prio_changed(cfqq);
1368 }
1369
1370 static inline void changed_ioprio(struct cfq_io_context *cic)
1371 {
1372         struct cfq_data *cfqd = cic->key;
1373         struct cfq_queue *cfqq;
1374         if (cfqd) {
1375                 spin_lock(cfqd->queue->queue_lock);
1376                 cfqq = cic->cfqq[ASYNC];
1377                 if (cfqq) {
1378                         struct cfq_queue *new_cfqq;
1379                         new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC,
1380                                                 cic->ioc->task, GFP_ATOMIC);
1381                         if (new_cfqq) {
1382                                 cic->cfqq[ASYNC] = new_cfqq;
1383                                 cfq_put_queue(cfqq);
1384                         }
1385                 }
1386                 cfqq = cic->cfqq[SYNC];
1387                 if (cfqq) {
1388                         cfq_mark_cfqq_prio_changed(cfqq);
1389                         cfq_init_prio_data(cfqq);
1390                 }
1391                 spin_unlock(cfqd->queue->queue_lock);
1392         }
1393 }
1394
1395 /*
1396  * callback from sys_ioprio_set, irqs are disabled
1397  */
1398 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1399 {
1400         struct cfq_io_context *cic;
1401         struct rb_node *n;
1402
1403         write_lock(&cfq_exit_lock);
1404
1405         n = rb_first(&ioc->cic_root);
1406         while (n != NULL) {
1407                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1408
1409                 changed_ioprio(cic);
1410                 n = rb_next(n);
1411         }
1412
1413         write_unlock(&cfq_exit_lock);
1414
1415         return 0;
1416 }
1417
1418 static struct cfq_queue *
1419 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1420               gfp_t gfp_mask)
1421 {
1422         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1423         struct cfq_queue *cfqq, *new_cfqq = NULL;
1424         unsigned short ioprio;
1425
1426 retry:
1427         ioprio = tsk->ioprio;
1428         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1429
1430         if (!cfqq) {
1431                 if (new_cfqq) {
1432                         cfqq = new_cfqq;
1433                         new_cfqq = NULL;
1434                 } else if (gfp_mask & __GFP_WAIT) {
1435                         spin_unlock_irq(cfqd->queue->queue_lock);
1436                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1437                         spin_lock_irq(cfqd->queue->queue_lock);
1438                         goto retry;
1439                 } else {
1440                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1441                         if (!cfqq)
1442                                 goto out;
1443                 }
1444
1445                 memset(cfqq, 0, sizeof(*cfqq));
1446
1447                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1448                 INIT_LIST_HEAD(&cfqq->cfq_list);
1449                 RB_CLEAR_ROOT(&cfqq->sort_list);
1450                 INIT_LIST_HEAD(&cfqq->fifo);
1451
1452                 cfqq->key = key;
1453                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1454                 atomic_set(&cfqq->ref, 0);
1455                 cfqq->cfqd = cfqd;
1456                 cfqq->service_last = 0;
1457                 /*
1458                  * set ->slice_left to allow preemption for a new process
1459                  */
1460                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1461                 cfq_mark_cfqq_idle_window(cfqq);
1462                 cfq_mark_cfqq_prio_changed(cfqq);
1463                 cfq_init_prio_data(cfqq);
1464         }
1465
1466         if (new_cfqq)
1467                 kmem_cache_free(cfq_pool, new_cfqq);
1468
1469         atomic_inc(&cfqq->ref);
1470 out:
1471         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1472         return cfqq;
1473 }
1474
1475 static void
1476 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1477 {
1478         read_lock(&cfq_exit_lock);
1479         rb_erase(&cic->rb_node, &ioc->cic_root);
1480         read_unlock(&cfq_exit_lock);
1481         kmem_cache_free(cfq_ioc_pool, cic);
1482         atomic_dec(&ioc_count);
1483 }
1484
1485 static struct cfq_io_context *
1486 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1487 {
1488         struct rb_node *n;
1489         struct cfq_io_context *cic;
1490         void *key = cfqd;
1491
1492 restart:
1493         n = ioc->cic_root.rb_node;
1494         while (n) {
1495                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1496                 if (unlikely(!cic->key)) {
1497                         cfq_drop_dead_cic(ioc, cic);
1498                         goto restart;
1499                 }
1500
1501                 if (key < cic->key)
1502                         n = n->rb_left;
1503                 else if (key > cic->key)
1504                         n = n->rb_right;
1505                 else
1506                         return cic;
1507         }
1508
1509         return NULL;
1510 }
1511
1512 static inline void
