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 #include <linux/blktrace_api.h>
  15
  16 /*
  17  * tunables
  18  */
  19 /* max queue in one round of service */
  20 static const int cfq_quantum = 4;
  21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  22 /* maximum backwards seek, in KiB */
  23 static const int cfq_back_max = 16 * 1024;
  24 /* penalty of a backwards seek */
  25 static const int cfq_back_penalty = 2;
  26 static const int cfq_slice_sync = HZ / 10;
  27 static int cfq_slice_async = HZ / 25;
  28 static const int cfq_slice_async_rq = 2;
  29 static int cfq_slice_idle = HZ / 125;
  30
  31 /*
  32  * offset from end of service tree
  33  */
  34 #define CFQ_IDLE_DELAY          (HZ / 5)
  35
  36 /*
  37  * below this threshold, we consider thinktime immediate
  38  */
  39 #define CFQ_MIN_TT              (2)
  40
  41 #define CFQ_SLICE_SCALE         (5)
  42 #define CFQ_HW_QUEUE_MIN        (5)
  43
  44 #define RQ_CIC(rq)              \
  45         ((struct cfq_io_context *) (rq)->elevator_private)
  46 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
  47
  48 static struct kmem_cache *cfq_pool;
  49 static struct kmem_cache *cfq_ioc_pool;
  50
  51 static DEFINE_PER_CPU(unsigned long, ioc_count);
  52 static struct completion *ioc_gone;
  53 static DEFINE_SPINLOCK(ioc_gone_lock);
  54
  55 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  56 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  57 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  58
  59 #define sample_valid(samples)   ((samples) > 80)
  60
  61 /*
  62  * Most of our rbtree usage is for sorting with min extraction, so
  63  * if we cache the leftmost node we don't have to walk down the tree
  64  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  65  * move this into the elevator for the rq sorting as well.
  66  */
  67 struct cfq_rb_root {
  68         struct rb_root rb;
  69         struct rb_node *left;
  70 };
  71 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, }
  72
  73 /*
  74  * Per block device queue structure
  75  */
  76 struct cfq_data {
  77         struct request_queue *queue;
  78
  79         /*
  80          * rr list of queues with requests and the count of them
  81          */
  82         struct cfq_rb_root service_tree;
  83
  84         /*
  85          * Each priority tree is sorted by next_request position.  These
  86          * trees are used when determining if two or more queues are
  87          * interleaving requests (see cfq_close_cooperator).
  88          */
  89         struct rb_root prio_trees[CFQ_PRIO_LISTS];
  90
  91         unsigned int busy_queues;
  92         /*
  93          * Used to track any pending rt requests so we can pre-empt current
  94          * non-RT cfqq in service when this value is non-zero.
  95          */
  96         unsigned int busy_rt_queues;
  97
  98         int rq_in_driver;
  99         int sync_flight;
 100
 101         /*
 102          * queue-depth detection
 103          */
 104         int rq_queued;
 105         int hw_tag;
 106         int hw_tag_samples;
 107         int rq_in_driver_peak;
 108
 109         /*
 110          * idle window management
 111          */
 112         struct timer_list idle_slice_timer;
 113         struct work_struct unplug_work;
 114
 115         struct cfq_queue *active_queue;
 116         struct cfq_io_context *active_cic;
 117
 118         /*
 119          * async queue for each priority case
 120          */
 121         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 122         struct cfq_queue *async_idle_cfqq;
 123
 124         sector_t last_position;
 125
 126         /*
 127          * tunables, see top of file
 128          */
 129         unsigned int cfq_quantum;
 130         unsigned int cfq_fifo_expire[2];
 131         unsigned int cfq_back_penalty;
 132         unsigned int cfq_back_max;
 133         unsigned int cfq_slice[2];
 134         unsigned int cfq_slice_async_rq;
 135         unsigned int cfq_slice_idle;
 136
 137         struct list_head cic_list;
 138 };
 139
 140 /*
 141  * Per process-grouping structure
 142  */
 143 struct cfq_queue {
 144         /* reference count */
 145         atomic_t ref;
 146         /* various state flags, see below */
 147         unsigned int flags;
 148         /* parent cfq_data */
 149         struct cfq_data *cfqd;
 150         /* service_tree member */
 151         struct rb_node rb_node;
 152         /* service_tree key */
 153         unsigned long rb_key;
 154         /* prio tree member */
 155         struct rb_node p_node;
 156         /* prio tree root we belong to, if any */
 157         struct rb_root *p_root;
 158         /* sorted list of pending requests */
 159         struct rb_root sort_list;
 160         /* if fifo isn't expired, next request to serve */
 161         struct request *next_rq;
 162         /* requests queued in sort_list */
 163         int queued[2];
 164         /* currently allocated requests */
 165         int allocated[2];
 166         /* fifo list of requests in sort_list */
 167         struct list_head fifo;
 168
 169         unsigned long slice_end;
 170         long slice_resid;
 171         unsigned int slice_dispatch;
 172
 173         /* pending metadata requests */
 174         int meta_pending;
 175         /* number of requests that are on the dispatch list or inside driver */
 176         int dispatched;
 177
 178         /* io prio of this group */
 179         unsigned short ioprio, org_ioprio;
 180         unsigned short ioprio_class, org_ioprio_class;
 181
 182         pid_t pid;
 183 };
 184
 185 enum cfqq_state_flags {
 186         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 187         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 188         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
 189         CFQ_CFQQ_FLAG_must_alloc,       /* must be allowed rq alloc */
 190         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 191         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 192         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 193         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 194         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 195         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 196         CFQ_CFQQ_FLAG_coop,             /* has done a coop jump of the queue */
 197 };
 198
 199 #define CFQ_CFQQ_FNS(name)                                              \
 200 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 201 {                                                                       \
 202         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 203 }                                                                       \
 204 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 205 {                                                                       \
 206         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 207 }                                                                       \
 208 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 209 {                                                                       \
 210         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 211 }
 212
 213 CFQ_CFQQ_FNS(on_rr);
 214 CFQ_CFQQ_FNS(wait_request);
 215 CFQ_CFQQ_FNS(must_dispatch);
 216 CFQ_CFQQ_FNS(must_alloc);
 217 CFQ_CFQQ_FNS(must_alloc_slice);
 218 CFQ_CFQQ_FNS(fifo_expire);
 219 CFQ_CFQQ_FNS(idle_window);
 220 CFQ_CFQQ_FNS(prio_changed);
 221 CFQ_CFQQ_FNS(slice_new);
 222 CFQ_CFQQ_FNS(sync);
 223 CFQ_CFQQ_FNS(coop);
 224 #undef CFQ_CFQQ_FNS
 225
 226 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 227         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
 228 #define cfq_log(cfqd, fmt, args...)     \
 229         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 230
 231 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 232 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
 233                                        struct io_context *, gfp_t);
 234 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
 235                                                 struct io_context *);
 236
 237 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
 238                                             int is_sync)
 239 {
 240         return cic->cfqq[!!is_sync];
 241 }
 242
 243 static inline void cic_set_cfqq(struct cfq_io_context *cic,
 244                                 struct cfq_queue *cfqq, int is_sync)
 245 {
 246         cic->cfqq[!!is_sync] = cfqq;
 247 }
 248
 249 /*
 250  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 251  * set (in which case it could also be direct WRITE).
 252  */
 253 static inline int cfq_bio_sync(struct bio *bio)
 254 {
 255         if (bio_data_dir(bio) == READ || bio_sync(bio))
 256                 return 1;
 257
 258         return 0;
 259 }
 260
 261 /*
 262  * scheduler run of queue, if there are requests pending and no one in the
 263  * driver that will restart queueing
 264  */
 265 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 266 {
 267         if (cfqd->busy_queues) {
 268                 cfq_log(cfqd, "schedule dispatch");
 269                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
 270         }
 271 }
 272
 273 static int cfq_queue_empty(struct request_queue *q)
 274 {
 275         struct cfq_data *cfqd = q->elevator->elevator_data;
 276
 277         return !cfqd->busy_queues;
 278 }
 279
 280 /*
 281  * Scale schedule slice based on io priority. Use the sync time slice only
 282  * if a queue is marked sync and has sync io queued. A sync queue with async
 283  * io only, should not get full sync slice length.
 284  */
 285 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
 286                                  unsigned short prio)
 287 {
 288         const int base_slice = cfqd->cfq_slice[sync];
 289
 290         WARN_ON(prio >= IOPRIO_BE_NR);
 291
 292         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 293 }
 294
 295 static inline int
 296 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 297 {
 298         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 299 }
 300
 301 static inline void
 302 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 303 {
 304         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 305         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
 306 }
 307
 308 /*
 309  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
 310  * isn't valid until the first request from the dispatch is activated
 311  * and the slice time set.
 312  */
 313 static inline int cfq_slice_used(struct cfq_queue *cfqq)
 314 {
 315         if (cfq_cfqq_slice_new(cfqq))
 316                 return 0;
 317         if (time_before(jiffies, cfqq->slice_end))
 318                 return 0;
 319
 320         return 1;
 321 }
 322
 323 /*
 324  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 325  * We choose the request that is closest to the head right now. Distance
 326  * behind the head is penalized and only allowed to a certain extent.
 327  */
 328 static struct request *
 329 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
 330 {
 331         sector_t last, s1, s2, d1 = 0, d2 = 0;
 332         unsigned long back_max;
 333 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 334 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 335         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 336
 337         if (rq1 == NULL || rq1 == rq2)
 338                 return rq2;
 339         if (rq2 == NULL)
 340                 return rq1;
 341
 342         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 343                 return rq1;
 344         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 345                 return rq2;
 346         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
 347                 return rq1;
 348         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
 349                 return rq2;
 350
 351         s1 = blk_rq_pos(rq1);
 352         s2 = blk_rq_pos(rq2);
 353
 354         last = cfqd->last_position;
 355
 356         /*
 357          * by definition, 1KiB is 2 sectors
 358          */
 359         back_max = cfqd->cfq_back_max * 2;
 360
 361         /*
 362          * Strict one way elevator _except_ in the case where we allow
 363          * short backward seeks which are biased as twice the cost of a
 364          * similar forward seek.
