block/cfq-iosched.c

   1 /*
   2  *  CFQ, or complete fairness queueing, disk scheduler.
   3  *
   4  *  Based on ideas from a previously unfinished io
   5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   6  *
   7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
   8  */
   9 #include <linux/module.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/elevator.h>
  12 #include <linux/rbtree.h>
  13 #include <linux/ioprio.h>
  14
  15 /*
  16  * tunables
  17  */
  18 static const int cfq_quantum = 4;               /* max queue in one round of service */
  19 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  20 static const int cfq_back_max = 16 * 1024;      /* maximum backwards seek, in KiB */
  21 static const int cfq_back_penalty = 2;          /* penalty of a backwards seek */
  22
  23 static const int cfq_slice_sync = HZ / 10;
  24 static int cfq_slice_async = HZ / 25;
  25 static const int cfq_slice_async_rq = 2;
  26 static int cfq_slice_idle = HZ / 125;
  27
  28 /*
  29  * offset from end of service tree
  30  */
  31 #define CFQ_IDLE_DELAY          (HZ / 5)
  32
  33 /*
  34  * below this threshold, we consider thinktime immediate
  35  */
  36 #define CFQ_MIN_TT              (2)
  37
  38 #define CFQ_SLICE_SCALE         (5)
  39
  40 #define RQ_CIC(rq)              ((struct cfq_io_context*)(rq)->elevator_private)
  41 #define RQ_CFQQ(rq)             ((rq)->elevator_private2)
  42
  43 static struct kmem_cache *cfq_pool;
  44 static struct kmem_cache *cfq_ioc_pool;
  45
  46 static DEFINE_PER_CPU(unsigned long, ioc_count);
  47 static struct completion *ioc_gone;
  48
  49 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  50 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  51 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  52
  53 #define ASYNC                   (0)
  54 #define SYNC                    (1)
  55
  56 #define sample_valid(samples)   ((samples) > 80)
  57
  58 /*
  59  * Most of our rbtree usage is for sorting with min extraction, so
  60  * if we cache the leftmost node we don't have to walk down the tree
  61  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  62  * move this into the elevator for the rq sorting as well.
  63  */
  64 struct cfq_rb_root {
  65         struct rb_root rb;
  66         struct rb_node *left;
  67 };
  68 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, }
  69
  70 /*
  71  * Per block device queue structure
  72  */
  73 struct cfq_data {
  74         struct request_queue *queue;
  75
  76         /*
  77          * rr list of queues with requests and the count of them
  78          */
  79         struct cfq_rb_root service_tree;
  80         unsigned int busy_queues;
  81
  82         int rq_in_driver;
  83         int sync_flight;
  84         int hw_tag;
  85
  86         /*
  87          * idle window management
  88          */
  89         struct timer_list idle_slice_timer;
  90         struct work_struct unplug_work;
  91
  92         struct cfq_queue *active_queue;
  93         struct cfq_io_context *active_cic;
  94
  95         /*
  96          * async queue for each priority case
  97          */
  98         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
  99         struct cfq_queue *async_idle_cfqq;
 100
 101         sector_t last_position;
 102         unsigned long last_end_request;
 103
 104         /*
 105          * tunables, see top of file
 106          */
 107         unsigned int cfq_quantum;
 108         unsigned int cfq_fifo_expire[2];
 109         unsigned int cfq_back_penalty;
 110         unsigned int cfq_back_max;
 111         unsigned int cfq_slice[2];
 112         unsigned int cfq_slice_async_rq;
 113         unsigned int cfq_slice_idle;
 114
 115         struct list_head cic_list;
 116 };
 117
 118 /*
 119  * Per process-grouping structure
 120  */
 121 struct cfq_queue {
 122         /* reference count */
 123         atomic_t ref;
 124         /* parent cfq_data */
 125         struct cfq_data *cfqd;
 126         /* service_tree member */
 127         struct rb_node rb_node;
 128         /* service_tree key */
 129         unsigned long rb_key;
 130         /* sorted list of pending requests */
 131         struct rb_root sort_list;
 132         /* if fifo isn't expired, next request to serve */
 133         struct request *next_rq;
 134         /* requests queued in sort_list */
 135         int queued[2];
 136         /* currently allocated requests */
 137         int allocated[2];
 138         /* pending metadata requests */
 139         int meta_pending;
 140         /* fifo list of requests in sort_list */
 141         struct list_head fifo;
 142
 143         unsigned long slice_end;
 144         long slice_resid;
 145
 146         /* number of requests that are on the dispatch list or inside driver */
 147         int dispatched;
 148
 149         /* io prio of this group */
 150         unsigned short ioprio, org_ioprio;
 151         unsigned short ioprio_class, org_ioprio_class;
 152
 153         /* various state flags, see below */
 154         unsigned int flags;
 155 };
 156
 157 enum cfqq_state_flags {
 158         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 159         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 160         CFQ_CFQQ_FLAG_must_alloc,       /* must be allowed rq alloc */
 161         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 162         CFQ_CFQQ_FLAG_must_dispatch,    /* must dispatch, even if expired */
 163         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 164         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 165         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 166         CFQ_CFQQ_FLAG_queue_new,        /* queue never been serviced */
 167         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 168         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 169 };
 170
 171 #define CFQ_CFQQ_FNS(name)                                              \
 172 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 173 {                                                                       \
 174         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 175 }                                                                       \
 176 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 177 {                                                                       \
 178         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 179 }                                                                       \
 180 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 181 {                                                                       \
 182         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 183 }
 184
 185 CFQ_CFQQ_FNS(on_rr);
 186 CFQ_CFQQ_FNS(wait_request);
 187 CFQ_CFQQ_FNS(must_alloc);
 188 CFQ_CFQQ_FNS(must_alloc_slice);
 189 CFQ_CFQQ_FNS(must_dispatch);
 190 CFQ_CFQQ_FNS(fifo_expire);
 191 CFQ_CFQQ_FNS(idle_window);
 192 CFQ_CFQQ_FNS(prio_changed);
 193 CFQ_CFQQ_FNS(queue_new);
 194 CFQ_CFQQ_FNS(slice_new);
 195 CFQ_CFQQ_FNS(sync);
 196 #undef CFQ_CFQQ_FNS
 197
 198 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 199 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
 200                                        struct io_context *, gfp_t);
 201 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
 202                                                 struct io_context *);
 203
 204 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
 205                                             int is_sync)
 206 {
 207         return cic->cfqq[!!is_sync];
 208 }
 209
 210 static inline void cic_set_cfqq(struct cfq_io_context *cic,
 211                                 struct cfq_queue *cfqq, int is_sync)
 212 {
 213         cic->cfqq[!!is_sync] = cfqq;
 214 }
 215
 216 /*
 217  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 218  * set (in which case it could also be direct WRITE).
 219  */
 220 static inline int cfq_bio_sync(struct bio *bio)
 221 {
 222         if (bio_data_dir(bio) == READ || bio_sync(bio))
 223                 return 1;
 224
 225         return 0;
 226 }
 227
 228 /*
 229  * scheduler run of queue, if there are requests pending and no one in the
 230  * driver that will restart queueing
 231  */
 232 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 233 {
 234         if (cfqd->busy_queues)
 235                 kblockd_schedule_work(&cfqd->unplug_work);
 236 }
 237
 238 static int cfq_queue_empty(struct request_queue *q)
 239 {
 240         struct cfq_data *cfqd = q->elevator->elevator_data;
 241
 242         return !cfqd->busy_queues;
 243 }
 244
 245 /*
 246  * Scale schedule slice based on io priority. Use the sync time slice only
 247  * if a queue is marked sync and has sync io queued. A sync queue with async
 248  * io only, should not get full sync slice length.
 249  */
 250 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
 251                                  unsigned short prio)
 252 {
 253         const int base_slice = cfqd->cfq_slice[sync];
 254
 255         WARN_ON(prio >= IOPRIO_BE_NR);
 256
 257         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 258 }
 259
 260 static inline int
 261 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 262 {
 263         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 264 }
 265
 266 static inline void
 267 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 268 {
 269         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 270 }
 271
 272 /*
 273  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
 274  * isn't valid until the first request from the dispatch is activated
 275  * and the slice time set.
 276  */
 277 static inline int cfq_slice_used(struct cfq_queue *cfqq)
 278 {
 279         if (cfq_cfqq_slice_new(cfqq))
 280                 return 0;
 281         if (time_before(jiffies, cfqq->slice_end))
 282                 return 0;
 283
 284         return 1;
 285 }
 286
 287 /*
 288  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 289  * We choose the request that is closest to the head right now. Distance
 290  * behind the head is penalized and only allowed to a certain extent.
