err.no Git - linux-2.6/blob - drivers/block/cfq-iosched.c

   1 /*
   2  *  linux/drivers/block/cfq-iosched.c
   3  *
   4  *  CFQ, or complete fairness queueing, disk scheduler.
   5  *
   6  *  Based on ideas from a previously unfinished io
   7  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   8  *
   9  *  Copyright (C) 2003 Jens Axboe <axboe@suse.de>
  10  */
  11 #include <linux/kernel.h>
  12 #include <linux/fs.h>
  13 #include <linux/blkdev.h>
  14 #include <linux/elevator.h>
  15 #include <linux/bio.h>
  16 #include <linux/config.h>
  17 #include <linux/module.h>
  18 #include <linux/slab.h>
  19 #include <linux/init.h>
  20 #include <linux/compiler.h>
  21 #include <linux/hash.h>
  22 #include <linux/rbtree.h>
  23 #include <linux/mempool.h>
  24 #include <linux/ioprio.h>
  25 #include <linux/writeback.h>
  26
  27 /*
  28  * tunables
  29  */
  30 static int cfq_quantum = 4;             /* max queue in one round of service */
  31 static int cfq_queued = 8;              /* minimum rq allocate limit per-queue*/
  32 static int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  33 static int cfq_back_max = 16 * 1024;    /* maximum backwards seek, in KiB */
  34 static int cfq_back_penalty = 2;        /* penalty of a backwards seek */
  35
  36 static int cfq_slice_sync = HZ / 10;
  37 static int cfq_slice_async = HZ / 50;
  38 static int cfq_slice_async_rq = 2;
  39 static int cfq_slice_idle = HZ / 50;
  40
  41 #define CFQ_IDLE_GRACE          (HZ / 10)
  42 #define CFQ_SLICE_SCALE         (5)
  43
  44 #define CFQ_KEY_ASYNC           (0)
  45
  46 /*
  47  * disable queueing at the driver/hardware level
  48  */
  49 static int cfq_max_depth = 1;
  50
  51 /*
  52  * for the hash of cfqq inside the cfqd
  53  */
  54 #define CFQ_QHASH_SHIFT         6
  55 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
  56 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
  57
  58 /*
  59  * for the hash of crq inside the cfqq
  60  */
  61 #define CFQ_MHASH_SHIFT         6
  62 #define CFQ_MHASH_BLOCK(sec)    ((sec) >> 3)
  63 #define CFQ_MHASH_ENTRIES       (1 << CFQ_MHASH_SHIFT)
  64 #define CFQ_MHASH_FN(sec)       hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
  65 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
  66 #define list_entry_hash(ptr)    hlist_entry((ptr), struct cfq_rq, hash)
  67
  68 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
  69 #define list_entry_fifo(ptr)    list_entry((ptr), struct request, queuelist)
  70
  71 #define RQ_DATA(rq)             (rq)->elevator_private
  72
  73 /*
  74  * rb-tree defines
  75  */
  76 #define RB_NONE                 (2)
  77 #define RB_EMPTY(node)          ((node)->rb_node == NULL)
  78 #define RB_CLEAR_COLOR(node)    (node)->rb_color = RB_NONE
  79 #define RB_CLEAR(node)          do {    \
  80         (node)->rb_parent = NULL;       \
  81         RB_CLEAR_COLOR((node));         \
  82         (node)->rb_right = NULL;        \
  83         (node)->rb_left = NULL;         \
  84 } while (0)
  85 #define RB_CLEAR_ROOT(root)     ((root)->rb_node = NULL)
  86 #define ON_RB(node)             ((node)->rb_color != RB_NONE)
  87 #define rb_entry_crq(node)      rb_entry((node), struct cfq_rq, rb_node)
  88 #define rq_rb_key(rq)           (rq)->sector
  89
  90 static kmem_cache_t *crq_pool;
  91 static kmem_cache_t *cfq_pool;
  92 static kmem_cache_t *cfq_ioc_pool;
  93
  94 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  95 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  96 #define cfq_class_be(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
  97 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  98
  99 #define cfq_cfqq_sync(cfqq)     ((cfqq)->key != CFQ_KEY_ASYNC)
 100
 101 /*
 102  * Per block device queue structure
 103  */
 104 struct cfq_data {
 105         atomic_t ref;
 106         request_queue_t *queue;
 107
 108         /*
 109          * rr list of queues with requests and the count of them
 110          */
 111         struct list_head rr_list[CFQ_PRIO_LISTS];
 112         struct list_head busy_rr;
 113         struct list_head cur_rr;
 114         struct list_head idle_rr;
 115         unsigned int busy_queues;
 116
 117         /*
 118          * non-ordered list of empty cfqq's
 119          */
 120         struct list_head empty_list;
 121
 122         /*
 123          * cfqq lookup hash
 124          */
 125         struct hlist_head *cfq_hash;
 126
 127         /*
 128          * global crq hash for all queues
 129          */
 130         struct hlist_head *crq_hash;
 131
 132         unsigned int max_queued;
 133
 134         mempool_t *crq_pool;
 135
 136         int rq_in_driver;
 137
 138         /*
 139          * schedule slice state info
 140          */
 141         /*
 142          * idle window management
 143          */
 144         struct timer_list idle_slice_timer;
 145         struct work_struct unplug_work;
 146
 147         struct cfq_queue *active_queue;
 148         struct cfq_io_context *active_cic;
 149         int cur_prio, cur_end_prio;
 150         unsigned int dispatch_slice;
 151
 152         struct timer_list idle_class_timer;
 153
 154         sector_t last_sector;
 155         unsigned long last_end_request;
 156
 157         unsigned int rq_starved;
 158
 159         /*
 160          * tunables, see top of file
 161          */
 162         unsigned int cfq_quantum;
 163         unsigned int cfq_queued;
 164         unsigned int cfq_fifo_expire[2];
 165         unsigned int cfq_back_penalty;
 166         unsigned int cfq_back_max;
 167         unsigned int cfq_slice[2];
 168         unsigned int cfq_slice_async_rq;
 169         unsigned int cfq_slice_idle;
 170         unsigned int cfq_max_depth;
 171 };
 172
 173 /*
 174  * Per process-grouping structure
 175  */
 176 struct cfq_queue {
 177         /* reference count */
 178         atomic_t ref;
 179         /* parent cfq_data */
 180         struct cfq_data *cfqd;
 181         /* cfqq lookup hash */
 182         struct hlist_node cfq_hash;
 183         /* hash key */
 184         unsigned int key;
 185         /* on either rr or empty list of cfqd */
 186         struct list_head cfq_list;
 187         /* sorted list of pending requests */
 188         struct rb_root sort_list;
 189         /* if fifo isn't expired, next request to serve */
 190         struct cfq_rq *next_crq;
 191         /* requests queued in sort_list */
 192         int queued[2];
 193         /* currently allocated requests */
 194         int allocated[2];
 195         /* fifo list of requests in sort_list */
 196         struct list_head fifo;
 197
 198         unsigned long slice_start;
 199         unsigned long slice_end;
 200         unsigned long slice_left;
 201         unsigned long service_last;
 202
 203         /* number of requests that have been handed to the driver */
 204         int in_flight;
 205
 206         /* io prio of this group */
 207         unsigned short ioprio, org_ioprio;
 208         unsigned short ioprio_class, org_ioprio_class;
 209
 210         /* whether queue is on rr (or empty) list */
 211         unsigned on_rr : 1;
 212         /* idle slice, waiting for new request submission */
 213         unsigned wait_request : 1;
 214         /* set when wait_request gets set, reset on first rq alloc */
 215         unsigned must_alloc : 1;
 216         /* only gets one must_alloc per slice */
 217         unsigned must_alloc_slice : 1;
 218         /* idle slice, request added, now waiting to dispatch it */
 219         unsigned must_dispatch : 1;
 220         /* fifo expire per-slice */
 221         unsigned fifo_expire : 1;
 222
 223         unsigned idle_window : 1;
 224         unsigned prio_changed : 1;
 225 };
 226
 227 struct cfq_rq {
 228         struct rb_node rb_node;
 229         sector_t rb_key;
 230         struct request *request;
 231         struct hlist_node hash;
 232
 233         struct cfq_queue *cfq_queue;
 234         struct cfq_io_context *io_context;
 235
 236         unsigned in_flight : 1;
 237         unsigned accounted : 1;
 238         unsigned is_sync   : 1;
 239         unsigned requeued  : 1;
 240 };
 241
 242 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int);
 243 static void cfq_dispatch_sort(request_queue_t *, struct cfq_rq *);
 244 static void cfq_put_cfqd(struct cfq_data *cfqd);
 245
 246 #define process_sync(tsk)       ((tsk)->flags & PF_SYNCWRITE)
 247
 248 /*
 249  * lots of deadline iosched dupes, can be abstracted later...
 250  */
 251 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
 252 {
 253         hlist_del_init(&crq->hash);
 254 }
 255
 256 static void cfq_remove_merge_hints(request_queue_t *q, struct cfq_rq *crq)
 257 {
 258         cfq_del_crq_hash(crq);
 259
 260         if (q->last_merge == crq->request)
 261                 q->last_merge = NULL;
 262 }
 263
 264 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
 265 {
 266         const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
 267
 268         hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
 269 }
 270
 271 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
 272 {
 273         struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
 274         struct hlist_node *entry, *next;
 275
 276         hlist_for_each_safe(entry, next, hash_list) {
 277                 struct cfq_rq *crq = list_entry_hash(entry);
 278                 struct request *__rq = crq->request;
 279
 280                 if (!rq_mergeable(__rq)) {
 281                         cfq_del_crq_hash(crq);
 282                         continue;
 283                 }
 284
 285                 if (rq_hash_key(__rq) == offset)
 286                         return __rq;
 287         }
 288
 289         return NULL;
 290 }
 291
 292 /*
 293  * Lifted from AS - choose which of crq1 and crq2 that is best served now.
 294  * We choose the request that is closest to the head right now. Distance
 295  * behind the head are penalized and only allowed to a certain extent.