 365          */
 366         if (s1 >= last)
 367                 d1 = s1 - last;
 368         else if (s1 + back_max >= last)
 369                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 370         else
 371                 wrap |= CFQ_RQ1_WRAP;
 372
 373         if (s2 >= last)
 374                 d2 = s2 - last;
 375         else if (s2 + back_max >= last)
 376                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 377         else
 378                 wrap |= CFQ_RQ2_WRAP;
 379
 380         /* Found required data */
 381
 382         /*
 383          * By doing switch() on the bit mask "wrap" we avoid having to
 384          * check two variables for all permutations: --> faster!
 385          */
 386         switch (wrap) {
 387         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 388                 if (d1 < d2)
 389                         return rq1;
 390                 else if (d2 < d1)
 391                         return rq2;
 392                 else {
 393                         if (s1 >= s2)
 394                                 return rq1;
 395                         else
 396                                 return rq2;
 397                 }
 398
 399         case CFQ_RQ2_WRAP:
 400                 return rq1;
 401         case CFQ_RQ1_WRAP:
 402                 return rq2;
 403         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
 404         default:
 405                 /*
 406                  * Since both rqs are wrapped,
 407                  * start with the one that's further behind head
 408                  * (--> only *one* back seek required),
 409                  * since back seek takes more time than forward.
 410                  */
 411                 if (s1 <= s2)
 412                         return rq1;
 413                 else
 414                         return rq2;
 415         }
 416 }
 417
 418 /*
 419  * The below is leftmost cache rbtree addon
 420  */
 421 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
 422 {
 423         if (!root->left)
 424                 root->left = rb_first(&root->rb);
 425
 426         if (root->left)
 427                 return rb_entry(root->left, struct cfq_queue, rb_node);
 428
 429         return NULL;
 430 }
 431
 432 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
 433 {
 434         rb_erase(n, root);
 435         RB_CLEAR_NODE(n);
 436 }
 437
 438 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
 439 {
 440         if (root->left == n)
 441                 root->left = NULL;
 442         rb_erase_init(n, &root->rb);
 443 }
 444
 445 /*
 446  * would be nice to take fifo expire time into account as well
 447  */
 448 static struct request *
 449 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 450                   struct request *last)
 451 {
 452         struct rb_node *rbnext = rb_next(&last->rb_node);
 453         struct rb_node *rbprev = rb_prev(&last->rb_node);
 454         struct request *next = NULL, *prev = NULL;
 455
 456         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
 457
 458         if (rbprev)
 459                 prev = rb_entry_rq(rbprev);
 460
 461         if (rbnext)
 462                 next = rb_entry_rq(rbnext);
 463         else {
 464                 rbnext = rb_first(&cfqq->sort_list);
 465                 if (rbnext && rbnext != &last->rb_node)
 466                         next = rb_entry_rq(rbnext);
 467         }
 468
 469         return cfq_choose_req(cfqd, next, prev);
 470 }
 471
 472 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
 473                                       struct cfq_queue *cfqq)
 474 {
 475         /*
 476          * just an approximation, should be ok.
 477          */
 478         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
 479                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
 480 }
 481
 482 /*
 483  * The cfqd->service_tree holds all pending cfq_queue's that have
 484  * requests waiting to be processed. It is sorted in the order that
 485  * we will service the queues.
 486  */
 487 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 488                                  int add_front)
 489 {
 490         struct rb_node **p, *parent;
 491         struct cfq_queue *__cfqq;
 492         unsigned long rb_key;
 493         int left;
 494
 495         if (cfq_class_idle(cfqq)) {
 496                 rb_key = CFQ_IDLE_DELAY;
 497                 parent = rb_last(&cfqd->service_tree.rb);
 498                 if (parent && parent != &cfqq->rb_node) {
 499                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
 500                         rb_key += __cfqq->rb_key;
 501                 } else
 502                         rb_key += jiffies;
 503         } else if (!add_front) {
 504                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
 505                 rb_key += cfqq->slice_resid;
 506                 cfqq->slice_resid = 0;
 507         } else
 508                 rb_key = 0;
 509
 510         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
 511                 /*
 512                  * same position, nothing more to do
 513                  */
 514                 if (rb_key == cfqq->rb_key)
 515                         return;
 516
 517                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
 518         }
 519
 520         left = 1;
 521         parent = NULL;
 522         p = &cfqd->service_tree.rb.rb_node;
 523         while (*p) {
 524                 struct rb_node **n;
 525
 526                 parent = *p;
 527                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
 528
 529                 /*
 530                  * sort RT queues first, we always want to give
 531                  * preference to them. IDLE queues goes to the back.
 532                  * after that, sort on the next service time.
 533                  */
 534                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
 535                         n = &(*p)->rb_left;
 536                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
 537                         n = &(*p)->rb_right;
 538                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
 539                         n = &(*p)->rb_left;
 540                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
 541                         n = &(*p)->rb_right;
 542                 else if (rb_key < __cfqq->rb_key)
 543                         n = &(*p)->rb_left;
 544                 else
 545                         n = &(*p)->rb_right;
 546
 547                 if (n == &(*p)->rb_right)
 548                         left = 0;
 549
 550                 p = n;
 551         }
 552
 553         if (left)
 554                 cfqd->service_tree.left = &cfqq->rb_node;
 555
 556         cfqq->rb_key = rb_key;
 557         rb_link_node(&cfqq->rb_node, parent, p);
 558         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
 559 }
 560
 561 static struct cfq_queue *
 562 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
 563                      sector_t sector, struct rb_node **ret_parent,
 564                      struct rb_node ***rb_link)
 565 {
 566         struct rb_node **p, *parent;
 567         struct cfq_queue *cfqq = NULL;
 568
 569         parent = NULL;
 570         p = &root->rb_node;
 571         while (*p) {
 572                 struct rb_node **n;
 573
 574                 parent = *p;
 575                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
 576
 577                 /*
 578                  * Sort strictly based on sector.  Smallest to the left,
 579                  * largest to the right.
 580                  */
 581                 if (sector > blk_rq_pos(cfqq->next_rq))
 582                         n = &(*p)->rb_right;
 583                 else if (sector < blk_rq_pos(cfqq->next_rq))
 584                         n = &(*p)->rb_left;
 585                 else
 586                         break;
 587                 p = n;
 588                 cfqq = NULL;
 589         }
 590
 591         *ret_parent = parent;
 592         if (rb_link)
 593                 *rb_link = p;
 594         return cfqq;
 595 }
 596
 597 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 598 {
 599         struct rb_node **p, *parent;
 600         struct cfq_queue *__cfqq;
 601
 602         if (cfqq->p_root) {
 603                 rb_erase(&cfqq->p_node, cfqq->p_root);
 604                 cfqq->p_root = NULL;
 605         }
 606
 607         if (cfq_class_idle(cfqq))
 608                 return;
 609         if (!cfqq->next_rq)
 610                 return;
 611
 612         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
 613         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
 614                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
 615         if (!__cfqq) {
 616                 rb_link_node(&cfqq->p_node, parent, p);
 617                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
 618         } else
 619                 cfqq->p_root = NULL;
 620 }
 621
 622 /*
 623  * Update cfqq's position in the service tree.
 624  */
 625 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 626 {
 627         /*
 628          * Resorting requires the cfqq to be on the RR list already.
 629          */
 630         if (cfq_cfqq_on_rr(cfqq)) {
 631                 cfq_service_tree_add(cfqd, cfqq, 0);
 632                 cfq_prio_tree_add(cfqd, cfqq);
 633         }
 634 }
 635
 636 /*
 637  * add to busy list of queues for service, trying to be fair in ordering
 638  * the pending list according to last request service
 639  */
 640 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 641 {
 642         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
 643         BUG_ON(cfq_cfqq_on_rr(cfqq));
 644         cfq_mark_cfqq_on_rr(cfqq);
 645         cfqd->busy_queues++;
 646         if (cfq_class_rt(cfqq))
 647                 cfqd->busy_rt_queues++;
 648
 649         cfq_resort_rr_list(cfqd, cfqq);
 650 }
 651
 652 /*
 653  * Called when the cfqq no longer has requests pending, remove it from
 654  * the service tree.
 655  */
 656 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 657 {
 658         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
 659         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 660         cfq_clear_cfqq_on_rr(cfqq);
 661
 662         if (!RB_EMPTY_NODE(&cfqq->rb_node))
 663                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
 664         if (cfqq->p_root) {
 665                 rb_erase(&cfqq->p_node, cfqq->p_root);
 666                 cfqq->p_root = NULL;
 667         }
 668
 669         BUG_ON(!cfqd->busy_queues);
 670         cfqd->busy_queues--;
 671         if (cfq_class_rt(cfqq))
 672                 cfqd->busy_rt_queues--;
 673 }
 674
 675 /*
 676  * rb tree support functions
 677  */
 678 static void cfq_del_rq_rb(struct request *rq)
 679 {
 680         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 681         struct cfq_data *cfqd = cfqq->cfqd;
 682         const int sync = rq_is_sync(rq);
 683
 684         BUG_ON(!cfqq->queued[sync]);
 685         cfqq->queued[sync]--;
 686
 687         elv_rb_del(&cfqq->sort_list, rq);
 688
 689         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
 690                 cfq_del_cfqq_rr(cfqd, cfqq);
 691 }
 692
 693 static void cfq_add_rq_rb(struct request *rq)
 694 {
 695         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 696         struct cfq_data *cfqd = cfqq->cfqd;
 697         struct request *__alias, *prev;
 698
 699         cfqq->queued[rq_is_sync(rq)]++;
 700
 701         /*
 702          * looks a little odd, but the first insert might return an alias.