 291  */
 292 static struct request *
 293 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
 294 {
 295         sector_t last, s1, s2, d1 = 0, d2 = 0;
 296         unsigned long back_max;
 297 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 298 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 299         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 300
 301         if (rq1 == NULL || rq1 == rq2)
 302                 return rq2;
 303         if (rq2 == NULL)
 304                 return rq1;
 305
 306         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 307                 return rq1;
 308         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 309                 return rq2;
 310         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
 311                 return rq1;
 312         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
 313                 return rq2;
 314
 315         s1 = rq1->sector;
 316         s2 = rq2->sector;
 317
 318         last = cfqd->last_position;
 319
 320         /*
 321          * by definition, 1KiB is 2 sectors
 322          */
 323         back_max = cfqd->cfq_back_max * 2;
 324
 325         /*
 326          * Strict one way elevator _except_ in the case where we allow
 327          * short backward seeks which are biased as twice the cost of a
 328          * similar forward seek.
 329          */
 330         if (s1 >= last)
 331                 d1 = s1 - last;
 332         else if (s1 + back_max >= last)
 333                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 334         else
 335                 wrap |= CFQ_RQ1_WRAP;
 336
 337         if (s2 >= last)
 338                 d2 = s2 - last;
 339         else if (s2 + back_max >= last)
 340                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 341         else
 342                 wrap |= CFQ_RQ2_WRAP;
 343
 344         /* Found required data */
 345
 346         /*
 347          * By doing switch() on the bit mask "wrap" we avoid having to
 348          * check two variables for all permutations: --> faster!
 349          */
 350         switch (wrap) {
 351         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 352                 if (d1 < d2)
 353                         return rq1;
 354                 else if (d2 < d1)
 355                         return rq2;
 356                 else {
 357                         if (s1 >= s2)
 358                                 return rq1;
 359                         else
 360                                 return rq2;
 361                 }
 362
 363         case CFQ_RQ2_WRAP:
 364                 return rq1;
 365         case CFQ_RQ1_WRAP:
 366                 return rq2;
 367         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
 368         default:
 369                 /*
 370                  * Since both rqs are wrapped,
 371                  * start with the one that's further behind head
 372                  * (--> only *one* back seek required),
 373                  * since back seek takes more time than forward.
 374                  */
 375                 if (s1 <= s2)
 376                         return rq1;
 377                 else
 378                         return rq2;
 379         }
 380 }
 381
 382 /*
 383  * The below is leftmost cache rbtree addon
 384  */
 385 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
 386 {
 387         if (!root->left)
 388                 root->left = rb_first(&root->rb);
 389
 390         if (root->left)
 391                 return rb_entry(root->left, struct cfq_queue, rb_node);
 392
 393         return NULL;
 394 }
 395
 396 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
 397 {
 398         if (root->left == n)
 399                 root->left = NULL;
 400
 401         rb_erase(n, &root->rb);
 402         RB_CLEAR_NODE(n);
 403 }
 404
 405 /*
 406  * would be nice to take fifo expire time into account as well
 407  */
 408 static struct request *
 409 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 410                   struct request *last)
 411 {
 412         struct rb_node *rbnext = rb_next(&last->rb_node);
 413         struct rb_node *rbprev = rb_prev(&last->rb_node);
 414         struct request *next = NULL, *prev = NULL;
 415
 416         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
 417
 418         if (rbprev)
 419                 prev = rb_entry_rq(rbprev);
 420
 421         if (rbnext)
 422                 next = rb_entry_rq(rbnext);
 423         else {
 424                 rbnext = rb_first(&cfqq->sort_list);
 425                 if (rbnext && rbnext != &last->rb_node)
 426                         next = rb_entry_rq(rbnext);
 427         }
 428
 429         return cfq_choose_req(cfqd, next, prev);
 430 }
 431
 432 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
 433                                       struct cfq_queue *cfqq)
 434 {
 435         /*
 436          * just an approximation, should be ok.
 437          */
 438         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
 439                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
 440 }
 441
 442 /*
 443  * The cfqd->service_tree holds all pending cfq_queue's that have
 444  * requests waiting to be processed. It is sorted in the order that
 445  * we will service the queues.
 446  */
 447 static void cfq_service_tree_add(struct cfq_data *cfqd,
 448                                     struct cfq_queue *cfqq, int add_front)
 449 {
 450         struct rb_node **p, *parent;
 451         struct cfq_queue *__cfqq;
 452         unsigned long rb_key;
 453         int left;
 454
 455         if (cfq_class_idle(cfqq)) {
 456                 rb_key = CFQ_IDLE_DELAY;
 457                 parent = rb_last(&cfqd->service_tree.rb);
 458                 if (parent && parent != &cfqq->rb_node) {
 459                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
 460                         rb_key += __cfqq->rb_key;
 461                 } else
 462                         rb_key += jiffies;
 463         } else if (!add_front) {
 464                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
 465                 rb_key += cfqq->slice_resid;
 466                 cfqq->slice_resid = 0;
 467         } else
 468                 rb_key = 0;
 469
 470         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
 471                 /*
 472                  * same position, nothing more to do
 473                  */
 474                 if (rb_key == cfqq->rb_key)
 475                         return;
 476
 477                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
 478         }
 479
 480         left = 1;
 481         parent = NULL;
 482         p = &cfqd->service_tree.rb.rb_node;
 483         while (*p) {
 484                 struct rb_node **n;
 485
 486                 parent = *p;
 487                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
 488
 489                 /*
 490                  * sort RT queues first, we always want to give
 491                  * preference to them. IDLE queues goes to the back.
 492                  * after that, sort on the next service time.
 493                  */
 494                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
 495                         n = &(*p)->rb_left;
 496                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
 497                         n = &(*p)->rb_right;
 498                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
 499                         n = &(*p)->rb_left;
 500                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
 501                         n = &(*p)->rb_right;
 502                 else if (rb_key < __cfqq->rb_key)
 503                         n = &(*p)->rb_left;
 504                 else
 505                         n = &(*p)->rb_right;
 506
 507                 if (n == &(*p)->rb_right)
 508                         left = 0;
 509
 510                 p = n;
 511         }
 512
 513         if (left)
 514                 cfqd->service_tree.left = &cfqq->rb_node;
 515
 516         cfqq->rb_key = rb_key;
 517         rb_link_node(&cfqq->rb_node, parent, p);
 518         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
 519 }
 520
 521 /*
 522  * Update cfqq's position in the service tree.
 523  */
 524 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 525 {
 526         /*
 527          * Resorting requires the cfqq to be on the RR list already.
 528          */
 529         if (cfq_cfqq_on_rr(cfqq))
 530                 cfq_service_tree_add(cfqd, cfqq, 0);
 531 }
 532
 533 /*
 534  * add to busy list of queues for service, trying to be fair in ordering
 535  * the pending list according to last request service
 536  */
 537 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 538 {
 539         BUG_ON(cfq_cfqq_on_rr(cfqq));
 540         cfq_mark_cfqq_on_rr(cfqq);
 541         cfqd->busy_queues++;
 542
 543         cfq_resort_rr_list(cfqd, cfqq);
 544 }
 545
 546 /*
 547  * Called when the cfqq no longer has requests pending, remove it from
 548  * the service tree.
 549  */
 550 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 551 {
 552         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 553         cfq_clear_cfqq_on_rr(cfqq);
 554
 555         if (!RB_EMPTY_NODE(&cfqq->rb_node))
 556                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
 557
 558         BUG_ON(!cfqd->busy_queues);
 559         cfqd->busy_queues--;
 560 }
 561
 562 /*
 563  * rb tree support functions
 564  */
 565 static void cfq_del_rq_rb(struct request *rq)
 566 {
 567         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 568         struct cfq_data *cfqd = cfqq->cfqd;
 569         const int sync = rq_is_sync(rq);
 570
 571         BUG_ON(!cfqq->queued[sync]);
 572         cfqq->queued[sync]--;
 573
 574         elv_rb_del(&cfqq->sort_list, rq);
 575
 576         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
 577                 cfq_del_cfqq_rr(cfqd, cfqq);
 578 }
 579
 580 static void cfq_add_rq_rb(struct request *rq)
 581 {
 582         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 583         struct cfq_data *cfqd = cfqq->cfqd;
 584         struct request *__alias;
 585
 586         cfqq->queued[rq_is_sync(rq)]++;
 587
 588         /*
 589          * looks a little odd, but the first insert might return an alias.