 296  */
 297 static struct cfq_rq *
 298 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
 299 {
 300         sector_t last, s1, s2, d1 = 0, d2 = 0;
 301         int r1_wrap = 0, r2_wrap = 0;   /* requests are behind the disk head */
 302         unsigned long back_max;
 303
 304         if (crq1 == NULL || crq1 == crq2)
 305                 return crq2;
 306         if (crq2 == NULL)
 307                 return crq1;
 308         if (crq1->requeued)
 309                 return crq1;
 310         if (crq2->requeued)
 311                 return crq2;
 312
 313         s1 = crq1->request->sector;
 314         s2 = crq2->request->sector;
 315
 316         last = cfqd->last_sector;
 317
 318         /*
 319          * by definition, 1KiB is 2 sectors
 320          */
 321         back_max = cfqd->cfq_back_max * 2;
 322
 323         /*
 324          * Strict one way elevator _except_ in the case where we allow
 325          * short backward seeks which are biased as twice the cost of a
 326          * similar forward seek.
 327          */
 328         if (s1 >= last)
 329                 d1 = s1 - last;
 330         else if (s1 + back_max >= last)
 331                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 332         else
 333                 r1_wrap = 1;
 334
 335         if (s2 >= last)
 336                 d2 = s2 - last;
 337         else if (s2 + back_max >= last)
 338                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 339         else
 340                 r2_wrap = 1;
 341
 342         /* Found required data */
 343         if (!r1_wrap && r2_wrap)
 344                 return crq1;
 345         else if (!r2_wrap && r1_wrap)
 346                 return crq2;
 347         else if (r1_wrap && r2_wrap) {
 348                 /* both behind the head */
 349                 if (s1 <= s2)
 350                         return crq1;
 351                 else
 352                         return crq2;
 353         }
 354
 355         /* Both requests in front of the head */
 356         if (d1 < d2)
 357                 return crq1;
 358         else if (d2 < d1)
 359                 return crq2;
 360         else {
 361                 if (s1 >= s2)
 362                         return crq1;
 363                 else
 364                         return crq2;
 365         }
 366 }
 367
 368 /*
 369  * would be nice to take fifo expire time into account as well
 370  */
 371 static struct cfq_rq *
 372 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 373                   struct cfq_rq *last)
 374 {
 375         struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
 376         struct rb_node *rbnext, *rbprev;
 377
 378         rbnext = NULL;
 379         if (ON_RB(&last->rb_node))
 380                 rbnext = rb_next(&last->rb_node);
 381         if (!rbnext) {
 382                 rbnext = rb_first(&cfqq->sort_list);
 383                 if (rbnext == &last->rb_node)
 384                         rbnext = NULL;
 385         }
 386
 387         rbprev = rb_prev(&last->rb_node);
 388
 389         if (rbprev)
 390                 crq_prev = rb_entry_crq(rbprev);
 391         if (rbnext)
 392                 crq_next = rb_entry_crq(rbnext);
 393
 394         return cfq_choose_req(cfqd, crq_next, crq_prev);
 395 }
 396
 397 static void cfq_update_next_crq(struct cfq_rq *crq)
 398 {
 399         struct cfq_queue *cfqq = crq->cfq_queue;
 400
 401         if (cfqq->next_crq == crq)
 402                 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
 403 }
 404
 405 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 406 {
 407         struct cfq_data *cfqd = cfqq->cfqd;
 408         struct list_head *list, *entry;
 409
 410         BUG_ON(!cfqq->on_rr);
 411
 412         list_del(&cfqq->cfq_list);
 413
 414         if (cfq_class_rt(cfqq))
 415                 list = &cfqd->cur_rr;
 416         else if (cfq_class_idle(cfqq))
 417                 list = &cfqd->idle_rr;
 418         else {
 419                 /*
 420                  * if cfqq has requests in flight, don't allow it to be
 421                  * found in cfq_set_active_queue before it has finished them.
 422                  * this is done to increase fairness between a process that
 423                  * has lots of io pending vs one that only generates one
 424                  * sporadically or synchronously
 425                  */
 426                 if (cfqq->in_flight)
 427                         list = &cfqd->busy_rr;
 428                 else
 429                         list = &cfqd->rr_list[cfqq->ioprio];
 430         }
 431
 432         /*
 433          * if queue was preempted, just add to front to be fair. busy_rr
 434          * isn't sorted.
 435          */
 436         if (preempted || list == &cfqd->busy_rr) {
 437                 list_add(&cfqq->cfq_list, list);
 438                 return;
 439         }
 440
 441         /*
 442          * sort by when queue was last serviced
 443          */
 444         entry = list;
 445         while ((entry = entry->prev) != list) {
 446                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
 447
 448                 if (!__cfqq->service_last)
 449                         break;
 450                 if (time_before(__cfqq->service_last, cfqq->service_last))
 451                         break;
 452         }
 453
 454         list_add(&cfqq->cfq_list, entry);
 455 }
 456
 457 /*
 458  * add to busy list of queues for service, trying to be fair in ordering
 459  * the pending list according to last request service
 460  */
 461 static inline void
 462 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq, int requeue)
 463 {
 464         BUG_ON(cfqq->on_rr);
 465         cfqq->on_rr = 1;
 466         cfqd->busy_queues++;
 467
 468         cfq_resort_rr_list(cfqq, requeue);
 469 }
 470
 471 static inline void
 472 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 473 {
 474         BUG_ON(!cfqq->on_rr);
 475         cfqq->on_rr = 0;
 476         list_move(&cfqq->cfq_list, &cfqd->empty_list);
 477
 478         BUG_ON(!cfqd->busy_queues);
 479         cfqd->busy_queues--;
 480 }
 481
 482 /*
 483  * rb tree support functions
 484  */
 485 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
 486 {
 487         struct cfq_queue *cfqq = crq->cfq_queue;
 488
 489         if (ON_RB(&crq->rb_node)) {
 490                 struct cfq_data *cfqd = cfqq->cfqd;
 491                 const int sync = crq->is_sync;
 492
 493                 BUG_ON(!cfqq->queued[sync]);
 494                 cfqq->queued[sync]--;
 495
 496                 cfq_update_next_crq(crq);
 497
 498                 rb_erase(&crq->rb_node, &cfqq->sort_list);
 499                 RB_CLEAR_COLOR(&crq->rb_node);
 500
 501                 if (cfqq->on_rr && RB_EMPTY(&cfqq->sort_list))
 502                         cfq_del_cfqq_rr(cfqd, cfqq);
 503         }
 504 }
 505
 506 static struct cfq_rq *
 507 __cfq_add_crq_rb(struct cfq_rq *crq)
 508 {
 509         struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
 510         struct rb_node *parent = NULL;
 511         struct cfq_rq *__crq;
 512
 513         while (*p) {
 514                 parent = *p;
 515                 __crq = rb_entry_crq(parent);
 516
 517                 if (crq->rb_key < __crq->rb_key)
 518                         p = &(*p)->rb_left;
 519                 else if (crq->rb_key > __crq->rb_key)
 520                         p = &(*p)->rb_right;
 521                 else
 522                         return __crq;
 523         }
 524
 525         rb_link_node(&crq->rb_node, parent, p);
 526         return NULL;
 527 }
 528
 529 static void cfq_add_crq_rb(struct cfq_rq *crq)
 530 {
 531         struct cfq_queue *cfqq = crq->cfq_queue;
 532         struct cfq_data *cfqd = cfqq->cfqd;
 533         struct request *rq = crq->request;
 534         struct cfq_rq *__alias;
 535
 536         crq->rb_key = rq_rb_key(rq);
 537         cfqq->queued[crq->is_sync]++;
 538
 539         /*
 540          * looks a little odd, but the first insert might return an alias.