 703          * if that happens, put the alias on the dispatch list
 704          */
 705         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
 706                 cfq_dispatch_insert(cfqd->queue, __alias);
 707
 708         if (!cfq_cfqq_on_rr(cfqq))
 709                 cfq_add_cfqq_rr(cfqd, cfqq);
 710
 711         /*
 712          * check if this request is a better next-serve candidate
 713          */
 714         prev = cfqq->next_rq;
 715         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
 716
 717         /*
 718          * adjust priority tree position, if ->next_rq changes
 719          */
 720         if (prev != cfqq->next_rq)
 721                 cfq_prio_tree_add(cfqd, cfqq);
 722
 723         BUG_ON(!cfqq->next_rq);
 724 }
 725
 726 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
 727 {
 728         elv_rb_del(&cfqq->sort_list, rq);
 729         cfqq->queued[rq_is_sync(rq)]--;
 730         cfq_add_rq_rb(rq);
 731 }
 732
 733 static struct request *
 734 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 735 {
 736         struct task_struct *tsk = current;
 737         struct cfq_io_context *cic;
 738         struct cfq_queue *cfqq;
 739
 740         cic = cfq_cic_lookup(cfqd, tsk->io_context);
 741         if (!cic)
 742                 return NULL;
 743
 744         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
 745         if (cfqq) {
 746                 sector_t sector = bio->bi_sector + bio_sectors(bio);
 747
 748                 return elv_rb_find(&cfqq->sort_list, sector);
 749         }
 750
 751         return NULL;
 752 }
 753
 754 static void cfq_activate_request(struct request_queue *q, struct request *rq)
 755 {
 756         struct cfq_data *cfqd = q->elevator->elevator_data;
 757
 758         cfqd->rq_in_driver++;
 759         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
 760                                                 cfqd->rq_in_driver);
 761
 762         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
 763 }
 764
 765 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
 766 {
 767         struct cfq_data *cfqd = q->elevator->elevator_data;
 768
 769         WARN_ON(!cfqd->rq_in_driver);
 770         cfqd->rq_in_driver--;
 771         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
 772                                                 cfqd->rq_in_driver);
 773 }
 774
 775 static void cfq_remove_request(struct request *rq)
 776 {
 777         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 778
 779         if (cfqq->next_rq == rq)
 780                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
 781
 782         list_del_init(&rq->queuelist);
 783         cfq_del_rq_rb(rq);
 784
 785         cfqq->cfqd->rq_queued--;
 786         if (rq_is_meta(rq)) {
 787                 WARN_ON(!cfqq->meta_pending);
 788                 cfqq->meta_pending--;
 789         }
 790 }
 791
 792 static int cfq_merge(struct request_queue *q, struct request **req,
 793                      struct bio *bio)
 794 {
 795         struct cfq_data *cfqd = q->elevator->elevator_data;
 796         struct request *__rq;
 797
 798         __rq = cfq_find_rq_fmerge(cfqd, bio);
 799         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 800                 *req = __rq;
 801                 return ELEVATOR_FRONT_MERGE;
 802         }
 803
 804         return ELEVATOR_NO_MERGE;
 805 }
 806
 807 static void cfq_merged_request(struct request_queue *q, struct request *req,
 808                                int type)
 809 {
 810         if (type == ELEVATOR_FRONT_MERGE) {
 811                 struct cfq_queue *cfqq = RQ_CFQQ(req);
 812
 813                 cfq_reposition_rq_rb(cfqq, req);
 814         }
 815 }
 816
 817 static void
 818 cfq_merged_requests(struct request_queue *q, struct request *rq,
 819                     struct request *next)
 820 {
 821         /*
 822          * reposition in fifo if next is older than rq
 823          */
 824         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 825             time_before(next->start_time, rq->start_time))
 826                 list_move(&rq->queuelist, &next->queuelist);
 827
 828         cfq_remove_request(next);
 829 }
 830
 831 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
 832                            struct bio *bio)
 833 {
 834         struct cfq_data *cfqd = q->elevator->elevator_data;
 835         struct cfq_io_context *cic;
 836         struct cfq_queue *cfqq;
 837
 838         /*
 839          * Disallow merge of a sync bio into an async request.
 840          */
 841         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
 842                 return 0;
 843
 844         /*
 845          * Lookup the cfqq that this bio will be queued with. Allow
 846          * merge only if rq is queued there.
 847          */
 848         cic = cfq_cic_lookup(cfqd, current->io_context);
 849         if (!cic)
 850                 return 0;
 851
 852         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
 853         if (cfqq == RQ_CFQQ(rq))
 854                 return 1;
 855
 856         return 0;
 857 }
 858
 859 static void __cfq_set_active_queue(struct cfq_data *cfqd,
 860                                    struct cfq_queue *cfqq)
 861 {
 862         if (cfqq) {
 863                 cfq_log_cfqq(cfqd, cfqq, "set_active");
 864                 cfqq->slice_end = 0;
 865                 cfqq->slice_dispatch = 0;
 866
 867                 cfq_clear_cfqq_wait_request(cfqq);
 868                 cfq_clear_cfqq_must_dispatch(cfqq);
 869                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 870                 cfq_clear_cfqq_fifo_expire(cfqq);
 871                 cfq_mark_cfqq_slice_new(cfqq);
 872
 873                 del_timer(&cfqd->idle_slice_timer);
 874         }
 875
 876         cfqd->active_queue = cfqq;
 877 }
 878
 879 /*
 880  * current cfqq expired its slice (or was too idle), select new one
 881  */
 882 static void
 883 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 884                     int timed_out)
 885 {
 886         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
 887
 888         if (cfq_cfqq_wait_request(cfqq))
 889                 del_timer(&cfqd->idle_slice_timer);
 890
 891         cfq_clear_cfqq_wait_request(cfqq);
 892
 893         /*
 894          * store what was left of this slice, if the queue idled/timed out
 895          */
 896         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
 897                 cfqq->slice_resid = cfqq->slice_end - jiffies;
 898                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
 899         }
 900
 901         cfq_resort_rr_list(cfqd, cfqq);
 902
 903         if (cfqq == cfqd->active_queue)
 904                 cfqd->active_queue = NULL;
 905
 906         if (cfqd->active_cic) {
 907                 put_io_context(cfqd->active_cic->ioc);
 908                 cfqd->active_cic = NULL;
 909         }
 910 }
 911
 912 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
 913 {
 914         struct cfq_queue *cfqq = cfqd->active_queue;
 915
 916         if (cfqq)
 917                 __cfq_slice_expired(cfqd, cfqq, timed_out);
 918 }
 919
 920 /*
 921  * Get next queue for service. Unless we have a queue preemption,
 922  * we'll simply select the first cfqq in the service tree.
 923  */
 924 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
 925 {
 926         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
 927                 return NULL;
 928
 929         return cfq_rb_first(&cfqd->service_tree);
 930 }
 931
 932 /*
 933  * Get and set a new active queue for service.
 934  */
 935 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
 936                                               struct cfq_queue *cfqq)
 937 {
 938         if (!cfqq) {
 939                 cfqq = cfq_get_next_queue(cfqd);
 940                 if (cfqq)
 941                         cfq_clear_cfqq_coop(cfqq);
 942         }
 943
 944         __cfq_set_active_queue(cfqd, cfqq);
 945         return cfqq;
 946 }
 947
 948 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
 949                                           struct request *rq)
 950 {
 951         if (blk_rq_pos(rq) >= cfqd->last_position)
 952                 return blk_rq_pos(rq) - cfqd->last_position;
 953         else
 954                 return cfqd->last_position - blk_rq_pos(rq);
 955 }
 956
 957 #define CIC_SEEK_THR    8 * 1024
 958 #define CIC_SEEKY(cic)  ((cic)->seek_mean > CIC_SEEK_THR)
 959
 960 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
 961 {
 962         struct cfq_io_context *cic = cfqd->active_cic;
 963         sector_t sdist = cic->seek_mean;
 964
 965         if (!sample_valid(cic->seek_samples))
 966                 sdist = CIC_SEEK_THR;
 967
 968         return cfq_dist_from_last(cfqd, rq) <= sdist;
 969 }
 970
 971 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
 972                                     struct cfq_queue *cur_cfqq)
 973 {
 974         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
 975         struct rb_node *parent, *node;
 976         struct cfq_queue *__cfqq;
 977         sector_t sector = cfqd->last_position;
 978
 979         if (RB_EMPTY_ROOT(root))
 980                 return NULL;
 981
 982         /*
 983          * First, if we find a request starting at the end of the last
 984          * request, choose it.
 985          */
 986         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
 987         if (__cfqq)
 988                 return __cfqq;
 989
 990         /*
 991          * If the exact sector wasn't found, the parent of the NULL leaf
 992          * will contain the closest sector.
 993          */
 994         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
 995         if (cfq_rq_close(cfqd, __cfqq->next_rq))
 996                 return __cfqq;
 997
 998         if (blk_rq_pos(__cfqq->next_rq) < sector)
 999                 node = rb_next(&__cfqq->p_node);
1000         else
1001                 node = rb_prev(&__cfqq->p_node);
1002         if (!node)
1003                 return NULL;
1004
1005         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1006         if (cfq_rq_close(cfqd, __cfqq->next_rq))
1007                 return __cfqq;
1008
1009         return NULL;
1010 }
1011
1012 /*
1013  * cfqd - obvious
1014  * cur_cfqq - passed in so that we don't decide that the current queue is
1015  *            closely cooperating with itself.
1016  *
1017  * So, basically we're assuming that that cur_cfqq has dispatched at least
1018  * one request, and that cfqd->last_position reflects a position on the disk
1019  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1020  * assumption.
1021  */
1022 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1023                                               struct cfq_queue *cur_cfqq,
1024                                               int probe)
1025 {
1026         struct cfq_queue *cfqq;
1027
1028         /*
1029          * A valid cfq_io_context is necessary to compare requests against
1030          * the seek_mean of the current cfqq.