 590          * if that happens, put the alias on the dispatch list
 591          */
 592         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
 593                 cfq_dispatch_insert(cfqd->queue, __alias);
 594
 595         if (!cfq_cfqq_on_rr(cfqq))
 596                 cfq_add_cfqq_rr(cfqd, cfqq);
 597
 598         /*
 599          * check if this request is a better next-serve candidate
 600          */
 601         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
 602         BUG_ON(!cfqq->next_rq);
 603 }
 604
 605 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
 606 {
 607         elv_rb_del(&cfqq->sort_list, rq);
 608         cfqq->queued[rq_is_sync(rq)]--;
 609         cfq_add_rq_rb(rq);
 610 }
 611
 612 static struct request *
 613 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 614 {
 615         struct task_struct *tsk = current;
 616         struct cfq_io_context *cic;
 617         struct cfq_queue *cfqq;
 618
 619         cic = cfq_cic_lookup(cfqd, tsk->io_context);
 620         if (!cic)
 621                 return NULL;
 622
 623         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
 624         if (cfqq) {
 625                 sector_t sector = bio->bi_sector + bio_sectors(bio);
 626
 627                 return elv_rb_find(&cfqq->sort_list, sector);
 628         }
 629
 630         return NULL;
 631 }
 632
 633 static void cfq_activate_request(struct request_queue *q, struct request *rq)
 634 {
 635         struct cfq_data *cfqd = q->elevator->elevator_data;
 636
 637         cfqd->rq_in_driver++;
 638
 639         /*
 640          * If the depth is larger 1, it really could be queueing. But lets
 641          * make the mark a little higher - idling could still be good for
 642          * low queueing, and a low queueing number could also just indicate
 643          * a SCSI mid layer like behaviour where limit+1 is often seen.
 644          */
 645         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
 646                 cfqd->hw_tag = 1;
 647
 648         cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
 649 }
 650
 651 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
 652 {
 653         struct cfq_data *cfqd = q->elevator->elevator_data;
 654
 655         WARN_ON(!cfqd->rq_in_driver);
 656         cfqd->rq_in_driver--;
 657 }
 658
 659 static void cfq_remove_request(struct request *rq)
 660 {
 661         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 662
 663         if (cfqq->next_rq == rq)
 664                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
 665
 666         list_del_init(&rq->queuelist);
 667         cfq_del_rq_rb(rq);
 668
 669         if (rq_is_meta(rq)) {
 670                 WARN_ON(!cfqq->meta_pending);
 671                 cfqq->meta_pending--;
 672         }
 673 }
 674
 675 static int cfq_merge(struct request_queue *q, struct request **req,
 676                      struct bio *bio)
 677 {
 678         struct cfq_data *cfqd = q->elevator->elevator_data;
 679         struct request *__rq;
 680
 681         __rq = cfq_find_rq_fmerge(cfqd, bio);
 682         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 683                 *req = __rq;
 684                 return ELEVATOR_FRONT_MERGE;
 685         }
 686
 687         return ELEVATOR_NO_MERGE;
 688 }
 689
 690 static void cfq_merged_request(struct request_queue *q, struct request *req,
 691                                int type)
 692 {
 693         if (type == ELEVATOR_FRONT_MERGE) {
 694                 struct cfq_queue *cfqq = RQ_CFQQ(req);
 695
 696                 cfq_reposition_rq_rb(cfqq, req);
 697         }
 698 }
 699
 700 static void
 701 cfq_merged_requests(struct request_queue *q, struct request *rq,
 702                     struct request *next)
 703 {
 704         /*
 705          * reposition in fifo if next is older than rq
 706          */
 707         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 708             time_before(next->start_time, rq->start_time))
 709                 list_move(&rq->queuelist, &next->queuelist);
 710
 711         cfq_remove_request(next);
 712 }
 713
 714 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
 715                            struct bio *bio)
 716 {
 717         struct cfq_data *cfqd = q->elevator->elevator_data;
 718         struct cfq_io_context *cic;
 719         struct cfq_queue *cfqq;
 720
 721         /*
 722          * Disallow merge of a sync bio into an async request.
 723          */
 724         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
 725                 return 0;
 726
 727         /*
 728          * Lookup the cfqq that this bio will be queued with. Allow
 729          * merge only if rq is queued there.
 730          */
 731         cic = cfq_cic_lookup(cfqd, current->io_context);
 732         if (!cic)
 733                 return 0;
 734
 735         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
 736         if (cfqq == RQ_CFQQ(rq))
 737                 return 1;
 738
 739         return 0;
 740 }
 741
 742 static void __cfq_set_active_queue(struct cfq_data *cfqd,
 743                                    struct cfq_queue *cfqq)
 744 {
 745         if (cfqq) {
 746                 cfqq->slice_end = 0;
 747                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 748                 cfq_clear_cfqq_fifo_expire(cfqq);
 749                 cfq_mark_cfqq_slice_new(cfqq);
 750                 cfq_clear_cfqq_queue_new(cfqq);
 751         }
 752
 753         cfqd->active_queue = cfqq;
 754 }
 755
 756 /*
 757  * current cfqq expired its slice (or was too idle), select new one
 758  */
 759 static void
 760 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 761                     int timed_out)
 762 {
 763         if (cfq_cfqq_wait_request(cfqq))
 764                 del_timer(&cfqd->idle_slice_timer);
 765
 766         cfq_clear_cfqq_must_dispatch(cfqq);
 767         cfq_clear_cfqq_wait_request(cfqq);
 768
 769         /*
 770          * store what was left of this slice, if the queue idled/timed out
 771          */
 772         if (timed_out && !cfq_cfqq_slice_new(cfqq))
 773                 cfqq->slice_resid = cfqq->slice_end - jiffies;
 774
 775         cfq_resort_rr_list(cfqd, cfqq);
 776
 777         if (cfqq == cfqd->active_queue)
 778                 cfqd->active_queue = NULL;
 779
 780         if (cfqd->active_cic) {
 781                 put_io_context(cfqd->active_cic->ioc);
 782                 cfqd->active_cic = NULL;
 783         }
 784 }
 785
 786 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
 787 {
 788         struct cfq_queue *cfqq = cfqd->active_queue;
 789
 790         if (cfqq)
 791                 __cfq_slice_expired(cfqd, cfqq, timed_out);
 792 }
 793
 794 /*
 795  * Get next queue for service. Unless we have a queue preemption,
 796  * we'll simply select the first cfqq in the service tree.
 797  */
 798 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
 799 {
 800         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
 801                 return NULL;
 802
 803         return cfq_rb_first(&cfqd->service_tree);
 804 }
 805
 806 /*
 807  * Get and set a new active queue for service.
 808  */
 809 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 810 {
 811         struct cfq_queue *cfqq;
 812
 813         cfqq = cfq_get_next_queue(cfqd);
 814         __cfq_set_active_queue(cfqd, cfqq);
 815         return cfqq;
 816 }
 817
 818 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
 819                                           struct request *rq)
 820 {
 821         if (rq->sector >= cfqd->last_position)
 822                 return rq->sector - cfqd->last_position;
 823         else
 824                 return cfqd->last_position - rq->sector;
 825 }
 826
 827 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
 828 {
 829         struct cfq_io_context *cic = cfqd->active_cic;
 830
 831         if (!sample_valid(cic->seek_samples))
 832                 return 0;
 833
 834         return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
 835 }
 836
 837 static int cfq_close_cooperator(struct cfq_data *cfq_data,
 838                                 struct cfq_queue *cfqq)
 839 {
 840         /*
 841          * We should notice if some of the queues are cooperating, eg
 842          * working closely on the same area of the disk. In that case,
 843          * we can group them together and don't waste time idling.
 844          */
 845         return 0;
 846 }
 847
 848 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
 849
 850 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
 851 {
 852         struct cfq_queue *cfqq = cfqd->active_queue;
 853         struct cfq_io_context *cic;
 854         unsigned long sl;
 855
 856         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
 857         WARN_ON(cfq_cfqq_slice_new(cfqq));
 858
 859         /*
 860          * idle is disabled, either manually or by past process history
 861          */
 862         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
 863                 return;
 864
 865         /*
 866          * task has exited, don't wait
 867          */
 868         cic = cfqd->active_cic;
 869         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
 870                 return;
 871
 872         /*
 873          * See if this prio level has a good candidate
 874          */
 875         if (cfq_close_cooperator(cfqd, cfqq) &&
 876             (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
 877                 return;
 878
 879         cfq_mark_cfqq_must_dispatch(cfqq);
 880         cfq_mark_cfqq_wait_request(cfqq);
 881
 882         /*
 883          * we don't want to idle for seeks, but we do want to allow
 884          * fair distribution of slice time for a process doing back-to-back
 885          * seeks. so allow a little bit of time for him to submit a new rq
 886          */
 887         sl = cfqd->cfq_slice_idle;
 888         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
 889                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
 890
 891         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 892 }
 893
 894 /*
 895  * Move request from internal lists to the request queue dispatch list.