1513 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1514              struct cfq_io_context *cic)
1515 {
1516         struct rb_node **p;
1517         struct rb_node *parent;
1518         struct cfq_io_context *__cic;
1519
1520         cic->ioc = ioc;
1521         cic->key = cfqd;
1522
1523         ioc->set_ioprio = cfq_ioc_set_ioprio;
1524 restart:
1525         parent = NULL;
1526         p = &ioc->cic_root.rb_node;
1527         while (*p) {
1528                 parent = *p;
1529                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1530                 if (unlikely(!__cic->key)) {
1531                         cfq_drop_dead_cic(ioc, cic);
1532                         goto restart;
1533                 }
1534
1535                 if (cic->key < __cic->key)
1536                         p = &(*p)->rb_left;
1537                 else if (cic->key > __cic->key)
1538                         p = &(*p)->rb_right;
1539                 else
1540                         BUG();
1541         }
1542
1543         read_lock(&cfq_exit_lock);
1544         rb_link_node(&cic->rb_node, parent, p);
1545         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1546         list_add(&cic->queue_list, &cfqd->cic_list);
1547         read_unlock(&cfq_exit_lock);
1548 }
1549
1550 /*
1551  * Setup general io context and cfq io context. There can be several cfq
1552  * io contexts per general io context, if this process is doing io to more
1553  * than one device managed by cfq.
1554  */
1555 static struct cfq_io_context *
1556 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1557 {
1558         struct io_context *ioc = NULL;
1559         struct cfq_io_context *cic;
1560
1561         might_sleep_if(gfp_mask & __GFP_WAIT);
1562
1563         ioc = get_io_context(gfp_mask);
1564         if (!ioc)
1565                 return NULL;
1566
1567         cic = cfq_cic_rb_lookup(cfqd, ioc);
1568         if (cic)
1569                 goto out;
1570
1571         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1572         if (cic == NULL)
1573                 goto err;
1574
1575         cfq_cic_link(cfqd, ioc, cic);
1576 out:
1577         return cic;
1578 err:
1579         put_io_context(ioc);
1580         return NULL;
1581 }
1582
1583 static void
1584 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1585 {
1586         unsigned long elapsed, ttime;
1587
1588         /*
1589          * if this context already has stuff queued, thinktime is from
1590          * last queue not last end
1591          */
1592 #if 0
1593         if (time_after(cic->last_end_request, cic->last_queue))
1594                 elapsed = jiffies - cic->last_end_request;
1595         else
1596                 elapsed = jiffies - cic->last_queue;
1597 #else
1598                 elapsed = jiffies - cic->last_end_request;
1599 #endif
1600
1601         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1602
1603         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1604         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1605         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1606 }
1607
1608 static void
1609 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1610                        struct cfq_rq *crq)
1611 {
1612         sector_t sdist;
1613         u64 total;
1614
1615         if (cic->last_request_pos < crq->request->sector)
1616                 sdist = crq->request->sector - cic->last_request_pos;
1617         else
1618                 sdist = cic->last_request_pos - crq->request->sector;
1619
1620         /*
1621          * Don't allow the seek distance to get too large from the
1622          * odd fragment, pagein, etc
1623          */
1624         if (cic->seek_samples <= 60) /* second&third seek */
1625                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1626         else
1627                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1628
1629         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1630         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1631         total = cic->seek_total + (cic->seek_samples/2);
1632         do_div(total, cic->seek_samples);
1633         cic->seek_mean = (sector_t)total;
1634 }
1635
1636 /*
1637  * Disable idle window if the process thinks too long or seeks so much that
1638  * it doesn't matter
1639  */
1640 static void
1641 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1642                        struct cfq_io_context *cic)
1643 {
1644         int enable_idle = cfq_cfqq_idle_window(cfqq);
1645
1646         if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1647                 enable_idle = 0;
1648         else if (sample_valid(cic->ttime_samples)) {
1649                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1650                         enable_idle = 0;
1651                 else
1652                         enable_idle = 1;
1653         }
1654
1655         if (enable_idle)
1656                 cfq_mark_cfqq_idle_window(cfqq);
1657         else
1658                 cfq_clear_cfqq_idle_window(cfqq);
1659 }
1660
1661
1662 /*
1663  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1664  * no or if we aren't sure, a 1 will cause a preempt.