 541          * if that happens, put the alias on the dispatch list
 542          */
 543         while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
 544                 cfq_dispatch_sort(cfqd->queue, __alias);
 545
 546         rb_insert_color(&crq->rb_node, &cfqq->sort_list);
 547
 548         if (!cfqq->on_rr)
 549                 cfq_add_cfqq_rr(cfqd, cfqq, crq->requeued);
 550
 551         /*
 552          * check if this request is a better next-serve candidate
 553          */
 554         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
 555 }
 556
 557 static inline void
 558 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
 559 {
 560         if (ON_RB(&crq->rb_node)) {
 561                 rb_erase(&crq->rb_node, &cfqq->sort_list);
 562                 cfqq->queued[crq->is_sync]--;
 563         }
 564
 565         cfq_add_crq_rb(crq);
 566 }
 567
 568 static struct request *cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector)
 569
 570 {
 571         struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, current->pid);
 572         struct rb_node *n;
 573
 574         if (!cfqq)
 575                 goto out;
 576
 577         n = cfqq->sort_list.rb_node;
 578         while (n) {
 579                 struct cfq_rq *crq = rb_entry_crq(n);
 580
 581                 if (sector < crq->rb_key)
 582                         n = n->rb_left;
 583                 else if (sector > crq->rb_key)
 584                         n = n->rb_right;
 585                 else
 586                         return crq->request;
 587         }
 588
 589 out:
 590         return NULL;
 591 }
 592
 593 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 594 {
 595         struct cfq_data *cfqd = q->elevator->elevator_data;
 596         struct cfq_rq *crq = RQ_DATA(rq);
 597
 598         if (crq) {
 599                 struct cfq_queue *cfqq = crq->cfq_queue;
 600
 601                 if (crq->accounted) {
 602                         crq->accounted = 0;
 603                         WARN_ON(!cfqd->rq_in_driver);
 604                         cfqd->rq_in_driver--;
 605                 }
 606                 if (crq->in_flight) {
 607                         crq->in_flight = 0;
 608                         WARN_ON(!cfqq->in_flight);
 609                         cfqq->in_flight--;
 610                 }
 611                 crq->requeued = 1;
 612         }
 613 }
 614
 615 /*
 616  * make sure the service time gets corrected on reissue of this request
 617  */
 618 static void cfq_requeue_request(request_queue_t *q, struct request *rq)
 619 {
 620         cfq_deactivate_request(q, rq);
 621         list_add(&rq->queuelist, &q->queue_head);
 622 }
 623
 624 static void cfq_remove_request(request_queue_t *q, struct request *rq)
 625 {
 626         struct cfq_rq *crq = RQ_DATA(rq);
 627
 628         if (crq) {
 629                 list_del_init(&rq->queuelist);
 630                 cfq_del_crq_rb(crq);
 631                 cfq_remove_merge_hints(q, crq);
 632
 633         }
 634 }
 635
 636 static int
 637 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 638 {
 639         struct cfq_data *cfqd = q->elevator->elevator_data;
 640         struct request *__rq;
 641         int ret;
 642
 643         ret = elv_try_last_merge(q, bio);
 644         if (ret != ELEVATOR_NO_MERGE) {
 645                 __rq = q->last_merge;
 646                 goto out_insert;
 647         }
 648
 649         __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
 650         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 651                 ret = ELEVATOR_BACK_MERGE;
 652                 goto out;
 653         }
 654
 655         __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio));
 656         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 657                 ret = ELEVATOR_FRONT_MERGE;
 658                 goto out;
 659         }
 660
 661         return ELEVATOR_NO_MERGE;
 662 out:
 663         q->last_merge = __rq;
 664 out_insert:
 665         *req = __rq;
 666         return ret;
 667 }
 668
 669 static void cfq_merged_request(request_queue_t *q, struct request *req)
 670 {
 671         struct cfq_data *cfqd = q->elevator->elevator_data;
 672         struct cfq_rq *crq = RQ_DATA(req);
 673
 674         cfq_del_crq_hash(crq);
 675         cfq_add_crq_hash(cfqd, crq);
 676
 677         if (ON_RB(&crq->rb_node) && (rq_rb_key(req) != crq->rb_key)) {
 678                 struct cfq_queue *cfqq = crq->cfq_queue;
 679
 680                 cfq_update_next_crq(crq);
 681                 cfq_reposition_crq_rb(cfqq, crq);
 682         }
 683
 684         q->last_merge = req;
 685 }
 686
 687 static void
 688 cfq_merged_requests(request_queue_t *q, struct request *rq,
 689                     struct request *next)
 690 {
 691         cfq_merged_request(q, rq);
 692
 693         /*
 694          * reposition in fifo if next is older than rq
 695          */
 696         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 697             time_before(next->start_time, rq->start_time))
 698                 list_move(&rq->queuelist, &next->queuelist);
 699
 700         cfq_remove_request(q, next);
 701 }
 702
 703 static inline void
 704 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 705 {
 706         if (cfqq) {
 707                 /*
 708                  * stop potential idle class queues waiting service
 709                  */
 710                 del_timer(&cfqd->idle_class_timer);
 711
 712                 cfqq->slice_start = jiffies;
 713                 cfqq->slice_end = 0;
 714                 cfqq->slice_left = 0;
 715                 cfqq->must_alloc_slice = 0;
 716                 cfqq->fifo_expire = 0;
 717         }
 718
 719         cfqd->active_queue = cfqq;
 720 }
 721
 722 /*
 723  * 0
 724  * 0,1
 725  * 0,1,2
 726  * 0,1,2,3
 727  * 0,1,2,3,4
 728  * 0,1,2,3,4,5
 729  * 0,1,2,3,4,5,6
 730  * 0,1,2,3,4,5,6,7
 731  */
 732 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 733 {
 734         int prio, wrap;
 735
 736         prio = -1;
 737         wrap = 0;
 738         do {
 739                 int p;
 740
 741                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 742                         if (!list_empty(&cfqd->rr_list[p])) {
 743                                 prio = p;
 744                                 break;
 745                         }
 746                 }
 747
 748                 if (prio != -1)
 749                         break;
 750                 cfqd->cur_prio = 0;
 751                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 752                         cfqd->cur_end_prio = 0;
 753                         if (wrap)
 754                                 break;
 755                         wrap = 1;
 756                 }
 757         } while (1);
 758
 759         if (unlikely(prio == -1))
 760                 return -1;
 761
 762         BUG_ON(prio >= CFQ_PRIO_LISTS);
 763
 764         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 765
 766         cfqd->cur_prio = prio + 1;
 767         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 768                 cfqd->cur_end_prio = cfqd->cur_prio;
 769                 cfqd->cur_prio = 0;
 770         }
 771         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 772                 cfqd->cur_prio = 0;
 773                 cfqd->cur_end_prio = 0;
 774         }
 775
 776         return prio;
 777 }
 778
 779 static void cfq_set_active_queue(struct cfq_data *cfqd)
 780 {
 781         struct cfq_queue *cfqq = NULL;
 782
 783         /*
 784          * if current list is non-empty, grab first entry. if it is empty,
 785          * get next prio level and grab first entry then if any are spliced
 786          */
 787         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
 788                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 789
 790         /*
 791          * if we have idle queues and no rt or be queues had pending
 792          * requests, either allow immediate service if the grace period
 793          * has passed or arm the idle grace timer
 794          */
 795         if (!cfqq && !list_empty(&cfqd->idle_rr)) {
 796                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 797
 798                 if (time_after_eq(jiffies, end))
 799                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 800                 else
 801                         mod_timer(&cfqd->idle_class_timer, end);
 802         }
 803
 804         __cfq_set_active_queue(cfqd, cfqq);
 805 }
 806
 807 /*
 808  * current cfqq expired its slice (or was too idle), select new one
 809  */
 810 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 811 {
 812         struct cfq_queue *cfqq = cfqd->active_queue;
 813
 814         if (cfqq) {
 815                 unsigned long now = jiffies;
 816
 817                 if (cfqq->wait_request)
 818                         del_timer(&cfqd->idle_slice_timer);
 819
 820                 if (!preempted && !cfqq->in_flight)
 821                         cfqq->service_last = now;
 822
 823                 cfqq->must_dispatch = 0;
 824                 cfqq->wait_request = 0;
 825
 826                 /*
 827                  * store what was left of this slice, if the queue idled out
 828                  * or was preempted
 829                  */
 830                 if (time_after(now, cfqq->slice_end))
 831                         cfqq->slice_left = now - cfqq->slice_end;
 832                 else
 833                         cfqq->slice_left = 0;
 834
 835                 if (cfqq->on_rr)
 836                         cfq_resort_rr_list(cfqq, preempted);
 837
 838                 cfqd->active_queue = NULL;
 839
 840                 if (cfqd->active_cic) {
 841                         put_io_context(cfqd->active_cic->ioc);
 842                         cfqd->active_cic = NULL;
 843                 }
 844         }
 845
 846         cfqd->dispatch_slice = 0;
 847 }
 848
 849 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 850
 851 {
 852         WARN_ON(!RB_EMPTY(&cfqq->sort_list));
 853         WARN_ON(cfqq != cfqd->active_queue);
 854
 855         /*
 856          * idle is disabled, either manually or by past process history
 857          */
 858         if (!cfqd->cfq_slice_idle)
 859                 return 0;
 860         if (!cfqq->idle_window)
 861                 return 0;
 862         /*
 863          * task has exited, don't wait
 864          */
 865         if (cfqd->active_cic && !cfqd->active_cic->ioc->task)
 866                 return 0;
 867
 868         cfqq->wait_request = 1;
 869         cfqq->must_alloc = 1;
 870
 871         if (!timer_pending(&cfqd->idle_slice_timer)) {
 872                 unsigned long slice_left = cfqq->slice_end - 1;
 873
 874                 cfqd->idle_slice_timer.expires = min(jiffies + cfqd->cfq_slice_idle, slice_left);
 875                 add_timer(&cfqd->idle_slice_timer);
 876         }
 877
 878         return 1;
 879 }
 880
 881 /*
 882  * we dispatch cfqd->cfq_quantum requests in total from the rr_list queues,
 883  * this function sector sorts the selected request to minimize seeks. we start
 884  * at cfqd->last_sector, not 0.