 896  */
 897 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
 898 {
 899         struct cfq_data *cfqd = q->elevator->elevator_data;
 900         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 901
 902         cfq_remove_request(rq);
 903         cfqq->dispatched++;
 904         elv_dispatch_sort(q, rq);
 905
 906         if (cfq_cfqq_sync(cfqq))
 907                 cfqd->sync_flight++;
 908 }
 909
 910 /*
 911  * return expired entry, or NULL to just start from scratch in rbtree
 912  */
 913 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
 914 {
 915         struct cfq_data *cfqd = cfqq->cfqd;
 916         struct request *rq;
 917         int fifo;
 918
 919         if (cfq_cfqq_fifo_expire(cfqq))
 920                 return NULL;
 921
 922         cfq_mark_cfqq_fifo_expire(cfqq);
 923
 924         if (list_empty(&cfqq->fifo))
 925                 return NULL;
 926
 927         fifo = cfq_cfqq_sync(cfqq);
 928         rq = rq_entry_fifo(cfqq->fifo.next);
 929
 930         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
 931                 return NULL;
 932
 933         return rq;
 934 }
 935
 936 static inline int
 937 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 938 {
 939         const int base_rq = cfqd->cfq_slice_async_rq;
 940
 941         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 942
 943         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
 944 }
 945
 946 /*
 947  * Select a queue for service. If we have a current active queue,
 948  * check whether to continue servicing it, or retrieve and set a new one.
 949  */
 950 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
 951 {
 952         struct cfq_queue *cfqq;
 953
 954         cfqq = cfqd->active_queue;
 955         if (!cfqq)
 956                 goto new_queue;
 957
 958         /*
 959          * The active queue has run out of time, expire it and select new.
 960          */
 961         if (cfq_slice_used(cfqq))
 962                 goto expire;
 963
 964         /*
 965          * The active queue has requests and isn't expired, allow it to
 966          * dispatch.
 967          */
 968         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
 969                 goto keep_queue;
 970
 971         /*
 972          * No requests pending. If the active queue still has requests in
 973          * flight or is idling for a new request, allow either of these
 974          * conditions to happen (or time out) before selecting a new queue.
 975          */
 976         if (timer_pending(&cfqd->idle_slice_timer) ||
 977             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
 978                 cfqq = NULL;
 979                 goto keep_queue;
 980         }
 981
 982 expire:
 983         cfq_slice_expired(cfqd, 0);
 984 new_queue:
 985         cfqq = cfq_set_active_queue(cfqd);
 986 keep_queue:
 987         return cfqq;
 988 }
 989
 990 /*
 991  * Dispatch some requests from cfqq, moving them to the request queue
 992  * dispatch list.
 993  */
 994 static int
 995 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 996                         int max_dispatch)
 997 {
 998         int dispatched = 0;
 999
1000         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1001
1002         do {
1003                 struct request *rq;
1004
1005                 /*
1006                  * follow expired path, else get first next available
1007                  */
1008                 if ((rq = cfq_check_fifo(cfqq)) == NULL)
1009                         rq = cfqq->next_rq;
1010
1011                 /*
1012                  * finally, insert request into driver dispatch list
1013                  */
1014                 cfq_dispatch_insert(cfqd->queue, rq);
1015
1016                 dispatched++;
1017
1018                 if (!cfqd->active_cic) {
1019                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1020                         cfqd->active_cic = RQ_CIC(rq);
1021                 }
1022
1023                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1024                         break;
1025
1026         } while (dispatched < max_dispatch);
1027
1028         /*
1029          * expire an async queue immediately if it has used up its slice. idle
1030          * queue always expire after 1 dispatch round.
1031          */
1032         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1033             dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1034             cfq_class_idle(cfqq))) {
1035                 cfqq->slice_end = jiffies + 1;
1036                 cfq_slice_expired(cfqd, 0);
1037         }
1038
1039         return dispatched;
1040 }
1041
1042 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1043 {
1044         int dispatched = 0;
1045
1046         while (cfqq->next_rq) {
1047                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1048                 dispatched++;
1049         }
1050
1051         BUG_ON(!list_empty(&cfqq->fifo));
1052         return dispatched;
1053 }
1054
1055 /*
1056  * Drain our current requests. Used for barriers and when switching
1057  * io schedulers on-the-fly.
1058  */
1059 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1060 {
1061         struct cfq_queue *cfqq;
1062         int dispatched = 0;
1063
1064         while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1065                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1066
1067         cfq_slice_expired(cfqd, 0);
1068
1069         BUG_ON(cfqd->busy_queues);
1070
1071         return dispatched;
1072 }
1073
1074 static int cfq_dispatch_requests(struct request_queue *q, int force)
1075 {
1076         struct cfq_data *cfqd = q->elevator->elevator_data;
1077         struct cfq_queue *cfqq;
1078         int dispatched;
1079
1080         if (!cfqd->busy_queues)
1081                 return 0;
1082
1083         if (unlikely(force))
1084                 return cfq_forced_dispatch(cfqd);
1085
1086         dispatched = 0;
1087         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1088                 int max_dispatch;
1089
1090                 max_dispatch = cfqd->cfq_quantum;
1091                 if (cfq_class_idle(cfqq))
1092                         max_dispatch = 1;
1093
1094                 if (cfqq->dispatched >= max_dispatch) {
1095                         if (cfqd->busy_queues > 1)
1096                                 break;
1097                         if (cfqq->dispatched >= 4 * max_dispatch)
1098                                 break;
1099                 }
1100
1101                 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1102                         break;
1103
1104                 cfq_clear_cfqq_must_dispatch(cfqq);
1105                 cfq_clear_cfqq_wait_request(cfqq);
1106                 del_timer(&cfqd->idle_slice_timer);
1107
1108                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1109         }
1110
1111         return dispatched;
1112 }
1113
1114 /*
1115  * task holds one reference to the queue, dropped when task exits. each rq
1116  * in-flight on this queue also holds a reference, dropped when rq is freed.
1117  *
1118  * queue lock must be held here.
1119  */
1120 static void cfq_put_queue(struct cfq_queue *cfqq)
1121 {
1122         struct cfq_data *cfqd = cfqq->cfqd;
1123
1124         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1125
1126         if (!atomic_dec_and_test(&cfqq->ref))
1127                 return;
1128
1129         BUG_ON(rb_first(&cfqq->sort_list));
1130         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1131         BUG_ON(cfq_cfqq_on_rr(cfqq));
1132
1133         if (unlikely(cfqd->active_queue == cfqq)) {
1134                 __cfq_slice_expired(cfqd, cfqq, 0);
1135                 cfq_schedule_dispatch(cfqd);
1136         }
1137
1138         kmem_cache_free(cfq_pool, cfqq);
1139 }
1140
1141 /*
1142  * Call func for each cic attached to this ioc. Returns number of cic's seen.
1143  */
1144 #define CIC_GANG_NR     16
1145 static unsigned int
1146 call_for_each_cic(struct io_context *ioc,
1147                   void (*func)(struct io_context *, struct cfq_io_context *))
1148 {
1149         struct cfq_io_context *cics[CIC_GANG_NR];
1150         unsigned long index = 0;
1151         unsigned int called = 0;
1152         int nr;
1153
1154         rcu_read_lock();
1155
1156         do {
1157                 int i;
1158
1159                 /*
1160                  * Perhaps there's a better way - this just gang lookups from
1161                  * 0 to the end, restarting after each CIC_GANG_NR from the
1162                  * last key + 1.
1163                  */
1164                 nr = radix_tree_gang_lookup(&ioc->radix_root, (void **) cics,
1165                                                 index, CIC_GANG_NR);
1166                 if (!nr)
1167                         break;
1168
1169                 called += nr;
1170                 index = 1 + (unsigned long) cics[nr - 1]->key;
1171
1172                 for (i = 0; i < nr; i++)
1173                         func(ioc, cics[i]);
1174         } while (nr == CIC_GANG_NR);
1175
1176         rcu_read_unlock();
1177
1178         return called;
1179 }
1180
1181 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1182 {
1183         unsigned long flags;
1184
1185         BUG_ON(!cic->dead_key);
1186
1187         spin_lock_irqsave(&ioc->lock, flags);
1188         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1189         spin_unlock_irqrestore(&ioc->lock, flags);
1190
1191         kmem_cache_free(cfq_ioc_pool, cic);
1192 }
1193
1194 static void cfq_free_io_context(struct io_context *ioc)
1195 {
1196         int freed;
1197
1198         /*
1199          * ioc->refcount is zero here, so no more cic's are allowed to be
1200          * linked into this ioc. So it should be ok to iterate over the known
1201          * list, we will see all cic's since no new ones are added.
