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@kernel.dk>
   8  */
   9 #include <linux/module.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/elevator.h>
  12 #include <linux/rbtree.h>
  13 #include <linux/ioprio.h>
  14
  15 /*
  16  * tunables
  17  */
  18 static const int cfq_quantum = 4;               /* max queue in one round of service */
  19 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  20 static const int cfq_back_max = 16 * 1024;      /* maximum backwards seek, in KiB */
  21 static const int cfq_back_penalty = 2;          /* penalty of a backwards seek */
  22
  23 static const int cfq_slice_sync = HZ / 10;
  24 static int cfq_slice_async = HZ / 25;
  25 static const int cfq_slice_async_rq = 2;
  26 static int cfq_slice_idle = HZ / 125;
  27
  28 /*
  29  * grace period before allowing idle class to get disk access
  30  */
  31 #define CFQ_IDLE_GRACE          (HZ / 10)
  32
  33 /*
  34  * below this threshold, we consider thinktime immediate
  35  */
  36 #define CFQ_MIN_TT              (2)
  37
  38 #define CFQ_SLICE_SCALE         (5)
  39
  40 #define RQ_CIC(rq)              ((struct cfq_io_context*)(rq)->elevator_private)
  41 #define RQ_CFQQ(rq)             ((rq)->elevator_private2)
  42
  43 static struct kmem_cache *cfq_pool;
  44 static struct kmem_cache *cfq_ioc_pool;
  45
  46 static DEFINE_PER_CPU(unsigned long, ioc_count);
  47 static struct completion *ioc_gone;
  48
  49 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  50 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  51 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  52
  53 #define ASYNC                   (0)
  54 #define SYNC                    (1)
  55
  56 #define sample_valid(samples)   ((samples) > 80)
  57
  58 /*
  59  * Most of our rbtree usage is for sorting with min extraction, so
  60  * if we cache the leftmost node we don't have to walk down the tree
  61  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  62  * move this into the elevator for the rq sorting as well.
  63  */
  64 struct cfq_rb_root {
  65         struct rb_root rb;
  66         struct rb_node *left;
  67 };
  68 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, }
  69
  70 /*
  71  * Per block device queue structure
  72  */
  73 struct cfq_data {
  74         request_queue_t *queue;
  75
  76         /*
  77          * rr list of queues with requests and the count of them
  78          */
  79         struct cfq_rb_root service_tree;
  80         unsigned int busy_queues;
  81
  82         int rq_in_driver;
  83         int sync_flight;
  84         int hw_tag;
  85
  86         /*
  87          * idle window management
  88          */
  89         struct timer_list idle_slice_timer;
  90         struct work_struct unplug_work;
  91
  92         struct cfq_queue *active_queue;
  93         struct cfq_io_context *active_cic;
  94
  95         /*
  96          * async queue for each priority case
  97          */
  98         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
  99         struct cfq_queue *async_idle_cfqq;
 100
 101         struct timer_list idle_class_timer;
 102
 103         sector_t last_position;
 104         unsigned long last_end_request;
 105
 106         /*
 107          * tunables, see top of file
 108          */
 109         unsigned int cfq_quantum;
 110         unsigned int cfq_fifo_expire[2];
 111         unsigned int cfq_back_penalty;
 112         unsigned int cfq_back_max;
 113         unsigned int cfq_slice[2];
 114         unsigned int cfq_slice_async_rq;
 115         unsigned int cfq_slice_idle;
 116
 117         struct list_head cic_list;
 118 };
 119
 120 /*
 121  * Per process-grouping structure
 122  */
 123 struct cfq_queue {
 124         /* reference count */
 125         atomic_t ref;
 126         /* parent cfq_data */
 127         struct cfq_data *cfqd;
 128         /* service_tree member */
 129         struct rb_node rb_node;
 130         /* service_tree key */
 131         unsigned long rb_key;
 132         /* sorted list of pending requests */
 133         struct rb_root sort_list;
 134         /* if fifo isn't expired, next request to serve */
 135         struct request *next_rq;
 136         /* requests queued in sort_list */
 137         int queued[2];
 138         /* currently allocated requests */
 139         int allocated[2];
 140         /* pending metadata requests */
 141         int meta_pending;
 142         /* fifo list of requests in sort_list */
 143         struct list_head fifo;
 144
 145         unsigned long slice_end;
 146         long slice_resid;
 147
 148         /* number of requests that are on the dispatch list or inside driver */
 149         int dispatched;
 150
 151         /* io prio of this group */
 152         unsigned short ioprio, org_ioprio;
 153         unsigned short ioprio_class, org_ioprio_class;
 154
 155         /* various state flags, see below */
 156         unsigned int flags;
 157 };
 158
 159 enum cfqq_state_flags {
 160         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 161         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 162         CFQ_CFQQ_FLAG_must_alloc,       /* must be allowed rq alloc */
 163         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 164         CFQ_CFQQ_FLAG_must_dispatch,    /* must dispatch, even if expired */
 165         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 166         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 167         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 168         CFQ_CFQQ_FLAG_queue_new,        /* queue never been serviced */
 169         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 170         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 171 };
 172
 173 #define CFQ_CFQQ_FNS(name)                                              \
 174 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 175 {                                                                       \
 176         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 177 }                                                                       \
 178 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 179 {                                                                       \
 180         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 181 }                                                                       \
 182 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 183 {                                                                       \
 184         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 185 }
 186
 187 CFQ_CFQQ_FNS(on_rr);
 188 CFQ_CFQQ_FNS(wait_request);
 189 CFQ_CFQQ_FNS(must_alloc);
 190 CFQ_CFQQ_FNS(must_alloc_slice);
 191 CFQ_CFQQ_FNS(must_dispatch);
 192 CFQ_CFQQ_FNS(fifo_expire);
 193 CFQ_CFQQ_FNS(idle_window);
 194 CFQ_CFQQ_FNS(prio_changed);
 195 CFQ_CFQQ_FNS(queue_new);
 196 CFQ_CFQQ_FNS(slice_new);
 197 CFQ_CFQQ_FNS(sync);
 198 #undef CFQ_CFQQ_FNS
 199
 200 static void cfq_dispatch_insert(request_queue_t *, struct request *);
 201 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
 202                                        struct task_struct *, gfp_t);
 203 static struct cfq_io_context *cfq_cic_rb_lookup(struct cfq_data *,
 204                                                 struct io_context *);
 205
 206 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
 207                                             int is_sync)
 208 {
 209         return cic->cfqq[!!is_sync];
 210 }
 211
 212 static inline void cic_set_cfqq(struct cfq_io_context *cic,
 213                                 struct cfq_queue *cfqq, int is_sync)
 214 {
 215         cic->cfqq[!!is_sync] = cfqq;
 216 }
 217
 218 /*
 219  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 220  * set (in which case it could also be direct WRITE).
 221  */
 222 static inline int cfq_bio_sync(struct bio *bio)
 223 {
 224         if (bio_data_dir(bio) == READ || bio_sync(bio))
 225                 return 1;
 226
 227         return 0;
 228 }
 229
 230 /*
 231  * scheduler run of queue, if there are requests pending and no one in the
 232  * driver that will restart queueing
 233  */
 234 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 235 {
 236         if (cfqd->busy_queues)
 237                 kblockd_schedule_work(&cfqd->unplug_work);
 238 }
 239
 240 static int cfq_queue_empty(request_queue_t *q)
 241 {
 242         struct cfq_data *cfqd = q->elevator->elevator_data;
 243
 244         return !cfqd->busy_queues;
 245 }
 246
 247 /*
 248  * Scale schedule slice based on io priority. Use the sync time slice only
 249  * if a queue is marked sync and has sync io queued. A sync queue with async
 250  * io only, should not get full sync slice length.
 251  */
 252 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
 253                                  unsigned short prio)
 254 {
 255         const int base_slice = cfqd->cfq_slice[sync];
 256
 257         WARN_ON(prio >= IOPRIO_BE_NR);
 258
 259         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 260 }
 261
 262 static inline int
 263 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 264 {
 265         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 266 }
 267
 268 static inline void
 269 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 270 {
 271         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 272 }
 273
 274 /*
 275  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
 276  * isn't valid until the first request from the dispatch is activated
 277  * and the slice time set.
 278  */
 279 static inline int cfq_slice_used(struct cfq_queue *cfqq)
 280 {
 281         if (cfq_cfqq_slice_new(cfqq))
 282                 return 0;
 283         if (time_before(jiffies, cfqq->slice_end))
 284                 return 0;
 285
 286         return 1;
 287 }
 288
 289 /*
 290  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 291  * We choose the request that is closest to the head right now. Distance
 292  * behind the head is penalized and only allowed to a certain extent.