1665  */
1666 static int
1667 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1668                    struct cfq_rq *crq)
1669 {
1670         struct cfq_queue *cfqq = cfqd->active_queue;
1671
1672         if (cfq_class_idle(new_cfqq))
1673                 return 0;
1674
1675         if (!cfqq)
1676                 return 1;
1677
1678         if (cfq_class_idle(cfqq))
1679                 return 1;
1680         if (!cfq_cfqq_wait_request(new_cfqq))
1681                 return 0;
1682         /*
1683          * if it doesn't have slice left, forget it
1684          */
1685         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1686                 return 0;
1687         if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1688                 return 1;
1689
1690         return 0;
1691 }
1692
1693 /*
1694  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1695  * let it have half of its nominal slice.
1696  */
1697 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1698 {
1699         struct cfq_queue *__cfqq, *next;
1700
1701         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1702                 cfq_resort_rr_list(__cfqq, 1);
1703
1704         if (!cfqq->slice_left)
1705                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1706
1707         cfqq->slice_end = cfqq->slice_left + jiffies;
1708         __cfq_slice_expired(cfqd, cfqq, 1);
1709         __cfq_set_active_queue(cfqd, cfqq);
1710 }
1711
1712 /*
1713  * should really be a ll_rw_blk.c helper
1714  */
1715 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1716 {
1717         request_queue_t *q = cfqd->queue;
1718
1719         if (!blk_queue_plugged(q))
1720                 q->request_fn(q);
1721         else
1722                 __generic_unplug_device(q);
1723 }
1724
1725 /*
1726  * Called when a new fs request (crq) is added (to cfqq). Check if there's
1727  * something we should do about it
1728  */
1729 static void
1730 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1731                  struct cfq_rq *crq)
1732 {
1733         struct cfq_io_context *cic;
1734
1735         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1736
1737         /*
1738          * we never wait for an async request and we don't allow preemption
1739          * of an async request. so just return early
1740          */
1741         if (!cfq_crq_is_sync(crq))
1742                 return;
1743
1744         cic = crq->io_context;
1745
1746         cfq_update_io_thinktime(cfqd, cic);
1747         cfq_update_io_seektime(cfqd, cic, crq);
1748         cfq_update_idle_window(cfqd, cfqq, cic);
1749
1750         cic->last_queue = jiffies;
1751         cic->last_request_pos = crq->request->sector + crq->request->nr_sectors;
1752
1753         if (cfqq == cfqd->active_queue) {
1754                 /*
1755                  * if we are waiting for a request for this queue, let it rip
1756                  * immediately and flag that we must not expire this queue
1757                  * just now
1758                  */
1759                 if (cfq_cfqq_wait_request(cfqq)) {
1760                         cfq_mark_cfqq_must_dispatch(cfqq);
1761                         del_timer(&cfqd->idle_slice_timer);
1762                         cfq_start_queueing(cfqd, cfqq);
1763                 }
1764         } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1765                 /*
1766                  * not the active queue - expire current slice if it is
1767                  * idle and has expired it's mean thinktime or this new queue
1768                  * has some old slice time left and is of higher priority
1769                  */
1770                 cfq_preempt_queue(cfqd, cfqq);
1771                 cfq_mark_cfqq_must_dispatch(cfqq);
1772                 cfq_start_queueing(cfqd, cfqq);
1773         }
1774 }
1775
1776 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1777 {
1778         struct cfq_data *cfqd = q->elevator->elevator_data;
1779         struct cfq_rq *crq = RQ_DATA(rq);
1780         struct cfq_queue *cfqq = crq->cfq_queue;
1781
1782         cfq_init_prio_data(cfqq);
1783
1784         cfq_add_crq_rb(crq);
1785
1786         list_add_tail(&rq->queuelist, &cfqq->fifo);
1787
1788         if (rq_mergeable(rq))
1789                 cfq_add_crq_hash(cfqd, crq);
1790
1791         cfq_crq_enqueued(cfqd, cfqq, crq);
1792 }
1793
1794 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1795 {
1796         struct cfq_rq *crq = RQ_DATA(rq);
1797         struct cfq_queue *cfqq = crq->cfq_queue;
1798         struct cfq_data *cfqd = cfqq->cfqd;
1799         const int sync = cfq_crq_is_sync(crq);
1800         unsigned long now;
1801
1802         now = jiffies;
1803
1804         WARN_ON(!cfqd->rq_in_driver);
1805         WARN_ON(!cfqq->on_dispatch[sync]);
1806         cfqd->rq_in_driver--;