1202          */
1203         freed = call_for_each_cic(ioc, cic_free_func);
1204
1205         elv_ioc_count_mod(ioc_count, -freed);
1206
1207         if (ioc_gone && !elv_ioc_count_read(ioc_count))
1208                 complete(ioc_gone);
1209 }
1210
1211 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1212 {
1213         if (unlikely(cfqq == cfqd->active_queue)) {
1214                 __cfq_slice_expired(cfqd, cfqq, 0);
1215                 cfq_schedule_dispatch(cfqd);
1216         }
1217
1218         cfq_put_queue(cfqq);
1219 }
1220
1221 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1222                                          struct cfq_io_context *cic)
1223 {
1224         list_del_init(&cic->queue_list);
1225
1226         /*
1227          * Make sure key == NULL is seen for dead queues
1228          */
1229         smp_wmb();
1230         cic->dead_key = (unsigned long) cic->key;
1231         cic->key = NULL;
1232
1233         if (cic->cfqq[ASYNC]) {
1234                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1235                 cic->cfqq[ASYNC] = NULL;
1236         }
1237
1238         if (cic->cfqq[SYNC]) {
1239                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1240                 cic->cfqq[SYNC] = NULL;
1241         }
1242 }
1243
1244 static void cfq_exit_single_io_context(struct io_context *ioc,
1245                                        struct cfq_io_context *cic)
1246 {
1247         struct cfq_data *cfqd = cic->key;
1248
1249         if (cfqd) {
1250                 struct request_queue *q = cfqd->queue;
1251                 unsigned long flags;
1252
1253                 spin_lock_irqsave(q->queue_lock, flags);
1254                 __cfq_exit_single_io_context(cfqd, cic);
1255                 spin_unlock_irqrestore(q->queue_lock, flags);
1256         }
1257 }
1258
1259 /*
1260  * The process that ioc belongs to has exited, we need to clean up
1261  * and put the internal structures we have that belongs to that process.
1262  */
1263 static void cfq_exit_io_context(struct io_context *ioc)
1264 {
1265         rcu_assign_pointer(ioc->ioc_data, NULL);
1266         call_for_each_cic(ioc, cfq_exit_single_io_context);
1267 }
1268
1269 static struct cfq_io_context *
1270 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1271 {
1272         struct cfq_io_context *cic;
1273
1274         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1275                                                         cfqd->queue->node);
1276         if (cic) {
1277                 cic->last_end_request = jiffies;
1278                 INIT_LIST_HEAD(&cic->queue_list);
1279                 cic->dtor = cfq_free_io_context;
1280                 cic->exit = cfq_exit_io_context;
1281                 elv_ioc_count_inc(ioc_count);
1282         }
1283
1284         return cic;
1285 }
1286
1287 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1288 {
1289         struct task_struct *tsk = current;
1290         int ioprio_class;
1291
1292         if (!cfq_cfqq_prio_changed(cfqq))
1293                 return;
1294
1295         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1296         switch (ioprio_class) {
1297                 default:
1298                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1299                 case IOPRIO_CLASS_NONE:
1300                         /*
1301                          * no prio set, place us in the middle of the BE classes
1302                          */
1303                         cfqq->ioprio = task_nice_ioprio(tsk);
1304                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1305                         break;
1306                 case IOPRIO_CLASS_RT:
1307                         cfqq->ioprio = task_ioprio(ioc);
1308                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1309                         break;
1310                 case IOPRIO_CLASS_BE:
1311                         cfqq->ioprio = task_ioprio(ioc);
1312                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1313                         break;
1314                 case IOPRIO_CLASS_IDLE:
1315                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1316                         cfqq->ioprio = 7;
1317                         cfq_clear_cfqq_idle_window(cfqq);
1318                         break;
1319         }
1320
1321         /*
1322          * keep track of original prio settings in case we have to temporarily
1323          * elevate the priority of this queue
1324          */
1325         cfqq->org_ioprio = cfqq->ioprio;
1326         cfqq->org_ioprio_class = cfqq->ioprio_class;
1327         cfq_clear_cfqq_prio_changed(cfqq);
1328 }
1329
1330 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1331 {
1332         struct cfq_data *cfqd = cic->key;
1333         struct cfq_queue *cfqq;
1334         unsigned long flags;
1335
1336         if (unlikely(!cfqd))
1337                 return;
1338
1339         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1340
1341         cfqq = cic->cfqq[ASYNC];
1342         if (cfqq) {
1343                 struct cfq_queue *new_cfqq;
1344                 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
1345                 if (new_cfqq) {
1346                         cic->cfqq[ASYNC] = new_cfqq;
1347                         cfq_put_queue(cfqq);
1348                 }
1349         }
1350
1351         cfqq = cic->cfqq[SYNC];
1352         if (cfqq)
1353                 cfq_mark_cfqq_prio_changed(cfqq);
1354
1355         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1356 }
1357
1358 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1359 {
1360         call_for_each_cic(ioc, changed_ioprio);
1361         ioc->ioprio_changed = 0;
1362 }
1363
1364 static struct cfq_queue *
1365 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1366                      struct io_context *ioc, gfp_t gfp_mask)
1367 {
1368         struct cfq_queue *cfqq, *new_cfqq = NULL;
1369         struct cfq_io_context *cic;
1370
1371 retry:
1372         cic = cfq_cic_lookup(cfqd, ioc);
1373         /* cic always exists here */
1374         cfqq = cic_to_cfqq(cic, is_sync);
1375
1376         if (!cfqq) {
1377                 if (new_cfqq) {
1378                         cfqq = new_cfqq;
1379                         new_cfqq = NULL;
1380                 } else if (gfp_mask & __GFP_WAIT) {
1381                         /*
1382                          * Inform the allocator of the fact that we will
1383                          * just repeat this allocation if it fails, to allow
1384                          * the allocator to do whatever it needs to attempt to
1385                          * free memory.
1386                          */
1387                         spin_unlock_irq(cfqd->queue->queue_lock);
1388                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1389                                         gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1390                                         cfqd->queue->node);
1391                         spin_lock_irq(cfqd->queue->queue_lock);
1392                         goto retry;
1393                 } else {
1394                         cfqq = kmem_cache_alloc_node(cfq_pool,
1395                                         gfp_mask | __GFP_ZERO,
1396                                         cfqd->queue->node);
1397                         if (!cfqq)
1398                                 goto out;
1399                 }
1400
1401                 RB_CLEAR_NODE(&cfqq->rb_node);
1402                 INIT_LIST_HEAD(&cfqq->fifo);
1403
1404                 atomic_set(&cfqq->ref, 0);
1405                 cfqq->cfqd = cfqd;
1406
1407                 cfq_mark_cfqq_prio_changed(cfqq);
1408                 cfq_mark_cfqq_queue_new(cfqq);
1409
1410                 cfq_init_prio_data(cfqq, ioc);
1411
1412                 if (is_sync) {
1413                         if (!cfq_class_idle(cfqq))
1414                                 cfq_mark_cfqq_idle_window(cfqq);
1415                         cfq_mark_cfqq_sync(cfqq);
1416                 }
1417         }
1418
1419         if (new_cfqq)
1420                 kmem_cache_free(cfq_pool, new_cfqq);
1421
1422 out:
1423         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1424         return cfqq;
1425 }
1426
1427 static struct cfq_queue **
1428 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1429 {
1430         switch(ioprio_class) {
1431         case IOPRIO_CLASS_RT:
1432                 return &cfqd->async_cfqq[0][ioprio];
1433         case IOPRIO_CLASS_BE:
1434                 return &cfqd->async_cfqq[1][ioprio];
1435         case IOPRIO_CLASS_IDLE:
1436                 return &cfqd->async_idle_cfqq;
1437         default:
1438                 BUG();
1439         }
1440 }
1441
1442 static struct cfq_queue *
1443 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1444               gfp_t gfp_mask)
1445 {
1446         const int ioprio = task_ioprio(ioc);
1447         const int ioprio_class = task_ioprio_class(ioc);
1448         struct cfq_queue **async_cfqq = NULL;
1449         struct cfq_queue *cfqq = NULL;
1450
1451         if (!is_sync) {
1452                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1453                 cfqq = *async_cfqq;
1454         }
1455
1456         if (!cfqq) {
1457                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1458                 if (!cfqq)
1459                         return NULL;
1460         }
1461
1462         /*
1463          * pin the queue now that it's allocated, scheduler exit will prune it
1464          */
1465         if (!is_sync && !(*async_cfqq)) {
1466                 atomic_inc(&cfqq->ref);
1467                 *async_cfqq = cfqq;
1468         }
1469
1470         atomic_inc(&cfqq->ref);
1471         return cfqq;
1472 }
1473
1474 static void cfq_cic_free(struct cfq_io_context *cic)
1475 {
1476         kmem_cache_free(cfq_ioc_pool, cic);
1477         elv_ioc_count_dec(ioc_count);
1478
1479         if (ioc_gone && !elv_ioc_count_read(ioc_count))
1480                 complete(ioc_gone);
1481 }
1482
1483 /*
1484  * We drop cfq io contexts lazily, so we may find a dead one.