 293  */
 294 static struct request *
 295 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
 296 {
 297         sector_t last, s1, s2, d1 = 0, d2 = 0;
 298         unsigned long back_max;
 299 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 300 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 301         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 302
 303         if (rq1 == NULL || rq1 == rq2)
 304                 return rq2;
 305         if (rq2 == NULL)
 306                 return rq1;
 307
 308         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 309                 return rq1;
 310         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 311                 return rq2;
 312         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
 313                 return rq1;
 314         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
 315                 return rq2;
 316
 317         s1 = rq1->sector;
 318         s2 = rq2->sector;
 319
 320         last = cfqd->last_position;
 321
 322         /*
 323          * by definition, 1KiB is 2 sectors
 324          */
 325         back_max = cfqd->cfq_back_max * 2;
 326
 327         /*
 328          * Strict one way elevator _except_ in the case where we allow
 329          * short backward seeks which are biased as twice the cost of a
 330          * similar forward seek.
 331          */
 332         if (s1 >= last)
 333                 d1 = s1 - last;
 334         else if (s1 + back_max >= last)
 335                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 336         else
 337                 wrap |= CFQ_RQ1_WRAP;
 338
 339         if (s2 >= last)
 340                 d2 = s2 - last;
 341         else if (s2 + back_max >= last)
 342                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 343         else
 344                 wrap |= CFQ_RQ2_WRAP;
 345
 346         /* Found required data */
 347
 348         /*
 349          * By doing switch() on the bit mask "wrap" we avoid having to
 350          * check two variables for all permutations: --> faster!
 351          */
 352         switch (wrap) {
 353         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 354                 if (d1 < d2)
 355                         return rq1;
 356                 else if (d2 < d1)
 357                         return rq2;
 358                 else {
 359                         if (s1 >= s2)
 360                                 return rq1;
 361                         else
 362                                 return rq2;
 363                 }
 364
 365         case CFQ_RQ2_WRAP:
 366                 return rq1;
 367         case CFQ_RQ1_WRAP:
 368                 return rq2;
 369         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
 370         default:
 371                 /*
 372                  * Since both rqs are wrapped,
 373                  * start with the one that's further behind head
 374                  * (--> only *one* back seek required),
 375                  * since back seek takes more time than forward.
 376                  */
 377                 if (s1 <= s2)
 378                         return rq1;
 379                 else
 380                         return rq2;
 381         }
 382 }
 383
 384 /*
 385  * The below is leftmost cache rbtree addon
 386  */
 387 static struct rb_node *cfq_rb_first(struct cfq_rb_root *root)
 388 {
 389         if (!root->left)
 390                 root->left = rb_first(&root->rb);
 391
 392         return root->left;
 393 }
 394
 395 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
 396 {
 397         if (root->left == n)
 398                 root->left = NULL;
 399
 400         rb_erase(n, &root->rb);
 401         RB_CLEAR_NODE(n);
 402 }
 403
 404 /*
 405  * would be nice to take fifo expire time into account as well
 406  */
 407 static struct request *
 408 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 409                   struct request *last)
 410 {
 411         struct rb_node *rbnext = rb_next(&last->rb_node);
 412         struct rb_node *rbprev = rb_prev(&last->rb_node);
 413         struct request *next = NULL, *prev = NULL;
 414
 415         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
 416
 417         if (rbprev)
 418                 prev = rb_entry_rq(rbprev);
 419
 420         if (rbnext)
 421                 next = rb_entry_rq(rbnext);
 422         else {
 423                 rbnext = rb_first(&cfqq->sort_list);
 424                 if (rbnext && rbnext != &last->rb_node)
 425                         next = rb_entry_rq(rbnext);
 426         }
 427
 428         return cfq_choose_req(cfqd, next, prev);
 429 }
 430
 431 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
 432                                       struct cfq_queue *cfqq)
 433 {
 434         /*
 435          * just an approximation, should be ok.
 436          */
 437         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
 438                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
 439 }
 440
 441 /*
 442  * The cfqd->service_tree holds all pending cfq_queue's that have
 443  * requests waiting to be processed. It is sorted in the order that
 444  * we will service the queues.
 445  */
 446 static void cfq_service_tree_add(struct cfq_data *cfqd,
 447                                     struct cfq_queue *cfqq, int add_front)
 448 {
 449         struct rb_node **p = &cfqd->service_tree.rb.rb_node;
 450         struct rb_node *parent = NULL;
 451         unsigned long rb_key;
 452         int left;
 453
 454         if (!add_front) {
 455                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
 456                 rb_key += cfqq->slice_resid;
 457                 cfqq->slice_resid = 0;
 458         } else
 459                 rb_key = 0;
 460
 461         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
 462                 /*
 463                  * same position, nothing more to do
 464                  */
 465                 if (rb_key == cfqq->rb_key)
 466                         return;
 467
 468                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
 469         }
 470
 471         left = 1;
 472         while (*p) {
 473                 struct cfq_queue *__cfqq;
 474                 struct rb_node **n;
 475
 476                 parent = *p;
 477                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
 478
 479                 /*
 480                  * sort RT queues first, we always want to give
 481                  * preference to them. IDLE queues goes to the back.
 482                  * after that, sort on the next service time.
 483                  */
 484                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
 485                         n = &(*p)->rb_left;
 486                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
 487                         n = &(*p)->rb_right;
 488                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
 489                         n = &(*p)->rb_left;
 490                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
 491                         n = &(*p)->rb_right;
 492                 else if (rb_key < __cfqq->rb_key)
 493                         n = &(*p)->rb_left;
 494                 else
 495                         n = &(*p)->rb_right;
 496
 497                 if (n == &(*p)->rb_right)
 498                         left = 0;
 499
 500                 p = n;
 501         }
 502
 503         if (left)
 504                 cfqd->service_tree.left = &cfqq->rb_node;
 505
 506         cfqq->rb_key = rb_key;
 507         rb_link_node(&cfqq->rb_node, parent, p);
 508         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
 509 }
 510
 511 /*
 512  * Update cfqq's position in the service tree.
 513  */
 514 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 515 {
 516         /*
 517          * Resorting requires the cfqq to be on the RR list already.
 518          */
 519         if (cfq_cfqq_on_rr(cfqq))
 520                 cfq_service_tree_add(cfqd, cfqq, 0);
 521 }
 522
 523 /*
 524  * add to busy list of queues for service, trying to be fair in ordering
 525  * the pending list according to last request service
 526  */
 527 static inline void
 528 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 529 {
 530         BUG_ON(cfq_cfqq_on_rr(cfqq));
 531         cfq_mark_cfqq_on_rr(cfqq);
 532         cfqd->busy_queues++;
 533
 534         cfq_resort_rr_list(cfqd, cfqq);
 535 }
 536
 537 /*
 538  * Called when the cfqq no longer has requests pending, remove it from
 539  * the service tree.
 540  */
 541 static inline void
 542 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 543 {
 544         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 545         cfq_clear_cfqq_on_rr(cfqq);
 546
 547         if (!RB_EMPTY_NODE(&cfqq->rb_node))
 548                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
 549
 550         BUG_ON(!cfqd->busy_queues);
 551         cfqd->busy_queues--;
 552 }
 553
 554 /*
 555  * rb tree support functions
 556  */
 557 static inline void cfq_del_rq_rb(struct request *rq)
 558 {
 559         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 560         struct cfq_data *cfqd = cfqq->cfqd;
 561         const int sync = rq_is_sync(rq);
 562
 563         BUG_ON(!cfqq->queued[sync]);
 564         cfqq->queued[sync]--;
 565
 566         elv_rb_del(&cfqq->sort_list, rq);
 567
 568         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
 569                 cfq_del_cfqq_rr(cfqd, cfqq);
 570 }
 571
 572 static void cfq_add_rq_rb(struct request *rq)
 573 {
 574         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 575         struct cfq_data *cfqd = cfqq->cfqd;
 576         struct request *__alias;
 577
 578         cfqq->queued[rq_is_sync(rq)]++;
 579
 580         /*
 581          * looks a little odd, but the first insert might return an alias.