 885  */
 886 static void cfq_dispatch_sort(request_queue_t *q, struct cfq_rq *crq)
 887 {
 888         struct cfq_data *cfqd = q->elevator->elevator_data;
 889         struct cfq_queue *cfqq = crq->cfq_queue;
 890         struct list_head *head = &q->queue_head, *entry = head;
 891         struct request *__rq;
 892         sector_t last;
 893
 894         list_del(&crq->request->queuelist);
 895
 896         last = cfqd->last_sector;
 897         list_for_each_entry_reverse(__rq, head, queuelist) {
 898                 struct cfq_rq *__crq = RQ_DATA(__rq);
 899
 900                 if (blk_barrier_rq(__rq))
 901                         break;
 902                 if (!blk_fs_request(__rq))
 903                         break;
 904                 if (__crq->requeued)
 905                         break;
 906
 907                 if (__rq->sector <= crq->request->sector)
 908                         break;
 909                 if (__rq->sector > last && crq->request->sector < last) {
 910                         last = crq->request->sector + crq->request->nr_sectors;
 911                         break;
 912                 }
 913                 entry = &__rq->queuelist;
 914         }
 915
 916         cfqd->last_sector = last;
 917
 918         cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
 919
 920         cfq_del_crq_rb(crq);
 921         cfq_remove_merge_hints(q, crq);
 922
 923         crq->in_flight = 1;
 924         crq->requeued = 0;
 925         cfqq->in_flight++;
 926         list_add_tail(&crq->request->queuelist, entry);
 927 }
 928
 929 /*
 930  * return expired entry, or NULL to just start from scratch in rbtree
 931  */
 932 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
 933 {
 934         struct cfq_data *cfqd = cfqq->cfqd;
 935         struct request *rq;
 936         struct cfq_rq *crq;
 937
 938         if (cfqq->fifo_expire)
 939                 return NULL;
 940
 941         if (!list_empty(&cfqq->fifo)) {
 942                 int fifo = cfq_cfqq_sync(cfqq);
 943
 944                 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
 945                 rq = crq->request;
 946                 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
 947                         cfqq->fifo_expire = 1;
 948                         return crq;
 949                 }
 950         }
 951
 952         return NULL;
 953 }
 954
 955 /*
 956  * Scale schedule slice based on io priority
 957  */
 958 static inline int
 959 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 960 {
 961         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 962
 963         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 964
 965         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
 966 }
 967
 968 static inline void
 969 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 970 {
 971         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 972 }
 973
 974 static inline int
 975 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 976 {
 977         const int base_rq = cfqd->cfq_slice_async_rq;
 978
 979         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 980
 981         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
 982 }
 983
 984 /*
 985  * get next queue for service
 986  */
 987 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd, int force)
 988 {
 989         unsigned long now = jiffies;
 990         struct cfq_queue *cfqq;
 991
 992         cfqq = cfqd->active_queue;
 993         if (!cfqq)
 994                 goto new_queue;
 995
 996         /*
 997          * slice has expired
 998          */
 999         if (!cfqq->must_dispatch && time_after(jiffies, cfqq->slice_end))
1000                 goto new_queue;
1001
1002         /*
1003          * if queue has requests, dispatch one. if not, check if
1004          * enough slice is left to wait for one
1005          */
1006         if (!RB_EMPTY(&cfqq->sort_list))
1007                 goto keep_queue;
1008         else if (!force && cfq_cfqq_sync(cfqq) &&
1009                  time_before(now, cfqq->slice_end)) {
1010                 if (cfq_arm_slice_timer(cfqd, cfqq))
1011                         return NULL;
1012         }
1013
1014 new_queue:
1015         cfq_slice_expired(cfqd, 0);
1016         cfq_set_active_queue(cfqd);
1017 keep_queue:
1018         return cfqd->active_queue;
1019 }
1020
1021 static int
1022 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1023                         int max_dispatch)
1024 {
1025         int dispatched = 0;
1026
1027         BUG_ON(RB_EMPTY(&cfqq->sort_list));
1028
1029         do {
1030                 struct cfq_rq *crq;
1031
1032                 /*
1033                  * follow expired path, else get first next available
1034                  */
1035                 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1036                         crq = cfqq->next_crq;
1037
1038                 /*
1039                  * finally, insert request into driver dispatch list
1040                  */
1041                 cfq_dispatch_sort(cfqd->queue, crq);
1042
1043                 cfqd->dispatch_slice++;
1044                 dispatched++;
1045
1046                 if (!cfqd->active_cic) {
1047                         atomic_inc(&crq->io_context->ioc->refcount);
1048                         cfqd->active_cic = crq->io_context;
1049                 }
1050
1051                 if (RB_EMPTY(&cfqq->sort_list))
1052                         break;
1053
1054         } while (dispatched < max_dispatch);
1055
1056         /*
1057          * if slice end isn't set yet, set it. if at least one request was
1058          * sync, use the sync time slice value
1059          */
1060         if (!cfqq->slice_end)
1061                 cfq_set_prio_slice(cfqd, cfqq);
1062
1063         /*
1064          * expire an async queue immediately if it has used up its slice. idle
1065          * queue always expire after 1 dispatch round.
1066          */
1067         if ((!cfq_cfqq_sync(cfqq) &&
1068             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1069             cfq_class_idle(cfqq))
1070                 cfq_slice_expired(cfqd, 0);
1071
1072         return dispatched;
1073 }
1074
1075 static int
1076 cfq_dispatch_requests(request_queue_t *q, int max_dispatch, int force)
1077 {
1078         struct cfq_data *cfqd = q->elevator->elevator_data;
1079         struct cfq_queue *cfqq;
1080
1081         if (!cfqd->busy_queues)
1082                 return 0;
1083
1084         cfqq = cfq_select_queue(cfqd, force);
1085         if (cfqq) {
1086                 cfqq->wait_request = 0;
1087                 cfqq->must_dispatch = 0;
1088                 del_timer(&cfqd->idle_slice_timer);
1089
1090                 if (cfq_class_idle(cfqq))
1091                         max_dispatch = 1;
1092
1093                 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1094         }
1095
1096         return 0;
1097 }
1098
1099 static inline void cfq_account_dispatch(struct cfq_rq *crq)
1100 {
1101         struct cfq_queue *cfqq = crq->cfq_queue;
1102         struct cfq_data *cfqd = cfqq->cfqd;
1103
1104         if (unlikely(!blk_fs_request(crq->request)))
1105                 return;
1106
1107         /*
1108          * accounted bit is necessary since some drivers will call
1109          * elv_next_request() many times for the same request (eg ide)
1110          */
1111         if (crq->accounted)
1112                 return;
1113
1114         crq->accounted = 1;
1115         cfqd->rq_in_driver++;
1116 }
1117
1118 static inline void
1119 cfq_account_completion(struct cfq_queue *cfqq, struct cfq_rq *crq)
1120 {
1121         struct cfq_data *cfqd = cfqq->cfqd;
1122         unsigned long now;
1123
1124         if (!crq->accounted)
1125                 return;
1126
1127         now = jiffies;
1128
1129         WARN_ON(!cfqd->rq_in_driver);
1130         cfqd->rq_in_driver--;
1131
1132         if (!cfq_class_idle(cfqq))
1133                 cfqd->last_end_request = now;
1134
1135         if (!cfqq->in_flight && cfqq->on_rr) {
1136                 cfqq->service_last = now;
1137                 cfq_resort_rr_list(cfqq, 0);
1138         }
1139
1140         if (crq->is_sync)
1141                 crq->io_context->last_end_request = now;
1142 }
1143
1144 static struct request *cfq_next_request(request_queue_t *q)
1145 {
1146         struct cfq_data *cfqd = q->elevator->elevator_data;
1147         struct request *rq;
1148
1149         if (!list_empty(&q->queue_head)) {
1150                 struct cfq_rq *crq;
1151 dispatch:
1152                 rq = list_entry_rq(q->queue_head.next);
1153
1154                 crq = RQ_DATA(rq);
1155                 if (crq) {
1156                         /*
1157                          * if idle window is disabled, allow queue buildup
1158                          */
1159                         if (!crq->in_flight && !crq->cfq_queue->idle_window &&
1160                             cfqd->rq_in_driver >= cfqd->cfq_max_depth)
1161                                 return NULL;
1162
1163                         cfq_remove_merge_hints(q, crq);
1164                         cfq_account_dispatch(crq);
1165                 }
1166
1167                 return rq;
1168         }
1169
1170         if (cfq_dispatch_requests(q, cfqd->cfq_quantum, 0))
1171                 goto dispatch;
1172
1173         return NULL;
1174 }
1175
1176 /*
1177  * task holds one reference to the queue, dropped when task exits. each crq
1178  * in-flight on this queue also holds a reference, dropped when crq is freed.
1179  *
1180  * queue lock must be held here.
1181  */
1182 static void cfq_put_queue(struct cfq_queue *cfqq)
1183 {
1184         struct cfq_data *cfqd = cfqq->cfqd;
1185
1186         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1187
1188         if (!atomic_dec_and_test(&cfqq->ref))
1189                 return;
1190
1191         BUG_ON(rb_first(&cfqq->sort_list));
1192         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1193         BUG_ON(cfqq->on_rr);
1194
1195         if (unlikely(cfqd->active_queue == cfqq)) {
1196                 cfq_slice_expired(cfqd, 0);
1197                 kblockd_schedule_work(&cfqd->unplug_work);
1198         }
1199
1200         cfq_put_cfqd(cfqq->cfqd);
1201
1202         /*
1203          * it's on the empty list and still hashed
1204          */
1205         list_del(&cfqq->cfq_list);
1206         hlist_del(&cfqq->cfq_hash);
1207         kmem_cache_free(cfq_pool, cfqq);
1208 }
1209
1210 static inline struct cfq_queue *
1211 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, const int hashval)
1212 {
1213         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1214         struct hlist_node *entry, *next;
1215
1216         hlist_for_each_safe(entry, next, hash_list) {
1217                 struct cfq_queue *__cfqq = list_entry_qhash(entry);
1218
1219                 if (__cfqq->key == key)
1220                         return __cfqq;
1221         }
1222
1223         return NULL;
1224 }
1225
1226 static struct cfq_queue *
1227 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key)
1228 {
1229         return __cfq_find_cfq_hash(cfqd, key, hash_long(key, CFQ_QHASH_SHIFT));
1230 }
1231
1232 static void cfq_free_io_context(struct cfq_io_context *cic)
1233 {
1234         struct cfq_io_context *__cic;
1235         struct list_head *entry, *next;
1236
1237         list_for_each_safe(entry, next, &cic->list) {
1238                 __cic = list_entry(entry, struct cfq_io_context, list);
1239                 kmem_cache_free(cfq_ioc_pool, __cic);
1240         }
1241
1242         kmem_cache_free(cfq_ioc_pool, cic);
1243 }
1244
1245 /*
1246  * Called with interrupts disabled
1247  */
1248 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1249 {
1250         struct cfq_data *cfqd = cic->cfqq->cfqd;
1251         request_queue_t *q = cfqd->queue;
1252
1253         WARN_ON(!irqs_disabled());
1254
1255         spin_lock(q->queue_lock);
1256
1257         if (unlikely(cic->cfqq == cfqd->active_queue)) {
1258                 cfq_slice_expired(cfqd, 0);
1259                 kblockd_schedule_work(&cfqd->unplug_work);