1031          */
1032         if (!cfqd->active_cic)
1033                 return NULL;
1034
1035         /*
1036          * We should notice if some of the queues are cooperating, eg
1037          * working closely on the same area of the disk. In that case,
1038          * we can group them together and don't waste time idling.
1039          */
1040         cfqq = cfqq_close(cfqd, cur_cfqq);
1041         if (!cfqq)
1042                 return NULL;
1043
1044         if (cfq_cfqq_coop(cfqq))
1045                 return NULL;
1046
1047         if (!probe)
1048                 cfq_mark_cfqq_coop(cfqq);
1049         return cfqq;
1050 }
1051
1052 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1053 {
1054         struct cfq_queue *cfqq = cfqd->active_queue;
1055         struct cfq_io_context *cic;
1056         unsigned long sl;
1057
1058         /*
1059          * SSD device without seek penalty, disable idling. But only do so
1060          * for devices that support queuing, otherwise we still have a problem
1061          * with sync vs async workloads.
1062          */
1063         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1064                 return;
1065
1066         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1067         WARN_ON(cfq_cfqq_slice_new(cfqq));
1068
1069         /*
1070          * idle is disabled, either manually or by past process history
1071          */
1072         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1073                 return;
1074
1075         /*
1076          * still requests with the driver, don't idle
1077          */
1078         if (cfqd->rq_in_driver)
1079                 return;
1080
1081         /*
1082          * task has exited, don't wait
1083          */
1084         cic = cfqd->active_cic;
1085         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1086                 return;
1087
1088         cfq_mark_cfqq_wait_request(cfqq);
1089
1090         /*
1091          * we don't want to idle for seeks, but we do want to allow
1092          * fair distribution of slice time for a process doing back-to-back
1093          * seeks. so allow a little bit of time for him to submit a new rq
1094          */
1095         sl = cfqd->cfq_slice_idle;
1096         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
1097                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1098
1099         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1100         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1101 }
1102
1103 /*
1104  * Move request from internal lists to the request queue dispatch list.
1105  */
1106 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1107 {
1108         struct cfq_data *cfqd = q->elevator->elevator_data;
1109         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1110
1111         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1112
1113         cfq_remove_request(rq);
1114         cfqq->dispatched++;
1115         elv_dispatch_sort(q, rq);
1116
1117         if (cfq_cfqq_sync(cfqq))
1118                 cfqd->sync_flight++;
1119 }
1120
1121 /*
1122  * return expired entry, or NULL to just start from scratch in rbtree
1123  */
1124 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1125 {
1126         struct cfq_data *cfqd = cfqq->cfqd;
1127         struct request *rq;
1128         int fifo;
1129
1130         if (cfq_cfqq_fifo_expire(cfqq))
1131                 return NULL;
1132
1133         cfq_mark_cfqq_fifo_expire(cfqq);
1134
1135         if (list_empty(&cfqq->fifo))
1136                 return NULL;
1137
1138         fifo = cfq_cfqq_sync(cfqq);
1139         rq = rq_entry_fifo(cfqq->fifo.next);
1140
1141         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
1142                 rq = NULL;
1143
1144         cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
1145         return rq;
1146 }
1147
1148 static inline int
1149 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1150 {
1151         const int base_rq = cfqd->cfq_slice_async_rq;
1152
1153         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1154
1155         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1156 }
1157
1158 /*
1159  * Select a queue for service. If we have a current active queue,
1160  * check whether to continue servicing it, or retrieve and set a new one.
1161  */
1162 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1163 {
1164         struct cfq_queue *cfqq, *new_cfqq = NULL;
1165
1166         cfqq = cfqd->active_queue;
1167         if (!cfqq)
1168                 goto new_queue;
1169
1170         /*
1171          * The active queue has run out of time, expire it and select new.
1172          */
1173         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1174                 goto expire;
1175
1176         /*
1177          * If we have a RT cfqq waiting, then we pre-empt the current non-rt
1178          * cfqq.
1179          */
1180         if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues) {
1181                 /*
1182                  * We simulate this as cfqq timed out so that it gets to bank
1183                  * the remaining of its time slice.
1184                  */
1185                 cfq_log_cfqq(cfqd, cfqq, "preempt");
1186                 cfq_slice_expired(cfqd, 1);
1187                 goto new_queue;
1188         }
1189
1190         /*
1191          * The active queue has requests and isn't expired, allow it to
1192          * dispatch.
1193          */
1194         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1195                 goto keep_queue;
1196
1197         /*
1198          * If another queue has a request waiting within our mean seek
1199          * distance, let it run.  The expire code will check for close
1200          * cooperators and put the close queue at the front of the service
1201          * tree.
1202          */
1203         new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1204         if (new_cfqq)
1205                 goto expire;
1206
1207         /*
1208          * No requests pending. If the active queue still has requests in
1209          * flight or is idling for a new request, allow either of these
1210          * conditions to happen (or time out) before selecting a new queue.
1211          */
1212         if (timer_pending(&cfqd->idle_slice_timer) ||
1213             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1214                 cfqq = NULL;
1215                 goto keep_queue;
1216         }
1217
1218 expire:
1219         cfq_slice_expired(cfqd, 0);
1220 new_queue:
1221         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1222 keep_queue:
1223         return cfqq;
1224 }
1225
1226 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1227 {
1228         int dispatched = 0;
1229
1230         while (cfqq->next_rq) {
1231                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1232                 dispatched++;
1233         }
1234
1235         BUG_ON(!list_empty(&cfqq->fifo));
1236         return dispatched;
1237 }
1238
1239 /*
1240  * Drain our current requests. Used for barriers and when switching
1241  * io schedulers on-the-fly.
1242  */
1243 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1244 {
1245         struct cfq_queue *cfqq;
1246         int dispatched = 0;
1247
1248         while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1249                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1250
1251         cfq_slice_expired(cfqd, 0);
1252
1253         BUG_ON(cfqd->busy_queues);
1254
1255         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1256         return dispatched;
1257 }
1258
1259 /*
1260  * Dispatch a request from cfqq, moving them to the request queue
1261  * dispatch list.
1262  */
1263 static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1264 {
1265         struct request *rq;
1266
1267         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1268
1269         /*
1270          * follow expired path, else get first next available
1271          */
1272         rq = cfq_check_fifo(cfqq);
1273         if (!rq)
1274                 rq = cfqq->next_rq;
1275
1276         /*
1277          * insert request into driver dispatch list
1278          */
1279         cfq_dispatch_insert(cfqd->queue, rq);
1280
1281         if (!cfqd->active_cic) {
1282                 struct cfq_io_context *cic = RQ_CIC(rq);
1283
1284                 atomic_long_inc(&cic->ioc->refcount);
1285                 cfqd->active_cic = cic;
1286         }
1287 }
1288
1289 /*
1290  * Find the cfqq that we need to service and move a request from that to the
1291  * dispatch list
1292  */
1293 static int cfq_dispatch_requests(struct request_queue *q, int force)
1294 {
1295         struct cfq_data *cfqd = q->elevator->elevator_data;
1296         struct cfq_queue *cfqq;
1297         unsigned int max_dispatch;
1298
1299         if (!cfqd->busy_queues)
1300                 return 0;
1301
1302         if (unlikely(force))
1303                 return cfq_forced_dispatch(cfqd);
1304
1305         cfqq = cfq_select_queue(cfqd);
1306         if (!cfqq)
1307                 return 0;
1308
1309         /*
1310          * If this is an async queue and we have sync IO in flight, let it wait
1311          */
1312         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1313                 return 0;
1314
1315         max_dispatch = cfqd->cfq_quantum;
1316         if (cfq_class_idle(cfqq))
1317                 max_dispatch = 1;
1318
1319         /*
1320          * Does this cfqq already have too much IO in flight?
1321          */
1322         if (cfqq->dispatched >= max_dispatch) {
1323                 /*
1324                  * idle queue must always only have a single IO in flight
1325                  */
1326                 if (cfq_class_idle(cfqq))
1327                         return 0;
1328
1329                 /*
1330                  * We have other queues, don't allow more IO from this one
1331                  */
1332                 if (cfqd->busy_queues > 1)
1333                         return 0;
1334
1335                 /*
1336                  * we are the only queue, allow up to 4 times of 'quantum'
1337                  */
1338                 if (cfqq->dispatched >= 4 * max_dispatch)
1339                         return 0;
1340         }
1341
1342         /*
1343          * Dispatch a request from this cfqq
1344          */
1345         cfq_dispatch_request(cfqd, cfqq);
1346         cfqq->slice_dispatch++;
1347         cfq_clear_cfqq_must_dispatch(cfqq);
1348
1349         /*
1350          * expire an async queue immediately if it has used up its slice. idle
1351          * queue always expire after 1 dispatch round.
1352          */
1353         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1354             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1355             cfq_class_idle(cfqq))) {
1356                 cfqq->slice_end = jiffies + 1;
1357                 cfq_slice_expired(cfqd, 0);
1358         }
1359
1360         cfq_log(cfqd, "dispatched a request");
1361         return 1;
1362 }
1363
1364 /*
1365  * task holds one reference to the queue, dropped when task exits. each rq
1366  * in-flight on this queue also holds a reference, dropped when rq is freed.
1367  *
1368  * queue lock must be held here.