 582          * if that happens, put the alias on the dispatch list
 583          */
 584         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
 585                 cfq_dispatch_insert(cfqd->queue, __alias);
 586
 587         if (!cfq_cfqq_on_rr(cfqq))
 588                 cfq_add_cfqq_rr(cfqd, cfqq);
 589
 590         /*
 591          * check if this request is a better next-serve candidate
 592          */
 593         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
 594         BUG_ON(!cfqq->next_rq);
 595 }
 596
 597 static inline void
 598 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
 599 {
 600         elv_rb_del(&cfqq->sort_list, rq);
 601         cfqq->queued[rq_is_sync(rq)]--;
 602         cfq_add_rq_rb(rq);
 603 }
 604
 605 static struct request *
 606 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 607 {
 608         struct task_struct *tsk = current;
 609         struct cfq_io_context *cic;
 610         struct cfq_queue *cfqq;
 611
 612         cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
 613         if (!cic)
 614                 return NULL;
 615
 616         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
 617         if (cfqq) {
 618                 sector_t sector = bio->bi_sector + bio_sectors(bio);
 619
 620                 return elv_rb_find(&cfqq->sort_list, sector);
 621         }
 622
 623         return NULL;
 624 }
 625
 626 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 627 {
 628         struct cfq_data *cfqd = q->elevator->elevator_data;
 629
 630         cfqd->rq_in_driver++;
 631
 632         /*
 633          * If the depth is larger 1, it really could be queueing. But lets
 634          * make the mark a little higher - idling could still be good for
 635          * low queueing, and a low queueing number could also just indicate
 636          * a SCSI mid layer like behaviour where limit+1 is often seen.
 637          */
 638         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
 639                 cfqd->hw_tag = 1;
 640
 641         cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
 642 }
 643
 644 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 645 {
 646         struct cfq_data *cfqd = q->elevator->elevator_data;
 647
 648         WARN_ON(!cfqd->rq_in_driver);
 649         cfqd->rq_in_driver--;
 650 }
 651
 652 static void cfq_remove_request(struct request *rq)
 653 {
 654         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 655
 656         if (cfqq->next_rq == rq)
 657                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
 658
 659         list_del_init(&rq->queuelist);
 660         cfq_del_rq_rb(rq);
 661
 662         if (rq_is_meta(rq)) {
 663                 WARN_ON(!cfqq->meta_pending);
 664                 cfqq->meta_pending--;
 665         }
 666 }
 667
 668 static int cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 669 {
 670         struct cfq_data *cfqd = q->elevator->elevator_data;
 671         struct request *__rq;
 672
 673         __rq = cfq_find_rq_fmerge(cfqd, bio);
 674         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 675                 *req = __rq;
 676                 return ELEVATOR_FRONT_MERGE;
 677         }
 678
 679         return ELEVATOR_NO_MERGE;
 680 }
 681
 682 static void cfq_merged_request(request_queue_t *q, struct request *req,
 683                                int type)
 684 {
 685         if (type == ELEVATOR_FRONT_MERGE) {
 686                 struct cfq_queue *cfqq = RQ_CFQQ(req);
 687
 688                 cfq_reposition_rq_rb(cfqq, req);
 689         }
 690 }
 691
 692 static void
 693 cfq_merged_requests(request_queue_t *q, struct request *rq,
 694                     struct request *next)
 695 {
 696         /*
 697          * reposition in fifo if next is older than rq
 698          */
 699         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 700             time_before(next->start_time, rq->start_time))
 701                 list_move(&rq->queuelist, &next->queuelist);
 702
 703         cfq_remove_request(next);
 704 }
 705
 706 static int cfq_allow_merge(request_queue_t *q, struct request *rq,
 707                            struct bio *bio)
 708 {
 709         struct cfq_data *cfqd = q->elevator->elevator_data;
 710         struct cfq_io_context *cic;
 711         struct cfq_queue *cfqq;
 712
 713         /*
 714          * Disallow merge of a sync bio into an async request.
 715          */
 716         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
 717                 return 0;
 718
 719         /*
 720          * Lookup the cfqq that this bio will be queued with. Allow
 721          * merge only if rq is queued there.
 722          */
 723         cic = cfq_cic_rb_lookup(cfqd, current->io_context);
 724         if (!cic)
 725                 return 0;
 726
 727         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
 728         if (cfqq == RQ_CFQQ(rq))
 729                 return 1;
 730
 731         return 0;
 732 }
 733
 734 static inline void
 735 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 736 {
 737         if (cfqq) {
 738                 /*
 739                  * stop potential idle class queues waiting service
 740                  */
 741                 del_timer(&cfqd->idle_class_timer);
 742
 743                 cfqq->slice_end = 0;
 744                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 745                 cfq_clear_cfqq_fifo_expire(cfqq);
 746                 cfq_mark_cfqq_slice_new(cfqq);
 747                 cfq_clear_cfqq_queue_new(cfqq);
 748         }
 749
 750         cfqd->active_queue = cfqq;
 751 }
 752
 753 /*
 754  * current cfqq expired its slice (or was too idle), select new one
 755  */
 756 static void
 757 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 758                     int timed_out)
 759 {
 760         if (cfq_cfqq_wait_request(cfqq))
 761                 del_timer(&cfqd->idle_slice_timer);
 762
 763         cfq_clear_cfqq_must_dispatch(cfqq);
 764         cfq_clear_cfqq_wait_request(cfqq);
 765
 766         /*
 767          * store what was left of this slice, if the queue idled/timed out
 768          */
 769         if (timed_out && !cfq_cfqq_slice_new(cfqq))
 770                 cfqq->slice_resid = cfqq->slice_end - jiffies;
 771
 772         cfq_resort_rr_list(cfqd, cfqq);
 773
 774         if (cfqq == cfqd->active_queue)
 775                 cfqd->active_queue = NULL;
 776
 777         if (cfqd->active_cic) {
 778                 put_io_context(cfqd->active_cic->ioc);
 779                 cfqd->active_cic = NULL;
 780         }
 781 }
 782
 783 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
 784 {
 785         struct cfq_queue *cfqq = cfqd->active_queue;
 786
 787         if (cfqq)
 788                 __cfq_slice_expired(cfqd, cfqq, timed_out);
 789 }
 790
 791 /*
 792  * Get next queue for service. Unless we have a queue preemption,
 793  * we'll simply select the first cfqq in the service tree.
 794  */
 795 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
 796 {
 797         struct cfq_queue *cfqq;
 798         struct rb_node *n;
 799
 800         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
 801                 return NULL;
 802
 803         n = cfq_rb_first(&cfqd->service_tree);
 804         cfqq = rb_entry(n, struct cfq_queue, rb_node);
 805
 806         if (cfq_class_idle(cfqq)) {
 807                 unsigned long end;
 808
 809                 /*
 810                  * if we have idle queues and no rt or be queues had
 811                  * pending requests, either allow immediate service if
 812                  * the grace period has passed or arm the idle grace
 813                  * timer
 814                  */
 815                 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 816                 if (time_before(jiffies, end)) {
 817                         mod_timer(&cfqd->idle_class_timer, end);
 818                         cfqq = NULL;
 819                 }
 820         }
 821
 822         return cfqq;
 823 }
 824
 825 /*
 826  * Get and set a new active queue for service.
 827  */
 828 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 829 {
 830         struct cfq_queue *cfqq;
 831
 832         cfqq = cfq_get_next_queue(cfqd);
 833         __cfq_set_active_queue(cfqd, cfqq);
 834         return cfqq;
 835 }
 836
 837 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
 838                                           struct request *rq)
 839 {
 840         if (rq->sector >= cfqd->last_position)
 841                 return rq->sector - cfqd->last_position;
 842         else
 843                 return cfqd->last_position - rq->sector;
 844 }
 845
 846 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
 847 {
 848         struct cfq_io_context *cic = cfqd->active_cic;
 849
 850         if (!sample_valid(cic->seek_samples))
 851                 return 0;
 852
 853         return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
 854 }
 855
 856 static int cfq_close_cooperator(struct cfq_data *cfq_data,
 857                                 struct cfq_queue *cfqq)
 858 {
 859         /*
 860          * We should notice if some of the queues are cooperating, eg
 861          * working closely on the same area of the disk. In that case,
 862          * we can group them together and don't waste time idling.