1260         }
1261
1262         cfq_put_queue(cic->cfqq);
1263         cic->cfqq = NULL;
1264         spin_unlock(q->queue_lock);
1265 }
1266
1267 /*
1268  * Another task may update the task cic list, if it is doing a queue lookup
1269  * on its behalf. cfq_cic_lock excludes such concurrent updates
1270  */
1271 static void cfq_exit_io_context(struct cfq_io_context *cic)
1272 {
1273         struct cfq_io_context *__cic;
1274         struct list_head *entry;
1275         unsigned long flags;
1276
1277         local_irq_save(flags);
1278
1279         /*
1280          * put the reference this task is holding to the various queues
1281          */
1282         list_for_each(entry, &cic->list) {
1283                 __cic = list_entry(entry, struct cfq_io_context, list);
1284                 cfq_exit_single_io_context(__cic);
1285         }
1286
1287         cfq_exit_single_io_context(cic);
1288         local_irq_restore(flags);
1289 }
1290
1291 static struct cfq_io_context *
1292 cfq_alloc_io_context(struct cfq_data *cfqd, int gfp_mask)
1293 {
1294         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1295
1296         if (cic) {
1297                 INIT_LIST_HEAD(&cic->list);
1298                 cic->cfqq = NULL;
1299                 cic->key = NULL;
1300                 cic->last_end_request = jiffies;
1301                 cic->ttime_total = 0;
1302                 cic->ttime_samples = 0;
1303                 cic->ttime_mean = 0;
1304                 cic->dtor = cfq_free_io_context;
1305                 cic->exit = cfq_exit_io_context;
1306         }
1307
1308         return cic;
1309 }
1310
1311 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1312 {
1313         struct task_struct *tsk = current;
1314         int ioprio_class;
1315
1316         if (!cfqq->prio_changed)
1317                 return;
1318
1319         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1320         switch (ioprio_class) {
1321                 default:
1322                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1323                 case IOPRIO_CLASS_NONE:
1324                         /*
1325                          * no prio set, place us in the middle of the BE classes
1326                          */
1327                         cfqq->ioprio = task_nice_ioprio(tsk);
1328                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1329                         break;
1330                 case IOPRIO_CLASS_RT:
1331                         cfqq->ioprio = task_ioprio(tsk);
1332                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1333                         break;
1334                 case IOPRIO_CLASS_BE:
1335                         cfqq->ioprio = task_ioprio(tsk);
1336                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1337                         break;
1338                 case IOPRIO_CLASS_IDLE:
1339                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1340                         cfqq->ioprio = 7;
1341                         cfqq->idle_window = 0;
1342                         break;
1343         }
1344
1345         /*
1346          * keep track of original prio settings in case we have to temporarily
1347          * elevate the priority of this queue
1348          */
1349         cfqq->org_ioprio = cfqq->ioprio;
1350         cfqq->org_ioprio_class = cfqq->ioprio_class;
1351
1352         if (cfqq->on_rr)
1353                 cfq_resort_rr_list(cfqq, 0);
1354
1355         cfqq->prio_changed = 0;
1356 }
1357
1358 static inline void changed_ioprio(struct cfq_queue *cfqq)
1359 {
1360         if (cfqq) {
1361                 struct cfq_data *cfqd = cfqq->cfqd;
1362
1363                 spin_lock(cfqd->queue->queue_lock);
1364                 cfqq->prio_changed = 1;
1365                 cfq_init_prio_data(cfqq);
1366                 spin_unlock(cfqd->queue->queue_lock);
1367         }
1368 }
1369
1370 /*
1371  * callback from sys_ioprio_set, irqs are disabled
1372  */
1373 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1374 {
1375         struct cfq_io_context *cic = ioc->cic;
1376
1377         changed_ioprio(cic->cfqq);
1378
1379         list_for_each_entry(cic, &cic->list, list)
1380                 changed_ioprio(cic->cfqq);
1381
1382         return 0;
1383 }
1384
1385 static struct cfq_queue *
1386 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, int gfp_mask)
1387 {
1388         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1389         struct cfq_queue *cfqq, *new_cfqq = NULL;
1390
1391 retry:
1392         cfqq = __cfq_find_cfq_hash(cfqd, key, hashval);
1393
1394         if (!cfqq) {
1395                 if (new_cfqq) {
1396                         cfqq = new_cfqq;
1397                         new_cfqq = NULL;
1398                 } else if (gfp_mask & __GFP_WAIT) {
1399                         spin_unlock_irq(cfqd->queue->queue_lock);
1400                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1401                         spin_lock_irq(cfqd->queue->queue_lock);
1402                         goto retry;
1403                 } else {
1404                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1405                         if (!cfqq)
1406                                 goto out;
1407                 }
1408
1409                 memset(cfqq, 0, sizeof(*cfqq));
1410
1411                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1412                 INIT_LIST_HEAD(&cfqq->cfq_list);
1413                 RB_CLEAR_ROOT(&cfqq->sort_list);
1414                 INIT_LIST_HEAD(&cfqq->fifo);
1415
1416                 cfqq->key = key;
1417                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1418                 atomic_set(&cfqq->ref, 0);
1419                 cfqq->cfqd = cfqd;
1420                 atomic_inc(&cfqd->ref);
1421                 cfqq->service_last = 0;
1422                 /*
1423                  * set ->slice_left to allow preemption for a new process
1424                  */
1425                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1426                 cfqq->idle_window = 1;
1427                 cfqq->ioprio = -1;
1428                 cfqq->ioprio_class = -1;
1429                 cfqq->prio_changed = 1;
1430         }
1431
1432         if (new_cfqq)
1433                 kmem_cache_free(cfq_pool, new_cfqq);
1434
1435         atomic_inc(&cfqq->ref);
1436 out:
1437         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1438         return cfqq;
1439 }
1440
1441 /*
1442  * Setup general io context and cfq io context. There can be several cfq
1443  * io contexts per general io context, if this process is doing io to more
1444  * than one device managed by cfq. Note that caller is holding a reference to
1445  * cfqq, so we don't need to worry about it disappearing
1446  */
1447 static struct cfq_io_context *
1448 cfq_get_io_context(struct cfq_data *cfqd, pid_t pid, int 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);
1456         if (!ioc)
1457                 return NULL;
1458
1459         if ((cic = ioc->cic) == NULL) {
1460                 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1461
1462                 if (cic == NULL)
1463                         goto err;
1464
1465                 /*
1466                  * manually increment generic io_context usage count, it
1467                  * cannot go away since we are already holding one ref to it
1468                  */
1469                 ioc->cic = cic;
1470                 ioc->set_ioprio = cfq_ioc_set_ioprio;
1471                 cic->ioc = ioc;
1472                 cic->key = cfqd;
1473                 atomic_inc(&cfqd->ref);
1474         } else {
1475                 struct cfq_io_context *__cic;
1476
1477                 /*
1478                  * the first cic on the list is actually the head itself
1479                  */
1480                 if (cic->key == cfqd)
1481                         goto out;
1482
1483                 /*
1484                  * cic exists, check if we already are there. linear search
1485                  * should be ok here, the list will usually not be more than
1486                  * 1 or a few entries long
1487                  */
1488                 list_for_each_entry(__cic, &cic->list, list) {
1489                         /*
1490                          * this process is already holding a reference to
1491                          * this queue, so no need to get one more
1492                          */
1493                         if (__cic->key == cfqd) {
1494                                 cic = __cic;
1495                                 goto out;
1496                         }
1497                 }
1498
1499                 /*
1500                  * nope, process doesn't have a cic assoicated with this
1501                  * cfqq yet. get a new one and add to list
1502                  */
1503                 __cic = cfq_alloc_io_context(cfqd, gfp_mask);
1504                 if (__cic == NULL)
1505                         goto err;
1506
1507                 __cic->ioc = ioc;
1508                 __cic->key = cfqd;
1509                 atomic_inc(&cfqd->ref);
1510                 list_add(&__cic->list, &cic->list);
1511                 cic = __cic;
1512         }
1513
1514 out:
1515         return cic;
1516 err:
1517         put_io_context(ioc);
1518         return NULL;
1519 }
1520
1521 static void
1522 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1523 {
1524         unsigned long elapsed, ttime;
1525
1526         /*
1527          * if this context already has stuff queued, thinktime is from
1528          * last queue not last end
1529          */
1530 #if 0
1531         if (time_after(cic->last_end_request, cic->last_queue))
1532                 elapsed = jiffies - cic->last_end_request;
1533         else
1534                 elapsed = jiffies - cic->last_queue;
1535 #else
1536                 elapsed = jiffies - cic->last_end_request;
1537 #endif
1538
1539         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1540
1541         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1542         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1543         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1544 }
1545
1546 #define sample_valid(samples)   ((samples) > 80)
1547
1548 /*
1549  * Disable idle window if the process thinks too long or seeks so much that
1550  * it doesn't matter
1551  */
1552 static void
1553 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1554                        struct cfq_io_context *cic)
1555 {
1556         int enable_idle = cfqq->idle_window;
1557
1558         if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1559                 enable_idle = 0;
1560         else if (sample_valid(cic->ttime_samples)) {
1561                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1562                         enable_idle = 0;
1563                 else
1564                         enable_idle = 1;
1565         }
1566
1567         cfqq->idle_window = enable_idle;
1568 }
1569
1570
1571 /*
1572  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1573  * no or if we aren't sure, a 1 will cause a preempt.
1574  */
1575 static int
1576 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1577                    struct cfq_rq *crq)
1578 {
1579         struct cfq_queue *cfqq = cfqd->active_queue;
1580
1581         if (cfq_class_idle(new_cfqq))
1582                 return 0;
1583
1584         if (!cfqq)
1585                 return 1;
1586
1587         if (cfq_class_idle(cfqq))
1588                 return 1;
1589         if (!new_cfqq->wait_request)
1590                 return 0;
1591         /*
1592          * if it doesn't have slice left, forget it
1593          */
1594         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1595                 return 0;
1596         if (crq->is_sync && !cfq_cfqq_sync(cfqq))
1597                 return 1;
1598
1599         return 0;
1600 }
1601
1602 /*
1603  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1604  * let it have half of its nominal slice.
