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, cfq_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 process-grouping structure
  75  */
  76 struct cfq_queue {
  77         /* reference count */
  78         atomic_t ref;
  79         /* various state flags, see below */
  80         unsigned int flags;
  81         /* parent cfq_data */
  82         struct cfq_data *cfqd;
  83         /* service_tree member */
  84         struct rb_node rb_node;
  85         /* service_tree key */
  86         unsigned long rb_key;
  87         /* prio tree member */
  88         struct rb_node p_node;
  89         /* prio tree root we belong to, if any */
  90         struct rb_root *p_root;
  91         /* sorted list of pending requests */
  92         struct rb_root sort_list;
  93         /* if fifo isn't expired, next request to serve */
  94         struct request *next_rq;
  95         /* requests queued in sort_list */
  96         int queued[2];
  97         /* currently allocated requests */
  98         int allocated[2];
  99         /* fifo list of requests in sort_list */
 100         struct list_head fifo;
 101
 102         unsigned long slice_end;
 103         long slice_resid;
 104         unsigned int slice_dispatch;
 105
 106         /* pending metadata requests */
 107         int meta_pending;
 108         /* number of requests that are on the dispatch list or inside driver */
 109         int dispatched;
 110
 111         /* io prio of this group */
 112         unsigned short ioprio, org_ioprio;
 113         unsigned short ioprio_class, org_ioprio_class;
 114
 115         pid_t pid;
 116 };
 117
 118 /*
 119  * Per block device queue structure
 120  */
 121 struct cfq_data {
 122         struct request_queue *queue;
 123
 124         /*
 125          * rr list of queues with requests and the count of them
 126          */
 127         struct cfq_rb_root service_tree;
 128
 129         /*
 130          * Each priority tree is sorted by next_request position.  These
 131          * trees are used when determining if two or more queues are
 132          * interleaving requests (see cfq_close_cooperator).
 133          */
 134         struct rb_root prio_trees[CFQ_PRIO_LISTS];
 135
 136         unsigned int busy_queues;
 137
 138         int rq_in_driver[2];
 139         int sync_flight;
 140
 141         /*
 142          * queue-depth detection
 143          */
 144         int rq_queued;
 145         int hw_tag;
 146         int hw_tag_samples;
 147         int rq_in_driver_peak;
 148
 149         /*
 150          * idle window management
 151          */
 152         struct timer_list idle_slice_timer;
 153         struct work_struct unplug_work;
 154
 155         struct cfq_queue *active_queue;
 156         struct cfq_io_context *active_cic;
 157
 158         /*
 159          * async queue for each priority case
 160          */
 161         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 162         struct cfq_queue *async_idle_cfqq;
 163
 164         sector_t last_position;
 165
 166         /*
 167          * tunables, see top of file
 168          */
 169         unsigned int cfq_quantum;
 170         unsigned int cfq_fifo_expire[2];
 171         unsigned int cfq_back_penalty;
 172         unsigned int cfq_back_max;
 173         unsigned int cfq_slice[2];
 174         unsigned int cfq_slice_async_rq;
 175         unsigned int cfq_slice_idle;
 176         unsigned int cfq_latency;
 177
 178         struct list_head cic_list;
 179
 180         /*
 181          * Fallback dummy cfqq for extreme OOM conditions
 182          */
 183         struct cfq_queue oom_cfqq;
 184
 185         unsigned long last_end_sync_rq;
 186 };
 187
 188 enum cfqq_state_flags {
 189         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 190         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 191         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
 192         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 193         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 194         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 195         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 196         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 197         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 198         CFQ_CFQQ_FLAG_coop,             /* has done a coop jump of the queue */
 199 };
 200
 201 #define CFQ_CFQQ_FNS(name)                                              \
 202 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 203 {                                                                       \
 204         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 205 }                                                                       \
 206 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 207 {                                                                       \
 208         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 209 }                                                                       \
 210 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 211 {                                                                       \
 212         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 213 }
 214
 215 CFQ_CFQQ_FNS(on_rr);
 216 CFQ_CFQQ_FNS(wait_request);
 217 CFQ_CFQQ_FNS(must_dispatch);
 218 CFQ_CFQQ_FNS(must_alloc_slice);
 219 CFQ_CFQQ_FNS(fifo_expire);
 220 CFQ_CFQQ_FNS(idle_window);
 221 CFQ_CFQQ_FNS(prio_changed);
 222 CFQ_CFQQ_FNS(slice_new);
 223 CFQ_CFQQ_FNS(sync);
 224 CFQ_CFQQ_FNS(coop);
 225 #undef CFQ_CFQQ_FNS
 226
 227 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 228         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
 229 #define cfq_log(cfqd, fmt, args...)     \
 230         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 231
 232 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 233 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
 234                                        struct io_context *, gfp_t);
 235 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
 236                                                 struct io_context *);
 237
 238 static inline int rq_in_driver(struct cfq_data *cfqd)
 239 {
 240         return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
 241 }
 242
 243 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
 244                                             bool is_sync)
 245 {
 246         return cic->cfqq[is_sync];
 247 }
 248
 249 static inline void cic_set_cfqq(struct cfq_io_context *cic,
 250                                 struct cfq_queue *cfqq, bool is_sync)
 251 {
 252         cic->cfqq[is_sync] = cfqq;
 253 }
 254
 255 /*
 256  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 257  * set (in which case it could also be direct WRITE).
 258  */
 259 static inline bool cfq_bio_sync(struct bio *bio)
 260 {
 261         return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
 262 }
 263
 264 /*
 265  * scheduler run of queue, if there are requests pending and no one in the
 266  * driver that will restart queueing
 267  */
 268 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 269 {
 270         if (cfqd->busy_queues) {
 271                 cfq_log(cfqd, "schedule dispatch");
 272                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
 273         }
 274 }
 275
 276 static int cfq_queue_empty(struct request_queue *q)
 277 {
 278         struct cfq_data *cfqd = q->elevator->elevator_data;
 279
 280         return !cfqd->busy_queues;
 281 }
 282
 283 /*
 284  * Scale schedule slice based on io priority. Use the sync time slice only
 285  * if a queue is marked sync and has sync io queued. A sync queue with async
 286  * io only, should not get full sync slice length.
 287  */
 288 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
 289                                  unsigned short prio)
 290 {
 291         const int base_slice = cfqd->cfq_slice[sync];
 292
 293         WARN_ON(prio >= IOPRIO_BE_NR);
 294
 295         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 296 }
 297
 298 static inline int
 299 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 300 {
 301         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 302 }
 303
 304 static inline void
 305 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 306 {
 307         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 308         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
 309 }
 310
 311 /*
 312  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
 313  * isn't valid until the first request from the dispatch is activated
 314  * and the slice time set.
 315  */
 316 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
 317 {
 318         if (cfq_cfqq_slice_new(cfqq))
 319                 return 0;
 320         if (time_before(jiffies, cfqq->slice_end))
 321                 return 0;
 322
 323         return 1;
 324 }
 325
 326 /*
 327  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 328  * We choose the request that is closest to the head right now. Distance
 329  * behind the head is penalized and only allowed to a certain extent.
 330  */
 331 static struct request *
 332 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
 333 {
 334         sector_t last, s1, s2, d1 = 0, d2 = 0;
 335         unsigned long back_max;
 336 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 337 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 338         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 339
 340         if (rq1 == NULL || rq1 == rq2)
 341                 return rq2;
 342         if (rq2 == NULL)
 343                 return rq1;
 344
 345         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 346                 return rq1;
 347         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 348                 return rq2;
 349         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
 350                 return rq1;
 351         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
 352                 return rq2;
 353
 354         s1 = blk_rq_pos(rq1);
 355         s2 = blk_rq_pos(rq2);
 356
 357         last = cfqd->last_position;
 358
 359         /*
 360          * by definition, 1KiB is 2 sectors
 361          */
 362         back_max = cfqd->cfq_back_max * 2;
 363
 364         /*
 365          * Strict one way elevator _except_ in the case where we allow
 366          * short backward seeks which are biased as twice the cost of a
 367          * similar forward seek.
 368          */
 369         if (s1 >= last)
 370                 d1 = s1 - last;
 371         else if (s1 + back_max >= last)
 372                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 373         else
 374                 wrap |= CFQ_RQ1_WRAP;
 375
 376         if (s2 >= last)
 377                 d2 = s2 - last;
 378         else if (s2 + back_max >= last)
 379                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 380         else
 381                 wrap |= CFQ_RQ2_WRAP;
 382
 383         /* Found required data */
 384
 385         /*
 386          * By doing switch() on the bit mask "wrap" we avoid having to
 387          * check two variables for all permutations: --> faster!
 388          */
 389         switch (wrap) {
 390         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 391                 if (d1 < d2)
 392                         return rq1;
 393                 else if (d2 < d1)
 394                         return rq2;
 395                 else {
 396                         if (s1 >= s2)
 397                                 return rq1;
 398                         else
 399                                 return rq2;
 400                 }
 401
 402         case CFQ_RQ2_WRAP:
 403                 return rq1;
 404         case CFQ_RQ1_WRAP:
 405                 return rq2;
 406         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
 407         default:
 408                 /*
 409                  * Since both rqs are wrapped,
 410                  * start with the one that's further behind head
 411                  * (--> only *one* back seek required),
 412                  * since back seek takes more time than forward.
