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