1605  */
1606 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1607 {
1608         struct cfq_queue *__cfqq, *next;
1609
1610         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1611                 cfq_resort_rr_list(__cfqq, 1);
1612
1613         if (!cfqq->slice_left)
1614                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1615
1616         cfqq->slice_end = cfqq->slice_left + jiffies;
1617         cfq_slice_expired(cfqd, 1);
1618         __cfq_set_active_queue(cfqd, cfqq);
1619 }
1620
1621 /*
1622  * should really be a ll_rw_blk.c helper
1623  */
1624 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1625 {
1626         request_queue_t *q = cfqd->queue;
1627
1628         if (!blk_queue_plugged(q))
1629                 q->request_fn(q);
1630         else
1631                 __generic_unplug_device(q);
1632 }
1633
1634 /*
1635  * Called when a new fs request (crq) is added (to cfqq). Check if there's
1636  * something we should do about it
1637  */
1638 static void
1639 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1640                  struct cfq_rq *crq)
1641 {
1642         const int sync = crq->is_sync;
1643
1644         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1645
1646         if (sync) {
1647                 struct cfq_io_context *cic = crq->io_context;
1648
1649                 cfq_update_io_thinktime(cfqd, cic);
1650                 cfq_update_idle_window(cfqd, cfqq, cic);
1651
1652                 cic->last_queue = jiffies;
1653         }
1654
1655         if (cfqq == cfqd->active_queue) {
1656                 /*
1657                  * if we are waiting for a request for this queue, let it rip
1658                  * immediately and flag that we must not expire this queue
1659                  * just now
1660                  */
1661                 if (cfqq->wait_request) {
1662                         cfqq->must_dispatch = 1;
1663                         del_timer(&cfqd->idle_slice_timer);
1664                         cfq_start_queueing(cfqd, cfqq);
1665                 }
1666         } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1667                 /*
1668                  * not the active queue - expire current slice if it is
1669                  * idle and has expired it's mean thinktime or this new queue
1670                  * has some old slice time left and is of higher priority
1671                  */
1672                 cfq_preempt_queue(cfqd, cfqq);
1673                 cfqq->must_dispatch = 1;
1674                 cfq_start_queueing(cfqd, cfqq);
1675         }
1676 }
1677
1678 static void cfq_enqueue(struct cfq_data *cfqd, struct request *rq)
1679 {
1680         struct cfq_rq *crq = RQ_DATA(rq);
1681         struct cfq_queue *cfqq = crq->cfq_queue;
1682
1683         cfq_init_prio_data(cfqq);
1684
1685         cfq_add_crq_rb(crq);
1686
1687         list_add_tail(&rq->queuelist, &cfqq->fifo);
1688
1689         if (rq_mergeable(rq)) {
1690                 cfq_add_crq_hash(cfqd, crq);
1691
1692                 if (!cfqd->queue->last_merge)
1693                         cfqd->queue->last_merge = rq;
1694         }
1695
1696         cfq_crq_enqueued(cfqd, cfqq, crq);
1697 }
1698
1699 static void
1700 cfq_insert_request(request_queue_t *q, struct request *rq, int where)
1701 {
1702         struct cfq_data *cfqd = q->elevator->elevator_data;
1703
1704         switch (where) {
1705                 case ELEVATOR_INSERT_BACK:
1706                         while (cfq_dispatch_requests(q, INT_MAX, 1))
1707                                 ;
1708                         list_add_tail(&rq->queuelist, &q->queue_head);
1709                         /*
1710                          * If we were idling with pending requests on
1711                          * inactive cfqqs, force dispatching will
1712                          * remove the idle timer and the queue won't
1713                          * be kicked by __make_request() afterward.
1714                          * Kick it here.
1715                          */
1716                         kblockd_schedule_work(&cfqd->unplug_work);
1717                         break;
1718                 case ELEVATOR_INSERT_FRONT:
1719                         list_add(&rq->queuelist, &q->queue_head);
1720                         break;
1721                 case ELEVATOR_INSERT_SORT:
1722                         BUG_ON(!blk_fs_request(rq));
1723                         cfq_enqueue(cfqd, rq);
1724                         break;
1725                 default:
1726                         printk("%s: bad insert point %d\n", __FUNCTION__,where);
1727                         return;
1728         }
1729 }
1730
1731 static inline int cfq_pending_requests(struct cfq_data *cfqd)
1732 {
1733         return !list_empty(&cfqd->queue->queue_head) || cfqd->busy_queues;
1734 }
1735
1736 static int cfq_queue_empty(request_queue_t *q)
1737 {
1738         struct cfq_data *cfqd = q->elevator->elevator_data;
1739
1740         return !cfq_pending_requests(cfqd);
1741 }
1742
1743 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1744 {
1745         struct cfq_rq *crq = RQ_DATA(rq);
1746         struct cfq_queue *cfqq;
1747
1748         if (unlikely(!blk_fs_request(rq)))
1749                 return;
1750
1751         cfqq = crq->cfq_queue;
1752
1753         if (crq->in_flight) {
1754                 WARN_ON(!cfqq->in_flight);
1755                 cfqq->in_flight--;
1756         }
1757
1758         cfq_account_completion(cfqq, crq);
1759 }
1760
1761 static struct request *
1762 cfq_former_request(request_queue_t *q, struct request *rq)
1763 {
1764         struct cfq_rq *crq = RQ_DATA(rq);
1765         struct rb_node *rbprev = rb_prev(&crq->rb_node);
1766
1767         if (rbprev)
1768                 return rb_entry_crq(rbprev)->request;
1769
1770         return NULL;
1771 }
1772
1773 static struct request *
1774 cfq_latter_request(request_queue_t *q, struct request *rq)
1775 {
1776         struct cfq_rq *crq = RQ_DATA(rq);
1777         struct rb_node *rbnext = rb_next(&crq->rb_node);
1778
1779         if (rbnext)
1780                 return rb_entry_crq(rbnext)->request;
1781
1782         return NULL;
1783 }
1784
1785 /*
1786  * we temporarily boost lower priority queues if they are holding fs exclusive
1787  * resources. they are boosted to normal prio (CLASS_BE/4)
1788  */
1789 static void cfq_prio_boost(struct cfq_queue *cfqq)
1790 {
1791         const int ioprio_class = cfqq->ioprio_class;
1792         const int ioprio = cfqq->ioprio;
1793
1794         if (has_fs_excl()) {
1795                 /*
1796                  * boost idle prio on transactions that would lock out other
1797                  * users of the filesystem
1798                  */
1799                 if (cfq_class_idle(cfqq))
1800                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1801                 if (cfqq->ioprio > IOPRIO_NORM)
1802                         cfqq->ioprio = IOPRIO_NORM;
1803         } else {
1804                 /*
1805                  * check if we need to unboost the queue
1806                  */
1807                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1808                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1809                 if (cfqq->ioprio != cfqq->org_ioprio)
1810                         cfqq->ioprio = cfqq->org_ioprio;
1811         }
1812
1813         /*
1814          * refile between round-robin lists if we moved the priority class
1815          */
1816         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1817             cfqq->on_rr)
1818                 cfq_resort_rr_list(cfqq, 0);
1819 }
1820
1821 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
1822 {
1823         if (rw == READ || process_sync(task))
1824                 return task->pid;
1825
1826         return CFQ_KEY_ASYNC;
1827 }
1828
1829 static inline int
1830 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1831                 struct task_struct *task, int rw)
1832 {
1833         if (cfqq->wait_request && cfqq->must_alloc)
1834                 return ELV_MQUEUE_MUST;
1835
1836         return ELV_MQUEUE_MAY;
1837 #if 0
1838         if (!cfqq || task->flags & PF_MEMALLOC)
1839                 return ELV_MQUEUE_MAY;
1840         if (!cfqq->allocated[rw] || cfqq->must_alloc) {
1841                 if (cfqq->wait_request)
1842                         return ELV_MQUEUE_MUST;
1843
1844                 /*
1845                  * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1846                  * can quickly flood the queue with writes from a single task
1847                  */
1848                 if (rw == READ || !cfqq->must_alloc_slice) {
1849                         cfqq->must_alloc_slice = 1;
1850                         return ELV_MQUEUE_MUST;
1851                 }
1852
1853                 return ELV_MQUEUE_MAY;
1854         }
1855         if (cfq_class_idle(cfqq))
1856                 return ELV_MQUEUE_NO;
1857         if (cfqq->allocated[rw] >= cfqd->max_queued) {
1858                 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1859                 int ret = ELV_MQUEUE_NO;
1860
1861                 if (ioc && ioc->nr_batch_requests)
1862                         ret = ELV_MQUEUE_MAY;
1863
1864                 put_io_context(ioc);
1865                 return ret;
1866         }
1867
1868         return ELV_MQUEUE_MAY;
1869 #endif
1870 }
1871
1872 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1873 {
1874         struct cfq_data *cfqd = q->elevator->elevator_data;
1875         struct task_struct *tsk = current;
1876         struct cfq_queue *cfqq;
1877
1878         /*
1879          * don't force setup of a queue from here, as a call to may_queue
1880          * does not necessarily imply that a request actually will be queued.
1881          * so just lookup a possibly existing queue, or return 'may queue'
1882          * if that fails
1883          */
1884         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw));
1885         if (cfqq) {
1886                 cfq_init_prio_data(cfqq);
1887                 cfq_prio_boost(cfqq);
1888
1889                 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1890         }
1891
1892         return ELV_MQUEUE_MAY;
1893 }
1894
1895 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1896 {
1897         struct cfq_data *cfqd = q->elevator->elevator_data;
1898         struct request_list *rl = &q->rq;
1899
1900         if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1901                 smp_mb();
1902                 if (waitqueue_active(&rl->wait[READ]))
1903                         wake_up(&rl->wait[READ]);
1904         }
1905
1906         if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
1907                 smp_mb();
1908                 if (waitqueue_active(&rl->wait[WRITE]))
1909                         wake_up(&rl->wait[WRITE]);
1910         }
1911 }
1912
1913 /*
1914  * queue lock held here
1915  */
1916 static void cfq_put_request(request_queue_t *q, struct request *rq)
1917 {
1918         struct cfq_data *cfqd = q->elevator->elevator_data;
1919         struct cfq_rq *crq = RQ_DATA(rq);
1920
1921         if (crq) {
1922                 struct cfq_queue *cfqq = crq->cfq_queue;
1923                 const int rw = rq_data_dir(rq);
1924
1925                 BUG_ON(!cfqq->allocated[rw]);
1926                 cfqq->allocated[rw]--;
1927
1928                 put_io_context(crq->io_context->ioc);
1929
1930                 mempool_free(crq, cfqd->crq_pool);
1931                 rq->elevator_private = NULL;
1932
1933                 cfq_check_waiters(q, cfqq);
1934                 cfq_put_queue(cfqq);
1935         }
1936 }
1937
1938 /*
1939  * Allocate cfq data structures associated with this request.