1369  */
1370 static void cfq_put_queue(struct cfq_queue *cfqq)
1371 {
1372         struct cfq_data *cfqd = cfqq->cfqd;
1373
1374         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1375
1376         if (!atomic_dec_and_test(&cfqq->ref))
1377                 return;
1378
1379         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1380         BUG_ON(rb_first(&cfqq->sort_list));
1381         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1382         BUG_ON(cfq_cfqq_on_rr(cfqq));
1383
1384         if (unlikely(cfqd->active_queue == cfqq)) {
1385                 __cfq_slice_expired(cfqd, cfqq, 0);
1386                 cfq_schedule_dispatch(cfqd);
1387         }
1388
1389         kmem_cache_free(cfq_pool, cfqq);
1390 }
1391
1392 /*
1393  * Must always be called with the rcu_read_lock() held
1394  */
1395 static void
1396 __call_for_each_cic(struct io_context *ioc,
1397                     void (*func)(struct io_context *, struct cfq_io_context *))
1398 {
1399         struct cfq_io_context *cic;
1400         struct hlist_node *n;
1401
1402         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1403                 func(ioc, cic);
1404 }
1405
1406 /*
1407  * Call func for each cic attached to this ioc.
1408  */
1409 static void
1410 call_for_each_cic(struct io_context *ioc,
1411                   void (*func)(struct io_context *, struct cfq_io_context *))
1412 {
1413         rcu_read_lock();
1414         __call_for_each_cic(ioc, func);
1415         rcu_read_unlock();
1416 }
1417
1418 static void cfq_cic_free_rcu(struct rcu_head *head)
1419 {
1420         struct cfq_io_context *cic;
1421
1422         cic = container_of(head, struct cfq_io_context, rcu_head);
1423
1424         kmem_cache_free(cfq_ioc_pool, cic);
1425         elv_ioc_count_dec(ioc_count);
1426
1427         if (ioc_gone) {
1428                 /*
1429                  * CFQ scheduler is exiting, grab exit lock and check
1430                  * the pending io context count. If it hits zero,
1431                  * complete ioc_gone and set it back to NULL
1432                  */
1433                 spin_lock(&ioc_gone_lock);
1434                 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1435                         complete(ioc_gone);
1436                         ioc_gone = NULL;
1437                 }
1438                 spin_unlock(&ioc_gone_lock);
1439         }
1440 }
1441
1442 static void cfq_cic_free(struct cfq_io_context *cic)
1443 {
1444         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1445 }
1446
1447 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1448 {
1449         unsigned long flags;
1450
1451         BUG_ON(!cic->dead_key);
1452
1453         spin_lock_irqsave(&ioc->lock, flags);
1454         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1455         hlist_del_rcu(&cic->cic_list);
1456         spin_unlock_irqrestore(&ioc->lock, flags);
1457
1458         cfq_cic_free(cic);
1459 }
1460
1461 /*
1462  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1463  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1464  * and ->trim() which is called with the task lock held
1465  */
1466 static void cfq_free_io_context(struct io_context *ioc)
1467 {
1468         /*
1469          * ioc->refcount is zero here, or we are called from elv_unregister(),
1470          * so no more cic's are allowed to be linked into this ioc.  So it
1471          * should be ok to iterate over the known list, we will see all cic's
1472          * since no new ones are added.
1473          */
1474         __call_for_each_cic(ioc, cic_free_func);
1475 }
1476
1477 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1478 {
1479         if (unlikely(cfqq == cfqd->active_queue)) {
1480                 __cfq_slice_expired(cfqd, cfqq, 0);
1481                 cfq_schedule_dispatch(cfqd);
1482         }
1483
1484         cfq_put_queue(cfqq);
1485 }
1486
1487 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1488                                          struct cfq_io_context *cic)
1489 {
1490         struct io_context *ioc = cic->ioc;
1491
1492         list_del_init(&cic->queue_list);
1493
1494         /*
1495          * Make sure key == NULL is seen for dead queues
1496          */
1497         smp_wmb();
1498         cic->dead_key = (unsigned long) cic->key;
1499         cic->key = NULL;
1500
1501         if (ioc->ioc_data == cic)
1502                 rcu_assign_pointer(ioc->ioc_data, NULL);
1503
1504         if (cic->cfqq[BLK_RW_ASYNC]) {
1505                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1506                 cic->cfqq[BLK_RW_ASYNC] = NULL;
1507         }
1508
1509         if (cic->cfqq[BLK_RW_SYNC]) {
1510                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1511                 cic->cfqq[BLK_RW_SYNC] = NULL;
1512         }
1513 }
1514
1515 static void cfq_exit_single_io_context(struct io_context *ioc,
1516                                        struct cfq_io_context *cic)
1517 {
1518         struct cfq_data *cfqd = cic->key;
1519
1520         if (cfqd) {
1521                 struct request_queue *q = cfqd->queue;
1522                 unsigned long flags;
1523
1524                 spin_lock_irqsave(q->queue_lock, flags);
1525
1526                 /*
1527                  * Ensure we get a fresh copy of the ->key to prevent
1528                  * race between exiting task and queue
1529                  */
1530                 smp_read_barrier_depends();
1531                 if (cic->key)
1532                         __cfq_exit_single_io_context(cfqd, cic);
1533
1534                 spin_unlock_irqrestore(q->queue_lock, flags);
1535         }
1536 }
1537
1538 /*
1539  * The process that ioc belongs to has exited, we need to clean up
1540  * and put the internal structures we have that belongs to that process.
1541  */
1542 static void cfq_exit_io_context(struct io_context *ioc)
1543 {
1544         call_for_each_cic(ioc, cfq_exit_single_io_context);
1545 }
1546
1547 static struct cfq_io_context *
1548 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1549 {
1550         struct cfq_io_context *cic;
1551
1552         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1553                                                         cfqd->queue->node);
1554         if (cic) {
1555                 cic->last_end_request = jiffies;
1556                 INIT_LIST_HEAD(&cic->queue_list);
1557                 INIT_HLIST_NODE(&cic->cic_list);
1558                 cic->dtor = cfq_free_io_context;
1559                 cic->exit = cfq_exit_io_context;
1560                 elv_ioc_count_inc(ioc_count);
1561         }
1562
1563         return cic;
1564 }
1565
1566 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1567 {
1568         struct task_struct *tsk = current;
1569         int ioprio_class;
1570
1571         if (!cfq_cfqq_prio_changed(cfqq))
1572                 return;
1573
1574         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1575         switch (ioprio_class) {
1576         default:
1577                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1578         case IOPRIO_CLASS_NONE:
1579                 /*
1580                  * no prio set, inherit CPU scheduling settings
1581                  */
1582                 cfqq->ioprio = task_nice_ioprio(tsk);
1583                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1584                 break;
1585         case IOPRIO_CLASS_RT:
1586                 cfqq->ioprio = task_ioprio(ioc);
1587                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1588                 break;
1589         case IOPRIO_CLASS_BE:
1590                 cfqq->ioprio = task_ioprio(ioc);
1591                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1592                 break;
1593         case IOPRIO_CLASS_IDLE:
1594                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1595                 cfqq->ioprio = 7;
1596                 cfq_clear_cfqq_idle_window(cfqq);
1597                 break;
1598         }
1599
1600         /*
1601          * keep track of original prio settings in case we have to temporarily
1602          * elevate the priority of this queue
1603          */
1604         cfqq->org_ioprio = cfqq->ioprio;
1605         cfqq->org_ioprio_class = cfqq->ioprio_class;
1606         cfq_clear_cfqq_prio_changed(cfqq);
1607 }
1608
1609 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1610 {
1611         struct cfq_data *cfqd = cic->key;
1612         struct cfq_queue *cfqq;
1613         unsigned long flags;
1614
1615         if (unlikely(!cfqd))
1616                 return;
1617
1618         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1619
1620         cfqq = cic->cfqq[BLK_RW_ASYNC];
1621         if (cfqq) {
1622                 struct cfq_queue *new_cfqq;
1623                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1624                                                 GFP_ATOMIC);
1625                 if (new_cfqq) {
1626                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1627                         cfq_put_queue(cfqq);
1628                 }
1629         }
1630
1631         cfqq = cic->cfqq[BLK_RW_SYNC];
1632         if (cfqq)
1633                 cfq_mark_cfqq_prio_changed(cfqq);
1634
1635         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1636 }
1637
1638 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1639 {
1640         call_for_each_cic(ioc, changed_ioprio);
1641         ioc->ioprio_changed = 0;
1642 }
1643
1644 static struct cfq_queue *
1645 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1646                      struct io_context *ioc, gfp_t gfp_mask)
1647 {
1648         struct cfq_queue *cfqq, *new_cfqq = NULL;
1649         struct cfq_io_context *cic;
1650
1651 retry:
1652         cic = cfq_cic_lookup(cfqd, ioc);
1653         /* cic always exists here */
1654         cfqq = cic_to_cfqq(cic, is_sync);
1655
1656         if (!cfqq) {
1657                 if (new_cfqq) {
1658                         cfqq = new_cfqq;
1659                         new_cfqq = NULL;
1660                 } else if (gfp_mask & __GFP_WAIT) {
1661                         /*
1662                          * Inform the allocator of the fact that we will
1663                          * just repeat this allocation if it fails, to allow
1664                          * the allocator to do whatever it needs to attempt to
1665                          * free memory.