1807         cfqq->on_dispatch[sync]--;
1808
1809         if (!cfq_class_idle(cfqq))
1810                 cfqd->last_end_request = now;
1811
1812         if (!cfq_cfqq_dispatched(cfqq)) {
1813                 if (cfq_cfqq_on_rr(cfqq)) {
1814                         cfqq->service_last = now;
1815                         cfq_resort_rr_list(cfqq, 0);
1816                 }
1817                 cfq_schedule_dispatch(cfqd);
1818         }
1819
1820         if (cfq_crq_is_sync(crq))
1821                 crq->io_context->last_end_request = now;
1822 }
1823
1824 static struct request *
1825 cfq_former_request(request_queue_t *q, struct request *rq)
1826 {
1827         struct cfq_rq *crq = RQ_DATA(rq);
1828         struct rb_node *rbprev = rb_prev(&crq->rb_node);
1829
1830         if (rbprev)
1831                 return rb_entry_crq(rbprev)->request;
1832
1833         return NULL;
1834 }
1835
1836 static struct request *
1837 cfq_latter_request(request_queue_t *q, struct request *rq)
1838 {
1839         struct cfq_rq *crq = RQ_DATA(rq);
1840         struct rb_node *rbnext = rb_next(&crq->rb_node);
1841
1842         if (rbnext)
1843                 return rb_entry_crq(rbnext)->request;
1844
1845         return NULL;
1846 }
1847
1848 /*
1849  * we temporarily boost lower priority queues if they are holding fs exclusive
1850  * resources. they are boosted to normal prio (CLASS_BE/4)
1851  */
1852 static void cfq_prio_boost(struct cfq_queue *cfqq)
1853 {
1854         const int ioprio_class = cfqq->ioprio_class;
1855         const int ioprio = cfqq->ioprio;
1856
1857         if (has_fs_excl()) {
1858                 /*
1859                  * boost idle prio on transactions that would lock out other
1860                  * users of the filesystem
1861                  */
1862                 if (cfq_class_idle(cfqq))
1863                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1864                 if (cfqq->ioprio > IOPRIO_NORM)
1865                         cfqq->ioprio = IOPRIO_NORM;
1866         } else {
1867                 /*
1868                  * check if we need to unboost the queue
1869                  */
1870                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1871                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1872                 if (cfqq->ioprio != cfqq->org_ioprio)
1873                         cfqq->ioprio = cfqq->org_ioprio;
1874         }
1875
1876         /*
1877          * refile between round-robin lists if we moved the priority class
1878          */
1879         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1880             cfq_cfqq_on_rr(cfqq))
1881                 cfq_resort_rr_list(cfqq, 0);
1882 }
1883
1884 static inline int
1885 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1886                 struct task_struct *task, int rw)
1887 {
1888 #if 1
1889         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1890             !cfq_cfqq_must_alloc_slice(cfqq)) {
1891                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1892                 return ELV_MQUEUE_MUST;
1893         }
1894
1895         return ELV_MQUEUE_MAY;
1896 #else
1897         if (!cfqq || task->flags & PF_MEMALLOC)
1898                 return ELV_MQUEUE_MAY;
1899         if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1900                 if (cfq_cfqq_wait_request(cfqq))
1901                         return ELV_MQUEUE_MUST;
1902
1903                 /*
1904                  * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1905                  * can quickly flood the queue with writes from a single task
1906                  */
1907                 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
1908                         cfq_mark_cfqq_must_alloc_slice(cfqq);
1909                         return ELV_MQUEUE_MUST;
1910                 }
1911
1912                 return ELV_MQUEUE_MAY;
1913         }
1914         if (cfq_class_idle(cfqq))
1915                 return ELV_MQUEUE_NO;
1916         if (cfqq->allocated[rw] >= cfqd->max_queued) {
1917                 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1918                 int ret = ELV_MQUEUE_NO;
1919
1920                 if (ioc && ioc->nr_batch_requests)
1921                         ret = ELV_MQUEUE_MAY;
1922
1923                 put_io_context(ioc);
1924                 return ret;
1925         }
1926
1927         return ELV_MQUEUE_MAY;
1928 #endif
1929 }
1930
1931 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1932 {
1933         struct cfq_data *cfqd = q->elevator->elevator_data;
1934         struct task_struct *tsk = current;
1935         struct cfq_queue *cfqq;
1936
1937         /*
1938          * don't force setup of a queue from here, as a call to may_queue
1939          * does not necessarily imply that a request actually will be queued.