 413                  */
 414                 if (s1 <= s2)
 415                         return rq1;
 416                 else
 417                         return rq2;
 418         }
 419 }
 420
 421 /*
 422  * The below is leftmost cache rbtree addon
 423  */
 424 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
 425 {
 426         if (!root->left)
 427                 root->left = rb_first(&root->rb);
 428
 429         if (root->left)
 430                 return rb_entry(root->left, struct cfq_queue, rb_node);
 431
 432         return NULL;
 433 }
 434
 435 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
 436 {
 437         rb_erase(n, root);
 438         RB_CLEAR_NODE(n);
 439 }
 440
 441 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
 442 {
 443         if (root->left == n)
 444                 root->left = NULL;
 445         rb_erase_init(n, &root->rb);
 446 }
 447
 448 /*
 449  * would be nice to take fifo expire time into account as well
 450  */
 451 static struct request *
 452 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 453                   struct request *last)
 454 {
 455         struct rb_node *rbnext = rb_next(&last->rb_node);
 456         struct rb_node *rbprev = rb_prev(&last->rb_node);
 457         struct request *next = NULL, *prev = NULL;
 458
 459         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
 460
 461         if (rbprev)
 462                 prev = rb_entry_rq(rbprev);
 463
 464         if (rbnext)
 465                 next = rb_entry_rq(rbnext);
 466         else {
 467                 rbnext = rb_first(&cfqq->sort_list);
 468                 if (rbnext && rbnext != &last->rb_node)
 469                         next = rb_entry_rq(rbnext);
 470         }
 471
 472         return cfq_choose_req(cfqd, next, prev);
 473 }
 474
 475 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
 476                                       struct cfq_queue *cfqq)
 477 {
 478         /*
 479          * just an approximation, should be ok.
 480          */
 481         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
 482                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
 483 }
 484
 485 /*
 486  * The cfqd->service_tree holds all pending cfq_queue's that have
 487  * requests waiting to be processed. It is sorted in the order that
 488  * we will service the queues.
 489  */
 490 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 491                                  bool add_front)
 492 {
 493         struct rb_node **p, *parent;
 494         struct cfq_queue *__cfqq;
 495         unsigned long rb_key;
 496         int left;
 497
 498         if (cfq_class_idle(cfqq)) {
 499                 rb_key = CFQ_IDLE_DELAY;
 500                 parent = rb_last(&cfqd->service_tree.rb);
 501                 if (parent && parent != &cfqq->rb_node) {
 502                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
 503                         rb_key += __cfqq->rb_key;
 504                 } else
 505                         rb_key += jiffies;
 506         } else if (!add_front) {
 507                 /*
 508                  * Get our rb key offset. Subtract any residual slice
 509                  * value carried from last service. A negative resid
 510                  * count indicates slice overrun, and this should position
 511                  * the next service time further away in the tree.
 512                  */
 513                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
 514                 rb_key -= cfqq->slice_resid;
 515                 cfqq->slice_resid = 0;
 516         } else {
 517                 rb_key = -HZ;
 518                 __cfqq = cfq_rb_first(&cfqd->service_tree);
 519                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
 520         }
 521
 522         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
 523                 /*
 524                  * same position, nothing more to do
 525                  */
 526                 if (rb_key == cfqq->rb_key)
 527                         return;
 528
 529                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
 530         }
 531
 532         left = 1;
 533         parent = NULL;
 534         p = &cfqd->service_tree.rb.rb_node;
 535         while (*p) {
 536                 struct rb_node **n;
 537
 538                 parent = *p;
 539                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
 540
 541                 /*
 542                  * sort RT queues first, we always want to give
 543                  * preference to them. IDLE queues goes to the back.
 544                  * after that, sort on the next service time.
 545                  */
 546                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
 547                         n = &(*p)->rb_left;
 548                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
 549                         n = &(*p)->rb_right;
 550                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
 551                         n = &(*p)->rb_left;
 552                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
 553                         n = &(*p)->rb_right;
 554                 else if (time_before(rb_key, __cfqq->rb_key))
 555                         n = &(*p)->rb_left;
 556                 else
 557                         n = &(*p)->rb_right;
 558
 559                 if (n == &(*p)->rb_right)
 560                         left = 0;
 561
 562                 p = n;
 563         }
 564
 565         if (left)
 566                 cfqd->service_tree.left = &cfqq->rb_node;
 567
 568         cfqq->rb_key = rb_key;
 569         rb_link_node(&cfqq->rb_node, parent, p);
 570         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
 571 }
 572
 573 static struct cfq_queue *
 574 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
 575                      sector_t sector, struct rb_node **ret_parent,
 576                      struct rb_node ***rb_link)
 577 {
 578         struct rb_node **p, *parent;
 579         struct cfq_queue *cfqq = NULL;
 580
 581         parent = NULL;
 582         p = &root->rb_node;
 583         while (*p) {
 584                 struct rb_node **n;
 585
 586                 parent = *p;
 587                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
 588
 589                 /*
 590                  * Sort strictly based on sector.  Smallest to the left,
 591                  * largest to the right.
 592                  */
 593                 if (sector > blk_rq_pos(cfqq->next_rq))
 594                         n = &(*p)->rb_right;
 595                 else if (sector < blk_rq_pos(cfqq->next_rq))
 596                         n = &(*p)->rb_left;
 597                 else
 598                         break;
 599                 p = n;
 600                 cfqq = NULL;
 601         }
 602
 603         *ret_parent = parent;
 604         if (rb_link)
 605                 *rb_link = p;
 606         return cfqq;
 607 }
 608
 609 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 610 {
 611         struct rb_node **p, *parent;
 612         struct cfq_queue *__cfqq;
 613
 614         if (cfqq->p_root) {
 615                 rb_erase(&cfqq->p_node, cfqq->p_root);
 616                 cfqq->p_root = NULL;
 617         }
 618
 619         if (cfq_class_idle(cfqq))
 620                 return;
 621         if (!cfqq->next_rq)
 622                 return;
 623
 624         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
 625         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
 626                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
 627         if (!__cfqq) {
 628                 rb_link_node(&cfqq->p_node, parent, p);
 629                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
 630         } else
 631                 cfqq->p_root = NULL;
 632 }
 633
 634 /*
 635  * Update cfqq's position in the service tree.
 636  */
 637 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 638 {
 639         /*
 640          * Resorting requires the cfqq to be on the RR list already.
 641          */
 642         if (cfq_cfqq_on_rr(cfqq)) {
 643                 cfq_service_tree_add(cfqd, cfqq, 0);
 644                 cfq_prio_tree_add(cfqd, cfqq);
 645         }
 646 }
 647
 648 /*
 649  * add to busy list of queues for service, trying to be fair in ordering
 650  * the pending list according to last request service
 651  */
 652 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 653 {
 654         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
 655         BUG_ON(cfq_cfqq_on_rr(cfqq));
 656         cfq_mark_cfqq_on_rr(cfqq);
 657         cfqd->busy_queues++;
 658
 659         cfq_resort_rr_list(cfqd, cfqq);
 660 }
 661
 662 /*
 663  * Called when the cfqq no longer has requests pending, remove it from
 664  * the service tree.
 665  */
 666 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 667 {
 668         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
 669         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 670         cfq_clear_cfqq_on_rr(cfqq);
 671
 672         if (!RB_EMPTY_NODE(&cfqq->rb_node))
 673                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
 674         if (cfqq->p_root) {
 675                 rb_erase(&cfqq->p_node, cfqq->p_root);
 676                 cfqq->p_root = NULL;
 677         }
 678
 679         BUG_ON(!cfqd->busy_queues);
 680         cfqd->busy_queues--;
 681 }
 682
 683 /*
 684  * rb tree support functions
 685  */
 686 static void cfq_del_rq_rb(struct request *rq)
 687 {
 688         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 689         struct cfq_data *cfqd = cfqq->cfqd;
 690         const int sync = rq_is_sync(rq);
 691
 692         BUG_ON(!cfqq->queued[sync]);
 693         cfqq->queued[sync]--;
 694
 695         elv_rb_del(&cfqq->sort_list, rq);
 696
 697         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
 698                 cfq_del_cfqq_rr(cfqd, cfqq);
 699 }
 700
 701 static void cfq_add_rq_rb(struct request *rq)
 702 {
 703         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 704         struct cfq_data *cfqd = cfqq->cfqd;
 705         struct request *__alias, *prev;
 706
 707         cfqq->queued[rq_is_sync(rq)]++;
 708
 709         /*
 710          * looks a little odd, but the first insert might return an alias.