1666                          */
1667                         spin_unlock_irq(cfqd->queue->queue_lock);
1668                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1669                                         gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1670                                         cfqd->queue->node);
1671                         spin_lock_irq(cfqd->queue->queue_lock);
1672                         goto retry;
1673                 } else {
1674                         cfqq = kmem_cache_alloc_node(cfq_pool,
1675                                         gfp_mask | __GFP_ZERO,
1676                                         cfqd->queue->node);
1677                         if (!cfqq)
1678                                 goto out;
1679                 }
1680
1681                 RB_CLEAR_NODE(&cfqq->rb_node);
1682                 RB_CLEAR_NODE(&cfqq->p_node);
1683                 INIT_LIST_HEAD(&cfqq->fifo);
1684
1685                 atomic_set(&cfqq->ref, 0);
1686                 cfqq->cfqd = cfqd;
1687
1688                 cfq_mark_cfqq_prio_changed(cfqq);
1689
1690                 cfq_init_prio_data(cfqq, ioc);
1691
1692                 if (is_sync) {
1693                         if (!cfq_class_idle(cfqq))
1694                                 cfq_mark_cfqq_idle_window(cfqq);
1695                         cfq_mark_cfqq_sync(cfqq);
1696                 }
1697                 cfqq->pid = current->pid;
1698                 cfq_log_cfqq(cfqd, cfqq, "alloced");
1699         }
1700
1701         if (new_cfqq)
1702                 kmem_cache_free(cfq_pool, new_cfqq);
1703
1704 out:
1705         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1706         return cfqq;
1707 }
1708
1709 static struct cfq_queue **
1710 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1711 {
1712         switch (ioprio_class) {
1713         case IOPRIO_CLASS_RT:
1714                 return &cfqd->async_cfqq[0][ioprio];
1715         case IOPRIO_CLASS_BE:
1716                 return &cfqd->async_cfqq[1][ioprio];
1717         case IOPRIO_CLASS_IDLE:
1718                 return &cfqd->async_idle_cfqq;
1719         default:
1720                 BUG();
1721         }
1722 }
1723
1724 static struct cfq_queue *
1725 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1726               gfp_t gfp_mask)
1727 {
1728         const int ioprio = task_ioprio(ioc);
1729         const int ioprio_class = task_ioprio_class(ioc);
1730         struct cfq_queue **async_cfqq = NULL;
1731         struct cfq_queue *cfqq = NULL;
1732
1733         if (!is_sync) {
1734                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1735                 cfqq = *async_cfqq;
1736         }
1737
1738         if (!cfqq) {
1739                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1740                 if (!cfqq)
1741                         return NULL;
1742         }
1743
1744         /*
1745          * pin the queue now that it's allocated, scheduler exit will prune it
1746          */
1747         if (!is_sync && !(*async_cfqq)) {
1748                 atomic_inc(&cfqq->ref);
1749                 *async_cfqq = cfqq;
1750         }
1751
1752         atomic_inc(&cfqq->ref);
1753         return cfqq;
1754 }
1755
1756 /*
1757  * We drop cfq io contexts lazily, so we may find a dead one.
1758  */
1759 static void
1760 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1761                   struct cfq_io_context *cic)
1762 {
1763         unsigned long flags;
1764
1765         WARN_ON(!list_empty(&cic->queue_list));
1766
1767         spin_lock_irqsave(&ioc->lock, flags);
1768
1769         BUG_ON(ioc->ioc_data == cic);
1770
1771         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1772         hlist_del_rcu(&cic->cic_list);
1773         spin_unlock_irqrestore(&ioc->lock, flags);
1774
1775         cfq_cic_free(cic);
1776 }
1777
1778 static struct cfq_io_context *
1779 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1780 {
1781         struct cfq_io_context *cic;
1782         unsigned long flags;
1783         void *k;
1784
1785         if (unlikely(!ioc))
1786                 return NULL;
1787
1788         rcu_read_lock();
1789
1790         /*
1791          * we maintain a last-hit cache, to avoid browsing over the tree
1792          */
1793         cic = rcu_dereference(ioc->ioc_data);
1794         if (cic && cic->key == cfqd) {
1795                 rcu_read_unlock();
1796                 return cic;
1797         }
1798
1799         do {
1800                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1801                 rcu_read_unlock();
1802                 if (!cic)
1803                         break;
1804                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1805                 k = cic->key;
1806                 if (unlikely(!k)) {
1807                         cfq_drop_dead_cic(cfqd, ioc, cic);
1808                         rcu_read_lock();
1809                         continue;
1810                 }
1811
1812                 spin_lock_irqsave(&ioc->lock, flags);
1813                 rcu_assign_pointer(ioc->ioc_data, cic);
1814                 spin_unlock_irqrestore(&ioc->lock, flags);
1815                 break;
1816         } while (1);
1817
1818         return cic;
1819 }
1820
1821 /*
1822  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1823  * the process specific cfq io context when entered from the block layer.
1824  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1825  */
1826 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1827                         struct cfq_io_context *cic, gfp_t gfp_mask)
1828 {
1829         unsigned long flags;
1830         int ret;
1831
1832         ret = radix_tree_preload(gfp_mask);
1833         if (!ret) {
1834                 cic->ioc = ioc;
1835                 cic->key = cfqd;
1836
1837                 spin_lock_irqsave(&ioc->lock, flags);
1838                 ret = radix_tree_insert(&ioc->radix_root,
1839                                                 (unsigned long) cfqd, cic);
1840                 if (!ret)
1841                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1842                 spin_unlock_irqrestore(&ioc->lock, flags);
1843
1844                 radix_tree_preload_end();
1845
1846                 if (!ret) {
1847                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1848                         list_add(&cic->queue_list, &cfqd->cic_list);
1849                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1850                 }
1851         }
1852
1853         if (ret)
1854                 printk(KERN_ERR "cfq: cic link failed!\n");
1855
1856         return ret;
1857 }
1858
1859 /*
1860  * Setup general io context and cfq io context. There can be several cfq
1861  * io contexts per general io context, if this process is doing io to more
1862  * than one device managed by cfq.
1863  */
1864 static struct cfq_io_context *
1865 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1866 {
1867         struct io_context *ioc = NULL;
1868         struct cfq_io_context *cic;
1869
1870         might_sleep_if(gfp_mask & __GFP_WAIT);
1871
1872         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1873         if (!ioc)
1874                 return NULL;
1875
1876         cic = cfq_cic_lookup(cfqd, ioc);
1877         if (cic)
1878                 goto out;
1879
1880         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1881         if (cic == NULL)
1882                 goto err;
1883
1884         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1885                 goto err_free;
1886
1887 out:
1888         smp_read_barrier_depends();
1889         if (unlikely(ioc->ioprio_changed))
1890                 cfq_ioc_set_ioprio(ioc);
1891
1892         return cic;
1893 err_free:
1894         cfq_cic_free(cic);
1895 err:
1896         put_io_context(ioc);
1897         return NULL;
1898 }
1899
1900 static void
1901 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1902 {
1903         unsigned long elapsed = jiffies - cic->last_end_request;
1904         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1905
1906         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1907         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1908         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1909 }
1910
1911 static void
1912 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1913                        struct request *rq)
1914 {
1915         sector_t sdist;
1916         u64 total;
1917
1918         if (!cic->last_request_pos)
1919                 sdist = 0;
1920         else if (cic->last_request_pos < blk_rq_pos(rq))
1921                 sdist = blk_rq_pos(rq) - cic->last_request_pos;
1922         else
1923                 sdist = cic->last_request_pos - blk_rq_pos(rq);
1924
1925         /*
1926          * Don't allow the seek distance to get too large from the
1927          * odd fragment, pagein, etc
1928          */
1929         if (cic->seek_samples <= 60) /* second&third seek */
1930                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1931         else
1932                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1933
1934         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1935         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1936         total = cic->seek_total + (cic->seek_samples/2);
1937         do_div(total, cic->seek_samples);
1938         cic->seek_mean = (sector_t)total;
1939 }
1940
1941 /*
1942  * Disable idle window if the process thinks too long or seeks so much that
1943  * it doesn't matter
1944  */
1945 static void
1946 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1947                        struct cfq_io_context *cic)
1948 {
1949         int old_idle, enable_idle;
1950
1951         /*
1952          * Don't idle for async or idle io prio class
1953          */
1954         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1955                 return;
1956
1957         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1958
1959         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1960             (cfqd->hw_tag && CIC_SEEKY(cic)))
1961                 enable_idle = 0;
1962         else if (sample_valid(cic->ttime_samples)) {
1963                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1964                         enable_idle = 0;
1965                 else
1966                         enable_idle = 1;
1967         }
1968
1969         if (old_idle != enable_idle) {
1970                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1971                 if (enable_idle)
1972                         cfq_mark_cfqq_idle_window(cfqq);
1973                 else
1974                         cfq_clear_cfqq_idle_window(cfqq);
1975         }
1976 }
1977
1978 /*
1979  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1980  * no or if we aren't sure, a 1 will cause a preempt.
1981  */
1982 static int
1983 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1984                    struct request *rq)
1985 {
1986         struct cfq_queue *cfqq;
1987
1988         cfqq = cfqd->active_queue;
1989         if (!cfqq)
1990                 return 0;
1991
1992         if (cfq_slice_used(cfqq))
1993                 return 1;
1994
1995         if (cfq_class_idle(new_cfqq))
1996                 return 0;
1997
1998         if (cfq_class_idle(cfqq))
1999                 return 1;
2000
2001         /*
2002          * if the new request is sync, but the currently running queue is
2003          * not, let the sync request have priority.
2004          */
2005         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2006                 return 1;
2007
2008         /*
2009          * So both queues are sync. Let the new request get disk time if
2010          * it's a metadata request and the current queue is doing regular IO.
2011          */
2012         if (rq_is_meta(rq) && !cfqq->meta_pending)
2013                 return 1;
2014
2015         /*
2016          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2017          */
2018         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2019                 return 1;
2020
2021         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2022                 return 0;
2023
2024         /*
2025          * if this request is as-good as one we would expect from the
2026          * current cfqq, let it preempt
2027          */
2028         if (cfq_rq_close(cfqd, rq))
2029                 return 1;
2030
2031         return 0;
2032 }
2033
2034 /*
2035  * cfqq preempts the active queue. if we allowed preempt with no slice left,
2036  * let it have half of its nominal slice.
2037  */
2038 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2039 {
2040         cfq_log_cfqq(cfqd, cfqq, "preempt");
2041         cfq_slice_expired(cfqd, 1);
2042
2043         /*
2044          * Put the new queue at the front of the of the current list,
2045          * so we know that it will be selected next.