1940          * so just lookup a possibly existing queue, or return 'may queue'
1941          * if that fails
1942          */
1943         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1944         if (cfqq) {
1945                 cfq_init_prio_data(cfqq);
1946                 cfq_prio_boost(cfqq);
1947
1948                 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1949         }
1950
1951         return ELV_MQUEUE_MAY;
1952 }
1953
1954 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1955 {
1956         struct cfq_data *cfqd = q->elevator->elevator_data;
1957         struct request_list *rl = &q->rq;
1958
1959         if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1960                 smp_mb();
1961                 if (waitqueue_active(&rl->wait[READ]))
1962                         wake_up(&rl->wait[READ]);
1963         }
1964
1965         if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
1966                 smp_mb();
1967                 if (waitqueue_active(&rl->wait[WRITE]))
1968                         wake_up(&rl->wait[WRITE]);
1969         }
1970 }
1971
1972 /*
1973  * queue lock held here
1974  */
1975 static void cfq_put_request(request_queue_t *q, struct request *rq)
1976 {
1977         struct cfq_data *cfqd = q->elevator->elevator_data;
1978         struct cfq_rq *crq = RQ_DATA(rq);
1979
1980         if (crq) {
1981                 struct cfq_queue *cfqq = crq->cfq_queue;
1982                 const int rw = rq_data_dir(rq);
1983
1984                 BUG_ON(!cfqq->allocated[rw]);
1985                 cfqq->allocated[rw]--;
1986
1987                 put_io_context(crq->io_context->ioc);
1988
1989                 mempool_free(crq, cfqd->crq_pool);
1990                 rq->elevator_private = NULL;
1991
1992                 cfq_check_waiters(q, cfqq);
1993                 cfq_put_queue(cfqq);
1994         }
1995 }
1996
1997 /*
1998  * Allocate cfq data structures associated with this request.
1999  */
2000 static int
2001 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
2002                 gfp_t gfp_mask)
2003 {
2004         struct cfq_data *cfqd = q->elevator->elevator_data;
2005         struct task_struct *tsk = current;
2006         struct cfq_io_context *cic;
2007         const int rw = rq_data_dir(rq);
2008         pid_t key = cfq_queue_pid(tsk, rw);
2009         struct cfq_queue *cfqq;
2010         struct cfq_rq *crq;
2011         unsigned long flags;
2012         int is_sync = key != CFQ_KEY_ASYNC;
2013
2014         might_sleep_if(gfp_mask & __GFP_WAIT);
2015
2016         cic = cfq_get_io_context(cfqd, gfp_mask);
2017
2018         spin_lock_irqsave(q->queue_lock, flags);
2019
2020         if (!cic)
2021                 goto queue_fail;
2022
2023         if (!cic->cfqq[is_sync]) {
2024                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
2025                 if (!cfqq)
2026                         goto queue_fail;
2027
2028                 cic->cfqq[is_sync] = cfqq;
2029         } else
2030                 cfqq = cic->cfqq[is_sync];
2031
2032         cfqq->allocated[rw]++;
2033         cfq_clear_cfqq_must_alloc(cfqq);
2034         cfqd->rq_starved = 0;
2035         atomic_inc(&cfqq->ref);
2036         spin_unlock_irqrestore(q->queue_lock, flags);
2037
2038         crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
2039         if (crq) {
2040                 RB_CLEAR(&crq->rb_node);
2041                 crq->rb_key = 0;
2042                 crq->request = rq;
2043                 INIT_HLIST_NODE(&crq->hash);
2044                 crq->cfq_queue = cfqq;
2045                 crq->io_context = cic;
2046
2047                 if (is_sync)
2048                         cfq_mark_crq_is_sync(crq);
2049                 else
2050                         cfq_clear_crq_is_sync(crq);
2051
2052                 rq->elevator_private = crq;
2053                 return 0;
2054         }
2055
2056         spin_lock_irqsave(q->queue_lock, flags);
2057         cfqq->allocated[rw]--;
2058         if (!(cfqq->allocated[0] + cfqq->allocated[1]))
2059                 cfq_mark_cfqq_must_alloc(cfqq);
2060         cfq_put_queue(cfqq);
2061 queue_fail:
2062         if (cic)
2063                 put_io_context(cic->ioc);
2064         /*
2065          * mark us rq allocation starved. we need to kickstart the process
2066          * ourselves if there are no pending requests that can do it for us.