 711          * if that happens, put the alias on the dispatch list
 712          */
 713         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
 714                 cfq_dispatch_insert(cfqd->queue, __alias);
 715
 716         if (!cfq_cfqq_on_rr(cfqq))
 717                 cfq_add_cfqq_rr(cfqd, cfqq);
 718
 719         /*
 720          * check if this request is a better next-serve candidate
 721          */
 722         prev = cfqq->next_rq;
 723         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
 724
 725         /*
 726          * adjust priority tree position, if ->next_rq changes
 727          */
 728         if (prev != cfqq->next_rq)
 729                 cfq_prio_tree_add(cfqd, cfqq);
 730
 731         BUG_ON(!cfqq->next_rq);
 732 }
 733
 734 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
 735 {
 736         elv_rb_del(&cfqq->sort_list, rq);
 737         cfqq->queued[rq_is_sync(rq)]--;
 738         cfq_add_rq_rb(rq);
 739 }
 740
 741 static struct request *
 742 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 743 {
 744         struct task_struct *tsk = current;
 745         struct cfq_io_context *cic;
 746         struct cfq_queue *cfqq;
 747
 748         cic = cfq_cic_lookup(cfqd, tsk->io_context);
 749         if (!cic)
 750                 return NULL;
 751
 752         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
 753         if (cfqq) {
 754                 sector_t sector = bio->bi_sector + bio_sectors(bio);
 755
 756                 return elv_rb_find(&cfqq->sort_list, sector);
 757         }
 758
 759         return NULL;
 760 }
 761
 762 static void cfq_activate_request(struct request_queue *q, struct request *rq)
 763 {
 764         struct cfq_data *cfqd = q->elevator->elevator_data;
 765
 766         cfqd->rq_in_driver[rq_is_sync(rq)]++;
 767         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
 768                                                 rq_in_driver(cfqd));
 769
 770         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
 771 }
 772
 773 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
 774 {
 775         struct cfq_data *cfqd = q->elevator->elevator_data;
 776         const int sync = rq_is_sync(rq);
 777
 778         WARN_ON(!cfqd->rq_in_driver[sync]);
 779         cfqd->rq_in_driver[sync]--;
 780         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
 781                                                 rq_in_driver(cfqd));
 782 }
 783
 784 static void cfq_remove_request(struct request *rq)
 785 {
 786         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 787
 788         if (cfqq->next_rq == rq)
 789                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
 790
 791         list_del_init(&rq->queuelist);
 792         cfq_del_rq_rb(rq);
 793
 794         cfqq->cfqd->rq_queued--;
 795         if (rq_is_meta(rq)) {
 796                 WARN_ON(!cfqq->meta_pending);
 797                 cfqq->meta_pending--;
 798         }
 799 }
 800
 801 static int cfq_merge(struct request_queue *q, struct request **req,
 802                      struct bio *bio)
 803 {
 804         struct cfq_data *cfqd = q->elevator->elevator_data;
 805         struct request *__rq;
 806
 807         __rq = cfq_find_rq_fmerge(cfqd, bio);
 808         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 809                 *req = __rq;
 810                 return ELEVATOR_FRONT_MERGE;
 811         }
 812
 813         return ELEVATOR_NO_MERGE;
 814 }
 815
 816 static void cfq_merged_request(struct request_queue *q, struct request *req,
 817                                int type)
 818 {
 819         if (type == ELEVATOR_FRONT_MERGE) {
 820                 struct cfq_queue *cfqq = RQ_CFQQ(req);
 821
 822                 cfq_reposition_rq_rb(cfqq, req);
 823         }
 824 }
 825
 826 static void
 827 cfq_merged_requests(struct request_queue *q, struct request *rq,
 828                     struct request *next)
 829 {
 830         /*
 831          * reposition in fifo if next is older than rq
 832          */
 833         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 834             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
 835                 list_move(&rq->queuelist, &next->queuelist);
 836                 rq_set_fifo_time(rq, rq_fifo_time(next));
 837         }
 838
 839         cfq_remove_request(next);
 840 }
 841
 842 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
 843                            struct bio *bio)
 844 {
 845         struct cfq_data *cfqd = q->elevator->elevator_data;
 846         struct cfq_io_context *cic;
 847         struct cfq_queue *cfqq;
 848
 849         /*
 850          * Disallow merge of a sync bio into an async request.
 851          */
 852         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
 853                 return false;
 854
 855         /*
 856          * Lookup the cfqq that this bio will be queued with. Allow
 857          * merge only if rq is queued there.
 858          */
 859         cic = cfq_cic_lookup(cfqd, current->io_context);
 860         if (!cic)
 861                 return false;
 862
 863         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
 864         return cfqq == RQ_CFQQ(rq);
 865 }
 866
 867 static void __cfq_set_active_queue(struct cfq_data *cfqd,
 868                                    struct cfq_queue *cfqq)
 869 {
 870         if (cfqq) {
 871                 cfq_log_cfqq(cfqd, cfqq, "set_active");
 872                 cfqq->slice_end = 0;
 873                 cfqq->slice_dispatch = 0;
 874
 875                 cfq_clear_cfqq_wait_request(cfqq);
 876                 cfq_clear_cfqq_must_dispatch(cfqq);
 877                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 878                 cfq_clear_cfqq_fifo_expire(cfqq);
 879                 cfq_mark_cfqq_slice_new(cfqq);
 880
 881                 del_timer(&cfqd->idle_slice_timer);
 882         }
 883
 884         cfqd->active_queue = cfqq;
 885 }
 886
 887 /*
 888  * current cfqq expired its slice (or was too idle), select new one
 889  */
 890 static void
 891 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 892                     bool timed_out)
 893 {
 894         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
 895
 896         if (cfq_cfqq_wait_request(cfqq))
 897                 del_timer(&cfqd->idle_slice_timer);
 898
 899         cfq_clear_cfqq_wait_request(cfqq);
 900
 901         /*
 902          * store what was left of this slice, if the queue idled/timed out
 903          */
 904         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
 905                 cfqq->slice_resid = cfqq->slice_end - jiffies;
 906                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
 907         }
 908
 909         cfq_resort_rr_list(cfqd, cfqq);
 910
 911         if (cfqq == cfqd->active_queue)
 912                 cfqd->active_queue = NULL;
 913
 914         if (cfqd->active_cic) {
 915                 put_io_context(cfqd->active_cic->ioc);
 916                 cfqd->active_cic = NULL;
 917         }
 918 }
 919
 920 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
 921 {
 922         struct cfq_queue *cfqq = cfqd->active_queue;
 923
 924         if (cfqq)
 925                 __cfq_slice_expired(cfqd, cfqq, timed_out);
 926 }
 927
 928 /*
 929  * Get next queue for service. Unless we have a queue preemption,
 930  * we'll simply select the first cfqq in the service tree.
 931  */
 932 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
 933 {
 934         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
 935                 return NULL;
 936
 937         return cfq_rb_first(&cfqd->service_tree);
 938 }
 939
 940 /*
 941  * Get and set a new active queue for service.
 942  */
 943 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
 944                                               struct cfq_queue *cfqq)
 945 {
 946         if (!cfqq) {
 947                 cfqq = cfq_get_next_queue(cfqd);
 948                 if (cfqq)
 949                         cfq_clear_cfqq_coop(cfqq);
 950         }
 951
 952         __cfq_set_active_queue(cfqd, cfqq);
 953         return cfqq;
 954 }
 955
 956 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
 957                                           struct request *rq)
 958 {
 959         if (blk_rq_pos(rq) >= cfqd->last_position)
 960                 return blk_rq_pos(rq) - cfqd->last_position;
 961         else
 962                 return cfqd->last_position - blk_rq_pos(rq);
 963 }
 964
 965 #define CIC_SEEK_THR    8 * 1024
 966 #define CIC_SEEKY(cic)  ((cic)->seek_mean > CIC_SEEK_THR)
 967
 968 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
 969 {
 970         struct cfq_io_context *cic = cfqd->active_cic;
 971         sector_t sdist = cic->seek_mean;
 972
 973         if (!sample_valid(cic->seek_samples))
 974                 sdist = CIC_SEEK_THR;
 975
 976         return cfq_dist_from_last(cfqd, rq) <= sdist;
 977 }
 978
 979 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
 980                                     struct cfq_queue *cur_cfqq)
 981 {
 982         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
 983         struct rb_node *parent, *node;
 984         struct cfq_queue *__cfqq;
 985         sector_t sector = cfqd->last_position;
 986
 987         if (RB_EMPTY_ROOT(root))
 988                 return NULL;
 989
 990         /*
 991          * First, if we find a request starting at the end of the last
 992          * request, choose it.
 993          */
 994         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
 995         if (__cfqq)
 996                 return __cfqq;
 997
 998         /*
 999          * If the exact sector wasn't found, the parent of the NULL leaf
1000          * will contain the closest sector.
1001          */
1002         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1003         if (cfq_rq_close(cfqd, __cfqq->next_rq))
1004                 return __cfqq;
1005
1006         if (blk_rq_pos(__cfqq->next_rq) < sector)
1007                 node = rb_next(&__cfqq->p_node);
1008         else
1009                 node = rb_prev(&__cfqq->p_node);
1010         if (!node)
1011                 return NULL;
1012
1013         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1014         if (cfq_rq_close(cfqd, __cfqq->next_rq))
1015                 return __cfqq;
1016
1017         return NULL;
1018 }
1019
1020 /*
1021  * cfqd - obvious
1022  * cur_cfqq - passed in so that we don't decide that the current queue is
1023  *            closely cooperating with itself.
1024  *
1025  * So, basically we're assuming that that cur_cfqq has dispatched at least
1026  * one request, and that cfqd->last_position reflects a position on the disk
1027  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1028  * assumption.
1029  */
1030 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1031                                               struct cfq_queue *cur_cfqq,
1032                                               bool probe)
1033 {
1034         struct cfq_queue *cfqq;
1035
1036         /*
1037          * A valid cfq_io_context is necessary to compare requests against
1038          * the seek_mean of the current cfqq.
1039          */
1040         if (!cfqd->active_cic)
1041                 return NULL;
1042
1043         /*
1044          * We should notice if some of the queues are cooperating, eg
1045          * working closely on the same area of the disk. In that case,
1046          * we can group them together and don't waste time idling.
1047          */
1048         cfqq = cfqq_close(cfqd, cur_cfqq);
1049         if (!cfqq)
1050                 return NULL;
1051
1052         if (cfq_cfqq_coop(cfqq))
1053                 return NULL;
1054
1055         if (!probe)
1056                 cfq_mark_cfqq_coop(cfqq);
1057         return cfqq;
1058 }
1059
1060 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1061 {
1062         struct cfq_queue *cfqq = cfqd->active_queue;
1063         struct cfq_io_context *cic;
1064         unsigned long sl;
1065
1066         /*
1067          * SSD device without seek penalty, disable idling. But only do so
1068          * for devices that support queuing, otherwise we still have a problem
1069          * with sync vs async workloads.