2046          */
2047         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2048
2049         cfq_service_tree_add(cfqd, cfqq, 1);
2050
2051         cfqq->slice_end = 0;
2052         cfq_mark_cfqq_slice_new(cfqq);
2053 }
2054
2055 /*
2056  * Called when a new fs request (rq) is added (to cfqq). Check if there's
2057  * something we should do about it
2058  */
2059 static void
2060 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2061                 struct request *rq)
2062 {
2063         struct cfq_io_context *cic = RQ_CIC(rq);
2064
2065         cfqd->rq_queued++;
2066         if (rq_is_meta(rq))
2067                 cfqq->meta_pending++;
2068
2069         cfq_update_io_thinktime(cfqd, cic);
2070         cfq_update_io_seektime(cfqd, cic, rq);
2071         cfq_update_idle_window(cfqd, cfqq, cic);
2072
2073         cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2074
2075         if (cfqq == cfqd->active_queue) {
2076                 /*
2077                  * Remember that we saw a request from this process, but
2078                  * don't start queuing just yet. Otherwise we risk seeing lots
2079                  * of tiny requests, because we disrupt the normal plugging
2080                  * and merging. If the request is already larger than a single
2081                  * page, let it rip immediately. For that case we assume that
2082                  * merging is already done. Ditto for a busy system that
2083                  * has other work pending, don't risk delaying until the
2084                  * idle timer unplug to continue working.
2085                  */
2086                 if (cfq_cfqq_wait_request(cfqq)) {
2087                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2088                             cfqd->busy_queues > 1) {
2089                                 del_timer(&cfqd->idle_slice_timer);
2090                         __blk_run_queue(cfqd->queue);
2091                         }
2092                         cfq_mark_cfqq_must_dispatch(cfqq);
2093                 }
2094         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2095                 /*
2096                  * not the active queue - expire current slice if it is
2097                  * idle and has expired it's mean thinktime or this new queue
2098                  * has some old slice time left and is of higher priority or
2099                  * this new queue is RT and the current one is BE
2100                  */
2101                 cfq_preempt_queue(cfqd, cfqq);
2102                 __blk_run_queue(cfqd->queue);
2103         }
2104 }
2105
2106 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2107 {
2108         struct cfq_data *cfqd = q->elevator->elevator_data;
2109         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2110
2111         cfq_log_cfqq(cfqd, cfqq, "insert_request");
2112         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2113
2114         cfq_add_rq_rb(rq);
2115
2116         list_add_tail(&rq->queuelist, &cfqq->fifo);
2117
2118         cfq_rq_enqueued(cfqd, cfqq, rq);
2119 }
2120
2121 /*
2122  * Update hw_tag based on peak queue depth over 50 samples under
2123  * sufficient load.
2124  */
2125 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2126 {
2127         if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
2128                 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
2129
2130         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2131             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
2132                 return;
2133
2134         if (cfqd->hw_tag_samples++ < 50)
2135                 return;
2136
2137         if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2138                 cfqd->hw_tag = 1;
2139         else
2140                 cfqd->hw_tag = 0;
2141
2142         cfqd->hw_tag_samples = 0;
2143         cfqd->rq_in_driver_peak = 0;
2144 }
2145
2146 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2147 {
2148         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2149         struct cfq_data *cfqd = cfqq->cfqd;
2150         const int sync = rq_is_sync(rq);
2151         unsigned long now;
2152
2153         now = jiffies;
2154         cfq_log_cfqq(cfqd, cfqq, "complete");
2155
2156         cfq_update_hw_tag(cfqd);
2157
2158         WARN_ON(!cfqd->rq_in_driver);
2159         WARN_ON(!cfqq->dispatched);
2160         cfqd->rq_in_driver--;
2161         cfqq->dispatched--;
2162
2163         if (cfq_cfqq_sync(cfqq))
2164                 cfqd->sync_flight--;
2165
2166         if (sync)
2167                 RQ_CIC(rq)->last_end_request = now;
2168
2169         /*
2170          * If this is the active queue, check if it needs to be expired,
2171          * or if we want to idle in case it has no pending requests.
2172          */
2173         if (cfqd->active_queue == cfqq) {
2174                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2175
2176                 if (cfq_cfqq_slice_new(cfqq)) {
2177                         cfq_set_prio_slice(cfqd, cfqq);
2178                         cfq_clear_cfqq_slice_new(cfqq);
2179                 }
2180                 /*
2181                  * If there are no requests waiting in this queue, and
2182                  * there are other queues ready to issue requests, AND
2183                  * those other queues are issuing requests within our
2184                  * mean seek distance, give them a chance to run instead
2185                  * of idling.
2186                  */
2187                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2188                         cfq_slice_expired(cfqd, 1);
2189                 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2190                          sync && !rq_noidle(rq))
2191                         cfq_arm_slice_timer(cfqd);
2192         }
2193
2194         if (!cfqd->rq_in_driver)
2195                 cfq_schedule_dispatch(cfqd);
2196 }
2197
2198 /*
2199  * we temporarily boost lower priority queues if they are holding fs exclusive
2200  * resources. they are boosted to normal prio (CLASS_BE/4)
2201  */
2202 static void cfq_prio_boost(struct cfq_queue *cfqq)
2203 {
2204         if (has_fs_excl()) {
2205                 /*
2206                  * boost idle prio on transactions that would lock out other
2207                  * users of the filesystem
2208                  */
2209                 if (cfq_class_idle(cfqq))
2210                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
2211                 if (cfqq->ioprio > IOPRIO_NORM)
2212                         cfqq->ioprio = IOPRIO_NORM;
2213         } else {
2214                 /*
2215                  * check if we need to unboost the queue
2216                  */
2217                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2218                         cfqq->ioprio_class = cfqq->org_ioprio_class;
2219                 if (cfqq->ioprio != cfqq->org_ioprio)
2220                         cfqq->ioprio = cfqq->org_ioprio;
2221         }
2222 }
2223
2224 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2225 {
2226         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
2227             !cfq_cfqq_must_alloc_slice(cfqq)) {
2228                 cfq_mark_cfqq_must_alloc_slice(cfqq);
2229                 return ELV_MQUEUE_MUST;
2230         }
2231
2232         return ELV_MQUEUE_MAY;
2233 }
2234
2235 static int cfq_may_queue(struct request_queue *q, int rw)
2236 {
2237         struct cfq_data *cfqd = q->elevator->elevator_data;
2238         struct task_struct *tsk = current;
2239         struct cfq_io_context *cic;
2240         struct cfq_queue *cfqq;
2241
2242         /*
2243          * don't force setup of a queue from here, as a call to may_queue
2244          * does not necessarily imply that a request actually will be queued.
2245          * so just lookup a possibly existing queue, or return 'may queue'
2246          * if that fails
2247          */
2248         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2249         if (!cic)
2250                 return ELV_MQUEUE_MAY;
2251
2252         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2253         if (cfqq) {
2254                 cfq_init_prio_data(cfqq, cic->ioc);
2255                 cfq_prio_boost(cfqq);
2256
2257                 return __cfq_may_queue(cfqq);
2258         }
2259
2260         return ELV_MQUEUE_MAY;
2261 }
2262
2263 /*
2264  * queue lock held here
2265  */
2266 static void cfq_put_request(struct request *rq)
2267 {
2268         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2269
2270         if (cfqq) {
2271                 const int rw = rq_data_dir(rq);
2272
2273                 BUG_ON(!cfqq->allocated[rw]);
2274                 cfqq->allocated[rw]--;
2275
2276                 put_io_context(RQ_CIC(rq)->ioc);
2277
2278                 rq->elevator_private = NULL;
2279                 rq->elevator_private2 = NULL;
2280
2281                 cfq_put_queue(cfqq);
2282         }
2283 }
2284
2285 /*
2286  * Allocate cfq data structures associated with this request.
