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