1070          */
1071         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1072                 return;
1073
1074         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1075         WARN_ON(cfq_cfqq_slice_new(cfqq));
1076
1077         /*
1078          * idle is disabled, either manually or by past process history
1079          */
1080         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1081                 return;
1082
1083         /*
1084          * still requests with the driver, don't idle
1085          */
1086         if (rq_in_driver(cfqd))
1087                 return;
1088
1089         /*
1090          * task has exited, don't wait
1091          */
1092         cic = cfqd->active_cic;
1093         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1094                 return;
1095
1096         /*
1097          * If our average think time is larger than the remaining time
1098          * slice, then don't idle. This avoids overrunning the allotted
1099          * time slice.
1100          */
1101         if (sample_valid(cic->ttime_samples) &&
1102             (cfqq->slice_end - jiffies < cic->ttime_mean))
1103                 return;
1104
1105         cfq_mark_cfqq_wait_request(cfqq);
1106
1107         /*
1108          * we don't want to idle for seeks, but we do want to allow
1109          * fair distribution of slice time for a process doing back-to-back
1110          * seeks. so allow a little bit of time for him to submit a new rq
1111          */
1112         sl = cfqd->cfq_slice_idle;
1113         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
1114                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1115
1116         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1117         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1118 }
1119
1120 /*
1121  * Move request from internal lists to the request queue dispatch list.
1122  */
1123 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1124 {
1125         struct cfq_data *cfqd = q->elevator->elevator_data;
1126         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1127
1128         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1129
1130         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1131         cfq_remove_request(rq);
1132         cfqq->dispatched++;
1133         elv_dispatch_sort(q, rq);
1134
1135         if (cfq_cfqq_sync(cfqq))
1136                 cfqd->sync_flight++;
1137 }
1138
1139 /*
1140  * return expired entry, or NULL to just start from scratch in rbtree
1141  */
1142 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1143 {
1144         struct request *rq = NULL;
1145
1146         if (cfq_cfqq_fifo_expire(cfqq))
1147                 return NULL;
1148
1149         cfq_mark_cfqq_fifo_expire(cfqq);
1150
1151         if (list_empty(&cfqq->fifo))
1152                 return NULL;
1153
1154         rq = rq_entry_fifo(cfqq->fifo.next);
1155         if (time_before(jiffies, rq_fifo_time(rq)))
1156                 rq = NULL;
1157
1158         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1159         return rq;
1160 }
1161
1162 static inline int
1163 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1164 {
1165         const int base_rq = cfqd->cfq_slice_async_rq;
1166
1167         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1168
1169         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1170 }
1171
1172 /*
1173  * Select a queue for service. If we have a current active queue,
1174  * check whether to continue servicing it, or retrieve and set a new one.
1175  */
1176 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1177 {
1178         struct cfq_queue *cfqq, *new_cfqq = NULL;
1179
1180         cfqq = cfqd->active_queue;
1181         if (!cfqq)
1182                 goto new_queue;
1183
1184         /*
1185          * The active queue has run out of time, expire it and select new.
1186          */
1187         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1188                 goto expire;
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 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1260 {
1261         unsigned int max_dispatch;
1262
1263         /*
1264          * Drain async requests before we start sync IO
1265          */
1266         if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1267                 return false;
1268
1269         /*
1270          * If this is an async queue and we have sync IO in flight, let it wait
1271          */
1272         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1273                 return false;
1274
1275         max_dispatch = cfqd->cfq_quantum;
1276         if (cfq_class_idle(cfqq))
1277                 max_dispatch = 1;
1278
1279         /*
1280          * Does this cfqq already have too much IO in flight?
1281          */
1282         if (cfqq->dispatched >= max_dispatch) {
1283                 /*
1284                  * idle queue must always only have a single IO in flight
1285                  */
1286                 if (cfq_class_idle(cfqq))
1287                         return false;
1288
1289                 /*
1290                  * We have other queues, don't allow more IO from this one
1291                  */
1292                 if (cfqd->busy_queues > 1)
1293                         return false;
1294
1295                 /*
1296                  * Sole queue user, allow bigger slice
1297                  */
1298                 max_dispatch *= 4;
1299         }
1300
1301         /*
1302          * Async queues must wait a bit before being allowed dispatch.
1303          * We also ramp up the dispatch depth gradually for async IO,
1304          * based on the last sync IO we serviced
1305          */
1306         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1307                 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1308                 unsigned int depth;
1309
1310                 depth = last_sync / cfqd->cfq_slice[1];
1311                 if (!depth && !cfqq->dispatched)
1312                         depth = 1;
1313                 if (depth < max_dispatch)
1314                         max_dispatch = depth;
1315         }
1316
1317         /*
1318          * If we're below the current max, allow a dispatch
1319          */
1320         return cfqq->dispatched < max_dispatch;
1321 }
1322
1323 /*
1324  * Dispatch a request from cfqq, moving them to the request queue
1325  * dispatch list.
1326  */
1327 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1328 {
1329         struct request *rq;
1330
1331         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1332
1333         if (!cfq_may_dispatch(cfqd, cfqq))
1334                 return false;
1335
1336         /*
1337          * follow expired path, else get first next available
1338          */
1339         rq = cfq_check_fifo(cfqq);
1340         if (!rq)
1341                 rq = cfqq->next_rq;
1342
1343         /*
1344          * insert request into driver dispatch list
1345          */
1346         cfq_dispatch_insert(cfqd->queue, rq);
1347
1348         if (!cfqd->active_cic) {
1349                 struct cfq_io_context *cic = RQ_CIC(rq);
1350
1351                 atomic_long_inc(&cic->ioc->refcount);
1352                 cfqd->active_cic = cic;
1353         }
1354
1355         return true;
1356 }
1357
1358 /*
1359  * Find the cfqq that we need to service and move a request from that to the
1360  * dispatch list
1361  */
1362 static int cfq_dispatch_requests(struct request_queue *q, int force)
1363 {
1364         struct cfq_data *cfqd = q->elevator->elevator_data;
1365         struct cfq_queue *cfqq;
1366
1367         if (!cfqd->busy_queues)
1368                 return 0;
1369
1370         if (unlikely(force))
1371                 return cfq_forced_dispatch(cfqd);
1372
1373         cfqq = cfq_select_queue(cfqd);
1374         if (!cfqq)
1375                 return 0;
1376
1377         /*
1378          * Dispatch a request from this cfqq, if it is allowed
1379          */
1380         if (!cfq_dispatch_request(cfqd, cfqq))
1381                 return 0;
1382
1383         cfqq->slice_dispatch++;
1384         cfq_clear_cfqq_must_dispatch(cfqq);
1385
1386         /*
1387          * expire an async queue immediately if it has used up its slice. idle
1388          * queue always expire after 1 dispatch round.
1389          */
1390         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1391             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1392             cfq_class_idle(cfqq))) {
1393                 cfqq->slice_end = jiffies + 1;
1394                 cfq_slice_expired(cfqd, 0);
1395         }
1396
1397         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1398         return 1;
1399 }
1400
1401 /*
1402  * task holds one reference to the queue, dropped when task exits. each rq
1403  * in-flight on this queue also holds a reference, dropped when rq is freed.
1404  *
1405  * queue lock must be held here.
1406  */
1407 static void cfq_put_queue(struct cfq_queue *cfqq)
1408 {
1409         struct cfq_data *cfqd = cfqq->cfqd;
1410
1411         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1412
1413         if (!atomic_dec_and_test(&cfqq->ref))
1414                 return;
1415
1416         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1417         BUG_ON(rb_first(&cfqq->sort_list));
1418         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1419         BUG_ON(cfq_cfqq_on_rr(cfqq));
1420
1421         if (unlikely(cfqd->active_queue == cfqq)) {
1422                 __cfq_slice_expired(cfqd, cfqq, 0);
1423                 cfq_schedule_dispatch(cfqd);
1424         }
1425
1426         kmem_cache_free(cfq_pool, cfqq);
1427 }
1428
1429 /*
1430  * Must always be called with the rcu_read_lock() held
1431  */
1432 static void
1433 __call_for_each_cic(struct io_context *ioc,
1434                     void (*func)(struct io_context *, struct cfq_io_context *))
1435 {
1436         struct cfq_io_context *cic;
1437         struct hlist_node *n;
1438
1439         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1440                 func(ioc, cic);
1441 }
1442
1443 /*
1444  * Call func for each cic attached to this ioc.
1445  */
1446 static void
1447 call_for_each_cic(struct io_context *ioc,
1448                   void (*func)(struct io_context *, struct cfq_io_context *))
1449 {
1450         rcu_read_lock();
1451         __call_for_each_cic(ioc, func);
1452         rcu_read_unlock();
1453 }
1454
1455 static void cfq_cic_free_rcu(struct rcu_head *head)
1456 {
1457         struct cfq_io_context *cic;
1458
1459         cic = container_of(head, struct cfq_io_context, rcu_head);
1460
1461         kmem_cache_free(cfq_ioc_pool, cic);
1462         elv_ioc_count_dec(cfq_ioc_count);
1463
1464         if (ioc_gone) {
1465                 /*
1466                  * CFQ scheduler is exiting, grab exit lock and check
1467                  * the pending io context count. If it hits zero,
1468                  * complete ioc_gone and set it back to NULL
1469                  */
1470                 spin_lock(&ioc_gone_lock);
1471                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1472                         complete(ioc_gone);
1473                         ioc_gone = NULL;
1474                 }
1475                 spin_unlock(&ioc_gone_lock);
1476         }
1477 }
1478
1479 static void cfq_cic_free(struct cfq_io_context *cic)
1480 {
1481         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1482 }
1483
1484 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1485 {
1486         unsigned long flags;
1487
1488         BUG_ON(!cic->dead_key);
1489
1490         spin_lock_irqsave(&ioc->lock, flags);
1491         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1492         hlist_del_rcu(&cic->cic_list);
1493         spin_unlock_irqrestore(&ioc->lock, flags);
1494
1495         cfq_cic_free(cic);
1496 }
1497
1498 /*
1499  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1500  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1501  * and ->trim() which is called with the task lock held
1502  */
1503 static void cfq_free_io_context(struct io_context *ioc)
1504 {
1505         /*
1506          * ioc->refcount is zero here, or we are called from elv_unregister(),
1507          * so no more cic's are allowed to be linked into this ioc.  So it
1508          * should be ok to iterate over the known list, we will see all cic's
1509          * since no new ones are added.