1940  */
1941 static int
1942 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
1943                 int gfp_mask)
1944 {
1945         struct cfq_data *cfqd = q->elevator->elevator_data;
1946         struct cfq_io_context *cic;
1947         const int rw = rq_data_dir(rq);
1948         struct cfq_queue *cfqq;
1949         struct cfq_rq *crq;
1950         unsigned long flags;
1951
1952         might_sleep_if(gfp_mask & __GFP_WAIT);
1953
1954         cic = cfq_get_io_context(cfqd, cfq_queue_pid(current, rw), gfp_mask);
1955
1956         spin_lock_irqsave(q->queue_lock, flags);
1957
1958         if (!cic)
1959                 goto queue_fail;
1960
1961         if (!cic->cfqq) {
1962                 cfqq = cfq_get_queue(cfqd, current->pid, gfp_mask);
1963                 if (!cfqq)
1964                         goto queue_fail;
1965
1966                 cic->cfqq = cfqq;
1967         } else
1968                 cfqq = cic->cfqq;
1969
1970         cfqq->allocated[rw]++;
1971         cfqq->must_alloc = 0;
1972         cfqd->rq_starved = 0;
1973         atomic_inc(&cfqq->ref);
1974         spin_unlock_irqrestore(q->queue_lock, flags);
1975
1976         crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
1977         if (crq) {
1978                 RB_CLEAR(&crq->rb_node);
1979                 crq->rb_key = 0;
1980                 crq->request = rq;
1981                 INIT_HLIST_NODE(&crq->hash);
1982                 crq->cfq_queue = cfqq;
1983                 crq->io_context = cic;
1984                 crq->in_flight = crq->accounted = 0;
1985                 crq->is_sync = (rw == READ || process_sync(current));
1986                 crq->requeued = 0;
1987                 rq->elevator_private = crq;
1988                 return 0;
1989         }
1990
1991         spin_lock_irqsave(q->queue_lock, flags);
1992         cfqq->allocated[rw]--;
1993         if (!(cfqq->allocated[0] + cfqq->allocated[1]))
1994                 cfqq->must_alloc = 1;
1995         cfq_put_queue(cfqq);
1996 queue_fail:
1997         if (cic)
1998                 put_io_context(cic->ioc);
1999         /*
2000          * mark us rq allocation starved. we need to kickstart the process
2001          * ourselves if there are no pending requests that can do it for us.
2002          * that would be an extremely rare OOM situation
2003          */
2004         cfqd->rq_starved = 1;
2005         kblockd_schedule_work(&cfqd->unplug_work);
2006         spin_unlock_irqrestore(q->queue_lock, flags);
2007         return 1;
2008 }
2009
2010 static void cfq_kick_queue(void *data)
2011 {
2012         request_queue_t *q = data;
2013         struct cfq_data *cfqd = q->elevator->elevator_data;
2014         unsigned long flags;
2015
2016         spin_lock_irqsave(q->queue_lock, flags);
2017
2018         if (cfqd->rq_starved) {
2019                 struct request_list *rl = &q->rq;
2020
2021                 /*
2022                  * we aren't guaranteed to get a request after this, but we
2023                  * have to be opportunistic
2024                  */
2025                 smp_mb();
2026                 if (waitqueue_active(&rl->wait[READ]))
2027                         wake_up(&rl->wait[READ]);
2028                 if (waitqueue_active(&rl->wait[WRITE]))
2029                         wake_up(&rl->wait[WRITE]);
2030         }
2031
2032         blk_remove_plug(q);
2033         q->request_fn(q);
2034         spin_unlock_irqrestore(q->queue_lock, flags);
2035 }
2036
2037 /*
2038  * Timer running if the active_queue is currently idling inside its time slice
2039  */
2040 static void cfq_idle_slice_timer(unsigned long data)
2041 {
2042         struct cfq_data *cfqd = (struct cfq_data *) data;
2043         struct cfq_queue *cfqq;
2044         unsigned long flags;
2045
2046         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2047
2048         if ((cfqq = cfqd->active_queue) != NULL) {
2049                 unsigned long now = jiffies;
2050
2051                 /*
2052                  * expired
2053                  */
2054                 if (time_after(now, cfqq->slice_end))
2055                         goto expire;
2056
2057                 /*
2058                  * only expire and reinvoke request handler, if there are
2059                  * other queues with pending requests
2060                  */
2061                 if (!cfq_pending_requests(cfqd)) {
2062                         cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2063                         add_timer(&cfqd->idle_slice_timer);
2064                         goto out_cont;
2065                 }
2066
2067                 /*
2068                  * not expired and it has a request pending, let it dispatch
2069                  */
2070                 if (!RB_EMPTY(&cfqq->sort_list)) {
2071                         cfqq->must_dispatch = 1;
2072                         goto out_kick;
2073                 }
2074         }
2075 expire:
2076         cfq_slice_expired(cfqd, 0);
2077 out_kick:
2078         if (cfq_pending_requests(cfqd))
2079                 kblockd_schedule_work(&cfqd->unplug_work);
2080 out_cont:
2081         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2082 }
2083
2084 /*
2085  * Timer running if an idle class queue is waiting for service
2086  */
2087 static void cfq_idle_class_timer(unsigned long data)
2088 {
2089         struct cfq_data *cfqd = (struct cfq_data *) data;
2090         unsigned long flags, end;
2091
2092         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2093
2094         /*
2095          * race with a non-idle queue, reset timer
2096          */
2097         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2098         if (!time_after_eq(jiffies, end)) {
2099                 cfqd->idle_class_timer.expires = end;
2100                 add_timer(&cfqd->idle_class_timer);
2101         } else
2102                 kblockd_schedule_work(&cfqd->unplug_work);
2103
2104         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2105 }
2106
2107
2108 static void cfq_put_cfqd(struct cfq_data *cfqd)
2109 {
2110         request_queue_t *q = cfqd->queue;
2111
2112         if (!atomic_dec_and_test(&cfqd->ref))
2113                 return;
2114
2115         blk_sync_queue(q);
2116
2117         blk_put_queue(q);
2118
2119         mempool_destroy(cfqd->crq_pool);
2120         kfree(cfqd->crq_hash);
2121         kfree(cfqd->cfq_hash);
2122         kfree(cfqd);
2123 }
2124
2125 static void cfq_exit_queue(elevator_t *e)
2126 {
2127         struct cfq_data *cfqd = e->elevator_data;
2128
2129         del_timer_sync(&cfqd->idle_slice_timer);
2130         del_timer_sync(&cfqd->idle_class_timer);
2131         cfq_put_cfqd(cfqd);
2132 }
2133
2134 static int cfq_init_queue(request_queue_t *q, elevator_t *e)
2135 {
2136         struct cfq_data *cfqd;
2137         int i;
2138
2139         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2140         if (!cfqd)
2141                 return -ENOMEM;
2142
2143         memset(cfqd, 0, sizeof(*cfqd));
2144
2145         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2146                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2147
2148         INIT_LIST_HEAD(&cfqd->busy_rr);
2149         INIT_LIST_HEAD(&cfqd->cur_rr);
2150         INIT_LIST_HEAD(&cfqd->idle_rr);
2151         INIT_LIST_HEAD(&cfqd->empty_list);
2152
2153         cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2154         if (!cfqd->crq_hash)
2155                 goto out_crqhash;
2156
2157         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2158         if (!cfqd->cfq_hash)
2159                 goto out_cfqhash;
2160
2161         cfqd->crq_pool = mempool_create(BLKDEV_MIN_RQ, mempool_alloc_slab, mempool_free_slab, crq_pool);
2162         if (!cfqd->crq_pool)
2163                 goto out_crqpool;
2164
2165         for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2166                 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2167         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2168                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2169
2170         e->elevator_data = cfqd;
2171
2172         cfqd->queue = q;
2173         atomic_inc(&q->refcnt);
2174
2175         cfqd->max_queued = q->nr_requests / 4;
2176         q->nr_batching = cfq_queued;
2177
2178         init_timer(&cfqd->idle_slice_timer);
2179         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2180         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2181
2182         init_timer(&cfqd->idle_class_timer);
2183         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2184         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2185
2186         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2187
2188         atomic_set(&cfqd->ref, 1);
2189
2190         cfqd->cfq_queued = cfq_queued;
2191         cfqd->cfq_quantum = cfq_quantum;
2192         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2193         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2194         cfqd->cfq_back_max = cfq_back_max;
2195         cfqd->cfq_back_penalty = cfq_back_penalty;
2196         cfqd->cfq_slice[0] = cfq_slice_async;
2197         cfqd->cfq_slice[1] = cfq_slice_sync;
2198         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2199         cfqd->cfq_slice_idle = cfq_slice_idle;
2200         cfqd->cfq_max_depth = cfq_max_depth;
2201         return 0;
2202 out_crqpool:
2203         kfree(cfqd->cfq_hash);
2204 out_cfqhash:
2205         kfree(cfqd->crq_hash);
2206 out_crqhash:
2207         kfree(cfqd);
2208         return -ENOMEM;
2209 }
2210
2211 static void cfq_slab_kill(void)
2212 {
2213         if (crq_pool)
2214                 kmem_cache_destroy(crq_pool);
2215         if (cfq_pool)
2216                 kmem_cache_destroy(cfq_pool);
2217         if (cfq_ioc_pool)
2218                 kmem_cache_destroy(cfq_ioc_pool);
2219 }
2220
2221 static int __init cfq_slab_setup(void)
2222 {
2223         crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2224                                         NULL, NULL);
2225         if (!crq_pool)
2226                 goto fail;
2227
2228         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2229                                         NULL, NULL);
2230         if (!cfq_pool)
2231                 goto fail;
2232
2233         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2234                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2235         if (!cfq_ioc_pool)
2236                 goto fail;
2237
2238         return 0;
2239 fail:
2240         cfq_slab_kill();
2241         return -ENOMEM;
2242 }
2243
2244 /*
2245  * sysfs parts below -->
2246  */
2247 struct cfq_fs_entry {