2067          * that would be an extremely rare OOM situation
2068          */
2069         cfqd->rq_starved = 1;
2070         cfq_schedule_dispatch(cfqd);
2071         spin_unlock_irqrestore(q->queue_lock, flags);
2072         return 1;
2073 }
2074
2075 static void cfq_kick_queue(void *data)
2076 {
2077         request_queue_t *q = data;
2078         struct cfq_data *cfqd = q->elevator->elevator_data;
2079         unsigned long flags;
2080
2081         spin_lock_irqsave(q->queue_lock, flags);
2082
2083         if (cfqd->rq_starved) {
2084                 struct request_list *rl = &q->rq;
2085
2086                 /*
2087                  * we aren't guaranteed to get a request after this, but we
2088                  * have to be opportunistic
2089                  */
2090                 smp_mb();
2091                 if (waitqueue_active(&rl->wait[READ]))
2092                         wake_up(&rl->wait[READ]);
2093                 if (waitqueue_active(&rl->wait[WRITE]))
2094                         wake_up(&rl->wait[WRITE]);
2095         }
2096
2097         blk_remove_plug(q);
2098         q->request_fn(q);
2099         spin_unlock_irqrestore(q->queue_lock, flags);
2100 }
2101
2102 /*
2103  * Timer running if the active_queue is currently idling inside its time slice
2104  */
2105 static void cfq_idle_slice_timer(unsigned long data)
2106 {
2107         struct cfq_data *cfqd = (struct cfq_data *) data;
2108         struct cfq_queue *cfqq;
2109         unsigned long flags;
2110
2111         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2112
2113         if ((cfqq = cfqd->active_queue) != NULL) {
2114                 unsigned long now = jiffies;
2115
2116                 /*
2117                  * expired
2118                  */
2119                 if (time_after(now, cfqq->slice_end))
2120                         goto expire;
2121
2122                 /*
2123                  * only expire and reinvoke request handler, if there are
2124                  * other queues with pending requests
2125                  */
2126                 if (!cfqd->busy_queues) {
2127                         cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2128                         add_timer(&cfqd->idle_slice_timer);
2129                         goto out_cont;
2130                 }
2131
2132                 /*
2133                  * not expired and it has a request pending, let it dispatch
2134                  */
2135                 if (!RB_EMPTY(&cfqq->sort_list)) {
2136                         cfq_mark_cfqq_must_dispatch(cfqq);
2137                         goto out_kick;
2138                 }
2139         }
2140 expire:
2141         cfq_slice_expired(cfqd, 0);
2142 out_kick:
2143         cfq_schedule_dispatch(cfqd);
2144 out_cont:
2145         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2146 }
2147
2148 /*
2149  * Timer running if an idle class queue is waiting for service
2150  */
2151 static void cfq_idle_class_timer(unsigned long data)
2152 {
2153         struct cfq_data *cfqd = (struct cfq_data *) data;
2154         unsigned long flags, end;
2155
2156         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2157
2158         /*
2159          * race with a non-idle queue, reset timer
2160          */
2161         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2162         if (!time_after_eq(jiffies, end)) {
2163                 cfqd->idle_class_timer.expires = end;
2164                 add_timer(&cfqd->idle_class_timer);
2165         } else
2166                 cfq_schedule_dispatch(cfqd);
2167
2168         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2169 }
2170
2171 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2172 {
2173         del_timer_sync(&cfqd->idle_slice_timer);
2174         del_timer_sync(&cfqd->idle_class_timer);
2175         blk_sync_queue(cfqd->queue);
2176 }
2177
2178 static void cfq_exit_queue(elevator_t *e)
2179 {
2180         struct cfq_data *cfqd = e->elevator_data;
2181         request_queue_t *q = cfqd->queue;
2182
2183         cfq_shutdown_timer_wq(cfqd);
2184
2185         write_lock(&cfq_exit_lock);
2186         spin_lock_irq(q->queue_lock);
2187
2188         if (cfqd->active_queue)
2189                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2190
2191         while (!list_empty(&cfqd->cic_list)) {
2192                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2193                                                         struct cfq_io_context,
2194                                                         queue_list);
2195                 if (cic->cfqq[ASYNC]) {
2196                         cfq_put_queue(cic->cfqq[ASYNC]);
2197                         cic->cfqq[ASYNC] = NULL;
2198                 }
2199                 if (cic->cfqq[SYNC]) {
2200                         cfq_put_queue(cic->cfqq[SYNC]);
2201                         cic->cfqq[SYNC] = NULL;
2202                 }
2203                 cic->key = NULL;
2204                 list_del_init(&cic->queue_list);
2205         }
2206
2207         spin_unlock_irq(q->queue_lock);
2208         write_unlock(&cfq_exit_lock);
2209
2210         cfq_shutdown_timer_wq(cfqd);
2211
2212         mempool_destroy(cfqd->crq_pool);