1485  */
1486 static void
1487 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1488                   struct cfq_io_context *cic)
1489 {
1490         unsigned long flags;
1491
1492         WARN_ON(!list_empty(&cic->queue_list));
1493
1494         spin_lock_irqsave(&ioc->lock, flags);
1495
1496         if (ioc->ioc_data == cic)
1497                 rcu_assign_pointer(ioc->ioc_data, NULL);
1498
1499         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1500         spin_unlock_irqrestore(&ioc->lock, flags);
1501
1502         cfq_cic_free(cic);
1503 }
1504
1505 static struct cfq_io_context *
1506 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1507 {
1508         struct cfq_io_context *cic;
1509         void *k;
1510
1511         if (unlikely(!ioc))
1512                 return NULL;
1513
1514         /*
1515          * we maintain a last-hit cache, to avoid browsing over the tree
1516          */
1517         cic = rcu_dereference(ioc->ioc_data);
1518         if (cic && cic->key == cfqd)
1519                 return cic;
1520
1521         do {
1522                 rcu_read_lock();
1523                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1524                 rcu_read_unlock();
1525                 if (!cic)
1526                         break;
1527                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1528                 k = cic->key;
1529                 if (unlikely(!k)) {
1530                         cfq_drop_dead_cic(cfqd, ioc, cic);
1531                         continue;
1532                 }
1533
1534                 rcu_assign_pointer(ioc->ioc_data, cic);
1535                 break;
1536         } while (1);
1537
1538         return cic;
1539 }
1540
1541 /*
1542  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1543  * the process specific cfq io context when entered from the block layer.
1544  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1545  */
1546 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1547                         struct cfq_io_context *cic, gfp_t gfp_mask)
1548 {
1549         unsigned long flags;
1550         int ret;
1551
1552         ret = radix_tree_preload(gfp_mask);
1553         if (!ret) {
1554                 cic->ioc = ioc;
1555                 cic->key = cfqd;
1556
1557                 spin_lock_irqsave(&ioc->lock, flags);
1558                 ret = radix_tree_insert(&ioc->radix_root,
1559                                                 (unsigned long) cfqd, cic);
1560                 spin_unlock_irqrestore(&ioc->lock, flags);
1561
1562                 radix_tree_preload_end();
1563
1564                 if (!ret) {
1565                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1566                         list_add(&cic->queue_list, &cfqd->cic_list);
1567                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1568                 }
1569         }
1570
1571         if (ret)
1572                 printk(KERN_ERR "cfq: cic link failed!\n");
1573
1574         return ret;
1575 }
1576
1577 /*
1578  * Setup general io context and cfq io context. There can be several cfq
1579  * io contexts per general io context, if this process is doing io to more
1580  * than one device managed by cfq.
1581  */
1582 static struct cfq_io_context *
1583 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1584 {
1585         struct io_context *ioc = NULL;
1586         struct cfq_io_context *cic;
1587
1588         might_sleep_if(gfp_mask & __GFP_WAIT);
1589
1590         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1591         if (!ioc)
1592                 return NULL;
1593
1594         cic = cfq_cic_lookup(cfqd, ioc);
1595         if (cic)
1596                 goto out;
1597
1598         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1599         if (cic == NULL)
1600                 goto err;
1601
1602         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1603                 goto err_free;
1604
1605 out:
1606         smp_read_barrier_depends();
1607         if (unlikely(ioc->ioprio_changed))
1608                 cfq_ioc_set_ioprio(ioc);
1609
1610         return cic;
1611 err_free:
1612         cfq_cic_free(cic);
1613 err:
1614         put_io_context(ioc);
1615         return NULL;
1616 }
1617
1618 static void
1619 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1620 {
1621         unsigned long elapsed = jiffies - cic->last_end_request;
1622         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1623
1624         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1625         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1626         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1627 }
1628
1629 static void
1630 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1631                        struct request *rq)
1632 {
1633         sector_t sdist;
1634         u64 total;
1635
1636         if (cic->last_request_pos < rq->sector)
1637                 sdist = rq->sector - cic->last_request_pos;
1638         else
1639                 sdist = cic->last_request_pos - rq->sector;
1640
1641         /*
1642          * Don't allow the seek distance to get too large from the
1643          * odd fragment, pagein, etc
1644          */
1645         if (cic->seek_samples <= 60) /* second&third seek */
1646                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1647         else
1648                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1649
1650         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1651         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1652         total = cic->seek_total + (cic->seek_samples/2);
1653         do_div(total, cic->seek_samples);
1654         cic->seek_mean = (sector_t)total;
1655 }
1656
1657 /*
1658  * Disable idle window if the process thinks too long or seeks so much that
1659  * it doesn't matter
1660  */
1661 static void
1662 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1663                        struct cfq_io_context *cic)
1664 {
1665         int enable_idle;
1666
1667         /*
1668          * Don't idle for async or idle io prio class
1669          */
1670         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1671                 return;
1672
1673         enable_idle = cfq_cfqq_idle_window(cfqq);
1674
1675         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1676             (cfqd->hw_tag && CIC_SEEKY(cic)))
1677                 enable_idle = 0;
1678         else if (sample_valid(cic->ttime_samples)) {
1679                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1680                         enable_idle = 0;
1681                 else
1682                         enable_idle = 1;
1683         }
1684
1685         if (enable_idle)
1686                 cfq_mark_cfqq_idle_window(cfqq);
1687         else
1688                 cfq_clear_cfqq_idle_window(cfqq);
1689 }
1690
1691 /*
1692  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1693  * no or if we aren't sure, a 1 will cause a preempt.
1694  */
1695 static int
1696 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1697                    struct request *rq)
1698 {
1699         struct cfq_queue *cfqq;
1700
1701         cfqq = cfqd->active_queue;
1702         if (!cfqq)
1703                 return 0;
1704
1705         if (cfq_slice_used(cfqq))
1706                 return 1;
1707
1708         if (cfq_class_idle(new_cfqq))
1709                 return 0;
1710
1711         if (cfq_class_idle(cfqq))
1712                 return 1;
1713
1714         /*
1715          * if the new request is sync, but the currently running queue is
1716          * not, let the sync request have priority.
1717          */
1718         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1719                 return 1;
1720
1721         /*
1722          * So both queues are sync. Let the new request get disk time if
1723          * it's a metadata request and the current queue is doing regular IO.
1724          */
1725         if (rq_is_meta(rq) && !cfqq->meta_pending)
1726                 return 1;
1727
1728         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1729                 return 0;
1730
1731         /*
1732          * if this request is as-good as one we would expect from the
1733          * current cfqq, let it preempt
1734          */
1735         if (cfq_rq_close(cfqd, rq))
1736                 return 1;
1737
1738         return 0;
1739 }
1740
1741 /*
1742  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1743  * let it have half of its nominal slice.
1744  */
1745 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1746 {
1747         cfq_slice_expired(cfqd, 1);
1748
1749         /*
1750          * Put the new queue at the front of the of the current list,
1751          * so we know that it will be selected next.