1510          */
1511         __call_for_each_cic(ioc, cic_free_func);
1512 }
1513
1514 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1515 {
1516         if (unlikely(cfqq == cfqd->active_queue)) {
1517                 __cfq_slice_expired(cfqd, cfqq, 0);
1518                 cfq_schedule_dispatch(cfqd);
1519         }
1520
1521         cfq_put_queue(cfqq);
1522 }
1523
1524 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1525                                          struct cfq_io_context *cic)
1526 {
1527         struct io_context *ioc = cic->ioc;
1528
1529         list_del_init(&cic->queue_list);
1530
1531         /*
1532          * Make sure key == NULL is seen for dead queues
1533          */
1534         smp_wmb();
1535         cic->dead_key = (unsigned long) cic->key;
1536         cic->key = NULL;
1537
1538         if (ioc->ioc_data == cic)
1539                 rcu_assign_pointer(ioc->ioc_data, NULL);
1540
1541         if (cic->cfqq[BLK_RW_ASYNC]) {
1542                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1543                 cic->cfqq[BLK_RW_ASYNC] = NULL;
1544         }
1545
1546         if (cic->cfqq[BLK_RW_SYNC]) {
1547                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1548                 cic->cfqq[BLK_RW_SYNC] = NULL;
1549         }
1550 }
1551
1552 static void cfq_exit_single_io_context(struct io_context *ioc,
1553                                        struct cfq_io_context *cic)
1554 {
1555         struct cfq_data *cfqd = cic->key;
1556
1557         if (cfqd) {
1558                 struct request_queue *q = cfqd->queue;
1559                 unsigned long flags;
1560
1561                 spin_lock_irqsave(q->queue_lock, flags);
1562
1563                 /*
1564                  * Ensure we get a fresh copy of the ->key to prevent
1565                  * race between exiting task and queue
1566                  */
1567                 smp_read_barrier_depends();
1568                 if (cic->key)
1569                         __cfq_exit_single_io_context(cfqd, cic);
1570
1571                 spin_unlock_irqrestore(q->queue_lock, flags);
1572         }
1573 }
1574
1575 /*
1576  * The process that ioc belongs to has exited, we need to clean up
1577  * and put the internal structures we have that belongs to that process.
1578  */
1579 static void cfq_exit_io_context(struct io_context *ioc)
1580 {
1581         call_for_each_cic(ioc, cfq_exit_single_io_context);
1582 }
1583
1584 static struct cfq_io_context *
1585 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1586 {
1587         struct cfq_io_context *cic;
1588
1589         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1590                                                         cfqd->queue->node);
1591         if (cic) {
1592                 cic->last_end_request = jiffies;
1593                 INIT_LIST_HEAD(&cic->queue_list);
1594                 INIT_HLIST_NODE(&cic->cic_list);
1595                 cic->dtor = cfq_free_io_context;
1596                 cic->exit = cfq_exit_io_context;
1597                 elv_ioc_count_inc(cfq_ioc_count);
1598         }
1599
1600         return cic;
1601 }
1602
1603 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1604 {
1605         struct task_struct *tsk = current;
1606         int ioprio_class;
1607
1608         if (!cfq_cfqq_prio_changed(cfqq))
1609                 return;
1610
1611         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1612         switch (ioprio_class) {
1613         default:
1614                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1615         case IOPRIO_CLASS_NONE:
1616                 /*
1617                  * no prio set, inherit CPU scheduling settings
1618                  */
1619                 cfqq->ioprio = task_nice_ioprio(tsk);
1620                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1621                 break;
1622         case IOPRIO_CLASS_RT:
1623                 cfqq->ioprio = task_ioprio(ioc);
1624                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1625                 break;
1626         case IOPRIO_CLASS_BE:
1627                 cfqq->ioprio = task_ioprio(ioc);
1628                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1629                 break;
1630         case IOPRIO_CLASS_IDLE:
1631                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1632                 cfqq->ioprio = 7;
1633                 cfq_clear_cfqq_idle_window(cfqq);
1634                 break;
1635         }
1636
1637         /*
1638          * keep track of original prio settings in case we have to temporarily
1639          * elevate the priority of this queue
1640          */
1641         cfqq->org_ioprio = cfqq->ioprio;
1642         cfqq->org_ioprio_class = cfqq->ioprio_class;
1643         cfq_clear_cfqq_prio_changed(cfqq);
1644 }
1645
1646 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1647 {
1648         struct cfq_data *cfqd = cic->key;
1649         struct cfq_queue *cfqq;
1650         unsigned long flags;
1651
1652         if (unlikely(!cfqd))
1653                 return;
1654
1655         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1656
1657         cfqq = cic->cfqq[BLK_RW_ASYNC];
1658         if (cfqq) {
1659                 struct cfq_queue *new_cfqq;
1660                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1661                                                 GFP_ATOMIC);
1662                 if (new_cfqq) {
1663                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1664                         cfq_put_queue(cfqq);
1665                 }
1666         }
1667
1668         cfqq = cic->cfqq[BLK_RW_SYNC];
1669         if (cfqq)
1670                 cfq_mark_cfqq_prio_changed(cfqq);
1671
1672         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1673 }
1674
1675 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1676 {
1677         call_for_each_cic(ioc, changed_ioprio);
1678         ioc->ioprio_changed = 0;
1679 }
1680
1681 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1682                           pid_t pid, bool is_sync)
1683 {
1684         RB_CLEAR_NODE(&cfqq->rb_node);
1685         RB_CLEAR_NODE(&cfqq->p_node);
1686         INIT_LIST_HEAD(&cfqq->fifo);
1687
1688         atomic_set(&cfqq->ref, 0);
1689         cfqq->cfqd = cfqd;
1690
1691         cfq_mark_cfqq_prio_changed(cfqq);
1692
1693         if (is_sync) {
1694                 if (!cfq_class_idle(cfqq))
1695                         cfq_mark_cfqq_idle_window(cfqq);
1696                 cfq_mark_cfqq_sync(cfqq);
1697         }
1698         cfqq->pid = pid;
1699 }
1700
1701 static struct cfq_queue *
1702 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
1703                      struct io_context *ioc, gfp_t gfp_mask)
1704 {
1705         struct cfq_queue *cfqq, *new_cfqq = NULL;
1706         struct cfq_io_context *cic;
1707
1708 retry:
1709         cic = cfq_cic_lookup(cfqd, ioc);
1710         /* cic always exists here */
1711         cfqq = cic_to_cfqq(cic, is_sync);
1712
1713         /*
1714          * Always try a new alloc if we fell back to the OOM cfqq
1715          * originally, since it should just be a temporary situation.
1716          */
1717         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1718                 cfqq = NULL;
1719                 if (new_cfqq) {
1720                         cfqq = new_cfqq;
1721                         new_cfqq = NULL;
1722                 } else if (gfp_mask & __GFP_WAIT) {
1723                         spin_unlock_irq(cfqd->queue->queue_lock);
1724                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1725                                         gfp_mask | __GFP_ZERO,
1726                                         cfqd->queue->node);
1727                         spin_lock_irq(cfqd->queue->queue_lock);
1728                         if (new_cfqq)
1729                                 goto retry;
1730                 } else {
1731                         cfqq = kmem_cache_alloc_node(cfq_pool,
1732                                         gfp_mask | __GFP_ZERO,
1733                                         cfqd->queue->node);
1734                 }
1735
1736                 if (cfqq) {
1737                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1738                         cfq_init_prio_data(cfqq, ioc);
1739                         cfq_log_cfqq(cfqd, cfqq, "alloced");
1740                 } else
1741                         cfqq = &cfqd->oom_cfqq;
1742         }
1743
1744         if (new_cfqq)
1745                 kmem_cache_free(cfq_pool, new_cfqq);
1746
1747         return cfqq;
1748 }
1749
1750 static struct cfq_queue **
1751 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1752 {
1753         switch (ioprio_class) {
1754         case IOPRIO_CLASS_RT:
1755                 return &cfqd->async_cfqq[0][ioprio];
1756         case IOPRIO_CLASS_BE:
1757                 return &cfqd->async_cfqq[1][ioprio];
1758         case IOPRIO_CLASS_IDLE:
1759                 return &cfqd->async_idle_cfqq;
1760         default:
1761                 BUG();
1762         }
1763 }
1764
1765 static struct cfq_queue *
1766 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
1767               gfp_t gfp_mask)
1768 {
1769         const int ioprio = task_ioprio(ioc);
1770         const int ioprio_class = task_ioprio_class(ioc);
1771         struct cfq_queue **async_cfqq = NULL;
1772         struct cfq_queue *cfqq = NULL;
1773
1774         if (!is_sync) {
1775                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1776                 cfqq = *async_cfqq;
1777         }
1778
1779         if (!cfqq)
1780                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1781
1782         /*
1783          * pin the queue now that it's allocated, scheduler exit will prune it
1784          */
1785         if (!is_sync && !(*async_cfqq)) {
1786                 atomic_inc(&cfqq->ref);
1787                 *async_cfqq = cfqq;
1788         }
1789
1790         atomic_inc(&cfqq->ref);
1791         return cfqq;
1792 }
1793
1794 /*
1795  * We drop cfq io contexts lazily, so we may find a dead one.