2248         struct attribute attr;
2249         ssize_t (*show)(struct cfq_data *, char *);
2250         ssize_t (*store)(struct cfq_data *, const char *, size_t);
2251 };
2252
2253 static ssize_t
2254 cfq_var_show(unsigned int var, char *page)
2255 {
2256         return sprintf(page, "%d\n", var);
2257 }
2258
2259 static ssize_t
2260 cfq_var_store(unsigned int *var, const char *page, size_t count)
2261 {
2262         char *p = (char *) page;
2263
2264         *var = simple_strtoul(p, &p, 10);
2265         return count;
2266 }
2267
2268 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2269 static ssize_t __FUNC(struct cfq_data *cfqd, char *page)                \
2270 {                                                                       \
2271         unsigned int __data = __VAR;                                    \
2272         if (__CONV)                                                     \
2273                 __data = jiffies_to_msecs(__data);                      \
2274         return cfq_var_show(__data, (page));                            \
2275 }
2276 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2277 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2278 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2279 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2280 SHOW_FUNCTION(cfq_back_max_show, cfqd->cfq_back_max, 0);
2281 SHOW_FUNCTION(cfq_back_penalty_show, cfqd->cfq_back_penalty, 0);
2282 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2283 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2284 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2285 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2286 SHOW_FUNCTION(cfq_max_depth_show, cfqd->cfq_max_depth, 0);
2287 #undef SHOW_FUNCTION
2288
2289 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2290 static ssize_t __FUNC(struct cfq_data *cfqd, const char *page, size_t count)    \
2291 {                                                                       \
2292         unsigned int __data;                                            \
2293         int ret = cfq_var_store(&__data, (page), count);                \
2294         if (__data < (MIN))                                             \
2295                 __data = (MIN);                                         \
2296         else if (__data > (MAX))                                        \
2297                 __data = (MAX);                                         \
2298         if (__CONV)                                                     \
2299                 *(__PTR) = msecs_to_jiffies(__data);                    \
2300         else                                                            \
2301                 *(__PTR) = __data;                                      \
2302         return ret;                                                     \
2303 }
2304 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2305 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2306 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2307 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2308 STORE_FUNCTION(cfq_back_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2309 STORE_FUNCTION(cfq_back_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2310 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2311 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2312 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2313 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2314 STORE_FUNCTION(cfq_max_depth_store, &cfqd->cfq_max_depth, 1, UINT_MAX, 0);
2315 #undef STORE_FUNCTION
2316
2317 static struct cfq_fs_entry cfq_quantum_entry = {
2318         .attr = {.name = "quantum", .mode = S_IRUGO | S_IWUSR },
2319         .show = cfq_quantum_show,
2320         .store = cfq_quantum_store,
2321 };
2322 static struct cfq_fs_entry cfq_queued_entry = {
2323         .attr = {.name = "queued", .mode = S_IRUGO | S_IWUSR },
2324         .show = cfq_queued_show,
2325         .store = cfq_queued_store,
2326 };
2327 static struct cfq_fs_entry cfq_fifo_expire_sync_entry = {
2328         .attr = {.name = "fifo_expire_sync", .mode = S_IRUGO | S_IWUSR },
2329         .show = cfq_fifo_expire_sync_show,
2330         .store = cfq_fifo_expire_sync_store,
2331 };
2332 static struct cfq_fs_entry cfq_fifo_expire_async_entry = {
2333         .attr = {.name = "fifo_expire_async", .mode = S_IRUGO | S_IWUSR },
2334         .show = cfq_fifo_expire_async_show,
2335         .store = cfq_fifo_expire_async_store,
2336 };
2337 static struct cfq_fs_entry cfq_back_max_entry = {
2338         .attr = {.name = "back_seek_max", .mode = S_IRUGO | S_IWUSR },
2339         .show = cfq_back_max_show,
2340         .store = cfq_back_max_store,
2341 };
2342 static struct cfq_fs_entry cfq_back_penalty_entry = {
2343         .attr = {.name = "back_seek_penalty", .mode = S_IRUGO | S_IWUSR },
2344         .show = cfq_back_penalty_show,
2345         .store = cfq_back_penalty_store,
2346 };
2347 static struct cfq_fs_entry cfq_slice_sync_entry = {
2348         .attr = {.name = "slice_sync", .mode = S_IRUGO | S_IWUSR },
2349         .show = cfq_slice_sync_show,
2350         .store = cfq_slice_sync_store,
2351 };
2352 static struct cfq_fs_entry cfq_slice_async_entry = {
2353         .attr = {.name = "slice_async", .mode = S_IRUGO | S_IWUSR },
2354         .show = cfq_slice_async_show,
2355         .store = cfq_slice_async_store,
2356 };
2357 static struct cfq_fs_entry cfq_slice_async_rq_entry = {
2358         .attr = {.name = "slice_async_rq", .mode = S_IRUGO | S_IWUSR },
2359         .show = cfq_slice_async_rq_show,
2360         .store = cfq_slice_async_rq_store,
2361 };
2362 static struct cfq_fs_entry cfq_slice_idle_entry = {
2363         .attr = {.name = "slice_idle", .mode = S_IRUGO | S_IWUSR },
2364         .show = cfq_slice_idle_show,
2365         .store = cfq_slice_idle_store,
2366 };
2367 static struct cfq_fs_entry cfq_max_depth_entry = {
2368         .attr = {.name = "max_depth", .mode = S_IRUGO | S_IWUSR },
2369         .show = cfq_max_depth_show,
2370         .store = cfq_max_depth_store,
2371 };
2372 static struct attribute *default_attrs[] = {
2373         &cfq_quantum_entry.attr,
2374         &cfq_queued_entry.attr,
2375         &cfq_fifo_expire_sync_entry.attr,
2376         &cfq_fifo_expire_async_entry.attr,
2377         &cfq_back_max_entry.attr,
2378         &cfq_back_penalty_entry.attr,
2379         &cfq_slice_sync_entry.attr,
2380         &cfq_slice_async_entry.attr,
2381         &cfq_slice_async_rq_entry.attr,
2382         &cfq_slice_idle_entry.attr,
2383         &cfq_max_depth_entry.attr,
2384         NULL,
2385 };
2386
2387 #define to_cfq(atr) container_of((atr), struct cfq_fs_entry, attr)
2388
2389 static ssize_t
2390 cfq_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
2391 {
2392         elevator_t *e = container_of(kobj, elevator_t, kobj);
2393         struct cfq_fs_entry *entry = to_cfq(attr);
2394
2395         if (!entry->show)
2396                 return -EIO;
2397
2398         return entry->show(e->elevator_data, page);
2399 }
2400
2401 static ssize_t
2402 cfq_attr_store(struct kobject *kobj, struct attribute *attr,
2403                const char *page, size_t length)
2404 {
2405         elevator_t *e = container_of(kobj, elevator_t, kobj);
2406         struct cfq_fs_entry *entry = to_cfq(attr);
2407
2408         if (!entry->store)
2409                 return -EIO;
2410
2411         return entry->store(e->elevator_data, page, length);
2412 }
2413
2414 static struct sysfs_ops cfq_sysfs_ops = {
2415         .show   = cfq_attr_show,
2416         .store  = cfq_attr_store,
2417 };
2418
2419 static struct kobj_type cfq_ktype = {
2420         .sysfs_ops      = &cfq_sysfs_ops,
2421         .default_attrs  = default_attrs,
2422 };
2423
2424 static struct elevator_type iosched_cfq = {
2425         .ops = {
2426                 .elevator_merge_fn =            cfq_merge,
2427                 .elevator_merged_fn =           cfq_merged_request,
2428                 .elevator_merge_req_fn =        cfq_merged_requests,
2429                 .elevator_next_req_fn =         cfq_next_request,
2430                 .elevator_add_req_fn =          cfq_insert_request,
2431                 .elevator_remove_req_fn =       cfq_remove_request,
2432                 .elevator_requeue_req_fn =      cfq_requeue_request,
2433                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2434                 .elevator_queue_empty_fn =      cfq_queue_empty,
2435                 .elevator_completed_req_fn =    cfq_completed_request,
2436                 .elevator_former_req_fn =       cfq_former_request,
2437                 .elevator_latter_req_fn =       cfq_latter_request,
2438                 .elevator_set_req_fn =          cfq_set_request,
2439                 .elevator_put_req_fn =          cfq_put_request,
2440                 .elevator_may_queue_fn =        cfq_may_queue,
2441                 .elevator_init_fn =             cfq_init_queue,
2442                 .elevator_exit_fn =             cfq_exit_queue,
2443         },
2444         .elevator_ktype =       &cfq_ktype,
2445         .elevator_name =        "cfq",
2446         .elevator_owner =       THIS_MODULE,
2447 };
2448
2449 static int __init cfq_init(void)
2450 {
2451         int ret;
2452
2453         /*
2454          * could be 0 on HZ < 1000 setups
2455          */
2456         if (!cfq_slice_async)
2457                 cfq_slice_async = 1;
2458         if (!cfq_slice_idle)
2459                 cfq_slice_idle = 1;
2460
2461         if (cfq_slab_setup())
2462                 return -ENOMEM;
2463
2464         ret = elv_register(&iosched_cfq);
2465         if (ret)
2466                 cfq_slab_kill();
2467
2468         return ret;
2469 }
2470
2471 static void __exit cfq_exit(void)
2472 {
2473         struct task_struct *g, *p;
2474         unsigned long flags;
2475
2476         read_lock_irqsave(&tasklist_lock, flags);
2477
2478         /*
2479          * iterate each process in the system, removing our io_context
2480          */
2481         do_each_thread(g, p) {
2482                 struct io_context *ioc = p->io_context;
2483
2484                 if (ioc && ioc->cic) {
2485                         ioc->cic->exit(ioc->cic);
2486                         cfq_free_io_context(ioc->cic);
2487                         ioc->cic = NULL;
2488                 }
2489         } while_each_thread(g, p);
2490
2491         read_unlock_irqrestore(&tasklist_lock, flags);
2492
2493         cfq_slab_kill();
2494         elv_unregister(&iosched_cfq);
2495 }
2496
2497 module_init(cfq_init);
2498 module_exit(cfq_exit);
2499
2500 MODULE_AUTHOR("Jens Axboe");
2501 MODULE_LICENSE("GPL");
2502 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");