 863          */
 864         return 0;
 865 }
 866
 867 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
 868
 869 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
 870 {
 871         struct cfq_queue *cfqq = cfqd->active_queue;
 872         struct cfq_io_context *cic;
 873         unsigned long sl;
 874
 875         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
 876         WARN_ON(cfq_cfqq_slice_new(cfqq));
 877
 878         /*
 879          * idle is disabled, either manually or by past process history
 880          */
 881         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
 882                 return;
 883
 884         /*
 885          * task has exited, don't wait
 886          */
 887         cic = cfqd->active_cic;
 888         if (!cic || !cic->ioc->task)
 889                 return;
 890
 891         /*
 892          * See if this prio level has a good candidate
 893          */
 894         if (cfq_close_cooperator(cfqd, cfqq) &&
 895             (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
 896                 return;
 897
 898         cfq_mark_cfqq_must_dispatch(cfqq);
 899         cfq_mark_cfqq_wait_request(cfqq);
 900
 901         /*
 902          * we don't want to idle for seeks, but we do want to allow
 903          * fair distribution of slice time for a process doing back-to-back
 904          * seeks. so allow a little bit of time for him to submit a new rq
 905          */
 906         sl = cfqd->cfq_slice_idle;
 907         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
 908                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
 909
 910         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 911 }
 912
 913 /*
 914  * Move request from internal lists to the request queue dispatch list.
 915  */
 916 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
 917 {
 918         struct cfq_data *cfqd = q->elevator->elevator_data;
 919         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 920
 921         cfq_remove_request(rq);
 922         cfqq->dispatched++;
 923         elv_dispatch_sort(q, rq);
 924
 925         if (cfq_cfqq_sync(cfqq))
 926                 cfqd->sync_flight++;
 927 }
 928
 929 /*
 930  * return expired entry, or NULL to just start from scratch in rbtree
 931  */
 932 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
 933 {
 934         struct cfq_data *cfqd = cfqq->cfqd;
 935         struct request *rq;
 936         int fifo;
 937
 938         if (cfq_cfqq_fifo_expire(cfqq))
 939                 return NULL;
 940
 941         cfq_mark_cfqq_fifo_expire(cfqq);
 942
 943         if (list_empty(&cfqq->fifo))
 944                 return NULL;
 945
 946         fifo = cfq_cfqq_sync(cfqq);
 947         rq = rq_entry_fifo(cfqq->fifo.next);
 948
 949         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
 950                 return NULL;
 951
 952         return rq;
 953 }
 954
 955 static inline int
 956 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 957 {
 958         const int base_rq = cfqd->cfq_slice_async_rq;
 959
 960         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 961
 962         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
 963 }
 964
 965 /*
 966  * Select a queue for service. If we have a current active queue,
 967  * check whether to continue servicing it, or retrieve and set a new one.
 968  */
 969 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
 970 {
 971         struct cfq_queue *cfqq;
 972
 973         cfqq = cfqd->active_queue;
 974         if (!cfqq)
 975                 goto new_queue;
 976
 977         /*
 978          * The active queue has run out of time, expire it and select new.
 979          */
 980         if (cfq_slice_used(cfqq))
 981                 goto expire;
 982
 983         /*
 984          * The active queue has requests and isn't expired, allow it to
 985          * dispatch.
 986          */
 987         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
 988                 goto keep_queue;
 989
 990         /*
 991          * No requests pending. If the active queue still has requests in
 992          * flight or is idling for a new request, allow either of these
 993          * conditions to happen (or time out) before selecting a new queue.
 994          */
 995         if (timer_pending(&cfqd->idle_slice_timer) ||
 996             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
 997                 cfqq = NULL;
 998                 goto keep_queue;
 999         }
1000
1001 expire:
1002         cfq_slice_expired(cfqd, 0);
1003 new_queue:
1004         cfqq = cfq_set_active_queue(cfqd);
1005 keep_queue:
1006         return cfqq;
1007 }
1008
1009 /*
1010  * Dispatch some requests from cfqq, moving them to the request queue
1011  * dispatch list.
1012  */
1013 static int
1014 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1015                         int max_dispatch)
1016 {
1017         int dispatched = 0;
1018
1019         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1020
1021         do {
1022                 struct request *rq;
1023
1024                 /*
1025                  * follow expired path, else get first next available
1026                  */
1027                 if ((rq = cfq_check_fifo(cfqq)) == NULL)
1028                         rq = cfqq->next_rq;
1029
1030                 /*
1031                  * finally, insert request into driver dispatch list
1032                  */
1033                 cfq_dispatch_insert(cfqd->queue, rq);
1034
1035                 dispatched++;
1036
1037                 if (!cfqd->active_cic) {
1038                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1039                         cfqd->active_cic = RQ_CIC(rq);
1040                 }
1041
1042                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1043                         break;
1044
1045         } while (dispatched < max_dispatch);
1046
1047         /*
1048          * expire an async queue immediately if it has used up its slice. idle
1049          * queue always expire after 1 dispatch round.
1050          */
1051         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1052             dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1053             cfq_class_idle(cfqq))) {
1054                 cfqq->slice_end = jiffies + 1;
1055                 cfq_slice_expired(cfqd, 0);
1056         }
1057
1058         return dispatched;
1059 }
1060
1061 static inline int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1062 {
1063         int dispatched = 0;
1064
1065         while (cfqq->next_rq) {
1066                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1067                 dispatched++;
1068         }
1069
1070         BUG_ON(!list_empty(&cfqq->fifo));
1071         return dispatched;
1072 }
1073
1074 /*
1075  * Drain our current requests. Used for barriers and when switching
1076  * io schedulers on-the-fly.
1077  */
1078 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1079 {
1080         int dispatched = 0;
1081         struct rb_node *n;
1082
1083         while ((n = cfq_rb_first(&cfqd->service_tree)) != NULL) {
1084                 struct cfq_queue *cfqq = rb_entry(n, struct cfq_queue, rb_node);
1085
1086                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1087         }
1088
1089         cfq_slice_expired(cfqd, 0);
1090
1091         BUG_ON(cfqd->busy_queues);
1092
1093         return dispatched;
1094 }
1095
1096 static int cfq_dispatch_requests(request_queue_t *q, int force)
1097 {
1098         struct cfq_data *cfqd = q->elevator->elevator_data;
1099         struct cfq_queue *cfqq;
1100         int dispatched;
1101
1102         if (!cfqd->busy_queues)
1103                 return 0;
1104
1105         if (unlikely(force))
1106                 return cfq_forced_dispatch(cfqd);
1107
1108         dispatched = 0;
1109         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1110                 int max_dispatch;
1111
1112                 max_dispatch = cfqd->cfq_quantum;
1113                 if (cfq_class_idle(cfqq))
1114                         max_dispatch = 1;
1115
1116                 if (cfqq->dispatched >= max_dispatch) {
1117                         if (cfqd->busy_queues > 1)
1118                                 break;
1119                         if (cfqq->dispatched >= 4 * max_dispatch)
1120                                 break;
1121                 }
1122
1123                 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1124                         break;
1125
1126                 cfq_clear_cfqq_must_dispatch(cfqq);
1127                 cfq_clear_cfqq_wait_request(cfqq);
1128                 del_timer(&cfqd->idle_slice_timer);
1129
1130                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1131         }
1132
1133         return dispatched;
1134 }
1135
1136 /*
1137  * task holds one reference to the queue, dropped when task exits. each rq
1138  * in-flight on this queue also holds a reference, dropped when rq is freed.
1139  *
1140  * queue lock must be held here.
1141  */
1142 static void cfq_put_queue(struct cfq_queue *cfqq)
1143 {
1144         struct cfq_data *cfqd = cfqq->cfqd;
1145
1146         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1147
1148         if (!atomic_dec_and_test(&cfqq->ref))
1149                 return;
1150
1151         BUG_ON(rb_first(&cfqq->sort_list));
1152         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1153         BUG_ON(cfq_cfqq_on_rr(cfqq));
1154
1155         if (unlikely(cfqd->active_queue == cfqq)) {
1156                 __cfq_slice_expired(cfqd, cfqq, 0);
1157                 cfq_schedule_dispatch(cfqd);
1158         }
1159
1160         kmem_cache_free(cfq_pool, cfqq);
1161 }
1162
1163 static void cfq_free_io_context(struct io_context *ioc)
1164 {
1165         struct cfq_io_context *__cic;
1166         struct rb_node *n;
1167         int freed = 0;
1168
1169         ioc->ioc_data = NULL;
1170
1171         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1172                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1173                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1174                 kmem_cache_free(cfq_ioc_pool, __cic);
1175                 freed++;
1176         }
1177
1178         elv_ioc_count_mod(ioc_count, -freed);
1179
1180         if (ioc_gone && !elv_ioc_count_read(ioc_count))
1181                 complete(ioc_gone);
1182 }
1183
1184 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1185 {
1186         if (unlikely(cfqq == cfqd->active_queue)) {
1187                 __cfq_slice_expired(cfqd, cfqq, 0);
1188                 cfq_schedule_dispatch(cfqd);
1189         }
1190
1191         cfq_put_queue(cfqq);
1192 }
1193
1194 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1195                                          struct cfq_io_context *cic)
1196 {
1197         list_del_init(&cic->queue_list);
1198         smp_wmb();
1199         cic->key = NULL;
1200
1201         if (cic->cfqq[ASYNC]) {
1202                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1203                 cic->cfqq[ASYNC] = NULL;
1204         }
1205
1206         if (cic->cfqq[SYNC]) {
1207                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1208                 cic->cfqq[SYNC] = NULL;
1209         }
1210 }
1211
1212 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1213 {
1214         struct cfq_data *cfqd = cic->key;
1215
1216         if (cfqd) {
1217                 request_queue_t *q = cfqd->queue;
1218
1219                 spin_lock_irq(q->queue_lock);
1220                 __cfq_exit_single_io_context(cfqd, cic);
1221                 spin_unlock_irq(q->queue_lock);
1222         }
1223 }
1224
1225 /*
1226  * The process that ioc belongs to has exited, we need to clean up
1227  * and put the internal structures we have that belongs to that process.