1796  */
1797 static void
1798 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1799                   struct cfq_io_context *cic)
1800 {
1801         unsigned long flags;
1802
1803         WARN_ON(!list_empty(&cic->queue_list));
1804
1805         spin_lock_irqsave(&ioc->lock, flags);
1806
1807         BUG_ON(ioc->ioc_data == cic);
1808
1809         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1810         hlist_del_rcu(&cic->cic_list);
1811         spin_unlock_irqrestore(&ioc->lock, flags);
1812
1813         cfq_cic_free(cic);
1814 }
1815
1816 static struct cfq_io_context *
1817 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1818 {
1819         struct cfq_io_context *cic;
1820         unsigned long flags;
1821         void *k;
1822
1823         if (unlikely(!ioc))
1824                 return NULL;
1825
1826         rcu_read_lock();
1827
1828         /*
1829          * we maintain a last-hit cache, to avoid browsing over the tree
1830          */
1831         cic = rcu_dereference(ioc->ioc_data);
1832         if (cic && cic->key == cfqd) {
1833                 rcu_read_unlock();
1834                 return cic;
1835         }
1836
1837         do {
1838                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1839                 rcu_read_unlock();
1840                 if (!cic)
1841                         break;
1842                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1843                 k = cic->key;
1844                 if (unlikely(!k)) {
1845                         cfq_drop_dead_cic(cfqd, ioc, cic);
1846                         rcu_read_lock();
1847                         continue;
1848                 }
1849
1850                 spin_lock_irqsave(&ioc->lock, flags);
1851                 rcu_assign_pointer(ioc->ioc_data, cic);
1852                 spin_unlock_irqrestore(&ioc->lock, flags);
1853                 break;
1854         } while (1);
1855
1856         return cic;
1857 }
1858
1859 /*
1860  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1861  * the process specific cfq io context when entered from the block layer.
1862  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1863  */
1864 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1865                         struct cfq_io_context *cic, gfp_t gfp_mask)
1866 {
1867         unsigned long flags;
1868         int ret;
1869
1870         ret = radix_tree_preload(gfp_mask);
1871         if (!ret) {
1872                 cic->ioc = ioc;
1873                 cic->key = cfqd;
1874
1875                 spin_lock_irqsave(&ioc->lock, flags);
1876                 ret = radix_tree_insert(&ioc->radix_root,
1877                                                 (unsigned long) cfqd, cic);
1878                 if (!ret)
1879                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1880                 spin_unlock_irqrestore(&ioc->lock, flags);
1881
1882                 radix_tree_preload_end();
1883
1884                 if (!ret) {
1885                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1886                         list_add(&cic->queue_list, &cfqd->cic_list);
1887                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1888                 }
1889         }
1890
1891         if (ret)
1892                 printk(KERN_ERR "cfq: cic link failed!\n");
1893
1894         return ret;
1895 }
1896
1897 /*
1898  * Setup general io context and cfq io context. There can be several cfq
1899  * io contexts per general io context, if this process is doing io to more
1900  * than one device managed by cfq.
1901  */
1902 static struct cfq_io_context *
1903 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1904 {
1905         struct io_context *ioc = NULL;
1906         struct cfq_io_context *cic;
1907
1908         might_sleep_if(gfp_mask & __GFP_WAIT);
1909
1910         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1911         if (!ioc)
1912                 return NULL;
1913
1914         cic = cfq_cic_lookup(cfqd, ioc);
1915         if (cic)
1916                 goto out;
1917
1918         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1919         if (cic == NULL)
1920                 goto err;
1921
1922         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1923                 goto err_free;
1924
1925 out:
1926         smp_read_barrier_depends();
1927         if (unlikely(ioc->ioprio_changed))
1928                 cfq_ioc_set_ioprio(ioc);
1929
1930         return cic;
1931 err_free:
1932         cfq_cic_free(cic);
1933 err:
1934         put_io_context(ioc);
1935         return NULL;
1936 }
1937
1938 static void
1939 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1940 {
1941         unsigned long elapsed = jiffies - cic->last_end_request;
1942         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1943
1944         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1945         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1946         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1947 }
1948
1949 static void
1950 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1951                        struct request *rq)
1952 {
1953         sector_t sdist;
1954         u64 total;
1955
1956         if (!cic->last_request_pos)
1957                 sdist = 0;
1958         else if (cic->last_request_pos < blk_rq_pos(rq))
1959                 sdist = blk_rq_pos(rq) - cic->last_request_pos;
1960         else
1961                 sdist = cic->last_request_pos - blk_rq_pos(rq);
1962
1963         /*
1964          * Don't allow the seek distance to get too large from the
1965          * odd fragment, pagein, etc
1966          */
1967         if (cic->seek_samples <= 60) /* second&third seek */
1968                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1969         else
1970                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1971
1972         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1973         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1974         total = cic->seek_total + (cic->seek_samples/2);
1975         do_div(total, cic->seek_samples);
1976         cic->seek_mean = (sector_t)total;
1977 }
1978
1979 /*
1980  * Disable idle window if the process thinks too long or seeks so much that
1981  * it doesn't matter
1982  */
1983 static void
1984 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1985                        struct cfq_io_context *cic)
1986 {
1987         int old_idle, enable_idle;
1988
1989         /*
1990          * Don't idle for async or idle io prio class
1991          */
1992         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1993                 return;
1994
1995         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1996
1997         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1998             (!cfqd->cfq_latency && cfqd->hw_tag && CIC_SEEKY(cic)))
1999                 enable_idle = 0;
2000         else if (sample_valid(cic->ttime_samples)) {
2001                 unsigned int slice_idle = cfqd->cfq_slice_idle;
2002                 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
2003                         slice_idle = msecs_to_jiffies(CFQ_MIN_TT);
2004                 if (cic->ttime_mean > slice_idle)
2005                         enable_idle = 0;
2006                 else
2007                         enable_idle = 1;
2008         }
2009
2010         if (old_idle != enable_idle) {
2011                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2012                 if (enable_idle)
2013                         cfq_mark_cfqq_idle_window(cfqq);
2014                 else
2015                         cfq_clear_cfqq_idle_window(cfqq);
2016         }
2017 }
2018
2019 /*
2020  * Check if new_cfqq should preempt the currently active queue. Return 0 for
2021  * no or if we aren't sure, a 1 will cause a preempt.
2022  */
2023 static bool
2024 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2025                    struct request *rq)
2026 {
2027         struct cfq_queue *cfqq;
2028
2029         cfqq = cfqd->active_queue;
2030         if (!cfqq)
2031                 return false;
2032
2033         if (cfq_slice_used(cfqq))
2034                 return true;
2035
2036         if (cfq_class_idle(new_cfqq))
2037                 return false;
2038
2039         if (cfq_class_idle(cfqq))
2040                 return true;
2041
2042         /*
2043          * if the new request is sync, but the currently running queue is
2044          * not, let the sync request have priority.
2045          */
2046         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2047                 return true;
2048
2049         /*
2050          * So both queues are sync. Let the new request get disk time if
2051          * it's a metadata request and the current queue is doing regular IO.
2052          */
2053         if (rq_is_meta(rq) && !cfqq->meta_pending)
2054                 return false;
2055
2056         /*
2057          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2058          */
2059         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2060                 return true;
2061
2062         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2063                 return false;
2064
2065         /*
2066          * if this request is as-good as one we would expect from the
2067          * current cfqq, let it preempt
2068          */
2069         if (cfq_rq_close(cfqd, rq))
2070                 return true;
2071
2072         return false;
2073 }
2074
2075 /*
2076  * cfqq preempts the active queue. if we allowed preempt with no slice left,
2077  * let it have half of its nominal slice.
2078  */
2079 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2080 {
2081         cfq_log_cfqq(cfqd, cfqq, "preempt");
2082         cfq_slice_expired(cfqd, 1);
2083
2084         /*
2085          * Put the new queue at the front of the of the current list,
2086          * so we know that it will be selected next.
2087          */
2088         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2089
2090         cfq_service_tree_add(cfqd, cfqq, 1);
2091
2092         cfqq->slice_end = 0;
2093         cfq_mark_cfqq_slice_new(cfqq);
2094 }
2095
2096 /*
2097  * Called when a new fs request (rq) is added (to cfqq). Check if there's
2098  * something we should do about it
2099  */
2100 static void
2101 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2102                 struct request *rq)
2103 {
2104         struct cfq_io_context *cic = RQ_CIC(rq);
2105
2106         cfqd->rq_queued++;
2107         if (rq_is_meta(rq))
2108                 cfqq->meta_pending++;
2109
2110         cfq_update_io_thinktime(cfqd, cic);
2111         cfq_update_io_seektime(cfqd, cic, rq);
2112         cfq_update_idle_window(cfqd, cfqq, cic);
2113
2114         cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2115
2116         if (cfqq == cfqd->active_queue) {
2117                 /*
2118                  * Remember that we saw a request from this process, but
2119                  * don't start queuing just yet. Otherwise we risk seeing lots
2120                  * of tiny requests, because we disrupt the normal plugging
2121                  * and merging. If the request is already larger than a single
2122                  * page, let it rip immediately. For that case we assume that
2123                  * merging is already done. Ditto for a busy system that
2124                  * has other work pending, don't risk delaying until the
2125                  * idle timer unplug to continue working.