2213         kfree(cfqd->crq_hash);
2214         kfree(cfqd->cfq_hash);
2215         kfree(cfqd);
2216 }
2217
2218 static int cfq_init_queue(request_queue_t *q, elevator_t *e)
2219 {
2220         struct cfq_data *cfqd;
2221         int i;
2222
2223         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2224         if (!cfqd)
2225                 return -ENOMEM;
2226
2227         memset(cfqd, 0, sizeof(*cfqd));
2228
2229         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2230                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2231
2232         INIT_LIST_HEAD(&cfqd->busy_rr);
2233         INIT_LIST_HEAD(&cfqd->cur_rr);
2234         INIT_LIST_HEAD(&cfqd->idle_rr);
2235         INIT_LIST_HEAD(&cfqd->empty_list);
2236         INIT_LIST_HEAD(&cfqd->cic_list);
2237
2238         cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2239         if (!cfqd->crq_hash)
2240                 goto out_crqhash;
2241
2242         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2243         if (!cfqd->cfq_hash)
2244                 goto out_cfqhash;
2245
2246         cfqd->crq_pool = mempool_create_slab_pool(BLKDEV_MIN_RQ, crq_pool);
2247         if (!cfqd->crq_pool)
2248                 goto out_crqpool;
2249
2250         for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2251                 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2252         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2253                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2254
2255         e->elevator_data = cfqd;
2256
2257         cfqd->queue = q;
2258
2259         cfqd->max_queued = q->nr_requests / 4;
2260         q->nr_batching = cfq_queued;
2261
2262         init_timer(&cfqd->idle_slice_timer);
2263         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2264         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2265
2266         init_timer(&cfqd->idle_class_timer);
2267         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2268         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2269
2270         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2271
2272         cfqd->cfq_queued = cfq_queued;
2273         cfqd->cfq_quantum = cfq_quantum;
2274         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2275         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2276         cfqd->cfq_back_max = cfq_back_max;
2277         cfqd->cfq_back_penalty = cfq_back_penalty;
2278         cfqd->cfq_slice[0] = cfq_slice_async;
2279         cfqd->cfq_slice[1] = cfq_slice_sync;
2280         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2281         cfqd->cfq_slice_idle = cfq_slice_idle;
2282
2283         return 0;
2284 out_crqpool:
2285         kfree(cfqd->cfq_hash);
2286 out_cfqhash:
2287         kfree(cfqd->crq_hash);
2288 out_crqhash:
2289         kfree(cfqd);
2290         return -ENOMEM;
2291 }
2292
2293 static void cfq_slab_kill(void)
2294 {
2295         if (crq_pool)
2296                 kmem_cache_destroy(crq_pool);
2297         if (cfq_pool)
2298                 kmem_cache_destroy(cfq_pool);
2299         if (cfq_ioc_pool)
2300                 kmem_cache_destroy(cfq_ioc_pool);
2301 }
2302
2303 static int __init cfq_slab_setup(void)
2304 {
2305         crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2306                                         NULL, NULL);
2307         if (!crq_pool)
2308                 goto fail;
2309
2310         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2311                                         NULL, NULL);
2312         if (!cfq_pool)
2313                 goto fail;
2314
2315         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2316                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2317         if (!cfq_ioc_pool)
2318                 goto fail;
2319
2320         return 0;
2321 fail:
2322         cfq_slab_kill();
2323         return -ENOMEM;
2324 }
2325
2326 /*
2327  * sysfs parts below -->
2328  */
2329
2330 static ssize_t
2331 cfq_var_show(unsigned int var, char *page)
2332 {
2333         return sprintf(page, "%d\n", var);
2334 }
2335
2336 static ssize_t
2337 cfq_var_store(unsigned int *var, const char *page, size_t count)
2338 {
2339         char *p = (char *) page;
2340
2341         *var = simple_strtoul(p, &p, 10);
2342         return count;
2343 }
2344
2345 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2346 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2347 {                                                                       \
2348         struct cfq_data *cfqd = e->elevator_data;                       \
2349         unsigned int __data = __VAR;                                    \
2350         if (__CONV)                                                     \
2351                 __data = jiffies_to_msecs(__data);                      \
2352         return cfq_var_show(__data, (page));                            \
2353 }