1228  */
1229 static void cfq_exit_io_context(struct io_context *ioc)
1230 {
1231         struct cfq_io_context *__cic;
1232         struct rb_node *n;
1233
1234         ioc->ioc_data = NULL;
1235
1236         /*
1237          * put the reference this task is holding to the various queues
1238          */
1239         n = rb_first(&ioc->cic_root);
1240         while (n != NULL) {
1241                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1242
1243                 cfq_exit_single_io_context(__cic);
1244                 n = rb_next(n);
1245         }
1246 }
1247
1248 static struct cfq_io_context *
1249 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1250 {
1251         struct cfq_io_context *cic;
1252
1253         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1254                                                         cfqd->queue->node);
1255         if (cic) {
1256                 cic->last_end_request = jiffies;
1257                 INIT_LIST_HEAD(&cic->queue_list);
1258                 cic->dtor = cfq_free_io_context;
1259                 cic->exit = cfq_exit_io_context;
1260                 elv_ioc_count_inc(ioc_count);
1261         }
1262
1263         return cic;
1264 }
1265
1266 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1267 {
1268         struct task_struct *tsk = current;
1269         int ioprio_class;
1270
1271         if (!cfq_cfqq_prio_changed(cfqq))
1272                 return;
1273
1274         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1275         switch (ioprio_class) {
1276                 default:
1277                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1278                 case IOPRIO_CLASS_NONE:
1279                         /*
1280                          * no prio set, place us in the middle of the BE classes
1281                          */
1282                         cfqq->ioprio = task_nice_ioprio(tsk);
1283                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1284                         break;
1285                 case IOPRIO_CLASS_RT:
1286                         cfqq->ioprio = task_ioprio(tsk);
1287                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1288                         break;
1289                 case IOPRIO_CLASS_BE:
1290                         cfqq->ioprio = task_ioprio(tsk);
1291                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1292                         break;
1293                 case IOPRIO_CLASS_IDLE:
1294                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1295                         cfqq->ioprio = 7;
1296                         cfq_clear_cfqq_idle_window(cfqq);
1297                         break;
1298         }
1299
1300         /*
1301          * keep track of original prio settings in case we have to temporarily
1302          * elevate the priority of this queue
1303          */
1304         cfqq->org_ioprio = cfqq->ioprio;
1305         cfqq->org_ioprio_class = cfqq->ioprio_class;
1306         cfq_clear_cfqq_prio_changed(cfqq);
1307 }
1308
1309 static inline void changed_ioprio(struct cfq_io_context *cic)
1310 {
1311         struct cfq_data *cfqd = cic->key;
1312         struct cfq_queue *cfqq;
1313         unsigned long flags;
1314
1315         if (unlikely(!cfqd))
1316                 return;
1317
1318         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1319
1320         cfqq = cic->cfqq[ASYNC];
1321         if (cfqq) {
1322                 struct cfq_queue *new_cfqq;
1323                 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc->task,
1324                                          GFP_ATOMIC);
1325                 if (new_cfqq) {
1326                         cic->cfqq[ASYNC] = new_cfqq;
1327                         cfq_put_queue(cfqq);
1328                 }
1329         }
1330
1331         cfqq = cic->cfqq[SYNC];
1332         if (cfqq)
1333                 cfq_mark_cfqq_prio_changed(cfqq);
1334
1335         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1336 }
1337
1338 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1339 {
1340         struct cfq_io_context *cic;
1341         struct rb_node *n;
1342
1343         ioc->ioprio_changed = 0;
1344
1345         n = rb_first(&ioc->cic_root);
1346         while (n != NULL) {
1347                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1348
1349                 changed_ioprio(cic);
1350                 n = rb_next(n);
1351         }
1352 }
1353
1354 static struct cfq_queue *
1355 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1356                      struct task_struct *tsk, gfp_t gfp_mask)
1357 {
1358         struct cfq_queue *cfqq, *new_cfqq = NULL;
1359         struct cfq_io_context *cic;
1360
1361 retry:
1362         cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
1363         /* cic always exists here */
1364         cfqq = cic_to_cfqq(cic, is_sync);
1365
1366         if (!cfqq) {
1367                 if (new_cfqq) {
1368                         cfqq = new_cfqq;
1369                         new_cfqq = NULL;
1370                 } else if (gfp_mask & __GFP_WAIT) {
1371                         /*
1372                          * Inform the allocator of the fact that we will
1373                          * just repeat this allocation if it fails, to allow
1374                          * the allocator to do whatever it needs to attempt to
1375                          * free memory.
1376                          */
1377                         spin_unlock_irq(cfqd->queue->queue_lock);
1378                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1379                                         gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1380                                         cfqd->queue->node);
1381                         spin_lock_irq(cfqd->queue->queue_lock);
1382                         goto retry;
1383                 } else {
1384                         cfqq = kmem_cache_alloc_node(cfq_pool,
1385                                         gfp_mask | __GFP_ZERO,
1386                                         cfqd->queue->node);
1387                         if (!cfqq)
1388                                 goto out;
1389                 }
1390
1391                 RB_CLEAR_NODE(&cfqq->rb_node);
1392                 INIT_LIST_HEAD(&cfqq->fifo);
1393
1394                 atomic_set(&cfqq->ref, 0);
1395                 cfqq->cfqd = cfqd;
1396
1397                 if (is_sync) {
1398                         cfq_mark_cfqq_idle_window(cfqq);
1399                         cfq_mark_cfqq_sync(cfqq);
1400                 }
1401
1402                 cfq_mark_cfqq_prio_changed(cfqq);
1403                 cfq_mark_cfqq_queue_new(cfqq);
1404
1405                 cfq_init_prio_data(cfqq);
1406         }
1407
1408         if (new_cfqq)
1409                 kmem_cache_free(cfq_pool, new_cfqq);
1410
1411 out:
1412         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1413         return cfqq;
1414 }
1415
1416 static struct cfq_queue **
1417 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1418 {
1419         switch(ioprio_class) {
1420         case IOPRIO_CLASS_RT:
1421                 return &cfqd->async_cfqq[0][ioprio];
1422         case IOPRIO_CLASS_BE:
1423                 return &cfqd->async_cfqq[1][ioprio];
1424         case IOPRIO_CLASS_IDLE:
1425                 return &cfqd->async_idle_cfqq;
1426         default:
1427                 BUG();
1428         }
1429 }
1430
1431 static struct cfq_queue *
1432 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct task_struct *tsk,
1433               gfp_t gfp_mask)
1434 {
1435         const int ioprio = task_ioprio(tsk);
1436         const int ioprio_class = task_ioprio_class(tsk);
1437         struct cfq_queue **async_cfqq = NULL;
1438         struct cfq_queue *cfqq = NULL;
1439
1440         if (!is_sync) {
1441                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1442                 cfqq = *async_cfqq;
1443         }
1444
1445         if (!cfqq)
1446                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, tsk, gfp_mask);
1447
1448         /*
1449          * pin the queue now that it's allocated, scheduler exit will prune it
1450          */
1451         if (!is_sync && !(*async_cfqq)) {
1452                 atomic_inc(&cfqq->ref);
1453                 *async_cfqq = cfqq;
1454         }
1455
1456         atomic_inc(&cfqq->ref);
1457         return cfqq;
1458 }
1459
1460 /*
1461  * We drop cfq io contexts lazily, so we may find a dead one.