2126                  */
2127                 if (cfq_cfqq_wait_request(cfqq)) {
2128                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2129                             cfqd->busy_queues > 1) {
2130                                 del_timer(&cfqd->idle_slice_timer);
2131                         __blk_run_queue(cfqd->queue);
2132                         }
2133                         cfq_mark_cfqq_must_dispatch(cfqq);
2134                 }
2135         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2136                 /*
2137                  * not the active queue - expire current slice if it is
2138                  * idle and has expired it's mean thinktime or this new queue
2139                  * has some old slice time left and is of higher priority or
2140                  * this new queue is RT and the current one is BE
2141                  */
2142                 cfq_preempt_queue(cfqd, cfqq);
2143                 __blk_run_queue(cfqd->queue);
2144         }
2145 }
2146
2147 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2148 {
2149         struct cfq_data *cfqd = q->elevator->elevator_data;
2150         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2151
2152         cfq_log_cfqq(cfqd, cfqq, "insert_request");
2153         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2154
2155         cfq_add_rq_rb(rq);
2156
2157         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2158         list_add_tail(&rq->queuelist, &cfqq->fifo);
2159
2160         cfq_rq_enqueued(cfqd, cfqq, rq);
2161 }
2162
2163 /*
2164  * Update hw_tag based on peak queue depth over 50 samples under
2165  * sufficient load.
2166  */
2167 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2168 {
2169         if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2170                 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2171
2172         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2173             rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2174                 return;
2175
2176         if (cfqd->hw_tag_samples++ < 50)
2177                 return;
2178
2179         if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2180                 cfqd->hw_tag = 1;
2181         else
2182                 cfqd->hw_tag = 0;
2183
2184         cfqd->hw_tag_samples = 0;
2185         cfqd->rq_in_driver_peak = 0;
2186 }
2187
2188 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2189 {
2190         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2191         struct cfq_data *cfqd = cfqq->cfqd;
2192         const int sync = rq_is_sync(rq);
2193         unsigned long now;
2194
2195         now = jiffies;
2196         cfq_log_cfqq(cfqd, cfqq, "complete");
2197
2198         cfq_update_hw_tag(cfqd);
2199
2200         WARN_ON(!cfqd->rq_in_driver[sync]);
2201         WARN_ON(!cfqq->dispatched);
2202         cfqd->rq_in_driver[sync]--;
2203         cfqq->dispatched--;
2204
2205         if (cfq_cfqq_sync(cfqq))
2206                 cfqd->sync_flight--;
2207
2208         if (sync) {
2209                 RQ_CIC(rq)->last_end_request = now;
2210                 cfqd->last_end_sync_rq = now;
2211         }
2212
2213         /*
2214          * If this is the active queue, check if it needs to be expired,
2215          * or if we want to idle in case it has no pending requests.
2216          */
2217         if (cfqd->active_queue == cfqq) {
2218                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2219
2220                 if (cfq_cfqq_slice_new(cfqq)) {
2221                         cfq_set_prio_slice(cfqd, cfqq);
2222                         cfq_clear_cfqq_slice_new(cfqq);
2223                 }
2224                 /*
2225                  * If there are no requests waiting in this queue, and
2226                  * there are other queues ready to issue requests, AND
2227                  * those other queues are issuing requests within our
2228                  * mean seek distance, give them a chance to run instead
2229                  * of idling.
2230                  */
2231                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2232                         cfq_slice_expired(cfqd, 1);
2233                 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2234                          sync && !rq_noidle(rq))
2235                         cfq_arm_slice_timer(cfqd);
2236         }
2237
2238         if (!rq_in_driver(cfqd))
2239                 cfq_schedule_dispatch(cfqd);
2240 }
2241
2242 /*
2243  * we temporarily boost lower priority queues if they are holding fs exclusive
2244  * resources. they are boosted to normal prio (CLASS_BE/4)
2245  */
2246 static void cfq_prio_boost(struct cfq_queue *cfqq)
2247 {
2248         if (has_fs_excl()) {
2249                 /*
2250                  * boost idle prio on transactions that would lock out other
2251                  * users of the filesystem
2252                  */
2253                 if (cfq_class_idle(cfqq))
2254                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
2255                 if (cfqq->ioprio > IOPRIO_NORM)
2256                         cfqq->ioprio = IOPRIO_NORM;
2257         } else {
2258                 /*
2259                  * check if we need to unboost the queue
2260                  */
2261                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2262                         cfqq->ioprio_class = cfqq->org_ioprio_class;
2263                 if (cfqq->ioprio != cfqq->org_ioprio)
2264                         cfqq->ioprio = cfqq->org_ioprio;
2265         }
2266 }
2267
2268 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2269 {
2270         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2271                 cfq_mark_cfqq_must_alloc_slice(cfqq);
2272                 return ELV_MQUEUE_MUST;
2273         }
2274
2275         return ELV_MQUEUE_MAY;
2276 }
2277
2278 static int cfq_may_queue(struct request_queue *q, int rw)
2279 {
2280         struct cfq_data *cfqd = q->elevator->elevator_data;
2281         struct task_struct *tsk = current;
2282         struct cfq_io_context *cic;
2283         struct cfq_queue *cfqq;
2284
2285         /*
2286          * don't force setup of a queue from here, as a call to may_queue
2287          * does not necessarily imply that a request actually will be queued.
2288          * so just lookup a possibly existing queue, or return 'may queue'
2289          * if that fails
2290          */
2291         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2292         if (!cic)
2293                 return ELV_MQUEUE_MAY;
2294
2295         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2296         if (cfqq) {
2297                 cfq_init_prio_data(cfqq, cic->ioc);
2298                 cfq_prio_boost(cfqq);
2299
2300                 return __cfq_may_queue(cfqq);
2301         }
2302
2303         return ELV_MQUEUE_MAY;
2304 }
2305
2306 /*
2307  * queue lock held here
2308  */
2309 static void cfq_put_request(struct request *rq)
2310 {
2311         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2312
2313         if (cfqq) {
2314                 const int rw = rq_data_dir(rq);
2315
2316                 BUG_ON(!cfqq->allocated[rw]);
2317                 cfqq->allocated[rw]--;
2318
2319                 put_io_context(RQ_CIC(rq)->ioc);
2320
2321                 rq->elevator_private = NULL;
2322                 rq->elevator_private2 = NULL;
2323
2324                 cfq_put_queue(cfqq);
2325         }
2326 }
2327
2328 /*
2329  * Allocate cfq data structures associated with this request.
2330  */
2331 static int
2332 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2333 {
2334         struct cfq_data *cfqd = q->elevator->elevator_data;
2335         struct cfq_io_context *cic;
2336         const int rw = rq_data_dir(rq);
2337         const bool is_sync = rq_is_sync(rq);
2338         struct cfq_queue *cfqq;
2339         unsigned long flags;
2340
2341         might_sleep_if(gfp_mask & __GFP_WAIT);
2342
2343         cic = cfq_get_io_context(cfqd, gfp_mask);
2344
2345         spin_lock_irqsave(q->queue_lock, flags);
2346
2347         if (!cic)
2348                 goto queue_fail;
2349
2350         cfqq = cic_to_cfqq(cic, is_sync);
2351         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2352                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2353                 cic_set_cfqq(cic, cfqq, is_sync);
2354         }
2355
2356         cfqq->allocated[rw]++;
2357         atomic_inc(&cfqq->ref);
2358
2359         spin_unlock_irqrestore(q->queue_lock, flags);
2360
2361         rq->elevator_private = cic;
2362         rq->elevator_private2 = cfqq;
2363         return 0;
2364
2365 queue_fail:
2366         if (cic)
2367                 put_io_context(cic->ioc);
2368
2369         cfq_schedule_dispatch(cfqd);
2370         spin_unlock_irqrestore(q->queue_lock, flags);
2371         cfq_log(cfqd, "set_request fail");
2372         return 1;
2373 }
2374
2375 static void cfq_kick_queue(struct work_struct *work)
2376 {
2377         struct cfq_data *cfqd =
2378                 container_of(work, struct cfq_data, unplug_work);
2379         struct request_queue *q = cfqd->queue;
2380
2381         spin_lock_irq(q->queue_lock);
2382         __blk_run_queue(cfqd->queue);
2383         spin_unlock_irq(q->queue_lock);
2384 }
2385
2386 /*
2387  * Timer running if the active_queue is currently idling inside its time slice
2388  */
2389 static void cfq_idle_slice_timer(unsigned long data)
2390 {
2391         struct cfq_data *cfqd = (struct cfq_data *) data;
2392         struct cfq_queue *cfqq;
2393         unsigned long flags;
2394         int timed_out = 1;
2395
2396         cfq_log(cfqd, "idle timer fired");
2397
2398         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2399
2400         cfqq = cfqd->active_queue;
2401         if (cfqq) {
2402                 timed_out = 0;
2403
2404                 /*
2405                  * We saw a request before the queue expired, let it through
2406                  */
2407                 if (cfq_cfqq_must_dispatch(cfqq))
2408                         goto out_kick;
2409
2410                 /*
2411                  * expired
2412                  */
2413                 if (cfq_slice_used(cfqq))
2414                         goto expire;
2415
2416                 /*
2417                  * only expire and reinvoke request handler, if there are
2418                  * other queues with pending requests
2419                  */
2420                 if (!cfqd->busy_queues)
2421                         goto out_cont;
2422
2423                 /*
2424                  * not expired and it has a request pending, let it dispatch
2425                  */
2426                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2427                         goto out_kick;
2428         }
2429 expire:
2430         cfq_slice_expired(cfqd, timed_out);
2431 out_kick:
2432         cfq_schedule_dispatch(cfqd);
2433 out_cont:
2434         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2435 }
2436
2437 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2438 {
2439         del_timer_sync(&cfqd->idle_slice_timer);
2440         cancel_work_sync(&cfqd->unplug_work);
2441 }
2442
2443 static void cfq_put_async_queues(struct cfq_data *cfqd)
2444 {
2445         int i;
2446
2447         for (i = 0; i < IOPRIO_BE_NR; i++) {
2448                 if (cfqd->async_cfqq[0][i])
2449                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2450                 if (cfqd->async_cfqq[1][i])
2451                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2452         }
2453
2454         if (cfqd->async_idle_cfqq)
2455                 cfq_put_queue(cfqd->async_idle_cfqq);
2456 }
2457
2458 static void cfq_exit_queue(struct elevator_queue *e)
2459 {
2460         struct cfq_data *cfqd = e->elevator_data;
2461         struct request_queue *q = cfqd->queue;
2462
2463         cfq_shutdown_timer_wq(cfqd);
2464
2465         spin_lock_irq(q->queue_lock);
2466
2467         if (cfqd->active_queue)
2468                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2469
2470         while (!list_empty(&cfqd->cic_list)) {
2471                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2472                                                         struct cfq_io_context,
2473                                                         queue_list);
2474
2475                 __cfq_exit_single_io_context(cfqd, cic);
2476         }
2477
2478         cfq_put_async_queues(cfqd);
2479
2480         spin_unlock_irq(q->queue_lock);
2481
2482         cfq_shutdown_timer_wq(cfqd);
2483
2484         kfree(cfqd);
2485 }
2486
2487 static void *cfq_init_queue(struct request_queue *q)
2488 {
2489         struct cfq_data *cfqd;
2490         int i;
2491
2492         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2493         if (!cfqd)
2494                 return NULL;
2495
2496         cfqd->service_tree = CFQ_RB_ROOT;
2497
2498         /*
2499          * Not strictly needed (since RB_ROOT just clears the node and we
2500          * zeroed cfqd on alloc), but better be safe in case someone decides
2501          * to add magic to the rb code
2502          */
2503         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2504                 cfqd->prio_trees[i] = RB_ROOT;
2505
2506         /*
2507          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2508          * Grab a permanent reference to it, so that the normal code flow
2509          * will not attempt to free it.