2287  */
2288 static int
2289 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2290 {
2291         struct cfq_data *cfqd = q->elevator->elevator_data;
2292         struct cfq_io_context *cic;
2293         const int rw = rq_data_dir(rq);
2294         const int is_sync = rq_is_sync(rq);
2295         struct cfq_queue *cfqq;
2296         unsigned long flags;
2297
2298         might_sleep_if(gfp_mask & __GFP_WAIT);
2299
2300         cic = cfq_get_io_context(cfqd, gfp_mask);
2301
2302         spin_lock_irqsave(q->queue_lock, flags);
2303
2304         if (!cic)
2305                 goto queue_fail;
2306
2307         cfqq = cic_to_cfqq(cic, is_sync);
2308         if (!cfqq) {
2309                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2310
2311                 if (!cfqq)
2312                         goto queue_fail;
2313
2314                 cic_set_cfqq(cic, cfqq, is_sync);
2315         }
2316
2317         cfqq->allocated[rw]++;
2318         cfq_clear_cfqq_must_alloc(cfqq);
2319         atomic_inc(&cfqq->ref);
2320
2321         spin_unlock_irqrestore(q->queue_lock, flags);
2322
2323         rq->elevator_private = cic;
2324         rq->elevator_private2 = cfqq;
2325         return 0;
2326
2327 queue_fail:
2328         if (cic)
2329                 put_io_context(cic->ioc);
2330
2331         cfq_schedule_dispatch(cfqd);
2332         spin_unlock_irqrestore(q->queue_lock, flags);
2333         cfq_log(cfqd, "set_request fail");
2334         return 1;
2335 }
2336
2337 static void cfq_kick_queue(struct work_struct *work)
2338 {
2339         struct cfq_data *cfqd =
2340                 container_of(work, struct cfq_data, unplug_work);
2341         struct request_queue *q = cfqd->queue;
2342
2343         spin_lock_irq(q->queue_lock);
2344         __blk_run_queue(cfqd->queue);
2345         spin_unlock_irq(q->queue_lock);
2346 }
2347
2348 /*
2349  * Timer running if the active_queue is currently idling inside its time slice
2350  */
2351 static void cfq_idle_slice_timer(unsigned long data)
2352 {
2353         struct cfq_data *cfqd = (struct cfq_data *) data;
2354         struct cfq_queue *cfqq;
2355         unsigned long flags;
2356         int timed_out = 1;
2357
2358         cfq_log(cfqd, "idle timer fired");
2359
2360         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2361
2362         cfqq = cfqd->active_queue;
2363         if (cfqq) {
2364                 timed_out = 0;
2365
2366                 /*
2367                  * We saw a request before the queue expired, let it through
2368                  */
2369                 if (cfq_cfqq_must_dispatch(cfqq))
2370                         goto out_kick;
2371
2372                 /*
2373                  * expired
2374                  */
2375                 if (cfq_slice_used(cfqq))
2376                         goto expire;
2377
2378                 /*
2379                  * only expire and reinvoke request handler, if there are
2380                  * other queues with pending requests
2381                  */
2382                 if (!cfqd->busy_queues)
2383                         goto out_cont;
2384
2385                 /*
2386                  * not expired and it has a request pending, let it dispatch
2387                  */
2388                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2389                         goto out_kick;
2390         }
2391 expire:
2392         cfq_slice_expired(cfqd, timed_out);
2393 out_kick:
2394         cfq_schedule_dispatch(cfqd);
2395 out_cont:
2396         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2397 }
2398
2399 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2400 {
2401         del_timer_sync(&cfqd->idle_slice_timer);
2402         cancel_work_sync(&cfqd->unplug_work);
2403 }
2404
2405 static void cfq_put_async_queues(struct cfq_data *cfqd)
2406 {
2407         int i;
2408
2409         for (i = 0; i < IOPRIO_BE_NR; i++) {
2410                 if (cfqd->async_cfqq[0][i])
2411                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2412                 if (cfqd->async_cfqq[1][i])
2413                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2414         }
2415
2416         if (cfqd->async_idle_cfqq)
2417                 cfq_put_queue(cfqd->async_idle_cfqq);
2418 }
2419
2420 static void cfq_exit_queue(struct elevator_queue *e)
2421 {
2422         struct cfq_data *cfqd = e->elevator_data;
2423         struct request_queue *q = cfqd->queue;
2424
2425         cfq_shutdown_timer_wq(cfqd);
2426
2427         spin_lock_irq(q->queue_lock);
2428
2429         if (cfqd->active_queue)
2430                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2431
2432         while (!list_empty(&cfqd->cic_list)) {
2433                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2434                                                         struct cfq_io_context,
2435                                                         queue_list);
2436
2437                 __cfq_exit_single_io_context(cfqd, cic);
2438         }
2439
2440         cfq_put_async_queues(cfqd);
2441
2442         spin_unlock_irq(q->queue_lock);
2443
2444         cfq_shutdown_timer_wq(cfqd);
2445
2446         kfree(cfqd);
2447 }
2448
2449 static void *cfq_init_queue(struct request_queue *q)
2450 {
2451         struct cfq_data *cfqd;
2452         int i;
2453
2454         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2455         if (!cfqd)
2456                 return NULL;
2457
2458         cfqd->service_tree = CFQ_RB_ROOT;
2459
2460         /*
2461          * Not strictly needed (since RB_ROOT just clears the node and we
2462          * zeroed cfqd on alloc), but better be safe in case someone decides
2463          * to add magic to the rb code
2464          */
2465         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2466                 cfqd->prio_trees[i] = RB_ROOT;
2467
2468         INIT_LIST_HEAD(&cfqd->cic_list);
2469
2470         cfqd->queue = q;
2471
2472         init_timer(&cfqd->idle_slice_timer);
2473         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2474         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2475
2476         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2477
2478         cfqd->cfq_quantum = cfq_quantum;
2479         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2480         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2481         cfqd->cfq_back_max = cfq_back_max;
2482         cfqd->cfq_back_penalty = cfq_back_penalty;
2483         cfqd->cfq_slice[0] = cfq_slice_async;
2484         cfqd->cfq_slice[1] = cfq_slice_sync;
2485         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2486         cfqd->cfq_slice_idle = cfq_slice_idle;
2487         cfqd->hw_tag = 1;
2488
2489         return cfqd;
2490 }
2491
2492 static void cfq_slab_kill(void)
2493 {
2494         /*
2495          * Caller already ensured that pending RCU callbacks are completed,
2496          * so we should have no busy allocations at this point.
2497          */
2498         if (cfq_pool)
2499                 kmem_cache_destroy(cfq_pool);
2500         if (cfq_ioc_pool)
2501                 kmem_cache_destroy(cfq_ioc_pool);
2502 }
2503
2504 static int __init cfq_slab_setup(void)
2505 {
2506         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2507         if (!cfq_pool)
2508                 goto fail;
2509
2510         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2511         if (!cfq_ioc_pool)
2512                 goto fail;
2513
2514         return 0;
2515 fail:
2516         cfq_slab_kill();
2517         return -ENOMEM;
2518 }
2519
2520 /*
2521  * sysfs parts below -->
2522  */
2523 static ssize_t
2524 cfq_var_show(unsigned int var, char *page)
2525 {
2526         return sprintf(page, "%d\n", var);
2527 }
2528
2529 static ssize_t
2530 cfq_var_store(unsigned int *var, const char *page, size_t count)
2531 {
2532         char *p = (char *) page;
2533
2534         *var = simple_strtoul(p, &p, 10);
2535         return count;
2536 }
2537
2538 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2539 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
2540 {                                                                       \
2541         struct cfq_data *cfqd = e->elevator_data;                       \
2542         unsigned int __data = __VAR;                                    \
2543         if (__CONV)                                                     \
2544                 __data = jiffies_to_msecs(__data);                      \
2545         return cfq_var_show(__data, (page));                            \
2546 }
2547 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2548 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2549 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2550 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2551 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2552 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2553 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2554 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2555 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2556 #undef SHOW_FUNCTION
2557
2558 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2559 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2560 {                                                                       \
2561         struct cfq_data *cfqd = e->elevator_data;                       \
2562         unsigned int __data;                                            \
2563         int ret = cfq_var_store(&__data, (page), count);                \
2564         if (__data < (MIN))                                             \
2565                 __data = (MIN);                                         \
2566         else if (__data > (MAX))                                        \
2567                 __data = (MAX);                                         \
2568         if (__CONV)                                                     \
2569                 *(__PTR) = msecs_to_jiffies(__data);                    \
2570         else                                                            \
2571                 *(__PTR) = __data;                                      \
2572         return ret;                                                     \
2573 }
2574 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2575 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2576                 UINT_MAX, 1);
2577 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2578                 UINT_MAX, 1);
2579 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2580 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2581                 UINT_MAX, 0);
2582 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2583 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2584 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2585 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2586                 UINT_MAX, 0);
2587 #undef STORE_FUNCTION
2588
2589 #define CFQ_ATTR(name) \
2590         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2591
2592 static struct elv_fs_entry cfq_attrs[] = {
2593         CFQ_ATTR(quantum),
2594         CFQ_ATTR(fifo_expire_sync),
2595         CFQ_ATTR(fifo_expire_async),
2596         CFQ_ATTR(back_seek_max),
2597         CFQ_ATTR(back_seek_penalty),
2598         CFQ_ATTR(slice_sync),
2599         CFQ_ATTR(slice_async),
2600         CFQ_ATTR(slice_async_rq),
2601         CFQ_ATTR(slice_idle),
2602         __ATTR_NULL
2603 };
2604
2605 static struct elevator_type iosched_cfq = {
2606         .ops = {
2607                 .elevator_merge_fn =            cfq_merge,
2608                 .elevator_merged_fn =           cfq_merged_request,
2609                 .elevator_merge_req_fn =        cfq_merged_requests,
2610                 .elevator_allow_merge_fn =      cfq_allow_merge,
2611                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2612                 .elevator_add_req_fn =          cfq_insert_request,
2613                 .elevator_activate_req_fn =     cfq_activate_request,
2614                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2615                 .elevator_queue_empty_fn =      cfq_queue_empty,
2616                 .elevator_completed_req_fn =    cfq_completed_request,
2617                 .elevator_former_req_fn =       elv_rb_former_request,
2618                 .elevator_latter_req_fn =       elv_rb_latter_request,
2619                 .elevator_set_req_fn =          cfq_set_request,
2620                 .elevator_put_req_fn =          cfq_put_request,
2621                 .elevator_may_queue_fn =        cfq_may_queue,
2622                 .elevator_init_fn =             cfq_init_queue,
2623                 .elevator_exit_fn =             cfq_exit_queue,
2624                 .trim =                         cfq_free_io_context,
2625         },
2626         .elevator_attrs =       cfq_attrs,
2627         .elevator_name =        "cfq",
2628         .elevator_owner =       THIS_MODULE,
2629 };
2630
2631 static int __init cfq_init(void)
2632 {
2633         /*
2634          * could be 0 on HZ < 1000 setups
2635          */
2636         if (!cfq_slice_async)
2637                 cfq_slice_async = 1;
2638         if (!cfq_slice_idle)
2639                 cfq_slice_idle = 1;
2640
2641         if (cfq_slab_setup())
2642                 return -ENOMEM;
2643
2644         elv_register(&iosched_cfq);
2645
2646         return 0;
2647 }
2648
2649 static void __exit cfq_exit(void)
2650 {
2651         DECLARE_COMPLETION_ONSTACK(all_gone);
2652         elv_unregister(&iosched_cfq);
2653         ioc_gone = &all_gone;
2654         /* ioc_gone's update must be visible before reading ioc_count */
2655         smp_wmb();
2656
2657         /*
2658          * this also protects us from entering cfq_slab_kill() with
2659          * pending RCU callbacks
2660          */
2661         if (elv_ioc_count_read(ioc_count))
2662                 wait_for_completion(&all_gone);
2663         cfq_slab_kill();
2664 }
2665
2666 module_init(cfq_init);
2667 module_exit(cfq_exit);
2668
2669 MODULE_AUTHOR("Jens Axboe");
2670 MODULE_LICENSE("GPL");
2671 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");