1462  */
1463 static void
1464 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1465 {
1466         WARN_ON(!list_empty(&cic->queue_list));
1467
1468         if (ioc->ioc_data == cic)
1469                 ioc->ioc_data = NULL;
1470
1471         rb_erase(&cic->rb_node, &ioc->cic_root);
1472         kmem_cache_free(cfq_ioc_pool, cic);
1473         elv_ioc_count_dec(ioc_count);
1474 }
1475
1476 static struct cfq_io_context *
1477 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1478 {
1479         struct rb_node *n;
1480         struct cfq_io_context *cic;
1481         void *k, *key = cfqd;
1482
1483         if (unlikely(!ioc))
1484                 return NULL;
1485
1486         /*
1487          * we maintain a last-hit cache, to avoid browsing over the tree
1488          */
1489         cic = ioc->ioc_data;
1490         if (cic && cic->key == cfqd)
1491                 return cic;
1492
1493 restart:
1494         n = ioc->cic_root.rb_node;
1495         while (n) {
1496                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1497                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1498                 k = cic->key;
1499                 if (unlikely(!k)) {
1500                         cfq_drop_dead_cic(ioc, cic);
1501                         goto restart;
1502                 }
1503
1504                 if (key < k)
1505                         n = n->rb_left;
1506                 else if (key > k)
1507                         n = n->rb_right;
1508                 else {
1509                         ioc->ioc_data = cic;
1510                         return cic;
1511                 }
1512         }
1513
1514         return NULL;
1515 }
1516
1517 static inline void
1518 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1519              struct cfq_io_context *cic)
1520 {
1521         struct rb_node **p;
1522         struct rb_node *parent;
1523         struct cfq_io_context *__cic;
1524         unsigned long flags;
1525         void *k;
1526
1527         cic->ioc = ioc;
1528         cic->key = cfqd;
1529
1530 restart:
1531         parent = NULL;
1532         p = &ioc->cic_root.rb_node;
1533         while (*p) {
1534                 parent = *p;
1535                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1536                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1537                 k = __cic->key;
1538                 if (unlikely(!k)) {
1539                         cfq_drop_dead_cic(ioc, __cic);
1540                         goto restart;
1541                 }
1542
1543                 if (cic->key < k)
1544                         p = &(*p)->rb_left;
1545                 else if (cic->key > k)
1546                         p = &(*p)->rb_right;
1547                 else
1548                         BUG();
1549         }
1550
1551         rb_link_node(&cic->rb_node, parent, p);
1552         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1553
1554         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1555         list_add(&cic->queue_list, &cfqd->cic_list);
1556         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1557 }
1558
1559 /*
1560  * Setup general io context and cfq io context. There can be several cfq
1561  * io contexts per general io context, if this process is doing io to more
1562  * than one device managed by cfq.
1563  */
1564 static struct cfq_io_context *
1565 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1566 {
1567         struct io_context *ioc = NULL;
1568         struct cfq_io_context *cic;
1569
1570         might_sleep_if(gfp_mask & __GFP_WAIT);
1571
1572         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1573         if (!ioc)
1574                 return NULL;
1575
1576         cic = cfq_cic_rb_lookup(cfqd, ioc);
1577         if (cic)
1578                 goto out;
1579
1580         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1581         if (cic == NULL)
1582                 goto err;
1583
1584         cfq_cic_link(cfqd, ioc, cic);
1585 out:
1586         smp_read_barrier_depends();
1587         if (unlikely(ioc->ioprio_changed))
1588                 cfq_ioc_set_ioprio(ioc);
1589
1590         return cic;
1591 err:
1592         put_io_context(ioc);
1593         return NULL;
1594 }
1595
1596 static void
1597 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1598 {
1599         unsigned long elapsed = jiffies - cic->last_end_request;
1600         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1601
1602         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1603         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1604         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1605 }
1606
1607 static void
1608 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1609                        struct request *rq)
1610 {
1611         sector_t sdist;
1612         u64 total;
1613
1614         if (cic->last_request_pos < rq->sector)
1615                 sdist = rq->sector - cic->last_request_pos;
1616         else
1617                 sdist = cic->last_request_pos - rq->sector;
1618
1619         /*
1620          * Don't allow the seek distance to get too large from the
1621          * odd fragment, pagein, etc
1622          */
1623         if (cic->seek_samples <= 60) /* second&third seek */
1624                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1625         else
1626                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1627
1628         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1629         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1630         total = cic->seek_total + (cic->seek_samples/2);
1631         do_div(total, cic->seek_samples);
1632         cic->seek_mean = (sector_t)total;
1633 }
1634
1635 /*
1636  * Disable idle window if the process thinks too long or seeks so much that
1637  * it doesn't matter
1638  */
1639 static void
1640 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1641                        struct cfq_io_context *cic)
1642 {
1643         int enable_idle;
1644
1645         if (!cfq_cfqq_sync(cfqq))
1646                 return;
1647
1648         enable_idle = cfq_cfqq_idle_window(cfqq);
1649
1650         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1651             (cfqd->hw_tag && CIC_SEEKY(cic)))
1652                 enable_idle = 0;
1653         else if (sample_valid(cic->ttime_samples)) {
1654                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1655                         enable_idle = 0;
1656                 else
1657                         enable_idle = 1;
1658         }
1659
1660         if (enable_idle)
1661                 cfq_mark_cfqq_idle_window(cfqq);
1662         else
1663                 cfq_clear_cfqq_idle_window(cfqq);
1664 }
1665
1666 /*
1667  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1668  * no or if we aren't sure, a 1 will cause a preempt.
1669  */
1670 static int
1671 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1672                    struct request *rq)
1673 {
1674         struct cfq_queue *cfqq;
1675
1676         cfqq = cfqd->active_queue;
1677         if (!cfqq)
1678                 return 0;
1679
1680         if (cfq_slice_used(cfqq))
1681                 return 1;
1682
1683         if (cfq_class_idle(new_cfqq))
1684                 return 0;
1685
1686         if (cfq_class_idle(cfqq))
1687                 return 1;
1688
1689         /*
1690          * if the new request is sync, but the currently running queue is
1691          * not, let the sync request have priority.
1692          */
1693         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1694                 return 1;
1695
1696         /*
1697          * So both queues are sync. Let the new request get disk time if
1698          * it's a metadata request and the current queue is doing regular IO.
1699          */
1700         if (rq_is_meta(rq) && !cfqq->meta_pending)
1701                 return 1;
1702
1703         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1704                 return 0;
1705
1706         /*
1707          * if this request is as-good as one we would expect from the
1708          * current cfqq, let it preempt
1709          */
1710         if (cfq_rq_close(cfqd, rq))
1711                 return 1;
1712
1713         return 0;
1714 }
1715
1716 /*
1717  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1718  * let it have half of its nominal slice.
1719  */
1720 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1721 {
1722         cfq_slice_expired(cfqd, 1);
1723
1724         /*
1725          * Put the new queue at the front of the of the current list,
1726          * so we know that it will be selected next.
1727          */
1728         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1729
1730         cfq_service_tree_add(cfqd, cfqq, 1);
1731
1732         cfqq->slice_end = 0;
1733         cfq_mark_cfqq_slice_new(cfqq);
1734 }
1735
1736 /*
1737  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1738  * something we should do about it
1739  */
1740 static void
1741 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1742                 struct request *rq)
1743 {
1744         struct cfq_io_context *cic = RQ_CIC(rq);
1745
1746         if (rq_is_meta(rq))
1747                 cfqq->meta_pending++;
1748
1749         cfq_update_io_thinktime(cfqd, cic);
1750         cfq_update_io_seektime(cfqd, cic, rq);
1751         cfq_update_idle_window(cfqd, cfqq, cic);
1752
1753         cic->last_request_pos = rq->sector + rq->nr_sectors;
1754
1755         if (cfqq == cfqd->active_queue) {
1756                 /*
1757                  * if we are waiting for a request for this queue, let it rip
1758                  * immediately and flag that we must not expire this queue
1759                  * just now
1760                  */
1761                 if (cfq_cfqq_wait_request(cfqq)) {
1762                         cfq_mark_cfqq_must_dispatch(cfqq);
1763                         del_timer(&cfqd->idle_slice_timer);
1764                         blk_start_queueing(cfqd->queue);
1765                 }
1766         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1767                 /*
1768                  * not the active queue - expire current slice if it is
1769                  * idle and has expired it's mean thinktime or this new queue
1770                  * has some old slice time left and is of higher priority
1771                  */
1772                 cfq_preempt_queue(cfqd, cfqq);
1773                 cfq_mark_cfqq_must_dispatch(cfqq);
1774                 blk_start_queueing(cfqd->queue);
1775         }
1776 }
1777
1778 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1779 {
1780         struct cfq_data *cfqd = q->elevator->elevator_data;
1781         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1782
1783         cfq_init_prio_data(cfqq);
1784
1785         cfq_add_rq_rb(rq);
1786
1787         list_add_tail(&rq->queuelist, &cfqq->fifo);
1788
1789         cfq_rq_enqueued(cfqd, cfqq, rq);
1790 }
1791
1792 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1793 {
1794         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1795         struct cfq_data *cfqd = cfqq->cfqd;
1796         const int sync = rq_is_sync(rq);
1797         unsigned long now;
1798
1799         now = jiffies;
1800
1801         WARN_ON(!cfqd->rq_in_driver);
1802         WARN_ON(!cfqq->dispatched);
1803         cfqd->rq_in_driver--;
1804         cfqq->dispatched--;
1805
1806         if (cfq_cfqq_sync(cfqq))
1807                 cfqd->sync_flight--;
1808
1809         if (!cfq_class_idle(cfqq))
1810                 cfqd->last_end_request = now;
1811
1812         if (sync)
1813                 RQ_CIC(rq)->last_end_request = now;
1814
1815         /*
1816          * If this is the active queue, check if it needs to be expired,
1817          * or if we want to idle in case it has no pending requests.