1752          */
1753         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1754
1755         cfq_service_tree_add(cfqd, cfqq, 1);
1756
1757         cfqq->slice_end = 0;
1758         cfq_mark_cfqq_slice_new(cfqq);
1759 }
1760
1761 /*
1762  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1763  * something we should do about it
1764  */
1765 static void
1766 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1767                 struct request *rq)
1768 {
1769         struct cfq_io_context *cic = RQ_CIC(rq);
1770
1771         if (rq_is_meta(rq))
1772                 cfqq->meta_pending++;
1773
1774         cfq_update_io_thinktime(cfqd, cic);
1775         cfq_update_io_seektime(cfqd, cic, rq);
1776         cfq_update_idle_window(cfqd, cfqq, cic);
1777
1778         cic->last_request_pos = rq->sector + rq->nr_sectors;
1779
1780         if (cfqq == cfqd->active_queue) {
1781                 /*
1782                  * if we are waiting for a request for this queue, let it rip
1783                  * immediately and flag that we must not expire this queue
1784                  * just now
1785                  */
1786                 if (cfq_cfqq_wait_request(cfqq)) {
1787                         cfq_mark_cfqq_must_dispatch(cfqq);
1788                         del_timer(&cfqd->idle_slice_timer);
1789                         blk_start_queueing(cfqd->queue);
1790                 }
1791         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1792                 /*
1793                  * not the active queue - expire current slice if it is
1794                  * idle and has expired it's mean thinktime or this new queue
1795                  * has some old slice time left and is of higher priority
1796                  */
1797                 cfq_preempt_queue(cfqd, cfqq);
1798                 cfq_mark_cfqq_must_dispatch(cfqq);
1799                 blk_start_queueing(cfqd->queue);
1800         }
1801 }
1802
1803 static void cfq_insert_request(struct request_queue *q, struct request *rq)
1804 {
1805         struct cfq_data *cfqd = q->elevator->elevator_data;
1806         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1807
1808         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
1809
1810         cfq_add_rq_rb(rq);
1811
1812         list_add_tail(&rq->queuelist, &cfqq->fifo);
1813
1814         cfq_rq_enqueued(cfqd, cfqq, rq);
1815 }
1816
1817 static void cfq_completed_request(struct request_queue *q, struct request *rq)
1818 {
1819         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1820         struct cfq_data *cfqd = cfqq->cfqd;
1821         const int sync = rq_is_sync(rq);
1822         unsigned long now;
1823
1824         now = jiffies;
1825
1826         WARN_ON(!cfqd->rq_in_driver);
1827         WARN_ON(!cfqq->dispatched);
1828         cfqd->rq_in_driver--;
1829         cfqq->dispatched--;
1830
1831         if (cfq_cfqq_sync(cfqq))
1832                 cfqd->sync_flight--;
1833
1834         if (!cfq_class_idle(cfqq))
1835                 cfqd->last_end_request = now;
1836
1837         if (sync)
1838                 RQ_CIC(rq)->last_end_request = now;
1839
1840         /*
1841          * If this is the active queue, check if it needs to be expired,
1842          * or if we want to idle in case it has no pending requests.
1843          */
1844         if (cfqd->active_queue == cfqq) {
1845                 if (cfq_cfqq_slice_new(cfqq)) {
1846                         cfq_set_prio_slice(cfqd, cfqq);
1847                         cfq_clear_cfqq_slice_new(cfqq);
1848                 }
1849                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
1850                         cfq_slice_expired(cfqd, 1);
1851                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1852                         cfq_arm_slice_timer(cfqd);
1853         }
1854
1855         if (!cfqd->rq_in_driver)
1856                 cfq_schedule_dispatch(cfqd);
1857 }
1858
1859 /*
1860  * we temporarily boost lower priority queues if they are holding fs exclusive
1861  * resources. they are boosted to normal prio (CLASS_BE/4)
1862  */
1863 static void cfq_prio_boost(struct cfq_queue *cfqq)
1864 {
1865         if (has_fs_excl()) {
1866                 /*
1867                  * boost idle prio on transactions that would lock out other
1868                  * users of the filesystem
1869                  */
1870                 if (cfq_class_idle(cfqq))
1871                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1872                 if (cfqq->ioprio > IOPRIO_NORM)
1873                         cfqq->ioprio = IOPRIO_NORM;
1874         } else {
1875                 /*
1876                  * check if we need to unboost the queue
1877                  */
1878                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1879                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1880                 if (cfqq->ioprio != cfqq->org_ioprio)
1881                         cfqq->ioprio = cfqq->org_ioprio;
1882         }
1883 }
1884
1885 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1886 {
1887         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1888             !cfq_cfqq_must_alloc_slice(cfqq)) {
1889                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1890                 return ELV_MQUEUE_MUST;
1891         }
1892
1893         return ELV_MQUEUE_MAY;
1894 }
1895
1896 static int cfq_may_queue(struct request_queue *q, int rw)
1897 {
1898         struct cfq_data *cfqd = q->elevator->elevator_data;
1899         struct task_struct *tsk = current;
1900         struct cfq_io_context *cic;
1901         struct cfq_queue *cfqq;
1902
1903         /*
1904          * don't force setup of a queue from here, as a call to may_queue
1905          * does not necessarily imply that a request actually will be queued.
1906          * so just lookup a possibly existing queue, or return 'may queue'
1907          * if that fails
1908          */
1909         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1910         if (!cic)
1911                 return ELV_MQUEUE_MAY;
1912
1913         cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
1914         if (cfqq) {
1915                 cfq_init_prio_data(cfqq, cic->ioc);
1916                 cfq_prio_boost(cfqq);
1917
1918                 return __cfq_may_queue(cfqq);
1919         }
1920
1921         return ELV_MQUEUE_MAY;
1922 }
1923
1924 /*
1925  * queue lock held here
1926  */
1927 static void cfq_put_request(struct request *rq)
1928 {
1929         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1930
1931         if (cfqq) {
1932                 const int rw = rq_data_dir(rq);
1933
1934                 BUG_ON(!cfqq->allocated[rw]);
1935                 cfqq->allocated[rw]--;
1936
1937                 put_io_context(RQ_CIC(rq)->ioc);
1938
1939                 rq->elevator_private = NULL;
1940                 rq->elevator_private2 = NULL;
1941
1942                 cfq_put_queue(cfqq);
1943         }
1944 }
1945
1946 /*
1947  * Allocate cfq data structures associated with this request.
1948  */
1949 static int
1950 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
1951 {
1952         struct cfq_data *cfqd = q->elevator->elevator_data;
1953         struct cfq_io_context *cic;
1954         const int rw = rq_data_dir(rq);
1955         const int is_sync = rq_is_sync(rq);
1956         struct cfq_queue *cfqq;
1957         unsigned long flags;
1958
1959         might_sleep_if(gfp_mask & __GFP_WAIT);
1960
1961         cic = cfq_get_io_context(cfqd, gfp_mask);
1962
1963         spin_lock_irqsave(q->queue_lock, flags);
1964
1965         if (!cic)
1966                 goto queue_fail;
1967
1968         cfqq = cic_to_cfqq(cic, is_sync);
1969         if (!cfqq) {
1970                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
1971
1972                 if (!cfqq)
1973                         goto queue_fail;
1974
1975                 cic_set_cfqq(cic, cfqq, is_sync);
1976         }
1977
1978         cfqq->allocated[rw]++;
1979         cfq_clear_cfqq_must_alloc(cfqq);
1980         atomic_inc(&cfqq->ref);
1981
1982         spin_unlock_irqrestore(q->queue_lock, flags);
1983
1984         rq->elevator_private = cic;
1985         rq->elevator_private2 = cfqq;
1986         return 0;
1987
1988 queue_fail:
1989         if (cic)
1990                 put_io_context(cic->ioc);
1991
1992         cfq_schedule_dispatch(cfqd);
1993         spin_unlock_irqrestore(q->queue_lock, flags);
1994         return 1;
1995 }
1996
1997 static void cfq_kick_queue(struct work_struct *work)
1998 {
1999         struct cfq_data *cfqd =
2000                 container_of(work, struct cfq_data, unplug_work);
2001         struct request_queue *q = cfqd->queue;
2002         unsigned long flags;
2003
2004         spin_lock_irqsave(q->queue_lock, flags);
2005         blk_start_queueing(q);
2006         spin_unlock_irqrestore(q->queue_lock, flags);
2007 }
2008
2009 /*
2010  * Timer running if the active_queue is currently idling inside its time slice
2011  */
2012 static void cfq_idle_slice_timer(unsigned long data)
2013 {
2014         struct cfq_data *cfqd = (struct cfq_data *) data;
2015         struct cfq_queue *cfqq;
2016         unsigned long flags;
2017         int timed_out = 1;
2018
2019         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2020
2021         if ((cfqq = cfqd->active_queue) != NULL) {
2022                 timed_out = 0;
2023
2024                 /*
2025                  * expired
2026                  */
2027                 if (cfq_slice_used(cfqq))
2028                         goto expire;
2029
2030                 /*
2031                  * only expire and reinvoke request handler, if there are
2032                  * other queues with pending requests
2033                  */
2034                 if (!cfqd->busy_queues)
2035                         goto out_cont;
2036
2037                 /*
2038                  * not expired and it has a request pending, let it dispatch
2039                  */
2040                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
2041                         cfq_mark_cfqq_must_dispatch(cfqq);
