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