1818          */
1819         if (cfqd->active_queue == cfqq) {
1820                 if (cfq_cfqq_slice_new(cfqq)) {
1821                         cfq_set_prio_slice(cfqd, cfqq);
1822                         cfq_clear_cfqq_slice_new(cfqq);
1823                 }
1824                 if (cfq_slice_used(cfqq))
1825                         cfq_slice_expired(cfqd, 1);
1826                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1827                         cfq_arm_slice_timer(cfqd);
1828         }
1829
1830         if (!cfqd->rq_in_driver)
1831                 cfq_schedule_dispatch(cfqd);
1832 }
1833
1834 /*
1835  * we temporarily boost lower priority queues if they are holding fs exclusive
1836  * resources. they are boosted to normal prio (CLASS_BE/4)
1837  */
1838 static void cfq_prio_boost(struct cfq_queue *cfqq)
1839 {
1840         if (has_fs_excl()) {
1841                 /*
1842                  * boost idle prio on transactions that would lock out other
1843                  * users of the filesystem
1844                  */
1845                 if (cfq_class_idle(cfqq))
1846                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1847                 if (cfqq->ioprio > IOPRIO_NORM)
1848                         cfqq->ioprio = IOPRIO_NORM;
1849         } else {
1850                 /*
1851                  * check if we need to unboost the queue
1852                  */
1853                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1854                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1855                 if (cfqq->ioprio != cfqq->org_ioprio)
1856                         cfqq->ioprio = cfqq->org_ioprio;
1857         }
1858 }
1859
1860 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1861 {
1862         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1863             !cfq_cfqq_must_alloc_slice(cfqq)) {
1864                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1865                 return ELV_MQUEUE_MUST;
1866         }
1867
1868         return ELV_MQUEUE_MAY;
1869 }
1870
1871 static int cfq_may_queue(request_queue_t *q, int rw)
1872 {
1873         struct cfq_data *cfqd = q->elevator->elevator_data;
1874         struct task_struct *tsk = current;
1875         struct cfq_io_context *cic;
1876         struct cfq_queue *cfqq;
1877
1878         /*
1879          * don't force setup of a queue from here, as a call to may_queue
1880          * does not necessarily imply that a request actually will be queued.
1881          * so just lookup a possibly existing queue, or return 'may queue'
1882          * if that fails
1883          */
1884         cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
1885         if (!cic)
1886                 return ELV_MQUEUE_MAY;
1887
1888         cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
1889         if (cfqq) {
1890                 cfq_init_prio_data(cfqq);
1891                 cfq_prio_boost(cfqq);
1892
1893                 return __cfq_may_queue(cfqq);
1894         }
1895
1896         return ELV_MQUEUE_MAY;
1897 }
1898
1899 /*
1900  * queue lock held here
1901  */
1902 static void cfq_put_request(struct request *rq)
1903 {
1904         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1905
1906         if (cfqq) {
1907                 const int rw = rq_data_dir(rq);
1908
1909                 BUG_ON(!cfqq->allocated[rw]);
1910                 cfqq->allocated[rw]--;
1911
1912                 put_io_context(RQ_CIC(rq)->ioc);
1913
1914                 rq->elevator_private = NULL;
1915                 rq->elevator_private2 = NULL;
1916
1917                 cfq_put_queue(cfqq);
1918         }
1919 }
1920
1921 /*
1922  * Allocate cfq data structures associated with this request.
1923  */
1924 static int
1925 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1926 {
1927         struct cfq_data *cfqd = q->elevator->elevator_data;
1928         struct task_struct *tsk = current;
1929         struct cfq_io_context *cic;
1930         const int rw = rq_data_dir(rq);
1931         const int is_sync = rq_is_sync(rq);
1932         struct cfq_queue *cfqq;
1933         unsigned long flags;
1934
1935         might_sleep_if(gfp_mask & __GFP_WAIT);
1936
1937         cic = cfq_get_io_context(cfqd, gfp_mask);
1938
1939         spin_lock_irqsave(q->queue_lock, flags);
1940
1941         if (!cic)
1942                 goto queue_fail;
1943
1944         cfqq = cic_to_cfqq(cic, is_sync);
1945         if (!cfqq) {
1946                 cfqq = cfq_get_queue(cfqd, is_sync, tsk, gfp_mask);
1947
1948                 if (!cfqq)
1949                         goto queue_fail;
1950
1951                 cic_set_cfqq(cic, cfqq, is_sync);
1952         }
1953
1954         cfqq->allocated[rw]++;
1955         cfq_clear_cfqq_must_alloc(cfqq);
1956         atomic_inc(&cfqq->ref);
1957
1958         spin_unlock_irqrestore(q->queue_lock, flags);
1959
1960         rq->elevator_private = cic;
1961         rq->elevator_private2 = cfqq;
1962         return 0;
1963
1964 queue_fail:
1965         if (cic)
1966                 put_io_context(cic->ioc);
1967
1968         cfq_schedule_dispatch(cfqd);
1969         spin_unlock_irqrestore(q->queue_lock, flags);
1970         return 1;
1971 }
1972
1973 static void cfq_kick_queue(struct work_struct *work)
1974 {
1975         struct cfq_data *cfqd =
1976                 container_of(work, struct cfq_data, unplug_work);
1977         request_queue_t *q = cfqd->queue;
1978         unsigned long flags;
1979
1980         spin_lock_irqsave(q->queue_lock, flags);
1981         blk_start_queueing(q);
1982         spin_unlock_irqrestore(q->queue_lock, flags);
1983 }
1984
1985 /*
1986  * Timer running if the active_queue is currently idling inside its time slice
1987  */
1988 static void cfq_idle_slice_timer(unsigned long data)
1989 {
1990         struct cfq_data *cfqd = (struct cfq_data *) data;
1991         struct cfq_queue *cfqq;
1992         unsigned long flags;
1993         int timed_out = 1;
1994
1995         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1996
1997         if ((cfqq = cfqd->active_queue) != NULL) {
1998                 timed_out = 0;
1999
2000                 /*
2001                  * expired
2002                  */
2003                 if (cfq_slice_used(cfqq))
2004                         goto expire;
2005
2006                 /*
2007                  * only expire and reinvoke request handler, if there are
2008                  * other queues with pending requests
2009                  */
2010                 if (!cfqd->busy_queues)
2011                         goto out_cont;
2012
2013                 /*
2014                  * not expired and it has a request pending, let it dispatch
2015                  */
2016                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
2017                         cfq_mark_cfqq_must_dispatch(cfqq);
2018                         goto out_kick;
2019                 }
2020         }
2021 expire:
2022         cfq_slice_expired(cfqd, timed_out);
2023 out_kick:
2024         cfq_schedule_dispatch(cfqd);
2025 out_cont:
2026         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2027 }
2028
2029 /*
2030  * Timer running if an idle class queue is waiting for service
2031  */
2032 static void cfq_idle_class_timer(unsigned long data)
2033 {
2034         struct cfq_data *cfqd = (struct cfq_data *) data;
2035         unsigned long flags, end;
2036
2037         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2038
2039         /*
2040          * race with a non-idle queue, reset timer
2041          */
2042         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2043         if (!time_after_eq(jiffies, end))
2044                 mod_timer(&cfqd->idle_class_timer, end);
2045         else
2046                 cfq_schedule_dispatch(cfqd);
2047
2048         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2049 }
2050
2051 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2052 {
2053         del_timer_sync(&cfqd->idle_slice_timer);
2054         del_timer_sync(&cfqd->idle_class_timer);
2055         blk_sync_queue(cfqd->queue);
2056 }
2057
2058 static void cfq_put_async_queues(struct cfq_data *cfqd)
2059 {
2060         int i;
2061
2062         for (i = 0; i < IOPRIO_BE_NR; i++) {
2063                 if (cfqd->async_cfqq[0][i])
2064                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2065                 if (cfqd->async_cfqq[1][i])
2066                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2067                 if (cfqd->async_idle_cfqq)
2068                         cfq_put_queue(cfqd->async_idle_cfqq);
2069         }
2070 }
2071
2072 static void cfq_exit_queue(elevator_t *e)
2073 {
2074         struct cfq_data *cfqd = e->elevator_data;
2075         request_queue_t *q = cfqd->queue;
2076
2077         cfq_shutdown_timer_wq(cfqd);
2078
2079         spin_lock_irq(q->queue_lock);
2080
2081         if (cfqd->active_queue)
2082                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2083