2042                         goto out_kick;
2043                 }
2044         }
2045 expire:
2046         cfq_slice_expired(cfqd, timed_out);
2047 out_kick:
2048         cfq_schedule_dispatch(cfqd);
2049 out_cont:
2050         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2051 }
2052
2053 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2054 {
2055         del_timer_sync(&cfqd->idle_slice_timer);
2056         kblockd_flush_work(&cfqd->unplug_work);
2057 }
2058
2059 static void cfq_put_async_queues(struct cfq_data *cfqd)
2060 {
2061         int i;
2062
2063         for (i = 0; i < IOPRIO_BE_NR; i++) {
2064                 if (cfqd->async_cfqq[0][i])
2065                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2066                 if (cfqd->async_cfqq[1][i])
2067                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2068         }
2069
2070         if (cfqd->async_idle_cfqq)
2071                 cfq_put_queue(cfqd->async_idle_cfqq);
2072 }
2073
2074 static void cfq_exit_queue(elevator_t *e)
2075 {
2076         struct cfq_data *cfqd = e->elevator_data;
2077         struct request_queue *q = cfqd->queue;
2078
2079         cfq_shutdown_timer_wq(cfqd);
2080
2081         spin_lock_irq(q->queue_lock);
2082
2083         if (cfqd->active_queue)
2084                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2085
2086         while (!list_empty(&cfqd->cic_list)) {
2087                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2088                                                         struct cfq_io_context,
2089                                                         queue_list);
2090
2091                 __cfq_exit_single_io_context(cfqd, cic);
2092         }
2093
2094         cfq_put_async_queues(cfqd);
2095
2096         spin_unlock_irq(q->queue_lock);
2097
2098         cfq_shutdown_timer_wq(cfqd);
2099
2100         kfree(cfqd);
2101 }
2102
2103 static void *cfq_init_queue(struct request_queue *q)
2104 {
2105         struct cfq_data *cfqd;
2106
2107         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2108         if (!cfqd)
2109                 return NULL;
2110
2111         cfqd->service_tree = CFQ_RB_ROOT;
2112         INIT_LIST_HEAD(&cfqd->cic_list);
2113
2114         cfqd->queue = q;
2115
2116         init_timer(&cfqd->idle_slice_timer);
2117         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2118         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2119
2120         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2121
2122         cfqd->last_end_request = jiffies;
2123         cfqd->cfq_quantum = cfq_quantum;
2124         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2125         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2126         cfqd->cfq_back_max = cfq_back_max;
2127         cfqd->cfq_back_penalty = cfq_back_penalty;
2128         cfqd->cfq_slice[0] = cfq_slice_async;
2129         cfqd->cfq_slice[1] = cfq_slice_sync;
2130         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2131         cfqd->cfq_slice_idle = cfq_slice_idle;
2132
2133         return cfqd;
2134 }
2135
2136 static void cfq_slab_kill(void)
2137 {
2138         if (cfq_pool)
2139                 kmem_cache_destroy(cfq_pool);
2140         if (cfq_ioc_pool)
2141                 kmem_cache_destroy(cfq_ioc_pool);
2142 }
2143
2144 static int __init cfq_slab_setup(void)
2145 {
2146         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2147         if (!cfq_pool)
2148                 goto fail;
2149
2150         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, SLAB_DESTROY_BY_RCU);
2151         if (!cfq_ioc_pool)
2152                 goto fail;
2153
2154         return 0;
2155 fail:
2156         cfq_slab_kill();
2157         return -ENOMEM;
2158 }
2159
2160 /*
2161  * sysfs parts below -->
2162  */
2163 static ssize_t
2164 cfq_var_show(unsigned int var, char *page)
2165 {
2166         return sprintf(page, "%d\n", var);
2167 }
2168
2169 static ssize_t
2170 cfq_var_store(unsigned int *var, const char *page, size_t count)
2171 {
2172         char *p = (char *) page;
2173
2174         *var = simple_strtoul(p, &p, 10);
2175         return count;
2176 }
2177
2178 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2179 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2180 {                                                                       \
2181         struct cfq_data *cfqd = e->elevator_data;                       \
2182         unsigned int __data = __VAR;                                    \
2183         if (__CONV)                                                     \
2184                 __data = jiffies_to_msecs(__data);                      \
2185         return cfq_var_show(__data, (page));                            \
2186 }
2187 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2188 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2189 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2190 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2191 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2192 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2193 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2194 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2195 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2196 #undef SHOW_FUNCTION
2197
2198 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2199 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2200 {                                                                       \
2201         struct cfq_data *cfqd = e->elevator_data;                       \
2202         unsigned int __data;                                            \
2203         int ret = cfq_var_store(&__data, (page), count);                \
2204         if (__data < (MIN))                                             \
2205                 __data = (MIN);                                         \
2206         else if (__data > (MAX))                                        \
2207                 __data = (MAX);                                         \
2208         if (__CONV)                                                     \
2209                 *(__PTR) = msecs_to_jiffies(__data);                    \
2210         else                                                            \
2211                 *(__PTR) = __data;                                      \
2212         return ret;                                                     \
2213 }
2214 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2215 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2216 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2217 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2218 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2219 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2220 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2221 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2222 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2223 #undef STORE_FUNCTION
2224
2225 #define CFQ_ATTR(name) \
2226         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2227
2228 static struct elv_fs_entry cfq_attrs[] = {
2229         CFQ_ATTR(quantum),
2230         CFQ_ATTR(fifo_expire_sync),
2231         CFQ_ATTR(fifo_expire_async),
2232         CFQ_ATTR(back_seek_max),
2233         CFQ_ATTR(back_seek_penalty),
2234         CFQ_ATTR(slice_sync),
2235         CFQ_ATTR(slice_async),
2236         CFQ_ATTR(slice_async_rq),
2237         CFQ_ATTR(slice_idle),
2238         __ATTR_NULL
2239 };
2240
2241 static struct elevator_type iosched_cfq = {
2242         .ops = {
2243                 .elevator_merge_fn =            cfq_merge,
2244                 .elevator_merged_fn =           cfq_merged_request,
2245                 .elevator_merge_req_fn =        cfq_merged_requests,
2246                 .elevator_allow_merge_fn =      cfq_allow_merge,
2247                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2248                 .elevator_add_req_fn =          cfq_insert_request,
2249                 .elevator_activate_req_fn =     cfq_activate_request,
2250                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2251                 .elevator_queue_empty_fn =      cfq_queue_empty,
2252                 .elevator_completed_req_fn =    cfq_completed_request,
2253                 .elevator_former_req_fn =       elv_rb_former_request,
2254                 .elevator_latter_req_fn =       elv_rb_latter_request,
2255                 .elevator_set_req_fn =          cfq_set_request,
2256                 .elevator_put_req_fn =          cfq_put_request,
2257                 .elevator_may_queue_fn =        cfq_may_queue,
2258                 .elevator_init_fn =             cfq_init_queue,
2259                 .elevator_exit_fn =             cfq_exit_queue,
2260                 .trim =                         cfq_free_io_context,
2261         },
2262         .elevator_attrs =       cfq_attrs,
2263         .elevator_name =        "cfq",
2264         .elevator_owner =       THIS_MODULE,
2265 };
2266
2267 static int __init cfq_init(void)
2268 {
2269         /*
2270          * could be 0 on HZ < 1000 setups
2271          */
2272         if (!cfq_slice_async)
2273                 cfq_slice_async = 1;
2274         if (!cfq_slice_idle)
2275                 cfq_slice_idle = 1;
2276
2277         if (cfq_slab_setup())
2278                 return -ENOMEM;
2279
2280         elv_register(&iosched_cfq);
2281
2282         return 0;
2283 }
2284
2285 static void __exit cfq_exit(void)
2286 {
2287         DECLARE_COMPLETION_ONSTACK(all_gone);
2288         elv_unregister(&iosched_cfq);
2289         ioc_gone = &all_gone;
2290         /* ioc_gone's update must be visible before reading ioc_count */
2291         smp_wmb();
2292         if (elv_ioc_count_read(ioc_count))
2293                 wait_for_completion(ioc_gone);
2294         synchronize_rcu();
2295         cfq_slab_kill();
2296 }
2297
2298 module_init(cfq_init);
2299 module_exit(cfq_exit);
2300
2301 MODULE_AUTHOR("Jens Axboe");
2302 MODULE_LICENSE("GPL");
2303 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");