2510          */
2511         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2512         atomic_inc(&cfqd->oom_cfqq.ref);
2513
2514         INIT_LIST_HEAD(&cfqd->cic_list);
2515
2516         cfqd->queue = q;
2517
2518         init_timer(&cfqd->idle_slice_timer);
2519         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2520         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2521
2522         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2523
2524         cfqd->cfq_quantum = cfq_quantum;
2525         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2526         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2527         cfqd->cfq_back_max = cfq_back_max;
2528         cfqd->cfq_back_penalty = cfq_back_penalty;
2529         cfqd->cfq_slice[0] = cfq_slice_async;
2530         cfqd->cfq_slice[1] = cfq_slice_sync;
2531         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2532         cfqd->cfq_slice_idle = cfq_slice_idle;
2533         cfqd->cfq_latency = 1;
2534         cfqd->hw_tag = 1;
2535         cfqd->last_end_sync_rq = jiffies;
2536         return cfqd;
2537 }
2538
2539 static void cfq_slab_kill(void)
2540 {
2541         /*
2542          * Caller already ensured that pending RCU callbacks are completed,
2543          * so we should have no busy allocations at this point.
2544          */
2545         if (cfq_pool)
2546                 kmem_cache_destroy(cfq_pool);
2547         if (cfq_ioc_pool)
2548                 kmem_cache_destroy(cfq_ioc_pool);
2549 }
2550
2551 static int __init cfq_slab_setup(void)
2552 {
2553         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2554         if (!cfq_pool)
2555                 goto fail;
2556
2557         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2558         if (!cfq_ioc_pool)
2559                 goto fail;
2560
2561         return 0;
2562 fail:
2563         cfq_slab_kill();
2564         return -ENOMEM;
2565 }
2566
2567 /*
2568  * sysfs parts below -->
2569  */
2570 static ssize_t
2571 cfq_var_show(unsigned int var, char *page)
2572 {
2573         return sprintf(page, "%d\n", var);
2574 }
2575
2576 static ssize_t
2577 cfq_var_store(unsigned int *var, const char *page, size_t count)
2578 {
2579         char *p = (char *) page;
2580
2581         *var = simple_strtoul(p, &p, 10);
2582         return count;
2583 }
2584
2585 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2586 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
2587 {                                                                       \
2588         struct cfq_data *cfqd = e->elevator_data;                       \
2589         unsigned int __data = __VAR;                                    \
2590         if (__CONV)                                                     \
2591                 __data = jiffies_to_msecs(__data);                      \
2592         return cfq_var_show(__data, (page));                            \
2593 }
2594 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2595 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2596 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2597 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2598 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2599 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2600 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2601 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2602 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2603 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
2604 #undef SHOW_FUNCTION
2605
2606 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2607 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2608 {                                                                       \
2609         struct cfq_data *cfqd = e->elevator_data;                       \
2610         unsigned int __data;                                            \
2611         int ret = cfq_var_store(&__data, (page), count);                \
2612         if (__data < (MIN))                                             \
2613                 __data = (MIN);                                         \
2614         else if (__data > (MAX))                                        \
2615                 __data = (MAX);                                         \
2616         if (__CONV)                                                     \
2617                 *(__PTR) = msecs_to_jiffies(__data);                    \
2618         else                                                            \
2619                 *(__PTR) = __data;                                      \
2620         return ret;                                                     \
2621 }
2622 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2623 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2624                 UINT_MAX, 1);
2625 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2626                 UINT_MAX, 1);
2627 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2628 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2629                 UINT_MAX, 0);
2630 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2631 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2632 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2633 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2634                 UINT_MAX, 0);
2635 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
2636 #undef STORE_FUNCTION
2637
2638 #define CFQ_ATTR(name) \
2639         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2640
2641 static struct elv_fs_entry cfq_attrs[] = {
2642         CFQ_ATTR(quantum),
2643         CFQ_ATTR(fifo_expire_sync),
2644         CFQ_ATTR(fifo_expire_async),
2645         CFQ_ATTR(back_seek_max),
2646         CFQ_ATTR(back_seek_penalty),
2647         CFQ_ATTR(slice_sync),
2648         CFQ_ATTR(slice_async),
2649         CFQ_ATTR(slice_async_rq),
2650         CFQ_ATTR(slice_idle),
2651         CFQ_ATTR(low_latency),
2652         __ATTR_NULL
2653 };
2654
2655 static struct elevator_type iosched_cfq = {
2656         .ops = {
2657                 .elevator_merge_fn =            cfq_merge,
2658                 .elevator_merged_fn =           cfq_merged_request,
2659                 .elevator_merge_req_fn =        cfq_merged_requests,
2660                 .elevator_allow_merge_fn =      cfq_allow_merge,
2661                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2662                 .elevator_add_req_fn =          cfq_insert_request,
2663                 .elevator_activate_req_fn =     cfq_activate_request,
2664                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2665                 .elevator_queue_empty_fn =      cfq_queue_empty,
2666                 .elevator_completed_req_fn =    cfq_completed_request,
2667                 .elevator_former_req_fn =       elv_rb_former_request,
2668                 .elevator_latter_req_fn =       elv_rb_latter_request,
2669                 .elevator_set_req_fn =          cfq_set_request,
2670                 .elevator_put_req_fn =          cfq_put_request,
2671                 .elevator_may_queue_fn =        cfq_may_queue,
2672                 .elevator_init_fn =             cfq_init_queue,
2673                 .elevator_exit_fn =             cfq_exit_queue,
2674                 .trim =                         cfq_free_io_context,
2675         },
2676         .elevator_attrs =       cfq_attrs,
2677         .elevator_name =        "cfq",
2678         .elevator_owner =       THIS_MODULE,
2679 };
2680
2681 static int __init cfq_init(void)
2682 {
2683         /*
2684          * could be 0 on HZ < 1000 setups
2685          */
2686         if (!cfq_slice_async)
2687                 cfq_slice_async = 1;
2688         if (!cfq_slice_idle)
2689                 cfq_slice_idle = 1;
2690
2691         if (cfq_slab_setup())
2692                 return -ENOMEM;
2693
2694         elv_register(&iosched_cfq);
2695
2696         return 0;
2697 }
2698
2699 static void __exit cfq_exit(void)
2700 {
2701         DECLARE_COMPLETION_ONSTACK(all_gone);
2702         elv_unregister(&iosched_cfq);
2703         ioc_gone = &all_gone;
2704         /* ioc_gone's update must be visible before reading ioc_count */
2705         smp_wmb();
2706
2707         /*
2708          * this also protects us from entering cfq_slab_kill() with
2709          * pending RCU callbacks
2710          */
2711         if (elv_ioc_count_read(cfq_ioc_count))
2712                 wait_for_completion(&all_gone);
2713         cfq_slab_kill();
2714 }
2715
2716 module_init(cfq_init);
2717 module_exit(cfq_exit);
2718
2719 MODULE_AUTHOR("Jens Axboe");
2720 MODULE_LICENSE("GPL");
2721 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");