2084         while (!list_empty(&cfqd->cic_list)) {
2085                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2086                                                         struct cfq_io_context,
2087                                                         queue_list);
2088
2089                 __cfq_exit_single_io_context(cfqd, cic);
2090         }
2091
2092         cfq_put_async_queues(cfqd);
2093
2094         spin_unlock_irq(q->queue_lock);
2095
2096         cfq_shutdown_timer_wq(cfqd);
2097
2098         kfree(cfqd);
2099 }
2100
2101 static void *cfq_init_queue(request_queue_t *q)
2102 {
2103         struct cfq_data *cfqd;
2104
2105         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2106         if (!cfqd)
2107                 return NULL;
2108
2109         cfqd->service_tree = CFQ_RB_ROOT;
2110         INIT_LIST_HEAD(&cfqd->cic_list);
2111
2112         cfqd->queue = q;
2113
2114         init_timer(&cfqd->idle_slice_timer);
2115         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2116         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2117
2118         init_timer(&cfqd->idle_class_timer);
2119         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2120         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2121
2122         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2123
2124         cfqd->cfq_quantum = cfq_quantum;
2125         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2126         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2127         cfqd->cfq_back_max = cfq_back_max;
2128         cfqd->cfq_back_penalty = cfq_back_penalty;
2129         cfqd->cfq_slice[0] = cfq_slice_async;
2130         cfqd->cfq_slice[1] = cfq_slice_sync;
2131         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2132         cfqd->cfq_slice_idle = cfq_slice_idle;
2133
2134         return cfqd;
2135 }
2136
2137 static void cfq_slab_kill(void)
2138 {
2139         if (cfq_pool)
2140                 kmem_cache_destroy(cfq_pool);
2141         if (cfq_ioc_pool)
2142                 kmem_cache_destroy(cfq_ioc_pool);
2143 }
2144
2145 static int __init cfq_slab_setup(void)
2146 {
2147         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2148         if (!cfq_pool)
2149                 goto fail;
2150
2151         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2152         if (!cfq_ioc_pool)
2153                 goto fail;
2154
2155         return 0;
2156 fail:
2157         cfq_slab_kill();
2158         return -ENOMEM;
2159 }
2160
2161 /*
2162  * sysfs parts below -->
2163  */
2164 static ssize_t
2165 cfq_var_show(unsigned int var, char *page)
2166 {
2167         return sprintf(page, "%d\n", var);
2168 }
2169
2170 static ssize_t
2171 cfq_var_store(unsigned int *var, const char *page, size_t count)
2172 {
2173         char *p = (char *) page;
2174
2175         *var = simple_strtoul(p, &p, 10);
2176         return count;
2177 }
2178
2179 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2180 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2181 {                                                                       \
2182         struct cfq_data *cfqd = e->elevator_data;                       \
2183         unsigned int __data = __VAR;                                    \
2184         if (__CONV)                                                     \
2185                 __data = jiffies_to_msecs(__data);                      \
2186         return cfq_var_show(__data, (page));                            \
2187 }
2188 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2189 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2190 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2191 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2192 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2193 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2194 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2195 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2196 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2197 #undef SHOW_FUNCTION
2198
2199 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2200 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2201 {                                                                       \
2202         struct cfq_data *cfqd = e->elevator_data;                       \
2203         unsigned int __data;                                            \
2204         int ret = cfq_var_store(&__data, (page), count);                \
2205         if (__data < (MIN))                                             \
2206                 __data = (MIN);                                         \
2207         else if (__data > (MAX))                                        \
2208                 __data = (MAX);                                         \
2209         if (__CONV)                                                     \
2210                 *(__PTR) = msecs_to_jiffies(__data);                    \
2211         else                                                            \
2212                 *(__PTR) = __data;                                      \
2213         return ret;                                                     \
2214 }
2215 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2216 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2217 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2218 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2219 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2220 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2221 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2222 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2223 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2224 #undef STORE_FUNCTION
2225
2226 #define CFQ_ATTR(name) \
2227         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2228
2229 static struct elv_fs_entry cfq_attrs[] = {
2230         CFQ_ATTR(quantum),
2231         CFQ_ATTR(fifo_expire_sync),
2232         CFQ_ATTR(fifo_expire_async),
2233         CFQ_ATTR(back_seek_max),
2234         CFQ_ATTR(back_seek_penalty),
2235         CFQ_ATTR(slice_sync),
2236         CFQ_ATTR(slice_async),
2237         CFQ_ATTR(slice_async_rq),
2238         CFQ_ATTR(slice_idle),
2239         __ATTR_NULL
2240 };
2241
2242 static struct elevator_type iosched_cfq = {
2243         .ops = {
2244                 .elevator_merge_fn =            cfq_merge,
2245                 .elevator_merged_fn =           cfq_merged_request,
2246                 .elevator_merge_req_fn =        cfq_merged_requests,
2247                 .elevator_allow_merge_fn =      cfq_allow_merge,
2248                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2249                 .elevator_add_req_fn =          cfq_insert_request,
2250                 .elevator_activate_req_fn =     cfq_activate_request,
2251                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2252                 .elevator_queue_empty_fn =      cfq_queue_empty,
2253                 .elevator_completed_req_fn =    cfq_completed_request,
2254                 .elevator_former_req_fn =       elv_rb_former_request,
2255                 .elevator_latter_req_fn =       elv_rb_latter_request,
2256                 .elevator_set_req_fn =          cfq_set_request,
2257                 .elevator_put_req_fn =          cfq_put_request,
2258                 .elevator_may_queue_fn =        cfq_may_queue,
2259                 .elevator_init_fn =             cfq_init_queue,
2260                 .elevator_exit_fn =             cfq_exit_queue,
2261                 .trim =                         cfq_free_io_context,
2262         },
2263         .elevator_attrs =       cfq_attrs,
2264         .elevator_name =        "cfq",
2265         .elevator_owner =       THIS_MODULE,
2266 };
2267
2268 static int __init cfq_init(void)
2269 {
2270         int ret;
2271
2272         /*
2273          * could be 0 on HZ < 1000 setups
2274          */
2275         if (!cfq_slice_async)
2276                 cfq_slice_async = 1;
2277         if (!cfq_slice_idle)
2278                 cfq_slice_idle = 1;
2279
2280         if (cfq_slab_setup())
2281                 return -ENOMEM;
2282
2283         ret = elv_register(&iosched_cfq);
2284         if (ret)
2285                 cfq_slab_kill();
2286
2287         return ret;
2288 }
2289
2290 static void __exit cfq_exit(void)
2291 {
2292         DECLARE_COMPLETION_ONSTACK(all_gone);
2293         elv_unregister(&iosched_cfq);
2294         ioc_gone = &all_gone;
2295         /* ioc_gone's update must be visible before reading ioc_count */
2296         smp_wmb();
2297         if (elv_ioc_count_read(ioc_count))
2298                 wait_for_completion(ioc_gone);
2299         synchronize_rcu();
2300         cfq_slab_kill();
2301 }
2302
2303 module_init(cfq_init);
2304 module_exit(cfq_exit);
2305
2306 MODULE_AUTHOR("Jens Axboe");
2307 MODULE_LICENSE("GPL");
2308 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");