2354 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2355 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2356 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2357 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2358 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2359 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2360 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2361 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2362 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2363 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2364 #undef SHOW_FUNCTION
2365
2366 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2367 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2368 {                                                                       \
2369         struct cfq_data *cfqd = e->elevator_data;                       \
2370         unsigned int __data;                                            \
2371         int ret = cfq_var_store(&__data, (page), count);                \
2372         if (__data < (MIN))                                             \
2373                 __data = (MIN);                                         \
2374         else if (__data > (MAX))                                        \
2375                 __data = (MAX);                                         \
2376         if (__CONV)                                                     \
2377                 *(__PTR) = msecs_to_jiffies(__data);                    \
2378         else                                                            \
2379                 *(__PTR) = __data;                                      \
2380         return ret;                                                     \
2381 }
2382 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2383 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2384 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2385 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2386 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2387 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2388 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2389 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2390 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2391 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2392 #undef STORE_FUNCTION
2393
2394 #define CFQ_ATTR(name) \
2395         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2396
2397 static struct elv_fs_entry cfq_attrs[] = {
2398         CFQ_ATTR(quantum),
2399         CFQ_ATTR(queued),
2400         CFQ_ATTR(fifo_expire_sync),
2401         CFQ_ATTR(fifo_expire_async),
2402         CFQ_ATTR(back_seek_max),
2403         CFQ_ATTR(back_seek_penalty),
2404         CFQ_ATTR(slice_sync),
2405         CFQ_ATTR(slice_async),
2406         CFQ_ATTR(slice_async_rq),
2407         CFQ_ATTR(slice_idle),
2408         __ATTR_NULL
2409 };
2410
2411 static struct elevator_type iosched_cfq = {
2412         .ops = {
2413                 .elevator_merge_fn =            cfq_merge,
2414                 .elevator_merged_fn =           cfq_merged_request,
2415                 .elevator_merge_req_fn =        cfq_merged_requests,
2416                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2417                 .elevator_add_req_fn =          cfq_insert_request,
2418                 .elevator_activate_req_fn =     cfq_activate_request,
2419                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2420                 .elevator_queue_empty_fn =      cfq_queue_empty,
2421                 .elevator_completed_req_fn =    cfq_completed_request,
2422                 .elevator_former_req_fn =       cfq_former_request,
2423                 .elevator_latter_req_fn =       cfq_latter_request,
2424                 .elevator_set_req_fn =          cfq_set_request,
2425                 .elevator_put_req_fn =          cfq_put_request,
2426                 .elevator_may_queue_fn =        cfq_may_queue,
2427                 .elevator_init_fn =             cfq_init_queue,
2428                 .elevator_exit_fn =             cfq_exit_queue,
2429                 .trim =                         cfq_trim,
2430         },
2431         .elevator_attrs =       cfq_attrs,
2432         .elevator_name =        "cfq",
2433         .elevator_owner =       THIS_MODULE,
2434 };
2435
2436 static int __init cfq_init(void)
2437 {
2438         int ret;
2439
2440         /*
2441          * could be 0 on HZ < 1000 setups
2442          */
2443         if (!cfq_slice_async)
2444                 cfq_slice_async = 1;
2445         if (!cfq_slice_idle)
2446                 cfq_slice_idle = 1;
2447
2448         if (cfq_slab_setup())
2449                 return -ENOMEM;
2450
2451         ret = elv_register(&iosched_cfq);
2452         if (ret)
2453                 cfq_slab_kill();
2454
2455         return ret;
2456 }
2457
2458 static void __exit cfq_exit(void)
2459 {
2460         DECLARE_COMPLETION(all_gone);
2461         elv_unregister(&iosched_cfq);
2462         ioc_gone = &all_gone;
2463         /* ioc_gone's update must be visible before reading ioc_count */
2464         smp_wmb();
2465         if (atomic_read(&ioc_count))
2466                 wait_for_completion(ioc_gone);
2467         synchronize_rcu();
2468         cfq_slab_kill();
2469 }
2470
2471 module_init(cfq_init);
2472 module_exit(cfq_exit);
2473
2474 MODULE_AUTHOR("Jens Axboe");
2475 MODULE_LICENSE("GPL");
2476 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");